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 < NR_MEM_SECTIONS) &&
 194                  (section_nr <= __highest_present_section_nr));
 195
 196         return -1;
 197 }
 198 #define for_each_present_section_nr(start, section_nr)          \
 199         for (section_nr = next_present_section_nr(start-1);     \
 200              ((section_nr >= 0) &&                              \
 201               (section_nr < NR_MEM_SECTIONS) &&                 \
 202               (section_nr <= __highest_present_section_nr));    \
 203              section_nr = next_present_section_nr(section_nr))
 204
 205 /* Record a memory area against a node. */
 206 void __init memory_present(int nid, unsigned long start, unsigned long end)
 207 {
 208         unsigned long pfn;
 209
 210 #ifdef CONFIG_SPARSEMEM_EXTREME
 211         if (unlikely(!mem_section)) {
 212                 unsigned long size, align;
 213
 214                 size = sizeof(struct mem_section) * NR_SECTION_ROOTS;
 215                 align = 1 << (INTERNODE_CACHE_SHIFT);
 216                 mem_section = memblock_virt_alloc(size, align);
 217         }
 218 #endif
 219
 220         start &= PAGE_SECTION_MASK;
 221         mminit_validate_memmodel_limits(&start, &end);
 222         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 223                 unsigned long section = pfn_to_section_nr(pfn);
 224                 struct mem_section *ms;
 225
 226                 sparse_index_init(section, nid);
 227                 set_section_nid(section, nid);
 228
 229                 ms = __nr_to_section(section);
 230                 if (!ms->section_mem_map) {
 231                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 232                                                         SECTION_IS_ONLINE;
 233                         section_mark_present(ms);
 234                 }
 235         }
 236 }
 237
 238 /*
 239  * Only used by the i386 NUMA architecures, but relatively
 240  * generic code.
 241  */
 242 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 243                                                      unsigned long end_pfn)
 244 {
 245         unsigned long pfn;
 246         unsigned long nr_pages = 0;
 247
 248         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 249         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 250                 if (nid != early_pfn_to_nid(pfn))
 251                         continue;
 252
 253                 if (pfn_present(pfn))
 254                         nr_pages += PAGES_PER_SECTION;
 255         }
 256
 257         return nr_pages * sizeof(struct page);
 258 }
 259
 260 /*
 261  * Subtle, we encode the real pfn into the mem_map such that
 262  * the identity pfn - section_mem_map will return the actual
 263  * physical page frame number.
 264  */
 265 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 266 {
 267         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 268 }
 269
 270 /*
 271  * Decode mem_map from the coded memmap
 272  */
 273 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 274 {
 275         /* mask off the extra low bits of information */
 276         coded_mem_map &= SECTION_MAP_MASK;
 277         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 278 }
 279
 280 static int __meminit sparse_init_one_section(struct mem_section *ms,
 281                 unsigned long pnum, struct page *mem_map,
 282                 unsigned long *pageblock_bitmap)
 283 {
 284         if (!present_section(ms))
 285                 return -EINVAL;
 286
 287         ms->section_mem_map &= ~SECTION_MAP_MASK;
 288         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 289                                                         SECTION_HAS_MEM_MAP;
 290         ms->pageblock_flags = pageblock_bitmap;
 291
 292         return 1;
 293 }
 294
 295 unsigned long usemap_size(void)
 296 {
 297         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 298 }
 299
 300 #ifdef CONFIG_MEMORY_HOTPLUG
 301 static unsigned long *__kmalloc_section_usemap(void)
 302 {
 303         return kmalloc(usemap_size(), GFP_KERNEL);
 304 }
 305 #endif /* CONFIG_MEMORY_HOTPLUG */
 306
 307 #ifdef CONFIG_MEMORY_HOTREMOVE
 308 static unsigned long * __init
 309 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 310                                          unsigned long size)
 311 {
 312         unsigned long goal, limit;
 313         unsigned long *p;
 314         int nid;
 315         /*
 316          * A page may contain usemaps for other sections preventing the
 317          * page being freed and making a section unremovable while
 318          * other sections referencing the usemap remain active. Similarly,
 319          * a pgdat can prevent a section being removed. If section A
 320          * contains a pgdat and section B contains the usemap, both
 321          * sections become inter-dependent. This allocates usemaps
 322          * from the same section as the pgdat where possible to avoid
 323          * this problem.
