Loading...
1/*
2 * sparse memory mappings.
3 */
4#include <linux/mm.h>
5#include <linux/slab.h>
6#include <linux/mmzone.h>
7#include <linux/bootmem.h>
8#include <linux/highmem.h>
9#include <linux/module.h>
10#include <linux/spinlock.h>
11#include <linux/vmalloc.h>
12#include "internal.h"
13#include <asm/dma.h>
14#include <asm/pgalloc.h>
15#include <asm/pgtable.h>
16
17/*
18 * Permanent SPARSEMEM data:
19 *
20 * 1) mem_section - memory sections, mem_map's for valid memory
21 */
22#ifdef CONFIG_SPARSEMEM_EXTREME
23struct mem_section *mem_section[NR_SECTION_ROOTS]
24 ____cacheline_internodealigned_in_smp;
25#else
26struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
27 ____cacheline_internodealigned_in_smp;
28#endif
29EXPORT_SYMBOL(mem_section);
30
31#ifdef NODE_NOT_IN_PAGE_FLAGS
32/*
33 * If we did not store the node number in the page then we have to
34 * do a lookup in the section_to_node_table in order to find which
35 * node the page belongs to.
36 */
37#if MAX_NUMNODES <= 256
38static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
39#else
40static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
41#endif
42
43int page_to_nid(const struct page *page)
44{
45 return section_to_node_table[page_to_section(page)];
46}
47EXPORT_SYMBOL(page_to_nid);
48
49static void set_section_nid(unsigned long section_nr, int nid)
50{
51 section_to_node_table[section_nr] = nid;
52}
53#else /* !NODE_NOT_IN_PAGE_FLAGS */
54static inline void set_section_nid(unsigned long section_nr, int nid)
55{
56}
57#endif
58
59#ifdef CONFIG_SPARSEMEM_EXTREME
60static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
61{
62 struct mem_section *section = NULL;
63 unsigned long array_size = SECTIONS_PER_ROOT *
64 sizeof(struct mem_section);
65
66 if (slab_is_available()) {
67 if (node_state(nid, N_HIGH_MEMORY))
68 section = kmalloc_node(array_size, GFP_KERNEL, nid);
69 else
70 section = kmalloc(array_size, GFP_KERNEL);
71 } else
72 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
73
74 if (section)
75 memset(section, 0, array_size);
76
77 return section;
78}
79
80static int __meminit sparse_index_init(unsigned long section_nr, int nid)
81{
82 static DEFINE_SPINLOCK(index_init_lock);
83 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
84 struct mem_section *section;
85 int ret = 0;
86
87 if (mem_section[root])
88 return -EEXIST;
89
90 section = sparse_index_alloc(nid);
91 if (!section)
92 return -ENOMEM;
93 /*
94 * This lock keeps two different sections from
95 * reallocating for the same index
96 */
97 spin_lock(&index_init_lock);
98
99 if (mem_section[root]) {
100 ret = -EEXIST;
101 goto out;
102 }
103
104 mem_section[root] = section;
105out:
106 spin_unlock(&index_init_lock);
107 return ret;
108}
109#else /* !SPARSEMEM_EXTREME */
110static inline int sparse_index_init(unsigned long section_nr, int nid)
111{
112 return 0;
113}
114#endif
115
116/*
117 * Although written for the SPARSEMEM_EXTREME case, this happens
118 * to also work for the flat array case because
119 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
120 */
121int __section_nr(struct mem_section* ms)
122{
123 unsigned long root_nr;
124 struct mem_section* root;
125
126 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
127 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
128 if (!root)
129 continue;
130
131 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
132 break;
133 }
134
135 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
136}
137
138/*
139 * During early boot, before section_mem_map is used for an actual
140 * mem_map, we use section_mem_map to store the section's NUMA
141 * node. This keeps us from having to use another data structure. The
142 * node information is cleared just before we store the real mem_map.
143 */
144static inline unsigned long sparse_encode_early_nid(int nid)
145{
146 return (nid << SECTION_NID_SHIFT);
147}
148
149static inline int sparse_early_nid(struct mem_section *section)
150{
151 return (section->section_mem_map >> SECTION_NID_SHIFT);
152}
153
154/* Validate the physical addressing limitations of the model */
155void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
156 unsigned long *end_pfn)
157{
158 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
159
160 /*
161 * Sanity checks - do not allow an architecture to pass
162 * in larger pfns than the maximum scope of sparsemem:
163 */
164 if (*start_pfn > max_sparsemem_pfn) {
165 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
166 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
167 *start_pfn, *end_pfn, max_sparsemem_pfn);
168 WARN_ON_ONCE(1);
169 *start_pfn = max_sparsemem_pfn;
170 *end_pfn = max_sparsemem_pfn;
171 } else if (*end_pfn > max_sparsemem_pfn) {
172 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
173 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
174 *start_pfn, *end_pfn, max_sparsemem_pfn);
175 WARN_ON_ONCE(1);
176 *end_pfn = max_sparsemem_pfn;
177 }
178}
179
180/* Record a memory area against a node. */
181void __init memory_present(int nid, unsigned long start, unsigned long end)
182{
183 unsigned long pfn;
184
185 start &= PAGE_SECTION_MASK;
186 mminit_validate_memmodel_limits(&start, &end);
187 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
188 unsigned long section = pfn_to_section_nr(pfn);
189 struct mem_section *ms;
190
191 sparse_index_init(section, nid);
192 set_section_nid(section, nid);
193
194 ms = __nr_to_section(section);
195 if (!ms->section_mem_map)
196 ms->section_mem_map = sparse_encode_early_nid(nid) |
197 SECTION_MARKED_PRESENT;
198 }
199}
200
201/*
202 * Only used by the i386 NUMA architecures, but relatively
203 * generic code.
