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 = alloc_bootmem_node(NODE_DATA(nid), array_size);
  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 = ___alloc_bootmem_node_nopanic(NODE_DATA(nid), size,
 283                                           SMP_CACHE_BYTES, goal, limit);
 284         if (!p && limit) {
 285                 limit = 0;
 286                 goto again;
 287         }
 288         return p;
 289 }
 290
 291 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 292 {
 293         unsigned long usemap_snr, pgdat_snr;
 294         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 295         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 296         struct pglist_data *pgdat = NODE_DATA(nid);
 297         int usemap_nid;
 298
 299         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 300         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 301         if (usemap_snr == pgdat_snr)
 302                 return;
 303
 304         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 305                 /* skip redundant message */
 306                 return;
 307
 308         old_usemap_snr = usemap_snr;
 309         old_pgdat_snr = pgdat_snr;
 310
 311         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 312         if (usemap_nid != nid) {
 313                 printk(KERN_INFO
 314                        "node %d must be removed before remove section %ld\n",
 315                        nid, usemap_snr);
 316                 return;
 317         }
 318         /*
 319          * There is a circular dependency.
 320          * Some platforms allow un-removable section because they will just
 321          * gather other removable sections for dynamic partitioning.
 322          * Just notify un-removable section's number here.
 323          */
 324         printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
 325                pgdat_snr, nid);
 326         printk(KERN_CONT
 327                " have a circular dependency on usemap and pgdat allocations\n");
 328 }
 329 #else
 330 static unsigned long * __init
 331 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 332                                          unsigned long size)
 333 {
 334         return alloc_bootmem_node_nopanic(pgdat, size);
 335 }
 336
 337 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 338 {
 339 }
 340 #endif /* CONFIG_MEMORY_HOTREMOVE */
 341
 342 static void __init sparse_early_usemaps_alloc_node(void *data,
 343                                  unsigned long pnum_begin,
 344                                  unsigned long pnum_end,
 345                                  unsigned long usemap_count, int nodeid)
 346 {
 347         void *usemap;
 348         unsigned long pnum;
 349         unsigned long **usemap_map = (unsigned long **)data;
 350         int size = usemap_size();
 351
 352         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 353                                                           size * usemap_count);
 354         if (!usemap) {
 355                 printk(KERN_WARNING "%s: allocation failed\n", __func__);
 356                 return;
 357         }
 358
 359         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 360                 if (!present_section_nr(pnum))
 361                         continue;
 362                 usemap_map[pnum] = usemap;
 363                 usemap += size;
 364                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 365         }
 366 }
 367
 368 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 369 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 370 {
 371         struct page *map;
 372         unsigned long size;
 373
 374         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 375         if (map)
 376                 return map;
 377
 378         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 379         map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
 380                                          PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 381         return map;
 382 }
 383 void __init sparse_mem_maps_populate_node(struct page **map_map,
 384                                           unsigned long pnum_begin,
 385                                           unsigned long pnum_end,
 386                                           unsigned long map_count, int nodeid)
 387 {
 388         void *map;
 389         unsigned long pnum;
 390         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 391
 392         map = alloc_remap(nodeid, size * map_count);
 393         if (map) {
 394                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 395                         if (!present_section_nr(pnum))
 396                                 continue;
 397                         map_map[pnum] = map;
 398                         map += size;
 399                 }
 400                 return;
 401         }
 402
 403         size = PAGE_ALIGN(size);
 404         map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
 405                                          PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 406         if (map) {
 407                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 408                         if (!present_section_nr(pnum))
 409                                 continue;
 410                         map_map[pnum] = map;
 411                         map += size;
 412                 }
 413                 return;
 414         }
 415
 416         /* fallback */
 417         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 418                 struct mem_section *ms;
 419
 420                 if (!present_section_nr(pnum))
 421                         continue;
 422                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 423                 if (map_map[pnum])
 424                         continue;
 425                 ms = __nr_to_section(pnum);
 426                 printk(KERN_ERR "%s: sparsemem memory map backing failed "
 427                         "some memory will not be available.\n", __func__);
 428                 ms->section_mem_map = 0;
 429         }
 430 }
 431 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 432
 433 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 434 static void __init sparse_early_mem_maps_alloc_node(void *data,
 435                                  unsigned long pnum_begin,
 436                                  unsigned long pnum_end,
 437                                  unsigned long map_count, int nodeid)
 438 {
 439         struct page **map_map = (struct page **)data;
 440         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 441                                          map_count, nodeid);
 442 }
 443 #else
 444 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 445 {
 446         struct page *map;
 447         struct mem_section *ms = __nr_to_section(pnum);
 448         int nid = sparse_early_nid(ms);
 449
 450         map = sparse_mem_map_populate(pnum, nid);
 451         if (map)
 452                 return map;
 453
 454         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 455                         "some memory will not be available.\n", __func__);
 456         ms->section_mem_map = 0;
 457         return NULL;
 458 }
 459 #endif
 460
 461 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 462 {
 463 }
 464
 465 /**
 466  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 467  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 468  */
 469 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 470                                         (void *, unsigned long, unsigned long,
