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[NR_SECTION_ROOTS]
  27         ____cacheline_internodealigned_in_smp;
  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_virt_alloc_node(array_size, nid);
  73
  74         return section;
  75 }
  76
  77 static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  78 {
  79         unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  80         struct mem_section *section;
  81
  82         if (mem_section[root])
  83                 return -EEXIST;
  84
  85         section = sparse_index_alloc(nid);
  86         if (!section)
  87                 return -ENOMEM;
  88
  89         mem_section[root] = section;
  90
  91         return 0;
  92 }
  93 #else /* !SPARSEMEM_EXTREME */
  94 static inline int sparse_index_init(unsigned long section_nr, int nid)
  95 {
  96         return 0;
  97 }
  98 #endif
  99
 100 #ifdef CONFIG_SPARSEMEM_EXTREME
 101 int __section_nr(struct mem_section* ms)
 102 {
 103         unsigned long root_nr;
 104         struct mem_section* root;
 105
 106         for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 107                 root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 108                 if (!root)
 109                         continue;
 110
 111                 if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 112                      break;
 113         }
 114
 115         VM_BUG_ON(root_nr == NR_SECTION_ROOTS);
 116
 117         return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 118 }
 119 #else
 120 int __section_nr(struct mem_section* ms)
 121 {
 122         return (int)(ms - mem_section[0]);
 123 }
 124 #endif
 125
 126 /*
 127  * During early boot, before section_mem_map is used for an actual
 128  * mem_map, we use section_mem_map to store the section's NUMA
 129  * node.  This keeps us from having to use another data structure.  The
 130  * node information is cleared just before we store the real mem_map.
 131  */
 132 static inline unsigned long sparse_encode_early_nid(int nid)
 133 {
 134         return (nid << SECTION_NID_SHIFT);
 135 }
 136
 137 static inline int sparse_early_nid(struct mem_section *section)
 138 {
 139         return (section->section_mem_map >> SECTION_NID_SHIFT);
 140 }
 141
 142 /* Validate the physical addressing limitations of the model */
 143 void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 144                                                 unsigned long *end_pfn)
 145 {
 146         unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 147
 148         /*
 149          * Sanity checks - do not allow an architecture to pass
 150          * in larger pfns than the maximum scope of sparsemem:
 151          */
 152         if (*start_pfn > max_sparsemem_pfn) {
 153                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 154                         "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 155                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 156                 WARN_ON_ONCE(1);
 157                 *start_pfn = max_sparsemem_pfn;
 158                 *end_pfn = max_sparsemem_pfn;
 159         } else if (*end_pfn > max_sparsemem_pfn) {
 160                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 161                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 162                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 163                 WARN_ON_ONCE(1);
 164                 *end_pfn = max_sparsemem_pfn;
 165         }
 166 }
 167
 168 /*
 169  * There are a number of times that we loop over NR_MEM_SECTIONS,
 170  * looking for section_present() on each.  But, when we have very
 171  * large physical address spaces, NR_MEM_SECTIONS can also be
 172  * very large which makes the loops quite long.
 173  *
 174  * Keeping track of this gives us an easy way to break out of
 175  * those loops early.
