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/compiler.h>
   9 #include <linux/highmem.h>
  10 #include <linux/export.h>
  11 #include <linux/spinlock.h>
  12 #include <linux/vmalloc.h>
  13
  14 #include "internal.h"
  15 #include <asm/dma.h>
  16 #include <asm/pgalloc.h>
  17 #include <asm/pgtable.h>
  18
  19 /*
  20  * Permanent SPARSEMEM data:
  21  *
  22  * 1) mem_section       - memory sections, mem_map's for valid memory
  23  */
  24 #ifdef CONFIG_SPARSEMEM_EXTREME
  25 struct mem_section *mem_section[NR_SECTION_ROOTS]
  26         ____cacheline_internodealigned_in_smp;
  27 #else
  28 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  29         ____cacheline_internodealigned_in_smp;
  30 #endif
  31 EXPORT_SYMBOL(mem_section);
  32
  33 #ifdef NODE_NOT_IN_PAGE_FLAGS
  34 /*
  35  * If we did not store the node number in the page then we have to
  36  * do a lookup in the section_to_node_table in order to find which
  37  * node the page belongs to.
  38  */
  39 #if MAX_NUMNODES <= 256
  40 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  41 #else
  42 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  43 #endif
  44
  45 int page_to_nid(const struct page *page)
  46 {
  47         return section_to_node_table[page_to_section(page)];
  48 }
  49 EXPORT_SYMBOL(page_to_nid);
  50
  51 static void set_section_nid(unsigned long section_nr, int nid)
  52 {
  53         section_to_node_table[section_nr] = nid;
  54 }
  55 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  56 static inline void set_section_nid(unsigned long section_nr, int nid)
  57 {
  58 }
  59 #endif
  60
  61 #ifdef CONFIG_SPARSEMEM_EXTREME
  62 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
  63 {
  64         struct mem_section *section = NULL;
  65         unsigned long array_size = SECTIONS_PER_ROOT *
  66                                    sizeof(struct mem_section);
  67
  68         if (slab_is_available()) {
  69                 if (node_state(nid, N_HIGH_MEMORY))
  70                         section = kzalloc_node(array_size, GFP_KERNEL, nid);
  71                 else
  72                         section = kzalloc(array_size, GFP_KERNEL);
  73         } else {
  74                 section = memblock_virt_alloc_node(array_size, nid);
  75         }
  76
  77         return section;
  78 }
  79
  80 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  81 {
  82         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  83         struct mem_section *section;
  84
  85         if (mem_section[root])
  86                 return -EEXIST;
  87
  88         section = sparse_index_alloc(nid);
  89         if (!section)
  90                 return -ENOMEM;
  91
  92         mem_section[root] = section;
  93
  94         return 0;
  95 }
  96 #else /* !SPARSEMEM_EXTREME */
  97 static inline int sparse_index_init(unsigned long section_nr, int nid)
  98 {
  99         return 0;
 100 }
 101 #endif
 102
 103 #ifdef CONFIG_SPARSEMEM_EXTREME
 104 int __section_nr(struct mem_section* ms)
 105 {
 106         unsigned long root_nr;
 107         struct mem_section* root;
 108
 109         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 110                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 111                 if (!root)
 112                         continue;
 113
 114                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 115                      break;
 116         }
 117
 118         VM_BUG_ON(root_nr == NR_SECTION_ROOTS);
 119
 120         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 121 }
 122 #else
 123 int __section_nr(struct mem_section* ms)
 124 {
 125         return (int)(ms - mem_section[0]);
 126 }
 127 #endif
 128
 129 /*
 130  * During early boot, before section_mem_map is used for an actual
 131  * mem_map, we use section_mem_map to store the section's NUMA
 132  * node.  This keeps us from having to use another data structure.  The
 133  * node information is cleared just before we store the real mem_map.
