diff mbox

[Part3,v4,37/38] x86, irq: Introduce mechanism to support different vector allocation policies

Message ID 1416901802-24211-38-git-send-email-jiang.liu@linux.intel.com
State Not Applicable
Headers show

Commit Message

Jiang Liu Nov. 25, 2014, 7:50 a.m. UTC 
Introduce mechanism to support different vector allocation policies,
so platform or user may choose the best suitable CPU vector allocation
policy. Currently two policies are supported:
1) allocate CPU vector from cpumask_of_node(dev_to_node(dev))
2) allocate from apic->target_cpus(), this is the default policy

Platform driver may call set_vector_alloc_policy() to choose the
preferred policies.

This mechanism may be used to support NumaConnect systems to allocate
CPU vectors from device local node.

We may also enhance to support per-cpu vector allocation if it's needed.

Signed-off-by: Jiang Liu <jiang.liu@linux.intel.com>
Cc: Daniel J Blueman <daniel@numascale.com>
---
 arch/x86/include/asm/hw_irq.h |   12 ++++++++
 arch/x86/kernel/apic/vector.c |   66 +++++++++++++++++++++++++++++++++++------
 2 files changed, 69 insertions(+), 9 deletions(-)

Comments

Thomas Gleixner Nov. 27, 2014, 10:44 a.m. UTC | #1 
On Tue, 25 Nov 2014, Jiang Liu wrote:
> Introduce mechanism to support different vector allocation policies,
> so platform or user may choose the best suitable CPU vector allocation

I've postponed this and the next patch for 3.20.

> policy. Currently two policies are supported:
> 1) allocate CPU vector from cpumask_of_node(dev_to_node(dev))
> 2) allocate from apic->target_cpus(), this is the default policy
> 
> Platform driver may call set_vector_alloc_policy() to choose the
> preferred policies.
> 
> This mechanism may be used to support NumaConnect systems to allocate
> CPU vectors from device local node.

So these policies are system wide and need to be selected at boot time?

Thanks,

	tglx

--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Jiang Liu Nov. 27, 2014, 4:22 p.m. UTC | #2 
On 2014/11/27 18:44, Thomas Gleixner wrote:
> On Tue, 25 Nov 2014, Jiang Liu wrote:
>> Introduce mechanism to support different vector allocation policies,
>> so platform or user may choose the best suitable CPU vector allocation
> 
> I've postponed this and the next patch for 3.20.
Hi Thomas,
	That's fine, thanks for merging all other patches for 3.19:)

> 
>> policy. Currently two policies are supported:
>> 1) allocate CPU vector from cpumask_of_node(dev_to_node(dev))
>> 2) allocate from apic->target_cpus(), this is the default policy
>>
>> Platform driver may call set_vector_alloc_policy() to choose the
>> preferred policies.
>>
>> This mechanism may be used to support NumaConnect systems to allocate
>> CPU vectors from device local node.
> 
> So these policies are system wide and need to be selected at boot time?
Yes, the polices are system wide.
It may be change at runtime, but seems no strong requirement yet so
haven't implemented the interface to change it at runtime.

Basically this idea is inspired by a patch set to allocate  CPU vector
from device local node for NumaConnect systems. I found it may also help
2-socket or 4-socket systems, so reimplemented in this way.
Thanks!
Gerry

> 
> Thanks,
> 
> 	tglx
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/arch/x86/include/asm/hw_irq.h b/arch/x86/include/asm/hw_irq.h
index e7ae6eb84934..61807eed48a3 100644
--- a/arch/x86/include/asm/hw_irq.h
+++ b/arch/x86/include/asm/hw_irq.h
@@ -170,6 +170,17 @@  enum {
 	X86_IRQ_ALLOC_CONTIGOUS_VECTORS		= 0x1,
 };
 
+enum {
+	/* Allocate vector from cpumask_of_node(dev_to_node(dev)) */
+	X86_VECTOR_POL_NODE = 0x1,
+	/* Allocate vector from apic->target_cpus() */
+	X86_VECTOR_POL_DEFAULT = 0x2,
+	/* Allocate vector from cpumask assigned by caller */
+	X86_VECTOR_POL_CALLER = 0x4,
+	X86_VECTOR_POL_MIN = X86_VECTOR_POL_NODE,
+	X86_VECTOR_POL_MAX = X86_VECTOR_POL_CALLER,
+};
+
 struct irq_cfg {
 	unsigned int		dest_apicid;
 	u8			vector;
@@ -185,6 +196,7 @@  extern struct irq_cfg *irq_cfg(unsigned int irq);
 extern struct irq_cfg *irqd_cfg(struct irq_data *irq_data);
 extern void lock_vector_lock(void);
 extern void unlock_vector_lock(void);
+extern void set_vector_alloc_policy(unsigned int policy);
 extern void setup_vector_irq(int cpu);
 #ifdef CONFIG_SMP
 extern void send_cleanup_vector(struct irq_cfg *);
diff --git a/arch/x86/kernel/apic/vector.c b/arch/x86/kernel/apic/vector.c
index 0e7c39beefed..16de8906ee1e 100644
--- a/arch/x86/kernel/apic/vector.c
+++ b/arch/x86/kernel/apic/vector.c
@@ -28,6 +28,8 @@  struct apic_chip_data {
 	u8			move_in_progress : 1;
 };
 
