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/bios_ebda.h>
  35 #include <asm/system.h>
  36 #include <asm/uaccess.h>
  37 #include <asm/pgtable.h>
  38 #include <asm/pgalloc.h>
  39 #include <asm/dma.h>
  40 #include <asm/fixmap.h>
  41 #include <asm/e820.h>
  42 #include <asm/apic.h>
  43 #include <asm/tlb.h>
  44 #include <asm/mmu_context.h>
  45 #include <asm/proto.h>
  46 #include <asm/smp.h>
  47 #include <asm/sections.h>
  48 #include <asm/kdebug.h>
  49 #include <asm/numa.h>
  50 #include <asm/cacheflush.h>
  51
  52 /*
  53  * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries.
  54  * The direct mapping extends to max_pfn_mapped, so that we can directly access
  55  * apertures, ACPI and other tables without having to play with fixmaps.
  56  */
  57 unsigned long max_low_pfn_mapped;
  58 unsigned long max_pfn_mapped;
  59
  60 static unsigned long dma_reserve __initdata;
  61
  62 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  63
  64 int direct_gbpages
  65 #ifdef CONFIG_DIRECT_GBPAGES
  66                                 = 1
  67 #endif
  68 ;
  69
  70 static int __init parse_direct_gbpages_off(char *arg)
  71 {
  72         direct_gbpages = 0;
  73         return 0;
  74 }
  75 early_param("nogbpages", parse_direct_gbpages_off);
  76
  77 static int __init parse_direct_gbpages_on(char *arg)
  78 {
  79         direct_gbpages = 1;
  80         return 0;
  81 }
  82 early_param("gbpages", parse_direct_gbpages_on);
  83
  84 /*
  85  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  86  * physical space so we can cache the place of the first one and move
  87  * around without checking the pgd every time.
  88  */
  89
  90 int after_bootmem;
  91
  92 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
  93 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  94
  95 static int do_not_nx __cpuinitdata;
  96
  97 /*
  98  * noexec=on|off
  99  * Control non-executable mappings for 64-bit processes.
 100  *
 101  * on   Enable (default)
 102  * off  Disable
 103  */
 104 static int __init nonx_setup(char *str)
 105 {
 106         if (!str)
 107                 return -EINVAL;
 108         if (!strncmp(str, "on", 2)) {
 109                 __supported_pte_mask |= _PAGE_NX;
 110                 do_not_nx = 0;
 111         } else if (!strncmp(str, "off", 3)) {
 112                 do_not_nx = 1;
 113                 __supported_pte_mask &= ~_PAGE_NX;
 114         }
 115         return 0;
 116 }
 117 early_param("noexec", nonx_setup);
 118
 119 void __cpuinit check_efer(void)
 120 {
 121         unsigned long efer;
 122
 123         rdmsrl(MSR_EFER, efer);
 124         if (!(efer & EFER_NX) || do_not_nx)
 125                 __supported_pte_mask &= ~_PAGE_NX;
 126 }
 127
 128 int force_personality32;
 129
 130 /*
 131  * noexec32=on|off
 132  * Control non executable heap for 32bit processes.
 133  * To control the stack too use noexec=off
 134  *
 135  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 136  * off  PROT_READ implies PROT_EXEC
 137  */
 138 static int __init nonx32_setup(char *str)
 139 {
 140         if (!strcmp(str, "on"))
 141                 force_personality32 &= ~READ_IMPLIES_EXEC;
 142         else if (!strcmp(str, "off"))
 143                 force_personality32 |= READ_IMPLIES_EXEC;
 144         return 1;
 145 }
 146 __setup("noexec32=", nonx32_setup);
 147
 148 /*
 149  * NOTE: This function is marked __ref because it calls __init function
 150  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 151  */
 152 static __ref void *spp_getpage(void)
 153 {
 154         void *ptr;
 155
 156         if (after_bootmem)
 157                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
 158         else
 159                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 160
 161         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 162                 panic("set_pte_phys: cannot allocate page data %s\n",
 163                         after_bootmem ? "after bootmem" : "");
 164         }
 165
 166         pr_debug("spp_getpage %p\n", ptr);
 167
 168         return ptr;
 169 }
 170
 171 void
 172 set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 173 {
 174         pud_t *pud;
 175         pmd_t *pmd;
 176         pte_t *pte;
 177
 178         pud = pud_page + pud_index(vaddr);
 179         if (pud_none(*pud)) {
 180                 pmd = (pmd_t *) spp_getpage();
 181                 pud_populate(&init_mm, pud, pmd);
 182                 if (pmd != pmd_offset(pud, 0)) {
 183                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 184                                 pmd, pmd_offset(pud, 0));
 185                         return;
 186                 }
 187         }
 188         pmd = pmd_offset(pud, vaddr);
 189         if (pmd_none(*pmd)) {
 190                 pte = (pte_t *) spp_getpage();
 191                 pmd_populate_kernel(&init_mm, pmd, pte);
 192                 if (pte != pte_offset_kernel(pmd, 0)) {
 193                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 194                         return;
 195                 }
 196         }
 197
 198         pte = pte_offset_kernel(pmd, vaddr);
 199         if (!pte_none(*pte) && pte_val(new_pte) &&
 200             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 201                 pte_ERROR(*pte);
 202         set_pte(pte, new_pte);
 203
 204         /*
 205          * It's enough to flush this one mapping.
