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                 section = kzalloc_node(array_size, GFP_KERNEL, nid);
  70         else
  71                 section = memblock_virt_alloc_node(array_size, nid);
  72
  73         return section;
  74 }
  75
  76 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  77 {
  78         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  79         struct mem_section *section;
  80
  81         if (mem_section[root])
  82                 return -EEXIST;
  83
  84         section = sparse_index_alloc(nid);
  85         if (!section)
  86                 return -ENOMEM;
  87
  88         mem_section[root] = section;
  89
  90         return 0;
  91 }
  92 #else /* !SPARSEMEM_EXTREME */
  93 static inline int sparse_index_init(unsigned long section_nr, int nid)
  94 {
  95         return 0;
  96 }
  97 #endif
  98
  99 #ifdef CONFIG_SPARSEMEM_EXTREME
 100 int __section_nr(struct mem_section* ms)
 101 {
 102         unsigned long root_nr;
 103         struct mem_section* root;
 104
 105         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 106                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 107                 if (!root)
 108                         continue;
 109
 110                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 111                      break;
 112         }
 113
 114         VM_BUG_ON(root_nr == NR_SECTION_ROOTS);
 115
 116         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 117 }
 118 #else
 119 int __section_nr(struct mem_section* ms)
 120 {
 121         return (int)(ms - mem_section[0]);
 122 }
 123 #endif
 124
 125 /*
 126  * During early boot, before section_mem_map is used for an actual
 127  * mem_map, we use section_mem_map to store the section's NUMA
 128  * node.  This keeps us from having to use another data structure.  The
 129  * node information is cleared just before we store the real mem_map.
 130  */
 131 static inline unsigned long sparse_encode_early_nid(int nid)
 132 {
 133         return (nid << SECTION_NID_SHIFT);
 134 }
 135
 136 static inline int sparse_early_nid(struct mem_section *section)
 137 {
 138         return (section->section_mem_map >> SECTION_NID_SHIFT);
 139 }
 140
 141 /* Validate the physical addressing limitations of the model */
 142 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 143                                                 unsigned long *end_pfn)
 144 {
 145         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 146
 147         /*
 148          * Sanity checks - do not allow an architecture to pass
 149          * in larger pfns than the maximum scope of sparsemem:
 150          */
 151         if (*start_pfn > max_sparsemem_pfn) {
 152                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 153                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 154                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 155                 WARN_ON_ONCE(1);
 156                 *start_pfn = max_sparsemem_pfn;
 157                 *end_pfn = max_sparsemem_pfn;
 158         } else if (*end_pfn > max_sparsemem_pfn) {
 159                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 160                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 161                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 162                 WARN_ON_ONCE(1);
 163                 *end_pfn = max_sparsemem_pfn;
 164         }
 165 }
 166
 167 /*
 168  * There are a number of times that we loop over NR_MEM_SECTIONS,
 169  * looking for section_present() on each.  But, when we have very
 170  * large physical address spaces, NR_MEM_SECTIONS can also be
 171  * very large which makes the loops quite long.
 172  *
 173  * Keeping track of this gives us an easy way to break out of
 174  * those loops early.
