Interrupts disturb real-time tasks on affined cpus.
To ensure CPU isolation for real-time tasks, interrupt handling must
be adjusted accordingly.
Non-managed interrupts can be configured from userspace,
while managed interrupts require adjustments in kernelspace.

Adjust status of managed interrupts according change
of housekeeping CPUs to support dynamic CPU isolation.

Signed-off-by: Costa Shulyupin <costa.shul@redhat.com>

---

The following code is a proof of concept to validate
and review the correctness of the approach to solving the problem.

C++ comments denote temporary comments.

v2:
- refactor irq_affinity_adjust():
- add more comments
- add managed_irq_isolate() derived from migrate_one_irq(),
  irq_needs_fixup() and irq_fixup_move_pending()
- use irq_set_affinity() instead of irq_set_affinity_locked
- Addressed Gleixner's comments:
- use `struct cpumask *` instead of `cpumask_var_t` in function signature
- remove locking in irq_affinity_adjust()

v1:
- https://lore.kernel.org/lkml/20240516190437.3545310-5-costa.shul@redhat.com/
---
 include/linux/irq.h      |  2 +
 kernel/cgroup/cpuset.c   |  1 +
 kernel/irq/cpuhotplug.c  | 95 ++++++++++++++++++++++++++++++++++++++++
 kernel/sched/isolation.c | 10 ++++-
 4 files changed, 107 insertions(+), 1 deletion(-)

diff --git a/include/linux/irq.h b/include/linux/irq.h
index fa711f80957b6..4eb2e765dbd95 100644
--- a/include/linux/irq.h
+++ b/include/linux/irq.h
@@ -619,6 +619,8 @@ extern int irq_affinity_online_cpu(unsigned int cpu);
 # define irq_affinity_online_cpu	NULL
 #endif
 
+int managed_irq_affinity_adjust(struct cpumask *enable_mask);
+
 #if defined(CONFIG_SMP) && defined(CONFIG_GENERIC_PENDING_IRQ)
 void __irq_move_irq(struct irq_data *data);
 static inline void irq_move_irq(struct irq_data *data)
diff --git a/kernel/cgroup/cpuset.c b/kernel/cgroup/cpuset.c
index bf60bdc973dd6..73b06b2cd91e3 100644
--- a/kernel/cgroup/cpuset.c
+++ b/kernel/cgroup/cpuset.c
@@ -90,6 +90,7 @@ static struct list_head remote_children;
 #define HOUSEKEEPING_FLAGS	(BIT(HK_TYPE_TIMER)  | BIT(HK_TYPE_RCU)  |\
 				 BIT(HK_TYPE_SCHED)  | BIT(HK_TYPE_MISC) |\
 				 BIT(HK_TYPE_DOMAIN) | BIT(HK_TYPE_WQ)   |\
+				 BIT(HK_TYPE_MANAGED_IRQ) |\
 				 BIT(HK_TYPE_KTHREAD))
 
 /*
diff --git a/kernel/irq/cpuhotplug.c b/kernel/irq/cpuhotplug.c
index ec2cdcd20bee7..adbe1f3e5bd22 100644
--- a/kernel/irq/cpuhotplug.c
+++ b/kernel/irq/cpuhotplug.c
@@ -252,3 +252,98 @@ int irq_affinity_online_cpu(unsigned int cpu)
 