 206          * (PGE mappings get flushed as well)
 207          */
 208         __flush_tlb_one(vaddr);
 209 }
 210
 211 void
 212 set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 213 {
 214         pgd_t *pgd;
 215         pud_t *pud_page;
 216
 217         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 218
 219         pgd = pgd_offset_k(vaddr);
 220         if (pgd_none(*pgd)) {
 221                 printk(KERN_ERR
 222                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 223                 return;
 224         }
 225         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 226         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 227 }
 228
 229 /*
 230  * Create large page table mappings for a range of physical addresses.
 231  */
 232 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 233                                                 pgprot_t prot)
 234 {
 235         pgd_t *pgd;
 236         pud_t *pud;
 237         pmd_t *pmd;
 238
 239         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 240         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 241                 pgd = pgd_offset_k((unsigned long)__va(phys));
 242                 if (pgd_none(*pgd)) {
 243                         pud = (pud_t *) spp_getpage();
 244                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 245                                                 _PAGE_USER));
 246                 }
 247                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 248                 if (pud_none(*pud)) {
 249                         pmd = (pmd_t *) spp_getpage();
 250                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 251                                                 _PAGE_USER));
 252                 }
 253                 pmd = pmd_offset(pud, phys);
 254                 BUG_ON(!pmd_none(*pmd));
 255                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 256         }
 257 }
 258
 259 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 260 {
 261         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 262 }
 263
 264 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 265 {
 266         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 267 }
 268
 269 /*
 270  * The head.S code sets up the kernel high mapping:
 271  *
 272  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 273  *
 274  * phys_addr holds the negative offset to the kernel, which is added
 275  * to the compile time generated pmds. This results in invalid pmds up
 276  * to the point where we hit the physaddr 0 mapping.
 277  *
 278  * We limit the mappings to the region from _text to _end.  _end is
 279  * rounded up to the 2MB boundary. This catches the invalid pmds as
 280  * well, as they are located before _text:
 281  */
 282 void __init cleanup_highmap(void)
 283 {
 284         unsigned long vaddr = __START_KERNEL_map;
 285         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
 286         pmd_t *pmd = level2_kernel_pgt;
 287         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 288
 289         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 290                 if (pmd_none(*pmd))
 291                         continue;
 292                 if (vaddr < (unsigned long) _text || vaddr > end)
 293                         set_pmd(pmd, __pmd(0));
 294         }
 295 }
 296
 297 static unsigned long __initdata table_start;
 298 static unsigned long __meminitdata table_end;
 299 static unsigned long __meminitdata table_top;
 300
 301 static __ref void *alloc_low_page(unsigned long *phys)
 302 {
 303         unsigned long pfn = table_end++;
 304         void *adr;
 305
 306         if (after_bootmem) {
 307                 adr = (void *)get_zeroed_page(GFP_ATOMIC);
 308                 *phys = __pa(adr);
 309
 310                 return adr;
 311         }
 312
 313         if (pfn >= table_top)
 314                 panic("alloc_low_page: ran out of memory");
 315
 316         adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
 317         memset(adr, 0, PAGE_SIZE);
 318         *phys  = pfn * PAGE_SIZE;
 319         return adr;
 320 }
 321
 322 static __ref void unmap_low_page(void *adr)
 323 {
 324         if (after_bootmem)
 325                 return;
 326
 327         early_iounmap(adr, PAGE_SIZE);
 328 }
 329
 330 static unsigned long __meminit
 331 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 332               pgprot_t prot)
 333 {
 334         unsigned pages = 0;
 335         unsigned long last_map_addr = end;
 336         int i;
 337
 338         pte_t *pte = pte_page + pte_index(addr);
 339
 340         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 341
 342                 if (addr >= end) {
 343                         if (!after_bootmem) {
 344                                 for(; i < PTRS_PER_PTE; i++, pte++)
 345                                         set_pte(pte, __pte(0));
 346                         }
 347                         break;
 348                 }
 349
 350                 /*
 351                  * We will re-use the existing mapping.
