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         /*
  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 static inline unsigned long next_present_section_nr(unsigned long section_nr)
 200 {
 201         do {
 202                 section_nr++;
 203                 if (present_section_nr(section_nr))
 204                         return section_nr;
 205         } while ((section_nr <= __highest_present_section_nr));
 206
 207         return -1;
 208 }
 209 #define for_each_present_section_nr(start, section_nr)          \
 210         for (section_nr = next_present_section_nr(start-1);     \
 211              ((section_nr != -1) &&                             \
 212               (section_nr <= __highest_present_section_nr));    \
 213              section_nr = next_present_section_nr(section_nr))
 214
 215 static inline unsigned long first_present_section_nr(void)
 216 {
 217         return next_present_section_nr(-1);
 218 }
 219
 220 void subsection_mask_set(unsigned long *map, unsigned long pfn,
 221                 unsigned long nr_pages)
 222 {
 223         int idx = subsection_map_index(pfn);
 224         int end = subsection_map_index(pfn + nr_pages - 1);
 225
 226         bitmap_set(map, idx, end - idx + 1);
 227 }
 228
 229 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 230 {
 231         int end_sec = pfn_to_section_nr(pfn + nr_pages - 1);
 232         unsigned long nr, start_sec = pfn_to_section_nr(pfn);
 233
 234         if (!nr_pages)
 235                 return;
 236
 237         for (nr = start_sec; nr <= end_sec; nr++) {
 238                 struct mem_section *ms;
 239                 unsigned long pfns;
 240
 241                 pfns = min(nr_pages, PAGES_PER_SECTION
 242                                 - (pfn & ~PAGE_SECTION_MASK));
 243                 ms = __nr_to_section(nr);
 244                 subsection_mask_set(ms->usage->subsection_map, pfn, pfns);
 245
 246                 pr_debug("%s: sec: %lu pfns: %lu set(%d, %d)\n", __func__, nr,
 247                                 pfns, subsection_map_index(pfn),
 248                                 subsection_map_index(pfn + pfns - 1));
 249
 250                 pfn += pfns;
 251                 nr_pages -= pfns;
 252         }
 253 }
 254
 255 /* Record a memory area against a node. */
 256 void __init memory_present(int nid, unsigned long start, unsigned long end)
 257 {
 258         unsigned long pfn;
 259
 260 #ifdef CONFIG_SPARSEMEM_EXTREME
 261         if (unlikely(!mem_section)) {
 262                 unsigned long size, align;
 263
 264                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
 265                 align = 1 << (INTERNODE_CACHE_SHIFT);
 266                 mem_section = memblock_alloc(size, align);
 267                 if (!mem_section)
 268                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 269                               __func__, size, align);
 270         }
 271 #endif
 272
 273         start &= PAGE_SECTION_MASK;
 274         mminit_validate_memmodel_limits(&start, &end);
 275         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 276                 unsigned long section = pfn_to_section_nr(pfn);
 277                 struct mem_section *ms;
 278
 279                 sparse_index_init(section, nid);
 280                 set_section_nid(section, nid);
 281
 282                 ms = __nr_to_section(section);
 283                 if (!ms->section_mem_map) {
 284                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 285                                                         SECTION_IS_ONLINE;
 286                         section_mark_present(ms);
 287                 }
 288         }
 289 }
 290
 291 /*
 292  * Mark all memblocks as present using memory_present(). This is a
 293  * convienence function that is useful for a number of arches
 294  * to mark all of the systems memory as present during initialization.
 295  */
 296 void __init memblocks_present(void)
 297 {
 298         struct memblock_region *reg;
 299
 300         for_each_memblock(memory, reg) {
 301                 memory_present(memblock_get_region_node(reg),
 302                                memblock_region_memory_base_pfn(reg),
 303                                memblock_region_memory_end_pfn(reg));
 304         }
 305 }
 306
 307 /*
 308  * Subtle, we encode the real pfn into the mem_map such that
 309  * the identity pfn - section_mem_map will return the actual
 310  * physical page frame number.
