]>
Commit | Line | Data |
---|---|---|
71ee73e7 | 1 | /* Common code for 32 and 64-bit NUMA */ |
a4106eae TH |
2 | #include <linux/kernel.h> |
3 | #include <linux/mm.h> | |
4 | #include <linux/string.h> | |
5 | #include <linux/init.h> | |
71ee73e7 | 6 | #include <linux/bootmem.h> |
a4106eae TH |
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> | |
14 | ||
15 | #include <asm/e820.h> | |
16 | #include <asm/proto.h> | |
17 | #include <asm/dma.h> | |
90321602 | 18 | #include <asm/acpi.h> |
a4106eae TH |
19 | #include <asm/amd_nb.h> |
20 | ||
21 | #include "numa_internal.h" | |
90321602 JB |
22 | |
23 | int __initdata numa_off; | |
e6df595b | 24 | nodemask_t numa_nodes_parsed __initdata; |
90321602 | 25 | |
a4106eae TH |
26 | struct pglist_data *node_data[MAX_NUMNODES] __read_mostly; |
27 | EXPORT_SYMBOL(node_data); | |
28 | ||
29 | static struct numa_meminfo numa_meminfo | |
30 | #ifndef CONFIG_MEMORY_HOTPLUG | |
31 | __initdata | |
32 | #endif | |
33 | ; | |
34 | ||
35 | static int numa_distance_cnt; | |
36 | static u8 *numa_distance; | |
a4106eae | 37 | |
90321602 JB |
38 | static __init int numa_setup(char *opt) |
39 | { | |
40 | if (!opt) | |
41 | return -EINVAL; | |
42 | if (!strncmp(opt, "off", 3)) | |
43 | numa_off = 1; | |
44 | #ifdef CONFIG_NUMA_EMU | |
45 | if (!strncmp(opt, "fake=", 5)) | |
46 | numa_emu_cmdline(opt + 5); | |
47 | #endif | |
48 | #ifdef CONFIG_ACPI_NUMA | |
49 | if (!strncmp(opt, "noacpi", 6)) | |
50 | acpi_numa = -1; | |
51 | #endif | |
52 | return 0; | |
53 | } | |
54 | early_param("numa", numa_setup); | |
71ee73e7 | 55 | |
71ee73e7 | 56 | /* |
bbc9e2f4 | 57 | * apicid, cpu, node mappings |
71ee73e7 | 58 | */ |
c4c60524 | 59 | s16 __apicid_to_node[MAX_LOCAL_APIC] = { |
bbc9e2f4 TH |
60 | [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE |
61 | }; | |
62 | ||
148f9bb8 | 63 | int numa_cpu_node(int cpu) |
6bd26273 TH |
64 | { |
65 | int apicid = early_per_cpu(x86_cpu_to_apicid, cpu); | |
66 | ||
67 | if (apicid != BAD_APICID) | |
68 | return __apicid_to_node[apicid]; | |
69 | return NUMA_NO_NODE; | |
70 | } | |
71 | ||
c032ef60 | 72 | cpumask_var_t node_to_cpumask_map[MAX_NUMNODES]; |
71ee73e7 RR |
73 | EXPORT_SYMBOL(node_to_cpumask_map); |
74 | ||
645a7919 TH |
75 | /* |
76 | * Map cpu index to node index | |
77 | */ | |
645a7919 | 78 | DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE); |
645a7919 TH |
79 | EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map); |
80 | ||
e13fe869 | 81 | void numa_set_node(int cpu, int node) |
645a7919 TH |
82 | { |
83 | int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map); | |
84 | ||
85 | /* early setting, no percpu area yet */ | |
86 | if (cpu_to_node_map) { | |
87 | cpu_to_node_map[cpu] = node; | |
88 | return; | |
89 | } | |
90 | ||
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); | |
94 | dump_stack(); | |
95 | return; | |
96 | } | |
97 | #endif | |
98 | per_cpu(x86_cpu_to_node_map, cpu) = node; | |
99 | ||
942670d0 | 100 | set_cpu_numa_node(cpu, node); |
645a7919 TH |
101 | } |
102 | ||
e13fe869 | 103 | void numa_clear_node(int cpu) |
645a7919 TH |
104 | { |
105 | numa_set_node(cpu, NUMA_NO_NODE); | |
106 | } | |
107 | ||
71ee73e7 RR |
108 | /* |
109 | * Allocate node_to_cpumask_map based on number of available nodes | |
110 | * Requires node_possible_map to be valid. | |
111 | * | |
9512938b | 112 | * Note: cpumask_of_node() is not valid until after this is done. |
71ee73e7 RR |
113 | * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.) |
114 | */ | |
115 | void __init setup_node_to_cpumask_map(void) | |
116 | { | |
d2ad351e | 117 | unsigned int node; |
71ee73e7 RR |
118 | |
119 | /* setup nr_node_ids if not done yet */ | |
d2ad351e CS |
120 | if (nr_node_ids == MAX_NUMNODES) |
121 | setup_nr_node_ids(); | |
71ee73e7 RR |
122 | |
123 | /* allocate the map */ | |
c032ef60 RR |
124 | for (node = 0; node < nr_node_ids; node++) |
