linux/mm/sparse.c
<<
>>
Prefs
   1/*
   2 * sparse memory mappings.
   3 */
   4#include <linux/mm.h>
   5#include <linux/slab.h>
   6#include <linux/mmzone.h>
   7#include <linux/bootmem.h>
   8#include <linux/highmem.h>
   9#include <linux/export.h>
  10#include <linux/spinlock.h>
  11#include <linux/vmalloc.h>
  12#include "internal.h"
  13#include <asm/dma.h>
  14#include <asm/pgalloc.h>
  15#include <asm/pgtable.h>
  16
  17/*
  18 * Permanent SPARSEMEM data:
  19 *
  20 * 1) mem_section       - memory sections, mem_map's for valid memory
  21 */
  22#ifdef CONFIG_SPARSEMEM_EXTREME
  23struct mem_section *mem_section[NR_SECTION_ROOTS]
  24        ____cacheline_internodealigned_in_smp;
  25#else
  26struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  27        ____cacheline_internodealigned_in_smp;
  28#endif
  29EXPORT_SYMBOL(mem_section);
  30
  31#ifdef NODE_NOT_IN_PAGE_FLAGS
  32/*
  33 * If we did not store the node number in the page then we have to
  34 * do a lookup in the section_to_node_table in order to find which
  35 * node the page belongs to.
  36 */
  37#if MAX_NUMNODES <= 256
  38static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  39#else
  40static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  41#endif
  42
  43int page_to_nid(const struct page *page)
  44{
  45        return section_to_node_table[page_to_section(page)];
  46}
  47EXPORT_SYMBOL(page_to_nid);
  48
  49static void set_section_nid(unsigned long section_nr, int nid)
  50{
  51        section_to_node_table[section_nr] = nid;
  52}
  53#else /* !NODE_NOT_IN_PAGE_FLAGS */
  54static inline void set_section_nid(unsigned long section_nr, int nid)
  55{
  56}
  57#endif
  58
  59#ifdef CONFIG_SPARSEMEM_EXTREME
  60static struct mem_section noinline __init_refok *sparse_index_alloc(int nid)
  61{
  62        struct mem_section *section = NULL;
  63        unsigned long array_size = SECTIONS_PER_ROOT *
  64                                   sizeof(struct mem_section);
  65
  66        if (slab_is_available()) {
  67                if (node_state(nid, N_HIGH_MEMORY))
  68                        section = kmalloc_node(array_size, GFP_KERNEL, nid);
  69                else
  70                        section = kmalloc(array_size, GFP_KERNEL);
  71        } else
  72                section = alloc_bootmem_node(NODE_DATA(nid), array_size);
  73
  74        if (section)
  75                memset(section, 0, array_size);
  76
  77        return section;
  78}
  79
  80static int __meminit sparse_index_init(unsigned long section_nr, int nid)
  81{
  82        static DEFINE_SPINLOCK(index_init_lock);
  83        unsigned long root = SECTION_NR_TO_ROOT(section_nr);
  84        struct mem_section *section;
  85        int ret = 0;
  86
  87        if (mem_section[root])
  88                return -EEXIST;
  89
  90        section = sparse_index_alloc(nid);
  91        if (!section)
  92                return -ENOMEM;
  93        /*
  94         * This lock keeps two different sections from
  95         * reallocating for the same index
  96         */
  97        spin_lock(&index_init_lock);
  98
  99        if (mem_section[root]) {
 100                ret = -EEXIST;
 101                goto out;
 102        }
 103
 104        mem_section[root] = section;
 105out:
 106        spin_unlock(&index_init_lock);
 107        return ret;
 108}
 109#else /* !SPARSEMEM_EXTREME */
 110static inline int sparse_index_init(unsigned long section_nr, int nid)
 111{
 112        return 0;
 113}
 114#endif
 115
 116/*
 117 * Although written for the SPARSEMEM_EXTREME case, this happens
 118 * to also work for the flat array case because
 119 * NR_SECTION_ROOTS==NR_MEM_SECTIONS.
