LoongArch: Add Non-Uniform Memory Access (NUMA) support

Add Non-Uniform Memory Access (NUMA) support for LoongArch. LoongArch
has 48-bit physical address, but the HyperTransport I/O bus only support
40-bit address, so we need a custom phys_to_dma() and dma_to_phys() to
extract the 4-bit node id (bit 44~47) from Loongson-3's 48-bit physical
address space and embed it into 40-bit. In the 40-bit dma address, node
id offset can be read from the LS7A_DMA_CFG register.

Reviewed-by: WANG Xuerui <[email protected]>
Reviewed-by: Jiaxun Yang <[email protected]>
Signed-off-by: Huacai Chen <[email protected]>

Author: chenhuacai

arch/loongarch/Kconfig

@@ -7,6 +7,7 @@ config LOONGARCH
 	select ARCH_ENABLE_MEMORY_HOTPLUG
 	select ARCH_ENABLE_MEMORY_HOTREMOVE
 	select ARCH_HAS_ACPI_TABLE_UPGRADE	if ACPI
+	select ARCH_HAS_PHYS_TO_DMA
 	select ARCH_HAS_PTE_SPECIAL
 	select ARCH_HAS_TICK_BROADCAST if GENERIC_CLOCKEVENTS_BROADCAST
 	select ARCH_INLINE_READ_LOCK if !PREEMPTION
@@ -41,6 +42,7 @@ config LOONGARCH
 	select ARCH_SUPPORTS_ACPI
 	select ARCH_SUPPORTS_ATOMIC_RMW
 	select ARCH_SUPPORTS_HUGETLBFS
+	select ARCH_SUPPORTS_NUMA_BALANCING
 	select ARCH_USE_BUILTIN_BSWAP
 	select ARCH_USE_CMPXCHG_LOCKREF
 	select ARCH_USE_QUEUED_RWLOCKS
@@ -91,12 +93,15 @@ config LOONGARCH
 	select HAVE_PERF_EVENTS
 	select HAVE_REGS_AND_STACK_ACCESS_API
 	select HAVE_RSEQ
+	select HAVE_SETUP_PER_CPU_AREA if NUMA
 	select HAVE_SYSCALL_TRACEPOINTS
 	select HAVE_TIF_NOHZ
 	select HAVE_VIRT_CPU_ACCOUNTING_GEN if !SMP
 	select IRQ_FORCED_THREADING
 	select IRQ_LOONGARCH_CPU
 	select MODULES_USE_ELF_RELA if MODULES
+	select NEED_PER_CPU_EMBED_FIRST_CHUNK
+	select NEED_PER_CPU_PAGE_FIRST_CHUNK
 	select OF
 	select OF_EARLY_FLATTREE
 	select PERF_USE_VMALLOC
@@ -105,6 +110,7 @@ config LOONGARCH
 	select SYSCTL_EXCEPTION_TRACE
 	select SWIOTLB
 	select TRACE_IRQFLAGS_SUPPORT
+	select USE_PERCPU_NUMA_NODE_ID
 	select ZONE_DMA32
 
 config 32BIT
@@ -335,6 +341,20 @@ config NR_CPUS
 	  This allows you to specify the maximum number of CPUs which this
 	  kernel will support.
 
+config NUMA
+	bool "NUMA Support"
+	select ACPI_NUMA if ACPI
+	help
+	  Say Y to compile the kernel with NUMA (Non-Uniform Memory Access)
+	  support.  This option improves performance on systems with more
+	  than one NUMA node; on single node systems it is generally better
+	  to leave it disabled.
+
+config NODES_SHIFT
+	int
+	default "6"
+	depends on NUMA
+
 config FORCE_MAX_ZONEORDER
 	int "Maximum zone order"
 	range 14 64 if PAGE_SIZE_64KB
@@ -381,6 +401,7 @@ config ARCH_SELECT_MEMORY_MODEL
 
 config ARCH_FLATMEM_ENABLE
 	def_bool y
+	depends on !NUMA
 
 config ARCH_SPARSEMEM_ENABLE
 	def_bool y

arch/loongarch/include/asm/bootinfo.h

@@ -13,6 +13,8 @@ const char *get_system_type(void);
 extern void init_environ(void);
 extern void memblock_init(void);
 extern void platform_init(void);
+extern void plat_swiotlb_setup(void);
+extern int __init init_numa_memory(void);
 
 struct loongson_board_info {
 	int bios_size;

arch/loongarch/include/asm/dma-direct.h

@@ -0,0 +1,11 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
+ */
+#ifndef _LOONGARCH_DMA_DIRECT_H
+#define _LOONGARCH_DMA_DIRECT_H
+
+dma_addr_t phys_to_dma(struct device *dev, phys_addr_t paddr);
+phys_addr_t dma_to_phys(struct device *dev, dma_addr_t daddr);
+
+#endif /* _LOONGARCH_DMA_DIRECT_H */

