arch/x86/mm/fault.c

   1 /*
   2  *  Copyright (C) 1995  Linus Torvalds
   3  *  Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
   4  *  Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
   5  */
   6 #include <linux/magic.h>                /* STACK_END_MAGIC              */
   7 #include <linux/sched.h>                /* test_thread_flag(), ...      */
   8 #include <linux/kdebug.h>               /* oops_begin/end, ...          */
   9 #include <linux/module.h>               /* search_exception_table       */
  10 #include <linux/bootmem.h>              /* max_low_pfn                  */
  11 #include <linux/kprobes.h>              /* __kprobes, ...               */
  12 #include <linux/mmiotrace.h>            /* kmmio_handler, ...           */
  13 #include <linux/perf_counter.h>         /* perf_swcounter_event         */
  14
  15 #include <asm/traps.h>                  /* dotraplinkage, ...           */
  16 #include <asm/pgalloc.h>                /* pgd_*(), ...                 */
  17
  18 /*
  19  * Page fault error code bits:
  20  *
  21  *   bit 0 ==    0: no page found       1: protection fault
  22  *   bit 1 ==    0: read access         1: write access
  23  *   bit 2 ==    0: kernel-mode access  1: user-mode access
  24  *   bit 3 ==                           1: use of reserved bit detected
  25  *   bit 4 ==                           1: fault was an instruction fetch
  26  */
  27 enum x86_pf_error_code {
  28
  29         PF_PROT         =               1 << 0,
  30         PF_WRITE        =               1 << 1,
  31         PF_USER         =               1 << 2,
  32         PF_RSVD         =               1 << 3,
  33         PF_INSTR        =               1 << 4,
  34 };
  35
  36 /*
  37  * Returns 0 if mmiotrace is disabled, or if the fault is not
  38  * handled by mmiotrace:
  39  */
  40 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
  41 {
  42         if (unlikely(is_kmmio_active()))
  43                 if (kmmio_handler(regs, addr) == 1)
  44                         return -1;
  45         return 0;
  46 }
  47
  48 static inline int notify_page_fault(struct pt_regs *regs)
  49 {
  50         int ret = 0;
  51
  52         /* kprobe_running() needs smp_processor_id() */
  53         if (kprobes_built_in() && !user_mode_vm(regs)) {
  54                 preempt_disable();
  55                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  56                         ret = 1;
  57                 preempt_enable();
  58         }
  59
  60         return ret;
  61 }
  62
  63 /*
  64  * Prefetch quirks:
  65  *
  66  * 32-bit mode:
  67  *
  68  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  69  *   Check that here and ignore it.
  70  *
  71  * 64-bit mode:
  72  *
  73  *   Sometimes the CPU reports invalid exceptions on prefetch.
  74  *   Check that here and ignore it.
  75  *
  76  * Opcode checker based on code by Richard Brunner.
  77  */
  78 static inline int
  79 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  80                       unsigned char opcode, int *prefetch)
  81 {
  82         unsigned char instr_hi = opcode & 0xf0;
  83         unsigned char instr_lo = opcode & 0x0f;
  84
  85         switch (instr_hi) {
  86         case 0x20:
  87         case 0x30:
  88                 /*
  89                  * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
  90                  * In X86_64 long mode, the CPU will signal invalid
  91                  * opcode if some of these prefixes are present so
  92                  * X86_64 will never get here anyway
  93                  */
  94                 return ((instr_lo & 7) == 0x6);
  95 #ifdef CONFIG_X86_64
  96         case 0x40:
  97                 /*
  98                  * In AMD64 long mode 0x40..0x4F are valid REX prefixes
  99                  * Need to figure out under what instruction mode the
 100                  * instruction was issued. Could check the LDT for lm,
 101                  * but for now it's good enough to assume that long
 102                  * mode only uses well known segments or kernel.
