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                 if (!section)
  74                         panic("%s: Failed to allocate %lu bytes nid=%d\n",
  75                               __func__, array_size, nid);
  76         }
  77
  78         return section;
  79 }
  80
  81 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  82 {
  83         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  84         struct mem_section *section;
  85
  86         if (mem_section[root])
  87                 return -EEXIST;
  88
  89         section = sparse_index_alloc(nid);
  90         if (!section)
  91                 return -ENOMEM;
  92
  93         mem_section[root] = section;
  94
  95         return 0;
  96 }
  97 #else /* !SPARSEMEM_EXTREME */
  98 static inline int sparse_index_init(unsigned long section_nr, int nid)
  99 {
 100         return 0;
 101 }
 102 #endif
 103
 104 #ifdef CONFIG_SPARSEMEM_EXTREME
 105 unsigned long __section_nr(struct mem_section *ms)
 106 {
 107         unsigned long root_nr;
 108         struct mem_section *root = NULL;
 109
 110         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 111                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 112                 if (!root)
 113                         continue;
 114
 115                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 116                      break;
 117         }
 118
 119         VM_BUG_ON(!root);
 120
 121         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 122 }
 123 #else
 124 unsigned long __section_nr(struct mem_section *ms)
 125 {
 126         return (unsigned long)(ms - mem_section[0]);
 127 }
 128 #endif
 129
 130 /*
 131  * During early boot, before section_mem_map is used for an actual
 132  * mem_map, we use section_mem_map to store the section's NUMA
 133  * node.  This keeps us from having to use another data structure.  The
 134  * node information is cleared just before we store the real mem_map.
 135  */
 136 static inline unsigned long sparse_encode_early_nid(int nid)
 137 {
 138         return (nid << SECTION_NID_SHIFT);
 139 }
 140
 141 static inline int sparse_early_nid(struct mem_section *section)
 142 {
 143         return (section->section_mem_map >> SECTION_NID_SHIFT);
 144 }
 145
 146 /* Validate the physical addressing limitations of the model */
 147 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 148                                                 unsigned long *end_pfn)
 149 {
 150         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 151
 152         /*
 153          * Sanity checks - do not allow an architecture to pass
 154          * in larger pfns than the maximum scope of sparsemem:
 155          */
 156         if (*start_pfn > max_sparsemem_pfn) {
 157                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 158                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 159                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 160                 WARN_ON_ONCE(1);
 161                 *start_pfn = max_sparsemem_pfn;
 162                 *end_pfn = max_sparsemem_pfn;
 163         } else if (*end_pfn > max_sparsemem_pfn) {
 164                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 165                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 166                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 167                 WARN_ON_ONCE(1);
 168                 *end_pfn = max_sparsemem_pfn;
 169         }
 170 }
 171
 172 /*
 173  * There are a number of times that we loop over NR_MEM_SECTIONS,
 174  * looking for section_present() on each.  But, when we have very
 175  * large physical address spaces, NR_MEM_SECTIONS can also be
 176  * very large which makes the loops quite long.
 177  *
 178  * Keeping track of this gives us an easy way to break out of
 179  * those loops early.