125 | alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]); | |
71ee73e7 | 126 | |
c032ef60 RR |
127 | /* cpumask_of_node() will now work */ |
128 | pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids); | |
71ee73e7 RR |
129 | } |
130 | ||
a4106eae TH |
131 | static int __init numa_add_memblk_to(int nid, u64 start, u64 end, |
132 | struct numa_meminfo *mi) | |
133 | { | |
134 | /* ignore zero length blks */ | |
135 | if (start == end) | |
136 | return 0; | |
137 | ||
138 | /* whine about and ignore invalid blks */ | |
139 | if (start > end || nid < 0 || nid >= MAX_NUMNODES) { | |
365811d6 BH |
140 | pr_warning("NUMA: Warning: invalid memblk node %d [mem %#010Lx-%#010Lx]\n", |
141 | nid, start, end - 1); | |
a4106eae TH |
142 | return 0; |
143 | } | |
144 | ||
145 | if (mi->nr_blks >= NR_NODE_MEMBLKS) { | |
146 | pr_err("NUMA: too many memblk ranges\n"); | |
147 | return -EINVAL; | |
148 | } | |
149 | ||
150 | mi->blk[mi->nr_blks].start = start; | |
151 | mi->blk[mi->nr_blks].end = end; | |
152 | mi->blk[mi->nr_blks].nid = nid; | |
153 | mi->nr_blks++; | |
154 | return 0; | |
155 | } | |
156 | ||
157 | /** | |
158 | * numa_remove_memblk_from - Remove one numa_memblk from a numa_meminfo | |
159 | * @idx: Index of memblk to remove | |
160 | * @mi: numa_meminfo to remove memblk from | |
161 | * | |
162 | * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and | |
163 | * decrementing @mi->nr_blks. | |
164 | */ | |
165 | void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi) | |
166 | { | |
167 | mi->nr_blks--; | |
168 | memmove(&mi->blk[idx], &mi->blk[idx + 1], | |
169 | (mi->nr_blks - idx) * sizeof(mi->blk[0])); | |
170 | } | |
171 | ||
172 | /** | |
173 | * numa_add_memblk - Add one numa_memblk to numa_meminfo | |
174 | * @nid: NUMA node ID of the new memblk | |
175 | * @start: Start address of the new memblk | |
176 | * @end: End address of the new memblk | |
177 | * | |
178 | * Add a new memblk to the default numa_meminfo. | |
179 | * | |
180 | * RETURNS: | |
181 | * 0 on success, -errno on failure. | |
182 | */ | |
183 | int __init numa_add_memblk(int nid, u64 start, u64 end) | |
184 | { | |
185 | return numa_add_memblk_to(nid, start, end, &numa_meminfo); | |
186 | } | |
187 | ||
a56bca80 YL |
188 | /* Initialize NODE_DATA for a node on the local memory */ |
189 | static void __init setup_node_data(int nid, u64 start, u64 end) | |
a4106eae | 190 | { |
a4106eae | 191 | const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE); |
38f3e1ca | 192 | u64 nd_pa; |
7888e96b | 193 | void *nd; |
a4106eae TH |
194 | int tnid; |
195 | ||
196 | /* | |
197 | * Don't confuse VM with a node that doesn't have the | |
198 | * minimum amount of memory: | |
199 | */ | |
200 | if (end && (end - start) < NODE_MIN_SIZE) | |
201 | return; | |
202 | ||
203 | start = roundup(start, ZONE_ALIGN); | |
204 | ||
365811d6 BH |
205 | printk(KERN_INFO "Initmem setup node %d [mem %#010Lx-%#010Lx]\n", |
206 | nid, start, end - 1); | |
a4106eae TH |
207 | |
208 | /* | |
07f4207a PA |
209 | * Allocate node data. Try node-local memory and then any node. |
210 | * Never allocate in DMA zone. | |
a4106eae | 211 | */ |
20e6926d | 212 | nd_pa = memblock_alloc_nid(nd_size, SMP_CACHE_BYTES, nid); |
07f4207a | 213 | if (!nd_pa) { |
f3d815cb LZ |
214 | nd_pa = __memblock_alloc_base(nd_size, SMP_CACHE_BYTES, |
215 | MEMBLOCK_ALLOC_ACCESSIBLE); | |
216 | if (!nd_pa) { | |
217 | pr_err("Cannot find %zu bytes in node %d\n", | |
218 | nd_size, nid); | |
219 | return; | |
220 | } | |
a4106eae | 221 | } |
07f4207a | 222 | nd = __va(nd_pa); |
a4106eae TH |
223 | |
224 | /* report and initialize */ | |
07f4207a PA |
225 | printk(KERN_INFO " NODE_DATA [mem %#010Lx-%#010Lx]\n", |
226 | nd_pa, nd_pa + nd_size - 1); | |
a4106eae | 227 | tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT); |
07f4207a | 228 | if (tnid != nid) |
a4106eae TH |
229 | printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid); |
230 | ||
