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 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 __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         int i = pmd_index(address);
 317
 318         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 319                 pmd_t *pmd = pmd_page + pmd_index(address);
 320
 321                 if (address >= end) {
 322                         if (!after_bootmem) {
 323                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 324                                         set_pmd(pmd, __pmd(0));
 325                         }
 326                         break;
 327                 }
 328
 329                 if (pmd_val(*pmd))
 330                         continue;
 331
 332                 set_pte((pte_t *)pmd,
 333                         pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 334         }
 335         return address;
 336 }
 337
 338 static unsigned long __meminit
 339 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
 340 {
 341         pmd_t *pmd = pmd_offset(pud, 0);
 342         unsigned long last_map_addr;
 343
 344         spin_lock(&init_mm.page_table_lock);
 345         last_map_addr = phys_pmd_init(pmd, address, end);
 346         spin_unlock(&init_mm.page_table_lock);
 347         __flush_tlb_all();
 348         return last_map_addr;
 349 }
 350
 351 static unsigned long __meminit
 352 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
 353 {
 354         unsigned long last_map_addr = end;
 355         int i = pud_index(addr);
 356
 357         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 358                 unsigned long pmd_phys;
 359                 pud_t *pud = pud_page + pud_index(addr);
 360                 pmd_t *pmd;
 361
 362                 if (addr >= end)
 363                         break;
 364
 365                 if (!after_bootmem &&
 366                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 367                         set_pud(pud, __pud(0));
 368                         continue;
 369                 }
 370
 371                 if (pud_val(*pud)) {
 372                         if (!pud_large(*pud))
 373                                 last_map_addr = phys_pmd_update(pud, addr, end);
 374                         continue;
 375                 }
 376
 377                 if (direct_gbpages) {
 378                         set_pte((pte_t *)pud,
 379                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 380                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 381                         continue;
 382                 }
 383
 384                 pmd = alloc_low_page(&pmd_phys);
 385
 386                 spin_lock(&init_mm.page_table_lock);
 387                 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
 388                 last_map_addr = phys_pmd_init(pmd, addr, end);
 389                 spin_unlock(&init_mm.page_table_lock);
 390
 391                 unmap_low_page(pmd);
 392         }
 393         __flush_tlb_all();
 394
 395         return last_map_addr >> PAGE_SHIFT;
 396 }
 397
 398 static void __init find_early_table_space(unsigned long end)
 399 {
 400         unsigned long puds, pmds, tables, start;
 401
 402         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 403         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 404         if (!direct_gbpages) {
 405                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 406                 tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 407         }
 408
 409         /*
 410          * RED-PEN putting page tables only on node 0 could
 411          * cause a hotspot and fill up ZONE_DMA. The page tables
 412          * need roughly 0.5KB per GB.
 413          */
 414         start = 0x8000;
 415         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 416         if (table_start == -1UL)
 417                 panic("Cannot find space for the kernel page tables");
 418
 419         table_start >>= PAGE_SHIFT;
 420         table_end = table_start;
 421
 422         early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
 423                 end, table_start << PAGE_SHIFT,
 424                 (table_start << PAGE_SHIFT) + tables);
 425 }
 426
 427 static void __init init_gbpages(void)
 428 {
 429         if (direct_gbpages && cpu_has_gbpages)
 430                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 431         else
 432                 direct_gbpages = 0;
 433 }
 434
 435 #ifdef CONFIG_MEMTEST_BOOTPARAM
 436
 437 static void __init memtest(unsigned long start_phys, unsigned long size,