 176  */
 177 int __highest_present_section_nr;
 178 static void section_mark_present(struct mem_section *ms)
 179 {
 180         int section_nr = __section_nr(ms);
 181
 182         if (section_nr > __highest_present_section_nr)
 183                 __highest_present_section_nr = section_nr;
 184
 185         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 186 }
 187
 188 static inline int next_present_section_nr(int section_nr)
 189 {
 190         do {
 191                 section_nr++;
 192                 if (present_section_nr(section_nr))
 193                         return section_nr;
 194         } while ((section_nr < NR_MEM_SECTIONS) &&
 195                  (section_nr <= __highest_present_section_nr));
 196
 197         return -1;
 198 }
 199 #define for_each_present_section_nr(start, section_nr)          \
 200         for (section_nr = next_present_section_nr(start-1);     \
 201              ((section_nr >= 0) &&                              \
 202               (section_nr < NR_MEM_SECTIONS) &&                 \
 203               (section_nr <= __highest_present_section_nr));    \
 204              section_nr = next_present_section_nr(section_nr))
 205
 206 /* Record a memory area against a node. */
 207 void __init memory_present(int nid, unsigned long start, unsigned long end)
 208 {
 209         unsigned long pfn;
 210
 211         start &= PAGE_SECTION_MASK;
 212         mminit_validate_memmodel_limits(&start, &end);
 213         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 214                 unsigned long section = pfn_to_section_nr(pfn);
 215                 struct mem_section *ms;
 216
 217                 sparse_index_init(section, nid);
 218                 set_section_nid(section, nid);
 219
 220                 ms = __nr_to_section(section);
 221                 if (!ms->section_mem_map) {
 222                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 223                                                         SECTION_IS_ONLINE;
 224                         section_mark_present(ms);
 225                 }
 226         }
 227 }
 228
 229 /*
 230  * Only used by the i386 NUMA architecures, but relatively
 231  * generic code.
 232  */
 233 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 234                                                      unsigned long end_pfn)
 235 {
 236         unsigned long pfn;
 237         unsigned long nr_pages = 0;
 238
 239         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 240         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 241                 if (nid != early_pfn_to_nid(pfn))
 242                         continue;
 243
 244                 if (pfn_present(pfn))
 245                         nr_pages += PAGES_PER_SECTION;
 246         }
 247
 248         return nr_pages * sizeof(struct page);
 249 }
 250
 251 /*
 252  * Subtle, we encode the real pfn into the mem_map such that
 253  * the identity pfn - section_mem_map will return the actual
 254  * physical page frame number.
 255  */
 256 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 257 {
 258         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 259 }
 260
 261 /*
 262  * Decode mem_map from the coded memmap
 263  */
 264 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 265 {
 266         /* mask off the extra low bits of information */
 267         coded_mem_map &= SECTION_MAP_MASK;
 268         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 269 }
 270
 271 static int __meminit sparse_init_one_section(struct mem_section *ms,
 272                 unsigned long pnum, struct page *mem_map,
 273                 unsigned long *pageblock_bitmap)
 274 {
 275         if (!present_section(ms))
 276                 return -EINVAL;
 277
 278         ms->section_mem_map &= ~SECTION_MAP_MASK;
 279         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 280                                                         SECTION_HAS_MEM_MAP;
 281         ms->pageblock_flags = pageblock_bitmap;
 282
 283         return 1;
 284 }
 285
 286 unsigned long usemap_size(void)
 287 {
 288         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 289 }
 290
 291 #ifdef CONFIG_MEMORY_HOTPLUG
 292 static unsigned long *__kmalloc_section_usemap(void)
 293 {
 294         return kmalloc(usemap_size(), GFP_KERNEL);
 295 }
 296 #endif /* CONFIG_MEMORY_HOTPLUG */
 297
 298 #ifdef CONFIG_MEMORY_HOTREMOVE
 299 static unsigned long * __init
 300 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 301                                          unsigned long size)
 302 {
 303         unsigned long goal, limit;
 304         unsigned long *p;
 305         int nid;
 306         /*
 307          * A page may contain usemaps for other sections preventing the
 308          * page being freed and making a section unremovable while
 309          * other sections referencing the usemap remain active. Similarly,
 310          * a pgdat can prevent a section being removed. If section A
 311          * contains a pgdat and section B contains the usemap, both
 312          * sections become inter-dependent. This allocates usemaps
 313          * from the same section as the pgdat where possible to avoid
 314          * this problem.
