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 max_arch_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 153         unsigned long pfn;
 154
 155         /*
 156          * Sanity checks - do not allow an architecture to pass
 157          * in larger pfns than the maximum scope of sparsemem:
 158          */
 159         if (start >= max_arch_pfn)
 160                 return;
 161         if (end >= max_arch_pfn)
 162                 end = max_arch_pfn;
 163
 164         start &= PAGE_SECTION_MASK;
 165         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 166                 unsigned long section = pfn_to_section_nr(pfn);
 167                 struct mem_section *ms;
 168
 169                 sparse_index_init(section, nid);
 170                 set_section_nid(section, nid);
 171
 172                 ms = __nr_to_section(section);
 173                 if (!ms->section_mem_map)
 174                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 175                                                         SECTION_MARKED_PRESENT;
 176         }
 177 }
 178
 179 /*
 180  * Only used by the i386 NUMA architecures, but relatively
 181  * generic code.
 182  */
 183 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 184                                                      unsigned long end_pfn)
 185 {
 186         unsigned long pfn;
 187         unsigned long nr_pages = 0;
 188
 189         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 190                 if (nid != early_pfn_to_nid(pfn))
 191                         continue;
 192
 193                 if (pfn_present(pfn))
 194                         nr_pages += PAGES_PER_SECTION;
 195         }
 196
 197         return nr_pages * sizeof(struct page);
 198 }
 199
 200 /*
 201  * Subtle, we encode the real pfn into the mem_map such that
 202  * the identity pfn - section_mem_map will return the actual
 203  * physical page frame number.
 204  */
 205 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 206 {
 207         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 208 }
 209
 210 /*
 211  * Decode mem_map from the coded memmap
 212  */
 213 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 214 {
 215         /* mask off the extra low bits of information */
 216         coded_mem_map &= SECTION_MAP_MASK;
 217         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 218 }
 219
 220 static int __meminit sparse_init_one_section(struct mem_section *ms,
 221                 unsigned long pnum, struct page *mem_map,
 222                 unsigned long *pageblock_bitmap)
 223 {
 224         if (!present_section(ms))
 225                 return -EINVAL;
 226
 227         ms->section_mem_map &= ~SECTION_MAP_MASK;
 228         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 229                                                         SECTION_HAS_MEM_MAP;
 230         ms->pageblock_flags = pageblock_bitmap;
 231
 232         return 1;
 233 }
 234
 235 unsigned long usemap_size(void)
 236 {
 237         unsigned long size_bytes;
 238         size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 239         size_bytes = roundup(size_bytes, sizeof(unsigned long));
 240         return size_bytes;
 241 }
 242
 243 #ifdef CONFIG_MEMORY_HOTPLUG
 244 static unsigned long *__kmalloc_section_usemap(void)
 245 {
 246         return kmalloc(usemap_size(), GFP_KERNEL);
 247 }
 248 #endif /* CONFIG_MEMORY_HOTPLUG */
 249
 250 static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum)
 251 {
 252         unsigned long *usemap, section_nr;
 253         struct mem_section *ms = __nr_to_section(pnum);
 254         int nid = sparse_early_nid(ms);
 255         struct pglist_data *pgdat = NODE_DATA(nid);
 256
 257         /*
 258          * Usemap's page can't be freed until freeing other sections
 259          * which use it. And, Pgdat has same feature.
 260          * If section A has pgdat and section B has usemap for other
 261          * sections (includes section A), both sections can't be removed,
 262          * because there is the dependency each other.
 263          * To solve above issue, this collects all usemap on the same section
 264          * which has pgdat.
 265          */
 266         section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 267         usemap = alloc_bootmem_section(usemap_size(), section_nr);
 268         if (usemap)
 269                 return usemap;
 270
 271         /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
 272         nid = 0;
 273
 274         printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
 275         return NULL;
 276 }
 277
 278 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 279 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 280 {
 281         struct page *map;
 282
 283         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 284         if (map)
 285                 return map;
 286
 287         map = alloc_bootmem_pages_node(NODE_DATA(nid),
 288                        PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION));
 289         return map;
 290 }
 291 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 292
 293 struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 294 {
 295         struct page *map;
 296         struct mem_section *ms = __nr_to_section(pnum);
 297         int nid = sparse_early_nid(ms);
 298
 299         map = sparse_mem_map_populate(pnum, nid);
 300         if (map)
 301                 return map;
 302
 303         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 304                         "some memory will not be available.\n", __FUNCTION__);
 305         ms->section_mem_map = 0;
 306         return NULL;
 307 }
 308
 309 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 310 {
 311 }
 312 /*
 313  * Allocate the accumulated non-linear sections, allocate a mem_map
 314  * for each and record the physical to section mapping.
 315  */
 316 void __init sparse_init(void)
 317 {
 318         unsigned long pnum;
 319         struct page *map;
 320         unsigned long *usemap;
 321         unsigned long **usemap_map;
 322         int size;
 323
 324         /*
 325          * map is using big page (aka 2M in x86 64 bit)
 326          * usemap is less one page (aka 24 bytes)
 327          * so alloc 2M (with 2M align) and 24 bytes in turn will
 328          * make next 2M slip to one more 2M later.
