linux/mm/sparse.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * sparse memory mappings.
   4 */
   5#include <linux/mm.h>
   6#include <linux/slab.h>
   7#include <linux/mmzone.h>
   8#include <linux/memblock.h>
   9#include <linux/compiler.h>
  10#include <linux/highmem.h>
  11#include <linux/export.h>
  12#include <linux/spinlock.h>
  13#include <linux/vmalloc.h>
  14
  15#include "internal.h"
  16#include <asm/dma.h>
  17#include <asm/pgalloc.h>
  18#include <asm/pgtable.h>
  19
  20/*
  21 * Permanent SPARSEMEM data:
  22 *
  23 * 1) mem_section       - memory sections, mem_map's for valid memory
  24 */
  25#ifdef CONFIG_SPARSEMEM_EXTREME
  26struct mem_section **mem_section;
  27#else
  28struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  29        ____cacheline_internodealigned_in_smp;
  30#endif
  31EXPORT_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
  40static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  41#else
  42static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  43#endif
  44
  45int page_to_nid(const struct page *page)
  46{
  47        return section_to_node_table[page_to_section(page)];
  48}
  49EXPORT_SYMBOL(page_to_nid);
  50
  51static 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 */
  56static inline void set_section_nid(unsigned long section_nr, int nid)
  57{
  58}
  59#endif
  60
  61#ifdef CONFIG_SPARSEMEM_EXTREME
  62static noinline struct mem_section __ref *sparse_index_alloc(int nid)
  63{
  64        struct mem_section *section = NULL;
  65        unsigned long array_size = SECTIONS_PER_ROOT *
  66                                   sizeof(struct mem_section);
  67
  68        if (slab_is_available()) {
  69                section = kzalloc_node(array_size, GFP_KERNEL, nid);
  70        } else {
  71                section = memblock_alloc_node(array_size, SMP_CACHE_BYTES,
  72                                              nid);
  73                if (!section)
  74                        panic("%s: Failed to allocate %lu bytes nid=%d\n",
  75                              __func__, array_size, nid);
  76        }
  77
  78        return section;
  79}
  80
  81static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  82{
  83        unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  84        struct mem_section *section;
  85
  86        if (mem_section[root])
  87                return -EEXIST;
  88
  89        section = sparse_index_alloc(nid);
  90        if (!section)
  91                return -ENOMEM;
  92
  93        mem_section[root] = section;
  94
  95        return 0;
  96}
  97#else /* !SPARSEMEM_EXTREME */
  98static inline int sparse_index_init(unsigned long section_nr, int nid)
  99{
 100        return 0;
 101}
 102#endif
 103
 104#ifdef CONFIG_SPARSEMEM_EXTREME
 105int __section_nr(struct mem_section* ms)
 106{
 107        unsigned long root_nr;
 108        struct mem_section *root = NULL;
 109
 110        for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 111                root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 112                if (!root)
 113                        continue;
 114
 115                if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 116                     break;
 117        }
 118
 119        VM_BUG_ON(!root);
 120
 121        return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 122}
 123#else
 124int __section_nr(struct mem_section* ms)
 125{
 126        return (int)(ms - mem_section[0]);
 127}
 128#endif
 129
 130/*
 131 * During early boot, before section_mem_map is used for an actual
 132 * mem_map, we use section_mem_map to store the section's NUMA
 133 * node.  This keeps us from having to use another data structure.  The
 134 * node information is cleared just before we store the real mem_map.