7888e96b | 231 | node_data[nid] = nd; |
a4106eae TH |
232 | memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); |
233 | NODE_DATA(nid)->node_id = nid; | |
234 | NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT; | |
235 | NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT; | |
236 | ||
237 | node_set_online(nid); | |
238 | } | |
239 | ||
240 | /** | |
241 | * numa_cleanup_meminfo - Cleanup a numa_meminfo | |
242 | * @mi: numa_meminfo to clean up | |
243 | * | |
244 | * Sanitize @mi by merging and removing unncessary memblks. Also check for | |
245 | * conflicts and clear unused memblks. | |
246 | * | |
247 | * RETURNS: | |
248 | * 0 on success, -errno on failure. | |
249 | */ | |
250 | int __init numa_cleanup_meminfo(struct numa_meminfo *mi) | |
251 | { | |
252 | const u64 low = 0; | |
38f3e1ca | 253 | const u64 high = PFN_PHYS(max_pfn); |
a4106eae TH |
254 | int i, j, k; |
255 | ||
e5a10c1b | 256 | /* first, trim all entries */ |
a4106eae TH |
257 | for (i = 0; i < mi->nr_blks; i++) { |
258 | struct numa_memblk *bi = &mi->blk[i]; | |
259 | ||
260 | /* make sure all blocks are inside the limits */ | |
261 | bi->start = max(bi->start, low); | |
262 | bi->end = min(bi->end, high); | |
263 | ||
264 | /* and there's no empty block */ | |
e5a10c1b | 265 | if (bi->start >= bi->end) |
a4106eae | 266 | numa_remove_memblk_from(i--, mi); |
e5a10c1b YL |
267 | } |
268 | ||
269 | /* merge neighboring / overlapping entries */ | |
270 | for (i = 0; i < mi->nr_blks; i++) { | |
271 | struct numa_memblk *bi = &mi->blk[i]; | |
a4106eae TH |
272 | |
273 | for (j = i + 1; j < mi->nr_blks; j++) { | |
274 | struct numa_memblk *bj = &mi->blk[j]; | |
38f3e1ca | 275 | u64 start, end; |
a4106eae TH |
276 | |
277 | /* | |
278 | * See whether there are overlapping blocks. Whine | |
279 | * about but allow overlaps of the same nid. They | |
280 | * will be merged below. | |
281 | */ | |
282 | if (bi->end > bj->start && bi->start < bj->end) { | |
283 | if (bi->nid != bj->nid) { | |
365811d6 BH |
284 | pr_err("NUMA: node %d [mem %#010Lx-%#010Lx] overlaps with node %d [mem %#010Lx-%#010Lx]\n", |
285 | bi->nid, bi->start, bi->end - 1, | |
286 | bj->nid, bj->start, bj->end - 1); | |
a4106eae TH |
287 | return -EINVAL; |
288 | } | |
365811d6 BH |
289 | pr_warning("NUMA: Warning: node %d [mem %#010Lx-%#010Lx] overlaps with itself [mem %#010Lx-%#010Lx]\n", |
290 | bi->nid, bi->start, bi->end - 1, | |
291 | bj->start, bj->end - 1); | |
a4106eae TH |
292 | } |
293 | ||
294 | /* | |
295 | * Join together blocks on the same node, holes | |
296 | * between which don't overlap with memory on other | |
297 | * nodes. | |
298 | */ | |
299 | if (bi->nid != bj->nid) | |
300 | continue; | |
e5a10c1b YL |
301 | start = min(bi->start, bj->start); |
302 | end = max(bi->end, bj->end); | |
a4106eae TH |
303 | for (k = 0; k < mi->nr_blks; k++) { |
304 | struct numa_memblk *bk = &mi->blk[k]; | |
305 | ||
306 | if (bi->nid == bk->nid) | |
307 | continue; | |
308 | if (start < bk->end && end > bk->start) | |
309 | break; | |
310 | } | |
311 | if (k < mi->nr_blks) | |
312 | continue; | |
365811d6 BH |
313 | printk(KERN_INFO "NUMA: Node %d [mem %#010Lx-%#010Lx] + [mem %#010Lx-%#010Lx] -> [mem %#010Lx-%#010Lx]\n", |
314 | bi->nid, bi->start, bi->end - 1, bj->start, | |
315 | bj->end - 1, start, end - 1); | |
a4106eae TH |
316 | bi->start = start; |
317 | bi->end = end; | |
318 | numa_remove_memblk_from(j--, mi); | |
319 | } | |
320 | } | |
321 | ||
e5a10c1b | 322 | /* clear unused ones */ |
a4106eae TH |
323 | for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) { |
324 | mi->blk[i].start = mi->blk[i].end = 0; | |
325 | mi->blk[i].nid = NUMA_NO_NODE; | |
326 | } | |
327 | ||
328 | return 0; | |
329 | } | |
330 | ||
331 | /* | |
332 | * Set nodes, which have memory in @mi, in *@nodemask. | |
333 | */ | |
334 | static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask, | |
335 | const struct numa_meminfo *mi) | |
