xref: /openbmc/linux/arch/loongarch/kernel/numa.c (revision da1d9caf)
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 struct pglist_data *node_data[MAX_NUMNODES];
31 unsigned char node_distances[MAX_NUMNODES][MAX_NUMNODES];
32 
33 EXPORT_SYMBOL(node_data);
34 EXPORT_SYMBOL(node_distances);
35 
36 static struct numa_meminfo numa_meminfo;
37 cpumask_t cpus_on_node[MAX_NUMNODES];
38 cpumask_t phys_cpus_on_node[MAX_NUMNODES];
39 EXPORT_SYMBOL(cpus_on_node);
40 
41 /*
42  * apicid, cpu, node mappings
43  */
44 s16 __cpuid_to_node[CONFIG_NR_CPUS] = {
45 	[0 ... CONFIG_NR_CPUS - 1] = NUMA_NO_NODE
46 };
47 EXPORT_SYMBOL(__cpuid_to_node);
48 
49 nodemask_t numa_nodes_parsed __initdata;
50 
51 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
52 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
53 EXPORT_SYMBOL(__per_cpu_offset);
54 
55 static int __init pcpu_cpu_to_node(int cpu)
56 {
57 	return early_cpu_to_node(cpu);
58 }
59 
60 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
61 {
62 	if (early_cpu_to_node(from) == early_cpu_to_node(to))
63 		return LOCAL_DISTANCE;
64 	else
65 		return REMOTE_DISTANCE;
66 }
67 
68 void __init pcpu_populate_pte(unsigned long addr)
69 {
70 	pgd_t *pgd = pgd_offset_k(addr);
71 	p4d_t *p4d = p4d_offset(pgd, addr);
72 	pud_t *pud;
73 	pmd_t *pmd;
74 
75 	if (p4d_none(*p4d)) {
76 		pud_t *new;
77 
78 		new = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
79 		pgd_populate(&init_mm, pgd, new);
80 #ifndef __PAGETABLE_PUD_FOLDED
81 		pud_init((unsigned long)new, (unsigned long)invalid_pmd_table);
82 #endif
83 	}
84 
85 	pud = pud_offset(p4d, addr);
86 	if (pud_none(*pud)) {
87 		pmd_t *new;
88 
89 		new = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
90 		pud_populate(&init_mm, pud, new);
91 #ifndef __PAGETABLE_PMD_FOLDED
92 		pmd_init((unsigned long)new, (unsigned long)invalid_pte_table);
93 #endif
94 	}
95 
96 	pmd = pmd_offset(pud, addr);
97 	if (!pmd_present(*pmd)) {
98 		pte_t *new;
99 
100 		new = memblock_alloc(PAGE_SIZE, PAGE_SIZE);
101 		pmd_populate_kernel(&init_mm, pmd, new);
102 	}
103 }
104 
105 void __init setup_per_cpu_areas(void)
106 {
107 	unsigned long delta;
108 	unsigned int cpu;
109 	int rc = -EINVAL;
110 
111 	if (pcpu_chosen_fc == PCPU_FC_AUTO) {
112 		if (nr_node_ids >= 8)
113 			pcpu_chosen_fc = PCPU_FC_PAGE;
114 		else
115 			pcpu_chosen_fc = PCPU_FC_EMBED;
116 	}
117 
118 	/*
119 	 * Always reserve area for module percpu variables.  That's
120 	 * what the legacy allocator did.
121 	 */
122 	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
123 		rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
124 					    PERCPU_DYNAMIC_RESERVE, PMD_SIZE,
125 					    pcpu_cpu_distance, pcpu_cpu_to_node);
126 		if (rc < 0)
127 			pr_warn("%s allocator failed (%d), falling back to page size\n",
128 				pcpu_fc_names[pcpu_chosen_fc], rc);
129 	}
130 	if (rc < 0)
131 		rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, pcpu_cpu_to_node);
132 	if (rc < 0)
133 		panic("cannot initialize percpu area (err=%d)", rc);
134 
135 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
136 	for_each_possible_cpu(cpu)
137 		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
138 }
139 #endif
140 
141 /*
142  * Get nodeid by logical cpu number.