 134  */
 135 static inline unsigned long sparse_encode_early_nid(int nid)
 136 {
 137         return (nid << SECTION_NID_SHIFT);
 138 }
 139
 140 static inline int sparse_early_nid(struct mem_section *section)
 141 {
 142         return (section->section_mem_map >> SECTION_NID_SHIFT);
 143 }
 144
 145 /* Validate the physical addressing limitations of the model */
 146 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 147                                                 unsigned long *end_pfn)
 148 {
 149         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 150
 151         /*
 152          * Sanity checks - do not allow an architecture to pass
 153          * in larger pfns than the maximum scope of sparsemem:
 154          */
 155         if (*start_pfn > max_sparsemem_pfn) {
 156                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 157                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 158                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 159                 WARN_ON_ONCE(1);
 160                 *start_pfn = max_sparsemem_pfn;
 161                 *end_pfn = max_sparsemem_pfn;
 162         } else if (*end_pfn > max_sparsemem_pfn) {
 163                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 164                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 165                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 166                 WARN_ON_ONCE(1);
 167                 *end_pfn = max_sparsemem_pfn;
 168         }
 169 }
 170
 171 /* Record a memory area against a node. */
 172 void __init memory_present(int nid, unsigned long start, unsigned long end)
 173 {
 174         unsigned long pfn;
 175
 176         start &= PAGE_SECTION_MASK;
 177         mminit_validate_memmodel_limits(&start, &end);
 178         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 179                 unsigned long section = pfn_to_section_nr(pfn);
 180                 struct mem_section *ms;
 181
 182                 sparse_index_init(section, nid);
 183                 set_section_nid(section, nid);
 184
 185                 ms = __nr_to_section(section);
 186                 if (!ms->section_mem_map)
 187                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 188                                                         SECTION_MARKED_PRESENT;
 189         }
 190 }
 191
 192 /*
 193  * Only used by the i386 NUMA architecures, but relatively
 194  * generic code.
 195  */
 196 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 197                                                      unsigned long end_pfn)
 198 {
 199         unsigned long pfn;
 200         unsigned long nr_pages = 0;
 201
 202         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 203         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 204                 if (nid != early_pfn_to_nid(pfn))
 205                         continue;
 206
 207                 if (pfn_present(pfn))
 208                         nr_pages += PAGES_PER_SECTION;
 209         }
 210
 211         return nr_pages * sizeof(struct page);
 212 }
 213
 214 /*
 215  * Subtle, we encode the real pfn into the mem_map such that
 216  * the identity pfn - section_mem_map will return the actual
 217  * physical page frame number.
 218  */
 219 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 220 {
 221         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 222 }
 223
 224 /*
 225  * Decode mem_map from the coded memmap
 226  */
 227 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 228 {
 229         /* mask off the extra low bits of information */
 230         coded_mem_map &= SECTION_MAP_MASK;
 231         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 232 }
 233
 234 static int __meminit sparse_init_one_section(struct mem_section *ms,
 235                 unsigned long pnum, struct page *mem_map,
 236                 unsigned long *pageblock_bitmap)
 237 {
 238         if (!present_section(ms))
 239                 return -EINVAL;
 240
 241         ms->section_mem_map &= ~SECTION_MAP_MASK;
 242         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 243                                                         SECTION_HAS_MEM_MAP;
 244         ms->pageblock_flags = pageblock_bitmap;
 245
 246         return 1;
 247 }
 248
 249 unsigned long usemap_size(void)
 250 {
 251         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 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 remain 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 = memblock_virt_alloc_try_nid_nopanic(size,
 284                                                 SMP_CACHE_BYTES, goal, limit,
 285                                                 nid);
 286         if (!p && limit) {
 287                 limit = 0;
 288                 goto again;
 289         }
 290         return p;
 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                 pr_info("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         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 326                 usemap_snr, pgdat_snr, nid);
 327 }
 328 #else
 329 static unsigned long * __init
 330 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 331                                          unsigned long size)
 332 {
 333         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 334 }
 335
 336 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 337 {
 338 }
 339 #endif /* CONFIG_MEMORY_HOTREMOVE */
 340
 341 static void __init sparse_early_usemaps_alloc_node(void *data,
 342                                  unsigned long pnum_begin,
 343                                  unsigned long pnum_end,
 344                                  unsigned long usemap_count, int nodeid)
 345 {
 346         void *usemap;
 347         unsigned long pnum;
 348         unsigned long **usemap_map = (unsigned long **)data;
