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