Re: [PATCH V14 12/15] irqchip: Add Loongson Extended I/O interrupt controller support
From: Jianmin Lv
Date: Sun Jul 10 2022 - 08:18:09 EST
On 2022/7/7 下午8:30, Marc Zyngier wrote:
On Sun, 03 Jul 2022 09:45:29 +0100,
Jianmin Lv <lvjianmin@xxxxxxxxxxx> wrote:
From: Huacai Chen <chenhuacai@xxxxxxxxxxx>
EIOINTC stands for "Extended I/O Interrupts" that described in Section
11.2 of "Loongson 3A5000 Processor Reference Manual". For more
information please refer Documentation/loongarch/irq-chip-model.rst.
Loongson-3A5000 has 4 cores per NUMA node, and each NUMA node has an
EIOINTC; while Loongson-3C5000 has 16 cores per NUMA node, and each NUMA
node has 4 EIOINTCs. In other words, 16 cores of one NUMA node in
Loongson-3C5000 are organized in 4 groups, each group connects to an
EIOINTC. We call the "group" here as an EIOINTC node, so each EIOINTC
node always includes 4 cores (both in Loongson-3A5000 and Loongson-
3C5000).
Co-developed-by: Jianmin Lv <lvjianmin@xxxxxxxxxxx>
Signed-off-by: Jianmin Lv <lvjianmin@xxxxxxxxxxx>
Signed-off-by: Huacai Chen <chenhuacai@xxxxxxxxxxx>
---
arch/loongarch/include/asm/irq.h | 11 +-
drivers/irqchip/Kconfig | 10 +
drivers/irqchip/Makefile | 1 +
drivers/irqchip/irq-loongson-eiointc.c | 413 +++++++++++++++++++++++++++++++++
include/linux/cpuhotplug.h | 1 +
5 files changed, 426 insertions(+), 10 deletions(-)
create mode 100644 drivers/irqchip/irq-loongson-eiointc.c
diff --git a/arch/loongarch/include/asm/irq.h b/arch/loongarch/include/asm/irq.h
index a07315a..a826ac4 100644
--- a/arch/loongarch/include/asm/irq.h
+++ b/arch/loongarch/include/asm/irq.h
@@ -87,15 +87,6 @@ struct acpi_vector_group {
extern int find_pch_pic(u32 gsi);
extern int eiointc_get_node(int id);
-static inline void eiointc_enable(void)
-{
- uint64_t misc;
-
- misc = iocsr_read64(LOONGARCH_IOCSR_MISC_FUNC);
- misc |= IOCSR_MISC_FUNC_EXT_IOI_EN;
- iocsr_write64(misc, LOONGARCH_IOCSR_MISC_FUNC);
-}
-
struct acpi_madt_lio_pic;
struct acpi_madt_eio_pic;
struct acpi_madt_ht_pic;
@@ -107,7 +98,7 @@ static inline void eiointc_enable(void)
int liointc_acpi_init(struct irq_domain *parent,
struct acpi_madt_lio_pic *acpi_liointc);
-struct irq_domain *eiointc_acpi_init(struct irq_domain *parent,
+int eiointc_acpi_init(struct irq_domain *parent,
struct acpi_madt_eio_pic *acpi_eiointc);
int htvec_acpi_init(struct irq_domain *parent,
diff --git a/drivers/irqchip/Kconfig b/drivers/irqchip/Kconfig
index c1d527f..e918579 100644
--- a/drivers/irqchip/Kconfig
+++ b/drivers/irqchip/Kconfig
@@ -555,6 +555,16 @@ config LOONGSON_LIOINTC
help
Support for the Loongson Local I/O Interrupt Controller.
+config LOONGSON_EIOINTC
+ bool "Loongson Extend I/O Interrupt Controller"
+ depends on LOONGARCH
+ depends on MACH_LOONGSON64
+ default MACH_LOONGSON64
+ select IRQ_DOMAIN_HIERARCHY
+ select GENERIC_IRQ_CHIP
+ help
+ Support for the Loongson3 Extend I/O Interrupt Vector Controller.
