xref: /openbmc/linux/arch/x86/mm/numa_64.c (revision f42b3800)
1 /*
2  * Generic VM initialization for x86-64 NUMA setups.
3  * Copyright 2002,2003 Andi Kleen, SuSE Labs.
4  */
5 #include <linux/kernel.h>
6 #include <linux/mm.h>
7 #include <linux/string.h>
8 #include <linux/init.h>
9 #include <linux/bootmem.h>
10 #include <linux/mmzone.h>
11 #include <linux/ctype.h>
12 #include <linux/module.h>
13 #include <linux/nodemask.h>
14 #include <linux/sched.h>
15 
16 #include <asm/e820.h>
17 #include <asm/proto.h>
18 #include <asm/dma.h>
19 #include <asm/numa.h>
20 #include <asm/acpi.h>
21 #include <asm/k8.h>
22 
23 #ifndef Dprintk
24 #define Dprintk(x...)
25 #endif
26 
27 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
28 EXPORT_SYMBOL(node_data);
29 
30 bootmem_data_t plat_node_bdata[MAX_NUMNODES];
31 
32 struct memnode memnode;
33 
34 #ifdef CONFIG_SMP
35 int x86_cpu_to_node_map_init[NR_CPUS] = {
36 	[0 ... NR_CPUS-1] = NUMA_NO_NODE
37 };
38 void *x86_cpu_to_node_map_early_ptr;
39 EXPORT_SYMBOL(x86_cpu_to_node_map_early_ptr);
40 #endif
41 DEFINE_PER_CPU(int, x86_cpu_to_node_map) = NUMA_NO_NODE;
42 EXPORT_PER_CPU_SYMBOL(x86_cpu_to_node_map);
43 
44 s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = {
45 	[0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
46 };
47 
48 cpumask_t node_to_cpumask_map[MAX_NUMNODES] __read_mostly;
49 EXPORT_SYMBOL(node_to_cpumask_map);
50 
51 int numa_off __initdata;
52 unsigned long __initdata nodemap_addr;
53 unsigned long __initdata nodemap_size;
54 
55 /*
56  * Given a shift value, try to populate memnodemap[]
57  * Returns :
58  * 1 if OK
59  * 0 if memnodmap[] too small (of shift too small)
60  * -1 if node overlap or lost ram (shift too big)
61  */
62 static int __init populate_memnodemap(const struct bootnode *nodes,
63 				      int numnodes, int shift)
64 {
65 	unsigned long addr, end;
66 	int i, res = -1;
67 
68 	memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize);
69 	for (i = 0; i < numnodes; i++) {
70 		addr = nodes[i].start;
71 		end = nodes[i].end;
72 		if (addr >= end)
73 			continue;
74 		if ((end >> shift) >= memnodemapsize)
75 			return 0;
76 		do {
77 			if (memnodemap[addr >> shift] != NUMA_NO_NODE)
78 				return -1;
79 			memnodemap[addr >> shift] = i;
80 			addr += (1UL << shift);
81 		} while (addr < end);
82 		res = 1;
83 	}
84 	return res;
85 }
86 
87 static int __init allocate_cachealigned_memnodemap(void)
88 {
89 	unsigned long addr;
90 
91 	memnodemap = memnode.embedded_map;
92 	if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map))
93 		return 0;
94 
95 	addr = 0x8000;
96 	nodemap_size = round_up(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES);
97 	nodemap_addr = find_e820_area(addr, end_pfn<<PAGE_SHIFT,
98 				      nodemap_size, L1_CACHE_BYTES);
99 	if (nodemap_addr == -1UL) {
100 		printk(KERN_ERR
101 		       "NUMA: Unable to allocate Memory to Node hash map\n");
102 		nodemap_addr = nodemap_size = 0;
103 		return -1;
104 	}
105 	memnodemap = phys_to_virt(nodemap_addr);
106 	reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP");
107 
108 	printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n",
109 	       nodemap_addr, nodemap_addr + nodemap_size);
110 	return 0;
111 }
112 
113 /*
114  * The LSB of all start and end addresses in the node map is the value of the
115  * maximum possible shift.
