mm/sparse.c

   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
  13 /*
  14  * Permanent SPARSEMEM data:
  15  *
  16  * 1) mem_section       - memory sections, mem_map's for valid memory
  17  */
  18 #ifdef CONFIG_SPARSEMEM_EXTREME
  19 struct mem_section *mem_section[NR_SECTION_ROOTS]
  20         ____cacheline_internodealigned_in_smp;
  21 #else
  22 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  23         ____cacheline_internodealigned_in_smp;
  24 #endif
  25 EXPORT_SYMBOL(mem_section);
  26
  27 #ifdef NODE_NOT_IN_PAGE_FLAGS
  28 /*
  29  * If we did not store the node number in the page then we have to
  30  * do a lookup in the section_to_node_table in order to find which
  31  * node the page belongs to.
  32  */
  33 #if MAX_NUMNODES <= 256
  34 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  35 #else
  36 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  37 #endif
  38
  39 int page_to_nid(struct page *page)
  40 {
  41         return section_to_node_table[page_to_section(page)];
  42 }
  43 EXPORT_SYMBOL(page_to_nid);
  44
  45 static void set_section_nid(unsigned long section_nr, int nid)
  46 {
  47         section_to_node_table[section_nr] = nid;
  48 }
  49 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  50 static inline void set_section_nid(unsigned long section_nr, int nid)
  51 {
  52 }
  53 #endif
  54
  55 #ifdef CONFIG_SPARSEMEM_EXTREME
  56 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
  57 {
  58         struct mem_section *section = NULL;
  59         unsigned long array_size = SECTIONS_PER_ROOT *
  60                                    sizeof(struct mem_section);
  61
  62         if (slab_is_available())
  63                 section = kmalloc_node(array_size, GFP_KERNEL, nid);
  64         else
  65                 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
  66
  67         if (section)
  68                 memset(section, 0, array_size);
  69
  70         return section;
  71 }
  72
  73 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  74 {
  75         static DEFINE_SPINLOCK(index_init_lock);
  76         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  77         struct mem_section *section;
  78         int ret = 0;
  79
  80         if (mem_section[root])
  81                 return -EEXIST;
  82
  83         section = sparse_index_alloc(nid);
  84         /*
  85          * This lock keeps two different sections from
  86          * reallocating for the same index
  87          */
  88         spin_lock(&index_init_lock);
  89
  90         if (mem_section[root]) {
  91                 ret = -EEXIST;
  92                 goto out;
  93         }
  94
  95         mem_section[root] = section;
  96 out:
  97         spin_unlock(&index_init_lock);
  98         return ret;
  99 }
 100 #else /* !SPARSEMEM_EXTREME */
 101 static inline int sparse_index_init(unsigned long section_nr, int nid)
 102 {
 103         return 0;
 104 }
 105 #endif
 106
 107 /*
 108  * Although written for the SPARSEMEM_EXTREME case, this happens
 109  * to also work for the flat array case becase
 110  * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
 111  */
 112 int __section_nr(struct mem_section* ms)
 113 {
 114         unsigned long root_nr;
 115         struct mem_section* root;
 116
 117         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 118                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 119                 if (!root)
 120                         continue;
 121
 122                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 123                      break;
 124         }
 125
 126         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 127 }
 128
 129 /*
 130  * During early boot, before section_mem_map is used for an actual
 131  * mem_map, we use section_mem_map to store the section's NUMA
 132  * node.  This keeps us from having to use another data structure.  The
 133  * node information is cleared just before we store the real mem_map.
 134  */
 135 static inline unsigned long sparse_encode_early_nid(int nid)
 136 {
 137         return (nid << SECTION_NID_SHIFT);
 138 }
 139
 140 static inline int sparse_early_nid(struct mem_section *section)
 141 {
 142         return (section->section_mem_map >> SECTION_NID_SHIFT);
 143 }
 144
 145 /* Record a memory area against a node. */
 146 void __init memory_present(int nid, unsigned long start, unsigned long end)
 147 {
 148         unsigned long pfn;
 149
 150         start &= PAGE_SECTION_MASK;
 151         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 152                 unsigned long section = pfn_to_section_nr(pfn);
 153                 struct mem_section *ms;
 154
 155                 sparse_index_init(section, nid);
 156                 set_section_nid(section, nid);
 157
 158                 ms = __nr_to_section(section);
 159                 if (!ms->section_mem_map)
 160                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 161                                                         SECTION_MARKED_PRESENT;
 162         }
 163 }
 164
 165 /*
 166  * Only used by the i386 NUMA architecures, but relatively
 167  * generic code.