 120 */
 121int __section_nr(struct mem_section* ms)
 122{
 123        unsigned long root_nr;
 124        struct mem_section* root;
 125
 126        for (root_nr = 0; root_nr < NR_SECTION_ROOTS; root_nr++) {
 127                root = __nr_to_section(root_nr * SECTIONS_PER_ROOT);
 128                if (!root)
 129                        continue;
 130
 131                if ((ms >= root) && (ms < (root + SECTIONS_PER_ROOT)))
 132                     break;
 133        }
 134
 135        return (root_nr * SECTIONS_PER_ROOT) + (ms - root);
 136}
 137
 138/*
 139 * During early boot, before section_mem_map is used for an actual
 140 * mem_map, we use section_mem_map to store the section's NUMA
 141 * node.  This keeps us from having to use another data structure.  The
 142 * node information is cleared just before we store the real mem_map.
 143 */
 144static inline unsigned long sparse_encode_early_nid(int nid)
 145{
 146        return (nid << SECTION_NID_SHIFT);
 147}
 148
 149static inline int sparse_early_nid(struct mem_section *section)
 150{
 151        return (section->section_mem_map >> SECTION_NID_SHIFT);
 152}
 153
 154/* Validate the physical addressing limitations of the model */
 155void __meminit mminit_validate_memmodel_limits(unsigned long *start_pfn,
 156                                                unsigned long *end_pfn)
 157{
 158        unsigned long max_sparsemem_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 159
 160        /*
 161         * Sanity checks - do not allow an architecture to pass
 162         * in larger pfns than the maximum scope of sparsemem:
 163         */
 164        if (*start_pfn > max_sparsemem_pfn) {
 165                mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 166                        "Start of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 167                        *start_pfn, *end_pfn, max_sparsemem_pfn);
 168                WARN_ON_ONCE(1);
 169                *start_pfn = max_sparsemem_pfn;
 170                *end_pfn = max_sparsemem_pfn;
 171        } else if (*end_pfn > max_sparsemem_pfn) {
 172                mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 173                        "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 174                        *start_pfn, *end_pfn, max_sparsemem_pfn);
 175                WARN_ON_ONCE(1);
 176                *end_pfn = max_sparsemem_pfn;
 177        }
 178}
 179
 180/* Record a memory area against a node. */
 181void __init memory_present(int nid, unsigned long start, unsigned long end)
 182{
 183        unsigned long pfn;
 184
 185        start &= PAGE_SECTION_MASK;
 186        mminit_validate_memmodel_limits(&start, &end);
 187        for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 188                unsigned long section = pfn_to_section_nr(pfn);
 189                struct mem_section *ms;
 190
 191                sparse_index_init(section, nid);
 192                set_section_nid(section, nid);
 193
 194                ms = __nr_to_section(section);
 195                if (!ms->section_mem_map)
 196                        ms->section_mem_map = sparse_encode_early_nid(nid) |
 197                                                        SECTION_MARKED_PRESENT;
 198        }
 199}
 200
 201/*
 202 * Only used by the i386 NUMA architecures, but relatively
 203 * generic code.
 204 */
 205unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 206                                                     unsigned long end_pfn)
 207{
 208        unsigned long pfn;
 209        unsigned long nr_pages = 0;
 210
 211        mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 212        for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 213                if (nid != early_pfn_to_nid(pfn))
 214                        continue;
 215
 216                if (pfn_present(pfn))
 217                        nr_pages += PAGES_PER_SECTION;
 218        }
 219
 220        return nr_pages * sizeof(struct page);
 221}
 222
 223/*
 224 * Subtle, we encode the real pfn into the mem_map such that
 225 * the identity pfn - section_mem_map will return the actual
 226 * physical page frame number.
