[PATCH v3 0/2] irqchip/econet-en751221: Support MIPS 34Kc VEIC mode

[PATCH v3 0/2] irqchip/econet-en751221: Support MIPS 34Kc VEIC mode

[Caleb James DeLisle]

MIPS 34Kc and 1004Kc have a Vectored External Interrupt mode, where
CPU interrupts are re-routed to the external interrupt controller,
which prioritizes them, renumbers them, and integrates them with its
own, creating a flat interrupt table.

This bypasses the CPU interrupt controller entirely, which is not ideal
for modern Linux which formalizes the interrupt hierarchy. The 1004Kc
standardized the interrupt controller (MIPS_GIC) so it can be viewed as
"part of the CPU" and it's tolerable for it to be more closely coupled
to arch/mips/* than a typical interrupt controller driver.

The 34Kc supports VEIC mode, but the interrupt controller is not
standardized, so it's different per-SoC. This creates a challenge of
writing a reasonably modular driver, given hardware that actually does
take over the interrupt hierarchy.

Ordinarily we wouldn't bother with VEIC on the 34Kc but it is required
for MIPS_MT_SMP, so without it you get a single thread processor. The
only other 34Kc device which has an in-tree DTS file is
realtek,rtl9302-soc, in OpenWrt there is this, realtek,rtl839x-soc, and
lantiq,xway. Of these, the realtek,rtl839x-soc has multi-thread support
via an out-of-tree patch to its interrupt controller, and the lantiq
is not 34Kc compliant in that it lacks the VEIC cpu flag, causing linux
to fall back on legacy interrupts despite CPU_MIPSR2_IRQ_EI being set.

This patchset aims to tackle this challenge in a way that is most
likely generalizable to other 34Kc interrupt controllers which are
facing the same problem.

Changes from v2:
* Fix goofup that changes were not actually in the patch...
* v2: https://lore.kernel.org/linux-mips/20260430150809.1827881-1-cjd@cjdns.fr/

Changes from v1:
* Address notes from Thomas Gleixner (Thank you very much!)
* Additional info in cover letter regarding lantiq,xway
* v1: https://lore.kernel.org/linux-mips/20260425123531.270548-1-cjd@cjdns.fr/

Caleb James DeLisle (2):
  dt-bindings: interrupt-controller: econet: Add CPU interrupt mapping
  irqchip/econet-en751221: Support MIPS 34Kc VEIC mode

 .../econet,en751221-intc.yaml                 |  20 ++
 drivers/irqchip/irq-econet-en751221.c         | 187 +++++++++++++++++-
 2 files changed, 199 insertions(+), 8 deletions(-)


base-commit: e8be82c2d77ec1bb0148406e54b105028a83537e
-- 
2.39.5

[PATCH v3 1/2] dt-bindings: interrupt-controller: econet: Add CPU interrupt mapping

[Caleb James DeLisle]

In MIPS VEIC mode (Vectored External Interrupt Controller), the
hardware stops directly dispatching CPU interrupts such as IPIs or CPU
performance counters, and instead it communicates them to the external
interrupt controller (the hardware described here) which prioritizes,
renumbers, and integrates them with its own hardware interrupt pins.
Interrupts from the external controller are then dispatched through a
different method via a dispatch table. In effect, the external
controller subsumes the CPU controller and becomes the root.

34K Manual (MD00534) Section 6.3.1.3 rev 1.13 page 136

Since there are interrupts which ought to be controlled by the CPU
controller driver - particularly the IPI interrupts - we create a
reverse mapping where those interrupts may be sent back to the CPU
intc when they are received. This maintains the fiction that there is
still a hierarchy, and keeps the DT the same no matter whether the
processor is in VEIC mode or not. The econet,cpu-interrupt-map is
optional and if omitted, it's assumed that no interrupts need to be
mapped.