 135 */
 136static inline unsigned long sparse_encode_early_nid(int nid)
 137{
 138        return (nid << SECTION_NID_SHIFT);
 139}
 140
 141static inline int sparse_early_nid(struct mem_section *section)
 142{
 143        return (section->section_mem_map >> SECTION_NID_SHIFT);
 144}
 145
 146/* Validate the physical addressing limitations of the model */
 147void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 148                                                unsigned long *end_pfn)
 149{
 150        unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 151
 152        /*
 153         * Sanity checks - do not allow an architecture to pass
 154         * in larger pfns than the maximum scope of sparsemem:
 155         */
 156        if (*start_pfn > max_sparsemem_pfn) {
 157                mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 158                        "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 159                        *start_pfn, *end_pfn, max_sparsemem_pfn);
 160                WARN_ON_ONCE(1);
 161                *start_pfn = max_sparsemem_pfn;
 162                *end_pfn = max_sparsemem_pfn;
 163        } else if (*end_pfn > max_sparsemem_pfn) {
 164                mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 165                        "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 166                        *start_pfn, *end_pfn, max_sparsemem_pfn);
 167                WARN_ON_ONCE(1);
 168                *end_pfn = max_sparsemem_pfn;
 169        }
 170}
 171
 172/*
 173 * There are a number of times that we loop over NR_MEM_SECTIONS,
 174 * looking for section_present() on each.  But, when we have very
 175 * large physical address spaces, NR_MEM_SECTIONS can also be
 176 * very large which makes the loops quite long.
 177 *
 178 * Keeping track of this gives us an easy way to break out of
 179 * those loops early.
 180 */
 181int __highest_present_section_nr;
 182static void section_mark_present(struct mem_section *ms)
 183{
 184        int section_nr = __section_nr(ms);
 185
 186        if (section_nr > __highest_present_section_nr)
 187                __highest_present_section_nr = section_nr;
 188
 189        ms->section_mem_map |= SECTION_MARKED_PRESENT;
 190}
 191
 192static inline int next_present_section_nr(int section_nr)
 193{
 194        do {
 195                section_nr++;
 196                if (present_section_nr(section_nr))
 197                        return section_nr;
 198        } while ((section_nr <= __highest_present_section_nr));
 199
 200        return -1;
 201}
 202#define for_each_present_section_nr(start, section_nr)          \
 203        for (section_nr = next_present_section_nr(start-1);     \
 204             ((section_nr != -1) &&                             \
 205              (section_nr <= __highest_present_section_nr));    \
 206             section_nr = next_present_section_nr(section_nr))
 207
 208static inline unsigned long first_present_section_nr(void)
 209{
 210        return next_present_section_nr(-1);
 211}
 212
 213/* Record a memory area against a node. */
 214void __init memory_present(int nid, unsigned long start, unsigned long end)
 215{
 216        unsigned long pfn;
 217
 218#ifdef CONFIG_SPARSEMEM_EXTREME
 219        if (unlikely(!mem_section)) {
 220                unsigned long size, align;
 221
 222                size = sizeof(struct mem_section*) * NR_SECTION_ROOTS;
 223                align = 1 << (INTERNODE_CACHE_SHIFT);
 224                mem_section = memblock_alloc(size, align);
 225                if (!mem_section)
 226                        panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
 227                              __func__, size, align);
 228        }
 229#endif
 230
 231        start &= PAGE_SECTION_MASK;
 232        mminit_validate_memmodel_limits(&start, &end);
 233        for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 234                unsigned long section = pfn_to_section_nr(pfn);
 235                struct mem_section *ms;
 236
 237                sparse_index_init(section, nid);
 238                set_section_nid(section, nid);
 239
 240                ms = __nr_to_section(section);
 241                if (!ms->section_mem_map) {
 242                        ms->section_mem_map = sparse_encode_early_nid(nid) |
 243                                                        SECTION_IS_ONLINE;
 244                        section_mark_present(ms);
 245                }
 246        }
 247}
 248
 249/*
 250 * Mark all memblocks as present using memory_present(). This is a
 251 * convienence function that is useful for a number of arches
 252 * to mark all of the systems memory as present during initialization.
