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 #include <asm/pgalloc.h>
  20 #include <asm/pgtable.h>
  21
  22 /*
  23  * Permanent SPARSEMEM data:
  24  *
  25  * 1) mem_section       - memory sections, mem_map's for valid memory
  26  */
  27 #ifdef CONFIG_SPARSEMEM_EXTREME
  28 struct mem_section **mem_section;
  29 #else
  30 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  31         ____cacheline_internodealigned_in_smp;
  32 #endif
  33 EXPORT_SYMBOL(mem_section);
  34
  35 #ifdef NODE_NOT_IN_PAGE_FLAGS
  36 /*
  37  * If we did not store the node number in the page then we have to
  38  * do a lookup in the section_to_node_table in order to find which
  39  * node the page belongs to.
  40  */
  41 #if MAX_NUMNODES <= 256
  42 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  43 #else
  44 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  45 #endif
  46
  47 int page_to_nid(const struct page *page)
  48 {
  49         return section_to_node_table[page_to_section(page)];
  50 }
  51 EXPORT_SYMBOL(page_to_nid);
  52
  53 static void set_section_nid(unsigned long section_nr, int nid)
  54 {
  55         section_to_node_table[section_nr] = nid;
  56 }
  57 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  58 static inline void set_section_nid(unsigned long section_nr, int nid)
  59 {
  60 }
  61 #endif
  62
  63 #ifdef CONFIG_SPARSEMEM_EXTREME
  64 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
  65 {
  66         struct mem_section *section = NULL;
  67         unsigned long array_size = SECTIONS_PER_ROOT *
  68                                    sizeof(struct mem_section);
  69
  70         if (slab_is_available()) {
  71                 section = kzalloc_node(array_size, GFP_KERNEL, nid);
  72         } else {
  73                 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
  74                                               nid);
  75                 if (!section)
  76                         panic("%s: Failed to allocate %lu bytes nid=%d\n",
  77                               __func__, array_size, nid);
  78         }
  79
  80         return section;
  81 }
  82
  83 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  84 {
  85         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  86         struct mem_section *section;
  87
  88         /*
  89          * An existing section is possible in the sub-section hotplug
  90          * case. First hot-add instantiates, follow-on hot-add reuses
  91          * the existing section.
  92          *
  93          * The mem_hotplug_lock resolves the apparent race below.
  94          */
  95         if (mem_section[root])
  96                 return 0;
  97
  98         section = sparse_index_alloc(nid);
  99         if (!section)
 100                 return -ENOMEM;
 101
 102         mem_section[root] = section;
 103
 104         return 0;
 105 }
 106 #else /* !SPARSEMEM_EXTREME */
 107 static inline int sparse_index_init(unsigned long section_nr, int nid)
 108 {
 109         return 0;
 110 }
 111 #endif
 112
 113 #ifdef CONFIG_SPARSEMEM_EXTREME
 114 unsigned long __section_nr(struct mem_section *ms)
 115 {
 116         unsigned long root_nr;
 117         struct mem_section *root = NULL;
 118
 119         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 120                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 121                 if (!root)
 122                         continue;
 123
 124                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 125                      break;
 126         }
 127
 128         VM_BUG_ON(!root);
 129
 130         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 131 }
 132 #else
 133 unsigned long __section_nr(struct mem_section *ms)
 134 {
 135         return (unsigned long)(ms - mem_section[0]);
 136 }
 137 #endif
 138
 139 /*
 140  * During early boot, before section_mem_map is used for an actual
 141  * mem_map, we use section_mem_map to store the section's NUMA
 142  * node.  This keeps us from having to use another data structure.  The
 143  * node information is cleared just before we store the real mem_map.