 438                                  unsigned pattern)
 439 {
 440         unsigned long i;
 441         unsigned long *start;
 442         unsigned long start_bad;
 443         unsigned long last_bad;
 444         unsigned long val;
 445         unsigned long start_phys_aligned;
 446         unsigned long count;
 447         unsigned long incr;
 448
 449         switch (pattern) {
 450         case 0:
 451                 val = 0UL;
 452                 break;
 453         case 1:
 454                 val = -1UL;
 455                 break;
 456         case 2:
 457                 val = 0x5555555555555555UL;
 458                 break;
 459         case 3:
 460                 val = 0xaaaaaaaaaaaaaaaaUL;
 461                 break;
 462         default:
 463                 return;
 464         }
 465
 466         incr = sizeof(unsigned long);
 467         start_phys_aligned = ALIGN(start_phys, incr);
 468         count = (size - (start_phys_aligned - start_phys))/incr;
 469         start = __va(start_phys_aligned);
 470         start_bad = 0;
 471         last_bad = 0;
 472
 473         for (i = 0; i < count; i++)
 474                 start[i] = val;
 475         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 476                 if (*start != val) {
 477                         if (start_phys_aligned == last_bad + incr) {
 478                                 last_bad += incr;
 479                         } else {
 480                                 if (start_bad) {
 481                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 482                                                 val, start_bad, last_bad + incr);
 483                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 484                                 }
 485                                 start_bad = last_bad = start_phys_aligned;
 486                         }
 487                 }
 488         }
 489         if (start_bad) {
 490                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 491                         val, start_bad, last_bad + incr);
 492                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 493         }
 494
 495 }
 496
 497 static int memtest_pattern __initdata = CONFIG_MEMTEST_BOOTPARAM_VALUE;
 498
 499 static int __init parse_memtest(char *arg)
 500 {
 501         if (arg)
 502                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 503         return 0;
 504 }
 505
 506 early_param("memtest", parse_memtest);
 507
 508 static void __init early_memtest(unsigned long start, unsigned long end)
 509 {
 510         u64 t_start, t_size;
 511         unsigned pattern;
 512
 513         if (!memtest_pattern)
 514                 return;
 515
 516         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 517         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 518                 t_start = start;
 519                 t_size = 0;
 520                 while (t_start < end) {
 521                         t_start = find_e820_area_size(t_start, &t_size, 1);
 522
 523                         /* done ? */
 524                         if (t_start >= end)
 525                                 break;
 526                         if (t_start + t_size > end)
 527                                 t_size = end - t_start;
 528
 529                         printk(KERN_CONT "\n  %016llx - %016llx pattern %d",
 530                                 t_start, t_start + t_size, pattern);
 531
 532                         memtest(t_start, t_size, pattern);
 533
 534                         t_start += t_size;
 535                 }
 536         }
 537         printk(KERN_CONT "\n");
 538 }
 539 #else
 540 static void __init early_memtest(unsigned long start, unsigned long end)
 541 {
 542 }
 543 #endif
 544
 545 /*
 546  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 547  * This runs before bootmem is initialized and gets pages directly from
 548  * the physical memory. To access them they are temporarily mapped.
 549  */
 550 unsigned long __init_refok init_memory_mapping(unsigned long start, unsigned long end)
 551 {
 552         unsigned long next, last_map_addr = end;
 553         unsigned long start_phys = start, end_phys = end;
 554
 555         printk(KERN_INFO "init_memory_mapping\n");
 556
 557         /*
 558          * Find space for the kernel direct mapping tables.
 559          *
 560          * Later we should allocate these tables in the local node of the
 561          * memory mapped. Unfortunately this is done currently before the
 562          * nodes are discovered.