 352                  * Xen for example has some special requirements, like mapping
 353                  * pagetable pages as RO. So assume someone who pre-setup
 354                  * these mappings are more intelligent.
 355                  */
 356                 if (pte_val(*pte))
 357                         continue;
 358
 359                 if (0)
 360                         printk("   pte=%p addr=%lx pte=%016lx\n",
 361                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 362                 pages++;
 363                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 364                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 365         }
 366
 367         update_page_count(PG_LEVEL_4K, pages);
 368
 369         return last_map_addr;
 370 }
 371
 372 static unsigned long __meminit
 373 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end,
 374                 pgprot_t prot)
 375 {
 376         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 377
 378         return phys_pte_init(pte, address, end, prot);
 379 }
 380
 381 static unsigned long __meminit
 382 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 383               unsigned long page_size_mask, pgprot_t prot)
 384 {
 385         unsigned long pages = 0;
 386         unsigned long last_map_addr = end;
 387
 388         int i = pmd_index(address);
 389
 390         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 391                 unsigned long pte_phys;
 392                 pmd_t *pmd = pmd_page + pmd_index(address);
 393                 pte_t *pte;
 394                 pgprot_t new_prot = prot;
 395
 396                 if (address >= end) {
 397                         if (!after_bootmem) {
 398                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 399                                         set_pmd(pmd, __pmd(0));
 400                         }
 401                         break;
 402                 }
 403
 404                 if (pmd_val(*pmd)) {
 405                         if (!pmd_large(*pmd)) {
 406                                 spin_lock(&init_mm.page_table_lock);
 407                                 last_map_addr = phys_pte_update(pmd, address,
 408                                                                 end, prot);
 409                                 spin_unlock(&init_mm.page_table_lock);
 410                                 continue;
 411                         }
 412                         /*
 413                          * If we are ok with PG_LEVEL_2M mapping, then we will
 414                          * use the existing mapping,
 415                          *
 416                          * Otherwise, we will split the large page mapping but
 417                          * use the same existing protection bits except for
 418                          * large page, so that we don't violate Intel's TLB
 419                          * Application note (317080) which says, while changing
 420                          * the page sizes, new and old translations should
 421                          * not differ with respect to page frame and
 422                          * attributes.
 423                          */
 424                         if (page_size_mask & (1 << PG_LEVEL_2M))
 425                                 continue;
 426                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 427                 }
 428
 429                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 430                         pages++;
 431                         spin_lock(&init_mm.page_table_lock);
 432                         set_pte((pte_t *)pmd,
 433                                 pfn_pte(address >> PAGE_SHIFT,
 434                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 435                         spin_unlock(&init_mm.page_table_lock);
 436                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 437                         continue;
 438                 }
 439
 440                 pte = alloc_low_page(&pte_phys);
 441                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 442                 unmap_low_page(pte);
 443
 444                 spin_lock(&init_mm.page_table_lock);
 445                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 446                 spin_unlock(&init_mm.page_table_lock);
 447         }
 448         update_page_count(PG_LEVEL_2M, pages);
 449         return last_map_addr;
 450 }
 451
 452 static unsigned long __meminit
 453 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 454                 unsigned long page_size_mask, pgprot_t prot)
 455 {
 456         pmd_t *pmd = pmd_offset(pud, 0);
 457         unsigned long last_map_addr;
 458
 459         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
 460         __flush_tlb_all();
 461         return last_map_addr;
 462 }
 463
 464 static unsigned long __meminit
 465 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 466                          unsigned long page_size_mask)
 467 {
 468         unsigned long pages = 0;
 469         unsigned long last_map_addr = end;
 470         int i = pud_index(addr);
 471
 472         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 473                 unsigned long pmd_phys;
 474                 pud_t *pud = pud_page + pud_index(addr);
 475                 pmd_t *pmd;
 476                 pgprot_t prot = PAGE_KERNEL;
 477
 478                 if (addr >= end)
 479                         break;
 480
 481                 if (!after_bootmem &&
 482                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 483                         set_pud(pud, __pud(0));
 484                         continue;
 485                 }
 486
 487                 if (pud_val(*pud)) {
 488                         if (!pud_large(*pud)) {
 489                                 last_map_addr = phys_pmd_update(pud, addr, end,
 490                                                          page_size_mask, prot);
 491                                 continue;
 492                         }
 493                         /*
 494                          * If we are ok with PG_LEVEL_1G mapping, then we will
 495                          * use the existing mapping.
