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         }
 168
 169         if (*end_pfn > max_sparsemem_pfn) {
 170                 mminit_dprintk(MMINIT_WARNING, "pfnvalidation",
 171                         "End of range %lu -> %lu exceeds SPARSEMEM max %lu\n",
 172                         *start_pfn, *end_pfn, max_sparsemem_pfn);
 173                 WARN_ON_ONCE(1);
 174                 *end_pfn = max_sparsemem_pfn;
 175         }
 176 }
 177
 178 /* Record a memory area against a node. */
 179 void __init memory_present(int nid, unsigned long start, unsigned long end)
 180 {
 181         unsigned long pfn;
 182
 183         start &= PAGE_SECTION_MASK;
 184         mminit_validate_memmodel_limits(&start, &end);
 185         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 186                 unsigned long section = pfn_to_section_nr(pfn);
 187                 struct mem_section *ms;
 188
 189                 sparse_index_init(section, nid);
 190                 set_section_nid(section, nid);
 191
 192                 ms = __nr_to_section(section);
 193                 if (!ms->section_mem_map)
 194                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 195                                                         SECTION_MARKED_PRESENT;
 196         }
 197 }
 198
 199 /*
 200  * Only used by the i386 NUMA architecures, but relatively
 201  * generic code.
 202  */
 203 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 204                                                      unsigned long end_pfn)
 205 {
 206         unsigned long pfn;
 207         unsigned long nr_pages = 0;
 208
 209         mminit_validate_memmodel_limits(&start_pfn, &end_pfn);
 210         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 211                 if (nid != early_pfn_to_nid(pfn))
 212                         continue;
 213
 214                 if (pfn_present(pfn))
 215                         nr_pages += PAGES_PER_SECTION;
 216         }
 217
 218         return nr_pages * sizeof(struct page);
 219 }
 220
 221 /*
 222  * Subtle, we encode the real pfn into the mem_map such that
 223  * the identity pfn - section_mem_map will return the actual
 224  * physical page frame number.
 225  */
 226 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 227 {
 228         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 229 }
 230
 231 /*
 232  * Decode mem_map from the coded memmap
 233  */
 234 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 235 {
 236         /* mask off the extra low bits of information */
 237         coded_mem_map &= SECTION_MAP_MASK;
 238         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 239 }
 240
 241 static int __meminit sparse_init_one_section(struct mem_section *ms,
 242                 unsigned long pnum, struct page *mem_map,
 243                 unsigned long *pageblock_bitmap)
 244 {
 245         if (!present_section(ms))
 246                 return -EINVAL;
 247
 248         ms->section_mem_map &= ~SECTION_MAP_MASK;
 249         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 250                                                         SECTION_HAS_MEM_MAP;
 251         ms->pageblock_flags = pageblock_bitmap;
 252
 253         return 1;
 254 }
 255
 256 unsigned long usemap_size(void)
 257 {
 258         unsigned long size_bytes;
 259         size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 260         size_bytes = roundup(size_bytes, sizeof(unsigned long));
 261         return size_bytes;
 262 }
 263
 264 #ifdef CONFIG_MEMORY_HOTPLUG
 265 static unsigned long *__kmalloc_section_usemap(void)
 266 {
 267         return kmalloc(usemap_size(), GFP_KERNEL);
 268 }
 269 #endif /* CONFIG_MEMORY_HOTPLUG */
 270
 271 #ifdef CONFIG_MEMORY_HOTREMOVE
 272 static unsigned long * __init
 273 sparse_early_usemap_alloc_pgdat_section(struct pglist_data *pgdat)
 274 {
 275         unsigned long section_nr;
 276
 277         /*
 278          * A page may contain usemaps for other sections preventing the
 279          * page being freed and making a section unremovable while
 280          * other sections referencing the usemap retmain active. Similarly,
 281          * a pgdat can prevent a section being removed. If section A
 282          * contains a pgdat and section B contains the usemap, both
 283          * sections become inter-dependent. This allocates usemaps
 284          * from the same section as the pgdat where possible to avoid
 285          * this problem.
 286          */
 287         section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 288         return alloc_bootmem_section(usemap_size(), section_nr);
 289 }
 290
 291 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 292 {
 293         unsigned long usemap_snr, pgdat_snr;
 294         static unsigned long old_usemap_snr = NR_MEM_SECTIONS;
 295         static unsigned long old_pgdat_snr = NR_MEM_SECTIONS;
 296         struct pglist_data *pgdat = NODE_DATA(nid);
 297         int usemap_nid;
 298
 299         usemap_snr = pfn_to_section_nr(__pa(usemap) >> PAGE_SHIFT);
 300         pgdat_snr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 301         if (usemap_snr == pgdat_snr)
 302                 return;
 303
 304         if (old_usemap_snr == usemap_snr && old_pgdat_snr == pgdat_snr)
 305                 /* skip redundant message */
 306                 return;
 307
 308         old_usemap_snr = usemap_snr;
 309         old_pgdat_snr = pgdat_snr;
 310
 311         usemap_nid = sparse_early_nid(__nr_to_section(usemap_snr));
 312         if (usemap_nid != nid) {
 313                 printk(KERN_INFO
 314                        "node %d must be removed before remove section %ld\n",
 315                        nid, usemap_snr);
 316                 return;
 317         }
 318         /*
 319          * There is a circular dependency.