 315          */
 316         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 317         limit = goal + (1UL << PA_SECTION_SHIFT);
 318         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 319 again:
 320         p = memblock_virt_alloc_try_nid_nopanic(size,
 321                                                 SMP_CACHE_BYTES, goal, limit,
 322                                                 nid);
 323         if (!p && limit) {
 324                 limit = 0;
 325                 goto again;
 326         }
 327         return p;
 328 }
 329
 330 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 331 {
 332         unsigned long usemap_snr, pgdat_snr;
 333         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 334         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 335         struct pglist_data *pgdat = NODE_DATA(nid);
 336         int usemap_nid;
 337
 338         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 339         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 340         if (usemap_snr == pgdat_snr)
 341                 return;
 342
 343         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 344                 /* skip redundant message */
 345                 return;
 346
 347         old_usemap_snr = usemap_snr;
 348         old_pgdat_snr = pgdat_snr;
 349
 350         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 351         if (usemap_nid != nid) {
 352                 pr_info("node %d must be removed before remove section %ld\n",
 353                         nid, usemap_snr);
 354                 return;
 355         }
 356         /*
 357          * There is a circular dependency.
 358          * Some platforms allow un-removable section because they will just
 359          * gather other removable sections for dynamic partitioning.
 360          * Just notify un-removable section's number here.
 361          */
 362         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 363                 usemap_snr, pgdat_snr, nid);
 364 }
 365 #else
 366 static unsigned long * __init
 367 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 368                                          unsigned long size)
 369 {
 370         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 371 }
 372
 373 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 374 {
 375 }
 376 #endif /* CONFIG_MEMORY_HOTREMOVE */
 377
 378 static void __init sparse_early_usemaps_alloc_node(void *data,
 379                                  unsigned long pnum_begin,
 380                                  unsigned long pnum_end,
 381                                  unsigned long usemap_count, int nodeid)
 382 {
 383         void *usemap;
 384         unsigned long pnum;
 385         unsigned long **usemap_map = (unsigned long **)data;
 386         int size = usemap_size();
 387
 388         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 389                                                           size * usemap_count);
 390         if (!usemap) {
 391                 pr_warn("%s: allocation failed\n", __func__);
 392                 return;
 393         }
 394
 395         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 396                 if (!present_section_nr(pnum))
 397                         continue;
 398                 usemap_map[pnum] = usemap;
 399                 usemap += size;
 400                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 401         }
 402 }
 403
 404 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 405 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 406 {
 407         struct page *map;
 408         unsigned long size;
 409
 410         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 411         if (map)
 412                 return map;
 413
 414         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 415         map = memblock_virt_alloc_try_nid(size,
 416                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 417                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 418         return map;
 419 }
 420 void __init sparse_mem_maps_populate_node(struct page **map_map,
 421                                           unsigned long pnum_begin,
 422                                           unsigned long pnum_end,
 423                                           unsigned long map_count, int nodeid)
 424 {
 425         void *map;
 426         unsigned long pnum;
 427         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 428
 429         map = alloc_remap(nodeid, size * map_count);
 430         if (map) {
 431                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 432                         if (!present_section_nr(pnum))
 433                                 continue;
 434                         map_map[pnum] = map;
 435                         map += size;
 436                 }
 437                 return;
 438         }
 439
 440         size = PAGE_ALIGN(size);
 441         map = memblock_virt_alloc_try_nid(size * map_count,
 442                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 443                                           BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 444         if (map) {