 168  */
 169 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 170                                                      unsigned long end_pfn)
 171 {
 172         unsigned long pfn;
 173         unsigned long nr_pages = 0;
 174
 175         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 176                 if (nid != early_pfn_to_nid(pfn))
 177                         continue;
 178
 179                 if (pfn_valid(pfn))
 180                         nr_pages += PAGES_PER_SECTION;
 181         }
 182
 183         return nr_pages * sizeof(struct page);
 184 }
 185
 186 /*
 187  * Subtle, we encode the real pfn into the mem_map such that
 188  * the identity pfn - section_mem_map will return the actual
 189  * physical page frame number.
 190  */
 191 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 192 {
 193         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 194 }
 195
 196 /*
 197  * We need this if we ever free the mem_maps.  While not implemented yet,
 198  * this function is included for parity with its sibling.
 199  */
 200 static __attribute((unused))
 201 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 202 {
 203         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 204 }
 205
 206 static int __meminit sparse_init_one_section(struct mem_section *ms,
 207                 unsigned long pnum, struct page *mem_map)
 208 {
 209         if (!valid_section(ms))
 210                 return -EINVAL;
 211
 212         ms->section_mem_map &= ~SECTION_MAP_MASK;
 213         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum);
 214
 215         return 1;
 216 }
 217
 218 __attribute__((weak)) __init
 219 void *alloc_bootmem_high_node(pg_data_t *pgdat, unsigned long size)
 220 {
 221         return NULL;
 222 }
 223
 224 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 225 {
 226         struct page *map;
 227         struct mem_section *ms = __nr_to_section(pnum);
 228         int nid = sparse_early_nid(ms);
 229
 230         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 231         if (map)
 232                 return map;
 233
 234         map = alloc_bootmem_high_node(NODE_DATA(nid),
 235                        sizeof(struct page) * PAGES_PER_SECTION);
 236         if (map)
 237                 return map;
 238
 239         map = alloc_bootmem_node(NODE_DATA(nid),
 240                         sizeof(struct page) * PAGES_PER_SECTION);
 241         if (map)
 242                 return map;
 243
 244         printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
 245         ms->section_mem_map = 0;
 246         return NULL;
 247 }
 248
 249 /*
 250  * Allocate the accumulated non-linear sections, allocate a mem_map
 251  * for each and record the physical to section mapping.
 252  */
 253 void __init sparse_init(void)
 254 {
 255         unsigned long pnum;
 256         struct page *map;
 257
 258         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 259                 if (!valid_section_nr(pnum))
 260                         continue;
 261
 262                 map = sparse_early_mem_map_alloc(pnum);
 263                 if (!map)
 264                         continue;
 265                 sparse_init_one_section(__nr_to_section(pnum), pnum, map);
 266         }
 267 }
 268
 269 #ifdef CONFIG_MEMORY_HOTPLUG
 270 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 271 {
 272         struct page *page, *ret;
 273         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 274
 275         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 276         if (page)
 277                 goto got_map_page;
 278
 279         ret = vmalloc(memmap_size);
 280         if (ret)
 281                 goto got_map_ptr;
 282
 283         return NULL;
 284 got_map_page:
 285         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 286 got_map_ptr:
 287         memset(ret, 0, memmap_size);
 288
 289         return ret;
 290 }
 291
 292 static int vaddr_in_vmalloc_area(void *addr)
 293 {
 294         if (addr >= (void *)VMALLOC_START &&
 295             addr < (void *)VMALLOC_END)
 296                 return 1;
 297         return 0;
 298 }
 299
 300 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 301 {
 302         if (vaddr_in_vmalloc_area(memmap))
 303                 vfree(memmap);
 304         else
 305                 free_pages((unsigned long)memmap,
 306                            get_order(sizeof(struct page) * nr_pages));
 307 }
 308
 309 /*
 310  * returns the number of sections whose mem_maps were properly
 311  * set.  If this is <=0, then that means that the passed-in
 312  * map was not consumed and must be freed.
 313  */
 314 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 315                            int nr_pages)
 316 {
 317         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 318         struct pglist_data *pgdat = zone->zone_pgdat;
 319         struct mem_section *ms;
 320         struct page *memmap;
 321         unsigned long flags;
 322         int ret;
 323
 324         /*
 325          * no locking for this, because it does its own
 326          * plus, it does a kmalloc
 327          */
 328         sparse_index_init(section_nr, pgdat->node_id);
 329         memmap = __kmalloc_section_memmap(nr_pages);
 330
 331         pgdat_resize_lock(pgdat, &flags);
 332
 333         ms = __pfn_to_section(start_pfn);
 334         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 335                 ret = -EEXIST;
 336                 goto out;
 337         }
 338         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 339
 340         ret = sparse_init_one_section(ms, section_nr, memmap);
 341
 342 out:
 343         pgdat_resize_unlock(pgdat, &flags);
 344         if (ret <= 0)
 345                 __kfree_section_memmap(memmap, nr_pages);
 346         return ret;
 347 }
 348 #endif