 227 */
 228static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 229{
 230        return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 231}
 232
 233/*
 234 * Decode mem_map from the coded memmap
 235 */
 236struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 237{
 238        /* mask off the extra low bits of information */
 239        coded_mem_map &= SECTION_MAP_MASK;
 240        return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 241}
 242
 243static int __meminit sparse_init_one_section(struct mem_section *ms,
 244                unsigned long pnum, struct page *mem_map,
 245                unsigned long *pageblock_bitmap)
 246{
 247        if (!present_section(ms))
 248                return -EINVAL;
 249
 250        ms->section_mem_map &= ~SECTION_MAP_MASK;
 251        ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 252                                                        SECTION_HAS_MEM_MAP;
 253        ms->pageblock_flags = pageblock_bitmap;
 254
 255        return 1;
 256}
 257
 258unsigned long usemap_size(void)
 259{
 260        unsigned long size_bytes;
 261        size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 262        size_bytes = roundup(size_bytes, sizeof(unsigned long));
 263        return size_bytes;
 264}
 265
 266#ifdef CONFIG_MEMORY_HOTPLUG
 267static unsigned long *__kmalloc_section_usemap(void)
 268{
 269        return kmalloc(usemap_size(), GFP_KERNEL);
 270}
 271#endif /* CONFIG_MEMORY_HOTPLUG */
 272
 273#ifdef CONFIG_MEMORY_HOTREMOVE
 274static unsigned long * __init
 275sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 276                                         unsigned long size)
 277{
 278        unsigned long goal, limit;
 279        unsigned long *p;
 280        int nid;
 281        /*
 282         * A page may contain usemaps for other sections preventing the
 283         * page being freed and making a section unremovable while
 284         * other sections referencing the usemap retmain active. Similarly,
 285         * a pgdat can prevent a section being removed. If section A
 286         * contains a pgdat and section B contains the usemap, both
 287         * sections become inter-dependent. This allocates usemaps
 288         * from the same section as the pgdat where possible to avoid
 289         * this problem.
 290         */
 291        goal = __pa(pgdat) & (PAGE_SECTION_MASK << PAGE_SHIFT);
 292        limit = goal + (1UL << PA_SECTION_SHIFT);
 293        nid = early_pfn_to_nid(goal >> PAGE_SHIFT);
 294again:
 295        p = ___alloc_bootmem_node_nopanic(NODE_DATA(nid), size,
 296                                          SMP_CACHE_BYTES, goal, limit);
 297        if (!p && limit) {
 298                limit = 0;
 299                goto again;
 300        }
 301        return p;
 302}
 303
 304static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 305{
 306        unsigned long usemap_snr, pgdat_snr;
 307        static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 308        static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 309        struct pglist_data *pgdat = NODE_DATA(nid);
 310        int usemap_nid;
 311
 312        usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 313        pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 314        if (usemap_snr == pgdat_snr)
 315                return;
 316
 317        if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 318                /* skip redundant message */
 319                return;
 320
 321        old_usemap_snr = usemap_snr;
 322        old_pgdat_snr = pgdat_snr;
 323
 324        usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 325        if (usemap_nid != nid) {
 326                printk(KERN_INFO
 327                       "node %d must be removed before remove section %ld\n",
 328                       nid, usemap_snr);
 329                return;
 330        }
 331        /*
 332         * There is a circular dependency.
 333         * Some platforms allow un-removable section because they will just
 334         * gather other removable sections for dynamic partitioning.
 335         * Just notify un-removable section's number here.
