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 = kzalloc_node(array_size, GFP_KERNEL, nid);
70 section = kzalloc(array_size, GFP_KERNEL);
72 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
78 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
80 static DEFINE_SPINLOCK(index_init_lock);
81 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
82 struct mem_section *section;
85 if (mem_section[root])
88 section = sparse_index_alloc(nid);
92 * This lock keeps two different sections from
93 * reallocating for the same index
95 spin_lock(&index_init_lock);
97 if (mem_section[root]) {
102 mem_section[root] = section;
104 spin_unlock(&index_init_lock);
107 #else /* !SPARSEMEM_EXTREME */
108 static inline int sparse_index_init(unsigned long section_nr, int nid)
115 * Although written for the SPARSEMEM_EXTREME case, this happens
116 * to also work for the flat array case because
117 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
119 int __section_nr(struct mem_section* ms)
121 unsigned long root_nr;
122 struct mem_section* root;
124 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
125 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
129 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
133 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
137 * During early boot, before section_mem_map is used for an actual
138 * mem_map, we use section_mem_map to store the section's NUMA
139 * node. This keeps us from having to use another data structure. The
140 * node information is cleared just before we store the real mem_map.
142 static inline unsigned long sparse_encode_early_nid(int nid)
144 return (nid << SECTION_NID_SHIFT);
147 static inline int sparse_early_nid(struct mem_section *section)
149 return (section->section_mem_map >> SECTION_NID_SHIFT);
152 /* Validate the physical addressing limitations of the model */
153 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
154 unsigned long *end_pfn)
156 unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
159 * Sanity checks - do not allow an architecture to pass
160 * in larger pfns than the maximum scope of sparsemem:
162 if (*start_pfn > max_sparsemem_pfn) {
163 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
164 "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
165 *start_pfn, *end_pfn, max_sparsemem_pfn);
167 *start_pfn = max_sparsemem_pfn;
168 *end_pfn = max_sparsemem_pfn;
169 } else if (*end_pfn > max_sparsemem_pfn) {
170 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
171 "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
172 *start_pfn, *end_pfn, max_sparsemem_pfn);
174 *end_pfn = max_sparsemem_pfn;
178 /* Record a memory area against a node. */
179 void __init memory_present(int nid, unsigned long start, unsigned long end)
183 start &= PAGE_SECTION_MASK;
184 mminit_validate_memmodel_limits(&start, &end);
185 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
186 unsigned long section = pfn_to_section_nr(pfn);
187 struct mem_section *ms;
189 sparse_index_init(section, nid);
190 set_section_nid(section, nid);
192 ms = __nr_to_section(section);
193 if (!ms->section_mem_map)
194 ms->section_mem_map = sparse_encode_early_nid(nid) |
195 SECTION_MARKED_PRESENT;
200 * Only used by the i386 NUMA architecures, but relatively
203 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
204 unsigned long end_pfn)
207 unsigned long nr_pages = 0;
209 mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
210 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
211 if (nid != early_pfn_to_nid(pfn))
214 if (pfn_present(pfn))
215 nr_pages += PAGES_PER_SECTION;
218 return nr_pages * sizeof(struct page);
222 * Subtle, we encode the real pfn into the mem_map such that
223 * the identity pfn - section_mem_map will return the actual
224 * physical page frame number.
