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 /*
 204  * Record how many memory sections are marked as present
 205  * during system bootup.
 206  */
 207 static int __initdata nr_present_sections;
 208
 209 /* Record a memory area against a node. */
 210 void __init memory_present(int nid, unsigned long start, unsigned long end)
 211 {
 212         unsigned long pfn;
 213
 214 #ifdef CONFIG_SPARSEMEM_EXTREME
 215         if (unlikely(!mem_section)) {
 216                 unsigned long size, align;
 217
 218                 size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
 219                 align = 1 << (INTERNODE_CACHE_SHIFT);
 220                 mem_section = memblock_virt_alloc(size, align);
 221         }
 222 #endif
 223
 224         start &= PAGE_SECTION_MASK;
 225         mminit_validate_memmodel_limits(&start, &end);
 226         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 227                 unsigned long section = pfn_to_section_nr(pfn);
 228                 struct mem_section *ms;
 229
 230                 sparse_index_init(section, nid);
 231                 set_section_nid(section, nid);
 232
 233                 ms = __nr_to_section(section);
 234                 if (!ms->section_mem_map) {
 235                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 236                                                         SECTION_IS_ONLINE;
 237                         section_mark_present(ms);
 238                         nr_present_sections++;
 239                 }
 240         }
 241 }
 242
 243 /*
 244  * Subtle, we encode the real pfn into the mem_map such that
 245  * the identity pfn - section_mem_map will return the actual
 246  * physical page frame number.
 247  */
 248 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 249 {
 250         unsigned long coded_mem_map =
 251                 (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 252         BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 253         BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 254         return coded_mem_map;
 255 }
 256
 257 /*
 258  * Decode mem_map from the coded memmap
 259  */
 260 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 261 {
 262         /* mask off the extra low bits of information */
 263         coded_mem_map &= SECTION_MAP_MASK;
 264         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 265 }
 266
 267 static void __meminit sparse_init_one_section(struct mem_section *ms,
 268                 unsigned long pnum, struct page *mem_map,
 269                 unsigned long *pageblock_bitmap)
 270 {
 271         ms->section_mem_map &= ~SECTION_MAP_MASK;
 272         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 273                                                         SECTION_HAS_MEM_MAP;
 274         ms->pageblock_flags = pageblock_bitmap;
 275 }
 276
 277 unsigned long usemap_size(void)
 278 {
 279         return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 280 }
 281
 282 #ifdef CONFIG_MEMORY_HOTPLUG
 283 static unsigned long *__kmalloc_section_usemap(void)
 284 {
 285         return kmalloc(usemap_size(), GFP_KERNEL);
 286 }
 287 #endif /* CONFIG_MEMORY_HOTPLUG */
 288
 289 #ifdef CONFIG_MEMORY_HOTREMOVE
 290 static unsigned long * __init
 291 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 292                                          unsigned long size)
 293 {
 294         unsigned long goal, limit;
 295         unsigned long *p;
 296         int nid;
 297         /*
 298          * A page may contain usemaps for other sections preventing the
 299          * page being freed and making a section unremovable while
 300          * other sections referencing the usemap remain active. Similarly,
 301          * a pgdat can prevent a section being removed. If section A
 302          * contains a pgdat and section B contains the usemap, both
 303          * sections become inter-dependent. This allocates usemaps
 304          * from the same section as the pgdat where possible to avoid
 305          * this problem.
