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> | |
72d7c3b3 | 10 | #include <linux/memblock.h> |
1da177e4 LT |
11 | #include <linux/mmzone.h> |
12 | #include <linux/ctype.h> | |
13 | #include <linux/module.h> | |
14 | #include <linux/nodemask.h> | |
3cc87e3f | 15 | #include <linux/sched.h> |
1da177e4 LT |
16 | |
17 | #include <asm/e820.h> | |
18 | #include <asm/proto.h> | |
19 | #include <asm/dma.h> | |
20 | #include <asm/numa.h> | |
21 | #include <asm/acpi.h> | |
23ac4ae8 | 22 | #include <asm/amd_nb.h> |
1da177e4 | 23 | |
6c231b7b | 24 | struct pglist_data *node_data[MAX_NUMNODES] __read_mostly; |
e3cfe529 TG |
25 | EXPORT_SYMBOL(node_data); |
26 | ||
dcf36bfa | 27 | struct memnode memnode; |
1da177e4 | 28 | |
864fc31e TG |
29 | static unsigned long __initdata nodemap_addr; |
30 | static unsigned long __initdata nodemap_size; | |
1da177e4 | 31 | |
529a3404 ED |
32 | /* |
33 | * Given a shift value, try to populate memnodemap[] | |
34 | * Returns : | |
35 | * 1 if OK | |
36 | * 0 if memnodmap[] too small (of shift too small) | |
37 | * -1 if node overlap or lost ram (shift too big) | |
38 | */ | |
e3cfe529 | 39 | static int __init populate_memnodemap(const struct bootnode *nodes, |
6ec6e0d9 | 40 | int numnodes, int shift, int *nodeids) |
1da177e4 | 41 | { |
529a3404 | 42 | unsigned long addr, end; |
e3cfe529 | 43 | int i, res = -1; |
b684664f | 44 | |
43238382 | 45 | memset(memnodemap, 0xff, sizeof(s16)*memnodemapsize); |
b684664f | 46 | for (i = 0; i < numnodes; i++) { |
529a3404 ED |
47 | addr = nodes[i].start; |
48 | end = nodes[i].end; | |
49 | if (addr >= end) | |
b684664f | 50 | continue; |
076422d2 | 51 | if ((end >> shift) >= memnodemapsize) |
529a3404 ED |
52 | return 0; |
53 | do { | |
43238382 | 54 | if (memnodemap[addr >> shift] != NUMA_NO_NODE) |
b684664f | 55 | return -1; |
6ec6e0d9 SS |
56 | |
57 | if (!nodeids) | |
58 | memnodemap[addr >> shift] = i; | |
59 | else | |
60 | memnodemap[addr >> shift] = nodeids[i]; | |
61 | ||
076422d2 | 62 | addr += (1UL << shift); |
529a3404 ED |
63 | } while (addr < end); |
64 | res = 1; | |
e3cfe529 | 65 | } |
529a3404 ED |
66 | return res; |
67 | } | |
68 | ||
076422d2 AS |
69 | static int __init allocate_cachealigned_memnodemap(void) |
70 | { | |
24a5da73 | 71 | unsigned long addr; |
076422d2 AS |
72 | |
73 | memnodemap = memnode.embedded_map; | |
316390b0 | 74 | if (memnodemapsize <= ARRAY_SIZE(memnode.embedded_map)) |
076422d2 | 75 | return 0; |
076422d2 | 76 | |
24a5da73 | 77 | addr = 0x8000; |
be3e89ee | 78 | nodemap_size = roundup(sizeof(s16) * memnodemapsize, L1_CACHE_BYTES); |
dbef7b56 | 79 | nodemap_addr = memblock_find_in_range(addr, get_max_mapped(), |
24a5da73 | 80 | nodemap_size, L1_CACHE_BYTES); |
a9ce6bc1 | 81 | if (nodemap_addr == MEMBLOCK_ERROR) { |
076422d2 AS |
82 | printk(KERN_ERR |
83 | "NUMA: Unable to allocate Memory to Node hash map\n"); | |
84 | nodemap_addr = nodemap_size = 0; | |
85 | return -1; | |
86 | } | |
24a5da73 | 87 | memnodemap = phys_to_virt(nodemap_addr); |
a9ce6bc1 | 88 | memblock_x86_reserve_range(nodemap_addr, nodemap_addr + nodemap_size, "MEMNODEMAP"); |
076422d2 AS |
89 | |
90 | printk(KERN_DEBUG "NUMA: Allocated memnodemap from %lx - %lx\n", | |
91 | nodemap_addr, nodemap_addr + nodemap_size); | |
92 | return 0; | |
93 | } | |
94 | ||
95 | /* | |
96 | * The LSB of all start and end addresses in the node map is the value of the | |
97 | * maximum possible shift. | |
98 | */ | |
e3cfe529 TG |
99 | static int __init extract_lsb_from_nodes(const struct bootnode *nodes, |