 144  */
 145 static inline unsigned long sparse_encode_early_nid(int nid)
 146 {
 147         return (nid << SECTION_NID_SHIFT);
 148 }
 149
 150 static inline int sparse_early_nid(struct mem_section *section)
 151 {
 152         return (section->section_mem_map >> SECTION_NID_SHIFT);
 153 }
 154
 155 /* Validate the physical addressing limitations of the model */
 156 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 157                                                 unsigned long *end_pfn)
 158 {
 159         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 160
 161         /*
 162          * Sanity checks - do not allow an architecture to pass
 163          * in larger pfns than the maximum scope of sparsemem:
 164          */
 165         if (*start_pfn > max_sparsemem_pfn) {
 166                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 167                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 168                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 169                 WARN_ON_ONCE(1);
 170                 *start_pfn = max_sparsemem_pfn;
 171                 *end_pfn = max_sparsemem_pfn;
 172         } else if (*end_pfn > max_sparsemem_pfn) {
 173                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 174                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 175                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 176                 WARN_ON_ONCE(1);
 177                 *end_pfn = max_sparsemem_pfn;
 178         }
 179 }
 180
 181 /*
 182  * There are a number of times that we loop over NR_MEM_SECTIONS,
 183  * looking for section_present() on each.  But, when we have very
 184  * large physical address spaces, NR_MEM_SECTIONS can also be
 185  * very large which makes the loops quite long.
 186  *
 187  * Keeping track of this gives us an easy way to break out of
 188  * those loops early.
 189  */
 190 unsigned long __highest_present_section_nr;
 191 static void section_mark_present(struct mem_section *ms)
 192 {
 193         unsigned long section_nr = __section_nr(ms);
 194
 195         if (section_nr > __highest_present_section_nr)
 196                 __highest_present_section_nr = section_nr;
 197
 198         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 199 }
 200
 201 static inline unsigned long next_present_section_nr(unsigned long section_nr)
 202 {
 203         do {
 204                 section_nr++;
 205                 if (present_section_nr(section_nr))
 206                         return section_nr;
 207         } while ((section_nr <= __highest_present_section_nr));
 208
 209         return -1;
 210 }
 211 #define for_each_present_section_nr(start, section_nr)          \
 212         for (section_nr = next_present_section_nr(start-1);     \
 213              ((section_nr != -1) &&                             \
 214               (section_nr <= __highest_present_section_nr));    \
 215              section_nr = next_present_section_nr(section_nr))
 216
 217 static inline unsigned long first_present_section_nr(void)
 218 {
 219         return next_present_section_nr(-1);
 220 }
 221
 222 static void subsection_mask_set(unsigned long *map, unsigned long pfn,
 223                 unsigned long nr_pages)
 224 {
 225         int idx = subsection_map_index(pfn);
 226         int end = subsection_map_index(pfn + nr_pages - 1);
 227
 228         bitmap_set(map, idx, end - idx + 1);
 229 }
 230
 231 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 232 {
 233         int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
 234         unsigned long nr, start_sec = pfn_to_section_nr(pfn);
 235
 236         if (!nr_pages)
 237                 return;
 238
 239         for (nr = start_sec; nr <= end_sec; nr++) {
 240                 struct mem_section *ms;
 241                 unsigned long pfns;
 242
 243                 pfns = min(nr_pages, PAGES_PER_SECTION
 244                                 - (pfn & ~PAGE_SECTION_MASK));
 245                 ms = __nr_to_section(nr);
 246                 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
 247
 248                 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
 249                                 pfns, subsection_map_index(pfn),
 250                                 subsection_map_index(pfn + pfns - 1));
 251
 252                 pfn += pfns;
 253                 nr_pages -= pfns;
 254         }
 255 }
 256
 257 /* Record a memory area against a node. */
 258 void __init memory_present(int nid, unsigned long start, unsigned long end)
 259 {
 260         unsigned long pfn;
 261
 262 #ifdef CONFIG_SPARSEMEM_EXTREME
 263         if (unlikely(!mem_section)) {
 264                 unsigned long size, align;
 265
 266                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
 267                 align = 1 << (INTERNODE_CACHE_SHIFT);
 268                 mem_section = memblock_alloc(size, align);
 269                 if (!mem_section)
 270                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 271                               __func__, size, align);
 272         }
 273 #endif
 274
 275         start &= PAGE_SECTION_MASK;
 276         mminit_validate_memmodel_limits(&start, &end);
 277         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 278                 unsigned long section = pfn_to_section_nr(pfn);
 279                 struct mem_section *ms;
 280
 281                 sparse_index_init(section, nid);
 282                 set_section_nid(section, nid);
 283
 284                 ms = __nr_to_section(section);
 285                 if (!ms->section_mem_map) {
 286                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 287                                                         SECTION_IS_ONLINE;
 288                         section_mark_present(ms);
 289                 }
 290         }
 291 }
 292
 293 /*
 294  * Mark all memblocks as present using memory_present(). This is a
 295  * convienence function that is useful for a number of arches
 296  * to mark all of the systems memory as present during initialization.