204 */
205unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
206 unsigned long end_pfn)
207{
208 unsigned long pfn;
209 unsigned long nr_pages = 0;
210
211 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
212 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
213 if (nid != early_pfn_to_nid(pfn))
214 continue;
215
216 if (pfn_present(pfn))
217 nr_pages += PAGES_PER_SECTION;
218 }
219
220 return nr_pages * sizeof(struct page);
221}
222
223/*
224 * Subtle, we encode the real pfn into the mem_map such that
225 * the identity pfn - section_mem_map will return the actual
226 * physical page frame number.
227 */
228static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
229{
230 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
231}
232
233/*
234 * Decode mem_map from the coded memmap
235 */
236struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
237{
238 /* mask off the extra low bits of information */
239 coded_mem_map &= SECTION_MAP_MASK;
240 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
241}
242
243static int __meminit sparse_init_one_section(struct mem_section *ms,
244 unsigned long pnum, struct page *mem_map,
245 unsigned long *pageblock_bitmap)
246{
247 if (!present_section(ms))
248 return -EINVAL;
249
250 ms->section_mem_map &= ~SECTION_MAP_MASK;
251 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
252 SECTION_HAS_MEM_MAP;
253 ms->pageblock_flags = pageblock_bitmap;
254
255 return 1;
256}
257
258unsigned long usemap_size(void)
259{
260 unsigned long size_bytes;
261 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
262 size_bytes = roundup(size_bytes, sizeof(unsigned long));
263 return size_bytes;
264}
265
266#ifdef CONFIG_MEMORY_HOTPLUG
267static unsigned long *__kmalloc_section_usemap(void)
268{
269 return kmalloc(usemap_size(), GFP_KERNEL);
270}
271#endif /* CONFIG_MEMORY_HOTPLUG */
272
273#ifdef CONFIG_MEMORY_HOTREMOVE
274static unsigned long * __init
275sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
276 unsigned long count)
277{
278 unsigned long section_nr;
279
280 /*
281 * A page may contain usemaps for other sections preventing the
282 * page being freed and making a section unremovable while
283 * other sections referencing the usemap retmain active. Similarly,
284 * a pgdat can prevent a section being removed. If section A
285 * contains a pgdat and section B contains the usemap, both
286 * sections become inter-dependent. This allocates usemaps
287 * from the same section as the pgdat where possible to avoid
288 * this problem.
289 */
290 section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
291 return alloc_bootmem_section(usemap_size() * count, section_nr);
292}
293
294static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
295{
296 unsigned long usemap_snr, pgdat_snr;
297 static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
298 static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
299 struct pglist_data *pgdat = NODE_DATA(nid);
300 int usemap_nid;
301
302 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
303 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
304 if (usemap_snr == pgdat_snr)
305 return;
306
307 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
308 /* skip redundant message */
309 return;
310
311 old_usemap_snr = usemap_snr;
312 old_pgdat_snr = pgdat_snr;
313
314 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
315 if (usemap_nid != nid) {
316 printk(KERN_INFO
317 "node %d must be removed before remove section %ld\n",
318 nid, usemap_snr);
319 return;
320 }
321 /*
322 * There is a circular dependency.
323 * Some platforms allow un-removable section because they will just
324 * gather other removable sections for dynamic partitioning.
325 * Just notify un-removable section's number here.
326 */
327 printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
328 pgdat_snr, nid);
329 printk(KERN_CONT
330 " have a circular dependency on usemap and pgdat allocations\n");
331}
332#else
333static unsigned long * __init
334sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
335 unsigned long count)
336{
337 return NULL;
338}
339
340static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
341{
342}
343#endif /* CONFIG_MEMORY_HOTREMOVE */
344
345static void __init sparse_early_usemaps_alloc_node(unsigned long**usemap_map,
346 unsigned long pnum_begin,
347 unsigned long pnum_end,
348 unsigned long usemap_count, int nodeid)
349{
350 void *usemap;
351 unsigned long pnum;
352 int size = usemap_size();
353
354 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
355 usemap_count);
356 if (usemap) {
357 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
358 if (!present_section_nr(pnum))
359 continue;
360 usemap_map[pnum] = usemap;
361 usemap += size;
362 }
363 return;
364 }
365
366 usemap = alloc_bootmem_node(NODE_DATA(nodeid), size * usemap_count);
367 if (usemap) {
368 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
369 if (!present_section_nr(pnum))
370 continue;
371 usemap_map[pnum] = usemap;
372 usemap += size;
373 check_usemap_section_nr(nodeid, usemap_map[pnum]);
374 }
375 return;
376 }
377
378 printk(KERN_WARNING "%s: allocation failed\n", __func__);
379}
380
381#ifndef CONFIG_SPARSEMEM_VMEMMAP
382struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
383{
384 struct page *map;
385 unsigned long size;
386
387 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
388 if (map)
389 return map;
390
391 size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
392 map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
