Skip to contentSkip to conversation

irqchip/gic-v4.1: Two bugfixes about doorbell affinity

PATCH In review Kunkun Jiang 2 participants 12 messages Jan 26, 2024 – Jan 31, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

v2 cover irqchip/gic-v4.1: Two bugfixes about doorbell affinity

There are two bugs in current doorbell affinity settings. May cause lost interrupts. See the patches commit message for details.

Kunkun Jiang (2):
irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity irqchip/gic-v4.1: Fix skipping VMOVP in cpu offline scenario

 drivers/irqchip/irq-gic-v3-its.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)
Signature · 1 line
2.33.0
#1 lore raw

PATCH v2 1/2 irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity

dd3f050a216e make an optimization, VMOVP can be skipped if moving VPE to a cpu whose RD is sharing its VPE table with the current one. But when skipping VMOVP, the affinity recorded in irq_data is still updated. This causes the doorbell affinity recorfed in the irq_data to be inconsistent with the actual.

In corner case, this may result in lost interrupts:
0. Each cpu die shares a VPE table and contains 32 CPUs die0(CPU0-31) die1(CPU32-63)...
1. VPE resides on CPU32, doorbell affinity to CPU32.
2. Move VPE to CPU33, skip VMOVP, doorbell still affinity to CPU32. The affinity recorded in irq_data is CPU33.
3. Manually offline CPU32 on the host side:
'echo 0 > /sys/devices/system/cpu/cpu32/online'
4. Core code cannot move the doorbell affinity to CPU32, since the record in irq_data is CPU33.
5. Subsequent doorbell interrupts will be lost.

So affinity recoreded in irq_data should not be updated when skipping VMOVP.

Fixes dd3f050a216e (irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP)
---
 drivers/irqchip/irq-gic-v3-its.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/irqchip/irq-gic-v3-its.c +2 -1
index d097001c1e3e..4b1dbb697959 100644
@@ -3850,8 +3850,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
 	its_send_vmovp(vpe);
 	its_vpe_db_proxy_move(vpe, from, cpu);
 
-out:
 	irq_data_update_effective_affinity(d, cpumask_of(cpu));
+
+out:
 	vpe_to_cpuid_unlock(vpe, flags);
 
 	return IRQ_SET_MASK_OK_DONE;
Signature · 1 line
2.33.0
maintainer
#2 lore raw

Re: [PATCH v2 1/2] irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity

Quoted 33 lines · Kunkun Jiang

dd3f050a216e make an optimization, VMOVP can be skipped if moving VPE to a cpu whose RD is sharing its VPE table with the current one. But when skipping VMOVP, the affinity recorded in irq_data is still updated. This causes the doorbell affinity recorfed in the irq_data to be inconsistent with the actual.

In corner case, this may result in lost interrupts:
0. Each cpu die shares a VPE table and contains 32 CPUs die0(CPU0-31) die1(CPU32-63)...
1. VPE resides on CPU32, doorbell affinity to CPU32.
2. Move VPE to CPU33, skip VMOVP, doorbell still affinity to CPU32. The affinity recorded in irq_data is CPU33.
3. Manually offline CPU32 on the host side:

   'echo 0 > /sys/devices/system/cpu/cpu32/online'

4. Core code cannot move the doorbell affinity to CPU32, since the record in irq_data is CPU33.
5. Subsequent doorbell interrupts will be lost.

So affinity recoreded in irq_data should not be updated when skipping VMOVP.

Fixes dd3f050a216e (irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP)
---
 drivers/irqchip/irq-gic-v3-its.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/irqchip/irq-gic-v3-its.c +0 -0
index d097001c1e3e..4b1dbb697959 100644

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3858

Quoted 10 lines
drivers/irqchip/irq-gic-v3-its.c +2 -1
@@ -3850,8 +3850,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
 	its_send_vmovp(vpe);
 	its_vpe_db_proxy_move(vpe, from, cpu);
 
-out:
 	irq_data_update_effective_affinity(d, cpumask_of(cpu));
+
+out:
 	vpe_to_cpuid_unlock(vpe, flags);
 
 	return IRQ_SET_MASK_OK_DONE;

This is becoming *very* annoying. I've already told you this patch was wrong, yet you keep sending it (twice in just over an hour).

