Skip to contentSkip to conversation

PCI: hv: Fix NUMA node assignment when kernel boots with custom NUMA topology

PATCH Superseded [email protected] 3 participants 3 messages Jan 19, 2022

Subsystems (4)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 PCI: hv: Fix NUMA node assignment when kernel boots with custom NUMA topology

From: Long Li <[email protected]>

When kernel boots with a NUMA topology with some NUMA nodes offline, the PCI driver should only set an online NUMA node on the device. This can happen during KDUMP where some NUMA nodes are not made online by the KDUMP kernel.

This patch also fixes the case where kernel is booting with "numa=off".

Change from v1:
Use numa_map_to_online_node() to assign a node to device (suggested by Michael Kelly <[email protected]>)

---
 drivers/pci/controller/pci-hyperv.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/pci/controller/pci-hyperv.c +9 -1
index 6c9efeefae1b..c7519add6f13 100644
@@ -2130,7 +2130,15 @@ static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus)
 			continue;
 
 		if (hv_dev->desc.flags & HV_PCI_DEVICE_FLAG_NUMA_AFFINITY)
-			set_dev_node(&dev->dev, hv_dev->desc.virtual_numa_node);
+			/*
+			 * The kernel may boot with some NUMA nodes offline
+			 * (e.g. in a KDUMP kernel) or with NUMA disabled via
+			 * "numa=off". In those cases, adjust the host provided
+			 * NUMA node to a valid NUMA node used by the kernel.
+			 */
+			set_dev_node(&dev->dev,
+				     numa_map_to_online_node(
+					     hv_dev->desc.virtual_numa_node));
 
 		put_pcichild(hv_dev);
 	}
Signature · 1 line
2.25.1
#1 lore raw

From: [email protected] <[email protected]> Sent: Tuesday, January 18, 2022 6:45 PM

Quoted 8 lines

When kernel boots with a NUMA topology with some NUMA nodes offline, the PCI driver should only set an online NUMA node on the device. This can happen during KDUMP where some NUMA nodes are not made online by the KDUMP kernel.

This patch also fixes the case where kernel is booting with "numa=off".

It seems like adding a "Fixes:" tag would be appropriate.

Quoted 18 lines

Change from v1:
Use numa_map_to_online_node() to assign a node to device (suggested by Michael Kelly <[email protected]>)

---
 drivers/pci/controller/pci-hyperv.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/pci/controller/pci-hyperv.c +0 -0
index 6c9efeefae1b..c7519add6f13 100644
@@ -2130,7 +2130,15 @@ static void hv_pci_assign_numa_node(struct
 hv_pcibus_device *hbus)
 			continue;
 
 		if (hv_dev->desc.flags & HV_PCI_DEVICE_FLAG_NUMA_AFFINITY)

drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct · line 2142

Quoted 10 lines
drivers/pci/controller/pci-hyperv.c +9 -1
-			set_dev_node(&dev->dev, hv_dev->desc.virtual_numa_node);
+			/*
+			 * The kernel may boot with some NUMA nodes offline
+			 * (e.g. in a KDUMP kernel) or with NUMA disabled via
+			 * "numa=off". In those cases, adjust the host provided
+			 * NUMA node to a valid NUMA node used by the kernel.
+			 */
+			set_dev_node(&dev->dev,
+				     numa_map_to_online_node(
+					     hv_dev->desc.virtual_numa_node));

Double-check me, but I think this approach has a flaw in that numa_map_to_online_node() doesn't check the input node for being out-of-range. The call to node_online() uses the input node to index into the node_states bitmap (of type nodemask_t), and could go off the end of the bitmap if the input node isn't validated to be less than MAX_NUMNODES (or nr_node_ids).

Michael

Quoted 5 lines
drivers/pci/controller/pci-hyperv.c +0 -1
 		put_pcichild(hv_dev);
 	}
--

2.25.1

#2 lore raw
Quoted 9 lines

Subject: RE: [Patch v2] PCI: hv: Fix NUMA node assignment when kernel boots with custom NUMA topology

From: [email protected] <[email protected]> Sent: Tuesday, January 18, 2022 6:45 PM

Quoted 4 lines

When kernel boots with a NUMA topology with some NUMA nodes offline, the PCI driver should only set an online NUMA node on the device. This can happen during KDUMP where some NUMA nodes are not made online by

Quoted 7 lines

the KDUMP kernel.

Quoted 4 lines

This patch also fixes the case where kernel is booting with "numa=off".

It seems like adding a "Fixes:" tag would be appropriate.

Okay, sending v3 with "Fixes:".

Quoted 31 lines
Quoted 16 lines

Change from v1:
Use numa_map_to_online_node() to assign a node to device (suggested by Michael Kelly <[email protected]>)

---
 drivers/pci/controller/pci-hyperv.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
drivers/pci/controller/pci-hyperv.c +0 -0
index 6c9efeefae1b..c7519add6f13 100644

drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct · line 2133

Quoted 4 lines
drivers/pci/controller/pci-hyperv.c +0 -0
@@ -2130,7 +2130,15 @@ static void hv_pci_assign_numa_node(struct
 hv_pcibus_device *hbus)
 			continue;
 
 		if (hv_dev->desc.flags &

HV_PCI_DEVICE_FLAG_NUMA_AFFINITY)

Quoted 10 lines
drivers/pci/controller/pci-hyperv.c +8 -1
-			set_dev_node(&dev->dev, hv_dev-
 desc.virtual_numa_node);
+			/*
+			 * The kernel may boot with some NUMA nodes offline
+			 * (e.g. in a KDUMP kernel) or with NUMA disabled via
+			 * "numa=off". In those cases, adjust the host provided
+			 * NUMA node to a valid NUMA node used by the kernel.
+			 */
+			set_dev_node(&dev->dev,
+				     numa_map_to_online_node(
Quoted 8 lines

drivers/pci/controller/pci-hyperv.c static void hv_pci_assign_numa_node(struct hv_pcibus_device *hbus) · line 2141

Quoted 1 line
drivers/pci/controller/pci-hyperv.c +1 -0
+					     hv_dev->desc.virtual_numa_node));

Double-check me, but I think this approach has a flaw in that numa_map_to_online_node() doesn't check the input node for being out-of- range. The call to node_online() uses the input node to index into the node_states bitmap (of type nodemask_t), and could go off the end of the bitmap if the input node isn't validated to be less than MAX_NUMNODES (or nr_node_ids).

Indeed, your concern is correct when "numa=off" is set.

Quoted 8 lines

Michael

Quoted 5 lines
drivers/pci/controller/pci-hyperv.c +0 -0
 		put_pcichild(hv_dev);

}
--
2.25.1