Commit | Line | Data |
---|---|---|
e3cfe529 | 1 | /* |
1da177e4 LT |
2 | * Generic VM initialization for x86-64 NUMA setups. |
3 | * Copyright 2002,2003 Andi Kleen, SuSE Labs. | |
e3cfe529 | 4 | */ |
1da177e4 LT |
5 | #include <linux/kernel.h> |
6 | #include <linux/mm.h> | |
7 | #include <linux/string.h> | |
8 | #include <linux/init.h> | |
9 | #include <linux/bootmem.h> | |
10 | #include <linux/mmzone.h> | |
11 | #include <linux/ctype.h> | |
12 | #include <linux/module.h> | |
13 | #include <linux/nodemask.h> | |
3cc87e3f | 14 | #include <linux/sched.h> |
1da177e4 LT |
15 | |
16 | #include <asm/e820.h> | |
17 | #include <asm/proto.h> | |
18 | #include <asm/dma.h> | |
19 | #include <asm/numa.h> | |
20 | #include <asm/acpi.h> | |
c9ff0342 | 21 | #include <asm/k8.h> |
1da177e4 | 22 | |
6c231b7b | 23 | struct pglist_data *node_data[MAX_NUMNODES] __read_mostly; |
e3cfe529 TG |
24 | EXPORT_SYMBOL(node_data); |
25 | ||
dcf36bfa | 26 | struct memnode memnode; |
1da177e4 | 27 | |
43238382 | 28 | s16 apicid_to_node[MAX_LOCAL_APIC] __cpuinitdata = { |
e3cfe529 | 29 | [0 ... MAX_LOCAL_APIC-1] = NUMA_NO_NODE |
3f098c26 | 30 | }; |
e3cfe529 | 31 | |
1da177e4 | 32 | int numa_off __initdata; |
864fc31e TG |
33 | static unsigned long __initdata nodemap_addr; |
34 | static unsigned long __initdata nodemap_size; | |
1da177e4 | 35 | |
6470aff6 BG |
36 | DEFINE_PER_CPU(int, node_number) = 0; |
37 | EXPORT_PER_CPU_SYMBOL(node_number); | |
38 | ||
39 | /* | |
40 | * Map cpu index to node index | |
41 | */ | |
42 | DEFINE_EARLY_PER_CPU(int, x86_cpu_to_node_map, NUMA_NO_NODE); | |
43 | EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_node_map); | |
44 | ||
529a3404 ED |
45 | /* |
46 | * Given a shift value, try to populate memnodemap[] | |
47 | * Returns : | |
48 | * 1 if OK | |
49 | * 0 if memnodmap[] too small (of shift too small) | |
50 | * -1 if node overlap or lost ram (shift too big) | |
51 | */ | |
e3cfe529 | 52 | static int __init populate_memnodemap(const struct bootnode *nodes, |
6ec6e0d9 | 53 | int numnodes, int shift, int *nodeids) |
1da177e4 | 54 | { |
529a3404 | 55 | unsigned long addr, end; |
e3cfe529 | 56 | int i, res = -1; |
b684664f | 57 | |
43238382 | 58 | memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize); |
b684664f | 59 | for (i = 0; i < numnodes; i++) { |
529a3404 ED |
60 | addr = nodes[i].start; |
61 | end = nodes[i].end; | |
62 | if (addr >= end) | |
b684664f | 63 | continue; |
076422d2 | 64 | if ((end >> shift) >= memnodemapsize) |
529a3404 ED |
65 | return 0; |
66 | do { | |
43238382 | 67 | if (memnodemap[addr >> shift] != NUMA_NO_NODE) |
b684664f | 68 | return -1; |
6ec6e0d9 SS |
69 | |
70 | if (!nodeids) | |
71 | memnodemap[addr >> shift] = i; | |
72 | else | |
73 | memnodemap[addr >> shift] = nodeids[i]; | |
74 | ||
076422d2 | 75 | addr += (1UL << shift); |
529a3404 ED |
76 | } while (addr < end); |
77 | res = 1; | |
e3cfe529 | 78 | } |
529a3404 ED |
79 | return res; |
80 | } | |
81 | ||
076422d2 AS |
82 | static int __init allocate_cachealigned_memnodemap(void) |
83 | { | |
24a5da73 | 84 | unsigned long addr; |
076422d2 AS |
85 | |
86 | memnodemap = memnode.embedded_map; | |
316390b0 | 87 | if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map)) |
076422d2 | 88 | return 0; |
076422d2 | 89 | |
24a5da73 | 90 | addr = 0x8000; |
be3e89ee | 91 | nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES); |
c987d12f | 92 | nodemap_addr = find_e820_area(addr, max_pfn<<PAGE_SHIFT, |
24a5da73 | 93 | nodemap_size, L1_CACHE_BYTES); |
076422d2 AS |
94 | if (nodemap_addr == -1UL) { |
95 | printk(KERN_ERR | |
96 | "NUMA: Unable to allocate Memory to Node hash map\n"); | |
97 | nodemap_addr = nodemap_size = 0; | |
