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_event.h>           /* perf_sw_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         info.si_addr_lsb = si_code == BUS_MCEERR_AR ? PAGE_SHIFT : 0;
 171
 172         force_sig_info(si_signo, &info, tsk);
 173 }
 174
 175 DEFINE_SPINLOCK(pgd_lock);
 176 LIST_HEAD(pgd_list);
 177
 178 #ifdef CONFIG_X86_32
 179 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 180 {
 181         unsigned index = pgd_index(address);
 182         pgd_t *pgd_k;
 183         pud_t *pud, *pud_k;
 184         pmd_t *pmd, *pmd_k;
 185
 186         pgd += index;
 187         pgd_k = init_mm.pgd + index;
 188
 189         if (!pgd_present(*pgd_k))
 190                 return NULL;
 191
 192         /*
 193          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 194          * and redundant with the set_pmd() on non-PAE. As would
 195          * set_pud.
 196          */
 197         pud = pud_offset(pgd, address);
 198         pud_k = pud_offset(pgd_k, address);
 199         if (!pud_present(*pud_k))
 200                 return NULL;
 201
 202         pmd = pmd_offset(pud, address);
 203         pmd_k = pmd_offset(pud_k, address);
 204         if (!pmd_present(*pmd_k))
 205                 return NULL;
 206
 207         if (!pmd_present(*pmd))
 208                 set_pmd(pmd, *pmd_k);
 209         else
 210                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 211
 212         return pmd_k;
 213 }
 214
 215 void vmalloc_sync_all(void)
 216 {
 217         unsigned long address;
 218
 219         if (SHARED_KERNEL_PMD)
 220                 return;
 221
 222         for (address = VMALLOC_START & PMD_MASK;
 223              address >= TASK_SIZE && address < FIXADDR_TOP;
 224              address += PMD_SIZE) {
 225
 226                 unsigned long flags;
 227                 struct page *page;
 228
 229                 spin_lock_irqsave(&pgd_lock, flags);
 230                 list_for_each_entry(page, &pgd_list, lru) {
 231                         if (!vmalloc_sync_one(page_address(page), address))
 232                                 break;
 233                 }
 234                 spin_unlock_irqrestore(&pgd_lock, flags);
 235         }
 236 }
 237
 238 /*
 239  * 32-bit:
 240  *
 241  *   Handle a fault on the vmalloc or module mapping area
 242  */
 243 static noinline int vmalloc_fault(unsigned long address)
 244 {
 245         unsigned long pgd_paddr;
 246         pmd_t *pmd_k;
 247         pte_t *pte_k;
 248
 249         /* Make sure we are in vmalloc area: */
 250         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 251                 return -1;
 252
 253         /*
 254          * Synchronize this task's top level page-table
 255          * with the 'reference' page table.
 256          *
 257          * Do _not_ use "current" here. We might be inside
 258          * an interrupt in the middle of a task switch..
 259          */
 260         pgd_paddr = read_cr3();
 261         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 262         if (!pmd_k)
 263                 return -1;
 264
 265         pte_k = pte_offset_kernel(pmd_k, address);
 266         if (!pte_present(*pte_k))
 267                 return -1;
 268
 269         return 0;
 270 }
 271
 272 /*
 273  * Did it hit the DOS screen memory VA from vm86 mode?
