2 * Generic VM initialization for x86-64 NUMA setups.
3 * Copyright 2002,2003 Andi Kleen, SuSE Labs.
5 #include <linux/kernel.h>
7 #include <linux/string.h>
8 #include <linux/init.h>
9 #include <linux/bootmem.h>
10 #include <linux/memblock.h>
11 #include <linux/mmzone.h>
12 #include <linux/ctype.h>
13 #include <linux/module.h>
14 #include <linux/nodemask.h>
15 #include <linux/sched.h>
16 #include <linux/acpi.h>
19 #include <asm/proto.h>
22 #include <asm/amd_nb.h>
24 #include "numa_internal.h"
26 struct pglist_data *node_data[MAX_NUMNODES] __read_mostly;
27 EXPORT_SYMBOL(node_data);
29 nodemask_t numa_nodes_parsed __initdata;
31 static struct numa_meminfo numa_meminfo
32 #ifndef CONFIG_MEMORY_HOTPLUG
37 static int numa_distance_cnt;
38 static u8 *numa_distance;
40 static int __init numa_add_memblk_to(int nid, u64 start, u64 end,
41 struct numa_meminfo *mi)
43 /* ignore zero length blks */
47 /* whine about and ignore invalid blks */
48 if (start > end || nid < 0 || nid >= MAX_NUMNODES) {
49 pr_warning("NUMA: Warning: invalid memblk node %d (%Lx-%Lx)\n",
54 if (mi->nr_blks >= NR_NODE_MEMBLKS) {
55 pr_err("NUMA: too many memblk ranges\n");
59 mi->blk[mi->nr_blks].start = start;
60 mi->blk[mi->nr_blks].end = end;
61 mi->blk[mi->nr_blks].nid = nid;
67 * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo
68 * @idx: Index of memblk to remove
69 * @mi: numa_meminfo to remove memblk from
71 * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and
72 * decrementing @mi->nr_blks.
74 void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi)
77 memmove(&mi->blk[idx], &mi->blk[idx + 1],
78 (mi->nr_blks - idx) * sizeof(mi->blk[0]));
82 * numa_add_memblk - Add one numa_memblk to numa_meminfo
83 * @nid: NUMA node ID of the new memblk
84 * @start: Start address of the new memblk
85 * @end: End address of the new memblk
87 * Add a new memblk to the default numa_meminfo.
90 * 0 on success, -errno on failure.
92 int __init numa_add_memblk(int nid, u64 start, u64 end)
94 return numa_add_memblk_to(nid, start, end, &numa_meminfo);
97 /* Initialize bootmem allocator for a node */
99 setup_node_bootmem(int nid, unsigned long start, unsigned long end)
101 const u64 nd_low = (u64)MAX_DMA_PFN << PAGE_SHIFT;
102 const u64 nd_high = (u64)max_pfn_mapped << PAGE_SHIFT;
103 const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE);
108 * Don't confuse VM with a node that doesn't have the
109 * minimum amount of memory:
111 if (end && (end - start) < NODE_MIN_SIZE)
114 start = roundup(start, ZONE_ALIGN);
116 printk(KERN_INFO "Initmem setup node %d %016lx-%016lx\n",
120 * Try to allocate node data on local node and then fall back to
121 * all nodes. Never allocate in DMA zone.
123 nd_pa = memblock_x86_find_in_range_node(nid, nd_low, nd_high,
124 nd_size, SMP_CACHE_BYTES);
125 if (nd_pa == MEMBLOCK_ERROR)
126 nd_pa = memblock_find_in_range(nd_low, nd_high,
127 nd_size, SMP_CACHE_BYTES);
128 if (nd_pa == MEMBLOCK_ERROR) {
129 pr_err("Cannot find %lu bytes in node %d\n", nd_size, nid);
132 memblock_x86_reserve_range(nd_pa, nd_pa + nd_size, "NODE_DATA");
134 /* report and initialize */
135 printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n",
136 nd_pa, nd_pa + nd_size - 1);
137 tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT);
139 printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid);
141 node_data[nid] = __va(nd_pa);
142 memset(NODE_DATA(nid), 0, sizeof(pg_data_t));
143 NODE_DATA(nid)->node_id = nid;
144 NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT;
145 NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT;
147 node_set_online(nid);
151 * numa_cleanup_meminfo - Cleanup a numa_meminfo
152 * @mi: numa_meminfo to clean up
154 * Sanitize @mi by merging and removing unncessary memblks. Also check for
155 * conflicts and clear unused memblks.
158 * 0 on success, -errno on failure.