 180  */
 181 unsigned long __highest_present_section_nr;
 182 static void section_mark_present(struct mem_section *ms)
 183 {
 184         unsigned long section_nr = __section_nr(ms);
 185
 186         if (section_nr > __highest_present_section_nr)
 187                 __highest_present_section_nr = section_nr;
 188
 189         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 190 }
 191
 192 static inline unsigned long next_present_section_nr(unsigned long section_nr)
 193 {
 194         do {
 195                 section_nr++;
 196                 if (present_section_nr(section_nr))
 197                         return section_nr;
 198         } while ((section_nr <= __highest_present_section_nr));
 199
 200         return -1;
 201 }
 202 #define for_each_present_section_nr(start, section_nr)          \
 203         for (section_nr = next_present_section_nr(start-1);     \
 204              ((section_nr != -1) &&                             \
 205               (section_nr <= __highest_present_section_nr));    \
 206              section_nr = next_present_section_nr(section_nr))
 207
 208 static inline unsigned long first_present_section_nr(void)
 209 {
 210         return next_present_section_nr(-1);
 211 }
 212
 213 void subsection_mask_set(unsigned long *map, unsigned long pfn,
 214                 unsigned long nr_pages)
 215 {
 216         int idx = subsection_map_index(pfn);
 217         int end = subsection_map_index(pfn + nr_pages - 1);
 218
 219         bitmap_set(map, idx, end - idx + 1);
 220 }
 221
 222 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 223 {
 224         int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
 225         int i, start_sec = pfn_to_section_nr(pfn);
 226
 227         if (!nr_pages)
 228                 return;
 229
 230         for (i = start_sec; i <= end_sec; i++) {
 231                 struct mem_section *ms;
 232                 unsigned long pfns;
 233
 234                 pfns = min(nr_pages, PAGES_PER_SECTION
 235                                 - (pfn & ~PAGE_SECTION_MASK));
 236                 ms = __nr_to_section(i);
 237                 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
 238
 239                 pr_debug("%s: sec: %d pfns: %ld set(%d, %d)\n", __func__, i,
 240                                 pfns, subsection_map_index(pfn),
 241                                 subsection_map_index(pfn + pfns - 1));
 242
 243                 pfn += pfns;
 244                 nr_pages -= pfns;
 245         }
 246 }
 247
 248 /* Record a memory area against a node. */
 249 void __init memory_present(int nid, unsigned long start, unsigned long end)
 250 {
 251         unsigned long pfn;
 252
 253 #ifdef CONFIG_SPARSEMEM_EXTREME
 254         if (unlikely(!mem_section)) {
 255                 unsigned long size, align;
 256
 257                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
 258                 align = 1 << (INTERNODE_CACHE_SHIFT);
 259                 mem_section = memblock_alloc(size, align);
 260                 if (!mem_section)
 261                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 262                               __func__, size, align);
 263         }
 264 #endif
 265
 266         start &= PAGE_SECTION_MASK;
 267         mminit_validate_memmodel_limits(&start, &end);
 268         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 269                 unsigned long section = pfn_to_section_nr(pfn);
 270                 struct mem_section *ms;
 271
 272                 sparse_index_init(section, nid);
 273                 set_section_nid(section, nid);
 274
 275                 ms = __nr_to_section(section);
 276                 if (!ms->section_mem_map) {
 277                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 278                                                         SECTION_IS_ONLINE;
 279                         section_mark_present(ms);
 280                 }
 281         }
 282 }
 283
 284 /*
 285  * Mark all memblocks as present using memory_present(). This is a
 286  * convienence function that is useful for a number of arches
 287  * to mark all of the systems memory as present during initialization.
 288  */
 289 void __init memblocks_present(void)
 290 {
 291         struct memblock_region *reg;
 292
 293         for_each_memblock(memory, reg) {
 294                 memory_present(memblock_get_region_node(reg),
 295                                memblock_region_memory_base_pfn(reg),
 296                                memblock_region_memory_end_pfn(reg));
 297         }
 298 }
 299
 300 /*
 301  * Subtle, we encode the real pfn into the mem_map such that
 302  * the identity pfn - section_mem_map will return the actual
 303  * physical page frame number.