 324          */
 325         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 326         limit = goal + (1UL << PA_SECTION_SHIFT);
 327         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 328 again:
 329         p = memblock_virt_alloc_try_nid_nopanic(size,
 330                                                 SMP_CACHE_BYTES, goal, limit,
 331                                                 nid);
 332         if (!p && limit) {
 333                 limit = 0;
 334                 goto again;
 335         }
 336         return p;
 337 }
 338
 339 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 340 {
 341         unsigned long usemap_snr, pgdat_snr;
 342         static unsigned long old_usemap_snr;
 343         static unsigned long old_pgdat_snr;
 344         struct pglist_data *pgdat = NODE_DATA(nid);
 345         int usemap_nid;
 346
 347         /* First call */
 348         if (!old_usemap_snr) {
 349                 old_usemap_snr = NR_MEM_SECTIONS;
 350                 old_pgdat_snr = NR_MEM_SECTIONS;
 351         }
 352
 353         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 354         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 355         if (usemap_snr == pgdat_snr)
 356                 return;
 357
 358         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 359                 /* skip redundant message */
 360                 return;
 361
 362         old_usemap_snr = usemap_snr;
 363         old_pgdat_snr = pgdat_snr;
 364
 365         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 366         if (usemap_nid != nid) {
 367                 pr_info("node %d must be removed before remove section %ld\n",
 368                         nid, usemap_snr);
 369                 return;
 370         }
 371         /*
 372          * There is a circular dependency.
 373          * Some platforms allow un-removable section because they will just
 374          * gather other removable sections for dynamic partitioning.
 375          * Just notify un-removable section's number here.
 376          */
 377         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 378                 usemap_snr, pgdat_snr, nid);
 379 }
 380 #else
 381 static unsigned long * __init
 382 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 383                                          unsigned long size)
 384 {
 385         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 386 }
 387
 388 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 389 {
 390 }
 391 #endif /* CONFIG_MEMORY_HOTREMOVE */
 392
 393 static void __init sparse_early_usemaps_alloc_node(void *data,
 394                                  unsigned long pnum_begin,
 395                                  unsigned long pnum_end,
 396                                  unsigned long usemap_count, int nodeid)
 397 {
 398         void *usemap;
 399         unsigned long pnum;
 400         unsigned long **usemap_map = (unsigned long **)data;
 401         int size = usemap_size();
 402
 403         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 404                                                           size * usemap_count);
 405         if (!usemap) {
 406                 pr_warn("%s: allocation failed\n", __func__);
 407                 return;
 408         }
 409
 410         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 411                 if (!present_section_nr(pnum))
 412                         continue;
 413                 usemap_map[pnum] = usemap;
 414                 usemap += size;
 415                 check_usemap_section_nr(nodeid, usemap_map[pnum]);
 416         }
 417 }
 418
 419 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 420 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 421 {
 422         struct page *map;
 423         unsigned long size;
 424
 425         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 426         if (map)
 427                 return map;
 428
 429         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 430         map = memblock_virt_alloc_try_nid(size,
 431                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 432                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 433         return map;
 434 }
 435 void __init sparse_mem_maps_populate_node(struct page **map_map,
 436                                           unsigned long pnum_begin,
 437                                           unsigned long pnum_end,
 438                                           unsigned long map_count, int nodeid)
 439 {
 440         void *map;
 441         unsigned long pnum;
 442         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 443
 444         map = alloc_remap(nodeid, size * map_count);
 445         if (map) {
 446                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 447                         if (!present_section_nr(pnum))
 448                                 continue;
 449                         map_map[pnum] = map;
 450                         map += size;
 451                 }
 452                 return;
 453         }
 454
 455         size = PAGE_ALIGN(size);
 456         map = memblock_virt_alloc_try_nid_raw(size * map_count,
