1 /* Common code for 32 and 64-bit NUMA */
2 #include <linux/kernel.h>
4 #include <linux/string.h>
5 #include <linux/init.h>
6 #include <linux/bootmem.h>
7 #include <linux/memblock.h>
8 #include <linux/mmzone.h>
9 #include <linux/ctype.h>
10 #include <linux/module.h>
11 #include <linux/nodemask.h>
12 #include <linux/sched.h>
13 #include <linux/topology.h>
16 #include <asm/proto.h>
19 #include <asm/amd_nb.h>
21 #include "numa_internal.h"
23 int __initdata numa_off;
24 nodemask_t numa_nodes_parsed __initdata;
26 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
27 EXPORT_SYMBOL(node_data);
29 static struct numa_meminfo numa_meminfo
30 #ifndef CONFIG_MEMORY_HOTPLUG
35 static int numa_distance_cnt;
36 static u8 *numa_distance;
38 static __init int numa_setup(char *opt)
42 if (!strncmp(opt, "off", 3))
44 #ifdef CONFIG_NUMA_EMU
45 if (!strncmp(opt, "fake=", 5))
46 numa_emu_cmdline(opt + 5);
48 #ifdef CONFIG_ACPI_NUMA
49 if (!strncmp(opt, "noacpi", 6))
54 early_param("numa", numa_setup);
57 * apicid, cpu, node mappings
59 s16 __apicid_to_node[MAX_LOCAL_APIC] = {
60 [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE
63 int __cpuinit numa_cpu_node(int cpu)
65 int apicid = early_per_cpu(x86_cpu_to_apicid, cpu);
67 if (apicid != BAD_APICID)
68 return __apicid_to_node[apicid];
72 cpumask_var_t node_to_cpumask_map[MAX_NUMNODES];
73 EXPORT_SYMBOL(node_to_cpumask_map);
76 * Map cpu index to node index
78 DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE);
79 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map);
81 void numa_set_node(int cpu, int node)
83 int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map);
85 /* early setting, no percpu area yet */
86 if (cpu_to_node_map) {
87 cpu_to_node_map[cpu] = node;
91 #ifdef CONFIG_DEBUG_PER_CPU_MAPS
92 if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) {
93 printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu);
98 per_cpu(x86_cpu_to_node_map, cpu) = node;
100 if (node != NUMA_NO_NODE)
101 set_cpu_numa_node(cpu, node);
104 void numa_clear_node(int cpu)
106 numa_set_node(cpu, NUMA_NO_NODE);
110 * Allocate node_to_cpumask_map based on number of available nodes
111 * Requires node_possible_map to be valid.
113 * Note: cpumask_of_node() is not valid until after this is done.
114 * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.)
116 void __init setup_node_to_cpumask_map(void)
118 unsigned int node, num = 0;
120 /* setup nr_node_ids if not done yet */
121 if (nr_node_ids == MAX_NUMNODES) {
122 for_each_node_mask(node, node_possible_map)
124 nr_node_ids = num + 1;
127 /* allocate the map */
128 for (node = 0; node < nr_node_ids; node++)
129 alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]);
131 /* cpumask_of_node() will now work */
132 pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids);
135 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
136 struct numa_meminfo *mi)
138 /* ignore zero length blks */
142 /* whine about and ignore invalid blks */
143 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
144 pr_warning("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n",
145 nid, start, end - 1);
149 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
150 pr_err("NUMA: too many memblk ranges\n");
154 mi->blk[mi->nr_blks].start = start;
155 mi->blk[mi->nr_blks].end = end;
156 mi->blk[mi->nr_blks].nid = nid;
162 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
163 * @idx: Index of memblk to remove
164 * @mi: numa_meminfo to remove memblk from
166 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
167 * decrementing @mi->nr_blks.
169 void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
172 memmove(&mi->blk[idx], &mi->blk[idx + 1],
173 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
177 * numa_add_memblk - Add one numa_memblk to numa_meminfo
178 * @nid: NUMA node ID of the new memblk
179 * @start: Start address of the new memblk
180 * @end: End address of the new memblk
182 * Add a new memblk to the default numa_meminfo.
