mm/sparse.c

   1 /*
   2  * sparse memory mappings.
   3  */
   4 #include <linux/mm.h>
   5 #include <linux/slab.h>
   6 #include <linux/mmzone.h>
   7 #include <linux/bootmem.h>
   8 #include <linux/highmem.h>
   9 #include <linux/export.h>
  10 #include <linux/spinlock.h>
  11 #include <linux/vmalloc.h>
  12 #include "internal.h"
  13 #include <asm/dma.h>
  14 #include <asm/pgalloc.h>
  15 #include <asm/pgtable.h>
  16
  17 /*
  18  * Permanent SPARSEMEM data:
  19  *
  20  * 1) mem_section       - memory sections, mem_map's for valid memory
  21  */
  22 #ifdef CONFIG_SPARSEMEM_EXTREME
  23 struct mem_section *mem_section[NR_SECTION_ROOTS]
  24         ____cacheline_internodealigned_in_smp;
  25 #else
  26 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  27         ____cacheline_internodealigned_in_smp;
  28 #endif
  29 EXPORT_SYMBOL(mem_section);
  30
  31 #ifdef NODE_NOT_IN_PAGE_FLAGS
  32 /*
  33  * If we did not store the node number in the page then we have to
  34  * do a lookup in the section_to_node_table in order to find which
  35  * node the page belongs to.
  36  */
  37 #if MAX_NUMNODES <= 256
  38 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  39 #else
  40 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  41 #endif
  42
  43 int page_to_nid(const struct page *page)
  44 {
  45         return section_to_node_table[page_to_section(page)];
  46 }
  47 EXPORT_SYMBOL(page_to_nid);
  48
  49 static void set_section_nid(unsigned long section_nr, int nid)
  50 {
  51         section_to_node_table[section_nr] = nid;
  52 }
  53 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  54 static inline void set_section_nid(unsigned long section_nr, int nid)
  55 {
  56 }
  57 #endif
  58
  59 #ifdef CONFIG_SPARSEMEM_EXTREME
  60 static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
  61 {
  62         struct mem_section *section = NULL;
  63         unsigned long array_size = SECTIONS_PER_ROOT *
  64                                    sizeof(struct mem_section);
  65
  66         if (slab_is_available()) {
  67                 if (node_state(nid, N_HIGH_MEMORY))
  68                         section = kzalloc_node(array_size, GFP_KERNEL, nid);
  69                 else
  70                         section = kzalloc(array_size, GFP_KERNEL);
  71         } else {
  72                 section = memblock_virt_alloc_node(array_size, nid);
  73         }
  74
  75         return section;
  76 }
  77
  78 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  79 {
  80         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  81         struct mem_section *section;
  82
  83         if (mem_section[root])
  84                 return -EEXIST;
  85
  86         section = sparse_index_alloc(nid);
  87         if (!section)
  88                 return -ENOMEM;
  89
  90         mem_section[root] = section;
  91
  92         return 0;
  93 }
  94 #else /* !SPARSEMEM_EXTREME */
  95 static inline int sparse_index_init(unsigned long section_nr, int nid)
  96 {
  97         return 0;
  98 }
  99 #endif
 100
 101 /*
 102  * Although written for the SPARSEMEM_EXTREME case, this happens
 103  * to also work for the flat array case because
 104  * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
 105  */
 106 int __section_nr(struct mem_section* ms)
 107 {
 108         unsigned long root_nr;
 109         struct mem_section* root;
 110
 111         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 112                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 113                 if (!root)
 114                         continue;
 115
 116                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 117                      break;
 118         }
 119
 120         VM_BUG_ON(root_nr == NR_SECTION_ROOTS);
 121
 122         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 123 }
 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 /* Record a memory area against a node. */
 168 void __init memory_present(int nid, unsigned long start, unsigned long end)
 169 {
 170         unsigned long pfn;
 171
 172         start &= PAGE_SECTION_MASK;
 173         mminit_validate_memmodel_limits(&start, &end);
 174         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 175                 unsigned long section = pfn_to_section_nr(pfn);
 176                 struct mem_section *ms;
 177
 178                 sparse_index_init(section, nid);
 179                 set_section_nid(section, nid);
 180
 181                 ms = __nr_to_section(section);
 182                 if (!ms->section_mem_map)
 183                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 184                                                         SECTION_MARKED_PRESENT;
 185         }
 186 }
 187
 188 /*
 189  * Only used by the i386 NUMA architecures, but relatively
 190  * generic code.
