Commit | Line | Data |
---|---|---|
d41dee36 AW |
1 | /* |
2 | * sparse memory mappings. | |
3 | */ | |
d41dee36 AW |
4 | #include <linux/mm.h> |
5 | #include <linux/mmzone.h> | |
6 | #include <linux/bootmem.h> | |
0b0acbec | 7 | #include <linux/highmem.h> |
d41dee36 | 8 | #include <linux/module.h> |
28ae55c9 | 9 | #include <linux/spinlock.h> |
0b0acbec | 10 | #include <linux/vmalloc.h> |
d41dee36 | 11 | #include <asm/dma.h> |
8f6aac41 CL |
12 | #include <asm/pgalloc.h> |
13 | #include <asm/pgtable.h> | |
d41dee36 AW |
14 | |
15 | /* | |
16 | * Permanent SPARSEMEM data: | |
17 | * | |
18 | * 1) mem_section - memory sections, mem_map's for valid memory | |
19 | */ | |
3e347261 | 20 | #ifdef CONFIG_SPARSEMEM_EXTREME |
802f192e | 21 | struct mem_section *mem_section[NR_SECTION_ROOTS] |
22fc6ecc | 22 | ____cacheline_internodealigned_in_smp; |
3e347261 BP |
23 | #else |
24 | struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT] | |
22fc6ecc | 25 | ____cacheline_internodealigned_in_smp; |
3e347261 BP |
26 | #endif |
27 | EXPORT_SYMBOL(mem_section); | |
28 | ||
89689ae7 CL |
29 | #ifdef NODE_NOT_IN_PAGE_FLAGS |
30 | /* | |
31 | * If we did not store the node number in the page then we have to | |
32 | * do a lookup in the section_to_node_table in order to find which | |
33 | * node the page belongs to. | |
34 | */ | |
35 | #if MAX_NUMNODES <= 256 | |
36 | static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned; | |
37 | #else | |
38 | static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned; | |
39 | #endif | |
40 | ||
25ba77c1 | 41 | int page_to_nid(struct page *page) |
89689ae7 CL |
42 | { |
43 | return section_to_node_table[page_to_section(page)]; | |
44 | } | |
45 | EXPORT_SYMBOL(page_to_nid); | |
85770ffe AW |
46 | |
47 | static void set_section_nid(unsigned long section_nr, int nid) | |
48 | { | |
49 | section_to_node_table[section_nr] = nid; | |
50 | } | |
51 | #else /* !NODE_NOT_IN_PAGE_FLAGS */ | |
52 | static inline void set_section_nid(unsigned long section_nr, int nid) | |
53 | { | |
54 | } | |
89689ae7 CL |
55 | #endif |
56 | ||
3e347261 | 57 | #ifdef CONFIG_SPARSEMEM_EXTREME |
577a32f6 | 58 | static struct mem_section noinline __init_refok *sparse_index_alloc(int nid) |
28ae55c9 DH |
59 | { |
60 | struct mem_section *section = NULL; | |
61 | unsigned long array_size = SECTIONS_PER_ROOT * | |
62 | sizeof(struct mem_section); | |
63 | ||
39d24e64 | 64 | if (slab_is_available()) |
46a66eec MK |
65 | section = kmalloc_node(array_size, GFP_KERNEL, nid); |
66 | else | |
67 | section = alloc_bootmem_node(NODE_DATA(nid), array_size); | |
28ae55c9 DH |
68 | |
69 | if (section) | |
70 | memset(section, 0, array_size); | |
71 | ||
72 | return section; | |
3e347261 | 73 | } |
802f192e | 74 | |
a3142c8e | 75 | static int __meminit sparse_index_init(unsigned long section_nr, int nid) |
802f192e | 76 | { |
34af946a | 77 | static DEFINE_SPINLOCK(index_init_lock); |
28ae55c9 DH |
78 | unsigned long root = SECTION_NR_TO_ROOT(section_nr); |
79 | struct mem_section *section; | |
80 | int ret = 0; | |
802f192e BP |
81 | |
82 | if (mem_section[root]) | |
28ae55c9 | 83 | return -EEXIST; |
3e347261 | 84 | |
28ae55c9 | 85 | section = sparse_index_alloc(nid); |
af0cd5a7 WC |
86 | if (!section) |
87 | return -ENOMEM; | |
28ae55c9 DH |
88 | /* |
89 | * This lock keeps two different sections from | |
90 | * reallocating for the same index | |
91 | */ | |
92 | spin_lock(&index_init_lock); | |
3e347261 | 93 | |
28ae55c9 DH |
94 | if (mem_section[root]) { |
95 | ret = -EEXIST; | |
96 | goto out; | |
97 | } | |
98 | ||
99 | mem_section[root] = section; | |