 	return 0;
 }
+
+/*
+ * managed_irq_isolate() - Deactivate managed interrupts if necessary
+ */
+// derived from migrate_one_irq, irq_needs_fixup, irq_fixup_move_pending
+static int managed_irq_isolate(struct irq_desc *desc)
+{
+	struct irq_data *d = irq_desc_get_irq_data(desc);
+	struct irq_chip *chip = irq_data_get_irq_chip(d);
+	const struct cpumask *a;
+	bool maskchip;
+	int err;
+
+	/*
+	 * Deactivate if:
+	 * - Interrupt is managed
+	 * - Interrupt is not per cpu
+	 * - Interrupt is started
+	 * - Effective affinity mask includes isolated CPUs
+	 */
+	if (!irqd_affinity_is_managed(d) || irqd_is_per_cpu(d) || !irqd_is_started(d)
+	    || cpumask_subset(irq_data_get_effective_affinity_mask(d),
+			      housekeeping_cpumask(HK_TYPE_MANAGED_IRQ)))
+		return 0;
+	// TBD: it is required?
+	/*
+	 * Complete an eventually pending irq move cleanup. If this
+	 * interrupt was moved in hard irq context, then the vectors need
+	 * to be cleaned up. It can't wait until this interrupt actually
+	 * happens and this CPU was involved.
+	 */
+	irq_force_complete_move(desc);
+
+	if (irqd_is_setaffinity_pending(d)) {
+		irqd_clr_move_pending(d);
+		if (cpumask_intersects(desc->pending_mask,
+		    housekeeping_cpumask(HK_TYPE_MANAGED_IRQ)))
+			a = irq_desc_get_pending_mask(desc);
+	} else {
+		a = irq_data_get_affinity_mask(d);
+	}
+
+	maskchip = chip->irq_mask && !irq_can_move_pcntxt(d) && !irqd_irq_masked(d);
+	if (maskchip)
+		chip->irq_mask(d);
+
+	if (!cpumask_intersects(a, housekeeping_cpumask(HK_TYPE_MANAGED_IRQ))) {
+		/*
+		 * Shut managed interrupt down and leave the affinity untouched.
+		 * The effective affinity is reset to the first online CPU.
+		 */
+		irqd_set_managed_shutdown(d);
+		irq_shutdown_and_deactivate(desc);
+		return 0;
+	}
+
+	/*
+	 * Do not set the force argument of irq_do_set_affinity() as this
+	 * disables the masking of offline CPUs from the supplied affinity
+	 * mask and therefore might keep/reassign the irq to the outgoing
+	 * CPU.
+	 */
+	err = irq_do_set_affinity(d, a, false);
+	if (err)
+		pr_warn_ratelimited("IRQ%u: set affinity failed(%d).\n",
+				    d->irq, err);
+
+	if (maskchip)
+		chip->irq_unmask(d);
+
+	return err;
+}
+
+/** managed_irq_affinity_adjust() - Deactivate of restore managed interrupts
+ * according to change of housekeeping cpumask.
+ *
+ * @enable_mask:	CPUs for which interrupts should be restored
+ */
+int managed_irq_affinity_adjust(struct cpumask *enable_mask)
+{
+	unsigned int irq;
+
+	for_each_active_irq(irq) {
+		struct irq_desc *desc = irq_to_desc(irq);
+		unsigned int cpu;
+
+		for_each_cpu(cpu, enable_mask)
+			irq_restore_affinity_of_irq(desc, cpu);
+		raw_spin_lock(&desc->lock);
+		managed_irq_isolate(desc);
+		raw_spin_unlock(&desc->lock);
+	}
+
+	return 0;
+}
diff --git a/kernel/sched/isolation.c b/kernel/sched/isolation.c
index 3f24921b929a0..cd72300ec8b99 100644
--- a/kernel/sched/isolation.c
+++ b/kernel/sched/isolation.c
@@ -130,6 +130,8 @@ static void __init housekeeping_setup_type(enum hk_type type,
  */
 static int housekeeping_update(enum hk_type type, const struct cpumask *update)
 {
+	int err = 0;
+
 	struct {
 		struct cpumask changed;
 		struct cpumask enable;
@@ -149,9 +151,15 @@ static int housekeeping_update(enum hk_type type, const struct cpumask *update)
 	if (!static_branch_unlikely(&housekeeping_overridden))
 		static_key_enable_cpuslocked(&housekeeping_overridden.key);
 
+	switch (type) {
+	case HK_TYPE_MANAGED_IRQ:
+		err = managed_irq_affinity_adjust(&masks->enable);
+		break;
+	default:
+	}
 	kfree(masks);
 
-	return 0;
+	return err;
 }
 
 static int __init housekeeping_setup(char *str, unsigned long flags)
-- 
2.45.0

On Mon, Sep 16 2024 at 15:20, Costa Shulyupin wrote:

> Interrupts disturb real-time tasks on affined cpus.
> To ensure CPU isolation for real-time tasks, interrupt handling must
> be adjusted accordingly.
> Non-managed interrupts can be configured from userspace,
> while managed interrupts require adjustments in kernelspace.
>
> Adjust status of managed interrupts according change
> of housekeeping CPUs to support dynamic CPU isolation.

What means 'adjust status' ?