 175  */
 176 int __highest_present_section_nr;
 177 static void section_mark_present(struct mem_section *ms)
 178 {
 179         int section_nr = __section_nr(ms);
 180
 181         if (section_nr > __highest_present_section_nr)
 182                 __highest_present_section_nr = section_nr;
 183
 184         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 185 }
 186
 187 static inline int next_present_section_nr(int section_nr)
 188 {
 189         do {
 190                 section_nr++;
 191                 if (present_section_nr(section_nr))
 192                         return section_nr;
 193         } while ((section_nr < NR_MEM_SECTIONS) &&
 194                  (section_nr <= __highest_present_section_nr));
 195
 196         return -1;
 197 }
 198 #define for_each_present_section_nr(start, section_nr)          \
 199         for (section_nr = next_present_section_nr(start-1);     \
 200              ((section_nr >= 0) &&                              \
 201               (section_nr < NR_MEM_SECTIONS) &&                 \
 202               (section_nr <= __highest_present_section_nr));    \
 203              section_nr = next_present_section_nr(section_nr))
 204
 205 /* Record a memory area against a node. */
 206 void __init memory_present(int nid, unsigned long start, unsigned long end)
 207 {
 208         unsigned long pfn;
 209
 210         start &= PAGE_SECTION_MASK;
 211         mminit_validate_memmodel_limits(&start, &end);
 212         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 213                 unsigned long section = pfn_to_section_nr(pfn);
 214                 struct mem_section *ms;
 215
 216                 sparse_index_init(section, nid);
 217                 set_section_nid(section, nid);
 218
 219                 ms = __nr_to_section(section);
 220                 if (!ms->section_mem_map) {
 221                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 222                                                         SECTION_IS_ONLINE;
 223                         section_mark_present(ms);
 224                 }
 225         }
 226 }
 227
 228 /*
 229  * Only used by the i386 NUMA architecures, but relatively
 230  * generic code.
 231  */
 232 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 233                                                      unsigned long end_pfn)
 234 {
 235         unsigned long pfn;
 236         unsigned long nr_pages = 0;
 237
 238         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 239         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 240                 if (nid != early_pfn_to_nid(pfn))
 241                         continue;
 242
 243                 if (pfn_present(pfn))
 244                         nr_pages += PAGES_PER_SECTION;
 245         }
 246
 247         return nr_pages * sizeof(struct page);
 248 }
 249
 250 /*
 251  * Subtle, we encode the real pfn into the mem_map such that
 252  * the identity pfn - section_mem_map will return the actual
 253  * physical page frame number.
 254  */
 255 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 256 {
 257         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 258 }
 259
 260 /*
 261  * Decode mem_map from the coded memmap
 262  */
 263 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 264 {
 265         /* mask off the extra low bits of information */
 266         coded_mem_map &= SECTION_MAP_MASK;
 267         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 268 }
 269
 270 static int __meminit sparse_init_one_section(struct mem_section *ms,
 271                 unsigned long pnum, struct page *mem_map,
 272                 unsigned long *pageblock_bitmap)
 273 {
 274         if (!present_section(ms))
 275                 return -EINVAL;
 276
 277         ms->section_mem_map &= ~SECTION_MAP_MASK;
 278         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 279                                                         SECTION_HAS_MEM_MAP;
 280         ms->pageblock_flags = pageblock_bitmap;
 281
 282         return 1;
 283 }
 284
 285 unsigned long usemap_size(void)
 286 {
 287         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 288 }
 289
 290 #ifdef CONFIG_MEMORY_HOTPLUG
 291 static unsigned long *__kmalloc_section_usemap(void)
 292 {
 293         return kmalloc(usemap_size(), GFP_KERNEL);
 294 }
 295 #endif /* CONFIG_MEMORY_HOTPLUG */
 296
 297 #ifdef CONFIG_MEMORY_HOTREMOVE
 298 static unsigned long * __init
 299 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 300                                          unsigned long size)
 301 {
 302         unsigned long goal, limit;
 303         unsigned long *p;
 304         int nid;
 305         /*
 306          * A page may contain usemaps for other sections preventing the
 307          * page being freed and making a section unremovable while
 308          * other sections referencing the usemap remain active. Similarly,
 309          * a pgdat can prevent a section being removed. If section A
 310          * contains a pgdat and section B contains the usemap, both
 311          * sections become inter-dependent. This allocates usemaps
 312          * from the same section as the pgdat where possible to avoid
 313          * this problem.