 349         int size = usemap_size();
 350
 351         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 352                                                           size * usemap_count);
 353         if (!usemap) {
 354                 pr_warn("%s: allocation failed\n", __func__);
 355                 return;
 356         }
 357
 358         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 359                 if (!present_section_nr(pnum))
 360                         continue;
 361                 usemap_map[pnum] = usemap;
 362                 usemap += size;
 363                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 364         }
 365 }
 366
 367 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 368 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 369 {
 370         struct page *map;
 371         unsigned long size;
 372
 373         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 374         if (map)
 375                 return map;
 376
 377         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 378         map = memblock_virt_alloc_try_nid(size,
 379                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 380                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 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 = memblock_virt_alloc_try_nid(size * map_count,
 405                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 406                                           BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 407         if (map) {
 408                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 409                         if (!present_section_nr(pnum))
 410                                 continue;
 411                         map_map[pnum] = map;
 412                         map += size;
 413                 }
 414                 return;
 415         }
 416
 417         /* fallback */
 418         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 419                 struct mem_section *ms;
 420
 421                 if (!present_section_nr(pnum))
 422                         continue;
 423                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 424                 if (map_map[pnum])
 425                         continue;
 426                 ms = __nr_to_section(pnum);
 427                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 428                        __func__);
 429                 ms->section_mem_map = 0;
 430         }
 431 }
 432 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 433
 434 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 435 static void __init sparse_early_mem_maps_alloc_node(void *data,
 436                                  unsigned long pnum_begin,
 437                                  unsigned long pnum_end,
 438                                  unsigned long map_count, int nodeid)
 439 {
 440         struct page **map_map = (struct page **)data;
 441         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 442                                          map_count, nodeid);
 443 }
 444 #else
 445 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 446 {
 447         struct page *map;
 448         struct mem_section *ms = __nr_to_section(pnum);
 449         int nid = sparse_early_nid(ms);
 450
 451         map = sparse_mem_map_populate(pnum, nid);
 452         if (map)
 453                 return map;
 454
 455         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 456                __func__);
 457         ms->section_mem_map = 0;
 458         return NULL;
 459 }
 460 #endif
 461
 462 void __weak __meminit vmemmap_populate_print_last(void)
 463 {
 464 }
 465
 466 /**
 467  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 468  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 469  */
 470 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 471                                         (void *, unsigned long, unsigned long,
 472                                         unsigned long, int), void *data)
 473 {
 474         unsigned long pnum;
 475         unsigned long map_count;
 476         int nodeid_begin = 0;
 477         unsigned long pnum_begin = 0;
 478
 479         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 480                 struct mem_section *ms;
 481
 482                 if (!present_section_nr(pnum))
 483                         continue;
 484                 ms = __nr_to_section(pnum);
 485                 nodeid_begin = sparse_early_nid(ms);
 486                 pnum_begin = pnum;
 487                 break;
 488         }
 489         map_count = 1;
 490         for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 491                 struct mem_section *ms;
 492                 int nodeid;
 493
 494                 if (!present_section_nr(pnum))
 495                         continue;
 496                 ms = __nr_to_section(pnum);
 497                 nodeid = sparse_early_nid(ms);
 498                 if (nodeid == nodeid_begin) {
 499                         map_count++;
 500                         continue;
 501                 }
 502                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 503                 alloc_func(data, pnum_begin, pnum,
 504                                                 map_count, nodeid_begin);
 505                 /* new start, update count etc*/
 506                 nodeid_begin = nodeid;
 507                 pnum_begin = pnum;
 508                 map_count = 1;
 509         }
 510         /* ok, last chunk */
 511         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
 512                                                 map_count, nodeid_begin);
 513 }
 514
 515 /*
 516  * Allocate the accumulated non-linear sections, allocate a mem_map
 517  * for each and record the physical to section mapping.