+
config LOONGSON_HTPIC
bool "Loongson3 HyperTransport PIC Controller"
depends on MACH_LOONGSON64 && MIPS
diff --git a/drivers/irqchip/Makefile b/drivers/irqchip/Makefile
index 242b8b3..0cfd4f0 100644
--- a/drivers/irqchip/Makefile
+++ b/drivers/irqchip/Makefile
@@ -104,6 +104,7 @@ obj-$(CONFIG_TI_SCI_INTR_IRQCHIP) += irq-ti-sci-intr.o
obj-$(CONFIG_TI_SCI_INTA_IRQCHIP) += irq-ti-sci-inta.o
obj-$(CONFIG_TI_PRUSS_INTC) += irq-pruss-intc.o
obj-$(CONFIG_LOONGSON_LIOINTC) += irq-loongson-liointc.o
+obj-$(CONFIG_LOONGSON_EIOINTC) += irq-loongson-eiointc.o
obj-$(CONFIG_LOONGSON_HTPIC) += irq-loongson-htpic.o
obj-$(CONFIG_LOONGSON_HTVEC) += irq-loongson-htvec.o
obj-$(CONFIG_LOONGSON_PCH_PIC) += irq-loongson-pch-pic.o
diff --git a/drivers/irqchip/irq-loongson-eiointc.c b/drivers/irqchip/irq-loongson-eiointc.c
new file mode 100644
index 0000000..9ddb9ac
--- /dev/null
+++ b/drivers/irqchip/irq-loongson-eiointc.c
@@ -0,0 +1,413 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Loongson Extend I/O Interrupt Controller support
+ *
+ * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
+ */
+
+#define pr_fmt(fmt) "eiointc: " fmt
+
+#include <linux/interrupt.h>
+#include <linux/irq.h>
+#include <linux/irqchip.h>
+#include <linux/irqdomain.h>
+#include <linux/irqchip/chained_irq.h>
+#include <linux/kernel.h>
+#include <linux/platform_device.h>
+#include <linux/of_address.h>
+#include <linux/of_irq.h>
+#include <linux/of_platform.h>
+
+#define EIOINTC_REG_NODEMAP 0x14a0
+#define EIOINTC_REG_IPMAP 0x14c0
+#define EIOINTC_REG_ENABLE 0x1600
+#define EIOINTC_REG_BOUNCE 0x1680
+#define EIOINTC_REG_ISR 0x1800
+#define EIOINTC_REG_ROUTE 0x1c00
+
+#define VEC_REG_COUNT 4
+#define VEC_COUNT_PER_REG 64
+#define VEC_COUNT (VEC_REG_COUNT * VEC_COUNT_PER_REG)
+#define VEC_REG_IDX(irq_id) ((irq_id) / VEC_COUNT_PER_REG)
+#define VEC_REG_BIT(irq_id) ((irq_id) % VEC_COUNT_PER_REG)
+#define EIOINTC_ALL_ENABLE 0xffffffff
+
+#define MAX_EIO_NODES (NR_CPUS / CORES_PER_EIO_NODE)
+
+static int nr_pics;
+
+struct eiointc_priv {
+ u32 node;
+ nodemask_t node_map;
+ cpumask_t cpuspan_map;
+ struct fwnode_handle *domain_handle;
+ struct irq_domain *eiointc_domain;
+};
+
+static struct eiointc_priv *eiointc_priv[MAX_IO_PICS];
+
+static void eiointc_enable(void)
+{
+ uint64_t misc;
+
+ misc = iocsr_read64(LOONGARCH_IOCSR_MISC_FUNC);
+ misc |= IOCSR_MISC_FUNC_EXT_IOI_EN;
+ iocsr_write64(misc, LOONGARCH_IOCSR_MISC_FUNC);
+}
+
+static int cpu_to_eio_node(int cpu)
+{
+ return cpu_logical_map(cpu) / CORES_PER_EIO_NODE;
+}
+
+static void eiointc_set_irq_route(int pos, unsigned int cpu, unsigned int mnode, nodemask_t *node_map)
+{
+ int i, node, cpu_node, route_node;
+ unsigned char coremap[MAX_EIO_NODES];
+ uint32_t pos_off, data, data_byte, data_mask;
+
+ pos_off = pos & ~3;
+ data_byte = pos & 3;
+ data_mask = ~BIT_MASK(data_byte) & 0xf;
+
+ memset(coremap, 0, sizeof(unsigned char) * MAX_EIO_NODES);
This would be better written as an initialisation of the array:
unsigned char coremap[MAX_EIO_NODES] = { };
+