 457                                               PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 458                                               BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 459         if (map) {
 460                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 461                         if (!present_section_nr(pnum))
 462                                 continue;
 463                         map_map[pnum] = map;
 464                         map += size;
 465                 }
 466                 return;
 467         }
 468
 469         /* fallback */
 470         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 471                 struct mem_section *ms;
 472
 473                 if (!present_section_nr(pnum))
 474                         continue;
 475                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 476                 if (map_map[pnum])
 477                         continue;
 478                 ms = __nr_to_section(pnum);
 479                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 480                        __func__);
 481                 ms->section_mem_map = 0;
 482         }
 483 }
 484 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 485
 486 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 487 static void __init sparse_early_mem_maps_alloc_node(void *data,
 488                                  unsigned long pnum_begin,
 489                                  unsigned long pnum_end,
 490                                  unsigned long map_count, int nodeid)
 491 {
 492         struct page **map_map = (struct page **)data;
 493         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 494                                          map_count, nodeid);
 495 }
 496 #else
 497 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 498 {
 499         struct page *map;
 500         struct mem_section *ms = __nr_to_section(pnum);
 501         int nid = sparse_early_nid(ms);
 502
 503         map = sparse_mem_map_populate(pnum, nid);
 504         if (map)
 505                 return map;
 506
 507         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 508                __func__);
 509         ms->section_mem_map = 0;
 510         return NULL;
 511 }
 512 #endif
 513
 514 void __weak __meminit vmemmap_populate_print_last(void)
 515 {
 516 }
 517
 518 /**
 519  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 520  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 521  */
 522 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 523                                         (void *, unsigned long, unsigned long,
 524                                         unsigned long, int), void *data)
 525 {
 526         unsigned long pnum;
 527         unsigned long map_count;
 528         int nodeid_begin = 0;
 529         unsigned long pnum_begin = 0;
 530
 531         for_each_present_section_nr(0, pnum) {
 532                 struct mem_section *ms;
 533
 534                 ms = __nr_to_section(pnum);
 535                 nodeid_begin = sparse_early_nid(ms);
 536                 pnum_begin = pnum;
 537                 break;
 538         }
 539         map_count = 1;
 540         for_each_present_section_nr(pnum_begin + 1, pnum) {
 541                 struct mem_section *ms;
 542                 int nodeid;
 543
 544                 ms = __nr_to_section(pnum);
 545                 nodeid = sparse_early_nid(ms);
 546                 if (nodeid == nodeid_begin) {
 547                         map_count++;
 548                         continue;
 549                 }
 550                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 551                 alloc_func(data, pnum_begin, pnum,
 552                                                 map_count, nodeid_begin);
 553                 /* new start, update count etc*/
 554                 nodeid_begin = nodeid;
 555                 pnum_begin = pnum;
 556                 map_count = 1;
 557         }
 558         /* ok, last chunk */
 559         alloc_func(data, pnum_begin, NR_MEM_SECTIONS,
 560                                                 map_count, nodeid_begin);
 561 }
 562
 563 /*
 564  * Allocate the accumulated non-linear sections, allocate a mem_map
 565  * for each and record the physical to section mapping.
 566  */
 567 void __init sparse_init(void)
 568 {
 569         unsigned long pnum;
 570         struct page *map;
 571         unsigned long *usemap;
 572         unsigned long **usemap_map;
 573         int size;
 574 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 575         int size2;
 576         struct page **map_map;
 577 #endif
 578
 579         /* see include/linux/mmzone.h 'struct mem_section' definition */
 580         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 581
 582         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 583         set_pageblock_order();
 584
 585         /*
 586          * map is using big page (aka 2M in x86 64 bit)
 587          * usemap is less one page (aka 24 bytes)
 588          * so alloc 2M (with 2M align) and 24 bytes in turn will
 589          * make next 2M slip to one more 2M later.
 590          * then in big system, the memory will have a lot of holes...
 591          * here try to allocate 2M pages continuously.
