mm/sparse.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * sparse memory mappings.
   4  */
   5 #include <linux/mm.h>
   6 #include <linux/slab.h>
   7 #include <linux/mmzone.h>
   8 #include <linux/bootmem.h>
   9 #include <linux/compiler.h>
  10 #include <linux/highmem.h>
  11 #include <linux/export.h>
  12 #include <linux/spinlock.h>
  13 #include <linux/vmalloc.h>
  14
  15 #include "internal.h"
  16 #include <asm/dma.h>
  17 #include <asm/pgalloc.h>
  18 #include <asm/pgtable.h>
  19
  20 /*
  21  * Permanent SPARSEMEM data:
  22  *
  23  * 1) mem_section       - memory sections, mem_map's for valid memory
  24  */
  25 #ifdef CONFIG_SPARSEMEM_EXTREME
  26 struct mem_section **mem_section;
  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 = NULL;
 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);
 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;
 333         static unsigned long old_pgdat_snr;
 334         struct pglist_data *pgdat = NODE_DATA(nid);
 335         int usemap_nid;
 336
 337         /* First call */
 338         if (!old_usemap_snr) {
 339                 old_usemap_snr = NR_MEM_SECTIONS;
 340                 old_pgdat_snr = NR_MEM_SECTIONS;
 341         }
 342
 343         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 344         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 345         if (usemap_snr == pgdat_snr)
 346                 return;
 347
 348         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 349                 /* skip redundant message */
 350                 return;
 351
 352         old_usemap_snr = usemap_snr;
 353         old_pgdat_snr = pgdat_snr;
 354
 355         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 356         if (usemap_nid != nid) {
 357                 pr_info("node %d must be removed before remove section %ld\n",
 358                         nid, usemap_snr);
 359                 return;
 360         }
 361         /*
 362          * There is a circular dependency.
 363          * Some platforms allow un-removable section because they will just
 364          * gather other removable sections for dynamic partitioning.
 365          * Just notify un-removable section's number here.
 366          */
 367         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 368                 usemap_snr, pgdat_snr, nid);
 369 }
 370 #else
 371 static unsigned long * __init
 372 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 373                                          unsigned long size)
 374 {
 375         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 376 }
 377
 378 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 379 {
 380 }
 381 #endif /* CONFIG_MEMORY_HOTREMOVE */
 382
 383 static void __init sparse_early_usemaps_alloc_node(void *data,
 384                                  unsigned long pnum_begin,
 385                                  unsigned long pnum_end,
 386                                  unsigned long usemap_count, int nodeid)
 387 {
 388         void *usemap;
 389         unsigned long pnum;
 390         unsigned long **usemap_map = (unsigned long **)data;
 391         int size = usemap_size();
 392
 393         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 394                                                           size * usemap_count);
 395         if (!usemap) {
 396                 pr_warn("%s: allocation failed\n", __func__);
 397                 return;
 398         }
 399
 400         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 401                 if (!present_section_nr(pnum))
 402                         continue;
 403                 usemap_map[pnum] = usemap;
 404                 usemap += size;
 405                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 406         }
 407 }
 408
 409 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 410 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 411 {
 412         struct page *map;
 413         unsigned long size;
 414
 415         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 416         if (map)
 417                 return map;
 418
 419         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 420         map = memblock_virt_alloc_try_nid(size,
 421                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 422                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 423         return map;
 424 }
 425 void __init sparse_mem_maps_populate_node(struct page **map_map,
 426                                           unsigned long pnum_begin,
 427                                           unsigned long pnum_end,
 428                                           unsigned long map_count, int nodeid)
 429 {
 430         void *map;
 431         unsigned long pnum;
 432         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 433
 434         map = alloc_remap(nodeid, size * map_count);
 435         if (map) {
 436                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 437                         if (!present_section_nr(pnum))
 438                                 continue;
 439                         map_map[pnum] = map;
 440                         map += size;
 441                 }
 442                 return;
 443         }
 444
 445         size = PAGE_ALIGN(size);