> +
> +/*
> + * managed_irq_isolate() - Deactivate managed interrupts if necessary
> + */
> +// derived from migrate_one_irq, irq_needs_fixup, irq_fixup_move_pending

If at all then this needs to be integrated with migrate_one_irq()

> +static int managed_irq_isolate(struct irq_desc *desc)
> +{
> +	struct irq_data *d = irq_desc_get_irq_data(desc);
> +	struct irq_chip *chip = irq_data_get_irq_chip(d);
> +	const struct cpumask *a;
> +	bool maskchip;
> +	int err;
> +
> +	/*
> +	 * Deactivate if:
> +	 * - Interrupt is managed
> +	 * - Interrupt is not per cpu
> +	 * - Interrupt is started
> +	 * - Effective affinity mask includes isolated CPUs
> +	 */
> +	if (!irqd_affinity_is_managed(d) || irqd_is_per_cpu(d) || !irqd_is_started(d)
> +	    || cpumask_subset(irq_data_get_effective_affinity_mask(d),
> +			      housekeeping_cpumask(HK_TYPE_MANAGED_IRQ)))
> +		return 0;
> +	// TBD: it is required?
> +	/*
> +	 * Complete an eventually pending irq move cleanup. If this
> +	 * interrupt was moved in hard irq context, then the vectors need
> +	 * to be cleaned up. It can't wait until this interrupt actually
> +	 * happens and this CPU was involved.
> +	 */
> +	irq_force_complete_move(desc);
> +
> +	if (irqd_is_setaffinity_pending(d)) {
> +		irqd_clr_move_pending(d);
> +		if (cpumask_intersects(desc->pending_mask,
> +		    housekeeping_cpumask(HK_TYPE_MANAGED_IRQ)))
> +			a = irq_desc_get_pending_mask(desc);
> +	} else {
> +		a = irq_data_get_affinity_mask(d);
> +	}
> +
> +	maskchip = chip->irq_mask && !irq_can_move_pcntxt(d) && !irqd_irq_masked(d);
> +	if (maskchip)
> +		chip->irq_mask(d);
> +
> +	if (!cpumask_intersects(a, housekeeping_cpumask(HK_TYPE_MANAGED_IRQ))) {
> +		/*
> +		 * Shut managed interrupt down and leave the affinity untouched.
> +		 * The effective affinity is reset to the first online CPU.
> +		 */
> +		irqd_set_managed_shutdown(d);
> +		irq_shutdown_and_deactivate(desc);
> +		return 0;

Seriously? The interrupt is active and the queue might have outstanding
requests which will never complete because the interrupt is taken away.

On CPU hotplug the related subsystem has shut down the device queue and
drained all outstanding requests. But none of this happens here.

> +	}
> +
> +	/*
> +	 * Do not set the force argument of irq_do_set_affinity() as this
> +	 * disables the masking of offline CPUs from the supplied affinity
> +	 * mask and therefore might keep/reassign the irq to the outgoing
> +	 * CPU.

Which outgoing CPU?

> +	 */
> +	err = irq_do_set_affinity(d, a, false);
> +	if (err)
> +		pr_warn_ratelimited("IRQ%u: set affinity failed(%d).\n",
> +				    d->irq, err);
> +
> +	if (maskchip)
> +		chip->irq_unmask(d);
> +
> +	return err;
> +}
> +
> +/** managed_irq_affinity_adjust() - Deactivate of restore managed interrupts
> + * according to change of housekeeping cpumask.
> + *
> + * @enable_mask:	CPUs for which interrupts should be restored
> + */
> +int managed_irq_affinity_adjust(struct cpumask *enable_mask)
> +{
> +	unsigned int irq;
> +
> +	for_each_active_irq(irq) {

What ensures that this iteration is safe?

> +		struct irq_desc *desc = irq_to_desc(irq);

And that the descriptor is valid?

> +		unsigned int cpu;
> +
> +		for_each_cpu(cpu, enable_mask)
> +			irq_restore_affinity_of_irq(desc, cpu);

And what protects irq_restore_affinity_of_irq() against other operations
on @desc?

> +		raw_spin_lock(&desc->lock);

What disables interrupts here in the runtime case?

> +		managed_irq_isolate(desc);
> +		raw_spin_unlock(&desc->lock);
> +	}
> +
> +	return 0;

That return value has which purpose?

None of this can work at runtime.

Thanks,

        tglx