 297  */
 298 void __init memblocks_present(void)
 299 {
 300         struct memblock_region *reg;
 301
 302         for_each_memblock(memory, reg) {
 303                 memory_present(memblock_get_region_node(reg),
 304                                memblock_region_memory_base_pfn(reg),
 305                                memblock_region_memory_end_pfn(reg));
 306         }
 307 }
 308
 309 /*
 310  * Subtle, we encode the real pfn into the mem_map such that
 311  * the identity pfn - section_mem_map will return the actual
 312  * physical page frame number.
 313  */
 314 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 315 {
 316         unsigned long coded_mem_map =
 317                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 318         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 319         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 320         return coded_mem_map;
 321 }
 322
 323 /*
 324  * Decode mem_map from the coded memmap
 325  */
 326 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 327 {
 328         /* mask off the extra low bits of information */
 329         coded_mem_map &= SECTION_MAP_MASK;
 330         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 331 }
 332
 333 static void __meminit sparse_init_one_section(struct mem_section *ms,
 334                 unsigned long pnum, struct page *mem_map,
 335                 struct mem_section_usage *usage, unsigned long flags)
 336 {
 337         ms->section_mem_map &= ~SECTION_MAP_MASK;
 338         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
 339                 | SECTION_HAS_MEM_MAP | flags;
 340         ms->usage = usage;
 341 }
 342
 343 static unsigned long usemap_size(void)
 344 {
 345         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 346 }
 347
 348 size_t mem_section_usage_size(void)
 349 {
 350         return sizeof(struct mem_section_usage) + usemap_size();
 351 }
 352
 353 #ifdef CONFIG_MEMORY_HOTREMOVE
 354 static struct mem_section_usage * __init
 355 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 356                                          unsigned long size)
 357 {
 358         struct mem_section_usage *usage;
 359         unsigned long goal, limit;
 360         int nid;
 361         /*
 362          * A page may contain usemaps for other sections preventing the
 363          * page being freed and making a section unremovable while
 364          * other sections referencing the usemap remain active. Similarly,
 365          * a pgdat can prevent a section being removed. If section A
 366          * contains a pgdat and section B contains the usemap, both
 367          * sections become inter-dependent. This allocates usemaps
 368          * from the same section as the pgdat where possible to avoid
 369          * this problem.
 370          */
 371         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 372         limit = goal + (1UL << PA_SECTION_SHIFT);
 373         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 374 again:
 375         usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 376         if (!usage && limit) {
 377                 limit = 0;
 378                 goto again;
 379         }
 380         return usage;
 381 }
 382
 383 static void __init check_usemap_section_nr(int nid,
 384                 struct mem_section_usage *usage)
 385 {
 386         unsigned long usemap_snr, pgdat_snr;
 387         static unsigned long old_usemap_snr;
 388         static unsigned long old_pgdat_snr;
 389         struct pglist_data *pgdat = NODE_DATA(nid);
 390         int usemap_nid;
 391
 392         /* First call */
 393         if (!old_usemap_snr) {
 394                 old_usemap_snr = NR_MEM_SECTIONS;
 395                 old_pgdat_snr = NR_MEM_SECTIONS;
 396         }
 397
 398         usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
 399         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 400         if (usemap_snr == pgdat_snr)
 401                 return;
 402
 403         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 404                 /* skip redundant message */
 405                 return;
 406
 407         old_usemap_snr = usemap_snr;
 408         old_pgdat_snr = pgdat_snr;
 409
 410         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 411         if (usemap_nid != nid) {
 412                 pr_info("node %d must be removed before remove section %ld\n",
 413                         nid, usemap_snr);
 414                 return;
 415         }
 416         /*
 417          * There is a circular dependency.
