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 | */ |
bbc9e2f4 TH |
59 | s16 __apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = { |
60 | [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE | |
61 | }; | |
62 | ||
6bd26273 TH |
63 | int __cpuinit numa_cpu_node(int cpu) |
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 | ||
81 | void __cpuinit numa_set_node(int cpu, int node) | |
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 | ||
100 | if (node != NUMA_NO_NODE) | |
101 | set_cpu_numa_node(cpu, node); | |
102 | } | |
103 | ||
104 | void __cpuinit numa_clear_node(int cpu) | |
105 | { | |
106 | numa_set_node(cpu, NUMA_NO_NODE); | |
107 | } | |
108 | ||
71ee73e7 RR |
109 | /* |
110 | * Allocate node_to_cpumask_map based on number of available nodes | |
111 | * Requires node_possible_map to be valid. | |
112 | * | |
113 | * Note: node_to_cpumask() is not valid until after this is done. | |
114 | * (Use CONFIG_DEBUG_PER_CPU_MAPS to check this.) | |
115 | */ | |
116 | void __init setup_node_to_cpumask_map(void) | |
117 | { | |
118 | unsigned int node, num = 0; | |
71ee73e7 RR |
119 | |
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) | |
123 | num = node; | |
124 | nr_node_ids = num + 1; | |
125 | } | |
126 | ||
127 | /* allocate the map */ | |
c032ef60 RR |
128 | for (node = 0; node < nr_node_ids; node++) |
129 | alloc_bootmem_cpumask_var(&node_to_cpumask_map[node]); | |
71ee73e7 | 130 | |
c032ef60 RR |
131 | /* cpumask_of_node() will now work */ |
132 | pr_debug("Node to cpumask map for %d nodes\n", nr_node_ids); | |
71ee73e7 RR |
133 | } |
134 | ||
a4106eae TH |
135 | static int __init numa_add_memblk_to(int nid, u64 start, u64 end, |
136 | struct numa_meminfo *mi) | |
137 | { | |
138 | /* ignore zero length blks */ | |
139 | if (start == end) | |
140 | return 0; | |
141 | ||
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 (%Lx-%Lx)\n", | |
145 | nid, start, end); | |
146 | return 0; | |
147 | } | |
148 | ||
149 | if (mi->nr_blks >= NR_NODE_MEMBLKS) { | |
150 | pr_err("NUMA: too many memblk ranges\n"); | |
151 | return -EINVAL; | |
152 | } | |
153 | ||
154 | mi->blk[mi->nr_blks].start = start; | |
155 | mi->blk[mi->nr_blks].end = end; | |
156 | mi->blk[mi->nr_blks].nid = nid; | |
157 | mi->nr_blks++; | |
158 | return 0; | |
159 | } | |
160 | ||
161 | /** | |
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 | |
165 | * | |
166 | * Remove @idx'th numa_memblk from @mi by shifting @mi->blk[] and | |
167 | * decrementing @mi->nr_blks. | |
168 | */ | |
169 | void __init numa_remove_memblk_from(int idx, struct numa_meminfo *mi) | |
170 | { | |
171 | mi->nr_blks--; | |
172 | memmove(&mi->blk[idx], &mi->blk[idx + 1], | |
173 | (mi->nr_blks - idx) * sizeof(mi->blk[0])); | |
174 | } | |
175 | ||
176 | /** | |
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 | |
181 | * | |
182 | * Add a new memblk to the default numa_meminfo. | |
183 | * | |
184 | * RETURNS: | |
185 | * 0 on success, -errno on failure. | |
186 | */ | |
187 | int __init numa_add_memblk(int nid, u64 start, u64 end) | |
188 | { | |
189 | return numa_add_memblk_to(nid, start, end, &numa_meminfo); | |
190 | } | |
191 | ||
a56bca80 YL |
192 | /* Initialize NODE_DATA for a node on the local memory */ |
193 | static void __init setup_node_data(int nid, u64 start, u64 end) | |