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#3 lore raw

Re: [PATCH v2 1/2] irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity

Hi Marc,

Quoted 43 lines · Marc Zyngier
Quoted 41 lines · Kunkun Jiang

dd3f050a216e make an optimization, VMOVP can be skipped if moving VPE to a cpu whose RD is sharing its VPE table with the current one. But when skipping VMOVP, the affinity recorded in irq_data is still updated. This causes the doorbell affinity recorfed in the irq_data to be inconsistent with the actual.

In corner case, this may result in lost interrupts:
0. Each cpu die shares a VPE table and contains 32 CPUs

    die0(CPU0-31) die1(CPU32-63)...

1. VPE resides on CPU32, doorbell affinity to CPU32.
2. Move VPE to CPU33, skip VMOVP, doorbell still affinity to CPU32. The affinity recorded in irq_data is CPU33.
3. Manually offline CPU32 on the host side:

    'echo 0 > /sys/devices/system/cpu/cpu32/online'

4. Core code cannot move the doorbell affinity to CPU32, since the record in irq_data is CPU33.
5. Subsequent doorbell interrupts will be lost.

So affinity recoreded in irq_data should not be updated when skipping VMOVP.

Fixes dd3f050a216e (irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP)
---
  drivers/irqchip/irq-gic-v3-its.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
drivers/irqchip/irq-gic-v3-its.c +2 -1
index d097001c1e3e..4b1dbb697959 100644
@@ -3850,8 +3850,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
  	its_send_vmovp(vpe);
  	its_vpe_db_proxy_move(vpe, from, cpu);
 
-out:
  	irq_data_update_effective_affinity(d, cpumask_of(cpu));
+
+out:
  	vpe_to_cpuid_unlock(vpe, flags);
 
Quoted 3 lines

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3858

Quoted 1 line
drivers/irqchip/irq-gic-v3-its.c +0 -0
  	return IRQ_SET_MASK_OK_DONE;

This is becoming *very* annoying. I've already told you this patch was wrong, yet you keep sending it (twice in just over an hour).

Sorry, I didn't see that you had replied to me, when I sent it for the second time.I have no intention of doing this.Sincerely apologize for you.

Please allow me to put your previous reply here for discussion:

Quoted 3 lines

That looks wrong. You are lying to the core code by saying that it's all OK, and yet haven't done*anything*. This stuff is obviously buggy, but I don't think this is right.

Make scene. I understand that the root cause of this problem is that there is a difference between the doorbell affinity recorded in the kernel and the actual hardware. So I'm just thinking about keeping both consistent.

Quoted 3 lines

In your example, you don't even solve the problem: if CPUs 32 and 33 are part of the same ITS affinity group, you won't issue a VMOVP either, so this doesn't fix anything.

No VMOVP is sent. But the affinity recorded in irq_data is correct. So when offline CPU32, core code can move doorbell to CPU0.The mask_val is the entire cpu range, except for the offline cpu.

Quoted 2 lines

At this stage, I think the VMOVP optimisation is wrong and that we should drop it.

This is the last resort. Maybe we can think of another way.

Kunkun Jiang

maintainer
#4 lore raw

Re: [PATCH v2 1/2] irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity

Quoted 4 lines · Kunkun Jiang
Quoted 3 lines

At this stage, I think the VMOVP optimisation is wrong and that we should drop it.

This is the last resort. Maybe we can think of another way.

I think the only other way is to never elide the VMOVP from set_affinity, but to only call into set_affinity when strictly necessary:

- when making the VPE resident on a CPU that isn't part of the same affinity group (VPE migration outside of the affinity group)

- when making the VPE non-resident on a CPU that isn't the doorbell delivery CPU *AND* that there is a need for a doorbell (VPE migration inside the affinity group)

Which means it is the responsibility of the upper layers of the stack to make these decisions, and that we can keep the low-level driver as simple as possible.

In the meantime, here's the patch I'm proposing to merge ASAP. Please let me know if that works for you.

	M.

From 00c338e91d3b6600f402f56cdb64c9d370f911c8 Mon Sep 17 00:00:00 2001
From: Marc Zyngier <[email protected]>
Date: Sun, 28 Jan 2024 13:05:24 +0000
Subject: [PATCH] irqchip/gic-v3-its: Fix GICv4.1 VPE affinity update

When updating the affinity of a VPE, we currently skip the VMOVP command if the two CPUs are part of the same VPE affinity.

But this is wrong, as the doorbell corresponding to this VPE is still delivered on the 'old' CPU, which screws up the balancing. Furthermore, offlining that 'old' CPU results in doorbell interrupts generated for this VPE being discarded.