 306          */
 307         goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 308         limit = goal + (1UL << PA_SECTION_SHIFT);
 309         nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 310 again:
 311         p = memblock_virt_alloc_try_nid_nopanic(size,
 312                                                 SMP_CACHE_BYTES, goal, limit,
 313                                                 nid);
 314         if (!p && limit) {
 315                 limit = 0;
 316                 goto again;
 317         }
 318         return p;
 319 }
 320
 321 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 322 {
 323         unsigned long usemap_snr, pgdat_snr;
 324         static unsigned long old_usemap_snr;
 325         static unsigned long old_pgdat_snr;
 326         struct pglist_data *pgdat = NODE_DATA(nid);
 327         int usemap_nid;
 328
 329         /* First call */
 330         if (!old_usemap_snr) {
 331                 old_usemap_snr = NR_MEM_SECTIONS;
 332                 old_pgdat_snr = NR_MEM_SECTIONS;
 333         }
 334
 335         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 336         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 337         if (usemap_snr == pgdat_snr)
 338                 return;
 339
 340         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 341                 /* skip redundant message */
 342                 return;
 343
 344         old_usemap_snr = usemap_snr;
 345         old_pgdat_snr = pgdat_snr;
 346
 347         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 348         if (usemap_nid != nid) {
 349                 pr_info("node %d must be removed before remove section %ld\n",
 350                         nid, usemap_snr);
 351                 return;
 352         }
 353         /*
 354          * There is a circular dependency.
 355          * Some platforms allow un-removable section because they will just
 356          * gather other removable sections for dynamic partitioning.
 357          * Just notify un-removable section's number here.
 358          */
 359         pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 360                 usemap_snr, pgdat_snr, nid);
 361 }
 362 #else
 363 static unsigned long * __init
 364 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 365                                          unsigned long size)
 366 {
 367         return memblock_virt_alloc_node_nopanic(size, pgdat->node_id);
 368 }
 369
 370 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 371 {
 372 }
 373 #endif /* CONFIG_MEMORY_HOTREMOVE */
 374
 375 static void __init sparse_early_usemaps_alloc_node(void *data,
 376                                  unsigned long pnum_begin,
 377                                  unsigned long pnum_end,
 378                                  unsigned long usemap_count, int nodeid)
 379 {
 380         void *usemap;
 381         unsigned long pnum;
 382         unsigned long **usemap_map = (unsigned long **)data;
 383         int size = usemap_size();
 384         int nr_consumed_maps = 0;
 385
 386         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 387                                                           size * usemap_count);
 388         if (!usemap) {
 389                 pr_warn("%s: allocation failed\n", __func__);
 390                 return;
 391         }
 392
 393         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 394                 if (!present_section_nr(pnum))
 395                         continue;
 396                 usemap_map[nr_consumed_maps] = usemap;
 397                 usemap += size;
 398                 check_usemap_section_nr(nodeid, usemap_map[nr_consumed_maps]);
 399                 nr_consumed_maps++;
 400         }
 401 }
 402
 403 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 404 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
 405                 struct vmem_altmap *altmap)
 406 {
 407         struct page *map;
 408         unsigned long size;
 409
 410         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 411         map = memblock_virt_alloc_try_nid(size,
 412                                           PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 413                                           BOOTMEM_ALLOC_ACCESSIBLE, nid);
 414         return map;
 415 }
 416 void __init sparse_mem_maps_populate_node(struct page **map_map,
 417                                           unsigned long pnum_begin,
 418                                           unsigned long pnum_end,
 419                                           unsigned long map_count, int nodeid)
 420 {
 421         void *map;
 422         unsigned long pnum;
 423         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 424         int nr_consumed_maps;
 425
 426         size = PAGE_ALIGN(size);
 427         map = memblock_virt_alloc_try_nid_raw(size * map_count,
 428                                               PAGE_SIZE, __pa(MAX_DMA_ADDRESS),
 429                                               BOOTMEM_ALLOC_ACCESSIBLE, nodeid);
 430         if (map) {
 431                 nr_consumed_maps = 0;
 432                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 433                         if (!present_section_nr(pnum))