393 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
394 return map;
395}
396void __init sparse_mem_maps_populate_node(struct page **map_map,
397 unsigned long pnum_begin,
398 unsigned long pnum_end,
399 unsigned long map_count, int nodeid)
400{
401 void *map;
402 unsigned long pnum;
403 unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
404
405 map = alloc_remap(nodeid, size * map_count);
406 if (map) {
407 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
408 if (!present_section_nr(pnum))
409 continue;
410 map_map[pnum] = map;
411 map += size;
412 }
413 return;
414 }
415
416 size = PAGE_ALIGN(size);
417 map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
418 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
419 if (map) {
420 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
421 if (!present_section_nr(pnum))
422 continue;
423 map_map[pnum] = map;
424 map += size;
425 }
426 return;
427 }
428
429 /* fallback */
430 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
431 struct mem_section *ms;
432
433 if (!present_section_nr(pnum))
434 continue;
435 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
436 if (map_map[pnum])
437 continue;
438 ms = __nr_to_section(pnum);
439 printk(KERN_ERR "%s: sparsemem memory map backing failed "
440 "some memory will not be available.\n", __func__);
441 ms->section_mem_map = 0;
442 }
443}
444#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
445
446#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
447static void __init sparse_early_mem_maps_alloc_node(struct page **map_map,
448 unsigned long pnum_begin,
449 unsigned long pnum_end,
450 unsigned long map_count, int nodeid)
451{
452 sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
453 map_count, nodeid);
454}
455#else
456static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
457{
458 struct page *map;
459 struct mem_section *ms = __nr_to_section(pnum);
460 int nid = sparse_early_nid(ms);
461
462 map = sparse_mem_map_populate(pnum, nid);
463 if (map)
464 return map;
465
466 printk(KERN_ERR "%s: sparsemem memory map backing failed "
467 "some memory will not be available.\n", __func__);
468 ms->section_mem_map = 0;
469 return NULL;
470}
471#endif
472
473void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
474{
475}
476
477/*
478 * Allocate the accumulated non-linear sections, allocate a mem_map
479 * for each and record the physical to section mapping.
480 */
481void __init sparse_init(void)
482{
483 unsigned long pnum;
484 struct page *map;
485 unsigned long *usemap;
486 unsigned long **usemap_map;
487 int size;
488 int nodeid_begin = 0;
489 unsigned long pnum_begin = 0;
490 unsigned long usemap_count;
491#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
492 unsigned long map_count;
493 int size2;
494 struct page **map_map;
495#endif
496
497 /*
498 * map is using big page (aka 2M in x86 64 bit)
499 * usemap is less one page (aka 24 bytes)
500 * so alloc 2M (with 2M align) and 24 bytes in turn will
501 * make next 2M slip to one more 2M later.
502 * then in big system, the memory will have a lot of holes...
503 * here try to allocate 2M pages continuously.
504 *
505 * powerpc need to call sparse_init_one_section right after each
506 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
507 */
508 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
509 usemap_map = alloc_bootmem(size);
510 if (!usemap_map)
511 panic("can not allocate usemap_map\n");
512
513 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
514 struct mem_section *ms;
515
516 if (!present_section_nr(pnum))
517 continue;
518 ms = __nr_to_section(pnum);
519 nodeid_begin = sparse_early_nid(ms);
520 pnum_begin = pnum;
521 break;
522 }
523 usemap_count = 1;
524 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
525 struct mem_section *ms;
526 int nodeid;
527
528 if (!present_section_nr(pnum))
529 continue;
530 ms = __nr_to_section(pnum);
531 nodeid = sparse_early_nid(ms);
532 if (nodeid == nodeid_begin) {
533 usemap_count++;
534 continue;
535 }
536 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
537 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, pnum,
538 usemap_count, nodeid_begin);
539 /* new start, update count etc*/
540 nodeid_begin = nodeid;
541 pnum_begin = pnum;
542 usemap_count = 1;
543 }
544 /* ok, last chunk */
545 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS,
546 usemap_count, nodeid_begin);
547
548#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
549 size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
550 map_map = alloc_bootmem(size2);
551 if (!map_map)
552 panic("can not allocate map_map\n");
553
554 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
555 struct mem_section *ms;
556
557 if (!present_section_nr(pnum))
558 continue;
559 ms = __nr_to_section(pnum);
560 nodeid_begin = sparse_early_nid(ms);
561 pnum_begin = pnum;
562 break;
563 }
564 map_count = 1;
565 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
566 struct mem_section *ms;
567 int nodeid;
568
569 if (!present_section_nr(pnum))
570 continue;
571 ms = __nr_to_section(pnum);
572 nodeid = sparse_early_nid(ms);
573 if (nodeid == nodeid_begin) {
574 map_count++;
575 continue;
576 }
577 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
578 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, pnum,
579 map_count, nodeid_begin);
580 /* new start, update count etc*/
581 nodeid_begin = nodeid;
582 pnum_begin = pnum;
583 map_count = 1;
584 }
585 /* ok, last chunk */
586 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS,
587 map_count, nodeid_begin);
