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 #include <linux/swap.h>
  15 #include <linux/swapops.h>
  16
  17 #include "internal.h"
  18 #include <asm/dma.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         /*
  87          * An existing section is possible in the sub-section hotplug
  88          * case. First hot-add instantiates, follow-on hot-add reuses
  89          * the existing section.
  90          *
  91          * The mem_hotplug_lock resolves the apparent race below.
  92          */
  93         if (mem_section[root])
  94                 return 0;
  95
  96         section = sparse_index_alloc(nid);
  97         if (!section)
  98                 return -ENOMEM;
  99
 100         mem_section[root] = section;
 101
 102         return 0;
 103 }
 104 #else /* !SPARSEMEM_EXTREME */
 105 static inline int sparse_index_init(unsigned long section_nr, int nid)
 106 {
 107         return 0;
 108 }
 109 #endif
 110
 111 #ifdef CONFIG_SPARSEMEM_EXTREME
 112 unsigned long __section_nr(struct mem_section *ms)
 113 {
 114         unsigned long root_nr;
 115         struct mem_section *root = NULL;
 116
 117         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 118                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 119                 if (!root)
 120                         continue;
 121
 122                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 123                      break;
 124         }
 125
 126         VM_BUG_ON(!root);
 127
 128         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 129 }
 130 #else
 131 unsigned long __section_nr(struct mem_section *ms)
 132 {
 133         return (unsigned long)(ms - mem_section[0]);
 134 }
 135 #endif
 136
 137 /*
 138  * During early boot, before section_mem_map is used for an actual
 139  * mem_map, we use section_mem_map to store the section's NUMA
 140  * node.  This keeps us from having to use another data structure.  The
 141  * node information is cleared just before we store the real mem_map.
 142  */
 143 static inline unsigned long sparse_encode_early_nid(int nid)
 144 {
 145         return (nid << SECTION_NID_SHIFT);
 146 }
 147
 148 static inline int sparse_early_nid(struct mem_section *section)
 149 {
 150         return (section->section_mem_map >> SECTION_NID_SHIFT);
 151 }
 152
 153 /* Validate the physical addressing limitations of the model */
 154 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 155                                                 unsigned long *end_pfn)
 156 {
 157         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 158
 159         /*
 160          * Sanity checks - do not allow an architecture to pass
 161          * in larger pfns than the maximum scope of sparsemem:
 162          */
 163         if (*start_pfn > max_sparsemem_pfn) {
 164                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 165                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 166                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 167                 WARN_ON_ONCE(1);
 168                 *start_pfn = max_sparsemem_pfn;
 169                 *end_pfn = max_sparsemem_pfn;
 170         } else if (*end_pfn > max_sparsemem_pfn) {
 171                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 172                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 173                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 174                 WARN_ON_ONCE(1);
 175                 *end_pfn = max_sparsemem_pfn;
 176         }
 177 }
 178
 179 /*
 180  * There are a number of times that we loop over NR_MEM_SECTIONS,
 181  * looking for section_present() on each.  But, when we have very
 182  * large physical address spaces, NR_MEM_SECTIONS can also be
 183  * very large which makes the loops quite long.
 184  *
 185  * Keeping track of this gives us an easy way to break out of
 186  * those loops early.
