net: mvneta: Statically assign queues to CPUs
authorMaxime Ripard <maxime.ripard@free-electrons.com>
Fri, 25 Sep 2015 16:09:38 +0000 (18:09 +0200)
committerDavid S. Miller <davem@davemloft.net>
Tue, 29 Sep 2015 18:51:41 +0000 (11:51 -0700)
Since the switch to per-CPU interrupts, we lost the ability to set which
CPU was going to receive our RX interrupt, which was now only the CPU on
which the mvneta_open function was run.

We can now assign our queues to their respective CPUs, and make sure only
this CPU is going to handle our traffic.

This also paves the road to be able to change that at runtime, and later on
to support RSS.

[gregory.clement@free-electrons.com]: hardened the CPU hotplug support.

Signed-off-by: Maxime Ripard <maxime.ripard@free-electrons.com>
Signed-off-by: Gregory CLEMENT <gregory.clement@free-electrons.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/marvell/mvneta.c

index 0702beb..dd6fe94 100644 (file)
@@ -32,6 +32,7 @@
 #include <linux/of_address.h>
 #include <linux/phy.h>
 #include <linux/clk.h>
+#include <linux/cpu.h>
 
 /* Registers */
 #define MVNETA_RXQ_CONFIG_REG(q)                (0x1400 + ((q) << 2))
@@ -306,6 +307,7 @@ struct mvneta_port {
        struct mvneta_rx_queue *rxqs;
        struct mvneta_tx_queue *txqs;
        struct net_device *dev;
+       struct notifier_block cpu_notifier;
 
        /* Core clock */
        struct clk *clk;
@@ -2057,7 +2059,6 @@ static irqreturn_t mvneta_isr(int irq, void *dev_id)
        struct mvneta_pcpu_port *port = (struct mvneta_pcpu_port *)dev_id;
 
        disable_percpu_irq(port->pp->dev->irq);
-
        napi_schedule(&port->napi);
 
        return IRQ_HANDLED;
@@ -2658,6 +2659,125 @@ static void mvneta_mdio_remove(struct mvneta_port *pp)
        pp->phy_dev = NULL;
 }
 
