Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * ACPI 3.0 based NUMA setup | |
3 | * Copyright 2004 Andi Kleen, SuSE Labs. | |
4 | * | |
5 | * Reads the ACPI SRAT table to figure out what memory belongs to which CPUs. | |
6 | * | |
7 | * Called from acpi_numa_init while reading the SRAT and SLIT tables. | |
8 | * Assumes all memory regions belonging to a single proximity domain | |
9 | * are in one chunk. Holes between them will be included in the node. | |
10 | */ | |
11 | ||
12 | #include <linux/kernel.h> | |
13 | #include <linux/acpi.h> | |
14 | #include <linux/mmzone.h> | |
15 | #include <linux/bitmap.h> | |
16 | #include <linux/module.h> | |
17 | #include <linux/topology.h> | |
68a3a7fe AK |
18 | #include <linux/bootmem.h> |
19 | #include <linux/mm.h> | |
1da177e4 LT |
20 | #include <asm/proto.h> |
21 | #include <asm/numa.h> | |
8a6fdd3e | 22 | #include <asm/e820.h> |
1da177e4 | 23 | |
c31fbb1a AK |
24 | int acpi_numa __initdata; |
25 | ||
1da177e4 LT |
26 | static struct acpi_table_slit *acpi_slit; |
27 | ||
28 | static nodemask_t nodes_parsed __initdata; | |
abe059e7 | 29 | static struct bootnode nodes[MAX_NUMNODES] __initdata; |
4942e998 | 30 | static struct bootnode nodes_add[MAX_NUMNODES]; |
68a3a7fe | 31 | static int found_add_area __initdata; |
fad7906d | 32 | int hotadd_percent __initdata = 0; |
1da177e4 | 33 | |
9391a3f9 AK |
34 | /* Too small nodes confuse the VM badly. Usually they result |
35 | from BIOS bugs. */ | |
36 | #define NODE_MIN_SIZE (4*1024*1024) | |
37 | ||
1da177e4 LT |
38 | static __init int setup_node(int pxm) |
39 | { | |
762834e8 | 40 | return acpi_map_pxm_to_node(pxm); |
1da177e4 LT |
41 | } |
42 | ||
43 | static __init int conflicting_nodes(unsigned long start, unsigned long end) | |
44 | { | |
45 | int i; | |
4b6a455c | 46 | for_each_node_mask(i, nodes_parsed) { |
abe059e7 | 47 | struct bootnode *nd = &nodes[i]; |
1da177e4 LT |
48 | if (nd->start == nd->end) |
49 | continue; | |
50 | if (nd->end > start && nd->start < end) | |
05d1fa4b | 51 | return i; |
1da177e4 | 52 | if (nd->end == end && nd->start == start) |
05d1fa4b | 53 | return i; |
1da177e4 LT |
54 | } |
55 | return -1; | |
56 | } | |
57 | ||
58 | static __init void cutoff_node(int i, unsigned long start, unsigned long end) | |
59 | { | |
abe059e7 | 60 | struct bootnode *nd = &nodes[i]; |
68a3a7fe AK |
61 | |
62 | if (found_add_area) | |
63 | return; | |
64 | ||
1da177e4 LT |
65 | if (nd->start < start) { |
66 | nd->start = start; | |
67 | if (nd->end < nd->start) | |
68 | nd->start = nd->end; | |
69 | } | |
70 | if (nd->end > end) { | |
1da177e4 LT |
71 | nd->end = end; |
72 | if (nd->start > nd->end) | |
73 | nd->start = nd->end; | |
74 | } | |
75 | } | |
76 | ||
77 | static __init void bad_srat(void) | |
78 | { | |
2bce2b54 | 79 | int i; |
1da177e4 LT |
80 | printk(KERN_ERR "SRAT: SRAT not used.\n"); |
81 | acpi_numa = -1; | |
fad7906d | 82 | found_add_area = 0; |
2bce2b54 AK |
83 | for (i = 0; i < MAX_LOCAL_APIC; i++) |
84 | apicid_to_node[i] = NUMA_NO_NODE; | |
68a3a7fe AK |
85 | for (i = 0; i < MAX_NUMNODES; i++) |
86 | nodes_add[i].start = nodes[i].end = 0; | |
5cb248ab | 87 | remove_all_active_ranges(); |
1da177e4 LT |
88 | } |
89 | ||
90 | static __init inline int srat_disabled(void) | |
91 | { | |
92 | return numa_off || acpi_numa < 0; | |
93 | } | |
94 | ||
1584b89c AK |
95 | /* |
96 | * A lot of BIOS fill in 10 (= no distance) everywhere. This messes | |
97 | * up the NUMA heuristics which wants the local node to have a smaller | |