Signed-off-by: Caleb James DeLisle <cjd@cjdns.fr>
---
 .../econet,en751221-intc.yaml                 | 20 +++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/Documentation/devicetree/bindings/interrupt-controller/econet,en751221-intc.yaml b/Documentation/devicetree/bindings/interrupt-controller/econet,en751221-intc.yaml
index 5536319c49c3..44c09785e6bb 100644
--- a/Documentation/devicetree/bindings/interrupt-controller/econet,en751221-intc.yaml
+++ b/Documentation/devicetree/bindings/interrupt-controller/econet,en751221-intc.yaml
@@ -52,6 +52,25 @@ properties:
         - description: primary per-CPU IRQ
         - description: shadow IRQ number
 
+  econet,cpu-interrupt-map:
+    $ref: /schemas/types.yaml#/definitions/uint32-matrix
+    description:
+      When running in VEIC mode, the hardware re-routes interrupts from the
+      CPU interrupt controller core to the "external" interrupt controller
+      (this device). It then prioritizes them and sends them back to the CPU
+      along with its own interrupts. The CPU hardware handles interrupts using
+      a special dispatch table (the normal interrupt handler is not invoked).
+      In this interrupt controller, the CPU interrupts are renumbered as they
+      are merged with this controller's own hardware interrupts.
+
+      This is the inverse of an interrupt-map, mapping which interrupts from
+      this controller must be routed back to the CPU interrupt domain for
+      correct handling there.
+    items:
+      items:
+        - description: The interrupt number as received in this controller
+        - description: The interrupt number to be dispatched on the CPU intc
+
 required:
   - compatible
   - reg
@@ -74,5 +93,6 @@ examples:
         interrupts = <2>;
 
         econet,shadow-interrupts = <7 2>, <8 3>, <13 12>, <30 29>;
+        econet,cpu-interrupt-map = <7 0>, <8 1>;
     };
 ...
-- 
2.39.5

[PATCH v3 2/2] irqchip/econet-en751221: Support MIPS 34Kc VEIC mode

[Caleb James DeLisle]

The Vectored External Interrupt Controller mode present in the MIPS
34Kc and 1004Kc causes the CPU to stop dispatching interrupts by the
normal code path and instead it sends those interrupts to the external
interrupt controller to be prioritized, renumbered, and sent back.
When they come back, they are handled through a different path using a
dispatch table, so plat_irq_dispatch never sees action.

This of course subverts the traditional intc hierarchy, and on the
1004Kc the interrupt controller is standardized (IRQ_GIC) so it can be
reasonably considered part of the CPU itself - and tighter coupling
between IRQ_GIC and arch/mips/* is tolerable. However on the 34Kc
the intc is defined by each SoC vendor, so it's required to have a
modular driver - but for a device which in fact ends up taking over the
entire interrupt system.

Let the DT describe which IRQs which come from the CPU and should be
routed back and handled by the CPU intc. These particularly include the
two IPI interrupts which would otherwise necessitate duplication of all
the IPI supporting infrastructure from the CPU intc.

Signed-off-by: Caleb James DeLisle <cjd@cjdns.fr>
---
 drivers/irqchip/irq-econet-en751221.c | 187 ++++++++++++++++++++++++--
 1 file changed, 179 insertions(+), 8 deletions(-)

diff --git a/drivers/irqchip/irq-econet-en751221.c b/drivers/irqchip/irq-econet-en751221.c
index d83d5eb12795..a4921528eaef 100644
--- a/drivers/irqchip/irq-econet-en751221.c
+++ b/drivers/irqchip/irq-econet-en751221.c
@@ -30,6 +30,8 @@
 #include <linux/irqchip.h>
 #include <linux/irqchip/chained_irq.h>
 
+#include <asm/setup.h>
+
 #define IRQ_COUNT		40
 
 #define NOT_PERCPU		0xff
@@ -41,15 +43,19 @@
 #define REG_PENDING1		0x54
 
 /**
- * @membase: Base address of the interrupt controller registers
- * @interrupt_shadows: Array of all interrupts, for each value,
- *	- NOT_PERCPU: This interrupt is not per-cpu, so it has no shadow
- *	- IS_SHADOW: This interrupt is a shadow of another per-cpu interrupt
- *	- else: This is a per-cpu interrupt whose shadow is the value
+ * @membase:		Base address of the interrupt controller registers
+ * @domain:		The irq_domain for direct dispatch
+ * @ipi_domain:		The irq_domain for inter-process dispatch
+ * @interrupt_shadows:	Array of all interrupts, for each value,
+ *	- NOT_PERCPU:	This interrupt is not per-cpu, so it has no shadow
+ *	- IS_SHADOW:	This interrupt is a shadow of another per-cpu interrupt
+ *	- else:		This is a per-cpu interrupt whose shadow is the value
  */
 static struct {
-	void __iomem	*membase;
-	u8		interrupt_shadows[IRQ_COUNT];
+	void __iomem		*membase;
+	struct irq_domain	*domain;
+	struct irq_domain	*ipi_domain;
+	u8			interrupt_shadows[IRQ_COUNT];
 } econet_intc __ro_after_init;
 
