When we use "nr_cpus=n" to hard limit the CPU number, cpu_to_node() is
not usable for CPUs exceeding the limit. Because cpu_to_node() depends
on the per-cpu area, and the per-cpu area is only applied on "possible"
CPUs. On the other hand, early_cpu_to_node() is always usable because it
uses plain arrays to record the cpu-node mapping, and the architectural
code ensures those arrays are correctly initialized. So here we can use
early_cpu_to_node() instead.

This change is mainly needed on multi-bridge machines. On these machines
there are multiple eiointc instances, each instance should get its NUMA
node correctly.

Note: we make this change for kdump and some other special usages where
"nr_cpus=n" is needed. However, if CONFIG_NR_CPUS is not big enough to
map all the nodes connect I/O bridges, then the machine can't boot still.

Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
---
V2: Fix build for !NUMA.
V3: Adjust header file order and update commit messages.
V4: Update commit messages again.

 drivers/irqchip/irq-loongson-eiointc.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/irqchip/irq-loongson-eiointc.c b/drivers/irqchip/irq-loongson-eiointc.c
index c7ddebf312ad..b1f2080be2be 100644
--- a/drivers/irqchip/irq-loongson-eiointc.c
+++ b/drivers/irqchip/irq-loongson-eiointc.c
@@ -15,6 +15,7 @@
 #include <linux/irqchip/chained_irq.h>
 #include <linux/kernel.h>
 #include <linux/syscore_ops.h>
+#include <asm/numa.h>
 
 #define EIOINTC_REG_NODEMAP	0x14a0
 #define EIOINTC_REG_IPMAP	0x14c0
@@ -339,7 +340,7 @@ static int __init pch_msi_parse_madt(union acpi_subtable_headers *header,
 	int node;
 
 	if (cpu_has_flatmode)
-		node = cpu_to_node(eiointc_priv[nr_pics - 1]->node * CORES_PER_EIO_NODE);
+		node = early_cpu_to_node(eiointc_priv[nr_pics - 1]->node * CORES_PER_EIO_NODE);
 	else
 		node = eiointc_priv[nr_pics - 1]->node;
 
@@ -431,7 +432,7 @@ int __init eiointc_acpi_init(struct irq_domain *parent,
 		goto out_free_handle;
 
 	if (cpu_has_flatmode)
-		node = cpu_to_node(acpi_eiointc->node * CORES_PER_EIO_NODE);
+		node = early_cpu_to_node(acpi_eiointc->node * CORES_PER_EIO_NODE);
 	else
 		node = acpi_eiointc->node;
 	acpi_set_vec_parent(node, priv->eiointc_domain, pch_group);
-- 
2.43.0

The following commit has been merged into the irq/urgent branch of tip:

Commit-ID:     2d64eaeeeda5659d52da1af79d237269ba3c2d2c
Gitweb:        https://git.kernel.org/tip/2d64eaeeeda5659d52da1af79d237269ba3c2d2c
Author:        Huacai Chen <chenhuacai@loongson.cn>
AuthorDate:    Sun, 23 Jun 2024 11:41:13 +08:00
Committer:     Thomas Gleixner <tglx@linutronix.de>
CommitterDate: Sun, 23 Jun 2024 17:09:26 +02:00

irqchip/loongson-eiointc: Use early_cpu_to_node() instead of cpu_to_node()

Multi-bridge machines required that all eiointc controllers in the system
are initialized, otherwise the system does not boot.

The initialization happens on the boot CPU during early boot and relies on
cpu_to_node() for identifying the individual nodes.

That works when the number of possible CPUs is large enough, but with a
command line limit, e.g. "nr_cpus=$N" for kdump, but fails when the CPUs
of the secondary nodes are not covered.

During early ACPI enumeration all CPU to node mappings are recorded up to
CONFIG_NR_CPUS. These are accessible via early_cpu_to_node() even in the
case that "nr_cpus=N" truncates the number of possible CPUs and only
provides the possible CPUs via cpu_to_node() translation.

Change the node lookup in the driver to use early_cpu_to_node() so that
even with a limitation on the number of possible CPUs all eointc instances
are initialized.

This can't obviously cure the case where CONFIG_NR_CPUS is too small.

[ tglx: Massaged changelog ]

Fixes: 64cc451e45e1 ("irqchip/loongson-eiointc: Fix incorrect use of acpi_get_vec_parent")
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20240623034113.1808727-1-chenhuacai@loongson.cn
---
 drivers/irqchip/irq-loongson-eiointc.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/irqchip/irq-loongson-eiointc.c b/drivers/irqchip/irq-loongson-eiointc.c
index c7ddebf..b1f2080 100644
--- a/drivers/irqchip/irq-loongson-eiointc.c
+++ b/drivers/irqchip/irq-loongson-eiointc.c
@@ -15,6 +15,7 @@
 #include <linux/irqchip/chained_irq.h>
 #include <linux/kernel.h>
 #include <linux/syscore_ops.h>
+#include <asm/numa.h>
 
 #define EIOINTC_REG_NODEMAP	0x14a0
 #define EIOINTC_REG_IPMAP	0x14c0
@@ -339,7 +340,7 @@ static int __init pch_msi_parse_madt(union acpi_subtable_headers *header,
 	int node;
 
 	if (cpu_has_flatmode)
-		node = cpu_to_node(eiointc_priv[nr_pics - 1]->node * CORES_PER_EIO_NODE);
+		node = early_cpu_to_node(eiointc_priv[nr_pics - 1]->node * CORES_PER_EIO_NODE);
 	else
 		node = eiointc_priv[nr_pics - 1]->node;
 
@@ -431,7 +432,7 @@ int __init eiointc_acpi_init(struct irq_domain *parent,
 		goto out_free_handle;
 
 	if (cpu_has_flatmode)
-		node = cpu_to_node(acpi_eiointc->node * CORES_PER_EIO_NODE);
+		node = early_cpu_to_node(acpi_eiointc->node * CORES_PER_EIO_NODE);
 	else
 		node = acpi_eiointc->node;
 	acpi_set_vec_parent(node, priv->eiointc_domain, pch_group);