mm/sparse.c

   1 /*
   2  * sparse memory mappings.
   3  */
   4 #include <linux/mm.h>
   5 #include <linux/slab.h>
   6 #include <linux/mmzone.h>
   7 #include <linux/bootmem.h>
   8 #include <linux/highmem.h>
   9 #include <linux/export.h>
  10 #include <linux/spinlock.h>
  11 #include <linux/vmalloc.h>
  12 #include "internal.h"
  13 #include <asm/dma.h>
  14 #include <asm/pgalloc.h>
  15 #include <asm/pgtable.h>
  16
  17 /*
  18  * Permanent SPARSEMEM data:
  19  *
  20  * 1) mem_section       - memory sections, mem_map's for valid memory
  21  */
  22 #ifdef CONFIG_SPARSEMEM_EXTREME
  23 struct mem_section *mem_section[NR_SECTION_ROOTS]
  24         ____cacheline_internodealigned_in_smp;
  25 #else
  26 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  27         ____cacheline_internodealigned_in_smp;
  28 #endif
  29 EXPORT_SYMBOL(mem_section);
  30
  31 #ifdef NODE_NOT_IN_PAGE_FLAGS
  32 /*
  33  * If we did not store the node number in the page then we have to
  34  * do a lookup in the section_to_node_table in order to find which
  35  * node the page belongs to.
  36  */
  37 #if MAX_NUMNODES <= 256
  38 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  39 #else
  40 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  41 #endif
  42
  43 int page_to_nid(const struct page *page)
  44 {
  45         return section_to_node_table[page_to_section(page)];
  46 }
  47 EXPORT_SYMBOL(page_to_nid);
  48
  49 static void set_section_nid(unsigned long section_nr, int nid)
  50 {
  51         section_to_node_table[section_nr] = nid;
  52 }
  53 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  54 static inline void set_section_nid(unsigned long section_nr, int nid)
  55 {
  56 }
  57 #endif
  58
  59 #ifdef CONFIG_SPARSEMEM_EXTREME
  60 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
  61 {
  62         struct mem_section *section = NULL;
  63         unsigned long array_size = SECTIONS_PER_ROOT *
  64                                    sizeof(struct mem_section);
  65
  66         if (slab_is_available()) {
  67                 if (node_state(nid, N_HIGH_MEMORY))
  68                         section = kzalloc_node(array_size, GFP_KERNEL, nid);
  69                 else
  70                         section = kzalloc(array_size, GFP_KERNEL);
  71         } else {
  72                 section = alloc_bootmem_node(NODE_DATA(nid), array_size);
  73         }
  74
  75         return section;
  76 }
  77
  78 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  79 {
  80         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  81         struct mem_section *section;
  82         int ret = 0;
  83
  84         if (mem_section[root])
  85                 return -EEXIST;
  86
  87         section = sparse_index_alloc(nid);
  88         if (!section)
  89                 return -ENOMEM;
  90
  91         mem_section[root] = section;
  92
  93         return ret;
  94 }
  95 #else /* !SPARSEMEM_EXTREME */
  96 static inline int sparse_index_init(unsigned long section_nr, int nid)
  97 {
  98         return 0;
  99 }
 100 #endif
 101
 102 /*
 103  * Although written for the SPARSEMEM_EXTREME case, this happens
 104  * to also work for the flat array case because
 105  * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
 106  */
 107 int __section_nr(struct mem_section* ms)
 108 {
 109         unsigned long root_nr;
 110         struct mem_section* root;
 111
 112         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 113                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 114                 if (!root)
 115                         continue;
 116
 117                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 118                      break;
 119         }
 120
 121         VM_BUG_ON(root_nr == NR_SECTION_ROOTS);
 122
 123         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 124 }
 125
 126 /*
 127  * During early boot, before section_mem_map is used for an actual
 128  * mem_map, we use section_mem_map to store the section's NUMA
 129  * node.  This keeps us from having to use another data structure.  The
 130  * node information is cleared just before we store the real mem_map.
