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         if (!section)
  87                 return -ENOMEM;
  88         /*
  89          * This lock keeps two different sections from
  90          * reallocating for the same index
  91          */
  92         spin_lock(&index_init_lock);
  93
  94         if (mem_section[root]) {
  95                 ret = -EEXIST;
  96                 goto out;
  97         }
  98
  99         mem_section[root] = section;
 100 out:
 101         spin_unlock(&index_init_lock);
 102         return ret;
 103 }
 104 #else /* !SPARSEMEM_EXTREME */
 105 static inline int sparse_index_init(unsigned long section_nr, int nid)
 106 {
 107         return 0;
 108 }
 109 #endif
 110
 111 /*
 112  * Although written for the SPARSEMEM_EXTREME case, this happens
 113  * to also work for the flat array case because
 114  * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
 115  */
 116 int __section_nr(struct mem_section* ms)
 117 {
 118         unsigned long root_nr;
 119         struct mem_section* root;
 120
 121         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 122                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 123                 if (!root)
 124                         continue;
 125
 126                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 127                      break;
 128         }
 129
 130         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 131 }
 132
 133 /*
 134  * During early boot, before section_mem_map is used for an actual
 135  * mem_map, we use section_mem_map to store the section's NUMA
 136  * node.  This keeps us from having to use another data structure.  The
 137  * node information is cleared just before we store the real mem_map.
 138  */
 139 static inline unsigned long sparse_encode_early_nid(int nid)
 140 {
 141         return (nid << SECTION_NID_SHIFT);
 142 }
 143
 144 static inline int sparse_early_nid(struct mem_section *section)
 145 {
 146         return (section->section_mem_map >> SECTION_NID_SHIFT);
 147 }
 148
 149 /* Record a memory area against a node. */
 150 void __init memory_present(int nid, unsigned long start, unsigned long end)
 151 {
 152         unsigned long pfn;
 153
 154         start &= PAGE_SECTION_MASK;
 155         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 156                 unsigned long section = pfn_to_section_nr(pfn);
 157                 struct mem_section *ms;
 158
 159                 sparse_index_init(section, nid);
 160                 set_section_nid(section, nid);
 161
 162                 ms = __nr_to_section(section);
 163                 if (!ms->section_mem_map)
 164                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 165                                                         SECTION_MARKED_PRESENT;
 166         }
 167 }
 168
 169 /*
 170  * Only used by the i386 NUMA architecures, but relatively
 171  * generic code.
 172  */
 173 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 174                                                      unsigned long end_pfn)
 175 {
 176         unsigned long pfn;
 177         unsigned long nr_pages = 0;
 178
 179         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 180                 if (nid != early_pfn_to_nid(pfn))
 181                         continue;
 182
 183                 if (pfn_present(pfn))
 184                         nr_pages += PAGES_PER_SECTION;
 185         }
 186
 187         return nr_pages * sizeof(struct page);
 188 }
 189
 190 /*
 191  * Subtle, we encode the real pfn into the mem_map such that
 192  * the identity pfn - section_mem_map will return the actual
 193  * physical page frame number.
 194  */
 195 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 196 {
 197         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 198 }
 199
 200 /*
 201  * We need this if we ever free the mem_maps.  While not implemented yet,
 202  * this function is included for parity with its sibling.
 203  */
 204 static __attribute((unused))
 205 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 206 {
 207         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 208 }
 209
 210 static int __meminit sparse_init_one_section(struct mem_section *ms,
 211                 unsigned long pnum, struct page *mem_map,
 212                 unsigned long *pageblock_bitmap)
 213 {
 214         if (!present_section(ms))
 215                 return -EINVAL;
 216
 217         ms->section_mem_map &= ~SECTION_MAP_MASK;
 218         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 219                                                         SECTION_HAS_MEM_MAP;
 220         ms->pageblock_flags = pageblock_bitmap;
 221
 222         return 1;
 223 }
 224
 225 static unsigned long usemap_size(void)
 226 {
 227         unsigned long size_bytes;
 228         size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 229         size_bytes = roundup(size_bytes, sizeof(unsigned long));
 230         return size_bytes;
 231 }
 232
 233 #ifdef CONFIG_MEMORY_HOTPLUG
 234 static unsigned long *__kmalloc_section_usemap(void)
 235 {
 236         return kmalloc(usemap_size(), GFP_KERNEL);
 237 }
 238 #endif /* CONFIG_MEMORY_HOTPLUG */
 239
 240 static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum)
 241 {
 242         unsigned long *usemap;
 243         struct mem_section *ms = __nr_to_section(pnum);
 244         int nid = sparse_early_nid(ms);
 245
 246         usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
 247         if (usemap)
 248                 return usemap;
 249
 250         /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
 251         nid = 0;
 252
 253         printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
 254         return NULL;
 255 }
 256
 257 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 258 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 259 {
 260         struct page *map;
 261
 262         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 263         if (map)
 264                 return map;
 265
 266         map = alloc_bootmem_node(NODE_DATA(nid),
 267                         sizeof(struct page) * PAGES_PER_SECTION);
 268         return map;
 269 }
 270 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 271
 272 struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 273 {
 274         struct page *map;
 275         struct mem_section *ms = __nr_to_section(pnum);
 276         int nid = sparse_early_nid(ms);
 277
 278         map = sparse_mem_map_populate(pnum, nid);
 279         if (map)
 280                 return map;
 281
 282         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 283                         "some memory will not be available.\n", __FUNCTION__);
 284         ms->section_mem_map = 0;
 285         return NULL;
 286 }
 287
 288 /*
 289  * Allocate the accumulated non-linear sections, allocate a mem_map
 290  * for each and record the physical to section mapping.
