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 /*
 172  * There are a number of times that we loop over NR_MEM_SECTIONS,
 173  * looking for section_present() on each.  But, when we have very
 174  * large physical address spaces, NR_MEM_SECTIONS can also be
 175  * very large which makes the loops quite long.
 176  *
 177  * Keeping track of this gives us an easy way to break out of
 178  * those loops early.
 179  */
 180 int __highest_present_section_nr;
 181 static void section_mark_present(struct mem_section *ms)
 182 {
 183         int section_nr = __section_nr(ms);
 184
 185         if (section_nr > __highest_present_section_nr)
 186                 __highest_present_section_nr = section_nr;
 187
 188         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 189 }
 190
 191 static inline int next_present_section_nr(int section_nr)
 192 {
 193         do {
 194                 section_nr++;
 195                 if (present_section_nr(section_nr))
 196                         return section_nr;
 197         } while ((section_nr < NR_MEM_SECTIONS) &&
 198                  (section_nr <= __highest_present_section_nr));
 199
 200         return -1;
 201 }
 202 #define for_each_present_section_nr(start, section_nr)          \
 203         for (section_nr = next_present_section_nr(start-1);     \
 204              ((section_nr >= 0) &&                              \
 205               (section_nr < NR_MEM_SECTIONS) &&                 \
 206               (section_nr <= __highest_present_section_nr));    \
 207              section_nr = next_present_section_nr(section_nr))
 208
 209 /* Record a memory area against a node. */
 210 void __init memory_present(int nid, unsigned long start, unsigned long end)
 211 {
 212         unsigned long pfn;
 213
 214         start &= PAGE_SECTION_MASK;
 215         mminit_validate_memmodel_limits(&start, &end);
 216         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 217                 unsigned long section = pfn_to_section_nr(pfn);
 218                 struct mem_section *ms;
 219
 220                 sparse_index_init(section, nid);
 221                 set_section_nid(section, nid);
 222
 223                 ms = __nr_to_section(section);
 224                 if (!ms->section_mem_map) {
 225                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 226                                                         SECTION_IS_ONLINE;
 227                         section_mark_present(ms);
 228                 }
 229         }
 230 }
 231
 232 /*
 233  * Only used by the i386 NUMA architecures, but relatively
 234  * generic code.
 235  */
 236 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 237                                                      unsigned long end_pfn)
 238 {
 239         unsigned long pfn;
 240         unsigned long nr_pages = 0;
 241
 242         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 243         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 244                 if (nid != early_pfn_to_nid(pfn))
 245                         continue;
 246
 247                 if (pfn_present(pfn))
 248                         nr_pages += PAGES_PER_SECTION;
 249         }
 250
 251         return nr_pages * sizeof(struct page);
 252 }
 253
 254 /*
 255  * Subtle, we encode the real pfn into the mem_map such that
 256  * the identity pfn - section_mem_map will return the actual
 257  * physical page frame number.
 258  */
 259 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 260 {
 261         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 262 }
 263
 264 /*
 265  * Decode mem_map from the coded memmap
 266  */
 267 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 268 {
 269         /* mask off the extra low bits of information */
 270         coded_mem_map &= SECTION_MAP_MASK;
 271         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 272 }
 273
 274 static int __meminit sparse_init_one_section(struct mem_section *ms,
 275                 unsigned long pnum, struct page *mem_map,
 276                 unsigned long *pageblock_bitmap)
 277 {
 278         if (!present_section(ms))
 279                 return -EINVAL;
 280
 281         ms->section_mem_map &= ~SECTION_MAP_MASK;
 282         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 283                                                         SECTION_HAS_MEM_MAP;
 284         ms->pageblock_flags = pageblock_bitmap;
 285
 286         return 1;
 287 }
 288
 289 unsigned long usemap_size(void)
 290 {
 291         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 292 }
 293
 294 #ifdef CONFIG_MEMORY_HOTPLUG
 295 static unsigned long *__kmalloc_section_usemap(void)
 296 {
 297         return kmalloc(usemap_size(), GFP_KERNEL);
 298 }
 299 #endif /* CONFIG_MEMORY_HOTPLUG */
 300
 301 #ifdef CONFIG_MEMORY_HOTREMOVE
 302 static unsigned long * __init
 303 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 304                                          unsigned long size)
 305 {
 306         unsigned long goal, limit;
 307         unsigned long *p;
 308         int nid;
 309         /*
 310          * A page may contain usemaps for other sections preventing the
 311          * page being freed and making a section unremovable while
 312          * other sections referencing the usemap remain active. Similarly,
 313          * a pgdat can prevent a section being removed. If section A
 314          * contains a pgdat and section B contains the usemap, both
 315          * sections become inter-dependent. This allocates usemaps
 316          * from the same section as the pgdat where possible to avoid
 317          * this problem.
