diff mbox

[v2,-tip,1/5] x86, MSI: Support multiple MSIs in presense of IRQ remapping

Message ID 034a9cbb168501954408f1004a28c827f3a4158b.1346653435.git.agordeev@redhat.com
State Not Applicable
Headers show

Commit Message

Alexander Gordeev Sept. 3, 2012, 9:17 a.m. UTC 
The MSI specification has several constraints in comparison with MSI-X,
most notable of them is the inability to configure MSIs independently.
As a result, it is impossible to dispatch interrupts from different
queues to different CPUs. This is largely devalues the support of
multiple MSIs in SMP systems.

Also, a necessity to allocate a contiguous block of vector numbers for
devices capable of multiple MSIs might cause a considerable pressure on
x86 interrupt vector allocator and could lead to fragmentation of the
interrupt vectors space.

This patch overcomes both drawbacks in presense of IRQ remapping and
lets devices take advantage of multiple queues and per-IRQ affinity
assignments.

Signed-off-by: Alexander Gordeev <agordeev@redhat.com>
---
 arch/x86/kernel/apic/io_apic.c |  166 +++++++++++++++++++++++++++++++++++++--
 include/linux/irq.h            |    6 ++
 kernel/irq/chip.c              |   30 +++++--
 kernel/irq/irqdesc.c           |   31 ++++++++
 4 files changed, 216 insertions(+), 17 deletions(-)

Comments

Yinghai Lu Sept. 3, 2012, 6:53 p.m. UTC | #1 
On Mon, Sep 3, 2012 at 2:17 AM, Alexander Gordeev <agordeev@redhat.com> wrote:
> The MSI specification has several constraints in comparison with MSI-X,
> most notable of them is the inability to configure MSIs independently.
> As a result, it is impossible to dispatch interrupts from different
> queues to different CPUs. This is largely devalues the support of
> multiple MSIs in SMP systems.
>
> Also, a necessity to allocate a contiguous block of vector numbers for
> devices capable of multiple MSIs might cause a considerable pressure on
> x86 interrupt vector allocator and could lead to fragmentation of the
> interrupt vectors space.
>
> This patch overcomes both drawbacks in presense of IRQ remapping and
> lets devices take advantage of multiple queues and per-IRQ affinity
> assignments.
>
> Signed-off-by: Alexander Gordeev <agordeev@redhat.com>
> ---
>  arch/x86/kernel/apic/io_apic.c |  166 +++++++++++++++++++++++++++++++++++++--
>  include/linux/irq.h            |    6 ++
>  kernel/irq/chip.c              |   30 +++++--
>  kernel/irq/irqdesc.c           |   31 ++++++++
>  4 files changed, 216 insertions(+), 17 deletions(-)
>
> diff --git a/arch/x86/kernel/apic/io_apic.c b/arch/x86/kernel/apic/io_apic.c
> index c265593..5fd2577 100644
> --- a/arch/x86/kernel/apic/io_apic.c
> +++ b/arch/x86/kernel/apic/io_apic.c
> @@ -305,6 +305,11 @@ static int alloc_irq_from(unsigned int from, int node)
>         return irq_alloc_desc_from(from, node);
>  }
>
> +static int alloc_irqs_from(unsigned int from, unsigned int count, int node)
> +{
> +       return irq_alloc_descs_from(from, count, node);
> +}
> +
>  static void free_irq_at(unsigned int at, struct irq_cfg *cfg)
>  {
>         free_irq_cfg(at, cfg);
> @@ -3039,6 +3044,55 @@ int create_irq(void)
>         return irq;
>  }
>
> +unsigned int create_irqs(unsigned int from, unsigned int count, int node)
> +{
> +       struct irq_cfg **cfg;
> +       unsigned long flags;
> +       int irq, i;
> +
> +       if (from < nr_irqs_gsi)
> +               from = nr_irqs_gsi;
> +
> +       cfg = kzalloc_node(count * sizeof(cfg[0]), GFP_KERNEL, node);
> +       if (!cfg)
> +               return 0;
> +
> +       irq = alloc_irqs_from(from, count, node);
> +       if (irq < 0)
> +               goto out_cfgs;
> +
> +       for (i = 0; i < count; i++) {
> +               cfg[i] = alloc_irq_cfg(irq + i, node);
> +               if (!cfg[i])
> +                       goto out_irqs;
> +       }
> +
> +       raw_spin_lock_irqsave(&vector_lock, flags);
> +       for (i = 0; i < count; i++)
> +               if (__assign_irq_vector(irq + i, cfg[i], apic->target_cpus()))
> +                       goto out_vecs;
> +       raw_spin_unlock_irqrestore(&vector_lock, flags);
> +
> +       for (i = 0; i < count; i++) {
> +               irq_set_chip_data(irq + i, cfg[i]);
> +               irq_clear_status_flags(irq + i, IRQ_NOREQUEST);
> +       }
> +
> +       kfree(cfg);
> +       return irq;
> +
> +out_vecs:
> +       for (; i; i--)
> +               __clear_irq_vector(irq + i - 1, cfg[i - 1]);
> +       raw_spin_unlock_irqrestore(&vector_lock, flags);
> +out_irqs:
> +       for (i = 0; i < count; i++)
> +               free_irq_at(irq + i, cfg[i]);
> +out_cfgs:
> +       kfree(cfg);
> +       return 0;
> +}
> +

