]>
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 */ |
7c43769a YL |
182 | void __init |
183 | setup_node_bootmem(int nodeid, unsigned long start, unsigned long end) | |
e3cfe529 | 184 | { |
886533a3 | 185 | unsigned long start_pfn, last_pfn, bootmap_pages, bootmap_size; |
7c43769a | 186 | const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE); |
e3cfe529 | 187 | unsigned long bootmap_start, nodedata_phys; |
a8062231 | 188 | void *bootmap; |
1a27fc0a | 189 | int nid; |
1da177e4 | 190 | |
4c31e92b YL |
191 | if (!end) |
192 | return; | |
193 | ||
7c43769a YL |
194 | /* |
195 | * Don't confuse VM with a node that doesn't have the | |
196 | * minimum amount of memory: | |
197 | */ | |
198 | if (end && (end - start) < NODE_MIN_SIZE) | |
199 | return; | |
200 | ||
be3e89ee | 201 | start = roundup(start, ZONE_ALIGN); |
1da177e4 | 202 | |
e3cfe529 TG |
203 | printk(KERN_INFO "Bootmem setup node %d %016lx-%016lx\n", nodeid, |
204 | start, end); | |
1da177e4 LT |
205 | |
206 | start_pfn = start >> PAGE_SHIFT; | |
886533a3 | 207 | last_pfn = end >> PAGE_SHIFT; |
1da177e4 | 208 | |
24a5da73 YL |
209 | node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size, |
210 | SMP_CACHE_BYTES); | |
a8062231 AK |
211 | if (node_data[nodeid] == NULL) |
212 | return; | |
213 | nodedata_phys = __pa(node_data[nodeid]); | |
6118f76f YL |
214 | printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys, |
215 | nodedata_phys + pgdat_size - 1); | |
1da177e4 | 216 | |
1da177e4 | 217 | memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t)); |
b61bfa3c | 218 | NODE_DATA(nodeid)->bdata = &bootmem_node_data[nodeid]; |
1da177e4 | 219 | NODE_DATA(nodeid)->node_start_pfn = start_pfn; |
886533a3 | 220 | NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn; |
1da177e4 | 221 | |
1a27fc0a YL |
222 | /* |
223 | * Find a place for the bootmem map | |
224 | * nodedata_phys could be on other nodes by alloc_bootmem, | |
225 | * so need to sure bootmap_start not to be small, otherwise | |
226 | * early_node_mem will get that with find_e820_area instead | |
227 | * of alloc_bootmem, that could clash with reserved range | |
228 | */ | |
886533a3 | 229 | bootmap_pages = bootmem_bootmap_pages(last_pfn - start_pfn); |
1a27fc0a YL |
230 | nid = phys_to_nid(nodedata_phys); |
231 | if (nid == nodeid) | |
be3e89ee | 232 | bootmap_start = roundup(nodedata_phys + pgdat_size, PAGE_SIZE); |
1a27fc0a | 233 | else |
be3e89ee | 234 | bootmap_start = roundup(start, PAGE_SIZE); |
24a5da73 | 235 | /* |
e9197bf0 | 236 | * SMP_CACHE_BYTES could be enough, but init_bootmem_node like |
24a5da73 YL |
237 | * to use that to align to PAGE_SIZE |
238 | */ | |
a8062231 | 239 | bootmap = early_node_mem(nodeid, bootmap_start, end, |
24a5da73 | 240 | bootmap_pages<<PAGE_SHIFT, PAGE_SIZE); |
a8062231 | 241 | if (bootmap == NULL) { |
021428ad YL |
242 | if (nodedata_phys < start || nodedata_phys >= end) { |
243 | /* | |
244 | * only need to free it if it is from other node | |
245 | * bootmem | |
246 | */ | |
247 | if (nid != nodeid) | |
248 | free_bootmem(nodedata_phys, pgdat_size); | |
249 | } | |
a8062231 AK |
250 | node_data[nodeid] = NULL; |
251 | return; | |
252 | } | |
253 | bootmap_start = __pa(bootmap); | |
e3cfe529 | 254 | |
1da177e4 | 255 | bootmap_size = init_bootmem_node(NODE_DATA(nodeid), |
e3cfe529 | 256 | bootmap_start >> PAGE_SHIFT, |
886533a3 | 257 | start_pfn, last_pfn); |
1da177e4 | 258 | |
6118f76f YL |
259 | printk(KERN_INFO " bootmap [%016lx - %016lx] pages %lx\n", |