 static DEFINE_RAW_SPINLOCK(irq_lock);
@@ -150,6 +156,56 @@ static void econet_intc_from_parent(struct irq_desc *desc)
 	chained_irq_exit(chip, desc);
 }
 
+/*
+ * When in VEIC mode, the CPU jumps to a handler in the vector table.
+ * The only way to know which interrupt is being triggered is from the vector table offset that
+ * has been jumped to. Reading REG_PENDING(0|1) will tell you which interrupts are currently
+ * pending in the intc, but that will not tell you which one the intc wants you to process
+ * right now. And if you are not processing the exact interrupt that the intc wants you to be
+ * processing, you might be on the wrong VPE. You can't tell which VPE any given REG_PENDING
+ * interrupt is intended for (shadow IRQ numbers are for masking only, they never flag as
+ * pending).
+ *
+ * Consequently, this little ritual of generating n handler functions and registering one per
+ * interrupt is unavoidable.
+ */
+#define X(irq) \
+	static void econet_irq_dispatch ## irq (void) \
+	{ \
+		do_domain_IRQ(econet_intc.domain, irq); \
+	}
+
+ X(0)  X(1)  X(2)  X(3)  X(4)  X(5)  X(6)  X(7)  X(8)  X(9)
+X(10) X(11) X(12) X(13) X(14) X(15) X(16) X(17) X(18) X(19)
+X(20) X(21) X(22) X(23) X(24) X(25) X(26) X(27) X(28) X(29)
+X(30) X(31) X(32) X(33) X(34) X(35) X(36) X(37) X(38) X(39)
+
+#undef X
+#define X(irq) econet_irq_dispatch ## irq,
+
+static void (* const econet_irq_dispatchers[])(void) = {
+	X(0)  X(1)  X(2)  X(3)  X(4)  X(5)  X(6)  X(7)  X(8)  X(9)
+	X(10) X(11) X(12) X(13) X(14) X(15) X(16) X(17) X(18) X(19)
+	X(20) X(21) X(22) X(23) X(24) X(25) X(26) X(27) X(28) X(29)
+	X(30) X(31) X(32) X(33) X(34) X(35) X(36) X(37) X(38) X(39)
+};
+
+/* Likewise, we do the same for the 2 IPI IRQs so that we can route them back */
+static void econet_cpu_dispatch0(void)
+{
+	do_domain_IRQ(econet_intc.ipi_domain, 0);
+}
+
+static void econet_cpu_dispatch1(void)
+{
+	do_domain_IRQ(econet_intc.ipi_domain, 1);
+}
+
+static void (* const econet_cpu_dispatchers[])(void) = {
+	econet_cpu_dispatch0,
+	econet_cpu_dispatch1,
+};
+
 static const struct irq_chip econet_irq_chip;
 
 static int econet_intc_map(struct irq_domain *d, u32 irq, irq_hw_number_t hwirq)
@@ -174,6 +230,10 @@ static int econet_intc_map(struct irq_domain *d, u32 irq, irq_hw_number_t hwirq)
 	}
 
 	irq_set_chip_data(irq, NULL);
+
+	if (cpu_has_veic)
+		set_vi_handler(hwirq + 1, econet_irq_dispatchers[hwirq]);
+
 	return 0;
 }
 
@@ -249,6 +309,101 @@ static int __init get_shadow_interrupts(struct device_node *node)
 	return 0;
 }
 