 253 */
 254void __init memblocks_present(void)
 255{
 256        struct memblock_region *reg;
 257
 258        for_each_memblock(memory, reg) {
 259                memory_present(memblock_get_region_node(reg),
 260                               memblock_region_memory_base_pfn(reg),
 261                               memblock_region_memory_end_pfn(reg));
 262        }
 263}
 264
 265/*
 266 * Subtle, we encode the real pfn into the mem_map such that
 267 * the identity pfn - section_mem_map will return the actual
 268 * physical page frame number.
 269 */
 270static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 271{
 272        unsigned long coded_mem_map =
 273                (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 274        BUILD_BUG_ON(SECTION_MAP_LAST_BIT > (1UL<<PFN_SECTION_SHIFT));
 275        BUG_ON(coded_mem_map & ~SECTION_MAP_MASK);
 276        return coded_mem_map;
 277}
 278
 279/*
 280 * Decode mem_map from the coded memmap
 281 */
 282struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 283{
 284        /* mask off the extra low bits of information */
 285        coded_mem_map &= SECTION_MAP_MASK;
 286        return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 287}
 288
 289static void __meminit sparse_init_one_section(struct mem_section *ms,
 290                unsigned long pnum, struct page *mem_map,
 291                unsigned long *pageblock_bitmap)
 292{
 293        ms->section_mem_map &= ~SECTION_MAP_MASK;
 294        ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 295                                                        SECTION_HAS_MEM_MAP;
 296        ms->pageblock_flags = pageblock_bitmap;
 297}
 298
 299unsigned long usemap_size(void)
 300{
 301        return BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS) * sizeof(unsigned long);
 302}
 303
 304#ifdef CONFIG_MEMORY_HOTPLUG
 305static unsigned long *__kmalloc_section_usemap(void)
 306{
 307        return kmalloc(usemap_size(), GFP_KERNEL);
 308}
 309#endif /* CONFIG_MEMORY_HOTPLUG */
 310
 311#ifdef CONFIG_MEMORY_HOTREMOVE
 312static unsigned long * __init
 313sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 314                                         unsigned long size)
 315{
 316        unsigned long goal, limit;
 317        unsigned long *p;
 318        int nid;
 319        /*
 320         * A page may contain usemaps for other sections preventing the
 321         * page being freed and making a section unremovable while
 322         * other sections referencing the usemap remain active. Similarly,
 323         * a pgdat can prevent a section being removed. If section A
 324         * contains a pgdat and section B contains the usemap, both
 325         * sections become inter-dependent. This allocates usemaps
 326         * from the same section as the pgdat where possible to avoid
 327         * this problem.
 328         */
 329        goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 330        limit = goal + (1UL << PA_SECTION_SHIFT);
 331        nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 332again:
 333        p = memblock_alloc_try_nid(size, SMP_CACHE_BYTES, goal, limit, nid);
 334        if (!p && limit) {
 335                limit = 0;
 336                goto again;
 337        }
 338        return p;
 339}
 340
 341static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 342{
 343        unsigned long usemap_snr, pgdat_snr;
 344        static unsigned long old_usemap_snr;
 345        static unsigned long old_pgdat_snr;
 346        struct pglist_data *pgdat = NODE_DATA(nid);
 347        int usemap_nid;
 348
 349        /* First call */
 350        if (!old_usemap_snr) {
 351                old_usemap_snr = NR_MEM_SECTIONS;
 352                old_pgdat_snr = NR_MEM_SECTIONS;
 353        }
 354
 355        usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 356        pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 357        if (usemap_snr == pgdat_snr)
 358                return;
 359
 360        if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 361                /* skip redundant message */
 362                return;
 363
 364        old_usemap_snr = usemap_snr;
 365        old_pgdat_snr = pgdat_snr;
 366
 367        usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 368        if (usemap_nid != nid) {
 369                pr_info("node %d must be removed before remove section %ld\n",
 370                        nid, usemap_snr);
 371                return;
 372        }
 373        /*
 374         * There is a circular dependency.
 375         * Some platforms allow un-removable section because they will just
 376         * gather other removable sections for dynamic partitioning.