 131  */
 132 static inline unsigned long sparse_encode_early_nid(int nid)
 133 {
 134         return (nid << SECTION_NID_SHIFT);
 135 }
 136
 137 static inline int sparse_early_nid(struct mem_section *section)
 138 {
 139         return (section->section_mem_map >> SECTION_NID_SHIFT);
 140 }
 141
 142 /* Validate the physical addressing limitations of the model */
 143 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 144                                                 unsigned long *end_pfn)
 145 {
 146         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 147
 148         /*
 149          * Sanity checks - do not allow an architecture to pass
 150          * in larger pfns than the maximum scope of sparsemem:
 151          */
 152         if (*start_pfn > max_sparsemem_pfn) {
 153                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 154                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 155                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 156                 WARN_ON_ONCE(1);
 157                 *start_pfn = max_sparsemem_pfn;
 158                 *end_pfn = max_sparsemem_pfn;
 159         } else if (*end_pfn > max_sparsemem_pfn) {
 160                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 161                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 162                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 163                 WARN_ON_ONCE(1);
 164                 *end_pfn = max_sparsemem_pfn;
 165         }
 166 }
 167
 168 /* Record a memory area against a node. */
 169 void __init memory_present(int nid, unsigned long start, unsigned long end)
 170 {
 171         unsigned long pfn;
 172
 173         start &= PAGE_SECTION_MASK;
 174         mminit_validate_memmodel_limits(&start, &end);
 175         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 176                 unsigned long section = pfn_to_section_nr(pfn);
 177                 struct mem_section *ms;
 178
 179                 sparse_index_init(section, nid);
 180                 set_section_nid(section, nid);
 181
 182                 ms = __nr_to_section(section);
 183                 if (!ms->section_mem_map)
 184                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 185                                                         SECTION_MARKED_PRESENT;
 186         }
 187 }
 188
 189 /*
 190  * Only used by the i386 NUMA architecures, but relatively
 191  * generic code.
 192  */
 193 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 194                                                      unsigned long end_pfn)
 195 {
 196         unsigned long pfn;
 197         unsigned long nr_pages = 0;
 198
 199         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 200         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 201                 if (nid != early_pfn_to_nid(pfn))
 202                         continue;
 203
 204                 if (pfn_present(pfn))
 205                         nr_pages += PAGES_PER_SECTION;
 206         }
 207
 208         return nr_pages * sizeof(struct page);
 209 }
 210
 211 /*
 212  * Subtle, we encode the real pfn into the mem_map such that
 213  * the identity pfn - section_mem_map will return the actual
 214  * physical page frame number.
 215  */
 216 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 217 {
 218         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 219 }
 220
 221 /*
 222  * Decode mem_map from the coded memmap
 223  */
 224 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 225 {
 226         /* mask off the extra low bits of information */
 227         coded_mem_map &= SECTION_MAP_MASK;
 228         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 229 }
 230
 231 static int __meminit sparse_init_one_section(struct mem_section *ms,
 232                 unsigned long pnum, struct page *mem_map,
 233                 unsigned long *pageblock_bitmap)
 234 {
 235         if (!present_section(ms))
 236                 return -EINVAL;
 237
 238         ms->section_mem_map &= ~SECTION_MAP_MASK;
 239         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 240                                                         SECTION_HAS_MEM_MAP;
 241         ms->pageblock_flags = pageblock_bitmap;
 242
 243         return 1;
 244 }
 245
 246 unsigned long usemap_size(void)
 247 {
 248         unsigned long size_bytes;
 249         size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 250         size_bytes = roundup(size_bytes, sizeof(unsigned long));
 251         return size_bytes;
 252 }
 253
 254 #ifdef CONFIG_MEMORY_HOTPLUG
 255 static unsigned long *__kmalloc_section_usemap(void)
 256 {
 257         return kmalloc(usemap_size(), GFP_KERNEL);
 258 }
 259 #endif /* CONFIG_MEMORY_HOTPLUG */
 260
 261 #ifdef CONFIG_MEMORY_HOTREMOVE
 262 static unsigned long * __init
 263 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 264                                          unsigned long size)
 265 {
 266         unsigned long goal, limit;
 267         unsigned long *p;
 268         int nid;
 269         /*
 270          * A page may contain usemaps for other sections preventing the
 271          * page being freed and making a section unremovable while
 272          * other sections referencing the usemap retmain active. Similarly,
 273          * a pgdat can prevent a section being removed. If section A
 274          * contains a pgdat and section B contains the usemap, both
 275          * sections become inter-dependent. This allocates usemaps
 276          * from the same section as the pgdat where possible to avoid
 277          * this problem.