 187  */
 188 unsigned long __highest_present_section_nr;
 189 static void section_mark_present(struct mem_section *ms)
 190 {
 191         unsigned long section_nr = __section_nr(ms);
 192
 193         if (section_nr > __highest_present_section_nr)
 194                 __highest_present_section_nr = section_nr;
 195
 196         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 197 }
 198
 199 #define for_each_present_section_nr(start, section_nr)          \
 200         for (section_nr = next_present_section_nr(start-1);     \
 201              ((section_nr != -1) &&                             \
 202               (section_nr <= __highest_present_section_nr));    \
 203              section_nr = next_present_section_nr(section_nr))
 204
 205 static inline unsigned long first_present_section_nr(void)
 206 {
 207         return next_present_section_nr(-1);
 208 }
 209
 210 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 211 static void subsection_mask_set(unsigned long *map, unsigned long pfn,
 212                 unsigned long nr_pages)
 213 {
 214         int idx = subsection_map_index(pfn);
 215         int end = subsection_map_index(pfn + nr_pages - 1);
 216
 217         bitmap_set(map, idx, end - idx + 1);
 218 }
 219
 220 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 221 {
 222         int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
 223         unsigned long nr, start_sec = pfn_to_section_nr(pfn);
 224
 225         if (!nr_pages)
 226                 return;
 227
 228         for (nr = start_sec; nr <= end_sec; nr++) {
 229                 struct mem_section *ms;
 230                 unsigned long pfns;
 231
 232                 pfns = min(nr_pages, PAGES_PER_SECTION
 233                                 - (pfn & ~PAGE_SECTION_MASK));
 234                 ms = __nr_to_section(nr);
 235                 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
 236
 237                 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
 238                                 pfns, subsection_map_index(pfn),
 239                                 subsection_map_index(pfn + pfns - 1));
 240
 241                 pfn += pfns;
 242                 nr_pages -= pfns;
 243         }
 244 }
 245 #else
 246 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 247 {
 248 }
 249 #endif
 250
 251 /* Record a memory area against a node. */
 252 static void __init memory_present(int nid, unsigned long start, unsigned long end)
 253 {
 254         unsigned long pfn;
 255
 256 #ifdef CONFIG_SPARSEMEM_EXTREME
 257         if (unlikely(!mem_section)) {
 258                 unsigned long size, align;
 259
 260                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
 261                 align = 1 << (INTERNODE_CACHE_SHIFT);
 262                 mem_section = memblock_alloc(size, align);
 263                 if (!mem_section)
 264                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 265                               __func__, size, align);
 266         }
 267 #endif
 268
 269         start &= PAGE_SECTION_MASK;
 270         mminit_validate_memmodel_limits(&start, &end);
 271         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 272                 unsigned long section = pfn_to_section_nr(pfn);
 273                 struct mem_section *ms;
 274
 275                 sparse_index_init(section, nid);
 276                 set_section_nid(section, nid);
 277
 278                 ms = __nr_to_section(section);
 279                 if (!ms->section_mem_map) {
 280                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 281                                                         SECTION_IS_ONLINE;
 282                         section_mark_present(ms);
 283                 }
 284         }
 285 }
 286
 287 /*
 288  * Mark all memblocks as present using memory_present().
 289  * This is a convenience function that is useful to mark all of the systems
 290  * memory as present during initialization.
 291  */
 292 static void __init memblocks_present(void)
 293 {
 294         unsigned long start, end;
 295         int i, nid;
 296
 297         for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
 298                 memory_present(nid, start, end);
 299 }
 300
 301 /*
 302  * Subtle, we encode the real pfn into the mem_map such that
 303  * the identity pfn - section_mem_map will return the actual
 304  * physical page frame number.
 305  */
 306 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 307 {
 308         unsigned long coded_mem_map =
 309                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 310         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 311         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 312         return coded_mem_map;
 313 }
 314
 315 /*
 316  * Decode mem_map from the coded memmap
 317  */
 318 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 319 {
 320         /* mask off the extra low bits of information */
 321         coded_mem_map &= SECTION_MAP_MASK;
 322         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 323 }
 324
 325 static void __meminit sparse_init_one_section(struct mem_section *ms,
 326                 unsigned long pnum, struct page *mem_map,
 327                 struct mem_section_usage *usage, unsigned long flags)
 328 {
 329         ms->section_mem_map &= ~SECTION_MAP_MASK;
 330         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
 331                 | SECTION_HAS_MEM_MAP | flags;
 332         ms->usage = usage;
 333 }
 334
 335 static unsigned long usemap_size(void)
 336 {
 337         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 338 }
 339
 340 size_t mem_section_usage_size(void)
 341 {
 342         return sizeof(struct mem_section_usage) + usemap_size();
 343 }
 344
 345 #ifdef CONFIG_MEMORY_HOTREMOVE
 346 static struct mem_section_usage * __init
 347 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 348                                          unsigned long size)
 349 {
 350         struct mem_section_usage *usage;
 351         unsigned long goal, limit;
 352         int nid;
 353         /*
 354          * A page may contain usemaps for other sections preventing the
 355          * page being freed and making a section unremovable while
 356          * other sections referencing the usemap remain active. Similarly,
 357          * a pgdat can prevent a section being removed. If section A
 358          * contains a pgdat and section B contains the usemap, both
 359          * sections become inter-dependent. This allocates usemaps
 360          * from the same section as the pgdat where possible to avoid
 361          * this problem.
