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