[PATCH] SGI XPC fails to load when cpu 0 is out of IRQ resources.

[PATCH] SGI XPC fails to load when cpu 0 is out of IRQ resources.

[Robin Holt]

On many of our larger systems, CPU 0 has had all of its IRQ resources
consumed before XPC loads.  Worse cases on machines with multiple
10 GigE cards and multiple IB cards have depleted the entire first
socket of IRQs.  That patch makes selecting the node upon which
IRQs are allocated (as well as all the other GRU Message Queue
structures) specifiable as a module load param and has a default
behavior of searching all nodes/cpus for an available resource.

Signed-off-by: Robin Holt <holt@sgi.com>
---
 drivers/misc/sgi-xp/xpc_uv.c |   66 ++++++++++++++++++++++++++++++++---------
 1 files changed, 51 insertions(+), 15 deletions(-)

diff --git a/drivers/misc/sgi-xp/xpc_uv.c b/drivers/misc/sgi-xp/xpc_uv.c
index 87b251a..f4398bb 100644
--- a/drivers/misc/sgi-xp/xpc_uv.c
+++ b/drivers/misc/sgi-xp/xpc_uv.c
@@ -59,6 +59,8 @@ static struct xpc_heartbeat_uv *xpc_heartbeat_uv;
 					 XPC_NOTIFY_MSG_SIZE_UV)
 #define XPC_NOTIFY_IRQ_NAME		"xpc_notify"
 
+static int xpc_mq_node = -1;
+
 static struct xpc_gru_mq_uv *xpc_activate_mq_uv;
 static struct xpc_gru_mq_uv *xpc_notify_mq_uv;
 
@@ -110,8 +112,6 @@ xpc_get_gru_mq_irq_uv(struct xpc_gru_mq_uv *mq, int cpu, char *irq_name)
 	mq->irq = uv_setup_irq(irq_name, cpu, mq->mmr_blade, mq->mmr_offset,
 			UV_AFFINITY_CPU);
 	if (mq->irq < 0) {
-		dev_err(xpc_part, "uv_setup_irq() returned error=%d\n",
-			-mq->irq);
 		return mq->irq;
 	}
 
@@ -1731,9 +1731,42 @@ static struct xpc_arch_operations xpc_arch_ops_uv = {
 	.notify_senders_of_disconnect = xpc_notify_senders_of_disconnect_uv,
 };
 
+static int
+xpc_init_mq_node(int nid)
+{
+	int cpu;
+
+	for_each_cpu(cpu, cpumask_of_node(nid)) {
+		xpc_activate_mq_uv = xpc_create_gru_mq_uv(XPC_ACTIVATE_MQ_SIZE_UV, nid,
+							  XPC_ACTIVATE_IRQ_NAME,
+							  xpc_handle_activate_IRQ_uv);
+		if (!IS_ERR(xpc_activate_mq_uv))
+			break;
+	}
+	if (IS_ERR(xpc_activate_mq_uv))
+		return PTR_ERR(xpc_activate_mq_uv);
+
+	for_each_cpu(cpu, cpumask_of_node(nid)) {
+		xpc_notify_mq_uv = xpc_create_gru_mq_uv(XPC_NOTIFY_MQ_SIZE_UV, nid,
+							XPC_NOTIFY_IRQ_NAME,
+							xpc_handle_notify_IRQ_uv);
+		if (!IS_ERR(xpc_notify_mq_uv))
+			break;
+	}
+	if (IS_ERR(xpc_notify_mq_uv)) {
+		xpc_destroy_gru_mq_uv(xpc_activate_mq_uv);
+		return PTR_ERR(xpc_notify_mq_uv);
+	}
+
+	return 0;
+}
+
 int
 xpc_init_uv(void)
 {
+	int nid;
+	int ret = 0;
+
 	xpc_arch_ops = xpc_arch_ops_uv;
 
 	if (sizeof(struct xpc_notify_mq_msghdr_uv) > XPC_MSG_HDR_MAX_SIZE) {
@@ -1742,21 +1775,21 @@ xpc_init_uv(void)
 		return -E2BIG;
 	}
 
-	xpc_activate_mq_uv = xpc_create_gru_mq_uv(XPC_ACTIVATE_MQ_SIZE_UV, 0,
-						  XPC_ACTIVATE_IRQ_NAME,
-						  xpc_handle_activate_IRQ_uv);
-	if (IS_ERR(xpc_activate_mq_uv))
-		return PTR_ERR(xpc_activate_mq_uv);
+	if (xpc_mq_node < 0)
+		for_each_online_node(nid) {
+			ret = xpc_init_mq_node(nid);
 
-	xpc_notify_mq_uv = xpc_create_gru_mq_uv(XPC_NOTIFY_MQ_SIZE_UV, 0,
-						XPC_NOTIFY_IRQ_NAME,
-						xpc_handle_notify_IRQ_uv);
-	if (IS_ERR(xpc_notify_mq_uv)) {
-		xpc_destroy_gru_mq_uv(xpc_activate_mq_uv);
-		return PTR_ERR(xpc_notify_mq_uv);
-	}
+			if (!ret)
+				break;
+		}
+	else
+		ret = xpc_init_mq_node(xpc_mq_node);
 