 592          *
 593          * powerpc need to call sparse_init_one_section right after each
 594          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 595          */
 596         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 597         usemap_map = memblock_virt_alloc(size, 0);
 598         if (!usemap_map)
 599                 panic("can not allocate usemap_map\n");
 600         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 601                                                         (void *)usemap_map);
 602
 603 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 604         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 605         map_map = memblock_virt_alloc(size2, 0);
 606         if (!map_map)
 607                 panic("can not allocate map_map\n");
 608         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 609                                                         (void *)map_map);
 610 #endif
 611
 612         for_each_present_section_nr(0, pnum) {
 613                 usemap = usemap_map[pnum];
 614                 if (!usemap)
 615                         continue;
 616
 617 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 618                 map = map_map[pnum];
 619 #else
 620                 map = sparse_early_mem_map_alloc(pnum);
 621 #endif
 622                 if (!map)
 623                         continue;
 624
 625                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 626                                                                 usemap);
 627         }
 628
 629         vmemmap_populate_print_last();
 630
 631 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 632         memblock_free_early(__pa(map_map), size2);
 633 #endif
 634         memblock_free_early(__pa(usemap_map), size);
 635 }
 636
 637 #ifdef CONFIG_MEMORY_HOTPLUG
 638
 639 /* Mark all memory sections within the pfn range as online */
 640 void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 641 {
 642         unsigned long pfn;
 643
 644         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 645                 unsigned long section_nr = pfn_to_section_nr(pfn);
 646                 struct mem_section *ms;
 647
 648                 /* onlining code should never touch invalid ranges */
 649                 if (WARN_ON(!valid_section_nr(section_nr)))
 650                         continue;
 651
 652                 ms = __nr_to_section(section_nr);
 653                 ms->section_mem_map |= SECTION_IS_ONLINE;
 654         }
 655 }
 656
 657 #ifdef CONFIG_MEMORY_HOTREMOVE
 658 /* Mark all memory sections within the pfn range as online */
 659 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 660 {
 661         unsigned long pfn;
 662
 663         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 664                 unsigned long section_nr = pfn_to_section_nr(start_pfn);
 665                 struct mem_section *ms;
 666
 667                 /*
 668                  * TODO this needs some double checking. Offlining code makes
 669                  * sure to check pfn_valid but those checks might be just bogus
 670                  */
 671                 if (WARN_ON(!valid_section_nr(section_nr)))
 672                         continue;
 673
 674                 ms = __nr_to_section(section_nr);
 675                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 676         }
 677 }
 678 #endif
 679
 680 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 681 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 682 {
 683         /* This will make the necessary allocations eventually. */
 684         return sparse_mem_map_populate(pnum, nid);
 685 }
 686 static void __kfree_section_memmap(struct page *memmap)
 687 {
 688         unsigned long start = (unsigned long)memmap;
 689         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 690
 691         vmemmap_free(start, end);
 692 }
 693 #ifdef CONFIG_MEMORY_HOTREMOVE
 694 static void free_map_bootmem(struct page *memmap)
 695 {
 696         unsigned long start = (unsigned long)memmap;
 697         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 698
 699         vmemmap_free(start, end);
 700 }
 701 #endif /* CONFIG_MEMORY_HOTREMOVE */
 702 #else
 703 static struct page *__kmalloc_section_memmap(void)
 704 {
 705         struct page *page, *ret;
 706         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 707
 708         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 709         if (page)
 710                 goto got_map_page;
 711
 712         ret = vmalloc(memmap_size);
 713         if (ret)
 714                 goto got_map_ptr;
 715
 716         return NULL;
 717 got_map_page:
 718         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 719 got_map_ptr:
 720
 721         return ret;
 722 }
 723
 724 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid)
 725 {
 726         return __kmalloc_section_memmap();
 727 }
 728
 729 static void __kfree_section_memmap(struct page *memmap)
 730 {
 731         if (is_vmalloc_addr(memmap))
 732                 vfree(memmap);
 733         else
 734                 free_pages((unsigned long)memmap,
 735                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 736 }
 737
 738 #ifdef CONFIG_MEMORY_HOTREMOVE
 739 static void free_map_bootmem(struct page *memmap)
 740 {
 741         unsigned long maps_section_nr, removing_section_nr, i;
 742         unsigned long magic, nr_pages;
 743         struct page *page = virt_to_page(memmap);
 744
 745         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 746                 >> PAGE_SHIFT;
 747
 748         for (i = 0; i < nr_pages; i++, page++) {
 749                 magic = (unsigned long) page->freelist;
 750
 751                 BUG_ON(magic == NODE_INFO);
 752
 753                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 754                 removing_section_nr = page_private(page);
 755
 756                 /*
 757                  * When this function is called, the removing section is
 758                  * logical offlined state. This means all pages are isolated
 759                  * from page allocator. If removing section's memmap is placed
 760                  * on the same section, it must not be freed.
