2 * sparse memory mappings.
4 #include <linux/config.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>
15 * Permanent SPARSEMEM data:
17 * 1) mem_section - memory sections, mem_map's for valid memory
19 #ifdef CONFIG_SPARSEMEM_EXTREME
20 struct mem_section *mem_section[NR_SECTION_ROOTS]
21 ____cacheline_internodealigned_in_smp;
23 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
24 ____cacheline_internodealigned_in_smp;
26 EXPORT_SYMBOL(mem_section);
28 #ifdef CONFIG_SPARSEMEM_EXTREME
29 static struct mem_section *sparse_index_alloc(int nid)
31 struct mem_section *section = NULL;
32 unsigned long array_size = SECTIONS_PER_ROOT *
33 sizeof(struct mem_section);
35 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
38 memset(section, 0, array_size);
43 static int sparse_index_init(unsigned long section_nr, int nid)
45 static spinlock_t index_init_lock = SPIN_LOCK_UNLOCKED;
46 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
47 struct mem_section *section;
50 if (mem_section[root])
53 section = sparse_index_alloc(nid);
55 * This lock keeps two different sections from
56 * reallocating for the same index
58 spin_lock(&index_init_lock);
60 if (mem_section[root]) {
65 mem_section[root] = section;
67 spin_unlock(&index_init_lock);
70 #else /* !SPARSEMEM_EXTREME */
71 static inline int sparse_index_init(unsigned long section_nr, int nid)
78 * Although written for the SPARSEMEM_EXTREME case, this happens
79 * to also work for the flat array case becase
80 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
82 int __section_nr(struct mem_section* ms)
84 unsigned long root_nr;
85 struct mem_section* root;
88 root_nr < NR_MEM_SECTIONS;
89 root_nr += SECTIONS_PER_ROOT) {
90 root = __nr_to_section(root_nr);
95 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
99 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
102 /* Record a memory area against a node. */
103 void memory_present(int nid, unsigned long start, unsigned long end)
107 start &= PAGE_SECTION_MASK;
108 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
109 unsigned long section = pfn_to_section_nr(pfn);
110 struct mem_section *ms;
112 sparse_index_init(section, nid);
114 ms = __nr_to_section(section);
115 if (!ms->section_mem_map)
116 ms->section_mem_map = SECTION_MARKED_PRESENT;
121 * Only used by the i386 NUMA architecures, but relatively
124 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
125 unsigned long end_pfn)
128 unsigned long nr_pages = 0;
130 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
131 if (nid != early_pfn_to_nid(pfn))
135 nr_pages += PAGES_PER_SECTION;
138 return nr_pages * sizeof(struct page);
142 * Subtle, we encode the real pfn into the mem_map such that
143 * the identity pfn - section_mem_map will return the actual
144 * physical page frame number.
146 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
148 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
152 * We need this if we ever free the mem_maps. While not implemented yet,
153 * this function is included for parity with its sibling.
155 static __attribute((unused))
156 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
158 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
161 static int sparse_init_one_section(struct mem_section *ms,
162 unsigned long pnum, struct page *mem_map)
164 if (!valid_section(ms))
167 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum);
172 static struct page *sparse_early_mem_map_alloc(unsigned long pnum)
175 int nid = early_pfn_to_nid(section_nr_to_pfn(pnum));
176 struct mem_section *ms = __nr_to_section(pnum);
178 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
182 map = alloc_bootmem_node(NODE_DATA(nid),
183 sizeof(struct page) * PAGES_PER_SECTION);
187 printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
188 ms->section_mem_map = 0;
192 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
194 struct page *page, *ret;
195 unsigned long memmap_size = sizeof(struct page) * nr_pages;
197 page = alloc_pages(GFP_KERNEL, get_order(memmap_size));
201 ret = vmalloc(memmap_size);
207 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
209 memset(ret, 0, memmap_size);
214 static int vaddr_in_vmalloc_area(void *addr)
216 if (addr >= (void *)VMALLOC_START &&
217 addr < (void *)VMALLOC_END)
222 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
224 if (vaddr_in_vmalloc_area(memmap))
227 free_pages((unsigned long)memmap,
228 get_order(sizeof(struct page) * nr_pages));
232 * Allocate the accumulated non-linear sections, allocate a mem_map
233 * for each and record the physical to section mapping.
235 void sparse_init(void)
240 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
241 if (!valid_section_nr(pnum))
244 map = sparse_early_mem_map_alloc(pnum);
247 sparse_init_one_section(__nr_to_section(pnum), pnum, map);
252 * returns the number of sections whose mem_maps were properly
253 * set. If this is <=0, then that means that the passed-in
254 * map was not consumed and must be freed.
256 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
259 unsigned long section_nr = pfn_to_section_nr(start_pfn);
260 struct pglist_data *pgdat = zone->zone_pgdat;
261 struct mem_section *ms;
267 * no locking for this, because it does its own
268 * plus, it does a kmalloc
270 sparse_index_init(section_nr, pgdat->node_id);
271 memmap = __kmalloc_section_memmap(nr_pages);
273 pgdat_resize_lock(pgdat, &flags);
275 ms = __pfn_to_section(start_pfn);
276 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
280 ms->section_mem_map |= SECTION_MARKED_PRESENT;
282 ret = sparse_init_one_section(ms, section_nr, memmap);
285 __kfree_section_memmap(memmap, nr_pages);
287 pgdat_resize_unlock(pgdat, &flags);