 274  */
 275 static inline void
 276 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 277                  struct task_struct *tsk)
 278 {
 279         unsigned long bit;
 280
 281         if (!v8086_mode(regs))
 282                 return;
 283
 284         bit = (address - 0xA0000) >> PAGE_SHIFT;
 285         if (bit < 32)
 286                 tsk->thread.screen_bitmap |= 1 << bit;
 287 }
 288
 289 static bool low_pfn(unsigned long pfn)
 290 {
 291         return pfn < max_low_pfn;
 292 }
 293
 294 static void dump_pagetable(unsigned long address)
 295 {
 296         pgd_t *base = __va(read_cr3());
 297         pgd_t *pgd = &base[pgd_index(address)];
 298         pmd_t *pmd;
 299         pte_t *pte;
 300
 301 #ifdef CONFIG_X86_PAE
 302         printk("*pdpt = %016Lx ", pgd_val(*pgd));
 303         if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
 304                 goto out;
 305 #endif
 306         pmd = pmd_offset(pud_offset(pgd, address), address);
 307         printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
 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 (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
 316                 goto out;
 317
 318         pte = pte_offset_kernel(pmd, address);
 319         printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
 320 out:
 321         printk("\n");
 322 }
 323
 324 #else /* CONFIG_X86_64: */
 325
 326 void vmalloc_sync_all(void)
 327 {
 328         unsigned long address;
 329
 330         for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
 331              address += PGDIR_SIZE) {
 332
 333                 const pgd_t *pgd_ref = pgd_offset_k(address);
 334                 unsigned long flags;
 335                 struct page *page;
 336
 337                 if (pgd_none(*pgd_ref))
 338                         continue;
 339
 340                 spin_lock_irqsave(&pgd_lock, flags);
 341                 list_for_each_entry(page, &pgd_list, lru) {
 342                         pgd_t *pgd;
 343                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 344                         if (pgd_none(*pgd))
 345                                 set_pgd(pgd, *pgd_ref);
 346                         else
 347                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 348                 }
 349                 spin_unlock_irqrestore(&pgd_lock, flags);
 350         }
 351 }
 352
 353 /*
 354  * 64-bit:
 355  *
 356  *   Handle a fault on the vmalloc area
 357  *
 358  * This assumes no large pages in there.
 359  */
 360 static noinline int vmalloc_fault(unsigned long address)
 361 {
 362         pgd_t *pgd, *pgd_ref;
 363         pud_t *pud, *pud_ref;
 364         pmd_t *pmd, *pmd_ref;
 365         pte_t *pte, *pte_ref;
 366
 367         /* Make sure we are in vmalloc area: */
 368         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 369                 return -1;
 370
 371         /*
 372          * Copy kernel mappings over when needed. This can also
 373          * happen within a race in page table update. In the later
 374          * case just flush:
 375          */
 376         pgd = pgd_offset(current->active_mm, address);
 377         pgd_ref = pgd_offset_k(address);
 378         if (pgd_none(*pgd_ref))
 379                 return -1;
 380
 381         if (pgd_none(*pgd))
 382                 set_pgd(pgd, *pgd_ref);
 383         else
 384                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 385
 386         /*
 387          * Below here mismatches are bugs because these lower tables
 388          * are shared:
 389          */
 390
 391         pud = pud_offset(pgd, address);
 392         pud_ref = pud_offset(pgd_ref, address);
 393         if (pud_none(*pud_ref))
 394                 return -1;
 395
 396         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 397                 BUG();
 398
 399         pmd = pmd_offset(pud, address);
 400         pmd_ref = pmd_offset(pud_ref, address);
 401         if (pmd_none(*pmd_ref))
 402                 return -1;
 403
 404         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 405                 BUG();
 406
 407         pte_ref = pte_offset_kernel(pmd_ref, address);
 408         if (!pte_present(*pte_ref))
 409                 return -1;
 410
 411         pte = pte_offset_kernel(pmd, address);
 412
 413         /*
 414          * Don't use pte_page here, because the mappings can point
 415          * outside mem_map, and the NUMA hash lookup cannot handle
 416          * that:
 417          */
 418         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 419                 BUG();
 420
 421         return 0;
 422 }
 423
 424 static const char errata93_warning[] =
 425 KERN_ERR
 426 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 427 "******* Working around it, but it may cause SEGVs or burn power.\n"