-	return 0;
+	if (ret < 0)
+		dev_err(xpc_part, "xpc_init_mq_node() returned error=%d\n",
+			-ret);
+
+	return ret;
 }
 
 void
@@ -1765,3 +1798,6 @@ xpc_exit_uv(void)
 	xpc_destroy_gru_mq_uv(xpc_notify_mq_uv);
 	xpc_destroy_gru_mq_uv(xpc_activate_mq_uv);
 }
+
+module_param(xpc_mq_node, int, 0);
+MODULE_PARM_DESC(xpc_mq_node, "Node number on which to allocate message queues.");
-- 
1.7.6.1

Re: [PATCH] SGI XPC fails to load when cpu 0 is out of IRQ resources.

[Andrew Morton]

On Fri, 3 Aug 2012 14:46:29 -0500
Robin Holt <holt@sgi.com> wrote:

> On many of our larger systems, CPU 0 has had all of its IRQ resources
> consumed before XPC loads.  Worse cases on machines with multiple
> 10 GigE cards and multiple IB cards have depleted the entire first
> socket of IRQs.  That patch makes selecting the node upon which
> IRQs are allocated (as well as all the other GRU Message Queue
> structures) specifiable as a module load param and has a default
> behavior of searching all nodes/cpus for an available resource.
> 

Is this problem serious enough to warrant a -stable backport?  If you
want it to appear in vendor kernels then I guess "yes".

> +static int
> +xpc_init_mq_node(int nid)
> +{
> +	int cpu;
> +
> +	for_each_cpu(cpu, cpumask_of_node(nid)) {
> +		xpc_activate_mq_uv = xpc_create_gru_mq_uv(XPC_ACTIVATE_MQ_SIZE_UV, nid,
> +							  XPC_ACTIVATE_IRQ_NAME,
> +							  xpc_handle_activate_IRQ_uv);
> +		if (!IS_ERR(xpc_activate_mq_uv))
> +			break;
> +	}
> +	if (IS_ERR(xpc_activate_mq_uv))
> +		return PTR_ERR(xpc_activate_mq_uv);
> +
> +	for_each_cpu(cpu, cpumask_of_node(nid)) {
> +		xpc_notify_mq_uv = xpc_create_gru_mq_uv(XPC_NOTIFY_MQ_SIZE_UV, nid,
> +							XPC_NOTIFY_IRQ_NAME,
> +							xpc_handle_notify_IRQ_uv);
> +		if (!IS_ERR(xpc_notify_mq_uv))
> +			break;
> +	}
> +	if (IS_ERR(xpc_notify_mq_uv)) {
> +		xpc_destroy_gru_mq_uv(xpc_activate_mq_uv);
> +		return PTR_ERR(xpc_notify_mq_uv);
> +	}
> +
> +	return 0;
> +}

This seems to take the optimistic approach to CPU hotplug ;)
get_online_cpus(), perhaps?

[PATCH] SGI XPC fails to load when cpu 0 is out of IRQ resources. -v2

[Robin Holt]

On many of our larger systems, CPU 0 has had all of its IRQ resources
consumed before XPC loads.  Worst cases on machines with multiple
10 GigE cards and multiple IB cards have depleted the entire first
socket of IRQs.  This patch makes selecting the node upon which
IRQs are allocated (as well as all the other GRU Message Queue
structures) specifiable as a module load param and has a default
behavior of searching all nodes/cpus for an available resources.

Signed-off-by: Robin Holt <holt@sgi.com>
Cc: stable@vger.kernel.org

---

-v2 - incorporate review comments from Andrew Morton.

 drivers/misc/sgi-xp/xpc_uv.c |   82 ++++++++++++++++++++++++++++++++----------
 1 files changed, 63 insertions(+), 19 deletions(-)

diff --git a/drivers/misc/sgi-xp/xpc_uv.c b/drivers/misc/sgi-xp/xpc_uv.c
index 87b251a..57a53c1 100644
--- a/drivers/misc/sgi-xp/xpc_uv.c
+++ b/drivers/misc/sgi-xp/xpc_uv.c
@@ -59,6 +59,8 @@ static struct xpc_heartbeat_uv *xpc_heartbeat_uv;
 					 XPC_NOTIFY_MSG_SIZE_UV)
 #define XPC_NOTIFY_IRQ_NAME		"xpc_notify"
 
+static int xpc_mq_node = -1;
+
 static struct xpc_gru_mq_uv *xpc_activate_mq_uv;
 static struct xpc_gru_mq_uv *xpc_notify_mq_uv;
 