The harsh reality is that we cannot easily elide VMOVP when a set_affinity request occurs. It needs to be obeyed, and if an optimisation is to be made, it is at the point where the affinity change request is made (such as in KVM).

Drop the VMOVP elision altogether, and only use the vpe_table_mask to try and stay within the same ITS affinity group if at all possible.

Fixes dd3f050a216e (irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP)
---
 drivers/irqchip/irq-gic-v3-its.c | 22 +++++++++++++---------
 1 file changed, 13 insertions(+), 9 deletions(-)
drivers/irqchip/irq-gic-v3-its.c +13 -9
index 250b4562f308..78aece62bff5 100644
@@ -3826,8 +3826,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
 				bool force)
 {
 	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
-	int from, cpu = cpumask_first(mask_val);
+	struct cpumask common;
 	unsigned long flags;
+	int from, cpu;
 
 	/*
 	 * Changing affinity is mega expensive, so let's be as lazy as
@@ -3843,19 +3844,22 @@ static int its_vpe_set_affinity(struct irq_data *d,
 	 * taken on any vLPI handling path that evaluates vpe->col_idx.
 	 */
 	from = vpe_to_cpuid_lock(vpe, &flags);
-	if (from == cpu)
-		goto out;
-
-	vpe->col_idx = cpu;
 
 	/*
-	 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
-	 * is sharing its VPE table with the current one.
+	 * If we are offered another CPU in the same GICv4.1 ITS
+	 * affinity, pick this one. Otherwise, any CPU will do.
 	 */
-	if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
-	    cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
+	if (gic_data_rdist_cpu(from)->vpe_table_mask &&
+	    cpumask_and(&common, mask_val, gic_data_rdist_cpu(from)->vpe_table_mask))
+		cpu = cpumask_first(&common);
+	else
+		cpu = cpumask_first(mask_val);
+
+	if (from == cpu)
 		goto out;
 
+	vpe->col_idx = cpu;
+
 	its_send_vmovp(vpe);
 	its_vpe_db_proxy_move(vpe, from, cpu);
 
Signature · 5 lines
2.39.2


-- 
Without deviation from the norm, progress is not possible.
#5 lore raw

Re: [PATCH v2 1/2] irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity

Hi Marc,

Quoted 16 lines · Marc Zyngier
Quoted 3 lines · Kunkun Jiang
Quoted 2 lines

At this stage, I think the VMOVP optimisation is wrong and that we should drop it.

This is the last resort. Maybe we can think of another way.

I think the only other way is to never elide the VMOVP from set_affinity, but to only call into set_affinity when strictly necessary:

- when making the VPE resident on a CPU that isn't part of the same affinity group (VPE migration outside of the affinity group)

- when making the VPE non-resident on a CPU that isn't the doorbell delivery CPU *AND* that there is a need for a doorbell (VPE migration inside the affinity group)

Quoted 6 lines

Which means it is the responsibility of the upper layers of the stack to make these decisions, and that we can keep the low-level driver as simple as possible.

In the meantime, here's the patch I'm proposing to merge ASAP. Please let me know if that works for you.

This patch works. But I still have some questions, which I will write down below.

Quoted 62 lines
	M.
Quoted 1 line

From 00c338e91d3b6600f402f56cdb64c9d370f911c8 Mon Sep 17 00:00:00 2001

From: Marc Zyngier <[email protected]>
Date: Sun, 28 Jan 2024 13:05:24 +0000
Subject: [PATCH] irqchip/gic-v3-its: Fix GICv4.1 VPE affinity update

When updating the affinity of a VPE, we currently skip the VMOVP command if the two CPUs are part of the same VPE affinity.

But this is wrong, as the doorbell corresponding to this VPE is still delivered on the 'old' CPU, which screws up the balancing. Furthermore, offlining that 'old' CPU results in doorbell interrupts generated for this VPE being discarded.

The harsh reality is that we cannot easily elide VMOVP when a set_affinity request occurs. It needs to be obeyed, and if an optimisation is to be made, it is at the point where the affinity change request is made (such as in KVM).

Drop the VMOVP elision altogether, and only use the vpe_table_mask to try and stay within the same ITS affinity group if at all possible.