 320          * Some platforms allow un-removable section because they will just
 321          * gather other removable sections for dynamic partitioning.
 322          * Just notify un-removable section's number here.
 323          */
 324         printk(KERN_INFO "Section %ld and %ld (node %d)", usemap_snr,
 325                pgdat_snr, nid);
 326         printk(KERN_CONT
 327                " have a circular dependency on usemap and pgdat allocations\n");
 328 }
 329 #else
 330 static unsigned long * __init
 331 sparse_early_usemap_alloc_pgdat_section(struct pglist_data *pgdat)
 332 {
 333         return NULL;
 334 }
 335
 336 static void __init check_usemap_section_nr(int nid, unsigned long *usemap)
 337 {
 338 }
 339 #endif /* CONFIG_MEMORY_HOTREMOVE */
 340
 341 static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum)
 342 {
 343         unsigned long *usemap;
 344         struct mem_section *ms = __nr_to_section(pnum);
 345         int nid = sparse_early_nid(ms);
 346
 347         usemap = sparse_early_usemap_alloc_pgdat_section(NODE_DATA(nid));
 348         if (usemap)
 349                 return usemap;
 350
 351         usemap = alloc_bootmem_node(NODE_DATA(nid), usemap_size());
 352         if (usemap) {
 353                 check_usemap_section_nr(nid, usemap);
 354                 return usemap;
 355         }
 356
 357         /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
 358         nid = 0;
 359
 360         printk(KERN_WARNING "%s: allocation failed\n", __func__);
 361         return NULL;
 362 }
 363
 364 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 365 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 366 {
 367         struct page *map;
 368
 369         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 370         if (map)
 371                 return map;
 372
 373         map = alloc_bootmem_pages_node(NODE_DATA(nid),
 374                        PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION));
 375         return map;
 376 }
 377 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 378
 379 static struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 380 {
 381         struct page *map;
 382         struct mem_section *ms = __nr_to_section(pnum);
 383         int nid = sparse_early_nid(ms);
 384
 385         map = sparse_mem_map_populate(pnum, nid);
 386         if (map)
 387                 return map;
 388
 389         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 390                         "some memory will not be available.\n", __func__);
 391         ms->section_mem_map = 0;
 392         return NULL;
 393 }
 394
 395 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 396 {
 397 }
 398 /*
 399  * Allocate the accumulated non-linear sections, allocate a mem_map
 400  * for each and record the physical to section mapping.
 401  */
 402 void __init sparse_init(void)
 403 {
 404         unsigned long pnum;
 405         struct page *map;
 406         unsigned long *usemap;
 407         unsigned long **usemap_map;
 408         int size;
 409
 410         /*
 411          * map is using big page (aka 2M in x86 64 bit)
 412          * usemap is less one page (aka 24 bytes)
 413          * so alloc 2M (with 2M align) and 24 bytes in turn will
 414          * make next 2M slip to one more 2M later.
 415          * then in big system, the memory will have a lot of holes...
 416          * here try to allocate 2M pages continously.
