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