98 | * distance than the others. | |
99 | * Do some quick checks here and only use the SLIT if it passes. | |
100 | */ | |
101 | static __init int slit_valid(struct acpi_table_slit *slit) | |
102 | { | |
103 | int i, j; | |
104 | int d = slit->localities; | |
105 | for (i = 0; i < d; i++) { | |
106 | for (j = 0; j < d; j++) { | |
107 | u8 val = slit->entry[d*i + j]; | |
108 | if (i == j) { | |
109 | if (val != 10) | |
110 | return 0; | |
111 | } else if (val <= 10) | |
112 | return 0; | |
113 | } | |
114 | } | |
115 | return 1; | |
116 | } | |
117 | ||
1da177e4 LT |
118 | /* Callback for SLIT parsing */ |
119 | void __init acpi_numa_slit_init(struct acpi_table_slit *slit) | |
120 | { | |
1584b89c AK |
121 | if (!slit_valid(slit)) { |
122 | printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n"); | |
123 | return; | |
124 | } | |
1da177e4 LT |
125 | acpi_slit = slit; |
126 | } | |
127 | ||
128 | /* Callback for Proximity Domain -> LAPIC mapping */ | |
129 | void __init | |
130 | acpi_numa_processor_affinity_init(struct acpi_table_processor_affinity *pa) | |
131 | { | |
132 | int pxm, node; | |
d22fe808 AK |
133 | if (srat_disabled()) |
134 | return; | |
fad7906d AK |
135 | if (pa->header.length != sizeof(struct acpi_table_processor_affinity)) { |
136 | bad_srat(); | |
d22fe808 AK |
137 | return; |
138 | } | |
139 | if (pa->flags.enabled == 0) | |
1da177e4 LT |
140 | return; |
141 | pxm = pa->proximity_domain; | |
142 | node = setup_node(pxm); | |
143 | if (node < 0) { | |
144 | printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm); | |
145 | bad_srat(); | |
146 | return; | |
147 | } | |
0b07e984 | 148 | apicid_to_node[pa->apic_id] = node; |
1da177e4 | 149 | acpi_numa = 1; |
0b07e984 AK |
150 | printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n", |
151 | pxm, pa->apic_id, node); | |
1da177e4 LT |
152 | } |
153 | ||
71efa8fd | 154 | #ifdef CONFIG_MEMORY_HOTPLUG_RESERVE |
68a3a7fe AK |
155 | /* |
156 | * Protect against too large hotadd areas that would fill up memory. | |
157 | */ | |
158 | static int hotadd_enough_memory(struct bootnode *nd) | |
159 | { | |
160 | static unsigned long allocated; | |
161 | static unsigned long last_area_end; | |
162 | unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT; | |
163 | long mem = pages * sizeof(struct page); | |
164 | unsigned long addr; | |
165 | unsigned long allowed; | |
166 | unsigned long oldpages = pages; | |
167 | ||
168 | if (mem < 0) | |
169 | return 0; | |
5cb248ab | 170 | allowed = (end_pfn - absent_pages_in_range(0, end_pfn)) * PAGE_SIZE; |
68a3a7fe AK |
171 | allowed = (allowed / 100) * hotadd_percent; |
172 | if (allocated + mem > allowed) { | |
fad7906d | 173 | unsigned long range; |
68a3a7fe AK |
174 | /* Give them at least part of their hotadd memory upto hotadd_percent |
175 | It would be better to spread the limit out | |
176 | over multiple hotplug areas, but that is too complicated | |
177 | right now */ | |
178 | if (allocated >= allowed) | |
179 | return 0; | |
fad7906d AK |
180 | range = allowed - allocated; |
181 | pages = (range / PAGE_SIZE); | |
68a3a7fe | 182 | mem = pages * sizeof(struct page); |
fad7906d | 183 | nd->end = nd->start + range; |
68a3a7fe AK |
184 | } |
185 | /* Not completely fool proof, but a good sanity check */ | |
186 | addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem); | |
187 | if (addr == -1UL) | |
188 | return 0; | |
189 | if (pages != oldpages) | |
190 | printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n", | |