 191  */
 192 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 193                                                      unsigned long end_pfn)
 194 {
 195         unsigned long pfn;
 196         unsigned long nr_pages = 0;
 197
 198         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 199         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 200                 if (nid != early_pfn_to_nid(pfn))
 201                         continue;
 202
 203                 if (pfn_present(pfn))
 204                         nr_pages += PAGES_PER_SECTION;
 205         }
 206
 207         return nr_pages * sizeof(struct page);
 208 }
 209
 210 /*
 211  * Subtle, we encode the real pfn into the mem_map such that
 212  * the identity pfn - section_mem_map will return the actual
 213  * physical page frame number.
 214  */
 215 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 216 {
 217         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 218 }
 219
 220 /*
 221  * Decode mem_map from the coded memmap
 222  */
 223 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 224 {
 225         /* mask off the extra low bits of information */
 226         coded_mem_map &= SECTION_MAP_MASK;
 227         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 228 }
 229
 230 static int __meminit sparse_init_one_section(struct mem_section *ms,
 231                 unsigned long pnum, struct page *mem_map,
 232                 unsigned long *pageblock_bitmap)
 233 {
 234         if (!present_section(ms))
 235                 return -EINVAL;
 236
 237         ms->section_mem_map &= ~SECTION_MAP_MASK;
 238         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 239                                                         SECTION_HAS_MEM_MAP;
 240         ms->pageblock_flags = pageblock_bitmap;
 241
 242         return 1;
 243 }
 244
 245 unsigned long usemap_size(void)
 246 {
 247         unsigned long size_bytes;
 248         size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 249         size_bytes = roundup(size_bytes, sizeof(unsigned long));
 250         return size_bytes;
 251 }
 252
 253 #ifdef CONFIG_MEMORY_HOTPLUG
 254 static unsigned long *__kmalloc_section_usemap(void)
 255 {
 256         return kmalloc(usemap_size(), GFP_KERNEL);
 257 }
 258 #endif /* CONFIG_MEMORY_HOTPLUG */
 259
 260 #ifdef CONFIG_MEMORY_HOTREMOVE
 261 static unsigned long * __init
 262 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 263                                          unsigned long size)
 264 {
 265         unsigned long goal, limit;
 266         unsigned long *p;
 267         int nid;
 268         /*
 269          * A page may contain usemaps for other sections preventing the
 270          * page being freed and making a section unremovable while
 271          * other sections referencing the usemap retmain active. Similarly,
 272          * a pgdat can prevent a section being removed. If section A
 273          * contains a pgdat and section B contains the usemap, both
 274          * sections become inter-dependent. This allocates usemaps
 275          * from the same section as the pgdat where possible to avoid
 276          * this problem.
 277          */
 278         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 279         limit = goal + (1UL << PA_SECTION_SHIFT);
 280         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 281 again:
 282         p = memblock_virt_alloc_try_nid_nopanic(size,
 283                                                 SMP_CACHE_BYTES, goal, limit,
 284                                                 nid);
 285         if (!p && limit) {
 286                 limit = 0;
 287                 goto again;
 288         }
 289         return p;
 290 }
 291
 292 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 293 {
 294         unsigned long usemap_snr, pgdat_snr;
 295         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 296         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 297         struct pglist_data *pgdat = NODE_DATA(nid);
 298         int usemap_nid;
 299
 300         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 301         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 302         if (usemap_snr == pgdat_snr)
 303                 return;
 304
 305         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 306                 /* skip redundant message */
 307                 return;
 308
 309         old_usemap_snr = usemap_snr;
 310         old_pgdat_snr = pgdat_snr;
 311
 312         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 313         if (usemap_nid != nid) {
 314                 printk(KERN_INFO
 315                        "node %d must be removed before remove section %ld\n",
 316                        nid, usemap_snr);
 317                 return;
 318         }
 319         /*
 320          * There is a circular dependency.
 321          * Some platforms allow un-removable section because they will just
 322          * gather other removable sections for dynamic partitioning.
 323          * Just notify un-removable section's number here.