 291  */
 292 void __init sparse_init(void)
 293 {
 294         unsigned long pnum;
 295         struct page *map;
 296         unsigned long *usemap;
 297
 298         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 299                 if (!present_section_nr(pnum))
 300                         continue;
 301
 302                 map = sparse_early_mem_map_alloc(pnum);
 303                 if (!map)
 304                         continue;
 305
 306                 usemap = sparse_early_usemap_alloc(pnum);
 307                 if (!usemap)
 308                         continue;
 309
 310                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 311                                                                 usemap);
 312         }
 313 }
 314
 315 #ifdef CONFIG_MEMORY_HOTPLUG
 316 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 317 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 318                                                  unsigned long nr_pages)
 319 {
 320         /* This will make the necessary allocations eventually. */
 321         return sparse_mem_map_populate(pnum, nid);
 322 }
 323 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 324 {
 325         return; /* XXX: Not implemented yet */
 326 }
 327 #else
 328 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 329 {
 330         struct page *page, *ret;
 331         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 332
 333         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 334         if (page)
 335                 goto got_map_page;
 336
 337         ret = vmalloc(memmap_size);
 338         if (ret)
 339                 goto got_map_ptr;
 340
 341         return NULL;
 342 got_map_page:
 343         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 344 got_map_ptr:
 345         memset(ret, 0, memmap_size);
 346
 347         return ret;
 348 }
 349
 350 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 351                                                   unsigned long nr_pages)
 352 {
 353         return __kmalloc_section_memmap(nr_pages);
 354 }
 355
 356 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 357 {
 358         if (is_vmalloc_addr(memmap))
 359                 vfree(memmap);
 360         else
 361                 free_pages((unsigned long)memmap,
 362                            get_order(sizeof(struct page) * nr_pages));
 363 }
 364 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 365
 366 /*
 367  * returns the number of sections whose mem_maps were properly
 368  * set.  If this is <=0, then that means that the passed-in
 369  * map was not consumed and must be freed.
 370  */
 371 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 372                            int nr_pages)
 373 {
 374         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 375         struct pglist_data *pgdat = zone->zone_pgdat;
 376         struct mem_section *ms;
 377         struct page *memmap;
 378         unsigned long *usemap;
 379         unsigned long flags;
 380         int ret;
 381
 382         /*
 383          * no locking for this, because it does its own
 384          * plus, it does a kmalloc
 385          */
 386         ret = sparse_index_init(section_nr, pgdat->node_id);
 387         if (ret < 0 && ret != -EEXIST)
 388                 return ret;
 389         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
 390         if (!memmap)
 391                 return -ENOMEM;
 392         usemap = __kmalloc_section_usemap();
 393         if (!usemap) {
 394                 __kfree_section_memmap(memmap, nr_pages);
 395                 return -ENOMEM;
 396         }
 397
 398         pgdat_resize_lock(pgdat, &flags);
 399
 400         ms = __pfn_to_section(start_pfn);
 401         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 402                 ret = -EEXIST;
 403                 goto out;
 404         }
 405
 406         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 407
 408         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 409
 410 out:
 411         pgdat_resize_unlock(pgdat, &flags);
 412         if (ret <= 0) {
 413                 kfree(usemap);
 414                 __kfree_section_memmap(memmap, nr_pages);
 415         }
 416         return ret;
 417 }
 418 #endif