 434                                 continue;
 435                         map_map[nr_consumed_maps] = map;
 436                         map += size;
 437                         nr_consumed_maps++;
 438                 }
 439                 return;
 440         }
 441
 442         /* fallback */
 443         nr_consumed_maps = 0;
 444         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 445                 struct mem_section *ms;
 446
 447                 if (!present_section_nr(pnum))
 448                         continue;
 449                 map_map[nr_consumed_maps] =
 450                                 sparse_mem_map_populate(pnum, nodeid, NULL);
 451                 if (map_map[nr_consumed_maps++])
 452                         continue;
 453                 ms = __nr_to_section(pnum);
 454                 pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 455                        __func__);
 456         }
 457 }
 458 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 459
 460 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 461 static void __init sparse_early_mem_maps_alloc_node(void *data,
 462                                  unsigned long pnum_begin,
 463                                  unsigned long pnum_end,
 464                                  unsigned long map_count, int nodeid)
 465 {
 466         struct page **map_map = (struct page **)data;
 467         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 468                                          map_count, nodeid);
 469 }
 470 #else
 471 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 472 {
 473         struct page *map;
 474         struct mem_section *ms = __nr_to_section(pnum);
 475         int nid = sparse_early_nid(ms);
 476
 477         map = sparse_mem_map_populate(pnum, nid, NULL);
 478         if (map)
 479                 return map;
 480
 481         pr_err("%s: sparsemem memory map backing failed some memory will not be available\n",
 482                __func__);
 483         return NULL;
 484 }
 485 #endif
 486
 487 void __weak __meminit vmemmap_populate_print_last(void)
 488 {
 489 }
 490
 491 /**
 492  *  alloc_usemap_and_memmap - memory alloction for pageblock flags and vmemmap
 493  *  @map: usemap_map for pageblock flags or mmap_map for vmemmap
 494  *  @unit_size: size of map unit
 495  */
 496 static void __init alloc_usemap_and_memmap(void (*alloc_func)
 497                                         (void *, unsigned long, unsigned long,
 498                                         unsigned long, int), void *data,
 499                                         int data_unit_size)
 500 {
 501         unsigned long pnum;
 502         unsigned long map_count;
 503         int nodeid_begin = 0;
 504         unsigned long pnum_begin = 0;
 505
 506         for_each_present_section_nr(0, pnum) {
 507                 struct mem_section *ms;
 508
 509                 ms = __nr_to_section(pnum);
 510                 nodeid_begin = sparse_early_nid(ms);
 511                 pnum_begin = pnum;
 512                 break;
 513         }
 514         map_count = 1;
 515         for_each_present_section_nr(pnum_begin + 1, pnum) {
 516                 struct mem_section *ms;
 517                 int nodeid;
 518
 519                 ms = __nr_to_section(pnum);
 520                 nodeid = sparse_early_nid(ms);
 521                 if (nodeid == nodeid_begin) {
 522                         map_count++;
 523                         continue;
 524                 }
 525                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 526                 alloc_func(data, pnum_begin, pnum,
 527                                                 map_count, nodeid_begin);
 528                 /* new start, update count etc*/
 529                 nodeid_begin = nodeid;
 530                 pnum_begin = pnum;
 531                 data += map_count * data_unit_size;
 532                 map_count = 1;
 533         }
 534         /* ok, last chunk */
 535         alloc_func(data, pnum_begin, __highest_present_section_nr+1,
 536                                                 map_count, nodeid_begin);
 537 }
 538
 539 /*
 540  * Allocate the accumulated non-linear sections, allocate a mem_map
 541  * for each and record the physical to section mapping.
 542  */
 543 void __init sparse_init(void)
 544 {
 545         unsigned long pnum;
 546         struct page *map;
 547         unsigned long *usemap;
 548         unsigned long **usemap_map;
 549         int size;
 550         int nr_consumed_maps = 0;
 551 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 552         int size2;
 553         struct page **map_map;
 554 #endif
 555
 556         /* see include/linux/mmzone.h 'struct mem_section' definition */
 557         BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 558
 559         /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 560         set_pageblock_order();
 561
 562         /*
 563          * map is using big page (aka 2M in x86 64 bit)
 564          * usemap is less one page (aka 24 bytes)
 565          * so alloc 2M (with 2M align) and 24 bytes in turn will
 566          * make next 2M slip to one more 2M later.