 336         */
 337        printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
 338               pgdat_snr, nid);
 339        printk(KERN_CONT
 340               " have a circular dependency on usemap and pgdat allocations\n");
 341}
 342#else
 343static unsigned long * __init
 344sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 345                                         unsigned long size)
 346{
 347        return alloc_bootmem_node_nopanic(pgdat, size);
 348}
 349
 350static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 351{
 352}
 353#endif /* CONFIG_MEMORY_HOTREMOVE */
 354
 355static void __init sparse_early_usemaps_alloc_node(unsigned long**usemap_map,
 356                                 unsigned long pnum_begin,
 357                                 unsigned long pnum_end,
 358                                 unsigned long usemap_count, int nodeid)
 359{
 360        void *usemap;
 361        unsigned long pnum;
 362        int size = usemap_size();
 363
 364        usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 365                                                          size * usemap_count);
 366        if (!usemap) {
 367                printk(KERN_WARNING "%s: allocation failed\n", __func__);
 368                return;
 369        }
 370
 371        for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 372                if (!present_section_nr(pnum))
 373                        continue;
 374                usemap_map[pnum] = usemap;
 375                usemap += size;
 376                check_usemap_section_nr(nodeid, usemap_map[pnum]);
 377        }
 378}
 379
 380#ifndef CONFIG_SPARSEMEM_VMEMMAP
 381struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 382{
 383        struct page *map;
 384        unsigned long size;
 385
 386        map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 387        if (map)
 388                return map;
 389
 390        size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 391        map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
 392                                         PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 393        return map;
 394}
 395void __init sparse_mem_maps_populate_node(struct page **map_map,
 396                                          unsigned long pnum_begin,
 397                                          unsigned long pnum_end,
 398                                          unsigned long map_count, int nodeid)
 399{
 400        void *map;
 401        unsigned long pnum;
 402        unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 403
 404        map = alloc_remap(nodeid, size * map_count);
 405        if (map) {
 406                for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 407                        if (!present_section_nr(pnum))
 408                                continue;
 409                        map_map[pnum] = map;
 410                        map += size;
 411                }
 412                return;
 413        }
 414
 415        size = PAGE_ALIGN(size);
 416        map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
 417                                         PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 418        if (map) {
 419                for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 420                        if (!present_section_nr(pnum))
 421                                continue;
 422                        map_map[pnum] = map;
 423                        map += size;
 424                }
 425                return;
 426        }
 427
 428        /* fallback */
 429        for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 430                struct mem_section *ms;
 431
 432                if (!present_section_nr(pnum))
 433                        continue;
 434                map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 435                if (map_map[pnum])
 436                        continue;
 437                ms = __nr_to_section(pnum);
 438                printk(KERN_ERR "%s: sparsemem memory map backing failed "
 439                        "some memory will not be available.\n", __func__);
 440                ms->section_mem_map = 0;
 441        }
 442}
 443#endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 444
 445#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 446static void __init sparse_early_mem_maps_alloc_node(struct page **map_map,
 447                                 unsigned long pnum_begin,
 448                                 unsigned long pnum_end,
 449                                 unsigned long map_count, int nodeid)
 450{
 451        sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 452                                         map_count, nodeid);
 453}
 454#else
 455static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 456{
 457        struct page *map;
 458        struct mem_section *ms = __nr_to_section(pnum);
 459        int nid = sparse_early_nid(ms);
 460
 461        map = sparse_mem_map_populate(pnum, nid);
 462        if (map)
 463                return map;
 464
 465        printk(KERN_ERR "%s: sparsemem memory map backing failed "
 466                        "some memory will not be available.\n", __func__);
 467        ms->section_mem_map = 0;
 468        return NULL;
 469}
 470#endif
 471
 472void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 473{
 474}
 475
 476/*
 477 * Allocate the accumulated non-linear sections, allocate a mem_map
 478 * for each and record the physical to section mapping.
 479 */
 480void __init sparse_init(void)
 481{
 482        unsigned long pnum;
 483        struct page *map;
 484        unsigned long *usemap;
 485        unsigned long **usemap_map;
 486        int size;
 487        int nodeid_begin = 0;
 488        unsigned long pnum_begin = 0;
 489        unsigned long usemap_count;
 490#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 491        unsigned long map_count;
 492        int size2;
 493        struct page **map_map;
 494#endif
 495
 496        /*
 497         * map is using big page (aka 2M in x86 64 bit)
 498         * usemap is less one page (aka 24 bytes)
 499         * so alloc 2M (with 2M align) and 24 bytes in turn will
 500         * make next 2M slip to one more 2M later.
 501         * then in big system, the memory will have a lot of holes...
 502         * here try to allocate 2M pages continuously.