Fixes dd3f050a216e (irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP)
---
  drivers/irqchip/irq-gic-v3-its.c | 22 +++++++++++++---------
  1 file changed, 13 insertions(+), 9 deletions(-)
drivers/irqchip/irq-gic-v3-its.c +4 -7
index 250b4562f308..78aece62bff5 100644
@@ -3826,8 +3826,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
  				bool force)
  {
  	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
-	int from, cpu = cpumask_first(mask_val);
+	struct cpumask common;
  	unsigned long flags;
+	int from, cpu;
 
  	/*
  	 * Changing affinity is mega expensive, so let's be as lazy as
@@ -3843,19 +3844,22 @@ static int its_vpe_set_affinity(struct irq_data *d,
  	 * taken on any vLPI handling path that evaluates vpe->col_idx.
  	 */
  	from = vpe_to_cpuid_lock(vpe, &flags);
-	if (from == cpu)
-		goto out;
-
-	vpe->col_idx = cpu;
 
  	/*
-	 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
-	 * is sharing its VPE table with the current one.
+	 * If we are offered another CPU in the same GICv4.1 ITS
+	 * affinity, pick this one. Otherwise, any CPU will do.
  	 */

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3859

Quoted 10 lines
drivers/irqchip/irq-gic-v3-its.c +7 -2
-	if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
-	    cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
+	if (gic_data_rdist_cpu(from)->vpe_table_mask &&
+	    cpumask_and(&common, mask_val, gic_data_rdist_cpu(from)->vpe_table_mask))
+		cpu = cpumask_first(&common);
+	else
+		cpu = cpumask_first(mask_val);
+
+	if (from == cpu)
  		goto out;

The meaning here should be that VMOVP will be skipped only if from and target are the same. Then why do we still to judge whether they are in the same GICv4.1 ITS affinity?

Kunkun Jiang

Quoted 6 lines
drivers/irqchip/irq-gic-v3-its.c +2 -0
+	vpe->col_idx = cpu;
+
  	its_send_vmovp(vpe);
  	its_vpe_db_proxy_move(vpe, from, cpu);
 
maintainer
#6 lore raw

Re: [PATCH v2 1/2] irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity

Quoted 94 lines · Kunkun Jiang

Hi Marc,

Quoted 16 lines · Marc Zyngier
Quoted 3 lines · Kunkun Jiang
Quoted 2 lines

At this stage, I think the VMOVP optimisation is wrong and that we should drop it.

This is the last resort. Maybe we can think of another way.

I think the only other way is to never elide the VMOVP from set_affinity, but to only call into set_affinity when strictly necessary:

- when making the VPE resident on a CPU that isn't part of the same affinity group (VPE migration outside of the affinity group)

- when making the VPE non-resident on a CPU that isn't the doorbell delivery CPU *AND* that there is a need for a doorbell (VPE migration inside the affinity group)

Quoted 6 lines

Which means it is the responsibility of the upper layers of the stack to make these decisions, and that we can keep the low-level driver as simple as possible.

In the meantime, here's the patch I'm proposing to merge ASAP. Please let me know if that works for you.

This patch works. But I still have some questions, which I will write down below.

Quoted 66 lines
	M.
Quoted 1 line

From 00c338e91d3b6600f402f56cdb64c9d370f911c8 Mon Sep 17 00:00:00 2001

From: Marc Zyngier <[email protected]>
Date: Sun, 28 Jan 2024 13:05:24 +0000
Subject: [PATCH] irqchip/gic-v3-its: Fix GICv4.1 VPE affinity update

When updating the affinity of a VPE, we currently skip the VMOVP command if the two CPUs are part of the same VPE affinity.

But this is wrong, as the doorbell corresponding to this VPE is still delivered on the 'old' CPU, which screws up the balancing. Furthermore, offlining that 'old' CPU results in doorbell interrupts generated for this VPE being discarded.

The harsh reality is that we cannot easily elide VMOVP when a set_affinity request occurs. It needs to be obeyed, and if an optimisation is to be made, it is at the point where the affinity change request is made (such as in KVM).

Drop the VMOVP elision altogether, and only use the vpe_table_mask to try and stay within the same ITS affinity group if at all possible.