100 | int numnodes) | |
529a3404 | 101 | { |
54413927 | 102 | int i, nodes_used = 0; |
076422d2 AS |
103 | unsigned long start, end; |
104 | unsigned long bitfield = 0, memtop = 0; | |
105 | ||
106 | for (i = 0; i < numnodes; i++) { | |
107 | start = nodes[i].start; | |
108 | end = nodes[i].end; | |
109 | if (start >= end) | |
110 | continue; | |
54413927 AS |
111 | bitfield |= start; |
112 | nodes_used++; | |
076422d2 AS |
113 | if (end > memtop) |
114 | memtop = end; | |
115 | } | |
54413927 AS |
116 | if (nodes_used <= 1) |
117 | i = 63; | |
118 | else | |
119 | i = find_first_bit(&bitfield, sizeof(unsigned long)*8); | |
076422d2 AS |
120 | memnodemapsize = (memtop >> i)+1; |
121 | return i; | |
122 | } | |
529a3404 | 123 | |
6ec6e0d9 SS |
124 | int __init compute_hash_shift(struct bootnode *nodes, int numnodes, |
125 | int *nodeids) | |
076422d2 AS |
126 | { |
127 | int shift; | |
529a3404 | 128 | |
076422d2 AS |
129 | shift = extract_lsb_from_nodes(nodes, numnodes); |
130 | if (allocate_cachealigned_memnodemap()) | |
131 | return -1; | |
6b050f80 | 132 | printk(KERN_DEBUG "NUMA: Using %d for the hash shift.\n", |
529a3404 ED |
133 | shift); |
134 | ||
6ec6e0d9 | 135 | if (populate_memnodemap(nodes, numnodes, shift, nodeids) != 1) { |
e3cfe529 TG |
136 | printk(KERN_INFO "Your memory is not aligned you need to " |
137 | "rebuild your kernel with a bigger NODEMAPSIZE " | |
138 | "shift=%d\n", shift); | |
529a3404 ED |
139 | return -1; |
140 | } | |
b684664f | 141 | return shift; |
1da177e4 LT |
142 | } |
143 | ||
f2dbcfa7 | 144 | int __meminit __early_pfn_to_nid(unsigned long pfn) |
bbfceef4 MT |
145 | { |
146 | return phys_to_nid(pfn << PAGE_SHIFT); | |
147 | } | |
bbfceef4 | 148 | |
e3cfe529 | 149 | static void * __init early_node_mem(int nodeid, unsigned long start, |
24a5da73 YL |
150 | unsigned long end, unsigned long size, |
151 | unsigned long align) | |
a8062231 | 152 | { |
cef625ee | 153 | unsigned long mem; |
e3cfe529 | 154 | |
cef625ee YL |
155 | /* |
156 | * put it on high as possible | |
157 | * something will go with NODE_DATA | |
158 | */ | |
159 | if (start < (MAX_DMA_PFN<<PAGE_SHIFT)) | |
160 | start = MAX_DMA_PFN<<PAGE_SHIFT; | |
161 | if (start < (MAX_DMA32_PFN<<PAGE_SHIFT) && | |
162 | end > (MAX_DMA32_PFN<<PAGE_SHIFT)) | |
163 | start = MAX_DMA32_PFN<<PAGE_SHIFT; | |
72d7c3b3 YL |
164 | mem = memblock_x86_find_in_range_node(nodeid, start, end, size, align); |
165 | if (mem != MEMBLOCK_ERROR) | |
a8062231 | 166 | return __va(mem); |
9347e0b0 | 167 | |
cef625ee YL |
168 | /* extend the search scope */ |
169 | end = max_pfn_mapped << PAGE_SHIFT; | |
419db274 YL |
170 | start = MAX_DMA_PFN << PAGE_SHIFT; |
171 | mem = memblock_find_in_range(start, end, size, align); | |
72d7c3b3 | 172 | if (mem != MEMBLOCK_ERROR) |
a8062231 | 173 | return __va(mem); |
9347e0b0 | 174 | |
1842f90c | 175 | printk(KERN_ERR "Cannot find %lu bytes in node %d\n", |
e3cfe529 | 176 | size, nodeid); |
1842f90c YL |
177 | |
178 | return NULL; | |
a8062231 AK |
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 | { |
08677214 | 185 | unsigned long start_pfn, last_pfn, nodedata_phys; |
7c43769a | 186 | const int pgdat_size = roundup(sizeof(pg_data_t), PAGE_SIZE); |
1a27fc0a | 187 | int nid; |
1da177e4 | 188 | |
4c31e92b YL |
189 | if (!end) |
190 | return; | |
191 | ||
7c43769a YL |
192 | /* |
193 | * Don't confuse VM with a node that doesn't have the | |
194 | * minimum amount of memory: | |
195 | */ | |
