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