 445                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 446                         if (!present_section_nr(pnum))
 447                                 continue;
 448                         map_map[pnum] = map;
 449                         map += size;
 450                 }
 451                 return;
 452         }
 453
 454         /* fallback */
 455         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 456                 struct mem_section *ms;
 457
 458                 if (!present_section_nr(pnum))
 459                         continue;
 460                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 461                 if (map_map[pnum])
 462                         continue;
 463                 ms = __nr_to_section(pnum);
 464                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 465                        __func__);
 466                 ms->section_mem_map = 0;
 467         }
 468 }
 469 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 470
 471 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 472 static void __init sparse_early_mem_maps_alloc_node(void *data,
 473                                  unsigned long pnum_begin,
 474                                  unsigned long pnum_end,
 475                                  unsigned long map_count, int nodeid)
 476 {
 477         struct page **map_map = (struct page **)data;
 478         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 479                                          map_count, nodeid);
 480 }
 481 #else
 482 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 483 {
 484         struct page *map;
 485         struct mem_section *ms = __nr_to_section(pnum);
 486         int nid = sparse_early_nid(ms);
 487
 488         map = sparse_mem_map_populate(pnum, nid);
 489         if (map)
 490                 return map;
 491
 492         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 493                __func__);
 494         ms->section_mem_map = 0;
 495         return NULL;
 496 }
 497 #endif
 498
 499 void __weak __meminit vmemmap_populate_print_last(void)
 500 {
 501 }
 502
 503 /**
 504  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 505  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 506  */
 507 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 508                                         (void *, unsigned long, unsigned long,
 509                                         unsigned long, int), void *data)
 510 {
 511         unsigned long pnum;
 512         unsigned long map_count;
 513         int nodeid_begin = 0;
 514         unsigned long pnum_begin = 0;
 515
 516         for_each_present_section_nr(0, pnum) {
 517                 struct mem_section *ms;
 518
 519                 ms = __nr_to_section(pnum);
 520                 nodeid_begin = sparse_early_nid(ms);
 521                 pnum_begin = pnum;
 522                 break;
 523         }
 524         map_count = 1;
 525         for_each_present_section_nr(pnum_begin + 1, pnum) {
 526                 struct mem_section *ms;
 527                 int nodeid;
 528
 529                 ms = __nr_to_section(pnum);
 530                 nodeid = sparse_early_nid(ms);
 531                 if (nodeid == nodeid_begin) {
 532                         map_count++;
 533                         continue;
 534                 }
 535                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 536                 alloc_func(data, pnum_begin, pnum,
 537                                                 map_count, nodeid_begin);
 538                 /* new start, update count etc*/
 539                 nodeid_begin = nodeid;
 540                 pnum_begin = pnum;
 541                 map_count = 1;
 542         }
 543         /* ok, last chunk */
 544         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
 545                                                 map_count, nodeid_begin);
 546 }
 547
 548 /*
 549  * Allocate the accumulated non-linear sections, allocate a mem_map
 550  * for each and record the physical to section mapping.
 551  */
 552 void __init sparse_init(void)
 553 {
 554         unsigned long pnum;
 555         struct page *map;
 556         unsigned long *usemap;
 557         unsigned long **usemap_map;
 558         int size;
 559 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 560         int size2;
 561         struct page **map_map;
 562 #endif
 563
 564         /* see include/linux/mmzone.h 'struct mem_section' definition */
 565         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 566
 567         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 568         set_pageblock_order();
 569
 570         /*
 571          * map is using big page (aka 2M in x86 64 bit)
 572          * usemap is less one page (aka 24 bytes)
 573          * so alloc 2M (with 2M align) and 24 bytes in turn will
 574          * make next 2M slip to one more 2M later.
 575          * then in big system, the memory will have a lot of holes...
 576          * here try to allocate 2M pages continuously.
