genirq: Allow per-cpu interrupt sharing for non-overlapping affinities

Interrupt sharing for percpu-devid interrupts is forbidden, and for good
reasons. These are interrupts generated *from* a CPU and handled by itself
(timer, for example). Nobody in their right mind would put two devices on
the same pin (and if they have, they get to keep the pieces...).

But this also prevents more benign cases, where devices are connected
to groups of CPUs, and for which the affinities are not overlapping.
Effectively, the only thing they share is the interrupt number, and
nothing else.

Tweak the definition of IRQF_SHARED applied to percpu_devid interrupts to
allow this particular use case. This results in extra validation at the
point of the interrupt being setup and freed, as well as a tiny bit of
extra complexity for interrupts at handling time (to pick the correct
irqaction).

Signed-off-by: Marc Zyngier <[e-mail]>
Signed-off-by: Thomas Gleixner <[e-mail]>
Tested-by: Will Deacon <[e-mail]>
Link: https://patch.msgid.link/[e-mail]

kernel/irq/chip.c

···
 void handle_percpu_devid_irq(struct irq_desc *desc)
 {
 	struct irq_chip *chip = irq_desc_get_chip(desc);
-	struct irqaction *action = desc->action;
 	unsigned int irq = irq_desc_get_irq(desc);
+	unsigned int cpu = smp_processor_id();
+	struct irqaction *action;
 	irqreturn_t res;
 
 	/*
···
 	if (chip->irq_ack)
 		chip->irq_ack(&desc->irq_data);
 
+	for (action = desc->action; action; action = action->next)
+		if (cpumask_test_cpu(cpu, action->affinity))
+			break;
+
 	if (likely(action)) {
 		trace_irq_handler_entry(irq, action);
 		res = action->handler(irq, raw_cpu_ptr(action->percpu_dev_id));
 		trace_irq_handler_exit(irq, action, res);
 	} else {
-		unsigned int cpu = smp_processor_id();
 		bool enabled = cpumask_test_cpu(cpu, desc->percpu_enabled);
 
 		if (enabled)

kernel/irq/manage.c

···
 	return 0;
 }
 
+static bool valid_percpu_irqaction(struct irqaction *old, struct irqaction *new)
+{
+	do {
+		if (cpumask_intersects(old->affinity, new->affinity) ||
+		    old->percpu_dev_id == new->percpu_dev_id)
+			return false;
+
+		old = old->next;
+	} while (old);
+
+	return true;
+}
+
 /*
  * Internal function to register an irqaction - typically used to
  * allocate special interrupts that are part of the architecture.
···
 	struct irqaction *old, **old_ptr;
 	unsigned long flags, thread_mask = 0;
 	int ret, nested, shared = 0;
+	bool per_cpu_devid;
 
 	if (!desc)
 		return -EINVAL;
···
 		return -ENOSYS;
 	if (!try_module_get(desc->owner))
 		return -ENODEV;
+
+	per_cpu_devid = irq_settings_is_per_cpu_devid(desc);
 
 	new->irq = irq;
 
···
 		 */
 		unsigned int oldtype;
 
-		if (irq_is_nmi(desc)) {
+		if (irq_is_nmi(desc) && !per_cpu_devid) {
 			pr_err("Invalid attempt to share NMI for %s (irq %d) on irqchip %s.\n",
+				new->name, irq, desc->irq_data.chip->name);
+			ret = -EINVAL;
+			goto out_unlock;
+		}
+
+		if (per_cpu_devid && !valid_percpu_irqaction(old, new)) {
+			pr_err("Overlapping affinities for %s (irq %d) on irqchip %s.\n",
 				new->name, irq, desc->irq_data.chip->name);
 			ret = -EINVAL;
 			goto out_unlock;
···
 		if (!(new->flags & IRQF_NO_AUTOEN) &&
 		    irq_settings_can_autoenable(desc)) {
 			irq_startup(desc, IRQ_RESEND, IRQ_START_COND);
-		} else {
+		} else if (!per_cpu_devid) {
 			/*
 			 * Shared interrupts do not go well with disabling
 			 * auto enable. The sharing interrupt might request
···
 static struct irqaction *__free_percpu_irq(unsigned int irq, void __percpu *dev_id)
 {
 	struct irq_desc *desc = irq_to_desc(irq);
-	struct irqaction *action;
+	struct irqaction *action, **action_ptr;
 
 	WARN(in_interrupt(), "Trying to free IRQ %d from IRQ context!\n", irq);
 
···
 		return NULL;
 
 	scoped_guard(raw_spinlock_irqsave, &desc->lock) {
-		action = desc->action;
-		if (!action || action->percpu_dev_id != dev_id) {
-			WARN(1, "Trying to free already-free IRQ %d\n", irq);
-			return NULL;
+		action_ptr = &desc->action;
+		for (;;) {
+			action = *action_ptr;
+
+			if (!action) {
+				WARN(1, "Trying to free already-free IRQ %d\n", irq);
+				return NULL;
+			}
+
+			if (action->percpu_dev_id == dev_id)
+				break;
+
+			action_ptr = &action->next;
 		}
 
-		if (!cpumask_empty(desc->percpu_enabled)) {
-			WARN(1, "percpu IRQ %d still enabled on CPU%d!\n",
-			     irq, cpumask_first(desc->percpu_enabled));
+		if (cpumask_intersects(desc->percpu_enabled, action->affinity)) {
+			WARN(1, "percpu IRQ %d still enabled on CPU%d!\n", irq,
+			     cpumask_first_and(desc->percpu_enabled, action->affinity));
 			return NULL;
 		}
 
 		/* Found it - now remove it from the list of entries: */
-		desc->action = NULL;
-		desc->istate &= ~IRQS_NMI;
+		*action_ptr = action->next;
+
+		/* Demote from NMI if we killed the last action */
+		if (!desc->action)
+			desc->istate &= ~IRQS_NMI;
 	}
 
 	unregister_handler_proc(irq, action);
···
 	action->name = devname;
 	action->percpu_dev_id = dev_id;
 	action->affinity = affinity;
+
+	/*
+	 * We allow some form of sharing for non-overlapping affinity
+	 * masks. Obviously, covering all CPUs prevents any sharing in
+	 * the first place.
+	 */
+	if (!cpumask_equal(affinity, cpu_possible_mask))
+		action->flags |= IRQF_SHARED;
 
 	return action;
 }