a4106eae | 194 | { |
38f3e1ca TH |
195 | const u64 nd_low = PFN_PHYS(MAX_DMA_PFN); |
196 | const u64 nd_high = PFN_PHYS(max_pfn_mapped); | |
a4106eae | 197 | const size_t nd_size = roundup(sizeof(pg_data_t), PAGE_SIZE); |
7888e96b | 198 | bool remapped = false; |
38f3e1ca | 199 | u64 nd_pa; |
7888e96b | 200 | void *nd; |
a4106eae TH |
201 | int tnid; |
202 | ||
203 | /* | |
204 | * Don't confuse VM with a node that doesn't have the | |
205 | * minimum amount of memory: | |
206 | */ | |
207 | if (end && (end - start) < NODE_MIN_SIZE) | |
208 | return; | |
209 | ||
7888e96b TH |
210 | /* initialize remap allocator before aligning to ZONE_ALIGN */ |
211 | init_alloc_remap(nid, start, end); | |
212 | ||
a4106eae TH |
213 | start = roundup(start, ZONE_ALIGN); |
214 | ||
38f3e1ca | 215 | printk(KERN_INFO "Initmem setup node %d %016Lx-%016Lx\n", |
a4106eae TH |
216 | nid, start, end); |
217 | ||
218 | /* | |
7888e96b TH |
219 | * Allocate node data. Try remap allocator first, node-local |
220 | * memory and then any node. Never allocate in DMA zone. | |
a4106eae | 221 | */ |
7888e96b TH |
222 | nd = alloc_remap(nid, nd_size); |
223 | if (nd) { | |
224 | nd_pa = __pa(nd); | |
225 | remapped = true; | |
226 | } else { | |
227 | nd_pa = memblock_x86_find_in_range_node(nid, nd_low, nd_high, | |
a4106eae | 228 | nd_size, SMP_CACHE_BYTES); |
7888e96b TH |
229 | if (nd_pa == MEMBLOCK_ERROR) |
230 | nd_pa = memblock_find_in_range(nd_low, nd_high, | |
231 | nd_size, SMP_CACHE_BYTES); | |
232 | if (nd_pa == MEMBLOCK_ERROR) { | |
233 | pr_err("Cannot find %zu bytes in node %d\n", | |
234 | nd_size, nid); | |
235 | return; | |
236 | } | |
237 | memblock_x86_reserve_range(nd_pa, nd_pa + nd_size, "NODE_DATA"); | |
238 | nd = __va(nd_pa); | |
a4106eae | 239 | } |
a4106eae TH |
240 | |
241 | /* report and initialize */ | |
7888e96b TH |
242 | printk(KERN_INFO " NODE_DATA [%016Lx - %016Lx]%s\n", |
243 | nd_pa, nd_pa + nd_size - 1, remapped ? " (remapped)" : ""); | |
a4106eae | 244 | tnid = early_pfn_to_nid(nd_pa >> PAGE_SHIFT); |
7888e96b | 245 | if (!remapped && tnid != nid) |
a4106eae TH |
246 | printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nid, tnid); |
247 | ||
7888e96b | 248 | node_data[nid] = nd; |
a4106eae TH |
249 | memset(NODE_DATA(nid), 0, sizeof(pg_data_t)); |
250 | NODE_DATA(nid)->node_id = nid; | |
251 | NODE_DATA(nid)->node_start_pfn = start >> PAGE_SHIFT; | |
252 | NODE_DATA(nid)->node_spanned_pages = (end - start) >> PAGE_SHIFT; | |
253 | ||
254 | node_set_online(nid); | |
255 | } | |
256 | ||
257 | /** | |
258 | * numa_cleanup_meminfo - Cleanup a numa_meminfo | |
259 | * @mi: numa_meminfo to clean up | |
260 | * | |
261 | * Sanitize @mi by merging and removing unncessary memblks. Also check for | |
262 | * conflicts and clear unused memblks. | |
263 | * | |
264 | * RETURNS: | |
265 | * 0 on success, -errno on failure. | |
266 | */ | |
267 | int __init numa_cleanup_meminfo(struct numa_meminfo *mi) | |
268 | { | |
269 | const u64 low = 0; | |
38f3e1ca | 270 | const u64 high = PFN_PHYS(max_pfn); |
a4106eae TH |
271 | int i, j, k; |
272 | ||
e5a10c1b | 273 | /* first, trim all entries */ |
a4106eae TH |
274 | for (i = 0; i < mi->nr_blks; i++) { |
275 | struct numa_memblk *bi = &mi->blk[i]; | |
276 | ||