 567          * then in big system, the memory will have a lot of holes...
 568          * here try to allocate 2M pages continuously.
 569          *
 570          * powerpc need to call sparse_init_one_section right after each
 571          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 572          */
 573         size = sizeof(unsigned long *) * nr_present_sections;
 574         usemap_map = memblock_virt_alloc(size, 0);
 575         if (!usemap_map)
 576                 panic("can not allocate usemap_map\n");
 577         alloc_usemap_and_memmap(sparse_early_usemaps_alloc_node,
 578                                 (void *)usemap_map,
 579                                 sizeof(usemap_map[0]));
 580
 581 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 582         size2 = sizeof(struct page *) * nr_present_sections;
 583         map_map = memblock_virt_alloc(size2, 0);
 584         if (!map_map)
 585                 panic("can not allocate map_map\n");
 586         alloc_usemap_and_memmap(sparse_early_mem_maps_alloc_node,
 587                                 (void *)map_map,
 588                                 sizeof(map_map[0]));
 589 #endif
 590
 591         /*
 592          * The number of present sections stored in nr_present_sections
 593          * are kept the same since mem sections are marked as present in
 594          * memory_present(). In this for loop, we need check which sections
 595          * failed to allocate memmap or usemap, then clear its
 596          * ->section_mem_map accordingly. During this process, we need
 597          * increase 'nr_consumed_maps' whether its allocation of memmap
 598          * or usemap failed or not, so that after we handle the i-th
 599          * memory section, can get memmap and usemap of (i+1)-th section
 600          * correctly.
 601          */
 602         for_each_present_section_nr(0, pnum) {
 603                 struct mem_section *ms;
 604
 605                 if (nr_consumed_maps >= nr_present_sections) {
 606                         pr_err("nr_consumed_maps goes beyond nr_present_sections\n");
 607                         break;
 608                 }
 609                 ms = __nr_to_section(pnum);
 610                 usemap = usemap_map[nr_consumed_maps];
 611                 if (!usemap) {
 612                         ms->section_mem_map = 0;
 613                         nr_consumed_maps++;
 614                         continue;
 615                 }
 616
 617 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 618                 map = map_map[nr_consumed_maps];
 619 #else
 620                 map = sparse_early_mem_map_alloc(pnum);
 621 #endif
 622                 if (!map) {
 623                         ms->section_mem_map = 0;
 624                         nr_consumed_maps++;
 625                         continue;
 626                 }
 627
 628                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 629                                                                 usemap);
 630                 nr_consumed_maps++;
 631         }
 632
 633         vmemmap_populate_print_last();
 634
 635 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 636         memblock_free_early(__pa(map_map), size2);
 637 #endif
 638         memblock_free_early(__pa(usemap_map), size);
 639 }
 640
 641 #ifdef CONFIG_MEMORY_HOTPLUG
 642
 643 /* Mark all memory sections within the pfn range as online */
 644 void online_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(pfn);
 650                 struct mem_section *ms;
 651
 652                 /* onlining code should never touch invalid ranges */
 653                 if (WARN_ON(!valid_section_nr(section_nr)))
 654                         continue;
 655
 656                 ms = __nr_to_section(section_nr);
 657                 ms->section_mem_map |= SECTION_IS_ONLINE;
 658         }
 659 }
 660
 661 #ifdef CONFIG_MEMORY_HOTREMOVE
 662 /* Mark all memory sections within the pfn range as online */
 663 void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 664 {
 665         unsigned long pfn;
 666
 667         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 668                 unsigned long section_nr = pfn_to_section_nr(pfn);