196 | if (end && (end - start) < NODE_MIN_SIZE) | |
197 | return; | |
198 | ||
be3e89ee | 199 | start = roundup(start, ZONE_ALIGN); |
1da177e4 | 200 | |
08677214 | 201 | printk(KERN_INFO "Initmem setup node %d %016lx-%016lx\n", nodeid, |
e3cfe529 | 202 | start, end); |
1da177e4 LT |
203 | |
204 | start_pfn = start >> PAGE_SHIFT; | |
886533a3 | 205 | last_pfn = end >> PAGE_SHIFT; |
1da177e4 | 206 | |
24a5da73 YL |
207 | node_data[nodeid] = early_node_mem(nodeid, start, end, pgdat_size, |
208 | SMP_CACHE_BYTES); | |
a8062231 AK |
209 | if (node_data[nodeid] == NULL) |
210 | return; | |
211 | nodedata_phys = __pa(node_data[nodeid]); | |
a9ce6bc1 | 212 | memblock_x86_reserve_range(nodedata_phys, nodedata_phys + pgdat_size, "NODE_DATA"); |
6118f76f YL |
213 | printk(KERN_INFO " NODE_DATA [%016lx - %016lx]\n", nodedata_phys, |
214 | nodedata_phys + pgdat_size - 1); | |
1842f90c YL |
215 | nid = phys_to_nid(nodedata_phys); |
216 | if (nid != nodeid) | |
217 | printk(KERN_INFO " NODE_DATA(%d) on node %d\n", nodeid, nid); | |
1da177e4 | 218 | |
1da177e4 | 219 | memset(NODE_DATA(nodeid), 0, sizeof(pg_data_t)); |
08677214 | 220 | NODE_DATA(nodeid)->node_id = nodeid; |
1da177e4 | 221 | NODE_DATA(nodeid)->node_start_pfn = start_pfn; |
886533a3 | 222 | NODE_DATA(nodeid)->node_spanned_pages = last_pfn - start_pfn; |
1da177e4 | 223 | |
1da177e4 | 224 | node_set_online(nodeid); |
e3cfe529 | 225 | } |
1da177e4 | 226 | |
1da177e4 | 227 | #ifdef CONFIG_NUMA_EMU |
53fee04f | 228 | /* Numa emulation */ |
adc19389 | 229 | static struct bootnode nodes[MAX_NUMNODES] __initdata; |
c1c3443c | 230 | static struct bootnode physnodes[MAX_NUMNODES] __cpuinitdata; |
864fc31e | 231 | static char *cmdline __initdata; |
1da177e4 | 232 | |
90321602 JB |
233 | void __init numa_emu_cmdline(char *str) |
234 | { | |
235 | cmdline = str; | |
236 | } | |
237 | ||
adc19389 | 238 | static int __init setup_physnodes(unsigned long start, unsigned long end, |
eec1d4fa | 239 | int acpi, int amd) |
adc19389 | 240 | { |
adc19389 DR |
241 | int ret = 0; |
242 | int i; | |
243 | ||
c1c3443c | 244 | memset(physnodes, 0, sizeof(physnodes)); |
adc19389 DR |
245 | #ifdef CONFIG_ACPI_NUMA |
246 | if (acpi) | |
a387e95a | 247 | acpi_get_nodes(physnodes, start, end); |
adc19389 | 248 | #endif |
eec1d4fa HR |
249 | #ifdef CONFIG_AMD_NUMA |
250 | if (amd) | |
a387e95a | 251 | amd_get_nodes(physnodes); |
adc19389 DR |
252 | #endif |
253 | /* | |
254 | * Basic sanity checking on the physical node map: there may be errors | |
eec1d4fa | 255 | * if the SRAT or AMD code incorrectly reported the topology or the mem= |
adc19389 DR |
256 | * kernel parameter is used. |
257 | */ | |
a387e95a | 258 | for (i = 0; i < MAX_NUMNODES; i++) { |
adc19389 DR |
259 | if (physnodes[i].start == physnodes[i].end) |
260 | continue; | |
261 | if (physnodes[i].start > end) { | |
262 | physnodes[i].end = physnodes[i].start; | |
263 | continue; | |
264 | } | |
265 | if (physnodes[i].end < start) { | |
266 | physnodes[i].start = physnodes[i].end; | |
267 | continue; | |
268 | } | |
269 | if (physnodes[i].start < start) | |
270 | physnodes[i].start = start; | |
271 | if (physnodes[i].end > end) | |
272 | physnodes[i].end = end; | |
adc19389 DR |
273 | ret++; |
274 | } | |
275 | ||
276 | /* | |
277 | * If no physical topology was detected, a single node is faked to cover | |
278 | * the entire address space. | |
279 | */ | |
280 | if (!ret) { | |
281 | physnodes[ret].start = start; | |
282 | physnodes[ret].end = end; | |