277 | /* make sure all blocks are inside the limits */ | |
278 | bi->start = max(bi->start, low); | |
279 | bi->end = min(bi->end, high); | |
280 | ||
281 | /* and there's no empty block */ | |
e5a10c1b | 282 | if (bi->start >= bi->end) |
a4106eae | 283 | numa_remove_memblk_from(i--, mi); |
e5a10c1b YL |
284 | } |
285 | ||
286 | /* merge neighboring / overlapping entries */ | |
287 | for (i = 0; i < mi->nr_blks; i++) { | |
288 | struct numa_memblk *bi = &mi->blk[i]; | |
a4106eae TH |
289 | |
290 | for (j = i + 1; j < mi->nr_blks; j++) { | |
291 | struct numa_memblk *bj = &mi->blk[j]; | |
38f3e1ca | 292 | u64 start, end; |
a4106eae TH |
293 | |
294 | /* | |
295 | * See whether there are overlapping blocks. Whine | |
296 | * about but allow overlaps of the same nid. They | |
297 | * will be merged below. | |
298 | */ | |
299 | if (bi->end > bj->start && bi->start < bj->end) { | |
300 | if (bi->nid != bj->nid) { | |
301 | pr_err("NUMA: node %d (%Lx-%Lx) overlaps with node %d (%Lx-%Lx)\n", | |
302 | bi->nid, bi->start, bi->end, | |
303 | bj->nid, bj->start, bj->end); | |
304 | return -EINVAL; | |
305 | } | |
306 | pr_warning("NUMA: Warning: node %d (%Lx-%Lx) overlaps with itself (%Lx-%Lx)\n", | |
307 | bi->nid, bi->start, bi->end, | |
308 | bj->start, bj->end); | |
309 | } | |
310 | ||
311 | /* | |
312 | * Join together blocks on the same node, holes | |
313 | * between which don't overlap with memory on other | |
314 | * nodes. | |
315 | */ | |
316 | if (bi->nid != bj->nid) | |
317 | continue; | |
e5a10c1b YL |
318 | start = min(bi->start, bj->start); |
319 | end = max(bi->end, bj->end); | |
a4106eae TH |
320 | for (k = 0; k < mi->nr_blks; k++) { |
321 | struct numa_memblk *bk = &mi->blk[k]; | |
322 | ||
323 | if (bi->nid == bk->nid) | |
324 | continue; | |
325 | if (start < bk->end && end > bk->start) | |
326 | break; | |
327 | } | |
328 | if (k < mi->nr_blks) | |
329 | continue; | |
38f3e1ca | 330 | printk(KERN_INFO "NUMA: Node %d [%Lx,%Lx) + [%Lx,%Lx) -> [%Lx,%Lx)\n", |
a4106eae TH |
331 | bi->nid, bi->start, bi->end, bj->start, bj->end, |
332 | start, end); | |
333 | bi->start = start; | |
334 | bi->end = end; | |
335 | numa_remove_memblk_from(j--, mi); | |
336 | } | |
337 | } | |
338 | ||
e5a10c1b | 339 | /* clear unused ones */ |
a4106eae TH |
340 | for (i = mi->nr_blks; i < ARRAY_SIZE(mi->blk); i++) { |
341 | mi->blk[i].start = mi->blk[i].end = 0; | |
342 | mi->blk[i].nid = NUMA_NO_NODE; | |
343 | } | |
344 | ||
345 | return 0; | |
346 | } | |
347 | ||
348 | /* | |
349 | * Set nodes, which have memory in @mi, in *@nodemask. | |
350 | */ | |
351 | static void __init numa_nodemask_from_meminfo(nodemask_t *nodemask, | |
352 | const struct numa_meminfo *mi) | |
353 | { | |
354 | int i; | |
355 | ||
356 | for (i = 0; i < ARRAY_SIZE(mi->blk); i++) | |
357 | if (mi->blk[i].start != mi->blk[i].end && | |
358 | mi->blk[i].nid != NUMA_NO_NODE) | |
359 | node_set(mi->blk[i].nid, *nodemask); | |
360 | } | |
361 | ||
362 | /** | |
363 | * numa_reset_distance - Reset NUMA distance table | |
364 | * | |
365 | * The current table is freed. The next numa_set_distance() call will | |
366 | * create a new one. | |
367 | */ | |
368 | void __init numa_reset_distance(void) | |
369 | { | |
370 | size_t size = numa_distance_cnt * numa_distance_cnt * sizeof(numa_distance[0]); | |
371 | ||