98 | return -1; | |
99 | } | |
24a5da73 | 100 | memnodemap = phys_to_virt(nodemap_addr); |
25eff8d4 | 101 | reserve_early(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP"); |
076422d2 AS |
102 | |
103 | printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n", | |
104 | nodemap_addr, nodemap_addr + nodemap_size); | |
105 | return 0; | |
106 | } | |
107 | ||
108 | /* | |
109 | * The LSB of all start and end addresses in the node map is the value of the | |
110 | * maximum possible shift. | |
111 | */ | |
e3cfe529 TG |
112 | static int __init extract_lsb_from_nodes(const struct bootnode *nodes, |
113 | int numnodes) | |
529a3404 | 114 | { |
54413927 | 115 | int i, nodes_used = 0; |
076422d2 AS |
116 | unsigned long start, end; |
117 | unsigned long bitfield = 0, memtop = 0; | |
118 | ||
119 | for (i = 0; i < numnodes; i++) { | |
120 | start = nodes[i].start; | |
121 | end = nodes[i].end; | |
122 | if (start >= end) | |
123 | continue; | |
54413927 AS |
124 | bitfield |= start; |
125 | nodes_used++; | |
076422d2 AS |
126 | if (end > memtop) |
127 | memtop = end; | |
128 | } | |
54413927 AS |
129 | if (nodes_used <= 1) |
130 | i = 63; | |
131 | else | |
132 | i = find_first_bit(&bitfield, sizeof(unsigned long)*8); | |
076422d2 AS |
133 | memnodemapsize = (memtop >> i)+1; |
134 | return i; | |
135 | } | |
529a3404 | 136 | |
6ec6e0d9 SS |
137 | int __init compute_hash_shift(struct bootnode *nodes, int numnodes, |
138 | int *nodeids) | |
076422d2 AS |
139 | { |
140 | int shift; | |
529a3404 | 141 | |
076422d2 AS |
142 | shift = extract_lsb_from_nodes(nodes, numnodes); |
143 | if (allocate_cachealigned_memnodemap()) | |
144 | return -1; | |
6b050f80 | 145 | printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n", |
529a3404 ED |
146 | shift); |
147 | ||
6ec6e0d9 | 148 | if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) { |
e3cfe529 TG |
149 | printk(KERN_INFO "Your memory is not aligned you need to " |
150 | "rebuild your kernel with a bigger NODEMAPSIZE " | |
151 | "shift=%d\n", shift); | |
529a3404 ED |
152 | return -1; |
153 | } | |
b684664f | 154 | return shift; |
1da177e4 LT |
155 | } |
156 | ||
f2dbcfa7 | 157 | int __meminit __early_pfn_to_nid(unsigned long pfn) |
bbfceef4 MT |
158 | { |
159 | return phys_to_nid(pfn << PAGE_SHIFT); | |
160 | } | |
bbfceef4 | 161 | |
e3cfe529 | 162 | static void * __init early_node_mem(int nodeid, unsigned long start, |
24a5da73 YL |
163 | unsigned long end, unsigned long size, |
164 | unsigned long align) | |
a8062231 | 165 | { |
24a5da73 | 166 | unsigned long mem = find_e820_area(start, end, size, align); |
a8062231 | 167 | void *ptr; |
e3cfe529 | 168 | |
9347e0b0 | 169 | if (mem != -1L) |
a8062231 | 170 | return __va(mem); |
9347e0b0 | 171 | |
24a5da73 | 172 | ptr = __alloc_bootmem_nopanic(size, align, __pa(MAX_DMA_ADDRESS)); |
83e83d54 | 173 | if (ptr == NULL) { |
a8062231 | 174 | printk(KERN_ERR "Cannot find %lu bytes in node %d\n", |
e3cfe529 | 175 | size, nodeid); |
a8062231 AK |
176 | return NULL; |
177 | } | |
178 | return ptr; | |
179 | } | |
180 | ||
1da177e4 | 181 | /* Initialize bootmem allocator for a node */ |
e3cfe529 TG |
182 | void __init setup_node_bootmem(int nodeid, unsigned long start, |
183 | unsigned long end) | |
184 | { | |
886533a3 | 185 | unsigned long start_pfn, last_pfn, bootmap_pages, bootmap_size; |
e3cfe529 | 186 | unsigned long bootmap_start, nodedata_phys; |
a8062231 | 187 | void *bootmap; |
be3e89ee | 188 | const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE); |
1a27fc0a | 189 | int nid; |
1da177e4 | 190 | |
4c31e92b YL |
191 | if (!end) |
192 | return; | |
193 | ||
be3e89ee | 194 | start = roundup(start, ZONE_ALIGN); |