 377         * Just notify un-removable section's number here.
 378         */
 379        pr_info("Section %ld and %ld (node %d) have a circular dependency on usemap and pgdat allocations\n",
 380                usemap_snr, pgdat_snr, nid);
 381}
 382#else
 383static unsigned long * __init
 384sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 385                                         unsigned long size)
 386{
 387        return memblock_alloc_node(size, SMP_CACHE_BYTES, pgdat->node_id);
 388}
 389
 390static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 391{
 392}
 393#endif /* CONFIG_MEMORY_HOTREMOVE */
 394
 395#ifdef CONFIG_SPARSEMEM_VMEMMAP
 396static unsigned long __init section_map_size(void)
 397{
 398        return ALIGN(sizeof(struct page) * PAGES_PER_SECTION, PMD_SIZE);
 399}
 400
 401#else
 402static unsigned long __init section_map_size(void)
 403{
 404        return PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 405}
 406
 407struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid,
 408                struct vmem_altmap *altmap)
 409{
 410        unsigned long size = section_map_size();
 411        struct page *map = sparse_buffer_alloc(size);
 412        phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 413
 414        if (map)
 415                return map;
 416
 417        map = memblock_alloc_try_nid(size,
 418                                          PAGE_SIZE, addr,
 419                                          MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 420        if (!map)
 421                panic("%s: Failed to allocate %lu bytes align=0x%lx nid=%d from=%pa\n",
 422                      __func__, size, PAGE_SIZE, nid, &addr);
 423
 424        return map;
 425}
 426#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 427
 428static void *sparsemap_buf __meminitdata;
 429static void *sparsemap_buf_end __meminitdata;
 430
 431static void __init sparse_buffer_init(unsigned long size, int nid)
 432{
 433        phys_addr_t addr = __pa(MAX_DMA_ADDRESS);
 434        WARN_ON(sparsemap_buf); /* forgot to call sparse_buffer_fini()? */
 435        sparsemap_buf =
 436                memblock_alloc_try_nid_raw(size, PAGE_SIZE,
 437                                                addr,
 438                                                MEMBLOCK_ALLOC_ACCESSIBLE, nid);
 439        sparsemap_buf_end = sparsemap_buf + size;
 440}
 441
 442static void __init sparse_buffer_fini(void)
 443{
 444        unsigned long size = sparsemap_buf_end - sparsemap_buf;
 445
 446        if (sparsemap_buf && size > 0)
 447                memblock_free_early(__pa(sparsemap_buf), size);
 448        sparsemap_buf = NULL;
 449}
 450
 451void * __meminit sparse_buffer_alloc(unsigned long size)
 452{
 453        void *ptr = NULL;
 454
 455        if (sparsemap_buf) {
 456                ptr = PTR_ALIGN(sparsemap_buf, size);
 457                if (ptr + size > sparsemap_buf_end)
 458                        ptr = NULL;
 459                else
 460                        sparsemap_buf = ptr + size;
 461        }
 462        return ptr;
 463}
 464
 465void __weak __meminit vmemmap_populate_print_last(void)
 466{
 467}
 468
 469/*
 470 * Initialize sparse on a specific node. The node spans [pnum_begin, pnum_end)
 471 * And number of present sections in this node is map_count.