 329          * then in big system, the memory will have a lot of holes...
 330          * here try to allocate 2M pages continously.
 331          *
 332          * powerpc need to call sparse_init_one_section right after each
 333          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 334          */
 335         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 336         usemap_map = alloc_bootmem(size);
 337         if (!usemap_map)
 338                 panic("can not allocate usemap_map\n");
 339
 340         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 341                 if (!present_section_nr(pnum))
 342                         continue;
 343                 usemap_map[pnum] = sparse_early_usemap_alloc(pnum);
 344         }
 345
 346         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 347                 if (!present_section_nr(pnum))
 348                         continue;
 349
 350                 usemap = usemap_map[pnum];
 351                 if (!usemap)
 352                         continue;
 353
 354                 map = sparse_early_mem_map_alloc(pnum);
 355                 if (!map)
 356                         continue;
 357
 358                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 359                                                                 usemap);
 360         }
 361
 362         vmemmap_populate_print_last();
 363
 364         free_bootmem(__pa(usemap_map), size);
 365 }
 366
 367 #ifdef CONFIG_MEMORY_HOTPLUG
 368 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 369 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 370                                                  unsigned long nr_pages)
 371 {
 372         /* This will make the necessary allocations eventually. */
 373         return sparse_mem_map_populate(pnum, nid);
 374 }
 375 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 376 {
 377         return; /* XXX: Not implemented yet */
 378 }
 379 #else
 380 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 381 {
 382         struct page *page, *ret;
 383         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 384
 385         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 386         if (page)
 387                 goto got_map_page;
 388
 389         ret = vmalloc(memmap_size);
 390         if (ret)
 391                 goto got_map_ptr;
 392
 393         return NULL;
 394 got_map_page:
 395         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 396 got_map_ptr:
 397         memset(ret, 0, memmap_size);
 398
 399         return ret;
 400 }
 401
 402 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 403                                                   unsigned long nr_pages)
 404 {
 405         return __kmalloc_section_memmap(nr_pages);
 406 }
 407
 408 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 409 {
 410         if (is_vmalloc_addr(memmap))
 411                 vfree(memmap);
 412         else
 413                 free_pages((unsigned long)memmap,
 414                            get_order(sizeof(struct page) * nr_pages));
 415 }
 416 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 417
 418 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 419 {
 420         if (!usemap)
 421                 return;
 422
 423         /*
 424          * Check to see if allocation came from hot-plug-add
 425          */
 426         if (PageSlab(virt_to_page(usemap))) {
 427                 kfree(usemap);
 428                 if (memmap)
 429                         __kfree_section_memmap(memmap, PAGES_PER_SECTION);
 430                 return;
 431         }
 432
 433         /*
 434          * TODO: Allocations came from bootmem - how do I free up ?
 435          */
 436         printk(KERN_WARNING "Not freeing up allocations from bootmem "
 437                         "- leaking memory\n");
 438 }
 439
 440 /*
 441  * returns the number of sections whose mem_maps were properly
 442  * set.  If this is <=0, then that means that the passed-in
 443  * map was not consumed and must be freed.
 444  */
 445 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 446                            int nr_pages)
 447 {
 448         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 449         struct pglist_data *pgdat = zone->zone_pgdat;
 450         struct mem_section *ms;
 451         struct page *memmap;
 452         unsigned long *usemap;
 453         unsigned long flags;
 454         int ret;
 455
 456         /*
 457          * no locking for this, because it does its own
 458          * plus, it does a kmalloc
 459          */
 460         ret = sparse_index_init(section_nr, pgdat->node_id);
 461         if (ret < 0 && ret != -EEXIST)
 462                 return ret;
 463         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
 464         if (!memmap)
 465                 return -ENOMEM;
 466         usemap = __kmalloc_section_usemap();
 467         if (!usemap) {
 468                 __kfree_section_memmap(memmap, nr_pages);
 469                 return -ENOMEM;
 470         }
 471
 472         pgdat_resize_lock(pgdat, &flags);
 473
 474         ms = __pfn_to_section(start_pfn);
 475         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 476                 ret = -EEXIST;
 477                 goto out;
 478         }
 479
 480         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 481
 482         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 483
 484 out:
 485         pgdat_resize_unlock(pgdat, &flags);
 486         if (ret <= 0) {
 487                 kfree(usemap);
 488                 __kfree_section_memmap(memmap, nr_pages);
 489         }
 490         return ret;
 491 }
 492
 493 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 494 {
 495         struct page *memmap = NULL;
 496         unsigned long *usemap = NULL;
 497
 498         if (ms->section_mem_map) {
 499                 usemap = ms->pageblock_flags;
 500                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 501                                                 __section_nr(ms));
 502                 ms->section_mem_map = 0;
 503                 ms->pageblock_flags = NULL;
 504         }
 505
 506         free_section_usemap(memmap, usemap);
 507 }
 508 #endif