116  */
117 static int __init extract_lsb_from_nodes(const struct bootnode *nodes,
118 					 int numnodes)
119 {
120 	int i, nodes_used = 0;
121 	unsigned long start, end;
122 	unsigned long bitfield = 0, memtop = 0;
123 
124 	for (i = 0; i < numnodes; i++) {
125 		start = nodes[i].start;
126 		end = nodes[i].end;
127 		if (start >= end)
128 			continue;
129 		bitfield |= start;
130 		nodes_used++;
131 		if (end > memtop)
132 			memtop = end;
133 	}
134 	if (nodes_used <= 1)
135 		i = 63;
136 	else
137 		i = find_first_bit(&bitfield, sizeof(unsigned long)*8);
138 	memnodemapsize = (memtop >> i)+1;
139 	return i;
140 }
141 
142 int __init compute_hash_shift(struct bootnode *nodes, int numnodes)
143 {
144 	int shift;
145 
146 	shift = extract_lsb_from_nodes(nodes, numnodes);
147 	if (allocate_cachealigned_memnodemap())
148 		return -1;
149 	printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n",
150 		shift);
151 
152 	if (populate_memnodemap(nodes, numnodes, shift) != 1) {
153 		printk(KERN_INFO "Your memory is not aligned you need to "
154 		       "rebuild your kernel with a bigger NODEMAPSIZE "
155 		       "shift=%d\n", shift);
156 		return -1;
157 	}
158 	return shift;
159 }
160 
161 int early_pfn_to_nid(unsigned long pfn)
162 {
163 	return phys_to_nid(pfn << PAGE_SHIFT);
164 }
165 
166 static void * __init early_node_mem(int nodeid, unsigned long start,
167 				    unsigned long end, unsigned long size,
168 				    unsigned long align)
169 {
170 	unsigned long mem = find_e820_area(start, end, size, align);
171 	void *ptr;
172 
173 	if (mem != -1L)
174 		return __va(mem);
175 
176 	ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS));
177 	if (ptr == NULL) {
178 		printk(KERN_ERR "Cannot find %lu bytes in node %d\n",
179 		       size, nodeid);
180 		return NULL;
181 	}
182 	return ptr;
183 }
184 
185 /* Initialize bootmem allocator for a node */
186 void __init setup_node_bootmem(int nodeid, unsigned long start,
187 			       unsigned long end)
188 {
189 	unsigned long start_pfn, end_pfn, bootmap_pages, bootmap_size;
190 	unsigned long bootmap_start, nodedata_phys;
191 	void *bootmap;
192 	const int pgdat_size = round_up(sizeof(pg_data_t), PAGE_SIZE);
193 
194 	start = round_up(start, ZONE_ALIGN);
195 
196 	printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid,
197 	       start, end);
198 
199 	start_pfn = start >> PAGE_SHIFT;
200 	end_pfn = end >> PAGE_SHIFT;
201 
202 	node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size,
203 					   SMP_CACHE_BYTES);
204 	if (node_data[nodeid] == NULL)
205 		return;
206 	nodedata_phys = __pa(node_data[nodeid]);
207 	printk(KERN_INFO "  NODE_DATA [%016lx - %016lx]\n", nodedata_phys,
208 		nodedata_phys + pgdat_size - 1);
209 
210 	memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t));
211 	NODE_DATA(nodeid)->bdata = &plat_node_bdata[nodeid];
212 	NODE_DATA(nodeid)->node_start_pfn = start_pfn;
213 	NODE_DATA(nodeid)->node_spanned_pages = end_pfn - start_pfn;
214 
215 	/* Find a place for the bootmem map */
216 	bootmap_pages = bootmem_bootmap_pages(end_pfn - start_pfn);
217 	bootmap_start = round_up(nodedata_phys + pgdat_size, PAGE_SIZE);
218 	/*
219 	 * SMP_CAHCE_BYTES could be enough, but init_bootmem_node like
220 	 * to use that to align to PAGE_SIZE
221 	 */
222 	bootmap = early_node_mem(nodeid, bootmap_start, end,
223 				 bootmap_pages<<PAGE_SHIFT, PAGE_SIZE);
224 	if (bootmap == NULL)  {
225 		if (nodedata_phys < start || nodedata_phys >= end)
226 			free_bootmem(nodedata_phys, pgdat_size);
227 		node_data[nodeid] = NULL;
228 		return;
229 	}
230 	bootmap_start = __pa(bootmap);
231 
232 	bootmap_size = init_bootmem_node(NODE_DATA(nodeid),
233 					 bootmap_start >> PAGE_SHIFT,
234 					 start_pfn, end_pfn);
235 
236 	printk(KERN_INFO "  bootmap [%016lx -  %016lx] pages %lx\n",
237 		 bootmap_start, bootmap_start + bootmap_size - 1,
238 		 bootmap_pages);
239 
240 	free_bootmem_with_active_regions(nodeid, end);
241 
242 	reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, pgdat_size,
243 			BOOTMEM_DEFAULT);
244 	reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start,
245 			bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT);
246 #ifdef CONFIG_ACPI_NUMA
247 	srat_reserve_add_area(nodeid);
248 #endif
249 	node_set_online(nodeid);
250 }
251 
252 /*
253  * There are unfortunately some poorly designed mainboards around that
254  * only connect memory to a single CPU. This breaks the 1:1 cpu->node
255  * mapping. To avoid this fill in the mapping for all possible CPUs,
256  * as the number of CPUs is not known yet. We round robin the existing
257  * nodes.