588#endif
589
590 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
591 if (!present_section_nr(pnum))
592 continue;
593
594 usemap = usemap_map[pnum];
595 if (!usemap)
596 continue;
597
598#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
599 map = map_map[pnum];
600#else
601 map = sparse_early_mem_map_alloc(pnum);
602#endif
603 if (!map)
604 continue;
605
606 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
607 usemap);
608 }
609
610 vmemmap_populate_print_last();
611
612#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
613 free_bootmem(__pa(map_map), size2);
614#endif
615 free_bootmem(__pa(usemap_map), size);
616}
617
618#ifdef CONFIG_MEMORY_HOTPLUG
619#ifdef CONFIG_SPARSEMEM_VMEMMAP
620static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
621 unsigned long nr_pages)
622{
623 /* This will make the necessary allocations eventually. */
624 return sparse_mem_map_populate(pnum, nid);
625}
626static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
627{
628 return; /* XXX: Not implemented yet */
629}
630static void free_map_bootmem(struct page *page, unsigned long nr_pages)
631{
632}
633#else
634static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
635{
636 struct page *page, *ret;
637 unsigned long memmap_size = sizeof(struct page) * nr_pages;
638
639 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
640 if (page)
641 goto got_map_page;
642
643 ret = vmalloc(memmap_size);
644 if (ret)
645 goto got_map_ptr;
646
647 return NULL;
648got_map_page:
649 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
650got_map_ptr:
651 memset(ret, 0, memmap_size);
652
653 return ret;
654}
655
656static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
657 unsigned long nr_pages)
658{
659 return __kmalloc_section_memmap(nr_pages);
660}
661
662static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
663{
664 if (is_vmalloc_addr(memmap))
665 vfree(memmap);
666 else
667 free_pages((unsigned long)memmap,
668 get_order(sizeof(struct page) * nr_pages));
669}
670
671static void free_map_bootmem(struct page *page, unsigned long nr_pages)
672{
673 unsigned long maps_section_nr, removing_section_nr, i;
674 unsigned long magic;
675
676 for (i = 0; i < nr_pages; i++, page++) {
677 magic = (unsigned long) page->lru.next;
678
679 BUG_ON(magic == NODE_INFO);
680
681 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
682 removing_section_nr = page->private;
683
684 /*
685 * When this function is called, the removing section is
686 * logical offlined state. This means all pages are isolated
687 * from page allocator. If removing section's memmap is placed
688 * on the same section, it must not be freed.
689 * If it is freed, page allocator may allocate it which will
690 * be removed physically soon.
691 */
692 if (maps_section_nr != removing_section_nr)
693 put_page_bootmem(page);
694 }
695}
696#endif /* CONFIG_SPARSEMEM_VMEMMAP */
697
698static void free_section_usemap(struct page *memmap, unsigned long *usemap)
699{
700 struct page *usemap_page;
701 unsigned long nr_pages;
702
703 if (!usemap)
704 return;
705
706 usemap_page = virt_to_page(usemap);
707 /*
708 * Check to see if allocation came from hot-plug-add
709 */
710 if (PageSlab(usemap_page)) {
711 kfree(usemap);
712 if (memmap)
713 __kfree_section_memmap(memmap, PAGES_PER_SECTION);
714 return;
715 }
716
717 /*
718 * The usemap came from bootmem. This is packed with other usemaps
719 * on the section which has pgdat at boot time. Just keep it as is now.
720 */
721
722 if (memmap) {
723 struct page *memmap_page;
724 memmap_page = virt_to_page(memmap);
725
726 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
727 >> PAGE_SHIFT;
728
729 free_map_bootmem(memmap_page, nr_pages);
730 }
731}
732
733/*
734 * returns the number of sections whose mem_maps were properly
735 * set. If this is <=0, then that means that the passed-in
736 * map was not consumed and must be freed.
737 */
738int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
739 int nr_pages)
740{
741 unsigned long section_nr = pfn_to_section_nr(start_pfn);
742 struct pglist_data *pgdat = zone->zone_pgdat;
743 struct mem_section *ms;
744 struct page *memmap;
745 unsigned long *usemap;
746 unsigned long flags;
747 int ret;
748
749 /*
750 * no locking for this, because it does its own
751 * plus, it does a kmalloc
752 */
753 ret = sparse_index_init(section_nr, pgdat->node_id);
754 if (ret < 0 && ret != -EEXIST)
755 return ret;
756 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
757 if (!memmap)
758 return -ENOMEM;
759 usemap = __kmalloc_section_usemap();
760 if (!usemap) {
761 __kfree_section_memmap(memmap, nr_pages);
762 return -ENOMEM;
763 }
764
765 pgdat_resize_lock(pgdat, &flags);
766
767 ms = __pfn_to_section(start_pfn);
768 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
769 ret = -EEXIST;
770 goto out;
771 }
772
773 ms->section_mem_map |= SECTION_MARKED_PRESENT;
774
775 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
776
777out:
778 pgdat_resize_unlock(pgdat, &flags);
779 if (ret <= 0) {
780 kfree(usemap);
781 __kfree_section_memmap(memmap, nr_pages);
782 }
783 return ret;
784}
785
786void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
787{
788 struct page *memmap = NULL;
789 unsigned long *usemap = NULL;
790
791 if (ms->section_mem_map) {
792 usemap = ms->pageblock_flags;
793 memmap = sparse_decode_mem_map(ms->section_mem_map,
794 __section_nr(ms));
795 ms->section_mem_map = 0;
796 ms->pageblock_flags = NULL;
797 }
798
799 free_section_usemap(memmap, usemap);
800}
801#endif
1/*
2 * sparse memory mappings.