Fixes dd3f050a216e (irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP)
---
  drivers/irqchip/irq-gic-v3-its.c | 22 +++++++++++++---------
  1 file changed, 13 insertions(+), 9 deletions(-)
drivers/irqchip/irq-gic-v3-its.c +8 -9
index 250b4562f308..78aece62bff5 100644
@@ -3826,8 +3826,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
  				bool force)
  {
  	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
-	int from, cpu = cpumask_first(mask_val);
+	struct cpumask common;
  	unsigned long flags;
+	int from, cpu;
    	/*
  	 * Changing affinity is mega expensive, so let's be as lazy as
@@ -3843,19 +3844,22 @@ static int its_vpe_set_affinity(struct irq_data *d,
  	 * taken on any vLPI handling path that evaluates vpe->col_idx.
  	 */
  	from = vpe_to_cpuid_lock(vpe, &flags);
-	if (from == cpu)
-		goto out;
-
-	vpe->col_idx = cpu;
    	/*
-	 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
-	 * is sharing its VPE table with the current one.
+	 * If we are offered another CPU in the same GICv4.1 ITS
+	 * affinity, pick this one. Otherwise, any CPU will do.
  	 */
-	if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
-	    cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
+	if (gic_data_rdist_cpu(from)->vpe_table_mask &&
+	    cpumask_and(&common, mask_val, gic_data_rdist_cpu(from)->vpe_table_mask))
+		cpu = cpumask_first(&common);
+	else
Quoted 8 lines

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3859

Quoted 4 lines
drivers/irqchip/irq-gic-v3-its.c +3 -0
+		cpu = cpumask_first(mask_val);
+
+	if (from == cpu)
  		goto out;

The meaning here should be that VMOVP will be skipped only if from and target are the same. Then why do we still to judge whether they are in the same GICv4.1 ITS affinity?

Having the same ITS affinity is a performance optimisation, as this minimises traffic between ITSs and keep cache locality. If the proposed affinity contains CPUs that are in the same affinity as current one, then it makes sense to stay there if possible in order to keep the benefit of the warm translation caches.

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#7 lore raw

Re: [PATCH v2 1/2] irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity

Hi Marc,

Quoted 102 lines · Marc Zyngier
Quoted 92 lines · Kunkun Jiang

Hi Marc,

Quoted 16 lines · Marc Zyngier
Quoted 3 lines · Kunkun Jiang
Quoted 2 lines

At this stage, I think the VMOVP optimisation is wrong and that we should drop it.

This is the last resort. Maybe we can think of another way.

I think the only other way is to never elide the VMOVP from set_affinity, but to only call into set_affinity when strictly necessary:

- when making the VPE resident on a CPU that isn't part of the same affinity group (VPE migration outside of the affinity group)

- when making the VPE non-resident on a CPU that isn't the doorbell delivery CPU *AND* that there is a need for a doorbell (VPE migration inside the affinity group)

Quoted 6 lines

Which means it is the responsibility of the upper layers of the stack to make these decisions, and that we can keep the low-level driver as simple as possible.

In the meantime, here's the patch I'm proposing to merge ASAP. Please let me know if that works for you.

This patch works. But I still have some questions, which I will write down below.

Quoted 65 lines
	M.
Quoted 1 line

From 00c338e91d3b6600f402f56cdb64c9d370f911c8 Mon Sep 17 00:00:00 2001

From: Marc Zyngier <[email protected]>
Date: Sun, 28 Jan 2024 13:05:24 +0000
Subject: [PATCH] irqchip/gic-v3-its: Fix GICv4.1 VPE affinity update

When updating the affinity of a VPE, we currently skip the VMOVP command if the two CPUs are part of the same VPE affinity.

But this is wrong, as the doorbell corresponding to this VPE is still delivered on the 'old' CPU, which screws up the balancing. Furthermore, offlining that 'old' CPU results in doorbell interrupts generated for this VPE being discarded.

The harsh reality is that we cannot easily elide VMOVP when a set_affinity request occurs. It needs to be obeyed, and if an optimisation is to be made, it is at the point where the affinity change request is made (such as in KVM).

Drop the VMOVP elision altogether, and only use the vpe_table_mask to try and stay within the same ITS affinity group if at all possible.

Fixes dd3f050a216e (irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP)
---
   drivers/irqchip/irq-gic-v3-its.c | 22 +++++++++++++---------
   1 file changed, 13 insertions(+), 9 deletions(-)
drivers/irqchip/irq-gic-v3-its.c +8 -9
index 250b4562f308..78aece62bff5 100644
@@ -3826,8 +3826,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
   				bool force)
   {
   	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
-	int from, cpu = cpumask_first(mask_val);
+	struct cpumask common;
   	unsigned long flags;
+	int from, cpu;
     	/*
   	 * Changing affinity is mega expensive, so let's be as lazy as
@@ -3843,19 +3844,22 @@ static int its_vpe_set_affinity(struct irq_data *d,
   	 * taken on any vLPI handling path that evaluates vpe->col_idx.
   	 */
   	from = vpe_to_cpuid_lock(vpe, &flags);
-	if (from == cpu)
-		goto out;
-
-	vpe->col_idx = cpu;
     	/*
-	 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
-	 * is sharing its VPE table with the current one.
+	 * If we are offered another CPU in the same GICv4.1 ITS
+	 * affinity, pick this one. Otherwise, any CPU will do.
   	 */
-	if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
-	    cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
+	if (gic_data_rdist_cpu(from)->vpe_table_mask &&
+	    cpumask_and(&common, mask_val, gic_data_rdist_cpu(from)->vpe_table_mask))
+		cpu = cpumask_first(&common);
+	else
Quoted 8 lines