260 | bootmap_start, bootmap_start + bootmap_size - 1, | |
261 | bootmap_pages); | |
262 | ||
5cb248ab | 263 | free_bootmem_with_active_regions(nodeid, end); |
1da177e4 | 264 | |
1a27fc0a YL |
265 | /* |
266 | * convert early reserve to bootmem reserve earlier | |
267 | * otherwise early_node_mem could use early reserved mem | |
268 | * on previous node | |
269 | */ | |
270 | early_res_to_bootmem(start, end); | |
271 | ||
272 | /* | |
273 | * in some case early_node_mem could use alloc_bootmem | |
274 | * to get range on other node, don't reserve that again | |
275 | */ | |
276 | if (nid != nodeid) | |
277 | printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid); | |
278 | else | |
279 | reserve_bootmem_node(NODE_DATA(nodeid), nodedata_phys, | |
280 | pgdat_size, BOOTMEM_DEFAULT); | |
281 | nid = phys_to_nid(bootmap_start); | |
282 | if (nid != nodeid) | |
283 | printk(KERN_INFO " bootmap(%d) on node %d\n", nodeid, nid); | |
284 | else | |
285 | reserve_bootmem_node(NODE_DATA(nodeid), bootmap_start, | |
286 | bootmap_pages<<PAGE_SHIFT, BOOTMEM_DEFAULT); | |
287 | ||
1da177e4 | 288 | node_set_online(nodeid); |
e3cfe529 | 289 | } |
1da177e4 | 290 | |
e3cfe529 TG |
291 | /* |
292 | * There are unfortunately some poorly designed mainboards around that | |
293 | * only connect memory to a single CPU. This breaks the 1:1 cpu->node | |
294 | * mapping. To avoid this fill in the mapping for all possible CPUs, | |
295 | * as the number of CPUs is not known yet. We round robin the existing | |
296 | * nodes. | |
297 | */ | |
1da177e4 LT |
298 | void __init numa_init_array(void) |
299 | { | |
300 | int rr, i; | |
e3cfe529 | 301 | |
85cc5135 | 302 | rr = first_node(node_online_map); |
168ef543 | 303 | for (i = 0; i < nr_cpu_ids; i++) { |
1ce35712 | 304 | if (early_cpu_to_node(i) != NUMA_NO_NODE) |
1da177e4 | 305 | continue; |
e3cfe529 | 306 | numa_set_node(i, rr); |
1da177e4 LT |
307 | rr = next_node(rr, node_online_map); |
308 | if (rr == MAX_NUMNODES) | |
309 | rr = first_node(node_online_map); | |
1da177e4 | 310 | } |
1da177e4 LT |
311 | } |
312 | ||
313 | #ifdef CONFIG_NUMA_EMU | |
53fee04f | 314 | /* Numa emulation */ |
adc19389 DR |
315 | static struct bootnode nodes[MAX_NUMNODES] __initdata; |
316 | static struct bootnode physnodes[MAX_NUMNODES] __initdata; | |
864fc31e | 317 | static char *cmdline __initdata; |
1da177e4 | 318 | |
adc19389 DR |
319 | static int __init setup_physnodes(unsigned long start, unsigned long end, |
320 | int acpi, int k8) | |
321 | { | |
322 | int nr_nodes = 0; | |
323 | int ret = 0; | |
324 | int i; | |
325 | ||
326 | #ifdef CONFIG_ACPI_NUMA | |
327 | if (acpi) | |
328 | nr_nodes = acpi_get_nodes(physnodes); | |
329 | #endif | |
330 | #ifdef CONFIG_K8_NUMA | |
331 | if (k8) | |
332 | nr_nodes = k8_get_nodes(physnodes); | |
333 | #endif | |
334 | /* | |
335 | * Basic sanity checking on the physical node map: there may be errors | |
336 | * if the SRAT or K8 incorrectly reported the topology or the mem= | |
337 | * kernel parameter is used. | |
338 | */ | |
339 | for (i = 0; i < nr_nodes; i++) { | |
340 | if (physnodes[i].start == physnodes[i].end) | |
341 | continue; | |
342 | if (physnodes[i].start > end) { | |
343 | physnodes[i].end = physnodes[i].start; | |
344 | continue; | |
345 | } | |
346 | if (physnodes[i].end < start) { | |
347 | physnodes[i].start = physnodes[i].end; | |
348 | continue; | |
349 | } | |
350 | if (physnodes[i].start < start) | |
351 | physnodes[i].start = start; | |
352 | if (physnodes[i].end > end) | |