 472 */
 473static void __init sparse_init_nid(int nid, unsigned long pnum_begin,
 474                                   unsigned long pnum_end,
 475                                   unsigned long map_count)
 476{
 477        unsigned long pnum, usemap_longs, *usemap;
 478        struct page *map;
 479
 480        usemap_longs = BITS_TO_LONGS(SECTION_BLOCKFLAGS_BITS);
 481        usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nid),
 482                                                          usemap_size() *
 483                                                          map_count);
 484        if (!usemap) {
 485                pr_err("%s: node[%d] usemap allocation failed", __func__, nid);
 486                goto failed;
 487        }
 488        sparse_buffer_init(map_count * section_map_size(), nid);
 489        for_each_present_section_nr(pnum_begin, pnum) {
 490                if (pnum >= pnum_end)
 491                        break;
 492
 493                map = sparse_mem_map_populate(pnum, nid, NULL);
 494                if (!map) {
 495                        pr_err("%s: node[%d] memory map backing failed. Some memory will not be available.",
 496                               __func__, nid);
 497                        pnum_begin = pnum;
 498                        goto failed;
 499                }
 500                check_usemap_section_nr(nid, usemap);
 501                sparse_init_one_section(__nr_to_section(pnum), pnum, map, usemap);
 502                usemap += usemap_longs;
 503        }
 504        sparse_buffer_fini();
 505        return;
 506failed:
 507        /* We failed to allocate, mark all the following pnums as not present */
 508        for_each_present_section_nr(pnum_begin, pnum) {
 509                struct mem_section *ms;
 510
 511                if (pnum >= pnum_end)
 512                        break;
 513                ms = __nr_to_section(pnum);
 514                ms->section_mem_map = 0;
 515        }
 516}
 517
 518/*
 519 * Allocate the accumulated non-linear sections, allocate a mem_map
 520 * for each and record the physical to section mapping.
 521 */
 522void __init sparse_init(void)
 523{
 524        unsigned long pnum_begin = first_present_section_nr();
 525        int nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 526        unsigned long pnum_end, map_count = 1;
 527
 528        /* Setup pageblock_order for HUGETLB_PAGE_SIZE_VARIABLE */
 529        set_pageblock_order();
 530
 531        for_each_present_section_nr(pnum_begin + 1, pnum_end) {
 532                int nid = sparse_early_nid(__nr_to_section(pnum_end));
 533
 534                if (nid == nid_begin) {
 535                        map_count++;
 536                        continue;
 537                }
 538                /* Init node with sections in range [pnum_begin, pnum_end) */
 539                sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 540                nid_begin = nid;
 541                pnum_begin = pnum_end;
 542                map_count = 1;
 543        }
 544        /* cover the last node */
 545        sparse_init_nid(nid_begin, pnum_begin, pnum_end, map_count);
 546        vmemmap_populate_print_last();
 547}
 548
 549#ifdef CONFIG_MEMORY_HOTPLUG
 550
 551/* Mark all memory sections within the pfn range as online */
 552void online_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 553{
 554        unsigned long pfn;
 555
 556        for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 557                unsigned long section_nr = pfn_to_section_nr(pfn);
 558                struct mem_section *ms;
 559
 560                /* onlining code should never touch invalid ranges */
 561                if (WARN_ON(!valid_section_nr(section_nr)))
 562                        continue;
 563
 564                ms = __nr_to_section(section_nr);
 565                ms->section_mem_map |= SECTION_IS_ONLINE;
 566        }
 567}
 568
 569#ifdef CONFIG_MEMORY_HOTREMOVE
 570/* Mark all memory sections within the pfn range as offline */
 571void offline_mem_sections(unsigned long start_pfn, unsigned long end_pfn)
 572{
 573        unsigned long pfn;
 574
 575        for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 576                unsigned long section_nr = pfn_to_section_nr(pfn);
