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 /* Record a memory area against a node. */
 151 void __init memory_present(int nid, unsigned long start, unsigned long end)
 152 {
 153         unsigned long max_arch_pfn = 1UL << (MAX_PHYSMEM_BITS-PAGE_SHIFT);
 154         unsigned long pfn;
 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 >= max_arch_pfn)
 161                 return;
 162         if (end >= max_arch_pfn)
 163                 end = max_arch_pfn;
 164
 165         start &= PAGE_SECTION_MASK;
 166         for (pfn = start; pfn < end; pfn += PAGES_PER_SECTION) {
 167                 unsigned long section = pfn_to_section_nr(pfn);
 168                 struct mem_section *ms;
 169
 170                 sparse_index_init(section, nid);
 171                 set_section_nid(section, nid);
 172
 173                 ms = __nr_to_section(section);
 174                 if (!ms->section_mem_map)
 175                         ms->section_mem_map = sparse_encode_early_nid(nid) |
 176                                                         SECTION_MARKED_PRESENT;
 177         }
 178 }
 179
 180 /*
 181  * Only used by the i386 NUMA architecures, but relatively
 182  * generic code.
 183  */
 184 unsigned long __init node_memmap_size_bytes(int nid, unsigned long start_pfn,
 185                                                      unsigned long end_pfn)
 186 {
 187         unsigned long pfn;
 188         unsigned long nr_pages = 0;
 189
 190         for (pfn = start_pfn; pfn < end_pfn; pfn += PAGES_PER_SECTION) {
 191                 if (nid != early_pfn_to_nid(pfn))
 192                         continue;
 193
 194                 if (pfn_present(pfn))
 195                         nr_pages += PAGES_PER_SECTION;
 196         }
 197
 198         return nr_pages * sizeof(struct page);
 199 }
 200
 201 /*
 202  * Subtle, we encode the real pfn into the mem_map such that
 203  * the identity pfn - section_mem_map will return the actual
 204  * physical page frame number.
 205  */
 206 static unsigned long sparse_encode_mem_map(struct page *mem_map, unsigned long pnum)
 207 {
 208         return (unsigned long)(mem_map - (section_nr_to_pfn(pnum)));
 209 }
 210
 211 /*
 212  * Decode mem_map from the coded memmap
 213  */
 214 struct page *sparse_decode_mem_map(unsigned long coded_mem_map, unsigned long pnum)
 215 {
 216         /* mask off the extra low bits of information */
 217         coded_mem_map &= SECTION_MAP_MASK;
 218         return ((struct page *)coded_mem_map) + section_nr_to_pfn(pnum);
 219 }
 220
 221 static int __meminit sparse_init_one_section(struct mem_section *ms,
 222                 unsigned long pnum, struct page *mem_map,
 223                 unsigned long *pageblock_bitmap)
 224 {
 225         if (!present_section(ms))
 226                 return -EINVAL;
 227
 228         ms->section_mem_map &= ~SECTION_MAP_MASK;
 229         ms->section_mem_map |= sparse_encode_mem_map(mem_map, pnum) |
 230                                                         SECTION_HAS_MEM_MAP;
 231         ms->pageblock_flags = pageblock_bitmap;
 232
 233         return 1;
 234 }
 235
 236 unsigned long usemap_size(void)
 237 {
 238         unsigned long size_bytes;
 239         size_bytes = roundup(SECTION_BLOCKFLAGS_BITS, 8) / 8;
 240         size_bytes = roundup(size_bytes, sizeof(unsigned long));
 241         return size_bytes;
 242 }
 243
 244 #ifdef CONFIG_MEMORY_HOTPLUG
 245 static unsigned long *__kmalloc_section_usemap(void)
 246 {
 247         return kmalloc(usemap_size(), GFP_KERNEL);
 248 }
 249 #endif /* CONFIG_MEMORY_HOTPLUG */
 250
 251 static unsigned long *__init sparse_early_usemap_alloc(unsigned long pnum)
 252 {
 253         unsigned long *usemap, section_nr;
 254         struct mem_section *ms = __nr_to_section(pnum);
 255         int nid = sparse_early_nid(ms);
 256         struct pglist_data *pgdat = NODE_DATA(nid);
 257
 258         /*
 259          * Usemap's page can't be freed until freeing other sections
 260          * which use it. And, Pgdat has same feature.
 261          * If section A has pgdat and section B has usemap for other
 262          * sections (includes section A), both sections can't be removed,
 263          * because there is the dependency each other.
 264          * To solve above issue, this collects all usemap on the same section
 265          * which has pgdat.
