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/bootmem.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_virt_alloc_node(array_size, nid);
  72
  73         return section;
  74 }
  75
  76 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  77 {
  78         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  79         struct mem_section *section;
  80
  81         if (mem_section[root])
  82                 return -EEXIST;
  83
  84         section = sparse_index_alloc(nid);
  85         if (!section)
  86                 return -ENOMEM;
  87
  88         mem_section[root] = section;
  89
  90         return 0;
  91 }
  92 #else /* !SPARSEMEM_EXTREME */
  93 static inline int sparse_index_init(unsigned long section_nr, int nid)
  94 {
  95         return 0;
  96 }
  97 #endif
  98
  99 #ifdef CONFIG_SPARSEMEM_EXTREME
 100 int __section_nr(struct mem_section* ms)
 101 {
 102         unsigned long root_nr;
 103         struct mem_section *root = NULL;
 104
 105         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 106                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 107                 if (!root)
 108                         continue;
 109
 110                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 111                      break;
 112         }
 113
 114         VM_BUG_ON(!root);
 115
 116         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 117 }
 118 #else
 119 int __section_nr(struct mem_section* ms)
 120 {
 121         return (int)(ms - mem_section[0]);
 122 }
 123 #endif
 124
 125 /*
 126  * During early boot, before section_mem_map is used for an actual
 127  * mem_map, we use section_mem_map to store the section's NUMA
 128  * node.  This keeps us from having to use another data structure.  The
 129  * node information is cleared just before we store the real mem_map.
 130  */
 131 static inline unsigned long sparse_encode_early_nid(int nid)
 132 {
 133         return (nid << SECTION_NID_SHIFT);
 134 }
 135
 136 static inline int sparse_early_nid(struct mem_section *section)
 137 {
 138         return (section->section_mem_map >> SECTION_NID_SHIFT);
 139 }
 140
 141 /* Validate the physical addressing limitations of the model */
 142 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 143                                                 unsigned long *end_pfn)
 144 {
 145         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 146
 147         /*
 148          * Sanity checks - do not allow an architecture to pass
 149          * in larger pfns than the maximum scope of sparsemem:
 150          */
 151         if (*start_pfn > max_sparsemem_pfn) {
 152                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 153                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 154                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 155                 WARN_ON_ONCE(1);
 156                 *start_pfn = max_sparsemem_pfn;
 157                 *end_pfn = max_sparsemem_pfn;
 158         } else if (*end_pfn > max_sparsemem_pfn) {
 159                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 160                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 161                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 162                 WARN_ON_ONCE(1);
 163                 *end_pfn = max_sparsemem_pfn;
 164         }
 165 }
 166
 167 /*
 168  * There are a number of times that we loop over NR_MEM_SECTIONS,
 169  * looking for section_present() on each.  But, when we have very
 170  * large physical address spaces, NR_MEM_SECTIONS can also be
 171  * very large which makes the loops quite long.
 172  *
 173  * Keeping track of this gives us an easy way to break out of
 174  * those loops early.
 175  */
 176 int __highest_present_section_nr;
 177 static void section_mark_present(struct mem_section *ms)
 178 {
 179         int section_nr = __section_nr(ms);
 180
 181         if (section_nr > __highest_present_section_nr)
 182                 __highest_present_section_nr = section_nr;
 183
 184         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 185 }
 186
 187 static inline int next_present_section_nr(int section_nr)
 188 {
 189         do {
 190                 section_nr++;
 191                 if (present_section_nr(section_nr))
 192                         return section_nr;
 193         } while ((section_nr <= __highest_present_section_nr));
 194
 195         return -1;
 196 }
 197 #define for_each_present_section_nr(start, section_nr)          \
 198         for (section_nr = next_present_section_nr(start-1);     \
 199              ((section_nr >= 0) &&                              \
 200               (section_nr <= __highest_present_section_nr));    \
 201              section_nr = next_present_section_nr(section_nr))
 202
 203 /* Record a memory area against a node. */
 204 void __init memory_present(int nid, unsigned long start, unsigned long end)
 205 {
 206         unsigned long pfn;
 207
 208 #ifdef CONFIG_SPARSEMEM_EXTREME
 209         if (unlikely(!mem_section)) {
 210                 unsigned long size, align;
 211
 212                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
 213                 align = 1 << (INTERNODE_CACHE_SHIFT);
 214                 mem_section = memblock_virt_alloc(size, align);
 215         }
 216 #endif
 217
 218         start &= PAGE_SECTION_MASK;
 219         mminit_validate_memmodel_limits(&start, &end);
 220         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 221                 unsigned long section = pfn_to_section_nr(pfn);
 222                 struct mem_section *ms;
 223
 224                 sparse_index_init(section, nid);
 225                 set_section_nid(section, nid);
 226
 227                 ms = __nr_to_section(section);
 228                 if (!ms->section_mem_map) {
 229                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 230                                                         SECTION_IS_ONLINE;
 231                         section_mark_present(ms);
 232                 }
 233         }
 234 }
 235
 236 /*
 237  * Subtle, we encode the real pfn into the mem_map such that
 238  * the identity pfn - section_mem_map will return the actual
 239  * physical page frame number.