353 | physnodes[i].end = end; | |
354 | } | |
355 | ||
356 | /* | |
357 | * Remove all nodes that have no memory or were truncated because of the | |
358 | * limited address range. | |
359 | */ | |
360 | for (i = 0; i < nr_nodes; i++) { | |
361 | if (physnodes[i].start == physnodes[i].end) | |
362 | continue; | |
363 | physnodes[ret].start = physnodes[i].start; | |
364 | physnodes[ret].end = physnodes[i].end; | |
365 | ret++; | |
366 | } | |
367 | ||
368 | /* | |
369 | * If no physical topology was detected, a single node is faked to cover | |
370 | * the entire address space. | |
371 | */ | |
372 | if (!ret) { | |
373 | physnodes[ret].start = start; | |
374 | physnodes[ret].end = end; | |
375 | ret = 1; | |
376 | } | |
377 | return ret; | |
378 | } | |
379 | ||
53fee04f | 380 | /* |
e3cfe529 TG |
381 | * Setups up nid to range from addr to addr + size. If the end |
382 | * boundary is greater than max_addr, then max_addr is used instead. | |
383 | * The return value is 0 if there is additional memory left for | |
384 | * allocation past addr and -1 otherwise. addr is adjusted to be at | |
385 | * the end of the node. | |
53fee04f | 386 | */ |
adc19389 | 387 | static int __init setup_node_range(int nid, u64 *addr, u64 size, u64 max_addr) |
53fee04f | 388 | { |
8b8ca80e DR |
389 | int ret = 0; |
390 | nodes[nid].start = *addr; | |
391 | *addr += size; | |
392 | if (*addr >= max_addr) { | |
393 | *addr = max_addr; | |
394 | ret = -1; | |
395 | } | |
396 | nodes[nid].end = *addr; | |
e3f1caee | 397 | node_set(nid, node_possible_map); |
8b8ca80e DR |
398 | printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid, |
399 | nodes[nid].start, nodes[nid].end, | |
400 | (nodes[nid].end - nodes[nid].start) >> 20); | |
401 | return ret; | |
53fee04f RS |
402 | } |
403 | ||
adc19389 DR |
404 | /* |
405 | * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr | |
406 | * to max_addr. The return value is the number of nodes allocated. | |
407 | */ | |
408 | static int __init split_nodes_interleave(u64 addr, u64 max_addr, | |
409 | int nr_phys_nodes, int nr_nodes) | |
410 | { | |
411 | nodemask_t physnode_mask = NODE_MASK_NONE; | |
412 | u64 size; | |
413 | int big; | |
414 | int ret = 0; | |
415 | int i; | |
416 | ||
417 | if (nr_nodes <= 0) | |
418 | return -1; | |
419 | if (nr_nodes > MAX_NUMNODES) { | |
420 | pr_info("numa=fake=%d too large, reducing to %d\n", | |
421 | nr_nodes, MAX_NUMNODES); | |
422 | nr_nodes = MAX_NUMNODES; | |
423 | } | |
424 | ||
425 | size = (max_addr - addr - e820_hole_size(addr, max_addr)) / nr_nodes; | |
426 | /* | |
427 | * Calculate the number of big nodes that can be allocated as a result | |
428 | * of consolidating the remainder. | |
429 | */ | |
68fd111e | 430 | big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) / |
adc19389 DR |
431 | FAKE_NODE_MIN_SIZE; |
432 | ||
433 | size &= FAKE_NODE_MIN_HASH_MASK; | |
434 | if (!size) { | |
435 | pr_err("Not enough memory for each node. " | |
436 | "NUMA emulation disabled.\n"); | |
437 | return -1; | |
438 | } | |
439 | ||
440 | for (i = 0; i < nr_phys_nodes; i++) | |
441 | if (physnodes[i].start != physnodes[i].end) | |
442 | node_set(i, physnode_mask); | |
443 | ||
444 | /* | |
445 | * Continue to fill physical nodes with fake nodes until there is no | |
446 | * memory left on any of them. | |
447 | */ | |
448 | while (nodes_weight(physnode_mask)) { | |
449 | for_each_node_mask(i, physnode_mask) { | |
450 | u64 end = physnodes[i].start + size; | |