 496                          *
 497                          * Otherwise, we will split the gbpage mapping but use
 498                          * the same existing protection  bits except for large
 499                          * page, so that we don't violate Intel's TLB
 500                          * Application note (317080) which says, while changing
 501                          * the page sizes, new and old translations should
 502                          * not differ with respect to page frame and
 503                          * attributes.
 504                          */
 505                         if (page_size_mask & (1 << PG_LEVEL_1G))
 506                                 continue;
 507                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 508                 }
 509
 510                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 511                         pages++;
 512                         spin_lock(&init_mm.page_table_lock);
 513                         set_pte((pte_t *)pud,
 514                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 515                         spin_unlock(&init_mm.page_table_lock);
 516                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 517                         continue;
 518                 }
 519
 520                 pmd = alloc_low_page(&pmd_phys);
 521                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 522                                               prot);
 523                 unmap_low_page(pmd);
 524
 525                 spin_lock(&init_mm.page_table_lock);
 526                 pud_populate(&init_mm, pud, __va(pmd_phys));
 527                 spin_unlock(&init_mm.page_table_lock);
 528         }
 529         __flush_tlb_all();
 530
 531         update_page_count(PG_LEVEL_1G, pages);
 532
 533         return last_map_addr;
 534 }
 535
 536 static unsigned long __meminit
 537 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 538                  unsigned long page_size_mask)
 539 {
 540         pud_t *pud;
 541
 542         pud = (pud_t *)pgd_page_vaddr(*pgd);
 543
 544         return phys_pud_init(pud, addr, end, page_size_mask);
 545 }
 546
 547 static void __init find_early_table_space(unsigned long end, int use_pse,
 548                                           int use_gbpages)
 549 {
 550         unsigned long puds, pmds, ptes, tables, start;
 551
 552         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 553         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
 554         if (use_gbpages) {
 555                 unsigned long extra;
 556                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
 557                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
 558         } else
 559                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 560         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
 561
 562         if (use_pse) {
 563                 unsigned long extra;
 564                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
 565                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
 566         } else
 567                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
 568         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
 569
 570         /*
 571          * RED-PEN putting page tables only on node 0 could
 572          * cause a hotspot and fill up ZONE_DMA. The page tables
 573          * need roughly 0.5KB per GB.
 574          */
 575         start = 0x8000;
 576         table_start = find_e820_area(start, end, tables, PAGE_SIZE);
 577         if (table_start == -1UL)
 578                 panic("Cannot find space for the kernel page tables");
 579
 580         table_start >>= PAGE_SHIFT;
 581         table_end = table_start;
 582         table_top = table_start + (tables >> PAGE_SHIFT);
 583
 584         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
 585                 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT);
 586 }
 587
 588 static void __init init_gbpages(void)
 589 {
 590         if (direct_gbpages && cpu_has_gbpages)
 591                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 592         else
 593                 direct_gbpages = 0;
 594 }
 595
 596 static unsigned long __init kernel_physical_mapping_init(unsigned long start,
 597                                                 unsigned long end,
 598                                                 unsigned long page_size_mask)
 599 {
 600
 601         unsigned long next, last_map_addr = end;
 602
 603         start = (unsigned long)__va(start);
 604         end = (unsigned long)__va(end);
 605
 606         for (; start < end; start = next) {
 607                 pgd_t *pgd = pgd_offset_k(start);
 608                 unsigned long pud_phys;
 609                 pud_t *pud;
 610
 611                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 612                 if (next > end)
 613                         next = end;
 614
 615                 if (pgd_val(*pgd)) {
 616                         last_map_addr = phys_pud_update(pgd, __pa(start),
 617                                                  __pa(end), page_size_mask);
 618                         continue;
 619                 }
 620
 621                 pud = alloc_low_page(&pud_phys);
 622                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 623                                                  page_size_mask);
 624                 unmap_low_page(pud);
 625
 626                 spin_lock(&init_mm.page_table_lock);
 627                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 628                 spin_unlock(&init_mm.page_table_lock);
 629         }
 630         __flush_tlb_all();
 631
 632         return last_map_addr;
 633 }
 634
 635 struct map_range {
 636         unsigned long start;
 637         unsigned long end;
 638         unsigned page_size_mask;
 639 };
 640
 641 #define NR_RANGE_MR 5
 642
 643 static int save_mr(struct map_range *mr, int nr_range,
 644                    unsigned long start_pfn, unsigned long end_pfn,
 645                    unsigned long page_size_mask)
 646 {
 647
 648         if (start_pfn < end_pfn) {
 649                 if (nr_range >= NR_RANGE_MR)
 650                         panic("run out of range for init_memory_mapping\n");
 651                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 652                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 653                 mr[nr_range].page_size_mask = page_size_mask;
 654                 nr_range++;
 655         }
 656
 657         return nr_range;
 658 }
 659
 660 /*
 661  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 662  * This runs before bootmem is initialized and gets pages directly from
 663  * the physical memory. To access them they are temporarily mapped.