100 | out: | |
101 | spin_unlock(&index_init_lock); | |
102 | return ret; | |
103 | } | |
104 | #else /* !SPARSEMEM_EXTREME */ | |
105 | static inline int sparse_index_init(unsigned long section_nr, int nid) | |
106 | { | |
107 | return 0; | |
802f192e | 108 | } |
28ae55c9 DH |
109 | #endif |
110 | ||
4ca644d9 DH |
111 | /* |
112 | * Although written for the SPARSEMEM_EXTREME case, this happens | |
cd881a6b | 113 | * to also work for the flat array case because |
4ca644d9 DH |
114 | * NR_SECTION_ROOTS==NR_MEM_SECTIONS. |
115 | */ | |
116 | int __section_nr(struct mem_section* ms) | |
117 | { | |
118 | unsigned long root_nr; | |
119 | struct mem_section* root; | |
120 | ||
12783b00 MK |
121 | for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) { |
122 | root = __nr_to_section(root_nr * SECTIONS_PER_ROOT); | |
4ca644d9 DH |
123 | if (!root) |
124 | continue; | |
125 | ||
126 | if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT))) | |
127 | break; | |
128 | } | |
129 | ||
130 | return (root_nr * SECTIONS_PER_ROOT) + (ms - root); | |
131 | } | |
132 | ||
30c253e6 AW |
133 | /* |
134 | * During early boot, before section_mem_map is used for an actual | |
135 | * mem_map, we use section_mem_map to store the section's NUMA | |
136 | * node. This keeps us from having to use another data structure. The | |
137 | * node information is cleared just before we store the real mem_map. | |
138 | */ | |
139 | static inline unsigned long sparse_encode_early_nid(int nid) | |
140 | { | |
141 | return (nid << SECTION_NID_SHIFT); | |
142 | } | |
143 | ||
144 | static inline int sparse_early_nid(struct mem_section *section) | |
145 | { | |
146 | return (section->section_mem_map >> SECTION_NID_SHIFT); | |
147 | } | |
148 | ||
d41dee36 | 149 | /* Record a memory area against a node. */ |
a3142c8e | 150 | void __init memory_present(int nid, unsigned long start, unsigned long end) |
d41dee36 | 151 | { |
bead9a3a | 152 | unsigned long max_arch_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT); |
d41dee36 AW |
153 | unsigned long pfn; |
154 | ||
bead9a3a IM |
155 | /* |
156 | * Sanity checks - do not allow an architecture to pass | |
157 | * in larger pfns than the maximum scope of sparsemem: | |
158 | */ | |
159 | if (start >= max_arch_pfn) | |
160 | return; | |
161 | if (end >= max_arch_pfn) | |
162 | end = max_arch_pfn; | |
163 | ||
d41dee36 AW |
164 | start &= PAGE_SECTION_MASK; |
165 | for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) { | |
166 | unsigned long section = pfn_to_section_nr(pfn); | |
802f192e BP |
167 | struct mem_section *ms; |
168 | ||
169 | sparse_index_init(section, nid); | |
85770ffe | 170 | set_section_nid(section, nid); |
802f192e BP |
171 | |
172 | ms = __nr_to_section(section); | |
173 | if (!ms->section_mem_map) | |
30c253e6 AW |
174 | ms->section_mem_map = sparse_encode_early_nid(nid) | |
175 | SECTION_MARKED_PRESENT; | |
d41dee36 AW |
176 | } |
177 | } | |
178 | ||
179 | /* | |
180 | * Only used by the i386 NUMA architecures, but relatively | |
181 | * generic code. | |
182 | */ | |
183 | unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn, | |
184 | unsigned long end_pfn) | |
185 | { | |
186 | unsigned long pfn; | |
187 | unsigned long nr_pages = 0; | |
188 | ||
189 | for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { | |
190 | if (nid != early_pfn_to_nid(pfn)) | |
191 | continue; | |
192 | ||
540557b9 | 193 | if (pfn_present(pfn)) |
d41dee36 AW |
194 | nr_pages += PAGES_PER_SECTION; |
195 | } | |
196 | ||
197 | return nr_pages * sizeof(struct page); | |
198 | } | |
199 | ||
29751f69 AW |
200 | /* |
