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; | |
1da177e4 LT |
96 | } |
97 | ||
98 | static __init inline int srat_disabled(void) | |
99 | { | |
100 | return numa_off || acpi_numa < 0; | |
101 | } | |
102 | ||
1584b89c AK |
103 | /* |
104 | * A lot of BIOS fill in 10 (= no distance) everywhere. This messes | |
105 | * up the NUMA heuristics which wants the local node to have a smaller | |
106 | * distance than the others. | |
107 | * Do some quick checks here and only use the SLIT if it passes. | |
108 | */ | |
109 | static __init int slit_valid(struct acpi_table_slit *slit) | |
110 | { | |
111 | int i, j; | |
112 | int d = slit->localities; | |
113 | for (i = 0; i < d; i++) { | |
114 | for (j = 0; j < d; j++) { | |
115 | u8 val = slit->entry[d*i + j]; | |
116 | if (i == j) { | |
117 | if (val != 10) | |
118 | return 0; | |
119 | } else if (val <= 10) | |
120 | return 0; | |
121 | } | |
122 | } | |
123 | return 1; | |
124 | } | |
125 | ||
1da177e4 LT |
126 | /* Callback for SLIT parsing */ |
127 | void __init acpi_numa_slit_init(struct acpi_table_slit *slit) | |
128 | { | |
1584b89c AK |
129 | if (!slit_valid(slit)) { |
130 | printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n"); | |
131 | return; | |
132 | } | |
1da177e4 LT |
133 | acpi_slit = slit; |
134 | } | |
135 | ||
136 | /* Callback for Proximity Domain -> LAPIC mapping */ | |
137 | void __init | |
138 | acpi_numa_processor_affinity_init(struct acpi_table_processor_affinity *pa) | |
139 | { | |
140 | int pxm, node; | |
d22fe808 AK |
141 | if (srat_disabled()) |
142 | return; | |
fad7906d AK |
143 | if (pa->header.length != sizeof(struct acpi_table_processor_affinity)) { |
144 | bad_srat(); | |
d22fe808 AK |
145 | return; |
146 | } | |
147 | if (pa->flags.enabled == 0) | |
1da177e4 LT |
148 | return; |
149 | pxm = pa->proximity_domain; | |
150 | node = setup_node(pxm); | |
151 | if (node < 0) { | |
152 | printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm); | |
153 | bad_srat(); | |
154 | return; | |
155 | } | |
0b07e984 | 156 | apicid_to_node[pa->apic_id] = node; |
1da177e4 | 157 | acpi_numa = 1; |
0b07e984 AK |
158 | printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n", |
159 | pxm, pa->apic_id, node); | |
1da177e4 LT |
160 | } |
161 | ||
68a3a7fe AK |
162 | #ifdef RESERVE_HOTADD |
163 | /* | |
164 | * Protect against too large hotadd areas that would fill up memory. | |
165 | */ | |
166 | static int hotadd_enough_memory(struct bootnode *nd) | |
167 | { | |
168 | static unsigned long allocated; | |
169 | static unsigned long last_area_end; | |
170 | unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT; | |
171 | long mem = pages * sizeof(struct page); | |
172 | unsigned long addr; | |
173 | unsigned long allowed; | |
174 | unsigned long oldpages = pages; | |
175 | ||
176 | if (mem < 0) | |
177 | return 0; | |
178 | allowed = (end_pfn - e820_hole_size(0, end_pfn)) * PAGE_SIZE; | |
179 | allowed = (allowed / 100) * hotadd_percent; | |
180 | if (allocated + mem > allowed) { | |
fad7906d | 181 | unsigned long range; |
68a3a7fe AK |
182 | /* Give them at least part of their hotadd memory upto hotadd_percent |
183 | It would be better to spread the limit out | |
184 | over multiple hotplug areas, but that is too complicated | |
