arch/x86/mm/init_64.c

   1 /*
   2  *  linux/arch/x86_64/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
   6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
   7  */
   8
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/string.h>
  14 #include <linux/types.h>
  15 #include <linux/ptrace.h>
  16 #include <linux/mman.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19 #include <linux/smp.h>
  20 #include <linux/init.h>
  21 #include <linux/initrd.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/proc_fs.h>
  25 #include <linux/pci.h>
  26 #include <linux/pfn.h>
  27 #include <linux/poison.h>
  28 #include <linux/dma-mapping.h>
  29 #include <linux/module.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/nmi.h>
  32
  33 #include <asm/processor.h>
  34 #include <asm/system.h>
  35 #include <asm/uaccess.h>
  36 #include <asm/pgtable.h>
  37 #include <asm/pgalloc.h>
  38 #include <asm/dma.h>
  39 #include <asm/fixmap.h>
  40 #include <asm/e820.h>
  41 #include <asm/apic.h>
  42 #include <asm/tlb.h>
  43 #include <asm/mmu_context.h>
  44 #include <asm/proto.h>
  45 #include <asm/smp.h>
  46 #include <asm/sections.h>
  47 #include <asm/kdebug.h>
  48 #include <asm/numa.h>
  49 #include <asm/cacheflush.h>
  50
  51 static unsigned long dma_reserve __initdata;
  52
  53 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  54
  55 int direct_gbpages __meminitdata
  56 #ifdef CONFIG_DIRECT_GBPAGES
  57                                 = 1
  58 #endif
  59 ;
  60
  61 static int __init parse_direct_gbpages_off(char *arg)
  62 {
  63         direct_gbpages = 0;
  64         return 0;
  65 }
  66 early_param("nogbpages", parse_direct_gbpages_off);
  67
  68 static int __init parse_direct_gbpages_on(char *arg)
  69 {
  70         direct_gbpages = 1;
  71         return 0;
  72 }
  73 early_param("gbpages", parse_direct_gbpages_on);
  74
  75 /*
  76  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  77  * physical space so we can cache the place of the first one and move
  78  * around without checking the pgd every time.
  79  */
  80
  81 void show_mem(void)
  82 {
  83         long i, total = 0, reserved = 0;
  84         long shared = 0, cached = 0;
  85         struct page *page;
  86         pg_data_t *pgdat;
  87
  88         printk(KERN_INFO "Mem-info:\n");
  89         show_free_areas();
  90         for_each_online_pgdat(pgdat) {
  91                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
  92                         /*
  93                          * This loop can take a while with 256 GB and
  94                          * 4k pages so defer the NMI watchdog:
  95                          */
  96                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
  97                                 touch_nmi_watchdog();
  98
  99                         if (!pfn_valid(pgdat->node_start_pfn + i))
 100                                 continue;
 101
 102                         page = pfn_to_page(pgdat->node_start_pfn + i);
 103                         total++;
 104                         if (PageReserved(page))
 105                                 reserved++;
 106                         else if (PageSwapCache(page))
 107                                 cached++;
 108                         else if (page_count(page))
 109                                 shared += page_count(page) - 1;
 110                 }
 111         }
 112         printk(KERN_INFO "%lu pages of RAM\n",          total);
 113         printk(KERN_INFO "%lu reserved pages\n",        reserved);
 114         printk(KERN_INFO "%lu pages shared\n",          shared);
 115         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 116 }
 117
 118 int after_bootmem;
 119
 120 static __init void *spp_getpage(void)
 121 {
 122         void *ptr;
 123
 124         if (after_bootmem)
 125                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 126         else
 127                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 128
 129         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 130                 panic("set_pte_phys: cannot allocate page data %s\n",
 131                         after_bootmem ? "after bootmem" : "");
 132         }
 133
 134         pr_debug("spp_getpage %p\n", ptr);
 135
 136         return ptr;
 137 }
 138
 139 static __init void
 140 set_pte_phys(unsigned long vaddr, unsigned long phys, pgprot_t prot)
 141 {
 142         pgd_t *pgd;
 143         pud_t *pud;
 144         pmd_t *pmd;
 145         pte_t *pte, new_pte;
 146
 147         pr_debug("set_pte_phys %lx to %lx\n", vaddr, phys);
 148
 149         pgd = pgd_offset_k(vaddr);
 150         if (pgd_none(*pgd)) {
 151                 printk(KERN_ERR
 152                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 153                 return;
 154         }
 155         pud = pud_offset(pgd, vaddr);
 156         if (pud_none(*pud)) {
 157                 pmd = (pmd_t *) spp_getpage();
 158                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
 159                 if (pmd != pmd_offset(pud, 0)) {
 160                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 161                                 pmd, pmd_offset(pud, 0));
 162                         return;
 163                 }
 164         }
 165         pmd = pmd_offset(pud, vaddr);
 166         if (pmd_none(*pmd)) {
 167                 pte = (pte_t *) spp_getpage();
 168                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
 169                 if (pte != pte_offset_kernel(pmd, 0)) {
 170                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 171                         return;
 172                 }
 173         }
 174         new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
 175
 176         pte = pte_offset_kernel(pmd, vaddr);
 177         if (!pte_none(*pte) && pte_val(new_pte) &&
 178             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 179                 pte_ERROR(*pte);
 180         set_pte(pte, new_pte);
 181
 182         /*
 183          * It's enough to flush this one mapping.