191 | pages << PAGE_SHIFT); | |
192 | last_area_end = addr + mem; | |
193 | allocated += mem; | |
194 | return 1; | |
195 | } | |
196 | ||
71efa8fd KM |
197 | static int update_end_of_memory(unsigned long end) |
198 | { | |
199 | found_add_area = 1; | |
200 | if ((end >> PAGE_SHIFT) > end_pfn) | |
201 | end_pfn = end >> PAGE_SHIFT; | |
202 | return 1; | |
203 | } | |
204 | ||
205 | static inline int save_add_info(void) | |
206 | { | |
207 | return hotadd_percent > 0; | |
208 | } | |
209 | #else | |
210 | int update_end_of_memory(unsigned long end) {return 0;} | |
211 | static int hotadd_enough_memory(struct bootnode *nd) {return 1;} | |
212 | #ifdef CONFIG_MEMORY_HOTPLUG_SPARSE | |
213 | static inline int save_add_info(void) {return 1;} | |
214 | #else | |
215 | static inline int save_add_info(void) {return 0;} | |
216 | #endif | |
217 | #endif | |
68a3a7fe | 218 | /* |
71efa8fd KM |
219 | * Update nodes_add and decide if to include add are in the zone. |
220 | * Both SPARSE and RESERVE need nodes_add infomation. | |
68a3a7fe AK |
221 | * This code supports one contigious hot add area per node. |
222 | */ | |
223 | static int reserve_hotadd(int node, unsigned long start, unsigned long end) | |
224 | { | |
225 | unsigned long s_pfn = start >> PAGE_SHIFT; | |
226 | unsigned long e_pfn = end >> PAGE_SHIFT; | |
71efa8fd | 227 | int ret = 0, changed = 0; |
68a3a7fe AK |
228 | struct bootnode *nd = &nodes_add[node]; |
229 | ||
230 | /* I had some trouble with strange memory hotadd regions breaking | |
231 | the boot. Be very strict here and reject anything unexpected. | |
232 | If you want working memory hotadd write correct SRATs. | |
233 | ||
234 | The node size check is a basic sanity check to guard against | |
235 | mistakes */ | |
236 | if ((signed long)(end - start) < NODE_MIN_SIZE) { | |
237 | printk(KERN_ERR "SRAT: Hotplug area too small\n"); | |
238 | return -1; | |
239 | } | |
240 | ||
241 | /* This check might be a bit too strict, but I'm keeping it for now. */ | |
5cb248ab | 242 | if (absent_pages_in_range(s_pfn, e_pfn) != e_pfn - s_pfn) { |
9c7cd687 MG |
243 | printk(KERN_ERR |
244 | "SRAT: Hotplug area %lu -> %lu has existing memory\n", | |
245 | s_pfn, e_pfn); | |
68a3a7fe AK |
246 | return -1; |
247 | } | |
248 | ||
249 | if (!hotadd_enough_memory(&nodes_add[node])) { | |
250 | printk(KERN_ERR "SRAT: Hotplug area too large\n"); | |
251 | return -1; | |
252 | } | |
253 | ||
254 | /* Looks good */ | |
255 | ||
68a3a7fe AK |
256 | if (nd->start == nd->end) { |
257 | nd->start = start; | |
258 | nd->end = end; | |
259 | changed = 1; | |
260 | } else { | |
261 | if (nd->start == end) { | |
262 | nd->start = start; | |
263 | changed = 1; | |
264 | } | |
265 | if (nd->end == start) { | |
266 | nd->end = end; | |
267 | changed = 1; | |
268 | } | |
269 | if (!changed) | |
270 | printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n"); | |
271 | } | |
272 | ||
71efa8fd | 273 | ret = update_end_of_memory(nd->end); |
68a3a7fe AK |
274 | |
275 | if (changed) | |
276 | printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end); | |
71efa8fd | 277 | return ret; |
68a3a7fe | 278 | } |
68a3a7fe | 279 | |
1da177e4 LT |
280 | /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */ |
281 | void __init | |
282 | acpi_numa_memory_affinity_init(struct acpi_table_memory_affinity *ma) | |
283 | { | |
68a3a7fe | 284 | struct bootnode *nd, oldnode; |
1da177e4 LT |
285 | unsigned long start, end; |
286 | int node, pxm; | |
287 | int i; | |
288 | ||
d22fe808 | 289 | if (srat_disabled()) |
