sparse.c revision 8f6aac419bd590f535fb110875a51f7db2b62b5b
1/* 2 * sparse memory mappings. 3 */ 4#include <linux/mm.h> 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> 11#include <asm/dma.h> 12#include <asm/pgalloc.h> 13#include <asm/pgtable.h> 14 15/* 16 * Permanent SPARSEMEM data: 17 * 18 * 1) mem_section - memory sections, mem_map's for valid memory 19 */ 20#ifdef CONFIG_SPARSEMEM_EXTREME 21struct mem_section *mem_section[NR_SECTION_ROOTS] 22 ____cacheline_internodealigned_in_smp; 23#else 24struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT] 25 ____cacheline_internodealigned_in_smp; 26#endif 27EXPORT_SYMBOL(mem_section); 28 29#ifdef NODE_NOT_IN_PAGE_FLAGS 30/* 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. 34 */ 35#if MAX_NUMNODES <= 256 36static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned; 37#else 38static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned; 39#endif 40 41int page_to_nid(struct page *page) 42{ 43 return section_to_node_table[page_to_section(page)]; 44} 45EXPORT_SYMBOL(page_to_nid); 46 47static void set_section_nid(unsigned long section_nr, int nid) 48{ 49 section_to_node_table[section_nr] = nid; 50} 51#else /* !NODE_NOT_IN_PAGE_FLAGS */ 52static inline void set_section_nid(unsigned long section_nr, int nid) 53{ 54} 55#endif 56 57#ifdef CONFIG_SPARSEMEM_EXTREME 58static struct mem_section noinline __init_refok *sparse_index_alloc(int nid) 59{ 60 struct mem_section *section = NULL; 61 unsigned long array_size = SECTIONS_PER_ROOT * 62 sizeof(struct mem_section); 63 64 if (slab_is_available()) 65 section = kmalloc_node(array_size, GFP_KERNEL, nid); 66 else 67 section = alloc_bootmem_node(NODE_DATA(nid), array_size); 68 69 if (section) 70 memset(section, 0, array_size); 71 72 return section; 73} 74 75static int __meminit sparse_index_init(unsigned long section_nr, int nid) 76{ 77 static DEFINE_SPINLOCK(index_init_lock); 78 unsigned long root = SECTION_NR_TO_ROOT(section_nr); 79 struct mem_section *section; 80 int ret = 0; 81 82 if (mem_section[root]) 83 return -EEXIST; 84 85 section = sparse_index_alloc(nid); 86 /* 87 * This lock keeps two different sections from 88 * reallocating for the same index 89 */ 90 spin_lock(&index_init_lock); 91 92 if (mem_section[root]) { 93 ret = -EEXIST; 94 goto out; 95 } 96 97 mem_section[root] = section; 98out: 99 spin_unlock(&index_init_lock); 100 return ret; 101} 102#else /* !SPARSEMEM_EXTREME */ 103static inline int sparse_index_init(unsigned long section_nr, int nid) 104{ 105 return 0; 106} 107#endif 108 109/* 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. 113 */ 114int __section_nr(struct mem_section* ms) 115{ 116 unsigned long root_nr; 117 struct mem_section* root; 118 119 for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) { 120 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT); 121 if (!root) 122 continue; 123 124 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT))) 125 break; 126 } 127 128 return (root_nr * SECTIONS_PER_ROOT) + (ms - root); 129} 130 131/* 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. 136 */ 137static inline unsigned long sparse_encode_early_nid(int nid) 138{ 139 return (nid << SECTION_NID_SHIFT); 140} 141 142static inline int sparse_early_nid(struct mem_section *section) 143{ 144 return (section->section_mem_map >> SECTION_NID_SHIFT); 145} 146 147/* Record a memory area against a node. */ 148void __init memory_present(int nid, unsigned long start, unsigned long end) 149{ 150 unsigned long pfn; 151 152 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; 156 157 sparse_index_init(section, nid); 158 set_section_nid(section, nid); 159 160 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; 164 } 165} 166 167/* 168 * Only used by the i386 NUMA architecures, but relatively 169 * generic code. 170 */ 171unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn, 172 unsigned long end_pfn) 173{ 174 unsigned long pfn; 175 unsigned long nr_pages = 0; 176 177 for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) { 178 if (nid != early_pfn_to_nid(pfn)) 179 continue; 180 181 if (pfn_present(pfn)) 182 nr_pages += PAGES_PER_SECTION; 183 } 184 185 return nr_pages * sizeof(struct page); 186} 187 188/* 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. 