git.oblomov.eu Git - linux-2.6/blob - 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 /*
  51  * PFN of last memory page.
  52  */
  53 unsigned long end_pfn;
  54
  55 /*
  56  * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
  57  * The direct mapping extends to max_pfn_mapped, so that we can directly access
  58  * apertures, ACPI and other tables without having to play with fixmaps.
  59  */
  60 unsigned long max_pfn_mapped;
  61
  62 static unsigned long dma_reserve __initdata;
  63
  64 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  65
  66 int direct_gbpages __meminitdata
  67 #ifdef CONFIG_DIRECT_GBPAGES
  68                                 = 1
  69 #endif
  70 ;
  71
  72 static int __init parse_direct_gbpages_off(char *arg)
  73 {
  74         direct_gbpages = 0;
  75         return 0;
  76 }
  77 early_param("nogbpages", parse_direct_gbpages_off);
  78
  79 static int __init parse_direct_gbpages_on(char *arg)
  80 {
  81         direct_gbpages = 1;
  82         return 0;
  83 }
  84 early_param("gbpages", parse_direct_gbpages_on);
  85
  86 /*
  87  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  88  * physical space so we can cache the place of the first one and move
  89  * around without checking the pgd every time.
  90  */
  91
  92 void show_mem(void)
  93 {
  94         long i, total = 0, reserved = 0;
  95         long shared = 0, cached = 0;
  96         struct page *page;
  97         pg_data_t *pgdat;
  98
  99         printk(KERN_INFO "Mem-info:\n");
 100         show_free_areas();
 101         for_each_online_pgdat(pgdat) {
 102                 for (i = 0; i < pgdat->node_spanned_pages; ++i) {
 103                         /*
 104                          * This loop can take a while with 256 GB and
 105                          * 4k pages so defer the NMI watchdog:
 106                          */
 107                         if (unlikely(i % MAX_ORDER_NR_PAGES == 0))
 108                                 touch_nmi_watchdog();
 109
 110                         if (!pfn_valid(pgdat->node_start_pfn + i))
 111                                 continue;
 112
 113                         page = pfn_to_page(pgdat->node_start_pfn + i);
 114                         total++;
 115                         if (PageReserved(page))
 116                                 reserved++;
 117                         else if (PageSwapCache(page))
 118                                 cached++;
 119                         else if (page_count(page))
 120                                 shared += page_count(page) - 1;
 121                 }
 122         }
 123         printk(KERN_INFO "%lu pages of RAM\n",          total);
 124         printk(KERN_INFO "%lu reserved pages\n",        reserved);
 125         printk(KERN_INFO "%lu pages shared\n",          shared);
 126         printk(KERN_INFO "%lu pages swap cached\n",     cached);
 127 }
 128
 129 int after_bootmem;
 130
 131 static __init void *spp_getpage(void)
 132 {
 133         void *ptr;
 134
 135         if (after_bootmem)
 136                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 137         else
 138                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 139
 140         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 141                 panic("set_pte_phys: cannot allocate page data %s\n",
 142                         after_bootmem ? "after bootmem" : "");
 143         }
 144
 145         pr_debug("spp_getpage %p\n", ptr);
 146
 147         return ptr;
 148 }
 149
 150 static __init void
 151 set_pte_phys(unsigned long vaddr, unsigned long phys, pgprot_t prot)
 152 {
 153         pgd_t *pgd;
 154         pud_t *pud;
 155         pmd_t *pmd;
 156         pte_t *pte, new_pte;
 157
 158         pr_debug("set_pte_phys %lx to %lx\n", vaddr, phys);
 159
 160         pgd = pgd_offset_k(vaddr);
 161         if (pgd_none(*pgd)) {
 162                 printk(KERN_ERR
 163                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 164                 return;
 165         }
 166         pud = pud_offset(pgd, vaddr);
 167         if (pud_none(*pud)) {
 168                 pmd = (pmd_t *) spp_getpage();
 169                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
 170                 if (pmd != pmd_offset(pud, 0)) {
 171                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 172                                 pmd, pmd_offset(pud, 0));
 173                         return;
 174                 }
 175         }
 176         pmd = pmd_offset(pud, vaddr);
 177         if (pmd_none(*pmd)) {
 178                 pte = (pte_t *) spp_getpage();
 179                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
 180                 if (pte != pte_offset_kernel(pmd, 0)) {
 181                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 182                         return;
 183                 }
 184         }
 185         new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
 186
 187         pte = pte_offset_kernel(pmd, vaddr);
 188         if (!pte_none(*pte) && pte_val(new_pte) &&
 189             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 190                 pte_ERROR(*pte);
 191         set_pte(pte, new_pte);
 192
 193         /*
 194          * It's enough to flush this one mapping.