 563          */
 564         if (!after_bootmem) {
 565                 init_gbpages();
 566                 find_early_table_space(end);
 567         }
 568
 569         start = (unsigned long)__va(start);
 570         end = (unsigned long)__va(end);
 571
 572         for (; start < end; start = next) {
 573                 pgd_t *pgd = pgd_offset_k(start);
 574                 unsigned long pud_phys;
 575                 pud_t *pud;
 576
 577                 if (after_bootmem)
 578                         pud = pud_offset(pgd, start & PGDIR_MASK);
 579                 else
 580                         pud = alloc_low_page(&pud_phys);
 581
 582                 next = start + PGDIR_SIZE;
 583                 if (next > end)
 584                         next = end;
 585                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next));
 586                 if (!after_bootmem)
 587                         set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
 588                 unmap_low_page(pud);
 589         }
 590
 591         if (!after_bootmem)
 592                 mmu_cr4_features = read_cr4();
 593         __flush_tlb_all();
 594
 595         if (!after_bootmem)
 596                 reserve_early(table_start << PAGE_SHIFT,
 597                                  table_end << PAGE_SHIFT, "PGTABLE");
 598
 599         if (!after_bootmem)
 600                 early_memtest(start_phys, end_phys);
 601
 602         return last_map_addr;
 603 }
 604
 605 #ifndef CONFIG_NUMA
 606 void __init paging_init(void)
 607 {
 608         unsigned long max_zone_pfns[MAX_NR_ZONES];
 609
 610         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 611         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 612         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 613         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 614
 615         memory_present(0, 0, end_pfn);
 616         sparse_init();
 617         free_area_init_nodes(max_zone_pfns);
 618 }
 619 #endif
 620
 621 /*
 622  * Memory hotplug specific functions
 623  */
 624 #ifdef CONFIG_MEMORY_HOTPLUG
 625 /*
 626  * Memory is added always to NORMAL zone. This means you will never get
 627  * additional DMA/DMA32 memory.
 628  */
 629 int arch_add_memory(int nid, u64 start, u64 size)
 630 {
 631         struct pglist_data *pgdat = NODE_DATA(nid);
 632         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 633         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 634         unsigned long nr_pages = size >> PAGE_SHIFT;
 635         int ret;
 636
 637         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 638         if (last_mapped_pfn > max_pfn_mapped)
 639                 max_pfn_mapped = last_mapped_pfn;
 640
 641         ret = __add_pages(zone, start_pfn, nr_pages);
 642         WARN_ON(1);
 643
 644         return ret;
 645 }
 646 EXPORT_SYMBOL_GPL(arch_add_memory);
 647
 648 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 649 int memory_add_physaddr_to_nid(u64 start)
 650 {
 651         return 0;
 652 }
 653 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 654 #endif
 655
 656 #endif /* CONFIG_MEMORY_HOTPLUG */
 657
 658 /*
 659  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 660  * is valid. The argument is a physical page number.
 661  *
 662  *
 663  * On x86, access has to be given to the first megabyte of ram because that area
 664  * contains bios code and data regions used by X and dosemu and similar apps.
 665  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 666  * mmio resources as well as potential bios/acpi data regions.
 667  */
 668 int devmem_is_allowed(unsigned long pagenr)
 669 {
 670         if (pagenr <= 256)
 671                 return 1;
 672         if (!page_is_ram(pagenr))
 673                 return 1;
 674         return 0;
 675 }
 676
 677
 678 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 679                          kcore_modules, kcore_vsyscall;
 680
 681 void __init mem_init(void)
 682 {
 683         long codesize, reservedpages, datasize, initsize;
 684
 685         pci_iommu_alloc();
 686
 687         /* clear_bss() already clear the empty_zero_page */
 688
 689         reservedpages = 0;
 690
 691         /* this will put all low memory onto the freelists */
 692 #ifdef CONFIG_NUMA
 693         totalram_pages = numa_free_all_bootmem();
 694 #else
 695         totalram_pages = free_all_bootmem();
 696 #endif
 697         reservedpages = end_pfn - totalram_pages -
 698                                         absent_pages_in_range(0, end_pfn);
 699         after_bootmem = 1;
 700
 701         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 702         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 703         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 704
 705         /* Register memory areas for /proc/kcore */
 706         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 707         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 708                    VMALLOC_END-VMALLOC_START);