arch/loongarch/include/asm/mmzone.h

@@ -0,0 +1,18 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Author: Huacai Chen ([email protected])
+ * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
+ */
+#ifndef _ASM_MMZONE_H_
+#define _ASM_MMZONE_H_
+
+#include <asm/page.h>
+#include <asm/numa.h>
+
+extern struct pglist_data *node_data[];
+
+#define NODE_DATA(nid)	(node_data[(nid)])
+
+extern void setup_zero_pages(void);
+
+#endif /* _ASM_MMZONE_H_ */

arch/loongarch/include/asm/numa.h

@@ -0,0 +1,67 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Author: Jianmin Lv <[email protected]>
+ *         Huacai Chen <[email protected]>
+ *
+ * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
+ */
+
+#ifndef _ASM_LOONGARCH_NUMA_H
+#define _ASM_LOONGARCH_NUMA_H
+
+#include <linux/nodemask.h>
+
+#define NODE_ADDRSPACE_SHIFT 44
+
+#define pa_to_nid(addr)		(((addr) & 0xf00000000000) >> NODE_ADDRSPACE_SHIFT)
+#define nid_to_addrbase(nid)	(_ULCAST_(nid) << NODE_ADDRSPACE_SHIFT)
+
+#ifdef CONFIG_NUMA
+
+extern int numa_off;
+extern s16 __cpuid_to_node[CONFIG_NR_CPUS];
+extern nodemask_t numa_nodes_parsed __initdata;
+
+struct numa_memblk {
+	u64			start;
+	u64			end;
+	int			nid;
+};
+
+#define NR_NODE_MEMBLKS		(MAX_NUMNODES*2)
+struct numa_meminfo {
+	int			nr_blks;
+	struct numa_memblk	blk[NR_NODE_MEMBLKS];
+};
+
+extern int __init numa_add_memblk(int nodeid, u64 start, u64 end);
+
+extern void __init early_numa_add_cpu(int cpuid, s16 node);
+extern void numa_add_cpu(unsigned int cpu);
+extern void numa_remove_cpu(unsigned int cpu);
+
+static inline void numa_clear_node(int cpu)
+{
+}
+
+static inline void set_cpuid_to_node(int cpuid, s16 node)
+{
+	__cpuid_to_node[cpuid] = node;
+}
+
+extern int early_cpu_to_node(int cpu);
+
+#else
+
+static inline void early_numa_add_cpu(int cpuid, s16 node)	{ }
+static inline void numa_add_cpu(unsigned int cpu)		{ }
+static inline void numa_remove_cpu(unsigned int cpu)		{ }
+
+static inline int early_cpu_to_node(int cpu)
+{
+	return 0;
+}
+
+#endif	/* CONFIG_NUMA */
+
+#endif	/* _ASM_LOONGARCH_NUMA_H */

arch/loongarch/include/asm/pgtable.h

@@ -541,6 +541,18 @@ static inline pmd_t pmdp_huge_get_and_clear(struct mm_struct *mm,
 
 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 
+#ifdef CONFIG_NUMA_BALANCING
+static inline long pte_protnone(pte_t pte)
+{
+	return (pte_val(pte) & _PAGE_PROTNONE);
+}
+
+static inline long pmd_protnone(pmd_t pmd)
+{
+	return (pmd_val(pmd) & _PAGE_PROTNONE);
+}
+#endif /* CONFIG_NUMA_BALANCING */
+
 /*
  * We provide our own get_unmapped area to cope with the virtual aliasing
  * constraints placed on us by the cache architecture.

arch/loongarch/include/asm/topology.h

@@ -7,6 +7,27 @@
 
 #include <linux/smp.h>
 
+#ifdef CONFIG_NUMA
+
+extern cpumask_t cpus_on_node[];
+
+#define cpumask_of_node(node)  (&cpus_on_node[node])
+
+struct pci_bus;
+extern int pcibus_to_node(struct pci_bus *);
+
+#define cpumask_of_pcibus(bus)	(cpu_online_mask)
+
+extern unsigned char node_distances[MAX_NUMNODES][MAX_NUMNODES];
+
+void numa_set_distance(int from, int to, int distance);
+
+#define node_distance(from, to)	(node_distances[(from)][(to)])
+
+#else
+#define pcibus_to_node(bus)	0
+#endif
+
 #ifdef CONFIG_SMP
 #define topology_physical_package_id(cpu)	(cpu_data[cpu].package)
 #define topology_core_id(cpu)			(cpu_data[cpu].core)

arch/loongarch/kernel/Makefile

@@ -20,4 +20,6 @@ obj-$(CONFIG_PROC_FS)		+= proc.o
 
 obj-$(CONFIG_SMP)		+= smp.o
 
+obj-$(CONFIG_NUMA)		+= numa.o
+
 CPPFLAGS_vmlinux.lds		:= $(KBUILD_CFLAGS)

arch/loongarch/kernel/acpi.c

@@ -14,6 +14,7 @@
 #include <linux/memblock.h>
 #include <linux/serial_core.h>
 #include <asm/io.h>
+#include <asm/numa.h>
 #include <asm/loongson.h>
 
 int acpi_disabled;
@@ -199,6 +200,79 @@ int __init acpi_boot_init(void)
 	return 0;
 }
 
