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 | |
68a3a7fe AK |
24 | #if (defined(CONFIG_ACPI_HOTPLUG_MEMORY) || \ |
25 | defined(CONFIG_ACPI_HOTPLUG_MEMORY_MODULE)) \ | |
26 | && !defined(CONFIG_MEMORY_HOTPLUG) | |
27 | #define RESERVE_HOTADD 1 | |
28 | #endif | |
29 | ||
1da177e4 LT |
30 | static struct acpi_table_slit *acpi_slit; |
31 | ||
32 | static nodemask_t nodes_parsed __initdata; | |
33 | static nodemask_t nodes_found __initdata; | |
abe059e7 | 34 | static struct bootnode nodes[MAX_NUMNODES] __initdata; |
68a3a7fe AK |
35 | static struct bootnode nodes_add[MAX_NUMNODES] __initdata; |
36 | static int found_add_area __initdata; | |
37 | int hotadd_percent __initdata = 10; | |
e4e94072 | 38 | static u8 pxm2node[256] = { [0 ... 255] = 0xff }; |
1da177e4 | 39 | |
9391a3f9 AK |
40 | /* Too small nodes confuse the VM badly. Usually they result |
41 | from BIOS bugs. */ | |
42 | #define NODE_MIN_SIZE (4*1024*1024) | |
43 | ||
05d1fa4b AK |
44 | static int node_to_pxm(int n); |
45 | ||
69e1a33f AK |
46 | int pxm_to_node(int pxm) |
47 | { | |
48 | if ((unsigned)pxm >= 256) | |
e4e94072 AK |
49 | return -1; |
50 | /* Extend 0xff to (int)-1 */ | |
51 | return (signed char)pxm2node[pxm]; | |
69e1a33f AK |
52 | } |
53 | ||
1da177e4 LT |
54 | static __init int setup_node(int pxm) |
55 | { | |
56 | unsigned node = pxm2node[pxm]; | |
57 | if (node == 0xff) { | |
58 | if (nodes_weight(nodes_found) >= MAX_NUMNODES) | |
59 | return -1; | |
60 | node = first_unset_node(nodes_found); | |
61 | node_set(node, nodes_found); | |
62 | pxm2node[pxm] = node; | |
63 | } | |
64 | return pxm2node[pxm]; | |
65 | } | |
66 | ||
67 | static __init int conflicting_nodes(unsigned long start, unsigned long end) | |
68 | { | |
69 | int i; | |
4b6a455c | 70 | for_each_node_mask(i, nodes_parsed) { |
abe059e7 | 71 | struct bootnode *nd = &nodes[i]; |
1da177e4 LT |
72 | if (nd->start == nd->end) |
73 | continue; | |
74 | if (nd->end > start && nd->start < end) | |
05d1fa4b | 75 | return i; |
1da177e4 | 76 | if (nd->end == end && nd->start == start) |
05d1fa4b | 77 | return i; |
1da177e4 LT |
78 | } |
79 | return -1; | |
80 | } | |
81 | ||
82 | static __init void cutoff_node(int i, unsigned long start, unsigned long end) | |
83 | { | |
abe059e7 | 84 | struct bootnode *nd = &nodes[i]; |
68a3a7fe AK |
85 | |
86 | if (found_add_area) | |
87 | return; | |
88 | ||
1da177e4 LT |
89 | if (nd->start < start) { |
90 | nd->start = start; | |
91 | if (nd->end < nd->start) | |
92 | nd->start = nd->end; | |
93 | } | |
94 | if (nd->end > end) { | |
1da177e4 LT |
95 | nd->end = end; |
96 | if (nd->start > nd->end) | |
97 | nd->start = nd->end; | |
98 | } | |
99 | } | |
100 | ||
101 | static __init void bad_srat(void) | |
102 | { | |
2bce2b54 | 103 | int i; |
1da177e4 LT |
104 | printk(KERN_ERR "SRAT: SRAT not used.\n"); |
105 | acpi_numa = -1; | |
2bce2b54 AK |
106 | for (i = 0; i < MAX_LOCAL_APIC; i++) |
