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 #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
  21 struct mem_section *mem_section[NR_SECTION_ROOTS]
  22         ____cacheline_internodealigned_in_smp;
  23 #else
  24 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  25         ____cacheline_internodealigned_in_smp;
  26 #endif
  27 EXPORT_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
  36 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  37 #else
  38 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  39 #endif
  40
  41 int page_to_nid(struct page *page)
  42 {
  43         return section_to_node_table[page_to_section(page)];
  44 }
  45 EXPORT_SYMBOL(page_to_nid);
  46
  47 static 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 */
  52 static inline void set_section_nid(unsigned long section_nr, int nid)
  53 {
  54 }
  55 #endif
  56
  57 #ifdef CONFIG_SPARSEMEM_EXTREME
  58 static 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
  75 static 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;
  98 out:
  99         spin_unlock(&index_init_lock);
 100         return ret;
 101 }
 102 #else /* !SPARSEMEM_EXTREME */
 103 static 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  */
 114 int __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  */
 137 static inline unsigned long sparse_encode_early_nid(int nid)
 138 {
 139         return (nid << SECTION_NID_SHIFT);
 140 }
 141
 142 static 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. */
 148 void __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  */
 171 unsigned 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  */
 193 static 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  */
 202 static __attribute((unused))
 203 struct 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
 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)
 211 {
 212         if (!present_section(ms))
 213                 return -EINVAL;
 214
 215         ms->section_mem_map &= ~SECTION_MAP_MASK;
 216         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 217                                                         SECTION_HAS_MEM_MAP;
 218         ms->pageblock_flags = pageblock_bitmap;
 219
 220         return 1;
 221 }
 222
 223 __attribute__((weak)) __init
 224 void *alloc_bootmem_high_node(pg_data_t *pgdat, unsigned long size)
 225 {
 226         return NULL;
 227 }
 228
 229 static unsigned long usemap_size(void)
 230 {
 231         unsigned long size_bytes;
 232         size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 233         size_bytes = roundup(size_bytes, sizeof(unsigned long));
 234         return size_bytes;
 235 }
 236
 237 #ifdef CONFIG_MEMORY_HOTPLUG
 238 static unsigned long *__kmalloc_section_usemap(void)
 239 {
 240         return kmalloc(usemap_size(), GFP_KERNEL);
 241 }
 242 #endif /* CONFIG_MEMORY_HOTPLUG */
 243
 244 static unsigned long *sparse_early_usemap_alloc(unsigned long pnum)
 245 {
 246         unsigned long *usemap;
 247         struct mem_section *ms = __nr_to_section(pnum);
 248         int nid = sparse_early_nid(ms);
 249
 250         usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
 251         if (usemap)
 252                 return usemap;
 253
 254         /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
 255         nid = 0;
 256
 257         printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
 258         return NULL;
 259 }
 260
 261 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 262 struct page __init *sparse_early_mem_map_populate(unsigned long pnum, int nid)
 263 {
 264         struct page *map;
 265
 266         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 267         if (map)
 268                 return map;
 269
 270         map = alloc_bootmem_high_node(NODE_DATA(nid),
 271                        sizeof(struct page) * PAGES_PER_SECTION);
 272         if (map)
 273                 return map;
 274
 275         map = alloc_bootmem_node(NODE_DATA(nid),
 276                         sizeof(struct page) * PAGES_PER_SECTION);
 277         return map;
 278 }
 279 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 280
 281 struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 282 {
 283         struct page *map;
 284         struct mem_section *ms = __nr_to_section(pnum);
 285         int nid = sparse_early_nid(ms);
 286
 287         map = sparse_early_mem_map_populate(pnum, nid);
 288         if (map)
 289                 return map;
 290
 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;
 294         return NULL;
 295 }
 296
 297 /*
 298  * Allocate the accumulated non-linear sections, allocate a mem_map
 299  * for each and record the physical to section mapping.
 300  */
 301 void __init sparse_init(void)
 302 {
 303         unsigned long pnum;
 304         struct page *map;
 305         unsigned long *usemap;
 306
 307         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 308                 if (!present_section_nr(pnum))
 309                         continue;
 310
 311                 map = sparse_early_mem_map_alloc(pnum);
 312                 if (!map)
 313                         continue;
 314
 315                 usemap = sparse_early_usemap_alloc(pnum);
 316                 if (!usemap)
 317                         continue;
 318
 319                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 320                                                                 usemap);
 321         }
 322 }
 323
 324 #ifdef CONFIG_MEMORY_HOTPLUG
 325 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 326 {
 327         struct page *page, *ret;
 328         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 329
 330         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 331         if (page)
 332                 goto got_map_page;
 333
 334         ret = vmalloc(memmap_size);
 335         if (ret)
 336                 goto got_map_ptr;
 337
 338         return NULL;
 339 got_map_page:
 340         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 341 got_map_ptr:
 342         memset(ret, 0, memmap_size);
 343
 344         return ret;
 345 }
 346
 347 static int vaddr_in_vmalloc_area(void *addr)
 348 {
 349         if (addr >= (void *)VMALLOC_START &&
 350             addr < (void *)VMALLOC_END)
 351                 return 1;
 352         return 0;
 353 }
 354
 355 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 356 {
 357         if (vaddr_in_vmalloc_area(memmap))
 358                 vfree(memmap);
 359         else
 360                 free_pages((unsigned long)memmap,
 361                            get_order(sizeof(struct page) * nr_pages));
 362 }
 363
 364 /*
 365  * returns the number of sections whose mem_maps were properly
 366  * set.  If this is <=0, then that means that the passed-in
 367  * map was not consumed and must be freed.
 368  */
 369 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 370                            int nr_pages)
 371 {
 372         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 373         struct pglist_data *pgdat = zone->zone_pgdat;
 374         struct mem_section *ms;
 375         struct page *memmap;
 376         unsigned long *usemap;
 377         unsigned long flags;
 378         int ret;
 379
 380         /*
 381          * no locking for this, because it does its own
 382          * plus, it does a kmalloc
 383          */
 384         sparse_index_init(section_nr, pgdat->node_id);
 385         memmap = __kmalloc_section_memmap(nr_pages);
 386         usemap = __kmalloc_section_usemap();
 387
 388         pgdat_resize_lock(pgdat, &flags);
 389
 390         ms = __pfn_to_section(start_pfn);
 391         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 392                 ret = -EEXIST;
 393                 goto out;
 394         }
 395
 396         if (!usemap) {
 397                 ret = -ENOMEM;
 398                 goto out;
 399         }
 400         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 401
 402         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 403
 404 out:
 405         pgdat_resize_unlock(pgdat, &flags);
 406         if (ret <= 0)
 407                 __kfree_section_memmap(memmap, nr_pages);
 408         return ret;
 409 }
 410 #endif