 103                  */
 104                 return (!user_mode(regs)) || (regs->cs == __USER_CS);
 105 #endif
 106         case 0x60:
 107                 /* 0x64 thru 0x67 are valid prefixes in all modes. */
 108                 return (instr_lo & 0xC) == 0x4;
 109         case 0xF0:
 110                 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 111                 return !instr_lo || (instr_lo>>1) == 1;
 112         case 0x00:
 113                 /* Prefetch instruction is 0x0F0D or 0x0F18 */
 114                 if (probe_kernel_address(instr, opcode))
 115                         return 0;
 116
 117                 *prefetch = (instr_lo == 0xF) &&
 118                         (opcode == 0x0D || opcode == 0x18);
 119                 return 0;
 120         default:
 121                 return 0;
 122         }
 123 }
 124
 125 static int
 126 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 127 {
 128         unsigned char *max_instr;
 129         unsigned char *instr;
 130         int prefetch = 0;
 131
 132         /*
 133          * If it was a exec (instruction fetch) fault on NX page, then
 134          * do not ignore the fault:
 135          */
 136         if (error_code & PF_INSTR)
 137                 return 0;
 138
 139         instr = (void *)convert_ip_to_linear(current, regs);
 140         max_instr = instr + 15;
 141
 142         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 143                 return 0;
 144
 145         while (instr < max_instr) {
 146                 unsigned char opcode;
 147
 148                 if (probe_kernel_address(instr, opcode))
 149                         break;
 150
 151                 instr++;
 152
 153                 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 154                         break;
 155         }
 156         return prefetch;
 157 }
 158
 159 static void
 160 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
 161                      struct task_struct *tsk)
 162 {
 163         siginfo_t info;
 164
 165         info.si_signo   = si_signo;
 166         info.si_errno   = 0;
 167         info.si_code    = si_code;
 168         info.si_addr    = (void __user *)address;
 169
 170         force_sig_info(si_signo, &info, tsk);
 171 }
 172
 173 DEFINE_SPINLOCK(pgd_lock);
 174 LIST_HEAD(pgd_list);
 175
 176 #ifdef CONFIG_X86_32
 177 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 178 {
 179         unsigned index = pgd_index(address);
 180         pgd_t *pgd_k;
 181         pud_t *pud, *pud_k;
 182         pmd_t *pmd, *pmd_k;
 183
 184         pgd += index;
 185         pgd_k = init_mm.pgd + index;
 186
 187         if (!pgd_present(*pgd_k))
 188                 return NULL;
 189
 190         /*
 191          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 192          * and redundant with the set_pmd() on non-PAE. As would
 193          * set_pud.
 194          */
 195         pud = pud_offset(pgd, address);
 196         pud_k = pud_offset(pgd_k, address);
 197         if (!pud_present(*pud_k))
 198                 return NULL;
 199
 200         pmd = pmd_offset(pud, address);
 201         pmd_k = pmd_offset(pud_k, address);
 202         if (!pmd_present(*pmd_k))
 203                 return NULL;
 204
 205         if (!pmd_present(*pmd))
 206                 set_pmd(pmd, *pmd_k);
 207         else
 208                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 209
 210         return pmd_k;
 211 }
 212
 213 void vmalloc_sync_all(void)
 214 {
 215         unsigned long address;
 216
 217         if (SHARED_KERNEL_PMD)
 218                 return;
 219
 220         for (address = VMALLOC_START & PMD_MASK;
 221              address >= TASK_SIZE && address < FIXADDR_TOP;
 222              address += PMD_SIZE) {
 223
 224                 unsigned long flags;
 225                 struct page *page;
 226
 227                 spin_lock_irqsave(&pgd_lock, flags);
 228                 list_for_each_entry(page, &pgd_list, lru) {
 229                         if (!vmalloc_sync_one(page_address(page), address))
 230                                 break;
 231                 }
 232                 spin_unlock_irqrestore(&pgd_lock, flags);
 233         }
 234 }
 235
 236 /*
 237  * 32-bit:
 238  *
 239  *   Handle a fault on the vmalloc or module mapping area
 240  */
 241 static noinline int vmalloc_fault(unsigned long address)
 242 {
 243         unsigned long pgd_paddr;
 244         pmd_t *pmd_k;
 245         pte_t *pte_k;
 246
 247         /* Make sure we are in vmalloc area: */
 248         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 249                 return -1;
 250
 251         /*
 252          * Synchronize this task's top level page-table
 253          * with the 'reference' page table.
 254          *
 255          * Do _not_ use "current" here. We might be inside
 256          * an interrupt in the middle of a task switch..
 257          */
 258         pgd_paddr = read_cr3();
 259         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 260         if (!pmd_k)
 261                 return -1;
 262
 263         pte_k = pte_offset_kernel(pmd_k, address);
 264         if (!pte_present(*pte_k))
 265                 return -1;
 266
 267         return 0;
 268 }
 269
 270 /*
 271  * Did it hit the DOS screen memory VA from vm86 mode?