1da177e4 | 195 | |
e3cfe529 TG |
196 | printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid, |
197 | start, end); | |
1da177e4 LT |
198 | |
199 | start_pfn = start >> PAGE_SHIFT; | |
886533a3 | 200 | last_pfn = end >> PAGE_SHIFT; |
1da177e4 | 201 | |
24a5da73 YL |
202 | node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size, |
203 | SMP_CACHE_BYTES); | |
a8062231 AK |
204 | if (node_data[nodeid] == NULL) |
205 | return; | |
206 | nodedata_phys = __pa(node_data[nodeid]); | |
6118f76f YL |
207 | printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys, |
208 | nodedata_phys + pgdat_size - 1); | |
1da177e4 | 209 | |
1da177e4 | 210 | memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t)); |
b61bfa3c | 211 | NODE_DATA(nodeid)->bdata = &bootmem_node_data[nodeid]; |
1da177e4 | 212 | NODE_DATA(nodeid)->node_start_pfn = start_pfn; |
886533a3 | 213 | NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn; |
1da177e4 | 214 | |
1a27fc0a YL |
215 | /* |
216 | * Find a place for the bootmem map | |
217 | * nodedata_phys could be on other nodes by alloc_bootmem, | |
218 | * so need to sure bootmap_start not to be small, otherwise | |
219 | * early_node_mem will get that with find_e820_area instead | |
220 | * of alloc_bootmem, that could clash with reserved range | |
221 | */ | |
886533a3 | 222 | bootmap_pages = bootmem_bootmap_pages(last_pfn - start_pfn); |
1a27fc0a YL |
223 | nid = phys_to_nid(nodedata_phys); |
224 | if (nid == nodeid) | |
be3e89ee | 225 | bootmap_start = roundup(nodedata_phys + pgdat_size, PAGE_SIZE); |
1a27fc0a | 226 | else |
be3e89ee | 227 | bootmap_start = roundup(start, PAGE_SIZE); |
24a5da73 | 228 | /* |
e9197bf0 | 229 | * SMP_CACHE_BYTES could be enough, but init_bootmem_node like |
24a5da73 YL |
230 | * to use that to align to PAGE_SIZE |
231 | */ | |
a8062231 | 232 | bootmap = early_node_mem(nodeid, bootmap_start, end, |
24a5da73 | 233 | bootmap_pages<<PAGE_SHIFT, PAGE_SIZE); |
a8062231 AK |
234 | if (bootmap == NULL) { |
235 | if (nodedata_phys < start || nodedata_phys >= end) | |
37bff62e | 236 | free_bootmem(nodedata_phys, pgdat_size); |
a8062231 AK |
237 | node_data[nodeid] = NULL; |
238 | return; | |
239 | } | |
240 | bootmap_start = __pa(bootmap); | |
e3cfe529 | 241 | |
1da177e4 | 242 | bootmap_size = init_bootmem_node(NODE_DATA(nodeid), |
e3cfe529 | 243 | bootmap_start >> PAGE_SHIFT, |
886533a3 | 244 | start_pfn, last_pfn); |
1da177e4 | 245 | |
6118f76f YL |
246 | printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n", |
247 | bootmap_start, bootmap_start + bootmap_size - 1, | |
248 | bootmap_pages); | |
249 | ||
5cb248ab | 250 | free_bootmem_with_active_regions(nodeid, end); |
1da177e4 | 251 | |
1a27fc0a YL |
252 | /* |
253 | * convert early reserve to bootmem reserve earlier | |
254 | * otherwise early_node_mem could use early reserved mem | |
255 | * on previous node | |
256 | */ | |
257 | early_res_to_bootmem(start, end); | |
258 | ||
259 | /* | |
260 | * in some case early_node_mem could use alloc_bootmem | |
261 | * to get range on other node, don't reserve that again | |
262 | */ | |
263 | if (nid != nodeid) | |
264 | printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid); | |
265 | else | |
266 | reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, | |
267 | pgdat_size, BOOTMEM_DEFAULT); | |
268 | nid = phys_to_nid(bootmap_start); | |
269 | if (nid != nodeid) | |
270 | printk(KERN_INFO " bootmap(%d) on node %d\n", nodeid, nid); | |
271 | else | |
272 | reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start, | |
273 | bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT); | |
274 | ||
68a3a7fe AK |
275 | #ifdef CONFIG_ACPI_NUMA |