201 | * Subtle, we encode the real pfn into the mem_map such that | |
202 | * the identity pfn - section_mem_map will return the actual | |
203 | * physical page frame number. | |
204 | */ | |
205 | static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum) | |
206 | { | |
207 | return (unsigned long)(mem_map - (section_nr_to_pfn(pnum))); | |
208 | } | |
209 | ||
210 | /* | |
ea01ea93 | 211 | * Decode mem_map from the coded memmap |
29751f69 | 212 | */ |
ea01ea93 | 213 | static |
29751f69 AW |
214 | struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum) |
215 | { | |
ea01ea93 BP |
216 | /* mask off the extra low bits of information */ |
217 | coded_mem_map &= SECTION_MAP_MASK; | |
29751f69 AW |
218 | return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum); |
219 | } | |
220 | ||
a3142c8e | 221 | static int __meminit sparse_init_one_section(struct mem_section *ms, |
5c0e3066 MG |
222 | unsigned long pnum, struct page *mem_map, |
223 | unsigned long *pageblock_bitmap) | |
29751f69 | 224 | { |
540557b9 | 225 | if (!present_section(ms)) |
29751f69 AW |
226 | return -EINVAL; |
227 | ||
30c253e6 | 228 | ms->section_mem_map &= ~SECTION_MAP_MASK; |
540557b9 AW |
229 | ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) | |
230 | SECTION_HAS_MEM_MAP; | |
5c0e3066 | 231 | ms->pageblock_flags = pageblock_bitmap; |
29751f69 AW |
232 | |
233 | return 1; | |
234 | } | |
235 | ||
5c0e3066 MG |
236 | static unsigned long usemap_size(void) |
237 | { | |
238 | unsigned long size_bytes; | |
239 | size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8; | |
240 | size_bytes = roundup(size_bytes, sizeof(unsigned long)); | |
241 | return size_bytes; | |
242 | } | |
243 | ||
244 | #ifdef CONFIG_MEMORY_HOTPLUG | |
245 | static unsigned long *__kmalloc_section_usemap(void) | |
246 | { | |
247 | return kmalloc(usemap_size(), GFP_KERNEL); | |
248 | } | |
249 | #endif /* CONFIG_MEMORY_HOTPLUG */ | |
250 | ||
a322f8ab | 251 | static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum) |
5c0e3066 MG |
252 | { |
253 | unsigned long *usemap; | |
254 | struct mem_section *ms = __nr_to_section(pnum); | |
255 | int nid = sparse_early_nid(ms); | |
256 | ||
257 | usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size()); | |
258 | if (usemap) | |
259 | return usemap; | |
260 | ||
261 | /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */ | |
262 | nid = 0; | |
263 | ||
264 | printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__); | |
265 | return NULL; | |
266 | } | |
267 | ||
8f6aac41 | 268 | #ifndef CONFIG_SPARSEMEM_VMEMMAP |
98f3cfc1 | 269 | struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid) |
29751f69 AW |
270 | { |
271 | struct page *map; | |
29751f69 AW |
272 | |
273 | map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION); | |
274 | if (map) | |
275 | return map; | |
276 | ||
277 | map = alloc_bootmem_node(NODE_DATA(nid), | |
278 | sizeof(struct page) * PAGES_PER_SECTION); | |
8f6aac41 CL |
279 | return map; |
280 | } | |
281 | #endif /* !CONFIG_SPARSEMEM_VMEMMAP */ | |
282 | ||
283 | struct page __init *sparse_early_mem_map_alloc(unsigned long pnum) | |
284 | { | |
285 | struct page *map; | |
286 | struct mem_section *ms = __nr_to_section(pnum); | |
287 | int nid = sparse_early_nid(ms); | |
288 | ||
98f3cfc1 | 289 | map = sparse_mem_map_populate(pnum, nid); |
29751f69 AW |
290 | if (map) |
291 | return map; | |
292 | ||
8f6aac41 CL |
293 | printk(KERN_ERR "%s: sparsemem memory map backing failed " |
294 | "some memory will not be available.\n", __FUNCTION__); | |
802f192e | 295 | ms->section_mem_map = 0; |
29751f69 AW |
296 | return NULL; |