 304  */
 305 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 306 {
 307         unsigned long coded_mem_map =
 308                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 309         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 310         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 311         return coded_mem_map;
 312 }
 313
 314 /*
 315  * Decode mem_map from the coded memmap
 316  */
 317 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 318 {
 319         /* mask off the extra low bits of information */
 320         coded_mem_map &= SECTION_MAP_MASK;
 321         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 322 }
 323
 324 static void __meminit sparse_init_one_section(struct mem_section *ms,
 325                 unsigned long pnum, struct page *mem_map,
 326                 struct mem_section_usage *usage, unsigned long flags)
 327 {
 328         ms->section_mem_map &= ~SECTION_MAP_MASK;
 329         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
 330                 | SECTION_HAS_MEM_MAP | flags;
 331         ms->usage = usage;
 332 }
 333
 334 static unsigned long usemap_size(void)
 335 {
 336         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 337 }
 338
 339 size_t mem_section_usage_size(void)
 340 {
 341         return sizeof(struct mem_section_usage) + usemap_size();
 342 }
 343
 344 #ifdef CONFIG_MEMORY_HOTREMOVE
 345 static struct mem_section_usage * __init
 346 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 347                                          unsigned long size)
 348 {
 349         struct mem_section_usage *usage;
 350         unsigned long goal, limit;
 351         int nid;
 352         /*
 353          * A page may contain usemaps for other sections preventing the
 354          * page being freed and making a section unremovable while
 355          * other sections referencing the usemap remain active. Similarly,
 356          * a pgdat can prevent a section being removed. If section A
 357          * contains a pgdat and section B contains the usemap, both
 358          * sections become inter-dependent. This allocates usemaps
 359          * from the same section as the pgdat where possible to avoid
 360          * this problem.
 361          */
 362         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 363         limit = goal + (1UL << PA_SECTION_SHIFT);
 364         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 365 again:
 366         usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 367         if (!usage && limit) {
 368                 limit = 0;
 369                 goto again;
 370         }
 371         return usage;
 372 }
 373
 374 static void __init check_usemap_section_nr(int nid,
 375                 struct mem_section_usage *usage)
 376 {
 377         unsigned long usemap_snr, pgdat_snr;
 378         static unsigned long old_usemap_snr;
 379         static unsigned long old_pgdat_snr;
 380         struct pglist_data *pgdat = NODE_DATA(nid);
 381         int usemap_nid;
 382
 383         /* First call */
 384         if (!old_usemap_snr) {
 385                 old_usemap_snr = NR_MEM_SECTIONS;
 386                 old_pgdat_snr = NR_MEM_SECTIONS;
 387         }
 388
 389         usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
 390         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 391         if (usemap_snr == pgdat_snr)
 392                 return;
 393
 394         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 395                 /* skip redundant message */
 396                 return;
 397
 398         old_usemap_snr = usemap_snr;
 399         old_pgdat_snr = pgdat_snr;
 400
 401         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 402         if (usemap_nid != nid) {
 403                 pr_info("node %d must be removed before remove section %ld\n",
 404                         nid, usemap_snr);
 405                 return;
 406         }
 407         /*
 408          * There is a circular dependency.
 409          * Some platforms allow un-removable section because they will just
 410          * gather other removable sections for dynamic partitioning.
 411          * Just notify un-removable section's number here.
 412          */
 413         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 414                 usemap_snr, pgdat_snr, nid);
 415 }
 416 #else
 417 static struct mem_section_usage * __init
 418 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 419                                          unsigned long size)
 420 {
 421         return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 422 }
 423
 424 static void __init check_usemap_section_nr(int nid,
 425                 struct mem_section_usage *usage)
 426 {
 427 }
