2 * sparse memory mappings.
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>
14 #include <asm/pgalloc.h>
15 #include <asm/pgtable.h>
18 * Permanent SPARSEMEM data:
20 * 1) mem_section - memory sections, mem_map's for valid memory
22 #ifdef CONFIG_SPARSEMEM_EXTREME
23 struct mem_section *mem_section[NR_SECTION_ROOTS]
24 ____cacheline_internodealigned_in_smp;
26 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
27 ____cacheline_internodealigned_in_smp;
29 EXPORT_SYMBOL(mem_section);
31 #ifdef NODE_NOT_IN_PAGE_FLAGS
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.
37 #if MAX_NUMNODES <= 256
38 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
40 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
43 int page_to_nid(const struct page *page)
45 return section_to_node_table[page_to_section(page)];
47 EXPORT_SYMBOL(page_to_nid);
49 static void set_section_nid(unsigned long section_nr, int nid)
51 section_to_node_table[section_nr] = nid;
53 #else /* !NODE_NOT_IN_PAGE_FLAGS */
54 static inline void set_section_nid(unsigned long section_nr, int nid)
59 #ifdef CONFIG_SPARSEMEM_EXTREME
60 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
62 struct mem_section *section = NULL;
63 unsigned long array_size = SECTIONS_PER_ROOT *
64 sizeof(struct mem_section);
66 if (slab_is_available()) {
67 if (node_state(nid, N_HIGH_MEMORY))
68 section = kmalloc_node(array_size, GFP_KERNEL, nid);
70 section = kmalloc(array_size, GFP_KERNEL);
72 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
75 memset(section, 0, array_size);
80 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
82 static DEFINE_SPINLOCK(index_init_lock);
83 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
84 struct mem_section *section;
87 if (mem_section[root])
90 section = sparse_index_alloc(nid);
94 * This lock keeps two different sections from
95 * reallocating for the same index
97 spin_lock(&index_init_lock);
99 if (mem_section[root]) {
104 mem_section[root] = section;
106 spin_unlock(&index_init_lock);
109 #else /* !SPARSEMEM_EXTREME */
110 static inline int sparse_index_init(unsigned long section_nr, int nid)
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.
121 int __section_nr(struct mem_section* ms)
123 unsigned long root_nr;
124 struct mem_section* root;
126 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
127 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
131 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
135 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
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.
144 static inline unsigned long sparse_encode_early_nid(int nid)
146 return (nid << SECTION_NID_SHIFT);
149 static inline int sparse_early_nid(struct mem_section *section)
151 return (section->section_mem_map >> SECTION_NID_SHIFT);
154 /* Validate the physical addressing limitations of the model */
155 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
156 unsigned long *end_pfn)
158 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
161 * Sanity checks - do not allow an architecture to pass
162 * in larger pfns than the maximum scope of sparsemem:
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);
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);
176 *end_pfn = max_sparsemem_pfn;
180 /* Record a memory area against a node. */
181 void __init memory_present(int nid, unsigned long start, unsigned long end)
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;
191 sparse_index_init(section, nid);
192 set_section_nid(section, nid);
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;
202 * Only used by the i386 NUMA architecures, but relatively
205 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
206 unsigned long end_pfn)
209 unsigned long nr_pages = 0;
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))
216 if (pfn_present(pfn))
217 nr_pages += PAGES_PER_SECTION;
220 return nr_pages * sizeof(struct page);
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.
228 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
230 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
234 * Decode mem_map from the coded memmap
236 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
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);
243 static int __meminit sparse_init_one_section(struct mem_section *ms,
244 unsigned long pnum, struct page *mem_map,
245 unsigned long *pageblock_bitmap)
247 if (!present_section(ms))
250 ms->section_mem_map &= ~SECTION_MAP_MASK;
251 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
253 ms->pageblock_flags = pageblock_bitmap;
258 unsigned long usemap_size(void)
260 unsigned long size_bytes;
261 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
262 size_bytes = roundup(size_bytes, sizeof(unsigned long));
266 #ifdef CONFIG_MEMORY_HOTPLUG
267 static unsigned long *__kmalloc_section_usemap(void)
269 return kmalloc(usemap_size(), GFP_KERNEL);
271 #endif /* CONFIG_MEMORY_HOTPLUG */
273 #ifdef CONFIG_MEMORY_HOTREMOVE
274 static unsigned long * __init
275 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
278 unsigned long section_nr;
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
290 section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
291 return alloc_bootmem_section(usemap_size() * count, section_nr);
294 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
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);
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)
307 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
308 /* skip redundant message */
311 old_usemap_snr = usemap_snr;
312 old_pgdat_snr = pgdat_snr;
314 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
315 if (usemap_nid != nid) {
317 "node %d must be removed before remove section %ld\n",
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.