276 | srat_reserve_add_area(nodeid); | |
277 | #endif | |
1da177e4 | 278 | node_set_online(nodeid); |
e3cfe529 | 279 | } |
1da177e4 | 280 | |
e3cfe529 TG |
281 | /* |
282 | * There are unfortunately some poorly designed mainboards around that | |
283 | * only connect memory to a single CPU. This breaks the 1:1 cpu->node | |
284 | * mapping. To avoid this fill in the mapping for all possible CPUs, | |
285 | * as the number of CPUs is not known yet. We round robin the existing | |
286 | * nodes. | |
287 | */ | |
1da177e4 LT |
288 | void __init numa_init_array(void) |
289 | { | |
290 | int rr, i; | |
e3cfe529 | 291 | |
85cc5135 | 292 | rr = first_node(node_online_map); |
168ef543 | 293 | for (i = 0; i < nr_cpu_ids; i++) { |
1ce35712 | 294 | if (early_cpu_to_node(i) != NUMA_NO_NODE) |
1da177e4 | 295 | continue; |
e3cfe529 | 296 | numa_set_node(i, rr); |
1da177e4 LT |
297 | rr = next_node(rr, node_online_map); |
298 | if (rr == MAX_NUMNODES) | |
299 | rr = first_node(node_online_map); | |
1da177e4 | 300 | } |
1da177e4 LT |
301 | } |
302 | ||
303 | #ifdef CONFIG_NUMA_EMU | |
53fee04f | 304 | /* Numa emulation */ |
864fc31e | 305 | static char *cmdline __initdata; |
1da177e4 | 306 | |
53fee04f | 307 | /* |
e3cfe529 TG |
308 | * Setups up nid to range from addr to addr + size. If the end |
309 | * boundary is greater than max_addr, then max_addr is used instead. | |
310 | * The return value is 0 if there is additional memory left for | |
311 | * allocation past addr and -1 otherwise. addr is adjusted to be at | |
312 | * the end of the node. | |
53fee04f | 313 | */ |
8b8ca80e DR |
314 | static int __init setup_node_range(int nid, struct bootnode *nodes, u64 *addr, |
315 | u64 size, u64 max_addr) | |
53fee04f | 316 | { |
8b8ca80e | 317 | int ret = 0; |
e3cfe529 | 318 | |
8b8ca80e DR |
319 | nodes[nid].start = *addr; |
320 | *addr += size; | |
321 | if (*addr >= max_addr) { | |
322 | *addr = max_addr; | |
323 | ret = -1; | |
324 | } | |
325 | nodes[nid].end = *addr; | |
e3f1caee | 326 | node_set(nid, node_possible_map); |
8b8ca80e DR |
327 | printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid, |
328 | nodes[nid].start, nodes[nid].end, | |
329 | (nodes[nid].end - nodes[nid].start) >> 20); | |
330 | return ret; | |
53fee04f RS |
331 | } |
332 | ||
8b8ca80e DR |
333 | /* |
334 | * Splits num_nodes nodes up equally starting at node_start. The return value | |
335 | * is the number of nodes split up and addr is adjusted to be at the end of the | |
336 | * last node allocated. | |
337 | */ | |
338 | static int __init split_nodes_equally(struct bootnode *nodes, u64 *addr, | |
339 | u64 max_addr, int node_start, | |
340 | int num_nodes) | |
1da177e4 | 341 | { |
8b8ca80e DR |
342 | unsigned int big; |
343 | u64 size; | |
344 | int i; | |
53fee04f | 345 | |
8b8ca80e DR |
346 | if (num_nodes <= 0) |
347 | return -1; | |
348 | if (num_nodes > MAX_NUMNODES) | |
349 | num_nodes = MAX_NUMNODES; | |
a7e96629 | 350 | size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) / |
8b8ca80e | 351 | num_nodes; |
53fee04f | 352 | /* |
8b8ca80e DR |
353 | * Calculate the number of big nodes that can be allocated as a result |
354 | * of consolidating the leftovers. | |
53fee04f | 355 | */ |
8b8ca80e DR |
356 | big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) / |
357 | FAKE_NODE_MIN_SIZE; | |
358 | ||
359 | /* Round down to nearest FAKE_NODE_MIN_SIZE. */ | |
360 | size &= FAKE_NODE_MIN_HASH_MASK; | |
361 | if (!size) { | |
362 | printk(KERN_ERR "Not enough memory for each node. " | |
363 | "NUMA emulation disabled.\n"); | |
364 | return -1; | |
53fee04f | 365 | } |
8b8ca80e DR |
366 | |
367 | for (i = node_start; i < num_nodes + node_start; i++) { | |