 664  */
 665 unsigned long __init_refok init_memory_mapping(unsigned long start,
 666                                                unsigned long end)
 667 {
 668         unsigned long last_map_addr = 0;
 669         unsigned long page_size_mask = 0;
 670         unsigned long start_pfn, end_pfn;
 671
 672         struct map_range mr[NR_RANGE_MR];
 673         int nr_range, i;
 674         int use_pse, use_gbpages;
 675
 676         printk(KERN_INFO "init_memory_mapping\n");
 677
 678         /*
 679          * Find space for the kernel direct mapping tables.
 680          *
 681          * Later we should allocate these tables in the local node of the
 682          * memory mapped. Unfortunately this is done currently before the
 683          * nodes are discovered.
 684          */
 685         if (!after_bootmem)
 686                 init_gbpages();
 687
 688 #ifdef CONFIG_DEBUG_PAGEALLOC
 689         /*
 690          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 691          * This will simplify cpa(), which otherwise needs to support splitting
 692          * large pages into small in interrupt context, etc.
 693          */
 694         use_pse = use_gbpages = 0;
 695 #else
 696         use_pse = cpu_has_pse;
 697         use_gbpages = direct_gbpages;
 698 #endif
 699
 700         if (use_gbpages)
 701                 page_size_mask |= 1 << PG_LEVEL_1G;
 702         if (use_pse)
 703                 page_size_mask |= 1 << PG_LEVEL_2M;
 704
 705         memset(mr, 0, sizeof(mr));
 706         nr_range = 0;
 707
 708         /* head if not big page alignment ?*/
 709         start_pfn = start >> PAGE_SHIFT;
 710         end_pfn = ((start + (PMD_SIZE - 1)) >> PMD_SHIFT)
 711                         << (PMD_SHIFT - PAGE_SHIFT);
 712         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 713
 714         /* big page (2M) range*/
 715         start_pfn = ((start + (PMD_SIZE - 1))>>PMD_SHIFT)
 716                          << (PMD_SHIFT - PAGE_SHIFT);
 717         end_pfn = ((start + (PUD_SIZE - 1))>>PUD_SHIFT)
 718                          << (PUD_SHIFT - PAGE_SHIFT);
 719         if (end_pfn > ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT)))
 720                 end_pfn = ((end>>PUD_SHIFT)<<(PUD_SHIFT - PAGE_SHIFT));
 721         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 722                         page_size_mask & (1<<PG_LEVEL_2M));
 723
 724         /* big page (1G) range */
 725         start_pfn = end_pfn;
 726         end_pfn = (end>>PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 727         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 728                                 page_size_mask &
 729                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 730
 731         /* tail is not big page (1G) alignment */
 732         start_pfn = end_pfn;
 733         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 734         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 735                         page_size_mask & (1<<PG_LEVEL_2M));
 736
 737         /* tail is not big page (2M) alignment */
 738         start_pfn = end_pfn;
 739         end_pfn = end>>PAGE_SHIFT;
 740         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 741
 742         /* try to merge same page size and continuous */
 743         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 744                 unsigned long old_start;
 745                 if (mr[i].end != mr[i+1].start ||
 746                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 747                         continue;
 748                 /* move it */
 749                 old_start = mr[i].start;
 750                 memmove(&mr[i], &mr[i+1],
 751                          (nr_range - 1 - i) * sizeof (struct map_range));
 752                 mr[i].start = old_start;
 753                 nr_range--;
 754         }
 755
 756         for (i = 0; i < nr_range; i++)
 757                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 758                                 mr[i].start, mr[i].end,
 759                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 760                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 761
 762         if (!after_bootmem)
 763                 find_early_table_space(end, use_pse, use_gbpages);
 764
 765         for (i = 0; i < nr_range; i++)
 766                 last_map_addr = kernel_physical_mapping_init(
 767                                         mr[i].start, mr[i].end,
 768                                         mr[i].page_size_mask);
 769
 770         if (!after_bootmem)
 771                 mmu_cr4_features = read_cr4();
 772         __flush_tlb_all();
 773
 774         if (!after_bootmem && table_end > table_start)
 775                 reserve_early(table_start << PAGE_SHIFT,
 776                                  table_end << PAGE_SHIFT, "PGTABLE");
 777
 778         printk(KERN_INFO "last_map_addr: %lx end: %lx\n",
 779                          last_map_addr, end);
 780
 781         if (!after_bootmem)
 782                 early_memtest(start, end);
 783
 784         return last_map_addr >> PAGE_SHIFT;
 785 }