 240  */
 241 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 242 {
 243         unsigned long coded_mem_map =
 244                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 245         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 246         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 247         return coded_mem_map;
 248 }
 249
 250 /*
 251  * Decode mem_map from the coded memmap
 252  */
 253 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 254 {
 255         /* mask off the extra low bits of information */
 256         coded_mem_map &= SECTION_MAP_MASK;
 257         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 258 }
 259
 260 static void __meminit sparse_init_one_section(struct mem_section *ms,
 261                 unsigned long pnum, struct page *mem_map,
 262                 unsigned long *pageblock_bitmap)
 263 {
 264         ms->section_mem_map &= ~SECTION_MAP_MASK;
 265         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 266                                                         SECTION_HAS_MEM_MAP;
 267         ms->pageblock_flags = pageblock_bitmap;
 268 }
 269
 270 unsigned long usemap_size(void)
 271 {
 272         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 273 }
 274
 275 #ifdef CONFIG_MEMORY_HOTPLUG
 276 static unsigned long *__kmalloc_section_usemap(void)
 277 {
 278         return kmalloc(usemap_size(), GFP_KERNEL);
 279 }
 280 #endif /* CONFIG_MEMORY_HOTPLUG */
 281
 282 #ifdef CONFIG_MEMORY_HOTREMOVE
 283 static unsigned long * __init
 284 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 285                                          unsigned long size)
 286 {
 287         unsigned long goal, limit;
 288         unsigned long *p;
 289         int nid;
 290         /*
 291          * A page may contain usemaps for other sections preventing the
 292          * page being freed and making a section unremovable while
 293          * other sections referencing the usemap remain active. Similarly,
 294          * a pgdat can prevent a section being removed. If section A
 295          * contains a pgdat and section B contains the usemap, both
 296          * sections become inter-dependent. This allocates usemaps
 297          * from the same section as the pgdat where possible to avoid
 298          * this problem.
 299          */
 300         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 301         limit = goal + (1UL << PA_SECTION_SHIFT);
 302         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 303 again:
 304         p = memblock_virt_alloc_try_nid_nopanic(size,
 305                                                 SMP_CACHE_BYTES, goal, limit,
 306                                                 nid);
 307         if (!p && limit) {
 308                 limit = 0;
 309                 goto again;
 310         }
 311         return p;
 312 }
 313
 314 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 315 {
 316         unsigned long usemap_snr, pgdat_snr;
 317         static unsigned long old_usemap_snr;
 318         static unsigned long old_pgdat_snr;
 319         struct pglist_data *pgdat = NODE_DATA(nid);
 320         int usemap_nid;
 321
 322         /* First call */
 323         if (!old_usemap_snr) {
 324                 old_usemap_snr = NR_MEM_SECTIONS;
 325                 old_pgdat_snr = NR_MEM_SECTIONS;
 326         }
 327
 328         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 329         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 330         if (usemap_snr == pgdat_snr)
 331                 return;
 332
 333         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 334                 /* skip redundant message */
 335                 return;
 336
 337         old_usemap_snr = usemap_snr;
 338         old_pgdat_snr = pgdat_snr;
 339
 340         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 341         if (usemap_nid != nid) {
 342                 pr_info("node %d must be removed before remove section %ld\n",
 343                         nid, usemap_snr);
 344                 return;
 345         }
 346         /*
 347          * There is a circular dependency.
 348          * Some platforms allow un-removable section because they will just
 349          * gather other removable sections for dynamic partitioning.
 350          * Just notify un-removable section's number here.