 362          */
 363         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 364         limit = goal + (1UL << PA_SECTION_SHIFT);
 365         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 366 again:
 367         usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 368         if (!usage && limit) {
 369                 limit = 0;
 370                 goto again;
 371         }
 372         return usage;
 373 }
 374
 375 static void __init check_usemap_section_nr(int nid,
 376                 struct mem_section_usage *usage)
 377 {
 378         unsigned long usemap_snr, pgdat_snr;
 379         static unsigned long old_usemap_snr;
 380         static unsigned long old_pgdat_snr;
 381         struct pglist_data *pgdat = NODE_DATA(nid);
 382         int usemap_nid;
 383
 384         /* First call */
 385         if (!old_usemap_snr) {
 386                 old_usemap_snr = NR_MEM_SECTIONS;
 387                 old_pgdat_snr = NR_MEM_SECTIONS;
 388         }
 389
 390         usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
 391         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 392         if (usemap_snr == pgdat_snr)
 393                 return;
 394
 395         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 396                 /* skip redundant message */
 397                 return;
 398
 399         old_usemap_snr = usemap_snr;
 400         old_pgdat_snr = pgdat_snr;
 401
 402         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 403         if (usemap_nid != nid) {
 404                 pr_info("node %d must be removed before remove section %ld\n",
 405                         nid, usemap_snr);
 406                 return;
 407         }
 408         /*
 409          * There is a circular dependency.
 410          * Some platforms allow un-removable section because they will just
 411          * gather other removable sections for dynamic partitioning.
 412          * Just notify un-removable section's number here.
 413          */
 414         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 415                 usemap_snr, pgdat_snr, nid);
 416 }
 417 #else
 418 static struct mem_section_usage * __init
 419 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 420                                          unsigned long size)
 421 {
 422         return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 423 }
 424
 425 static void __init check_usemap_section_nr(int nid,
 426                 struct mem_section_usage *usage)
 427 {
 428 }
 429 #endif /* CONFIG_MEMORY_HOTREMOVE */
 430
 431 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 432 static unsigned long __init section_map_size(void)
 433 {
 434         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 435 }
 436
 437 #else
 438 static unsigned long __init section_map_size(void)
 439 {
 440         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 441 }
 442
 443 struct page __init *__populate_section_memmap(unsigned long pfn,
 444                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 445 {
 446         unsigned long size = section_map_size();
 447         struct page *map = sparse_buffer_alloc(size);
 448         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 449
 450         if (map)
 451                 return map;
 452
 453         map = memblock_alloc_try_nid_raw(size, 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 inline void __meminit sparse_buffer_free(unsigned long size)
 467 {
 468         WARN_ON(!sparsemap_buf || size == 0);
 469         memblock_free_early(__pa(sparsemap_buf), size);
 470 }
 471
 472 static void __init sparse_buffer_init(unsigned long size, int nid)
 473 {
 474         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 475         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 476         /*
 477          * Pre-allocated buffer is mainly used by __populate_section_memmap
 478          * and we want it to be properly aligned to the section size - this is
 479          * especially the case for VMEMMAP which maps memmap to PMDs
 480          */
 481         sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
 482                                         addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 483         sparsemap_buf_end = sparsemap_buf + size;
 484 }
 485
 486 static void __init sparse_buffer_fini(void)
 487 {
 488         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 489
 490         if (sparsemap_buf && size > 0)
 491                 sparse_buffer_free(size);
 492         sparsemap_buf = NULL;
 493 }
 494
 495 void * __meminit sparse_buffer_alloc(unsigned long size)
 496 {
 497         void *ptr = NULL;
 498
 499         if (sparsemap_buf) {
 500                 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
 501                 if (ptr + size > sparsemap_buf_end)
 502                         ptr = NULL;
 503                 else {
 504                         /* Free redundant aligned space */
 505                         if ((unsigned long)(ptr - sparsemap_buf) > 0)
 506                                 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
 507                         sparsemap_buf = ptr + size;
 508                 }
 509         }
 510         return ptr;
 511 }
 512
 513 void __weak __meminit vmemmap_populate_print_last(void)
 514 {
 515 }
 516
 517 /*
 518  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 519  * And number of present sections in this node is map_count.