You may update create_irq_nr to be __create_irq_nr, and it could take
extra count.

and later have create_irq_nr to be __create_irq_nr(,1,)
and create_irqs to be __create_irq_nr(,count,)
....

BTW, in short, how much performance benefits for adding 500 lines code?

Thanks

Yinghai
--
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
Alexander Gordeev Sept. 4, 2012, 9:32 a.m. UTC | #2 
On Mon, Sep 03, 2012 at 11:53:39AM -0700, Yinghai Lu wrote:
> On Mon, Sep 3, 2012 at 2:17 AM, Alexander Gordeev <agordeev@redhat.com> wrote:
> You may update create_irq_nr to be __create_irq_nr, and it could take
> extra count.
> 
> and later have create_irq_nr to be __create_irq_nr(,1,)
> and create_irqs to be __create_irq_nr(,count,)
> ....

Indeed. Will do.

> BTW, in short, how much performance benefits for adding 500 lines code?

Unfortunatelly, I do not have a short answer here. There are three types
of performance this series deals with - I'll try to summarize:

- devices - 3 SATA HDDs generate roughly one interrupt every 273 us while
  it get handled in less than 5 us. So there is/could be no increase here;

- the hardware context interrupt handler - its performance dropped 2.5 times
  (little bit more in fact) at the expense of increase of 1.3 times in
  overall interrupt handling time (hardware context + threaded context);

- overall system performance - I *assume* it should increase, because:
  (a) AHCI interrupt handlers keep local interrupts disabled 2.5 times less
  (b) separate AHCI IRQs become subjects of IRQ balancing
  (c) threaded handlers are per-device, per-CPU (well, up to irqbalanced)
      and executed with local interrupts enabled;
diff mbox

Patch

diff --git a/arch/x86/kernel/apic/io_apic.c b/arch/x86/kernel/apic/io_apic.c
index c265593..5fd2577 100644
--- a/arch/x86/kernel/apic/io_apic.c
+++ b/arch/x86/kernel/apic/io_apic.c
@@ -305,6 +305,11 @@  static int alloc_irq_from(unsigned int from, int node)
 	return irq_alloc_desc_from(from, node);
 }
 
+static int alloc_irqs_from(unsigned int from, unsigned int count, int node)
+{
+	return irq_alloc_descs_from(from, count, node);
+}
+
 static void free_irq_at(unsigned int at, struct irq_cfg *cfg)
 {
 	free_irq_cfg(at, cfg);
@@ -3039,6 +3044,55 @@  int create_irq(void)
 	return irq;
 }
 