3 */
4#include <linux/mm.h>
5#include <linux/slab.h>
6#include <linux/mmzone.h>
7#include <linux/bootmem.h>
8#include <linux/highmem.h>
9#include <linux/export.h>
10#include <linux/spinlock.h>
11#include <linux/vmalloc.h>
12#include "internal.h"
13#include <asm/dma.h>
14#include <asm/pgalloc.h>
15#include <asm/pgtable.h>
16
17/*
18 * Permanent SPARSEMEM data:
19 *
20 * 1) mem_section - memory sections, mem_map's for valid memory
21 */
22#ifdef CONFIG_SPARSEMEM_EXTREME
23struct mem_section *mem_section[NR_SECTION_ROOTS]
24 ____cacheline_internodealigned_in_smp;
25#else
26struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
27 ____cacheline_internodealigned_in_smp;
28#endif
29EXPORT_SYMBOL(mem_section);
30
31#ifdef NODE_NOT_IN_PAGE_FLAGS
32/*
33 * If we did not store the node number in the page then we have to
34 * do a lookup in the section_to_node_table in order to find which
35 * node the page belongs to.
36 */
37#if MAX_NUMNODES <= 256
38static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
39#else
40static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
41#endif
42
43int page_to_nid(const struct page *page)
44{
45 return section_to_node_table[page_to_section(page)];
46}
47EXPORT_SYMBOL(page_to_nid);
48
49static void set_section_nid(unsigned long section_nr, int nid)
50{
51 section_to_node_table[section_nr] = nid;
52}
53#else /* !NODE_NOT_IN_PAGE_FLAGS */
54static inline void set_section_nid(unsigned long section_nr, int nid)
55{
56}
57#endif
58
59#ifdef CONFIG_SPARSEMEM_EXTREME
60static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
61{
62 struct mem_section *section = NULL;
63 unsigned long array_size = SECTIONS_PER_ROOT *
64 sizeof(struct mem_section);
65
66 if (slab_is_available()) {
67 if (node_state(nid, N_HIGH_MEMORY))
68 section = kmalloc_node(array_size, GFP_KERNEL, nid);
69 else
70 section = kmalloc(array_size, GFP_KERNEL);
71 } else
72 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
73
74 if (section)
75 memset(section, 0, array_size);
76
77 return section;
78}
79
80static int __meminit sparse_index_init(unsigned long section_nr, int nid)
81{
82 static DEFINE_SPINLOCK(index_init_lock);
83 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
84 struct mem_section *section;
85 int ret = 0;
86
87 if (mem_section[root])
88 return -EEXIST;
89
90 section = sparse_index_alloc(nid);
91 if (!section)
92 return -ENOMEM;
93 /*
94 * This lock keeps two different sections from
95 * reallocating for the same index
96 */
97 spin_lock(&index_init_lock);
98
99 if (mem_section[root]) {
100 ret = -EEXIST;
101 goto out;
102 }
103
104 mem_section[root] = section;
105out:
106 spin_unlock(&index_init_lock);
107 return ret;
108}
109#else /* !SPARSEMEM_EXTREME */
110static inline int sparse_index_init(unsigned long section_nr, int nid)
111{
112 return 0;
113}
114#endif
115
116/*
117 * Although written for the SPARSEMEM_EXTREME case, this happens
118 * to also work for the flat array case because
119 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
120 */
121int __section_nr(struct mem_section* ms)
122{
123 unsigned long root_nr;
124 struct mem_section* root;
125
126 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
127 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
128 if (!root)
129 continue;
130
131 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
132 break;
133 }
134
135 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
136}
137
138/*
139 * During early boot, before section_mem_map is used for an actual
140 * mem_map, we use section_mem_map to store the section's NUMA
141 * node. This keeps us from having to use another data structure. The
142 * node information is cleared just before we store the real mem_map.
143 */
144static inline unsigned long sparse_encode_early_nid(int nid)
145{
146 return (nid << SECTION_NID_SHIFT);
147}
148
149static inline int sparse_early_nid(struct mem_section *section)
150{
151 return (section->section_mem_map >> SECTION_NID_SHIFT);
152}
153
154/* Validate the physical addressing limitations of the model */
155void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
156 unsigned long *end_pfn)
157{
158 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
159
160 /*
161 * Sanity checks - do not allow an architecture to pass
162 * in larger pfns than the maximum scope of sparsemem:
163 */
164 if (*start_pfn > max_sparsemem_pfn) {
165 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
166 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
167 *start_pfn, *end_pfn, max_sparsemem_pfn);
168 WARN_ON_ONCE(1);
169 *start_pfn = max_sparsemem_pfn;
170 *end_pfn = max_sparsemem_pfn;
171 } else if (*end_pfn > max_sparsemem_pfn) {
172 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
173 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
174 *start_pfn, *end_pfn, max_sparsemem_pfn);
175 WARN_ON_ONCE(1);
176 *end_pfn = max_sparsemem_pfn;
177 }
178}
179
180/* Record a memory area against a node. */
181void __init memory_present(int nid, unsigned long start, unsigned long end)
182{
183 unsigned long pfn;
184
185 start &= PAGE_SECTION_MASK;
186 mminit_validate_memmodel_limits(&start, &end);
187 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
188 unsigned long section = pfn_to_section_nr(pfn);
189 struct mem_section *ms;
190
191 sparse_index_init(section, nid);
192 set_section_nid(section, nid);
193
194 ms = __nr_to_section(section);
195 if (!ms->section_mem_map)
196 ms->section_mem_map = sparse_encode_early_nid(nid) |
197 SECTION_MARKED_PRESENT;
198 }
199}
200
201/*
202 * Only used by the i386 NUMA architecures, but relatively
203 * generic code.