143  * __cpuid_to_node maps phyical cpu id to node, so we
144  * should use cpu_logical_map(cpu) to index it.
145  *
146  * This routine is only used in early phase during
147  * booting, after setup_per_cpu_areas calling and numa_node
148  * initialization, cpu_to_node will be used instead.
149  */
150 int early_cpu_to_node(int cpu)
151 {
152 	int physid = cpu_logical_map(cpu);
153 
154 	if (physid < 0)
155 		return NUMA_NO_NODE;
156 
157 	return __cpuid_to_node[physid];
158 }
159 
160 void __init early_numa_add_cpu(int cpuid, s16 node)
161 {
162 	int cpu = __cpu_number_map[cpuid];
163 
164 	if (cpu < 0)
165 		return;
166 
167 	cpumask_set_cpu(cpu, &cpus_on_node[node]);
168 	cpumask_set_cpu(cpuid, &phys_cpus_on_node[node]);
169 }
170 
171 void numa_add_cpu(unsigned int cpu)
172 {
173 	int nid = cpu_to_node(cpu);
174 	cpumask_set_cpu(cpu, &cpus_on_node[nid]);
175 }
176 
177 void numa_remove_cpu(unsigned int cpu)
178 {
179 	int nid = cpu_to_node(cpu);
180 	cpumask_clear_cpu(cpu, &cpus_on_node[nid]);
181 }
182 
183 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
184 				     struct numa_meminfo *mi)
185 {
186 	/* ignore zero length blks */
187 	if (start == end)
188 		return 0;
189 
190 	/* whine about and ignore invalid blks */
191 	if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
192 		pr_warn("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
193 			   nid, start, end - 1);
194 		return 0;
195 	}
196 
197 	if (mi->nr_blks >= NR_NODE_MEMBLKS) {
198 		pr_err("NUMA: too many memblk ranges\n");
199 		return -EINVAL;
200 	}
201 
202 	mi->blk[mi->nr_blks].start = PFN_ALIGN(start);
203 	mi->blk[mi->nr_blks].end = PFN_ALIGN(end - PAGE_SIZE + 1);
204 	mi->blk[mi->nr_blks].nid = nid;
205 	mi->nr_blks++;
206 	return 0;
207 }
208 
209 /**
210  * numa_add_memblk - Add one numa_memblk to numa_meminfo
211  * @nid: NUMA node ID of the new memblk
212  * @start: Start address of the new memblk
213  * @end: End address of the new memblk
214  *
215  * Add a new memblk to the default numa_meminfo.
216  *
217  * RETURNS:
218  * 0 on success, -errno on failure.
219  */
220 int __init numa_add_memblk(int nid, u64 start, u64 end)
221 {
222 	return numa_add_memblk_to(nid, start, end, &numa_meminfo);
223 }
224 
225 static void __init alloc_node_data(int nid)
226 {
227 	void *nd;
228 	unsigned long nd_pa;
229 	size_t nd_sz = roundup(sizeof(pg_data_t), PAGE_SIZE);
230 
231 	nd_pa = memblock_phys_alloc_try_nid(nd_sz, SMP_CACHE_BYTES, nid);
232 	if (!nd_pa) {
233 		pr_err("Cannot find %zu Byte for node_data (initial node: %d)\n", nd_sz, nid);
234 		return;
235 	}
236 
237 	nd = __va(nd_pa);
238 
239 	node_data[nid] = nd;
240 	memset(nd, 0, sizeof(pg_data_t));
241 }
242 
243 static void __init node_mem_init(unsigned int node)
244 {
245 	unsigned long start_pfn, end_pfn;
246 	unsigned long node_addrspace_offset;
247 
248 	node_addrspace_offset = nid_to_addrbase(node);
249 	pr_info("Node%d's addrspace_offset is 0x%lx\n",
250 			node, node_addrspace_offset);
251 
252 	get_pfn_range_for_nid(node, &start_pfn, &end_pfn);
253 	pr_info("Node%d: start_pfn=0x%lx, end_pfn=0x%lx\n",
254 		node, start_pfn, end_pfn);
255 
256 	alloc_node_data(node);
257 }
258 
259 #ifdef CONFIG_ACPI_NUMA
260 
261 /*
262  * Sanity check to catch more bad NUMA configurations (they are amazingly
263  * common).  Make sure the nodes cover all memory.