 577          *
 578          * powerpc need to call sparse_init_one_section right after each
 579          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 580          */
 581         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 582         usemap_map = memblock_virt_alloc(size, 0);
 583         if (!usemap_map)
 584                 panic("can not allocate usemap_map\n");
 585         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 586                                                         (void *)usemap_map);
 587
 588 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 589         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 590         map_map = memblock_virt_alloc(size2, 0);
 591         if (!map_map)
 592                 panic("can not allocate map_map\n");
 593         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 594                                                         (void *)map_map);
 595 #endif
 596
 597         for_each_present_section_nr(0, pnum) {
 598                 usemap = usemap_map[pnum];
 599                 if (!usemap)
 600                         continue;
 601
 602 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 603                 map = map_map[pnum];
 604 #else
 605                 map = sparse_early_mem_map_alloc(pnum);
 606 #endif
 607                 if (!map)
 608                         continue;
 609
 610                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 611                                                                 usemap);
 612         }
 613
 614         vmemmap_populate_print_last();
 615
 616 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 617         memblock_free_early(__pa(map_map), size2);
 618 #endif
 619         memblock_free_early(__pa(usemap_map), size);
 620 }
 621
 622 #ifdef CONFIG_MEMORY_HOTPLUG
 623
 624 /* Mark all memory sections within the pfn range as online */
 625 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 626 {
 627         unsigned long pfn;
 628
 629         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 630                 unsigned long section_nr = pfn_to_section_nr(pfn);
 631                 struct mem_section *ms;
 632
 633                 /* onlining code should never touch invalid ranges */
 634                 if (WARN_ON(!valid_section_nr(section_nr)))
 635                         continue;
 636
 637                 ms = __nr_to_section(section_nr);
 638                 ms->section_mem_map |= SECTION_IS_ONLINE;
 639         }
 640 }
 641
 642 #ifdef CONFIG_MEMORY_HOTREMOVE
 643 /* Mark all memory sections within the pfn range as online */
 644 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 645 {
 646         unsigned long pfn;
 647
 648         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 649                 unsigned long section_nr = pfn_to_section_nr(start_pfn);
 650                 struct mem_section *ms;
 651
 652                 /*
 653                  * TODO this needs some double checking. Offlining code makes
 654                  * sure to check pfn_valid but those checks might be just bogus
 655                  */
 656                 if (WARN_ON(!valid_section_nr(section_nr)))
 657                         continue;
 658
 659                 ms = __nr_to_section(section_nr);
 660                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 661         }
 662 }
 663 #endif
 664
 665 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 666 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 667 {
 668         /* This will make the necessary allocations eventually. */
 669         return sparse_mem_map_populate(pnum, nid);
 670 }
 671 static void __kfree_section_memmap(struct page *memmap)
 672 {
 673         unsigned long start = (unsigned long)memmap;
 674         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 675
 676         vmemmap_free(start, end);
 677 }
 678 #ifdef CONFIG_MEMORY_HOTREMOVE
 679 static void free_map_bootmem(struct page *memmap)
 680 {
 681         unsigned long start = (unsigned long)memmap;
 682         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 683
 684         vmemmap_free(start, end);
 685 }
 686 #endif /* CONFIG_MEMORY_HOTREMOVE */
 687 #else
 688 static struct page *__kmalloc_section_memmap(void)
 689 {
 690         struct page *page, *ret;
 691         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 692
 693         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 694         if (page)
 695                 goto got_map_page;
 696
 697         ret = vmalloc(memmap_size);
 698         if (ret)
 699                 goto got_map_ptr;
 700
 701         return NULL;
 702 got_map_page:
 703         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 704 got_map_ptr:
 705
 706         return ret;
 707 }
 708
 709 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 710 {
 711         return __kmalloc_section_memmap();
 712 }
 713
 714 static void __kfree_section_memmap(struct page *memmap)
 715 {
 716         if (is_vmalloc_addr(memmap))
 717                 vfree(memmap);
 718         else
 719                 free_pages((unsigned long)memmap,
 720                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 721 }
 722
 723 #ifdef CONFIG_MEMORY_HOTREMOVE
 724 static void free_map_bootmem(struct page *memmap)
 725 {
 726         unsigned long maps_section_nr, removing_section_nr, i;
 727         unsigned long magic, nr_pages;
 728         struct page *page = virt_to_page(memmap);
 729
 730         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 731                 >> PAGE_SHIFT;
 732
 733         for (i = 0; i < nr_pages; i++, page++) {
 734                 magic = (unsigned long) page->freelist;
 735
 736                 BUG_ON(magic == NODE_INFO);
 737
 738                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 739                 removing_section_nr = page_private(page);
 740
 741                 /*
 742                  * When this function is called, the removing section is
 743                  * logical offlined state. This means all pages are isolated
 744                  * from page allocator. If removing section's memmap is placed
 745                  * on the same section, it must not be freed.
 746                  * If it is freed, page allocator may allocate it which will
 747                  * be removed physically soon.