 318          */
 319         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 320         limit = goal + (1UL << PA_SECTION_SHIFT);
 321         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 322 again:
 323         p = memblock_virt_alloc_try_nid_nopanic(size,
 324                                                 SMP_CACHE_BYTES, goal, limit,
 325                                                 nid);
 326         if (!p && limit) {
 327                 limit = 0;
 328                 goto again;
 329         }
 330         return p;
 331 }
 332
 333 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 334 {
 335         unsigned long usemap_snr, pgdat_snr;
 336         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 337         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 338         struct pglist_data *pgdat = NODE_DATA(nid);
 339         int usemap_nid;
 340
 341         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 342         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 343         if (usemap_snr == pgdat_snr)
 344                 return;
 345
 346         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 347                 /* skip redundant message */
 348                 return;
 349
 350         old_usemap_snr = usemap_snr;
 351         old_pgdat_snr = pgdat_snr;
 352
 353         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 354         if (usemap_nid != nid) {
 355                 pr_info("node %d must be removed before remove section %ld\n",
 356                         nid, usemap_snr);
 357                 return;
 358         }
 359         /*
 360          * There is a circular dependency.
 361          * Some platforms allow un-removable section because they will just
 362          * gather other removable sections for dynamic partitioning.
 363          * Just notify un-removable section's number here.
 364          */
 365         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 366                 usemap_snr, pgdat_snr, nid);
 367 }
 368 #else
 369 static unsigned long * __init
 370 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 371                                          unsigned long size)
 372 {
 373         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 374 }
 375
 376 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 377 {
 378 }
 379 #endif /* CONFIG_MEMORY_HOTREMOVE */
 380
 381 static void __init sparse_early_usemaps_alloc_node(void *data,
 382                                  unsigned long pnum_begin,
 383                                  unsigned long pnum_end,
 384                                  unsigned long usemap_count, int nodeid)
 385 {
 386         void *usemap;
 387         unsigned long pnum;
 388         unsigned long **usemap_map = (unsigned long **)data;
 389         int size = usemap_size();
 390
 391         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 392                                                           size * usemap_count);
 393         if (!usemap) {
 394                 pr_warn("%s: allocation failed\n", __func__);
 395                 return;
 396         }
 397
 398         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 399                 if (!present_section_nr(pnum))
 400                         continue;
 401                 usemap_map[pnum] = usemap;
 402                 usemap += size;
 403                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 404         }
 405 }
 406
 407 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 408 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 409 {
 410         struct page *map;
 411         unsigned long size;
 412
 413         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 414         if (map)
 415                 return map;
 416
 417         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 418         map = memblock_virt_alloc_try_nid(size,
 419                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 420                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 421         return map;
 422 }
 423 void __init sparse_mem_maps_populate_node(struct page **map_map,
 424                                           unsigned long pnum_begin,
 425                                           unsigned long pnum_end,
 426                                           unsigned long map_count, int nodeid)
 427 {
 428         void *map;
 429         unsigned long pnum;
 430         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 431
 432         map = alloc_remap(nodeid, size * map_count);
 433         if (map) {
 434                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 435                         if (!present_section_nr(pnum))
 436                                 continue;
 437                         map_map[pnum] = map;
 438                         map += size;
 439                 }
 440                 return;
 441         }
 442
 443         size = PAGE_ALIGN(size);
 444         map = memblock_virt_alloc_try_nid(size * map_count,
 445                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 446                                           BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 447         if (map) {