 195          * (PGE mappings get flushed as well)
 196          */
 197         __flush_tlb_one(vaddr);
 198 }
 199
 200 /*
 201  * The head.S code sets up the kernel high mapping:
 202  *
 203  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 204  *
 205  * phys_addr holds the negative offset to the kernel, which is added
 206  * to the compile time generated pmds. This results in invalid pmds up
 207  * to the point where we hit the physaddr 0 mapping.
 208  *
 209  * We limit the mappings to the region from _text to _end.  _end is
 210  * rounded up to the 2MB boundary. This catches the invalid pmds as
 211  * well, as they are located before _text:
 212  */
 213 void __init cleanup_highmap(void)
 214 {
 215         unsigned long vaddr = __START_KERNEL_map;
 216         unsigned long end = round_up((unsigned long)_end, PMD_SIZE) - 1;
 217         pmd_t *pmd = level2_kernel_pgt;
 218         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 219
 220         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 221                 if (pmd_none(*pmd))
 222                         continue;
 223                 if (vaddr < (unsigned long) _text || vaddr > end)
 224                         set_pmd(pmd, __pmd(0));
 225         }
 226 }
 227
 228 /* NOTE: this is meant to be run only at boot */
 229 void __init __set_fixmap(enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
 230 {
 231         unsigned long address = __fix_to_virt(idx);
 232
 233         if (idx >= __end_of_fixed_addresses) {
 234                 printk(KERN_ERR "Invalid __set_fixmap\n");
 235                 return;
 236         }
 237         set_pte_phys(address, phys, prot);
 238 }
 239
 240 static unsigned long __initdata table_start;
 241 static unsigned long __meminitdata table_end;
 242
 243 static __meminit void *alloc_low_page(unsigned long *phys)
 244 {
 245         unsigned long pfn = table_end++;
 246         void *adr;
 247
 248         if (after_bootmem) {
 249                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 250                 *phys = __pa(adr);
 251
 252                 return adr;
 253         }
 254
 255         if (pfn >= end_pfn)
 256                 panic("alloc_low_page: ran out of memory");
 257
 258         adr = early_ioremap(pfn * PAGE_SIZE, PAGE_SIZE);
 259         memset(adr, 0, PAGE_SIZE);
 260         *phys  = pfn * PAGE_SIZE;
 261         return adr;
 262 }
 263
 264 static __meminit void unmap_low_page(void *adr)
 265 {
 266         if (after_bootmem)
 267                 return;
 268
 269         early_iounmap(adr, PAGE_SIZE);
 270 }
 271
 272 /* Must run before zap_low_mappings */
 273 __meminit void *early_ioremap(unsigned long addr, unsigned long size)
 274 {
 275         pmd_t *pmd, *last_pmd;
 276         unsigned long vaddr;
 277         int i, pmds;
 278
 279         pmds = ((addr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 280         vaddr = __START_KERNEL_map;
 281         pmd = level2_kernel_pgt;
 282         last_pmd = level2_kernel_pgt + PTRS_PER_PMD - 1;
 283
 284         for (; pmd <= last_pmd; pmd++, vaddr += PMD_SIZE) {
 285                 for (i = 0; i < pmds; i++) {
 286                         if (pmd_present(pmd[i]))
 287                                 goto continue_outer_loop;
 288                 }
 289                 vaddr += addr & ~PMD_MASK;
 290                 addr &= PMD_MASK;
 291
 292                 for (i = 0; i < pmds; i++, addr += PMD_SIZE)
 293                         set_pmd(pmd+i, __pmd(addr | __PAGE_KERNEL_LARGE_EXEC));
 294                 __flush_tlb_all();
 295
 296                 return (void *)vaddr;
 297 continue_outer_loop:
 298                 ;
 299         }
 300         printk(KERN_ERR "early_ioremap(0x%lx, %lu) failed\n", addr, size);
 301
 302         return NULL;
 303 }
 304
 305 /*
 306  * To avoid virtual aliases later:
 307  */
 308 __meminit void early_iounmap(void *addr, unsigned long size)
 309 {
 310         unsigned long vaddr;
 311         pmd_t *pmd;
 312         int i, pmds;
 313