+ /* Calculate node and coremap of target irq */
+ cpu_node = cpu_logical_map(cpu) / CORES_PER_EIO_NODE;
+ coremap[cpu_node] |= BIT(cpu_logical_map(cpu) % CORES_PER_EIO_NODE);
+
+ for_each_online_cpu(i) {
+ node = cpu_to_eio_node(i);
+ if (!node_isset(node, *node_map))
+ continue;
+
+ /* EIO node 0 is in charge of inter-node interrupt dispatch */
+ route_node = (node == mnode) ? cpu_node : node;
+ data = ((coremap[node] | (route_node << 4)) << (data_byte * 8));
+ csr_any_send(EIOINTC_REG_ROUTE + pos_off, data, data_mask, node * CORES_PER_EIO_NODE);
+ }
But the coremap array doesn't make much sense. You only initialise
*one* byte in it, and the rest is always 0. Either it is completely
irrelevant and you should get rid of it, or this is buggy and you
should fix it.
Yes, thanks, Marc, it's confused code here because the definition of
coremap was no changed when adjust the function previously, a byte typed
variable is enough.
+}
+
+static DEFINE_RAW_SPINLOCK(affinity_lock);
+
+static int eiointc_set_irq_affinity(struct irq_data *d, const struct cpumask *affinity, bool force)
+{
+ unsigned int cpu;
+ unsigned long flags;
+ uint32_t vector, regaddr;
+ struct cpumask intersect_affinity;
+ struct eiointc_priv *priv = d->domain->host_data;
+
+ if (!IS_ENABLED(CONFIG_SMP))
+ return -EPERM;
Is it even legal to build this architecture with !SMP?
It's legal in our internal repo for single core CPU, but now it's not
allowed to build with !SMP, I'll remove it in next version, and take it
into account when supporting single core CPU in future.
+
+ raw_spin_lock_irqsave(&affinity_lock, flags);
+
+ cpumask_and(&intersect_affinity, affinity, cpu_online_mask);
+ cpumask_and(&intersect_affinity, &intersect_affinity, &priv->cpuspan_map);
+
+ if (cpumask_empty(&intersect_affinity)) {
+ raw_spin_unlock_irqrestore(&affinity_lock, flags);
+ return -EINVAL;
+ }
+ cpu = cpumask_first(&intersect_affinity);
+
+ if (!d->parent_data)
+ vector = d->hwirq;
+ else
+ vector = d->parent_data->hwirq;
+
+ regaddr = EIOINTC_REG_ENABLE + ((vector >> 5) << 2);
+
+ /* Mask target vector */
+ csr_any_send(regaddr, EIOINTC_ALL_ENABLE & (~BIT(vector & 0x1F)), 0x0, 0);
+ /* Set route for target vector */
+ eiointc_set_irq_route(vector, cpu, priv->node, &priv->node_map);
+ /* Unmask target vector */
+ csr_any_send(regaddr, EIOINTC_ALL_ENABLE, 0x0, 0);
+
+ irq_data_update_effective_affinity(d, cpumask_of(cpu));
+
+ raw_spin_unlock_irqrestore(&affinity_lock, flags);
+
+ return IRQ_SET_MASK_OK;
+}
+
+static int eiointc_index(int node)
+{
+ int i;
+
+ for (i = 0; i < nr_pics; i++) {
+ if (node_isset(node, eiointc_priv[i]->node_map))
+ return i;
+ }
+
+ return -1;
+}
+
+static int eiointc_router_init(unsigned int cpu)
+{
+ int i, bit;
+ uint32_t data;
+ uint32_t node = cpu_to_eio_node(cpu);
+ uint32_t index = eiointc_index(node);
+
+ if (index < 0) {
+ pr_err("Error: invalid nodemap!\n");
+ return -1;
+ }
+
+ if ((cpu_logical_map(cpu) % CORES_PER_EIO_NODE) == 0) {
+ eiointc_enable();
+
+ for (i = 0; i < VEC_COUNT / 32; i++) {