 184          * (PGE mappings get flushed as well)
 185          */
 186         __flush_tlb_one(vaddr);
 187 }
 188
 189 /*
 190  * The head.S code sets up the kernel high mapping:
 191  *
 192  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 193  *
 194  * phys_addr holds the negative offset to the kernel, which is added
 195  * to the compile time generated pmds. This results in invalid pmds up
 196  * to the point where we hit the physaddr 0 mapping.
 197  *
 198  * We limit the mappings to the region from _text to _end.  _end is
 199  * rounded up to the 2MB boundary. This catches the invalid pmds as
 200  * well, as they are located before _text:
 201  */
 202 void __init cleanup_highmap(void)
 203 {
 204         unsigned long vaddr = __START_KERNEL_map;
 205         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 206         pmd_t *pmd = level2_kernel_pgt;
 207         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 208
 209         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 210                 if (pmd_none(*pmd))
 211                         continue;
 212                 if (vaddr < (unsigned long) _text || vaddr > end)
 213                         set_pmd(pmd, __pmd(0));
 214         }
 215 }
 216
 217 /* NOTE: this is meant to be run only at boot */
 218 void __init __set_fixmap(enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
 219 {
 220         unsigned long address = __fix_to_virt(idx);
 221
 222         if (idx >= __end_of_fixed_addresses) {
 223                 printk(KERN_ERR "Invalid __set_fixmap\n");
 224                 return;
 225         }
 226         set_pte_phys(address, phys, prot);
 227 }
 228
 229 static unsigned long __initdata table_start;
 230 static unsigned long __meminitdata table_end;
 231
 232 static __meminit void *alloc_low_page(unsigned long *phys)
 233 {
 234         unsigned long pfn = table_end++;
 235         void *adr;
 236
 237         if (after_bootmem) {
 238                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 239                 *phys = __pa(adr);
 240
 241                 return adr;
 242         }
 243
 244         if (pfn >= end_pfn)
 245                 panic("alloc_low_page: ran out of memory");
 246
 247         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 248         memset(adr, 0, PAGE_SIZE);
 249         *phys  = pfn * PAGE_SIZE;
 250         return adr;
 251 }
 252
 253 static __meminit void unmap_low_page(void *adr)
 254 {
 255         if (after_bootmem)
 256                 return;
 257
 258         early_iounmap(adr, PAGE_SIZE);
 259 }
 260
 261 /* Must run before zap_low_mappings */
 262 __meminit void *early_ioremap(unsigned long addr, unsigned long size)
 263 {
 264         pmd_t *pmd, *last_pmd;
 265         unsigned long vaddr;
 266         int i, pmds;
 267
 268         pmds = ((addr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 269         vaddr = __START_KERNEL_map;
 270         pmd = level2_kernel_pgt;
 271         last_pmd = level2_kernel_pgt + PTRS_PER_PMD - 1;
 272
 273         for (; pmd <= last_pmd; pmd++, vaddr += PMD_SIZE) {
 274                 for (i = 0; i < pmds; i++) {
 275                         if (pmd_present(pmd[i]))
 276                                 goto continue_outer_loop;
 277                 }
 278                 vaddr += addr & ~PMD_MASK;
 279                 addr &= PMD_MASK;
 280
 281                 for (i = 0; i < pmds; i++, addr += PMD_SIZE)
 282                         set_pmd(pmd+i, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
 283                 __flush_tlb_all();
 284
 285                 return (void *)vaddr;
 286 continue_outer_loop:
 287                 ;
 288         }
 289         printk(KERN_ERR "early_ioremap(0x%lx, %lu) failed\n", addr, size);
 290
 291         return NULL;
 292 }
 293
 294 /*
 295  * To avoid virtual aliases later:
 296  */
 297 __meminit void early_iounmap(void *addr, unsigned long size)
 298 {
 299         unsigned long vaddr;
 300         pmd_t *pmd;
 301         int i, pmds;
 302
 303         vaddr = (unsigned long)addr;
 304         pmds = ((vaddr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 305         pmd = level2_kernel_pgt + pmd_index(vaddr);