226 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
228 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
232 * Decode mem_map from the coded memmap
234 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
236 /* mask off the extra low bits of information */
237 coded_mem_map &= SECTION_MAP_MASK;
238 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
241 static int __meminit sparse_init_one_section(struct mem_section *ms,
242 unsigned long pnum, struct page *mem_map,
243 unsigned long *pageblock_bitmap)
245 if (!present_section(ms))
248 ms->section_mem_map &= ~SECTION_MAP_MASK;
249 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
251 ms->pageblock_flags = pageblock_bitmap;
256 unsigned long usemap_size(void)
258 unsigned long size_bytes;
259 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
260 size_bytes = roundup(size_bytes, sizeof(unsigned long));
264 #ifdef CONFIG_MEMORY_HOTPLUG
265 static unsigned long *__kmalloc_section_usemap(void)
267 return kmalloc(usemap_size(), GFP_KERNEL);
269 #endif /* CONFIG_MEMORY_HOTPLUG */
271 #ifdef CONFIG_MEMORY_HOTREMOVE
272 static unsigned long * __init
273 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
276 unsigned long goal, limit;
280 * A page may contain usemaps for other sections preventing the
281 * page being freed and making a section unremovable while
282 * other sections referencing the usemap retmain active. Similarly,
283 * a pgdat can prevent a section being removed. If section A
284 * contains a pgdat and section B contains the usemap, both
285 * sections become inter-dependent. This allocates usemaps
286 * from the same section as the pgdat where possible to avoid
289 goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
290 limit = goal + (1UL << PA_SECTION_SHIFT);
291 nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
293 p = ___alloc_bootmem_node_nopanic(NODE_DATA(nid), size,
294 SMP_CACHE_BYTES, goal, limit);
302 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
304 unsigned long usemap_snr, pgdat_snr;
305 static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
306 static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
307 struct pglist_data *pgdat = NODE_DATA(nid);
310 usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
311 pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
312 if (usemap_snr == pgdat_snr)
315 if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
316 /* skip redundant message */
319 old_usemap_snr = usemap_snr;
320 old_pgdat_snr = pgdat_snr;
322 usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
323 if (usemap_nid != nid) {
325 "node %d must be removed before remove section %ld\n",
330 * There is a circular dependency.
331 * Some platforms allow un-removable section because they will just
332 * gather other removable sections for dynamic partitioning.
333 * Just notify un-removable section's number here.
335 printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
338 " have a circular dependency on usemap and pgdat allocations\n");
341 static unsigned long * __init
342 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
345 return alloc_bootmem_node_nopanic(pgdat, size);
348 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
351 #endif /* CONFIG_MEMORY_HOTREMOVE */
353 static void __init sparse_early_usemaps_alloc_node(unsigned long**usemap_map,
354 unsigned long pnum_begin,
355 unsigned long pnum_end,
356 unsigned long usemap_count, int nodeid)
360 int size = usemap_size();
362 usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
363 size * usemap_count);
365 printk(KERN_WARNING "%s: allocation failed\n", __func__);
369 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
370 if (!present_section_nr(pnum))
372 usemap_map[pnum] = usemap;
374 check_usemap_section_nr(nodeid, usemap_map[pnum]);
378 #ifndef CONFIG_SPARSEMEM_VMEMMAP
379 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
384 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
388 size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
389 map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
390 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
393 void __init sparse_mem_maps_populate_node(struct page **map_map,
394 unsigned long pnum_begin,
395 unsigned long pnum_end,
396 unsigned long map_count, int nodeid)
400 unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
402 map = alloc_remap(nodeid, size * map_count);
404 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
405 if (!present_section_nr(pnum))
413 size = PAGE_ALIGN(size);
414 map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
415 PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
417 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
418 if (!present_section_nr(pnum))
427 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
428 struct mem_section *ms;
430 if (!present_section_nr(pnum))
432 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
435 ms = __nr_to_section(pnum);
436 printk(KERN_ERR "%s: sparsemem memory map backing failed "
437 "some memory will not be available.\n", __func__);
438 ms->section_mem_map = 0;
441 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
443 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
444 static void __init sparse_early_mem_maps_alloc_node(struct page **map_map,
445 unsigned long pnum_begin,
446 unsigned long pnum_end,
447 unsigned long map_count, int nodeid)
449 sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
453 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
456 struct mem_section *ms = __nr_to_section(pnum);
457 int nid = sparse_early_nid(ms);
459 map = sparse_mem_map_populate(pnum, nid);
463 printk(KERN_ERR "%s: sparsemem memory map backing failed "
464 "some memory will not be available.\n", __func__);
465 ms->section_mem_map = 0;
470 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
475 * Allocate the accumulated non-linear sections, allocate a mem_map
476 * for each and record the physical to section mapping.
478 void __init sparse_init(void)
482 unsigned long *usemap;
483 unsigned long **usemap_map;
485 int nodeid_begin = 0;
486 unsigned long pnum_begin = 0;
487 unsigned long usemap_count;
488 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
489 unsigned long map_count;
491 struct page **map_map;
494 /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
495 set_pageblock_order();
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.