 418          * Some platforms allow un-removable section because they will just
 419          * gather other removable sections for dynamic partitioning.
 420          * Just notify un-removable section's number here.
 421          */
 422         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 423                 usemap_snr, pgdat_snr, nid);
 424 }
 425 #else
 426 static struct mem_section_usage * __init
 427 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 428                                          unsigned long size)
 429 {
 430         return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 431 }
 432
 433 static void __init check_usemap_section_nr(int nid,
 434                 struct mem_section_usage *usage)
 435 {
 436 }
 437 #endif /* CONFIG_MEMORY_HOTREMOVE */
 438
 439 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 440 static unsigned long __init section_map_size(void)
 441 {
 442         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 443 }
 444
 445 #else
 446 static unsigned long __init section_map_size(void)
 447 {
 448         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 449 }
 450
 451 struct page __init *__populate_section_memmap(unsigned long pfn,
 452                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 453 {
 454         unsigned long size = section_map_size();
 455         struct page *map = sparse_buffer_alloc(size);
 456         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 457
 458         if (map)
 459                 return map;
 460
 461         map = memblock_alloc_try_nid_raw(size, size, addr,
 462                                           MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 463         if (!map)
 464                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 465                       __func__, size, PAGE_SIZE, nid, &addr);
 466
 467         return map;
 468 }
 469 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 470
 471 static void *sparsemap_buf __meminitdata;
 472 static void *sparsemap_buf_end __meminitdata;
 473
 474 static inline void __meminit sparse_buffer_free(unsigned long size)
 475 {
 476         WARN_ON(!sparsemap_buf || size == 0);
 477         memblock_free_early(__pa(sparsemap_buf), size);
 478 }
 479
 480 static void __init sparse_buffer_init(unsigned long size, int nid)
 481 {
 482         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 483         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 484         /*
 485          * Pre-allocated buffer is mainly used by __populate_section_memmap
 486          * and we want it to be properly aligned to the section size - this is
 487          * especially the case for VMEMMAP which maps memmap to PMDs
 488          */
 489         sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
 490                                         addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 491         sparsemap_buf_end = sparsemap_buf + size;
 492 }
 493
 494 static void __init sparse_buffer_fini(void)
 495 {
 496         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 497
 498         if (sparsemap_buf && size > 0)
 499                 sparse_buffer_free(size);
 500         sparsemap_buf = NULL;
 501 }
 502
 503 void * __meminit sparse_buffer_alloc(unsigned long size)
 504 {
 505         void *ptr = NULL;
 506
 507         if (sparsemap_buf) {
 508                 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
 509                 if (ptr + size > sparsemap_buf_end)
 510                         ptr = NULL;
 511                 else {
 512                         /* Free redundant aligned space */
 513                         if ((unsigned long)(ptr - sparsemap_buf) > 0)
 514                                 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
 515                         sparsemap_buf = ptr + size;
 516                 }
 517         }
 518         return ptr;
 519 }
 520
 521 void __weak __meminit vmemmap_populate_print_last(void)
 522 {
 523 }
 524
 525 /*
 526  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 527  * And number of present sections in this node is map_count.
 528  */
 529 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 530                                    unsigned long pnum_end,
 531                                    unsigned long map_count)
 532 {
 533         struct mem_section_usage *usage;
 534         unsigned long pnum;
 535         struct page *map;
 536
 537         usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 538                         mem_section_usage_size() * map_count);
 539         if (!usage) {
 540                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 541                 goto failed;
 542         }
 543         sparse_buffer_init(map_count * section_map_size(), nid);
 544         for_each_present_section_nr(pnum_begin, pnum) {
 545                 unsigned long pfn = section_nr_to_pfn(pnum);
 546
 547                 if (pnum >= pnum_end)
 548                         break;
 549
 550                 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
 551                                 nid, NULL);
 552                 if (!map) {
 553                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 554                                __func__, nid);
 555                         pnum_begin = pnum;
 556                         goto failed;
 557                 }
 558                 check_usemap_section_nr(nid, usage);
 559                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
 560                                 SECTION_IS_EARLY);
 561                 usage = (void *) usage + mem_section_usage_size();
 562         }
 563         sparse_buffer_fini();
 564         return;
 565 failed:
 566         /* We failed to allocate, mark all the following pnums as not present */
 567         for_each_present_section_nr(pnum_begin, pnum) {
 568                 struct mem_section *ms;
 569
 570                 if (pnum >= pnum_end)
 571                         break;
 572                 ms = __nr_to_section(pnum);
 573                 ms->section_mem_map = 0;
 574         }
 575 }
 576
 577 /*
 578  * Allocate the accumulated non-linear sections, allocate a mem_map
 579  * for each and record the physical to section mapping.