 272  */
 273 static inline void
 274 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 275                  struct task_struct *tsk)
 276 {
 277         unsigned long bit;
 278
 279         if (!v8086_mode(regs))
 280                 return;
 281
 282         bit = (address - 0xA0000) >> PAGE_SHIFT;
 283         if (bit < 32)
 284                 tsk->thread.screen_bitmap |= 1 << bit;
 285 }
 286
 287 static void dump_pagetable(unsigned long address)
 288 {
 289         __typeof__(pte_val(__pte(0))) page;
 290
 291         page = read_cr3();
 292         page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 293
 294 #ifdef CONFIG_X86_PAE
 295         printk("*pdpt = %016Lx ", page);
 296         if ((page >> PAGE_SHIFT) < max_low_pfn
 297             && page & _PAGE_PRESENT) {
 298                 page &= PAGE_MASK;
 299                 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 300                                                         & (PTRS_PER_PMD - 1)];
 301                 printk(KERN_CONT "*pde = %016Lx ", page);
 302                 page &= ~_PAGE_NX;
 303         }
 304 #else
 305         printk("*pde = %08lx ", page);
 306 #endif
 307
 308         /*
 309          * We must not directly access the pte in the highpte
 310          * case if the page table is located in highmem.
 311          * And let's rather not kmap-atomic the pte, just in case
 312          * it's allocated already:
 313          */
 314         if ((page >> PAGE_SHIFT) < max_low_pfn
 315             && (page & _PAGE_PRESENT)
 316             && !(page & _PAGE_PSE)) {
 317
 318                 page &= PAGE_MASK;
 319                 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 320                                                         & (PTRS_PER_PTE - 1)];
 321                 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 322         }
 323
 324         printk("\n");
 325 }
 326
 327 #else /* CONFIG_X86_64: */
 328
 329 void vmalloc_sync_all(void)
 330 {
 331         unsigned long address;
 332
 333         for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
 334              address += PGDIR_SIZE) {
 335
 336                 const pgd_t *pgd_ref = pgd_offset_k(address);
 337                 unsigned long flags;
 338                 struct page *page;
 339
 340                 if (pgd_none(*pgd_ref))
 341                         continue;
 342
 343                 spin_lock_irqsave(&pgd_lock, flags);
 344                 list_for_each_entry(page, &pgd_list, lru) {
 345                         pgd_t *pgd;
 346                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 347                         if (pgd_none(*pgd))
 348                                 set_pgd(pgd, *pgd_ref);
 349                         else
 350                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 351                 }
 352                 spin_unlock_irqrestore(&pgd_lock, flags);
 353         }
 354 }
 355
 356 /*
 357  * 64-bit:
 358  *
 359  *   Handle a fault on the vmalloc area
 360  *
 361  * This assumes no large pages in there.
 362  */
 363 static noinline int vmalloc_fault(unsigned long address)
 364 {
 365         pgd_t *pgd, *pgd_ref;
 366         pud_t *pud, *pud_ref;
 367         pmd_t *pmd, *pmd_ref;
 368         pte_t *pte, *pte_ref;
 369
 370         /* Make sure we are in vmalloc area: */
 371         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 372                 return -1;
 373
 374         /*
 375          * Copy kernel mappings over when needed. This can also
 376          * happen within a race in page table update. In the later
 377          * case just flush:
 378          */
 379         pgd = pgd_offset(current->active_mm, address);
 380         pgd_ref = pgd_offset_k(address);
 381         if (pgd_none(*pgd_ref))
 382                 return -1;
 383
 384         if (pgd_none(*pgd))
 385                 set_pgd(pgd, *pgd_ref);
 386         else
 387                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 388
 389         /*
 390          * Below here mismatches are bugs because these lower tables
 391          * are shared:
 392          */
 393
 394         pud = pud_offset(pgd, address);
 395         pud_ref = pud_offset(pgd_ref, address);
 396         if (pud_none(*pud_ref))