 446         map = memblock_virt_alloc_try_nid(size * map_count,
 447                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 448                                           BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 449         if (map) {
 450                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 451                         if (!present_section_nr(pnum))
 452                                 continue;
 453                         map_map[pnum] = map;
 454                         map += size;
 455                 }
 456                 return;
 457         }
 458
 459         /* fallback */
 460         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 461                 struct mem_section *ms;
 462
 463                 if (!present_section_nr(pnum))
 464                         continue;
 465                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 466                 if (map_map[pnum])
 467                         continue;
 468                 ms = __nr_to_section(pnum);
 469                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 470                        __func__);
 471                 ms->section_mem_map = 0;
 472         }
 473 }
 474 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 475
 476 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 477 static void __init sparse_early_mem_maps_alloc_node(void *data,
 478                                  unsigned long pnum_begin,
 479                                  unsigned long pnum_end,
 480                                  unsigned long map_count, int nodeid)
 481 {
 482         struct page **map_map = (struct page **)data;
 483         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 484                                          map_count, nodeid);
 485 }
 486 #else
 487 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 488 {
 489         struct page *map;
 490         struct mem_section *ms = __nr_to_section(pnum);
 491         int nid = sparse_early_nid(ms);
 492
 493         map = sparse_mem_map_populate(pnum, nid);
 494         if (map)
 495                 return map;
 496
 497         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 498                __func__);
 499         ms->section_mem_map = 0;
 500         return NULL;
 501 }
 502 #endif
 503
 504 void __weak __meminit vmemmap_populate_print_last(void)
 505 {
 506 }
 507
 508 /**
 509  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 510  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 511  */
 512 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 513                                         (void *, unsigned long, unsigned long,
 514                                         unsigned long, int), void *data)
 515 {
 516         unsigned long pnum;
 517         unsigned long map_count;
 518         int nodeid_begin = 0;
 519         unsigned long pnum_begin = 0;
 520
 521         for_each_present_section_nr(0, pnum) {
 522                 struct mem_section *ms;
 523
 524                 ms = __nr_to_section(pnum);
 525                 nodeid_begin = sparse_early_nid(ms);
 526                 pnum_begin = pnum;
 527                 break;
 528         }
 529         map_count = 1;
 530         for_each_present_section_nr(pnum_begin + 1, pnum) {
 531                 struct mem_section *ms;
 532                 int nodeid;
 533
 534                 ms = __nr_to_section(pnum);
 535                 nodeid = sparse_early_nid(ms);
 536                 if (nodeid == nodeid_begin) {
 537                         map_count++;
 538                         continue;
 539                 }
 540                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 541                 alloc_func(data, pnum_begin, pnum,
 542                                                 map_count, nodeid_begin);
 543                 /* new start, update count etc*/
 544                 nodeid_begin = nodeid;
 545                 pnum_begin = pnum;
 546                 map_count = 1;
 547         }
 548         /* ok, last chunk */
 549         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
 550                                                 map_count, nodeid_begin);
 551 }
 552
 553 /*
 554  * Allocate the accumulated non-linear sections, allocate a mem_map
 555  * for each and record the physical to section mapping.
 556  */
 557 void __init sparse_init(void)
 558 {
 559         unsigned long pnum;
 560         struct page *map;
 561         unsigned long *usemap;
 562         unsigned long **usemap_map;
 563         int size;
 564 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 565         int size2;
 566         struct page **map_map;
 567 #endif
 568
 569         /* see include/linux/mmzone.h 'struct mem_section' definition */
 570         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 571
 572         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 573         set_pageblock_order();
 574
 575         /*
 576          * map is using big page (aka 2M in x86 64 bit)
 577          * usemap is less one page (aka 24 bytes)
 578          * so alloc 2M (with 2M align) and 24 bytes in turn will
 579          * make next 2M slip to one more 2M later.
 580          * then in big system, the memory will have a lot of holes...
 581          * here try to allocate 2M pages continuously.