+ data = (((1 << (i * 2 + 1)) << 16) | (1 << (i * 2)));
+ iocsr_write32(data, EIOINTC_REG_NODEMAP + i * 4);
+ }
+
+ for (i = 0; i < VEC_COUNT / 32 / 4; i++) {
+ bit = BIT(1 + index); /* Route to IP[1 + index] */
+ data = bit | (bit << 8) | (bit << 16) | (bit << 24);
+ iocsr_write32(data, EIOINTC_REG_IPMAP + i * 4);
+ }
+
+ for (i = 0; i < VEC_COUNT / 4; i++) {
+ /* Route to Node-0 Core-0 */
+ if (index == 0)
+ bit = BIT(cpu_logical_map(0));
+ else
+ bit = (eiointc_priv[index]->node << 4) | 1;
+
+ data = bit | (bit << 8) | (bit << 16) | (bit << 24);
+ iocsr_write32(data, EIOINTC_REG_ROUTE + i * 4);
+ }
+
+ for (i = 0; i < VEC_COUNT / 32; i++) {
+ data = 0xffffffff;
+ iocsr_write32(data, EIOINTC_REG_ENABLE + i * 4);
+ iocsr_write32(data, EIOINTC_REG_BOUNCE + i * 4);
+ }
+ }
+
+ return 0;
+}
+
+static void eiointc_irq_dispatch(struct irq_desc *desc)
+{
+ int i;
+ u64 pending;
+ bool handled = false;
+ struct irq_chip *chip = irq_desc_get_chip(desc);
+ struct eiointc_priv *priv = irq_desc_get_handler_data(desc);
+
+ chained_irq_enter(chip, desc);
+
+ for (i = 0; i < VEC_REG_COUNT; i++) {
+ pending = iocsr_read64(EIOINTC_REG_ISR + (i << 3));
+ iocsr_write64(pending, EIOINTC_REG_ISR + (i << 3));
+ while (pending) {
+ int bit = __ffs(pending);
+ int irq = bit + VEC_COUNT_PER_REG * i;
+
+ generic_handle_domain_irq(priv->eiointc_domain, irq);
+ pending &= ~BIT(bit);
+ handled = true;
+ }
+ }
+
+ if (!handled)
+ spurious_interrupt();
+
+ chained_irq_exit(chip, desc);
+}
+
+static void eiointc_ack_irq(struct irq_data *d)
+{
+ if (d->parent_data)
+ irq_chip_ack_parent(d);
+}
+
+static void eiointc_mask_irq(struct irq_data *d)
+{
+ if (d->parent_data)
+ irq_chip_mask_parent(d);
+}
+
+static void eiointc_unmask_irq(struct irq_data *d)
+{
+ if (d->parent_data)
+ irq_chip_unmask_parent(d);
+}
It is pretty unclear to me under what circumstances some interrupts
have a parent, and some don't.
t seems that parent_data here is always NULL because:
- 'd' passed in is irq_data of eiointc when calling the eiointc
irqchip's ops function.
- eiointc has no parent domain.
The code(should be cleaned) is here because previous incorrect use to
eiointc_set_irq_affinity in resume code(which has been cleaned from the
patch). I'll clean it in next version.
So, we have noting to do here, leaving some empty unmask/mask.
+
+static struct irq_chip eiointc_irq_chip = {
+ .name = "EIOINTC",
+ .irq_ack = eiointc_ack_irq,
+ .irq_mask = eiointc_mask_irq,
+ .irq_unmask = eiointc_unmask_irq,
+ .irq_set_affinity = eiointc_set_irq_affinity,
Under what circumstances can an interrupt be moved to another CPU?
Given that you have an EIO per 4-cpu group, I can see how the EIO can
move the affinity. But can you move an interrupt to a CPU belonging to
another EOIO? If you can, is any *other* interrupt affected by the
affinity change?
This is specially troubling as in general, chained irq handlers cannot
move interrupts without breaking the affinity of other interrupts, and
I'd like to know where we stand here.
