xref: /linux/arch/loongarch/kernel/numa.c (revision 3a39d672e7f48b8d6b91a09afa4b55352773b4b5)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Author:  Xiang Gao <gaoxiang@loongson.cn>
4  *          Huacai Chen <chenhuacai@loongson.cn>
5  *
6  * Copyright (C) 2020-2022 Loongson Technology Corporation Limited
7  */
8 #include <linux/init.h>
9 #include <linux/kernel.h>
10 #include <linux/mm.h>
11 #include <linux/mmzone.h>
12 #include <linux/export.h>
13 #include <linux/nodemask.h>
14 #include <linux/swap.h>
15 #include <linux/memblock.h>
16 #include <linux/pfn.h>
17 #include <linux/acpi.h>
18 #include <linux/efi.h>
19 #include <linux/irq.h>
20 #include <linux/pci.h>
21 #include <asm/bootinfo.h>
22 #include <asm/loongson.h>
23 #include <asm/numa.h>
24 #include <asm/page.h>
25 #include <asm/pgalloc.h>
26 #include <asm/sections.h>
27 #include <asm/time.h>
28 
29 int numa_off;
30 unsigned char node_distances[MAX_NUMNODES][MAX_NUMNODES];
31 EXPORT_SYMBOL(node_distances);
32 
33 static struct numa_meminfo numa_meminfo;
34 cpumask_t cpus_on_node[MAX_NUMNODES];
35 cpumask_t phys_cpus_on_node[MAX_NUMNODES];
36 EXPORT_SYMBOL(cpus_on_node);
37 
38 /*
39  * apicid, cpu, node mappings
40  */
41 s16 __cpuid_to_node[CONFIG_NR_CPUS] = {
42 	[0 ... CONFIG_NR_CPUS - 1] = NUMA_NO_NODE
43 };
44 EXPORT_SYMBOL(__cpuid_to_node);
45 
46 nodemask_t numa_nodes_parsed __initdata;
47 
48 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
49 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
50 EXPORT_SYMBOL(__per_cpu_offset);
51 
pcpu_cpu_to_node(int cpu)52 static int __init pcpu_cpu_to_node(int cpu)
53 {
54 	return early_cpu_to_node(cpu);
55 }
56 
pcpu_cpu_distance(unsigned int from,unsigned int to)57 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
58 {
59 	if (early_cpu_to_node(from) == early_cpu_to_node(to))
60 		return LOCAL_DISTANCE;
61 	else
62 		return REMOTE_DISTANCE;
63 }
64 
pcpu_populate_pte(unsigned long addr)65 void __init pcpu_populate_pte(unsigned long addr)
66 {
67 	populate_kernel_pte(addr);
68 }
69 
setup_per_cpu_areas(void)70 void __init setup_per_cpu_areas(void)
71 {
72 	unsigned long delta;
73 	unsigned int cpu;
74 	int rc = -EINVAL;
75 
76 	if (pcpu_chosen_fc == PCPU_FC_AUTO) {
77 		if (nr_node_ids >= 8)
78 			pcpu_chosen_fc = PCPU_FC_PAGE;
79 		else
80 			pcpu_chosen_fc = PCPU_FC_EMBED;
81 	}
82 
83 	/*
84 	 * Always reserve area for module percpu variables.  That's
85 	 * what the legacy allocator did.
86 	 */
87 	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
88 		rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
89 					    PERCPU_DYNAMIC_RESERVE, PMD_SIZE,
90 					    pcpu_cpu_distance, pcpu_cpu_to_node);
91 		if (rc < 0)
92 			pr_warn("%s allocator failed (%d), falling back to page size\n",
93 				pcpu_fc_names[pcpu_chosen_fc], rc);
94 	}
95 	if (rc < 0)
96 		rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, pcpu_cpu_to_node);
97 	if (rc < 0)
98 		panic("cannot initialize percpu area (err=%d)", rc);
99 
100 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
101 	for_each_possible_cpu(cpu)
102 		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
103 }
104 #endif
105 
106 /*
107  * Get nodeid by logical cpu number.
108  * __cpuid_to_node maps phyical cpu id to node, so we
109  * should use cpu_logical_map(cpu) to index it.
110  *
111  * This routine is only used in early phase during
112  * booting, after setup_per_cpu_areas calling and numa_node
113  * initialization, cpu_to_node will be used instead.
