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 #include <linux/bootmem_info.h>
  17
  18 #include "internal.h"
  19 #include <asm/dma.h>
  20
  21 /*
  22  * Permanent SPARSEMEM data:
  23  *
  24  * 1) mem_section       - memory sections, mem_map's for valid memory
  25  */
  26 #ifdef CONFIG_SPARSEMEM_EXTREME
  27 struct mem_section **mem_section;
  28 #else
  29 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  30         ____cacheline_internodealigned_in_smp;
  31 #endif
  32 EXPORT_SYMBOL(mem_section);
  33
  34 #ifdef NODE_NOT_IN_PAGE_FLAGS
  35 /*
  36  * If we did not store the node number in the page then we have to
  37  * do a lookup in the section_to_node_table in order to find which
  38  * node the page belongs to.
  39  */
  40 #if MAX_NUMNODES <= 256
  41 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  42 #else
  43 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  44 #endif
  45
  46 int page_to_nid(const struct page *page)
  47 {
  48         return section_to_node_table[page_to_section(page)];
  49 }
  50 EXPORT_SYMBOL(page_to_nid);
  51
  52 static void set_section_nid(unsigned long section_nr, int nid)
  53 {
  54         section_to_node_table[section_nr] = nid;
  55 }
  56 #else /* !NODE_NOT_IN_PAGE_FLAGS */
  57 static inline void set_section_nid(unsigned long section_nr, int nid)
  58 {
  59 }
  60 #endif
  61
  62 #ifdef CONFIG_SPARSEMEM_EXTREME
  63 static noinline struct mem_section __ref *sparse_index_alloc(int nid)
  64 {
  65         struct mem_section *section = NULL;
  66         unsigned long array_size = SECTIONS_PER_ROOT *
  67                                    sizeof(struct mem_section);
  68
  69         if (slab_is_available()) {
  70                 section = kzalloc_node(array_size, GFP_KERNEL, nid);
  71         } else {
  72                 section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
  73                                               nid);
  74                 if (!section)
  75                         panic("%s: Failed to allocate %lu bytes nid=%d\n",
  76                               __func__, array_size, nid);
  77         }
  78
  79         return section;
  80 }
  81
  82 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  83 {
  84         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  85         struct mem_section *section;
  86
  87         /*
  88          * An existing section is possible in the sub-section hotplug
  89          * case. First hot-add instantiates, follow-on hot-add reuses
  90          * the existing section.
  91          *
  92          * The mem_hotplug_lock resolves the apparent race below.
  93          */
  94         if (mem_section[root])
  95                 return 0;
  96
  97         section = sparse_index_alloc(nid);
  98         if (!section)
  99                 return -ENOMEM;
 100
 101         mem_section[root] = section;
 102
 103         return 0;
 104 }
 105 #else /* !SPARSEMEM_EXTREME */
 106 static inline int sparse_index_init(unsigned long section_nr, int nid)
 107 {
 108         return 0;
 109 }
 110 #endif
 111
 112 #ifdef CONFIG_SPARSEMEM_EXTREME
 113 unsigned long __section_nr(struct mem_section *ms)
 114 {
 115         unsigned long root_nr;
 116         struct mem_section *root = NULL;
 117
 118         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 119                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 120                 if (!root)
 121                         continue;
 122
 123                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 124                      break;
 125         }
 126
 127         VM_BUG_ON(!root);
 128
 129         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 130 }
 131 #else
 132 unsigned long __section_nr(struct mem_section *ms)
 133 {
 134         return (unsigned long)(ms - mem_section[0]);
 135 }
 136 #endif
 137
 138 /*
 139  * During early boot, before section_mem_map is used for an actual
 140  * mem_map, we use section_mem_map to store the section's NUMA
 141  * node.  This keeps us from having to use another data structure.  The
 142  * node information is cleared just before we store the real mem_map.