 580  */
 581 void __init sparse_init(void)
 582 {
 583         unsigned long pnum_begin = first_present_section_nr();
 584         int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 585         unsigned long pnum_end, map_count = 1;
 586
 587         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 588         set_pageblock_order();
 589
 590         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 591                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 592
 593                 if (nid == nid_begin) {
 594                         map_count++;
 595                         continue;
 596                 }
 597                 /* Init node with sections in range [pnum_begin, pnum_end) */
 598                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 599                 nid_begin = nid;
 600                 pnum_begin = pnum_end;
 601                 map_count = 1;
 602         }
 603         /* cover the last node */
 604         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 605         vmemmap_populate_print_last();
 606 }
 607
 608 #ifdef CONFIG_MEMORY_HOTPLUG
 609
 610 /* Mark all memory sections within the pfn range as online */
 611 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 612 {
 613         unsigned long pfn;
 614
 615         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 616                 unsigned long section_nr = pfn_to_section_nr(pfn);
 617                 struct mem_section *ms;
 618
 619                 /* onlining code should never touch invalid ranges */
 620                 if (WARN_ON(!valid_section_nr(section_nr)))
 621                         continue;
 622
 623                 ms = __nr_to_section(section_nr);
 624                 ms->section_mem_map |= SECTION_IS_ONLINE;
 625         }
 626 }
 627
 628 #ifdef CONFIG_MEMORY_HOTREMOVE
 629 /* Mark all memory sections within the pfn range as offline */
 630 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 631 {
 632         unsigned long pfn;
 633
 634         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 635                 unsigned long section_nr = pfn_to_section_nr(pfn);
 636                 struct mem_section *ms;
 637
 638                 /*
 639                  * TODO this needs some double checking. Offlining code makes
 640                  * sure to check pfn_valid but those checks might be just bogus
 641                  */
 642                 if (WARN_ON(!valid_section_nr(section_nr)))
 643                         continue;
 644
 645                 ms = __nr_to_section(section_nr);
 646                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 647         }
 648 }
 649 #endif
 650
 651 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 652 static struct page * __meminit populate_section_memmap(unsigned long pfn,
 653                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 654 {
 655         return __populate_section_memmap(pfn, nr_pages, nid, altmap);
 656 }
 657
 658 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 659                 struct vmem_altmap *altmap)
 660 {
 661         unsigned long start = (unsigned long) pfn_to_page(pfn);
 662         unsigned long end = start + nr_pages * sizeof(struct page);
 663
 664         vmemmap_free(start, end, altmap);
 665 }
 666 static void free_map_bootmem(struct page *memmap)
 667 {
 668         unsigned long start = (unsigned long)memmap;
 669         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 670
 671         vmemmap_free(start, end, NULL);
 672 }
 673 #else
 674 struct page * __meminit populate_section_memmap(unsigned long pfn,
 675                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 676 {
 677         struct page *page, *ret;
 678         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 679
 680         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 681         if (page)
 682                 goto got_map_page;
 683
 684         ret = vmalloc(memmap_size);
 685         if (ret)
 686                 goto got_map_ptr;
 687
 688         return NULL;
 689 got_map_page:
 690         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 691 got_map_ptr:
 692
 693         return ret;
 694 }
 695
 696 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 697                 struct vmem_altmap *altmap)
 698 {
 699         struct page *memmap = pfn_to_page(pfn);
 700
 701         if (is_vmalloc_addr(memmap))
 702                 vfree(memmap);
 703         else
 704                 free_pages((unsigned long)memmap,
 705                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 706 }
 707
 708 static void free_map_bootmem(struct page *memmap)
 709 {
 710         unsigned long maps_section_nr, removing_section_nr, i;
 711         unsigned long magic, nr_pages;
 712         struct page *page = virt_to_page(memmap);
 713
 714         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 715                 >> PAGE_SHIFT;
 716
 717         for (i = 0; i < nr_pages; i++, page++) {
 718                 magic = (unsigned long) page->freelist;
 719
 720                 BUG_ON(magic == NODE_INFO);
 721
 722                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 723                 removing_section_nr = page_private(page);
 724
 725                 /*
 726                  * When this function is called, the removing section is
 727                  * logical offlined state. This means all pages are isolated
 728                  * from page allocator. If removing section's memmap is placed
 729                  * on the same section, it must not be freed.
