drivers/perf: arm_pmu: Request PMU SPIs with IRQF_PER_CPU

Since the PMU register interface is banked per CPU, CPU PMU interrrupts
cannot be handled by a CPU other than the one with the PMU asserting the
interrupt. This means that migrating PMU SPIs, as we do during a CPU
hotplug operation doesn't make any sense and can lead to the IRQ being
disabled entirely if we route a spurious IRQ to the new affinity target.

This has been observed in practice on AMD Seattle, where CPUs on the
non-boot cluster appear to take a spurious PMU IRQ when coming online,
which is routed to CPU0 where it cannot be handled.

This patch passes IRQF_PERCPU for PMU SPIs and forcefully sets their
affinity prior to requesting them, ensuring that they cannot
be migrated during hotplug events. This interacts badly with the DB8500
erratum workaround that ping-pongs the interrupt affinity from the handler,
so we avoid passing IRQF_PERCPU in that case by allowing the IRQ flags
to be overridden in the platdata.

Fixes: 3cf7ee9 ("drivers/perf: arm_pmu: move irq request/free into probe")
Cc: Mark Rutland <[email protected]>
Cc: Linus Walleij <[email protected]>
Signed-off-by: Will Deacon <[email protected]>

Author: wildea01

arch/arm/mach-ux500/cpu-db8500.c

@@ -133,6 +133,7 @@ static irqreturn_t db8500_pmu_handler(int irq, void *dev, irq_handler_t handler)
 
 static struct arm_pmu_platdata db8500_pmu_platdata = {
 	.handle_irq		= db8500_pmu_handler,
+	.irq_flags		= IRQF_NOBALANCING | IRQF_NO_THREAD,
 };
 
 static struct of_dev_auxdata u8500_auxdata_lookup[] __initdata = {

drivers/perf/arm_pmu.c

@@ -569,22 +569,41 @@ int armpmu_request_irq(struct arm_pmu *armpmu, int cpu)
 		if (irq != other_irq) {
 			pr_warn("mismatched PPIs detected.\n");
 			err = -EINVAL;
+			goto err_out;
 		}
 	} else {
-		err = request_irq(irq, handler,
-				  IRQF_NOBALANCING | IRQF_NO_THREAD, "arm-pmu",
+		struct arm_pmu_platdata *platdata = armpmu_get_platdata(armpmu);
+		unsigned long irq_flags;
+
+		err = irq_force_affinity(irq, cpumask_of(cpu));
+
+		if (err && num_possible_cpus() > 1) {
+			pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
+				irq, cpu);
+			goto err_out;
+		}
+
+		if (platdata && platdata->irq_flags) {
+			irq_flags = platdata->irq_flags;
+		} else {
+			irq_flags = IRQF_PERCPU |
+				    IRQF_NOBALANCING |
+				    IRQF_NO_THREAD;
+		}
+
+		err = request_irq(irq, handler, irq_flags, "arm-pmu",
 				  per_cpu_ptr(&hw_events->percpu_pmu, cpu));
 	}
 
-	if (err) {
-		pr_err("unable to request IRQ%d for ARM PMU counters\n",
-			irq);
-		return err;
-	}
+	if (err)
+		goto err_out;
 
 	cpumask_set_cpu(cpu, &armpmu->active_irqs);
-
 	return 0;
+
+err_out:
+	pr_err("unable to request IRQ%d for ARM PMU counters\n", irq);
+	return err;
 }
 
 int armpmu_request_irqs(struct arm_pmu *armpmu)
@@ -628,12 +647,6 @@ static int arm_perf_starting_cpu(unsigned int cpu, struct hlist_node *node)
 			enable_percpu_irq(irq, IRQ_TYPE_NONE);
 			return 0;
 		}
-
-		if (irq_force_affinity(irq, cpumask_of(cpu)) &&
-		    num_possible_cpus() > 1) {
-			pr_warn("unable to set irq affinity (irq=%d, cpu=%u)\n",
-				irq, cpu);
-		}
 	}
 
 	return 0;

include/linux/perf/arm_pmu.h

@@ -24,10 +24,14 @@
  *	interrupt and passed the address of the low level handler,
  *	and can be used to implement any platform specific handling
  *	before or after calling it.
+ *
+ * @irq_flags: if non-zero, these flags will be passed to request_irq
+ *             when requesting interrupts for this PMU device.
  */
 struct arm_pmu_platdata {
 	irqreturn_t (*handle_irq)(int irq, void *dev,
 				  irq_handler_t pmu_handler);
+	unsigned long irq_flags;
 };
 
 #ifdef CONFIG_ARM_PMU