 306
 307         for (i = 0; i < pmds; i++)
 308                 pmd_clear(pmd + i);
 309
 310         __flush_tlb_all();
 311 }
 312
 313 static unsigned long __meminit
 314 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
 315 {
 316         unsigned long pages = 0;
 317
 318         int i = pmd_index(address);
 319
 320         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 321                 pmd_t *pmd = pmd_page + pmd_index(address);
 322
 323                 if (address >= end) {
 324                         if (!after_bootmem) {
 325                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 326                                         set_pmd(pmd, __pmd(0));
 327                         }
 328                         break;
 329                 }
 330
 331                 if (pmd_val(*pmd))
 332                         continue;
 333
 334                 pages++;
 335                 set_pte((pte_t *)pmd,
 336                         pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 337         }
 338         update_page_count(PG_LEVEL_2M, pages);
 339         return address;
 340 }
 341
 342 static unsigned long __meminit
 343 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
 344 {
 345         pmd_t *pmd = pmd_offset(pud, 0);
 346         unsigned long last_map_addr;
 347
 348         spin_lock(&init_mm.page_table_lock);
 349         last_map_addr = phys_pmd_init(pmd, address, end);
 350         spin_unlock(&init_mm.page_table_lock);
 351         __flush_tlb_all();
 352         return last_map_addr;
 353 }
 354
 355 static unsigned long __meminit
 356 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
 357 {
 358         unsigned long pages = 0;
 359         unsigned long last_map_addr = end;
 360         int i = pud_index(addr);
 361
 362         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 363                 unsigned long pmd_phys;
 364                 pud_t *pud = pud_page + pud_index(addr);
 365                 pmd_t *pmd;
 366
 367                 if (addr >= end)
 368                         break;
 369
 370                 if (!after_bootmem &&
 371                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 372                         set_pud(pud, __pud(0));
 373                         continue;
 374                 }
 375
 376                 if (pud_val(*pud)) {
 377                         if (!pud_large(*pud))
 378                                 last_map_addr = phys_pmd_update(pud, addr, end);
 379                         continue;
 380                 }
 381
 382                 if (direct_gbpages) {
 383                         pages++;
 384                         set_pte((pte_t *)pud,
 385                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 386                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 387                         continue;
 388                 }
 389
 390                 pmd = alloc_low_page(&pmd_phys);
 391
 392                 spin_lock(&init_mm.page_table_lock);
 393                 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
 394                 last_map_addr = phys_pmd_init(pmd, addr, end);
 395                 spin_unlock(&init_mm.page_table_lock);
 396
 397                 unmap_low_page(pmd);
 398         }
 399         __flush_tlb_all();
 400         update_page_count(PG_LEVEL_1G, pages);
 401
 402         return last_map_addr >> PAGE_SHIFT;
 403 }
 404
 405 static void __init find_early_table_space(unsigned long end)
 406 {
 407         unsigned long puds, pmds, tables, start;
 408
 409         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 410         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 411         if (!direct_gbpages) {
 412                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 413                 tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 414         }
 415
 416         /*
 417          * RED-PEN putting page tables only on node 0 could
 418          * cause a hotspot and fill up ZONE_DMA. The page tables
 419          * need roughly 0.5KB per GB.