 520  */
 521 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 522                                    unsigned long pnum_end,
 523                                    unsigned long map_count)
 524 {
 525         struct mem_section_usage *usage;
 526         unsigned long pnum;
 527         struct page *map;
 528
 529         usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 530                         mem_section_usage_size() * map_count);
 531         if (!usage) {
 532                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 533                 goto failed;
 534         }
 535         sparse_buffer_init(map_count * section_map_size(), nid);
 536         for_each_present_section_nr(pnum_begin, pnum) {
 537                 unsigned long pfn = section_nr_to_pfn(pnum);
 538
 539                 if (pnum >= pnum_end)
 540                         break;
 541
 542                 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
 543                                 nid, NULL);
 544                 if (!map) {
 545                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 546                                __func__, nid);
 547                         pnum_begin = pnum;
 548                         goto failed;
 549                 }
 550                 check_usemap_section_nr(nid, usage);
 551                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
 552                                 SECTION_IS_EARLY);
 553                 usage = (void *) usage + mem_section_usage_size();
 554         }
 555         sparse_buffer_fini();
 556         return;
 557 failed:
 558         /* We failed to allocate, mark all the following pnums as not present */
 559         for_each_present_section_nr(pnum_begin, pnum) {
 560                 struct mem_section *ms;
 561
 562                 if (pnum >= pnum_end)
 563                         break;
 564                 ms = __nr_to_section(pnum);
 565                 ms->section_mem_map = 0;
 566         }
 567 }
 568
 569 /*
 570  * Allocate the accumulated non-linear sections, allocate a mem_map
 571  * for each and record the physical to section mapping.
 572  */
 573 void __init sparse_init(void)
 574 {
 575         unsigned long pnum_end, pnum_begin, map_count = 1;
 576         int nid_begin;
 577
 578         memblocks_present();
 579
 580         pnum_begin = first_present_section_nr();
 581         nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 582
 583         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 584         set_pageblock_order();
 585
 586         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 587                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 588
 589                 if (nid == nid_begin) {
 590                         map_count++;
 591                         continue;
 592                 }
 593                 /* Init node with sections in range [pnum_begin, pnum_end) */
 594                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 595                 nid_begin = nid;
 596                 pnum_begin = pnum_end;
 597                 map_count = 1;
 598         }
 599         /* cover the last node */
 600         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 601         vmemmap_populate_print_last();
 602 }
 603
 604 #ifdef CONFIG_MEMORY_HOTPLUG
 605
 606 /* Mark all memory sections within the pfn range as online */
 607 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 608 {
 609         unsigned long pfn;
 610
 611         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 612                 unsigned long section_nr = pfn_to_section_nr(pfn);
 613                 struct mem_section *ms;
 614
 615                 /* onlining code should never touch invalid ranges */
 616                 if (WARN_ON(!valid_section_nr(section_nr)))
 617                         continue;
 618
 619                 ms = __nr_to_section(section_nr);
 620                 ms->section_mem_map |= SECTION_IS_ONLINE;
 621         }
 622 }
 623
 624 #ifdef CONFIG_MEMORY_HOTREMOVE
 625 /* Mark all memory sections within the pfn range as offline */
 626 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 627 {
 628         unsigned long pfn;
 629
 630         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 631                 unsigned long section_nr = pfn_to_section_nr(pfn);
 632                 struct mem_section *ms;
 633
 634                 /*
 635                  * TODO this needs some double checking. Offlining code makes
 636                  * sure to check pfn_valid but those checks might be just bogus
 637                  */
 638                 if (WARN_ON(!valid_section_nr(section_nr)))
 639                         continue;
 640
 641                 ms = __nr_to_section(section_nr);
 642                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 643         }
 644 }
 645 #endif
 646
 647 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 648 static struct page * __meminit populate_section_memmap(unsigned long pfn,
 649                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 650 {
 651         return __populate_section_memmap(pfn, nr_pages, nid, altmap);
 652 }
 653
 654 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 655                 struct vmem_altmap *altmap)
 656 {
 657         unsigned long start = (unsigned long) pfn_to_page(pfn);
 658         unsigned long end = start + nr_pages * sizeof(struct page);
 659
 660         vmemmap_free(start, end, altmap);
 661 }
 662 static void free_map_bootmem(struct page *memmap)
 663 {
 664         unsigned long start = (unsigned long)memmap;
 665         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 666
 667         vmemmap_free(start, end, NULL);
 668 }
 669
 670 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 671 {
 672         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 673         DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
 674         struct mem_section *ms = __pfn_to_section(pfn);
 675         unsigned long *subsection_map = ms->usage