372 | /* numa_distance could be 1LU marking allocation failure, test cnt */ | |
373 | if (numa_distance_cnt) | |
374 | memblock_x86_free_range(__pa(numa_distance), | |
375 | __pa(numa_distance) + size); | |
376 | numa_distance_cnt = 0; | |
377 | numa_distance = NULL; /* enable table creation */ | |
378 | } | |
379 | ||
380 | static int __init numa_alloc_distance(void) | |
381 | { | |
382 | nodemask_t nodes_parsed; | |
383 | size_t size; | |
384 | int i, j, cnt = 0; | |
385 | u64 phys; | |
386 | ||
387 | /* size the new table and allocate it */ | |
388 | nodes_parsed = numa_nodes_parsed; | |
389 | numa_nodemask_from_meminfo(&nodes_parsed, &numa_meminfo); | |
390 | ||
391 | for_each_node_mask(i, nodes_parsed) | |
392 | cnt = i; | |
393 | cnt++; | |
394 | size = cnt * cnt * sizeof(numa_distance[0]); | |
395 | ||
38f3e1ca | 396 | phys = memblock_find_in_range(0, PFN_PHYS(max_pfn_mapped), |
a4106eae TH |
397 | size, PAGE_SIZE); |
398 | if (phys == MEMBLOCK_ERROR) { | |
399 | pr_warning("NUMA: Warning: can't allocate distance table!\n"); | |
400 | /* don't retry until explicitly reset */ | |
401 | numa_distance = (void *)1LU; | |
402 | return -ENOMEM; | |
403 | } | |
404 | memblock_x86_reserve_range(phys, phys + size, "NUMA DIST"); | |
405 | ||
406 | numa_distance = __va(phys); | |
407 | numa_distance_cnt = cnt; | |
408 | ||
409 | /* fill with the default distances */ | |
410 | for (i = 0; i < cnt; i++) | |
411 | for (j = 0; j < cnt; j++) | |
412 | numa_distance[i * cnt + j] = i == j ? | |
413 | LOCAL_DISTANCE : REMOTE_DISTANCE; | |
414 | printk(KERN_DEBUG "NUMA: Initialized distance table, cnt=%d\n", cnt); | |
415 | ||
416 | return 0; | |
417 | } | |
418 | ||
419 | /** | |
420 | * numa_set_distance - Set NUMA distance from one NUMA to another | |
421 | * @from: the 'from' node to set distance | |
422 | * @to: the 'to' node to set distance | |
423 | * @distance: NUMA distance | |
424 | * | |
425 | * Set the distance from node @from to @to to @distance. If distance table | |
426 | * doesn't exist, one which is large enough to accommodate all the currently | |
427 | * known nodes will be created. | |
428 | * | |
429 | * If such table cannot be allocated, a warning is printed and further | |
430 | * calls are ignored until the distance table is reset with | |
431 | * numa_reset_distance(). | |
432 | * | |
433 | * If @from or @to is higher than the highest known node at the time of | |
434 | * table creation or @distance doesn't make sense, the call is ignored. | |
435 | * This is to allow simplification of specific NUMA config implementations. | |
436 | */ | |
437 | void __init numa_set_distance(int from, int to, int distance) | |
438 | { | |
439 | if (!numa_distance && numa_alloc_distance() < 0) | |
440 | return; | |
441 | ||
442 | if (from >= numa_distance_cnt || to >= numa_distance_cnt) { | |
443 | printk_once(KERN_DEBUG "NUMA: Debug: distance out of bound, from=%d to=%d distance=%d\n", | |
444 | from, to, distance); | |
445 | return; | |
446 | } | |
447 | ||
448 | if ((u8)distance != distance || | |
449 | (from == to && distance != LOCAL_DISTANCE)) { | |
450 | pr_warn_once("NUMA: Warning: invalid distance parameter, from=%d to=%d distance=%d\n", | |
451 | from, to, distance); | |
452 | return; | |
453 | } | |
454 | ||
455 | numa_distance[from * numa_distance_cnt + to] = distance; | |