 311  */
 312 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 313 {
 314         unsigned long coded_mem_map =
 315                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 316         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 317         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 318         return coded_mem_map;
 319 }
 320
 321 /*
 322  * Decode mem_map from the coded memmap
 323  */
 324 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 325 {
 326         /* mask off the extra low bits of information */
 327         coded_mem_map &= SECTION_MAP_MASK;
 328         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 329 }
 330
 331 static void __meminit sparse_init_one_section(struct mem_section *ms,
 332                 unsigned long pnum, struct page *mem_map,
 333                 struct mem_section_usage *usage, unsigned long flags)
 334 {
 335         ms->section_mem_map &= ~SECTION_MAP_MASK;
 336         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
 337                 | SECTION_HAS_MEM_MAP | flags;
 338         ms->usage = usage;
 339 }
 340
 341 static unsigned long usemap_size(void)
 342 {
 343         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 344 }
 345
 346 size_t mem_section_usage_size(void)
 347 {
 348         return sizeof(struct mem_section_usage) + usemap_size();
 349 }
 350
 351 #ifdef CONFIG_MEMORY_HOTREMOVE
 352 static struct mem_section_usage * __init
 353 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 354                                          unsigned long size)
 355 {
 356         struct mem_section_usage *usage;
 357         unsigned long goal, limit;
 358         int nid;
 359         /*
 360          * A page may contain usemaps for other sections preventing the
 361          * page being freed and making a section unremovable while
 362          * other sections referencing the usemap remain active. Similarly,
 363          * a pgdat can prevent a section being removed. If section A
 364          * contains a pgdat and section B contains the usemap, both
 365          * sections become inter-dependent. This allocates usemaps
 366          * from the same section as the pgdat where possible to avoid
 367          * this problem.
 368          */
 369         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 370         limit = goal + (1UL << PA_SECTION_SHIFT);
 371         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 372 again:
 373         usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 374         if (!usage && limit) {
 375                 limit = 0;
 376                 goto again;
 377         }
 378         return usage;
 379 }
 380
 381 static void __init check_usemap_section_nr(int nid,
 382                 struct mem_section_usage *usage)
 383 {
 384         unsigned long usemap_snr, pgdat_snr;
 385         static unsigned long old_usemap_snr;
 386         static unsigned long old_pgdat_snr;
 387         struct pglist_data *pgdat = NODE_DATA(nid);
 388         int usemap_nid;
 389
 390         /* First call */
 391         if (!old_usemap_snr) {
 392                 old_usemap_snr = NR_MEM_SECTIONS;
 393                 old_pgdat_snr = NR_MEM_SECTIONS;
 394         }
 395
 396         usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
 397         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 398         if (usemap_snr == pgdat_snr)
 399                 return;
 400
 401         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 402                 /* skip redundant message */
 403                 return;
 404
 405         old_usemap_snr = usemap_snr;
 406         old_pgdat_snr = pgdat_snr;
 407
 408         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 409         if (usemap_nid != nid) {
 410                 pr_info("node %d must be removed before remove section %ld\n",
 411                         nid, usemap_snr);
 412                 return;
 413         }
 414         /*
 415          * There is a circular dependency.
 416          * Some platforms allow un-removable section because they will just
 417          * gather other removable sections for dynamic partitioning.
 418          * Just notify un-removable section's number here.
 419          */
 420         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 421                 usemap_snr, pgdat_snr, nid);
 422 }
 423 #else
 424 static struct mem_section_usage * __init
 425 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 426                                          unsigned long size)
 427 {
 428         return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 429 }
 430
 431 static void __init check_usemap_section_nr(int nid,
 432                 struct mem_section_usage *usage)
 433 {
 434 }
 435 #endif /* CONFIG_MEMORY_HOTREMOVE */
 436
 437 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 438 static unsigned long __init section_map_size(void)
 439 {
 440         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 441 }
 442
 443 #else
 444 static unsigned long __init section_map_size(void)
 445 {
 446         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 447 }
 448
 449 struct page __init *__populate_section_memmap(unsigned long pfn,
 450                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 451 {
 452         unsigned long size = section_map_size();
 453         struct page *map = sparse_buffer_alloc(size);
 454         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 455
 456         if (map)
 457                 return map;
 458
 459         map = memblock_alloc_try_nid(size,
 460                                           PAGE_SIZE, addr,
 461                                           MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 462         if (!map)
 463                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 464                       __func__, size, PAGE_SIZE, nid, &addr);
 465
 466         return map;
 467 }
 468 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 469
 470 static void *sparsemap_buf __meminitdata;
 471 static void *sparsemap_buf_end __meminitdata;
 472
 473 static void __init sparse_buffer_init(unsigned long size, int nid)
 474 {
 475         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 476         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 477         sparsemap_buf =
 478                 memblock_alloc_try_nid_raw(size, PAGE_SIZE,
 479                                                 addr,
 480                                                 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 481         sparsemap_buf_end = sparsemap_buf + size;
 482 }
 483
 484 static void __init sparse_buffer_fini(void)
 485 {
 486         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 487
 488         if (sparsemap_buf && size > 0)
 489                 memblock_free_early(__pa(sparsemap_buf), size);
 490         sparsemap_buf = NULL;
 491 }
 492
 493 void * __meminit sparse_buffer_alloc(unsigned long size)
 494 {
 495         void *ptr = NULL;
 496
 497         if (sparsemap_buf) {
 498                 ptr = PTR_ALIGN(sparsemap_buf, size);
 499                 if (ptr + size > sparsemap_buf_end)
 500                         ptr = NULL;
 501                 else
 502                         sparsemap_buf = ptr + size;
 503         }
 504         return ptr;
 505 }
 506
 507 void __weak __meminit vmemmap_populate_print_last(void)
 508 {
 509 }
 510
 511 /*
 512  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 513  * And number of present sections in this node is map_count.