 351          */
 352         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 353                 usemap_snr, pgdat_snr, nid);
 354 }
 355 #else
 356 static unsigned long * __init
 357 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 358                                          unsigned long size)
 359 {
 360         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 361 }
 362
 363 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 364 {
 365 }
 366 #endif /* CONFIG_MEMORY_HOTREMOVE */
 367
 368 static void __init sparse_early_usemaps_alloc_node(void *data,
 369                                  unsigned long pnum_begin,
 370                                  unsigned long pnum_end,
 371                                  unsigned long usemap_count, int nodeid)
 372 {
 373         void *usemap;
 374         unsigned long pnum;
 375         unsigned long **usemap_map = (unsigned long **)data;
 376         int size = usemap_size();
 377
 378         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 379                                                           size * usemap_count);
 380         if (!usemap) {
 381                 pr_warn("%s: allocation failed\n", __func__);
 382                 return;
 383         }
 384
 385         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 386                 if (!present_section_nr(pnum))
 387                         continue;
 388                 usemap_map[pnum] = usemap;
 389                 usemap += size;
 390                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 391         }
 392 }
 393
 394 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 395 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
 396                 struct vmem_altmap *altmap)
 397 {
 398         struct page *map;
 399         unsigned long size;
 400
 401         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 402         map = memblock_virt_alloc_try_nid(size,
 403                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 404                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 405         return map;
 406 }
 407 void __init sparse_mem_maps_populate_node(struct page **map_map,
 408                                           unsigned long pnum_begin,
 409                                           unsigned long pnum_end,
 410                                           unsigned long map_count, int nodeid)
 411 {
 412         void *map;
 413         unsigned long pnum;
 414         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 415
 416         size = PAGE_ALIGN(size);
 417         map = memblock_virt_alloc_try_nid_raw(size * map_count,
 418                                               PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 419                                               BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 420         if (map) {
 421                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 422                         if (!present_section_nr(pnum))
 423                                 continue;
 424                         map_map[pnum] = map;
 425                         map += size;
 426                 }
 427                 return;
 428         }
 429
 430         /* fallback */
 431         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 432                 struct mem_section *ms;
 433
 434                 if (!present_section_nr(pnum))
 435                         continue;
 436                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid, NULL);
 437                 if (map_map[pnum])
 438                         continue;
 439                 ms = __nr_to_section(pnum);
 440                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 441                        __func__);
 442                 ms->section_mem_map = 0;
 443         }
 444 }
 445 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 446
 447 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 448 static void __init sparse_early_mem_maps_alloc_node(void *data,
 449                                  unsigned long pnum_begin,
 450                                  unsigned long pnum_end,
 451                                  unsigned long map_count, int nodeid)
 452 {
 453         struct page **map_map = (struct page **)data;
 454         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 455                                          map_count, nodeid);
 456 }
 457 #else
 458 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 459 {
 460         struct page *map;
 461         struct mem_section *ms = __nr_to_section(pnum);
 462         int nid = sparse_early_nid(ms);
 463
 464         map = sparse_mem_map_populate(pnum, nid, NULL);
 465         if (map)
 466                 return map;
 467
 468         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 469                __func__);
 470         ms->section_mem_map = 0;
 471         return NULL;
 472 }
 473 #endif
 474
 475 void __weak __meminit vmemmap_populate_print_last(void)
 476 {
 477 }
 478
 479 /**
 480  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 481  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 482  */
 483 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 484                                         (void *, unsigned long, unsigned long,
 485                                         unsigned long, int), void *data)
 486 {
 487         unsigned long pnum;
 488         unsigned long map_count;
 489         int nodeid_begin = 0;
 490         unsigned long pnum_begin = 0;
 491
 492         for_each_present_section_nr(0, pnum) {
 493                 struct mem_section *ms;
 494
 495                 ms = __nr_to_section(pnum);
 496                 nodeid_begin = sparse_early_nid(ms);
 497                 pnum_begin = pnum;
 498                 break;
 499         }
 500         map_count = 1;
 501         for_each_present_section_nr(pnum_begin + 1, pnum) {
 502                 struct mem_section *ms;
 503                 int nodeid;
 504
 505                 ms = __nr_to_section(pnum);
 506                 nodeid = sparse_early_nid(ms);
 507                 if (nodeid == nodeid_begin) {
 508                         map_count++;
 509                         continue;
 510                 }
 511                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 512                 alloc_func(data, pnum_begin, pnum,
 513                                                 map_count, nodeid_begin);
 514                 /* new start, update count etc*/
 515                 nodeid_begin = nodeid;
 516                 pnum_begin = pnum;
 517                 map_count = 1;
 518         }
 519         /* ok, last chunk */
 520         alloc_func(data, pnum_begin, __highest_present_section_nr+1,
 521                                                 map_count, nodeid_begin);
 522 }
 523
 524 /*
 525  * Allocate the accumulated non-linear sections, allocate a mem_map
 526  * for each and record the physical to section mapping.