 730                  * If it is freed, page allocator may allocate it which will
 731                  * be removed physically soon.
 732                  */
 733                 if (maps_section_nr != removing_section_nr)
 734                         put_page_bootmem(page);
 735         }
 736 }
 737 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 738
 739 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
 740                 struct vmem_altmap *altmap)
 741 {
 742         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 743         DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
 744         struct mem_section *ms = __pfn_to_section(pfn);
 745         bool section_is_early = early_section(ms);
 746         struct page *memmap = NULL;
 747         unsigned long *subsection_map = ms->usage
 748                 ? &ms->usage->subsection_map[0] : NULL;
 749
 750         subsection_mask_set(map, pfn, nr_pages);
 751         if (subsection_map)
 752                 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
 753
 754         if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
 755                                 "section already deactivated (%#lx + %ld)\n",
 756                                 pfn, nr_pages))
 757                 return;
 758
 759         /*
 760          * There are 3 cases to handle across two configurations
 761          * (SPARSEMEM_VMEMMAP={y,n}):
 762          *
 763          * 1/ deactivation of a partial hot-added section (only possible
 764          * in the SPARSEMEM_VMEMMAP=y case).
 765          *    a/ section was present at memory init
 766          *    b/ section was hot-added post memory init
 767          * 2/ deactivation of a complete hot-added section
 768          * 3/ deactivation of a complete section from memory init
 769          *
 770          * For 1/, when subsection_map does not empty we will not be
 771          * freeing the usage map, but still need to free the vmemmap
 772          * range.
 773          *
 774          * For 2/ and 3/ the SPARSEMEM_VMEMMAP={y,n} cases are unified
 775          */
 776         bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
 777         if (bitmap_empty(subsection_map, SUBSECTIONS_PER_SECTION)) {
 778                 unsigned long section_nr = pfn_to_section_nr(pfn);
 779
 780                 /*
 781                  * When removing an early section, the usage map is kept (as the
 782                  * usage maps of other sections fall into the same page). It
 783                  * will be re-used when re-adding the section - which is then no
 784                  * longer an early section. If the usage map is PageReserved, it
 785                  * was allocated during boot.
 786                  */
 787                 if (!PageReserved(virt_to_page(ms->usage))) {
 788                         kfree(ms->usage);
 789                         ms->usage = NULL;
 790                 }
 791                 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 792                 ms->section_mem_map = sparse_encode_mem_map(NULL, section_nr);
 793         }
 794
 795         if (section_is_early && memmap)
 796                 free_map_bootmem(memmap);
 797         else
 798                 depopulate_section_memmap(pfn, nr_pages, altmap);
 799 }
 800
 801 static struct page * __meminit section_activate(int nid, unsigned long pfn,
 802                 unsigned long nr_pages, struct vmem_altmap *altmap)
 803 {
 804         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 805         struct mem_section *ms = __pfn_to_section(pfn);
 806         struct mem_section_usage *usage = NULL;
 807         unsigned long *subsection_map;
 808         struct page *memmap;
 809         int rc = 0;
 810
 811         subsection_mask_set(map, pfn, nr_pages);
 812
 813         if (!ms->usage) {
 814                 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
 815                 if (!usage)
 816                         return ERR_PTR(-ENOMEM);
 817                 ms->usage = usage;
 818         }
 819         subsection_map = &ms->usage->subsection_map[0];
 820
 821         if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
 822                 rc = -EINVAL;
 823         else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
 824                 rc = -EEXIST;
 825         else
 826                 bitmap_or(subsection_map, map, subsection_map,
 827                                 SUBSECTIONS_PER_SECTION);
 828
 829         if (rc) {
 830                 if (usage)
 831                         ms->usage = NULL;
 832                 kfree(usage);
 833                 return ERR_PTR(rc);
 834         }
 835
 836         /*
 837          * The early init code does not consider partially populated
 838          * initial sections, it simply assumes that memory will never be
 839          * referenced.  If we hot-add memory into such a section then we
 840          * do not need to populate the memmap and can simply reuse what
 841          * is already there.