 143  */
 144 static inline unsigned long sparse_encode_early_nid(int nid)
 145 {
 146         return (nid << SECTION_NID_SHIFT);
 147 }
 148
 149 static inline int sparse_early_nid(struct mem_section *section)
 150 {
 151         return (section->section_mem_map >> SECTION_NID_SHIFT);
 152 }
 153
 154 /* Validate the physical addressing limitations of the model */
 155 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 156                                                 unsigned long *end_pfn)
 157 {
 158         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 159
 160         /*
 161          * Sanity checks - do not allow an architecture to pass
 162          * in larger pfns than the maximum scope of sparsemem:
 163          */
 164         if (*start_pfn > max_sparsemem_pfn) {
 165                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 166                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 167                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 168                 WARN_ON_ONCE(1);
 169                 *start_pfn = max_sparsemem_pfn;
 170                 *end_pfn = max_sparsemem_pfn;
 171         } else if (*end_pfn > max_sparsemem_pfn) {
 172                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 173                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 174                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 175                 WARN_ON_ONCE(1);
 176                 *end_pfn = max_sparsemem_pfn;
 177         }
 178 }
 179
 180 /*
 181  * There are a number of times that we loop over NR_MEM_SECTIONS,
 182  * looking for section_present() on each.  But, when we have very
 183  * large physical address spaces, NR_MEM_SECTIONS can also be
 184  * very large which makes the loops quite long.
 185  *
 186  * Keeping track of this gives us an easy way to break out of
 187  * those loops early.
 188  */
 189 unsigned long __highest_present_section_nr;
 190 static void section_mark_present(struct mem_section *ms)
 191 {
 192         unsigned long section_nr = __section_nr(ms);
 193
 194         if (section_nr > __highest_present_section_nr)
 195                 __highest_present_section_nr = section_nr;
 196
 197         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 198 }
 199
 200 #define for_each_present_section_nr(start, section_nr)          \
 201         for (section_nr = next_present_section_nr(start-1);     \
 202              ((section_nr != -1) &&                             \
 203               (section_nr <= __highest_present_section_nr));    \
 204              section_nr = next_present_section_nr(section_nr))
 205
 206 static inline unsigned long first_present_section_nr(void)
 207 {
 208         return next_present_section_nr(-1);
 209 }
 210
 211 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 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 #else
 247 void __init subsection_map_init(unsigned long pfn, unsigned long nr_pages)
 248 {
 249 }
 250 #endif
 251
 252 /* Record a memory area against a node. */
 253 static void __init memory_present(int nid, unsigned long start, unsigned long end)
 254 {
 255         unsigned long pfn;
 256
 257 #ifdef CONFIG_SPARSEMEM_EXTREME
 258         if (unlikely(!mem_section)) {
 259                 unsigned long size, align;
 260
 261                 size = sizeof(struct mem_section *) * NR_SECTION_ROOTS;
 262                 align = 1 << (INTERNODE_CACHE_SHIFT);
 263                 mem_section = memblock_alloc(size, align);
 264                 if (!mem_section)
 265                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 266                               __func__, size, align);
 267         }
 268 #endif
 269
 270         start &= PAGE_SECTION_MASK;
 271         mminit_validate_memmodel_limits(&start, &end);
 272         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 273                 unsigned long section = pfn_to_section_nr(pfn);
 274                 struct mem_section *ms;
 275
 276                 sparse_index_init(section, nid);
 277                 set_section_nid(section, nid);
 278
 279                 ms = __nr_to_section(section);
 280                 if (!ms->section_mem_map) {
 281                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 282                                                         SECTION_IS_ONLINE;
 283                         section_mark_present(ms);
 284                 }
 285         }
 286 }
 287
 288 /*
 289  * Mark all memblocks as present using memory_present().
 290  * This is a convenience function that is useful to mark all of the systems
 291  * memory as present during initialization.