 397                 return -1;
 398
 399         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 400                 BUG();
 401
 402         pmd = pmd_offset(pud, address);
 403         pmd_ref = pmd_offset(pud_ref, address);
 404         if (pmd_none(*pmd_ref))
 405                 return -1;
 406
 407         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 408                 BUG();
 409
 410         pte_ref = pte_offset_kernel(pmd_ref, address);
 411         if (!pte_present(*pte_ref))
 412                 return -1;
 413
 414         pte = pte_offset_kernel(pmd, address);
 415
 416         /*
 417          * Don't use pte_page here, because the mappings can point
 418          * outside mem_map, and the NUMA hash lookup cannot handle
 419          * that:
 420          */
 421         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 422                 BUG();
 423
 424         return 0;
 425 }
 426
 427 static const char errata93_warning[] =
 428 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 429 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 430 KERN_ERR "******* Please consider a BIOS update.\n"
 431 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 432
 433 /*
 434  * No vm86 mode in 64-bit mode:
 435  */
 436 static inline void
 437 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 438                  struct task_struct *tsk)
 439 {
 440 }
 441
 442 static int bad_address(void *p)
 443 {
 444         unsigned long dummy;
 445
 446         return probe_kernel_address((unsigned long *)p, dummy);
 447 }
 448
 449 static void dump_pagetable(unsigned long address)
 450 {
 451         pgd_t *pgd;
 452         pud_t *pud;
 453         pmd_t *pmd;
 454         pte_t *pte;
 455
 456         pgd = (pgd_t *)read_cr3();
 457
 458         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 459
 460         pgd += pgd_index(address);
 461         if (bad_address(pgd))
 462                 goto bad;
 463
 464         printk("PGD %lx ", pgd_val(*pgd));
 465
 466         if (!pgd_present(*pgd))
 467                 goto out;
 468
 469         pud = pud_offset(pgd, address);
 470         if (bad_address(pud))
 471                 goto bad;
 472
 473         printk("PUD %lx ", pud_val(*pud));
 474         if (!pud_present(*pud) || pud_large(*pud))
 475                 goto out;
 476
 477         pmd = pmd_offset(pud, address);
 478         if (bad_address(pmd))
 479                 goto bad;
 480
 481         printk("PMD %lx ", pmd_val(*pmd));
 482         if (!pmd_present(*pmd) || pmd_large(*pmd))
 483                 goto out;
 484
 485         pte = pte_offset_kernel(pmd, address);
 486         if (bad_address(pte))
 487                 goto bad;
 488
 489         printk("PTE %lx", pte_val(*pte));
 490 out:
 491         printk("\n");
 492         return;
 493 bad:
 494         printk("BAD\n");
 495 }
 496
 497 #endif /* CONFIG_X86_64 */
 498
 499 /*
 500  * Workaround for K8 erratum #93 & buggy BIOS.
 501  *
 502  * BIOS SMM functions are required to use a specific workaround
 503  * to avoid corruption of the 64bit RIP register on C stepping K8.
 504  *
 505  * A lot of BIOS that didn't get tested properly miss this.
 506  *
 507  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 508  * Try to work around it here.
 509  *
 510  * Note we only handle faults in kernel here.
 511  * Does nothing on 32-bit.
 512  */
 513 static int is_errata93(struct pt_regs *regs, unsigned long address)
 514 {
 515 #ifdef CONFIG_X86_64
 516         if (address != regs->ip)
 517                 return 0;
 518
 519         if ((address >> 32) != 0)
 520                 return 0;
 521
 522         address |= 0xffffffffUL << 32;
 523         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 524             (address >= MODULES_VADDR && address <= MODULES_END)) {
 525                 printk_once(errata93_warning);
 526                 regs->ip = address;
 527                 return 1;
 528         }
 529 #endif
 530         return 0;
 531 }
 532
 533 /*
 534  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 535  * to illegal addresses >4GB.
 536  *
 537  * We catch this in the page fault handler because these addresses
 538  * are not reachable. Just detect this case and return.  Any code
 539  * segment in LDT is compatibility mode.