 428 "******* Please consider a BIOS update.\n"
 429 "******* Disabling USB legacy in the BIOS may also help.\n";
 430
 431 /*
 432  * No vm86 mode in 64-bit mode:
 433  */
 434 static inline void
 435 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 436                  struct task_struct *tsk)
 437 {
 438 }
 439
 440 static int bad_address(void *p)
 441 {
 442         unsigned long dummy;
 443
 444         return probe_kernel_address((unsigned long *)p, dummy);
 445 }
 446
 447 static void dump_pagetable(unsigned long address)
 448 {
 449         pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
 450         pgd_t *pgd = base + pgd_index(address);
 451         pud_t *pud;
 452         pmd_t *pmd;
 453         pte_t *pte;
 454
 455         if (bad_address(pgd))
 456                 goto bad;
 457
 458         printk("PGD %lx ", pgd_val(*pgd));
 459
 460         if (!pgd_present(*pgd))
 461                 goto out;
 462
 463         pud = pud_offset(pgd, address);
 464         if (bad_address(pud))
 465                 goto bad;
 466
 467         printk("PUD %lx ", pud_val(*pud));
 468         if (!pud_present(*pud) || pud_large(*pud))
 469                 goto out;
 470
 471         pmd = pmd_offset(pud, address);
 472         if (bad_address(pmd))
 473                 goto bad;
 474
 475         printk("PMD %lx ", pmd_val(*pmd));
 476         if (!pmd_present(*pmd) || pmd_large(*pmd))
 477                 goto out;
 478
 479         pte = pte_offset_kernel(pmd, address);
 480         if (bad_address(pte))
 481                 goto bad;
 482
 483         printk("PTE %lx", pte_val(*pte));
 484 out:
 485         printk("\n");
 486         return;
 487 bad:
 488         printk("BAD\n");
 489 }
 490
 491 #endif /* CONFIG_X86_64 */
 492
 493 /*
 494  * Workaround for K8 erratum #93 & buggy BIOS.
 495  *
 496  * BIOS SMM functions are required to use a specific workaround
 497  * to avoid corruption of the 64bit RIP register on C stepping K8.
 498  *
 499  * A lot of BIOS that didn't get tested properly miss this.
 500  *
 501  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 502  * Try to work around it here.
 503  *
 504  * Note we only handle faults in kernel here.
 505  * Does nothing on 32-bit.
 506  */
 507 static int is_errata93(struct pt_regs *regs, unsigned long address)
 508 {
 509 #ifdef CONFIG_X86_64
 510         if (address != regs->ip)
 511                 return 0;
 512
 513         if ((address >> 32) != 0)
 514                 return 0;
 515
 516         address |= 0xffffffffUL << 32;
 517         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 518             (address >= MODULES_VADDR && address <= MODULES_END)) {
 519                 printk_once(errata93_warning);
 520                 regs->ip = address;
 521                 return 1;
 522         }
 523 #endif
 524         return 0;
 525 }
 526
 527 /*
 528  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 529  * to illegal addresses >4GB.
 530  *
 531  * We catch this in the page fault handler because these addresses
 532  * are not reachable. Just detect this case and return.  Any code
 533  * segment in LDT is compatibility mode.
 534  */
 535 static int is_errata100(struct pt_regs *regs, unsigned long address)
 536 {
 537 #ifdef CONFIG_X86_64
 538         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 539                 return 1;
 540 #endif
 541         return 0;
 542 }
 543
 544 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 545 {
 546 #ifdef CONFIG_X86_F00F_BUG
 547         unsigned long nr;
 548
 549         /*
 550          * Pentium F0 0F C7 C8 bug workaround:
 551          */
 552         if (boot_cpu_data.f00f_bug) {
 553                 nr = (address - idt_descr.address) >> 3;
 554
 555                 if (nr == 6) {
 556                         do_invalid_op(regs, 0);
 557                         return 1;
 558                 }
 559         }
 560 #endif
 561         return 0;
 562 }
 563
 564 static const char nx_warning[] = KERN_CRIT
 565 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
 566
 567 static void
 568 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 569                 unsigned long address)
 570 {
 571         if (!oops_may_print())
 572                 return;
 573
 574         if (error_code & PF_INSTR) {
 575                 unsigned int level;
 576
 577                 pte_t *pte = lookup_address(address, &level);
 578
 579                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 580                         printk(nx_warning, current_uid());
 581         }
 582
 583         printk(KERN_ALERT "BUG: unable to handle kernel ");
 584         if (address < PAGE_SIZE)
 585                 printk(KERN_CONT "NULL pointer dereference");
 586         else
 587                 printk(KERN_CONT "paging request");
 588