 676                 ? &ms->usage->subsection_map[0] : NULL;
 677
 678         subsection_mask_set(map, pfn, nr_pages);
 679         if (subsection_map)
 680                 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
 681
 682         if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
 683                                 "section already deactivated (%#lx + %ld)\n",
 684                                 pfn, nr_pages))
 685                 return -EINVAL;
 686
 687         bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
 688         return 0;
 689 }
 690
 691 static bool is_subsection_map_empty(struct mem_section *ms)
 692 {
 693         return bitmap_empty(&ms->usage->subsection_map[0],
 694                             SUBSECTIONS_PER_SECTION);
 695 }
 696
 697 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 698 {
 699         struct mem_section *ms = __pfn_to_section(pfn);
 700         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 701         unsigned long *subsection_map;
 702         int rc = 0;
 703
 704         subsection_mask_set(map, pfn, nr_pages);
 705
 706         subsection_map = &ms->usage->subsection_map[0];
 707
 708         if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
 709                 rc = -EINVAL;
 710         else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
 711                 rc = -EEXIST;
 712         else
 713                 bitmap_or(subsection_map, map, subsection_map,
 714                                 SUBSECTIONS_PER_SECTION);
 715
 716         return rc;
 717 }
 718 #else
 719 struct page * __meminit populate_section_memmap(unsigned long pfn,
 720                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 721 {
 722         return kvmalloc_node(array_size(sizeof(struct page),
 723                                         PAGES_PER_SECTION), GFP_KERNEL, nid);
 724 }
 725
 726 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 727                 struct vmem_altmap *altmap)
 728 {
 729         kvfree(pfn_to_page(pfn));
 730 }
 731
 732 static void free_map_bootmem(struct page *memmap)
 733 {
 734         unsigned long maps_section_nr, removing_section_nr, i;
 735         unsigned long magic, nr_pages;
 736         struct page *page = virt_to_page(memmap);
 737
 738         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 739                 >> PAGE_SHIFT;
 740
 741         for (i = 0; i < nr_pages; i++, page++) {
 742                 magic = (unsigned long) page->freelist;
 743
 744                 BUG_ON(magic == NODE_INFO);
 745
 746                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 747                 removing_section_nr = page_private(page);
 748
 749                 /*
 750                  * When this function is called, the removing section is
 751                  * logical offlined state. This means all pages are isolated
 752                  * from page allocator. If removing section's memmap is placed
 753                  * on the same section, it must not be freed.
 754                  * If it is freed, page allocator may allocate it which will
 755                  * be removed physically soon.
 756                  */
 757                 if (maps_section_nr != removing_section_nr)
 758                         put_page_bootmem(page);
 759         }
 760 }
 761
 762 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 763 {
 764         return 0;
 765 }
 766
 767 static bool is_subsection_map_empty(struct mem_section *ms)
 768 {
 769         return true;
 770 }
 771
 772 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 773 {
 774         return 0;
 775 }
 776 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 777
 778 /*
 779  * To deactivate a memory region, there are 3 cases to handle across
 780  * two configurations (SPARSEMEM_VMEMMAP={y,n}):
 781  *
 782  * 1. deactivation of a partial hot-added section (only possible in
 783  *    the SPARSEMEM_VMEMMAP=y case).
 784  *      a) section was present at memory init.
 785  *      b) section was hot-added post memory init.
 786  * 2. deactivation of a complete hot-added section.
 787  * 3. deactivation of a complete section from memory init.
 788  *
 789  * For 1, when subsection_map does not empty we will not be freeing the
 790  * usage map, but still need to free the vmemmap range.
 791  *
 792  * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
 793  */
 794 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
 795                 struct vmem_altmap *altmap)
 796 {
 797         struct mem_section *ms = __pfn_to_section(pfn);
 798         bool section_is_early = early_section(ms);
 799         struct page *memmap = NULL;
 800         bool empty;
 801
 802         if (clear_subsection_map(pfn, nr_pages))
 803                 return;
 804
 805         empty = is_subsection_map_empty(ms);
 806         if (empty) {
 807                 unsigned long section_nr = pfn_to_section_nr(pfn);
 808
 809                 /*
 810                  * When removing an early section, the usage map is kept (as the
 811                  * usage maps of other sections fall into the same page). It
 812                  * will be re-used when re-adding the section - which is then no
 813                  * longer an early section. If the usage map is PageReserved, it
 814                  * was allocated during boot.
 815                  */
 816                 if (!PageReserved(virt_to_page(ms->usage))) {
 817                         kfree(ms->usage);
 818                         ms->usage = NULL;
 819                 }
 820                 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 821                 /*
 822                  * Mark the section invalid so that valid_section()
 823                  * return false. This prevents code from dereferencing
 824                  * ms->usage array.
 825                  */
 826                 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
 827         }
 828
 829         /*
 830          * The memmap of early sections is always fully populated. See
 831          * section_activate() and pfn_valid() .