 540  */
 541 static int is_errata100(struct pt_regs *regs, unsigned long address)
 542 {
 543 #ifdef CONFIG_X86_64
 544         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 545                 return 1;
 546 #endif
 547         return 0;
 548 }
 549
 550 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 551 {
 552 #ifdef CONFIG_X86_F00F_BUG
 553         unsigned long nr;
 554
 555         /*
 556          * Pentium F0 0F C7 C8 bug workaround:
 557          */
 558         if (boot_cpu_data.f00f_bug) {
 559                 nr = (address - idt_descr.address) >> 3;
 560
 561                 if (nr == 6) {
 562                         do_invalid_op(regs, 0);
 563                         return 1;
 564                 }
 565         }
 566 #endif
 567         return 0;
 568 }
 569
 570 static const char nx_warning[] = KERN_CRIT
 571 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
 572
 573 static void
 574 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 575                 unsigned long address)
 576 {
 577         if (!oops_may_print())
 578                 return;
 579
 580         if (error_code & PF_INSTR) {
 581                 unsigned int level;
 582
 583                 pte_t *pte = lookup_address(address, &level);
 584
 585                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 586                         printk(nx_warning, current_uid());
 587         }
 588
 589         printk(KERN_ALERT "BUG: unable to handle kernel ");
 590         if (address < PAGE_SIZE)
 591                 printk(KERN_CONT "NULL pointer dereference");
 592         else
 593                 printk(KERN_CONT "paging request");
 594
 595         printk(KERN_CONT " at %p\n", (void *) address);
 596         printk(KERN_ALERT "IP:");
 597         printk_address(regs->ip, 1);
 598
 599         dump_pagetable(address);
 600 }
 601
 602 static noinline void
 603 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 604             unsigned long address)
 605 {
 606         struct task_struct *tsk;
 607         unsigned long flags;
 608         int sig;
 609
 610         flags = oops_begin();
 611         tsk = current;
 612         sig = SIGKILL;
 613
 614         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 615                tsk->comm, address);
 616         dump_pagetable(address);
 617
 618         tsk->thread.cr2         = address;
 619         tsk->thread.trap_no     = 14;
 620         tsk->thread.error_code  = error_code;
 621
 622         if (__die("Bad pagetable", regs, error_code))
 623                 sig = 0;
 624
 625         oops_end(flags, regs, sig);
 626 }
 627
 628 static noinline void
 629 no_context(struct pt_regs *regs, unsigned long error_code,
 630            unsigned long address)
 631 {
 632         struct task_struct *tsk = current;
 633         unsigned long *stackend;
 634         unsigned long flags;
 635         int sig;
 636
 637         /* Are we prepared to handle this kernel fault? */
 638         if (fixup_exception(regs))
 639                 return;
 640
 641         /*
 642          * 32-bit:
 643          *
 644          *   Valid to do another page fault here, because if this fault
 645          *   had been triggered by is_prefetch fixup_exception would have
 646          *   handled it.
 647          *
 648          * 64-bit:
 649          *
 650          *   Hall of shame of CPU/BIOS bugs.
 651          */
 652         if (is_prefetch(regs, error_code, address))
 653                 return;
 654
 655         if (is_errata93(regs, address))
 656                 return;
 657
 658         /*
 659          * Oops. The kernel tried to access some bad page. We'll have to
 660          * terminate things with extreme prejudice:
 661          */
 662         flags = oops_begin();
 663
 664         show_fault_oops(regs, error_code, address);
 665
 666         stackend = end_of_stack(tsk);
 667         if (*stackend != STACK_END_MAGIC)
 668                 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
 669
 670         tsk->thread.cr2         = address;
 671         tsk->thread.trap_no     = 14;
 672         tsk->thread.error_code  = error_code;
 673
 674         sig = SIGKILL;
 675         if (__die("Oops", regs, error_code))
 676                 sig = 0;
 677
 678         /* Executive summary in case the body of the oops scrolled away */
 679         printk(KERN_EMERG "CR2: %016lx\n", address);
 680
 681         oops_end(flags, regs, sig);
 682 }
 683
 684 /*
 685  * Print out info about fatal segfaults, if the show_unhandled_signals
 686  * sysctl is set:
 687  */
 688 static inline void
 689 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 690                 unsigned long address, struct task_struct *tsk)
 691 {
 692         if (!unhandled_signal(tsk, SIGSEGV))
 693                 return;
 694
 695         if (!printk_ratelimit())
 696                 return;
 697
 698         printk(KERN_CONT "%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 699                 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 700                 tsk->comm, task_pid_nr(tsk), address,
 701                 (void *)regs->ip, (void *)regs->sp, error_code);
 702