368 | u64 end = *addr + size; | |
e3cfe529 | 369 | |
53fee04f RS |
370 | if (i < big) |
371 | end += FAKE_NODE_MIN_SIZE; | |
372 | /* | |
8b8ca80e DR |
373 | * The final node can have the remaining system RAM. Other |
374 | * nodes receive roughly the same amount of available pages. | |
53fee04f | 375 | */ |
8b8ca80e DR |
376 | if (i == num_nodes + node_start - 1) |
377 | end = max_addr; | |
378 | else | |
a7e96629 | 379 | while (end - *addr - e820_hole_size(*addr, end) < |
8b8ca80e DR |
380 | size) { |
381 | end += FAKE_NODE_MIN_SIZE; | |
382 | if (end > max_addr) { | |
383 | end = max_addr; | |
384 | break; | |
385 | } | |
386 | } | |
387 | if (setup_node_range(i, nodes, addr, end - *addr, max_addr) < 0) | |
388 | break; | |
389 | } | |
390 | return i - node_start + 1; | |
391 | } | |
392 | ||
382591d5 DR |
393 | /* |
394 | * Splits the remaining system RAM into chunks of size. The remaining memory is | |
395 | * always assigned to a final node and can be asymmetric. Returns the number of | |
396 | * nodes split. | |
397 | */ | |
398 | static int __init split_nodes_by_size(struct bootnode *nodes, u64 *addr, | |
399 | u64 max_addr, int node_start, u64 size) | |
400 | { | |
401 | int i = node_start; | |
402 | size = (size << 20) & FAKE_NODE_MIN_HASH_MASK; | |
403 | while (!setup_node_range(i++, nodes, addr, size, max_addr)) | |
404 | ; | |
405 | return i - node_start; | |
406 | } | |
407 | ||
8b8ca80e | 408 | /* |
886533a3 | 409 | * Sets up the system RAM area from start_pfn to last_pfn according to the |
8b8ca80e DR |
410 | * numa=fake command-line option. |
411 | */ | |
f46bdf2d MT |
412 | static struct bootnode nodes[MAX_NUMNODES] __initdata; |
413 | ||
886533a3 | 414 | static int __init numa_emulation(unsigned long start_pfn, unsigned long last_pfn) |
8b8ca80e | 415 | { |
e3cfe529 | 416 | u64 size, addr = start_pfn << PAGE_SHIFT; |
886533a3 | 417 | u64 max_addr = last_pfn << PAGE_SHIFT; |
e3cfe529 | 418 | int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i; |
8b8ca80e DR |
419 | |
420 | memset(&nodes, 0, sizeof(nodes)); | |
421 | /* | |
422 | * If the numa=fake command-line is just a single number N, split the | |
423 | * system RAM into N fake nodes. | |
424 | */ | |
425 | if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) { | |
e3cfe529 TG |
426 | long n = simple_strtol(cmdline, NULL, 0); |
427 | ||
428 | num_nodes = split_nodes_equally(nodes, &addr, max_addr, 0, n); | |
8b8ca80e DR |
429 | if (num_nodes < 0) |
430 | return num_nodes; | |
431 | goto out; | |
432 | } | |
433 | ||
434 | /* Parse the command line. */ | |
382591d5 | 435 | for (coeff_flag = 0; ; cmdline++) { |
8b8ca80e DR |
436 | if (*cmdline && isdigit(*cmdline)) { |
437 | num = num * 10 + *cmdline - '0'; | |
438 | continue; | |
53fee04f | 439 | } |
382591d5 DR |
440 | if (*cmdline == '*') { |
441 | if (num > 0) | |
442 | coeff = num; | |
443 | coeff_flag = 1; | |
444 | } | |
8b8ca80e | 445 | if (!*cmdline || *cmdline == ',') { |
382591d5 DR |
446 | if (!coeff_flag) |
447 | coeff = 1; | |
8b8ca80e DR |
448 | /* |
449 | * Round down to the nearest FAKE_NODE_MIN_SIZE. | |
450 | * Command-line coefficients are in megabytes. | |
451 | */ | |
452 | size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK; | |
382591d5 | 453 | if (size) |
8b8ca80e DR |
454 | for (i = 0; i < coeff; i++, num_nodes++) |
455 | if (setup_node_range(num_nodes, nodes, | |
456 | &addr, size, max_addr) < 0) | |
457 | goto done; | |
382591d5 DR |
458 | if (!*cmdline) |
459 | break; | |
460 | coeff_flag = 0; | |
461 | coeff = -1; | |
53fee04f | 462 | } |
8b8ca80e DR |
463 | num = 0; |
464 | } | |
465 | done: | |
466 | if (!num_nodes) | |
467 | return -1; | |