258  */
259 void __init numa_init_array(void)
260 {
261 	int rr, i;
262 
263 	rr = first_node(node_online_map);
264 	for (i = 0; i < NR_CPUS; i++) {
265 		if (early_cpu_to_node(i) != NUMA_NO_NODE)
266 			continue;
267 		numa_set_node(i, rr);
268 		rr = next_node(rr, node_online_map);
269 		if (rr == MAX_NUMNODES)
270 			rr = first_node(node_online_map);
271 	}
272 }
273 
274 #ifdef CONFIG_NUMA_EMU
275 /* Numa emulation */
276 char *cmdline __initdata;
277 
278 /*
279  * Setups up nid to range from addr to addr + size.  If the end
280  * boundary is greater than max_addr, then max_addr is used instead.
281  * The return value is 0 if there is additional memory left for
282  * allocation past addr and -1 otherwise.  addr is adjusted to be at
283  * the end of the node.
284  */
285 static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr,
286 				   u64 size, u64 max_addr)
287 {
288 	int ret = 0;
289 
290 	nodes[nid].start = *addr;
291 	*addr += size;
292 	if (*addr >= max_addr) {
293 		*addr = max_addr;
294 		ret = -1;
295 	}
296 	nodes[nid].end = *addr;
297 	node_set(nid, node_possible_map);
298 	printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid,
299 	       nodes[nid].start, nodes[nid].end,
300 	       (nodes[nid].end - nodes[nid].start) >> 20);
301 	return ret;
302 }
303 
304 /*
305  * Splits num_nodes nodes up equally starting at node_start.  The return value
306  * is the number of nodes split up and addr is adjusted to be at the end of the
307  * last node allocated.
308  */
309 static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr,
310 				      u64 max_addr, int node_start,
311 				      int num_nodes)
312 {
313 	unsigned int big;
314 	u64 size;
315 	int i;
316 
317 	if (num_nodes <= 0)
318 		return -1;
319 	if (num_nodes > MAX_NUMNODES)
320 		num_nodes = MAX_NUMNODES;
321 	size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) /
322 	       num_nodes;
323 	/*
324 	 * Calculate the number of big nodes that can be allocated as a result
325 	 * of consolidating the leftovers.
326 	 */
327 	big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) /
328 	      FAKE_NODE_MIN_SIZE;
329 
330 	/* Round down to nearest FAKE_NODE_MIN_SIZE. */
331 	size &= FAKE_NODE_MIN_HASH_MASK;
332 	if (!size) {
333 		printk(KERN_ERR "Not enough memory for each node.  "
334 		       "NUMA emulation disabled.\n");
335 		return -1;
336 	}
337 
338 	for (i = node_start; i < num_nodes + node_start; i++) {
339 		u64 end = *addr + size;
340 
341 		if (i < big)
342 			end += FAKE_NODE_MIN_SIZE;
343 		/*
344 		 * The final node can have the remaining system RAM.  Other
345 		 * nodes receive roughly the same amount of available pages.
346 		 */
347 		if (i == num_nodes + node_start - 1)
348 			end = max_addr;
349 		else
350 			while (end - *addr - e820_hole_size(*addr, end) <
351 			       size) {
352 				end += FAKE_NODE_MIN_SIZE;
353 				if (end > max_addr) {
354 					end = max_addr;
355 					break;
356 				}
357 			}
358 		if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0)
359 			break;
360 	}
361 	return i - node_start + 1;
362 }
363 
364 /*
365  * Splits the remaining system RAM into chunks of size.  The remaining memory is
366  * always assigned to a final node and can be asymmetric.  Returns the number of
367  * nodes split.
368  */
369 static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr,
370 				      u64 max_addr, int node_start, u64 size)
371 {
372 	int i = node_start;
373 	size = (size << 20) & FAKE_NODE_MIN_HASH_MASK;
374 	while (!setup_node_range(i++, nodes, addr, size, max_addr))
375 		;
376 	return i - node_start;
377 }
378 
379 /*
380  * Sets up the system RAM area from start_pfn to end_pfn according to the
381  * numa=fake command-line option.
