xref: /linux/arch/x86/mm/numa.c (revision c94cd9508b1335b949fd13ebd269313c65492df0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Common code for 32 and 64-bit NUMA */
3 #include <linux/acpi.h>
4 #include <linux/kernel.h>
5 #include <linux/mm.h>
6 #include <linux/of.h>
7 #include <linux/string.h>
8 #include <linux/init.h>
9 #include <linux/memblock.h>
10 #include <linux/mmzone.h>
11 #include <linux/ctype.h>
12 #include <linux/nodemask.h>
13 #include <linux/sched.h>
14 #include <linux/topology.h>
15 #include <linux/sort.h>
16 #include <linux/numa_memblks.h>
17 
18 #include <asm/e820/api.h>
19 #include <asm/proto.h>
20 #include <asm/dma.h>
21 #include <asm/amd_nb.h>
22 
23 #include "numa_internal.h"
24 
25 int numa_off;
26 
27 static __init int numa_setup(char *opt)
28 {
29 	if (!opt)
30 		return -EINVAL;
31 	if (!strncmp(opt, "off", 3))
32 		numa_off = 1;
33 	if (!strncmp(opt, "fake=", 5))
34 		return numa_emu_cmdline(opt + 5);
35 	if (!strncmp(opt, "noacpi", 6))
36 		disable_srat();
37 	if (!strncmp(opt, "nohmat", 6))
38 		disable_hmat();
39 	return 0;
40 }
41 early_param("numa", numa_setup);
42 
43 /*
44  * apicid, cpu, node mappings
45  */
46 s16 __apicid_to_node[MAX_LOCAL_APIC] = {
47 	[0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
48 };
49 
50 int numa_cpu_node(int cpu)
51 {
52 	u32 apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
53 
54 	if (apicid != BAD_APICID)
55 		return __apicid_to_node[apicid];
56 	return NUMA_NO_NODE;
57 }
58 
59 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
60 EXPORT_SYMBOL(node_to_cpumask_map);
61 
62 /*
63  * Map cpu index to node index
64  */
65 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
66 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
67 
68 void numa_set_node(int cpu, int node)
69 {
70 	int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
71 
72 	/* early setting, no percpu area yet */
73 	if (cpu_to_node_map) {
74 		cpu_to_node_map[cpu] = node;
75 		return;
76 	}
77 
78 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
79 	if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
80 		printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
81 		dump_stack();
82 		return;
83 	}
84 #endif
85 	per_cpu(x86_cpu_to_node_map, cpu) = node;
86 
87 	set_cpu_numa_node(cpu, node);
88 }
89 
90 void numa_clear_node(int cpu)
91 {
92 	numa_set_node(cpu, NUMA_NO_NODE);
93 }
94 
95 /*
96  * Allocate node_to_cpumask_map based on number of available nodes
97  * Requires node_possible_map to be valid.
98  *
99  * Note: cpumask_of_node() is not valid until after this is done.
100  * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
101  */
102 void __init setup_node_to_cpumask_map(void)
103 {
104 	unsigned int node;
105 
106 	/* setup nr_node_ids if not done yet */
107 	if (nr_node_ids == MAX_NUMNODES)
108 		setup_nr_node_ids();
109 
110 	/* allocate the map */
111 	for (node = 0; node < nr_node_ids; node++)
112 		alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
113 
114 	/* cpumask_of_node() will now work */
115 	pr_debug("Node to cpumask map for %u nodes\n", nr_node_ids);
116 }
117 
118 static int __init numa_register_nodes(void)
119 {
120 	int nid;
121 
122 	if (!memblock_validate_numa_coverage(SZ_1M))
123 		return -EINVAL;
124 
125 	/* Finally register nodes. */
126 	for_each_node_mask(nid, node_possible_map) {
127 		unsigned long start_pfn, end_pfn;
128 
129 		/*
130 		 * Note, get_pfn_range_for_nid() depends on
131 		 * memblock_set_node() having already happened
132 		 */
133 		get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
134 		if (start_pfn >= end_pfn)
135 			continue;
136 
137 		alloc_node_data(nid);
138 		node_set_online(nid);
139 	}
140 
141 	/* Dump memblock with node info and return. */
142 	memblock_dump_all();
143 	return 0;
144 }
145 
146 /*
147  * There are unfortunately some poorly designed mainboards around that
148  * only connect memory to a single CPU. This breaks the 1:1 cpu->node
149  * mapping. To avoid this fill in the mapping for all possible CPUs,
150  * as the number of CPUs is not known yet. We round robin the existing
151  * nodes.
