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 #include <asm/kmemcheck.h>              /* kmemcheck_*(), ...           */
  18
  19 /*
  20  * Page fault error code bits:
  21  *
  22  *   bit 0 ==    0: no page found       1: protection fault
  23  *   bit 1 ==    0: read access         1: write access
  24  *   bit 2 ==    0: kernel-mode access  1: user-mode access
  25  *   bit 3 ==                           1: use of reserved bit detected
  26  *   bit 4 ==                           1: fault was an instruction fetch
  27  */
  28 enum x86_pf_error_code {
  29
  30         PF_PROT         =               1 << 0,
  31         PF_WRITE        =               1 << 1,
  32         PF_USER         =               1 << 2,
  33         PF_RSVD         =               1 << 3,
  34         PF_INSTR        =               1 << 4,
  35 };
  36
  37 /*
  38  * Returns 0 if mmiotrace is disabled, or if the fault is not
  39  * handled by mmiotrace:
  40  */
  41 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
  42 {
  43         if (unlikely(is_kmmio_active()))
  44                 if (kmmio_handler(regs, addr) == 1)
  45                         return -1;
  46         return 0;
  47 }
  48
  49 static inline int notify_page_fault(struct pt_regs *regs)
  50 {
  51         int ret = 0;
  52
  53         /* kprobe_running() needs smp_processor_id() */
  54         if (kprobes_built_in() && !user_mode_vm(regs)) {
  55                 preempt_disable();
  56                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  57                         ret = 1;
  58                 preempt_enable();
  59         }
  60
  61         return ret;
  62 }
  63
  64 /*
  65  * Prefetch quirks:
  66  *
  67  * 32-bit mode:
  68  *
  69  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  70  *   Check that here and ignore it.
  71  *
  72  * 64-bit mode:
  73  *
  74  *   Sometimes the CPU reports invalid exceptions on prefetch.
  75  *   Check that here and ignore it.
  76  *
  77  * Opcode checker based on code by Richard Brunner.
  78  */
  79 static inline int
  80 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  81                       unsigned char opcode, int *prefetch)
  82 {
  83         unsigned char instr_hi = opcode & 0xf0;
  84         unsigned char instr_lo = opcode & 0x0f;
  85
  86         switch (instr_hi) {
  87         case 0x20:
  88         case 0x30:
  89                 /*
  90                  * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
  91                  * In X86_64 long mode, the CPU will signal invalid
  92                  * opcode if some of these prefixes are present so
  93                  * X86_64 will never get here anyway
  94                  */
  95                 return ((instr_lo & 7) == 0x6);
  96 #ifdef CONFIG_X86_64
  97         case 0x40:
  98                 /*
  99                  * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 100                  * Need to figure out under what instruction mode the
 101                  * instruction was issued. Could check the LDT for lm,
 102                  * but for now it's good enough to assume that long
 103                  * mode only uses well known segments or kernel.
 104                  */
 105                 return (!user_mode(regs)) || (regs->cs == __USER_CS);
 106 #endif
 107         case 0x60:
 108                 /* 0x64 thru 0x67 are valid prefixes in all modes. */
 109                 return (instr_lo & 0xC) == 0x4;
 110         case 0xF0:
 111                 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 112                 return !instr_lo || (instr_lo>>1) == 1;
 113         case 0x00:
 114                 /* Prefetch instruction is 0x0F0D or 0x0F18 */
 115                 if (probe_kernel_address(instr, opcode))
 116                         return 0;
 117
 118                 *prefetch = (instr_lo == 0xF) &&
 119                         (opcode == 0x0D || opcode == 0x18);
 120                 return 0;
 121         default:
 122                 return 0;
 123         }
 124 }
 125
 126 static int
 127 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 128 {
 129         unsigned char *max_instr;
 130         unsigned char *instr;
 131         int prefetch = 0;
 132
 133         /*
 134          * If it was a exec (instruction fetch) fault on NX page, then
 135          * do not ignore the fault:
 136          */
 137         if (error_code & PF_INSTR)
 138                 return 0;
 139
 140         instr = (void *)convert_ip_to_linear(current, regs);
 141         max_instr = instr + 15;
 142
 143         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 144                 return 0;
 145
 146         while (instr < max_instr) {
 147                 unsigned char opcode;
 148
 149                 if (probe_kernel_address(instr, opcode))
 150                         break;
 151
 152                 instr++;
 153
 154                 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 155                         break;
 156         }
 157         return prefetch;
 158 }
 159
 160 static void
 161 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
 162                      struct task_struct *tsk)
 163 {
 164         siginfo_t info;
 165
 166         info.si_signo   = si_signo;
 167         info.si_errno   = 0;
 168         info.si_code    = si_code;
 169         info.si_addr    = (void __user *)address;
 170
 171         force_sig_info(si_signo, &info, tsk);
 172 }
 173
 174 DEFINE_SPINLOCK(pgd_lock);
 175 LIST_HEAD(pgd_list);
 176
 177 #ifdef CONFIG_X86_32
 178 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 179 {
 180         unsigned index = pgd_index(address);
 181         pgd_t *pgd_k;
 182         pud_t *pud, *pud_k;
 183         pmd_t *pmd, *pmd_k;
 184
 185         pgd += index;
 186         pgd_k = init_mm.pgd + index;
 187
 188         if (!pgd_present(*pgd_k))
 189                 return NULL;
 190
 191         /*
 192          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 193          * and redundant with the set_pmd() on non-PAE. As would
 194          * set_pud.