Each node(4-core cpu) has an EIO irq controller, the controllers in each
node of a nodes-group in a system are looked as an EIO instance and
configured according to the affinity of an EIO vector.
Although each EIO has 256 vectors, the nodes-group containing the EIO
instance has 256 vectors too, because all the vector in the same
position in an EIO of each node is responsible for only one I/O
interrupt. Any vector of the EIO instance in the group can be routed to
any cpu core of an node of the nodes-group, and each EIO has a routing
register for the same vector, which is required to be configured when
the vector's affinity is changed.
An I/O irq of PCH-PIC or PCH-MSI is mapped to one vector of the EIO
instance, when the irq affinity of an vector is required to change, we
just change routing register of the vector in each EIO hardware, which
will not affect other vectors(I/O irqs).
The flow of an irq affinity setting for PCH-PIC or PCH-MSI is following:
- an irq of PCH-PIC:
do_set_affinity
-->irqchip ops->irq_set_affinity(d，...) PCH-PIC domain
-->irqchip ops->irq_set_affinity(d，...) EIOINTC domain
- an irq of PCH-MSI:
do_set_affinity
-->irqchip ops->irq_set_affinity(d，...) PCH-MSI domain
-->irqchip ops->irq_set_affinity(d，...) MIDDLE domain
-->irqchip ops->irq_set_affinity(d，...) EIOINTC domain
+};
+
+static int eiointc_domain_alloc(struct irq_domain *domain, unsigned int virq,
+ unsigned int nr_irqs, void *arg)
+{
+ int ret;
+ unsigned int i, type;
+ unsigned long hwirq = 0;
+ struct eiointc *priv = domain->host_data;
+
+ ret = irq_domain_translate_onecell(domain, arg, &hwirq, &type);
+ if (ret)
+ return ret;
+
+ for (i = 0; i < nr_irqs; i++) {
+ irq_domain_set_info(domain, virq + i, hwirq + i, &eiointc_irq_chip,
+ priv, handle_edge_irq, NULL, NULL);
+ }
+
+ return 0;
+}
+
+static void eiointc_domain_free(struct irq_domain *domain, unsigned int virq,
+ unsigned int nr_irqs)
+{
+ int i;
+
+ for (i = 0; i < nr_irqs; i++) {
+ struct irq_data *d = irq_domain_get_irq_data(domain, virq + i);
+
+ irq_set_handler(virq + i, NULL);
+ irq_domain_reset_irq_data(d);
+ }
+}
+
+static const struct irq_domain_ops eiointc_domain_ops = {
+ .translate = irq_domain_translate_onecell,
+ .alloc = eiointc_domain_alloc,
+ .free = eiointc_domain_free,
+};
+
+static void acpi_set_vec_parent(int node, struct irq_domain *parent, struct acpi_vector_group *vec_group)
+{
+ int i;
+
+ if (cpu_has_flatmode)
+ node = cpu_to_node(node * CORES_PER_EIO_NODE);
+
+ for (i = 0; i < MAX_IO_PICS; i++) {
+ if (node == vec_group[i].node) {
+ vec_group[i].parent = parent;
+ return;
+ }
+ }
+}
+
+struct irq_domain *acpi_get_vec_parent(int node, struct acpi_vector_group *vec_group)
+{
+ int i;
+
+ for (i = 0; i < MAX_IO_PICS; i++) {
+ if (node == vec_group[i].node)
+ return vec_group[i].parent;
+ }
+ return NULL;
+}
+
+static int __init
+pch_pic_parse_madt(union acpi_subtable_headers *header,
+ const unsigned long end)
+{
+ struct acpi_madt_bio_pic *pchpic_entry = (struct acpi_madt_bio_pic *)header;
+ unsigned int node = (pchpic_entry->address >> 44) & 0xf;
+ struct irq_domain *parent = acpi_get_vec_parent(node, pch_group);
+
+ if (parent)
+ return pch_pic_acpi_init(parent, pchpic_entry);
+
+ return -EINVAL;
+}
+
+static int __init
+pch_msi_parse_madt(union acpi_subtable_headers *header,
+ const unsigned long end)
+{
+ struct acpi_madt_msi_pic *pchmsi_entry = (struct acpi_madt_msi_pic *)header;
+ struct irq_domain *parent = acpi_get_vec_parent(eiointc_priv[nr_pics - 1]->node, msi_group);
+
+ if (parent)
+ return pch_msi_acpi_init(parent, pchmsi_entry);
+
+ return -EINVAL;
+}
+
+static int __init acpi_cascade_irqdomain_init(void)
+{
+ acpi_table_parse_madt(ACPI_MADT_TYPE_BIO_PIC,
+ pch_pic_parse_madt, 0);
+ acpi_table_parse_madt(ACPI_MADT_TYPE_MSI_PIC,
+ pch_msi_parse_madt, 1);
+ return 0;
+}
+
+int __init eiointc_acpi_init(struct irq_domain *parent,
+ struct acpi_madt_eio_pic *acpi_eiointc)
+{
+ int i, parent_irq;
+ unsigned long node_map;
+ struct eiointc_priv *priv;
+
+ if (!acpi_eiointc)
+ return -EINVAL;
How can this happen?