 292  */
 293 static void __init memblocks_present(void)
 294 {
 295         unsigned long start, end;
 296         int i, nid;
 297
 298         for_each_mem_pfn_range(i, MAX_NUMNODES, &start, &end, &nid)
 299                 memory_present(nid, start, end);
 300 }
 301
 302 /*
 303  * Subtle, we encode the real pfn into the mem_map such that
 304  * the identity pfn - section_mem_map will return the actual
 305  * physical page frame number.
 306  */
 307 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 308 {
 309         unsigned long coded_mem_map =
 310                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 311         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 312         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 313         return coded_mem_map;
 314 }
 315
 316 #ifdef CONFIG_MEMORY_HOTPLUG
 317 /*
 318  * Decode mem_map from the coded memmap
 319  */
 320 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 321 {
 322         /* mask off the extra low bits of information */
 323         coded_mem_map &= SECTION_MAP_MASK;
 324         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 325 }
 326 #endif /* CONFIG_MEMORY_HOTPLUG */
 327
 328 static void __meminit sparse_init_one_section(struct mem_section *ms,
 329                 unsigned long pnum, struct page *mem_map,
 330                 struct mem_section_usage *usage, unsigned long flags)
 331 {
 332         ms->section_mem_map &= ~SECTION_MAP_MASK;
 333         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum)
 334                 | SECTION_HAS_MEM_MAP | flags;
 335         ms->usage = usage;
 336 }
 337
 338 static unsigned long usemap_size(void)
 339 {
 340         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 341 }
 342
 343 size_t mem_section_usage_size(void)
 344 {
 345         return sizeof(struct mem_section_usage) + usemap_size();
 346 }
 347
 348 static inline phys_addr_t pgdat_to_phys(struct pglist_data *pgdat)
 349 {
 350 #ifndef CONFIG_NUMA
 351         return __pa_symbol(pgdat);
 352 #else
 353         return __pa(pgdat);
 354 #endif
 355 }
 356
 357 #ifdef CONFIG_MEMORY_HOTREMOVE
 358 static struct mem_section_usage * __init
 359 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 360                                          unsigned long size)
 361 {
 362         struct mem_section_usage *usage;
 363         unsigned long goal, limit;
 364         int nid;
 365         /*
 366          * A page may contain usemaps for other sections preventing the
 367          * page being freed and making a section unremovable while
 368          * other sections referencing the usemap remain active. Similarly,
 369          * a pgdat can prevent a section being removed. If section A
 370          * contains a pgdat and section B contains the usemap, both
 371          * sections become inter-dependent. This allocates usemaps
 372          * from the same section as the pgdat where possible to avoid
 373          * this problem.
 374          */
 375         goal = pgdat_to_phys(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 376         limit = goal + (1UL << PA_SECTION_SHIFT);
 377         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 378 again:
 379         usage = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 380         if (!usage && limit) {
 381                 limit = 0;
 382                 goto again;
 383         }
 384         return usage;
 385 }
 386
 387 static void __init check_usemap_section_nr(int nid,
 388                 struct mem_section_usage *usage)
 389 {
 390         unsigned long usemap_snr, pgdat_snr;
 391         static unsigned long old_usemap_snr;
 392         static unsigned long old_pgdat_snr;
 393         struct pglist_data *pgdat = NODE_DATA(nid);
 394         int usemap_nid;
 395
 396         /* First call */
 397         if (!old_usemap_snr) {
 398                 old_usemap_snr = NR_MEM_SECTIONS;
 399                 old_pgdat_snr = NR_MEM_SECTIONS;
 400         }
 401
 402         usemap_snr = pfn_to_section_nr(__pa(usage) >> PAGE_SHIFT);
 403         pgdat_snr = pfn_to_section_nr(pgdat_to_phys(pgdat) >> PAGE_SHIFT);
 404         if (usemap_snr == pgdat_snr)
 405                 return;
 406
 407         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 408                 /* skip redundant message */
 409                 return;
 410
 411         old_usemap_snr = usemap_snr;
 412         old_pgdat_snr = pgdat_snr;
 413
 414         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 415         if (usemap_nid != nid) {
 416                 pr_info("node %d must be removed before remove section %ld\n",
 417                         nid, usemap_snr);
 418                 return;
 419         }
 420         /*
 421          * There is a circular dependency.
