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         if (mem_section[root])
  87                 return -EEXIST;
  88
  89         section = sparse_index_alloc(nid);
  90         if (!section)
  91                 return -ENOMEM;
  92
  93         mem_section[root] = section;
  94
  95         return 0;
  96 }
  97 #else /* !SPARSEMEM_EXTREME */
  98 static inline int sparse_index_init(unsigned long section_nr, int nid)
  99 {
 100         return 0;
 101 }
 102 #endif
 103
 104 #ifdef CONFIG_SPARSEMEM_EXTREME
 105 int __section_nr(struct mem_section* ms)
 106 {
 107         unsigned long root_nr;
 108         struct mem_section *root = NULL;
 109
 110         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 111                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 112                 if (!root)
 113                         continue;
 114
 115                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 116                      break;
 117         }
 118
 119         VM_BUG_ON(!root);
 120
 121         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 122 }
 123 #else
 124 int __section_nr(struct mem_section* ms)
 125 {
 126         return (int)(ms - mem_section[0]);
 127 }
 128 #endif
 129
 130 /*
 131  * During early boot, before section_mem_map is used for an actual
 132  * mem_map, we use section_mem_map to store the section's NUMA
 133  * node.  This keeps us from having to use another data structure.  The
 134  * node information is cleared just before we store the real mem_map.
 135  */
 136 static inline unsigned long sparse_encode_early_nid(int nid)
 137 {
 138         return (nid << SECTION_NID_SHIFT);
 139 }
 140
 141 static inline int sparse_early_nid(struct mem_section *section)
 142 {
 143         return (section->section_mem_map >> SECTION_NID_SHIFT);
 144 }
 145
 146 /* Validate the physical addressing limitations of the model */
 147 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 148                                                 unsigned long *end_pfn)
 149 {
 150         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 151
 152         /*
 153          * Sanity checks - do not allow an architecture to pass
 154          * in larger pfns than the maximum scope of sparsemem:
 155          */
 156         if (*start_pfn > max_sparsemem_pfn) {
 157                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 158                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 159                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 160                 WARN_ON_ONCE(1);
 161                 *start_pfn = max_sparsemem_pfn;
 162                 *end_pfn = max_sparsemem_pfn;
 163         } else if (*end_pfn > max_sparsemem_pfn) {
 164                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 165                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 166                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 167                 WARN_ON_ONCE(1);
 168                 *end_pfn = max_sparsemem_pfn;
 169         }
 170 }
 171
 172 /*
 173  * There are a number of times that we loop over NR_MEM_SECTIONS,
 174  * looking for section_present() on each.  But, when we have very
 175  * large physical address spaces, NR_MEM_SECTIONS can also be
 176  * very large which makes the loops quite long.
 177  *
 178  * Keeping track of this gives us an easy way to break out of
 179  * those loops early.
 180  */
 181 int __highest_present_section_nr;
 182 static void section_mark_present(struct mem_section *ms)
 183 {
 184         int section_nr = __section_nr(ms);
 185
 186         if (section_nr > __highest_present_section_nr)
 187                 __highest_present_section_nr = section_nr;
 188
 189         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 190 }
 191
 192 static inline int next_present_section_nr(int section_nr)
 193 {
 194         do {
 195                 section_nr++;
 196                 if (present_section_nr(section_nr))
 197                         return section_nr;
 198         } while ((section_nr <= __highest_present_section_nr));
 199
 200         return -1;
 201 }
 202 #define for_each_present_section_nr(start, section_nr)          \
 203         for (section_nr = next_present_section_nr(start-1);     \
 204              ((section_nr != -1) &&                             \
 205               (section_nr <= __highest_present_section_nr));    \
 206              section_nr = next_present_section_nr(section_nr))
 207
 208 static inline unsigned long first_present_section_nr(void)
 209 {
 210         return next_present_section_nr(-1);
 211 }
 212
 213 /* Record a memory area against a node. */
 214 void __init memory_present(int nid, unsigned long start, unsigned long end)
 215 {
 216         unsigned long pfn;
 217
 218 #ifdef CONFIG_SPARSEMEM_EXTREME
 219         if (unlikely(!mem_section)) {
 220                 unsigned long size, align;
 221
 222                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
 223                 align = 1 << (INTERNODE_CACHE_SHIFT);
 224                 mem_section = memblock_alloc(size, align);
 225                 if (!mem_section)
 226                         panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 227                               __func__, size, align);
 228         }
 229 #endif
 230
 231         start &= PAGE_SECTION_MASK;
 232         mminit_validate_memmodel_limits(&start, &end);
 233         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 234                 unsigned long section = pfn_to_section_nr(pfn);
 235                 struct mem_section *ms;
 236
 237                 sparse_index_init(section, nid);
 238                 set_section_nid(section, nid);
 239
 240                 ms = __nr_to_section(section);
 241                 if (!ms->section_mem_map) {
 242                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 243                                                         SECTION_IS_ONLINE;
 244                         section_mark_present(ms);
 245                 }
 246         }
 247 }
 248
 249 /*
 250  * Mark all memblocks as present using memory_present(). This is a
 251  * convienence function that is useful for a number of arches
 252  * to mark all of the systems memory as present during initialization.