160 int __init numa_cleanup_meminfo(struct numa_meminfo *mi)
163 const u64 high = (u64)max_pfn << PAGE_SHIFT;
166 for (i = 0; i < mi->nr_blks; i++) {
167 struct numa_memblk *bi = &mi->blk[i];
169 /* make sure all blocks are inside the limits */
170 bi->start = max(bi->start, low);
171 bi->end = min(bi->end, high);
173 /* and there's no empty block */
174 if (bi->start >= bi->end) {
175 numa_remove_memblk_from(i--, mi);
179 for (j = i + 1; j < mi->nr_blks; j++) {
180 struct numa_memblk *bj = &mi->blk[j];
181 unsigned long start, end;
184 * See whether there are overlapping blocks. Whine
185 * about but allow overlaps of the same nid. They
186 * will be merged below.
188 if (bi->end > bj->start && bi->start < bj->end) {
189 if (bi->nid != bj->nid) {
190 pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n",
191 bi->nid, bi->start, bi->end,
192 bj->nid, bj->start, bj->end);
195 pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n",
196 bi->nid, bi->start, bi->end,
201 * Join together blocks on the same node, holes
202 * between which don't overlap with memory on other
205 if (bi->nid != bj->nid)
207 start = max(min(bi->start, bj->start), low);
208 end = min(max(bi->end, bj->end), high);
209 for (k = 0; k < mi->nr_blks; k++) {
210 struct numa_memblk *bk = &mi->blk[k];
212 if (bi->nid == bk->nid)
214 if (start < bk->end && end > bk->start)
219 printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%lx,%lx)\n",
220 bi->nid, bi->start, bi->end, bj->start, bj->end,
224 numa_remove_memblk_from(j--, mi);
228 for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) {
229 mi->blk[i].start = mi->blk[i].end = 0;
230 mi->blk[i].nid = NUMA_NO_NODE;
237 * Set nodes, which have memory in @mi, in *@nodemask.
239 static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask,
240 const struct numa_meminfo *mi)
244 for (i = 0; i < ARRAY_SIZE(mi->blk); i++)
245 if (mi->blk[i].start != mi->blk[i].end &&
246 mi->blk[i].nid != NUMA_NO_NODE)
247 node_set(mi->blk[i].nid, *nodemask);
251 * numa_reset_distance - Reset NUMA distance table
253 * The current table is freed. The next numa_set_distance() call will
256 void __init numa_reset_distance(void)
258 size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]);
260 /* numa_distance could be 1LU marking allocation failure, test cnt */
261 if (numa_distance_cnt)
262 memblock_x86_free_range(__pa(numa_distance),
263 __pa(numa_distance) + size);
264 numa_distance_cnt = 0;
265 numa_distance = NULL; /* enable table creation */
268 static int __init numa_alloc_distance(void)
270 nodemask_t nodes_parsed;
275 /* size the new table and allocate it */
276 nodes_parsed = numa_nodes_parsed;
277 numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo);
279 for_each_node_mask(i, nodes_parsed)
282 size = cnt * cnt * sizeof(numa_distance[0]);
284 phys = memblock_find_in_range(0, (u64)max_pfn_mapped << PAGE_SHIFT,
286 if (phys == MEMBLOCK_ERROR) {
287 pr_warning("NUMA: Warning: can't allocate distance table!\n");
288 /* don't retry until explicitly reset */
289 numa_distance = (void *)1LU;
292 memblock_x86_reserve_range(phys, phys + size, "NUMA DIST");
294 numa_distance = __va(phys);
295 numa_distance_cnt = cnt;
297 /* fill with the default distances */
298 for (i = 0; i < cnt; i++)
299 for (j = 0; j < cnt; j++)
300 numa_distance[i * cnt + j] = i == j ?
301 LOCAL_DISTANCE : REMOTE_DISTANCE;
302 printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt);
308 * numa_set_distance - Set NUMA distance from one NUMA to another
309 * @from: the 'from' node to set distance
310 * @to: the 'to' node to set distance
311 * @distance: NUMA distance
313 * Set the distance from node @from to @to to @distance. If distance table
314 * doesn't exist, one which is large enough to accommodate all the currently
315 * known nodes will be created.
317 * If such table cannot be allocated, a warning is printed and further
318 * calls are ignored until the distance table is reset with
319 * numa_reset_distance().
321 * If @from or @to is higher than the highest known node at the time of
322 * table creation or @distance doesn't make sense, the call is ignored.
323 * This is to allow simplification of specific NUMA config implementations.
325 void __init numa_set_distance(int from, int to, int distance)
327 if (!numa_distance && numa_alloc_distance() < 0)
330 if (from >= numa_distance_cnt || to >= numa_distance_cnt) {
331 printk_once(KERN_DEBUG "NUMA: Debug: distance out of bound, from=%d to=%d distance=%d\n",
336 if ((u8)distance != distance ||
337 (from == to && distance != LOCAL_DISTANCE)) {
338 pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n",
343 numa_distance[from * numa_distance_cnt + to] = distance;
346 int __node_distance(int from, int to)
348 if (from >= numa_distance_cnt || to >= numa_distance_cnt)
349 return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE;
350 return numa_distance[from * numa_distance_cnt + to];
352 EXPORT_SYMBOL(__node_distance);
355 * Sanity check to catch more bad NUMA configurations (they are amazingly
356 * common). Make sure the nodes cover all memory.