 527  */
 528 void __init sparse_init(void)
 529 {
 530         unsigned long pnum;
 531         struct page *map;
 532         unsigned long *usemap;
 533         unsigned long **usemap_map;
 534         int size;
 535 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 536         int size2;
 537         struct page **map_map;
 538 #endif
 539
 540         /* see include/linux/mmzone.h 'struct mem_section' definition */
 541         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 542
 543         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 544         set_pageblock_order();
 545
 546         /*
 547          * map is using big page (aka 2M in x86 64 bit)
 548          * usemap is less one page (aka 24 bytes)
 549          * so alloc 2M (with 2M align) and 24 bytes in turn will
 550          * make next 2M slip to one more 2M later.
 551          * then in big system, the memory will have a lot of holes...
 552          * here try to allocate 2M pages continuously.
 553          *
 554          * powerpc need to call sparse_init_one_section right after each
 555          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 556          */
 557         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 558         usemap_map = memblock_virt_alloc(size, 0);
 559         if (!usemap_map)
 560                 panic("can not allocate usemap_map\n");
 561         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 562                                                         (void *)usemap_map);
 563
 564 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 565         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 566         map_map = memblock_virt_alloc(size2, 0);
 567         if (!map_map)
 568                 panic("can not allocate map_map\n");
 569         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 570                                                         (void *)map_map);
 571 #endif
 572
 573         for_each_present_section_nr(0, pnum) {
 574                 usemap = usemap_map[pnum];
 575                 if (!usemap)
 576                         continue;
 577
 578 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 579                 map = map_map[pnum];
 580 #else
 581                 map = sparse_early_mem_map_alloc(pnum);
 582 #endif
 583                 if (!map)
 584                         continue;
 585
 586                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 587                                                                 usemap);
 588         }
 589
 590         vmemmap_populate_print_last();
 591
 592 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 593         memblock_free_early(__pa(map_map), size2);
 594 #endif
 595         memblock_free_early(__pa(usemap_map), size);
 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 online */
 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 inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 643                 struct vmem_altmap *altmap)
 644 {
 645         /* This will make the necessary allocations eventually. */
 646         return sparse_mem_map_populate(pnum, nid, altmap);
 647 }
 648 static void __kfree_section_memmap(struct page *memmap,
 649                 struct vmem_altmap *altmap)
 650 {
 651         unsigned long start = (unsigned long)memmap;
 652         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 653
 654         vmemmap_free(start, end, altmap);
 655 }
 656 #ifdef CONFIG_MEMORY_HOTREMOVE
 657 static void free_map_bootmem(struct page *memmap)
 658 {
 659         unsigned long start = (unsigned long)memmap;
 660         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 661
 662         vmemmap_free(start, end, NULL);
 663 }
 664 #endif /* CONFIG_MEMORY_HOTREMOVE */
 665 #else
 666 static struct page *__kmalloc_section_memmap(void)
 667 {
 668         struct page *page, *ret;
 669         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 670
 671         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 672         if (page)
 673                 goto got_map_page;
 674
 675         ret = vmalloc(memmap_size);
 676         if (ret)
 677                 goto got_map_ptr;
 678
 679         return NULL;
 680 got_map_page:
 681         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 682 got_map_ptr:
 683
 684         return ret;
 685 }
 686
 687 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 688                 struct vmem_altmap *altmap)
 689 {
 690         return __kmalloc_section_memmap();
 691 }
 692
 693 static void __kfree_section_memmap(struct page *memmap,
 694                 struct vmem_altmap *altmap)
 695 {
 696         if (is_vmalloc_addr(memmap))
 697                 vfree(memmap);
 698         else
 699                 free_pages((unsigned long)memmap,
 700                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 701 }
 702
 703 #ifdef CONFIG_MEMORY_HOTREMOVE
 704 static void free_map_bootmem(struct page *memmap)
 705 {
 706         unsigned long maps_section_nr, removing_section_nr, i;
 707         unsigned long magic, nr_pages;
 708         struct page *page = virt_to_page(memmap);
 709
 710         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 711                 >> PAGE_SHIFT;
 712
 713         for (i = 0; i < nr_pages; i++, page++) {
 714                 magic = (unsigned long) page->freelist;
 715
 716                 BUG_ON(magic == NODE_INFO);
 717
 718                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 719                 removing_section_nr = page_private(page);
 720
 721                 /*
 722                  * When this function is called, the removing section is
 723                  * logical offlined state. This means all pages are isolated
 724                  * from page allocator. If removing section's memmap is placed
 725                  * on the same section, it must not be freed.
 726                  * If it is freed, page allocator may allocate it which will
 727                  * be removed physically soon.