 314          */
 315         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 316         limit = goal + (1UL << PA_SECTION_SHIFT);
 317         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 318 again:
 319         p = memblock_virt_alloc_try_nid_nopanic(size,
 320                                                 SMP_CACHE_BYTES, goal, limit,
 321                                                 nid);
 322         if (!p && limit) {
 323                 limit = 0;
 324                 goto again;
 325         }
 326         return p;
 327 }
 328
 329 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 330 {
 331         unsigned long usemap_snr, pgdat_snr;
 332         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 333         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 334         struct pglist_data *pgdat = NODE_DATA(nid);
 335         int usemap_nid;
 336
 337         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 338         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 339         if (usemap_snr == pgdat_snr)
 340                 return;
 341
 342         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 343                 /* skip redundant message */
 344                 return;
 345
 346         old_usemap_snr = usemap_snr;
 347         old_pgdat_snr = pgdat_snr;
 348
 349         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 350         if (usemap_nid != nid) {
 351                 pr_info("node %d must be removed before remove section %ld\n",
 352                         nid, usemap_snr);
 353                 return;
 354         }
 355         /*
 356          * There is a circular dependency.
 357          * Some platforms allow un-removable section because they will just
 358          * gather other removable sections for dynamic partitioning.
 359          * Just notify un-removable section's number here.
 360          */
 361         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 362                 usemap_snr, pgdat_snr, nid);
 363 }
 364 #else
 365 static unsigned long * __init
 366 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 367                                          unsigned long size)
 368 {
 369         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 370 }
 371
 372 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 373 {
 374 }
 375 #endif /* CONFIG_MEMORY_HOTREMOVE */
 376
 377 static void __init sparse_early_usemaps_alloc_node(void *data,
 378                                  unsigned long pnum_begin,
 379                                  unsigned long pnum_end,
 380                                  unsigned long usemap_count, int nodeid)
 381 {
 382         void *usemap;
 383         unsigned long pnum;
 384         unsigned long **usemap_map = (unsigned long **)data;
 385         int size = usemap_size();
 386
 387         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 388                                                           size * usemap_count);
 389         if (!usemap) {
 390                 pr_warn("%s: allocation failed\n", __func__);
 391                 return;
 392         }
 393
 394         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 395                 if (!present_section_nr(pnum))
 396                         continue;
 397                 usemap_map[pnum] = usemap;
 398                 usemap += size;
 399                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 400         }
 401 }
 402
 403 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 404 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 405 {
 406         struct page *map;
 407         unsigned long size;
 408
 409         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 410         if (map)
 411                 return map;
 412
 413         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 414         map = memblock_virt_alloc_try_nid(size,
 415                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 416                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 417         return map;
 418 }
 419 void __init sparse_mem_maps_populate_node(struct page **map_map,
 420                                           unsigned long pnum_begin,
 421                                           unsigned long pnum_end,
 422                                           unsigned long map_count, int nodeid)
 423 {
 424         void *map;
 425         unsigned long pnum;
 426         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 427
 428         map = alloc_remap(nodeid, size * map_count);
 429         if (map) {
 430                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 431                         if (!present_section_nr(pnum))
 432                                 continue;
 433                         map_map[pnum] = map;
 434                         map += size;
 435                 }
 436                 return;
 437         }
 438
 439         size = PAGE_ALIGN(size);
 440         map = memblock_virt_alloc_try_nid(size * map_count,
 441                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 442                                           BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 443         if (map) {