336 | { | |
337 | int i; | |
338 | ||
339 | for (i = 0; i < ARRAY_SIZE(mi->blk); i++) | |
340 | if (mi->blk[i].start != mi->blk[i].end && | |
341 | mi->blk[i].nid != NUMA_NO_NODE) | |
342 | node_set(mi->blk[i].nid, *nodemask); | |
343 | } | |
344 | ||
345 | /** | |
346 | * numa_reset_distance - Reset NUMA distance table | |
347 | * | |
348 | * The current table is freed. The next numa_set_distance() call will | |
349 | * create a new one. | |
350 | */ | |
351 | void __init numa_reset_distance(void) | |
352 | { | |
353 | size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]); | |
354 | ||
355 | /* numa_distance could be 1LU marking allocation failure, test cnt */ | |
356 | if (numa_distance_cnt) | |
24aa0788 | 357 | memblock_free(__pa(numa_distance), size); |
a4106eae TH |
358 | numa_distance_cnt = 0; |
359 | numa_distance = NULL; /* enable table creation */ | |
360 | } | |
361 | ||
362 | static int __init numa_alloc_distance(void) | |
363 | { | |
364 | nodemask_t nodes_parsed; | |
365 | size_t size; | |
366 | int i, j, cnt = 0; | |
367 | u64 phys; | |
368 | ||
369 | /* size the new table and allocate it */ | |
370 | nodes_parsed = numa_nodes_parsed; | |
371 | numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo); | |
372 | ||
373 | for_each_node_mask(i, nodes_parsed) | |
374 | cnt = i; | |
375 | cnt++; | |
376 | size = cnt * cnt * sizeof(numa_distance[0]); | |
377 | ||
38f3e1ca | 378 | phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped), |
a4106eae | 379 | size, PAGE_SIZE); |
1f5026a7 | 380 | if (!phys) { |
a4106eae TH |
381 | pr_warning("NUMA: Warning: can't allocate distance table!\n"); |
382 | /* don't retry until explicitly reset */ | |
383 | numa_distance = (void *)1LU; | |
384 | return -ENOMEM; | |
385 | } | |
24aa0788 | 386 | memblock_reserve(phys, size); |
a4106eae TH |
387 | |
388 | numa_distance = __va(phys); | |
389 | numa_distance_cnt = cnt; | |
390 | ||
391 | /* fill with the default distances */ | |
392 | for (i = 0; i < cnt; i++) | |
393 | for (j = 0; j < cnt; j++) | |
394 | numa_distance[i * cnt + j] = i == j ? | |
395 | LOCAL_DISTANCE : REMOTE_DISTANCE; | |
396 | printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt); | |
397 | ||
398 | return 0; | |
399 | } | |
400 | ||
401 | /** | |
402 | * numa_set_distance - Set NUMA distance from one NUMA to another | |
403 | * @from: the 'from' node to set distance | |
404 | * @to: the 'to' node to set distance | |
405 | * @distance: NUMA distance | |
406 | * | |
407 | * Set the distance from node @from to @to to @distance. If distance table | |
408 | * doesn't exist, one which is large enough to accommodate all the currently | |
409 | * known nodes will be created. | |
410 | * | |
411 | * If such table cannot be allocated, a warning is printed and further | |
412 | * calls are ignored until the distance table is reset with | |
413 | * numa_reset_distance(). | |
414 | * | |
54eed6cb PH |
415 | * If @from or @to is higher than the highest known node or lower than zero |
416 | * at the time of table creation or @distance doesn't make sense, the call | |
417 | * is ignored. | |
a4106eae TH |
418 | * This is to allow simplification of specific NUMA config implementations. |
419 | */ | |
420 | void __init numa_set_distance(int from, int to, int distance) | |
421 | { | |
422 | if (!numa_distance && numa_alloc_distance() < 0) | |
423 | return; | |
424 | ||
54eed6cb PH |
425 | if (from >= numa_distance_cnt || to >= numa_distance_cnt || |
426 | from < 0 || to < 0) { | |
427 | pr_warn_once("NUMA: Warning: node ids are out of bound, from=%d to=%d distance=%d\n", | |
a4106eae TH |
428 | from, to, distance); |
429 | return; | |
430 | } | |
431 | ||
432 | if ((u8)distance != distance || | |
433 | (from == to && distance != LOCAL_DISTANCE)) { | |
434 | pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n", | |
435 | from, to, distance); | |
436 | return; | |
437 | } | |
438 | ||