Yes, it's should be cleaned. In previours version, acpi_eiointc is a
global variable and will be set outside this file. But now, it's not
global any more, so I'll clean it, thanks.
+
+ priv = kzalloc(sizeof(*priv), GFP_KERNEL);
+ if (!priv)
+ return -ENOMEM;
+
+ priv->domain_handle = irq_domain_alloc_fwnode((phys_addr_t *)acpi_eiointc);
+ if (!priv->domain_handle) {
+ pr_err("Unable to allocate domain handle\n");
+ goto out_free_priv;
+ }
+
+ priv->node = acpi_eiointc->node;
+ node_map = acpi_eiointc->node_map ? : -1ULL;
+
+ for_each_possible_cpu(i) {
+ if (node_map & (1ULL << cpu_to_eio_node(i))) {
+ node_set(cpu_to_eio_node(i), priv->node_map);
+ cpumask_or(&priv->cpuspan_map, &priv->cpuspan_map, cpumask_of(i));
+ }
+ }
+
+ /* Setup IRQ domain */
+ priv->eiointc_domain = irq_domain_create_linear(priv->domain_handle, VEC_COUNT,
+ &eiointc_domain_ops, priv);
+ if (!priv->eiointc_domain) {
+ pr_err("loongson-eiointc: cannot add IRQ domain\n");
+ goto out_free_handle;
+ }
+
+ eiointc_priv[nr_pics++] = priv;
+
+ eiointc_router_init(0);
+
+ parent_irq = irq_create_mapping(parent, acpi_eiointc->cascade);
+ irq_set_chained_handler_and_data(parent_irq, eiointc_irq_dispatch, priv);
+
+ cpuhp_setup_state_nocalls(CPUHP_AP_IRQ_LOONGARCH_STARTING,
+ "irqchip/loongarch/intc:starting",
+ eiointc_router_init, NULL);
+
+ acpi_set_vec_parent(acpi_eiointc->node, priv->eiointc_domain, pch_group);
+ acpi_set_vec_parent(acpi_eiointc->node, priv->eiointc_domain, msi_group);
+ acpi_cascade_irqdomain_init();
+
+ return 0;
+
+out_free_handle:
+ irq_domain_free_fwnode(priv->domain_handle);
+ priv->domain_handle = NULL;
+out_free_priv:
+ kfree(priv);
+
+ return -ENOMEM;
+}
diff --git a/include/linux/cpuhotplug.h b/include/linux/cpuhotplug.h
index 19f0dbf..de662f3 100644
--- a/include/linux/cpuhotplug.h
+++ b/include/linux/cpuhotplug.h
@@ -151,6 +151,7 @@ enum cpuhp_state {
CPUHP_AP_IRQ_BCM2836_STARTING,
CPUHP_AP_IRQ_MIPS_GIC_STARTING,
CPUHP_AP_IRQ_RISCV_STARTING,
+ CPUHP_AP_IRQ_LOONGARCH_STARTING,
CPUHP_AP_IRQ_SIFIVE_PLIC_STARTING,
CPUHP_AP_ARM_MVEBU_COHERENCY,
CPUHP_AP_MICROCODE_LOADER,
--
1.8.3.1
Thanks,
M.