 444                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 445                         if (!present_section_nr(pnum))
 446                                 continue;
 447                         map_map[pnum] = map;
 448                         map += size;
 449                 }
 450                 return;
 451         }
 452
 453         /* fallback */
 454         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 455                 struct mem_section *ms;
 456
 457                 if (!present_section_nr(pnum))
 458                         continue;
 459                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 460                 if (map_map[pnum])
 461                         continue;
 462                 ms = __nr_to_section(pnum);
 463                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 464                        __func__);
 465                 ms->section_mem_map = 0;
 466         }
 467 }
 468 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 469
 470 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 471 static void __init sparse_early_mem_maps_alloc_node(void *data,
 472                                  unsigned long pnum_begin,
 473                                  unsigned long pnum_end,
 474                                  unsigned long map_count, int nodeid)
 475 {
 476         struct page **map_map = (struct page **)data;
 477         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 478                                          map_count, nodeid);
 479 }
 480 #else
 481 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 482 {
 483         struct page *map;
 484         struct mem_section *ms = __nr_to_section(pnum);
 485         int nid = sparse_early_nid(ms);
 486
 487         map = sparse_mem_map_populate(pnum, nid);
 488         if (map)
 489                 return map;
 490
 491         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 492                __func__);
 493         ms->section_mem_map = 0;
 494         return NULL;
 495 }
 496 #endif
 497
 498 void __weak __meminit vmemmap_populate_print_last(void)
 499 {
 500 }
 501
 502 /**
 503  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 504  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 505  */
 506 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 507                                         (void *, unsigned long, unsigned long,
 508                                         unsigned long, int), void *data)
 509 {
 510         unsigned long pnum;
 511         unsigned long map_count;
 512         int nodeid_begin = 0;
 513         unsigned long pnum_begin = 0;
 514
 515         for_each_present_section_nr(0, pnum) {
 516                 struct mem_section *ms;
 517
 518                 ms = __nr_to_section(pnum);
 519                 nodeid_begin = sparse_early_nid(ms);
 520                 pnum_begin = pnum;
 521                 break;
 522         }
 523         map_count = 1;
 524         for_each_present_section_nr(pnum_begin + 1, pnum) {
 525                 struct mem_section *ms;
 526                 int nodeid;
 527
 528                 ms = __nr_to_section(pnum);
 529                 nodeid = sparse_early_nid(ms);
 530                 if (nodeid == nodeid_begin) {
 531                         map_count++;
 532                         continue;
 533                 }
 534                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 535                 alloc_func(data, pnum_begin, pnum,
 536                                                 map_count, nodeid_begin);
 537                 /* new start, update count etc*/
 538                 nodeid_begin = nodeid;
 539                 pnum_begin = pnum;
 540                 map_count = 1;
 541         }
 542         /* ok, last chunk */
 543         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
 544                                                 map_count, nodeid_begin);
 545 }
 546
 547 /*
 548  * Allocate the accumulated non-linear sections, allocate a mem_map
 549  * for each and record the physical to section mapping.
 550  */
 551 void __init sparse_init(void)
 552 {
 553         unsigned long pnum;
 554         struct page *map;
 555         unsigned long *usemap;
 556         unsigned long **usemap_map;
 557         int size;
 558 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 559         int size2;
 560         struct page **map_map;
 561 #endif
 562
 563         /* see include/linux/mmzone.h 'struct mem_section' definition */
 564         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 565
 566         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 567         set_pageblock_order();
 568
 569         /*
 570          * map is using big page (aka 2M in x86 64 bit)
 571          * usemap is less one page (aka 24 bytes)
 572          * so alloc 2M (with 2M align) and 24 bytes in turn will
 573          * make next 2M slip to one more 2M later.
 574          * then in big system, the memory will have a lot of holes...
 575          * here try to allocate 2M pages continuously.