 314         vaddr = (unsigned long)addr;
 315         pmds = ((vaddr & ~PMD_MASK) + size + ~PMD_MASK) / PMD_SIZE;
 316         pmd = level2_kernel_pgt + pmd_index(vaddr);
 317
 318         for (i = 0; i < pmds; i++)
 319                 pmd_clear(pmd + i);
 320
 321         __flush_tlb_all();
 322 }
 323
 324 static unsigned long __meminit
 325 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end)
 326 {
 327         int i = pmd_index(address);
 328
 329         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 330                 pmd_t *pmd = pmd_page + pmd_index(address);
 331
 332                 if (address >= end) {
 333                         if (!after_bootmem) {
 334                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 335                                         set_pmd(pmd, __pmd(0));
 336                         }
 337                         break;
 338                 }
 339
 340                 if (pmd_val(*pmd))
 341                         continue;
 342
 343                 set_pte((pte_t *)pmd,
 344                         pfn_pte(address >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 345         }
 346         return address;
 347 }
 348
 349 static unsigned long __meminit
 350 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end)
 351 {
 352         pmd_t *pmd = pmd_offset(pud, 0);
 353         unsigned long last_map_addr;
 354
 355         spin_lock(&init_mm.page_table_lock);
 356         last_map_addr = phys_pmd_init(pmd, address, end);
 357         spin_unlock(&init_mm.page_table_lock);
 358         __flush_tlb_all();
 359         return last_map_addr;
 360 }
 361
 362 static unsigned long __meminit
 363 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end)
 364 {
 365         unsigned long last_map_addr = end;
 366         int i = pud_index(addr);
 367
 368         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 369                 unsigned long pmd_phys;
 370                 pud_t *pud = pud_page + pud_index(addr);
 371                 pmd_t *pmd;
 372
 373                 if (addr >= end)
 374                         break;
 375
 376                 if (!after_bootmem &&
 377                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 378                         set_pud(pud, __pud(0));
 379                         continue;
 380                 }
 381
 382                 if (pud_val(*pud)) {
 383                         if (!pud_large(*pud))
 384                                 last_map_addr = phys_pmd_update(pud, addr, end);
 385                         continue;
 386                 }
 387
 388                 if (direct_gbpages) {
 389                         set_pte((pte_t *)pud,
 390                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 391                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 392                         continue;
 393                 }
 394
 395                 pmd = alloc_low_page(&pmd_phys);
 396
 397                 spin_lock(&init_mm.page_table_lock);
 398                 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
 399                 last_map_addr = phys_pmd_init(pmd, addr, end);
 400                 spin_unlock(&init_mm.page_table_lock);
 401
 402                 unmap_low_page(pmd);
 403         }
 404         __flush_tlb_all();
 405
 406         return last_map_addr >> PAGE_SHIFT;
 407 }
 408
 409 static void __init find_early_table_space(unsigned long end)
 410 {
 411         unsigned long puds, pmds, tables, start;
 412
 413         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 414         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE);
 415         if (!direct_gbpages) {
 416                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 417                 tables += round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 418         }
 419
 420         /*
 421          * RED-PEN putting page tables only on node 0 could
 422          * cause a hotspot and fill up ZONE_DMA. The page tables
 423          * need roughly 0.5KB per GB.