 448                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 449                         if (!present_section_nr(pnum))
 450                                 continue;
 451                         map_map[pnum] = map;
 452                         map += size;
 453                 }
 454                 return;
 455         }
 456
 457         /* fallback */
 458         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 459                 struct mem_section *ms;
 460
 461                 if (!present_section_nr(pnum))
 462                         continue;
 463                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 464                 if (map_map[pnum])
 465                         continue;
 466                 ms = __nr_to_section(pnum);
 467                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 468                        __func__);
 469                 ms->section_mem_map = 0;
 470         }
 471 }
 472 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 473
 474 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 475 static void __init sparse_early_mem_maps_alloc_node(void *data,
 476                                  unsigned long pnum_begin,
 477                                  unsigned long pnum_end,
 478                                  unsigned long map_count, int nodeid)
 479 {
 480         struct page **map_map = (struct page **)data;
 481         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 482                                          map_count, nodeid);
 483 }
 484 #else
 485 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 486 {
 487         struct page *map;
 488         struct mem_section *ms = __nr_to_section(pnum);
 489         int nid = sparse_early_nid(ms);
 490
 491         map = sparse_mem_map_populate(pnum, nid);
 492         if (map)
 493                 return map;
 494
 495         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 496                __func__);
 497         ms->section_mem_map = 0;
 498         return NULL;
 499 }
 500 #endif
 501
 502 void __weak __meminit vmemmap_populate_print_last(void)
 503 {
 504 }
 505
 506 /**
 507  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 508  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 509  */
 510 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 511                                         (void *, unsigned long, unsigned long,
 512                                         unsigned long, int), void *data)
 513 {
 514         unsigned long pnum;
 515         unsigned long map_count;
 516         int nodeid_begin = 0;
 517         unsigned long pnum_begin = 0;
 518
 519         for_each_present_section_nr(0, pnum) {
 520                 struct mem_section *ms;
 521
 522                 ms = __nr_to_section(pnum);
 523                 nodeid_begin = sparse_early_nid(ms);
 524                 pnum_begin = pnum;
 525                 break;
 526         }
 527         map_count = 1;
 528         for_each_present_section_nr(pnum_begin + 1, pnum) {
 529                 struct mem_section *ms;
 530                 int nodeid;
 531
 532                 ms = __nr_to_section(pnum);
 533                 nodeid = sparse_early_nid(ms);
 534                 if (nodeid == nodeid_begin) {
 535                         map_count++;
 536                         continue;
 537                 }
 538                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 539                 alloc_func(data, pnum_begin, pnum,
 540                                                 map_count, nodeid_begin);
 541                 /* new start, update count etc*/
 542                 nodeid_begin = nodeid;
 543                 pnum_begin = pnum;
 544                 map_count = 1;
 545         }
 546         /* ok, last chunk */
 547         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
 548                                                 map_count, nodeid_begin);
 549 }
 550
 551 /*
 552  * Allocate the accumulated non-linear sections, allocate a mem_map
 553  * for each and record the physical to section mapping.
 554  */
 555 void __init sparse_init(void)
 556 {
 557         unsigned long pnum;
 558         struct page *map;
 559         unsigned long *usemap;
 560         unsigned long **usemap_map;
 561         int size;
 562 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 563         int size2;
 564         struct page **map_map;
 565 #endif
 566
 567         /* see include/linux/mmzone.h 'struct mem_section' definition */
 568         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 569
 570         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 571         set_pageblock_order();
 572
 573         /*
 574          * map is using big page (aka 2M in x86 64 bit)
 575          * usemap is less one page (aka 24 bytes)
 576          * so alloc 2M (with 2M align) and 24 bytes in turn will
 577          * make next 2M slip to one more 2M later.
 578          * then in big system, the memory will have a lot of holes...
 579          * here try to allocate 2M pages continuously.