 703         print_vma_addr(KERN_CONT " in ", regs->ip);
 704
 705         printk(KERN_CONT "\n");
 706 }
 707
 708 static void
 709 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 710                        unsigned long address, int si_code)
 711 {
 712         struct task_struct *tsk = current;
 713
 714         /* User mode accesses just cause a SIGSEGV */
 715         if (error_code & PF_USER) {
 716                 /*
 717                  * It's possible to have interrupts off here:
 718                  */
 719                 local_irq_enable();
 720
 721                 /*
 722                  * Valid to do another page fault here because this one came
 723                  * from user space:
 724                  */
 725                 if (is_prefetch(regs, error_code, address))
 726                         return;
 727
 728                 if (is_errata100(regs, address))
 729                         return;
 730
 731                 if (unlikely(show_unhandled_signals))
 732                         show_signal_msg(regs, error_code, address, tsk);
 733
 734                 /* Kernel addresses are always protection faults: */
 735                 tsk->thread.cr2         = address;
 736                 tsk->thread.error_code  = error_code | (address >= TASK_SIZE);
 737                 tsk->thread.trap_no     = 14;
 738
 739                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 740
 741                 return;
 742         }
 743
 744         if (is_f00f_bug(regs, address))
 745                 return;
 746
 747         no_context(regs, error_code, address);
 748 }
 749
 750 static noinline void
 751 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 752                      unsigned long address)
 753 {
 754         __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 755 }
 756
 757 static void
 758 __bad_area(struct pt_regs *regs, unsigned long error_code,
 759            unsigned long address, int si_code)
 760 {
 761         struct mm_struct *mm = current->mm;
 762
 763         /*
 764          * Something tried to access memory that isn't in our memory map..
 765          * Fix it, but check if it's kernel or user first..
 766          */
 767         up_read(&mm->mmap_sem);
 768
 769         __bad_area_nosemaphore(regs, error_code, address, si_code);
 770 }
 771
 772 static noinline void
 773 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 774 {
 775         __bad_area(regs, error_code, address, SEGV_MAPERR);
 776 }
 777
 778 static noinline void
 779 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 780                       unsigned long address)
 781 {
 782         __bad_area(regs, error_code, address, SEGV_ACCERR);
 783 }
 784
 785 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
 786 static void
 787 out_of_memory(struct pt_regs *regs, unsigned long error_code,
 788               unsigned long address)
 789 {
 790         /*
 791          * We ran out of memory, call the OOM killer, and return the userspace
 792          * (which will retry the fault, or kill us if we got oom-killed):
 793          */
 794         up_read(&current->mm->mmap_sem);
 795
 796         pagefault_out_of_memory();
 797 }
 798
 799 static void
 800 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 801 {
 802         struct task_struct *tsk = current;
 803         struct mm_struct *mm = tsk->mm;
 804
 805         up_read(&mm->mmap_sem);
 806
 807         /* Kernel mode? Handle exceptions or die: */
 808         if (!(error_code & PF_USER))
 809                 no_context(regs, error_code, address);
 810
 811         /* User-space => ok to do another page fault: */
 812         if (is_prefetch(regs, error_code, address))
 813                 return;
 814
 815         tsk->thread.cr2         = address;
 816         tsk->thread.error_code  = error_code;
 817         tsk->thread.trap_no     = 14;
 818
 819         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 820 }
 821
 822 static noinline void
 823 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 824                unsigned long address, unsigned int fault)
 825 {
 826         if (fault & VM_FAULT_OOM) {
 827                 out_of_memory(regs, error_code, address);
 828         } else {
 829                 if (fault & VM_FAULT_SIGBUS)
 830                         do_sigbus(regs, error_code, address);
 831                 else
 832                         BUG();
 833         }
 834 }
 835
 836 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 837 {
 838         if ((error_code & PF_WRITE) && !pte_write(*pte))
 839                 return 0;
 840
 841         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 842                 return 0;
 843
 844         return 1;
 845 }
 846
 847 /*
 848  * Handle a spurious fault caused by a stale TLB entry.
 849  *
 850  * This allows us to lazily refresh the TLB when increasing the
 851  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
 852  * eagerly is very expensive since that implies doing a full
 853  * cross-processor TLB flush, even if no stale TLB entries exist
 854  * on other processors.
 855  *
 856  * There are no security implications to leaving a stale TLB when
 857  * increasing the permissions on a page.