+#ifdef CONFIG_ACPI_NUMA
+
+static __init int setup_node(int pxm)
+{
+	return acpi_map_pxm_to_node(pxm);
+}
+
+/*
+ * Callback for SLIT parsing.  pxm_to_node() returns NUMA_NO_NODE for
+ * I/O localities since SRAT does not list them.  I/O localities are
+ * not supported at this point.
+ */
+unsigned int numa_distance_cnt;
+
+static inline unsigned int get_numa_distances_cnt(struct acpi_table_slit *slit)
+{
+	return slit->locality_count;
+}
+
+void __init numa_set_distance(int from, int to, int distance)
+{
+	if ((u8)distance != distance || (from == to && distance != LOCAL_DISTANCE)) {
+		pr_warn_once("Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
+				from, to, distance);
+		return;
+	}
+
+	node_distances[from][to] = distance;
+}
+
+/* Callback for Proximity Domain -> CPUID mapping */
+void __init
+acpi_numa_processor_affinity_init(struct acpi_srat_cpu_affinity *pa)
+{
+	int pxm, node;
+
+	if (srat_disabled())
+		return;
+	if (pa->header.length != sizeof(struct acpi_srat_cpu_affinity)) {
+		bad_srat();
+		return;
+	}
+	if ((pa->flags & ACPI_SRAT_CPU_ENABLED) == 0)
+		return;
+	pxm = pa->proximity_domain_lo;
+	if (acpi_srat_revision >= 2) {
+		pxm |= (pa->proximity_domain_hi[0] << 8);
+		pxm |= (pa->proximity_domain_hi[1] << 16);
+		pxm |= (pa->proximity_domain_hi[2] << 24);
+	}
+	node = setup_node(pxm);
+	if (node < 0) {
+		pr_err("SRAT: Too many proximity domains %x\n", pxm);
+		bad_srat();
+		return;
+	}
+
+	if (pa->apic_id >= CONFIG_NR_CPUS) {
+		pr_info("SRAT: PXM %u -> CPU 0x%02x -> Node %u skipped apicid that is too big\n",
+				pxm, pa->apic_id, node);
+		return;
+	}
+
+	early_numa_add_cpu(pa->apic_id, node);
+
+	set_cpuid_to_node(pa->apic_id, node);
+	node_set(node, numa_nodes_parsed);
+	pr_info("SRAT: PXM %u -> CPU 0x%02x -> Node %u\n", pxm, pa->apic_id, node);
+}
+
+void __init acpi_numa_arch_fixup(void) {}
+#endif
+
 void __init arch_reserve_mem_area(acpi_physical_address addr, size_t size)
 {
 	memblock_reserve(addr, size);
@@ -208,6 +282,22 @@ void __init arch_reserve_mem_area(acpi_physical_address addr, size_t size)
 
 #include <acpi/processor.h>
 
+static int __ref acpi_map_cpu2node(acpi_handle handle, int cpu, int physid)
+{
+#ifdef CONFIG_ACPI_NUMA
+	int nid;
+
+	nid = acpi_get_node(handle);
+	if (nid != NUMA_NO_NODE) {
+		set_cpuid_to_node(physid, nid);
+		node_set(nid, numa_nodes_parsed);
+		set_cpu_numa_node(cpu, nid);
+		cpumask_set_cpu(cpu, cpumask_of_node(nid));
+	}
+#endif
+	return 0;
+}
+
 int acpi_map_cpu(acpi_handle handle, phys_cpuid_t physid, u32 acpi_id, int *pcpu)
 {
 	int cpu;
@@ -218,6 +308,8 @@ int acpi_map_cpu(acpi_handle handle, phys_cpuid_t physid, u32 acpi_id, int *pcpu
 		return cpu;
 	}
 
+	acpi_map_cpu2node(handle, cpu, physid);
+
 	*pcpu = cpu;
 
 	return 0;
@@ -226,6 +318,9 @@ EXPORT_SYMBOL(acpi_map_cpu);
 
 int acpi_unmap_cpu(int cpu)
 {
+#ifdef CONFIG_ACPI_NUMA
+	set_cpuid_to_node(cpu_logical_map(cpu), NUMA_NO_NODE);
+#endif
 	set_cpu_present(cpu, false);
 	num_processors--;