 576          *
 577          * powerpc need to call sparse_init_one_section right after each
 578          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 579          */
 580         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 581         usemap_map = memblock_virt_alloc(size, 0);
 582         if (!usemap_map)
 583                 panic("can not allocate usemap_map\n");
 584         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 585                                                         (void *)usemap_map);
 586
 587 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 588         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 589         map_map = memblock_virt_alloc(size2, 0);
 590         if (!map_map)
 591                 panic("can not allocate map_map\n");
 592         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 593                                                         (void *)map_map);
 594 #endif
 595
 596         for_each_present_section_nr(0, pnum) {
 597                 usemap = usemap_map[pnum];
 598                 if (!usemap)
 599                         continue;
 600
 601 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 602                 map = map_map[pnum];
 603 #else
 604                 map = sparse_early_mem_map_alloc(pnum);
 605 #endif
 606                 if (!map)
 607                         continue;
 608
 609                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 610                                                                 usemap);
 611         }
 612
 613         vmemmap_populate_print_last();
 614
 615 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 616         memblock_free_early(__pa(map_map), size2);
 617 #endif
 618         memblock_free_early(__pa(usemap_map), size);
 619 }
 620
 621 #ifdef CONFIG_MEMORY_HOTPLUG
 622
 623 /* Mark all memory sections within the pfn range as online */
 624 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 625 {
 626         unsigned long pfn;
 627
 628         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 629                 unsigned long section_nr = pfn_to_section_nr(start_pfn);
 630                 struct mem_section *ms;
 631
 632                 /* onlining code should never touch invalid ranges */
 633                 if (WARN_ON(!valid_section_nr(section_nr)))
 634                         continue;
 635
 636                 ms = __nr_to_section(section_nr);
 637                 ms->section_mem_map |= SECTION_IS_ONLINE;
 638         }
 639 }
 640
 641 #ifdef CONFIG_MEMORY_HOTREMOVE
 642 /* Mark all memory sections within the pfn range as online */
 643 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 644 {
 645         unsigned long pfn;
 646
 647         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 648                 unsigned long section_nr = pfn_to_section_nr(start_pfn);
 649                 struct mem_section *ms;
 650
 651                 /*
 652                  * TODO this needs some double checking. Offlining code makes
 653                  * sure to check pfn_valid but those checks might be just bogus
 654                  */
 655                 if (WARN_ON(!valid_section_nr(section_nr)))
 656                         continue;
 657
 658                 ms = __nr_to_section(section_nr);
 659                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 660         }
 661 }
 662 #endif
 663
 664 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 665 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 666 {
 667         /* This will make the necessary allocations eventually. */
 668         return sparse_mem_map_populate(pnum, nid);
 669 }
 670 static void __kfree_section_memmap(struct page *memmap)
 671 {
 672         unsigned long start = (unsigned long)memmap;
 673         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 674
 675         vmemmap_free(start, end);
 676 }
 677 #ifdef CONFIG_MEMORY_HOTREMOVE
 678 static void free_map_bootmem(struct page *memmap)
 679 {
 680         unsigned long start = (unsigned long)memmap;
 681         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 682
 683         vmemmap_free(start, end);
 684 }
 685 #endif /* CONFIG_MEMORY_HOTREMOVE */
 686 #else
 687 static struct page *__kmalloc_section_memmap(void)
 688 {
 689         struct page *page, *ret;
 690         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 691
 692         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 693         if (page)
 694                 goto got_map_page;
 695
 696         ret = vmalloc(memmap_size);
 697         if (ret)
 698                 goto got_map_ptr;
 699
 700         return NULL;
 701 got_map_page:
 702         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 703 got_map_ptr:
 704
 705         return ret;
 706 }
 707
 708 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 709 {
 710         return __kmalloc_section_memmap();
 711 }
 712
 713 static void __kfree_section_memmap(struct page *memmap)
 714 {
 715         if (is_vmalloc_addr(memmap))
 716                 vfree(memmap);
 717         else
 718                 free_pages((unsigned long)memmap,
 719                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 720 }
 721
 722 #ifdef CONFIG_MEMORY_HOTREMOVE
 723 static void free_map_bootmem(struct page *memmap)
 724 {
 725         unsigned long maps_section_nr, removing_section_nr, i;
 726         unsigned long magic, nr_pages;
 727         struct page *page = virt_to_page(memmap);
 728
 729         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 730                 >> PAGE_SHIFT;
 731
 732         for (i = 0; i < nr_pages; i++, page++) {
 733                 magic = (unsigned long) page->freelist;
 734
 735                 BUG_ON(magic == NODE_INFO);
 736
 737                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 738                 removing_section_nr = page_private(page);
 739
 740                 /*
 741                  * When this function is called, the removing section is
 742                  * logical offlined state. This means all pages are isolated
 743                  * from page allocator. If removing section's memmap is placed
 744                  * on the same section, it must not be freed.
 745                  * If it is freed, page allocator may allocate it which will
 746                  * be removed physically soon.