204 */
205unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
206 unsigned long end_pfn)
207{
208 unsigned long pfn;
209 unsigned long nr_pages = 0;
210
211 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
212 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
213 if (nid != early_pfn_to_nid(pfn))
214 continue;
215
216 if (pfn_present(pfn))
217 nr_pages += PAGES_PER_SECTION;
218 }
219
220 return nr_pages * sizeof(struct page);
221}
222
223/*
224 * Subtle, we encode the real pfn into the mem_map such that
225 * the identity pfn - section_mem_map will return the actual
226 * physical page frame number.
227 */
228static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
229{
230 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
231}
232
233/*
234 * Decode mem_map from the coded memmap
235 */
236struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
237{
238 /* mask off the extra low bits of information */
239 coded_mem_map &= SECTION_MAP_MASK;
240 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
241}
242
243static int __meminit sparse_init_one_section(struct mem_section *ms,
244 unsigned long pnum, struct page *mem_map,
245 unsigned long *pageblock_bitmap)
246{
247 if (!present_section(ms))
248 return -EINVAL;
249
250 ms->section_mem_map &= ~SECTION_MAP_MASK;
251 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
252 SECTION_HAS_MEM_MAP;
253 ms->pageblock_flags = pageblock_bitmap;
254
255 return 1;
256}
257
258unsigned long usemap_size(void)
259{
260 unsigned long size_bytes;
261 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
262 size_bytes = roundup(size_bytes, sizeof(unsigned long));
263 return size_bytes;
264}
265
266#ifdef CONFIG_MEMORY_HOTPLUG
267static unsigned long *__kmalloc_section_usemap(void)
268{
269 return kmalloc(usemap_size(), GFP_KERNEL);
270}
271#endif /* CONFIG_MEMORY_HOTPLUG */
272
273#ifdef CONFIG_MEMORY_HOTREMOVE
274static unsigned long * __init
275sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
276 unsigned long size)
277{
278 unsigned long goal, limit;
279 unsigned long *p;
280 int nid;
281 /*
282 * A page may contain usemaps for other sections preventing the
283 * page being freed and making a section unremovable while
284 * other sections referencing the usemap retmain active. Similarly,
285 * a pgdat can prevent a section being removed. If section A
286 * contains a pgdat and section B contains the usemap, both
287 * sections become inter-dependent. This allocates usemaps
288 * from the same section as the pgdat where possible to avoid
289 * this problem.
290 */
291 goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
292 limit = goal + (1UL << PA_SECTION_SHIFT);
293 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
294again:
295 p = ___alloc_bootmem_node_nopanic(NODE_DATA(nid), size,
296 SMP_CACHE_BYTES, goal, limit);
297 if (!p && limit) {
298 limit = 0;
299 goto again;
300 }
301 return p;
302}
303
304static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
305{
306 unsigned long usemap_snr, pgdat_snr;
307 static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
308 static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
309 struct pglist_data *pgdat = NODE_DATA(nid);
310 int usemap_nid;
311
312 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
313 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
314 if (usemap_snr == pgdat_snr)
315 return;
316
317 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
318 /* skip redundant message */
319 return;
320
321 old_usemap_snr = usemap_snr;
322 old_pgdat_snr = pgdat_snr;
323
324 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
325 if (usemap_nid != nid) {
326 printk(KERN_INFO
327 "node %d must be removed before remove section %ld\n",
328 nid, usemap_snr);
329 return;
330 }
331 /*
332 * There is a circular dependency.
333 * Some platforms allow un-removable section because they will just
334 * gather other removable sections for dynamic partitioning.
335 * Just notify un-removable section's number here.