1da177e4 | 290 | return; |
d22fe808 AK |
291 | if (ma->header.length != sizeof(struct acpi_table_memory_affinity)) { |
292 | bad_srat(); | |
293 | return; | |
294 | } | |
295 | if (ma->flags.enabled == 0) | |
296 | return; | |
71efa8fd | 297 | if (ma->flags.hot_pluggable && !save_add_info()) |
68a3a7fe | 298 | return; |
d22fe808 AK |
299 | start = ma->base_addr_lo | ((u64)ma->base_addr_hi << 32); |
300 | end = start + (ma->length_lo | ((u64)ma->length_hi << 32)); | |
1da177e4 LT |
301 | pxm = ma->proximity_domain; |
302 | node = setup_node(pxm); | |
303 | if (node < 0) { | |
304 | printk(KERN_ERR "SRAT: Too many proximity domains.\n"); | |
305 | bad_srat(); | |
306 | return; | |
307 | } | |
1da177e4 | 308 | i = conflicting_nodes(start, end); |
05d1fa4b AK |
309 | if (i == node) { |
310 | printk(KERN_WARNING | |
311 | "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n", | |
312 | pxm, start, end, nodes[i].start, nodes[i].end); | |
313 | } else if (i >= 0) { | |
1da177e4 | 314 | printk(KERN_ERR |
05d1fa4b AK |
315 | "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n", |
316 | pxm, start, end, node_to_pxm(i), | |
317 | nodes[i].start, nodes[i].end); | |
1da177e4 LT |
318 | bad_srat(); |
319 | return; | |
320 | } | |
321 | nd = &nodes[node]; | |
68a3a7fe | 322 | oldnode = *nd; |
1da177e4 LT |
323 | if (!node_test_and_set(node, nodes_parsed)) { |
324 | nd->start = start; | |
325 | nd->end = end; | |
326 | } else { | |
327 | if (start < nd->start) | |
328 | nd->start = start; | |
329 | if (nd->end < end) | |
330 | nd->end = end; | |
331 | } | |
68a3a7fe | 332 | |
1da177e4 LT |
333 | printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm, |
334 | nd->start, nd->end); | |
5cb248ab MG |
335 | e820_register_active_regions(node, nd->start >> PAGE_SHIFT, |
336 | nd->end >> PAGE_SHIFT); | |
fb01439c MG |
337 | push_node_boundaries(node, nd->start >> PAGE_SHIFT, |
338 | nd->end >> PAGE_SHIFT); | |
68a3a7fe | 339 | |
71efa8fd | 340 | if (ma->flags.hot_pluggable && !reserve_hotadd(node, start, end) < 0) { |
68a3a7fe AK |
341 | /* Ignore hotadd region. Undo damage */ |
342 | printk(KERN_NOTICE "SRAT: Hotplug region ignored\n"); | |
343 | *nd = oldnode; | |
344 | if ((nd->start | nd->end) == 0) | |
345 | node_clear(node, nodes_parsed); | |
346 | } | |
1da177e4 LT |
347 | } |
348 | ||
8a6fdd3e AK |
349 | /* Sanity check to catch more bad SRATs (they are amazingly common). |
350 | Make sure the PXMs cover all memory. */ | |
351 | static int nodes_cover_memory(void) | |
352 | { | |
353 | int i; | |
354 | unsigned long pxmram, e820ram; | |
355 | ||
356 | pxmram = 0; | |
357 | for_each_node_mask(i, nodes_parsed) { | |
358 | unsigned long s = nodes[i].start >> PAGE_SHIFT; | |
359 | unsigned long e = nodes[i].end >> PAGE_SHIFT; | |
360 | pxmram += e - s; | |
5cb248ab | 361 | pxmram -= absent_pages_in_range(s, e); |
68a3a7fe AK |
362 | if ((long)pxmram < 0) |
363 | pxmram = 0; | |
8a6fdd3e AK |
364 | } |
365 | ||
5cb248ab | 366 | e820ram = end_pfn - absent_pages_in_range(0, end_pfn); |
fdb9df94 AK |
367 | /* We seem to lose 3 pages somewhere. Allow a bit of slack. */ |
368 | if ((long)(e820ram - pxmram) >= 1*1024*1024) { | |
8a6fdd3e AK |
369 | printk(KERN_ERR |
370 | "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n", | |
371 | (pxmram << PAGE_SHIFT) >> 20, | |
372 | (e820ram << PAGE_SHIFT) >> 20); | |
373 | return 0; | |
374 | } | |
375 | return 1; | |
376 | } | |
377 | ||
9391a3f9 AK |
378 | static void unparse_node(int node) |
379 | { | |