 253  */
 254 void __init memblocks_present(void)
 255 {
 256         struct memblock_region *reg;
 257
 258         for_each_memblock(memory, reg) {
 259                 memory_present(memblock_get_region_node(reg),
 260                                memblock_region_memory_base_pfn(reg),
 261                                memblock_region_memory_end_pfn(reg));
 262         }
 263 }
 264
 265 /*
 266  * Subtle, we encode the real pfn into the mem_map such that
 267  * the identity pfn - section_mem_map will return the actual
 268  * physical page frame number.
 269  */
 270 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 271 {
 272         unsigned long coded_mem_map =
 273                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 274         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 275         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 276         return coded_mem_map;
 277 }
 278
 279 /*
 280  * Decode mem_map from the coded memmap
 281  */
 282 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 283 {
 284         /* mask off the extra low bits of information */
 285         coded_mem_map &= SECTION_MAP_MASK;
 286         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 287 }
 288
 289 static void __meminit sparse_init_one_section(struct mem_section *ms,
 290                 unsigned long pnum, struct page *mem_map,
 291                 unsigned long *pageblock_bitmap)
 292 {
 293         ms->section_mem_map &= ~SECTION_MAP_MASK;
 294         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 295                                                         SECTION_HAS_MEM_MAP;
 296         ms->pageblock_flags = pageblock_bitmap;
 297 }
 298
 299 unsigned long usemap_size(void)
 300 {
 301         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 302 }
 303
 304 #ifdef CONFIG_MEMORY_HOTPLUG
 305 static unsigned long *__kmalloc_section_usemap(void)
 306 {
 307         return kmalloc(usemap_size(), GFP_KERNEL);
 308 }
 309 #endif /* CONFIG_MEMORY_HOTPLUG */
 310
 311 #ifdef CONFIG_MEMORY_HOTREMOVE
 312 static unsigned long * __init
 313 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 314                                          unsigned long size)
 315 {
 316         unsigned long goal, limit;
 317         unsigned long *p;
 318         int nid;
 319         /*
 320          * A page may contain usemaps for other sections preventing the
 321          * page being freed and making a section unremovable while
 322          * other sections referencing the usemap remain active. Similarly,
 323          * a pgdat can prevent a section being removed. If section A
 324          * contains a pgdat and section B contains the usemap, both
 325          * sections become inter-dependent. This allocates usemaps
 326          * from the same section as the pgdat where possible to avoid
 327          * this problem.
 328          */
 329         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 330         limit = goal + (1UL << PA_SECTION_SHIFT);
 331         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 332 again:
 333         p = memblock_alloc_try_nid_nopanic(size,
 334                                                 SMP_CACHE_BYTES, goal, limit,
 335                                                 nid);
 336         if (!p && limit) {
 337                 limit = 0;
 338                 goto again;
 339         }
 340         return p;
 341 }
 342
 343 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 344 {
 345         unsigned long usemap_snr, pgdat_snr;
 346         static unsigned long old_usemap_snr;
 347         static unsigned long old_pgdat_snr;
 348         struct pglist_data *pgdat = NODE_DATA(nid);
 349         int usemap_nid;
 350
 351         /* First call */
 352         if (!old_usemap_snr) {
 353                 old_usemap_snr = NR_MEM_SECTIONS;
 354                 old_pgdat_snr = NR_MEM_SECTIONS;
 355         }
 356
 357         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 358         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 359         if (usemap_snr == pgdat_snr)
 360                 return;
 361
 362         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 363                 /* skip redundant message */
 364                 return;
 365
 366         old_usemap_snr = usemap_snr;
 367         old_pgdat_snr = pgdat_snr;
 368
 369         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 370         if (usemap_nid != nid) {
 371                 pr_info("node %d must be removed before remove section %ld\n",
 372                         nid, usemap_snr);
 373                 return;
 374         }
 375         /*
 376          * There is a circular dependency.