192 */ 193static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum) 194{ 195 return (unsigned long)(mem_map - (section_nr_to_pfn(pnum))); 196} 197 198/* 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. 201 */ 202static __attribute((unused)) 203struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum) 204{ 205 return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum); 206} 207 208static int __meminit sparse_init_one_section(struct mem_section *ms, 209 unsigned long pnum, struct page *mem_map) 210{ 211 if (!present_section(ms)) 212 return -EINVAL; 213 214 ms->section_mem_map &= ~SECTION_MAP_MASK; 215 ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) | 216 SECTION_HAS_MEM_MAP; 217 218 return 1; 219} 220 221__attribute__((weak)) __init 222void *alloc_bootmem_high_node(pg_data_t *pgdat, unsigned long size) 223{ 224 return NULL; 225} 226 227#ifndef CONFIG_SPARSEMEM_VMEMMAP 228struct page __init *sparse_early_mem_map_populate(unsigned long pnum, int nid) 229{ 230 struct page *map; 231 232 map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION); 233 if (map) 234 return map; 235 236 map = alloc_bootmem_high_node(NODE_DATA(nid), 237 sizeof(struct page) * PAGES_PER_SECTION); 238 if (map) 239 return map; 240 241 map = alloc_bootmem_node(NODE_DATA(nid), 242 sizeof(struct page) * PAGES_PER_SECTION); 243 return map; 244} 245#endif /* !CONFIG_SPARSEMEM_VMEMMAP */ 246 247struct page __init *sparse_early_mem_map_alloc(unsigned long pnum) 248{ 249 struct page *map; 250 struct mem_section *ms = __nr_to_section(pnum); 251 int nid = sparse_early_nid(ms); 252 253 map = sparse_early_mem_map_populate(pnum, nid); 254 if (map) 255 return map; 256 257 printk(KERN_ERR "%s: sparsemem memory map backing failed " 258 "some memory will not be available.\n", __FUNCTION__); 259 ms->section_mem_map = 0; 260 return NULL; 261} 262 263/* 264 * Allocate the accumulated non-linear sections, allocate a mem_map 265 * for each and record the physical to section mapping. 266 */ 267void __init sparse_init(void) 268{ 269 unsigned long pnum; 270 struct page *map; 271 272 for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) { 273 if (!present_section_nr(pnum)) 274 continue; 275 276 map = sparse_early_mem_map_alloc(pnum); 277 if (!map) 278 continue; 279 sparse_init_one_section(__nr_to_section(pnum), pnum, map); 280 } 281} 282 283#ifdef CONFIG_MEMORY_HOTPLUG 284static struct page *__kmalloc_section_memmap(unsigned long nr_pages) 285{ 286 struct page *page, *ret; 287 unsigned long memmap_size = sizeof(struct page) * nr_pages; 288 289 page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size)); 290 if (page) 291 goto got_map_page; 292 293 ret = vmalloc(memmap_size); 294 if (ret) 295 goto got_map_ptr; 296 297 return NULL; 298got_map_page: 299 ret = (struct page *)pfn_to_kaddr(page_to_pfn(page)); 300got_map_ptr: 301 memset(ret, 0, memmap_size); 302 303 return ret; 304} 305 306static int vaddr_in_vmalloc_area(void *addr) 307{ 308 if (addr >= (void *)VMALLOC_START && 309 addr < (void *)VMALLOC_END) 310 return 1; 311 return 0; 312} 313 314static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages) 315{ 316 if (vaddr_in_vmalloc_area(memmap)) 317 vfree(memmap); 318 else 319 free_pages((unsigned long)memmap, 320 get_order(sizeof(struct page) * nr_pages)); 321} 322 323/* 324 * returns the number of sections whose mem_maps were properly 325 * set. If this is <=0, then that means that the passed-in 326 * map was not consumed and must be freed. 327 */ 328int sparse_add_one_section(struct zone *zone, unsigned long start_pfn, 329 int nr_pages) 330{ 331 unsigned long section_nr = pfn_to_section_nr(start_pfn); 332 struct pglist_data *pgdat = zone->zone_pgdat; 333 struct mem_section *ms; 334 struct page *memmap; 335 unsigned long flags; 336 int ret; 337 338 /* 339 * no locking for this, because it does its own 340 * plus, it does a kmalloc 341 */ 342 sparse_index_init(section_nr, pgdat->node_id); 343 memmap = __kmalloc_section_memmap(nr_pages); 344 345 pgdat_resize_lock(pgdat, &flags); 346 347 ms = __pfn_to_section(start_pfn); 348 if (ms->section_mem_map & SECTION_MARKED_PRESENT) { 349 ret = -EEXIST; 350 goto out; 351 } 352 ms->section_mem_map |= SECTION_MARKED_PRESENT; 353 354 ret = sparse_init_one_section(ms, section_nr, memmap); 355 356out: 357 pgdat_resize_unlock(pgdat, &flags); 358 if (ret <= 0) 359 __kfree_section_memmap(memmap, nr_pages); 360 return ret; 361} 362#endif 363