327 printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
330 " have a circular dependency on usemap and pgdat allocations\n");
333 static unsigned long * __init
334 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
340 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
343 #endif /* CONFIG_MEMORY_HOTREMOVE */
345 static 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)
352 int size = usemap_size();
354 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
357 usemap = alloc_bootmem_node(NODE_DATA(nodeid), size * usemap_count);
359 printk(KERN_WARNING "%s: allocation failed\n", __func__);
364 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
365 if (!present_section_nr(pnum))
367 usemap_map[pnum] = usemap;
369 check_usemap_section_nr(nodeid, usemap_map[pnum]);
373 #ifndef CONFIG_SPARSEMEM_VMEMMAP
374 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
379 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
383 size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
384 map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
385 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
388 void __init sparse_mem_maps_populate_node(struct page **map_map,
389 unsigned long pnum_begin,
390 unsigned long pnum_end,
391 unsigned long map_count, int nodeid)
395 unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
397 map = alloc_remap(nodeid, size * map_count);
399 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
400 if (!present_section_nr(pnum))
408 size = PAGE_ALIGN(size);
409 map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
410 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
412 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
413 if (!present_section_nr(pnum))
422 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
423 struct mem_section *ms;
425 if (!present_section_nr(pnum))
427 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
430 ms = __nr_to_section(pnum);
431 printk(KERN_ERR "%s: sparsemem memory map backing failed "
432 "some memory will not be available.\n", __func__);
433 ms->section_mem_map = 0;
436 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
438 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
439 static void __init sparse_early_mem_maps_alloc_node(struct page **map_map,
440 unsigned long pnum_begin,
441 unsigned long pnum_end,
442 unsigned long map_count, int nodeid)
444 sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
448 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
451 struct mem_section *ms = __nr_to_section(pnum);
452 int nid = sparse_early_nid(ms);
454 map = sparse_mem_map_populate(pnum, nid);
458 printk(KERN_ERR "%s: sparsemem memory map backing failed "
459 "some memory will not be available.\n", __func__);
460 ms->section_mem_map = 0;
465 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
470 * Allocate the accumulated non-linear sections, allocate a mem_map
471 * for each and record the physical to section mapping.
473 void __init sparse_init(void)
477 unsigned long *usemap;
478 unsigned long **usemap_map;
480 int nodeid_begin = 0;
481 unsigned long pnum_begin = 0;
482 unsigned long usemap_count;
483 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
484 unsigned long map_count;
486 struct page **map_map;
490 * map is using big page (aka 2M in x86 64 bit)
491 * usemap is less one page (aka 24 bytes)
492 * so alloc 2M (with 2M align) and 24 bytes in turn will
493 * make next 2M slip to one more 2M later.
494 * then in big system, the memory will have a lot of holes...
495 * here try to allocate 2M pages continuously.