 428 #endif /* CONFIG_MEMORY_HOTREMOVE */
 429
 430 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 431 static unsigned long __init section_map_size(void)
 432 {
 433         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 434 }
 435
 436 #else
 437 static unsigned long __init section_map_size(void)
 438 {
 439         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 440 }
 441
 442 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
 443                 struct vmem_altmap *altmap)
 444 {
 445         unsigned long size = section_map_size();
 446         struct page *map = sparse_buffer_alloc(size);
 447         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 448
 449         if (map)
 450                 return map;
 451
 452         map = memblock_alloc_try_nid(size,
 453                                           PAGE_SIZE, addr,
 454                                           MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 455         if (!map)
 456                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 457                       __func__, size, PAGE_SIZE, nid, &addr);
 458
 459         return map;
 460 }
 461 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 462
 463 static void *sparsemap_buf __meminitdata;
 464 static void *sparsemap_buf_end __meminitdata;
 465
 466 static void __init sparse_buffer_init(unsigned long size, int nid)
 467 {
 468         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 469         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 470         sparsemap_buf =
 471                 memblock_alloc_try_nid_raw(size, PAGE_SIZE,
 472                                                 addr,
 473                                                 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 474         sparsemap_buf_end = sparsemap_buf + size;
 475 }
 476
 477 static void __init sparse_buffer_fini(void)
 478 {
 479         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 480
 481         if (sparsemap_buf && size > 0)
 482                 memblock_free_early(__pa(sparsemap_buf), size);
 483         sparsemap_buf = NULL;
 484 }
 485
 486 void * __meminit sparse_buffer_alloc(unsigned long size)
 487 {
 488         void *ptr = NULL;
 489
 490         if (sparsemap_buf) {
 491                 ptr = PTR_ALIGN(sparsemap_buf, size);
 492                 if (ptr + size > sparsemap_buf_end)
 493                         ptr = NULL;
 494                 else
 495                         sparsemap_buf = ptr + size;
 496         }
 497         return ptr;
 498 }
 499
 500 void __weak __meminit vmemmap_populate_print_last(void)
 501 {
 502 }
 503
 504 /*
 505  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 506  * And number of present sections in this node is map_count.
 507  */
 508 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 509                                    unsigned long pnum_end,
 510                                    unsigned long map_count)
 511 {
 512         struct mem_section_usage *usage;
 513         unsigned long pnum;
 514         struct page *map;
 515
 516         usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 517                         mem_section_usage_size() * map_count);
 518         if (!usage) {
 519                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 520                 goto failed;
 521         }
 522         sparse_buffer_init(map_count * section_map_size(), nid);
 523         for_each_present_section_nr(pnum_begin, pnum) {
 524                 if (pnum >= pnum_end)
 525                         break;
 526
 527                 map = sparse_mem_map_populate(pnum, nid, NULL);
 528                 if (!map) {
 529                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 530                                __func__, nid);
 531                         pnum_begin = pnum;
 532                         goto failed;
 533                 }
 534                 check_usemap_section_nr(nid, usage);
 535                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
 536                                 SECTION_IS_EARLY);
 537                 usage = (void *) usage + mem_section_usage_size();
 538         }
 539         sparse_buffer_fini();
 540         return;
 541 failed:
 542         /* We failed to allocate, mark all the following pnums as not present */
 543         for_each_present_section_nr(pnum_begin, pnum) {
 544                 struct mem_section *ms;
 545
 546                 if (pnum >= pnum_end)
 547                         break;
 548                 ms = __nr_to_section(pnum);
 549                 ms->section_mem_map = 0;
 550         }
 551 }
 552
 553 /*
 554  * Allocate the accumulated non-linear sections, allocate a mem_map
 555  * for each and record the physical to section mapping.