 577                struct mem_section *ms;
 578
 579                /*
 580                 * TODO this needs some double checking. Offlining code makes
 581                 * sure to check pfn_valid but those checks might be just bogus
 582                 */
 583                if (WARN_ON(!valid_section_nr(section_nr)))
 584                        continue;
 585
 586                ms = __nr_to_section(section_nr);
 587                ms->section_mem_map &= ~SECTION_IS_ONLINE;
 588        }
 589}
 590#endif
 591
 592#ifdef CONFIG_SPARSEMEM_VMEMMAP
 593static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 594                struct vmem_altmap *altmap)
 595{
 596        /* This will make the necessary allocations eventually. */
 597        return sparse_mem_map_populate(pnum, nid, altmap);
 598}
 599static void __kfree_section_memmap(struct page *memmap,
 600                struct vmem_altmap *altmap)
 601{
 602        unsigned long start = (unsigned long)memmap;
 603        unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 604
 605        vmemmap_free(start, end, altmap);
 606}
 607#ifdef CONFIG_MEMORY_HOTREMOVE
 608static void free_map_bootmem(struct page *memmap)
 609{
 610        unsigned long start = (unsigned long)memmap;
 611        unsigned long end = (unsigned long)(memmap + PAGES_PER_SECTION);
 612
 613        vmemmap_free(start, end, NULL);
 614}
 615#endif /* CONFIG_MEMORY_HOTREMOVE */
 616#else
 617static struct page *__kmalloc_section_memmap(void)
 618{
 619        struct page *page, *ret;
 620        unsigned long memmap_size = sizeof(struct page) * PAGES_PER_SECTION;
 621
 622        page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 623        if (page)
 624                goto got_map_page;
 625
 626        ret = vmalloc(memmap_size);
 627        if (ret)
 628                goto got_map_ptr;
 629
 630        return NULL;
 631got_map_page:
 632        ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 633got_map_ptr:
 634
 635        return ret;
 636}
 637
 638static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 639                struct vmem_altmap *altmap)
 640{
 641        return __kmalloc_section_memmap();
 642}
 643
 644static void __kfree_section_memmap(struct page *memmap,
 645                struct vmem_altmap *altmap)
 646{
 647        if (is_vmalloc_addr(memmap))
 648                vfree(memmap);
 649        else
 650                free_pages((unsigned long)memmap,
 651                           get_order(sizeof(struct page) * PAGES_PER_SECTION));
 652}
 653
 654#ifdef CONFIG_MEMORY_HOTREMOVE
 655static void free_map_bootmem(struct page *memmap)
 656{
 657        unsigned long maps_section_nr, removing_section_nr, i;
 658        unsigned long magic, nr_pages;
 659        struct page *page = virt_to_page(memmap);
 660
 661        nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 662                >> PAGE_SHIFT;
 663
 664        for (i = 0; i < nr_pages; i++, page++) {
 665                magic = (unsigned long) page->freelist;
 666
 667                BUG_ON(magic == NODE_INFO);
 668
 669                maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 670                removing_section_nr = page_private(page);
 671
 672                /*
 673                 * When this function is called, the removing section is
 674                 * logical offlined state. This means all pages are isolated
 675                 * from page allocator. If removing section's memmap is placed
 676                 * on the same section, it must not be freed.
 677                 * If it is freed, page allocator may allocate it which will
 678                 * be removed physically soon.
 679                 */
 680                if (maps_section_nr != removing_section_nr)
 681                        put_page_bootmem(page);
 682        }
 683}
 684#endif /* CONFIG_MEMORY_HOTREMOVE */
 685#endif /* CONFIG_SPARSEMEM_VMEMMAP */
 686
 687/**
 688 * sparse_add_one_section - add a memory section
 689 * @nid: The node to add section on
 690 * @start_pfn: start pfn of the memory range
 691 * @altmap: device page map
 692 *
 693 * This is only intended for hotplug.
 694 *
 695 * Return:
 696 * * 0          - On success.
 697 * * -EEXIST    - Section has been present.
 698 * * -ENOMEM    - Out of memory.