336 */
337 printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
338 pgdat_snr, nid);
339 printk(KERN_CONT
340 " have a circular dependency on usemap and pgdat allocations\n");
341}
342#else
343static unsigned long * __init
344sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
345 unsigned long size)
346{
347 return alloc_bootmem_node_nopanic(pgdat, size);
348}
349
350static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
351{
352}
353#endif /* CONFIG_MEMORY_HOTREMOVE */
354
355static void __init sparse_early_usemaps_alloc_node(unsigned long**usemap_map,
356 unsigned long pnum_begin,
357 unsigned long pnum_end,
358 unsigned long usemap_count, int nodeid)
359{
360 void *usemap;
361 unsigned long pnum;
362 int size = usemap_size();
363
364 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
365 size * usemap_count);
366 if (!usemap) {
367 printk(KERN_WARNING "%s: allocation failed\n", __func__);
368 return;
369 }
370
371 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
372 if (!present_section_nr(pnum))
373 continue;
374 usemap_map[pnum] = usemap;
375 usemap += size;
376 check_usemap_section_nr(nodeid, usemap_map[pnum]);
377 }
378}
379
380#ifndef CONFIG_SPARSEMEM_VMEMMAP
381struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
382{
383 struct page *map;
384 unsigned long size;
385
386 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
387 if (map)
388 return map;
389
390 size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
391 map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
392 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
393 return map;
394}
395void __init sparse_mem_maps_populate_node(struct page **map_map,
396 unsigned long pnum_begin,
397 unsigned long pnum_end,
398 unsigned long map_count, int nodeid)
399{
400 void *map;
401 unsigned long pnum;
402 unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
403
404 map = alloc_remap(nodeid, size * map_count);
405 if (map) {
406 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
407 if (!present_section_nr(pnum))
408 continue;
409 map_map[pnum] = map;
410 map += size;
411 }
412 return;
413 }
414
415 size = PAGE_ALIGN(size);
416 map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
417 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
418 if (map) {
419 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
420 if (!present_section_nr(pnum))
421 continue;
422 map_map[pnum] = map;
423 map += size;
424 }
425 return;
426 }
427
428 /* fallback */
429 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
430 struct mem_section *ms;
431
432 if (!present_section_nr(pnum))
433 continue;
434 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
435 if (map_map[pnum])
436 continue;
437 ms = __nr_to_section(pnum);
438 printk(KERN_ERR "%s: sparsemem memory map backing failed "
439 "some memory will not be available.\n", __func__);
440 ms->section_mem_map = 0;
441 }
442}
443#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
444
445#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
446static void __init sparse_early_mem_maps_alloc_node(struct page **map_map,
447 unsigned long pnum_begin,
448 unsigned long pnum_end,
449 unsigned long map_count, int nodeid)
450{
451 sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
452 map_count, nodeid);
453}
454#else
455static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
456{
457 struct page *map;
458 struct mem_section *ms = __nr_to_section(pnum);
459 int nid = sparse_early_nid(ms);
460
461 map = sparse_mem_map_populate(pnum, nid);
462 if (map)
463 return map;
464
465 printk(KERN_ERR "%s: sparsemem memory map backing failed "
466 "some memory will not be available.\n", __func__);
467 ms->section_mem_map = 0;
468 return NULL;
469}
470#endif
471
472void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
473{
474}
475
476/*
477 * Allocate the accumulated non-linear sections, allocate a mem_map
478 * for each and record the physical to section mapping.
479 */
480void __init sparse_init(void)
481{
482 unsigned long pnum;
483 struct page *map;
484 unsigned long *usemap;
485 unsigned long **usemap_map;
486 int size;
487 int nodeid_begin = 0;
488 unsigned long pnum_begin = 0;
489 unsigned long usemap_count;
490#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
491 unsigned long map_count;
492 int size2;
493 struct page **map_map;
494#endif
495
496 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
497 set_pageblock_order();
498
499 /*
500 * map is using big page (aka 2M in x86 64 bit)
501 * usemap is less one page (aka 24 bytes)
502 * so alloc 2M (with 2M align) and 24 bytes in turn will
503 * make next 2M slip to one more 2M later.
504 * then in big system, the memory will have a lot of holes...
505 * here try to allocate 2M pages continuously.
506 *
507 * powerpc need to call sparse_init_one_section right after each
508 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
509 */
510 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
511 usemap_map = alloc_bootmem(size);
512 if (!usemap_map)
513 panic("can not allocate usemap_map\n");
514
515 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
516 struct mem_section *ms;
517
518 if (!present_section_nr(pnum))
519 continue;
520 ms = __nr_to_section(pnum);
521 nodeid_begin = sparse_early_nid(ms);
522 pnum_begin = pnum;
523 break;
524 }
525 usemap_count = 1;
526 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
527 struct mem_section *ms;
528 int nodeid;
529
530 if (!present_section_nr(pnum))
531 continue;
532 ms = __nr_to_section(pnum);
533 nodeid = sparse_early_nid(ms);
534 if (nodeid == nodeid_begin) {
535 usemap_count++;
536 continue;
537 }
538 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
539 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, pnum,
540 usemap_count, nodeid_begin);
541 /* new start, update count etc*/
542 nodeid_begin = nodeid;
543 pnum_begin = pnum;
544 usemap_count = 1;
545 }
546 /* ok, last chunk */
547 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS,
548 usemap_count, nodeid_begin);
549
550#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
551 size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
552 map_map = alloc_bootmem(size2);
553 if (!map_map)
554 panic("can not allocate map_map\n");
555
556 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
557 struct mem_section *ms;
558