 195          */
 196         pud = pud_offset(pgd, address);
 197         pud_k = pud_offset(pgd_k, address);
 198         if (!pud_present(*pud_k))
 199                 return NULL;
 200
 201         pmd = pmd_offset(pud, address);
 202         pmd_k = pmd_offset(pud_k, address);
 203         if (!pmd_present(*pmd_k))
 204                 return NULL;
 205
 206         if (!pmd_present(*pmd))
 207                 set_pmd(pmd, *pmd_k);
 208         else
 209                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 210
 211         return pmd_k;
 212 }
 213
 214 void vmalloc_sync_all(void)
 215 {
 216         unsigned long address;
 217
 218         if (SHARED_KERNEL_PMD)
 219                 return;
 220
 221         for (address = VMALLOC_START & PMD_MASK;
 222              address >= TASK_SIZE && address < FIXADDR_TOP;
 223              address += PMD_SIZE) {
 224
 225                 unsigned long flags;
 226                 struct page *page;
 227
 228                 spin_lock_irqsave(&pgd_lock, flags);
 229                 list_for_each_entry(page, &pgd_list, lru) {
 230                         if (!vmalloc_sync_one(page_address(page), address))
 231                                 break;
 232                 }
 233                 spin_unlock_irqrestore(&pgd_lock, flags);
 234         }
 235 }
 236
 237 /*
 238  * 32-bit:
 239  *
 240  *   Handle a fault on the vmalloc or module mapping area
 241  */
 242 static noinline int vmalloc_fault(unsigned long address)
 243 {
 244         unsigned long pgd_paddr;
 245         pmd_t *pmd_k;
 246         pte_t *pte_k;
 247
 248         /* Make sure we are in vmalloc area: */
 249         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 250                 return -1;
 251
 252         /*
 253          * Synchronize this task's top level page-table
 254          * with the 'reference' page table.
 255          *
 256          * Do _not_ use "current" here. We might be inside
 257          * an interrupt in the middle of a task switch..
 258          */
 259         pgd_paddr = read_cr3();
 260         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 261         if (!pmd_k)
 262                 return -1;
 263
 264         pte_k = pte_offset_kernel(pmd_k, address);
 265         if (!pte_present(*pte_k))
 266                 return -1;
 267
 268         return 0;
 269 }
 270
 271 /*
 272  * Did it hit the DOS screen memory VA from vm86 mode?
 273  */
 274 static inline void
 275 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 276                  struct task_struct *tsk)
 277 {
 278         unsigned long bit;
 279
 280         if (!v8086_mode(regs))
 281                 return;
 282
 283         bit = (address - 0xA0000) >> PAGE_SHIFT;
 284         if (bit < 32)
 285                 tsk->thread.screen_bitmap |= 1 << bit;
 286 }
 287
 288 static void dump_pagetable(unsigned long address)
 289 {
 290         __typeof__(pte_val(__pte(0))) page;
 291
 292         page = read_cr3();
 293         page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 294
 295 #ifdef CONFIG_X86_PAE
 296         printk("*pdpt = %016Lx ", page);
 297         if ((page >> PAGE_SHIFT) < max_low_pfn
 298             && page & _PAGE_PRESENT) {
 299                 page &= PAGE_MASK;
 300                 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 301                                                         & (PTRS_PER_PMD - 1)];
 302                 printk(KERN_CONT "*pde = %016Lx ", page);
 303                 page &= ~_PAGE_NX;
 304         }
 305 #else
 306         printk("*pde = %08lx ", page);
 307 #endif
 308
 309         /*
 310          * We must not directly access the pte in the highpte
 311          * case if the page table is located in highmem.