 514  */
 515 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 516                                    unsigned long pnum_end,
 517                                    unsigned long map_count)
 518 {
 519         struct mem_section_usage *usage;
 520         unsigned long pnum;
 521         struct page *map;
 522
 523         usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 524                         mem_section_usage_size() * map_count);
 525         if (!usage) {
 526                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 527                 goto failed;
 528         }
 529         sparse_buffer_init(map_count * section_map_size(), nid);
 530         for_each_present_section_nr(pnum_begin, pnum) {
 531                 unsigned long pfn = section_nr_to_pfn(pnum);
 532
 533                 if (pnum >= pnum_end)
 534                         break;
 535
 536                 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
 537                                 nid, NULL);
 538                 if (!map) {
 539                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 540                                __func__, nid);
 541                         pnum_begin = pnum;
 542                         goto failed;
 543                 }
 544                 check_usemap_section_nr(nid, usage);
 545                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
 546                                 SECTION_IS_EARLY);
 547                 usage = (void *) usage + mem_section_usage_size();
 548         }
 549         sparse_buffer_fini();
 550         return;
 551 failed:
 552         /* We failed to allocate, mark all the following pnums as not present */
 553         for_each_present_section_nr(pnum_begin, pnum) {
 554                 struct mem_section *ms;
 555
 556                 if (pnum >= pnum_end)
 557                         break;
 558                 ms = __nr_to_section(pnum);
 559                 ms->section_mem_map = 0;
 560         }
 561 }
 562
 563 /*
 564  * Allocate the accumulated non-linear sections, allocate a mem_map
 565  * for each and record the physical to section mapping.
 566  */
 567 void __init sparse_init(void)
 568 {
 569         unsigned long pnum_begin = first_present_section_nr();
 570         int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 571         unsigned long pnum_end, map_count = 1;
 572
 573         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 574         set_pageblock_order();
 575
 576         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 577                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 578
 579                 if (nid == nid_begin) {
 580                         map_count++;
 581                         continue;
 582                 }
 583                 /* Init node with sections in range [pnum_begin, pnum_end) */
 584                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 585                 nid_begin = nid;
 586                 pnum_begin = pnum_end;
 587                 map_count = 1;
 588         }
 589         /* cover the last node */
 590         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 591         vmemmap_populate_print_last();
 592 }
 593
 594 #ifdef CONFIG_MEMORY_HOTPLUG
 595
 596 /* Mark all memory sections within the pfn range as online */
 597 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 598 {
 599         unsigned long pfn;
 600
 601         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 602                 unsigned long section_nr = pfn_to_section_nr(pfn);
 603                 struct mem_section *ms;
 604
 605                 /* onlining code should never touch invalid ranges */
 606                 if (WARN_ON(!valid_section_nr(section_nr)))
 607                         continue;
 608
 609                 ms = __nr_to_section(section_nr);
 610                 ms->section_mem_map |= SECTION_IS_ONLINE;
 611         }
 612 }
 613
 614 #ifdef CONFIG_MEMORY_HOTREMOVE
 615 /* Mark all memory sections within the pfn range as offline */
 616 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 617 {
 618         unsigned long pfn;
 619
 620         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 621                 unsigned long section_nr = pfn_to_section_nr(pfn);
 622                 struct mem_section *ms;
 623
 624                 /*
 625                  * TODO this needs some double checking. Offlining code makes
 626                  * sure to check pfn_valid but those checks might be just bogus
 627                  */
 628                 if (WARN_ON(!valid_section_nr(section_nr)))
 629                         continue;
 630
 631                 ms = __nr_to_section(section_nr);
 632                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 633         }
 634 }
 635 #endif
 636