264  */
265 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
266 {
267 	int i;
268 	u64 numaram, biosram;
269 
270 	numaram = 0;
271 	for (i = 0; i < mi->nr_blks; i++) {
272 		u64 s = mi->blk[i].start >> PAGE_SHIFT;
273 		u64 e = mi->blk[i].end >> PAGE_SHIFT;
274 
275 		numaram += e - s;
276 		numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
277 		if ((s64)numaram < 0)
278 			numaram = 0;
279 	}
280 	max_pfn = max_low_pfn;
281 	biosram = max_pfn - absent_pages_in_range(0, max_pfn);
282 
283 	BUG_ON((s64)(biosram - numaram) >= (1 << (20 - PAGE_SHIFT)));
284 	return true;
285 }
286 
287 static void __init add_node_intersection(u32 node, u64 start, u64 size, u32 type)
288 {
289 	static unsigned long num_physpages;
290 
291 	num_physpages += (size >> PAGE_SHIFT);
292 	pr_info("Node%d: mem_type:%d, mem_start:0x%llx, mem_size:0x%llx Bytes\n",
293 		node, type, start, size);
294 	pr_info("       start_pfn:0x%llx, end_pfn:0x%llx, num_physpages:0x%lx\n",
295 		start >> PAGE_SHIFT, (start + size) >> PAGE_SHIFT, num_physpages);
296 	memblock_set_node(start, size, &memblock.memory, node);
297 }
298 
299 /*
300  * add_numamem_region
301  *
302  * Add a uasable memory region described by BIOS. The
303  * routine gets each intersection between BIOS's region
304  * and node's region, and adds them into node's memblock
305  * pool.
306  *
307  */
308 static void __init add_numamem_region(u64 start, u64 end, u32 type)
309 {
310 	u32 i;
311 	u64 ofs = start;
312 
313 	if (start >= end) {
314 		pr_debug("Invalid region: %016llx-%016llx\n", start, end);
315 		return;
316 	}
317 
318 	for (i = 0; i < numa_meminfo.nr_blks; i++) {
319 		struct numa_memblk *mb = &numa_meminfo.blk[i];
320 
321 		if (ofs > mb->end)
322 			continue;
323 
324 		if (end > mb->end) {
325 			add_node_intersection(mb->nid, ofs, mb->end - ofs, type);
326 			ofs = mb->end;
327 		} else {
328 			add_node_intersection(mb->nid, ofs, end - ofs, type);
329 			break;
330 		}
331 	}
332 }
333 
334 static void __init init_node_memblock(void)
335 {
336 	u32 mem_type;
337 	u64 mem_end, mem_start, mem_size;
338 	efi_memory_desc_t *md;
339 
340 	/* Parse memory information and activate */
341 	for_each_efi_memory_desc(md) {
342 		mem_type = md->type;
343 		mem_start = md->phys_addr;
344 		mem_size = md->num_pages << EFI_PAGE_SHIFT;
345 		mem_end = mem_start + mem_size;
346 
347 		switch (mem_type) {
348 		case EFI_LOADER_CODE:
349 		case EFI_LOADER_DATA:
350 		case EFI_BOOT_SERVICES_CODE:
351 		case EFI_BOOT_SERVICES_DATA:
352 		case EFI_PERSISTENT_MEMORY:
353 		case EFI_CONVENTIONAL_MEMORY:
354 			add_numamem_region(mem_start, mem_end, mem_type);
355 			break;
356 		case EFI_PAL_CODE:
357 		case EFI_UNUSABLE_MEMORY:
358 		case EFI_ACPI_RECLAIM_MEMORY:
359 			add_numamem_region(mem_start, mem_end, mem_type);
360 			fallthrough;
361 		case EFI_RESERVED_TYPE:
362 		case EFI_RUNTIME_SERVICES_CODE:
363 		case EFI_RUNTIME_SERVICES_DATA:
364 		case EFI_MEMORY_MAPPED_IO:
365 		case EFI_MEMORY_MAPPED_IO_PORT_SPACE:
366 			pr_info("Resvd: mem_type:%d, mem_start:0x%llx, mem_size:0x%llx Bytes\n",
367 					mem_type, mem_start, mem_size);
368 			break;
369 		}
370 	}
371 }
372 
373 static void __init numa_default_distance(void)
374 {
375 	int row, col;
376 
377 	for (row = 0; row < MAX_NUMNODES; row++)
378 		for (col = 0; col < MAX_NUMNODES; col++) {
379 			if (col == row)
380 				node_distances[row][col] = LOCAL_DISTANCE;
381 			else
382 				/* We assume that one node per package here!
