mm/sparse.c

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