+/**
+ * econet_cpu_init() - configure routing of CPU interrupts to the correct domain.
+ * @node: The devicetree node of this interrupt controller.
+ *
+ * Interrupts that originate from the CPU are unconditionally unmasked here and are re-routed back
+ * to the IPI irq_domain in the CPU intc. Masking still takes place but the CPU intc is in charge
+ * of it, using the mask bits of the c0_status register.
+ *
+ * Note that because IP2 ... IP7 are repurposed as Interrupt Priority Level, only the two IPI
+ * interrupts are actually supported.
+ */
+static int __init econet_cpu_init(struct device_node *node)
+{
+	const char *field = "econet,cpu-interrupt-map";
+	struct device_node *parent_intc;
+	int map_size;
+	u32 mask;
+
+	map_size = of_property_count_u32_elems(node, field);
+
+	if (map_size <= 0) {
+		return 0;
+	} else if (map_size % 2) {
+		pr_err("%pOF: %s count is odd, ignoring\n", node, field);
+		return 0;
+	}
+
+	u32 *maps __free(kfree) = kmalloc_array(map_size, sizeof(u32), GFP_KERNEL);
+	if (!maps)
+		return -ENOMEM;
+
+	if (of_property_read_u32_array(node, field, maps, map_size)) {
+		pr_err("%pOF: Failed to read %s\n", node, field);
+		return -EINVAL;
+	}
+
+	/* Validation */
+	for (int i = 0; i < map_size; i += 2) {
+		u32 receive = maps[i];
+		u32 dispatch = maps[i + 1];
+		u8 shadow;
+
+		if (receive >= IRQ_COUNT) {
+			pr_err("%pOF: Entry %d:%d in %s (%u) is out of bounds\n",
+			       node, i, 0, field, receive);
+			return -EINVAL;
+		}
+
+		shadow = econet_intc.interrupt_shadows[receive];
+		if (shadow != NOT_PERCPU && shadow >= IRQ_COUNT) {
+			pr_err("%pOF: Entry %d:%d in %s (%u) has invalid shadow (%d)\n",
+			       node, i, 0, field, receive, shadow);
+			return -EINVAL;
+		}
+
+		if (dispatch >= ARRAY_SIZE(econet_cpu_dispatchers)) {
+			pr_err("%pOF: Entry %d:%d in %s (%u) is out of bounds only IPI interrupts "
+			       "are supported\n",
+			       node, i, 1, field, dispatch);
+			return -EINVAL;
+		}
+	}
+
+	parent_intc = of_irq_find_parent(node);
+	if (!parent_intc) {
+		pr_err("%pOF: Failed to find parent %s\n", node, "IRQ device");
+		return -ENODEV;
+	}
+
+	econet_intc.ipi_domain = irq_find_matching_host(parent_intc, DOMAIN_BUS_IPI);
+	if (!econet_intc.ipi_domain) {
+		pr_err("%pOF: Failed to find parent %s\n", node, "IPI domain");
+		return -ENODEV;
+	}
+
+	mask = 0;
+	for (int i = 0; i < map_size; i += 2) {
+		u32 receive = maps[i];
+		u32 dispatch = maps[i + 1];
+		u8 shadow;
+
+		set_vi_handler(receive + 1, econet_cpu_dispatchers[dispatch]);
+
+		mask |= BIT(receive);
+
+		shadow = econet_intc.interrupt_shadows[receive];
+		if (shadow != NOT_PERCPU)
+			mask |= BIT(shadow);
+	}
+
+	econet_wreg(REG_MASK0, mask, mask);
+
+	return 0;
+}
+
 static int __init econet_intc_of_init(struct device_node *node, struct device_node *parent)
 {
 	struct irq_domain *domain;
@@ -294,7 +449,23 @@ static int __init econet_intc_of_init(struct device_node *node, struct device_no
 		goto err_unmap;
 	}
 
-	irq_set_chained_handler_and_data(irq, econet_intc_from_parent, domain);
+	/*
+	 * 34K Manual (MD00534) Section 6.3.1.3 rev 1.13 page 136:
+	 * In VEIC mode, IP2 ... IP7 are repurposed as Interrupt Priority Level. The controller
+	 * will filter incoming interrupts whose priority is lower than the IPL number. Therefore
+	 * we must not set any of these bits. We avoid setting IP2 by not actually chaining this
+	 * intc to the CPU intc.
+	 */
+	if (cpu_has_veic) {
+		ret = econet_cpu_init(node);
+
+		if (ret)
+			return ret;
+	} else {
+		irq_set_chained_handler_and_data(irq, econet_intc_from_parent, domain);
+	}
+
+	econet_intc.domain = domain;
 
 	return 0;
 
-- 
2.39.5