 471                                         unsigned long, int), void *data)
 472 {
 473         unsigned long pnum;
 474         unsigned long map_count;
 475         int nodeid_begin = 0;
 476         unsigned long pnum_begin = 0;
 477
 478         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 479                 struct mem_section *ms;
 480
 481                 if (!present_section_nr(pnum))
 482                         continue;
 483                 ms = __nr_to_section(pnum);
 484                 nodeid_begin = sparse_early_nid(ms);
 485                 pnum_begin = pnum;
 486                 break;
 487         }
 488         map_count = 1;
 489         for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 490                 struct mem_section *ms;
 491                 int nodeid;
 492
 493                 if (!present_section_nr(pnum))
 494                         continue;
 495                 ms = __nr_to_section(pnum);
 496                 nodeid = sparse_early_nid(ms);
 497                 if (nodeid == nodeid_begin) {
 498                         map_count++;
 499                         continue;
 500                 }
 501                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 502                 alloc_func(data, pnum_begin, pnum,
 503                                                 map_count, nodeid_begin);
 504                 /* new start, update count etc*/
 505                 nodeid_begin = nodeid;
 506                 pnum_begin = pnum;
 507                 map_count = 1;
 508         }
 509         /* ok, last chunk */
 510         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
 511                                                 map_count, nodeid_begin);
 512 }
 513
 514 /*
 515  * Allocate the accumulated non-linear sections, allocate a mem_map
 516  * for each and record the physical to section mapping.
 517  */
 518 void __init sparse_init(void)
 519 {
 520         unsigned long pnum;
 521         struct page *map;
 522         unsigned long *usemap;
 523         unsigned long **usemap_map;
 524         int size;
 525 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 526         int size2;
 527         struct page **map_map;
 528 #endif
 529
 530         /* see include/linux/mmzone.h 'struct mem_section' definition */
 531         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 532
 533         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 534         set_pageblock_order();
 535
 536         /*
 537          * map is using big page (aka 2M in x86 64 bit)
 538          * usemap is less one page (aka 24 bytes)
 539          * so alloc 2M (with 2M align) and 24 bytes in turn will
 540          * make next 2M slip to one more 2M later.
 541          * then in big system, the memory will have a lot of holes...
 542          * here try to allocate 2M pages continuously.
 543          *
 544          * powerpc need to call sparse_init_one_section right after each
 545          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 546          */
 547         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 548         usemap_map = alloc_bootmem(size);
 549         if (!usemap_map)
 550                 panic("can not allocate usemap_map\n");
 551         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 552                                                         (void *)usemap_map);
 553
 554 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 555         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 556         map_map = alloc_bootmem(size2);
 557         if (!map_map)
 558                 panic("can not allocate map_map\n");
 559         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 560                                                         (void *)map_map);
 561 #endif
 562
 563         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 564                 if (!present_section_nr(pnum))
 565                         continue;
 566
 567                 usemap = usemap_map[pnum];
 568                 if (!usemap)
 569                         continue;
 570
 571 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 572                 map = map_map[pnum];
 573 #else
 574                 map = sparse_early_mem_map_alloc(pnum);
 575 #endif
 576                 if (!map)
 577                         continue;
 578
 579                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 580                                                                 usemap);
 581         }
 582
 583         vmemmap_populate_print_last();
 584
 585 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 586         free_bootmem(__pa(map_map), size2);
 587 #endif
 588         free_bootmem(__pa(usemap_map), size);
 589 }
 590
 591 #ifdef CONFIG_MEMORY_HOTPLUG
 592 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 593 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 594 {
 595         /* This will make the necessary allocations eventually. */
 596         return sparse_mem_map_populate(pnum, nid);
 597 }
 598 static void __kfree_section_memmap(struct page *memmap)
 599 {
 600         unsigned long start = (unsigned long)memmap;
 601         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 602
 603         vmemmap_free(start, end);
 604 }
 605 #ifdef CONFIG_MEMORY_HOTREMOVE
 606 static void free_map_bootmem(struct page *memmap)
 607 {
 608         unsigned long start = (unsigned long)memmap;
 609         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 610
 611         vmemmap_free(start, end);
 612 }
 613 #endif /* CONFIG_MEMORY_HOTREMOVE */
 614 #else
 615 static struct page *__kmalloc_section_memmap(void)
 616 {
 617         struct page *page, *ret;
 618         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 619
 620         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 621         if (page)
 622                 goto got_map_page;
 623
 624         ret = vmalloc(memmap_size);
 625         if (ret)
 626                 goto got_map_ptr;
 627
 628         return NULL;
 629 got_map_page:
 630         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 631 got_map_ptr:
 632
 633         return ret;
 634 }
 635
 636 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 637 {
 638         return __kmalloc_section_memmap();
 639 }
 640
 641 static void __kfree_section_memmap(struct page *memmap)
 642 {
 643         if (is_vmalloc_addr(memmap))
 644                 vfree(memmap);
 645         else
 646                 free_pages((unsigned long)memmap,
 647                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 648 }
 649
 650 #ifdef CONFIG_MEMORY_HOTREMOVE
 651 static void free_map_bootmem(struct page *memmap)
 652 {
 653         unsigned long maps_section_nr, removing_section_nr, i;
 654         unsigned long magic, nr_pages;
 655         struct page *page = virt_to_page(memmap);
 656
 657         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 658                 >> PAGE_SHIFT;
 659
 660         for (i = 0; i < nr_pages; i++, page++) {
 661                 magic = (unsigned long) page->lru.next;
 662
 663                 BUG_ON(magic == NODE_INFO);
 664
 665                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 666                 removing_section_nr = page->private;
 667
 668                 /*
 669                  * When this function is called, the removing section is
 670                  * logical offlined state. This means all pages are isolated
 671                  * from page allocator. If removing section's memmap is placed
 672                  * on the same section, it must not be freed.