 786
 787 #ifndef CONFIG_NUMA
 788 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn)
 789 {
 790         unsigned long bootmap_size, bootmap;
 791
 792         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 793         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 794                                  PAGE_SIZE);
 795         if (bootmap == -1L)
 796                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 797         /* don't touch min_low_pfn */
 798         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 799                                          0, end_pfn);
 800         e820_register_active_regions(0, start_pfn, end_pfn);
 801         free_bootmem_with_active_regions(0, end_pfn);
 802         early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT);
 803         reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT);
 804 }
 805
 806 void __init paging_init(void)
 807 {
 808         unsigned long max_zone_pfns[MAX_NR_ZONES];
 809
 810         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 811         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 812         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 813         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 814
 815         memory_present(0, 0, max_pfn);
 816         sparse_init();
 817         free_area_init_nodes(max_zone_pfns);
 818 }
 819 #endif
 820
 821 /*
 822  * Memory hotplug specific functions
 823  */
 824 #ifdef CONFIG_MEMORY_HOTPLUG
 825 /*
 826  * Memory is added always to NORMAL zone. This means you will never get
 827  * additional DMA/DMA32 memory.
 828  */
 829 int arch_add_memory(int nid, u64 start, u64 size)
 830 {
 831         struct pglist_data *pgdat = NODE_DATA(nid);
 832         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 833         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 834         unsigned long nr_pages = size >> PAGE_SHIFT;
 835         int ret;
 836
 837         last_mapped_pfn = init_memory_mapping(start, start + size-1);
 838         if (last_mapped_pfn > max_pfn_mapped)
 839                 max_pfn_mapped = last_mapped_pfn;
 840
 841         ret = __add_pages(zone, start_pfn, nr_pages);
 842         WARN_ON(1);
 843
 844         return ret;
 845 }
 846 EXPORT_SYMBOL_GPL(arch_add_memory);
 847
 848 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 849 int memory_add_physaddr_to_nid(u64 start)
 850 {
 851         return 0;
 852 }
 853 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 854 #endif
 855
 856 #endif /* CONFIG_MEMORY_HOTPLUG */
 857
 858 /*
 859  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 860  * is valid. The argument is a physical page number.
 861  *
 862  *
 863  * On x86, access has to be given to the first megabyte of ram because that area
 864  * contains bios code and data regions used by X and dosemu and similar apps.
 865  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 866  * mmio resources as well as potential bios/acpi data regions.
 867  */
 868 int devmem_is_allowed(unsigned long pagenr)
 869 {
 870         if (pagenr <= 256)
 871                 return 1;
 872         if (!page_is_ram(pagenr))
 873                 return 1;
 874         return 0;
 875 }
 876
 877
 878 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel,
 879                          kcore_modules, kcore_vsyscall;
 880
 881 void __init mem_init(void)
 882 {
 883         long codesize, reservedpages, datasize, initsize;
 884
 885         start_periodic_check_for_corruption();
 886
 887         pci_iommu_alloc();
 888
 889         /* clear_bss() already clear the empty_zero_page */
 890
 891         reservedpages = 0;
 892
 893         /* this will put all low memory onto the freelists */
 894 #ifdef CONFIG_NUMA
 895         totalram_pages = numa_free_all_bootmem();
 896 #else
 897         totalram_pages = free_all_bootmem();
 898 #endif
 899         reservedpages = max_pfn - totalram_pages -
 900                                         absent_pages_in_range(0, max_pfn);
 901         after_bootmem = 1;
 902
 903         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 904         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 905         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 906
 907         /* Register memory areas for /proc/kcore */
 908         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 909         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 910                    VMALLOC_END-VMALLOC_START);
 911         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 912         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 913         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 914                                  VSYSCALL_END - VSYSCALL_START);
 915
 916         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 917                                 "%ldk reserved, %ldk data, %ldk init)\n",
 918                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 919                 max_pfn << (PAGE_SHIFT-10),
 920                 codesize >> 10,
 921                 reservedpages << (PAGE_SHIFT-10),