185 | right now */ | |
186 | if (allocated >= allowed) | |
187 | return 0; | |
fad7906d AK |
188 | range = allowed - allocated; |
189 | pages = (range / PAGE_SIZE); | |
68a3a7fe | 190 | mem = pages * sizeof(struct page); |
fad7906d | 191 | nd->end = nd->start + range; |
68a3a7fe AK |
192 | } |
193 | /* Not completely fool proof, but a good sanity check */ | |
194 | addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem); | |
195 | if (addr == -1UL) | |
196 | return 0; | |
197 | if (pages != oldpages) | |
198 | printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n", | |
199 | pages << PAGE_SHIFT); | |
200 | last_area_end = addr + mem; | |
201 | allocated += mem; | |
202 | return 1; | |
203 | } | |
204 | ||
205 | /* | |
206 | * It is fine to add this area to the nodes data it will be used later | |
207 | * This code supports one contigious hot add area per node. | |
208 | */ | |
209 | static int reserve_hotadd(int node, unsigned long start, unsigned long end) | |
210 | { | |
211 | unsigned long s_pfn = start >> PAGE_SHIFT; | |
212 | unsigned long e_pfn = end >> PAGE_SHIFT; | |
213 | int changed = 0; | |
214 | struct bootnode *nd = &nodes_add[node]; | |
215 | ||
216 | /* I had some trouble with strange memory hotadd regions breaking | |
217 | the boot. Be very strict here and reject anything unexpected. | |
218 | If you want working memory hotadd write correct SRATs. | |
219 | ||
220 | The node size check is a basic sanity check to guard against | |
221 | mistakes */ | |
222 | if ((signed long)(end - start) < NODE_MIN_SIZE) { | |
223 | printk(KERN_ERR "SRAT: Hotplug area too small\n"); | |
224 | return -1; | |
225 | } | |
226 | ||
227 | /* This check might be a bit too strict, but I'm keeping it for now. */ | |
228 | if (e820_hole_size(s_pfn, e_pfn) != e_pfn - s_pfn) { | |
229 | printk(KERN_ERR "SRAT: Hotplug area has existing memory\n"); | |
230 | return -1; | |
231 | } | |
232 | ||
233 | if (!hotadd_enough_memory(&nodes_add[node])) { | |
234 | printk(KERN_ERR "SRAT: Hotplug area too large\n"); | |
235 | return -1; | |
236 | } | |
237 | ||
238 | /* Looks good */ | |
239 | ||
240 | found_add_area = 1; | |
241 | if (nd->start == nd->end) { | |
242 | nd->start = start; | |
243 | nd->end = end; | |
244 | changed = 1; | |
245 | } else { | |
246 | if (nd->start == end) { | |
247 | nd->start = start; | |
248 | changed = 1; | |
249 | } | |
250 | if (nd->end == start) { | |
251 | nd->end = end; | |
252 | changed = 1; | |
253 | } | |
254 | if (!changed) | |
255 | printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n"); | |
256 | } | |
257 | ||
258 | if ((nd->end >> PAGE_SHIFT) > end_pfn) | |
259 | end_pfn = nd->end >> PAGE_SHIFT; | |
260 | ||
261 | if (changed) | |
262 | printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end); | |
263 | return 0; | |
264 | } | |
265 | #endif | |
266 | ||
1da177e4 LT |
267 | /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */ |
268 | void __init | |
269 | acpi_numa_memory_affinity_init(struct acpi_table_memory_affinity *ma) | |
270 | { | |
68a3a7fe | 271 | struct bootnode *nd, oldnode; |
1da177e4 LT |
272 | unsigned long start, end; |
273 | int node, pxm; | |
274 | int i; | |
275 | ||
d22fe808 | 276 | if (srat_disabled()) |
1da177e4 | 277 | return; |