 424          */
 425         start = 0x8000;
 426         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 427         if (table_start == -1UL)
 428                 panic("Cannot find space for the kernel page tables");
 429
 430         table_start >>= PAGE_SHIFT;
 431         table_end = table_start;
 432
 433         early_printk("kernel direct mapping tables up to %lx @ %lx-%lx\n",
 434                 end, table_start << PAGE_SHIFT,
 435                 (table_start << PAGE_SHIFT) + tables);
 436 }
 437
 438 static void __init init_gbpages(void)
 439 {
 440         if (direct_gbpages && cpu_has_gbpages)
 441                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 442         else
 443                 direct_gbpages = 0;
 444 }
 445
 446 #ifdef CONFIG_MEMTEST_BOOTPARAM
 447
 448 static void __init memtest(unsigned long start_phys, unsigned long size,
 449                                  unsigned pattern)
 450 {
 451         unsigned long i;
 452         unsigned long *start;
 453         unsigned long start_bad;
 454         unsigned long last_bad;
 455         unsigned long val;
 456         unsigned long start_phys_aligned;
 457         unsigned long count;
 458         unsigned long incr;
 459
 460         switch (pattern) {
 461         case 0:
 462                 val = 0UL;
 463                 break;
 464         case 1:
 465                 val = -1UL;
 466                 break;
 467         case 2:
 468                 val = 0x5555555555555555UL;
 469                 break;
 470         case 3:
 471                 val = 0xaaaaaaaaaaaaaaaaUL;
 472                 break;
 473         default:
 474                 return;
 475         }
 476
 477         incr = sizeof(unsigned long);
 478         start_phys_aligned = ALIGN(start_phys, incr);
 479         count = (size - (start_phys_aligned - start_phys))/incr;
 480         start = __va(start_phys_aligned);
 481         start_bad = 0;
 482         last_bad = 0;
 483
 484         for (i = 0; i < count; i++)
 485                 start[i] = val;
 486         for (i = 0; i < count; i++, start++, start_phys_aligned += incr) {
 487                 if (*start != val) {
 488                         if (start_phys_aligned == last_bad + incr) {
 489                                 last_bad += incr;
 490                         } else {
 491                                 if (start_bad) {
 492                                         printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 493                                                 val, start_bad, last_bad + incr);
 494                                         reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 495                                 }
 496                                 start_bad = last_bad = start_phys_aligned;
 497                         }
 498                 }
 499         }
 500         if (start_bad) {
 501                 printk(KERN_CONT "\n  %016lx bad mem addr %016lx - %016lx reserved",
 502                         val, start_bad, last_bad + incr);
 503                 reserve_early(start_bad, last_bad - start_bad, "BAD RAM");
 504         }
 505
 506 }
 507
 508 static int memtest_pattern __initdata = CONFIG_MEMTEST_BOOTPARAM_VALUE;
 509
 510 static int __init parse_memtest(char *arg)
 511 {
 512         if (arg)
 513                 memtest_pattern = simple_strtoul(arg, NULL, 0);
 514         return 0;
 515 }
 516
 517 early_param("memtest", parse_memtest);
 518
 519 static void __init early_memtest(unsigned long start, unsigned long end)
 520 {
 521         u64 t_start, t_size;
 522         unsigned pattern;
 523
 524         if (!memtest_pattern)
 525                 return;
 526
 527         printk(KERN_INFO "early_memtest: pattern num %d", memtest_pattern);
 528         for (pattern = 0; pattern < memtest_pattern; pattern++) {
 529                 t_start = start;
 530                 t_size = 0;
 531                 while (t_start < end) {
 532                         t_start = find_e820_area_size(t_start, &t_size, 1);
 533
 534                         /* done ? */
 535                         if (t_start >= end)
 536                                 break;
 537                         if (t_start + t_size > end)
 538                                 t_size = end - t_start;
 539
 540                         printk(KERN_CONT "\n  %016llx - %016llx pattern %d",
 541                                 (unsigned long long)t_start,
 542                                 (unsigned long long)t_start + t_size, pattern);
 543
 544                         memtest(t_start, t_size, pattern);
 545
 546                         t_start += t_size;
 547                 }
 548         }
 549         printk(KERN_CONT "\n");
 550 }
 551 #else
 552 static void __init early_memtest(unsigned long start, unsigned long end)
 553 {
 554 }
 555 #endif
 556
 557 /*
 558  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 559  * This runs before bootmem is initialized and gets pages directly from
 560  * the physical memory. To access them they are temporarily mapped.
 561  */
 562 unsigned long __init_refok init_memory_mapping(unsigned long start, unsigned long end)
 563 {
 564         unsigned long next, last_map_addr = end;
 565         unsigned long start_phys = start, end_phys = end;
 566
 567         printk(KERN_INFO "init_memory_mapping\n");
 568
 569         /*
 570          * Find space for the kernel direct mapping tables.
 571          *
 572          * Later we should allocate these tables in the local node of the
 573          * memory mapped. Unfortunately this is done currently before the
 574          * nodes are discovered.