297 | } | |
298 | ||
c2b91e2e YL |
299 | void __attribute__((weak)) __meminit vmemmap_populate_print_last(void) |
300 | { | |
301 | } | |
193faea9 SR |
302 | /* |
303 | * Allocate the accumulated non-linear sections, allocate a mem_map | |
304 | * for each and record the physical to section mapping. | |
305 | */ | |
306 | void __init sparse_init(void) | |
307 | { | |
308 | unsigned long pnum; | |
309 | struct page *map; | |
5c0e3066 | 310 | unsigned long *usemap; |
e123dd3f YL |
311 | unsigned long **usemap_map; |
312 | int size; | |
313 | ||
314 | /* | |
315 | * map is using big page (aka 2M in x86 64 bit) | |
316 | * usemap is less one page (aka 24 bytes) | |
317 | * so alloc 2M (with 2M align) and 24 bytes in turn will | |
318 | * make next 2M slip to one more 2M later. | |
319 | * then in big system, the memory will have a lot of holes... | |
320 | * here try to allocate 2M pages continously. | |
321 | * | |
322 | * powerpc need to call sparse_init_one_section right after each | |
323 | * sparse_early_mem_map_alloc, so allocate usemap_map at first. | |
324 | */ | |
325 | size = sizeof(unsigned long *) * NR_MEM_SECTIONS; | |
326 | usemap_map = alloc_bootmem(size); | |
327 | if (!usemap_map) | |
328 | panic("can not allocate usemap_map\n"); | |
193faea9 SR |
329 | |
330 | for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) { | |
540557b9 | 331 | if (!present_section_nr(pnum)) |
193faea9 | 332 | continue; |
e123dd3f YL |
333 | usemap_map[pnum] = sparse_early_usemap_alloc(pnum); |
334 | } | |
193faea9 | 335 | |
e123dd3f YL |
336 | for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) { |
337 | if (!present_section_nr(pnum)) | |
193faea9 | 338 | continue; |
5c0e3066 | 339 | |
e123dd3f | 340 | usemap = usemap_map[pnum]; |
5c0e3066 MG |
341 | if (!usemap) |
342 | continue; | |
343 | ||
e123dd3f YL |
344 | map = sparse_early_mem_map_alloc(pnum); |
345 | if (!map) | |
346 | continue; | |
347 | ||
5c0e3066 MG |
348 | sparse_init_one_section(__nr_to_section(pnum), pnum, map, |
349 | usemap); | |
193faea9 | 350 | } |
e123dd3f | 351 | |
c2b91e2e YL |
352 | vmemmap_populate_print_last(); |
353 | ||
e123dd3f | 354 | free_bootmem(__pa(usemap_map), size); |
193faea9 SR |
355 | } |
356 | ||
357 | #ifdef CONFIG_MEMORY_HOTPLUG | |
98f3cfc1 YG |
358 | #ifdef CONFIG_SPARSEMEM_VMEMMAP |
359 | static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid, | |
360 | unsigned long nr_pages) | |
361 | { | |
362 | /* This will make the necessary allocations eventually. */ | |
363 | return sparse_mem_map_populate(pnum, nid); | |
364 | } | |
365 | static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages) | |
366 | { | |
367 | return; /* XXX: Not implemented yet */ | |
368 | } | |
369 | #else | |
0b0acbec DH |
370 | static struct page *__kmalloc_section_memmap(unsigned long nr_pages) |
371 | { | |
372 | struct page *page, *ret; | |
373 | unsigned long memmap_size = sizeof(struct page) * nr_pages; | |
374 | ||
f2d0aa5b | 375 | page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size)); |
0b0acbec DH |
376 | if (page) |
377 | goto got_map_page; | |
378 | ||
379 | ret = vmalloc(memmap_size); | |
380 | if (ret) | |
381 | goto got_map_ptr; | |
382 | ||
383 | return NULL; | |
384 | got_map_page: | |
385 | ret = (struct page *)pfn_to_kaddr(page_to_pfn(page)); | |
386 | got_map_ptr: | |
387 | memset(ret, 0, memmap_size); | |
388 | ||
389 | return ret; | |
390 | } | |
391 | ||
98f3cfc1 YG |
392 | static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid, |
393 | unsigned long nr_pages) | |
394 | { | |
395 | return __kmalloc_section_memmap(nr_pages); | |
396 | } | |