451 | u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN); | |
452 | ||
453 | if (ret < big) | |
454 | end += FAKE_NODE_MIN_SIZE; | |
455 | ||
456 | /* | |
457 | * Continue to add memory to this fake node if its | |
458 | * non-reserved memory is less than the per-node size. | |
459 | */ | |
460 | while (end - physnodes[i].start - | |
461 | e820_hole_size(physnodes[i].start, end) < size) { | |
462 | end += FAKE_NODE_MIN_SIZE; | |
463 | if (end > physnodes[i].end) { | |
464 | end = physnodes[i].end; | |
465 | break; | |
466 | } | |
467 | } | |
468 | ||
469 | /* | |
470 | * If there won't be at least FAKE_NODE_MIN_SIZE of | |
471 | * non-reserved memory in ZONE_DMA32 for the next node, | |
472 | * this one must extend to the boundary. | |
473 | */ | |
474 | if (end < dma32_end && dma32_end - end - | |
475 | e820_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE) | |
476 | end = dma32_end; | |
477 | ||
478 | /* | |
479 | * If there won't be enough non-reserved memory for the | |
480 | * next node, this one must extend to the end of the | |
481 | * physical node. | |
482 | */ | |
483 | if (physnodes[i].end - end - | |
484 | e820_hole_size(end, physnodes[i].end) < size) | |
485 | end = physnodes[i].end; | |
486 | ||
487 | /* | |
488 | * Avoid allocating more nodes than requested, which can | |
489 | * happen as a result of rounding down each node's size | |
490 | * to FAKE_NODE_MIN_SIZE. | |
491 | */ | |
492 | if (nodes_weight(physnode_mask) + ret >= nr_nodes) | |
493 | end = physnodes[i].end; | |
494 | ||
495 | if (setup_node_range(ret++, &physnodes[i].start, | |
496 | end - physnodes[i].start, | |
497 | physnodes[i].end) < 0) | |
498 | node_clear(i, physnode_mask); | |
499 | } | |
500 | } | |
501 | return ret; | |
502 | } | |
503 | ||
8b8ca80e DR |
504 | /* |
505 | * Splits num_nodes nodes up equally starting at node_start. The return value | |
506 | * is the number of nodes split up and addr is adjusted to be at the end of the | |
507 | * last node allocated. | |
508 | */ | |
adc19389 | 509 | static int __init split_nodes_equally(u64 *addr, u64 max_addr, int node_start, |
8b8ca80e | 510 | int num_nodes) |
1da177e4 | 511 | { |
8b8ca80e DR |
512 | unsigned int big; |
513 | u64 size; | |
514 | int i; | |
53fee04f | 515 | |
8b8ca80e DR |
516 | if (num_nodes <= 0) |
517 | return -1; | |
518 | if (num_nodes > MAX_NUMNODES) | |
519 | num_nodes = MAX_NUMNODES; | |
a7e96629 | 520 | size = (max_addr - *addr - e820_hole_size(*addr, max_addr)) / |
8b8ca80e | 521 | num_nodes; |
53fee04f | 522 | /* |
8b8ca80e DR |
523 | * Calculate the number of big nodes that can be allocated as a result |
524 | * of consolidating the leftovers. | |
53fee04f | 525 | */ |
8b8ca80e DR |
526 | big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * num_nodes) / |
527 | FAKE_NODE_MIN_SIZE; | |
528 | ||
529 | /* Round down to nearest FAKE_NODE_MIN_SIZE. */ | |
530 | size &= FAKE_NODE_MIN_HASH_MASK; | |
531 | if (!size) { | |
532 | printk(KERN_ERR "Not enough memory for each node. " | |
533 | "NUMA emulation disabled.\n"); | |
534 | return -1; | |
53fee04f | 535 | } |
8b8ca80e DR |
536 | |
537 | for (i = node_start; i < num_nodes + node_start; i++) { | |
538 | u64 end = *addr + size; | |
e3cfe529 | 539 | |
53fee04f RS |
540 | if (i < big) |
541 | end += FAKE_NODE_MIN_SIZE; | |
542 | /* | |
8b8ca80e DR |
543 | * The final node can have the remaining system RAM. Other |
544 | * nodes receive roughly the same amount of available pages. | |
53fee04f | 545 | */ |
8b8ca80e DR |