152  */
153 static void __init numa_init_array(void)
154 {
155 	int rr, i;
156 
157 	rr = first_node(node_online_map);
158 	for (i = 0; i < nr_cpu_ids; i++) {
159 		if (early_cpu_to_node(i) != NUMA_NO_NODE)
160 			continue;
161 		numa_set_node(i, rr);
162 		rr = next_node_in(rr, node_online_map);
163 	}
164 }
165 
166 static int __init numa_init(int (*init_func)(void))
167 {
168 	int i;
169 	int ret;
170 
171 	for (i = 0; i < MAX_LOCAL_APIC; i++)
172 		set_apicid_to_node(i, NUMA_NO_NODE);
173 
174 	ret = numa_memblks_init(init_func, /* memblock_force_top_down */ true);
175 	if (ret < 0)
176 		return ret;
177 
178 	ret = numa_register_nodes();
179 	if (ret < 0)
180 		return ret;
181 
182 	for (i = 0; i < nr_cpu_ids; i++) {
183 		int nid = early_cpu_to_node(i);
184 
185 		if (nid == NUMA_NO_NODE)
186 			continue;
187 		if (!node_online(nid))
188 			numa_clear_node(i);
189 	}
190 	numa_init_array();
191 
192 	return 0;
193 }
194 
195 /**
196  * dummy_numa_init - Fallback dummy NUMA init
197  *
198  * Used if there's no underlying NUMA architecture, NUMA initialization
199  * fails, or NUMA is disabled on the command line.
200  *
201  * Must online at least one node and add memory blocks that cover all
202  * allowed memory.  This function must not fail.
203  */
204 static int __init dummy_numa_init(void)
205 {
206 	printk(KERN_INFO "%s\n",
207 	       numa_off ? "NUMA turned off" : "No NUMA configuration found");
208 	printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
209 	       0LLU, PFN_PHYS(max_pfn) - 1);
210 
211 	node_set(0, numa_nodes_parsed);
212 	numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
213 
214 	return 0;
215 }
216 
217 /**
218  * x86_numa_init - Initialize NUMA
219  *
220  * Try each configured NUMA initialization method until one succeeds.  The
221  * last fallback is dummy single node config encompassing whole memory and
222  * never fails.
223  */
224 void __init x86_numa_init(void)
225 {
226 	if (!numa_off) {
227 #ifdef CONFIG_ACPI_NUMA
228 		if (!numa_init(x86_acpi_numa_init))
229 			return;
230 #endif
231 #ifdef CONFIG_AMD_NUMA
232 		if (!numa_init(amd_numa_init))
233 			return;
234 #endif
235 		if (acpi_disabled && !numa_init(of_numa_init))
236 			return;
237 	}
238 
239 	numa_init(dummy_numa_init);
240 }
241 
242 
243 /*
244  * A node may exist which has one or more Generic Initiators but no CPUs and no
245  * memory.
246  *
247  * This function must be called after init_cpu_to_node(), to ensure that any
248  * memoryless CPU nodes have already been brought online, and before the
249  * node_data[nid] is needed for zone list setup in build_all_zonelists().
250  *
251  * When this function is called, any nodes containing either memory and/or CPUs
252  * will already be online and there is no need to do anything extra, even if
253  * they also contain one or more Generic Initiators.
254  */
255 void __init init_gi_nodes(void)
256 {
257 	int nid;
258 
259 	/*
260 	 * Exclude this node from
261 	 * bringup_nonboot_cpus
262 	 *  cpu_up
263 	 *   __try_online_node
264 	 *    register_one_node
265 	 * because node_subsys is not initialized yet.
266 	 * TODO remove dependency on node_online
267 	 */
268 	for_each_node_state(nid, N_GENERIC_INITIATOR)
269 		if (!node_online(nid))
270 			node_set_online(nid);
271 }
272 
273 /*
274  * Setup early cpu_to_node.
275  *
276  * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
277  * and apicid_to_node[] tables have valid entries for a CPU.
278  * This means we skip cpu_to_node[] initialisation for NUMA
279  * emulation and faking node case (when running a kernel compiled
280  * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
281  * is already initialized in a round robin manner at numa_init_array,
282  * prior to this call, and this initialization is good enough
283  * for the fake NUMA cases.
284  *
285  * Called before the per_cpu areas are setup.
286  */
287 void __init init_cpu_to_node(void)
288 {
289 	int cpu;
290 	u32 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
291 
292 	BUG_ON(cpu_to_apicid == NULL);
293 
294 	for_each_possible_cpu(cpu) {
295 		int node = numa_cpu_node(cpu);
296 
297 		if (node == NUMA_NO_NODE)
298 			continue;
299 
300 		/*
301 		 * Exclude this node from
302 		 * bringup_nonboot_cpus
303 		 *  cpu_up
304 		 *   __try_online_node
305 		 *    register_one_node
306 		 * because node_subsys is not initialized yet.