 842          */
 843         if (nr_pages < PAGES_PER_SECTION && early_section(ms))
 844                 return pfn_to_page(pfn);
 845
 846         memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
 847         if (!memmap) {
 848                 section_deactivate(pfn, nr_pages, altmap);
 849                 return ERR_PTR(-ENOMEM);
 850         }
 851
 852         return memmap;
 853 }
 854
 855 /**
 856  * sparse_add_section - add a memory section, or populate an existing one
 857  * @nid: The node to add section on
 858  * @start_pfn: start pfn of the memory range
 859  * @nr_pages: number of pfns to add in the section
 860  * @altmap: device page map
 861  *
 862  * This is only intended for hotplug.
 863  *
 864  * Return:
 865  * * 0          - On success.
 866  * * -EEXIST    - Section has been present.
 867  * * -ENOMEM    - Out of memory.
 868  */
 869 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
 870                 unsigned long nr_pages, struct vmem_altmap *altmap)
 871 {
 872         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 873         struct mem_section *ms;
 874         struct page *memmap;
 875         int ret;
 876
 877         ret = sparse_index_init(section_nr, nid);
 878         if (ret < 0)
 879                 return ret;
 880
 881         memmap = section_activate(nid, start_pfn, nr_pages, altmap);
 882         if (IS_ERR(memmap))
 883                 return PTR_ERR(memmap);
 884
 885         /*
 886          * Poison uninitialized struct pages in order to catch invalid flags
 887          * combinations.
 888          */
 889         page_init_poison(pfn_to_page(start_pfn), sizeof(struct page) * nr_pages);
 890
 891         ms = __nr_to_section(section_nr);
 892         set_section_nid(section_nr, nid);
 893         section_mark_present(ms);
 894
 895         /* Align memmap to section boundary in the subsection case */
 896         if (section_nr_to_pfn(section_nr) != start_pfn)
 897                 memmap = pfn_to_kaddr(section_nr_to_pfn(section_nr));
 898         sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
 899
 900         return 0;
 901 }
 902
 903 #ifdef CONFIG_MEMORY_FAILURE
 904 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 905 {
 906         int i;
 907
 908         /*
 909          * A further optimization is to have per section refcounted
 910          * num_poisoned_pages.  But that would need more space per memmap, so
 911          * for now just do a quick global check to speed up this routine in the
 912          * absence of bad pages.
 913          */
 914         if (atomic_long_read(&num_poisoned_pages) == 0)
 915                 return;
 916
 917         for (i = 0; i < nr_pages; i++) {
 918                 if (PageHWPoison(&memmap[i])) {
 919                         num_poisoned_pages_dec();
 920                         ClearPageHWPoison(&memmap[i]);
 921                 }
 922         }
 923 }
 924 #else
 925 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 926 {
 927 }
 928 #endif
 929
 930 void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
 931                 unsigned long nr_pages, unsigned long map_offset,
 932                 struct vmem_altmap *altmap)
 933 {
 934         clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
 935                         nr_pages - map_offset);
 936         section_deactivate(pfn, nr_pages, altmap);
 937 }
 938 #endif /* CONFIG_MEMORY_HOTPLUG */