456 | } | |
457 | ||
458 | int __node_distance(int from, int to) | |
459 | { | |
460 | if (from >= numa_distance_cnt || to >= numa_distance_cnt) | |
461 | return from == to ? LOCAL_DISTANCE : REMOTE_DISTANCE; | |
462 | return numa_distance[from * numa_distance_cnt + to]; | |
463 | } | |
464 | EXPORT_SYMBOL(__node_distance); | |
465 | ||
466 | /* | |
467 | * Sanity check to catch more bad NUMA configurations (they are amazingly | |
468 | * common). Make sure the nodes cover all memory. | |
469 | */ | |
470 | static bool __init numa_meminfo_cover_memory(const struct numa_meminfo *mi) | |
471 | { | |
38f3e1ca | 472 | u64 numaram, e820ram; |
a4106eae TH |
473 | int i; |
474 | ||
475 | numaram = 0; | |
476 | for (i = 0; i < mi->nr_blks; i++) { | |
38f3e1ca TH |
477 | u64 s = mi->blk[i].start >> PAGE_SHIFT; |
478 | u64 e = mi->blk[i].end >> PAGE_SHIFT; | |
a4106eae TH |
479 | numaram += e - s; |
480 | numaram -= __absent_pages_in_range(mi->blk[i].nid, s, e); | |
38f3e1ca | 481 | if ((s64)numaram < 0) |
a4106eae TH |
482 | numaram = 0; |
483 | } | |
484 | ||
485 | e820ram = max_pfn - (memblock_x86_hole_size(0, | |
38f3e1ca | 486 | PFN_PHYS(max_pfn)) >> PAGE_SHIFT); |
a4106eae | 487 | /* We seem to lose 3 pages somewhere. Allow 1M of slack. */ |
38f3e1ca TH |
488 | if ((s64)(e820ram - numaram) >= (1 << (20 - PAGE_SHIFT))) { |
489 | printk(KERN_ERR "NUMA: nodes only cover %LuMB of your %LuMB e820 RAM. Not used.\n", | |
a4106eae TH |
490 | (numaram << PAGE_SHIFT) >> 20, |
491 | (e820ram << PAGE_SHIFT) >> 20); | |
492 | return false; | |
493 | } | |
494 | return true; | |
495 | } | |
496 | ||
497 | static int __init numa_register_memblks(struct numa_meminfo *mi) | |
498 | { | |
1e01979c | 499 | unsigned long uninitialized_var(pfn_align); |
a4106eae TH |
500 | int i, nid; |
501 | ||
502 | /* Account for nodes with cpus and no memory */ | |
503 | node_possible_map = numa_nodes_parsed; | |
504 | numa_nodemask_from_meminfo(&node_possible_map, mi); | |
505 | if (WARN_ON(nodes_empty(node_possible_map))) | |
506 | return -EINVAL; | |
507 | ||
508 | for (i = 0; i < mi->nr_blks; i++) | |
509 | memblock_x86_register_active_regions(mi->blk[i].nid, | |
510 | mi->blk[i].start >> PAGE_SHIFT, | |
511 | mi->blk[i].end >> PAGE_SHIFT); | |
512 | ||
513 | /* for out of order entries */ | |
514 | sort_node_map(); | |
1e01979c TH |
515 | |
516 | /* | |
517 | * If sections array is gonna be used for pfn -> nid mapping, check | |
518 | * whether its granularity is fine enough. | |
519 | */ | |
520 | #ifdef NODE_NOT_IN_PAGE_FLAGS | |
521 | pfn_align = node_map_pfn_alignment(); | |
522 | if (pfn_align && pfn_align < PAGES_PER_SECTION) { | |
523 | printk(KERN_WARNING "Node alignment %LuMB < min %LuMB, rejecting NUMA config\n", | |
524 | PFN_PHYS(pfn_align) >> 20, | |
525 | PFN_PHYS(PAGES_PER_SECTION) >> 20); | |
526 | return -EINVAL; | |
527 | } | |
528 | #endif | |
a4106eae TH |
529 | if (!numa_meminfo_cover_memory(mi)) |
530 | return -EINVAL; | |
531 | ||
532 | /* Finally register nodes. */ | |
533 | for_each_node_mask(nid, node_possible_map) { | |
38f3e1ca | 534 | u64 start = PFN_PHYS(max_pfn); |
a4106eae TH |
535 | u64 end = 0; |
536 | ||
537 | for (i = 0; i < mi->nr_blks; i++) { | |
538 | if (nid != mi->blk[i].nid) | |
539 | continue; | |
540 | start = min(mi->blk[i].start, start); | |