14694d73 | 468 | /* Fill remainder of system RAM, if appropriate. */ |
8b8ca80e | 469 | if (addr < max_addr) { |
382591d5 DR |
470 | if (coeff_flag && coeff < 0) { |
471 | /* Split remaining nodes into num-sized chunks */ | |
472 | num_nodes += split_nodes_by_size(nodes, &addr, max_addr, | |
473 | num_nodes, num); | |
474 | goto out; | |
475 | } | |
14694d73 DR |
476 | switch (*(cmdline - 1)) { |
477 | case '*': | |
478 | /* Split remaining nodes into coeff chunks */ | |
479 | if (coeff <= 0) | |
480 | break; | |
481 | num_nodes += split_nodes_equally(nodes, &addr, max_addr, | |
482 | num_nodes, coeff); | |
483 | break; | |
484 | case ',': | |
485 | /* Do not allocate remaining system RAM */ | |
486 | break; | |
487 | default: | |
488 | /* Give one final node */ | |
489 | setup_node_range(num_nodes, nodes, &addr, | |
490 | max_addr - addr, max_addr); | |
491 | num_nodes++; | |
492 | } | |
8b8ca80e DR |
493 | } |
494 | out: | |
6ec6e0d9 | 495 | memnode_shift = compute_hash_shift(nodes, num_nodes, NULL); |
8b8ca80e DR |
496 | if (memnode_shift < 0) { |
497 | memnode_shift = 0; | |
498 | printk(KERN_ERR "No NUMA hash function found. NUMA emulation " | |
499 | "disabled.\n"); | |
500 | return -1; | |
501 | } | |
502 | ||
503 | /* | |
504 | * We need to vacate all active ranges that may have been registered by | |
1c05f093 DR |
505 | * SRAT and set acpi_numa to -1 so that srat_disabled() always returns |
506 | * true. NUMA emulation has succeeded so we will not scan ACPI nodes. | |
8b8ca80e DR |
507 | */ |
508 | remove_all_active_ranges(); | |
1c05f093 DR |
509 | #ifdef CONFIG_ACPI_NUMA |
510 | acpi_numa = -1; | |
511 | #endif | |
e3f1caee | 512 | for_each_node_mask(i, node_possible_map) { |
5cb248ab MG |
513 | e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT, |
514 | nodes[i].end >> PAGE_SHIFT); | |
e3cfe529 | 515 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); |
5cb248ab | 516 | } |
3484d798 | 517 | acpi_fake_nodes(nodes, num_nodes); |
e3cfe529 TG |
518 | numa_init_array(); |
519 | return 0; | |
1da177e4 | 520 | } |
8b8ca80e | 521 | #endif /* CONFIG_NUMA_EMU */ |
1da177e4 | 522 | |
1f75d7e3 | 523 | void __init initmem_init(unsigned long start_pfn, unsigned long last_pfn) |
e3cfe529 | 524 | { |
1da177e4 LT |
525 | int i; |
526 | ||
e3f1caee | 527 | nodes_clear(node_possible_map); |
b7ad149d | 528 | nodes_clear(node_online_map); |
e3f1caee | 529 | |
1da177e4 | 530 | #ifdef CONFIG_NUMA_EMU |
886533a3 | 531 | if (cmdline && !numa_emulation(start_pfn, last_pfn)) |
e3cfe529 | 532 | return; |
e3f1caee | 533 | nodes_clear(node_possible_map); |
b7ad149d | 534 | nodes_clear(node_online_map); |
1da177e4 LT |
535 | #endif |
536 | ||
537 | #ifdef CONFIG_ACPI_NUMA | |
538 | if (!numa_off && !acpi_scan_nodes(start_pfn << PAGE_SHIFT, | |
886533a3 | 539 | last_pfn << PAGE_SHIFT)) |
e3cfe529 | 540 | return; |
e3f1caee | 541 | nodes_clear(node_possible_map); |
b7ad149d | 542 | nodes_clear(node_online_map); |
1da177e4 LT |
543 | #endif |
544 | ||
545 | #ifdef CONFIG_K8_NUMA | |
e3cfe529 | 546 | if (!numa_off && !k8_scan_nodes(start_pfn<<PAGE_SHIFT, |
886533a3 | 547 | last_pfn<<PAGE_SHIFT)) |
1da177e4 | 548 | return; |
e3f1caee | 549 | nodes_clear(node_possible_map); |
b7ad149d | 550 | nodes_clear(node_online_map); |
1da177e4 LT |
551 | #endif |
552 | printk(KERN_INFO "%s\n", | |
553 | numa_off ? "NUMA turned off" : "No NUMA configuration found"); | |
554 | ||
e3cfe529 | 555 | printk(KERN_INFO "Faking a node at %016lx-%016lx\n", |
1da177e4 | 556 | start_pfn << PAGE_SHIFT, |
886533a3 | 557 | last_pfn << PAGE_SHIFT); |
e3cfe529 TG |
558 | /* setup dummy node covering all memory */ |
559 | memnode_shift = 63; | |