 420          */
 421         start = 0x8000;
 422         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 423         if (table_start == -1UL)
 424                 panic("Cannot find space for the kernel page tables");
 425
 426         table_start >>= PAGE_SHIFT;
 427         table_end = table_start;
 428
 429         early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
 430                 end, table_start << PAGE_SHIFT,
 431                 (table_start << PAGE_SHIFT) + tables);
 432 }
 433
 434 static void __init init_gbpages(void)
 435 {
 436         if (direct_gbpages && cpu_has_gbpages)
 437                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 438         else
 439                 direct_gbpages = 0;
 440 }
 441
 442 #ifdef CONFIG_MEMTEST
 443
 444 static void __init memtest(unsigned long start_phys, unsigned long size,
 445                                  unsigned pattern)
 446 {
 447         unsigned long i;
 448         unsigned long *start;
 449         unsigned long start_bad;
 450         unsigned long last_bad;
 451         unsigned long val;
 452         unsigned long start_phys_aligned;
 453         unsigned long count;
 454         unsigned long incr;
 455
 456         switch (pattern) {
 457         case 0:
 458                 val = 0UL;
 459                 break;
 460         case 1:
 461                 val = -1UL;
 462                 break;
 463         case 2:
 464                 val = 0x5555555555555555UL;
 465                 break;
 466         case 3:
 467                 val = 0xaaaaaaaaaaaaaaaaUL;
 468                 break;
 469         default:
 470                 return;
 471         }
 472
 473         incr = sizeof(unsigned long);
 474         start_phys_aligned = ALIGN(start_phys, incr);
 475         count = (size - (start_phys_aligned - start_phys))/incr;
 476         start = __va(start_phys_aligned);
 477         start_bad = 0;
 478         last_bad = 0;
 479
 480         for (i = 0; i < count; i++)
 481                 start[i] = val;
 482         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 483                 if (*start != val) {
 484                         if (start_phys_aligned == last_bad + incr) {
 485                                 last_bad += incr;
 486                         } else {
 487                                 if (start_bad) {
 488                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 489                                                 val, start_bad, last_bad + incr);
 490                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 491                                 }
 492                                 start_bad = last_bad = start_phys_aligned;
 493                         }
 494                 }
 495         }
 496         if (start_bad) {
 497                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 498                         val, start_bad, last_bad + incr);
 499                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 500         }
 501
 502 }
 503
 504 /* default is disabled */
 505 static int memtest_pattern __initdata;
 506
 507 static int __init parse_memtest(char *arg)
 508 {
 509         if (arg)
 510                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 511         return 0;
 512 }
 513
 514 early_param("memtest", parse_memtest);
 515
 516 static void __init early_memtest(unsigned long start, unsigned long end)
 517 {
 518         u64 t_start, t_size;
 519         unsigned pattern;
 520
 521         if (!memtest_pattern)
 522                 return;
 523
 524         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 525         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 526                 t_start = start;
 527                 t_size = 0;
 528                 while (t_start < end) {
 529                         t_start = find_e820_area_size(t_start, &t_size, 1);
 530
 531                         /* done ? */
 532                         if (t_start >= end)
 533                                 break;
 534                         if (t_start + t_size > end)
 535                                 t_size = end - t_start;
 536
 537                         printk(KERN_CONT "\n  %016llx - %016llx pattern %d",
 538                                 (unsigned long long)t_start,
 539                                 (unsigned long long)t_start + t_size, pattern);
 540
 541                         memtest(t_start, t_size, pattern);
 542
 543                         t_start += t_size;
 544                 }
 545         }
 546         printk(KERN_CONT "\n");
 547 }
 548 #else
 549 static void __init early_memtest(unsigned long start, unsigned long end)
 550 {
 551 }
 552 #endif
 553
 554 /*
 555  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 556  * This runs before bootmem is initialized and gets pages directly from
 557  * the physical memory. To access them they are temporarily mapped.
 558  */
 559 unsigned long __init_refok init_memory_mapping(unsigned long start, unsigned long end)
 560 {
 561         unsigned long next, last_map_addr = end;
 562         unsigned long start_phys = start, end_phys = end;
 563
 564         printk(KERN_INFO "init_memory_mapping\n");
 565
 566         /*
 567          * Find space for the kernel direct mapping tables.