 312          * And let's rather not kmap-atomic the pte, just in case
 313          * it's allocated already:
 314          */
 315         if ((page >> PAGE_SHIFT) < max_low_pfn
 316             && (page & _PAGE_PRESENT)
 317             && !(page & _PAGE_PSE)) {
 318
 319                 page &= PAGE_MASK;
 320                 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 321                                                         & (PTRS_PER_PTE - 1)];
 322                 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 323         }
 324
 325         printk("\n");
 326 }
 327
 328 #else /* CONFIG_X86_64: */
 329
 330 void vmalloc_sync_all(void)
 331 {
 332         unsigned long address;
 333
 334         for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
 335              address += PGDIR_SIZE) {
 336
 337                 const pgd_t *pgd_ref = pgd_offset_k(address);
 338                 unsigned long flags;
 339                 struct page *page;
 340
 341                 if (pgd_none(*pgd_ref))
 342                         continue;
 343
 344                 spin_lock_irqsave(&pgd_lock, flags);
 345                 list_for_each_entry(page, &pgd_list, lru) {
 346                         pgd_t *pgd;
 347                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 348                         if (pgd_none(*pgd))
 349                                 set_pgd(pgd, *pgd_ref);
 350                         else
 351                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 352                 }
 353                 spin_unlock_irqrestore(&pgd_lock, flags);
 354         }
 355 }
 356
 357 /*
 358  * 64-bit:
 359  *
 360  *   Handle a fault on the vmalloc area
 361  *
 362  * This assumes no large pages in there.
 363  */
 364 static noinline int vmalloc_fault(unsigned long address)
 365 {
 366         pgd_t *pgd, *pgd_ref;
 367         pud_t *pud, *pud_ref;
 368         pmd_t *pmd, *pmd_ref;
 369         pte_t *pte, *pte_ref;
 370
 371         /* Make sure we are in vmalloc area: */
 372         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 373                 return -1;
 374
 375         /*
 376          * Copy kernel mappings over when needed. This can also
 377          * happen within a race in page table update. In the later
 378          * case just flush:
 379          */
 380         pgd = pgd_offset(current->active_mm, address);
 381         pgd_ref = pgd_offset_k(address);
 382         if (pgd_none(*pgd_ref))
 383                 return -1;
 384
 385         if (pgd_none(*pgd))
 386                 set_pgd(pgd, *pgd_ref);
 387         else
 388                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 389
 390         /*
 391          * Below here mismatches are bugs because these lower tables
 392          * are shared:
 393          */
 394
 395         pud = pud_offset(pgd, address);
 396         pud_ref = pud_offset(pgd_ref, address);
 397         if (pud_none(*pud_ref))
 398                 return -1;
 399
 400         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 401                 BUG();
 402
 403         pmd = pmd_offset(pud, address);
 404         pmd_ref = pmd_offset(pud_ref, address);
 405         if (pmd_none(*pmd_ref))
 406                 return -1;
 407
 408         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 409                 BUG();
 410
 411         pte_ref = pte_offset_kernel(pmd_ref, address);
 412         if (!pte_present(*pte_ref))
 413                 return -1;
 414
 415         pte = pte_offset_kernel(pmd, address);
 416
 417         /*
 418          * Don't use pte_page here, because the mappings can point