382  */
383 static int __init numa_emulation(unsigned long start_pfn, unsigned long end_pfn)
384 {
385 	struct bootnode nodes[MAX_NUMNODES];
386 	u64 size, addr = start_pfn << PAGE_SHIFT;
387 	u64 max_addr = end_pfn << PAGE_SHIFT;
388 	int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i;
389 
390 	memset(&nodes, 0, sizeof(nodes));
391 	/*
392 	 * If the numa=fake command-line is just a single number N, split the
393 	 * system RAM into N fake nodes.
394 	 */
395 	if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) {
396 		long n = simple_strtol(cmdline, NULL, 0);
397 
398 		num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n);
399 		if (num_nodes < 0)
400 			return num_nodes;
401 		goto out;
402 	}
403 
404 	/* Parse the command line. */
405 	for (coeff_flag = 0; ; cmdline++) {
406 		if (*cmdline && isdigit(*cmdline)) {
407 			num = num * 10 + *cmdline - '0';
408 			continue;
409 		}
410 		if (*cmdline == '*') {
411 			if (num > 0)
412 				coeff = num;
413 			coeff_flag = 1;
414 		}
415 		if (!*cmdline || *cmdline == ',') {
416 			if (!coeff_flag)
417 				coeff = 1;
418 			/*
419 			 * Round down to the nearest FAKE_NODE_MIN_SIZE.
420 			 * Command-line coefficients are in megabytes.
421 			 */
422 			size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK;
423 			if (size)
424 				for (i = 0; i < coeff; i++, num_nodes++)
425 					if (setup_node_range(num_nodes, nodes,
426 						&addr, size, max_addr) < 0)
427 						goto done;
428 			if (!*cmdline)
429 				break;
430 			coeff_flag = 0;
431 			coeff = -1;
432 		}
433 		num = 0;
434 	}
435 done:
436 	if (!num_nodes)
437 		return -1;
438 	/* Fill remainder of system RAM, if appropriate. */
439 	if (addr < max_addr) {
440 		if (coeff_flag && coeff < 0) {
441 			/* Split remaining nodes into num-sized chunks */
442 			num_nodes += split_nodes_by_size(nodes, &addr, max_addr,
443 							 num_nodes, num);
444 			goto out;
445 		}
446 		switch (*(cmdline - 1)) {
447 		case '*':
448 			/* Split remaining nodes into coeff chunks */
449 			if (coeff <= 0)
450 				break;
451 			num_nodes += split_nodes_equally(nodes, &addr, max_addr,
452 							 num_nodes, coeff);
453 			break;
454 		case ',':
455 			/* Do not allocate remaining system RAM */
456 			break;
457 		default:
458 			/* Give one final node */
459 			setup_node_range(num_nodes, nodes, &addr,
460 					 max_addr - addr, max_addr);
461 			num_nodes++;
462 		}
463 	}
464 out:
465 	memnode_shift = compute_hash_shift(nodes, num_nodes);
466 	if (memnode_shift < 0) {
467 		memnode_shift = 0;
468 		printk(KERN_ERR "No NUMA hash function found.  NUMA emulation "
469 		       "disabled.\n");
470 		return -1;
471 	}
472 
473 	/*
474 	 * We need to vacate all active ranges that may have been registered by
475 	 * SRAT and set acpi_numa to -1 so that srat_disabled() always returns
476 	 * true.  NUMA emulation has succeeded so we will not scan ACPI nodes.