185 * 0 on success, -errno on failure.
187 int __init numa_add_memblk(int nid, u64 start, u64 end)
189 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
192 /* Initialize NODE_DATA for a node on the local memory */
193 static void __init setup_node_data(int nid, u64 start, u64 end)
195 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
201 * Don't confuse VM with a node that doesn't have the
202 * minimum amount of memory:
204 if (end && (end - start) < NODE_MIN_SIZE)
207 start = roundup(start, ZONE_ALIGN);
209 printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n",
210 nid, start, end - 1);
213 * Allocate node data. Try node-local memory and then any node.
214 * Never allocate in DMA zone.
216 nd_pa = memblock_alloc_try_nid(nd_size, SMP_CACHE_BYTES, nid);
218 pr_err("Cannot find %zu bytes in any node\n", nd_size);
223 /* report and initialize */
224 printk(KERN_INFO " NODE_DATA [mem %#010Lx-%#010Lx]\n",
225 nd_pa, nd_pa + nd_size - 1);
226 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
228 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
231 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
232 NODE_DATA(nid)->node_id = nid;
233 NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
234 NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
236 node_set_online(nid);
240 * numa_cleanup_meminfo - Cleanup a numa_meminfo
241 * @mi: numa_meminfo to clean up
243 * Sanitize @mi by merging and removing unncessary memblks. Also check for
244 * conflicts and clear unused memblks.
247 * 0 on success, -errno on failure.
249 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
252 const u64 high = PFN_PHYS(max_pfn);
255 /* first, trim all entries */
256 for (i = 0; i < mi->nr_blks; i++) {
257 struct numa_memblk *bi = &mi->blk[i];
259 /* make sure all blocks are inside the limits */
260 bi->start = max(bi->start, low);
261 bi->end = min(bi->end, high);
263 /* and there's no empty block */
264 if (bi->start >= bi->end)
265 numa_remove_memblk_from(i--, mi);
268 /* merge neighboring / overlapping entries */
269 for (i = 0; i < mi->nr_blks; i++) {
270 struct numa_memblk *bi = &mi->blk[i];
272 for (j = i + 1; j < mi->nr_blks; j++) {
273 struct numa_memblk *bj = &mi->blk[j];
277 * See whether there are overlapping blocks. Whine
278 * about but allow overlaps of the same nid. They
279 * will be merged below.
281 if (bi->end > bj->start && bi->start < bj->end) {
282 if (bi->nid != bj->nid) {
283 pr_err("NUMA: node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n",
284 bi->nid, bi->start, bi->end - 1,
285 bj->nid, bj->start, bj->end - 1);
288 pr_warning("NUMA: Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n",
289 bi->nid, bi->start, bi->end - 1,
290 bj->start, bj->end - 1);
294 * Join together blocks on the same node, holes
295 * between which don't overlap with memory on other
298 if (bi->nid != bj->nid)
300 start = min(bi->start, bj->start);
301 end = max(bi->end, bj->end);
302 for (k = 0; k < mi->nr_blks; k++) {
303 struct numa_memblk *bk = &mi->blk[k];
305 if (bi->nid == bk->nid)
307 if (start < bk->end && end > bk->start)
312 printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n",
313 bi->nid, bi->start, bi->end - 1, bj->start,
314 bj->end - 1, start, end - 1);
317 numa_remove_memblk_from(j--, mi);
321 /* clear unused ones */
322 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
323 mi->blk[i].start = mi->blk[i].end = 0;
324 mi->blk[i].nid = NUMA_NO_NODE;
331 * Set nodes, which have memory in @mi, in *@nodemask.