 419          * outside mem_map, and the NUMA hash lookup cannot handle
 420          * that:
 421          */
 422         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 423                 BUG();
 424
 425         return 0;
 426 }
 427
 428 static const char errata93_warning[] =
 429 KERN_ERR
 430 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 431 "******* Working around it, but it may cause SEGVs or burn power.\n"
 432 "******* Please consider a BIOS update.\n"
 433 "******* Disabling USB legacy in the BIOS may also help.\n";
 434
 435 /*
 436  * No vm86 mode in 64-bit mode:
 437  */
 438 static inline void
 439 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 440                  struct task_struct *tsk)
 441 {
 442 }
 443
 444 static int bad_address(void *p)
 445 {
 446         unsigned long dummy;
 447
 448         return probe_kernel_address((unsigned long *)p, dummy);
 449 }
 450
 451 static void dump_pagetable(unsigned long address)
 452 {
 453         pgd_t *pgd;
 454         pud_t *pud;
 455         pmd_t *pmd;
 456         pte_t *pte;
 457
 458         pgd = (pgd_t *)read_cr3();
 459
 460         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 461
 462         pgd += pgd_index(address);
 463         if (bad_address(pgd))
 464                 goto bad;
 465
 466         printk("PGD %lx ", pgd_val(*pgd));
 467
 468         if (!pgd_present(*pgd))
 469                 goto out;
 470
 471         pud = pud_offset(pgd, address);
 472         if (bad_address(pud))
 473                 goto bad;
 474
 475         printk("PUD %lx ", pud_val(*pud));
 476         if (!pud_present(*pud) || pud_large(*pud))
 477                 goto out;
 478
 479         pmd = pmd_offset(pud, address);
 480         if (bad_address(pmd))
 481                 goto bad;
 482
 483         printk("PMD %lx ", pmd_val(*pmd));
 484         if (!pmd_present(*pmd) || pmd_large(*pmd))
 485                 goto out;
 486
 487         pte = pte_offset_kernel(pmd, address);
 488         if (bad_address(pte))
 489                 goto bad;
 490
 491         printk("PTE %lx", pte_val(*pte));
 492 out:
 493         printk("\n");
 494         return;
 495 bad:
 496         printk("BAD\n");
 497 }
 498
 499 #endif /* CONFIG_X86_64 */
 500
 501 /*
 502  * Workaround for K8 erratum #93 & buggy BIOS.
 503  *
 504  * BIOS SMM functions are required to use a specific workaround
 505  * to avoid corruption of the 64bit RIP register on C stepping K8.
 506  *
 507  * A lot of BIOS that didn't get tested properly miss this.
 508  *
 509  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 510  * Try to work around it here.
 511  *
 512  * Note we only handle faults in kernel here.
 513  * Does nothing on 32-bit.
 514  */
 515 static int is_errata93(struct pt_regs *regs, unsigned long address)
 516 {
 517 #ifdef CONFIG_X86_64
 518         if (address != regs->ip)
 519                 return 0;
 520
 521         if ((address >> 32) != 0)
 522                 return 0;
 523
 524         address |= 0xffffffffUL << 32;
 525         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 526             (address >= MODULES_VADDR && address <= MODULES_END)) {
 527                 printk_once(errata93_warning);
 528                 regs->ip = address;
 529                 return 1;
 530         }
 531 #endif
 532         return 0;
 533 }
 534
 535 /*
 536  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 537  * to illegal addresses >4GB.
 538  *
 539  * We catch this in the page fault handler because these addresses
 540  * are not reachable. Just detect this case and return.  Any code
 541  * segment in LDT is compatibility mode.