 568          *
 569          * Later we should allocate these tables in the local node of the
 570          * memory mapped. Unfortunately this is done currently before the
 571          * nodes are discovered.
 572          */
 573         if (!after_bootmem) {
 574                 init_gbpages();
 575                 find_early_table_space(end);
 576         }
 577
 578         start = (unsigned long)__va(start);
 579         end = (unsigned long)__va(end);
 580
 581         for (; start < end; start = next) {
 582                 pgd_t *pgd = pgd_offset_k(start);
 583                 unsigned long pud_phys;
 584                 pud_t *pud;
 585
 586                 if (after_bootmem)
 587                         pud = pud_offset(pgd, start & PGDIR_MASK);
 588                 else
 589                         pud = alloc_low_page(&pud_phys);
 590
 591                 next = start + PGDIR_SIZE;
 592                 if (next > end)
 593                         next = end;
 594                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next));
 595                 if (!after_bootmem)
 596                         set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
 597                 unmap_low_page(pud);
 598         }
 599
 600         if (!after_bootmem)
 601                 mmu_cr4_features = read_cr4();
 602         __flush_tlb_all();
 603
 604         if (!after_bootmem)
 605                 reserve_early(table_start << PAGE_SHIFT,
 606                                  table_end << PAGE_SHIFT, "PGTABLE");
 607
 608         if (!after_bootmem)
 609                 early_memtest(start_phys, end_phys);
 610
 611         return last_map_addr;
 612 }
 613
 614 #ifndef CONFIG_NUMA
 615 void __init paging_init(void)
 616 {
 617         unsigned long max_zone_pfns[MAX_NR_ZONES];
 618
 619         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 620         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 621         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 622         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 623
 624         memory_present(0, 0, end_pfn);
 625         sparse_init();
 626         free_area_init_nodes(max_zone_pfns);
 627 }
 628 #endif
 629
 630 /*
 631  * Memory hotplug specific functions
 632  */
 633 #ifdef CONFIG_MEMORY_HOTPLUG
 634 /*
 635  * Memory is added always to NORMAL zone. This means you will never get
 636  * additional DMA/DMA32 memory.
 637  */
 638 int arch_add_memory(int nid, u64 start, u64 size)
 639 {
 640         struct pglist_data *pgdat = NODE_DATA(nid);
 641         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 642         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 643         unsigned long nr_pages = size >> PAGE_SHIFT;
 644         int ret;
 645
 646         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 647         if (last_mapped_pfn > max_pfn_mapped)
 648                 max_pfn_mapped = last_mapped_pfn;
 649
 650         ret = __add_pages(zone, start_pfn, nr_pages);
 651         WARN_ON(1);
 652
 653         return ret;
 654 }
 655 EXPORT_SYMBOL_GPL(arch_add_memory);
 656
 657 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 658 int memory_add_physaddr_to_nid(u64 start)
 659 {
 660         return 0;
 661 }
 662 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 663 #endif
 664
 665 #endif /* CONFIG_MEMORY_HOTPLUG */
 666
 667 /*
 668  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 669  * is valid. The argument is a physical page number.
 670  *
 671  *
 672  * On x86, access has to be given to the first megabyte of ram because that area
 673  * contains bios code and data regions used by X and dosemu and similar apps.
 674  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 675  * mmio resources as well as potential bios/acpi data regions.
