mm/sparse.c

   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/module.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
  23 struct mem_section *mem_section[NR_SECTION_ROOTS]
  24         ____cacheline_internodealigned_in_smp;
  25 #else
  26 struct mem_section mem_section[NR_SECTION_ROOTS][SECTIONS_PER_ROOT]
  27         ____cacheline_internodealigned_in_smp;
  28 #endif
  29 EXPORT_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
  38 static u8 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  39 #else
  40 static u16 section_to_node_table[NR_MEM_SECTIONS] __cacheline_aligned;
  41 #endif
  42
  43 int page_to_nid(struct page *page)
  44 {
  45         return section_to_node_table[page_to_section(page)];
  46 }
  47 EXPORT_SYMBOL(page_to_nid);
  48
  49 static 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 */
  54 static inline void set_section_nid(unsigned long section_nr, int nid)
  55 {
  56 }
  57 #endif
  58
  59 #ifdef CONFIG_SPARSEMEM_EXTREME
  60 static 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
  80 static 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;
 105 out:
 106         spin_unlock(&index_init_lock);
 107         return ret;
 108 }
 109 #else /* !SPARSEMEM_EXTREME */
 110 static 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  */
 121 int __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  */
 144 static inline unsigned long sparse_encode_early_nid(int nid)
 145 {
 146         return (nid << SECTION_NID_SHIFT);
 147 }
 148
 149 static 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 */
 155 void __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. */
 181 void __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  */
 205 unsigned 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  */
 228 static 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  */
 236 struct 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
 243 static 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
 258 unsigned 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
 267 static 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
 274 static unsigned long * __init
 275 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 276                                          unsigned long count)
 277 {
 278         unsigned long section_nr;
 279
 280         /*
 281          * A page may contain usemaps for other sections preventing the
 282          * page being freed and making a section unremovable while
 283          * other sections referencing the usemap retmain active. Similarly,
 284          * a pgdat can prevent a section being removed. If section A
 285          * contains a pgdat and section B contains the usemap, both
 286          * sections become inter-dependent. This allocates usemaps
 287          * from the same section as the pgdat where possible to avoid
 288          * this problem.
 289          */
 290         section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 291         return alloc_bootmem_section(usemap_size() * count, section_nr);
 292 }
 293
 294 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 295 {
 296         unsigned long usemap_snr, pgdat_snr;
 297         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 298         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 299         struct pglist_data *pgdat = NODE_DATA(nid);
 300         int usemap_nid;
 301
 302         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 303         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 304         if (usemap_snr == pgdat_snr)
 305                 return;
 306
 307         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 308                 /* skip redundant message */
 309                 return;
 310
 311         old_usemap_snr = usemap_snr;
 312         old_pgdat_snr = pgdat_snr;
 313
 314         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 315         if (usemap_nid != nid) {
 316                 printk(KERN_INFO
 317                        "node %d must be removed before remove section %ld\n",
 318                        nid, usemap_snr);
 319                 return;
 320         }
 321         /*
 322          * There is a circular dependency.
 323          * Some platforms allow un-removable section because they will just
 324          * gather other removable sections for dynamic partitioning.
 325          * Just notify un-removable section's number here.