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3859

Quoted 4 lines
drivers/irqchip/irq-gic-v3-its.c +3 -0
+		cpu = cpumask_first(mask_val);
+
+	if (from == cpu)
   		goto out;

The meaning here should be that VMOVP will be skipped only if from and target are the same. Then why do we still to judge whether they are in the same GICv4.1 ITS affinity?

Quoted 5 lines

Having the same ITS affinity is a performance optimisation, as this minimises traffic between ITSs and keep cache locality. If the proposed affinity contains CPUs that are in the same affinity as current one, then it makes sense to stay there if possible in order to keep the benefit of the warm translation caches.

Ok, I see. In another case, is it possible to skip VMOVP? The from belongs to common, but it is not the first one.

Kunkun Jiang

maintainer
#8 lore raw

Re: [PATCH v2 1/2] irqchip/gic-v4.1: Fix GICv4.1 doorbell affinity

Quoted 106 lines · Kunkun Jiang

Hi Marc,

Quoted 102 lines · Marc Zyngier
Quoted 100 lines · Kunkun Jiang

Hi Marc,

Quoted 16 lines · Marc Zyngier
Quoted 3 lines · Kunkun Jiang
Quoted 2 lines

At this stage, I think the VMOVP optimisation is wrong and that we should drop it.

This is the last resort. Maybe we can think of another way.

I think the only other way is to never elide the VMOVP from set_affinity, but to only call into set_affinity when strictly necessary:

- when making the VPE resident on a CPU that isn't part of the same affinity group (VPE migration outside of the affinity group)

- when making the VPE non-resident on a CPU that isn't the doorbell delivery CPU *AND* that there is a need for a doorbell (VPE migration inside the affinity group)

Quoted 6 lines

Which means it is the responsibility of the upper layers of the stack to make these decisions, and that we can keep the low-level driver as simple as possible.

In the meantime, here's the patch I'm proposing to merge ASAP. Please let me know if that works for you.

This patch works. But I still have some questions, which I will write down below.

Quoted 59 lines
	M.
Quoted 1 line

From 00c338e91d3b6600f402f56cdb64c9d370f911c8 Mon Sep 17 00:00:00 2001

From: Marc Zyngier <[email protected]>
Date: Sun, 28 Jan 2024 13:05:24 +0000
Subject: [PATCH] irqchip/gic-v3-its: Fix GICv4.1 VPE affinity update

When updating the affinity of a VPE, we currently skip the VMOVP command if the two CPUs are part of the same VPE affinity.

But this is wrong, as the doorbell corresponding to this VPE is still delivered on the 'old' CPU, which screws up the balancing. Furthermore, offlining that 'old' CPU results in doorbell interrupts generated for this VPE being discarded.

The harsh reality is that we cannot easily elide VMOVP when a set_affinity request occurs. It needs to be obeyed, and if an optimisation is to be made, it is at the point where the affinity change request is made (such as in KVM).

Drop the VMOVP elision altogether, and only use the vpe_table_mask to try and stay within the same ITS affinity group if at all possible.

Fixes dd3f050a216e (irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP)
---
   drivers/irqchip/irq-gic-v3-its.c | 22 +++++++++++++---------
   1 file changed, 13 insertions(+), 9 deletions(-)
drivers/irqchip/irq-gic-v3-its.c +4 -7
index 250b4562f308..78aece62bff5 100644
@@ -3826,8 +3826,9 @@ static int its_vpe_set_affinity(struct irq_data *d,
   				bool force)
   {
   	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
-	int from, cpu = cpumask_first(mask_val);
+	struct cpumask common;
   	unsigned long flags;
+	int from, cpu;
     	/*
   	 * Changing affinity is mega expensive, so let's be as lazy as
@@ -3843,19 +3844,22 @@ static int its_vpe_set_affinity(struct irq_data *d,
   	 * taken on any vLPI handling path that evaluates vpe->col_idx.
   	 */
   	from = vpe_to_cpuid_lock(vpe, &flags);
-	if (from == cpu)
-		goto out;
-
-	vpe->col_idx = cpu;
     	/*
-	 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
-	 * is sharing its VPE table with the current one.
+	 * If we are offered another CPU in the same GICv4.1 ITS
+	 * affinity, pick this one. Otherwise, any CPU will do.
   	 */