 422          * Some platforms allow un-removable section because they will just
 423          * gather other removable sections for dynamic partitioning.
 424          * Just notify un-removable section's number here.
 425          */
 426         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 427                 usemap_snr, pgdat_snr, nid);
 428 }
 429 #else
 430 static struct mem_section_usage * __init
 431 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 432                                          unsigned long size)
 433 {
 434         return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 435 }
 436
 437 static void __init check_usemap_section_nr(int nid,
 438                 struct mem_section_usage *usage)
 439 {
 440 }
 441 #endif /* CONFIG_MEMORY_HOTREMOVE */
 442
 443 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 444 static unsigned long __init section_map_size(void)
 445 {
 446         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 447 }
 448
 449 #else
 450 static unsigned long __init section_map_size(void)
 451 {
 452         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 453 }
 454
 455 struct page __init *__populate_section_memmap(unsigned long pfn,
 456                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 457 {
 458         unsigned long size = section_map_size();
 459         struct page *map = sparse_buffer_alloc(size);
 460         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 461
 462         if (map)
 463                 return map;
 464
 465         map = memblock_alloc_try_nid_raw(size, size, addr,
 466                                           MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 467         if (!map)
 468                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 469                       __func__, size, PAGE_SIZE, nid, &addr);
 470
 471         return map;
 472 }
 473 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 474
 475 static void *sparsemap_buf __meminitdata;
 476 static void *sparsemap_buf_end __meminitdata;
 477
 478 static inline void __meminit sparse_buffer_free(unsigned long size)
 479 {
 480         WARN_ON(!sparsemap_buf || size == 0);
 481         memblock_free_early(__pa(sparsemap_buf), size);
 482 }
 483
 484 static void __init sparse_buffer_init(unsigned long size, int nid)
 485 {
 486         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 487         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 488         /*
 489          * Pre-allocated buffer is mainly used by __populate_section_memmap
 490          * and we want it to be properly aligned to the section size - this is
 491          * especially the case for VMEMMAP which maps memmap to PMDs
 492          */
 493         sparsemap_buf = memblock_alloc_exact_nid_raw(size, section_map_size(),
 494                                         addr, MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 495         sparsemap_buf_end = sparsemap_buf + size;
 496 }
 497
 498 static void __init sparse_buffer_fini(void)
 499 {
 500         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 501
 502         if (sparsemap_buf && size > 0)
 503                 sparse_buffer_free(size);
 504         sparsemap_buf = NULL;
 505 }
 506
 507 void * __meminit sparse_buffer_alloc(unsigned long size)
 508 {
 509         void *ptr = NULL;
 510
 511         if (sparsemap_buf) {
 512                 ptr = (void *) roundup((unsigned long)sparsemap_buf, size);
 513                 if (ptr + size > sparsemap_buf_end)
 514                         ptr = NULL;
 515                 else {
 516                         /* Free redundant aligned space */
 517                         if ((unsigned long)(ptr - sparsemap_buf) > 0)
 518                                 sparse_buffer_free((unsigned long)(ptr - sparsemap_buf));
 519                         sparsemap_buf = ptr + size;
 520                 }
 521         }
 522         return ptr;
 523 }
 524
 525 void __weak __meminit vmemmap_populate_print_last(void)
 526 {
 527 }
 528
 529 /*
 530  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 531  * And number of present sections in this node is map_count.