 266          */
 267         section_nr = pfn_to_section_nr(__pa(pgdat) >> PAGE_SHIFT);
 268         usemap = alloc_bootmem_section(usemap_size(), section_nr);
 269         if (usemap)
 270                 return usemap;
 271
 272         /* Stupid: suppress gcc warning for SPARSEMEM && !NUMA */
 273         nid = 0;
 274
 275         printk(KERN_WARNING "%s: allocation failed\n", __FUNCTION__);
 276         return NULL;
 277 }
 278
 279 #ifndef CONFIG_SPARSEMEM_VMEMMAP
 280 struct page __init *sparse_mem_map_populate(unsigned long pnum, int nid)
 281 {
 282         struct page *map;
 283
 284         map = alloc_remap(nid, sizeof(struct page) * PAGES_PER_SECTION);
 285         if (map)
 286                 return map;
 287
 288         map = alloc_bootmem_pages_node(NODE_DATA(nid),
 289                        PAGE_ALIGN(sizeof(struct page) * PAGES_PER_SECTION));
 290         return map;
 291 }
 292 #endif /* !CONFIG_SPARSEMEM_VMEMMAP */
 293
 294 struct page __init *sparse_early_mem_map_alloc(unsigned long pnum)
 295 {
 296         struct page *map;
 297         struct mem_section *ms = __nr_to_section(pnum);
 298         int nid = sparse_early_nid(ms);
 299
 300         map = sparse_mem_map_populate(pnum, nid);
 301         if (map)
 302                 return map;
 303
 304         printk(KERN_ERR "%s: sparsemem memory map backing failed "
 305                         "some memory will not be available.\n", __FUNCTION__);
 306         ms->section_mem_map = 0;
 307         return NULL;
 308 }
 309
 310 void __attribute__((weak)) __meminit vmemmap_populate_print_last(void)
 311 {
 312 }
 313 /*
 314  * Allocate the accumulated non-linear sections, allocate a mem_map
 315  * for each and record the physical to section mapping.
 316  */
 317 void __init sparse_init(void)
 318 {
 319         unsigned long pnum;
 320         struct page *map;
 321         unsigned long *usemap;
 322         unsigned long **usemap_map;
 323         int size;
 324
 325         /*
 326          * map is using big page (aka 2M in x86 64 bit)
 327          * usemap is less one page (aka 24 bytes)
 328          * so alloc 2M (with 2M align) and 24 bytes in turn will
 329          * make next 2M slip to one more 2M later.
 330          * then in big system, the memory will have a lot of holes...
 331          * here try to allocate 2M pages continously.
 332          *
 333          * powerpc need to call sparse_init_one_section right after each
 334          * sparse_early_mem_map_alloc, so allocate usemap_map at first.
 335          */
 336         size = sizeof(unsigned long *) * NR_MEM_SECTIONS;
 337         usemap_map = alloc_bootmem(size);
 338         if (!usemap_map)
 339                 panic("can not allocate usemap_map\n");
 340
 341         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 342                 if (!present_section_nr(pnum))
 343                         continue;
 344                 usemap_map[pnum] = sparse_early_usemap_alloc(pnum);
 345         }
 346
 347         for (pnum = 0; pnum < NR_MEM_SECTIONS; pnum++) {
 348                 if (!present_section_nr(pnum))
 349                         continue;
 350
 351                 usemap = usemap_map[pnum];
 352                 if (!usemap)
 353                         continue;
 354
 355                 map = sparse_early_mem_map_alloc(pnum);
 356                 if (!map)
 357                         continue;
 358
 359                 sparse_init_one_section(__nr_to_section(pnum), pnum, map,
 360                                                                 usemap);
 361         }
 362
 363         vmemmap_populate_print_last();
 364
 365         free_bootmem(__pa(usemap_map), size);
 366 }
 367
 368 #ifdef CONFIG_MEMORY_HOTPLUG
 369 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 370 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 371                                                  unsigned long nr_pages)
 372 {
 373         /* This will make the necessary allocations eventually. */
 374         return sparse_mem_map_populate(pnum, nid);
 375 }
 376 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 377 {
 378         return; /* XXX: Not implemented yet */
 379 }
 380 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 381 {
 382 }
 383 #else
 384 static struct page *__kmalloc_section_memmap(unsigned long nr_pages)
 385 {
 386         struct page *page, *ret;
 387         unsigned long memmap_size = sizeof(struct page) * nr_pages;
 388
 389         page = alloc_pages(GFP_KERNEL|__GFP_NOWARN, get_order(memmap_size));
 390         if (page)
 391                 goto got_map_page;
 392
 393         ret = vmalloc(memmap_size);
 394         if (ret)
 395                 goto got_map_ptr;
 396
 397         return NULL;
 398 got_map_page:
 399         ret = (struct page *)pfn_to_kaddr(page_to_pfn(page));
 400 got_map_ptr:
 401         memset(ret, 0, memmap_size);
 402
 403         return ret;
 404 }
 405
 406 static inline struct page *kmalloc_section_memmap(unsigned long pnum, int nid,
 407                                                   unsigned long nr_pages)
 408 {
 409         return __kmalloc_section_memmap(nr_pages);
 410 }
 411
 412 static void __kfree_section_memmap(struct page *memmap, unsigned long nr_pages)
 413 {
 414         if (is_vmalloc_addr(memmap))
 415                 vfree(memmap);
 416         else
 417                 free_pages((unsigned long)memmap,
 418                            get_order(sizeof(struct page) * nr_pages));
 419 }
 420
 421 static void free_map_bootmem(struct page *page, unsigned long nr_pages)
 422 {
 423         unsigned long maps_section_nr, removing_section_nr, i;
 424         int magic;
 425
 426         for (i = 0; i < nr_pages; i++, page++) {
 427                 magic = atomic_read(&page->_mapcount);
 428
 429                 BUG_ON(magic == NODE_INFO);
 430
 431                 maps_section_nr = pfn_to_section_nr(page_to_pfn(page));
 432                 removing_section_nr = page->private;
 433
 434                 /*
 435                  * When this function is called, the removing section is
 436                  * logical offlined state. This means all pages are isolated
 437                  * from page allocator. If removing section's memmap is placed
 438                  * on the same section, it must not be freed.