439 | numa_distance[from * numa_distance_cnt + to] = distance; | |
440 | } | |
441 | ||
442 | int __node_distance(int from, int to) | |
443 | { | |
444 | if (from >= numa_distance_cnt || to >= numa_distance_cnt) | |
445 | return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE; | |
446 | return numa_distance[from * numa_distance_cnt + to]; | |
447 | } | |
448 | EXPORT_SYMBOL(__node_distance); | |
449 | ||
450 | /* | |
451 | * Sanity check to catch more bad NUMA configurations (they are amazingly | |
452 | * common). Make sure the nodes cover all memory. | |
453 | */ | |
454 | static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi) | |
455 | { | |
38f3e1ca | 456 | u64 numaram, e820ram; |
a4106eae TH |
457 | int i; |
458 | ||
459 | numaram = 0; | |
460 | for (i = 0; i < mi->nr_blks; i++) { | |
38f3e1ca TH |
461 | u64 s = mi->blk[i].start >> PAGE_SHIFT; |
462 | u64 e = mi->blk[i].end >> PAGE_SHIFT; | |
a4106eae TH |
463 | numaram += e - s; |
464 | numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e); | |
38f3e1ca | 465 | if ((s64)numaram < 0) |
a4106eae TH |
466 | numaram = 0; |
467 | } | |
468 | ||
474b881b TH |
469 | e820ram = max_pfn - absent_pages_in_range(0, max_pfn); |
470 | ||
a4106eae | 471 | /* We seem to lose 3 pages somewhere. Allow 1M of slack. */ |
38f3e1ca TH |
472 | if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) { |
473 | printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n", | |
a4106eae TH |
474 | (numaram << PAGE_SHIFT) >> 20, |
475 | (e820ram << PAGE_SHIFT) >> 20); | |
476 | return false; | |
477 | } | |
478 | return true; | |
479 | } | |
480 | ||
481 | static int __init numa_register_memblks(struct numa_meminfo *mi) | |
482 | { | |
1e01979c | 483 | unsigned long uninitialized_var(pfn_align); |
a4106eae TH |
484 | int i, nid; |
485 | ||
486 | /* Account for nodes with cpus and no memory */ | |
487 | node_possible_map = numa_nodes_parsed; | |
488 | numa_nodemask_from_meminfo(&node_possible_map, mi); | |
489 | if (WARN_ON(nodes_empty(node_possible_map))) | |
490 | return -EINVAL; | |
491 | ||
0608f70c TH |
492 | for (i = 0; i < mi->nr_blks; i++) { |
493 | struct numa_memblk *mb = &mi->blk[i]; | |
e7e8de59 TC |
494 | memblock_set_node(mb->start, mb->end - mb->start, |
495 | &memblock.memory, mb->nid); | |
a0acda91 TC |
496 | |
497 | /* | |
498 | * At this time, all memory regions reserved by memblock are | |
499 | * used by the kernel. Set the nid in memblock.reserved will | |
500 | * mark out all the nodes the kernel resides in. | |
501 | */ | |
502 | memblock_set_node(mb->start, mb->end - mb->start, | |
503 | &memblock.reserved, mb->nid); | |
0608f70c | 504 | } |
1e01979c TH |
505 | |
506 | /* | |
507 | * If sections array is gonna be used for pfn -> nid mapping, check | |
508 | * whether its granularity is fine enough. | |
509 | */ | |
510 | #ifdef NODE_NOT_IN_PAGE_FLAGS | |
511 | pfn_align = node_map_pfn_alignment(); | |
512 | if (pfn_align && pfn_align < PAGES_PER_SECTION) { | |
513 | printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n", | |
514 | PFN_PHYS(pfn_align) >> 20, | |
515 | PFN_PHYS(PAGES_PER_SECTION) >> 20); | |
516 | return -EINVAL; | |
517 | } | |
518 | #endif | |
a4106eae TH |
519 | if (!numa_meminfo_cover_memory(mi)) |
520 | return -EINVAL; | |
521 | ||
522 | /* Finally register nodes. */ | |
523 | for_each_node_mask(nid, node_possible_map) { | |
38f3e1ca | 524 | u64 start = PFN_PHYS(max_pfn); |
a4106eae TH |
525 | u64 end = 0; |
526 | ||
527 | for (i = 0; i < mi->nr_blks; i++) { | |
528 | if (nid != mi->blk[i].nid) | |
529 | continue; | |
530 | start = min(mi->blk[i].start, start); | |
531 | end = max(mi->blk[i].end, end); | |
532 | } | |
533 | ||
534 | if (start < end) | |
a56bca80 | 535 | setup_node_data(nid, start, end); |
a4106eae TH |
536 | } |
537 | ||
0608f70c TH |
538 | /* Dump memblock with node info and return. */ |
539 | memblock_dump_all(); | |
a4106eae TH |
540 | return 0; |