397 | ||
0b0acbec DH |
398 | static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages) |
399 | { | |
9e2779fa | 400 | if (is_vmalloc_addr(memmap)) |
0b0acbec DH |
401 | vfree(memmap); |
402 | else | |
403 | free_pages((unsigned long)memmap, | |
404 | get_order(sizeof(struct page) * nr_pages)); | |
405 | } | |
98f3cfc1 | 406 | #endif /* CONFIG_SPARSEMEM_VMEMMAP */ |
0b0acbec | 407 | |
ea01ea93 BP |
408 | static void free_section_usemap(struct page *memmap, unsigned long *usemap) |
409 | { | |
410 | if (!usemap) | |
411 | return; | |
412 | ||
413 | /* | |
414 | * Check to see if allocation came from hot-plug-add | |
415 | */ | |
416 | if (PageSlab(virt_to_page(usemap))) { | |
417 | kfree(usemap); | |
418 | if (memmap) | |
419 | __kfree_section_memmap(memmap, PAGES_PER_SECTION); | |
420 | return; | |
421 | } | |
422 | ||
423 | /* | |
424 | * TODO: Allocations came from bootmem - how do I free up ? | |
425 | */ | |
426 | printk(KERN_WARNING "Not freeing up allocations from bootmem " | |
427 | "- leaking memory\n"); | |
428 | } | |
429 | ||
29751f69 AW |
430 | /* |
431 | * returns the number of sections whose mem_maps were properly | |
432 | * set. If this is <=0, then that means that the passed-in | |
433 | * map was not consumed and must be freed. | |
434 | */ | |
0b0acbec DH |
435 | int sparse_add_one_section(struct zone *zone, unsigned long start_pfn, |
436 | int nr_pages) | |
29751f69 | 437 | { |
0b0acbec DH |
438 | unsigned long section_nr = pfn_to_section_nr(start_pfn); |
439 | struct pglist_data *pgdat = zone->zone_pgdat; | |
440 | struct mem_section *ms; | |
441 | struct page *memmap; | |
5c0e3066 | 442 | unsigned long *usemap; |
0b0acbec DH |
443 | unsigned long flags; |
444 | int ret; | |
29751f69 | 445 | |
0b0acbec DH |
446 | /* |
447 | * no locking for this, because it does its own | |
448 | * plus, it does a kmalloc | |
449 | */ | |
bbd06825 WC |
450 | ret = sparse_index_init(section_nr, pgdat->node_id); |
451 | if (ret < 0 && ret != -EEXIST) | |
452 | return ret; | |
98f3cfc1 | 453 | memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages); |
bbd06825 WC |
454 | if (!memmap) |
455 | return -ENOMEM; | |
5c0e3066 | 456 | usemap = __kmalloc_section_usemap(); |
bbd06825 WC |
457 | if (!usemap) { |
458 | __kfree_section_memmap(memmap, nr_pages); | |
459 | return -ENOMEM; | |
460 | } | |
0b0acbec DH |
461 | |
462 | pgdat_resize_lock(pgdat, &flags); | |
29751f69 | 463 | |
0b0acbec DH |
464 | ms = __pfn_to_section(start_pfn); |
465 | if (ms->section_mem_map & SECTION_MARKED_PRESENT) { | |
466 | ret = -EEXIST; | |
467 | goto out; | |
468 | } | |
5c0e3066 | 469 | |
29751f69 AW |
470 | ms->section_mem_map |= SECTION_MARKED_PRESENT; |
471 | ||
5c0e3066 | 472 | ret = sparse_init_one_section(ms, section_nr, memmap, usemap); |
0b0acbec | 473 | |
0b0acbec DH |
474 | out: |
475 | pgdat_resize_unlock(pgdat, &flags); | |
bbd06825 WC |
476 | if (ret <= 0) { |
477 | kfree(usemap); | |
46a66eec | 478 | __kfree_section_memmap(memmap, nr_pages); |
bbd06825 | 479 | } |
0b0acbec | 480 | return ret; |
29751f69 | 481 | } |
ea01ea93 BP |
482 | |
483 | void sparse_remove_one_section(struct zone *zone, struct mem_section *ms) | |
484 | { | |
485 | struct page *memmap = NULL; | |
486 | unsigned long *usemap = NULL; | |
487 | ||
488 | if (ms->section_mem_map) { | |
489 | usemap = ms->pageblock_flags; | |
490 | memmap = sparse_decode_mem_map(ms->section_mem_map, | |
491 | __section_nr(ms)); | |
492 | ms->section_mem_map = 0; | |
493 | ms->pageblock_flags = NULL; | |
494 | } | |
495 | ||
496 | free_section_usemap(memmap, usemap); | |
497 | } | |
a3142c8e | 498 | #endif |