 542  */
 543 static int is_errata100(struct pt_regs *regs, unsigned long address)
 544 {
 545 #ifdef CONFIG_X86_64
 546         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 547                 return 1;
 548 #endif
 549         return 0;
 550 }
 551
 552 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 553 {
 554 #ifdef CONFIG_X86_F00F_BUG
 555         unsigned long nr;
 556
 557         /*
 558          * Pentium F0 0F C7 C8 bug workaround:
 559          */
 560         if (boot_cpu_data.f00f_bug) {
 561                 nr = (address - idt_descr.address) >> 3;
 562
 563                 if (nr == 6) {
 564                         do_invalid_op(regs, 0);
 565                         return 1;
 566                 }
 567         }
 568 #endif
 569         return 0;
 570 }
 571
 572 static const char nx_warning[] = KERN_CRIT
 573 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
 574
 575 static void
 576 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 577                 unsigned long address)
 578 {
 579         if (!oops_may_print())
 580                 return;
 581
 582         if (error_code & PF_INSTR) {
 583                 unsigned int level;
 584
 585                 pte_t *pte = lookup_address(address, &level);
 586
 587                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 588                         printk(nx_warning, current_uid());
 589         }
 590
 591         printk(KERN_ALERT "BUG: unable to handle kernel ");
 592         if (address < PAGE_SIZE)
 593                 printk(KERN_CONT "NULL pointer dereference");
 594         else
 595                 printk(KERN_CONT "paging request");
 596
 597         printk(KERN_CONT " at %p\n", (void *) address);
 598         printk(KERN_ALERT "IP:");
 599         printk_address(regs->ip, 1);
 600
 601         dump_pagetable(address);
 602 }
 603
 604 static noinline void
 605 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 606             unsigned long address)
 607 {
 608         struct task_struct *tsk;
 609         unsigned long flags;
 610         int sig;
 611
 612         flags = oops_begin();
 613         tsk = current;
 614         sig = SIGKILL;
 615
 616         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 617                tsk->comm, address);
 618         dump_pagetable(address);
 619
 620         tsk->thread.cr2         = address;
 621         tsk->thread.trap_no     = 14;
 622         tsk->thread.error_code  = error_code;
 623
 624         if (__die("Bad pagetable", regs, error_code))
 625                 sig = 0;
 626
 627         oops_end(flags, regs, sig);
 628 }
 629
 630 static noinline void
 631 no_context(struct pt_regs *regs, unsigned long error_code,
 632            unsigned long address)
 633 {
 634         struct task_struct *tsk = current;
 635         unsigned long *stackend;
 636         unsigned long flags;
 637         int sig;
 638
 639         /* Are we prepared to handle this kernel fault? */
 640         if (fixup_exception(regs))
 641                 return;
 642
 643         /*
 644          * 32-bit:
 645          *
 646          *   Valid to do another page fault here, because if this fault
 647          *   had been triggered by is_prefetch fixup_exception would have
 648          *   handled it.
 649          *
 650          * 64-bit:
 651          *
 652          *   Hall of shame of CPU/BIOS bugs.
 653          */
 654         if (is_prefetch(regs, error_code, address))
 655                 return;
 656
 657         if (is_errata93(regs, address))
 658                 return;
 659
 660         /*
 661          * Oops. The kernel tried to access some bad page. We'll have to
 662          * terminate things with extreme prejudice:
 663          */
 664         flags = oops_begin();
 665
 666         show_fault_oops(regs, error_code, address);
 667
 668         stackend = end_of_stack(tsk);
 669         if (*stackend != STACK_END_MAGIC)
 670                 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
 671
 672         tsk->thread.cr2         = address;
 673         tsk->thread.trap_no     = 14;
 674         tsk->thread.error_code  = error_code;
 675
 676         sig = SIGKILL;
 677         if (__die("Oops", regs, error_code))
 678                 sig = 0;
 679
 680         /* Executive summary in case the body of the oops scrolled away */
 681         printk(KERN_EMERG "CR2: %016lx\n", address);
 682
 683         oops_end(flags, regs, sig);
 684 }
 685
 686 /*
 687  * Print out info about fatal segfaults, if the show_unhandled_signals
 688  * sysctl is set:
 689  */
 690 static inline void
 691 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 692                 unsigned long address, struct task_struct *tsk)
 693 {
 694         if (!unhandled_signal(tsk, SIGSEGV))
 695                 return;
 696
 697         if (!printk_ratelimit())
 698                 return;
 699
 700         printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 701                 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 702                 tsk->comm, task_pid_nr(tsk), address,
 703                 (void *)regs->ip, (void *)regs->sp, error_code);
 704
 705         print_vma_addr(KERN_CONT " in ", regs->ip);
 706
 707         printk(KERN_CONT "\n");
 708 }
 709
 710 static void
 711 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 712                        unsigned long address, int si_code)
 713 {
 714         struct task_struct *tsk = current;
 715
 716         /* User mode accesses just cause a SIGSEGV */
 717         if (error_code & PF_USER) {
 718                 /*
 719                  * It's possible to have interrupts off here:
 720                  */
 721                 local_irq_enable();
 722
 723                 /*
 724                  * Valid to do another page fault here because this one came
 725                  * from user space:
 726                  */
 727                 if (is_prefetch(regs, error_code, address))
 728                         return;
 729
 730                 if (is_errata100(regs, address))
 731                         return;
 732
 733                 if (unlikely(show_unhandled_signals))
 734                         show_signal_msg(regs, error_code, address, tsk);
 735
 736                 /* Kernel addresses are always protection faults: */
 737                 tsk->thread.cr2         = address;
 738                 tsk->thread.error_code  = error_code | (address >= TASK_SIZE);
 739                 tsk->thread.trap_no     = 14;
 740
 741                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 742
 743                 return;
 744         }
 745
 746         if (is_f00f_bug(regs, address))
 747                 return;
 748
 749         no_context(regs, error_code, address);
 750 }
 751
 752 static noinline void
 753 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 754                      unsigned long address)
 755 {
 756         __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 757 }
 758
 759 static void
 760 __bad_area(struct pt_regs *regs, unsigned long error_code,
 761            unsigned long address, int si_code)
 762 {
 763         struct mm_struct *mm = current->mm;
 764
 765         /*
 766          * Something tried to access memory that isn't in our memory map..