541 | end = max(mi->blk[i].end, end); | |
542 | } | |
543 | ||
544 | if (start < end) | |
a56bca80 | 545 | setup_node_data(nid, start, end); |
a4106eae TH |
546 | } |
547 | ||
548 | return 0; | |
549 | } | |
a4106eae | 550 | |
8db78cc4 TH |
551 | /* |
552 | * There are unfortunately some poorly designed mainboards around that | |
553 | * only connect memory to a single CPU. This breaks the 1:1 cpu->node | |
554 | * mapping. To avoid this fill in the mapping for all possible CPUs, | |
555 | * as the number of CPUs is not known yet. We round robin the existing | |
556 | * nodes. | |
557 | */ | |
752d4f37 | 558 | static void __init numa_init_array(void) |
8db78cc4 TH |
559 | { |
560 | int rr, i; | |
561 | ||
562 | rr = first_node(node_online_map); | |
563 | for (i = 0; i < nr_cpu_ids; i++) { | |
564 | if (early_cpu_to_node(i) != NUMA_NO_NODE) | |
565 | continue; | |
566 | numa_set_node(i, rr); | |
567 | rr = next_node(rr, node_online_map); | |
568 | if (rr == MAX_NUMNODES) | |
569 | rr = first_node(node_online_map); | |
570 | } | |
571 | } | |
572 | ||
a4106eae TH |
573 | static int __init numa_init(int (*init_func)(void)) |
574 | { | |
575 | int i; | |
576 | int ret; | |
577 | ||
578 | for (i = 0; i < MAX_LOCAL_APIC; i++) | |
579 | set_apicid_to_node(i, NUMA_NO_NODE); | |
580 | ||
581 | nodes_clear(numa_nodes_parsed); | |
582 | nodes_clear(node_possible_map); | |
583 | nodes_clear(node_online_map); | |
584 | memset(&numa_meminfo, 0, sizeof(numa_meminfo)); | |
585 | remove_all_active_ranges(); | |
586 | numa_reset_distance(); | |
587 | ||
588 | ret = init_func(); | |
589 | if (ret < 0) | |
590 | return ret; | |
591 | ret = numa_cleanup_meminfo(&numa_meminfo); | |
592 | if (ret < 0) | |
593 | return ret; | |
594 | ||
595 | numa_emulation(&numa_meminfo, numa_distance_cnt); | |
596 | ||
597 | ret = numa_register_memblks(&numa_meminfo); | |
598 | if (ret < 0) | |
599 | return ret; | |
600 | ||
601 | for (i = 0; i < nr_cpu_ids; i++) { | |
602 | int nid = early_cpu_to_node(i); | |
603 | ||
604 | if (nid == NUMA_NO_NODE) | |
605 | continue; | |
606 | if (!node_online(nid)) | |
607 | numa_clear_node(i); | |
608 | } | |
609 | numa_init_array(); | |
610 | return 0; | |
611 | } | |
612 | ||
613 | /** | |
614 | * dummy_numa_init - Fallback dummy NUMA init | |
615 | * | |
616 | * Used if there's no underlying NUMA architecture, NUMA initialization | |
617 | * fails, or NUMA is disabled on the command line. | |
618 | * | |
619 | * Must online at least one node and add memory blocks that cover all | |
620 | * allowed memory. This function must not fail. | |
621 | */ | |
622 | static int __init dummy_numa_init(void) | |
623 | { | |
624 | printk(KERN_INFO "%s\n", | |
625 | numa_off ? "NUMA turned off" : "No NUMA configuration found"); | |
38f3e1ca TH |
626 | printk(KERN_INFO "Faking a node at %016Lx-%016Lx\n", |
627 | 0LLU, PFN_PHYS(max_pfn)); | |
a4106eae TH |
628 | |
629 | node_set(0, numa_nodes_parsed); | |
38f3e1ca | 630 | numa_add_memblk(0, 0, PFN_PHYS(max_pfn)); |
a4106eae TH |
631 | |
632 | return 0; | |
633 | } | |
634 | ||
635 | /** | |
636 | * x86_numa_init - Initialize NUMA | |
637 | * | |
638 | * Try each configured NUMA initialization method until one succeeds. The | |
639 | * last fallback is dummy single node config encomapssing whole memory and | |