 582          *
 583          * powerpc need to call sparse_init_one_section right after each
 584          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 585          */
 586         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 587         usemap_map = memblock_virt_alloc(size, 0);
 588         if (!usemap_map)
 589                 panic("can not allocate usemap_map\n");
 590         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 591                                                         (void *)usemap_map);
 592
 593 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 594         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 595         map_map = memblock_virt_alloc(size2, 0);
 596         if (!map_map)
 597                 panic("can not allocate map_map\n");
 598         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 599                                                         (void *)map_map);
 600 #endif
 601
 602         for_each_present_section_nr(0, pnum) {
 603                 usemap = usemap_map[pnum];
 604                 if (!usemap)
 605                         continue;
 606
 607 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 608                 map = map_map[pnum];
 609 #else
 610                 map = sparse_early_mem_map_alloc(pnum);
 611 #endif
 612                 if (!map)
 613                         continue;
 614
 615                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 616                                                                 usemap);
 617         }
 618
 619         vmemmap_populate_print_last();
 620
 621 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 622         memblock_free_early(__pa(map_map), size2);
 623 #endif
 624         memblock_free_early(__pa(usemap_map), size);
 625 }
 626
 627 #ifdef CONFIG_MEMORY_HOTPLUG
 628
 629 /* Mark all memory sections within the pfn range as online */
 630 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 631 {
 632         unsigned long pfn;
 633
 634         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 635                 unsigned long section_nr = pfn_to_section_nr(pfn);
 636                 struct mem_section *ms;
 637
 638                 /* onlining code should never touch invalid ranges */
 639                 if (WARN_ON(!valid_section_nr(section_nr)))
 640                         continue;
 641
 642                 ms = __nr_to_section(section_nr);
 643                 ms->section_mem_map |= SECTION_IS_ONLINE;
 644         }
 645 }
 646
 647 #ifdef CONFIG_MEMORY_HOTREMOVE
 648 /* Mark all memory sections within the pfn range as online */
 649 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 650 {
 651         unsigned long pfn;
 652
 653         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 654                 unsigned long section_nr = pfn_to_section_nr(start_pfn);
 655                 struct mem_section *ms;
 656
 657                 /*
 658                  * TODO this needs some double checking. Offlining code makes
 659                  * sure to check pfn_valid but those checks might be just bogus
 660                  */
 661                 if (WARN_ON(!valid_section_nr(section_nr)))
 662                         continue;
 663
 664                 ms = __nr_to_section(section_nr);
 665                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 666         }
 667 }
 668 #endif
 669
 670 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 671 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 672 {
 673         /* This will make the necessary allocations eventually. */
 674         return sparse_mem_map_populate(pnum, nid);
 675 }
 676 static void __kfree_section_memmap(struct page *memmap)
 677 {
 678         unsigned long start = (unsigned long)memmap;
 679         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 680
 681         vmemmap_free(start, end);
 682 }
 683 #ifdef CONFIG_MEMORY_HOTREMOVE
 684 static void free_map_bootmem(struct page *memmap)
 685 {
 686         unsigned long start = (unsigned long)memmap;
 687         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 688
 689         vmemmap_free(start, end);
 690 }
 691 #endif /* CONFIG_MEMORY_HOTREMOVE */
 692 #else
 693 static struct page *__kmalloc_section_memmap(void)
 694 {
 695         struct page *page, *ret;
 696         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 697
 698         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 699         if (page)
 700                 goto got_map_page;
 701
 702         ret = vmalloc(memmap_size);
 703         if (ret)
 704                 goto got_map_ptr;
 705
 706         return NULL;
 707 got_map_page:
 708         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 709 got_map_ptr:
 710
 711         return ret;
 712 }
 713
 714 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 715 {
 716         return __kmalloc_section_memmap();
 717 }
 718
 719 static void __kfree_section_memmap(struct page *memmap)
 720 {
 721         if (is_vmalloc_addr(memmap))
 722                 vfree(memmap);
 723         else
 724                 free_pages((unsigned long)memmap,
 725                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 726 }
 727
 728 #ifdef CONFIG_MEMORY_HOTREMOVE
 729 static void free_map_bootmem(struct page *memmap)
 730 {
 731         unsigned long maps_section_nr, removing_section_nr, i;
 732         unsigned long magic, nr_pages;
 733         struct page *page = virt_to_page(memmap);
 734
 735         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 736                 >> PAGE_SHIFT;
 737
 738         for (i = 0; i < nr_pages; i++, page++) {
 739                 magic = (unsigned long) page->freelist;
 740
 741                 BUG_ON(magic == NODE_INFO);
 742
 743                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 744                 removing_section_nr = page_private(page);
 745
 746                 /*
 747                  * When this function is called, the removing section is
 748                  * logical offlined state. This means all pages are isolated
 749                  * from page allocator. If removing section's memmap is placed
 750                  * on the same section, it must not be freed.