283 | ret = 1; | |
284 | } | |
285 | return ret; | |
286 | } | |
287 | ||
f51bf307 DR |
288 | static void __init fake_physnodes(int acpi, int amd, int nr_nodes) |
289 | { | |
290 | int i; | |
291 | ||
292 | BUG_ON(acpi && amd); | |
293 | #ifdef CONFIG_ACPI_NUMA | |
294 | if (acpi) | |
295 | acpi_fake_nodes(nodes, nr_nodes); | |
296 | #endif | |
297 | #ifdef CONFIG_AMD_NUMA | |
298 | if (amd) | |
299 | amd_fake_nodes(nodes, nr_nodes); | |
300 | #endif | |
301 | if (!acpi && !amd) | |
302 | for (i = 0; i < nr_cpu_ids; i++) | |
303 | numa_set_node(i, 0); | |
304 | } | |
305 | ||
53fee04f | 306 | /* |
e3cfe529 TG |
307 | * Setups up nid to range from addr to addr + size. If the end |
308 | * boundary is greater than max_addr, then max_addr is used instead. | |
309 | * The return value is 0 if there is additional memory left for | |
310 | * allocation past addr and -1 otherwise. addr is adjusted to be at | |
311 | * the end of the node. | |
53fee04f | 312 | */ |
adc19389 | 313 | static int __init setup_node_range(int nid, u64 *addr, u64 size, u64 max_addr) |
53fee04f | 314 | { |
8b8ca80e DR |
315 | int ret = 0; |
316 | nodes[nid].start = *addr; | |
317 | *addr += size; | |
318 | if (*addr >= max_addr) { | |
319 | *addr = max_addr; | |
320 | ret = -1; | |
321 | } | |
322 | nodes[nid].end = *addr; | |
e3f1caee | 323 | node_set(nid, node_possible_map); |
8b8ca80e DR |
324 | printk(KERN_INFO "Faking node %d at %016Lx-%016Lx (%LuMB)\n", nid, |
325 | nodes[nid].start, nodes[nid].end, | |
326 | (nodes[nid].end - nodes[nid].start) >> 20); | |
327 | return ret; | |
53fee04f RS |
328 | } |
329 | ||
adc19389 DR |
330 | /* |
331 | * Sets up nr_nodes fake nodes interleaved over physical nodes ranging from addr | |
332 | * to max_addr. The return value is the number of nodes allocated. | |
333 | */ | |
c1c3443c | 334 | static int __init split_nodes_interleave(u64 addr, u64 max_addr, int nr_nodes) |
adc19389 DR |
335 | { |
336 | nodemask_t physnode_mask = NODE_MASK_NONE; | |
337 | u64 size; | |
338 | int big; | |
339 | int ret = 0; | |
340 | int i; | |
341 | ||
342 | if (nr_nodes <= 0) | |
343 | return -1; | |
344 | if (nr_nodes > MAX_NUMNODES) { | |
345 | pr_info("numa=fake=%d too large, reducing to %d\n", | |
346 | nr_nodes, MAX_NUMNODES); | |
347 | nr_nodes = MAX_NUMNODES; | |
348 | } | |
349 | ||
a9ce6bc1 | 350 | size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) / nr_nodes; |
adc19389 DR |
351 | /* |
352 | * Calculate the number of big nodes that can be allocated as a result | |
353 | * of consolidating the remainder. | |
354 | */ | |
68fd111e | 355 | big = ((size & ~FAKE_NODE_MIN_HASH_MASK) * nr_nodes) / |
adc19389 DR |
356 | FAKE_NODE_MIN_SIZE; |
357 | ||
358 | size &= FAKE_NODE_MIN_HASH_MASK; | |
359 | if (!size) { | |
360 | pr_err("Not enough memory for each node. " | |
361 | "NUMA emulation disabled.\n"); | |
362 | return -1; | |
363 | } | |
364 | ||
c1c3443c | 365 | for (i = 0; i < MAX_NUMNODES; i++) |
adc19389 DR |
366 | if (physnodes[i].start != physnodes[i].end) |
367 | node_set(i, physnode_mask); | |
368 | ||
369 | /* | |
370 | * Continue to fill physical nodes with fake nodes until there is no | |
371 | * memory left on any of them. | |
372 | */ | |
373 | while (nodes_weight(physnode_mask)) { | |
374 | for_each_node_mask(i, physnode_mask) { | |
375 | u64 end = physnodes[i].start + size; | |
376 | u64 dma32_end = PFN_PHYS(MAX_DMA32_PFN); | |
377 | ||
378 | if (ret < big) | |
379 | end += FAKE_NODE_MIN_SIZE; | |
380 | ||
381 | /* | |