 532  */
 533 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 534                                    unsigned long pnum_end,
 535                                    unsigned long map_count)
 536 {
 537         struct mem_section_usage *usage;
 538         unsigned long pnum;
 539         struct page *map;
 540
 541         usage = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 542                         mem_section_usage_size() * map_count);
 543         if (!usage) {
 544                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 545                 goto failed;
 546         }
 547         sparse_buffer_init(map_count * section_map_size(), nid);
 548         for_each_present_section_nr(pnum_begin, pnum) {
 549                 unsigned long pfn = section_nr_to_pfn(pnum);
 550
 551                 if (pnum >= pnum_end)
 552                         break;
 553
 554                 map = __populate_section_memmap(pfn, PAGES_PER_SECTION,
 555                                 nid, NULL);
 556                 if (!map) {
 557                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 558                                __func__, nid);
 559                         pnum_begin = pnum;
 560                         sparse_buffer_fini();
 561                         goto failed;
 562                 }
 563                 check_usemap_section_nr(nid, usage);
 564                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usage,
 565                                 SECTION_IS_EARLY);
 566                 usage = (void *) usage + mem_section_usage_size();
 567         }
 568         sparse_buffer_fini();
 569         return;
 570 failed:
 571         /* We failed to allocate, mark all the following pnums as not present */
 572         for_each_present_section_nr(pnum_begin, pnum) {
 573                 struct mem_section *ms;
 574
 575                 if (pnum >= pnum_end)
 576                         break;
 577                 ms = __nr_to_section(pnum);
 578                 ms->section_mem_map = 0;
 579         }
 580 }
 581
 582 /*
 583  * Allocate the accumulated non-linear sections, allocate a mem_map
 584  * for each and record the physical to section mapping.
 585  */
 586 void __init sparse_init(void)
 587 {
 588         unsigned long pnum_end, pnum_begin, map_count = 1;
 589         int nid_begin;
 590
 591         memblocks_present();
 592
 593         pnum_begin = first_present_section_nr();
 594         nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 595
 596         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 597         set_pageblock_order();
 598
 599         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 600                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 601
 602                 if (nid == nid_begin) {
 603                         map_count++;
 604                         continue;
 605                 }
 606                 /* Init node with sections in range [pnum_begin, pnum_end) */
 607                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 608                 nid_begin = nid;
 609                 pnum_begin = pnum_end;
 610                 map_count = 1;
 611         }
 612         /* cover the last node */
 613         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 614         vmemmap_populate_print_last();
 615 }
 616
 617 #ifdef CONFIG_MEMORY_HOTPLUG
 618
 619 /* Mark all memory sections within the pfn range as online */
 620 void online_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                 /* onlining code should never touch invalid ranges */
 629                 if (WARN_ON(!valid_section_nr(section_nr)))
 630                         continue;
 631
 632                 ms = __nr_to_section(section_nr);
 633                 ms->section_mem_map |= SECTION_IS_ONLINE;
 634         }
 635 }
 636
 637 /* Mark all memory sections within the pfn range as offline */
 638 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 639 {
 640         unsigned long pfn;
 641
 642         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 643                 unsigned long section_nr = pfn_to_section_nr(pfn);
 644                 struct mem_section *ms;
 645
 646                 /*
 647                  * TODO this needs some double checking. Offlining code makes
 648                  * sure to check pfn_valid but those checks might be just bogus
 649                  */
 650                 if (WARN_ON(!valid_section_nr(section_nr)))
 651                         continue;
 652