d22fe808 AK |
278 | if (ma->header.length != sizeof(struct acpi_table_memory_affinity)) { |
279 | bad_srat(); | |
280 | return; | |
281 | } | |
282 | if (ma->flags.enabled == 0) | |
283 | return; | |
68a3a7fe AK |
284 | if (ma->flags.hot_pluggable && hotadd_percent == 0) |
285 | return; | |
d22fe808 AK |
286 | start = ma->base_addr_lo | ((u64)ma->base_addr_hi << 32); |
287 | end = start + (ma->length_lo | ((u64)ma->length_hi << 32)); | |
1da177e4 LT |
288 | pxm = ma->proximity_domain; |
289 | node = setup_node(pxm); | |
290 | if (node < 0) { | |
291 | printk(KERN_ERR "SRAT: Too many proximity domains.\n"); | |
292 | bad_srat(); | |
293 | return; | |
294 | } | |
1da177e4 | 295 | i = conflicting_nodes(start, end); |
05d1fa4b AK |
296 | if (i == node) { |
297 | printk(KERN_WARNING | |
298 | "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n", | |
299 | pxm, start, end, nodes[i].start, nodes[i].end); | |
300 | } else if (i >= 0) { | |
1da177e4 | 301 | printk(KERN_ERR |
05d1fa4b AK |
302 | "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n", |
303 | pxm, start, end, node_to_pxm(i), | |
304 | nodes[i].start, nodes[i].end); | |
1da177e4 LT |
305 | bad_srat(); |
306 | return; | |
307 | } | |
308 | nd = &nodes[node]; | |
68a3a7fe | 309 | oldnode = *nd; |
1da177e4 LT |
310 | if (!node_test_and_set(node, nodes_parsed)) { |
311 | nd->start = start; | |
312 | nd->end = end; | |
313 | } else { | |
314 | if (start < nd->start) | |
315 | nd->start = start; | |
316 | if (nd->end < end) | |
317 | nd->end = end; | |
318 | } | |
68a3a7fe | 319 | |
1da177e4 LT |
320 | printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm, |
321 | nd->start, nd->end); | |
68a3a7fe AK |
322 | |
323 | #ifdef RESERVE_HOTADD | |
324 | if (ma->flags.hot_pluggable && reserve_hotadd(node, start, end) < 0) { | |
325 | /* Ignore hotadd region. Undo damage */ | |
326 | printk(KERN_NOTICE "SRAT: Hotplug region ignored\n"); | |
327 | *nd = oldnode; | |
328 | if ((nd->start | nd->end) == 0) | |
329 | node_clear(node, nodes_parsed); | |
330 | } | |
331 | #endif | |
1da177e4 LT |
332 | } |
333 | ||
8a6fdd3e AK |
334 | /* Sanity check to catch more bad SRATs (they are amazingly common). |
335 | Make sure the PXMs cover all memory. */ | |
336 | static int nodes_cover_memory(void) | |
337 | { | |
338 | int i; | |
339 | unsigned long pxmram, e820ram; | |
340 | ||
341 | pxmram = 0; | |
342 | for_each_node_mask(i, nodes_parsed) { | |
343 | unsigned long s = nodes[i].start >> PAGE_SHIFT; | |
344 | unsigned long e = nodes[i].end >> PAGE_SHIFT; | |
345 | pxmram += e - s; | |
346 | pxmram -= e820_hole_size(s, e); | |
68a3a7fe AK |
347 | pxmram -= nodes_add[i].end - nodes_add[i].start; |
348 | if ((long)pxmram < 0) | |
349 | pxmram = 0; | |
8a6fdd3e AK |
350 | } |
351 | ||
352 | e820ram = end_pfn - e820_hole_size(0, end_pfn); | |
fdb9df94 AK |
353 | /* We seem to lose 3 pages somewhere. Allow a bit of slack. */ |
354 | if ((long)(e820ram - pxmram) >= 1*1024*1024) { | |
8a6fdd3e AK |
355 | printk(KERN_ERR |
356 | "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n", | |
357 | (pxmram << PAGE_SHIFT) >> 20, | |
358 | (e820ram << PAGE_SHIFT) >> 20); | |
359 | return 0; | |
360 | } | |
361 | return 1; | |
362 | } | |
363 | ||
9391a3f9 AK |
364 | static void unparse_node(int node) |