107 | apicid_to_node[i] = NUMA_NO_NODE; | |
68a3a7fe AK |
108 | for (i = 0; i < MAX_NUMNODES; i++) |
109 | nodes_add[i].start = nodes[i].end = 0; | |
1da177e4 LT |
110 | } |
111 | ||
112 | static __init inline int srat_disabled(void) | |
113 | { | |
114 | return numa_off || acpi_numa < 0; | |
115 | } | |
116 | ||
1584b89c AK |
117 | /* |
118 | * A lot of BIOS fill in 10 (= no distance) everywhere. This messes | |
119 | * up the NUMA heuristics which wants the local node to have a smaller | |
120 | * distance than the others. | |
121 | * Do some quick checks here and only use the SLIT if it passes. | |
122 | */ | |
123 | static __init int slit_valid(struct acpi_table_slit *slit) | |
124 | { | |
125 | int i, j; | |
126 | int d = slit->localities; | |
127 | for (i = 0; i < d; i++) { | |
128 | for (j = 0; j < d; j++) { | |
129 | u8 val = slit->entry[d*i + j]; | |
130 | if (i == j) { | |
131 | if (val != 10) | |
132 | return 0; | |
133 | } else if (val <= 10) | |
134 | return 0; | |
135 | } | |
136 | } | |
137 | return 1; | |
138 | } | |
139 | ||
1da177e4 LT |
140 | /* Callback for SLIT parsing */ |
141 | void __init acpi_numa_slit_init(struct acpi_table_slit *slit) | |
142 | { | |
1584b89c AK |
143 | if (!slit_valid(slit)) { |
144 | printk(KERN_INFO "ACPI: SLIT table looks invalid. Not used.\n"); | |
145 | return; | |
146 | } | |
1da177e4 LT |
147 | acpi_slit = slit; |
148 | } | |
149 | ||
150 | /* Callback for Proximity Domain -> LAPIC mapping */ | |
151 | void __init | |
152 | acpi_numa_processor_affinity_init(struct acpi_table_processor_affinity *pa) | |
153 | { | |
154 | int pxm, node; | |
d22fe808 AK |
155 | if (srat_disabled()) |
156 | return; | |
157 | if (pa->header.length != sizeof(struct acpi_table_processor_affinity)) { bad_srat(); | |
158 | return; | |
159 | } | |
160 | if (pa->flags.enabled == 0) | |
1da177e4 LT |
161 | return; |
162 | pxm = pa->proximity_domain; | |
163 | node = setup_node(pxm); | |
164 | if (node < 0) { | |
165 | printk(KERN_ERR "SRAT: Too many proximity domains %x\n", pxm); | |
166 | bad_srat(); | |
167 | return; | |
168 | } | |
0b07e984 | 169 | apicid_to_node[pa->apic_id] = node; |
1da177e4 | 170 | acpi_numa = 1; |
0b07e984 AK |
171 | printk(KERN_INFO "SRAT: PXM %u -> APIC %u -> Node %u\n", |
172 | pxm, pa->apic_id, node); | |
1da177e4 LT |
173 | } |
174 | ||
68a3a7fe AK |
175 | #ifdef RESERVE_HOTADD |
176 | /* | |
177 | * Protect against too large hotadd areas that would fill up memory. | |
178 | */ | |
179 | static int hotadd_enough_memory(struct bootnode *nd) | |
180 | { | |
181 | static unsigned long allocated; | |
182 | static unsigned long last_area_end; | |
183 | unsigned long pages = (nd->end - nd->start) >> PAGE_SHIFT; | |
184 | long mem = pages * sizeof(struct page); | |
185 | unsigned long addr; | |
186 | unsigned long allowed; | |
187 | unsigned long oldpages = pages; | |
188 | ||
189 | if (mem < 0) | |
190 | return 0; | |