 709         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 710         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 711         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 712                                  VSYSCALL_END - VSYSCALL_START);
 713
 714         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 715                                 "%ldk reserved, %ldk data, %ldk init)\n",
 716                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 717                 end_pfn << (PAGE_SHIFT-10),
 718                 codesize >> 10,
 719                 reservedpages << (PAGE_SHIFT-10),
 720                 datasize >> 10,
 721                 initsize >> 10);
 722
 723         cpa_init();
 724 }
 725
 726 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 727 {
 728         unsigned long addr = begin;
 729
 730         if (addr >= end)
 731                 return;
 732
 733         /*
 734          * If debugging page accesses then do not free this memory but
 735          * mark them not present - any buggy init-section access will
 736          * create a kernel page fault:
 737          */
 738 #ifdef CONFIG_DEBUG_PAGEALLOC
 739         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 740                 begin, PAGE_ALIGN(end));
 741         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 742 #else
 743         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 744
 745         for (; addr < end; addr += PAGE_SIZE) {
 746                 ClearPageReserved(virt_to_page(addr));
 747                 init_page_count(virt_to_page(addr));
 748                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 749                         POISON_FREE_INITMEM, PAGE_SIZE);
 750                 free_page(addr);
 751                 totalram_pages++;
 752         }
 753 #endif
 754 }
 755
 756 void free_initmem(void)
 757 {
 758         free_init_pages("unused kernel memory",
 759                         (unsigned long)(&__init_begin),
 760                         (unsigned long)(&__init_end));
 761 }
 762
 763 #ifdef CONFIG_DEBUG_RODATA
 764 const int rodata_test_data = 0xC3;
 765 EXPORT_SYMBOL_GPL(rodata_test_data);
 766
 767 void mark_rodata_ro(void)
 768 {
 769         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 770
 771         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 772                (end - start) >> 10);
 773         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 774
 775         /*
 776          * The rodata section (but not the kernel text!) should also be
 777          * not-executable.
 778          */
 779         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 780         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 781
 782         rodata_test();
 783
 784 #ifdef CONFIG_CPA_DEBUG
 785         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 786         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 787
 788         printk(KERN_INFO "Testing CPA: again\n");
 789         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 790 #endif
 791 }
 792
 793 #endif
 794
 795 #ifdef CONFIG_BLK_DEV_INITRD
 796 void free_initrd_mem(unsigned long start, unsigned long end)
 797 {
 798         free_init_pages("initrd memory", start, end);
 799 }
 800 #endif
 801
 802 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
 803 {
 804 #ifdef CONFIG_NUMA
 805         int nid, next_nid;
 806 #endif
 807         unsigned long pfn = phys >> PAGE_SHIFT;
 808
 809         if (pfn >= end_pfn) {
 810                 /*
 811                  * This can happen with kdump kernels when accessing
 812                  * firmware tables:
 813                  */
 814                 if (pfn < max_pfn_mapped)
 815                         return;
 816
 817                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
 818                                 phys, len);
 819                 return;
 820         }
 821
 822         /* Should check here against the e820 map to avoid double free */
 823 #ifdef CONFIG_NUMA
 824         nid = phys_to_nid(phys);
 825         next_nid = phys_to_nid(phys + len - 1);
 826         if (nid == next_nid)
 827                 reserve_bootmem_node(NODE_DATA(nid), phys, len, BOOTMEM_DEFAULT);
 828         else
 829                 reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 830 #else
 831         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 832 #endif
 833
 834         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 835                 dma_reserve += len / PAGE_SIZE;
 836                 set_dma_reserve(dma_reserve);
 837         }
 838 }
 839
 840 int kern_addr_valid(unsigned long addr)
 841 {
 842         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 843         pgd_t *pgd;
 844         pud_t *pud;
 845         pmd_t *pmd;
 846         pte_t *pte;
 847
 848         if (above != 0 && above != -1UL)
 849                 return 0;
 850