 326          */
 327         printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
 328                pgdat_snr, nid);
 329         printk(KERN_CONT
 330                " have a circular dependency on usemap and pgdat allocations\n");
 331 }
 332 #else
 333 static unsigned long * __init
 334 sparse_early_usemaps_alloc_pgdat_section(struct pglist_data *pgdat,
 335                                          unsigned long count)
 336 {
 337         return NULL;
 338 }
 339
 340 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 341 {
 342 }
 343 #endif /* CONFIG_MEMORY_HOTREMOVE */
 344
 345 static void __init sparse_early_usemaps_alloc_node(unsigned long**usemap_map,
 346                                  unsigned long pnum_begin,
 347                                  unsigned long pnum_end,
 348                                  unsigned long usemap_count, int nodeid)
 349 {
 350         void *usemap;
 351         unsigned long pnum;
 352         int size = usemap_size();
 353
 354         usemap = sparse_early_usemaps_alloc_pgdat_section(NODE_DATA(nodeid),
 355                                                                  usemap_count);
 356         if (usemap) {
 357                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 358                         if (!present_section_nr(pnum))
 359                                 continue;
 360                         usemap_map[pnum] = usemap;
 361                         usemap += size;
 362                 }
 363                 return;
 364         }
 365
 366         usemap = alloc_bootmem_node(NODE_DATA(nodeid), size * usemap_count);
 367         if (usemap) {
 368                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 369                         if (!present_section_nr(pnum))
 370                                 continue;
 371                         usemap_map[pnum] = usemap;
 372                         usemap += size;
 373                         check_usemap_section_nr(nodeid, usemap_map[pnum]);
 374                 }
 375                 return;
 376         }
 377
 378         printk(KERN_WARNING "%s: allocation failed\n", __func__);
 379 }
 380
 381 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 382 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 383 {
 384         struct page *map;
 385         unsigned long size;
 386
 387         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 388         if (map)
 389                 return map;
 390
 391         size = PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION);
 392         map = __alloc_bootmem_node_high(NODE_DATA(nid), size,
 393                                          PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 394         return map;
 395 }
 396 void __init sparse_mem_maps_populate_node(struct page **map_map,
 397                                           unsigned long pnum_begin,
 398                                           unsigned long pnum_end,
 399                                           unsigned long map_count, int nodeid)
 400 {
 401         void *map;
 402         unsigned long pnum;
 403         unsigned long size = sizeof(struct page) * PAGES_PER_SECTION;
 404
 405         map = alloc_remap(nodeid, size * map_count);
 406         if (map) {
 407                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 408                         if (!present_section_nr(pnum))
 409                                 continue;
 410                         map_map[pnum] = map;
 411                         map += size;
 412                 }
 413                 return;
 414         }
 415
 416         size = PAGE_ALIGN(size);
 417         map = __alloc_bootmem_node_high(NODE_DATA(nodeid), size * map_count,
 418                                          PAGE_SIZE, __pa(MAX_DMA_ADDRESS));
 419         if (map) {
 420                 for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 421                         if (!present_section_nr(pnum))
 422                                 continue;
 423                         map_map[pnum] = map;
 424                         map += size;
 425                 }
 426                 return;
 427         }
 428
 429         /* fallback */
 430         for (pnum = pnum_begin; pnum < pnum_end; pnum++) {
 431                 struct mem_section *ms;
 432
 433                 if (!present_section_nr(pnum))
 434                         continue;
 435                 map_map[pnum] = sparse_mem_map_populate(pnum, nodeid);
 436                 if (map_map[pnum])
 437                         continue;
 438                 ms = __nr_to_section(pnum);
 439                 printk(KERN_ERR "%s: sparsemem memory map backing failed "
 440                         "some memory will not be available.\n", __func__);
 441                 ms->section_mem_map = 0;
 442         }
 443 }
 444 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 445
 446 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 447 static void __init sparse_early_mem_maps_alloc_node(struct page **map_map,
 448                                  unsigned long pnum_begin,
 449                                  unsigned long pnum_end,
 450                                  unsigned long map_count, int nodeid)
 451 {
 452         sparse_mem_maps_populate_node(map_map, pnum_begin, pnum_end,
 453                                          map_count, nodeid);
 454 }
 455 #else
 456 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 457 {
 458         struct page *map;
 459         struct mem_section *ms = __nr_to_section(pnum);
 460         int nid = sparse_early_nid(ms);
 461
 462         map = sparse_mem_map_populate(pnum, nid);
 463         if (map)
 464                 return map;
 465
 466         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 467                         "some memory will not be available.\n", __func__);
 468         ms->section_mem_map = 0;
 469         return NULL;
 470 }
 471 #endif
 472
 473 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 474 {
 475 }
 476
 477 /*
 478  * Allocate the accumulated non-linear sections, allocate a mem_map
 479  * for each and record the physical to section mapping.
 480  */
 481 void __init sparse_init(void)
 482 {
 483         unsigned long pnum;
 484         struct page *map;
 485         unsigned long *usemap;
 486         unsigned long **usemap_map;
 487         int size;
 488         int nodeid_begin = 0;
 489         unsigned long pnum_begin = 0;
 490         unsigned long usemap_count;
 491 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 492         unsigned long map_count;
 493         int size2;
 494         struct page **map_map;
 495 #endif
 496
 497         /*
 498          * map is using big page (aka 2M in x86 64 bit)
 499          * usemap is less one page (aka 24 bytes)
 500          * so alloc 2M (with 2M align) and 24 bytes in turn will
 501          * make next 2M slip to one more 2M later.
 502          * then in big system, the memory will have a lot of holes...
 503          * here try to allocate 2M pages continously.
