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>
12 #include <asm/pgalloc.h>
13 #include <asm/pgtable.h>
16 * Permanent SPARSEMEM data:
18 * 1) mem_section - memory sections, mem_map's for valid memory
20 #ifdef CONFIG_SPARSEMEM_EXTREME
21 struct mem_section *mem_section[NR_SECTION_ROOTS]
22 ____cacheline_internodealigned_in_smp;
24 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
25 ____cacheline_internodealigned_in_smp;
27 EXPORT_SYMBOL(mem_section);
29 #ifdef NODE_NOT_IN_PAGE_FLAGS
31 * If we did not store the node number in the page then we have to
32 * do a lookup in the section_to_node_table in order to find which
33 * node the page belongs to.
35 #if MAX_NUMNODES <= 256
36 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
38 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
41 int page_to_nid(struct page *page)
43 return section_to_node_table[page_to_section(page)];
45 EXPORT_SYMBOL(page_to_nid);
47 static void set_section_nid(unsigned long section_nr, int nid)
49 section_to_node_table[section_nr] = nid;
51 #else /* !NODE_NOT_IN_PAGE_FLAGS */
52 static inline void set_section_nid(unsigned long section_nr, int nid)
57 #ifdef CONFIG_SPARSEMEM_EXTREME
58 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
60 struct mem_section *section = NULL;
61 unsigned long array_size = SECTIONS_PER_ROOT *
62 sizeof(struct mem_section);
64 if (slab_is_available())
65 section = kmalloc_node(array_size, GFP_KERNEL, nid);
67 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
70 memset(section, 0, array_size);
75 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
77 static DEFINE_SPINLOCK(index_init_lock);
78 unsigned long root = SECTION_NR_TO_ROOT(section_nr);
79 struct mem_section *section;
82 if (mem_section[root])
85 section = sparse_index_alloc(nid);
87 * This lock keeps two different sections from
88 * reallocating for the same index
90 spin_lock(&index_init_lock);
92 if (mem_section[root]) {
97 mem_section[root] = section;
99 spin_unlock(&index_init_lock);
102 #else /* !SPARSEMEM_EXTREME */
103 static inline int sparse_index_init(unsigned long section_nr, int nid)
110 * Although written for the SPARSEMEM_EXTREME case, this happens
111 * to also work for the flat array case because
112 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
114 int __section_nr(struct mem_section* ms)
116 unsigned long root_nr;
117 struct mem_section* root;
119 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
120 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
124 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
128 return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
132 * During early boot, before section_mem_map is used for an actual
133 * mem_map, we use section_mem_map to store the section's NUMA
134 * node. This keeps us from having to use another data structure. The
135 * node information is cleared just before we store the real mem_map.
137 static inline unsigned long sparse_encode_early_nid(int nid)
139 return (nid << SECTION_NID_SHIFT);
142 static inline int sparse_early_nid(struct mem_section *section)
144 return (section->section_mem_map >> SECTION_NID_SHIFT);
147 /* Record a memory area against a node. */
148 void __init memory_present(int nid, unsigned long start, unsigned long end)
152 start &= PAGE_SECTION_MASK;
153 for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
154 unsigned long section = pfn_to_section_nr(pfn);
155 struct mem_section *ms;
157 sparse_index_init(section, nid);
158 set_section_nid(section, nid);
160 ms = __nr_to_section(section);
161 if (!ms->section_mem_map)
162 ms->section_mem_map = sparse_encode_early_nid(nid) |
163 SECTION_MARKED_PRESENT;
168 * Only used by the i386 NUMA architecures, but relatively
171 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
172 unsigned long end_pfn)
175 unsigned long nr_pages = 0;
177 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
178 if (nid != early_pfn_to_nid(pfn))
181 if (pfn_present(pfn))
182 nr_pages += PAGES_PER_SECTION;
185 return nr_pages * sizeof(struct page);
189 * Subtle, we encode the real pfn into the mem_map such that
190 * the identity pfn - section_mem_map will return the actual
191 * physical page frame number.
193 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
195 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
199 * We need this if we ever free the mem_maps. While not implemented yet,
200 * this function is included for parity with its sibling.