 669                 struct mem_section *ms;
 670
 671                 /*
 672                  * TODO this needs some double checking. Offlining code makes
 673                  * sure to check pfn_valid but those checks might be just bogus
 674                  */
 675                 if (WARN_ON(!valid_section_nr(section_nr)))
 676                         continue;
 677
 678                 ms = __nr_to_section(section_nr);
 679                 ms->section_mem_map &= ~SECTION_IS_ONLINE;
 680         }
 681 }
 682 #endif
 683
 684 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 685 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 686                 struct vmem_altmap *altmap)
 687 {
 688         /* This will make the necessary allocations eventually. */
 689         return sparse_mem_map_populate(pnum, nid, altmap);
 690 }
 691 static void __kfree_section_memmap(struct page *memmap,
 692                 struct vmem_altmap *altmap)
 693 {
 694         unsigned long start = (unsigned long)memmap;
 695         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 696
 697         vmemmap_free(start, end, altmap);
 698 }
 699 #ifdef CONFIG_MEMORY_HOTREMOVE
 700 static void free_map_bootmem(struct page *memmap)
 701 {
 702         unsigned long start = (unsigned long)memmap;
 703         unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 704
 705         vmemmap_free(start, end, NULL);
 706 }
 707 #endif /* CONFIG_MEMORY_HOTREMOVE */
 708 #else
 709 static struct page *__kmalloc_section_memmap(void)
 710 {
 711         struct page *page, *ret;
 712         unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 713
 714         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 715         if (page)
 716                 goto got_map_page;
 717
 718         ret = vmalloc(memmap_size);
 719         if (ret)
 720                 goto got_map_ptr;
 721
 722         return NULL;
 723 got_map_page:
 724         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 725 got_map_ptr:
 726
 727         return ret;
 728 }
 729
 730 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 731                 struct vmem_altmap *altmap)
 732 {
 733         return __kmalloc_section_memmap();
 734 }
 735
 736 static void __kfree_section_memmap(struct page *memmap,
 737                 struct vmem_altmap *altmap)
 738 {
 739         if (is_vmalloc_addr(memmap))
 740                 vfree(memmap);
 741         else
 742                 free_pages((unsigned long)memmap,
 743                            get_order(sizeof(struct page) * PAGES_PER_SECTION));
 744 }
 745
 746 #ifdef CONFIG_MEMORY_HOTREMOVE
 747 static void free_map_bootmem(struct page *memmap)
 748 {
 749         unsigned long maps_section_nr, removing_section_nr, i;
 750         unsigned long magic, nr_pages;
 751         struct page *page = virt_to_page(memmap);
 752
 753         nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 754                 >> PAGE_SHIFT;
 755
 756         for (i = 0; i < nr_pages; i++, page++) {
 757                 magic = (unsigned long) page->freelist;
 758
 759                 BUG_ON(magic == NODE_INFO);
 760
 761                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 762                 removing_section_nr = page_private(page);
 763
 764                 /*
 765                  * When this function is called, the removing section is
 766                  * logical offlined state. This means all pages are isolated
 767                  * from page allocator. If removing section's memmap is placed
 768                  * on the same section, it must not be freed.
 769                  * If it is freed, page allocator may allocate it which will
 770                  * be removed physically soon.
 771                  */
 772                 if (maps_section_nr != removing_section_nr)
 773                         put_page_bootmem(page);
 774         }
 775 }
 776 #endif /* CONFIG_MEMORY_HOTREMOVE */
 777 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 778
 779 /*
 780  * returns the number of sections whose mem_maps were properly
 781  * set.  If this is <=0, then that means that the passed-in
 782  * map was not consumed and must be freed.