 767          * Fix it, but check if it's kernel or user first..
 768          */
 769         up_read(&mm->mmap_sem);
 770
 771         __bad_area_nosemaphore(regs, error_code, address, si_code);
 772 }
 773
 774 static noinline void
 775 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 776 {
 777         __bad_area(regs, error_code, address, SEGV_MAPERR);
 778 }
 779
 780 static noinline void
 781 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 782                       unsigned long address)
 783 {
 784         __bad_area(regs, error_code, address, SEGV_ACCERR);
 785 }
 786
 787 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
 788 static void
 789 out_of_memory(struct pt_regs *regs, unsigned long error_code,
 790               unsigned long address)
 791 {
 792         /*
 793          * We ran out of memory, call the OOM killer, and return the userspace
 794          * (which will retry the fault, or kill us if we got oom-killed):
 795          */
 796         up_read(&current->mm->mmap_sem);
 797
 798         pagefault_out_of_memory();
 799 }
 800
 801 static void
 802 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 803 {
 804         struct task_struct *tsk = current;
 805         struct mm_struct *mm = tsk->mm;
 806
 807         up_read(&mm->mmap_sem);
 808
 809         /* Kernel mode? Handle exceptions or die: */
 810         if (!(error_code & PF_USER))
 811                 no_context(regs, error_code, address);
 812
 813         /* User-space => ok to do another page fault: */
 814         if (is_prefetch(regs, error_code, address))
 815                 return;
 816
 817         tsk->thread.cr2         = address;
 818         tsk->thread.error_code  = error_code;
 819         tsk->thread.trap_no     = 14;
 820
 821         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 822 }
 823
 824 static noinline void
 825 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 826                unsigned long address, unsigned int fault)
 827 {
 828         if (fault & VM_FAULT_OOM) {
 829                 out_of_memory(regs, error_code, address);
 830         } else {
 831                 if (fault & VM_FAULT_SIGBUS)
 832                         do_sigbus(regs, error_code, address);
 833                 else
 834                         BUG();
 835         }
 836 }
 837
 838 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 839 {
 840         if ((error_code & PF_WRITE) && !pte_write(*pte))
 841                 return 0;
 842
 843         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 844                 return 0;
 845
 846         return 1;
 847 }
 848
 849 /*
 850  * Handle a spurious fault caused by a stale TLB entry.
 851  *
 852  * This allows us to lazily refresh the TLB when increasing the
 853  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
 854  * eagerly is very expensive since that implies doing a full
 855  * cross-processor TLB flush, even if no stale TLB entries exist
 856  * on other processors.
 857  *
 858  * There are no security implications to leaving a stale TLB when
 859  * increasing the permissions on a page.