382 | * Continue to add memory to this fake node if its | |
383 | * non-reserved memory is less than the per-node size. | |
384 | */ | |
385 | while (end - physnodes[i].start - | |
a9ce6bc1 | 386 | memblock_x86_hole_size(physnodes[i].start, end) < size) { |
adc19389 DR |
387 | end += FAKE_NODE_MIN_SIZE; |
388 | if (end > physnodes[i].end) { | |
389 | end = physnodes[i].end; | |
390 | break; | |
391 | } | |
392 | } | |
393 | ||
394 | /* | |
395 | * If there won't be at least FAKE_NODE_MIN_SIZE of | |
396 | * non-reserved memory in ZONE_DMA32 for the next node, | |
397 | * this one must extend to the boundary. | |
398 | */ | |
399 | if (end < dma32_end && dma32_end - end - | |
a9ce6bc1 | 400 | memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE) |
adc19389 DR |
401 | end = dma32_end; |
402 | ||
403 | /* | |
404 | * If there won't be enough non-reserved memory for the | |
405 | * next node, this one must extend to the end of the | |
406 | * physical node. | |
407 | */ | |
408 | if (physnodes[i].end - end - | |
a9ce6bc1 | 409 | memblock_x86_hole_size(end, physnodes[i].end) < size) |
adc19389 DR |
410 | end = physnodes[i].end; |
411 | ||
412 | /* | |
413 | * Avoid allocating more nodes than requested, which can | |
414 | * happen as a result of rounding down each node's size | |
415 | * to FAKE_NODE_MIN_SIZE. | |
416 | */ | |
417 | if (nodes_weight(physnode_mask) + ret >= nr_nodes) | |
418 | end = physnodes[i].end; | |
419 | ||
420 | if (setup_node_range(ret++, &physnodes[i].start, | |
421 | end - physnodes[i].start, | |
422 | physnodes[i].end) < 0) | |
423 | node_clear(i, physnode_mask); | |
424 | } | |
425 | } | |
426 | return ret; | |
427 | } | |
428 | ||
8df5bb34 DR |
429 | /* |
430 | * Returns the end address of a node so that there is at least `size' amount of | |
431 | * non-reserved memory or `max_addr' is reached. | |
432 | */ | |
433 | static u64 __init find_end_of_node(u64 start, u64 max_addr, u64 size) | |
434 | { | |
435 | u64 end = start + size; | |
436 | ||
a9ce6bc1 | 437 | while (end - start - memblock_x86_hole_size(start, end) < size) { |
8df5bb34 DR |
438 | end += FAKE_NODE_MIN_SIZE; |
439 | if (end > max_addr) { | |
440 | end = max_addr; | |
441 | break; | |
442 | } | |
443 | } | |
444 | return end; | |
445 | } | |
446 | ||
447 | /* | |
448 | * Sets up fake nodes of `size' interleaved over physical nodes ranging from | |
449 | * `addr' to `max_addr'. The return value is the number of nodes allocated. | |
450 | */ | |
451 | static int __init split_nodes_size_interleave(u64 addr, u64 max_addr, u64 size) | |
452 | { | |
453 | nodemask_t physnode_mask = NODE_MASK_NONE; | |
454 | u64 min_size; | |
455 | int ret = 0; | |
456 | int i; | |
457 | ||
458 | if (!size) | |
459 | return -1; | |
460 | /* | |
461 | * The limit on emulated nodes is MAX_NUMNODES, so the size per node is | |
462 | * increased accordingly if the requested size is too small. This | |
463 | * creates a uniform distribution of node sizes across the entire | |
464 | * machine (but not necessarily over physical nodes). | |
465 | */ | |
a9ce6bc1 | 466 | min_size = (max_addr - addr - memblock_x86_hole_size(addr, max_addr)) / |
8df5bb34 DR |
467 | MAX_NUMNODES; |
468 | min_size = max(min_size, FAKE_NODE_MIN_SIZE); | |
469 | if ((min_size & FAKE_NODE_MIN_HASH_MASK) < min_size) | |
470 | min_size = (min_size + FAKE_NODE_MIN_SIZE) & | |
471 | FAKE_NODE_MIN_HASH_MASK; | |
472 | if (size < min_size) { | |
473 | pr_err("Fake node size %LuMB too small, increasing to %LuMB\n", | |