 653                 ms = __nr_to_section(section_nr);
 654                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 655         }
 656 }
 657
 658 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 659 static struct page * __meminit populate_section_memmap(unsigned long pfn,
 660                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 661 {
 662         return __populate_section_memmap(pfn, nr_pages, nid, altmap);
 663 }
 664
 665 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 666                 struct vmem_altmap *altmap)
 667 {
 668         unsigned long start = (unsigned long) pfn_to_page(pfn);
 669         unsigned long end = start + nr_pages * sizeof(struct page);
 670
 671         vmemmap_free(start, end, altmap);
 672 }
 673 static void free_map_bootmem(struct page *memmap)
 674 {
 675         unsigned long start = (unsigned long)memmap;
 676         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 677
 678         vmemmap_free(start, end, NULL);
 679 }
 680
 681 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 682 {
 683         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 684         DECLARE_BITMAP(tmp, SUBSECTIONS_PER_SECTION) = { 0 };
 685         struct mem_section *ms = __pfn_to_section(pfn);
 686         unsigned long *subsection_map = ms->usage
 687                 ? &ms->usage->subsection_map[0] : NULL;
 688
 689         subsection_mask_set(map, pfn, nr_pages);
 690         if (subsection_map)
 691                 bitmap_and(tmp, map, subsection_map, SUBSECTIONS_PER_SECTION);
 692
 693         if (WARN(!subsection_map || !bitmap_equal(tmp, map, SUBSECTIONS_PER_SECTION),
 694                                 "section already deactivated (%#lx + %ld)\n",
 695                                 pfn, nr_pages))
 696                 return -EINVAL;
 697
 698         bitmap_xor(subsection_map, map, subsection_map, SUBSECTIONS_PER_SECTION);
 699         return 0;
 700 }
 701
 702 static bool is_subsection_map_empty(struct mem_section *ms)
 703 {
 704         return bitmap_empty(&ms->usage->subsection_map[0],
 705                             SUBSECTIONS_PER_SECTION);
 706 }
 707
 708 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 709 {
 710         struct mem_section *ms = __pfn_to_section(pfn);
 711         DECLARE_BITMAP(map, SUBSECTIONS_PER_SECTION) = { 0 };
 712         unsigned long *subsection_map;
 713         int rc = 0;
 714
 715         subsection_mask_set(map, pfn, nr_pages);
 716
 717         subsection_map = &ms->usage->subsection_map[0];
 718
 719         if (bitmap_empty(map, SUBSECTIONS_PER_SECTION))
 720                 rc = -EINVAL;
 721         else if (bitmap_intersects(map, subsection_map, SUBSECTIONS_PER_SECTION))
 722                 rc = -EEXIST;
 723         else
 724                 bitmap_or(subsection_map, map, subsection_map,
 725                                 SUBSECTIONS_PER_SECTION);
 726
 727         return rc;
 728 }
 729 #else
 730 struct page * __meminit populate_section_memmap(unsigned long pfn,
 731                 unsigned long nr_pages, int nid, struct vmem_altmap *altmap)
 732 {
 733         return kvmalloc_node(array_size(sizeof(struct page),
 734                                         PAGES_PER_SECTION), GFP_KERNEL, nid);
 735 }
 736
 737 static void depopulate_section_memmap(unsigned long pfn, unsigned long nr_pages,
 738                 struct vmem_altmap *altmap)
 739 {
 740         kvfree(pfn_to_page(pfn));
 741 }
 742
 743 static void free_map_bootmem(struct page *memmap)
 744 {
 745         unsigned long maps_section_nr, removing_section_nr, i;
 746         unsigned long magic, nr_pages;
 747         struct page *page = virt_to_page(memmap);
 748
 749         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 750                 >> PAGE_SHIFT;
 751
 752         for (i = 0; i < nr_pages; i++, page++) {
 753                 magic = (unsigned long) page->freelist;
 754
 755                 BUG_ON(magic == NODE_INFO);
 756
 757                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 758                 removing_section_nr = page_private(page);
 759
 760                 /*
 761                  * When this function is called, the removing section is
 762                  * logical offlined state. This means all pages are isolated
 763                  * from page allocator. If removing section's memmap is placed
 764                  * on the same section, it must not be freed.
 765                  * If it is freed, page allocator may allocate it which will
 766                  * be removed physically soon.