 676  */
 677 int devmem_is_allowed(unsigned long pagenr)
 678 {
 679         if (pagenr <= 256)
 680                 return 1;
 681         if (!page_is_ram(pagenr))
 682                 return 1;
 683         return 0;
 684 }
 685
 686
 687 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 688                          kcore_modules, kcore_vsyscall;
 689
 690 void __init mem_init(void)
 691 {
 692         long codesize, reservedpages, datasize, initsize;
 693
 694         pci_iommu_alloc();
 695
 696         /* clear_bss() already clear the empty_zero_page */
 697
 698         reservedpages = 0;
 699
 700         /* this will put all low memory onto the freelists */
 701 #ifdef CONFIG_NUMA
 702         totalram_pages = numa_free_all_bootmem();
 703 #else
 704         totalram_pages = free_all_bootmem();
 705 #endif
 706         reservedpages = end_pfn - totalram_pages -
 707                                         absent_pages_in_range(0, end_pfn);
 708         after_bootmem = 1;
 709
 710         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 711         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 712         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 713
 714         /* Register memory areas for /proc/kcore */
 715         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 716         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 717                    VMALLOC_END-VMALLOC_START);
 718         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 719         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 720         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 721                                  VSYSCALL_END - VSYSCALL_START);
 722
 723         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 724                                 "%ldk reserved, %ldk data, %ldk init)\n",
 725                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 726                 end_pfn << (PAGE_SHIFT-10),
 727                 codesize >> 10,
 728                 reservedpages << (PAGE_SHIFT-10),
 729                 datasize >> 10,
 730                 initsize >> 10);
 731
 732         cpa_init();
 733 }
 734
 735 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 736 {
 737         unsigned long addr = begin;
 738
 739         if (addr >= end)
 740                 return;
 741
 742         /*
 743          * If debugging page accesses then do not free this memory but
 744          * mark them not present - any buggy init-section access will
 745          * create a kernel page fault:
 746          */
 747 #ifdef CONFIG_DEBUG_PAGEALLOC
 748         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 749                 begin, PAGE_ALIGN(end));
 750         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 751 #else
 752         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 753
 754         for (; addr < end; addr += PAGE_SIZE) {
 755                 ClearPageReserved(virt_to_page(addr));
 756                 init_page_count(virt_to_page(addr));
 757                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 758                         POISON_FREE_INITMEM, PAGE_SIZE);
 759                 free_page(addr);
 760                 totalram_pages++;
 761         }
 762 #endif
 763 }
 764
 765 void free_initmem(void)
 766 {
 767         free_init_pages("unused kernel memory",
 768                         (unsigned long)(&__init_begin),
 769                         (unsigned long)(&__init_end));
 770 }
 771
 772 #ifdef CONFIG_DEBUG_RODATA
 773 const int rodata_test_data = 0xC3;
 774 EXPORT_SYMBOL_GPL(rodata_test_data);
 775
 776 void mark_rodata_ro(void)
 777 {
 778         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 779
 780         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 781                (end - start) >> 10);
 782         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 783
 784         /*
 785          * The rodata section (but not the kernel text!) should also be
 786          * not-executable.
 787          */
 788         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 789         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 790
 791         rodata_test();
 792
 793 #ifdef CONFIG_CPA_DEBUG
 794         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 795         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 796
 797         printk(KERN_INFO "Testing CPA: again\n");
 798         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 799 #endif
 800 }
 801
 802 #endif
 803
 804 #ifdef CONFIG_BLK_DEV_INITRD
 805 void free_initrd_mem(unsigned long start, unsigned long end)
 806 {
 807         free_init_pages("initrd memory", start, end);
 808 }
 809 #endif
 810
 811 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
 812 {
 813 #ifdef CONFIG_NUMA
 814         int nid, next_nid;
 815 #endif
 816         unsigned long pfn = phys >> PAGE_SHIFT;
 817
 818         if (pfn >= end_pfn) {
 819                 /*
 820                  * This can happen with kdump kernels when accessing
 821                  * firmware tables:
 822                  */
 823                 if (pfn < max_pfn_mapped)
 824                         return;
 825
 826                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
 827                                 phys, len);
 828                 return;
 829         }
 830
 831         /* Should check here against the e820 map to avoid double free */
 832 #ifdef CONFIG_NUMA
 833         nid = phys_to_nid(phys);
 834         next_nid = phys_to_nid(phys + len - 1);
 835         if (nid == next_nid)
 836                 reserve_bootmem_node(NODE_DATA(nid), phys, len, BOOTMEM_DEFAULT);
