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 <git@xen0n.name>
Reviewed-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>

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

+/* 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

+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Author: Huacai Chen (chenhuacai@loongson.cn)
+ * 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

+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Author: Jianmin Lv <lvjianmin@loongson.cn>
+ *         Huacai Chen <chenhuacai@loongson.cn>
+ *
+ * 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--;
 

arch/loongarch/kernel/dma.c

+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
+ */
+#include <linux/init.h>
+#include <linux/dma-direct.h>
+#include <linux/dma-mapping.h>
+#include <linux/dma-map-ops.h>
+#include <linux/swiotlb.h>
+
+#include <asm/bootinfo.h>
+#include <asm/dma.h>
+#include <asm/loongson.h>
+
+/*
+ * We extract 4bit node id (bit 44~47) from Loongson-3's
+ * 48bit physical address space and embed it into 40bit.
+ */
+
+static int node_id_offset;
+
+dma_addr_t phys_to_dma(struct device *dev, phys_addr_t paddr)
+{
+	long nid = (paddr >> 44) & 0xf;
+
+	return ((nid << 44) ^ paddr) | (nid << node_id_offset);
+}
+
+phys_addr_t dma_to_phys(struct device *dev, dma_addr_t daddr)
+{
+	long nid = (daddr >> node_id_offset) & 0xf;
+
+	return ((nid << node_id_offset) ^ daddr) | (nid << 44);
+}
+
+void __init plat_swiotlb_setup(void)
+{
+	swiotlb_init(true, SWIOTLB_VERBOSE);
+	node_id_offset = ((readl(LS7A_DMA_CFG) & LS7A_DMA_NODE_MASK) >> LS7A_DMA_NODE_SHF) + 36;
+}

arch/loongarch/kernel/module.c

@@ -11,6 +11,7 @@
 #include <linux/moduleloader.h>
 #include <linux/elf.h>
 #include <linux/mm.h>
+#include <linux/numa.h>
 #include <linux/vmalloc.h>
 #include <linux/slab.h>
 #include <linux/fs.h>

arch/loongarch/kernel/setup.c

@@ -35,6 +35,7 @@
 #include <asm/dma.h>
 #include <asm/efi.h>
 #include <asm/loongson.h>
+#include <asm/numa.h>
 #include <asm/pgalloc.h>
 #include <asm/sections.h>
 #include <asm/setup.h>
@@ -185,7 +186,10 @@ static int __init early_parse_mem(char *p)
 		return -EINVAL;
 	}
 
-	memblock_add(start, size);
+	if (!IS_ENABLED(CONFIG_NUMA))
+		memblock_add(start, size);
+	else
+		memblock_add_node(start, size, pa_to_nid(start), MEMBLOCK_NONE);
 
 	return 0;
 }
@@ -203,6 +207,9 @@ void __init platform_init(void)
 	acpi_boot_init();
 #endif
 
+#ifdef CONFIG_NUMA
+	init_numa_memory();
+#endif
 	dmi_setup();
 	smbios_parse();
 	pr_info("The BIOS Version: %s\n", b_info.bios_version);
@@ -241,7 +248,7 @@ static void __init arch_mem_init(char **cmdline_p)
 	sparse_init();
 	memblock_set_bottom_up(true);
 
-	swiotlb_init(true, SWIOTLB_VERBOSE);
+	plat_swiotlb_setup();
 
 	dma_contiguous_reserve(PFN_PHYS(max_low_pfn));
 

arch/loongarch/kernel/smp.c

@@ -25,6 +25,7 @@
 #include <asm/idle.h>
 #include <asm/loongson.h>
 #include <asm/mmu_context.h>
+#include <asm/numa.h>
 #include <asm/processor.h>
 #include <asm/setup.h>
 #include <asm/time.h>
@@ -222,6 +223,9 @@ void loongson3_init_secondary(void)
 
 	iocsr_write32(0xffffffff, LOONGARCH_IOCSR_IPI_EN);
 