 728                  */
 729                 if (maps_section_nr != removing_section_nr)
 730                         put_page_bootmem(page);
 731         }
 732 }
 733 #endif /* CONFIG_MEMORY_HOTREMOVE */
 734 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 735
 736 /*
 737  * returns the number of sections whose mem_maps were properly
 738  * set.  If this is <=0, then that means that the passed-in
 739  * map was not consumed and must be freed.
 740  */
 741 int __meminit sparse_add_one_section(struct pglist_data *pgdat,
 742                 unsigned long start_pfn, struct vmem_altmap *altmap)
 743 {
 744         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 745         struct mem_section *ms;
 746         struct page *memmap;
 747         unsigned long *usemap;
 748         unsigned long flags;
 749         int ret;
 750
 751         /*
 752          * no locking for this, because it does its own
 753          * plus, it does a kmalloc
 754          */
 755         ret = sparse_index_init(section_nr, pgdat->node_id);
 756         if (ret < 0 && ret != -EEXIST)
 757                 return ret;
 758         ret = 0;
 759         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, altmap);
 760         if (!memmap)
 761                 return -ENOMEM;
 762         usemap = __kmalloc_section_usemap();
 763         if (!usemap) {
 764                 __kfree_section_memmap(memmap, altmap);
 765                 return -ENOMEM;
 766         }
 767
 768         pgdat_resize_lock(pgdat, &flags);
 769
 770         ms = __pfn_to_section(start_pfn);
 771         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 772                 ret = -EEXIST;
 773                 goto out;
 774         }
 775
 776 #ifdef CONFIG_DEBUG_VM
 777         /*
 778          * Poison uninitialized struct pages in order to catch invalid flags
 779          * combinations.
 780          */
 781         memset(memmap, PAGE_POISON_PATTERN, sizeof(struct page) * PAGES_PER_SECTION);
 782 #endif
 783
 784         section_mark_present(ms);
 785         sparse_init_one_section(ms, section_nr, memmap, usemap);
 786
 787 out:
 788         pgdat_resize_unlock(pgdat, &flags);
 789         if (ret < 0) {
 790                 kfree(usemap);
 791                 __kfree_section_memmap(memmap, altmap);
 792         }
 793         return ret;
 794 }
 795
 796 #ifdef CONFIG_MEMORY_HOTREMOVE
 797 #ifdef CONFIG_MEMORY_FAILURE
 798 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 799 {
 800         int i;
 801
 802         if (!memmap)
 803                 return;
 804
 805         for (i = 0; i < nr_pages; i++) {
 806                 if (PageHWPoison(&memmap[i])) {
 807                         atomic_long_sub(1, &num_poisoned_pages);
 808                         ClearPageHWPoison(&memmap[i]);
 809                 }
 810         }
 811 }
 812 #else
 813 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 814 {
 815 }
 816 #endif
 817
 818 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
 819                 struct vmem_altmap *altmap)
 820 {
 821         struct page *usemap_page;
 822
 823         if (!usemap)
 824                 return;
 825
 826         usemap_page = virt_to_page(usemap);
 827         /*
 828          * Check to see if allocation came from hot-plug-add
 829          */
 830         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 831                 kfree(usemap);
 832                 if (memmap)
 833                         __kfree_section_memmap(memmap, altmap);
 834                 return;
 835         }
 836
 837         /*
 838          * The usemap came from bootmem. This is packed with other usemaps
 839          * on the section which has pgdat at boot time. Just keep it as is now.
 840          */
 841
 842         if (memmap)
 843                 free_map_bootmem(memmap);
 844 }
 845
 846 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 847                 unsigned long map_offset, struct vmem_altmap *altmap)
 848 {
 849         struct page *memmap = NULL;
 850         unsigned long *usemap = NULL, flags;
 851         struct pglist_data *pgdat = zone->zone_pgdat;
 852
 853         pgdat_resize_lock(pgdat, &flags);
 854         if (ms->section_mem_map) {
 855                 usemap = ms->pageblock_flags;
 856                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 857                                                 __section_nr(ms));
 858                 ms->section_mem_map = 0;
 859                 ms->pageblock_flags = NULL;
 860         }
 861         pgdat_resize_unlock(pgdat, &flags);
 862
 863         clear_hwpoisoned_pages(memmap + map_offset,
 864                         PAGES_PER_SECTION - map_offset);
 865         free_section_usemap(memmap, usemap, altmap);
 866 }
 867 #endif /* CONFIG_MEMORY_HOTREMOVE */
 868 #endif /* CONFIG_MEMORY_HOTPLUG */