 837         else
 838                 reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 839 #else
 840         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 841 #endif
 842
 843         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 844                 dma_reserve += len / PAGE_SIZE;
 845                 set_dma_reserve(dma_reserve);
 846         }
 847 }
 848
 849 int kern_addr_valid(unsigned long addr)
 850 {
 851         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 852         pgd_t *pgd;
 853         pud_t *pud;
 854         pmd_t *pmd;
 855         pte_t *pte;
 856
 857         if (above != 0 && above != -1UL)
 858                 return 0;
 859
 860         pgd = pgd_offset_k(addr);
 861         if (pgd_none(*pgd))
 862                 return 0;
 863
 864         pud = pud_offset(pgd, addr);
 865         if (pud_none(*pud))
 866                 return 0;
 867
 868         pmd = pmd_offset(pud, addr);
 869         if (pmd_none(*pmd))
 870                 return 0;
 871
 872         if (pmd_large(*pmd))
 873                 return pfn_valid(pmd_pfn(*pmd));
 874
 875         pte = pte_offset_kernel(pmd, addr);
 876         if (pte_none(*pte))
 877                 return 0;
 878
 879         return pfn_valid(pte_pfn(*pte));
 880 }
 881
 882 /*
 883  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 884  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 885  * not need special handling anymore:
 886  */
 887 static struct vm_area_struct gate_vma = {
 888         .vm_start       = VSYSCALL_START,
 889         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 890         .vm_page_prot   = PAGE_READONLY_EXEC,
 891         .vm_flags       = VM_READ | VM_EXEC
 892 };
 893
 894 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 895 {
 896 #ifdef CONFIG_IA32_EMULATION
 897         if (test_tsk_thread_flag(tsk, TIF_IA32))
 898                 return NULL;
 899 #endif
 900         return &gate_vma;
 901 }
 902
 903 int in_gate_area(struct task_struct *task, unsigned long addr)
 904 {
 905         struct vm_area_struct *vma = get_gate_vma(task);
 906
 907         if (!vma)
 908                 return 0;
 909
 910         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 911 }
 912
 913 /*
 914  * Use this when you have no reliable task/vma, typically from interrupt
 915  * context. It is less reliable than using the task's vma and may give
 916  * false positives:
 917  */
 918 int in_gate_area_no_task(unsigned long addr)
 919 {
 920         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 921 }
 922
 923 const char *arch_vma_name(struct vm_area_struct *vma)
 924 {
 925         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 926                 return "[vdso]";
 927         if (vma == &gate_vma)
 928                 return "[vsyscall]";
 929         return NULL;
 930 }
 931
 932 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 933 /*
 934  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 935  */
 936 static long __meminitdata addr_start, addr_end;
 937 static void __meminitdata *p_start, *p_end;
 938 static int __meminitdata node_start;
 939
 940 int __meminit
 941 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 942 {
 943         unsigned long addr = (unsigned long)start_page;
 944         unsigned long end = (unsigned long)(start_page + size);
 945         unsigned long next;
 946         pgd_t *pgd;
 947         pud_t *pud;
 948         pmd_t *pmd;
 949
 950         for (; addr < end; addr = next) {
 951                 next = pmd_addr_end(addr, end);
 952
 953                 pgd = vmemmap_pgd_populate(addr, node);
 954                 if (!pgd)
 955                         return -ENOMEM;
 956
 957                 pud = vmemmap_pud_populate(pgd, addr, node);
 958                 if (!pud)
 959                         return -ENOMEM;
 960
 961                 pmd = pmd_offset(pud, addr);
 962                 if (pmd_none(*pmd)) {
 963                         pte_t entry;
 964                         void *p;
 965
 966                         p = vmemmap_alloc_block(PMD_SIZE, node);
 967                         if (!p)
 968                                 return -ENOMEM;
 969
 970                         entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 971                                                         PAGE_KERNEL_LARGE);
 972                         set_pmd(pmd, __pmd(pte_val(entry)));
 973
 974                         /* check to see if we have contiguous blocks */
 975                         if (p_end != p || node_start != node) {
 976                                 if (p_start)
 977                                         printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 978                                                 addr_start, addr_end-1, p_start, p_end-1, node_start);
 979                                 addr_start = addr;
 980                                 node_start = node;
 981                                 p_start = p;
 982                         }
 983                         addr_end = addr + PMD_SIZE;
 984                         p_end = p + PMD_SIZE;
 985                 } else {
 986                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 987                 }
 988         }
 989         return 0;
 990 }
 991
 992 void __meminit vmemmap_populate_print_last(void)
 993 {
 994         if (p_start) {
 995                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 996                         addr_start, addr_end-1, p_start, p_end-1, node_start);
 997                 p_start = NULL;
 998                 p_end = NULL;
 999                 node_start = 0;
1000         }
1001 }
1002 #endif