474 | size >> 20, min_size >> 20); | |
475 | size = min_size; | |
476 | } | |
477 | size &= FAKE_NODE_MIN_HASH_MASK; | |
478 | ||
479 | for (i = 0; i < MAX_NUMNODES; i++) | |
480 | if (physnodes[i].start != physnodes[i].end) | |
481 | node_set(i, physnode_mask); | |
482 | /* | |
483 | * Fill physical nodes with fake nodes of size until there is no memory | |
484 | * left on any of them. | |
485 | */ | |
486 | while (nodes_weight(physnode_mask)) { | |
487 | for_each_node_mask(i, physnode_mask) { | |
488 | u64 dma32_end = MAX_DMA32_PFN << PAGE_SHIFT; | |
489 | u64 end; | |
490 | ||
491 | end = find_end_of_node(physnodes[i].start, | |
492 | physnodes[i].end, size); | |
493 | /* | |
494 | * If there won't be at least FAKE_NODE_MIN_SIZE of | |
495 | * non-reserved memory in ZONE_DMA32 for the next node, | |
496 | * this one must extend to the boundary. | |
497 | */ | |
498 | if (end < dma32_end && dma32_end - end - | |
a9ce6bc1 | 499 | memblock_x86_hole_size(end, dma32_end) < FAKE_NODE_MIN_SIZE) |
8df5bb34 DR |
500 | end = dma32_end; |
501 | ||
502 | /* | |
503 | * If there won't be enough non-reserved memory for the | |
504 | * next node, this one must extend to the end of the | |
505 | * physical node. | |
506 | */ | |
507 | if (physnodes[i].end - end - | |
a9ce6bc1 | 508 | memblock_x86_hole_size(end, physnodes[i].end) < size) |
8df5bb34 DR |
509 | end = physnodes[i].end; |
510 | ||
511 | /* | |
512 | * Setup the fake node that will be allocated as bootmem | |
513 | * later. If setup_node_range() returns non-zero, there | |
514 | * is no more memory available on this physical node. | |
515 | */ | |
516 | if (setup_node_range(ret++, &physnodes[i].start, | |
517 | end - physnodes[i].start, | |
518 | physnodes[i].end) < 0) | |
519 | node_clear(i, physnode_mask); | |
520 | } | |
521 | } | |
522 | return ret; | |
523 | } | |
524 | ||
8b8ca80e | 525 | /* |
886533a3 | 526 | * Sets up the system RAM area from start_pfn to last_pfn according to the |
8b8ca80e DR |
527 | * numa=fake command-line option. |
528 | */ | |
adc19389 | 529 | static int __init numa_emulation(unsigned long start_pfn, |
eec1d4fa | 530 | unsigned long last_pfn, int acpi, int amd) |
8b8ca80e | 531 | { |
ca2107c9 | 532 | u64 addr = start_pfn << PAGE_SHIFT; |
886533a3 | 533 | u64 max_addr = last_pfn << PAGE_SHIFT; |
ca2107c9 DR |
534 | int num_nodes; |
535 | int i; | |
8b8ca80e | 536 | |
8df5bb34 DR |
537 | /* |
538 | * If the numa=fake command-line contains a 'M' or 'G', it represents | |
ca2107c9 DR |
539 | * the fixed node size. Otherwise, if it is just a single number N, |
540 | * split the system RAM into N fake nodes. | |
8df5bb34 DR |
541 | */ |
542 | if (strchr(cmdline, 'M') || strchr(cmdline, 'G')) { | |
ca2107c9 DR |
543 | u64 size; |
544 | ||
8df5bb34 DR |
545 | size = memparse(cmdline, &cmdline); |
546 | num_nodes = split_nodes_size_interleave(addr, max_addr, size); | |
ca2107c9 DR |
547 | } else { |
548 | unsigned long n; | |
8df5bb34 | 549 | |
ca2107c9 | 550 | n = simple_strtoul(cmdline, NULL, 0); |
c1c3443c | 551 | num_nodes = split_nodes_interleave(addr, max_addr, n); |
8b8ca80e DR |
552 | } |
553 | ||
ca2107c9 DR |
554 | if (num_nodes < 0) |
555 | return num_nodes; | |
6ec6e0d9 | 556 | memnode_shift = compute_hash_shift(nodes, num_nodes, NULL); |
8b8ca80e DR |
557 | if (memnode_shift < 0) { |
558 | memnode_shift = 0; | |
559 | printk(KERN_ERR "No NUMA hash function found. NUMA emulation " | |
560 | "disabled.\n"); | |
561 | return -1; | |
562 | } | |
563 | ||