 761                  * If it is freed, page allocator may allocate it which will
 762                  * be removed physically soon.
 763                  */
 764                 if (maps_section_nr != removing_section_nr)
 765                         put_page_bootmem(page);
 766         }
 767 }
 768 #endif /* CONFIG_MEMORY_HOTREMOVE */
 769 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 770
 771 /*
 772  * returns the number of sections whose mem_maps were properly
 773  * set.  If this is <=0, then that means that the passed-in
 774  * map was not consumed and must be freed.
 775  */
 776 int __meminit sparse_add_one_section(struct pglist_data *pgdat, unsigned long start_pfn)
 777 {
 778         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 779         struct mem_section *ms;
 780         struct page *memmap;
 781         unsigned long *usemap;
 782         unsigned long flags;
 783         int ret;
 784
 785         /*
 786          * no locking for this, because it does its own
 787          * plus, it does a kmalloc
 788          */
 789         ret = sparse_index_init(section_nr, pgdat->node_id);
 790         if (ret < 0 && ret != -EEXIST)
 791                 return ret;
 792         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id);
 793         if (!memmap)
 794                 return -ENOMEM;
 795         usemap = __kmalloc_section_usemap();
 796         if (!usemap) {
 797                 __kfree_section_memmap(memmap);
 798                 return -ENOMEM;
 799         }
 800
 801         pgdat_resize_lock(pgdat, &flags);
 802
 803         ms = __pfn_to_section(start_pfn);
 804         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 805                 ret = -EEXIST;
 806                 goto out;
 807         }
 808
 809         memset(memmap, 0, sizeof(struct page) * PAGES_PER_SECTION);
 810
 811         section_mark_present(ms);
 812
 813         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 814
 815 out:
 816         pgdat_resize_unlock(pgdat, &flags);
 817         if (ret <= 0) {
 818                 kfree(usemap);
 819                 __kfree_section_memmap(memmap);
 820         }
 821         return ret;
 822 }
 823
 824 #ifdef CONFIG_MEMORY_HOTREMOVE
 825 #ifdef CONFIG_MEMORY_FAILURE
 826 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 827 {
 828         int i;
 829
 830         if (!memmap)
 831                 return;
 832
 833         for (i = 0; i < nr_pages; i++) {
 834                 if (PageHWPoison(&memmap[i])) {
 835                         atomic_long_sub(1, &num_poisoned_pages);
 836                         ClearPageHWPoison(&memmap[i]);
 837                 }
 838         }
 839 }
 840 #else
 841 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 842 {
 843 }
 844 #endif
 845
 846 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 847 {
 848         struct page *usemap_page;
 849
 850         if (!usemap)
 851                 return;
 852
 853         usemap_page = virt_to_page(usemap);
 854         /*
 855          * Check to see if allocation came from hot-plug-add
 856          */
 857         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 858                 kfree(usemap);
 859                 if (memmap)
 860                         __kfree_section_memmap(memmap);
 861                 return;
 862         }
 863
 864         /*
 865          * The usemap came from bootmem. This is packed with other usemaps
 866          * on the section which has pgdat at boot time. Just keep it as is now.
 867          */
 868
 869         if (memmap)
 870                 free_map_bootmem(memmap);
 871 }
 872
 873 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 874                 unsigned long map_offset)
 875 {
 876         struct page *memmap = NULL;
 877         unsigned long *usemap = NULL, flags;
 878         struct pglist_data *pgdat = zone->zone_pgdat;
 879
 880         pgdat_resize_lock(pgdat, &flags);
 881         if (ms->section_mem_map) {
 882                 usemap = ms->pageblock_flags;
 883                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 884                                                 __section_nr(ms));
 885                 ms->section_mem_map = 0;
 886                 ms->pageblock_flags = NULL;
 887         }
 888         pgdat_resize_unlock(pgdat, &flags);
 889
 890         clear_hwpoisoned_pages(memmap + map_offset,
 891                         PAGES_PER_SECTION - map_offset);
 892         free_section_usemap(memmap, usemap);
 893 }
 894 #endif /* CONFIG_MEMORY_HOTREMOVE */
 895 #endif /* CONFIG_MEMORY_HOTPLUG */