 767                  */
 768                 if (maps_section_nr != removing_section_nr)
 769                         put_page_bootmem(page);
 770         }
 771 }
 772
 773 static int clear_subsection_map(unsigned long pfn, unsigned long nr_pages)
 774 {
 775         return 0;
 776 }
 777
 778 static bool is_subsection_map_empty(struct mem_section *ms)
 779 {
 780         return true;
 781 }
 782
 783 static int fill_subsection_map(unsigned long pfn, unsigned long nr_pages)
 784 {
 785         return 0;
 786 }
 787 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 788
 789 /*
 790  * To deactivate a memory region, there are 3 cases to handle across
 791  * two configurations (SPARSEMEM_VMEMMAP={y,n}):
 792  *
 793  * 1. deactivation of a partial hot-added section (only possible in
 794  *    the SPARSEMEM_VMEMMAP=y case).
 795  *      a) section was present at memory init.
 796  *      b) section was hot-added post memory init.
 797  * 2. deactivation of a complete hot-added section.
 798  * 3. deactivation of a complete section from memory init.
 799  *
 800  * For 1, when subsection_map does not empty we will not be freeing the
 801  * usage map, but still need to free the vmemmap range.
 802  *
 803  * For 2 and 3, the SPARSEMEM_VMEMMAP={y,n} cases are unified
 804  */
 805 static void section_deactivate(unsigned long pfn, unsigned long nr_pages,
 806                 struct vmem_altmap *altmap)
 807 {
 808         struct mem_section *ms = __pfn_to_section(pfn);
 809         bool section_is_early = early_section(ms);
 810         struct page *memmap = NULL;
 811         bool empty;
 812
 813         if (clear_subsection_map(pfn, nr_pages))
 814                 return;
 815
 816         empty = is_subsection_map_empty(ms);
 817         if (empty) {
 818                 unsigned long section_nr = pfn_to_section_nr(pfn);
 819
 820                 /*
 821                  * When removing an early section, the usage map is kept (as the
 822                  * usage maps of other sections fall into the same page). It
 823                  * will be re-used when re-adding the section - which is then no
 824                  * longer an early section. If the usage map is PageReserved, it
 825                  * was allocated during boot.
 826                  */
 827                 if (!PageReserved(virt_to_page(ms->usage))) {
 828                         kfree(ms->usage);
 829                         ms->usage = NULL;
 830                 }
 831                 memmap = sparse_decode_mem_map(ms->section_mem_map, section_nr);
 832                 /*
 833                  * Mark the section invalid so that valid_section()
 834                  * return false. This prevents code from dereferencing
 835                  * ms->usage array.
 836                  */
 837                 ms->section_mem_map &= ~SECTION_HAS_MEM_MAP;
 838         }
 839
 840         /*
 841          * The memmap of early sections is always fully populated. See
 842          * section_activate() and pfn_valid() .
 843          */
 844         if (!section_is_early)
 845                 depopulate_section_memmap(pfn, nr_pages, altmap);
 846         else if (memmap)
 847                 free_map_bootmem(memmap);
 848
 849         if (empty)
 850                 ms->section_mem_map = (unsigned long)NULL;
 851 }
 852
 853 static struct page * __meminit section_activate(int nid, unsigned long pfn,
 854                 unsigned long nr_pages, struct vmem_altmap *altmap)
 855 {
 856         struct mem_section *ms = __pfn_to_section(pfn);
 857         struct mem_section_usage *usage = NULL;
 858         struct page *memmap;
 859         int rc = 0;
 860
 861         if (!ms->usage) {
 862                 usage = kzalloc(mem_section_usage_size(), GFP_KERNEL);
 863                 if (!usage)
 864                         return ERR_PTR(-ENOMEM);
 865                 ms->usage = usage;
 866         }
 867
 868         rc = fill_subsection_map(pfn, nr_pages);
 869         if (rc) {
 870                 if (usage)
 871                         ms->usage = NULL;
 872                 kfree(usage);
 873                 return ERR_PTR(rc);
 874         }
 875
 876         /*
 877          * The early init code does not consider partially populated
 878          * initial sections, it simply assumes that memory will never be
 879          * referenced.  If we hot-add memory into such a section then we
 880          * do not need to populate the memmap and can simply reuse what
 881          * is already there.