 278          */
 279         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 280         limit = goal + (1UL << PA_SECTION_SHIFT);
 281         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 282 again:
 283         p = ___alloc_bootmem_node_nopanic(NODE_DATA(nid), size,
 284                                           SMP_CACHE_BYTES, goal, limit);
 285         if (!p && limit) {
 286                 limit = 0;
 287                 goto again;
 288         }
 289         return p;
 290 }
 291
 292 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 293 {
 294         unsigned long usemap_snr, pgdat_snr;
 295         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 296         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 297         struct pglist_data *pgdat = NODE_DATA(nid);
 298         int usemap_nid;
 299
 300         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 301         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 302         if (usemap_snr == pgdat_snr)
 303                 return;
 304
 305         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 306                 /* skip redundant message */
 307                 return;
 308
 309         old_usemap_snr = usemap_snr;
 310         old_pgdat_snr = pgdat_snr;
 311
 312         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 313         if (usemap_nid != nid) {
 314                 printk(KERN_INFO
 315                        "node %d must be removed before remove section %ld\n",
 316                        nid, usemap_snr);
 317                 return;
 318         }
 319         /*
 320          * There is a circular dependency.
 321          * Some platforms allow un-removable section because they will just
 322          * gather other removable sections for dynamic partitioning.
 323          * Just notify un-removable section's number here.
 324          */
 325         printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
 326                pgdat_snr, nid);
 327         printk(KERN_CONT
 328                " have a circular dependency on usemap and pgdat allocations\n");
 329 }
 330 #else
 331 static unsigned long * __init
 332 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 333                                          unsigned long size)
 334 {
 335         return alloc_bootmem_node_nopanic(pgdat, size);
 336 }
 337
 338 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 339 {
 340 }
 341 #endif /* CONFIG_MEMORY_HOTREMOVE */
 342
 343 static void __init sparse_early_usemaps_alloc_node(unsigned long**usemap_map,
 344                                  unsigned long pnum_begin,
 345                                  unsigned long pnum_end,
 346                                  unsigned long usemap_count, int nodeid)
 347 {
 348         void *usemap;
 349         unsigned long pnum;
 350         int size = usemap_size();
 351
 352         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 353                                                           size * usemap_count);
 354         if (!usemap) {
 355                 printk(KERN_WARNING "%s: allocation failed\n", __func__);
 356                 return;
 357         }
 358
 359         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 360                 if (!present_section_nr(pnum))
 361                         continue;
 362                 usemap_map[pnum] = usemap;
 363                 usemap += size;
 364                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 365         }
 366 }
 367
 368 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 369 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 370 {
 371         struct page *map;
 372         unsigned long size;
 373
 374         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 375         if (map)
 376                 return map;
 377
 378         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 379         map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
 380                                          PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 381         return map;
 382 }
 383 void __init sparse_mem_maps_populate_node(struct page **map_map,
 384                                           unsigned long pnum_begin,
 385                                           unsigned long pnum_end,
 386                                           unsigned long map_count, int nodeid)
 387 {
 388         void *map;
 389         unsigned long pnum;
 390         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 391
 392         map = alloc_remap(nodeid, size * map_count);
 393         if (map) {
 394                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 395                         if (!present_section_nr(pnum))
 396                                 continue;
 397                         map_map[pnum] = map;
 398                         map += size;
 399                 }
 400                 return;
 401         }
 402
 403         size = PAGE_ALIGN(size);
 404         map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
 405                                          PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 406         if (map) {