202 static __attribute((unused))
203 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
205 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
208 static int __meminit sparse_init_one_section(struct mem_section *ms,
209 unsigned long pnum, struct page *mem_map,
210 unsigned long *pageblock_bitmap)
212 if (!present_section(ms))
215 ms->section_mem_map &= ~SECTION_MAP_MASK;
216 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
218 ms->pageblock_flags = pageblock_bitmap;
223 __attribute__((weak)) __init
224 void *alloc_bootmem_high_node(pg_data_t *pgdat, unsigned long size)
229 static unsigned long usemap_size(void)
231 unsigned long size_bytes;
232 size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
233 size_bytes = roundup(size_bytes, sizeof(unsigned long));
237 #ifdef CONFIG_MEMORY_HOTPLUG
238 static unsigned long *__kmalloc_section_usemap(void)
240 return kmalloc(usemap_size(), GFP_KERNEL);
242 #endif /* CONFIG_MEMORY_HOTPLUG */
244 static unsigned long *sparse_early_usemap_alloc(unsigned long pnum)
246 unsigned long *usemap;
247 struct mem_section *ms = __nr_to_section(pnum);
248 int nid = sparse_early_nid(ms);
250 usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
254 /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
257 printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
261 #ifndef CONFIG_SPARSEMEM_VMEMMAP
262 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
266 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
270 map = alloc_bootmem_high_node(NODE_DATA(nid),
271 sizeof(struct page) * PAGES_PER_SECTION);
275 map = alloc_bootmem_node(NODE_DATA(nid),
276 sizeof(struct page) * PAGES_PER_SECTION);
279 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
281 struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
284 struct mem_section *ms = __nr_to_section(pnum);
285 int nid = sparse_early_nid(ms);
287 map = sparse_mem_map_populate(pnum, nid);
291 printk(KERN_ERR "%s: sparsemem memory map backing failed "
292 "some memory will not be available.\n", __FUNCTION__);
293 ms->section_mem_map = 0;
298 * Allocate the accumulated non-linear sections, allocate a mem_map
299 * for each and record the physical to section mapping.
301 void __init sparse_init(void)
305 unsigned long *usemap;
307 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
308 if (!present_section_nr(pnum))
311 map = sparse_early_mem_map_alloc(pnum);
315 usemap = sparse_early_usemap_alloc(pnum);
319 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
324 #ifdef CONFIG_MEMORY_HOTPLUG
325 #ifdef CONFIG_SPARSEMEM_VMEMMAP
326 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
327 unsigned long nr_pages)
329 /* This will make the necessary allocations eventually. */
330 return sparse_mem_map_populate(pnum, nid);
332 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
334 return; /* XXX: Not implemented yet */
337 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
339 struct page *page, *ret;
340 unsigned long memmap_size = sizeof(struct page) * nr_pages;
342 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
346 ret = vmalloc(memmap_size);
352 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
354 memset(ret, 0, memmap_size);
359 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
360 unsigned long nr_pages)
362 return __kmalloc_section_memmap(nr_pages);
365 static int vaddr_in_vmalloc_area(void *addr)
367 if (addr >= (void *)VMALLOC_START &&
368 addr < (void *)VMALLOC_END)
373 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
375 if (vaddr_in_vmalloc_area(memmap))
378 free_pages((unsigned long)memmap,
379 get_order(sizeof(struct page) * nr_pages));
381 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
384 * returns the number of sections whose mem_maps were properly
385 * set. If this is <=0, then that means that the passed-in
386 * map was not consumed and must be freed.
388 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
391 unsigned long section_nr = pfn_to_section_nr(start_pfn);
392 struct pglist_data *pgdat = zone->zone_pgdat;
393 struct mem_section *ms;
395 unsigned long *usemap;
400 * no locking for this, because it does its own
401 * plus, it does a kmalloc
403 sparse_index_init(section_nr, pgdat->node_id);
404 memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
405 usemap = __kmalloc_section_usemap();
407 pgdat_resize_lock(pgdat, &flags);
409 ms = __pfn_to_section(start_pfn);
410 if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
419 ms->section_mem_map |= SECTION_MARKED_PRESENT;
421 ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
424 pgdat_resize_unlock(pgdat, &flags);
426 __kfree_section_memmap(memmap, nr_pages);