 439                  * If it is freed, page allocator may allocate it which will
 440                  * be removed physically soon.
 441                  */
 442                 if (maps_section_nr != removing_section_nr)
 443                         put_page_bootmem(page);
 444         }
 445 }
 446 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
 447
 448 static void free_section_usemap(struct page *memmap, unsigned long *usemap)
 449 {
 450         struct page *usemap_page;
 451         unsigned long nr_pages;
 452
 453         if (!usemap)
 454                 return;
 455
 456         usemap_page = virt_to_page(usemap);
 457         /*
 458          * Check to see if allocation came from hot-plug-add
 459          */
 460         if (PageSlab(usemap_page)) {
 461                 kfree(usemap);
 462                 if (memmap)
 463                         __kfree_section_memmap(memmap, PAGES_PER_SECTION);
 464                 return;
 465         }
 466
 467         /*
 468          * The usemap came from bootmem. This is packed with other usemaps
 469          * on the section which has pgdat at boot time. Just keep it as is now.
 470          */
 471
 472         if (memmap) {
 473                 struct page *memmap_page;
 474                 memmap_page = virt_to_page(memmap);
 475
 476                 nr_pages = PAGE_ALIGN(PAGES_PER_SECTION * sizeof(struct page))
 477                         >> PAGE_SHIFT;
 478
 479                 free_map_bootmem(memmap_page, nr_pages);
 480         }
 481 }
 482
 483 /*
 484  * returns the number of sections whose mem_maps were properly
 485  * set.  If this is <=0, then that means that the passed-in
 486  * map was not consumed and must be freed.
 487  */
 488 int sparse_add_one_section(struct zone *zone, unsigned long start_pfn,
 489                            int nr_pages)
 490 {
 491         unsigned long section_nr = pfn_to_section_nr(start_pfn);
 492         struct pglist_data *pgdat = zone->zone_pgdat;
 493         struct mem_section *ms;
 494         struct page *memmap;
 495         unsigned long *usemap;
 496         unsigned long flags;
 497         int ret;
 498
 499         /*
 500          * no locking for this, because it does its own
 501          * plus, it does a kmalloc
 502          */
 503         ret = sparse_index_init(section_nr, pgdat->node_id);
 504         if (ret < 0 && ret != -EEXIST)
 505                 return ret;
 506         memmap = kmalloc_section_memmap(section_nr, pgdat->node_id, nr_pages);
 507         if (!memmap)
 508                 return -ENOMEM;
 509         usemap = __kmalloc_section_usemap();
 510         if (!usemap) {
 511                 __kfree_section_memmap(memmap, nr_pages);
 512                 return -ENOMEM;
 513         }
 514
 515         pgdat_resize_lock(pgdat, &flags);
 516
 517         ms = __pfn_to_section(start_pfn);
 518         if (ms->section_mem_map & SECTION_MARKED_PRESENT) {
 519                 ret = -EEXIST;
 520                 goto out;
 521         }
 522
 523         ms->section_mem_map |= SECTION_MARKED_PRESENT;
 524
 525         ret = sparse_init_one_section(ms, section_nr, memmap, usemap);
 526
 527 out:
 528         pgdat_resize_unlock(pgdat, &flags);
 529         if (ret <= 0) {
 530                 kfree(usemap);
 531                 __kfree_section_memmap(memmap, nr_pages);
 532         }
 533         return ret;
 534 }
 535
 536 void sparse_remove_one_section(struct zone *zone, struct mem_section *ms)
 537 {
 538         struct page *memmap = NULL;
 539         unsigned long *usemap = NULL;
 540
 541         if (ms->section_mem_map) {
 542                 usemap = ms->pageblock_flags;
 543                 memmap = sparse_decode_mem_map(ms->section_mem_map,
 544                                                 __section_nr(ms));
 545                 ms->section_mem_map = 0;
 546                 ms->pageblock_flags = NULL;
 547         }
 548
 549         free_section_usemap(memmap, usemap);
 550 }
 551 #endif