2 * sparse memory mappings.
5 #include <linux/mmzone.h>
6 #include <linux/bootmem.h>
7 #include <linux/highmem.h>
8 #include <linux/module.h>
9 #include <linux/spinlock.h>
10 #include <linux/vmalloc.h>
13 #include <asm/pgalloc.h>
14 #include <asm/pgtable.h>
17 * Permanent SPARSEMEM data:
19 * 1) mem_section - memory sections, mem_map's for valid memory
21 #ifdef CONFIG_SPARSEMEM_EXTREME
22 struct mem_section *mem_section[NR_SECTION_ROOTS]
23 ____cacheline_internodealigned_in_smp;
25 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
26 ____cacheline_internodealigned_in_smp;
28 EXPORT_SYMBOL(mem_section);
30 #ifdef NODE_NOT_IN_PAGE_FLAGS
32 * If we did not store the node number in the page then we have to
33 * do a lookup in the section_to_node_table in order to find which
34 * node the page belongs to.
36 #if MAX_NUMNODES <= 256
37 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
39 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
42 int page_to_nid(struct page *page)
44 return section_to_node_table[page_to_section(page)];
46 EXPORT_SYMBOL(page_to_nid);
48 static void set_section_nid(unsigned long section_nr, int nid)
50 section_to_node_table[section_nr] = nid;
52 #else /* !NODE_NOT_IN_PAGE_FLAGS */
53 static inline void set_section_nid(unsigned long section_nr, int nid)
58 #ifdef CONFIG_SPARSEMEM_EXTREME
59 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
61 struct mem_section *section = NULL;
62 unsigned long array_size = SECTIONS_PER_ROOT *
63 sizeof(struct mem_section);
65 if (slab_is_available())
66 section = kmalloc_node(array_size, GFP_KERNEL, nid);
68 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
71 memset(section, 0, array_size);
76 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
78 static DEFINE_SPINLOCK(index_init_lock);
79 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
80 struct mem_section *section;
83 if (mem_section[root])
86 section = sparse_index_alloc(nid);
90 * This lock keeps two different sections from
91 * reallocating for the same index
93 spin_lock(&index_init_lock);
95 if (mem_section[root]) {
100 mem_section[root] = section;
102 spin_unlock(&index_init_lock);
105 #else /* !SPARSEMEM_EXTREME */
106 static inline int sparse_index_init(unsigned long section_nr, int nid)
113 * Although written for the SPARSEMEM_EXTREME case, this happens
114 * to also work for the flat array case because
115 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
117 int __section_nr(struct mem_section* ms)
119 unsigned long root_nr;
120 struct mem_section* root;
122 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
123 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
127 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
131 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
135 * During early boot, before section_mem_map is used for an actual
136 * mem_map, we use section_mem_map to store the section's NUMA
137 * node. This keeps us from having to use another data structure. The
138 * node information is cleared just before we store the real mem_map.
140 static inline unsigned long sparse_encode_early_nid(int nid)
142 return (nid << SECTION_NID_SHIFT);
145 static inline int sparse_early_nid(struct mem_section *section)
147 return (section->section_mem_map >> SECTION_NID_SHIFT);
150 /* Record a memory area against a node. */
151 void __init memory_present(int nid, unsigned long start, unsigned long end)
153 unsigned long max_arch_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
157 * Sanity checks - do not allow an architecture to pass
158 * in larger pfns than the maximum scope of sparsemem:
160 if (start >= max_arch_pfn)
162 if (end >= max_arch_pfn)
165 start &= PAGE_SECTION_MASK;
166 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
167 unsigned long section = pfn_to_section_nr(pfn);
168 struct mem_section *ms;
170 sparse_index_init(section, nid);
171 set_section_nid(section, nid);
173 ms = __nr_to_section(section);
174 if (!ms->section_mem_map)
175 ms->section_mem_map = sparse_encode_early_nid(nid) |
176 SECTION_MARKED_PRESENT;
181 * Only used by the i386 NUMA architecures, but relatively
184 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
185 unsigned long end_pfn)
188 unsigned long nr_pages = 0;
190 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
191 if (nid != early_pfn_to_nid(pfn))
194 if (pfn_present(pfn))
195 nr_pages += PAGES_PER_SECTION;
198 return nr_pages * sizeof(struct page);
202 * Subtle, we encode the real pfn into the mem_map such that
203 * the identity pfn - section_mem_map will return the actual
204 * physical page frame number.