 503         *
 504         * powerpc need to call sparse_init_one_section right after each
 505         * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 506         */
 507        size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 508        usemap_map = alloc_bootmem(size);
 509        if (!usemap_map)
 510                panic("can not allocate usemap_map\n");
 511
 512        for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 513                struct mem_section *ms;
 514
 515                if (!present_section_nr(pnum))
 516                        continue;
 517                ms = __nr_to_section(pnum);
 518                nodeid_begin = sparse_early_nid(ms);
 519                pnum_begin = pnum;
 520                break;
 521        }
 522        usemap_count = 1;
 523        for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 524                struct mem_section *ms;
 525                int nodeid;
 526
 527                if (!present_section_nr(pnum))
 528                        continue;
 529                ms = __nr_to_section(pnum);
 530                nodeid = sparse_early_nid(ms);
 531                if (nodeid == nodeid_begin) {
 532                        usemap_count++;
 533                        continue;
 534                }
 535                /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 536                sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, pnum,
 537                                                 usemap_count, nodeid_begin);
 538                /* new start, update count etc*/
 539                nodeid_begin = nodeid;
 540                pnum_begin = pnum;
 541                usemap_count = 1;
 542        }
 543        /* ok, last chunk */
 544        sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS,
 545                                         usemap_count, nodeid_begin);
 546
 547#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 548        size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 549        map_map = alloc_bootmem(size2);
 550        if (!map_map)
 551                panic("can not allocate map_map\n");
 552
 553        for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 554                struct mem_section *ms;
 555
 556                if (!present_section_nr(pnum))
 557                        continue;
 558                ms = __nr_to_section(pnum);
 559                nodeid_begin = sparse_early_nid(ms);
 560                pnum_begin = pnum;
 561                break;
 562        }
 563        map_count = 1;
 564        for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 565                struct mem_section *ms;
 566                int nodeid;
 567
 568                if (!present_section_nr(pnum))
 569                        continue;
 570                ms = __nr_to_section(pnum);
 571                nodeid = sparse_early_nid(ms);
 572                if (nodeid == nodeid_begin) {
 573                        map_count++;
 574                        continue;
 575                }
 576                /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 577                sparse_early_mem_maps_alloc_node(map_map, pnum_begin, pnum,
 578                                                 map_count, nodeid_begin);
 579                /* new start, update count etc*/
 580                nodeid_begin = nodeid;
 581                pnum_begin = pnum;
 582                map_count = 1;
 583        }
 584        /* ok, last chunk */
 585        sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS,
 586                                         map_count, nodeid_begin);
 587#endif
 588
 589        for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 590                if (!present_section_nr(pnum))
 591                        continue;
 592
 593                usemap = usemap_map[pnum];
 594                if (!usemap)
 595                        continue;
 596
 597#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 598                map = map_map[pnum];
 599#else
 600                map = sparse_early_mem_map_alloc(pnum);
 601#endif
 602                if (!map)
 603                        continue;
 604
 605                sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 606                                                                usemap);
 607        }
 608
 609        vmemmap_populate_print_last();
 610
 611#ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 612        free_bootmem(__pa(map_map), size2);
 613#endif
 614        free_bootmem(__pa(usemap_map), size);
 615}
 616
 617#ifdef CONFIG_MEMORY_HOTPLUG
 618#ifdef CONFIG_SPARSEMEM_VMEMMAP
 619static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 620                                                 unsigned long nr_pages)
 621{
 622        /* This will make the necessary allocations eventually. */
 623        return sparse_mem_map_populate(pnum, nid);
 624}
 625static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 626{
 627        return; /* XXX: Not implemented yet */
 628}
 629static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 630{
 631}
 632#else
 633static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 634{
 635        struct page *page, *ret;
 636        unsigned long memmap_size = sizeof(struct page) * nr_pages;
 637
 638        page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 639        if (page)
 640                goto got_map_page;
 641
 642        ret = vmalloc(memmap_size);
 643        if (ret)
 644                goto got_map_ptr;
 645
 646        return NULL;
 647got_map_page:
 648        ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 649got_map_ptr:
 650        memset(ret, 0, memmap_size);
 651
 652        return ret;
 653}
 654
 655static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 656                                                  unsigned long nr_pages)
 657{
 658        return __kmalloc_section_memmap(nr_pages);
 659}
 660
 661static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 662{
 663        if (is_vmalloc_addr(memmap))
 664                vfree(memmap);
 665        else
 666                free_pages((unsigned long)memmap,
 667                           get_order(sizeof(struct page) * nr_pages));
 668}
 669
 670static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 671{
 672        unsigned long maps_section_nr, removing_section_nr, i;
 673        unsigned long magic;
 674
 675        for (i = 0; i < nr_pages; i++, page++) {
 676                magic = (unsigned long) page->lru.next;
 677
 678                BUG_ON(magic == NODE_INFO);
 679
 680                maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 681                removing_section_nr = page->private;
 682
 683                /*
 684                 * When this function is called, the removing section is
 685                 * logical offlined state. This means all pages are isolated
 686                 * from page allocator. If removing section's memmap is placed
 687                 * on the same section, it must not be freed.