541 | } | |
a4106eae | 542 | |
8db78cc4 TH |
543 | /* |
544 | * There are unfortunately some poorly designed mainboards around that | |
545 | * only connect memory to a single CPU. This breaks the 1:1 cpu->node | |
546 | * mapping. To avoid this fill in the mapping for all possible CPUs, | |
547 | * as the number of CPUs is not known yet. We round robin the existing | |
548 | * nodes. | |
549 | */ | |
752d4f37 | 550 | static void __init numa_init_array(void) |
8db78cc4 TH |
551 | { |
552 | int rr, i; | |
553 | ||
554 | rr = first_node(node_online_map); | |
555 | for (i = 0; i < nr_cpu_ids; i++) { | |
556 | if (early_cpu_to_node(i) != NUMA_NO_NODE) | |
557 | continue; | |
558 | numa_set_node(i, rr); | |
559 | rr = next_node(rr, node_online_map); | |
560 | if (rr == MAX_NUMNODES) | |
561 | rr = first_node(node_online_map); | |
562 | } | |
563 | } | |
564 | ||
a0acda91 TC |
565 | static void __init numa_clear_kernel_node_hotplug(void) |
566 | { | |
567 | int i, nid; | |
568 | nodemask_t numa_kernel_nodes; | |
569 | unsigned long start, end; | |
570 | struct memblock_type *type = &memblock.reserved; | |
571 | ||
572 | /* Mark all kernel nodes. */ | |
573 | for (i = 0; i < type->cnt; i++) | |
574 | node_set(type->regions[i].nid, numa_kernel_nodes); | |
575 | ||
576 | /* Clear MEMBLOCK_HOTPLUG flag for memory in kernel nodes. */ | |
577 | for (i = 0; i < numa_meminfo.nr_blks; i++) { | |
578 | nid = numa_meminfo.blk[i].nid; | |
579 | if (!node_isset(nid, numa_kernel_nodes)) | |
580 | continue; | |
581 | ||
582 | start = numa_meminfo.blk[i].start; | |
583 | end = numa_meminfo.blk[i].end; | |
584 | ||
585 | memblock_clear_hotplug(start, end - start); | |
586 | } | |
587 | } | |
588 | ||
a4106eae TH |
589 | static int __init numa_init(int (*init_func)(void)) |
590 | { | |
591 | int i; | |
592 | int ret; | |
593 | ||
594 | for (i = 0; i < MAX_LOCAL_APIC; i++) | |
595 | set_apicid_to_node(i, NUMA_NO_NODE); | |
596 | ||
20e6926d | 597 | nodes_clear(numa_nodes_parsed); |
a4106eae TH |
598 | nodes_clear(node_possible_map); |
599 | nodes_clear(node_online_map); | |
20e6926d | 600 | memset(&numa_meminfo, 0, sizeof(numa_meminfo)); |
e7e8de59 TC |
601 | WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.memory, |
602 | MAX_NUMNODES)); | |
a0acda91 TC |
603 | WARN_ON(memblock_set_node(0, ULLONG_MAX, &memblock.reserved, |
604 | MAX_NUMNODES)); | |
05d1d8cb TC |
605 | /* In case that parsing SRAT failed. */ |
606 | WARN_ON(memblock_clear_hotplug(0, ULLONG_MAX)); | |
a4106eae TH |
607 | numa_reset_distance(); |
608 | ||
609 | ret = init_func(); | |
610 | if (ret < 0) | |
611 | return ret; | |
c5320926 TC |
612 | |
613 | /* | |
614 | * We reset memblock back to the top-down direction | |
615 | * here because if we configured ACPI_NUMA, we have | |
616 | * parsed SRAT in init_func(). It is ok to have the | |
617 | * reset here even if we did't configure ACPI_NUMA | |
618 | * or acpi numa init fails and fallbacks to dummy | |
619 | * numa init. | |
620 | */ | |
621 | memblock_set_bottom_up(false); | |
622 | ||
a4106eae TH |
623 | ret = numa_cleanup_meminfo(&numa_meminfo); |
624 | if (ret < 0) | |
625 | return ret; | |
626 | ||
627 | numa_emulation(&numa_meminfo, numa_distance_cnt); | |
628 | ||
629 | ret = numa_register_memblks(&numa_meminfo); | |
630 | if (ret < 0) | |
631 | return ret; | |
632 | ||
633 | for (i = 0; i < nr_cpu_ids; i++) { | |
634 | int nid = early_cpu_to_node(i); | |
635 | ||
636 | if (nid == NUMA_NO_NODE) | |
637 | continue; | |
638 | if (!node_online(nid)) | |
639 | numa_clear_node(i); | |
640 | } | |
641 | numa_init_array(); | |
a0acda91 TC |
642 | |
643 | /* | |
644 | * At very early time, the kernel have to use some memory such as | |
645 | * loading the kernel image. We cannot prevent this anyway. So any | |
646 | * node the kernel resides in should be un-hotpluggable. | |
647 | * | |
648 | * And when we come here, numa_init() won't fail. | |
649 | */ | |