114  */
early_cpu_to_node(int cpu)115 int early_cpu_to_node(int cpu)
116 {
117 	int physid = cpu_logical_map(cpu);
118 
119 	if (physid < 0)
120 		return NUMA_NO_NODE;
121 
122 	return __cpuid_to_node[physid];
123 }
124 
early_numa_add_cpu(int cpuid,s16 node)125 void __init early_numa_add_cpu(int cpuid, s16 node)
126 {
127 	int cpu = __cpu_number_map[cpuid];
128 
129 	if (cpu < 0)
130 		return;
131 
132 	cpumask_set_cpu(cpu, &cpus_on_node[node]);
133 	cpumask_set_cpu(cpuid, &phys_cpus_on_node[node]);
134 }
135 
numa_add_cpu(unsigned int cpu)136 void numa_add_cpu(unsigned int cpu)
137 {
138 	int nid = cpu_to_node(cpu);
139 	cpumask_set_cpu(cpu, &cpus_on_node[nid]);
140 }
141 
numa_remove_cpu(unsigned int cpu)142 void numa_remove_cpu(unsigned int cpu)
143 {
144 	int nid = cpu_to_node(cpu);
145 	cpumask_clear_cpu(cpu, &cpus_on_node[nid]);
146 }
147 
numa_add_memblk_to(int nid,u64 start,u64 end,struct numa_meminfo * mi)148 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
149 				     struct numa_meminfo *mi)
150 {
151 	/* ignore zero length blks */
152 	if (start == end)
153 		return 0;
154 
155 	/* whine about and ignore invalid blks */
156 	if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
157 		pr_warn("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
158 			   nid, start, end - 1);
159 		return 0;
160 	}
161 
162 	if (mi->nr_blks >= NR_NODE_MEMBLKS) {
163 		pr_err("NUMA: too many memblk ranges\n");
164 		return -EINVAL;
165 	}
166 
167 	mi->blk[mi->nr_blks].start = PFN_ALIGN(start);
168 	mi->blk[mi->nr_blks].end = PFN_ALIGN(end - PAGE_SIZE + 1);
169 	mi->blk[mi->nr_blks].nid = nid;
170 	mi->nr_blks++;
171 	return 0;
172 }
173 
174 /**
175  * numa_add_memblk - Add one numa_memblk to numa_meminfo
176  * @nid: NUMA node ID of the new memblk
177  * @start: Start address of the new memblk
178  * @end: End address of the new memblk
179  *
180  * Add a new memblk to the default numa_meminfo.
181  *
182  * RETURNS:
183  * 0 on success, -errno on failure.
184  */
numa_add_memblk(int nid,u64 start,u64 end)185 int __init numa_add_memblk(int nid, u64 start, u64 end)
186 {
187 	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
188 }
189 
node_mem_init(unsigned int node)190 static void __init node_mem_init(unsigned int node)
191 {
192 	unsigned long start_pfn, end_pfn;
193 	unsigned long node_addrspace_offset;
194 
195 	node_addrspace_offset = nid_to_addrbase(node);
196 	pr_info("Node%d's addrspace_offset is 0x%lx\n",
197 			node, node_addrspace_offset);
198 
199 	get_pfn_range_for_nid(node, &start_pfn, &end_pfn);
200 	pr_info("Node%d: start_pfn=0x%lx, end_pfn=0x%lx\n",
201 		node, start_pfn, end_pfn);
202 
203 	alloc_node_data(node);
204 }
205 
206 #ifdef CONFIG_ACPI_NUMA
207 
add_node_intersection(u32 node,u64 start,u64 size,u32 type)208 static void __init add_node_intersection(u32 node, u64 start, u64 size, u32 type)
209 {
210 	static unsigned long num_physpages;
211 
212 	num_physpages += (size >> PAGE_SHIFT);
213 	pr_info("Node%d: mem_type:%d, mem_start:0x%llx, mem_size:0x%llx Bytes\n",
214 		node, type, start, size);
215 	pr_info("       start_pfn:0x%llx, end_pfn:0x%llx, num_physpages:0x%lx\n",
216 		start >> PAGE_SHIFT, (start + size) >> PAGE_SHIFT, num_physpages);
217 	memblock_set_node(start, size, &memblock.memory, node);
218 }
219 
220 /*
221  * add_numamem_region
222  *
223  * Add a uasable memory region described by BIOS. The
224  * routine gets each intersection between BIOS's region
225  * and node's region, and adds them into node's memblock
226  * pool.