640 | * never fails. | |
641 | */ | |
642 | void __init x86_numa_init(void) | |
643 | { | |
644 | if (!numa_off) { | |
bd6709a9 TH |
645 | #ifdef CONFIG_X86_NUMAQ |
646 | if (!numa_init(numaq_numa_init)) | |
647 | return; | |
648 | #endif | |
a4106eae TH |
649 | #ifdef CONFIG_ACPI_NUMA |
650 | if (!numa_init(x86_acpi_numa_init)) | |
651 | return; | |
652 | #endif | |
653 | #ifdef CONFIG_AMD_NUMA | |
654 | if (!numa_init(amd_numa_init)) | |
655 | return; | |
656 | #endif | |
657 | } | |
658 | ||
659 | numa_init(dummy_numa_init); | |
660 | } | |
a4106eae | 661 | |
8db78cc4 TH |
662 | static __init int find_near_online_node(int node) |
663 | { | |
664 | int n, val; | |
665 | int min_val = INT_MAX; | |
666 | int best_node = -1; | |
667 | ||
668 | for_each_online_node(n) { | |
669 | val = node_distance(node, n); | |
670 | ||
671 | if (val < min_val) { | |
672 | min_val = val; | |
673 | best_node = n; | |
674 | } | |
675 | } | |
676 | ||
677 | return best_node; | |
678 | } | |
679 | ||
680 | /* | |
681 | * Setup early cpu_to_node. | |
682 | * | |
683 | * Populate cpu_to_node[] only if x86_cpu_to_apicid[], | |
684 | * and apicid_to_node[] tables have valid entries for a CPU. | |
685 | * This means we skip cpu_to_node[] initialisation for NUMA | |
686 | * emulation and faking node case (when running a kernel compiled | |
687 | * for NUMA on a non NUMA box), which is OK as cpu_to_node[] | |
688 | * is already initialized in a round robin manner at numa_init_array, | |
689 | * prior to this call, and this initialization is good enough | |
690 | * for the fake NUMA cases. | |
691 | * | |
692 | * Called before the per_cpu areas are setup. | |
693 | */ | |
694 | void __init init_cpu_to_node(void) | |
695 | { | |
696 | int cpu; | |
697 | u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid); | |
698 | ||
699 | BUG_ON(cpu_to_apicid == NULL); | |
700 | ||
701 | for_each_possible_cpu(cpu) { | |
702 | int node = numa_cpu_node(cpu); | |
703 | ||
704 | if (node == NUMA_NO_NODE) | |
705 | continue; | |
706 | if (!node_online(node)) | |
707 | node = find_near_online_node(node); | |
708 | numa_set_node(cpu, node); | |
709 | } | |
710 | } | |
711 | ||
de2d9445 TH |
712 | #ifndef CONFIG_DEBUG_PER_CPU_MAPS |
713 | ||
714 | # ifndef CONFIG_NUMA_EMU | |
715 | void __cpuinit numa_add_cpu(int cpu) | |
716 | { | |
717 | cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); | |
718 | } | |
719 | ||
720 | void __cpuinit numa_remove_cpu(int cpu) | |
721 | { | |
722 | cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); | |
723 | } | |
724 | # endif /* !CONFIG_NUMA_EMU */ | |
725 | ||
726 | #else /* !CONFIG_DEBUG_PER_CPU_MAPS */ | |
645a7919 TH |
727 | |
728 | int __cpu_to_node(int cpu) | |
729 | { | |
730 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) { | |
731 | printk(KERN_WARNING | |
732 | "cpu_to_node(%d): usage too early!\n", cpu); | |
733 | dump_stack(); | |
734 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
735 | } | |
736 | return per_cpu(x86_cpu_to_node_map, cpu); | |
737 | } | |
738 | EXPORT_SYMBOL(__cpu_to_node); | |
739 | ||
740 | /* | |
741 | * Same function as cpu_to_node() but used if called before the | |
742 | * per_cpu areas are setup. | |
743 | */ | |
744 | int early_cpu_to_node(int cpu) | |
745 | { | |