 673                  * If it is freed, page allocator may allocate it which will
 674                  * be removed physically soon.
 675                  */
 676                 if (maps_section_nr != removing_section_nr)
 677                         put_page_bootmem(page);
 678         }
 679 }
 680 #endif /* CONFIG_MEMORY_HOTREMOVE */
 681 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 682
 683 /*
 684  * returns the number of sections whose mem_maps were properly
 685  * set.  If this is <=0, then that means that the passed-in
 686  * map was not consumed and must be freed.
 687  */
 688 int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn)
 689 {
 690         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 691         struct pglist_data *pgdat = zone->zone_pgdat;
 692         struct mem_section *ms;
 693         struct page *memmap;
 694         unsigned long *usemap;
 695         unsigned long flags;
 696         int ret;
 697
 698         /*
 699          * no locking for this, because it does its own
 700          * plus, it does a kmalloc
 701          */
 702         ret = sparse_index_init(section_nr, pgdat->node_id);
 703         if (ret < 0 && ret != -EEXIST)
 704                 return ret;
 705         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
 706         if (!memmap)
 707                 return -ENOMEM;
 708         usemap = __kmalloc_section_usemap();
 709         if (!usemap) {
 710                 __kfree_section_memmap(memmap);
 711                 return -ENOMEM;
 712         }
 713
 714         pgdat_resize_lock(pgdat, &flags);
 715
 716         ms = __pfn_to_section(start_pfn);
 717         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 718                 ret = -EEXIST;
 719                 goto out;
 720         }
 721
 722         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 723
 724         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 725
 726         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 727
 728 out:
 729         pgdat_resize_unlock(pgdat, &flags);
 730         if (ret <= 0) {
 731                 kfree(usemap);
 732                 __kfree_section_memmap(memmap);
 733         }
 734         return ret;
 735 }
 736
 737 #ifdef CONFIG_MEMORY_HOTREMOVE
 738 #ifdef CONFIG_MEMORY_FAILURE
 739 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 740 {
 741         int i;
 742
 743         if (!memmap)
 744                 return;
 745
 746         for (i = 0; i < PAGES_PER_SECTION; i++) {
 747                 if (PageHWPoison(&memmap[i])) {
 748                         atomic_long_sub(1, &num_poisoned_pages);
 749                         ClearPageHWPoison(&memmap[i]);
 750                 }
 751         }
 752 }
 753 #else
 754 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 755 {
 756 }
 757 #endif
 758
 759 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 760 {
 761         struct page *usemap_page;
 762
 763         if (!usemap)
 764                 return;
 765
 766         usemap_page = virt_to_page(usemap);
 767         /*
 768          * Check to see if allocation came from hot-plug-add
 769          */
 770         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 771                 kfree(usemap);
 772                 if (memmap)
 773                         __kfree_section_memmap(memmap);
 774                 return;
 775         }
 776
 777         /*
 778          * The usemap came from bootmem. This is packed with other usemaps
 779          * on the section which has pgdat at boot time. Just keep it as is now.
 780          */
 781
 782         if (memmap)
 783                 free_map_bootmem(memmap);
 784 }
 785
 786 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 787 {
 788         struct page *memmap = NULL;
 789         unsigned long *usemap = NULL, flags;
 790         struct pglist_data *pgdat = zone->zone_pgdat;
 791
 792         pgdat_resize_lock(pgdat, &flags);
 793         if (ms->section_mem_map) {
 794                 usemap = ms->pageblock_flags;
 795                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 796                                                 __section_nr(ms));
 797                 ms->section_mem_map = 0;
 798                 ms->pageblock_flags = NULL;
 799         }
 800         pgdat_resize_unlock(pgdat, &flags);
 801
 802         clear_hwpoisoned_pages(memmap, PAGES_PER_SECTION);
 803         free_section_usemap(memmap, usemap);
 804 }
 805 #endif /* CONFIG_MEMORY_HOTREMOVE */
 806 #endif /* CONFIG_MEMORY_HOTPLUG */