227  *
228  */
add_numamem_region(u64 start,u64 end,u32 type)229 static void __init add_numamem_region(u64 start, u64 end, u32 type)
230 {
231 	u32 i;
232 	u64 ofs = start;
233 
234 	if (start >= end) {
235 		pr_debug("Invalid region: %016llx-%016llx\n", start, end);
236 		return;
237 	}
238 
239 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
240 		struct numa_memblk *mb = &numa_meminfo.blk[i];
241 
242 		if (ofs > mb->end)
243 			continue;
244 
245 		if (end > mb->end) {
246 			add_node_intersection(mb->nid, ofs, mb->end - ofs, type);
247 			ofs = mb->end;
248 		} else {
249 			add_node_intersection(mb->nid, ofs, end - ofs, type);
250 			break;
251 		}
252 	}
253 }
254 
init_node_memblock(void)255 static void __init init_node_memblock(void)
256 {
257 	u32 mem_type;
258 	u64 mem_end, mem_start, mem_size;
259 	efi_memory_desc_t *md;
260 
261 	/* Parse memory information and activate */
262 	for_each_efi_memory_desc(md) {
263 		mem_type = md->type;
264 		mem_start = md->phys_addr;
265 		mem_size = md->num_pages << EFI_PAGE_SHIFT;
266 		mem_end = mem_start + mem_size;
267 
268 		switch (mem_type) {
269 		case EFI_LOADER_CODE:
270 		case EFI_LOADER_DATA:
271 		case EFI_BOOT_SERVICES_CODE:
272 		case EFI_BOOT_SERVICES_DATA:
273 		case EFI_PERSISTENT_MEMORY:
274 		case EFI_CONVENTIONAL_MEMORY:
275 			add_numamem_region(mem_start, mem_end, mem_type);
276 			break;
277 		case EFI_PAL_CODE:
278 		case EFI_UNUSABLE_MEMORY:
279 		case EFI_ACPI_RECLAIM_MEMORY:
280 			add_numamem_region(mem_start, mem_end, mem_type);
281 			fallthrough;
282 		case EFI_RESERVED_TYPE:
283 		case EFI_RUNTIME_SERVICES_CODE:
284 		case EFI_RUNTIME_SERVICES_DATA:
285 		case EFI_MEMORY_MAPPED_IO:
286 		case EFI_MEMORY_MAPPED_IO_PORT_SPACE:
287 			pr_info("Resvd: mem_type:%d, mem_start:0x%llx, mem_size:0x%llx Bytes\n",
288 					mem_type, mem_start, mem_size);
289 			break;
290 		}
291 	}
292 }
293 
numa_default_distance(void)294 static void __init numa_default_distance(void)
295 {
296 	int row, col;
297 
298 	for (row = 0; row < MAX_NUMNODES; row++)
299 		for (col = 0; col < MAX_NUMNODES; col++) {
300 			if (col == row)
301 				node_distances[row][col] = LOCAL_DISTANCE;
302 			else
303 				/* We assume that one node per package here!
304 				 *
305 				 * A SLIT should be used for multiple nodes
306 				 * per package to override default setting.
307 				 */
308 				node_distances[row][col] = REMOTE_DISTANCE;
309 	}
310 }
311 
312 /*
313  * fake_numa_init() - For Non-ACPI systems
314  * Return: 0 on success, -errno on failure.
315  */
fake_numa_init(void)316 static int __init fake_numa_init(void)
317 {
318 	phys_addr_t start = memblock_start_of_DRAM();
319 	phys_addr_t end = memblock_end_of_DRAM() - 1;
320 
321 	node_set(0, numa_nodes_parsed);
322 	pr_info("Faking a node at [mem %pap-%pap]\n", &start, &end);
323 
324 	return numa_add_memblk(0, start, end + 1);
325 }
326 
init_numa_memory(void)327 int __init init_numa_memory(void)
328 {
329 	int i;
330 	int ret;
331 	int node;
332 
333 	for (i = 0; i < NR_CPUS; i++)
334 		set_cpuid_to_node(i, NUMA_NO_NODE);
335 
336 	numa_default_distance();
337 	nodes_clear(numa_nodes_parsed);
338 	nodes_clear(node_possible_map);
339 	nodes_clear(node_online_map);
340 	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
341 
342 	/* Parse SRAT and SLIT if provided by firmware. */
343 	ret = acpi_disabled ? fake_numa_init() : acpi_numa_init();
344 	if (ret < 0)
345 		return ret;
346 
347 	node_possible_map = numa_nodes_parsed;
348 	if (WARN_ON(nodes_empty(node_possible_map)))
349 		return -EINVAL;
350 
351 	init_node_memblock();
352 	if (!memblock_validate_numa_coverage(SZ_1M))
353 		return -EINVAL;
354 
355 	for_each_node_mask(node, node_possible_map) {
356 		node_mem_init(node);
357 		node_set_online(node);
358 	}
359 	max_low_pfn = PHYS_PFN(memblock_end_of_DRAM());
360 
361 	setup_nr_node_ids();
362 	loongson_sysconf.nr_nodes = nr_node_ids;
363 	loongson_sysconf.cores_per_node = cpumask_weight(&phys_cpus_on_node[0]);
364 
365 	return 0;
366 }
367 
368 #endif
369 
paging_init(void)370 void __init paging_init(void)
371 {
372 	unsigned int node;
373 	unsigned long zones_size[MAX_NR_ZONES] = {0, };
374 
375 	for_each_online_node(node) {
376 		unsigned long start_pfn, end_pfn;
377 
378 		get_pfn_range_for_nid(node, &start_pfn, &end_pfn);
379 
380 		if (end_pfn > max_low_pfn)
381 			max_low_pfn = end_pfn;
382 	}
383 #ifdef CONFIG_ZONE_DMA32
384 	zones_size[ZONE_DMA32] = MAX_DMA32_PFN;
385 #endif
386 	zones_size[ZONE_NORMAL] = max_low_pfn;
387 	free_area_init(zones_size);
388 }
389 
mem_init(void)390 void __init mem_init(void)
391 {
392 	high_memory = (void *) __va(max_low_pfn << PAGE_SHIFT);
393 	memblock_free_all();
394 }
395 
pcibus_to_node(struct pci_bus * bus)396 int pcibus_to_node(struct pci_bus *bus)
397 {
398 	return dev_to_node(&bus->dev);
399 }
400 EXPORT_SYMBOL(pcibus_to_node);
401