 580          *
 581          * powerpc need to call sparse_init_one_section right after each
 582          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 583          */
 584         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 585         usemap_map = memblock_virt_alloc(size, 0);
 586         if (!usemap_map)
 587                 panic("can not allocate usemap_map\n");
 588         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 589                                                         (void *)usemap_map);
 590
 591 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 592         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 593         map_map = memblock_virt_alloc(size2, 0);
 594         if (!map_map)
 595                 panic("can not allocate map_map\n");
 596         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 597                                                         (void *)map_map);
 598 #endif
 599
 600         for_each_present_section_nr(0, pnum) {
 601                 usemap = usemap_map[pnum];
 602                 if (!usemap)
 603                         continue;
 604
 605 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 606                 map = map_map[pnum];
 607 #else
 608                 map = sparse_early_mem_map_alloc(pnum);
 609 #endif
 610                 if (!map)
 611                         continue;
 612
 613                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 614                                                                 usemap);
 615         }
 616
 617         vmemmap_populate_print_last();
 618
 619 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 620         memblock_free_early(__pa(map_map), size2);
 621 #endif
 622         memblock_free_early(__pa(usemap_map), size);
 623 }
 624
 625 #ifdef CONFIG_MEMORY_HOTPLUG
 626
 627 /* Mark all memory sections within the pfn range as online */
 628 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 629 {
 630         unsigned long pfn;
 631
 632         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 633                 unsigned long section_nr = pfn_to_section_nr(pfn);
 634                 struct mem_section *ms;
 635
 636                 /* onlining code should never touch invalid ranges */
 637                 if (WARN_ON(!valid_section_nr(section_nr)))
 638                         continue;
 639
 640                 ms = __nr_to_section(section_nr);
 641                 ms->section_mem_map |= SECTION_IS_ONLINE;
 642         }
 643 }
 644
 645 #ifdef CONFIG_MEMORY_HOTREMOVE
 646 /* Mark all memory sections within the pfn range as online */
 647 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 648 {
 649         unsigned long pfn;
 650
 651         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 652                 unsigned long section_nr = pfn_to_section_nr(start_pfn);
 653                 struct mem_section *ms;
 654
 655                 /*
 656                  * TODO this needs some double checking. Offlining code makes
 657                  * sure to check pfn_valid but those checks might be just bogus
 658                  */
 659                 if (WARN_ON(!valid_section_nr(section_nr)))
 660                         continue;
 661
 662                 ms = __nr_to_section(section_nr);
 663                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 664         }
 665 }
 666 #endif
 667
 668 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 669 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 670 {
 671         /* This will make the necessary allocations eventually. */
 672         return sparse_mem_map_populate(pnum, nid);
 673 }
 674 static void __kfree_section_memmap(struct page *memmap)
 675 {
 676         unsigned long start = (unsigned long)memmap;
 677         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 678
 679         vmemmap_free(start, end);
 680 }
 681 #ifdef CONFIG_MEMORY_HOTREMOVE
 682 static void free_map_bootmem(struct page *memmap)
 683 {
 684         unsigned long start = (unsigned long)memmap;
 685         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 686
 687         vmemmap_free(start, end);
 688 }
 689 #endif /* CONFIG_MEMORY_HOTREMOVE */
 690 #else
 691 static struct page *__kmalloc_section_memmap(void)
 692 {
 693         struct page *page, *ret;
 694         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 695
 696         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 697         if (page)
 698                 goto got_map_page;
 699
 700         ret = vmalloc(memmap_size);
 701         if (ret)
 702                 goto got_map_ptr;
 703
 704         return NULL;
 705 got_map_page:
 706         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 707 got_map_ptr:
 708
 709         return ret;
 710 }
 711
 712 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 713 {
 714         return __kmalloc_section_memmap();
 715 }
 716
 717 static void __kfree_section_memmap(struct page *memmap)
 718 {
 719         if (is_vmalloc_addr(memmap))
 720                 vfree(memmap);
 721         else
 722                 free_pages((unsigned long)memmap,
 723                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 724 }
 725
 726 #ifdef CONFIG_MEMORY_HOTREMOVE
 727 static void free_map_bootmem(struct page *memmap)
 728 {
 729         unsigned long maps_section_nr, removing_section_nr, i;
 730         unsigned long magic, nr_pages;
 731         struct page *page = virt_to_page(memmap);
 732
 733         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 734                 >> PAGE_SHIFT;
 735
 736         for (i = 0; i < nr_pages; i++, page++) {
 737                 magic = (unsigned long) page->freelist;
 738
 739                 BUG_ON(magic == NODE_INFO);
 740
 741                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 742                 removing_section_nr = page_private(page);
 743
 744                 /*
 745                  * When this function is called, the removing section is
 746                  * logical offlined state. This means all pages are isolated
 747                  * from page allocator. If removing section's memmap is placed
 748                  * on the same section, it must not be freed.