 504          *
 505          * powerpc need to call sparse_init_one_section right after each
 506          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 507          */
 508         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 509         usemap_map = alloc_bootmem(size);
 510         if (!usemap_map)
 511                 panic("can not allocate usemap_map\n");
 512
 513         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 514                 struct mem_section *ms;
 515
 516                 if (!present_section_nr(pnum))
 517                         continue;
 518                 ms = __nr_to_section(pnum);
 519                 nodeid_begin = sparse_early_nid(ms);
 520                 pnum_begin = pnum;
 521                 break;
 522         }
 523         usemap_count = 1;
 524         for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 525                 struct mem_section *ms;
 526                 int nodeid;
 527
 528                 if (!present_section_nr(pnum))
 529                         continue;
 530                 ms = __nr_to_section(pnum);
 531                 nodeid = sparse_early_nid(ms);
 532                 if (nodeid == nodeid_begin) {
 533                         usemap_count++;
 534                         continue;
 535                 }
 536                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 537                 sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, pnum,
 538                                                  usemap_count, nodeid_begin);
 539                 /* new start, update count etc*/
 540                 nodeid_begin = nodeid;
 541                 pnum_begin = pnum;
 542                 usemap_count = 1;
 543         }
 544         /* ok, last chunk */
 545         sparse_early_usemaps_alloc_node(usemap_map, pnum_begin, NR_MEM_SECTIONS,
 546                                          usemap_count, nodeid_begin);
 547
 548 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 549         size2 = sizeof(struct page *) * NR_MEM_SECTIONS;
 550         map_map = alloc_bootmem(size2);
 551         if (!map_map)
 552                 panic("can not allocate map_map\n");
 553
 554         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 555                 struct mem_section *ms;
 556
 557                 if (!present_section_nr(pnum))
 558                         continue;
 559                 ms = __nr_to_section(pnum);
 560                 nodeid_begin = sparse_early_nid(ms);
 561                 pnum_begin = pnum;
 562                 break;
 563         }
 564         map_count = 1;
 565         for (pnum = pnum_begin + 1; pnum < NR_MEM_SECTIONS; pnum++) {
 566                 struct mem_section *ms;
 567                 int nodeid;
 568
 569                 if (!present_section_nr(pnum))
 570                         continue;
 571                 ms = __nr_to_section(pnum);
 572                 nodeid = sparse_early_nid(ms);
 573                 if (nodeid == nodeid_begin) {
 574                         map_count++;
 575                         continue;
 576                 }
 577                 /* ok, we need to take cake of from pnum_begin to pnum - 1*/
 578                 sparse_early_mem_maps_alloc_node(map_map, pnum_begin, pnum,
 579                                                  map_count, nodeid_begin);
 580                 /* new start, update count etc*/
 581                 nodeid_begin = nodeid;
 582                 pnum_begin = pnum;
 583                 map_count = 1;
 584         }
 585         /* ok, last chunk */
 586         sparse_early_mem_maps_alloc_node(map_map, pnum_begin, NR_MEM_SECTIONS,
 587                                          map_count, nodeid_begin);
 588 #endif
 589
 590         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 591                 if (!present_section_nr(pnum))
 592                         continue;
 593
 594                 usemap = usemap_map[pnum];
 595                 if (!usemap)
 596                         continue;
 597
 598 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 599                 map = map_map[pnum];
 600 #else
 601                 map = sparse_early_mem_map_alloc(pnum);
 602 #endif
 603                 if (!map)
 604                         continue;
 605
 606                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 607                                                                 usemap);
 608         }
 609
 610         vmemmap_populate_print_last();
 611
 612 #ifdef CONFIG_SPARSEMEM_ALLOC_MEM_MAP_TOGETHER
 613         free_bootmem(__pa(map_map), size2);
 614 #endif
 615         free_bootmem(__pa(usemap_map), size);
 616 }
 617
 618 #ifdef CONFIG_MEMORY_HOTPLUG
 619 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 620 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 621                                                  unsigned long nr_pages)
 622 {
 623         /* This will make the necessary allocations eventually. */
 624         return sparse_mem_map_populate(pnum, nid);
 625 }
 626 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 627 {
 628         return; /* XXX: Not implemented yet */
 629 }
 630 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 631 {
 632 }
 633 #else
 634 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 635 {
 636         struct page *page, *ret;
 637         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 638
 639         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 640         if (page)
 641                 goto got_map_page;
 642
 643         ret = vmalloc(memmap_size);
 644         if (ret)
 645                 goto got_map_ptr;
 646
 647         return NULL;
 648 got_map_page:
 649         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 650 got_map_ptr:
 651         memset(ret, 0, memmap_size);
 652
 653         return ret;
 654 }
 655
 656 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 657                                                   unsigned long nr_pages)
 658 {
 659         return __kmalloc_section_memmap(nr_pages);
 660 }
 661
 662 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 663 {
 664         if (is_vmalloc_addr(memmap))
 665                 vfree(memmap);
 666         else
 667                 free_pages((unsigned long)memmap,
 668                            get_order(sizeof(struct page) * nr_pages));
 669 }
 670
 671 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 672 {
 673         unsigned long maps_section_nr, removing_section_nr, i;
 674         unsigned long magic;
 675
 676         for (i = 0; i < nr_pages; i++, page++) {
 677                 magic = (unsigned long) page->lru.next;
 678
 679                 BUG_ON(magic == NODE_INFO);
 680
 681                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 682                 removing_section_nr = page->private;
 683
 684                 /*
 685                  * When this function is called, the removing section is
 686                  * logical offlined state. This means all pages are isolated
 687                  * from page allocator. If removing section's memmap is placed
 688                  * on the same section, it must not be freed.