 860  */
 861 static noinline int
 862 spurious_fault(unsigned long error_code, unsigned long address)
 863 {
 864         pgd_t *pgd;
 865         pud_t *pud;
 866         pmd_t *pmd;
 867         pte_t *pte;
 868         int ret;
 869
 870         /* Reserved-bit violation or user access to kernel space? */
 871         if (error_code & (PF_USER | PF_RSVD))
 872                 return 0;
 873
 874         pgd = init_mm.pgd + pgd_index(address);
 875         if (!pgd_present(*pgd))
 876                 return 0;
 877
 878         pud = pud_offset(pgd, address);
 879         if (!pud_present(*pud))
 880                 return 0;
 881
 882         if (pud_large(*pud))
 883                 return spurious_fault_check(error_code, (pte_t *) pud);
 884
 885         pmd = pmd_offset(pud, address);
 886         if (!pmd_present(*pmd))
 887                 return 0;
 888
 889         if (pmd_large(*pmd))
 890                 return spurious_fault_check(error_code, (pte_t *) pmd);
 891
 892         pte = pte_offset_kernel(pmd, address);
 893         if (!pte_present(*pte))
 894                 return 0;
 895
 896         ret = spurious_fault_check(error_code, pte);
 897         if (!ret)
 898                 return 0;
 899
 900         /*
 901          * Make sure we have permissions in PMD.
 902          * If not, then there's a bug in the page tables:
 903          */
 904         ret = spurious_fault_check(error_code, (pte_t *) pmd);
 905         WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
 906
 907         return ret;
 908 }
 909
 910 int show_unhandled_signals = 1;
 911
 912 static inline int
 913 access_error(unsigned long error_code, int write, struct vm_area_struct *vma)
 914 {
 915         if (write) {
 916                 /* write, present and write, not present: */
 917                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 918                         return 1;
 919                 return 0;
 920         }
 921
 922         /* read, present: */
 923         if (unlikely(error_code & PF_PROT))
 924                 return 1;
 925
 926         /* read, not present: */
 927         if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 928                 return 1;
 929
 930         return 0;
 931 }
 932
 933 static int fault_in_kernel_space(unsigned long address)
 934 {
 935         return address >= TASK_SIZE_MAX;
 936 }
 937
 938 /*
 939  * This routine handles page faults.  It determines the address,
 940  * and the problem, and then passes it off to one of the appropriate
 941  * routines.
 942  */
 943 dotraplinkage void __kprobes
 944 do_page_fault(struct pt_regs *regs, unsigned long error_code)
 945 {
 946         struct vm_area_struct *vma;
 947         struct task_struct *tsk;
 948         unsigned long address;
 949         struct mm_struct *mm;
 950         int write;
 951         int fault;
 952
 953         tsk = current;
 954         mm = tsk->mm;
 955
 956         /* Get the faulting address: */
 957         address = read_cr2();
 958
 959         /*
 960          * Detect and handle instructions that would cause a page fault for
 961          * both a tracked kernel page and a userspace page.
 962          */
 963         if (kmemcheck_active(regs))
 964                 kmemcheck_hide(regs);
 965         prefetchw(&mm->mmap_sem);
 966
 967         if (unlikely(kmmio_fault(regs, address)))
 968                 return;
 969
 970         /*
 971          * We fault-in kernel-space virtual memory on-demand. The
 972          * 'reference' page table is init_mm.pgd.
 973          *
 974          * NOTE! We MUST NOT take any locks for this case. We may
 975          * be in an interrupt or a critical region, and should
 976          * only copy the information from the master page table,
 977          * nothing more.
 978          *
 979          * This verifies that the fault happens in kernel space
 980          * (error_code & 4) == 0, and that the fault was not a
 981          * protection error (error_code & 9) == 0.
 982          */
 983         if (unlikely(fault_in_kernel_space(address))) {
 984                 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
 985                         if (vmalloc_fault(address) >= 0)
 986                                 return;
 987
 988                         if (kmemcheck_fault(regs, address, error_code))
 989                                 return;
 990                 }
 991
 992                 /* Can handle a stale RO->RW TLB: */
 993                 if (spurious_fault(error_code, address))
 994                         return;
 995
 996                 /* kprobes don't want to hook the spurious faults: */
 997                 if (notify_page_fault(regs))
 998                         return;
 999                 /*
1000                  * Don't take the mm semaphore here. If we fixup a prefetch
1001                  * fault we could otherwise deadlock:
1002                  */
1003                 bad_area_nosemaphore(regs, error_code, address);
1004
1005                 return;
1006         }
1007
1008         /* kprobes don't want to hook the spurious faults: */
1009         if (unlikely(notify_page_fault(regs)))
1010                 return;
1011         /*
1012          * It's safe to allow irq's after cr2 has been saved and the
1013          * vmalloc fault has been handled.