 688                 * If it is freed, page allocator may allocate it which will
 689                 * be removed physically soon.
 690                 */
 691                if (maps_section_nr != removing_section_nr)
 692                        put_page_bootmem(page);
 693        }
 694}
 695#endif /* CONFIG_SPARSEMEM_VMEMMAP */
 696
 697static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 698{
 699        struct page *usemap_page;
 700        unsigned long nr_pages;
 701
 702        if (!usemap)
 703                return;
 704
 705        usemap_page = virt_to_page(usemap);
 706        /*
 707         * Check to see if allocation came from hot-plug-add
 708         */
 709        if (PageSlab(usemap_page)) {
 710                kfree(usemap);
 711                if (memmap)
 712                        __kfree_section_memmap(memmap, PAGES_PER_SECTION);
 713                return;
 714        }
 715
 716        /*
 717         * The usemap came from bootmem. This is packed with other usemaps
 718         * on the section which has pgdat at boot time. Just keep it as is now.
 719         */
 720
 721        if (memmap) {
 722                struct page *memmap_page;
 723                memmap_page = virt_to_page(memmap);
 724
 725                nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 726                        >> PAGE_SHIFT;
 727
 728                free_map_bootmem(memmap_page, nr_pages);
 729        }
 730}
 731
 732/*
 733 * returns the number of sections whose mem_maps were properly
 734 * set.  If this is <=0, then that means that the passed-in
 735 * map was not consumed and must be freed.
 736 */
 737int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 738                           int nr_pages)
 739{
 740        unsigned long section_nr = pfn_to_section_nr(start_pfn);
 741        struct pglist_data *pgdat = zone->zone_pgdat;
 742        struct mem_section *ms;
 743        struct page *memmap;
 744        unsigned long *usemap;
 745        unsigned long flags;
 746        int ret;
 747
 748        /*
 749         * no locking for this, because it does its own
 750         * plus, it does a kmalloc
 751         */
 752        ret = sparse_index_init(section_nr, pgdat->node_id);
 753        if (ret < 0 && ret != -EEXIST)
 754                return ret;
 755        memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
 756        if (!memmap)
 757                return -ENOMEM;
 758        usemap = __kmalloc_section_usemap();
 759        if (!usemap) {
 760                __kfree_section_memmap(memmap, nr_pages);
 761                return -ENOMEM;
 762        }
 763
 764        pgdat_resize_lock(pgdat, &flags);
 765
 766        ms = __pfn_to_section(start_pfn);
 767        if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 768                ret = -EEXIST;
 769                goto out;
 770        }
 771
 772        ms->section_mem_map |= SECTION_MARKED_PRESENT;
 773
 774        ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 775
 776out:
 777        pgdat_resize_unlock(pgdat, &flags);
 778        if (ret <= 0) {
 779                kfree(usemap);
 780                __kfree_section_memmap(memmap, nr_pages);
 781        }
 782        return ret;
 783}
 784
 785void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 786{
 787        struct page *memmap = NULL;
 788        unsigned long *usemap = NULL;
 789
 790        if (ms->section_mem_map) {
 791                usemap = ms->pageblock_flags;
 792                memmap = sparse_decode_mem_map(ms->section_mem_map,
 793                                                __section_nr(ms));
 794                ms->section_mem_map = 0;
 795                ms->pageblock_flags = NULL;
 796        }
 797
 798        free_section_usemap(memmap, usemap);
 799}
 800#endif
 801