559 if (!present_section_nr(pnum))
560 continue;
561 ms = __nr_to_section(pnum);
562 nodeid_begin = sparse_early_nid(ms);
563 pnum_begin = pnum;
564 break;
565 }
566 map_count = 1;
567 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
568 struct mem_section *ms;
569 int nodeid;
570
571 if (!present_section_nr(pnum))
572 continue;
573 ms = __nr_to_section(pnum);
574 nodeid = sparse_early_nid(ms);
575 if (nodeid == nodeid_begin) {
576 map_count++;
577 continue;
578 }
579 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
580 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, pnum,
581 map_count, nodeid_begin);
582 /* new start, update count etc*/
583 nodeid_begin = nodeid;
584 pnum_begin = pnum;
585 map_count = 1;
586 }
587 /* ok, last chunk */
588 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS,
589 map_count, nodeid_begin);
590#endif
591
592 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
593 if (!present_section_nr(pnum))
594 continue;
595
596 usemap = usemap_map[pnum];
597 if (!usemap)
598 continue;
599
600#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
601 map = map_map[pnum];
602#else
603 map = sparse_early_mem_map_alloc(pnum);
604#endif
605 if (!map)
606 continue;
607
608 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
609 usemap);
610 }
611
612 vmemmap_populate_print_last();
613
614#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
615 free_bootmem(__pa(map_map), size2);
616#endif
617 free_bootmem(__pa(usemap_map), size);
618}
619
620#ifdef CONFIG_MEMORY_HOTPLUG
621#ifdef CONFIG_SPARSEMEM_VMEMMAP
622static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
623 unsigned long nr_pages)
624{
625 /* This will make the necessary allocations eventually. */
626 return sparse_mem_map_populate(pnum, nid);
627}
628static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
629{
630 return; /* XXX: Not implemented yet */
631}
632static void free_map_bootmem(struct page *page, unsigned long nr_pages)
633{
634}
635#else
636static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
637{
638 struct page *page, *ret;
639 unsigned long memmap_size = sizeof(struct page) * nr_pages;
640
641 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
642 if (page)
643 goto got_map_page;
644
645 ret = vmalloc(memmap_size);
646 if (ret)
647 goto got_map_ptr;
648
649 return NULL;
650got_map_page:
651 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
652got_map_ptr:
653 memset(ret, 0, memmap_size);
654
655 return ret;
656}
657
658static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
659 unsigned long nr_pages)
660{
661 return __kmalloc_section_memmap(nr_pages);
662}
663
664static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
665{
666 if (is_vmalloc_addr(memmap))
667 vfree(memmap);
668 else
669 free_pages((unsigned long)memmap,
670 get_order(sizeof(struct page) * nr_pages));
671}
672
673static void free_map_bootmem(struct page *page, unsigned long nr_pages)
674{
675 unsigned long maps_section_nr, removing_section_nr, i;
676 unsigned long magic;
677
678 for (i = 0; i < nr_pages; i++, page++) {
679 magic = (unsigned long) page->lru.next;
680
681 BUG_ON(magic == NODE_INFO);
682
683 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
684 removing_section_nr = page->private;
685
686 /*
687 * When this function is called, the removing section is
688 * logical offlined state. This means all pages are isolated
689 * from page allocator. If removing section's memmap is placed
690 * on the same section, it must not be freed.
691 * If it is freed, page allocator may allocate it which will
692 * be removed physically soon.
693 */
694 if (maps_section_nr != removing_section_nr)
695 put_page_bootmem(page);
696 }
697}
698#endif /* CONFIG_SPARSEMEM_VMEMMAP */
699
700static void free_section_usemap(struct page *memmap, unsigned long *usemap)
701{
702 struct page *usemap_page;
703 unsigned long nr_pages;
704
705 if (!usemap)
706 return;
707
708 usemap_page = virt_to_page(usemap);
709 /*
710 * Check to see if allocation came from hot-plug-add
711 */
712 if (PageSlab(usemap_page)) {
713 kfree(usemap);
714 if (memmap)
715 __kfree_section_memmap(memmap, PAGES_PER_SECTION);
716 return;
717 }
718
719 /*
720 * The usemap came from bootmem. This is packed with other usemaps
721 * on the section which has pgdat at boot time. Just keep it as is now.
722 */
723
724 if (memmap) {
725 struct page *memmap_page;
726 memmap_page = virt_to_page(memmap);
727
728 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
729 >> PAGE_SHIFT;
730
731 free_map_bootmem(memmap_page, nr_pages);
732 }
733}
734
735/*
736 * returns the number of sections whose mem_maps were properly
737 * set. If this is <=0, then that means that the passed-in
738 * map was not consumed and must be freed.
739 */
740int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
741 int nr_pages)
742{
743 unsigned long section_nr = pfn_to_section_nr(start_pfn);
744 struct pglist_data *pgdat = zone->zone_pgdat;
745 struct mem_section *ms;
746 struct page *memmap;
747 unsigned long *usemap;
748 unsigned long flags;
749 int ret;
750
751 /*
752 * no locking for this, because it does its own
753 * plus, it does a kmalloc
754 */
755 ret = sparse_index_init(section_nr, pgdat->node_id);
756 if (ret < 0 && ret != -EEXIST)
757 return ret;
758 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
759 if (!memmap)
760 return -ENOMEM;
761 usemap = __kmalloc_section_usemap();
762 if (!usemap) {
763 __kfree_section_memmap(memmap, nr_pages);
764 return -ENOMEM;
765 }
766
767 pgdat_resize_lock(pgdat, &flags);
768
769 ms = __pfn_to_section(start_pfn);
770 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
771 ret = -EEXIST;
772 goto out;
773 }
774
775 ms->section_mem_map |= SECTION_MARKED_PRESENT;
776
777 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
778
779out:
780 pgdat_resize_unlock(pgdat, &flags);
781 if (ret <= 0) {
782 kfree(usemap);
783 __kfree_section_memmap(memmap, nr_pages);
784 }
785 return ret;
786}
787
788void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
789{
790 struct page *memmap = NULL;
791 unsigned long *usemap = NULL;
792
793 if (ms->section_mem_map) {
794 usemap = ms->pageblock_flags;
795 memmap = sparse_decode_mem_map(ms->section_mem_map,
796 __section_nr(ms));
797 ms->section_mem_map = 0;
798 ms->pageblock_flags = NULL;
799 }
800
801 free_section_usemap(memmap, usemap);
802}
803#endif