380 | int i; | |
381 | node_clear(node, nodes_parsed); | |
382 | for (i = 0; i < MAX_LOCAL_APIC; i++) { | |
383 | if (apicid_to_node[i] == node) | |
384 | apicid_to_node[i] = NUMA_NO_NODE; | |
385 | } | |
386 | } | |
387 | ||
1da177e4 LT |
388 | void __init acpi_numa_arch_fixup(void) {} |
389 | ||
390 | /* Use the information discovered above to actually set up the nodes. */ | |
391 | int __init acpi_scan_nodes(unsigned long start, unsigned long end) | |
392 | { | |
393 | int i; | |
8a6fdd3e | 394 | |
e58e0d03 | 395 | /* First clean up the node list */ |
9391a3f9 | 396 | for (i = 0; i < MAX_NUMNODES; i++) { |
68a3a7fe | 397 | cutoff_node(i, start, end); |
0d015324 | 398 | if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) { |
9391a3f9 | 399 | unparse_node(i); |
0d015324 DY |
400 | node_set_offline(i); |
401 | } | |
e58e0d03 AK |
402 | } |
403 | ||
9391a3f9 AK |
404 | if (acpi_numa <= 0) |
405 | return -1; | |
406 | ||
8a6fdd3e AK |
407 | if (!nodes_cover_memory()) { |
408 | bad_srat(); | |
409 | return -1; | |
410 | } | |
411 | ||
2aed711a | 412 | memnode_shift = compute_hash_shift(nodes, MAX_NUMNODES); |
1da177e4 LT |
413 | if (memnode_shift < 0) { |
414 | printk(KERN_ERR | |
415 | "SRAT: No NUMA node hash function found. Contact maintainer\n"); | |
416 | bad_srat(); | |
417 | return -1; | |
418 | } | |
e58e0d03 AK |
419 | |
420 | /* Finally register nodes */ | |
421 | for_each_node_mask(i, nodes_parsed) | |
1da177e4 | 422 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); |
a8062231 AK |
423 | /* Try again in case setup_node_bootmem missed one due |
424 | to missing bootmem */ | |
425 | for_each_node_mask(i, nodes_parsed) | |
426 | if (!node_online(i)) | |
427 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); | |
428 | ||
1da177e4 LT |
429 | for (i = 0; i < NR_CPUS; i++) { |
430 | if (cpu_to_node[i] == NUMA_NO_NODE) | |
431 | continue; | |
432 | if (!node_isset(cpu_to_node[i], nodes_parsed)) | |
69d81fcd | 433 | numa_set_node(i, NUMA_NO_NODE); |
1da177e4 LT |
434 | } |
435 | numa_init_array(); | |
436 | return 0; | |
437 | } | |
438 | ||
68a3a7fe AK |
439 | void __init srat_reserve_add_area(int nodeid) |
440 | { | |
441 | if (found_add_area && nodes_add[nodeid].end) { | |
442 | u64 total_mb; | |
443 | ||
444 | printk(KERN_INFO "SRAT: Reserving hot-add memory space " | |
445 | "for node %d at %Lx-%Lx\n", | |
446 | nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end); | |
447 | total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start) | |
448 | >> PAGE_SHIFT; | |
449 | total_mb *= sizeof(struct page); | |
450 | total_mb >>= 20; | |
451 | printk(KERN_INFO "SRAT: This will cost you %Lu MB of " | |
452 | "pre-allocated memory.\n", (unsigned long long)total_mb); | |
453 | reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start, | |
454 | nodes_add[nodeid].end - nodes_add[nodeid].start); | |
455 | } | |
456 | } | |
457 | ||
1da177e4 LT |
458 | int __node_distance(int a, int b) |
459 | { | |
460 | int index; | |
461 | ||
462 | if (!acpi_slit) | |
463 | return a == b ? 10 : 20; | |
464 | index = acpi_slit->localities * node_to_pxm(a); | |
465 | return acpi_slit->entry[index + node_to_pxm(b)]; | |
466 | } | |
467 | ||
468 | EXPORT_SYMBOL(__node_distance); | |
4942e998 KM |
469 | |
470 | int memory_add_physaddr_to_nid(u64 start) | |
471 | { | |
472 | int i, ret = 0; | |
473 | ||
474 | for_each_node(i) | |
475 | if (nodes_add[i].start <= start && nodes_add[i].end > start) | |
476 | ret = i; | |
477 | ||
478 | return ret; | |
479 | } |