383 				 *
384 				 * A SLIT should be used for multiple nodes
385 				 * per package to override default setting.
386 				 */
387 				node_distances[row][col] = REMOTE_DISTANCE;
388 	}
389 }
390 
391 int __init init_numa_memory(void)
392 {
393 	int i;
394 	int ret;
395 	int node;
396 
397 	for (i = 0; i < NR_CPUS; i++)
398 		set_cpuid_to_node(i, NUMA_NO_NODE);
399 
400 	numa_default_distance();
401 	nodes_clear(numa_nodes_parsed);
402 	nodes_clear(node_possible_map);
403 	nodes_clear(node_online_map);
404 	memset(&numa_meminfo, 0, sizeof(numa_meminfo));
405 
406 	/* Parse SRAT and SLIT if provided by firmware. */
407 	ret = acpi_numa_init();
408 	if (ret < 0)
409 		return ret;
410 
411 	node_possible_map = numa_nodes_parsed;
412 	if (WARN_ON(nodes_empty(node_possible_map)))
413 		return -EINVAL;
414 
415 	init_node_memblock();
416 	if (numa_meminfo_cover_memory(&numa_meminfo) == false)
417 		return -EINVAL;
418 
419 	for_each_node_mask(node, node_possible_map) {
420 		node_mem_init(node);
421 		node_set_online(node);
422 	}
423 	max_low_pfn = PHYS_PFN(memblock_end_of_DRAM());
424 
425 	setup_nr_node_ids();
426 	loongson_sysconf.nr_nodes = nr_node_ids;
427 	loongson_sysconf.cores_per_node = cpumask_weight(&phys_cpus_on_node[0]);
428 
429 	return 0;
430 }
431 
432 EXPORT_SYMBOL(init_numa_memory);
433 #endif
434 
435 void __init paging_init(void)
436 {
437 	unsigned int node;
438 	unsigned long zones_size[MAX_NR_ZONES] = {0, };
439 
440 	for_each_online_node(node) {
441 		unsigned long start_pfn, end_pfn;
442 
443 		get_pfn_range_for_nid(node, &start_pfn, &end_pfn);
444 
445 		if (end_pfn > max_low_pfn)
446 			max_low_pfn = end_pfn;
447 	}
448 #ifdef CONFIG_ZONE_DMA32
449 	zones_size[ZONE_DMA32] = MAX_DMA32_PFN;
450 #endif
451 	zones_size[ZONE_NORMAL] = max_low_pfn;
452 	free_area_init(zones_size);
453 }
454 
455 void __init mem_init(void)
456 {
457 	high_memory = (void *) __va(get_num_physpages() << PAGE_SHIFT);
458 	memblock_free_all();
459 	setup_zero_pages();	/* This comes from node 0 */
460 }
461 
462 int pcibus_to_node(struct pci_bus *bus)
463 {
464 	return dev_to_node(&bus->dev);
465 }
466 EXPORT_SYMBOL(pcibus_to_node);
467