 748                  */
 749                 if (maps_section_nr != removing_section_nr)
 750                         put_page_bootmem(page);
 751         }
 752 }
 753 #endif /* CONFIG_MEMORY_HOTREMOVE */
 754 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 755
 756 /*
 757  * returns the number of sections whose mem_maps were properly
 758  * set.  If this is <=0, then that means that the passed-in
 759  * map was not consumed and must be freed.
 760  */
 761 int __meminit sparse_add_one_section(struct pglist_data *pgdat, unsigned long start_pfn)
 762 {
 763         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 764         struct mem_section *ms;
 765         struct page *memmap;
 766         unsigned long *usemap;
 767         unsigned long flags;
 768         int ret;
 769
 770         /*
 771          * no locking for this, because it does its own
 772          * plus, it does a kmalloc
 773          */
 774         ret = sparse_index_init(section_nr, pgdat->node_id);
 775         if (ret < 0 && ret != -EEXIST)
 776                 return ret;
 777         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
 778         if (!memmap)
 779                 return -ENOMEM;
 780         usemap = __kmalloc_section_usemap();
 781         if (!usemap) {
 782                 __kfree_section_memmap(memmap);
 783                 return -ENOMEM;
 784         }
 785
 786         pgdat_resize_lock(pgdat, &flags);
 787
 788         ms = __pfn_to_section(start_pfn);
 789         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 790                 ret = -EEXIST;
 791                 goto out;
 792         }
 793
 794         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 795
 796         section_mark_present(ms);
 797
 798         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 799
 800 out:
 801         pgdat_resize_unlock(pgdat, &flags);
 802         if (ret <= 0) {
 803                 kfree(usemap);
 804                 __kfree_section_memmap(memmap);
 805         }
 806         return ret;
 807 }
 808
 809 #ifdef CONFIG_MEMORY_HOTREMOVE
 810 #ifdef CONFIG_MEMORY_FAILURE
 811 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 812 {
 813         int i;
 814
 815         if (!memmap)
 816                 return;
 817
 818         for (i = 0; i < nr_pages; i++) {
 819                 if (PageHWPoison(&memmap[i])) {
 820                         atomic_long_sub(1, &num_poisoned_pages);
 821                         ClearPageHWPoison(&memmap[i]);
 822                 }
 823         }
 824 }
 825 #else
 826 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 827 {
 828 }
 829 #endif
 830
 831 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 832 {
 833         struct page *usemap_page;
 834
 835         if (!usemap)
 836                 return;
 837
 838         usemap_page = virt_to_page(usemap);
 839         /*
 840          * Check to see if allocation came from hot-plug-add
 841          */
 842         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 843                 kfree(usemap);
 844                 if (memmap)
 845                         __kfree_section_memmap(memmap);
 846                 return;
 847         }
 848
 849         /*
 850          * The usemap came from bootmem. This is packed with other usemaps
 851          * on the section which has pgdat at boot time. Just keep it as is now.
 852          */
 853
 854         if (memmap)
 855                 free_map_bootmem(memmap);
 856 }
 857
 858 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 859                 unsigned long map_offset)
 860 {
 861         struct page *memmap = NULL;
 862         unsigned long *usemap = NULL, flags;
 863         struct pglist_data *pgdat = zone->zone_pgdat;
 864
 865         pgdat_resize_lock(pgdat, &flags);
 866         if (ms->section_mem_map) {
 867                 usemap = ms->pageblock_flags;
 868                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 869                                                 __section_nr(ms));
 870                 ms->section_mem_map = 0;
 871                 ms->pageblock_flags = NULL;
 872         }
 873         pgdat_resize_unlock(pgdat, &flags);
 874
 875         clear_hwpoisoned_pages(memmap + map_offset,
 876                         PAGES_PER_SECTION - map_offset);
 877         free_section_usemap(memmap, usemap);
 878 }
 879 #endif /* CONFIG_MEMORY_HOTREMOVE */
 880 #endif /* CONFIG_MEMORY_HOTPLUG */