206 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
208 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
212 * Decode mem_map from the coded memmap
214 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
216 /* mask off the extra low bits of information */
217 coded_mem_map &= SECTION_MAP_MASK;
218 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
221 static int __meminit sparse_init_one_section(struct mem_section *ms,
222 unsigned long pnum, struct page *mem_map,
223 unsigned long *pageblock_bitmap)
225 if (!present_section(ms))
228 ms->section_mem_map &= ~SECTION_MAP_MASK;
229 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
231 ms->pageblock_flags = pageblock_bitmap;
236 unsigned long usemap_size(void)
238 unsigned long size_bytes;
239 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
240 size_bytes = roundup(size_bytes, sizeof(unsigned long));
244 #ifdef CONFIG_MEMORY_HOTPLUG
245 static unsigned long *__kmalloc_section_usemap(void)
247 return kmalloc(usemap_size(), GFP_KERNEL);
249 #endif /* CONFIG_MEMORY_HOTPLUG */
251 static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum)
253 unsigned long *usemap;
254 struct mem_section *ms = __nr_to_section(pnum);
255 int nid = sparse_early_nid(ms);
257 usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
261 /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
264 printk(KERN_WARNING "%s: allocation failed\n", __func__);
268 #ifndef CONFIG_SPARSEMEM_VMEMMAP
269 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
273 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
277 map = alloc_bootmem_pages_node(NODE_DATA(nid),
278 PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION));
281 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
283 struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
286 struct mem_section *ms = __nr_to_section(pnum);
287 int nid = sparse_early_nid(ms);
289 map = sparse_mem_map_populate(pnum, nid);
293 printk(KERN_ERR "%s: sparsemem memory map backing failed "
294 "some memory will not be available.\n", __func__);
295 ms->section_mem_map = 0;
299 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
303 * Allocate the accumulated non-linear sections, allocate a mem_map
304 * for each and record the physical to section mapping.
306 void __init sparse_init(void)
310 unsigned long *usemap;
311 unsigned long **usemap_map;
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.
322 * powerpc need to call sparse_init_one_section right after each
323 * sparse_early_mem_map_alloc, so allocate usemap_map at first.
325 size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
326 usemap_map = alloc_bootmem(size);
328 panic("can not allocate usemap_map\n");
330 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
331 if (!present_section_nr(pnum))
333 usemap_map[pnum] = sparse_early_usemap_alloc(pnum);
336 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
337 if (!present_section_nr(pnum))
340 usemap = usemap_map[pnum];
344 map = sparse_early_mem_map_alloc(pnum);
348 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
352 vmemmap_populate_print_last();
354 free_bootmem(__pa(usemap_map), size);
357 #ifdef CONFIG_MEMORY_HOTPLUG
358 #ifdef CONFIG_SPARSEMEM_VMEMMAP
359 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
360 unsigned long nr_pages)
362 /* This will make the necessary allocations eventually. */
363 return sparse_mem_map_populate(pnum, nid);
365 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
367 return; /* XXX: Not implemented yet */
369 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
373 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
375 struct page *page, *ret;
376 unsigned long memmap_size = sizeof(struct page) * nr_pages;
378 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
382 ret = vmalloc(memmap_size);
388 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
390 memset(ret, 0, memmap_size);
395 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
396 unsigned long nr_pages)
398 return __kmalloc_section_memmap(nr_pages);
401 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
403 if (is_vmalloc_addr(memmap))
406 free_pages((unsigned long)memmap,
407 get_order(sizeof(struct page) * nr_pages));
410 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
412 unsigned long maps_section_nr, removing_section_nr, i;
415 for (i = 0; i < nr_pages; i++, page++) {
416 magic = atomic_read(&page->_mapcount);
418 BUG_ON(magic == NODE_INFO);
420 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
421 removing_section_nr = page->private;
424 * When this function is called, the removing section is
425 * logical offlined state. This means all pages are isolated
426 * from page allocator. If removing section's memmap is placed
427 * on the same section, it must not be freed.
428 * If it is freed, page allocator may allocate it which will
429 * be removed physically soon.
431 if (maps_section_nr != removing_section_nr)
432 put_page_bootmem(page);
435 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
437 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
439 struct page *usemap_page;
440 unsigned long nr_pages;
445 usemap_page = virt_to_page(usemap);
447 * Check to see if allocation came from hot-plug-add
449 if (PageSlab(usemap_page)) {
452 __kfree_section_memmap(memmap, PAGES_PER_SECTION);
457 * The usemap came from bootmem. This is packed with other usemaps
458 * on the section which has pgdat at boot time. Just keep it as is now.
462 struct page *memmap_page;
463 memmap_page = virt_to_page(memmap);
465 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
468 free_map_bootmem(memmap_page, nr_pages);
473 * returns the number of sections whose mem_maps were properly
474 * set. If this is <=0, then that means that the passed-in
475 * map was not consumed and must be freed.
477 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
480 unsigned long section_nr = pfn_to_section_nr(start_pfn);
481 struct pglist_data *pgdat = zone->zone_pgdat;
482 struct mem_section *ms;
484 unsigned long *usemap;
489 * no locking for this, because it does its own
490 * plus, it does a kmalloc
492 ret = sparse_index_init(section_nr, pgdat->node_id);
493 if (ret < 0 && ret != -EEXIST)
495 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
498 usemap = __kmalloc_section_usemap();
500 __kfree_section_memmap(memmap, nr_pages);
504 pgdat_resize_lock(pgdat, &flags);
506 ms = __pfn_to_section(start_pfn);
507 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
512 ms->section_mem_map |= SECTION_MARKED_PRESENT;
514 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
517 pgdat_resize_unlock(pgdat, &flags);
520 __kfree_section_memmap(memmap, nr_pages);
525 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
527 struct page *memmap = NULL;
528 unsigned long *usemap = NULL;
530 if (ms->section_mem_map) {
531 usemap = ms->pageblock_flags;
532 memmap = sparse_decode_mem_map(ms->section_mem_map,
534 ms->section_mem_map = 0;
535 ms->pageblock_flags = NULL;
538 free_section_usemap(memmap, usemap);