 637 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 638 static struct page *populate_section_memmap(unsigned long pfn,
 639                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 640 {
 641         return __populate_section_memmap(pfn, nr_pages, nid, altmap);
 642 }
 643
 644 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 645                 struct vmem_altmap *altmap)
 646 {
 647         unsigned long start = (unsigned long) pfn_to_page(pfn);
 648         unsigned long end = start + nr_pages * sizeof(struct page);
 649
 650         vmemmap_free(start, end, altmap);
 651 }
 652 static void free_map_bootmem(struct page *memmap)
 653 {
 654         unsigned long start = (unsigned long)memmap;
 655         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 656
 657         vmemmap_free(start, end, NULL);
 658 }
 659 #else
 660 struct page *populate_section_memmap(unsigned long pfn,
 661                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 662 {
 663         struct page *page, *ret;
 664         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 665
 666         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 667         if (page)
 668                 goto got_map_page;
 669
 670         ret = vmalloc(memmap_size);
 671         if (ret)
 672                 goto got_map_ptr;
 673
 674         return NULL;
 675 got_map_page:
 676         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 677 got_map_ptr:
 678
 679         return ret;
 680 }
 681
 682 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 683                 struct vmem_altmap *altmap)
 684 {
 685         struct page *memmap = pfn_to_page(pfn);
 686
 687         if (is_vmalloc_addr(memmap))
 688                 vfree(memmap);
 689         else
 690                 free_pages((unsigned long)memmap,
 691                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 692 }
 693
 694 static void free_map_bootmem(struct page *memmap)
 695 {
 696         unsigned long maps_section_nr, removing_section_nr, i;
 697         unsigned long magic, nr_pages;
 698         struct page *page = virt_to_page(memmap);
 699
 700         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 701                 >> PAGE_SHIFT;
 702
 703         for (i = 0; i < nr_pages; i++, page++) {
 704                 magic = (unsigned long) page->freelist;
 705
 706                 BUG_ON(magic == NODE_INFO);
 707
 708                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 709                 removing_section_nr = page_private(page);
 710
 711                 /*
 712                  * When this function is called, the removing section is
 713                  * logical offlined state. This means all pages are isolated
 714                  * from page allocator. If removing section's memmap is placed
 715                  * on the same section, it must not be freed.
 716                  * If it is freed, page allocator may allocate it which will
 717                  * be removed physically soon.
 718                  */
 719                 if (maps_section_nr != removing_section_nr)
 720                         put_page_bootmem(page);
 721         }
 722 }
 723 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 724
 725 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
 726                 struct vmem_altmap *altmap)
 727 {
 728         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 729         DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
 730         struct mem_section *ms = __pfn_to_section(pfn);
 731         bool section_is_early = early_section(ms);
 732         struct page *memmap = NULL;
 733         unsigned long *subsection_map = ms->usage
 734                 ? &ms->usage->subsection_map[0] : NULL;
 735
 736         subsection_mask_set(map, pfn, nr_pages);
 737         if (subsection_map)
 738                 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
 739
 740         if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
 741                                 "section already deactivated (%#lx + %ld)\n",
 742                                 pfn, nr_pages))
 743                 return;
 744
 745         /*
 746          * There are 3 cases to handle across two configurations
 747          * (SPARSEMEM_VMEMMAP={y,n}):
 748          *
 749          * 1/ deactivation of a partial hot-added section (only possible
 750          * in the SPARSEMEM_VMEMMAP=y case).
 751          *    a/ section was present at memory init
 752          *    b/ section was hot-added post memory init
 753          * 2/ deactivation of a complete hot-added section
 754          * 3/ deactivation of a complete section from memory init
 755          *
 756          * For 1/, when subsection_map does not empty we will not be
 757          * freeing the usage map, but still need to free the vmemmap
 758          * range.