 556  */
 557 void __init sparse_init(void)
 558 {
 559         unsigned long pnum_begin = first_present_section_nr();
 560         int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 561         unsigned long pnum_end, map_count = 1;
 562
 563         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 564         set_pageblock_order();
 565
 566         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 567                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 568
 569                 if (nid == nid_begin) {
 570                         map_count++;
 571                         continue;
 572                 }
 573                 /* Init node with sections in range [pnum_begin, pnum_end) */
 574                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 575                 nid_begin = nid;
 576                 pnum_begin = pnum_end;
 577                 map_count = 1;
 578         }
 579         /* cover the last node */
 580         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 581         vmemmap_populate_print_last();
 582 }
 583
 584 #ifdef CONFIG_MEMORY_HOTPLUG
 585
 586 /* Mark all memory sections within the pfn range as online */
 587 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 588 {
 589         unsigned long pfn;
 590
 591         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 592                 unsigned long section_nr = pfn_to_section_nr(pfn);
 593                 struct mem_section *ms;
 594
 595                 /* onlining code should never touch invalid ranges */
 596                 if (WARN_ON(!valid_section_nr(section_nr)))
 597                         continue;
 598
 599                 ms = __nr_to_section(section_nr);
 600                 ms->section_mem_map |= SECTION_IS_ONLINE;
 601         }
 602 }
 603
 604 #ifdef CONFIG_MEMORY_HOTREMOVE
 605 /* Mark all memory sections within the pfn range as offline */
 606 void offline_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                 /*
 615                  * TODO this needs some double checking. Offlining code makes
 616                  * sure to check pfn_valid but those checks might be just bogus
 617                  */
 618                 if (WARN_ON(!valid_section_nr(section_nr)))
 619                         continue;
 620
 621                 ms = __nr_to_section(section_nr);
 622                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 623         }
 624 }
 625 #endif
 626
 627 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 628 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 629                 struct vmem_altmap *altmap)
 630 {
 631         /* This will make the necessary allocations eventually. */
 632         return sparse_mem_map_populate(pnum, nid, altmap);
 633 }
 634 static void __kfree_section_memmap(struct page *memmap,
 635                 struct vmem_altmap *altmap)
 636 {
 637         unsigned long start = (unsigned long)memmap;
 638         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 639
 640         vmemmap_free(start, end, altmap);
 641 }
 642 static void free_map_bootmem(struct page *memmap)
 643 {
 644         unsigned long start = (unsigned long)memmap;
 645         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 646
 647         vmemmap_free(start, end, NULL);
 648 }
 649 #else
 650 static struct page *__kmalloc_section_memmap(void)
 651 {
 652         struct page *page, *ret;
 653         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 654
 655         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 656         if (page)
 657                 goto got_map_page;
 658
 659         ret = vmalloc(memmap_size);
 660         if (ret)
 661                 goto got_map_ptr;
 662
 663         return NULL;
 664 got_map_page:
 665         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 666 got_map_ptr:
 667
 668         return ret;
 669 }
 670
 671 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 672                 struct vmem_altmap *altmap)
 673 {
 674         return __kmalloc_section_memmap();
 675 }
 676
 677 static void __kfree_section_memmap(struct page *memmap,
 678                 struct vmem_altmap *altmap)
 679 {
 680         if (is_vmalloc_addr(memmap))
 681                 vfree(memmap);
 682         else
 683                 free_pages((unsigned long)memmap,
 684                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 685 }
 686
 687 static void free_map_bootmem(struct page *memmap)
 688 {
 689         unsigned long maps_section_nr, removing_section_nr, i;
 690         unsigned long magic, nr_pages;
 691         struct page *page = virt_to_page(memmap);
 692
 693         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 694                 >> PAGE_SHIFT;
 695
 696         for (i = 0; i < nr_pages; i++, page++) {
 697                 magic = (unsigned long) page->freelist;
 698
 699                 BUG_ON(magic == NODE_INFO);
 700
 701                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 702                 removing_section_nr = page_private(page);
 703
 704                 /*
 705                  * When this function is called, the removing section is
 706                  * logical offlined state. This means all pages are isolated
 707                  * from page allocator. If removing section's memmap is placed
 708                  * on the same section, it must not be freed.
 709                  * If it is freed, page allocator may allocate it which will
 710                  * be removed physically soon.
 711                  */
 712                 if (maps_section_nr != removing_section_nr)
 713                         put_page_bootmem(page);
 714         }
 715 }
 716 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 717
 718 /**
 719  * sparse_add_one_section - add a memory section
 720  * @nid: The node to add section on
 721  * @start_pfn: start pfn of the memory range
 722  * @altmap: device page map
 723  *
 724  * This is only intended for hotplug.
 725  *
 726  * Return:
 727  * * 0          - On success.
 728  * * -EEXIST    - Section has been present.
 729  * * -ENOMEM    - Out of memory.