 575          */
 576         if (!after_bootmem) {
 577                 init_gbpages();
 578                 find_early_table_space(end);
 579         }
 580
 581         start = (unsigned long)__va(start);
 582         end = (unsigned long)__va(end);
 583
 584         for (; start < end; start = next) {
 585                 pgd_t *pgd = pgd_offset_k(start);
 586                 unsigned long pud_phys;
 587                 pud_t *pud;
 588
 589                 if (after_bootmem)
 590                         pud = pud_offset(pgd, start & PGDIR_MASK);
 591                 else
 592                         pud = alloc_low_page(&pud_phys);
 593
 594                 next = start + PGDIR_SIZE;
 595                 if (next > end)
 596                         next = end;
 597                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next));
 598                 if (!after_bootmem)
 599                         set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
 600                 unmap_low_page(pud);
 601         }
 602
 603         if (!after_bootmem)
 604                 mmu_cr4_features = read_cr4();
 605         __flush_tlb_all();
 606
 607         if (!after_bootmem)
 608                 reserve_early(table_start << PAGE_SHIFT,
 609                                  table_end << PAGE_SHIFT, "PGTABLE");
 610
 611         if (!after_bootmem)
 612                 early_memtest(start_phys, end_phys);
 613
 614         return last_map_addr;
 615 }
 616
 617 #ifndef CONFIG_NUMA
 618 void __init paging_init(void)
 619 {
 620         unsigned long max_zone_pfns[MAX_NR_ZONES];
 621
 622         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 623         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 624         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 625         max_zone_pfns[ZONE_NORMAL] = end_pfn;
 626
 627         memory_present(0, 0, end_pfn);
 628         sparse_init();
 629         free_area_init_nodes(max_zone_pfns);
 630 }
 631 #endif
 632
 633 /*
 634  * Memory hotplug specific functions
 635  */
 636 #ifdef CONFIG_MEMORY_HOTPLUG
 637 /*
 638  * Memory is added always to NORMAL zone. This means you will never get
 639  * additional DMA/DMA32 memory.
 640  */
 641 int arch_add_memory(int nid, u64 start, u64 size)
 642 {
 643         struct pglist_data *pgdat = NODE_DATA(nid);
 644         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 645         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 646         unsigned long nr_pages = size >> PAGE_SHIFT;
 647         int ret;
 648
 649         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 650         if (last_mapped_pfn > max_pfn_mapped)
 651                 max_pfn_mapped = last_mapped_pfn;
 652
 653         ret = __add_pages(zone, start_pfn, nr_pages);
 654         WARN_ON(1);
 655
 656         return ret;
 657 }
 658 EXPORT_SYMBOL_GPL(arch_add_memory);
 659
 660 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 661 int memory_add_physaddr_to_nid(u64 start)
 662 {
 663         return 0;
 664 }
 665 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 666 #endif
 667
 668 #endif /* CONFIG_MEMORY_HOTPLUG */
 669
 670 /*
 671  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 672  * is valid. The argument is a physical page number.
 673  *
 674  *
 675  * On x86, access has to be given to the first megabyte of ram because that area
 676  * contains bios code and data regions used by X and dosemu and similar apps.
 677  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 678  * mmio resources as well as potential bios/acpi data regions.