365 | { | |
366 | int i; | |
367 | node_clear(node, nodes_parsed); | |
368 | for (i = 0; i < MAX_LOCAL_APIC; i++) { | |
369 | if (apicid_to_node[i] == node) | |
370 | apicid_to_node[i] = NUMA_NO_NODE; | |
371 | } | |
372 | } | |
373 | ||
1da177e4 LT |
374 | void __init acpi_numa_arch_fixup(void) {} |
375 | ||
376 | /* Use the information discovered above to actually set up the nodes. */ | |
377 | int __init acpi_scan_nodes(unsigned long start, unsigned long end) | |
378 | { | |
379 | int i; | |
8a6fdd3e | 380 | |
e58e0d03 | 381 | /* First clean up the node list */ |
9391a3f9 | 382 | for (i = 0; i < MAX_NUMNODES; i++) { |
68a3a7fe | 383 | cutoff_node(i, start, end); |
0d015324 | 384 | if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) { |
9391a3f9 | 385 | unparse_node(i); |
0d015324 DY |
386 | node_set_offline(i); |
387 | } | |
e58e0d03 AK |
388 | } |
389 | ||
9391a3f9 AK |
390 | if (acpi_numa <= 0) |
391 | return -1; | |
392 | ||
8a6fdd3e AK |
393 | if (!nodes_cover_memory()) { |
394 | bad_srat(); | |
395 | return -1; | |
396 | } | |
397 | ||
2aed711a | 398 | memnode_shift = compute_hash_shift(nodes, MAX_NUMNODES); |
1da177e4 LT |
399 | if (memnode_shift < 0) { |
400 | printk(KERN_ERR | |
401 | "SRAT: No NUMA node hash function found. Contact maintainer\n"); | |
402 | bad_srat(); | |
403 | return -1; | |
404 | } | |
e58e0d03 AK |
405 | |
406 | /* Finally register nodes */ | |
407 | for_each_node_mask(i, nodes_parsed) | |
1da177e4 | 408 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); |
a8062231 AK |
409 | /* Try again in case setup_node_bootmem missed one due |
410 | to missing bootmem */ | |
411 | for_each_node_mask(i, nodes_parsed) | |
412 | if (!node_online(i)) | |
413 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); | |
414 | ||
1da177e4 LT |
415 | for (i = 0; i < NR_CPUS; i++) { |
416 | if (cpu_to_node[i] == NUMA_NO_NODE) | |
417 | continue; | |
418 | if (!node_isset(cpu_to_node[i], nodes_parsed)) | |
69d81fcd | 419 | numa_set_node(i, NUMA_NO_NODE); |
1da177e4 LT |
420 | } |
421 | numa_init_array(); | |
422 | return 0; | |
423 | } | |
424 | ||
68a3a7fe AK |
425 | void __init srat_reserve_add_area(int nodeid) |
426 | { | |
427 | if (found_add_area && nodes_add[nodeid].end) { | |
428 | u64 total_mb; | |
429 | ||
430 | printk(KERN_INFO "SRAT: Reserving hot-add memory space " | |
431 | "for node %d at %Lx-%Lx\n", | |
432 | nodeid, nodes_add[nodeid].start, nodes_add[nodeid].end); | |
433 | total_mb = (nodes_add[nodeid].end - nodes_add[nodeid].start) | |
434 | >> PAGE_SHIFT; | |
435 | total_mb *= sizeof(struct page); | |
436 | total_mb >>= 20; | |
437 | printk(KERN_INFO "SRAT: This will cost you %Lu MB of " | |
438 | "pre-allocated memory.\n", (unsigned long long)total_mb); | |
439 | reserve_bootmem_node(NODE_DATA(nodeid), nodes_add[nodeid].start, | |
440 | nodes_add[nodeid].end - nodes_add[nodeid].start); | |
441 | } | |
442 | } | |
443 | ||
1da177e4 LT |
444 | int __node_distance(int a, int b) |
445 | { | |
446 | int index; | |
447 | ||
448 | if (!acpi_slit) | |
449 | return a == b ? 10 : 20; | |
450 | index = acpi_slit->localities * node_to_pxm(a); | |
451 | return acpi_slit->entry[index + node_to_pxm(b)]; | |
452 | } | |
453 | ||
454 | EXPORT_SYMBOL(__node_distance); |