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