333 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
334 const struct numa_meminfo *mi)
338 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
339 if (mi->blk[i].start != mi->blk[i].end &&
340 mi->blk[i].nid != NUMA_NO_NODE)
341 node_set(mi->blk[i].nid, *nodemask);
345 * numa_reset_distance - Reset NUMA distance table
347 * The current table is freed. The next numa_set_distance() call will
350 void __init numa_reset_distance(void)
352 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
354 /* numa_distance could be 1LU marking allocation failure, test cnt */
355 if (numa_distance_cnt)
356 memblock_free(__pa(numa_distance), size);
357 numa_distance_cnt = 0;
358 numa_distance = NULL; /* enable table creation */
361 static int __init numa_alloc_distance(void)
363 nodemask_t nodes_parsed;
368 /* size the new table and allocate it */
369 nodes_parsed = numa_nodes_parsed;
370 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
372 for_each_node_mask(i, nodes_parsed)
375 size = cnt * cnt * sizeof(numa_distance[0]);
377 phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped),
380 pr_warning("NUMA: Warning: can't allocate distance table!\n");
381 /* don't retry until explicitly reset */
382 numa_distance = (void *)1LU;
385 memblock_reserve(phys, size);
387 numa_distance = __va(phys);
388 numa_distance_cnt = cnt;
390 /* fill with the default distances */
391 for (i = 0; i < cnt; i++)
392 for (j = 0; j < cnt; j++)
393 numa_distance[i * cnt + j] = i == j ?
394 LOCAL_DISTANCE : REMOTE_DISTANCE;
395 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
401 * numa_set_distance - Set NUMA distance from one NUMA to another
402 * @from: the 'from' node to set distance
403 * @to: the 'to' node to set distance
404 * @distance: NUMA distance
406 * Set the distance from node @from to @to to @distance. If distance table
407 * doesn't exist, one which is large enough to accommodate all the currently
408 * known nodes will be created.
410 * If such table cannot be allocated, a warning is printed and further
411 * calls are ignored until the distance table is reset with
412 * numa_reset_distance().
414 * If @from or @to is higher than the highest known node or lower than zero
415 * at the time of table creation or @distance doesn't make sense, the call
417 * This is to allow simplification of specific NUMA config implementations.
419 void __init numa_set_distance(int from, int to, int distance)
421 if (!numa_distance && numa_alloc_distance() < 0)
424 if (from >= numa_distance_cnt || to >= numa_distance_cnt ||
425 from < 0 || to < 0) {
426 pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n",
431 if ((u8)distance != distance ||
432 (from == to && distance != LOCAL_DISTANCE)) {
433 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
438 numa_distance[from * numa_distance_cnt + to] = distance;
441 int __node_distance(int from, int to)
443 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
444 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
445 return numa_distance[from * numa_distance_cnt + to];
447 EXPORT_SYMBOL(__node_distance);
450 * Sanity check to catch more bad NUMA configurations (they are amazingly
451 * common). Make sure the nodes cover all memory.
453 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
455 u64 numaram, e820ram;
459 for (i = 0; i < mi->nr_blks; i++) {
460 u64 s = mi->blk[i].start >> PAGE_SHIFT;
461 u64 e = mi->blk[i].end >> PAGE_SHIFT;
463 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
464 if ((s64)numaram < 0)
468 e820ram = max_pfn - absent_pages_in_range(0, max_pfn);
470 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
471 if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
472 printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n",
473 (numaram << PAGE_SHIFT) >> 20,
474 (e820ram << PAGE_SHIFT) >> 20);
480 static int __init numa_register_memblks(struct numa_meminfo *mi)
482 unsigned long uninitialized_var(pfn_align);
485 /* Account for nodes with cpus and no memory */
486 node_possible_map = numa_nodes_parsed;
487 numa_nodemask_from_meminfo(&node_possible_map, mi);
488 if (WARN_ON(nodes_empty(node_possible_map)))
491 for (i = 0; i < mi->nr_blks; i++) {
492 struct numa_memblk *mb = &mi->blk[i];
493 memblock_set_node(mb->start, mb->end - mb->start, mb->nid);
497 * If sections array is gonna be used for pfn -> nid mapping, check
498 * whether its granularity is fine enough.