 324          */
 325         printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
 326                pgdat_snr, nid);
 327         printk(KERN_CONT
 328                " have a circular dependency on usemap and pgdat allocations\n");
 329 }
 330 #else
 331 static unsigned long * __init
 332 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 333                                          unsigned long size)
 334 {
 335         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 336 }
 337
 338 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 339 {
 340 }
 341 #endif /* CONFIG_MEMORY_HOTREMOVE */
 342
 343 static void __init sparse_early_usemaps_alloc_node(void *data,
 344                                  unsigned long pnum_begin,
 345                                  unsigned long pnum_end,
 346                                  unsigned long usemap_count, int nodeid)
 347 {
 348         void *usemap;
 349         unsigned long pnum;
 350         unsigned long **usemap_map = (unsigned long **)data;
 351         int size = usemap_size();
 352
 353         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 354                                                           size * usemap_count);
 355         if (!usemap) {
 356                 printk(KERN_WARNING "%s: allocation failed\n", __func__);
 357                 return;
 358         }
 359
 360         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 361                 if (!present_section_nr(pnum))
 362                         continue;
 363                 usemap_map[pnum] = usemap;
 364                 usemap += size;
 365                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 366         }
 367 }
 368
 369 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 370 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 371 {
 372         struct page *map;
 373         unsigned long size;
 374
 375         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 376         if (map)
 377                 return map;
 378
 379         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 380         map = memblock_virt_alloc_try_nid(size,
 381                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 382                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 383         return map;
 384 }
 385 void __init sparse_mem_maps_populate_node(struct page **map_map,
 386                                           unsigned long pnum_begin,
 387                                           unsigned long pnum_end,
 388                                           unsigned long map_count, int nodeid)
 389 {
 390         void *map;
 391         unsigned long pnum;
 392         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 393
 394         map = alloc_remap(nodeid, size * map_count);
 395         if (map) {
 396                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 397                         if (!present_section_nr(pnum))
 398                                 continue;
 399                         map_map[pnum] = map;
 400                         map += size;
 401                 }
 402                 return;
 403         }
 404
 405         size = PAGE_ALIGN(size);
 406         map = memblock_virt_alloc_try_nid(size * map_count,
 407                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 408                                           BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 409         if (map) {
 410                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 411                         if (!present_section_nr(pnum))
 412                                 continue;
 413                         map_map[pnum] = map;
 414                         map += size;
 415                 }
 416                 return;
 417         }
 418
 419         /* fallback */
 420         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 421                 struct mem_section *ms;
 422
 423                 if (!present_section_nr(pnum))
 424                         continue;
 425                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 426                 if (map_map[pnum])
 427                         continue;
 428                 ms = __nr_to_section(pnum);
 429                 printk(KERN_ERR "%s: sparsemem memory map backing failed "
 430                         "some memory will not be available.\n", __func__);
 431                 ms->section_mem_map = 0;
 432         }
 433 }
 434 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 435
 436 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 437 static void __init sparse_early_mem_maps_alloc_node(void *data,
 438                                  unsigned long pnum_begin,
 439                                  unsigned long pnum_end,
 440                                  unsigned long map_count, int nodeid)
 441 {
 442         struct page **map_map = (struct page **)data;
 443         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 444                                          map_count, nodeid);
 445 }
 446 #else
 447 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 448 {
 449         struct page *map;
 450         struct mem_section *ms = __nr_to_section(pnum);
 451         int nid = sparse_early_nid(ms);
 452
 453         map = sparse_mem_map_populate(pnum, nid);
 454         if (map)
 455                 return map;
 456
 457         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 458                         "some memory will not be available.\n", __func__);
 459         ms->section_mem_map = 0;
 460         return NULL;
 461 }
 462 #endif
 463
 464 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 465 {
 466 }
 467
 468 /**
 469  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 470  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 471  */
 472 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 473                                         (void *, unsigned long, unsigned long,
 474                                         unsigned long, int), void *data)
 475 {
 476         unsigned long pnum;
 477         unsigned long map_count;
 478         int nodeid_begin = 0;
 479         unsigned long pnum_begin = 0;
 480
 481         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 482                 struct mem_section *ms;
 483
 484                 if (!present_section_nr(pnum))
 485                         continue;
 486                 ms = __nr_to_section(pnum);
 487                 nodeid_begin = sparse_early_nid(ms);
 488                 pnum_begin = pnum;
 489                 break;
 490         }
 491         map_count = 1;
 492         for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 493                 struct mem_section *ms;
 494                 int nodeid;
 495
 496                 if (!present_section_nr(pnum))
 497                         continue;
 498                 ms = __nr_to_section(pnum);
 499                 nodeid = sparse_early_nid(ms);
 500                 if (nodeid == nodeid_begin) {
 501                         map_count++;
 502                         continue;
 503                 }
 504                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 505                 alloc_func(data, pnum_begin, pnum,
 506                                                 map_count, nodeid_begin);
 507                 /* new start, update count etc*/
 508                 nodeid_begin = nodeid;
 509                 pnum_begin = pnum;
 510                 map_count = 1;
 511         }
 512         /* ok, last chunk */
 513         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
 514                                                 map_count, nodeid_begin);
 515 }
 516
 517 /*
 518  * Allocate the accumulated non-linear sections, allocate a mem_map
 519  * for each and record the physical to section mapping.