 377          * Some platforms allow un-removable section because they will just
 378          * gather other removable sections for dynamic partitioning.
 379          * Just notify un-removable section's number here.
 380          */
 381         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 382                 usemap_snr, pgdat_snr, nid);
 383 }
 384 #else
 385 static unsigned long * __init
 386 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 387                                          unsigned long size)
 388 {
 389         return memblock_alloc_node_nopanic(size, pgdat->node_id);
 390 }
 391
 392 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 393 {
 394 }
 395 #endif /* CONFIG_MEMORY_HOTREMOVE */
 396
 397 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 398 static unsigned long __init section_map_size(void)
 399 {
 400         return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 401 }
 402
 403 #else
 404 static unsigned long __init section_map_size(void)
 405 {
 406         return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 407 }
 408
 409 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
 410                 struct vmem_altmap *altmap)
 411 {
 412         unsigned long size = section_map_size();
 413         struct page *map = sparse_buffer_alloc(size);
 414         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 415
 416         if (map)
 417                 return map;
 418
 419         map = memblock_alloc_try_nid(size,
 420                                           PAGE_SIZE, addr,
 421                                           MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 422         if (!map)
 423                 panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 424                       __func__, size, PAGE_SIZE, nid, &addr);
 425
 426         return map;
 427 }
 428 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 429
 430 static void *sparsemap_buf __meminitdata;
 431 static void *sparsemap_buf_end __meminitdata;
 432
 433 static void __init sparse_buffer_init(unsigned long size, int nid)
 434 {
 435         phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 436         WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 437         sparsemap_buf =
 438                 memblock_alloc_try_nid_raw(size, PAGE_SIZE,
 439                                                 addr,
 440                                                 MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 441         sparsemap_buf_end = sparsemap_buf + size;
 442 }
 443
 444 static void __init sparse_buffer_fini(void)
 445 {
 446         unsigned long size = sparsemap_buf_end - sparsemap_buf;
 447
 448         if (sparsemap_buf && size > 0)
 449                 memblock_free_early(__pa(sparsemap_buf), size);
 450         sparsemap_buf = NULL;
 451 }
 452
 453 void * __meminit sparse_buffer_alloc(unsigned long size)
 454 {
 455         void *ptr = NULL;
 456
 457         if (sparsemap_buf) {
 458                 ptr = PTR_ALIGN(sparsemap_buf, size);
 459                 if (ptr + size > sparsemap_buf_end)
 460                         ptr = NULL;
 461                 else
 462                         sparsemap_buf = ptr + size;
 463         }
 464         return ptr;
 465 }
 466
 467 void __weak __meminit vmemmap_populate_print_last(void)
 468 {
 469 }
 470
 471 /*
 472  * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 473  * And number of present sections in this node is map_count.
 474  */
 475 static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 476                                    unsigned long pnum_end,
 477                                    unsigned long map_count)
 478 {
 479         unsigned long pnum, usemap_longs, *usemap;
 480         struct page *map;
 481
 482         usemap_longs = BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS);
 483         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 484                                                           usemap_size() *
 485                                                           map_count);
 486         if (!usemap) {
 487                 pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 488                 goto failed;
 489         }
 490         sparse_buffer_init(map_count * section_map_size(), nid);
 491         for_each_present_section_nr(pnum_begin, pnum) {
 492                 if (pnum >= pnum_end)
 493                         break;
 494
 495                 map = sparse_mem_map_populate(pnum, nid, NULL);
 496                 if (!map) {
 497                         pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 498                                __func__, nid);
 499                         pnum_begin = pnum;
 500                         goto failed;
 501                 }
 502                 check_usemap_section_nr(nid, usemap);
 503                 sparse_init_one_section(__nr_to_section(pnum), pnum, map, usemap);
 504                 usemap += usemap_longs;
 505         }
 506         sparse_buffer_fini();
 507         return;
 508 failed:
 509         /* We failed to allocate, mark all the following pnums as not present */
 510         for_each_present_section_nr(pnum_begin, pnum) {
 511                 struct mem_section *ms;
 512
 513                 if (pnum >= pnum_end)
 514                         break;
 515                 ms = __nr_to_section(pnum);
 516                 ms->section_mem_map = 0;
 517         }
 518 }
 519
 520 /*
 521  * Allocate the accumulated non-linear sections, allocate a mem_map
 522  * for each and record the physical to section mapping.