500 #ifdef NODE_NOT_IN_PAGE_FLAGS
501 pfn_align = node_map_pfn_alignment();
502 if (pfn_align && pfn_align < PAGES_PER_SECTION) {
503 printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n",
504 PFN_PHYS(pfn_align) >> 20,
505 PFN_PHYS(PAGES_PER_SECTION) >> 20);
509 if (!numa_meminfo_cover_memory(mi))
512 /* Finally register nodes. */
513 for_each_node_mask(nid, node_possible_map) {
514 u64 start = PFN_PHYS(max_pfn);
517 for (i = 0; i < mi->nr_blks; i++) {
518 if (nid != mi->blk[i].nid)
520 start = min(mi->blk[i].start, start);
521 end = max(mi->blk[i].end, end);
525 setup_node_data(nid, start, end);
528 /* Dump memblock with node info and return. */
534 * There are unfortunately some poorly designed mainboards around that
535 * only connect memory to a single CPU. This breaks the 1:1 cpu->node
536 * mapping. To avoid this fill in the mapping for all possible CPUs,
537 * as the number of CPUs is not known yet. We round robin the existing
540 static void __init numa_init_array(void)
544 rr = first_node(node_online_map);
545 for (i = 0; i < nr_cpu_ids; i++) {
546 if (early_cpu_to_node(i) != NUMA_NO_NODE)
548 numa_set_node(i, rr);
549 rr = next_node(rr, node_online_map);
550 if (rr == MAX_NUMNODES)
551 rr = first_node(node_online_map);
555 static int __init numa_init(int (*init_func)(void))
560 for (i = 0; i < MAX_LOCAL_APIC; i++)
561 set_apicid_to_node(i, NUMA_NO_NODE);
564 * Do not clear numa_nodes_parsed or zero numa_meminfo here, because
565 * SRAT was parsed earlier in early_parse_srat().
567 nodes_clear(node_possible_map);
568 nodes_clear(node_online_map);
569 WARN_ON(memblock_set_node(0, ULLONG_MAX, MAX_NUMNODES));
570 numa_reset_distance();
575 ret = numa_cleanup_meminfo(&numa_meminfo);
579 numa_emulation(&numa_meminfo, numa_distance_cnt);
581 ret = numa_register_memblks(&numa_meminfo);
585 for (i = 0; i < nr_cpu_ids; i++) {
586 int nid = early_cpu_to_node(i);
588 if (nid == NUMA_NO_NODE)
590 if (!node_online(nid))
598 * dummy_numa_init - Fallback dummy NUMA init
600 * Used if there's no underlying NUMA architecture, NUMA initialization
601 * fails, or NUMA is disabled on the command line.
603 * Must online at least one node and add memory blocks that cover all
604 * allowed memory. This function must not fail.
606 static int __init dummy_numa_init(void)
608 printk(KERN_INFO "%s\n",
609 numa_off ? "NUMA turned off" : "No NUMA configuration found");
610 printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n",
611 0LLU, PFN_PHYS(max_pfn) - 1);
613 node_set(0, numa_nodes_parsed);
614 numa_add_memblk(0, 0, PFN_PHYS(max_pfn));
620 * x86_numa_init - Initialize NUMA
622 * Try each configured NUMA initialization method until one succeeds. The
623 * last fallback is dummy single node config encomapssing whole memory and
626 void __init x86_numa_init(void)
629 #ifdef CONFIG_X86_NUMAQ
630 if (!numa_init(numaq_numa_init))
633 #ifdef CONFIG_ACPI_NUMA
634 if (!numa_init(x86_acpi_numa_init))
637 #ifdef CONFIG_AMD_NUMA
638 if (!numa_init(amd_numa_init))
643 numa_init(dummy_numa_init);
646 static __init int find_near_online_node(int node)
649 int min_val = INT_MAX;
652 for_each_online_node(n) {
653 val = node_distance(node, n);
665 * Setup early cpu_to_node.
667 * Populate cpu_to_node[] only if x86_cpu_to_apicid[],
668 * and apicid_to_node[] tables have valid entries for a CPU.