564 | /* | |
adc19389 DR |
565 | * We need to vacate all active ranges that may have been registered for |
566 | * the e820 memory map. | |
8b8ca80e DR |
567 | */ |
568 | remove_all_active_ranges(); | |
1411e0ec | 569 | for_each_node_mask(i, node_possible_map) |
a9ce6bc1 | 570 | memblock_x86_register_active_regions(i, nodes[i].start >> PAGE_SHIFT, |
5cb248ab | 571 | nodes[i].end >> PAGE_SHIFT); |
1411e0ec YL |
572 | init_memory_mapping_high(); |
573 | for_each_node_mask(i, node_possible_map) | |
e3cfe529 | 574 | setup_node_bootmem(i, nodes[i].start, nodes[i].end); |
c1c3443c | 575 | setup_physnodes(addr, max_addr, acpi, amd); |
f51bf307 | 576 | fake_physnodes(acpi, amd, num_nodes); |
e3cfe529 TG |
577 | numa_init_array(); |
578 | return 0; | |
1da177e4 | 579 | } |
8b8ca80e | 580 | #endif /* CONFIG_NUMA_EMU */ |
1da177e4 | 581 | |
86ef4dbf | 582 | void __init initmem_init(int acpi, int amd) |
e3cfe529 | 583 | { |
1da177e4 LT |
584 | int i; |
585 | ||
e3f1caee | 586 | nodes_clear(node_possible_map); |
b7ad149d | 587 | nodes_clear(node_online_map); |
e3f1caee | 588 | |
1da177e4 | 589 | #ifdef CONFIG_NUMA_EMU |
86ef4dbf TH |
590 | setup_physnodes(0, max_pfn << PAGE_SHIFT, acpi, amd); |
591 | if (cmdline && !numa_emulation(0, max_pfn, acpi, amd)) | |
e3cfe529 | 592 | return; |
86ef4dbf | 593 | setup_physnodes(0, max_pfn << PAGE_SHIFT, acpi, amd); |
e3f1caee | 594 | nodes_clear(node_possible_map); |
b7ad149d | 595 | nodes_clear(node_online_map); |
1da177e4 LT |
596 | #endif |
597 | ||
598 | #ifdef CONFIG_ACPI_NUMA | |
940fed2e | 599 | if (!numa_off && acpi && !acpi_scan_nodes()) |
e3cfe529 | 600 | return; |
e3f1caee | 601 | nodes_clear(node_possible_map); |
b7ad149d | 602 | nodes_clear(node_online_map); |
1da177e4 LT |
603 | #endif |
604 | ||
eec1d4fa HR |
605 | #ifdef CONFIG_AMD_NUMA |
606 | if (!numa_off && amd && !amd_scan_nodes()) | |
1da177e4 | 607 | return; |
e3f1caee | 608 | nodes_clear(node_possible_map); |
b7ad149d | 609 | nodes_clear(node_online_map); |
1da177e4 LT |
610 | #endif |
611 | printk(KERN_INFO "%s\n", | |
612 | numa_off ? "NUMA turned off" : "No NUMA configuration found"); | |
613 | ||
e3cfe529 | 614 | printk(KERN_INFO "Faking a node at %016lx-%016lx\n", |
86ef4dbf | 615 | 0LU, max_pfn << PAGE_SHIFT); |
e3cfe529 TG |
616 | /* setup dummy node covering all memory */ |
617 | memnode_shift = 63; | |
076422d2 | 618 | memnodemap = memnode.embedded_map; |
1da177e4 | 619 | memnodemap[0] = 0; |
1da177e4 | 620 | node_set_online(0); |
e3f1caee | 621 | node_set(0, node_possible_map); |
7d36b7bc TH |
622 | for (i = 0; i < MAX_LOCAL_APIC; i++) |
623 | set_apicid_to_node(i, NUMA_NO_NODE); | |
86ef4dbf | 624 | memblock_x86_register_active_regions(0, 0, max_pfn); |
1411e0ec | 625 | init_memory_mapping_high(); |
86ef4dbf | 626 | setup_node_bootmem(0, 0, max_pfn << PAGE_SHIFT); |
7d36b7bc | 627 | numa_init_array(); |
69d81fcd AK |
628 | } |
629 | ||
e3cfe529 TG |
630 | unsigned long __init numa_free_all_bootmem(void) |
631 | { | |
1da177e4 | 632 | unsigned long pages = 0; |
e3cfe529 TG |
633 | int i; |
634 | ||
635 | for_each_online_node(i) | |
1da177e4 | 636 | pages += free_all_bootmem_node(NODE_DATA(i)); |
e3cfe529 | 637 | |
08677214 | 638 | pages += free_all_memory_core_early(MAX_NUMNODES); |
08677214 | 639 | |
1da177e4 | 640 | return pages; |
e3cfe529 | 641 | } |
1da177e4 | 642 | |
bbc9e2f4 | 643 | int __cpuinit numa_cpu_node(int cpu) |
d9c2d5ac | 644 | { |
bbc9e2f4 | 645 | int apicid = early_per_cpu(x86_cpu_to_apicid, cpu); |