191 | allowed = (end_pfn - e820_hole_size(0, end_pfn)) * PAGE_SIZE; | |
192 | allowed = (allowed / 100) * hotadd_percent; | |
193 | if (allocated + mem > allowed) { | |
194 | /* Give them at least part of their hotadd memory upto hotadd_percent | |
195 | It would be better to spread the limit out | |
196 | over multiple hotplug areas, but that is too complicated | |
197 | right now */ | |
198 | if (allocated >= allowed) | |
199 | return 0; | |
200 | pages = (allowed - allocated + mem) / sizeof(struct page); | |
201 | mem = pages * sizeof(struct page); | |
202 | nd->end = nd->start + pages*PAGE_SIZE; | |
203 | } | |
204 | /* Not completely fool proof, but a good sanity check */ | |
205 | addr = find_e820_area(last_area_end, end_pfn<<PAGE_SHIFT, mem); | |
206 | if (addr == -1UL) | |
207 | return 0; | |
208 | if (pages != oldpages) | |
209 | printk(KERN_NOTICE "SRAT: Hotadd area limited to %lu bytes\n", | |
210 | pages << PAGE_SHIFT); | |
211 | last_area_end = addr + mem; | |
212 | allocated += mem; | |
213 | return 1; | |
214 | } | |
215 | ||
216 | /* | |
217 | * It is fine to add this area to the nodes data it will be used later | |
218 | * This code supports one contigious hot add area per node. | |
219 | */ | |
220 | static int reserve_hotadd(int node, unsigned long start, unsigned long end) | |
221 | { | |
222 | unsigned long s_pfn = start >> PAGE_SHIFT; | |
223 | unsigned long e_pfn = end >> PAGE_SHIFT; | |
224 | int changed = 0; | |
225 | struct bootnode *nd = &nodes_add[node]; | |
226 | ||
227 | /* I had some trouble with strange memory hotadd regions breaking | |
228 | the boot. Be very strict here and reject anything unexpected. | |
229 | If you want working memory hotadd write correct SRATs. | |
230 | ||
231 | The node size check is a basic sanity check to guard against | |
232 | mistakes */ | |
233 | if ((signed long)(end - start) < NODE_MIN_SIZE) { | |
234 | printk(KERN_ERR "SRAT: Hotplug area too small\n"); | |
235 | return -1; | |
236 | } | |
237 | ||
238 | /* This check might be a bit too strict, but I'm keeping it for now. */ | |
239 | if (e820_hole_size(s_pfn, e_pfn) != e_pfn - s_pfn) { | |
240 | printk(KERN_ERR "SRAT: Hotplug area has existing memory\n"); | |
241 | return -1; | |
242 | } | |
243 | ||
244 | if (!hotadd_enough_memory(&nodes_add[node])) { | |
245 | printk(KERN_ERR "SRAT: Hotplug area too large\n"); | |
246 | return -1; | |
247 | } | |
248 | ||
249 | /* Looks good */ | |
250 | ||
251 | found_add_area = 1; | |
252 | if (nd->start == nd->end) { | |
253 | nd->start = start; | |
254 | nd->end = end; | |
255 | changed = 1; | |
256 | } else { | |
257 | if (nd->start == end) { | |
258 | nd->start = start; | |
259 | changed = 1; | |
260 | } | |
261 | if (nd->end == start) { | |
262 | nd->end = end; | |
263 | changed = 1; | |
264 | } | |
265 | if (!changed) | |
266 | printk(KERN_ERR "SRAT: Hotplug zone not continuous. Partly ignored\n"); | |