307 		 * TODO remove dependency on node_online
308 		 */
309 		if (!node_online(node))
310 			node_set_online(node);
311 
312 		numa_set_node(cpu, node);
313 	}
314 }
315 
316 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
317 
318 # ifndef CONFIG_NUMA_EMU
319 void numa_add_cpu(unsigned int cpu)
320 {
321 	cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
322 }
323 
324 void numa_remove_cpu(unsigned int cpu)
325 {
326 	cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
327 }
328 # endif	/* !CONFIG_NUMA_EMU */
329 
330 #else	/* !CONFIG_DEBUG_PER_CPU_MAPS */
331 
332 int __cpu_to_node(int cpu)
333 {
334 	if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
335 		printk(KERN_WARNING
336 			"cpu_to_node(%d): usage too early!\n", cpu);
337 		dump_stack();
338 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
339 	}
340 	return per_cpu(x86_cpu_to_node_map, cpu);
341 }
342 EXPORT_SYMBOL(__cpu_to_node);
343 
344 /*
345  * Same function as cpu_to_node() but used if called before the
346  * per_cpu areas are setup.
347  */
348 int early_cpu_to_node(int cpu)
349 {
350 	if (early_per_cpu_ptr(x86_cpu_to_node_map))
351 		return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
352 
353 	if (!cpu_possible(cpu)) {
354 		printk(KERN_WARNING
355 			"early_cpu_to_node(%d): no per_cpu area!\n", cpu);
356 		dump_stack();
357 		return NUMA_NO_NODE;
358 	}
359 	return per_cpu(x86_cpu_to_node_map, cpu);
360 }
361 
362 void debug_cpumask_set_cpu(unsigned int cpu, int node, bool enable)
363 {
364 	struct cpumask *mask;
365 
366 	if (node == NUMA_NO_NODE) {
367 		/* early_cpu_to_node() already emits a warning and trace */
368 		return;
369 	}
370 	mask = node_to_cpumask_map[node];
371 	if (!cpumask_available(mask)) {
372 		pr_err("node_to_cpumask_map[%i] NULL\n", node);
373 		dump_stack();
374 		return;
375 	}
376 
377 	if (enable)
378 		cpumask_set_cpu(cpu, mask);
379 	else
380 		cpumask_clear_cpu(cpu, mask);
381 
382 	printk(KERN_DEBUG "%s cpu %d node %d: mask now %*pbl\n",
383 		enable ? "numa_add_cpu" : "numa_remove_cpu",
384 		cpu, node, cpumask_pr_args(mask));
385 	return;
386 }
387 
388 # ifndef CONFIG_NUMA_EMU
389 static void numa_set_cpumask(int cpu, bool enable)
390 {
391 	debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
392 }
393 
394 void numa_add_cpu(unsigned int cpu)
395 {
396 	numa_set_cpumask(cpu, true);
397 }
398 
399 void numa_remove_cpu(unsigned int cpu)
400 {
401 	numa_set_cpumask(cpu, false);
402 }
403 # endif	/* !CONFIG_NUMA_EMU */
404 
405 /*
406  * Returns a pointer to the bitmask of CPUs on Node 'node'.
407  */
408 const struct cpumask *cpumask_of_node(int node)
409 {
410 	if ((unsigned)node >= nr_node_ids) {
411 		printk(KERN_WARNING
412 			"cpumask_of_node(%d): (unsigned)node >= nr_node_ids(%u)\n",
413 			node, nr_node_ids);
414 		dump_stack();
415 		return cpu_none_mask;
416 	}
417 	if (!cpumask_available(node_to_cpumask_map[node])) {
418 		printk(KERN_WARNING
419 			"cpumask_of_node(%d): no node_to_cpumask_map!\n",
420 			node);
421 		dump_stack();
422 		return cpu_online_mask;
423 	}
424 	return node_to_cpumask_map[node];
425 }
426 EXPORT_SYMBOL(cpumask_of_node);
427 
428 #endif	/* !CONFIG_DEBUG_PER_CPU_MAPS */
429 
430 #ifdef CONFIG_NUMA_EMU
431 void __init numa_emu_update_cpu_to_node(int *emu_nid_to_phys,
432 					unsigned int nr_emu_nids)
433 {
434 	int i, j;
435 
436 	/*
437 	 * Transform __apicid_to_node table to use emulated nids by
438 	 * reverse-mapping phys_nid.  The maps should always exist but fall
439 	 * back to zero just in case.
440 	 */
441 	for (i = 0; i < ARRAY_SIZE(__apicid_to_node); i++) {
442 		if (__apicid_to_node[i] == NUMA_NO_NODE)
443 			continue;
444 		for (j = 0; j < nr_emu_nids; j++)
445 			if (__apicid_to_node[i] == emu_nid_to_phys[j])
446 				break;
447 		__apicid_to_node[i] = j < nr_emu_nids ? j : 0;
448 	}
449 }
450 
451 u64 __init numa_emu_dma_end(void)
452 {
453 	return PFN_PHYS(MAX_DMA32_PFN);
454 }
455 #endif /* CONFIG_NUMA_EMU */
456