546 | if (i == num_nodes + node_start - 1) |
547 | end = max_addr; | |
548 | else | |
a7e96629 | 549 | while (end - *addr - e820_hole_size(*addr, end) < |
8b8ca80e DR |
550 | size) { |
551 | end += FAKE_NODE_MIN_SIZE; | |
552 | if (end > max_addr) { | |
553 | end = max_addr; | |
554 | break; | |
555 | } | |
556 | } | |
adc19389 | 557 | if (setup_node_range(i, addr, end - *addr, max_addr) < 0) |
8b8ca80e DR |
558 | break; |
559 | } | |
560 | return i - node_start + 1; | |
561 | } | |
562 | ||
382591d5 DR |
563 | /* |
564 | * Splits the remaining system RAM into chunks of size. The remaining memory is | |
565 | * always assigned to a final node and can be asymmetric. Returns the number of | |
566 | * nodes split. | |
567 | */ | |
adc19389 DR |
568 | static int __init split_nodes_by_size(u64 *addr, u64 max_addr, int node_start, |
569 | u64 size) | |
382591d5 DR |
570 | { |
571 | int i = node_start; | |
572 | size = (size << 20) & FAKE_NODE_MIN_HASH_MASK; | |
adc19389 | 573 | while (!setup_node_range(i++, addr, size, max_addr)) |
382591d5 DR |
574 | ; |
575 | return i - node_start; | |
576 | } | |
577 | ||
8b8ca80e | 578 | /* |
886533a3 | 579 | * Sets up the system RAM area from start_pfn to last_pfn according to the |
8b8ca80e DR |
580 | * numa=fake command-line option. |
581 | */ | |
adc19389 DR |
582 | static int __init numa_emulation(unsigned long start_pfn, |
583 | unsigned long last_pfn, int acpi, int k8) | |
8b8ca80e | 584 | { |
e3cfe529 | 585 | u64 size, addr = start_pfn << PAGE_SHIFT; |
886533a3 | 586 | u64 max_addr = last_pfn << PAGE_SHIFT; |
e3cfe529 | 587 | int num_nodes = 0, num = 0, coeff_flag, coeff = -1, i; |
adc19389 | 588 | int num_phys_nodes; |
8b8ca80e | 589 | |
adc19389 | 590 | num_phys_nodes = setup_physnodes(addr, max_addr, acpi, k8); |
8b8ca80e DR |
591 | /* |
592 | * If the numa=fake command-line is just a single number N, split the | |
593 | * system RAM into N fake nodes. | |
594 | */ | |
595 | if (!strchr(cmdline, '*') && !strchr(cmdline, ',')) { | |
e3cfe529 TG |
596 | long n = simple_strtol(cmdline, NULL, 0); |
597 | ||
adc19389 DR |
598 | num_nodes = split_nodes_interleave(addr, max_addr, |
599 | num_phys_nodes, n); | |
8b8ca80e DR |
600 | if (num_nodes < 0) |
601 | return num_nodes; | |
602 | goto out; | |
603 | } | |
604 | ||
605 | /* Parse the command line. */ | |
382591d5 | 606 | for (coeff_flag = 0; ; cmdline++) { |
8b8ca80e DR |
607 | if (*cmdline && isdigit(*cmdline)) { |
608 | num = num * 10 + *cmdline - '0'; | |
609 | continue; | |
53fee04f | 610 | } |
382591d5 DR |
611 | if (*cmdline == '*') { |
612 | if (num > 0) | |
613 | coeff = num; | |
614 | coeff_flag = 1; | |
615 | } | |
8b8ca80e | 616 | if (!*cmdline || *cmdline == ',') { |
382591d5 DR |
617 | if (!coeff_flag) |
618 | coeff = 1; | |
8b8ca80e DR |
619 | /* |
620 | * Round down to the nearest FAKE_NODE_MIN_SIZE. | |
621 | * Command-line coefficients are in megabytes. | |
622 | */ | |
623 | size = ((u64)num << 20) & FAKE_NODE_MIN_HASH_MASK; | |
382591d5 | 624 | if (size) |
8b8ca80e | 625 | for (i = 0; i < coeff; i++, num_nodes++) |
adc19389 DR |
626 | if (setup_node_range(num_nodes, &addr, |
627 | size, max_addr) < 0) | |
8b8ca80e | 628 | goto done; |
382591d5 DR |
629 | if (!*cmdline) |
630 | break; | |
631 | coeff_flag = 0; | |
632 | coeff = -1; | |
53fee04f | 633 | } |
8b8ca80e DR |
634 | num = 0; |
635 | } | |
636 | done: | |
637 | if (!num_nodes) | |
638 | return -1; | |
14694d73 | 639 | /* Fill remainder of system RAM, if appropriate. */ |