746 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) | |
747 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
748 | ||
749 | if (!cpu_possible(cpu)) { | |
750 | printk(KERN_WARNING | |
751 | "early_cpu_to_node(%d): no per_cpu area!\n", cpu); | |
752 | dump_stack(); | |
753 | return NUMA_NO_NODE; | |
754 | } | |
755 | return per_cpu(x86_cpu_to_node_map, cpu); | |
756 | } | |
757 | ||
7a6c6547 | 758 | void debug_cpumask_set_cpu(int cpu, int node, bool enable) |
de2d9445 | 759 | { |
de2d9445 TH |
760 | struct cpumask *mask; |
761 | char buf[64]; | |
762 | ||
14392fd3 DR |
763 | if (node == NUMA_NO_NODE) { |
764 | /* early_cpu_to_node() already emits a warning and trace */ | |
7a6c6547 | 765 | return; |
14392fd3 | 766 | } |
de2d9445 TH |
767 | mask = node_to_cpumask_map[node]; |
768 | if (!mask) { | |
769 | pr_err("node_to_cpumask_map[%i] NULL\n", node); | |
770 | dump_stack(); | |
7a6c6547 | 771 | return; |
de2d9445 TH |
772 | } |
773 | ||
7a6c6547 DR |
774 | if (enable) |
775 | cpumask_set_cpu(cpu, mask); | |
776 | else | |
777 | cpumask_clear_cpu(cpu, mask); | |
778 | ||
de2d9445 TH |
779 | cpulist_scnprintf(buf, sizeof(buf), mask); |
780 | printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n", | |
781 | enable ? "numa_add_cpu" : "numa_remove_cpu", | |
782 | cpu, node, buf); | |
7a6c6547 | 783 | return; |
de2d9445 TH |
784 | } |
785 | ||
786 | # ifndef CONFIG_NUMA_EMU | |
7a6c6547 | 787 | static void __cpuinit numa_set_cpumask(int cpu, bool enable) |
de2d9445 | 788 | { |
7a6c6547 | 789 | debug_cpumask_set_cpu(cpu, early_cpu_to_node(cpu), enable); |
de2d9445 TH |
790 | } |
791 | ||
792 | void __cpuinit numa_add_cpu(int cpu) | |
793 | { | |
7a6c6547 | 794 | numa_set_cpumask(cpu, true); |
de2d9445 TH |
795 | } |
796 | ||
797 | void __cpuinit numa_remove_cpu(int cpu) | |
798 | { | |
7a6c6547 | 799 | numa_set_cpumask(cpu, false); |
de2d9445 TH |
800 | } |
801 | # endif /* !CONFIG_NUMA_EMU */ | |
802 | ||
71ee73e7 RR |
803 | /* |
804 | * Returns a pointer to the bitmask of CPUs on Node 'node'. | |
805 | */ | |
73e907de | 806 | const struct cpumask *cpumask_of_node(int node) |
71ee73e7 | 807 | { |
71ee73e7 RR |
808 | if (node >= nr_node_ids) { |
809 | printk(KERN_WARNING | |
810 | "cpumask_of_node(%d): node > nr_node_ids(%d)\n", | |
811 | node, nr_node_ids); | |
812 | dump_stack(); | |
813 | return cpu_none_mask; | |
814 | } | |
c032ef60 RR |
815 | if (node_to_cpumask_map[node] == NULL) { |
816 | printk(KERN_WARNING | |
817 | "cpumask_of_node(%d): no node_to_cpumask_map!\n", | |
818 | node); | |
819 | dump_stack(); | |
820 | return cpu_online_mask; | |
821 | } | |
0b966252 | 822 | return node_to_cpumask_map[node]; |
71ee73e7 RR |
823 | } |
824 | EXPORT_SYMBOL(cpumask_of_node); | |
645a7919 | 825 | |
de2d9445 | 826 | #endif /* !CONFIG_DEBUG_PER_CPU_MAPS */ |
a4106eae | 827 | |
bd6709a9 | 828 | #ifdef CONFIG_MEMORY_HOTPLUG |
a4106eae TH |
829 | int memory_add_physaddr_to_nid(u64 start) |
830 | { | |
831 | struct numa_meminfo *mi = &numa_meminfo; | |
832 | int nid = mi->blk[0].nid; | |
833 | int i; | |
834 | ||
835 | for (i = 0; i < mi->nr_blks; i++) | |
836 | if (mi->blk[i].start <= start && mi->blk[i].end > start) | |
837 | nid = mi->blk[i].nid; | |
838 | return nid; | |
839 | } | |
840 | EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); | |
841 | #endif |