 518  */
 519 void __init sparse_init(void)
 520 {
 521         unsigned long pnum;
 522         struct page *map;
 523         unsigned long *usemap;
 524         unsigned long **usemap_map;
 525         int size;
 526 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 527         int size2;
 528         struct page **map_map;
 529 #endif
 530
 531         /* see include/linux/mmzone.h 'struct mem_section' definition */
 532         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 533
 534         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 535         set_pageblock_order();
 536
 537         /*
 538          * map is using big page (aka 2M in x86 64 bit)
 539          * usemap is less one page (aka 24 bytes)
 540          * so alloc 2M (with 2M align) and 24 bytes in turn will
 541          * make next 2M slip to one more 2M later.
 542          * then in big system, the memory will have a lot of holes...
 543          * here try to allocate 2M pages continuously.
 544          *
 545          * powerpc need to call sparse_init_one_section right after each
 546          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 547          */
 548         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 549         usemap_map = memblock_virt_alloc(size, 0);
 550         if (!usemap_map)
 551                 panic("can not allocate usemap_map\n");
 552         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 553                                                         (void *)usemap_map);
 554
 555 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 556         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 557         map_map = memblock_virt_alloc(size2, 0);
 558         if (!map_map)
 559                 panic("can not allocate map_map\n");
 560         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 561                                                         (void *)map_map);
 562 #endif
 563
 564         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 565                 if (!present_section_nr(pnum))
 566                         continue;
 567
 568                 usemap = usemap_map[pnum];
 569                 if (!usemap)
 570                         continue;
 571
 572 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 573                 map = map_map[pnum];
 574 #else
 575                 map = sparse_early_mem_map_alloc(pnum);
 576 #endif
 577                 if (!map)
 578                         continue;
 579
 580                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 581                                                                 usemap);
 582         }
 583
 584         vmemmap_populate_print_last();
 585
 586 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 587         memblock_free_early(__pa(map_map), size2);
 588 #endif
 589         memblock_free_early(__pa(usemap_map), size);
 590 }
 591
 592 #ifdef CONFIG_MEMORY_HOTPLUG
 593 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 594 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 595 {
 596         /* This will make the necessary allocations eventually. */
 597         return sparse_mem_map_populate(pnum, nid);
 598 }
 599 static void __kfree_section_memmap(struct page *memmap)
 600 {
 601         unsigned long start = (unsigned long)memmap;
 602         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 603
 604         vmemmap_free(start, end);
 605 }
 606 #ifdef CONFIG_MEMORY_HOTREMOVE
 607 static void free_map_bootmem(struct page *memmap)
 608 {
 609         unsigned long start = (unsigned long)memmap;
 610         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 611
 612         vmemmap_free(start, end);
 613 }
 614 #endif /* CONFIG_MEMORY_HOTREMOVE */
 615 #else
 616 static struct page *__kmalloc_section_memmap(void)
 617 {
 618         struct page *page, *ret;
 619         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 620
 621         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 622         if (page)
 623                 goto got_map_page;
 624
 625         ret = vmalloc(memmap_size);
 626         if (ret)
 627                 goto got_map_ptr;
 628
 629         return NULL;
 630 got_map_page:
 631         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 632 got_map_ptr:
 633
 634         return ret;
 635 }
 636
 637 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 638 {
 639         return __kmalloc_section_memmap();
 640 }
 641
 642 static void __kfree_section_memmap(struct page *memmap)
 643 {
 644         if (is_vmalloc_addr(memmap))
 645                 vfree(memmap);
 646         else
 647                 free_pages((unsigned long)memmap,
 648                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 649 }
 650
 651 #ifdef CONFIG_MEMORY_HOTREMOVE
 652 static void free_map_bootmem(struct page *memmap)
 653 {
 654         unsigned long maps_section_nr, removing_section_nr, i;
 655         unsigned long magic, nr_pages;
 656         struct page *page = virt_to_page(memmap);
 657
 658         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 659                 >> PAGE_SHIFT;
 660
 661         for (i = 0; i < nr_pages; i++, page++) {
 662                 magic = (unsigned long) page->freelist;
 663
 664                 BUG_ON(magic == NODE_INFO);
 665
 666                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 667                 removing_section_nr = page_private(page);
 668
 669                 /*
 670                  * When this function is called, the removing section is
 671                  * logical offlined state. This means all pages are isolated
 672                  * from page allocator. If removing section's memmap is placed
 673                  * on the same section, it must not be freed.