 858  */
 859 static noinline int
 860 spurious_fault(unsigned long error_code, unsigned long address)
 861 {
 862         pgd_t *pgd;
 863         pud_t *pud;
 864         pmd_t *pmd;
 865         pte_t *pte;
 866         int ret;
 867
 868         /* Reserved-bit violation or user access to kernel space? */
 869         if (error_code & (PF_USER | PF_RSVD))
 870                 return 0;
 871
 872         pgd = init_mm.pgd + pgd_index(address);
 873         if (!pgd_present(*pgd))
 874                 return 0;
 875
 876         pud = pud_offset(pgd, address);
 877         if (!pud_present(*pud))
 878                 return 0;
 879
 880         if (pud_large(*pud))
 881                 return spurious_fault_check(error_code, (pte_t *) pud);
 882
 883         pmd = pmd_offset(pud, address);
 884         if (!pmd_present(*pmd))
 885                 return 0;
 886
 887         if (pmd_large(*pmd))
 888                 return spurious_fault_check(error_code, (pte_t *) pmd);
 889
 890         pte = pte_offset_kernel(pmd, address);
 891         if (!pte_present(*pte))
 892                 return 0;
 893
 894         ret = spurious_fault_check(error_code, pte);
 895         if (!ret)
 896                 return 0;
 897
 898         /*
 899          * Make sure we have permissions in PMD.
 900          * If not, then there's a bug in the page tables:
 901          */
 902         ret = spurious_fault_check(error_code, (pte_t *) pmd);
 903         WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
 904
 905         return ret;
 906 }
 907
 908 int show_unhandled_signals = 1;
 909
 910 static inline int
 911 access_error(unsigned long error_code, int write, struct vm_area_struct *vma)
 912 {
 913         if (write) {
 914                 /* write, present and write, not present: */
 915                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 916                         return 1;
 917                 return 0;
 918         }
 919
 920         /* read, present: */
 921         if (unlikely(error_code & PF_PROT))
 922                 return 1;
 923
 924         /* read, not present: */
 925         if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 926                 return 1;
 927
 928         return 0;
 929 }
 930
 931 static int fault_in_kernel_space(unsigned long address)
 932 {
 933         return address >= TASK_SIZE_MAX;
 934 }
 935
 936 /*
 937  * This routine handles page faults.  It determines the address,
 938  * and the problem, and then passes it off to one of the appropriate
 939  * routines.
 940  */
 941 dotraplinkage void __kprobes
 942 do_page_fault(struct pt_regs *regs, unsigned long error_code)
 943 {
 944         struct vm_area_struct *vma;
 945         struct task_struct *tsk;
 946         unsigned long address;
 947         struct mm_struct *mm;
 948         int write;
 949         int fault;
 950
 951         tsk = current;
 952         mm = tsk->mm;
 953
 954         prefetchw(&mm->mmap_sem);
 955
 956         /* Get the faulting address: */
 957         address = read_cr2();
 958
 959         if (unlikely(kmmio_fault(regs, address)))
 960                 return;
 961
 962         /*
 963          * We fault-in kernel-space virtual memory on-demand. The
 964          * 'reference' page table is init_mm.pgd.
 965          *
 966          * NOTE! We MUST NOT take any locks for this case. We may
 967          * be in an interrupt or a critical region, and should
 968          * only copy the information from the master page table,
 969          * nothing more.
 970          *
 971          * This verifies that the fault happens in kernel space
 972          * (error_code & 4) == 0, and that the fault was not a
 973          * protection error (error_code & 9) == 0.
 974          */
 975         if (unlikely(fault_in_kernel_space(address))) {
 976                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 977                     vmalloc_fault(address) >= 0)
 978                         return;
 979
 980                 /* Can handle a stale RO->RW TLB: */
 981                 if (spurious_fault(error_code, address))
 982                         return;
 983
 984                 /* kprobes don't want to hook the spurious faults: */
 985                 if (notify_page_fault(regs))
 986                         return;
 987                 /*
 988                  * Don't take the mm semaphore here. If we fixup a prefetch
 989                  * fault we could otherwise deadlock:
 990                  */
 991                 bad_area_nosemaphore(regs, error_code, address);
 992
 993                 return;
 994         }
 995
 996         /* kprobes don't want to hook the spurious faults: */
 997         if (unlikely(notify_page_fault(regs)))
 998                 return;
 999         /*
1000          * It's safe to allow irq's after cr2 has been saved and the
1001          * vmalloc fault has been handled.