076422d2 | 560 | memnodemap = memnode.embedded_map; |
1da177e4 | 561 | memnodemap[0] = 0; |
1da177e4 | 562 | node_set_online(0); |
e3f1caee | 563 | node_set(0, node_possible_map); |
168ef543 | 564 | for (i = 0; i < nr_cpu_ids; i++) |
69d81fcd | 565 | numa_set_node(i, 0); |
886533a3 TG |
566 | e820_register_active_regions(0, start_pfn, last_pfn); |
567 | setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT); | |
69d81fcd AK |
568 | } |
569 | ||
e3cfe529 TG |
570 | unsigned long __init numa_free_all_bootmem(void) |
571 | { | |
1da177e4 | 572 | unsigned long pages = 0; |
e3cfe529 TG |
573 | int i; |
574 | ||
575 | for_each_online_node(i) | |
1da177e4 | 576 | pages += free_all_bootmem_node(NODE_DATA(i)); |
e3cfe529 | 577 | |
1da177e4 | 578 | return pages; |
e3cfe529 | 579 | } |
1da177e4 LT |
580 | |
581 | void __init paging_init(void) | |
e3cfe529 | 582 | { |
6391af17 | 583 | unsigned long max_zone_pfns[MAX_NR_ZONES]; |
e3cfe529 | 584 | |
6391af17 MG |
585 | memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); |
586 | max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN; | |
587 | max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN; | |
c987d12f | 588 | max_zone_pfns[ZONE_NORMAL] = max_pfn; |
d3ee871e | 589 | |
f0a5a58a BP |
590 | sparse_memory_present_with_active_regions(MAX_NUMNODES); |
591 | sparse_init(); | |
d3ee871e | 592 | |
5cb248ab | 593 | free_area_init_nodes(max_zone_pfns); |
e3cfe529 | 594 | } |
1da177e4 | 595 | |
2c8c0e6b | 596 | static __init int numa_setup(char *opt) |
e3cfe529 | 597 | { |
2c8c0e6b AK |
598 | if (!opt) |
599 | return -EINVAL; | |
e3cfe529 | 600 | if (!strncmp(opt, "off", 3)) |
1da177e4 LT |
601 | numa_off = 1; |
602 | #ifdef CONFIG_NUMA_EMU | |
8b8ca80e DR |
603 | if (!strncmp(opt, "fake=", 5)) |
604 | cmdline = opt + 5; | |
1da177e4 LT |
605 | #endif |
606 | #ifdef CONFIG_ACPI_NUMA | |
e3cfe529 TG |
607 | if (!strncmp(opt, "noacpi", 6)) |
608 | acpi_numa = -1; | |
609 | if (!strncmp(opt, "hotadd=", 7)) | |
68a3a7fe | 610 | hotadd_percent = simple_strtoul(opt+7, NULL, 10); |
1da177e4 | 611 | #endif |
2c8c0e6b | 612 | return 0; |
e3cfe529 | 613 | } |
2c8c0e6b AK |
614 | early_param("numa", numa_setup); |
615 | ||
23ca4bba | 616 | #ifdef CONFIG_NUMA |
05b3cbd8 RT |
617 | /* |
618 | * Setup early cpu_to_node. | |
619 | * | |
620 | * Populate cpu_to_node[] only if x86_cpu_to_apicid[], | |
621 | * and apicid_to_node[] tables have valid entries for a CPU. | |
622 | * This means we skip cpu_to_node[] initialisation for NUMA | |
623 | * emulation and faking node case (when running a kernel compiled | |
624 | * for NUMA on a non NUMA box), which is OK as cpu_to_node[] | |
625 | * is already initialized in a round robin manner at numa_init_array, | |
626 | * prior to this call, and this initialization is good enough | |
627 | * for the fake NUMA cases. | |
23ca4bba MT |
628 | * |
629 | * Called before the per_cpu areas are setup. | |
05b3cbd8 RT |
630 | */ |
631 | void __init init_cpu_to_node(void) | |
632 | { | |
23ca4bba MT |
633 | int cpu; |
634 | u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid); | |
e3cfe529 | 635 | |
23ca4bba MT |
636 | BUG_ON(cpu_to_apicid == NULL); |
637 | ||
638 | for_each_possible_cpu(cpu) { | |
7c9e92b6 | 639 | int node; |
23ca4bba | 640 | u16 apicid = cpu_to_apicid[cpu]; |
e3cfe529 | 641 | |
05b3cbd8 RT |
642 | if (apicid == BAD_APICID) |
643 | continue; | |
7c9e92b6 YL |
644 | node = apicid_to_node[apicid]; |
645 | if (node == NUMA_NO_NODE) | |
05b3cbd8 | 646 | continue; |
7c9e92b6 YL |
647 | if (!node_online(node)) |
648 | continue; | |
23ca4bba | 649 | numa_set_node(cpu, node); |
05b3cbd8 RT |
650 | } |
651 | } | |
23ca4bba | 652 | #endif |
05b3cbd8 | 653 | |