650 | numa_clear_kernel_node_hotplug(); | |
651 | ||
a4106eae TH |
652 | return 0; |
653 | } | |
654 | ||
655 | /** | |
656 | * dummy_numa_init - Fallback dummy NUMA init | |
657 | * | |
658 | * Used if there's no underlying NUMA architecture, NUMA initialization | |
659 | * fails, or NUMA is disabled on the command line. | |
660 | * | |
661 | * Must online at least one node and add memory blocks that cover all | |
662 | * allowed memory. This function must not fail. | |
663 | */ | |
664 | static int __init dummy_numa_init(void) | |
665 | { | |
666 | printk(KERN_INFO "%s\n", | |
667 | numa_off ? "NUMA turned off" : "No NUMA configuration found"); | |
365811d6 BH |
668 | printk(KERN_INFO "Faking a node at [mem %#018Lx-%#018Lx]\n", |
669 | 0LLU, PFN_PHYS(max_pfn) - 1); | |
a4106eae TH |
670 | |
671 | node_set(0, numa_nodes_parsed); | |
38f3e1ca | 672 | numa_add_memblk(0, 0, PFN_PHYS(max_pfn)); |
a4106eae TH |
673 | |
674 | return 0; | |
675 | } | |
676 | ||
677 | /** | |
678 | * x86_numa_init - Initialize NUMA | |
679 | * | |
680 | * Try each configured NUMA initialization method until one succeeds. The | |
681 | * last fallback is dummy single node config encomapssing whole memory and | |
682 | * never fails. | |
683 | */ | |
684 | void __init x86_numa_init(void) | |
685 | { | |
686 | if (!numa_off) { | |
bd6709a9 TH |
687 | #ifdef CONFIG_X86_NUMAQ |
688 | if (!numa_init(numaq_numa_init)) | |
689 | return; | |
690 | #endif | |
a4106eae TH |
691 | #ifdef CONFIG_ACPI_NUMA |
692 | if (!numa_init(x86_acpi_numa_init)) | |
693 | return; | |
694 | #endif | |
695 | #ifdef CONFIG_AMD_NUMA | |
696 | if (!numa_init(amd_numa_init)) | |
697 | return; | |
698 | #endif | |
699 | } | |
700 | ||
701 | numa_init(dummy_numa_init); | |
702 | } | |
a4106eae | 703 | |
8db78cc4 TH |
704 | static __init int find_near_online_node(int node) |
705 | { | |
706 | int n, val; | |
707 | int min_val = INT_MAX; | |
708 | int best_node = -1; | |
709 | ||
710 | for_each_online_node(n) { | |
711 | val = node_distance(node, n); | |
712 | ||
713 | if (val < min_val) { | |
714 | min_val = val; | |
715 | best_node = n; | |
716 | } | |
717 | } | |
718 | ||
719 | return best_node; | |
720 | } | |
721 | ||
722 | /* | |
723 | * Setup early cpu_to_node. | |
724 | * | |
725 | * Populate cpu_to_node[] only if x86_cpu_to_apicid[], | |
726 | * and apicid_to_node[] tables have valid entries for a CPU. | |
727 | * This means we skip cpu_to_node[] initialisation for NUMA | |
728 | * emulation and faking node case (when running a kernel compiled | |
729 | * for NUMA on a non NUMA box), which is OK as cpu_to_node[] | |
730 | * is already initialized in a round robin manner at numa_init_array, | |
731 | * prior to this call, and this initialization is good enough | |
732 | * for the fake NUMA cases. | |
733 | * | |
734 | * Called before the per_cpu areas are setup. | |
735 | */ | |
736 | void __init init_cpu_to_node(void) | |
737 | { | |
738 | int cpu; | |
739 | u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid); | |
740 | ||
741 | BUG_ON(cpu_to_apicid == NULL); | |
742 | ||
743 | for_each_possible_cpu(cpu) { | |
744 | int node = numa_cpu_node(cpu); | |
745 | ||
746 | if (node == NUMA_NO_NODE) | |
747 | continue; | |
748 | if (!node_online(node)) | |
749 | node = find_near_online_node(node); | |
750 | numa_set_node(cpu, node); | |
751 | } | |
752 | } | |
753 | ||
de2d9445 TH |
754 | #ifndef CONFIG_DEBUG_PER_CPU_MAPS |
755 | ||
756 | # ifndef CONFIG_NUMA_EMU | |
148f9bb8 | 757 | void numa_add_cpu(int cpu) |
de2d9445 TH |
758 | { |
759 | cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); | |
760 | } | |
761 | ||
148f9bb8 | 762 | void numa_remove_cpu(int cpu) |
de2d9445 TH |
763 | { |
764 | cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); | |
765 | } | |
766 | # endif /* !CONFIG_NUMA_EMU */ | |
767 | ||
768 | #else /* !CONFIG_DEBUG_PER_CPU_MAPS */ | |
645a7919 TH |
769 | |
770 | int __cpu_to_node(int cpu) | |