267 | } | |
268 | ||
269 | if ((nd->end >> PAGE_SHIFT) > end_pfn) | |
270 | end_pfn = nd->end >> PAGE_SHIFT; | |
271 | ||
272 | if (changed) | |
273 | printk(KERN_INFO "SRAT: hot plug zone found %Lx - %Lx\n", nd->start, nd->end); | |
274 | return 0; | |
275 | } | |
276 | #endif | |
277 | ||
1da177e4 LT |
278 | /* Callback for parsing of the Proximity Domain <-> Memory Area mappings */ |
279 | void __init | |
280 | acpi_numa_memory_affinity_init(struct acpi_table_memory_affinity *ma) | |
281 | { | |
68a3a7fe | 282 | struct bootnode *nd, oldnode; |
1da177e4 LT |
283 | unsigned long start, end; |
284 | int node, pxm; | |
285 | int i; | |
286 | ||
d22fe808 | 287 | if (srat_disabled()) |
1da177e4 | 288 | return; |
d22fe808 AK |
289 | if (ma->header.length != sizeof(struct acpi_table_memory_affinity)) { |
290 | bad_srat(); | |
291 | return; | |
292 | } | |
293 | if (ma->flags.enabled == 0) | |
294 | return; | |
68a3a7fe AK |
295 | if (ma->flags.hot_pluggable && hotadd_percent == 0) |
296 | return; | |
d22fe808 AK |
297 | start = ma->base_addr_lo | ((u64)ma->base_addr_hi << 32); |
298 | end = start + (ma->length_lo | ((u64)ma->length_hi << 32)); | |
1da177e4 LT |
299 | pxm = ma->proximity_domain; |
300 | node = setup_node(pxm); | |
301 | if (node < 0) { | |
302 | printk(KERN_ERR "SRAT: Too many proximity domains.\n"); | |
303 | bad_srat(); | |
304 | return; | |
305 | } | |
1da177e4 | 306 | i = conflicting_nodes(start, end); |
05d1fa4b AK |
307 | if (i == node) { |
308 | printk(KERN_WARNING | |
309 | "SRAT: Warning: PXM %d (%lx-%lx) overlaps with itself (%Lx-%Lx)\n", | |
310 | pxm, start, end, nodes[i].start, nodes[i].end); | |
311 | } else if (i >= 0) { | |
1da177e4 | 312 | printk(KERN_ERR |
05d1fa4b AK |
313 | "SRAT: PXM %d (%lx-%lx) overlaps with PXM %d (%Lx-%Lx)\n", |
314 | pxm, start, end, node_to_pxm(i), | |
315 | nodes[i].start, nodes[i].end); | |
1da177e4 LT |
316 | bad_srat(); |
317 | return; | |
318 | } | |
319 | nd = &nodes[node]; | |
68a3a7fe | 320 | oldnode = *nd; |
1da177e4 LT |
321 | if (!node_test_and_set(node, nodes_parsed)) { |
322 | nd->start = start; | |
323 | nd->end = end; | |
324 | } else { | |
325 | if (start < nd->start) | |
326 | nd->start = start; | |
327 | if (nd->end < end) | |
328 | nd->end = end; | |
329 | } | |
68a3a7fe | 330 | |
1da177e4 LT |
331 | printk(KERN_INFO "SRAT: Node %u PXM %u %Lx-%Lx\n", node, pxm, |
332 | nd->start, nd->end); | |
68a3a7fe AK |
333 | |
334 | #ifdef RESERVE_HOTADD | |
335 | if (ma->flags.hot_pluggable && reserve_hotadd(node, start, end) < 0) { | |
336 | /* Ignore hotadd region. Undo damage */ | |
337 | printk(KERN_NOTICE "SRAT: Hotplug region ignored\n"); | |
338 | *nd = oldnode; | |
339 | if ((nd->start | nd->end) == 0) | |
340 | node_clear(node, nodes_parsed); | |
341 | } | |
342 | #endif | |
1da177e4 LT |
343 | } |
344 | ||
8a6fdd3e AK |
345 | /* Sanity check to catch more bad SRATs (they are amazingly common). |