8b8ca80e | 640 | if (addr < max_addr) { |
382591d5 DR |
641 | if (coeff_flag && coeff < 0) { |
642 | /* Split remaining nodes into num-sized chunks */ | |
adc19389 | 643 | num_nodes += split_nodes_by_size(&addr, max_addr, |
382591d5 DR |
644 | num_nodes, num); |
645 | goto out; | |
646 | } | |
14694d73 DR |
647 | switch (*(cmdline - 1)) { |
648 | case '*': | |
649 | /* Split remaining nodes into coeff chunks */ | |
650 | if (coeff <= 0) | |
651 | break; | |
adc19389 | 652 | num_nodes += split_nodes_equally(&addr, max_addr, |
14694d73 DR |
653 | num_nodes, coeff); |
654 | break; | |
655 | case ',': | |
656 | /* Do not allocate remaining system RAM */ | |
657 | break; | |
658 | default: | |
659 | /* Give one final node */ | |
adc19389 DR |
660 | setup_node_range(num_nodes, &addr, max_addr - addr, |
661 | max_addr); | |
14694d73 DR |
662 | num_nodes++; |
663 | } | |
8b8ca80e DR |
664 | } |
665 | out: | |
6ec6e0d9 | 666 | memnode_shift = compute_hash_shift(nodes, num_nodes, NULL); |
8b8ca80e DR |
667 | if (memnode_shift < 0) { |
668 | memnode_shift = 0; | |
669 | printk(KERN_ERR "No NUMA hash function found. NUMA emulation " | |
670 | "disabled.\n"); | |
671 | return -1; | |
672 | } | |
673 | ||
674 | /* | |
adc19389 DR |
675 | * We need to vacate all active ranges that may have been registered for |
676 | * the e820 memory map. | |
8b8ca80e DR |
677 | */ |
678 | remove_all_active_ranges(); | |
e3f1caee | 679 | for_each_node_mask(i, node_possible_map) { |
5cb248ab MG |
680 | e820_register_active_regions(i, nodes[i].start >> PAGE_SHIFT, |
681 | nodes[i].end >> PAGE_SHIFT); | |
e3cfe529 | 682 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); |
5cb248ab | 683 | } |
3484d798 | 684 | acpi_fake_nodes(nodes, num_nodes); |
e3cfe529 TG |
685 | numa_init_array(); |
686 | return 0; | |
1da177e4 | 687 | } |
8b8ca80e | 688 | #endif /* CONFIG_NUMA_EMU */ |
1da177e4 | 689 | |
8ee2debc DR |
690 | void __init initmem_init(unsigned long start_pfn, unsigned long last_pfn, |
691 | int acpi, int k8) | |
e3cfe529 | 692 | { |
1da177e4 LT |
693 | int i; |
694 | ||
e3f1caee | 695 | nodes_clear(node_possible_map); |
b7ad149d | 696 | nodes_clear(node_online_map); |
e3f1caee | 697 | |
1da177e4 | 698 | #ifdef CONFIG_NUMA_EMU |
adc19389 | 699 | if (cmdline && !numa_emulation(start_pfn, last_pfn, acpi, k8)) |
e3cfe529 | 700 | return; |
e3f1caee | 701 | nodes_clear(node_possible_map); |
b7ad149d | 702 | nodes_clear(node_online_map); |
1da177e4 LT |
703 | #endif |
704 | ||
705 | #ifdef CONFIG_ACPI_NUMA | |
8716273c DR |
706 | if (!numa_off && acpi && !acpi_scan_nodes(start_pfn << PAGE_SHIFT, |
707 | last_pfn << PAGE_SHIFT)) | |
e3cfe529 | 708 | return; |
e3f1caee | 709 | nodes_clear(node_possible_map); |
b7ad149d | 710 | nodes_clear(node_online_map); |
1da177e4 LT |
711 | #endif |
712 | ||
713 | #ifdef CONFIG_K8_NUMA | |
8ee2debc | 714 | if (!numa_off && k8 && !k8_scan_nodes()) |
1da177e4 | 715 | return; |
e3f1caee | 716 | nodes_clear(node_possible_map); |
b7ad149d | 717 | nodes_clear(node_online_map); |
1da177e4 LT |
718 | #endif |
719 | printk(KERN_INFO "%s\n", | |
720 | numa_off ? "NUMA turned off" : "No NUMA configuration found"); | |
721 | ||
e3cfe529 | 722 | printk(KERN_INFO "Faking a node at %016lx-%016lx\n", |
1da177e4 | 723 | start_pfn << PAGE_SHIFT, |
886533a3 | 724 | last_pfn << PAGE_SHIFT); |
e3cfe529 TG |
725 | /* setup dummy node covering all memory */ |
726 | memnode_shift = 63; | |
076422d2 | 727 | memnodemap = memnode.embedded_map; |
1da177e4 | 728 | memnodemap[0] = 0; |