358 static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi)
360 unsigned long numaram, e820ram;
364 for (i = 0; i < mi->nr_blks; i++) {
365 unsigned long s = mi->blk[i].start >> PAGE_SHIFT;
366 unsigned long e = mi->blk[i].end >> PAGE_SHIFT;
368 numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e);
369 if ((long)numaram < 0)
373 e820ram = max_pfn - (memblock_x86_hole_size(0,
374 max_pfn << PAGE_SHIFT) >> PAGE_SHIFT);
375 /* We seem to lose 3 pages somewhere. Allow 1M of slack. */
376 if ((long)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) {
377 printk(KERN_ERR "NUMA: nodes only cover %luMB of your %luMB e820 RAM. Not used.\n",
378 (numaram << PAGE_SHIFT) >> 20,
379 (e820ram << PAGE_SHIFT) >> 20);
385 static int __init numa_register_memblks(struct numa_meminfo *mi)
389 /* Account for nodes with cpus and no memory */
390 node_possible_map = numa_nodes_parsed;
391 numa_nodemask_from_meminfo(&node_possible_map, mi);
392 if (WARN_ON(nodes_empty(node_possible_map)))
395 for (i = 0; i < mi->nr_blks; i++)
396 memblock_x86_register_active_regions(mi->blk[i].nid,
397 mi->blk[i].start >> PAGE_SHIFT,
398 mi->blk[i].end >> PAGE_SHIFT);
400 /* for out of order entries */
402 if (!numa_meminfo_cover_memory(mi))
405 /* Finally register nodes. */
406 for_each_node_mask(nid, node_possible_map) {
407 u64 start = (u64)max_pfn << PAGE_SHIFT;
410 for (i = 0; i < mi->nr_blks; i++) {
411 if (nid != mi->blk[i].nid)
413 start = min(mi->blk[i].start, start);
414 end = max(mi->blk[i].end, end);
418 setup_node_bootmem(nid, start, end);
425 * dummy_numma_init - Fallback dummy NUMA init
427 * Used if there's no underlying NUMA architecture, NUMA initialization
428 * fails, or NUMA is disabled on the command line.
430 * Must online at least one node and add memory blocks that cover all
431 * allowed memory. This function must not fail.
433 static int __init dummy_numa_init(void)
435 printk(KERN_INFO "%s\n",
436 numa_off ? "NUMA turned off" : "No NUMA configuration found");
437 printk(KERN_INFO "Faking a node at %016lx-%016lx\n",
438 0LU, max_pfn << PAGE_SHIFT);
440 node_set(0, numa_nodes_parsed);
441 numa_add_memblk(0, 0, (u64)max_pfn << PAGE_SHIFT);
446 static int __init numa_init(int (*init_func)(void))
451 for (i = 0; i < MAX_LOCAL_APIC; i++)
452 set_apicid_to_node(i, NUMA_NO_NODE);
454 nodes_clear(numa_nodes_parsed);
455 nodes_clear(node_possible_map);
456 nodes_clear(node_online_map);
457 memset(&numa_meminfo, 0, sizeof(numa_meminfo));
458 remove_all_active_ranges();
459 numa_reset_distance();
464 ret = numa_cleanup_meminfo(&numa_meminfo);
468 numa_emulation(&numa_meminfo, numa_distance_cnt);
470 ret = numa_register_memblks(&numa_meminfo);
474 for (i = 0; i < nr_cpu_ids; i++) {
475 int nid = early_cpu_to_node(i);
477 if (nid == NUMA_NO_NODE)
479 if (!node_online(nid))
486 void __init initmem_init(void)
489 #ifdef CONFIG_ACPI_NUMA
490 if (!numa_init(x86_acpi_numa_init))
493 #ifdef CONFIG_AMD_NUMA
494 if (!numa_init(amd_numa_init))
499 numa_init(dummy_numa_init);
502 unsigned long __init numa_free_all_bootmem(void)
504 unsigned long pages = 0;
507 for_each_online_node(i)
508 pages += free_all_bootmem_node(NODE_DATA(i));
510 pages += free_all_memory_core_early(MAX_NUMNODES);
515 #ifdef CONFIG_MEMORY_HOTPLUG
516 int memory_add_physaddr_to_nid(u64 start)
518 struct numa_meminfo *mi = &numa_meminfo;
519 int nid = mi->blk[0].nid;
522 for (i = 0; i < mi->nr_blks; i++)
523 if (mi->blk[i].start <= start && mi->blk[i].end > start)
524 nid = mi->blk[i].nid;
527 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);