d9c2d5ac | 646 | |
bbc9e2f4 TH |
647 | if (apicid != BAD_APICID) |
648 | return __apicid_to_node[apicid]; | |
649 | return NUMA_NO_NODE; | |
d9c2d5ac YL |
650 | } |
651 | ||
05b3cbd8 | 652 | /* |
de2d9445 TH |
653 | * UGLINESS AHEAD: Currently, CONFIG_NUMA_EMU is 64bit only and makes use |
654 | * of 64bit specific data structures. The distinction is artificial and | |
655 | * should be removed. numa_{add|remove}_cpu() are implemented in numa.c | |
656 | * for both 32 and 64bit when CONFIG_NUMA_EMU is disabled but here when | |
657 | * enabled. | |
05b3cbd8 | 658 | * |
de2d9445 TH |
659 | * NUMA emulation is planned to be made generic and the following and other |
660 | * related code should be moved to numa.c. | |
05b3cbd8 | 661 | */ |
de2d9445 TH |
662 | #ifdef CONFIG_NUMA_EMU |
663 | # ifndef CONFIG_DEBUG_PER_CPU_MAPS | |
c1c3443c DR |
664 | void __cpuinit numa_add_cpu(int cpu) |
665 | { | |
666 | unsigned long addr; | |
bbc9e2f4 | 667 | int physnid, nid; |
c1c3443c | 668 | |
bbc9e2f4 | 669 | nid = numa_cpu_node(cpu); |
c1c3443c DR |
670 | if (nid == NUMA_NO_NODE) |
671 | nid = early_cpu_to_node(cpu); | |
672 | BUG_ON(nid == NUMA_NO_NODE || !node_online(nid)); | |
673 | ||
674 | /* | |
675 | * Use the starting address of the emulated node to find which physical | |
676 | * node it is allocated on. | |
677 | */ | |
678 | addr = node_start_pfn(nid) << PAGE_SHIFT; | |
679 | for (physnid = 0; physnid < MAX_NUMNODES; physnid++) | |
680 | if (addr >= physnodes[physnid].start && | |
681 | addr < physnodes[physnid].end) | |
682 | break; | |
683 | ||
684 | /* | |
685 | * Map the cpu to each emulated node that is allocated on the physical | |
686 | * node of the cpu's apic id. | |
687 | */ | |
688 | for_each_online_node(nid) { | |
689 | addr = node_start_pfn(nid) << PAGE_SHIFT; | |
690 | if (addr >= physnodes[physnid].start && | |
691 | addr < physnodes[physnid].end) | |
692 | cpumask_set_cpu(cpu, node_to_cpumask_map[nid]); | |
693 | } | |
694 | } | |
695 | ||
696 | void __cpuinit numa_remove_cpu(int cpu) | |
697 | { | |
698 | int i; | |
699 | ||
700 | for_each_online_node(i) | |
701 | cpumask_clear_cpu(cpu, node_to_cpumask_map[i]); | |
702 | } | |
de2d9445 | 703 | # else /* !CONFIG_DEBUG_PER_CPU_MAPS */ |
6470aff6 BG |
704 | static void __cpuinit numa_set_cpumask(int cpu, int enable) |
705 | { | |
706 | int node = early_cpu_to_node(cpu); | |
73e907de | 707 | struct cpumask *mask; |
c1c3443c | 708 | int i; |
6470aff6 | 709 | |
14392fd3 DR |
710 | if (node == NUMA_NO_NODE) { |
711 | /* early_cpu_to_node() already emits a warning and trace */ | |
712 | return; | |
713 | } | |
c1c3443c DR |
714 | for_each_online_node(i) { |
715 | unsigned long addr; | |
6470aff6 | 716 | |
c1c3443c DR |
717 | addr = node_start_pfn(i) << PAGE_SHIFT; |
718 | if (addr < physnodes[node].start || | |
719 | addr >= physnodes[node].end) | |
720 | continue; | |
d906f0eb DR |
721 | mask = debug_cpumask_set_cpu(cpu, enable); |
722 | if (!mask) | |
c1c3443c | 723 | return; |
c1c3443c DR |
724 | |
725 | if (enable) | |
726 | cpumask_set_cpu(cpu, mask); | |
727 | else | |
728 | cpumask_clear_cpu(cpu, mask); | |
c1c3443c | 729 | } |
6470aff6 BG |
730 | } |
731 | ||
732 | void __cpuinit numa_add_cpu(int cpu) | |
733 | { | |
734 | numa_set_cpumask(cpu, 1); | |
735 | } | |
736 | ||
737 | void __cpuinit numa_remove_cpu(int cpu) | |
738 | { | |
739 | numa_set_cpumask(cpu, 0); | |
740 | } | |
de2d9445 TH |
741 | # endif /* !CONFIG_DEBUG_PER_CPU_MAPS */ |
742 | #endif /* CONFIG_NUMA_EMU */ |