 407                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 408                         if (!present_section_nr(pnum))
 409                                 continue;
 410                         map_map[pnum] = map;
 411                         map += size;
 412                 }
 413                 return;
 414         }
 415
 416         /* fallback */
 417         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 418                 struct mem_section *ms;
 419
 420                 if (!present_section_nr(pnum))
 421                         continue;
 422                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 423                 if (map_map[pnum])
 424                         continue;
 425                 ms = __nr_to_section(pnum);
 426                 printk(KERN_ERR "%s: sparsemem memory map backing failed "
 427                         "some memory will not be available.\n", __func__);
 428                 ms->section_mem_map = 0;
 429         }
 430 }
 431 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 432
 433 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 434 static void __init sparse_early_mem_maps_alloc_node(struct page **map_map,
 435                                  unsigned long pnum_begin,
 436                                  unsigned long pnum_end,
 437                                  unsigned long map_count, int nodeid)
 438 {
 439         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 440                                          map_count, nodeid);
 441 }
 442 #else
 443 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 444 {
 445         struct page *map;
 446         struct mem_section *ms = __nr_to_section(pnum);
 447         int nid = sparse_early_nid(ms);
 448
 449         map = sparse_mem_map_populate(pnum, nid);
 450         if (map)
 451                 return map;
 452
 453         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 454                         "some memory will not be available.\n", __func__);
 455         ms->section_mem_map = 0;
 456         return NULL;
 457 }
 458 #endif
 459
 460 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 461 {
 462 }
 463
 464 /*
 465  * Allocate the accumulated non-linear sections, allocate a mem_map
 466  * for each and record the physical to section mapping.
 467  */
 468 void __init sparse_init(void)
 469 {
 470         unsigned long pnum;
 471         struct page *map;
 472         unsigned long *usemap;
 473         unsigned long **usemap_map;
 474         int size;
 475         int nodeid_begin = 0;
 476         unsigned long pnum_begin = 0;
 477         unsigned long usemap_count;
 478 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 479         unsigned long map_count;
 480         int size2;
 481         struct page **map_map;
 482 #endif
 483
 484         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 485         set_pageblock_order();
 486
 487         /*
 488          * map is using big page (aka 2M in x86 64 bit)
 489          * usemap is less one page (aka 24 bytes)
 490          * so alloc 2M (with 2M align) and 24 bytes in turn will
 491          * make next 2M slip to one more 2M later.
 492          * then in big system, the memory will have a lot of holes...
 493          * here try to allocate 2M pages continuously.
 494          *
 495          * powerpc need to call sparse_init_one_section right after each
 496          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 497          */
 498         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 499         usemap_map = alloc_bootmem(size);
 500         if (!usemap_map)
 501                 panic("can not allocate usemap_map\n");
 502
 503         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 504                 struct mem_section *ms;
 505
 506                 if (!present_section_nr(pnum))
 507                         continue;
 508                 ms = __nr_to_section(pnum);
 509                 nodeid_begin = sparse_early_nid(ms);
 510                 pnum_begin = pnum;
 511                 break;
 512         }
 513         usemap_count = 1;
 514         for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 515                 struct mem_section *ms;
 516                 int nodeid;
 517
 518                 if (!present_section_nr(pnum))
 519                         continue;
 520                 ms = __nr_to_section(pnum);
 521                 nodeid = sparse_early_nid(ms);
 522                 if (nodeid == nodeid_begin) {
 523                         usemap_count++;
 524                         continue;
 525                 }
 526                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 527                 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, pnum,
 528                                                  usemap_count, nodeid_begin);
 529                 /* new start, update count etc*/
 530                 nodeid_begin = nodeid;
 531                 pnum_begin = pnum;
 532                 usemap_count = 1;
 533         }
 534         /* ok, last chunk */
 535         sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS,
 536                                          usemap_count, nodeid_begin);
 537