 851         pgd = pgd_offset_k(addr);
 852         if (pgd_none(*pgd))
 853                 return 0;
 854
 855         pud = pud_offset(pgd, addr);
 856         if (pud_none(*pud))
 857                 return 0;
 858
 859         pmd = pmd_offset(pud, addr);
 860         if (pmd_none(*pmd))
 861                 return 0;
 862
 863         if (pmd_large(*pmd))
 864                 return pfn_valid(pmd_pfn(*pmd));
 865
 866         pte = pte_offset_kernel(pmd, addr);
 867         if (pte_none(*pte))
 868                 return 0;
 869
 870         return pfn_valid(pte_pfn(*pte));
 871 }
 872
 873 /*
 874  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 875  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 876  * not need special handling anymore:
 877  */
 878 static struct vm_area_struct gate_vma = {
 879         .vm_start       = VSYSCALL_START,
 880         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 881         .vm_page_prot   = PAGE_READONLY_EXEC,
 882         .vm_flags       = VM_READ | VM_EXEC
 883 };
 884
 885 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 886 {
 887 #ifdef CONFIG_IA32_EMULATION
 888         if (test_tsk_thread_flag(tsk, TIF_IA32))
 889                 return NULL;
 890 #endif
 891         return &gate_vma;
 892 }
 893
 894 int in_gate_area(struct task_struct *task, unsigned long addr)
 895 {
 896         struct vm_area_struct *vma = get_gate_vma(task);
 897
 898         if (!vma)
 899                 return 0;
 900
 901         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 902 }
 903
 904 /*
 905  * Use this when you have no reliable task/vma, typically from interrupt
 906  * context. It is less reliable than using the task's vma and may give
 907  * false positives:
 908  */
 909 int in_gate_area_no_task(unsigned long addr)
 910 {
 911         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 912 }
 913
 914 const char *arch_vma_name(struct vm_area_struct *vma)
 915 {
 916         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 917                 return "[vdso]";
 918         if (vma == &gate_vma)
 919                 return "[vsyscall]";
 920         return NULL;
 921 }
 922
 923 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 924 /*
 925  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 926  */
 927 static long __meminitdata addr_start, addr_end;
 928 static void __meminitdata *p_start, *p_end;
 929 static int __meminitdata node_start;
 930
 931 int __meminit
 932 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 933 {
 934         unsigned long addr = (unsigned long)start_page;
 935         unsigned long end = (unsigned long)(start_page + size);
 936         unsigned long next;
 937         pgd_t *pgd;
 938         pud_t *pud;
 939         pmd_t *pmd;
 940
 941         for (; addr < end; addr = next) {
 942                 next = pmd_addr_end(addr, end);
 943
 944                 pgd = vmemmap_pgd_populate(addr, node);
 945                 if (!pgd)
 946                         return -ENOMEM;
 947
 948                 pud = vmemmap_pud_populate(pgd, addr, node);
 949                 if (!pud)
 950                         return -ENOMEM;
 951
 952                 pmd = pmd_offset(pud, addr);
 953                 if (pmd_none(*pmd)) {
 954                         pte_t entry;
 955                         void *p;
 956
 957                         p = vmemmap_alloc_block(PMD_SIZE, node);
 958                         if (!p)
 959                                 return -ENOMEM;
 960
 961                         entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 962                                                         PAGE_KERNEL_LARGE);
 963                         set_pmd(pmd, __pmd(pte_val(entry)));
 964
 965                         /* check to see if we have contiguous blocks */
 966                         if (p_end != p || node_start != node) {
 967                                 if (p_start)
 968                                         printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 969                                                 addr_start, addr_end-1, p_start, p_end-1, node_start);
 970                                 addr_start = addr;
 971                                 node_start = node;
 972                                 p_start = p;
 973                         }
 974                         addr_end = addr + PMD_SIZE;
 975                         p_end = p + PMD_SIZE;
 976                 } else {
 977                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 978                 }
 979         }
 980         return 0;
 981 }
 982
 983 void __meminit vmemmap_populate_print_last(void)
 984 {
 985         if (p_start) {
 986                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 987                         addr_start, addr_end-1, p_start, p_end-1, node_start);
 988                 p_start = NULL;
 989                 p_end = NULL;
 990                 node_start = 0;
 991         }
 992 }
 993 #endif