+#ifdef CONFIG_NUMA
+	numa_add_cpu(cpu);
+#endif
 	per_cpu(cpu_state, cpu) = CPU_ONLINE;
 	cpu_data[cpu].core =
 		     cpu_logical_map(cpu) % loongson_sysconf.cores_per_package;
@@ -254,6 +258,9 @@ int loongson3_cpu_disable(void)
 	if (io_master(cpu))
 		return -EBUSY;
 
+#ifdef CONFIG_NUMA
+	numa_remove_cpu(cpu);
+#endif
 	set_cpu_online(cpu, false);
 	calculate_cpu_foreign_map();
 	local_irq_save(flags);
@@ -478,14 +485,36 @@ void calculate_cpu_foreign_map(void)
 /* Preload SMP state for boot cpu */
 void smp_prepare_boot_cpu(void)
 {
-	unsigned int cpu;
+	unsigned int cpu, node, rr_node;
 
 	set_cpu_possible(0, true);
 	set_cpu_online(0, true);
 	set_my_cpu_offset(per_cpu_offset(0));
 
-	for_each_possible_cpu(cpu)
-		set_cpu_numa_node(cpu, 0);
+	rr_node = first_node(node_online_map);
+	for_each_possible_cpu(cpu) {
+		node = early_cpu_to_node(cpu);
+
+		/*
+		 * The mapping between present cpus and nodes has been
+		 * built during MADT and SRAT parsing.
+		 *
+		 * If possible cpus = present cpus here, early_cpu_to_node
+		 * will return valid node.
+		 *
+		 * If possible cpus > present cpus here (e.g. some possible
+		 * cpus will be added by cpu-hotplug later), for possible but
+		 * not present cpus, early_cpu_to_node will return NUMA_NO_NODE,
+		 * and we just map them to online nodes in round-robin way.
+		 * Once hotplugged, new correct mapping will be built for them.
+		 */
+		if (node != NUMA_NO_NODE)
+			set_cpu_numa_node(cpu, node);
+		else {
+			set_cpu_numa_node(cpu, rr_node);
+			rr_node = next_node_in(rr_node, node_online_map);
+		}
+	}
 }
 
 /* called from main before smp_init() */
@@ -651,17 +680,10 @@ void flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned l
 		on_each_cpu_mask(mm_cpumask(mm), flush_tlb_range_ipi, &fd, 1);
 	} else {
 		unsigned int cpu;
-		int exec = vma->vm_flags & VM_EXEC;
 
 		for_each_online_cpu(cpu) {
-			/*
-			 * flush_cache_range() will only fully flush icache if
-			 * the VMA is executable, otherwise we must invalidate
-			 * ASID without it appearing to has_valid_asid() as if
-			 * mm has been completely unused by that CPU.
-			 */
 			if (cpu != smp_processor_id() && cpu_context(cpu, mm))
-				cpu_context(cpu, mm) = !exec;
+				cpu_context(cpu, mm) = 0;
 		}
 		local_flush_tlb_range(vma, start, end);
 	}
@@ -706,14 +728,8 @@ void flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 		unsigned int cpu;
 
 		for_each_online_cpu(cpu) {
-			/*
-			 * flush_cache_page() only does partial flushes, so
-			 * invalidate ASID without it appearing to
-			 * has_valid_asid() as if mm has been completely unused
-			 * by that CPU.
-			 */
 			if (cpu != smp_processor_id() && cpu_context(cpu, vma->vm_mm))
-				cpu_context(cpu, vma->vm_mm) = 1;
+				cpu_context(cpu, vma->vm_mm) = 0;
 		}
 		local_flush_tlb_page(vma, page);
 	}

arch/loongarch/kernel/traps.c

@@ -630,7 +630,7 @@ asmlinkage void noinstr do_vint(struct pt_regs *regs, unsigned long sp)
 	irqentry_exit(regs, state);
 }
 
-extern void tlb_init(void);
+extern void tlb_init(int cpu);
 extern void cache_error_setup(void);
 
 unsigned long eentry;
@@ -669,7 +669,7 @@ void per_cpu_trap_init(int cpu)
 		for (i = 0; i < 64; i++)
 			set_handler(i * VECSIZE, handle_reserved, VECSIZE);
 