 679  */
 680 int devmem_is_allowed(unsigned long pagenr)
 681 {
 682         if (pagenr <= 256)
 683                 return 1;
 684         if (!page_is_ram(pagenr))
 685                 return 1;
 686         return 0;
 687 }
 688
 689
 690 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 691                          kcore_modules, kcore_vsyscall;
 692
 693 void __init mem_init(void)
 694 {
 695         long codesize, reservedpages, datasize, initsize;
 696
 697         pci_iommu_alloc();
 698
 699         /* clear_bss() already clear the empty_zero_page */
 700
 701         reservedpages = 0;
 702
 703         /* this will put all low memory onto the freelists */
 704 #ifdef CONFIG_NUMA
 705         totalram_pages = numa_free_all_bootmem();
 706 #else
 707         totalram_pages = free_all_bootmem();
 708 #endif
 709         reservedpages = end_pfn - totalram_pages -
 710                                         absent_pages_in_range(0, end_pfn);
 711         after_bootmem = 1;
 712
 713         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 714         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 715         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 716
 717         /* Register memory areas for /proc/kcore */
 718         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 719         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 720                    VMALLOC_END-VMALLOC_START);
 721         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 722         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 723         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 724                                  VSYSCALL_END - VSYSCALL_START);
 725
 726         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 727                                 "%ldk reserved, %ldk data, %ldk init)\n",
 728                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 729                 end_pfn << (PAGE_SHIFT-10),
 730                 codesize >> 10,
 731                 reservedpages << (PAGE_SHIFT-10),
 732                 datasize >> 10,
 733                 initsize >> 10);
 734
 735         cpa_init();
 736 }
 737
 738 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 739 {
 740         unsigned long addr = begin;
 741
 742         if (addr >= end)
 743                 return;
 744
 745         /*
 746          * If debugging page accesses then do not free this memory but
 747          * mark them not present - any buggy init-section access will
 748          * create a kernel page fault:
 749          */
 750 #ifdef CONFIG_DEBUG_PAGEALLOC
 751         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 752                 begin, PAGE_ALIGN(end));
 753         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 754 #else
 755         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 756
 757         for (; addr < end; addr += PAGE_SIZE) {
 758                 ClearPageReserved(virt_to_page(addr));
 759                 init_page_count(virt_to_page(addr));
 760                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 761                         POISON_FREE_INITMEM, PAGE_SIZE);
 762                 free_page(addr);
 763                 totalram_pages++;
 764         }
 765 #endif
 766 }
 767
 768 void free_initmem(void)
 769 {
 770         free_init_pages("unused kernel memory",
 771                         (unsigned long)(&__init_begin),
 772                         (unsigned long)(&__init_end));
 773 }
 774
 775 #ifdef CONFIG_DEBUG_RODATA
 776 const int rodata_test_data = 0xC3;
 777 EXPORT_SYMBOL_GPL(rodata_test_data);
 778
 779 void mark_rodata_ro(void)
 780 {
 781         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 782
 783         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 784                (end - start) >> 10);
 785         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 786
 787         /*
 788          * The rodata section (but not the kernel text!) should also be
 789          * not-executable.
 790          */
 791         start = ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 792         set_memory_nx(start, (end - start) >> PAGE_SHIFT);
 793
 794         rodata_test();
 795
 796 #ifdef CONFIG_CPA_DEBUG
 797         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 798         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 799
 800         printk(KERN_INFO "Testing CPA: again\n");
 801         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 802 #endif
 803 }
 804
 805 #endif
 806
 807 #ifdef CONFIG_BLK_DEV_INITRD
 808 void free_initrd_mem(unsigned long start, unsigned long end)
 809 {
 810         free_init_pages("initrd memory", start, end);
 811 }
 812 #endif
 813
 814 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 815                                    int flags)
 816 {
 817 #ifdef CONFIG_NUMA
 818         int nid, next_nid;
 819 #endif
 820         unsigned long pfn = phys >> PAGE_SHIFT;
 821         int ret;
 822
 823         if (pfn >= end_pfn) {
 824                 /*
 825                  * This can happen with kdump kernels when accessing
 826                  * firmware tables:
 827                  */
 828                 if (pfn < max_pfn_mapped)
 829                         return -EFAULT;
 830
 831                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %u\n",
 832                                 phys, len);
 833                 return -EFAULT;
 834         }
 835
 836         /* Should check here against the e820 map to avoid double free */
 837 #ifdef CONFIG_NUMA
 838         nid = phys_to_nid(phys);
 839         next_nid = phys_to_nid(phys + len - 1);
 840         if (nid == next_nid)
 841                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 842         else