 520  */
 521 void __init sparse_init(void)
 522 {
 523         unsigned long pnum;
 524         struct page *map;
 525         unsigned long *usemap;
 526         unsigned long **usemap_map;
 527         int size;
 528 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 529         int size2;
 530         struct page **map_map;
 531 #endif
 532
 533         /* see include/linux/mmzone.h 'struct mem_section' definition */
 534         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 535
 536         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 537         set_pageblock_order();
 538
 539         /*
 540          * map is using big page (aka 2M in x86 64 bit)
 541          * usemap is less one page (aka 24 bytes)
 542          * so alloc 2M (with 2M align) and 24 bytes in turn will
 543          * make next 2M slip to one more 2M later.
 544          * then in big system, the memory will have a lot of holes...
 545          * here try to allocate 2M pages continuously.
 546          *
 547          * powerpc need to call sparse_init_one_section right after each
 548          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 549          */
 550         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 551         usemap_map = memblock_virt_alloc(size, 0);
 552         if (!usemap_map)
 553                 panic("can not allocate usemap_map\n");
 554         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 555                                                         (void *)usemap_map);
 556
 557 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 558         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 559         map_map = memblock_virt_alloc(size2, 0);
 560         if (!map_map)
 561                 panic("can not allocate map_map\n");
 562         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 563                                                         (void *)map_map);
 564 #endif
 565
 566         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 567                 if (!present_section_nr(pnum))
 568                         continue;
 569
 570                 usemap = usemap_map[pnum];
 571                 if (!usemap)
 572                         continue;
 573
 574 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 575                 map = map_map[pnum];
 576 #else
 577                 map = sparse_early_mem_map_alloc(pnum);
 578 #endif
 579                 if (!map)
 580                         continue;
 581
 582                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 583                                                                 usemap);
 584         }
 585
 586         vmemmap_populate_print_last();
 587
 588 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 589         memblock_free_early(__pa(map_map), size2);
 590 #endif
 591         memblock_free_early(__pa(usemap_map), size);
 592 }
 593
 594 #ifdef CONFIG_MEMORY_HOTPLUG
 595 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 596 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 597 {
 598         /* This will make the necessary allocations eventually. */
 599         return sparse_mem_map_populate(pnum, nid);
 600 }
 601 static void __kfree_section_memmap(struct page *memmap)
 602 {
 603         unsigned long start = (unsigned long)memmap;
 604         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 605
 606         vmemmap_free(start, end);
 607 }
 608 #ifdef CONFIG_MEMORY_HOTREMOVE
 609 static void free_map_bootmem(struct page *memmap)
 610 {
 611         unsigned long start = (unsigned long)memmap;
 612         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 613
 614         vmemmap_free(start, end);
 615 }
 616 #endif /* CONFIG_MEMORY_HOTREMOVE */
 617 #else
 618 static struct page *__kmalloc_section_memmap(void)
 619 {
 620         struct page *page, *ret;
 621         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 622
 623         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 624         if (page)
 625                 goto got_map_page;
 626
 627         ret = vmalloc(memmap_size);
 628         if (ret)
 629                 goto got_map_ptr;
 630
 631         return NULL;
 632 got_map_page:
 633         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 634 got_map_ptr:
 635
 636         return ret;
 637 }
 638
 639 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 640 {
 641         return __kmalloc_section_memmap();
 642 }
 643
 644 static void __kfree_section_memmap(struct page *memmap)
 645 {
 646         if (is_vmalloc_addr(memmap))
 647                 vfree(memmap);
 648         else
 649                 free_pages((unsigned long)memmap,
 650                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 651 }
 652
 653 #ifdef CONFIG_MEMORY_HOTREMOVE
 654 static void free_map_bootmem(struct page *memmap)
 655 {
 656         unsigned long maps_section_nr, removing_section_nr, i;
 657         unsigned long magic, nr_pages;
 658         struct page *page = virt_to_page(memmap);
 659
 660         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 661                 >> PAGE_SHIFT;
 662
 663         for (i = 0; i < nr_pages; i++, page++) {
 664                 magic = (unsigned long) page->lru.next;
 665
 666                 BUG_ON(magic == NODE_INFO);
 667
 668                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 669                 removing_section_nr = page->private;
 670
 671                 /*
 672                  * When this function is called, the removing section is
 673                  * logical offlined state. This means all pages are isolated
 674                  * from page allocator. If removing section's memmap is placed
 675                  * on the same section, it must not be freed.