+static void mvneta_percpu_enable(void *arg)
+{
+       struct mvneta_port *pp = arg;
+
+       enable_percpu_irq(pp->dev->irq, IRQ_TYPE_NONE);
+}
+
+static void mvneta_percpu_disable(void *arg)
+{
+       struct mvneta_port *pp = arg;
+
+       disable_percpu_irq(pp->dev->irq);
+}
+
+static void mvneta_percpu_elect(struct mvneta_port *pp)
+{
+       int online_cpu_idx, cpu, i = 0;
+
+       online_cpu_idx = rxq_def % num_online_cpus();
+
+       for_each_online_cpu(cpu) {
+               if (i == online_cpu_idx)
+                       /* Enable per-CPU interrupt on the one CPU we
+                        * just elected
+                        */
+                       smp_call_function_single(cpu, mvneta_percpu_enable,
+                                               pp, true);
+               else
+                       /* Disable per-CPU interrupt on all the other CPU */
+                       smp_call_function_single(cpu, mvneta_percpu_disable,
+                                               pp, true);
+               i++;
+       }
+};
+
+static int mvneta_percpu_notifier(struct notifier_block *nfb,
+                                 unsigned long action, void *hcpu)
+{
+       struct mvneta_port *pp = container_of(nfb, struct mvneta_port,
+                                             cpu_notifier);
+       int cpu = (unsigned long)hcpu, other_cpu;
+       struct mvneta_pcpu_port *port = per_cpu_ptr(pp->ports, cpu);
+
+       switch (action) {
+       case CPU_ONLINE:
+       case CPU_ONLINE_FROZEN:
+               netif_tx_stop_all_queues(pp->dev);
+
+               /* We have to synchronise on tha napi of each CPU
+                * except the one just being waked up
+                */
+               for_each_online_cpu(other_cpu) {
+                       if (other_cpu != cpu) {
+                               struct mvneta_pcpu_port *other_port =
+                                       per_cpu_ptr(pp->ports, other_cpu);
+
+                               napi_synchronize(&other_port->napi);
+                       }
+               }
+
+               /* Mask all ethernet port interrupts */
+               mvreg_write(pp, MVNETA_INTR_NEW_MASK, 0);
+               mvreg_write(pp, MVNETA_INTR_OLD_MASK, 0);
+               mvreg_write(pp, MVNETA_INTR_MISC_MASK, 0);
+               napi_enable(&port->napi);
+
+               /* Enable per-CPU interrupt on the one CPU we care
+                * about.
+                */
+               mvneta_percpu_elect(pp);
+
+               /* Unmask all ethernet port interrupts */
+               mvreg_write(pp, MVNETA_INTR_NEW_MASK,
+                       MVNETA_RX_INTR_MASK(rxq_number) |
+                       MVNETA_TX_INTR_MASK(txq_number) |
+                       MVNETA_MISCINTR_INTR_MASK);
+               mvreg_write(pp, MVNETA_INTR_MISC_MASK,
+                       MVNETA_CAUSE_PHY_STATUS_CHANGE |
+                       MVNETA_CAUSE_LINK_CHANGE |
+                       MVNETA_CAUSE_PSC_SYNC_CHANGE);
+               netif_tx_start_all_queues(pp->dev);
+               break;
+       case CPU_DOWN_PREPARE:
+       case CPU_DOWN_PREPARE_FROZEN:
+               netif_tx_stop_all_queues(pp->dev);
+               /* Mask all ethernet port interrupts */
+               mvreg_write(pp, MVNETA_INTR_NEW_MASK, 0);
+               mvreg_write(pp, MVNETA_INTR_OLD_MASK, 0);
+               mvreg_write(pp, MVNETA_INTR_MISC_MASK, 0);
+
+               napi_synchronize(&port->napi);
+               napi_disable(&port->napi);
+               /* Disable per-CPU interrupts on the CPU that is
+                * brought down.
+                */
+               smp_call_function_single(cpu, mvneta_percpu_disable,
+                                        pp, true);
+
+               break;
+       case CPU_DEAD:
+       case CPU_DEAD_FROZEN:
+               /* Check if a new CPU must be elected now this on is down */
+               mvneta_percpu_elect(pp);
+               /* Unmask all ethernet port interrupts */
+               mvreg_write(pp, MVNETA_INTR_NEW_MASK,
+                       MVNETA_RX_INTR_MASK(rxq_number) |
+                       MVNETA_TX_INTR_MASK(txq_number) |
+                       MVNETA_MISCINTR_INTR_MASK);
+               mvreg_write(pp, MVNETA_INTR_MISC_MASK,
+                       MVNETA_CAUSE_PHY_STATUS_CHANGE |
+                       MVNETA_CAUSE_LINK_CHANGE |
+                       MVNETA_CAUSE_PSC_SYNC_CHANGE);
+               netif_tx_start_all_queues(pp->dev);
+               break;
+       }
+
+       return NOTIFY_OK;
+}
+
 static int mvneta_open(struct net_device *dev)
 {
        struct mvneta_port *pp = netdev_priv(dev);
@@ -2683,6 +2803,22 @@ static int mvneta_open(struct net_device *dev)
                goto err_cleanup_txqs;
        }
 
+       /* Even though the documentation says that request_percpu_irq
+        * doesn't enable the interrupts automatically, it actually
+        * does so on the local CPU.
+        *
+        * Make sure it's disabled.
+        */
+       mvneta_percpu_disable(pp);
+
+       /* Elect a CPU to handle our RX queue interrupt */
+       mvneta_percpu_elect(pp);
+
+       /* Register a CPU notifier to handle the case where our CPU
+        * might be taken offline.
+        */
+       register_cpu_notifier(&pp->cpu_notifier);
+
        /* In default link is down */
        netif_carrier_off(pp->dev);
 
@@ -2709,9 +2845,13 @@ err_cleanup_rxqs:
 static int mvneta_stop(struct net_device *dev)
 {
        struct mvneta_port *pp = netdev_priv(dev);
+       int cpu;
 
        mvneta_stop_dev(pp);
        mvneta_mdio_remove(pp);
+       unregister_cpu_notifier(&pp->cpu_notifier);
+       for_each_present_cpu(cpu)
+               smp_call_function_single(cpu, mvneta_percpu_disable, pp, true);
        free_percpu_irq(dev->irq, pp->ports);
        mvneta_cleanup_rxqs(pp);
        mvneta_cleanup_txqs(pp);
@@ -3051,6 +3191,7 @@ static int mvneta_probe(struct platform_device *pdev)
        err = of_property_read_string(dn, "managed", &managed);
        pp->use_inband_status = (err == 0 &&
                                 strcmp(managed, "in-band-status") == 0);
+       pp->cpu_notifier.notifier_call = mvneta_percpu_notifier;
 
        pp->clk = devm_clk_get(&pdev->dev, NULL);
        if (IS_ERR(pp->clk)) {