669 * This means we skip cpu_to_node[] initialisation for NUMA
670 * emulation and faking node case (when running a kernel compiled
671 * for NUMA on a non NUMA box), which is OK as cpu_to_node[]
672 * is already initialized in a round robin manner at numa_init_array,
673 * prior to this call, and this initialization is good enough
674 * for the fake NUMA cases.
676 * Called before the per_cpu areas are setup.
678 void __init init_cpu_to_node(void)
681 u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
683 BUG_ON(cpu_to_apicid == NULL);
685 for_each_possible_cpu(cpu) {
686 int node = numa_cpu_node(cpu);
688 if (node == NUMA_NO_NODE)
690 if (!node_online(node))
691 node = find_near_online_node(node);
692 numa_set_node(cpu, node);
696 #ifndef CONFIG_DEBUG_PER_CPU_MAPS
698 # ifndef CONFIG_NUMA_EMU
699 void __cpuinit numa_add_cpu(int cpu)
701 cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
704 void __cpuinit numa_remove_cpu(int cpu)
706 cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]);
708 # endif /* !CONFIG_NUMA_EMU */
710 #else /* !CONFIG_DEBUG_PER_CPU_MAPS */
712 int __cpu_to_node(int cpu)
714 if (early_per_cpu_ptr(x86_cpu_to_node_map)) {
716 "cpu_to_node(%d): usage too early!\n", cpu);
718 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
720 return per_cpu(x86_cpu_to_node_map, cpu);
722 EXPORT_SYMBOL(__cpu_to_node);
725 * Same function as cpu_to_node() but used if called before the
726 * per_cpu areas are setup.
728 int early_cpu_to_node(int cpu)
730 if (early_per_cpu_ptr(x86_cpu_to_node_map))
731 return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu];
733 if (!cpu_possible(cpu)) {
735 "early_cpu_to_node(%d): no per_cpu area!\n", cpu);
739 return per_cpu(x86_cpu_to_node_map, cpu);
742 void debug_cpumask_set_cpu(int cpu, int node, bool enable)
744 struct cpumask *mask;
747 if (node == NUMA_NO_NODE) {
748 /* early_cpu_to_node() already emits a warning and trace */
751 mask = node_to_cpumask_map[node];
753 pr_err("node_to_cpumask_map[%i] NULL\n", node);
759 cpumask_set_cpu(cpu, mask);
761 cpumask_clear_cpu(cpu, mask);
763 cpulist_scnprintf(buf, sizeof(buf), mask);
764 printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n",
765 enable ? "numa_add_cpu" : "numa_remove_cpu",
770 # ifndef CONFIG_NUMA_EMU
771 static void __cpuinit numa_set_cpumask(int cpu, bool enable)
773 debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable);
776 void __cpuinit numa_add_cpu(int cpu)
778 numa_set_cpumask(cpu, true);
781 void __cpuinit numa_remove_cpu(int cpu)
783 numa_set_cpumask(cpu, false);
785 # endif /* !CONFIG_NUMA_EMU */
788 * Returns a pointer to the bitmask of CPUs on Node 'node'.
790 const struct cpumask *cpumask_of_node(int node)
792 if (node >= nr_node_ids) {
794 "cpumask_of_node(%d): node > nr_node_ids(%d)\n",
797 return cpu_none_mask;
799 if (node_to_cpumask_map[node] == NULL) {
801 "cpumask_of_node(%d): no node_to_cpumask_map!\n",
804 return cpu_online_mask;
806 return node_to_cpumask_map[node];
808 EXPORT_SYMBOL(cpumask_of_node);
810 #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */
812 #ifdef CONFIG_MEMORY_HOTPLUG
813 int memory_add_physaddr_to_nid(u64 start)
815 struct numa_meminfo *mi = &numa_meminfo;
816 int nid = mi->blk[0].nid;
819 for (i = 0; i < mi->nr_blks; i++)
820 if (mi->blk[i].start <= start && mi->blk[i].end > start)
821 nid = mi->blk[i].nid;
824 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);