cf050132 | 654 | |
6470aff6 BG |
655 | void __cpuinit numa_set_node(int cpu, int node) |
656 | { | |
657 | int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map); | |
658 | ||
659 | /* early setting, no percpu area yet */ | |
660 | if (cpu_to_node_map) { | |
661 | cpu_to_node_map[cpu] = node; | |
662 | return; | |
663 | } | |
664 | ||
665 | #ifdef CONFIG_DEBUG_PER_CPU_MAPS | |
44581a28 | 666 | if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) { |
6470aff6 BG |
667 | printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu); |
668 | dump_stack(); | |
669 | return; | |
670 | } | |
671 | #endif | |
672 | per_cpu(x86_cpu_to_node_map, cpu) = node; | |
673 | ||
674 | if (node != NUMA_NO_NODE) | |
675 | per_cpu(node_number, cpu) = node; | |
676 | } | |
677 | ||
678 | void __cpuinit numa_clear_node(int cpu) | |
679 | { | |
680 | numa_set_node(cpu, NUMA_NO_NODE); | |
681 | } | |
682 | ||
683 | #ifndef CONFIG_DEBUG_PER_CPU_MAPS | |
684 | ||
685 | void __cpuinit numa_add_cpu(int cpu) | |
686 | { | |
c032ef60 | 687 | cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); |
6470aff6 BG |
688 | } |
689 | ||
690 | void __cpuinit numa_remove_cpu(int cpu) | |
691 | { | |
c032ef60 | 692 | cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); |
6470aff6 BG |
693 | } |
694 | ||
695 | #else /* CONFIG_DEBUG_PER_CPU_MAPS */ | |
696 | ||
697 | /* | |
698 | * --------- debug versions of the numa functions --------- | |
699 | */ | |
700 | static void __cpuinit numa_set_cpumask(int cpu, int enable) | |
701 | { | |
702 | int node = early_cpu_to_node(cpu); | |
73e907de | 703 | struct cpumask *mask; |
6470aff6 BG |
704 | char buf[64]; |
705 | ||
c032ef60 RR |
706 | mask = node_to_cpumask_map[node]; |
707 | if (mask == NULL) { | |
708 | printk(KERN_ERR "node_to_cpumask_map[%i] NULL\n", node); | |
6470aff6 BG |
709 | dump_stack(); |
710 | return; | |
711 | } | |
712 | ||
6470aff6 | 713 | if (enable) |
c032ef60 | 714 | cpumask_set_cpu(cpu, mask); |
6470aff6 | 715 | else |
c032ef60 | 716 | cpumask_clear_cpu(cpu, mask); |
6470aff6 BG |
717 | |
718 | cpulist_scnprintf(buf, sizeof(buf), mask); | |
719 | printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n", | |
720 | enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf); | |
721 | } | |
722 | ||
723 | void __cpuinit numa_add_cpu(int cpu) | |
724 | { | |
725 | numa_set_cpumask(cpu, 1); | |
726 | } | |
727 | ||
728 | void __cpuinit numa_remove_cpu(int cpu) | |
729 | { | |
730 | numa_set_cpumask(cpu, 0); | |
731 | } | |
732 | ||
733 | int cpu_to_node(int cpu) | |
734 | { | |
735 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) { | |
736 | printk(KERN_WARNING | |
737 | "cpu_to_node(%d): usage too early!\n", cpu); | |
738 | dump_stack(); | |
739 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
740 | } | |
741 | return per_cpu(x86_cpu_to_node_map, cpu); | |
742 | } | |
743 | EXPORT_SYMBOL(cpu_to_node); | |
744 | ||
745 | /* | |
746 | * Same function as cpu_to_node() but used if called before the | |
747 | * per_cpu areas are setup. | |
748 | */ | |
749 | int early_cpu_to_node(int cpu) | |
750 | { | |
751 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) | |
752 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
753 | ||
44581a28 | 754 | if (!cpu_possible(cpu)) { |
6470aff6 BG |
755 | printk(KERN_WARNING |
756 | "early_cpu_to_node(%d): no per_cpu area!\n", cpu); | |
757 | dump_stack(); | |
758 | return NUMA_NO_NODE; | |
759 | } | |
760 | return per_cpu(x86_cpu_to_node_map, cpu); | |
761 | } | |
762 | ||
6470aff6 BG |
763 | /* |
764 | * --------- end of debug versions of the numa functions --------- | |
765 | */ | |
766 | ||
767 | #endif /* CONFIG_DEBUG_PER_CPU_MAPS */ |