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3858

Quoted 10 lines
drivers/irqchip/irq-gic-v3-its.c +7 -2
-	if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
-	    cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
+	if (gic_data_rdist_cpu(from)->vpe_table_mask &&
+	    cpumask_and(&common, mask_val, gic_data_rdist_cpu(from)->vpe_table_mask))
+		cpu = cpumask_first(&common);
+	else
+		cpu = cpumask_first(mask_val);
+
+	if (from == cpu)
   		goto out;

The meaning here should be that VMOVP will be skipped only if from and target are the same. Then why do we still to judge whether they are in the same GICv4.1 ITS affinity?

Quoted 7 lines
Quoted 5 lines

Having the same ITS affinity is a performance optimisation, as this minimises traffic between ITSs and keep cache locality. If the proposed affinity contains CPUs that are in the same affinity as current one, then it makes sense to stay there if possible in order to keep the benefit of the warm translation caches.

Ok, I see. In another case, is it possible to skip VMOVP? The from belongs to common, but it is not the first one.

You could have something like

		if (cpumask_test_cpu(from, &common))
			cpu = from;
		else
			cpu = cpumask_first(&common);

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#9 lore raw

PATCH v2 2/2 irqchip/gic-v4.1: Fix skipping VMOVP in cpu offline scenario

In reply to Kunkun Jiang

commit dd3f050a216e ("irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP") make an optimization, VMOVP can be skipped if moving VPE to a cpu whose RD is sharing its VPE table with the current one.

In the cpu offline scenario, the mask_val is the entire cpu range, except for the offline cpu. Therefore, the first cpu is CPU0. In corner case, this may result in lost interrupts:
0. Each cpu die shares a VPE table and contains 32 CPUs die0(CPU0-31) die1(CPU32-63)...
1. VPE resides on CPU32, doorbell affinity to CPU32.
2. Move VPE to CPU16, doorbell affinity to CPU16.
3. Manually offline CPU16 on the host side:
'echo 0 > /sys/devices/system/cpu/cpu16/online'
4. VMOVP will be skipped.
5. Subsequent doorbell interrupts will be lost.

So VMOVP cannot be skipped when the affinity CPU is not in mask_val.

Fixes dd3f050a216e ("irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP")
---
 drivers/irqchip/irq-gic-v3-its.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/irqchip/irq-gic-v3-its.c +3 -1
index 4b1dbb697959..bfb922f16bc6 100644
@@ -3817,7 +3817,7 @@ static int its_vpe_set_affinity(struct irq_data *d,
 				bool force)
 {
 	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
-	int from, cpu = cpumask_first(mask_val);
+	int cur, from, cpu = cpumask_first(mask_val);
 	unsigned long flags;
 
 	/*
@@ -3839,11 +3839,13 @@ static int its_vpe_set_affinity(struct irq_data *d,
 
 	vpe->col_idx = cpu;
 
+	cur = cpumask_first(irq_data_get_effective_affinity_mask(d));
 	/*
 	 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
 	 * is sharing its VPE table with the current one.
 	 */
 	if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
+	    cpumask_test_cpu(cur, mask_val) &&
 	    cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
 		goto out;
 
Signature · 1 line
2.33.0
maintainer
#10 lore raw

Re: [PATCH v2 2/2] irqchip/gic-v4.1: Fix skipping VMOVP in cpu offline scenario

Quoted 39 lines · Kunkun Jiang

commit dd3f050a216e ("irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP") make an optimization, VMOVP can be skipped if moving VPE to a cpu whose RD is sharing its VPE table with the current one.

In the cpu offline scenario, the mask_val is the entire cpu range, except for the offline cpu. Therefore, the first cpu is CPU0. In corner case, this may result in lost interrupts:
0. Each cpu die shares a VPE table and contains 32 CPUs die0(CPU0-31) die1(CPU32-63)...
1. VPE resides on CPU32, doorbell affinity to CPU32.
2. Move VPE to CPU16, doorbell affinity to CPU16.
3. Manually offline CPU16 on the host side:

   'echo 0 > /sys/devices/system/cpu/cpu16/online'

4. VMOVP will be skipped.
5. Subsequent doorbell interrupts will be lost.

So VMOVP cannot be skipped when the affinity CPU is not in mask_val.