 538 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 539         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 540         map_map = alloc_bootmem(size2);
 541         if (!map_map)
 542                 panic("can not allocate map_map\n");
 543
 544         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 545                 struct mem_section *ms;
 546
 547                 if (!present_section_nr(pnum))
 548                         continue;
 549                 ms = __nr_to_section(pnum);
 550                 nodeid_begin = sparse_early_nid(ms);
 551                 pnum_begin = pnum;
 552                 break;
 553         }
 554         map_count = 1;
 555         for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 556                 struct mem_section *ms;
 557                 int nodeid;
 558
 559                 if (!present_section_nr(pnum))
 560                         continue;
 561                 ms = __nr_to_section(pnum);
 562                 nodeid = sparse_early_nid(ms);
 563                 if (nodeid == nodeid_begin) {
 564                         map_count++;
 565                         continue;
 566                 }
 567                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 568                 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, pnum,
 569                                                  map_count, nodeid_begin);
 570                 /* new start, update count etc*/
 571                 nodeid_begin = nodeid;
 572                 pnum_begin = pnum;
 573                 map_count = 1;
 574         }
 575         /* ok, last chunk */
 576         sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS,
 577                                          map_count, nodeid_begin);
 578 #endif
 579
 580         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 581                 if (!present_section_nr(pnum))
 582                         continue;
 583
 584                 usemap = usemap_map[pnum];
 585                 if (!usemap)
 586                         continue;
 587
 588 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 589                 map = map_map[pnum];
 590 #else
 591                 map = sparse_early_mem_map_alloc(pnum);
 592 #endif
 593                 if (!map)
 594                         continue;
 595
 596                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 597                                                                 usemap);
 598         }
 599
 600         vmemmap_populate_print_last();
 601
 602 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 603         free_bootmem(__pa(map_map), size2);
 604 #endif
 605         free_bootmem(__pa(usemap_map), size);
 606 }
 607
 608 #ifdef CONFIG_MEMORY_HOTPLUG
 609 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 610 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 611                                                  unsigned long nr_pages)
 612 {
 613         /* This will make the necessary allocations eventually. */
 614         return sparse_mem_map_populate(pnum, nid);
 615 }
 616 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 617 {
 618         return; /* XXX: Not implemented yet */
 619 }
 620 static void free_map_bootmem(struct page *memmap, unsigned long nr_pages)
 621 {
 622 }
 623 #else
 624 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 625 {
 626         struct page *page, *ret;
 627         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 628
 629         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 630         if (page)
 631                 goto got_map_page;
 632
 633         ret = vmalloc(memmap_size);
 634         if (ret)
 635                 goto got_map_ptr;
 636
 637         return NULL;
 638 got_map_page:
 639         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 640 got_map_ptr:
 641
 642         return ret;
 643 }
 644
 645 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 646                                                   unsigned long nr_pages)
 647 {
 648         return __kmalloc_section_memmap(nr_pages);
 649 }
 650
 651 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 652 {
 653         if (is_vmalloc_addr(memmap))
 654                 vfree(memmap);
 655         else
 656                 free_pages((unsigned long)memmap,
 657                            get_order(sizeof(struct page) * nr_pages));
 658 }
 659
 660 static void free_map_bootmem(struct page *memmap, unsigned long nr_pages)
 661 {
 662         unsigned long maps_section_nr, removing_section_nr, i;
 663         unsigned long magic;
 664         struct page *page = virt_to_page(memmap);
 665
 666         for (i = 0; i < nr_pages; i++, page++) {
 667                 magic = (unsigned long) page->lru.next;
 668
 669                 BUG_ON(magic == NODE_INFO);
 670
 671                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 672                 removing_section_nr = page->private;
 673
 674                 /*
 675                  * When this function is called, the removing section is
 676                  * logical offlined state. This means all pages are isolated
 677                  * from page allocator. If removing section's memmap is placed
 678                  * on the same section, it must not be freed.