 759          *
 760          * For 2/ and 3/ the SPARSEMEM_VMEMMAP={y,n} cases are unified
 761          */
 762         bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
 763         if (bitmap_empty(subsection_map, SUBSECTIONS_PER_SECTION)) {
 764                 unsigned long section_nr = pfn_to_section_nr(pfn);
 765
 766                 if (!section_is_early) {
 767                         kfree(ms->usage);
 768                         ms->usage = NULL;
 769                 }
 770                 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 771                 ms->section_mem_map = sparse_encode_mem_map(NULL, section_nr);
 772         }
 773
 774         if (section_is_early && memmap)
 775                 free_map_bootmem(memmap);
 776         else
 777                 depopulate_section_memmap(pfn, nr_pages, altmap);
 778 }
 779
 780 static struct page * __meminit section_activate(int nid, unsigned long pfn,
 781                 unsigned long nr_pages, struct vmem_altmap *altmap)
 782 {
 783         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 784         struct mem_section *ms = __pfn_to_section(pfn);
 785         struct mem_section_usage *usage = NULL;
 786         unsigned long *subsection_map;
 787         struct page *memmap;
 788         int rc = 0;
 789
 790         subsection_mask_set(map, pfn, nr_pages);
 791
 792         if (!ms->usage) {
 793                 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
 794                 if (!usage)
 795                         return ERR_PTR(-ENOMEM);
 796                 ms->usage = usage;
 797         }
 798         subsection_map = &ms->usage->subsection_map[0];
 799
 800         if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
 801                 rc = -EINVAL;
 802         else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
 803                 rc = -EEXIST;
 804         else
 805                 bitmap_or(subsection_map, map, subsection_map,
 806                                 SUBSECTIONS_PER_SECTION);
 807
 808         if (rc) {
 809                 if (usage)
 810                         ms->usage = NULL;
 811                 kfree(usage);
 812                 return ERR_PTR(rc);
 813         }
 814
 815         /*
 816          * The early init code does not consider partially populated
 817          * initial sections, it simply assumes that memory will never be
 818          * referenced.  If we hot-add memory into such a section then we
 819          * do not need to populate the memmap and can simply reuse what
 820          * is already there.
 821          */
 822         if (nr_pages < PAGES_PER_SECTION && early_section(ms))
 823                 return pfn_to_page(pfn);
 824
 825         memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
 826         if (!memmap) {
 827                 section_deactivate(pfn, nr_pages, altmap);
 828                 return ERR_PTR(-ENOMEM);
 829         }
 830
 831         return memmap;
 832 }
 833
 834 /**
 835  * sparse_add_section - add a memory section, or populate an existing one
 836  * @nid: The node to add section on
 837  * @start_pfn: start pfn of the memory range
 838  * @nr_pages: number of pfns to add in the section
 839  * @altmap: device page map
 840  *
 841  * This is only intended for hotplug.
 842  *
 843  * Return:
 844  * * 0          - On success.
 845  * * -EEXIST    - Section has been present.
 846  * * -ENOMEM    - Out of memory.
 847  */
 848 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
 849                 unsigned long nr_pages, struct vmem_altmap *altmap)
 850 {
 851         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 852         struct mem_section *ms;
 853         struct page *memmap;
 854         int ret;
 855
 856         ret = sparse_index_init(section_nr, nid);
 857         if (ret < 0)
 858                 return ret;
 859
 860         memmap = section_activate(nid, start_pfn, nr_pages, altmap);
 861         if (IS_ERR(memmap))
 862                 return PTR_ERR(memmap);
 863
 864         /*
 865          * Poison uninitialized struct pages in order to catch invalid flags
 866          * combinations.
 867          */
 868         page_init_poison(pfn_to_page(start_pfn), sizeof(struct page) * nr_pages);
 869
 870         ms = __pfn_to_section(start_pfn);
 871         set_section_nid(section_nr, nid);
 872         section_mark_present(ms);
 873
 874         /* Align memmap to section boundary in the subsection case */
 875         if (section_nr_to_pfn(section_nr) != start_pfn)
 876                 memmap = pfn_to_kaddr(section_nr_to_pfn(section_nr));
 877         sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
 878
 879         return 0;
 880 }
 881
 882 #ifdef CONFIG_MEMORY_FAILURE
 883 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 884 {
 885         int i;
 886
 887         if (!memmap)
 888                 return;
 889
 890         /*
 891          * A further optimization is to have per section refcounted
 892          * num_poisoned_pages.  But that would need more space per memmap, so
 893          * for now just do a quick global check to speed up this routine in the
 894          * absence of bad pages.
 895          */
 896         if (atomic_long_read(&num_poisoned_pages) == 0)
 897                 return;
 898
 899         for (i = 0; i < nr_pages; i++) {
 900                 if (PageHWPoison(&memmap[i])) {
 901                         atomic_long_sub(1, &num_poisoned_pages);
 902                         ClearPageHWPoison(&memmap[i]);
 903                 }
 904         }
 905 }
 906 #else
 907 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 908 {
 909 }
 910 #endif
 911
 912 void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
 913                 unsigned long nr_pages, unsigned long map_offset,
 914                 struct vmem_altmap *altmap)
 915 {
 916         clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
 917                         nr_pages - map_offset);
 918         section_deactivate(pfn, nr_pages, altmap);
 919 }
 920 #endif /* CONFIG_MEMORY_HOTPLUG */