+unsigned int create_irqs(unsigned int from, unsigned int count, int node)
+{
+	struct irq_cfg **cfg;
+	unsigned long flags;
+	int irq, i;
+
+	if (from < nr_irqs_gsi)
+		from = nr_irqs_gsi;
+
+	cfg = kzalloc_node(count * sizeof(cfg[0]), GFP_KERNEL, node);
+	if (!cfg)
+		return 0;
+
+	irq = alloc_irqs_from(from, count, node);
+	if (irq < 0)
+		goto out_cfgs;
+
+	for (i = 0; i < count; i++) {
+		cfg[i] = alloc_irq_cfg(irq + i, node);
+		if (!cfg[i])
+			goto out_irqs;
+	}
+
+	raw_spin_lock_irqsave(&vector_lock, flags);
+	for (i = 0; i < count; i++)
+		if (__assign_irq_vector(irq + i, cfg[i], apic->target_cpus()))
+			goto out_vecs;
+	raw_spin_unlock_irqrestore(&vector_lock, flags);
+
+	for (i = 0; i < count; i++) {
+		irq_set_chip_data(irq + i, cfg[i]);
+		irq_clear_status_flags(irq + i, IRQ_NOREQUEST);
+	}
+
+	kfree(cfg);
+	return irq;
+
+out_vecs:
+	for (; i; i--)
+		__clear_irq_vector(irq + i - 1, cfg[i - 1]);
+	raw_spin_unlock_irqrestore(&vector_lock, flags);
+out_irqs:
+	for (i = 0; i < count; i++)
+		free_irq_at(irq + i, cfg[i]);
+out_cfgs:
+	kfree(cfg);
+	return 0;
+}
+
 void destroy_irq(unsigned int irq)
 {
 	struct irq_cfg *cfg = irq_get_chip_data(irq);
@@ -3054,6 +3108,27 @@  void destroy_irq(unsigned int irq)
 	free_irq_at(irq, cfg);
 }
 
+static inline void destroy_irqs(unsigned int irq, unsigned int count)
+{
+	unsigned int i;
+	for (i = 0; i < count; i++)
+		destroy_irq(irq + i);
+}
+
+static inline int
+can_create_pow_of_two_irqs(unsigned int from, unsigned int count)
+{
+	if ((count > 1) && (count % 2))
+		return -EINVAL;
+
+	for (; count; count = count / 2) {
+		if (!irq_can_alloc_irqs(from, count))
+			return count;
+	}
+
+	return -ENOSPC;
+}
+
 /*
  * MSI message composition
  */
@@ -3145,18 +3220,25 @@  static struct irq_chip msi_chip = {
 	.irq_retrigger		= ioapic_retrigger_irq,
 };
 
-static int setup_msi_irq(struct pci_dev *dev, struct msi_desc *msidesc, int irq)
+static int setup_msi_irq(struct pci_dev *dev, struct msi_desc *msidesc,
+			 unsigned int irq_base, unsigned int irq_offset)
 {
 	struct irq_chip *chip = &msi_chip;
 	struct msi_msg msg;
+	unsigned int irq = irq_base + irq_offset;
 	int ret;
 
 	ret = msi_compose_msg(dev, irq, &msg, -1);
 	if (ret < 0)
 		return ret;
 
-	irq_set_msi_desc(irq, msidesc);
-	write_msi_msg(irq, &msg);
+	irq_set_msi_desc_off(irq_base, irq_offset, msidesc);
+
+	/* MSI-X message is written per-IRQ, the offset is always 0.
+	 * MSI message denotes a contiguous group of IRQs, written for 0th IRQ.
+	 */
+	if (!irq_offset)
+		write_msi_msg(irq, &msg);
 
 	if (irq_remapped(irq_get_chip_data(irq))) {
 		irq_set_status_flags(irq, IRQ_MOVE_PCNTXT);
@@ -3170,16 +3252,12 @@  static int setup_msi_irq(struct pci_dev *dev, struct msi_desc *msidesc, int irq)
 	return 0;
 }
 
-int native_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
+int setup_msix_irqs(struct pci_dev *dev, int nvec)
 {
 	int node, ret, sub_handle, index = 0;
 	unsigned int irq, irq_want;
 	struct msi_desc *msidesc;
 
-	/* x86 doesn't support multiple MSI yet */
-	if (type == PCI_CAP_ID_MSI && nvec > 1)
-		return 1;
-
 	node = dev_to_node(&dev->dev);
 	irq_want = nr_irqs_gsi;
 	sub_handle = 0;
@@ -3208,7 +3286,7 @@  int native_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
 				goto error;
 		}
 no_ir:
-		ret = setup_msi_irq(dev, msidesc, irq);
+		ret = setup_msi_irq(dev, msidesc, irq, 0);
 		if (ret < 0)
 			goto error;
 		sub_handle++;
@@ -3220,6 +3298,76 @@  error:
 	return ret;
 }
 