 922                 datasize >> 10,
 923                 initsize >> 10);
 924 }
 925
 926 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 927 {
 928         unsigned long addr = begin;
 929
 930         if (addr >= end)
 931                 return;
 932
 933         /*
 934          * If debugging page accesses then do not free this memory but
 935          * mark them not present - any buggy init-section access will
 936          * create a kernel page fault:
 937          */
 938 #ifdef CONFIG_DEBUG_PAGEALLOC
 939         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 940                 begin, PAGE_ALIGN(end));
 941         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 942 #else
 943         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 944
 945         for (; addr < end; addr += PAGE_SIZE) {
 946                 ClearPageReserved(virt_to_page(addr));
 947                 init_page_count(virt_to_page(addr));
 948                 memset((void *)(addr & ~(PAGE_SIZE-1)),
 949                         POISON_FREE_INITMEM, PAGE_SIZE);
 950                 free_page(addr);
 951                 totalram_pages++;
 952         }
 953 #endif
 954 }
 955
 956 void free_initmem(void)
 957 {
 958         free_init_pages("unused kernel memory",
 959                         (unsigned long)(&__init_begin),
 960                         (unsigned long)(&__init_end));
 961 }
 962
 963 #ifdef CONFIG_DEBUG_RODATA
 964 const int rodata_test_data = 0xC3;
 965 EXPORT_SYMBOL_GPL(rodata_test_data);
 966
 967 void mark_rodata_ro(void)
 968 {
 969         unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata);
 970         unsigned long rodata_start =
 971                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 972
 973 #ifdef CONFIG_DYNAMIC_FTRACE
 974         /* Dynamic tracing modifies the kernel text section */
 975         start = rodata_start;
 976 #endif
 977
 978         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 979                (end - start) >> 10);
 980         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 981
 982         /*
 983          * The rodata section (but not the kernel text!) should also be
 984          * not-executable.
 985          */
 986         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 987
 988         rodata_test();
 989
 990 #ifdef CONFIG_CPA_DEBUG
 991         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 992         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 993
 994         printk(KERN_INFO "Testing CPA: again\n");
 995         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 996 #endif
 997 }
 998
 999 #endif
1000
1001 #ifdef CONFIG_BLK_DEV_INITRD
1002 void free_initrd_mem(unsigned long start, unsigned long end)
1003 {
1004         free_init_pages("initrd memory", start, end);
1005 }
1006 #endif
1007
1008 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
1009                                    int flags)
1010 {
1011 #ifdef CONFIG_NUMA
1012         int nid, next_nid;
1013         int ret;
1014 #endif
1015         unsigned long pfn = phys >> PAGE_SHIFT;
1016
1017         if (pfn >= max_pfn) {
1018                 /*
1019                  * This can happen with kdump kernels when accessing
1020                  * firmware tables:
1021                  */
1022                 if (pfn < max_pfn_mapped)
1023                         return -EFAULT;
1024
1025                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
1026                                 phys, len);
1027                 return -EFAULT;
1028         }
1029
1030         /* Should check here against the e820 map to avoid double free */
1031 #ifdef CONFIG_NUMA
1032         nid = phys_to_nid(phys);
1033         next_nid = phys_to_nid(phys + len - 1);
1034         if (nid == next_nid)
1035                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
1036         else
1037                 ret = reserve_bootmem(phys, len, flags);
1038
1039         if (ret != 0)
1040                 return ret;
1041
1042 #else
1043         reserve_bootmem(phys, len, BOOTMEM_DEFAULT);
1044 #endif
1045
1046         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
1047                 dma_reserve += len / PAGE_SIZE;
1048                 set_dma_reserve(dma_reserve);
1049         }
1050
1051         return 0;
1052 }
1053
1054 int kern_addr_valid(unsigned long addr)
1055 {
1056         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
1057         pgd_t *pgd;
1058         pud_t *pud;
1059         pmd_t *pmd;
1060         pte_t *pte;
1061
1062         if (above != 0 && above != -1UL)
1063                 return 0;
1064
1065         pgd = pgd_offset_k(addr);
1066         if (pgd_none(*pgd))
1067                 return 0;
1068
1069         pud = pud_offset(pgd, addr);
1070         if (pud_none(*pud))
1071                 return 0;
1072
1073         pmd = pmd_offset(pud, addr);
1074         if (pmd_none(*pmd))
1075                 return 0;
1076
1077         if (pmd_large(*pmd))
1078                 return pfn_valid(pmd_pfn(*pmd));
1079
1080         pte = pte_offset_kernel(pmd, addr);
1081         if (pte_none(*pte))
1082                 return 0;