@@ -109,11 +111,8 @@ xpc_get_gru_mq_irq_uv(struct xpc_gru_mq_uv *mq, int cpu, char *irq_name)
 #if defined CONFIG_X86_64
 	mq->irq = uv_setup_irq(irq_name, cpu, mq->mmr_blade, mq->mmr_offset,
 			UV_AFFINITY_CPU);
-	if (mq->irq < 0) {
-		dev_err(xpc_part, "uv_setup_irq() returned error=%d\n",
-			-mq->irq);
+	if (mq->irq < 0)
 		return mq->irq;
-	}
 
 	mq->mmr_value = uv_read_global_mmr64(mmr_pnode, mq->mmr_offset);
 
@@ -238,8 +237,9 @@ xpc_create_gru_mq_uv(unsigned int mq_size, int cpu, char *irq_name,
 	mq->mmr_blade = uv_cpu_to_blade_id(cpu);
 
 	nid = cpu_to_node(cpu);
-	page = alloc_pages_exact_node(nid, GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
-				pg_order);
+	page = alloc_pages_exact_node(nid,
+				      GFP_KERNEL | __GFP_ZERO | GFP_THISNODE,
+				      pg_order);
 	if (page == NULL) {
 		dev_err(xpc_part, "xpc_create_gru_mq_uv() failed to alloc %d "
 			"bytes of memory on nid=%d for GRU mq\n", mq_size, nid);
@@ -1731,9 +1731,50 @@ static struct xpc_arch_operations xpc_arch_ops_uv = {
 	.notify_senders_of_disconnect = xpc_notify_senders_of_disconnect_uv,
 };
 
+static int
+xpc_init_mq_node(int nid)
+{
+	int cpu;
+
+	get_online_cpus();
+
+	for_each_cpu(cpu, cpumask_of_node(nid)) {
+		xpc_activate_mq_uv =
+			xpc_create_gru_mq_uv(XPC_ACTIVATE_MQ_SIZE_UV, nid,
+					     XPC_ACTIVATE_IRQ_NAME,
+					     xpc_handle_activate_IRQ_uv);
+		if (!IS_ERR(xpc_activate_mq_uv))
+			break;
+	}
+	if (IS_ERR(xpc_activate_mq_uv)) {
+		put_online_cpus();
+		return PTR_ERR(xpc_activate_mq_uv);
+	}
+
+	for_each_cpu(cpu, cpumask_of_node(nid)) {
+		xpc_notify_mq_uv =
+			xpc_create_gru_mq_uv(XPC_NOTIFY_MQ_SIZE_UV, nid,
+					     XPC_NOTIFY_IRQ_NAME,
+					     xpc_handle_notify_IRQ_uv);
+		if (!IS_ERR(xpc_notify_mq_uv))
+			break;
+	}
+	if (IS_ERR(xpc_notify_mq_uv)) {
+		xpc_destroy_gru_mq_uv(xpc_activate_mq_uv);
+		put_online_cpus();
+		return PTR_ERR(xpc_notify_mq_uv);
+	}
+
+	put_online_cpus();
+	return 0;
+}
+
 int
 xpc_init_uv(void)
 {
+	int nid;
+	int ret = 0;
+
 	xpc_arch_ops = xpc_arch_ops_uv;
 
 	if (sizeof(struct xpc_notify_mq_msghdr_uv) > XPC_MSG_HDR_MAX_SIZE) {
@@ -1742,21 +1783,21 @@ xpc_init_uv(void)
 		return -E2BIG;
 	}
 
-	xpc_activate_mq_uv = xpc_create_gru_mq_uv(XPC_ACTIVATE_MQ_SIZE_UV, 0,
-						  XPC_ACTIVATE_IRQ_NAME,
-						  xpc_handle_activate_IRQ_uv);
-	if (IS_ERR(xpc_activate_mq_uv))
-		return PTR_ERR(xpc_activate_mq_uv);
+	if (xpc_mq_node < 0)
+		for_each_online_node(nid) {
+			ret = xpc_init_mq_node(nid);
 
-	xpc_notify_mq_uv = xpc_create_gru_mq_uv(XPC_NOTIFY_MQ_SIZE_UV, 0,
-						XPC_NOTIFY_IRQ_NAME,
-						xpc_handle_notify_IRQ_uv);
-	if (IS_ERR(xpc_notify_mq_uv)) {
-		xpc_destroy_gru_mq_uv(xpc_activate_mq_uv);
-		return PTR_ERR(xpc_notify_mq_uv);
-	}
+			if (!ret)
+				break;
+		}
+	else
+		ret = xpc_init_mq_node(xpc_mq_node);
 
-	return 0;
+	if (ret < 0)
+		dev_err(xpc_part, "xpc_init_mq_node() returned error=%d\n",
+			-ret);
+
+	return ret;
 }
 
 void
@@ -1765,3 +1806,6 @@ xpc_exit_uv(void)
 	xpc_destroy_gru_mq_uv(xpc_notify_mq_uv);
 	xpc_destroy_gru_mq_uv(xpc_activate_mq_uv);
 }
+
+module_param(xpc_mq_node, int, 0);
+MODULE_PARM_DESC(xpc_mq_node, "Node number on which to allocate message queues.");
-- 
1.7.6.1