 751                  * If it is freed, page allocator may allocate it which will
 752                  * be removed physically soon.
 753                  */
 754                 if (maps_section_nr != removing_section_nr)
 755                         put_page_bootmem(page);
 756         }
 757 }
 758 #endif /* CONFIG_MEMORY_HOTREMOVE */
 759 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 760
 761 /*
 762  * returns the number of sections whose mem_maps were properly
 763  * set.  If this is <=0, then that means that the passed-in
 764  * map was not consumed and must be freed.
 765  */
 766 int __meminit sparse_add_one_section(struct pglist_data *pgdat, unsigned long start_pfn)
 767 {
 768         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 769         struct mem_section *ms;
 770         struct page *memmap;
 771         unsigned long *usemap;
 772         unsigned long flags;
 773         int ret;
 774
 775         /*
 776          * no locking for this, because it does its own
 777          * plus, it does a kmalloc
 778          */
 779         ret = sparse_index_init(section_nr, pgdat->node_id);
 780         if (ret < 0 && ret != -EEXIST)
 781                 return ret;
 782         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
 783         if (!memmap)
 784                 return -ENOMEM;
 785         usemap = __kmalloc_section_usemap();
 786         if (!usemap) {
 787                 __kfree_section_memmap(memmap);
 788                 return -ENOMEM;
 789         }
 790
 791         pgdat_resize_lock(pgdat, &flags);
 792
 793         ms = __pfn_to_section(start_pfn);
 794         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 795                 ret = -EEXIST;
 796                 goto out;
 797         }
 798
 799         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 800
 801         section_mark_present(ms);
 802
 803         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 804
 805 out:
 806         pgdat_resize_unlock(pgdat, &flags);
 807         if (ret <= 0) {
 808                 kfree(usemap);
 809                 __kfree_section_memmap(memmap);
 810         }
 811         return ret;
 812 }
 813
 814 #ifdef CONFIG_MEMORY_HOTREMOVE
 815 #ifdef CONFIG_MEMORY_FAILURE
 816 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 817 {
 818         int i;
 819
 820         if (!memmap)
 821                 return;
 822
 823         for (i = 0; i < nr_pages; i++) {
 824                 if (PageHWPoison(&memmap[i])) {
 825                         atomic_long_sub(1, &num_poisoned_pages);
 826                         ClearPageHWPoison(&memmap[i]);
 827                 }
 828         }
 829 }
 830 #else
 831 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 832 {
 833 }
 834 #endif
 835
 836 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 837 {
 838         struct page *usemap_page;
 839
 840         if (!usemap)
 841                 return;
 842
 843         usemap_page = virt_to_page(usemap);
 844         /*
 845          * Check to see if allocation came from hot-plug-add
 846          */
 847         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 848                 kfree(usemap);
 849                 if (memmap)
 850                         __kfree_section_memmap(memmap);
 851                 return;
 852         }
 853
 854         /*
 855          * The usemap came from bootmem. This is packed with other usemaps
 856          * on the section which has pgdat at boot time. Just keep it as is now.
 857          */
 858
 859         if (memmap)
 860                 free_map_bootmem(memmap);
 861 }
 862
 863 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 864                 unsigned long map_offset)
 865 {
 866         struct page *memmap = NULL;
 867         unsigned long *usemap = NULL, flags;
 868         struct pglist_data *pgdat = zone->zone_pgdat;
 869
 870         pgdat_resize_lock(pgdat, &flags);
 871         if (ms->section_mem_map) {
 872                 usemap = ms->pageblock_flags;
 873                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 874                                                 __section_nr(ms));
 875                 ms->section_mem_map = 0;
 876                 ms->pageblock_flags = NULL;
 877         }
 878         pgdat_resize_unlock(pgdat, &flags);
 879
 880         clear_hwpoisoned_pages(memmap + map_offset,
 881                         PAGES_PER_SECTION - map_offset);
 882         free_section_usemap(memmap, usemap);
 883 }
 884 #endif /* CONFIG_MEMORY_HOTREMOVE */
 885 #endif /* CONFIG_MEMORY_HOTPLUG */