 679                  * If it is freed, page allocator may allocate it which will
 680                  * be removed physically soon.
 681                  */
 682                 if (maps_section_nr != removing_section_nr)
 683                         put_page_bootmem(page);
 684         }
 685 }
 686 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 687
 688 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 689 {
 690         struct page *usemap_page;
 691         unsigned long nr_pages;
 692
 693         if (!usemap)
 694                 return;
 695
 696         usemap_page = virt_to_page(usemap);
 697         /*
 698          * Check to see if allocation came from hot-plug-add
 699          */
 700         if (PageSlab(usemap_page)) {
 701                 kfree(usemap);
 702                 if (memmap)
 703                         __kfree_section_memmap(memmap, PAGES_PER_SECTION);
 704                 return;
 705         }
 706
 707         /*
 708          * The usemap came from bootmem. This is packed with other usemaps
 709          * on the section which has pgdat at boot time. Just keep it as is now.
 710          */
 711
 712         if (memmap) {
 713                 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 714                         >> PAGE_SHIFT;
 715
 716                 free_map_bootmem(memmap, nr_pages);
 717         }
 718 }
 719
 720 /*
 721  * returns the number of sections whose mem_maps were properly
 722  * set.  If this is <=0, then that means that the passed-in
 723  * map was not consumed and must be freed.
 724  */
 725 int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 726                            int nr_pages)
 727 {
 728         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 729         struct pglist_data *pgdat = zone->zone_pgdat;
 730         struct mem_section *ms;
 731         struct page *memmap;
 732         unsigned long *usemap;
 733         unsigned long flags;
 734         int ret;
 735
 736         /*
 737          * no locking for this, because it does its own
 738          * plus, it does a kmalloc
 739          */
 740         ret = sparse_index_init(section_nr, pgdat->node_id);
 741         if (ret < 0 && ret != -EEXIST)
 742                 return ret;
 743         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
 744         if (!memmap)
 745                 return -ENOMEM;
 746         usemap = __kmalloc_section_usemap();
 747         if (!usemap) {
 748                 __kfree_section_memmap(memmap, nr_pages);
 749                 return -ENOMEM;
 750         }
 751
 752         pgdat_resize_lock(pgdat, &flags);
 753
 754         ms = __pfn_to_section(start_pfn);
 755         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 756                 ret = -EEXIST;
 757                 goto out;
 758         }
 759
 760         memset(memmap, 0, sizeof(struct page) * nr_pages);
 761
 762         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 763
 764         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 765
 766 out:
 767         pgdat_resize_unlock(pgdat, &flags);
 768         if (ret <= 0) {
 769                 kfree(usemap);
 770                 __kfree_section_memmap(memmap, nr_pages);
 771         }
 772         return ret;
 773 }
 774
 775 #ifdef CONFIG_MEMORY_FAILURE
 776 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 777 {
 778         int i;
 779
 780         if (!memmap)
 781                 return;
 782
 783         for (i = 0; i < PAGES_PER_SECTION; i++) {
 784                 if (PageHWPoison(&memmap[i])) {
 785                         atomic_long_sub(1, &mce_bad_pages);
 786                         ClearPageHWPoison(&memmap[i]);
 787                 }
 788         }
 789 }
 790 #else
 791 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 792 {
 793 }
 794 #endif
 795
 796 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 797 {
 798         struct page *memmap = NULL;
 799         unsigned long *usemap = NULL;
 800
 801         if (ms->section_mem_map) {
 802                 usemap = ms->pageblock_flags;
 803                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 804                                                 __section_nr(ms));
 805                 ms->section_mem_map = 0;
 806                 ms->pageblock_flags = NULL;
 807         }
 808
 809         clear_hwpoisoned_pages(memmap, PAGES_PER_SECTION);
 810         free_section_usemap(memmap, usemap);
 811 }
 812 #endif