346 | Make sure the PXMs cover all memory. */ | |
347 | static int nodes_cover_memory(void) | |
348 | { | |
349 | int i; | |
350 | unsigned long pxmram, e820ram; | |
351 | ||
352 | pxmram = 0; | |
353 | for_each_node_mask(i, nodes_parsed) { | |
354 | unsigned long s = nodes[i].start >> PAGE_SHIFT; | |
355 | unsigned long e = nodes[i].end >> PAGE_SHIFT; | |
356 | pxmram += e - s; | |
357 | pxmram -= e820_hole_size(s, e); | |
68a3a7fe AK |
358 | pxmram -= nodes_add[i].end - nodes_add[i].start; |
359 | if ((long)pxmram < 0) | |
360 | pxmram = 0; | |
8a6fdd3e AK |
361 | } |
362 | ||
363 | e820ram = end_pfn - e820_hole_size(0, end_pfn); | |
fdb9df94 AK |
364 | /* We seem to lose 3 pages somewhere. Allow a bit of slack. */ |
365 | if ((long)(e820ram - pxmram) >= 1*1024*1024) { | |
8a6fdd3e AK |
366 | printk(KERN_ERR |
367 | "SRAT: PXMs only cover %luMB of your %luMB e820 RAM. Not used.\n", | |
368 | (pxmram << PAGE_SHIFT) >> 20, | |
369 | (e820ram << PAGE_SHIFT) >> 20); | |
370 | return 0; | |
371 | } | |
372 | return 1; | |
373 | } | |
374 | ||
9391a3f9 AK |
375 | static void unparse_node(int node) |
376 | { | |
377 | int i; | |
378 | node_clear(node, nodes_parsed); | |
379 | for (i = 0; i < MAX_LOCAL_APIC; i++) { | |
380 | if (apicid_to_node[i] == node) | |
381 | apicid_to_node[i] = NUMA_NO_NODE; | |
382 | } | |
383 | } | |
384 | ||
1da177e4 LT |
385 | void __init acpi_numa_arch_fixup(void) {} |
386 | ||
387 | /* Use the information discovered above to actually set up the nodes. */ | |
388 | int __init acpi_scan_nodes(unsigned long start, unsigned long end) | |
389 | { | |
390 | int i; | |
8a6fdd3e | 391 | |
e58e0d03 | 392 | /* First clean up the node list */ |
9391a3f9 | 393 | for (i = 0; i < MAX_NUMNODES; i++) { |
68a3a7fe | 394 | cutoff_node(i, start, end); |
9391a3f9 AK |
395 | if ((nodes[i].end - nodes[i].start) < NODE_MIN_SIZE) |
396 | unparse_node(i); | |
e58e0d03 AK |
397 | } |
398 | ||
9391a3f9 AK |
399 | if (acpi_numa <= 0) |
400 | return -1; | |
401 | ||
8a6fdd3e AK |
402 | if (!nodes_cover_memory()) { |
403 | bad_srat(); | |
404 | return -1; | |
405 | } | |
406 | ||
2aed711a | 407 | memnode_shift = compute_hash_shift(nodes, MAX_NUMNODES); |
1da177e4 LT |
408 | if (memnode_shift < 0) { |
409 | printk(KERN_ERR | |
410 | "SRAT: No NUMA node hash function found. Contact maintainer\n"); | |
411 | bad_srat(); | |
412 | return -1; | |
413 | } | |
e58e0d03 AK |
414 | |
415 | /* Finally register nodes */ | |
416 | for_each_node_mask(i, nodes_parsed) | |
1da177e4 | 417 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); |
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 | ||
05d1fa4b | 428 | static int node_to_pxm(int n) |
1da177e4 LT |
429 | { |
430 | int i; | |
431 | if (pxm2node[n] == n) | |
432 | return n; | |
433 | for (i = 0; i < 256; i++) | |
434 | if (pxm2node[i] == n) | |
435 | return i; | |
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); |