1002          *
1003          * User-mode registers count as a user access even for any
1004          * potential system fault or CPU buglet:
1005          */
1006         if (user_mode_vm(regs)) {
1007                 local_irq_enable();
1008                 error_code |= PF_USER;
1009         } else {
1010                 if (regs->flags & X86_EFLAGS_IF)
1011                         local_irq_enable();
1012         }
1013
1014         if (unlikely(error_code & PF_RSVD))
1015                 pgtable_bad(regs, error_code, address);
1016
1017         perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1018
1019         /*
1020          * If we're in an interrupt, have no user context or are running
1021          * in an atomic region then we must not take the fault:
1022          */
1023         if (unlikely(in_atomic() || !mm)) {
1024                 bad_area_nosemaphore(regs, error_code, address);
1025                 return;
1026         }
1027
1028         /*
1029          * When running in the kernel we expect faults to occur only to
1030          * addresses in user space.  All other faults represent errors in
1031          * the kernel and should generate an OOPS.  Unfortunately, in the
1032          * case of an erroneous fault occurring in a code path which already
1033          * holds mmap_sem we will deadlock attempting to validate the fault
1034          * against the address space.  Luckily the kernel only validly
1035          * references user space from well defined areas of code, which are
1036          * listed in the exceptions table.
1037          *
1038          * As the vast majority of faults will be valid we will only perform
1039          * the source reference check when there is a possibility of a
1040          * deadlock. Attempt to lock the address space, if we cannot we then
1041          * validate the source. If this is invalid we can skip the address
1042          * space check, thus avoiding the deadlock:
1043          */
1044         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1045                 if ((error_code & PF_USER) == 0 &&
1046                     !search_exception_tables(regs->ip)) {
1047                         bad_area_nosemaphore(regs, error_code, address);
1048                         return;
1049                 }
1050                 down_read(&mm->mmap_sem);
1051         } else {
1052                 /*
1053                  * The above down_read_trylock() might have succeeded in
1054                  * which case we'll have missed the might_sleep() from
1055                  * down_read():
1056                  */
1057                 might_sleep();
1058         }
1059
1060         vma = find_vma(mm, address);
1061         if (unlikely(!vma)) {
1062                 bad_area(regs, error_code, address);
1063                 return;
1064         }
1065         if (likely(vma->vm_start <= address))
1066                 goto good_area;
1067         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1068                 bad_area(regs, error_code, address);
1069                 return;
1070         }
1071         if (error_code & PF_USER) {
1072                 /*
1073                  * Accessing the stack below %sp is always a bug.
1074                  * The large cushion allows instructions like enter
1075                  * and pusha to work. ("enter $65535, $31" pushes
1076                  * 32 pointers and then decrements %sp by 65535.)
1077                  */
1078                 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1079                         bad_area(regs, error_code, address);
1080                         return;
1081                 }
1082         }
1083         if (unlikely(expand_stack(vma, address))) {
1084                 bad_area(regs, error_code, address);
1085                 return;
1086         }
1087
1088         /*
1089          * Ok, we have a good vm_area for this memory access, so
1090          * we can handle it..
1091          */
1092 good_area:
1093         write = error_code & PF_WRITE;
1094
1095         if (unlikely(access_error(error_code, write, vma))) {
1096                 bad_area_access_error(regs, error_code, address);
1097                 return;
1098         }
1099
1100         /*
1101          * If for any reason at all we couldn't handle the fault,
1102          * make sure we exit gracefully rather than endlessly redo
1103          * the fault:
1104          */
1105         fault = handle_mm_fault(mm, vma, address, write);
1106
1107         if (unlikely(fault & VM_FAULT_ERROR)) {
1108                 mm_fault_error(regs, error_code, address, fault);
1109                 return;
1110         }
1111
1112         if (fault & VM_FAULT_MAJOR) {
1113                 tsk->maj_flt++;
1114                 perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1115                                      regs, address);
1116         } else {
1117                 tsk->min_flt++;
1118                 perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1119                                      regs, address);
1120         }
1121
1122         check_v8086_mode(regs, address, tsk);
1123
1124         up_read(&mm->mmap_sem);
1125 }