1da177e4 | 729 | node_set_online(0); |
e3f1caee | 730 | node_set(0, node_possible_map); |
168ef543 | 731 | for (i = 0; i < nr_cpu_ids; i++) |
69d81fcd | 732 | numa_set_node(i, 0); |
886533a3 TG |
733 | e820_register_active_regions(0, start_pfn, last_pfn); |
734 | setup_node_bootmem(0, start_pfn << PAGE_SHIFT, last_pfn << PAGE_SHIFT); | |
69d81fcd AK |
735 | } |
736 | ||
e3cfe529 TG |
737 | unsigned long __init numa_free_all_bootmem(void) |
738 | { | |
1da177e4 | 739 | unsigned long pages = 0; |
e3cfe529 TG |
740 | int i; |
741 | ||
742 | for_each_online_node(i) | |
1da177e4 | 743 | pages += free_all_bootmem_node(NODE_DATA(i)); |
e3cfe529 | 744 | |
1da177e4 | 745 | return pages; |
e3cfe529 | 746 | } |
1da177e4 | 747 | |
2c8c0e6b | 748 | static __init int numa_setup(char *opt) |
e3cfe529 | 749 | { |
2c8c0e6b AK |
750 | if (!opt) |
751 | return -EINVAL; | |
e3cfe529 | 752 | if (!strncmp(opt, "off", 3)) |
1da177e4 LT |
753 | numa_off = 1; |
754 | #ifdef CONFIG_NUMA_EMU | |
8b8ca80e DR |
755 | if (!strncmp(opt, "fake=", 5)) |
756 | cmdline = opt + 5; | |
1da177e4 LT |
757 | #endif |
758 | #ifdef CONFIG_ACPI_NUMA | |
e3cfe529 TG |
759 | if (!strncmp(opt, "noacpi", 6)) |
760 | acpi_numa = -1; | |
1da177e4 | 761 | #endif |
2c8c0e6b | 762 | return 0; |
e3cfe529 | 763 | } |
2c8c0e6b AK |
764 | early_param("numa", numa_setup); |
765 | ||
23ca4bba | 766 | #ifdef CONFIG_NUMA |
d9c2d5ac YL |
767 | |
768 | static __init int find_near_online_node(int node) | |
769 | { | |
770 | int n, val; | |
771 | int min_val = INT_MAX; | |
772 | int best_node = -1; | |
773 | ||
774 | for_each_online_node(n) { | |
775 | val = node_distance(node, n); | |
776 | ||
777 | if (val < min_val) { | |
778 | min_val = val; | |
779 | best_node = n; | |
780 | } | |
781 | } | |
782 | ||
783 | return best_node; | |
784 | } | |
785 | ||
05b3cbd8 RT |
786 | /* |
787 | * Setup early cpu_to_node. | |
788 | * | |
789 | * Populate cpu_to_node[] only if x86_cpu_to_apicid[], | |
790 | * and apicid_to_node[] tables have valid entries for a CPU. | |
791 | * This means we skip cpu_to_node[] initialisation for NUMA | |
792 | * emulation and faking node case (when running a kernel compiled | |
793 | * for NUMA on a non NUMA box), which is OK as cpu_to_node[] | |
794 | * is already initialized in a round robin manner at numa_init_array, | |
795 | * prior to this call, and this initialization is good enough | |
796 | * for the fake NUMA cases. | |
23ca4bba MT |
797 | * |
798 | * Called before the per_cpu areas are setup. | |
05b3cbd8 RT |
799 | */ |
800 | void __init init_cpu_to_node(void) | |
801 | { | |
23ca4bba MT |
802 | int cpu; |
803 | u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid); | |
e3cfe529 | 804 | |
23ca4bba MT |
805 | BUG_ON(cpu_to_apicid == NULL); |
806 | ||
807 | for_each_possible_cpu(cpu) { | |
7c9e92b6 | 808 | int node; |
23ca4bba | 809 | u16 apicid = cpu_to_apicid[cpu]; |
e3cfe529 | 810 | |
05b3cbd8 RT |
811 | if (apicid == BAD_APICID) |
812 | continue; | |
7c9e92b6 YL |
813 | node = apicid_to_node[apicid]; |
814 | if (node == NUMA_NO_NODE) | |
05b3cbd8 | 815 | continue; |
7c9e92b6 | 816 | if (!node_online(node)) |
d9c2d5ac | 817 | node = find_near_online_node(node); |
23ca4bba | 818 | numa_set_node(cpu, node); |
05b3cbd8 RT |
819 | } |
820 | } | |
23ca4bba | 821 | #endif |
05b3cbd8 | 822 | |
cf050132 | 823 | |
6470aff6 BG |
824 | void __cpuinit numa_set_node(int cpu, int node) |
825 | { | |
826 | int *cpu_to_node_map = early_per_cpu_ptr(x86_cpu_to_node_map); | |
827 | ||