1014          *
1015          * User-mode registers count as a user access even for any
1016          * potential system fault or CPU buglet:
1017          */
1018         if (user_mode_vm(regs)) {
1019                 local_irq_enable();
1020                 error_code |= PF_USER;
1021         } else {
1022                 if (regs->flags & X86_EFLAGS_IF)
1023                         local_irq_enable();
1024         }
1025
1026         if (unlikely(error_code & PF_RSVD))
1027                 pgtable_bad(regs, error_code, address);
1028
1029         perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1030
1031         /*
1032          * If we're in an interrupt, have no user context or are running
1033          * in an atomic region then we must not take the fault:
1034          */
1035         if (unlikely(in_atomic() || !mm)) {
1036                 bad_area_nosemaphore(regs, error_code, address);
1037                 return;
1038         }
1039
1040         /*
1041          * When running in the kernel we expect faults to occur only to
1042          * addresses in user space.  All other faults represent errors in
1043          * the kernel and should generate an OOPS.  Unfortunately, in the
1044          * case of an erroneous fault occurring in a code path which already
1045          * holds mmap_sem we will deadlock attempting to validate the fault
1046          * against the address space.  Luckily the kernel only validly
1047          * references user space from well defined areas of code, which are
1048          * listed in the exceptions table.
1049          *
1050          * As the vast majority of faults will be valid we will only perform
1051          * the source reference check when there is a possibility of a
1052          * deadlock. Attempt to lock the address space, if we cannot we then
1053          * validate the source. If this is invalid we can skip the address
1054          * space check, thus avoiding the deadlock:
1055          */
1056         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1057                 if ((error_code & PF_USER) == 0 &&
1058                     !search_exception_tables(regs->ip)) {
1059                         bad_area_nosemaphore(regs, error_code, address);
1060                         return;
1061                 }
1062                 down_read(&mm->mmap_sem);
1063         } else {
1064                 /*
1065                  * The above down_read_trylock() might have succeeded in
1066                  * which case we'll have missed the might_sleep() from
1067                  * down_read():
1068                  */
1069                 might_sleep();
1070         }
1071
1072         vma = find_vma(mm, address);
1073         if (unlikely(!vma)) {
1074                 bad_area(regs, error_code, address);
1075                 return;
1076         }
1077         if (likely(vma->vm_start <= address))
1078                 goto good_area;
1079         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1080                 bad_area(regs, error_code, address);
1081                 return;
1082         }
1083         if (error_code & PF_USER) {
1084                 /*
1085                  * Accessing the stack below %sp is always a bug.
1086                  * The large cushion allows instructions like enter
1087                  * and pusha to work. ("enter $65535, $31" pushes
1088                  * 32 pointers and then decrements %sp by 65535.)
1089                  */
1090                 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1091                         bad_area(regs, error_code, address);
1092                         return;
1093                 }
1094         }
1095         if (unlikely(expand_stack(vma, address))) {
1096                 bad_area(regs, error_code, address);
1097                 return;
1098         }
1099
1100         /*
1101          * Ok, we have a good vm_area for this memory access, so
1102          * we can handle it..
1103          */
1104 good_area:
1105         write = error_code & PF_WRITE;
1106
1107         if (unlikely(access_error(error_code, write, vma))) {
1108                 bad_area_access_error(regs, error_code, address);
1109                 return;
1110         }
1111
1112         /*
1113          * If for any reason at all we couldn't handle the fault,
1114          * make sure we exit gracefully rather than endlessly redo
1115          * the fault:
1116          */
1117         fault = handle_mm_fault(mm, vma, address, write ? FAULT_FLAG_WRITE : 0);
1118
1119         if (unlikely(fault & VM_FAULT_ERROR)) {
1120                 mm_fault_error(regs, error_code, address, fault);
1121                 return;
1122         }
1123
1124         if (fault & VM_FAULT_MAJOR) {
1125                 tsk->maj_flt++;
1126                 perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1127                                      regs, address);
1128         } else {
1129                 tsk->min_flt++;
1130                 perf_swcounter_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1131                                      regs, address);
1132         }
1133
1134         check_v8086_mode(regs, address, tsk);
1135
1136         up_read(&mm->mmap_sem);
1137 }