-	tlb_init();
+	tlb_init(cpu);
 	cpu_cache_init();
 }
 

arch/loongarch/mm/init.c

@@ -84,6 +84,7 @@ int __ref page_is_ram(unsigned long pfn)
 	return memblock_is_memory(addr) && !memblock_is_reserved(addr);
 }
 
+#ifndef CONFIG_NUMA
 void __init paging_init(void)
 {
 	unsigned long max_zone_pfns[MAX_NR_ZONES];
@@ -107,6 +108,7 @@ void __init mem_init(void)
 	memblock_free_all();
 	setup_zero_pages();	/* Setup zeroed pages.  */
 }
+#endif /* !CONFIG_NUMA */
 
 void __ref free_initmem(void)
 {
@@ -129,6 +131,17 @@ int arch_add_memory(int nid, u64 start, u64 size, struct mhp_params *params)
 	return ret;
 }
 
+#ifdef CONFIG_NUMA
+int memory_add_physaddr_to_nid(u64 start)
+{
+	int nid;
+
+	nid = pa_to_nid(start);
+	return nid;
+}
+EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
+#endif
+
 #ifdef CONFIG_MEMORY_HOTREMOVE
 void arch_remove_memory(u64 start, u64 size, struct vmem_altmap *altmap)
 {

arch/loongarch/mm/tlb.c

@@ -250,15 +250,18 @@ static void output_pgtable_bits_defines(void)
 	pr_debug("\n");
 }
 
-void setup_tlb_handler(void)
-{
-	static int run_once = 0;
+#ifdef CONFIG_NUMA
+static unsigned long pcpu_handlers[NR_CPUS];
+#endif
+extern long exception_handlers[VECSIZE * 128 / sizeof(long)];
 
+void setup_tlb_handler(int cpu)
+{
 	setup_ptwalker();
 	output_pgtable_bits_defines();
 
 	/* The tlb handlers are generated only once */
-	if (!run_once) {
+	if (cpu == 0) {
 		memcpy((void *)tlbrentry, handle_tlb_refill, 0x80);
 		local_flush_icache_range(tlbrentry, tlbrentry + 0x80);
 		set_handler(EXCCODE_TLBI * VECSIZE, handle_tlb_load, VECSIZE);
@@ -268,15 +271,35 @@ void setup_tlb_handler(void)
 		set_handler(EXCCODE_TLBNR * VECSIZE, handle_tlb_protect, VECSIZE);
 		set_handler(EXCCODE_TLBNX * VECSIZE, handle_tlb_protect, VECSIZE);
 		set_handler(EXCCODE_TLBPE * VECSIZE, handle_tlb_protect, VECSIZE);
-		run_once++;
 	}
+#ifdef CONFIG_NUMA
+	else {
+		void *addr;
+		struct page *page;
+		const int vec_sz = sizeof(exception_handlers);
+
+		if (pcpu_handlers[cpu])
+			return;
+
+		page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, get_order(vec_sz));
+		if (!page)
+			return;
+
+		addr = page_address(page);
+		pcpu_handlers[cpu] = virt_to_phys(addr);
+		memcpy((void *)addr, (void *)eentry, vec_sz);
+		local_flush_icache_range((unsigned long)addr, (unsigned long)addr + vec_sz);
+		csr_write64(pcpu_handlers[cpu], LOONGARCH_CSR_TLBRENTRY);
+		csr_write64(pcpu_handlers[cpu] + 80*VECSIZE, LOONGARCH_CSR_TLBRENTRY);
+	}
+#endif
 }
 
-void tlb_init(void)
+void tlb_init(int cpu)
 {
 	write_csr_pagesize(PS_DEFAULT_SIZE);
 	write_csr_stlbpgsize(PS_DEFAULT_SIZE);
 	write_csr_tlbrefill_pagesize(PS_DEFAULT_SIZE);
-	setup_tlb_handler();
+	setup_tlb_handler(cpu);
 	local_flush_tlb_all();
 }