828 | /* early setting, no percpu area yet */ | |
829 | if (cpu_to_node_map) { | |
830 | cpu_to_node_map[cpu] = node; | |
831 | return; | |
832 | } | |
833 | ||
834 | #ifdef CONFIG_DEBUG_PER_CPU_MAPS | |
44581a28 | 835 | if (cpu >= nr_cpu_ids || !cpu_possible(cpu)) { |
6470aff6 BG |
836 | printk(KERN_ERR "numa_set_node: invalid cpu# (%d)\n", cpu); |
837 | dump_stack(); | |
838 | return; | |
839 | } | |
840 | #endif | |
841 | per_cpu(x86_cpu_to_node_map, cpu) = node; | |
842 | ||
843 | if (node != NUMA_NO_NODE) | |
844 | per_cpu(node_number, cpu) = node; | |
845 | } | |
846 | ||
847 | void __cpuinit numa_clear_node(int cpu) | |
848 | { | |
849 | numa_set_node(cpu, NUMA_NO_NODE); | |
850 | } | |
851 | ||
852 | #ifndef CONFIG_DEBUG_PER_CPU_MAPS | |
853 | ||
854 | void __cpuinit numa_add_cpu(int cpu) | |
855 | { | |
c032ef60 | 856 | cpumask_set_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); |
6470aff6 BG |
857 | } |
858 | ||
859 | void __cpuinit numa_remove_cpu(int cpu) | |
860 | { | |
c032ef60 | 861 | cpumask_clear_cpu(cpu, node_to_cpumask_map[early_cpu_to_node(cpu)]); |
6470aff6 BG |
862 | } |
863 | ||
864 | #else /* CONFIG_DEBUG_PER_CPU_MAPS */ | |
865 | ||
866 | /* | |
867 | * --------- debug versions of the numa functions --------- | |
868 | */ | |
869 | static void __cpuinit numa_set_cpumask(int cpu, int enable) | |
870 | { | |
871 | int node = early_cpu_to_node(cpu); | |
73e907de | 872 | struct cpumask *mask; |
6470aff6 BG |
873 | char buf[64]; |
874 | ||
c032ef60 RR |
875 | mask = node_to_cpumask_map[node]; |
876 | if (mask == NULL) { | |
877 | printk(KERN_ERR "node_to_cpumask_map[%i] NULL\n", node); | |
6470aff6 BG |
878 | dump_stack(); |
879 | return; | |
880 | } | |
881 | ||
6470aff6 | 882 | if (enable) |
c032ef60 | 883 | cpumask_set_cpu(cpu, mask); |
6470aff6 | 884 | else |
c032ef60 | 885 | cpumask_clear_cpu(cpu, mask); |
6470aff6 BG |
886 | |
887 | cpulist_scnprintf(buf, sizeof(buf), mask); | |
888 | printk(KERN_DEBUG "%s cpu %d node %d: mask now %s\n", | |
889 | enable ? "numa_add_cpu" : "numa_remove_cpu", cpu, node, buf); | |
890 | } | |
891 | ||
892 | void __cpuinit numa_add_cpu(int cpu) | |
893 | { | |
894 | numa_set_cpumask(cpu, 1); | |
895 | } | |
896 | ||
897 | void __cpuinit numa_remove_cpu(int cpu) | |
898 | { | |
899 | numa_set_cpumask(cpu, 0); | |
900 | } | |
901 | ||
902 | int cpu_to_node(int cpu) | |
903 | { | |
904 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) { | |
905 | printk(KERN_WARNING | |
906 | "cpu_to_node(%d): usage too early!\n", cpu); | |
907 | dump_stack(); | |
908 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
909 | } | |
910 | return per_cpu(x86_cpu_to_node_map, cpu); | |
911 | } | |
912 | EXPORT_SYMBOL(cpu_to_node); | |
913 | ||
914 | /* | |
915 | * Same function as cpu_to_node() but used if called before the | |
916 | * per_cpu areas are setup. | |
917 | */ | |
918 | int early_cpu_to_node(int cpu) | |
919 | { | |
920 | if (early_per_cpu_ptr(x86_cpu_to_node_map)) | |
921 | return early_per_cpu_ptr(x86_cpu_to_node_map)[cpu]; | |
922 | ||
44581a28 | 923 | if (!cpu_possible(cpu)) { |
6470aff6 BG |
924 | printk(KERN_WARNING |
925 | "early_cpu_to_node(%d): no per_cpu area!\n", cpu); | |
926 | dump_stack(); | |
927 | return NUMA_NO_NODE; | |
928 | } | |
929 | return per_cpu(x86_cpu_to_node_map, cpu); | |
930 | } | |
931 | ||
6470aff6 BG |
932 | /* |
933 | * --------- end of debug versions of the numa functions --------- | |
934 | */ | |
935 | ||
936 | #endif /* CONFIG_DEBUG_PER_CPU_MAPS */ |