 749                  * If it is freed, page allocator may allocate it which will
 750                  * be removed physically soon.
 751                  */
 752                 if (maps_section_nr != removing_section_nr)
 753                         put_page_bootmem(page);
 754         }
 755 }
 756 #endif /* CONFIG_MEMORY_HOTREMOVE */
 757 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 758
 759 /*
 760  * returns the number of sections whose mem_maps were properly
 761  * set.  If this is <=0, then that means that the passed-in
 762  * map was not consumed and must be freed.
 763  */
 764 int __meminit sparse_add_one_section(struct pglist_data *pgdat, unsigned long start_pfn)
 765 {
 766         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 767         struct mem_section *ms;
 768         struct page *memmap;
 769         unsigned long *usemap;
 770         unsigned long flags;
 771         int ret;
 772
 773         /*
 774          * no locking for this, because it does its own
 775          * plus, it does a kmalloc
 776          */
 777         ret = sparse_index_init(section_nr, pgdat->node_id);
 778         if (ret < 0 && ret != -EEXIST)
 779                 return ret;
 780         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
 781         if (!memmap)
 782                 return -ENOMEM;
 783         usemap = __kmalloc_section_usemap();
 784         if (!usemap) {
 785                 __kfree_section_memmap(memmap);
 786                 return -ENOMEM;
 787         }
 788
 789         pgdat_resize_lock(pgdat, &flags);
 790
 791         ms = __pfn_to_section(start_pfn);
 792         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 793                 ret = -EEXIST;
 794                 goto out;
 795         }
 796
 797         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 798
 799         section_mark_present(ms);
 800
 801         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 802
 803 out:
 804         pgdat_resize_unlock(pgdat, &flags);
 805         if (ret <= 0) {
 806                 kfree(usemap);
 807                 __kfree_section_memmap(memmap);
 808         }
 809         return ret;
 810 }
 811
 812 #ifdef CONFIG_MEMORY_HOTREMOVE
 813 #ifdef CONFIG_MEMORY_FAILURE
 814 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 815 {
 816         int i;
 817
 818         if (!memmap)
 819                 return;
 820
 821         for (i = 0; i < nr_pages; i++) {
 822                 if (PageHWPoison(&memmap[i])) {
 823                         atomic_long_sub(1, &num_poisoned_pages);
 824                         ClearPageHWPoison(&memmap[i]);
 825                 }
 826         }
 827 }
 828 #else
 829 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 830 {
 831 }
 832 #endif
 833
 834 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 835 {
 836         struct page *usemap_page;
 837
 838         if (!usemap)
 839                 return;
 840
 841         usemap_page = virt_to_page(usemap);
 842         /*
 843          * Check to see if allocation came from hot-plug-add
 844          */
 845         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 846                 kfree(usemap);
 847                 if (memmap)
 848                         __kfree_section_memmap(memmap);
 849                 return;
 850         }
 851
 852         /*
 853          * The usemap came from bootmem. This is packed with other usemaps
 854          * on the section which has pgdat at boot time. Just keep it as is now.
 855          */
 856
 857         if (memmap)
 858                 free_map_bootmem(memmap);
 859 }
 860
 861 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 862                 unsigned long map_offset)
 863 {
 864         struct page *memmap = NULL;
 865         unsigned long *usemap = NULL, flags;
 866         struct pglist_data *pgdat = zone->zone_pgdat;
 867
 868         pgdat_resize_lock(pgdat, &flags);
 869         if (ms->section_mem_map) {
 870                 usemap = ms->pageblock_flags;
 871                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 872                                                 __section_nr(ms));
 873                 ms->section_mem_map = 0;
 874                 ms->pageblock_flags = NULL;
 875         }
 876         pgdat_resize_unlock(pgdat, &flags);
 877
 878         clear_hwpoisoned_pages(memmap + map_offset,
 879                         PAGES_PER_SECTION - map_offset);
 880         free_section_usemap(memmap, usemap);
 881 }
 882 #endif /* CONFIG_MEMORY_HOTREMOVE */
 883 #endif /* CONFIG_MEMORY_HOTPLUG */