497 * powerpc need to call sparse_init_one_section right after each
498 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
500 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
501 usemap_map = alloc_bootmem(size);
503 panic("can not allocate usemap_map\n");
505 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
506 struct mem_section *ms;
508 if (!present_section_nr(pnum))
510 ms = __nr_to_section(pnum);
511 nodeid_begin = sparse_early_nid(ms);
516 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
517 struct mem_section *ms;
520 if (!present_section_nr(pnum))
522 ms = __nr_to_section(pnum);
523 nodeid = sparse_early_nid(ms);
524 if (nodeid == nodeid_begin) {
528 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
529 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, pnum,
530 usemap_count, nodeid_begin);
531 /* new start, update count etc*/
532 nodeid_begin = nodeid;
537 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS,
538 usemap_count, nodeid_begin);
540 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
541 size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
542 map_map = alloc_bootmem(size2);
544 panic("can not allocate map_map\n");
546 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
547 struct mem_section *ms;
549 if (!present_section_nr(pnum))
551 ms = __nr_to_section(pnum);
552 nodeid_begin = sparse_early_nid(ms);
557 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
558 struct mem_section *ms;
561 if (!present_section_nr(pnum))
563 ms = __nr_to_section(pnum);
564 nodeid = sparse_early_nid(ms);
565 if (nodeid == nodeid_begin) {
569 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
570 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, pnum,
571 map_count, nodeid_begin);
572 /* new start, update count etc*/
573 nodeid_begin = nodeid;
578 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS,
579 map_count, nodeid_begin);
582 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
583 if (!present_section_nr(pnum))
586 usemap = usemap_map[pnum];
590 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
593 map = sparse_early_mem_map_alloc(pnum);
598 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
602 vmemmap_populate_print_last();
604 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
605 free_bootmem(__pa(map_map), size2);
607 free_bootmem(__pa(usemap_map), size);
610 #ifdef CONFIG_MEMORY_HOTPLUG
611 #ifdef CONFIG_SPARSEMEM_VMEMMAP
612 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
613 unsigned long nr_pages)
615 /* This will make the necessary allocations eventually. */
616 return sparse_mem_map_populate(pnum, nid);
618 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
620 return; /* XXX: Not implemented yet */
622 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
626 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
628 struct page *page, *ret;
629 unsigned long memmap_size = sizeof(struct page) * nr_pages;
631 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
635 ret = vmalloc(memmap_size);
641 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
643 memset(ret, 0, memmap_size);
648 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
649 unsigned long nr_pages)
651 return __kmalloc_section_memmap(nr_pages);
654 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
656 if (is_vmalloc_addr(memmap))
659 free_pages((unsigned long)memmap,
660 get_order(sizeof(struct page) * nr_pages));
663 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
665 unsigned long maps_section_nr, removing_section_nr, i;
668 for (i = 0; i < nr_pages; i++, page++) {
669 magic = (unsigned long) page->lru.next;
671 BUG_ON(magic == NODE_INFO);
673 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
674 removing_section_nr = page->private;
677 * When this function is called, the removing section is
678 * logical offlined state. This means all pages are isolated
679 * from page allocator. If removing section's memmap is placed
680 * on the same section, it must not be freed.
681 * If it is freed, page allocator may allocate it which will
682 * be removed physically soon.
684 if (maps_section_nr != removing_section_nr)
685 put_page_bootmem(page);
688 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
690 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
692 struct page *usemap_page;
693 unsigned long nr_pages;
698 usemap_page = virt_to_page(usemap);
700 * Check to see if allocation came from hot-plug-add
702 if (PageSlab(usemap_page)) {
705 __kfree_section_memmap(memmap, PAGES_PER_SECTION);
710 * The usemap came from bootmem. This is packed with other usemaps
711 * on the section which has pgdat at boot time. Just keep it as is now.
715 struct page *memmap_page;
716 memmap_page = virt_to_page(memmap);
718 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
721 free_map_bootmem(memmap_page, nr_pages);
726 * returns the number of sections whose mem_maps were properly
727 * set. If this is <=0, then that means that the passed-in
728 * map was not consumed and must be freed.
730 int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
733 unsigned long section_nr = pfn_to_section_nr(start_pfn);
734 struct pglist_data *pgdat = zone->zone_pgdat;
735 struct mem_section *ms;
737 unsigned long *usemap;
742 * no locking for this, because it does its own
743 * plus, it does a kmalloc
745 ret = sparse_index_init(section_nr, pgdat->node_id);
746 if (ret < 0 && ret != -EEXIST)
748 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
751 usemap = __kmalloc_section_usemap();
753 __kfree_section_memmap(memmap, nr_pages);
757 pgdat_resize_lock(pgdat, &flags);
759 ms = __pfn_to_section(start_pfn);
760 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
765 ms->section_mem_map |= SECTION_MARKED_PRESENT;
767 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
770 pgdat_resize_unlock(pgdat, &flags);
773 __kfree_section_memmap(memmap, nr_pages);
778 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
780 struct page *memmap = NULL;
781 unsigned long *usemap = NULL;
783 if (ms->section_mem_map) {
784 usemap = ms->pageblock_flags;
785 memmap = sparse_decode_mem_map(ms->section_mem_map,
787 ms->section_mem_map = 0;
788 ms->pageblock_flags = NULL;
791 free_section_usemap(memmap, usemap);