 832          */
 833         if (!section_is_early)
 834                 depopulate_section_memmap(pfn, nr_pages, altmap);
 835         else if (memmap)
 836                 free_map_bootmem(memmap);
 837
 838         if (empty)
 839                 ms->section_mem_map = (unsigned long)NULL;
 840 }
 841
 842 static struct page * __meminit section_activate(int nid, unsigned long pfn,
 843                 unsigned long nr_pages, struct vmem_altmap *altmap)
 844 {
 845         struct mem_section *ms = __pfn_to_section(pfn);
 846         struct mem_section_usage *usage = NULL;
 847         struct page *memmap;
 848         int rc = 0;
 849
 850         if (!ms->usage) {
 851                 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
 852                 if (!usage)
 853                         return ERR_PTR(-ENOMEM);
 854                 ms->usage = usage;
 855         }
 856
 857         rc = fill_subsection_map(pfn, nr_pages);
 858         if (rc) {
 859                 if (usage)
 860                         ms->usage = NULL;
 861                 kfree(usage);
 862                 return ERR_PTR(rc);
 863         }
 864
 865         /*
 866          * The early init code does not consider partially populated
 867          * initial sections, it simply assumes that memory will never be
 868          * referenced.  If we hot-add memory into such a section then we
 869          * do not need to populate the memmap and can simply reuse what
 870          * is already there.
 871          */
 872         if (nr_pages < PAGES_PER_SECTION && early_section(ms))
 873                 return pfn_to_page(pfn);
 874
 875         memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
 876         if (!memmap) {
 877                 section_deactivate(pfn, nr_pages, altmap);
 878                 return ERR_PTR(-ENOMEM);
 879         }
 880
 881         return memmap;
 882 }
 883
 884 /**
 885  * sparse_add_section - add a memory section, or populate an existing one
 886  * @nid: The node to add section on
 887  * @start_pfn: start pfn of the memory range
 888  * @nr_pages: number of pfns to add in the section
 889  * @altmap: device page map
 890  *
 891  * This is only intended for hotplug.
 892  *
 893  * Note that only VMEMMAP supports sub-section aligned hotplug,
 894  * the proper alignment and size are gated by check_pfn_span().
 895  *
 896  *
 897  * Return:
 898  * * 0          - On success.
 899  * * -EEXIST    - Section has been present.
 900  * * -ENOMEM    - Out of memory.
 901  */
 902 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
 903                 unsigned long nr_pages, struct vmem_altmap *altmap)
 904 {
 905         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 906         struct mem_section *ms;
 907         struct page *memmap;
 908         int ret;
 909
 910         ret = sparse_index_init(section_nr, nid);
 911         if (ret < 0)
 912                 return ret;
 913
 914         memmap = section_activate(nid, start_pfn, nr_pages, altmap);
 915         if (IS_ERR(memmap))
 916                 return PTR_ERR(memmap);
 917
 918         /*
 919          * Poison uninitialized struct pages in order to catch invalid flags
 920          * combinations.
 921          */
 922         page_init_poison(memmap, sizeof(struct page) * nr_pages);
 923
 924         ms = __nr_to_section(section_nr);
 925         set_section_nid(section_nr, nid);
 926         section_mark_present(ms);
 927
 928         /* Align memmap to section boundary in the subsection case */
 929         if (section_nr_to_pfn(section_nr) != start_pfn)
 930                 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
 931         sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
 932
 933         return 0;
 934 }
 935
 936 #ifdef CONFIG_MEMORY_FAILURE
 937 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 938 {
 939         int i;
 940
 941         /*
 942          * A further optimization is to have per section refcounted
 943          * num_poisoned_pages.  But that would need more space per memmap, so
 944          * for now just do a quick global check to speed up this routine in the
 945          * absence of bad pages.
 946          */
 947         if (atomic_long_read(&num_poisoned_pages) == 0)
 948                 return;
 949
 950         for (i = 0; i < nr_pages; i++) {
 951                 if (PageHWPoison(&memmap[i])) {
 952                         num_poisoned_pages_dec();
 953                         ClearPageHWPoison(&memmap[i]);
 954                 }
 955         }
 956 }
 957 #else
 958 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 959 {
 960 }
 961 #endif
 962
 963 void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
 964                 unsigned long nr_pages, unsigned long map_offset,
 965                 struct vmem_altmap *altmap)
 966 {
 967         clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
 968                         nr_pages - map_offset);
 969         section_deactivate(pfn, nr_pages, altmap);
 970 }
 971 #endif /* CONFIG_MEMORY_HOTPLUG */