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