 882          */
 883         if (nr_pages < PAGES_PER_SECTION && early_section(ms))
 884                 return pfn_to_page(pfn);
 885
 886         memmap = populate_section_memmap(pfn, nr_pages, nid, altmap);
 887         if (!memmap) {
 888                 section_deactivate(pfn, nr_pages, altmap);
 889                 return ERR_PTR(-ENOMEM);
 890         }
 891
 892         return memmap;
 893 }
 894
 895 /**
 896  * sparse_add_section - add a memory section, or populate an existing one
 897  * @nid: The node to add section on
 898  * @start_pfn: start pfn of the memory range
 899  * @nr_pages: number of pfns to add in the section
 900  * @altmap: device page map
 901  *
 902  * This is only intended for hotplug.
 903  *
 904  * Note that only VMEMMAP supports sub-section aligned hotplug,
 905  * the proper alignment and size are gated by check_pfn_span().
 906  *
 907  *
 908  * Return:
 909  * * 0          - On success.
 910  * * -EEXIST    - Section has been present.
 911  * * -ENOMEM    - Out of memory.
 912  */
 913 int __meminit sparse_add_section(int nid, unsigned long start_pfn,
 914                 unsigned long nr_pages, struct vmem_altmap *altmap)
 915 {
 916         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 917         struct mem_section *ms;
 918         struct page *memmap;
 919         int ret;
 920
 921         ret = sparse_index_init(section_nr, nid);
 922         if (ret < 0)
 923                 return ret;
 924
 925         memmap = section_activate(nid, start_pfn, nr_pages, altmap);
 926         if (IS_ERR(memmap))
 927                 return PTR_ERR(memmap);
 928
 929         /*
 930          * Poison uninitialized struct pages in order to catch invalid flags
 931          * combinations.
 932          */
 933         page_init_poison(memmap, sizeof(struct page) * nr_pages);
 934
 935         ms = __nr_to_section(section_nr);
 936         set_section_nid(section_nr, nid);
 937         section_mark_present(ms);
 938
 939         /* Align memmap to section boundary in the subsection case */
 940         if (section_nr_to_pfn(section_nr) != start_pfn)
 941                 memmap = pfn_to_page(section_nr_to_pfn(section_nr));
 942         sparse_init_one_section(ms, section_nr, memmap, ms->usage, 0);
 943
 944         return 0;
 945 }
 946
 947 #ifdef CONFIG_MEMORY_FAILURE
 948 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 949 {
 950         int i;
 951
 952         /*
 953          * A further optimization is to have per section refcounted
 954          * num_poisoned_pages.  But that would need more space per memmap, so
 955          * for now just do a quick global check to speed up this routine in the
 956          * absence of bad pages.
 957          */
 958         if (atomic_long_read(&num_poisoned_pages) == 0)
 959                 return;
 960
 961         for (i = 0; i < nr_pages; i++) {
 962                 if (PageHWPoison(&memmap[i])) {
 963                         num_poisoned_pages_dec();
 964                         ClearPageHWPoison(&memmap[i]);
 965                 }
 966         }
 967 }
 968 #else
 969 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 970 {
 971 }
 972 #endif
 973
 974 void sparse_remove_section(struct mem_section *ms, unsigned long pfn,
 975                 unsigned long nr_pages, unsigned long map_offset,
 976                 struct vmem_altmap *altmap)
 977 {
 978         clear_hwpoisoned_pages(pfn_to_page(pfn) + map_offset,
 979                         nr_pages - map_offset);
 980         section_deactivate(pfn, nr_pages, altmap);
 981 }
 982 #endif /* CONFIG_MEMORY_HOTPLUG */