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