 523  */
 524 void __init sparse_init(void)
 525 {
 526         unsigned long pnum_begin = first_present_section_nr();
 527         int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 528         unsigned long pnum_end, map_count = 1;
 529
 530         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 531         set_pageblock_order();
 532
 533         for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 534                 int nid = sparse_early_nid(__nr_to_section(pnum_end));
 535
 536                 if (nid == nid_begin) {
 537                         map_count++;
 538                         continue;
 539                 }
 540                 /* Init node with sections in range [pnum_begin, pnum_end) */
 541                 sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 542                 nid_begin = nid;
 543                 pnum_begin = pnum_end;
 544                 map_count = 1;
 545         }
 546         /* cover the last node */
 547         sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 548         vmemmap_populate_print_last();
 549 }
 550
 551 #ifdef CONFIG_MEMORY_HOTPLUG
 552
 553 /* Mark all memory sections within the pfn range as online */
 554 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 555 {
 556         unsigned long pfn;
 557
 558         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 559                 unsigned long section_nr = pfn_to_section_nr(pfn);
 560                 struct mem_section *ms;
 561
 562                 /* onlining code should never touch invalid ranges */
 563                 if (WARN_ON(!valid_section_nr(section_nr)))
 564                         continue;
 565
 566                 ms = __nr_to_section(section_nr);
 567                 ms->section_mem_map |= SECTION_IS_ONLINE;
 568         }
 569 }
 570
 571 #ifdef CONFIG_MEMORY_HOTREMOVE
 572 /* Mark all memory sections within the pfn range as online */
 573 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 574 {
 575         unsigned long pfn;
 576
 577         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 578                 unsigned long section_nr = pfn_to_section_nr(pfn);
 579                 struct mem_section *ms;
 580
 581                 /*
 582                  * TODO this needs some double checking. Offlining code makes
 583                  * sure to check pfn_valid but those checks might be just bogus
 584                  */
 585                 if (WARN_ON(!valid_section_nr(section_nr)))
 586                         continue;
 587
 588                 ms = __nr_to_section(section_nr);
 589                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 590         }
 591 }
 592 #endif
 593
 594 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 595 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 596                 struct vmem_altmap *altmap)
 597 {
 598         /* This will make the necessary allocations eventually. */
 599         return sparse_mem_map_populate(pnum, nid, altmap);
 600 }
 601 static void __kfree_section_memmap(struct page *memmap,
 602                 struct vmem_altmap *altmap)
 603 {
 604         unsigned long start = (unsigned long)memmap;
 605         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 606
 607         vmemmap_free(start, end, altmap);
 608 }
 609 #ifdef CONFIG_MEMORY_HOTREMOVE
 610 static void free_map_bootmem(struct page *memmap)
 611 {
 612         unsigned long start = (unsigned long)memmap;
 613         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 614
 615         vmemmap_free(start, end, NULL);
 616 }
 617 #endif /* CONFIG_MEMORY_HOTREMOVE */
 618 #else
 619 static struct page *__kmalloc_section_memmap(void)
 620 {
 621         struct page *page, *ret;
 622         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 623
 624         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 625         if (page)
 626                 goto got_map_page;
 627
 628         ret = vmalloc(memmap_size);
 629         if (ret)
 630                 goto got_map_ptr;
 631
 632         return NULL;
 633 got_map_page:
 634         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 635 got_map_ptr:
 636
 637         return ret;
 638 }
 639
 640 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 641                 struct vmem_altmap *altmap)
 642 {
 643         return __kmalloc_section_memmap();
 644 }
 645
 646 static void __kfree_section_memmap(struct page *memmap,
 647                 struct vmem_altmap *altmap)
 648 {
 649         if (is_vmalloc_addr(memmap))
 650                 vfree(memmap);
 651         else
 652                 free_pages((unsigned long)memmap,
 653                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 654 }
 655
 656 #ifdef CONFIG_MEMORY_HOTREMOVE
 657 static void free_map_bootmem(struct page *memmap)
 658 {
 659         unsigned long maps_section_nr, removing_section_nr, i;
 660         unsigned long magic, nr_pages;
 661         struct page *page = virt_to_page(memmap);
 662
 663         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 664                 >> PAGE_SHIFT;
 665
 666         for (i = 0; i < nr_pages; i++, page++) {
 667                 magic = (unsigned long) page->freelist;
 668
 669                 BUG_ON(magic == NODE_INFO);
 670
 671                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 672                 removing_section_nr = page_private(page);
 673
 674                 /*
 675                  * When this function is called, the removing section is
 676                  * logical offlined state. This means all pages are isolated
 677                  * from page allocator. If removing section's memmap is placed
 678                  * on the same section, it must not be freed.
