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