+int setup_msi_irqs(struct pci_dev *dev, int nvec)
+{
+	int node, ret, sub_handle, index = 0;
+	unsigned int irq;
+	struct msi_desc *msidesc;
+
+	if (nvec > 1 && !irq_remapping_enabled)
+		return 1;
+
+	nvec = __roundup_pow_of_two(nvec);
+	ret = can_create_pow_of_two_irqs(nr_irqs_gsi, nvec);
+	if (ret != nvec)
+		return ret;
+
+	WARN_ON(!list_is_singular(&dev->msi_list));
+	msidesc = list_entry(dev->msi_list.next, struct msi_desc, list);
+	WARN_ON(msidesc->irq);
+	WARN_ON(msidesc->msi_attrib.multiple);
+
+	node = dev_to_node(&dev->dev);
+	irq = create_irqs(nr_irqs_gsi, nvec, node);
+	if (irq == 0)
+		return -ENOSPC;
+
+	if (!irq_remapping_enabled) {
+		ret = setup_msi_irq(dev, msidesc, irq, 0);
+		if (ret < 0)
+			goto error;
+		return 0;
+	}
+
+	msidesc->msi_attrib.multiple = ilog2(nvec);
+	for (sub_handle = 0; sub_handle < nvec; sub_handle++) {
+		if (!sub_handle) {
+			index = msi_alloc_remapped_irq(dev, irq, nvec);
+			if (index < 0) {
+				ret = index;
+				goto error;
+			}
+		} else {
+			ret = msi_setup_remapped_irq(dev, irq + sub_handle,
+						     index, sub_handle);
+			if (ret < 0)
+				goto error;
+		}
+		ret = setup_msi_irq(dev, msidesc, irq, sub_handle);
+		if (ret < 0)
+			goto error;
+	}
+	return 0;
+
+error:
+	destroy_irqs(irq, nvec);
+
+	/* Restore altered MSI descriptor fields and prevent just destroyed
+	 * IRQs from tearing down again in default_teardown_msi_irqs()
+	 */
+	msidesc->irq = 0;
+	msidesc->msi_attrib.multiple = 0;
+
+	return ret;
+}
+
+int native_setup_msi_irqs(struct pci_dev *dev, int nvec, int type)
+{
+	if (type == PCI_CAP_ID_MSI)
+		return setup_msi_irqs(dev, nvec);
+	return setup_msix_irqs(dev, nvec);
+}
+
 void native_teardown_msi_irq(unsigned int irq)
 {
 	destroy_irq(irq);
diff --git a/include/linux/irq.h b/include/linux/irq.h
index 216b0ba..c3ba39f 100644
--- a/include/linux/irq.h
+++ b/include/linux/irq.h
@@ -522,6 +522,8 @@  extern int irq_set_handler_data(unsigned int irq, void *data);
 extern int irq_set_chip_data(unsigned int irq, void *data);
 extern int irq_set_irq_type(unsigned int irq, unsigned int type);
 extern int irq_set_msi_desc(unsigned int irq, struct msi_desc *entry);
+extern int irq_set_msi_desc_off(unsigned int irq_base, unsigned int irq_offset,
+				struct msi_desc *entry);
 extern struct irq_data *irq_get_irq_data(unsigned int irq);
 
 static inline struct irq_chip *irq_get_chip(unsigned int irq)
@@ -584,8 +586,12 @@  int __irq_alloc_descs(int irq, unsigned int from, unsigned int cnt, int node,
 #define irq_alloc_desc_from(from, node)		\
 	irq_alloc_descs(-1, from, 1, node)
 