 679                  * If it is freed, page allocator may allocate it which will
 680                  * be removed physically soon.
 681                  */
 682                 if (maps_section_nr != removing_section_nr)
 683                         put_page_bootmem(page);
 684         }
 685 }
 686 #endif /* CONFIG_MEMORY_HOTREMOVE */
 687 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 688
 689 /*
 690  * returns the number of sections whose mem_maps were properly
 691  * set.  If this is <=0, then that means that the passed-in
 692  * map was not consumed and must be freed.
 693  */
 694 int __meminit sparse_add_one_section(int nid, unsigned long start_pfn,
 695                                      struct vmem_altmap *altmap)
 696 {
 697         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 698         struct mem_section *ms;
 699         struct page *memmap;
 700         unsigned long *usemap;
 701         int ret;
 702
 703         /*
 704          * no locking for this, because it does its own
 705          * plus, it does a kmalloc
 706          */
 707         ret = sparse_index_init(section_nr, nid);
 708         if (ret < 0 && ret != -EEXIST)
 709                 return ret;
 710         ret = 0;
 711         memmap = kmalloc_section_memmap(section_nr, nid, altmap);
 712         if (!memmap)
 713                 return -ENOMEM;
 714         usemap = __kmalloc_section_usemap();
 715         if (!usemap) {
 716                 __kfree_section_memmap(memmap, altmap);
 717                 return -ENOMEM;
 718         }
 719
 720         ms = __pfn_to_section(start_pfn);
 721         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 722                 ret = -EEXIST;
 723                 goto out;
 724         }
 725
 726         /*
 727          * Poison uninitialized struct pages in order to catch invalid flags
 728          * combinations.
 729          */
 730         page_init_poison(memmap, sizeof(struct page) * PAGES_PER_SECTION);
 731
 732         section_mark_present(ms);
 733         sparse_init_one_section(ms, section_nr, memmap, usemap);
 734
 735 out:
 736         if (ret < 0) {
 737                 kfree(usemap);
 738                 __kfree_section_memmap(memmap, altmap);
 739         }
 740         return ret;
 741 }
 742
 743 #ifdef CONFIG_MEMORY_HOTREMOVE
 744 #ifdef CONFIG_MEMORY_FAILURE
 745 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 746 {
 747         int i;
 748
 749         if (!memmap)
 750                 return;
 751
 752         /*
 753          * A further optimization is to have per section refcounted
 754          * num_poisoned_pages.  But that would need more space per memmap, so
 755          * for now just do a quick global check to speed up this routine in the
 756          * absence of bad pages.
 757          */
 758         if (atomic_long_read(&num_poisoned_pages) == 0)
 759                 return;
 760
 761         for (i = 0; i < nr_pages; i++) {
 762                 if (PageHWPoison(&memmap[i])) {
 763                         atomic_long_sub(1, &num_poisoned_pages);
 764                         ClearPageHWPoison(&memmap[i]);
 765                 }
 766         }
 767 }
 768 #else
 769 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 770 {
 771 }
 772 #endif
 773
 774 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
 775                 struct vmem_altmap *altmap)
 776 {
 777         struct page *usemap_page;
 778
 779         if (!usemap)
 780                 return;
 781
 782         usemap_page = virt_to_page(usemap);
 783         /*
 784          * Check to see if allocation came from hot-plug-add
 785          */
 786         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 787                 kfree(usemap);
 788                 if (memmap)
 789                         __kfree_section_memmap(memmap, altmap);
 790                 return;
 791         }
 792
 793         /*
 794          * The usemap came from bootmem. This is packed with other usemaps
 795          * on the section which has pgdat at boot time. Just keep it as is now.
 796          */
 797
 798         if (memmap)
 799                 free_map_bootmem(memmap);
 800 }
 801
 802 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 803                 unsigned long map_offset, struct vmem_altmap *altmap)
 804 {
 805         struct page *memmap = NULL;
 806         unsigned long *usemap = NULL;
 807
 808         if (ms->section_mem_map) {
 809                 usemap = ms->pageblock_flags;
 810                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 811                                                 __section_nr(ms));
 812                 ms->section_mem_map = 0;
 813                 ms->pageblock_flags = NULL;
 814         }
 815
 816         clear_hwpoisoned_pages(memmap + map_offset,
 817                         PAGES_PER_SECTION - map_offset);
 818         free_section_usemap(memmap, usemap, altmap);
 819 }
 820 #endif /* CONFIG_MEMORY_HOTREMOVE */
 821 #endif /* CONFIG_MEMORY_HOTPLUG */