1083
1084         return pfn_valid(pte_pfn(*pte));
1085 }
1086
1087 /*
1088  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
1089  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
1090  * not need special handling anymore:
1091  */
1092 static struct vm_area_struct gate_vma = {
1093         .vm_start       = VSYSCALL_START,
1094         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
1095         .vm_page_prot   = PAGE_READONLY_EXEC,
1096         .vm_flags       = VM_READ | VM_EXEC
1097 };
1098
1099 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
1100 {
1101 #ifdef CONFIG_IA32_EMULATION
1102         if (test_tsk_thread_flag(tsk, TIF_IA32))
1103                 return NULL;
1104 #endif
1105         return &gate_vma;
1106 }
1107
1108 int in_gate_area(struct task_struct *task, unsigned long addr)
1109 {
1110         struct vm_area_struct *vma = get_gate_vma(task);
1111
1112         if (!vma)
1113                 return 0;
1114
1115         return (addr >= vma->vm_start) && (addr < vma->vm_end);
1116 }
1117
1118 /*
1119  * Use this when you have no reliable task/vma, typically from interrupt
1120  * context. It is less reliable than using the task's vma and may give
1121  * false positives:
1122  */
1123 int in_gate_area_no_task(unsigned long addr)
1124 {
1125         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
1126 }
1127
1128 const char *arch_vma_name(struct vm_area_struct *vma)
1129 {
1130         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
1131                 return "[vdso]";
1132         if (vma == &gate_vma)
1133                 return "[vsyscall]";
1134         return NULL;
1135 }
1136
1137 #ifdef CONFIG_SPARSEMEM_VMEMMAP
1138 /*
1139  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
1140  */
1141 static long __meminitdata addr_start, addr_end;
1142 static void __meminitdata *p_start, *p_end;
1143 static int __meminitdata node_start;
1144
1145 int __meminit
1146 vmemmap_populate(struct page *start_page, unsigned long size, int node)
1147 {
1148         unsigned long addr = (unsigned long)start_page;
1149         unsigned long end = (unsigned long)(start_page + size);
1150         unsigned long next;
1151         pgd_t *pgd;
1152         pud_t *pud;
1153         pmd_t *pmd;
1154
1155         for (; addr < end; addr = next) {
1156                 void *p = NULL;
1157
1158                 pgd = vmemmap_pgd_populate(addr, node);
1159                 if (!pgd)
1160                         return -ENOMEM;
1161
1162                 pud = vmemmap_pud_populate(pgd, addr, node);
1163                 if (!pud)
1164                         return -ENOMEM;
1165
1166                 if (!cpu_has_pse) {
1167                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1168                         pmd = vmemmap_pmd_populate(pud, addr, node);
1169
1170                         if (!pmd)
1171                                 return -ENOMEM;
1172
1173                         p = vmemmap_pte_populate(pmd, addr, node);
1174
1175                         if (!p)
1176                                 return -ENOMEM;
1177
1178                         addr_end = addr + PAGE_SIZE;
1179                         p_end = p + PAGE_SIZE;
1180                 } else {
1181                         next = pmd_addr_end(addr, end);
1182
1183                         pmd = pmd_offset(pud, addr);
1184                         if (pmd_none(*pmd)) {
1185                                 pte_t entry;
1186
1187                                 p = vmemmap_alloc_block(PMD_SIZE, node);
1188                                 if (!p)
1189                                         return -ENOMEM;
1190
1191                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1192                                                 PAGE_KERNEL_LARGE);
1193                                 set_pmd(pmd, __pmd(pte_val(entry)));
1194
1195                                 /* check to see if we have contiguous blocks */
1196                                 if (p_end != p || node_start != node) {
1197                                         if (p_start)
1198                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1199                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1200                                         addr_start = addr;
1201                                         node_start = node;
1202                                         p_start = p;
1203                                 }
1204
1205                                 addr_end = addr + PMD_SIZE;
1206                                 p_end = p + PMD_SIZE;
1207                         } else
1208                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1209                 }
1210
1211         }
1212         return 0;
1213 }
1214
1215 void __meminit vmemmap_populate_print_last(void)
1216 {
1217         if (p_start) {
1218                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1219                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1220                 p_start = NULL;
1221                 p_end = NULL;
1222                 node_start = 0;
1223         }
1224 }
1225 #endif