Fixes dd3f050a216e ("irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP")
---
 drivers/irqchip/irq-gic-v3-its.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/irqchip/irq-gic-v3-its.c +1 -1
index 4b1dbb697959..bfb922f16bc6 100644
@@ -3817,7 +3817,7 @@ static int its_vpe_set_affinity(struct irq_data *d,
 				bool force)
 {
 	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
-	int from, cpu = cpumask_first(mask_val);
+	int cur, from, cpu = cpumask_first(mask_val);
 	unsigned long flags;
 
 	/*

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3842

Quoted 4 lines
drivers/irqchip/irq-gic-v3-its.c +1 -0
@@ -3839,11 +3839,13 @@ static int its_vpe_set_affinity(struct irq_data *d,
 
 	vpe->col_idx = cpu;
 
+	cur = cpumask_first(irq_data_get_effective_affinity_mask(d));

When is this not equal to 'from'?

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3850

Quoted 8 lines
drivers/irqchip/irq-gic-v3-its.c +1 -0
 	/*
 	 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
 	 * is sharing its VPE table with the current one.
 	 */
 	if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
+	    cpumask_test_cpu(cur, mask_val) &&
 	    cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
 		goto out;

This looks utterly pointless to me.

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#11 lore raw

Re: [PATCH v2 2/2] irqchip/gic-v4.1: Fix skipping VMOVP in cpu offline scenario

Hi Marc,

Quoted 38 lines · Marc Zyngier
Quoted 36 lines · Kunkun Jiang

commit dd3f050a216e ("irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP") make an optimization, VMOVP can be skipped if moving VPE to a cpu whose RD is sharing its VPE table with the current one.

In the cpu offline scenario, the mask_val is the entire cpu range, except for the offline cpu. Therefore, the first cpu is CPU0. In corner case, this may result in lost interrupts:
0. Each cpu die shares a VPE table and contains 32 CPUs

    die0(CPU0-31) die1(CPU32-63)...
1. VPE resides on CPU32, doorbell affinity to CPU32.
2. Move VPE to CPU16, doorbell affinity to CPU16.
3. Manually offline CPU16 on the host side:
    'echo 0 > /sys/devices/system/cpu/cpu16/online'

4. VMOVP will be skipped.
5. Subsequent doorbell interrupts will be lost.

So VMOVP cannot be skipped when the affinity CPU is not in mask_val.

Fixes dd3f050a216e ("irqchip/gic-v4.1: Implement the v4.1 flavour of VMOVP")
---
  drivers/irqchip/irq-gic-v3-its.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)
drivers/irqchip/irq-gic-v3-its.c +1 -1
index 4b1dbb697959..bfb922f16bc6 100644
@@ -3817,7 +3817,7 @@ static int its_vpe_set_affinity(struct irq_data *d,
  				bool force)
  {
  	struct its_vpe *vpe = irq_data_get_irq_chip_data(d);
-	int from, cpu = cpumask_first(mask_val);
+	int cur, from, cpu = cpumask_first(mask_val);
  	unsigned long flags;
 
Quoted 7 lines

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3842

Quoted 6 lines
drivers/irqchip/irq-gic-v3-its.c +1 -0
  	/*
@@ -3839,11 +3839,13 @@ static int its_vpe_set_affinity(struct irq_data *d,
 
  	vpe->col_idx = cpu;
 
+	cur = cpumask_first(irq_data_get_effective_affinity_mask(d));

When is this not equal to 'from'?

It's based on my last patch. When skipping VMOVP, it is not equal to 'from'.

Quoted 9 lines

drivers/irqchip/irq-gic-v3-its.c static int its_vpe_set_affinity(struct irq_data *d, · line 3850

Quoted 8 lines
drivers/irqchip/irq-gic-v3-its.c +1 -0
  	/*
  	 * GICv4.1 allows us to skip VMOVP if moving to a cpu whose RD
  	 * is sharing its VPE table with the current one.
  	 */
  	if (gic_data_rdist_cpu(cpu)->vpe_table_mask &&
+	    cpumask_test_cpu(cur, mask_val) &&
  	    cpumask_test_cpu(from, gic_data_rdist_cpu(cpu)->vpe_table_mask))
  		goto out;

This looks utterly pointless to me.

Yes, there are some problems here. I'll think about it more carefully.

Kunkun Jiang