+static unsigned int x86_vector_alloc_policy = X86_VECTOR_POL_DEFAULT |
+					      X86_VECTOR_POL_CALLER;
 struct irq_domain *x86_vector_domain;
 static DEFINE_RAW_SPINLOCK(vector_lock);
 static struct irq_chip lapic_controller;
@@ -71,6 +73,12 @@  struct irq_cfg *irq_cfg(unsigned int irq)
 	return irqd_cfg(irq_get_irq_data(irq));
 }
 
+void set_vector_alloc_policy(unsigned int policy)
+{
+	if (!WARN_ON((policy & (X86_VECTOR_POL_MAX - 1)) == 0))
+		x86_vector_alloc_policy = policy | X86_VECTOR_POL_CALLER;
+}
+
 static struct apic_chip_data *alloc_apic_chip_data(int node)
 {
 	struct apic_chip_data *data;
@@ -259,12 +267,6 @@  void copy_irq_alloc_info(struct irq_alloc_info *dst, struct irq_alloc_info *src)
 		memset(dst, 0, sizeof(*dst));
 }
 
-static inline const struct cpumask *
-irq_alloc_info_get_mask(struct irq_alloc_info *info)
-{
-	return (!info || !info->mask) ? apic->target_cpus() : info->mask;
-}
-
 static void x86_vector_free_irqs(struct irq_domain *domain,
 				 unsigned int virq, unsigned int nr_irqs)
 {
@@ -285,13 +287,59 @@  static void x86_vector_free_irqs(struct irq_domain *domain,
 	}
 }
 
+static int assign_irq_vector_policy(int irq, int node,
+				    struct apic_chip_data *data,
+				    struct irq_alloc_info *info)
+{
+	int err = -EBUSY;
+	unsigned int policy;
+	const struct cpumask *mask;
+
+	if (info && info->mask)
+		policy = X86_VECTOR_POL_CALLER;
+	else
+		policy = X86_VECTOR_POL_MIN;
+
+	for (; policy <= X86_VECTOR_POL_MAX; policy <<= 1) {
+		if (!(x86_vector_alloc_policy & policy))
+			continue;
+
+		switch (policy) {
+		case X86_VECTOR_POL_NODE:
+			if (node >= 0)
+				mask = cpumask_of_node(node);
+			else
+				mask = NULL;
+			break;
+		case X86_VECTOR_POL_DEFAULT:
+			mask = apic->target_cpus();
+			break;
+		case X86_VECTOR_POL_CALLER:
+			if (info && info->mask)
+				mask = info->mask;
+			else
+				mask = NULL;
+			break;
+		default:
+			mask = NULL;
+			break;
+		}
+		if (mask) {
+			err = assign_irq_vector(irq, data, mask);
+			if (!err)
+				return 0;
+		}
+	}
+
+	return err;
+}
+
 static int x86_vector_alloc_irqs(struct irq_domain *domain, unsigned int virq,
 				 unsigned int nr_irqs, void *arg)
 {
 	int i, err;
 	struct apic_chip_data *data;
 	struct irq_data *irq_data;
-	const struct cpumask *mask;
 	struct irq_alloc_info *info = arg;
 
 	if (disable_apic)
@@ -301,7 +349,6 @@  static int x86_vector_alloc_irqs(struct irq_domain *domain, unsigned int virq,
 	if ((info->flags & X86_IRQ_ALLOC_CONTIGOUS_VECTORS) && nr_irqs > 1)
 		return -ENOSYS;
 
-	mask = irq_alloc_info_get_mask(info);
 	for (i = 0; i < nr_irqs; i++) {
 		irq_data = irq_domain_get_irq_data(domain, virq + i);
 		BUG_ON(!irq_data);
@@ -319,7 +366,8 @@  static int x86_vector_alloc_irqs(struct irq_domain *domain, unsigned int virq,
 		irq_data->chip = &lapic_controller;
 		irq_data->chip_data = data;
 		irq_data->hwirq = virq + i;
-		err = assign_irq_vector(virq, data, mask);
+		err = assign_irq_vector_policy(virq, irq_data->node, data,
+					       info);
 		if (err)
 			goto error;
 	}