 730  */
 731 int __meminit sparse_add_one_section(int nid, unsigned long start_pfn,
 732                                      struct vmem_altmap *altmap)
 733 {
 734         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 735         struct mem_section_usage *usage;
 736         struct mem_section *ms;
 737         struct page *memmap;
 738         int ret;
 739
 740         /*
 741          * no locking for this, because it does its own
 742          * plus, it does a kmalloc
 743          */
 744         ret = sparse_index_init(section_nr, nid);
 745         if (ret < 0 && ret != -EEXIST)
 746                 return ret;
 747         ret = 0;
 748         memmap = kmalloc_section_memmap(section_nr, nid, altmap);
 749         if (!memmap)
 750                 return -ENOMEM;
 751         usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
 752         if (!usage) {
 753                 __kfree_section_memmap(memmap, altmap);
 754                 return -ENOMEM;
 755         }
 756
 757         ms = __pfn_to_section(start_pfn);
 758         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 759                 ret = -EEXIST;
 760                 goto out;
 761         }
 762
 763         /*
 764          * Poison uninitialized struct pages in order to catch invalid flags
 765          * combinations.
 766          */
 767         page_init_poison(memmap, sizeof(struct page) * PAGES_PER_SECTION);
 768
 769         set_section_nid(section_nr, nid);
 770         section_mark_present(ms);
 771         sparse_init_one_section(ms, section_nr, memmap, usage, 0);
 772
 773 out:
 774         if (ret < 0) {
 775                 kfree(usage);
 776                 __kfree_section_memmap(memmap, altmap);
 777         }
 778         return ret;
 779 }
 780
 781 #ifdef CONFIG_MEMORY_FAILURE
 782 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 783 {
 784         int i;
 785
 786         if (!memmap)
 787                 return;
 788
 789         /*
 790          * A further optimization is to have per section refcounted
 791          * num_poisoned_pages.  But that would need more space per memmap, so
 792          * for now just do a quick global check to speed up this routine in the
 793          * absence of bad pages.
 794          */
 795         if (atomic_long_read(&num_poisoned_pages) == 0)
 796                 return;
 797
 798         for (i = 0; i < nr_pages; i++) {
 799                 if (PageHWPoison(&memmap[i])) {
 800                         atomic_long_sub(1, &num_poisoned_pages);
 801                         ClearPageHWPoison(&memmap[i]);
 802                 }
 803         }
 804 }
 805 #else
 806 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 807 {
 808 }
 809 #endif
 810
 811 static void free_section_usage(struct mem_section *ms, struct page *memmap,
 812                 struct mem_section_usage *usage, struct vmem_altmap *altmap)
 813 {
 814         if (!usage)
 815                 return;
 816
 817         /*
 818          * Check to see if allocation came from hot-plug-add
 819          */
 820         if (!early_section(ms)) {
 821                 kfree(usage);
 822                 if (memmap)
 823                         __kfree_section_memmap(memmap, altmap);
 824                 return;
 825         }
 826
 827         /*
 828          * The usemap came from bootmem. This is packed with other usemaps
 829          * on the section which has pgdat at boot time. Just keep it as is now.
 830          */
 831
 832         if (memmap)
 833                 free_map_bootmem(memmap);
 834 }
 835
 836 void sparse_remove_one_section(struct mem_section *ms, unsigned long map_offset,
 837                                struct vmem_altmap *altmap)
 838 {
 839         struct page *memmap = NULL;
 840         struct mem_section_usage *usage = NULL;
 841
 842         if (ms->section_mem_map) {
 843                 usage = ms->usage;
 844                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 845                                                 __section_nr(ms));
 846                 ms->section_mem_map = 0;
 847                 ms->usage = NULL;
 848         }
 849
 850         clear_hwpoisoned_pages(memmap + map_offset,
 851                         PAGES_PER_SECTION - map_offset);
 852         free_section_usage(ms, memmap, usage, altmap);
 853 }
 854 #endif /* CONFIG_MEMORY_HOTPLUG */