 783  */
 784 int __meminit sparse_add_one_section(struct pglist_data *pgdat,
 785                 unsigned long start_pfn, struct vmem_altmap *altmap)
 786 {
 787         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 788         struct mem_section *ms;
 789         struct page *memmap;
 790         unsigned long *usemap;
 791         unsigned long flags;
 792         int ret;
 793
 794         /*
 795          * no locking for this, because it does its own
 796          * plus, it does a kmalloc
 797          */
 798         ret = sparse_index_init(section_nr, pgdat->node_id);
 799         if (ret < 0 && ret != -EEXIST)
 800                 return ret;
 801         ret = 0;
 802         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, altmap);
 803         if (!memmap)
 804                 return -ENOMEM;
 805         usemap = __kmalloc_section_usemap();
 806         if (!usemap) {
 807                 __kfree_section_memmap(memmap, altmap);
 808                 return -ENOMEM;
 809         }
 810
 811         pgdat_resize_lock(pgdat, &flags);
 812
 813         ms = __pfn_to_section(start_pfn);
 814         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 815                 ret = -EEXIST;
 816                 goto out;
 817         }
 818
 819 #ifdef CONFIG_DEBUG_VM
 820         /*
 821          * Poison uninitialized struct pages in order to catch invalid flags
 822          * combinations.
 823          */
 824         memset(memmap, PAGE_POISON_PATTERN, sizeof(struct page) * PAGES_PER_SECTION);
 825 #endif
 826
 827         section_mark_present(ms);
 828         sparse_init_one_section(ms, section_nr, memmap, usemap);
 829
 830 out:
 831         pgdat_resize_unlock(pgdat, &flags);
 832         if (ret < 0) {
 833                 kfree(usemap);
 834                 __kfree_section_memmap(memmap, altmap);
 835         }
 836         return ret;
 837 }
 838
 839 #ifdef CONFIG_MEMORY_HOTREMOVE
 840 #ifdef CONFIG_MEMORY_FAILURE
 841 static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 842 {
 843         int i;
 844
 845         if (!memmap)
 846                 return;
 847
 848         for (i = 0; i < nr_pages; i++) {
 849                 if (PageHWPoison(&memmap[i])) {
 850                         atomic_long_sub(1, &num_poisoned_pages);
 851                         ClearPageHWPoison(&memmap[i]);
 852                 }
 853         }
 854 }
 855 #else
 856 static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 857 {
 858 }
 859 #endif
 860
 861 static void free_section_usemap(struct page *memmap, unsigned long *usemap,
 862                 struct vmem_altmap *altmap)
 863 {
 864         struct page *usemap_page;
 865
 866         if (!usemap)
 867                 return;
 868
 869         usemap_page = virt_to_page(usemap);
 870         /*
 871          * Check to see if allocation came from hot-plug-add
 872          */
 873         if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 874                 kfree(usemap);
 875                 if (memmap)
 876                         __kfree_section_memmap(memmap, altmap);
 877                 return;
 878         }
 879
 880         /*
 881          * The usemap came from bootmem. This is packed with other usemaps
 882          * on the section which has pgdat at boot time. Just keep it as is now.
 883          */
 884
 885         if (memmap)
 886                 free_map_bootmem(memmap);
 887 }
 888
 889 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 890                 unsigned long map_offset, struct vmem_altmap *altmap)
 891 {
 892         struct page *memmap = NULL;
 893         unsigned long *usemap = NULL, flags;
 894         struct pglist_data *pgdat = zone->zone_pgdat;
 895
 896         pgdat_resize_lock(pgdat, &flags);
 897         if (ms->section_mem_map) {
 898                 usemap = ms->pageblock_flags;
 899                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 900                                                 __section_nr(ms));
 901                 ms->section_mem_map = 0;
 902                 ms->pageblock_flags = NULL;
 903         }
 904         pgdat_resize_unlock(pgdat, &flags);
 905
 906         clear_hwpoisoned_pages(memmap + map_offset,
 907                         PAGES_PER_SECTION - map_offset);
 908         free_section_usemap(memmap, usemap, altmap);
 909 }
 910 #endif /* CONFIG_MEMORY_HOTREMOVE */
 911 #endif /* CONFIG_MEMORY_HOTPLUG */