 747                  */
 748                 if (maps_section_nr != removing_section_nr)
 749                         put_page_bootmem(page);
 750         }
 751 }
 752 #endif /* CONFIG_MEMORY_HOTREMOVE */
 753 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 754
 755 /*
 756  * returns the number of sections whose mem_maps were properly
 757  * set.  If this is <=0, then that means that the passed-in
 758  * map was not consumed and must be freed.
 759  */
 760 int __meminit sparse_add_one_section(struct pglist_data *pgdat, unsigned long start_pfn)
 761 {
 762         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 763         struct mem_section *ms;
 764         struct page *memmap;
 765         unsigned long *usemap;
 766         unsigned long flags;
 767         int ret;
 768
 769         /*
 770          * no locking for this, because it does its own
 771          * plus, it does a kmalloc
 772          */
 773         ret = sparse_index_init(section_nr, pgdat->node_id);
 774         if (ret < 0 && ret != -EEXIST)
 775                 return ret;
 776         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
 777         if (!memmap)
 778                 return -ENOMEM;
 779         usemap = __kmalloc_section_usemap();
 780         if (!usemap) {
 781                 __kfree_section_memmap(memmap);
 782                 return -ENOMEM;
 783         }
 784
 785         pgdat_resize_lock(pgdat, &flags);
 786
 787         ms = __pfn_to_section(start_pfn);
 788         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 789                 ret = -EEXIST;
 790                 goto out;
 791         }
 792
 793         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 794
 795         section_mark_present(ms);
 796
 797         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 798
 799 out:
 800         pgdat_resize_unlock(pgdat, &flags);
 801         if (ret <= 0) {
 802                 kfree(usemap);
 803                 __kfree_section_memmap(memmap);
 804         }
 805         return ret;
 806 }
 807
 808 #ifdef CONFIG_MEMORY_HOTREMOVE
 809 #ifdef CONFIG_MEMORY_FAILURE
 810 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 811 {
 812         int i;
 813
 814         if (!memmap)
 815                 return;
 816
 817         for (i = 0; i < nr_pages; i++) {
 818                 if (PageHWPoison(&memmap[i])) {
 819                         atomic_long_sub(1, &num_poisoned_pages);
 820                         ClearPageHWPoison(&memmap[i]);
 821                 }
 822         }
 823 }
 824 #else
 825 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 826 {
 827 }
 828 #endif
 829
 830 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 831 {
 832         struct page *usemap_page;
 833
 834         if (!usemap)
 835                 return;
 836
 837         usemap_page = virt_to_page(usemap);
 838         /*
 839          * Check to see if allocation came from hot-plug-add
 840          */
 841         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 842                 kfree(usemap);
 843                 if (memmap)
 844                         __kfree_section_memmap(memmap);
 845                 return;
 846         }
 847
 848         /*
 849          * The usemap came from bootmem. This is packed with other usemaps
 850          * on the section which has pgdat at boot time. Just keep it as is now.
 851          */
 852
 853         if (memmap)
 854                 free_map_bootmem(memmap);
 855 }
 856
 857 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 858                 unsigned long map_offset)
 859 {
 860         struct page *memmap = NULL;
 861         unsigned long *usemap = NULL, flags;
 862         struct pglist_data *pgdat = zone->zone_pgdat;
 863
 864         pgdat_resize_lock(pgdat, &flags);
 865         if (ms->section_mem_map) {
 866                 usemap = ms->pageblock_flags;
 867                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 868                                                 __section_nr(ms));
 869                 ms->section_mem_map = 0;
 870                 ms->pageblock_flags = NULL;
 871         }
 872         pgdat_resize_unlock(pgdat, &flags);
 873
 874         clear_hwpoisoned_pages(memmap + map_offset,
 875                         PAGES_PER_SECTION - map_offset);
 876         free_section_usemap(memmap, usemap);
 877 }
 878 #endif /* CONFIG_MEMORY_HOTREMOVE */
 879 #endif /* CONFIG_MEMORY_HOTPLUG */