771 | { | |
772 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) { | |
773 | printk(KERN_WARNING | |
774 | "cpu_to_node(%d): usage too early!\n", cpu); | |
775 | dump_stack(); | |
776 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
777 | } | |
778 | return per_cpu(x86_cpu_to_node_map, cpu); | |
779 | } | |
780 | EXPORT_SYMBOL(__cpu_to_node); | |
781 | ||
782 | /* | |
783 | * Same function as cpu_to_node() but used if called before the | |
784 | * per_cpu areas are setup. | |
785 | */ | |
786 | int early_cpu_to_node(int cpu) | |
787 | { | |
788 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) | |
789 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
790 | ||
791 | if (!cpu_possible(cpu)) { | |
792 | printk(KERN_WARNING | |
793 | "early_cpu_to_node(%d): no per_cpu area!\n", cpu); | |
794 | dump_stack(); | |
795 | return NUMA_NO_NODE; | |
796 | } | |
797 | return per_cpu(x86_cpu_to_node_map, cpu); | |
798 | } | |
799 | ||
7a6c6547 | 800 | void debug_cpumask_set_cpu(int cpu, int node, bool enable) |
de2d9445 | 801 | { |
de2d9445 TH |
802 | struct cpumask *mask; |
803 | char buf[64]; | |
804 | ||
14392fd3 DR |
805 | if (node == NUMA_NO_NODE) { |
806 | /* early_cpu_to_node() already emits a warning and trace */ | |
7a6c6547 | 807 | return; |
14392fd3 | 808 | } |
de2d9445 TH |
809 | mask = node_to_cpumask_map[node]; |
810 | if (!mask) { | |
811 | pr_err("node_to_cpumask_map[%i] NULL\n", node); | |
812 | dump_stack(); | |
7a6c6547 | 813 | return; |
de2d9445 TH |
814 | } |
815 | ||
7a6c6547 DR |
816 | if (enable) |
817 | cpumask_set_cpu(cpu, mask); | |
818 | else | |
819 | cpumask_clear_cpu(cpu, mask); | |
820 | ||
de2d9445 TH |
821 | cpulist_scnprintf(buf, sizeof(buf), mask); |
822 | printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n", | |
823 | enable ? "numa_add_cpu" : "numa_remove_cpu", | |
824 | cpu, node, buf); | |
7a6c6547 | 825 | return; |
de2d9445 TH |
826 | } |
827 | ||
828 | # ifndef CONFIG_NUMA_EMU | |
148f9bb8 | 829 | static void numa_set_cpumask(int cpu, bool enable) |
de2d9445 | 830 | { |
7a6c6547 | 831 | debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable); |
de2d9445 TH |
832 | } |
833 | ||
148f9bb8 | 834 | void numa_add_cpu(int cpu) |
de2d9445 | 835 | { |
7a6c6547 | 836 | numa_set_cpumask(cpu, true); |
de2d9445 TH |
837 | } |
838 | ||
148f9bb8 | 839 | void numa_remove_cpu(int cpu) |
de2d9445 | 840 | { |
7a6c6547 | 841 | numa_set_cpumask(cpu, false); |
de2d9445 TH |
842 | } |
843 | # endif /* !CONFIG_NUMA_EMU */ | |
844 | ||
71ee73e7 RR |
845 | /* |
846 | * Returns a pointer to the bitmask of CPUs on Node 'node'. | |
847 | */ | |
73e907de | 848 | const struct cpumask *cpumask_of_node(int node) |
71ee73e7 | 849 | { |
71ee73e7 RR |
850 | if (node >= nr_node_ids) { |
851 | printk(KERN_WARNING | |
852 | "cpumask_of_node(%d): node > nr_node_ids(%d)\n", | |
853 | node, nr_node_ids); | |
854 | dump_stack(); | |
855 | return cpu_none_mask; | |
856 | } | |
c032ef60 RR |
857 | if (node_to_cpumask_map[node] == NULL) { |
858 | printk(KERN_WARNING | |
859 | "cpumask_of_node(%d): no node_to_cpumask_map!\n", | |
860 | node); | |
861 | dump_stack(); | |
862 | return cpu_online_mask; | |
863 | } | |
0b966252 | 864 | return node_to_cpumask_map[node]; |
71ee73e7 RR |
865 | } |
866 | EXPORT_SYMBOL(cpumask_of_node); | |
645a7919 | 867 | |
de2d9445 | 868 | #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */ |
a4106eae | 869 | |
bd6709a9 | 870 | #ifdef CONFIG_MEMORY_HOTPLUG |
a4106eae TH |
871 | int memory_add_physaddr_to_nid(u64 start) |
872 | { | |
873 | struct numa_meminfo *mi = &numa_meminfo; | |
874 | int nid = mi->blk[0].nid; | |
875 | int i; | |
876 | ||
877 | for (i = 0; i < mi->nr_blks; i++) | |
878 | if (mi->blk[i].start <= start && mi->blk[i].end > start) | |
879 | nid = mi->blk[i].nid; | |
880 | return nid; | |
881 | } | |
882 | EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); | |
883 | #endif |