 676                  * If it is freed, page allocator may allocate it which will
 677                  * be removed physically soon.
 678                  */
 679                 if (maps_section_nr != removing_section_nr)
 680                         put_page_bootmem(page);
 681         }
 682 }
 683 #endif /* CONFIG_MEMORY_HOTREMOVE */
 684 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 685
 686 /*
 687  * returns the number of sections whose mem_maps were properly
 688  * set.  If this is <=0, then that means that the passed-in
 689  * map was not consumed and must be freed.
 690  */
 691 int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn)
 692 {
 693         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 694         struct pglist_data *pgdat = zone->zone_pgdat;
 695         struct mem_section *ms;
 696         struct page *memmap;
 697         unsigned long *usemap;
 698         unsigned long flags;
 699         int ret;
 700
 701         /*
 702          * no locking for this, because it does its own
 703          * plus, it does a kmalloc
 704          */
 705         ret = sparse_index_init(section_nr, pgdat->node_id);
 706         if (ret < 0 && ret != -EEXIST)
 707                 return ret;
 708         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
 709         if (!memmap)
 710                 return -ENOMEM;
 711         usemap = __kmalloc_section_usemap();
 712         if (!usemap) {
 713                 __kfree_section_memmap(memmap);
 714                 return -ENOMEM;
 715         }
 716
 717         pgdat_resize_lock(pgdat, &flags);
 718
 719         ms = __pfn_to_section(start_pfn);
 720         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 721                 ret = -EEXIST;
 722                 goto out;
 723         }
 724
 725         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 726
 727         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 728
 729         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 730
 731 out:
 732         pgdat_resize_unlock(pgdat, &flags);
 733         if (ret <= 0) {
 734                 kfree(usemap);
 735                 __kfree_section_memmap(memmap);
 736         }
 737         return ret;
 738 }
 739
 740 #ifdef CONFIG_MEMORY_HOTREMOVE
 741 #ifdef CONFIG_MEMORY_FAILURE
 742 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 743 {
 744         int i;
 745
 746         if (!memmap)
 747                 return;
 748
 749         for (i = 0; i < PAGES_PER_SECTION; i++) {
 750                 if (PageHWPoison(&memmap[i])) {
 751                         atomic_long_sub(1, &num_poisoned_pages);
 752                         ClearPageHWPoison(&memmap[i]);
 753                 }
 754         }
 755 }
 756 #else
 757 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 758 {
 759 }
 760 #endif
 761
 762 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 763 {
 764         struct page *usemap_page;
 765
 766         if (!usemap)
 767                 return;
 768
 769         usemap_page = virt_to_page(usemap);
 770         /*
 771          * Check to see if allocation came from hot-plug-add
 772          */
 773         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 774                 kfree(usemap);
 775                 if (memmap)
 776                         __kfree_section_memmap(memmap);
 777                 return;
 778         }
 779
 780         /*
 781          * The usemap came from bootmem. This is packed with other usemaps
 782          * on the section which has pgdat at boot time. Just keep it as is now.
 783          */
 784
 785         if (memmap)
 786                 free_map_bootmem(memmap);
 787 }
 788
 789 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 790 {
 791         struct page *memmap = NULL;
 792         unsigned long *usemap = NULL, flags;
 793         struct pglist_data *pgdat = zone->zone_pgdat;
 794
 795         pgdat_resize_lock(pgdat, &flags);
 796         if (ms->section_mem_map) {
 797                 usemap = ms->pageblock_flags;
 798                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 799                                                 __section_nr(ms));
 800                 ms->section_mem_map = 0;
 801                 ms->pageblock_flags = NULL;
 802         }
 803         pgdat_resize_unlock(pgdat, &flags);
 804
 805         clear_hwpoisoned_pages(memmap, PAGES_PER_SECTION);
 806         free_section_usemap(memmap, usemap);
 807 }
 808 #endif /* CONFIG_MEMORY_HOTREMOVE */
 809 #endif /* CONFIG_MEMORY_HOTPLUG */