 699 */
 700int __meminit sparse_add_one_section(int nid, unsigned long start_pfn,
 701                                     struct vmem_altmap *altmap)
 702{
 703        unsigned long section_nr = pfn_to_section_nr(start_pfn);
 704        struct mem_section *ms;
 705        struct page *memmap;
 706        unsigned long *usemap;
 707        int ret;
 708
 709        /*
 710         * no locking for this, because it does its own
 711         * plus, it does a kmalloc
 712         */
 713        ret = sparse_index_init(section_nr, nid);
 714        if (ret < 0 && ret != -EEXIST)
 715                return ret;
 716        ret = 0;
 717        memmap = kmalloc_section_memmap(section_nr, nid, altmap);
 718        if (!memmap)
 719                return -ENOMEM;
 720        usemap = __kmalloc_section_usemap();
 721        if (!usemap) {
 722                __kfree_section_memmap(memmap, altmap);
 723                return -ENOMEM;
 724        }
 725
 726        ms = __pfn_to_section(start_pfn);
 727        if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 728                ret = -EEXIST;
 729                goto out;
 730        }
 731
 732        /*
 733         * Poison uninitialized struct pages in order to catch invalid flags
 734         * combinations.
 735         */
 736        page_init_poison(memmap, sizeof(struct page) * PAGES_PER_SECTION);
 737
 738        section_mark_present(ms);
 739        sparse_init_one_section(ms, section_nr, memmap, usemap);
 740
 741out:
 742        if (ret < 0) {
 743                kfree(usemap);
 744                __kfree_section_memmap(memmap, altmap);
 745        }
 746        return ret;
 747}
 748
 749#ifdef CONFIG_MEMORY_HOTREMOVE
 750#ifdef CONFIG_MEMORY_FAILURE
 751static void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 752{
 753        int i;
 754
 755        if (!memmap)
 756                return;
 757
 758        /*
 759         * A further optimization is to have per section refcounted
 760         * num_poisoned_pages.  But that would need more space per memmap, so
 761         * for now just do a quick global check to speed up this routine in the
 762         * absence of bad pages.
 763         */
 764        if (atomic_long_read(&num_poisoned_pages) == 0)
 765                return;
 766
 767        for (i = 0; i < nr_pages; i++) {
 768                if (PageHWPoison(&memmap[i])) {
 769                        atomic_long_sub(1, &num_poisoned_pages);
 770                        ClearPageHWPoison(&memmap[i]);
 771                }
 772        }
 773}
 774#else
 775static inline void clear_hwpoisoned_pages(struct page *memmap, int nr_pages)
 776{
 777}
 778#endif
 779
 780static void free_section_usemap(struct page *memmap, unsigned long *usemap,
 781                struct vmem_altmap *altmap)
 782{
 783        struct page *usemap_page;
 784
 785        if (!usemap)
 786                return;
 787
 788        usemap_page = virt_to_page(usemap);
 789        /*
 790         * Check to see if allocation came from hot-plug-add
 791         */
 792        if (PageSlab(usemap_page) || PageCompound(usemap_page)) {
 793                kfree(usemap);
 794                if (memmap)
 795                        __kfree_section_memmap(memmap, altmap);
 796                return;
 797        }
 798
 799        /*
 800         * The usemap came from bootmem. This is packed with other usemaps
 801         * on the section which has pgdat at boot time. Just keep it as is now.
 802         */
 803
 804        if (memmap)
 805                free_map_bootmem(memmap);
 806}
 807
 808void sparse_remove_one_section(struct zone *zone, struct mem_section *ms,
 809                unsigned long map_offset, struct vmem_altmap *altmap)
 810{
 811        struct page *memmap = NULL;
 812        unsigned long *usemap = NULL;
 813
 814        if (ms->section_mem_map) {
 815                usemap = ms->pageblock_flags;
 816                memmap = sparse_decode_mem_map(ms->section_mem_map,
 817                                                __section_nr(ms));
 818                ms->section_mem_map = 0;
 819                ms->pageblock_flags = NULL;
 820        }
 821
 822        clear_hwpoisoned_pages(memmap + map_offset,
 823                        PAGES_PER_SECTION - map_offset);
 824        free_section_usemap(memmap, usemap, altmap);
 825}
 826#endif /* CONFIG_MEMORY_HOTREMOVE */
 827#endif /* CONFIG_MEMORY_HOTPLUG */
 828