 689                  * If it is freed, page allocator may allocate it which will
 690                  * be removed physically soon.
 691                  */
 692                 if (maps_section_nr != removing_section_nr)
 693                         put_page_bootmem(page);
 694         }
 695 }
 696 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 697
 698 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 699 {
 700         struct page *usemap_page;
 701         unsigned long nr_pages;
 702
 703         if (!usemap)
 704                 return;
 705
 706         usemap_page = virt_to_page(usemap);
 707         /*
 708          * Check to see if allocation came from hot-plug-add
 709          */
 710         if (PageSlab(usemap_page)) {
 711                 kfree(usemap);
 712                 if (memmap)
 713                         __kfree_section_memmap(memmap, PAGES_PER_SECTION);
 714                 return;
 715         }
 716
 717         /*
 718          * The usemap came from bootmem. This is packed with other usemaps
 719          * on the section which has pgdat at boot time. Just keep it as is now.
 720          */
 721
 722         if (memmap) {
 723                 struct page *memmap_page;
 724                 memmap_page = virt_to_page(memmap);
 725
 726                 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 727                         >> PAGE_SHIFT;
 728
 729                 free_map_bootmem(memmap_page, nr_pages);
 730         }
 731 }
 732
 733 /*
 734  * returns the number of sections whose mem_maps were properly
 735  * set.  If this is <=0, then that means that the passed-in
 736  * map was not consumed and must be freed.
 737  */
 738 int __meminit sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 739                            int nr_pages)
 740 {
 741         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 742         struct pglist_data *pgdat = zone->zone_pgdat;
 743         struct mem_section *ms;
 744         struct page *memmap;
 745         unsigned long *usemap;
 746         unsigned long flags;
 747         int ret;
 748
 749         /*
 750          * no locking for this, because it does its own
 751          * plus, it does a kmalloc
 752          */
 753         ret = sparse_index_init(section_nr, pgdat->node_id);
 754         if (ret < 0 && ret != -EEXIST)
 755                 return ret;
 756         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
 757         if (!memmap)
 758                 return -ENOMEM;
 759         usemap = __kmalloc_section_usemap();
 760         if (!usemap) {
 761                 __kfree_section_memmap(memmap, nr_pages);
 762                 return -ENOMEM;
 763         }
 764
 765         pgdat_resize_lock(pgdat, &flags);
 766
 767         ms = __pfn_to_section(start_pfn);
 768         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 769                 ret = -EEXIST;
 770                 goto out;
 771         }
 772
 773         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 774
 775         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 776
 777 out:
 778         pgdat_resize_unlock(pgdat, &flags);
 779         if (ret <= 0) {
 780                 kfree(usemap);
 781                 __kfree_section_memmap(memmap, nr_pages);
 782         }
 783         return ret;
 784 }
 785
 786 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 787 {
 788         struct page *memmap = NULL;
 789         unsigned long *usemap = NULL;
 790
 791         if (ms->section_mem_map) {
 792                 usemap = ms->pageblock_flags;
 793                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 794                                                 __section_nr(ms));
 795                 ms->section_mem_map = 0;
 796                 ms->pageblock_flags = NULL;
 797         }
 798
 799         free_section_usemap(memmap, usemap);
 800 }
 801 #endif