 843                 ret = reserve_bootmem(phys, len, flags);
 844
 845         if (ret != 0)
 846                 return ret;
 847
 848 #else
 849         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
 850 #endif
 851
 852         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 853                 dma_reserve += len / PAGE_SIZE;
 854                 set_dma_reserve(dma_reserve);
 855         }
 856
 857         return 0;
 858 }
 859
 860 int kern_addr_valid(unsigned long addr)
 861 {
 862         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 863         pgd_t *pgd;
 864         pud_t *pud;
 865         pmd_t *pmd;
 866         pte_t *pte;
 867
 868         if (above != 0 && above != -1UL)
 869                 return 0;
 870
 871         pgd = pgd_offset_k(addr);
 872         if (pgd_none(*pgd))
 873                 return 0;
 874
 875         pud = pud_offset(pgd, addr);
 876         if (pud_none(*pud))
 877                 return 0;
 878
 879         pmd = pmd_offset(pud, addr);
 880         if (pmd_none(*pmd))
 881                 return 0;
 882
 883         if (pmd_large(*pmd))
 884                 return pfn_valid(pmd_pfn(*pmd));
 885
 886         pte = pte_offset_kernel(pmd, addr);
 887         if (pte_none(*pte))
 888                 return 0;
 889
 890         return pfn_valid(pte_pfn(*pte));
 891 }
 892
 893 /*
 894  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 895  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 896  * not need special handling anymore:
 897  */
 898 static struct vm_area_struct gate_vma = {
 899         .vm_start       = VSYSCALL_START,
 900         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 901         .vm_page_prot   = PAGE_READONLY_EXEC,
 902         .vm_flags       = VM_READ | VM_EXEC
 903 };
 904
 905 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 906 {
 907 #ifdef CONFIG_IA32_EMULATION
 908         if (test_tsk_thread_flag(tsk, TIF_IA32))
 909                 return NULL;
 910 #endif
 911         return &gate_vma;
 912 }
 913
 914 int in_gate_area(struct task_struct *task, unsigned long addr)
 915 {
 916         struct vm_area_struct *vma = get_gate_vma(task);
 917
 918         if (!vma)
 919                 return 0;
 920
 921         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 922 }
 923
 924 /*
 925  * Use this when you have no reliable task/vma, typically from interrupt
 926  * context. It is less reliable than using the task's vma and may give
 927  * false positives:
 928  */
 929 int in_gate_area_no_task(unsigned long addr)
 930 {
 931         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 932 }
 933
 934 const char *arch_vma_name(struct vm_area_struct *vma)
 935 {
 936         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 937                 return "[vdso]";
 938         if (vma == &gate_vma)
 939                 return "[vsyscall]";
 940         return NULL;
 941 }
 942
 943 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 944 /*
 945  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 946  */
 947 static long __meminitdata addr_start, addr_end;
 948 static void __meminitdata *p_start, *p_end;
 949 static int __meminitdata node_start;
 950
 951 int __meminit
 952 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 953 {
 954         unsigned long addr = (unsigned long)start_page;
 955         unsigned long end = (unsigned long)(start_page + size);
 956         unsigned long next;
 957         pgd_t *pgd;
 958         pud_t *pud;
 959         pmd_t *pmd;
 960
 961         for (; addr < end; addr = next) {
 962                 next = pmd_addr_end(addr, end);
 963
 964                 pgd = vmemmap_pgd_populate(addr, node);
 965                 if (!pgd)
 966                         return -ENOMEM;
 967
 968                 pud = vmemmap_pud_populate(pgd, addr, node);
 969                 if (!pud)
 970                         return -ENOMEM;
 971
 972                 pmd = pmd_offset(pud, addr);
 973                 if (pmd_none(*pmd)) {
 974                         pte_t entry;
 975                         void *p;
 976
 977                         p = vmemmap_alloc_block(PMD_SIZE, node);
 978                         if (!p)
 979                                 return -ENOMEM;
 980
 981                         entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 982                                                         PAGE_KERNEL_LARGE);
 983                         set_pmd(pmd, __pmd(pte_val(entry)));
 984
 985                         /* check to see if we have contiguous blocks */
 986                         if (p_end != p || node_start != node) {
 987                                 if (p_start)
 988                                         printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 989                                                 addr_start, addr_end-1, p_start, p_end-1, node_start);
 990                                 addr_start = addr;
 991                                 node_start = node;
 992                                 p_start = p;
 993                         }
 994                         addr_end = addr + PMD_SIZE;
 995                         p_end = p + PMD_SIZE;
 996                 } else {
 997                         vmemmap_verify((pte_t *)pmd, node, addr, next);
 998                 }
 999         }
1000         return 0;
1001 }
1002
1003 void __meminit vmemmap_populate_print_last(void)
1004 {
1005         if (p_start) {
1006                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1007                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1008                 p_start = NULL;
1009                 p_end = NULL;
1010                 node_start = 0;
1011         }
1012 }
1013 #endif