505 * powerpc need to call sparse_init_one_section right after each
506 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
508 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
509 usemap_map = alloc_bootmem(size);
511 panic("can not allocate usemap_map\n");
513 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
514 struct mem_section *ms;
516 if (!present_section_nr(pnum))
518 ms = __nr_to_section(pnum);
519 nodeid_begin = sparse_early_nid(ms);
524 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
525 struct mem_section *ms;
528 if (!present_section_nr(pnum))
530 ms = __nr_to_section(pnum);
531 nodeid = sparse_early_nid(ms);
532 if (nodeid == nodeid_begin) {
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;
545 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS,
546 usemap_count, nodeid_begin);
548 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
549 size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
550 map_map = alloc_bootmem(size2);
552 panic("can not allocate map_map\n");
554 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
555 struct mem_section *ms;
557 if (!present_section_nr(pnum))
559 ms = __nr_to_section(pnum);
560 nodeid_begin = sparse_early_nid(ms);
565 for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
566 struct mem_section *ms;
569 if (!present_section_nr(pnum))
571 ms = __nr_to_section(pnum);
572 nodeid = sparse_early_nid(ms);
573 if (nodeid == nodeid_begin) {
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;
586 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS,
587 map_count, nodeid_begin);
590 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
591 if (!present_section_nr(pnum))
594 usemap = usemap_map[pnum];
598 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
601 map = sparse_early_mem_map_alloc(pnum);
606 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
610 vmemmap_populate_print_last();
612 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
613 free_bootmem(__pa(map_map), size2);
615 free_bootmem(__pa(usemap_map), size);
618 #ifdef CONFIG_MEMORY_HOTPLUG
619 #ifdef CONFIG_SPARSEMEM_VMEMMAP
620 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
621 unsigned long nr_pages)
623 /* This will make the necessary allocations eventually. */
624 return sparse_mem_map_populate(pnum, nid);
626 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
628 return; /* XXX: Not implemented yet */
630 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
634 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
636 struct page *page, *ret;
637 unsigned long memmap_size = sizeof(struct page) * nr_pages;
639 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
643 ret = vmalloc(memmap_size);
649 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
651 memset(ret, 0, memmap_size);
656 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
657 unsigned long nr_pages)
659 return __kmalloc_section_memmap(nr_pages);
662 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
664 if (is_vmalloc_addr(memmap))
667 free_pages((unsigned long)memmap,
668 get_order(sizeof(struct page) * nr_pages));
671 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
673 unsigned long maps_section_nr, removing_section_nr, i;
676 for (i = 0; i < nr_pages; i++, page++) {
677 magic = (unsigned long) page->lru.next;
679 BUG_ON(magic == NODE_INFO);
681 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
682 removing_section_nr = page->private;
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.
692 if (maps_section_nr != removing_section_nr)
693 put_page_bootmem(page);
696 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
698 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
700 struct page *usemap_page;
701 unsigned long nr_pages;
706 usemap_page = virt_to_page(usemap);
708 * Check to see if allocation came from hot-plug-add
710 if (PageSlab(usemap_page)) {
713 __kfree_section_memmap(memmap, PAGES_PER_SECTION);
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.
723 struct page *memmap_page;
724 memmap_page = virt_to_page(memmap);
726 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
729 free_map_bootmem(memmap_page, nr_pages);
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.
738 int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
741 unsigned long section_nr = pfn_to_section_nr(start_pfn);
742 struct pglist_data *pgdat = zone->zone_pgdat;
743 struct mem_section *ms;
745 unsigned long *usemap;
750 * no locking for this, because it does its own
751 * plus, it does a kmalloc
753 ret = sparse_index_init(section_nr, pgdat->node_id);
754 if (ret < 0 && ret != -EEXIST)
756 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
759 usemap = __kmalloc_section_usemap();
761 __kfree_section_memmap(memmap, nr_pages);
765 pgdat_resize_lock(pgdat, &flags);
767 ms = __pfn_to_section(start_pfn);
768 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
773 ms->section_mem_map |= SECTION_MARKED_PRESENT;
775 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
778 pgdat_resize_unlock(pgdat, &flags);
781 __kfree_section_memmap(memmap, nr_pages);
786 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
788 struct page *memmap = NULL;
789 unsigned long *usemap = NULL;
791 if (ms->section_mem_map) {
792 usemap = ms->pageblock_flags;
793 memmap = sparse_decode_mem_map(ms->section_mem_map,
795 ms->section_mem_map = 0;
796 ms->pageblock_flags = NULL;
799 free_section_usemap(memmap, usemap);