 674                  * If it is freed, page allocator may allocate it which will
 675                  * be removed physically soon.
 676                  */
 677                 if (maps_section_nr != removing_section_nr)
 678                         put_page_bootmem(page);
 679         }
 680 }
 681 #endif /* CONFIG_MEMORY_HOTREMOVE */
 682 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 683
 684 /*
 685  * returns the number of sections whose mem_maps were properly
 686  * set.  If this is <=0, then that means that the passed-in
 687  * map was not consumed and must be freed.
 688  */
 689 int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn)
 690 {
 691         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 692         struct pglist_data *pgdat = zone->zone_pgdat;
 693         struct mem_section *ms;
 694         struct page *memmap;
 695         unsigned long *usemap;
 696         unsigned long flags;
 697         int ret;
 698
 699         /*
 700          * no locking for this, because it does its own
 701          * plus, it does a kmalloc
 702          */
 703         ret = sparse_index_init(section_nr, pgdat->node_id);
 704         if (ret < 0 && ret != -EEXIST)
 705                 return ret;
 706         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
 707         if (!memmap)
 708                 return -ENOMEM;
 709         usemap = __kmalloc_section_usemap();
 710         if (!usemap) {
 711                 __kfree_section_memmap(memmap);
 712                 return -ENOMEM;
 713         }
 714
 715         pgdat_resize_lock(pgdat, &flags);
 716
 717         ms = __pfn_to_section(start_pfn);
 718         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 719                 ret = -EEXIST;
 720                 goto out;
 721         }
 722
 723         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 724
 725         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 726
 727         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 728
 729 out:
 730         pgdat_resize_unlock(pgdat, &flags);
 731         if (ret <= 0) {
 732                 kfree(usemap);
 733                 __kfree_section_memmap(memmap);
 734         }
 735         return ret;
 736 }
 737
 738 #ifdef CONFIG_MEMORY_HOTREMOVE
 739 #ifdef CONFIG_MEMORY_FAILURE
 740 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 741 {
 742         int i;
 743
 744         if (!memmap)
 745                 return;
 746
 747         for (i = 0; i < nr_pages; i++) {
 748                 if (PageHWPoison(&memmap[i])) {
 749                         atomic_long_sub(1, &num_poisoned_pages);
 750                         ClearPageHWPoison(&memmap[i]);
 751                 }
 752         }
 753 }
 754 #else
 755 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 756 {
 757 }
 758 #endif
 759
 760 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 761 {
 762         struct page *usemap_page;
 763
 764         if (!usemap)
 765                 return;
 766
 767         usemap_page = virt_to_page(usemap);
 768         /*
 769          * Check to see if allocation came from hot-plug-add
 770          */
 771         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 772                 kfree(usemap);
 773                 if (memmap)
 774                         __kfree_section_memmap(memmap);
 775                 return;
 776         }
 777
 778         /*
 779          * The usemap came from bootmem. This is packed with other usemaps
 780          * on the section which has pgdat at boot time. Just keep it as is now.
 781          */
 782
 783         if (memmap)
 784                 free_map_bootmem(memmap);
 785 }
 786
 787 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 788                 unsigned long map_offset)
 789 {
 790         struct page *memmap = NULL;
 791         unsigned long *usemap = NULL, flags;
 792         struct pglist_data *pgdat = zone->zone_pgdat;
 793
 794         pgdat_resize_lock(pgdat, &flags);
 795         if (ms->section_mem_map) {
 796                 usemap = ms->pageblock_flags;
 797                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 798                                                 __section_nr(ms));
 799                 ms->section_mem_map = 0;
 800                 ms->pageblock_flags = NULL;
 801         }
 802         pgdat_resize_unlock(pgdat, &flags);
 803
 804         clear_hwpoisoned_pages(memmap + map_offset,
 805                         PAGES_PER_SECTION - map_offset);
 806         free_section_usemap(memmap, usemap);
 807 }
 808 #endif /* CONFIG_MEMORY_HOTREMOVE */
 809 #endif /* CONFIG_MEMORY_HOTPLUG */