+#define irq_alloc_descs_from(from, cnt, node)	\
+	irq_alloc_descs(-1, from, cnt, node)
+
 void irq_free_descs(unsigned int irq, unsigned int cnt);
 int irq_reserve_irqs(unsigned int from, unsigned int cnt);
+int irq_can_alloc_irqs(unsigned int from, unsigned int cnt);
 
 static inline void irq_free_desc(unsigned int irq)
 {
diff --git a/kernel/irq/chip.c b/kernel/irq/chip.c
index 57d86d0..2230389 100644
--- a/kernel/irq/chip.c
+++ b/kernel/irq/chip.c
@@ -90,27 +90,41 @@  int irq_set_handler_data(unsigned int irq, void *data)
 EXPORT_SYMBOL(irq_set_handler_data);
 
 /**
- *	irq_set_msi_desc - set MSI descriptor data for an irq
- *	@irq:	Interrupt number
- *	@entry:	Pointer to MSI descriptor data
+ *	irq_set_msi_desc_off - set MSI descriptor data for an irq at offset
+ *	@irq_base:	Interrupt number base
+ *	@irq_offset:	Interrupt number offset
+ *	@entry:		Pointer to MSI descriptor data
  *
- *	Set the MSI descriptor entry for an irq
+ *	Set the MSI descriptor entry for an irq at offset
  */
-int irq_set_msi_desc(unsigned int irq, struct msi_desc *entry)
+int irq_set_msi_desc_off(unsigned int irq_base, unsigned int irq_offset,
+			 struct msi_desc *entry)
 {
 	unsigned long flags;
-	struct irq_desc *desc = irq_get_desc_lock(irq, &flags, IRQ_GET_DESC_CHECK_GLOBAL);
+	struct irq_desc *desc = irq_get_desc_lock(irq_base + irq_offset, &flags, IRQ_GET_DESC_CHECK_GLOBAL);
 
 	if (!desc)
 		return -EINVAL;
 	desc->irq_data.msi_desc = entry;
-	if (entry)
-		entry->irq = irq;
+	if (entry && !irq_offset)
+		entry->irq = irq_base;
 	irq_put_desc_unlock(desc, flags);
 	return 0;
 }
 
 /**
+ *	irq_set_msi_desc - set MSI descriptor data for an irq
+ *	@irq:	Interrupt number
+ *	@entry:	Pointer to MSI descriptor data
+ *
+ *	Set the MSI descriptor entry for an irq
+ */
+int irq_set_msi_desc(unsigned int irq, struct msi_desc *entry)
+{
+	return irq_set_msi_desc_off(irq, 0, entry);
+}
+
+/**
  *	irq_set_chip_data - set irq chip data for an irq
  *	@irq:	Interrupt number
  *	@data:	Pointer to chip specific data
diff --git a/kernel/irq/irqdesc.c b/kernel/irq/irqdesc.c
index 192a302..8287b78 100644
--- a/kernel/irq/irqdesc.c
+++ b/kernel/irq/irqdesc.c
@@ -210,6 +210,13 @@  static int irq_expand_nr_irqs(unsigned int nr)
 	return 0;
 }
 
+static int irq_can_expand_nr_irqs(unsigned int nr)
+{
+	if (nr > IRQ_BITMAP_BITS)
+		return -ENOMEM;
+	return 0;
+}
+
 int __init early_irq_init(void)
 {
 	int i, initcnt, node = first_online_node;
@@ -414,6 +421,30 @@  int irq_reserve_irqs(unsigned int from, unsigned int cnt)
 }
 
 /**
+ * irq_can_alloc_irqs - checks if a range of irqs could be allocated
+ * @from:	check from irq number
+ * @cnt:	number of irqs to check
+ *
+ * Returns 0 on success or an appropriate error code
+ */
+int irq_can_alloc_irqs(unsigned int from, unsigned int cnt)
+{
+	unsigned int start;
+	int ret = 0;
+
+	if (!cnt)
+		return -EINVAL;
+
+	mutex_lock(&sparse_irq_lock);
+	start = bitmap_find_next_zero_area(allocated_irqs, IRQ_BITMAP_BITS,
+					   from, cnt, 0);
+	mutex_unlock(&sparse_irq_lock);
+	if (start + cnt > nr_irqs)
+		ret = irq_can_expand_nr_irqs(start + cnt);
+	return ret;
+}
+
+/**
  * irq_get_next_irq - get next allocated irq number
  * @offset:	where to start the search
  *