 417          *
 418          * powerpc need to call sparse_init_one_section right after each
 419          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 420          */
 421         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 422         usemap_map = alloc_bootmem(size);
 423         if (!usemap_map)
 424                 panic("can not allocate usemap_map\n");
 425
 426         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 427                 if (!present_section_nr(pnum))
 428                         continue;
 429                 usemap_map[pnum] = sparse_early_usemap_alloc(pnum);
 430         }
 431
 432         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 433                 if (!present_section_nr(pnum))
 434                         continue;
 435
 436                 usemap = usemap_map[pnum];
 437                 if (!usemap)
 438                         continue;
 439
 440                 map = sparse_early_mem_map_alloc(pnum);
 441                 if (!map)
 442                         continue;
 443
 444                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 445                                                                 usemap);
 446         }
 447
 448         vmemmap_populate_print_last();
 449
 450         free_bootmem(__pa(usemap_map), size);
 451 }
 452
 453 #ifdef CONFIG_MEMORY_HOTPLUG
 454 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 455 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 456                                                  unsigned long nr_pages)
 457 {
 458         /* This will make the necessary allocations eventually. */
 459         return sparse_mem_map_populate(pnum, nid);
 460 }
 461 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 462 {
 463         return; /* XXX: Not implemented yet */
 464 }
 465 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 466 {
 467 }
 468 #else
 469 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 470 {
 471         struct page *page, *ret;
 472         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 473
 474         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 475         if (page)
 476                 goto got_map_page;
 477
 478         ret = vmalloc(memmap_size);
 479         if (ret)
 480                 goto got_map_ptr;
 481
 482         return NULL;
 483 got_map_page:
 484         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 485 got_map_ptr:
 486         memset(ret, 0, memmap_size);
 487
 488         return ret;
 489 }
 490
 491 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 492                                                   unsigned long nr_pages)
 493 {
 494         return __kmalloc_section_memmap(nr_pages);
 495 }
 496
 497 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 498 {
 499         if (is_vmalloc_addr(memmap))
 500                 vfree(memmap);
 501         else
 502                 free_pages((unsigned long)memmap,
 503                            get_order(sizeof(struct page) * nr_pages));
 504 }
 505
 506 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 507 {
 508         unsigned long maps_section_nr, removing_section_nr, i;
 509         int magic;
 510
 511         for (i = 0; i < nr_pages; i++, page++) {
 512                 magic = atomic_read(&page->_mapcount);
 513
 514                 BUG_ON(magic == NODE_INFO);
 515
 516                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 517                 removing_section_nr = page->private;
 518
 519                 /*
 520                  * When this function is called, the removing section is
 521                  * logical offlined state. This means all pages are isolated
 522                  * from page allocator. If removing section's memmap is placed
 523                  * on the same section, it must not be freed.
 524                  * If it is freed, page allocator may allocate it which will
 525                  * be removed physically soon.
 526                  */
 527                 if (maps_section_nr != removing_section_nr)
 528                         put_page_bootmem(page);
 529         }
 530 }
 531 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 532
 533 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 534 {
 535         struct page *usemap_page;
 536         unsigned long nr_pages;
 537
 538         if (!usemap)
 539                 return;
 540
 541         usemap_page = virt_to_page(usemap);
 542         /*
 543          * Check to see if allocation came from hot-plug-add
 544          */
 545         if (PageSlab(usemap_page)) {
 546                 kfree(usemap);
 547                 if (memmap)
 548                         __kfree_section_memmap(memmap, PAGES_PER_SECTION);
 549                 return;
 550         }
 551
 552         /*
 553          * The usemap came from bootmem. This is packed with other usemaps
 554          * on the section which has pgdat at boot time. Just keep it as is now.
 555          */
 556
 557         if (memmap) {
 558                 struct page *memmap_page;
 559                 memmap_page = virt_to_page(memmap);
 560
 561                 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 562                         >> PAGE_SHIFT;
 563
 564                 free_map_bootmem(memmap_page, nr_pages);
 565         }
 566 }
 567
 568 /*
 569  * returns the number of sections whose mem_maps were properly
 570  * set.  If this is <=0, then that means that the passed-in
 571  * map was not consumed and must be freed.
 572  */
 573 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 574                            int nr_pages)
 575 {
 576         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 577         struct pglist_data *pgdat = zone->zone_pgdat;
 578         struct mem_section *ms;
 579         struct page *memmap;
 580         unsigned long *usemap;
 581         unsigned long flags;
 582         int ret;
 583
 584         /*
 585          * no locking for this, because it does its own
 586          * plus, it does a kmalloc
 587          */
 588         ret = sparse_index_init(section_nr, pgdat->node_id);
 589         if (ret < 0 && ret != -EEXIST)
 590                 return ret;
 591         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
 592         if (!memmap)
 593                 return -ENOMEM;
 594         usemap = __kmalloc_section_usemap();
 595         if (!usemap) {
 596                 __kfree_section_memmap(memmap, nr_pages);
 597                 return -ENOMEM;
 598         }
 599
 600         pgdat_resize_lock(pgdat, &flags);
 601
 602         ms = __pfn_to_section(start_pfn);
 603         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 604                 ret = -EEXIST;
 605                 goto out;
 606         }
 607
 608         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 609
 610         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 611
 612 out:
 613         pgdat_resize_unlock(pgdat, &flags);
 614         if (ret <= 0) {
 615                 kfree(usemap);
 616                 __kfree_section_memmap(memmap, nr_pages);
 617         }
 618         return ret;
 619 }
 620
 621 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 622 {
 623         struct page *memmap = NULL;
 624         unsigned long *usemap = NULL;
 625
 626         if (ms->section_mem_map) {
 627                 usemap = ms->pageblock_flags;
 628                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 629                                                 __section_nr(ms));
 630                 ms->section_mem_map = 0;
 631                 ms->pageblock_flags = NULL;
 632         }
 633
 634         free_section_usemap(memmap, usemap);
 635 }
 636 #endif