477 	 */
478 	remove_all_active_ranges();
479 #ifdef CONFIG_ACPI_NUMA
480 	acpi_numa = -1;
481 #endif
482 	for_each_node_mask(i, node_possible_map) {
483 		e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT,
484 						nodes[i].end >> PAGE_SHIFT);
485 		setup_node_bootmem(i, nodes[i].start, nodes[i].end);
486 	}
487 	acpi_fake_nodes(nodes, num_nodes);
488 	numa_init_array();
489 	return 0;
490 }
491 #endif /* CONFIG_NUMA_EMU */
492 
493 void __init numa_initmem_init(unsigned long start_pfn, unsigned long end_pfn)
494 {
495 	int i;
496 
497 	nodes_clear(node_possible_map);
498 	nodes_clear(node_online_map);
499 
500 #ifdef CONFIG_NUMA_EMU
501 	if (cmdline && !numa_emulation(start_pfn, end_pfn))
502 		return;
503 	nodes_clear(node_possible_map);
504 	nodes_clear(node_online_map);
505 #endif
506 
507 #ifdef CONFIG_ACPI_NUMA
508 	if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT,
509 					  end_pfn << PAGE_SHIFT))
510 		return;
511 	nodes_clear(node_possible_map);
512 	nodes_clear(node_online_map);
513 #endif
514 
515 #ifdef CONFIG_K8_NUMA
516 	if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT,
517 					end_pfn<<PAGE_SHIFT))
518 		return;
519 	nodes_clear(node_possible_map);
520 	nodes_clear(node_online_map);
521 #endif
522 	printk(KERN_INFO "%s\n",
523 	       numa_off ? "NUMA turned off" : "No NUMA configuration found");
524 
525 	printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
526 	       start_pfn << PAGE_SHIFT,
527 	       end_pfn << PAGE_SHIFT);
528 	/* setup dummy node covering all memory */
529 	memnode_shift = 63;
530 	memnodemap = memnode.embedded_map;
531 	memnodemap[0] = 0;
532 	node_set_online(0);
533 	node_set(0, node_possible_map);
534 	for (i = 0; i < NR_CPUS; i++)
535 		numa_set_node(i, 0);
536 	/* cpumask_of_cpu() may not be available during early startup */
537 	memset(&node_to_cpumask_map[0], 0, sizeof(node_to_cpumask_map[0]));
538 	cpu_set(0, node_to_cpumask_map[0]);
539 	e820_register_active_regions(0, start_pfn, end_pfn);
540 	setup_node_bootmem(0, start_pfn << PAGE_SHIFT, end_pfn << PAGE_SHIFT);
541 }
542 
543 __cpuinit void numa_add_cpu(int cpu)
544 {
545 	set_bit(cpu,
546 		(unsigned long *)&node_to_cpumask_map[early_cpu_to_node(cpu)]);
547 }
548 
549 void __cpuinit numa_set_node(int cpu, int node)
550 {
551 	int *cpu_to_node_map = x86_cpu_to_node_map_early_ptr;
552 
553 	if(cpu_to_node_map)
554 		cpu_to_node_map[cpu] = node;
555 	else if(per_cpu_offset(cpu))
556 		per_cpu(x86_cpu_to_node_map, cpu) = node;
557 	else
558 		Dprintk(KERN_INFO "Setting node for non-present cpu %d\n", cpu);
559 }
560 
561 unsigned long __init numa_free_all_bootmem(void)
562 {
563 	unsigned long pages = 0;
564 	int i;
565 
566 	for_each_online_node(i)
567 		pages += free_all_bootmem_node(NODE_DATA(i));
568 
569 	return pages;
570 }
571 
572 void __init paging_init(void)
573 {
574 	unsigned long max_zone_pfns[MAX_NR_ZONES];
575 
576 	memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
577 	max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
578 	max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
579 	max_zone_pfns[ZONE_NORMAL] = end_pfn;
580 
581 	sparse_memory_present_with_active_regions(MAX_NUMNODES);
582 	sparse_init();
583 
584 	free_area_init_nodes(max_zone_pfns);
585 }
586 
587 static __init int numa_setup(char *opt)
588 {
589 	if (!opt)
590 		return -EINVAL;
591 	if (!strncmp(opt, "off", 3))
592 		numa_off = 1;
593 #ifdef CONFIG_NUMA_EMU
594 	if (!strncmp(opt, "fake=", 5))
595 		cmdline = opt + 5;
596 #endif
597 #ifdef CONFIG_ACPI_NUMA
598 	if (!strncmp(opt, "noacpi", 6))
599 		acpi_numa = -1;
600 	if (!strncmp(opt, "hotadd=", 7))
601 		hotadd_percent = simple_strtoul(opt+7, NULL, 10);
602 #endif
603 	return 0;
604 }
605 early_param("numa", numa_setup);
606 
607 /*
608  * Setup early cpu_to_node.
609  *
610  * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
611  * and apicid_to_node[] tables have valid entries for a CPU.
612  * This means we skip cpu_to_node[] initialisation for NUMA
613  * emulation and faking node case (when running a kernel compiled
614  * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
615  * is already initialized in a round robin manner at numa_init_array,
616  * prior to this call, and this initialization is good enough
617  * for the fake NUMA cases.
618  */
619 void __init init_cpu_to_node(void)
620 {
621 	int i;
622 
623 	for (i = 0; i < NR_CPUS; i++) {
624 		int node;
625 		u16 apicid = x86_cpu_to_apicid_init[i];
626 
627 		if (apicid == BAD_APICID)
628 			continue;
629 		node = apicid_to_node[apicid];
630 		if (node == NUMA_NO_NODE)
631 			continue;
632 		if (!node_online(node))
633 			continue;
634 		numa_set_node(i, node);
635 	}
636 }
637 
638 
639