 589         printk(KERN_CONT " at %p\n", (void *) address);
 590         printk(KERN_ALERT "IP:");
 591         printk_address(regs->ip, 1);
 592
 593         dump_pagetable(address);
 594 }
 595
 596 static noinline void
 597 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 598             unsigned long address)
 599 {
 600         struct task_struct *tsk;
 601         unsigned long flags;
 602         int sig;
 603
 604         flags = oops_begin();
 605         tsk = current;
 606         sig = SIGKILL;
 607
 608         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 609                tsk->comm, address);
 610         dump_pagetable(address);
 611
 612         tsk->thread.cr2         = address;
 613         tsk->thread.trap_no     = 14;
 614         tsk->thread.error_code  = error_code;
 615
 616         if (__die("Bad pagetable", regs, error_code))
 617                 sig = 0;
 618
 619         oops_end(flags, regs, sig);
 620 }
 621
 622 static noinline void
 623 no_context(struct pt_regs *regs, unsigned long error_code,
 624            unsigned long address)
 625 {
 626         struct task_struct *tsk = current;
 627         unsigned long *stackend;
 628         unsigned long flags;
 629         int sig;
 630
 631         /* Are we prepared to handle this kernel fault? */
 632         if (fixup_exception(regs))
 633                 return;
 634
 635         /*
 636          * 32-bit:
 637          *
 638          *   Valid to do another page fault here, because if this fault
 639          *   had been triggered by is_prefetch fixup_exception would have
 640          *   handled it.
 641          *
 642          * 64-bit:
 643          *
 644          *   Hall of shame of CPU/BIOS bugs.
 645          */
 646         if (is_prefetch(regs, error_code, address))
 647                 return;
 648
 649         if (is_errata93(regs, address))
 650                 return;
 651
 652         /*
 653          * Oops. The kernel tried to access some bad page. We'll have to
 654          * terminate things with extreme prejudice:
 655          */
 656         flags = oops_begin();
 657
 658         show_fault_oops(regs, error_code, address);
 659
 660         stackend = end_of_stack(tsk);
 661         if (*stackend != STACK_END_MAGIC)
 662                 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
 663
 664         tsk->thread.cr2         = address;
 665         tsk->thread.trap_no     = 14;
 666         tsk->thread.error_code  = error_code;
 667
 668         sig = SIGKILL;
 669         if (__die("Oops", regs, error_code))
 670                 sig = 0;
 671
 672         /* Executive summary in case the body of the oops scrolled away */
 673         printk(KERN_EMERG "CR2: %016lx\n", address);
 674
 675         oops_end(flags, regs, sig);
 676 }
 677
 678 /*
 679  * Print out info about fatal segfaults, if the show_unhandled_signals
 680  * sysctl is set:
 681  */
 682 static inline void
 683 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 684                 unsigned long address, struct task_struct *tsk)
 685 {
 686         if (!unhandled_signal(tsk, SIGSEGV))
 687                 return;
 688
 689         if (!printk_ratelimit())
 690                 return;
 691
 692         printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 693                 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 694                 tsk->comm, task_pid_nr(tsk), address,
 695                 (void *)regs->ip, (void *)regs->sp, error_code);
 696
 697         print_vma_addr(KERN_CONT " in ", regs->ip);
 698
 699         printk(KERN_CONT "\n");
 700 }
 701
 702 static void
 703 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 704                        unsigned long address, int si_code)
 705 {
 706         struct task_struct *tsk = current;
 707
 708         /* User mode accesses just cause a SIGSEGV */
 709         if (error_code & PF_USER) {
 710                 /*
 711                  * It's possible to have interrupts off here:
 712                  */
 713                 local_irq_enable();
 714
 715                 /*
 716                  * Valid to do another page fault here because this one came
 717                  * from user space:
 718                  */
 719                 if (is_prefetch(regs, error_code, address))
 720                         return;
 721
 722                 if (is_errata100(regs, address))
 723                         return;
 724
 725                 if (unlikely(show_unhandled_signals))
 726                         show_signal_msg(regs, error_code, address, tsk);
 727
 728                 /* Kernel addresses are always protection faults: */
 729                 tsk->thread.cr2         = address;
 730                 tsk->thread.error_code  = error_code | (address >= TASK_SIZE);
 731                 tsk->thread.trap_no     = 14;
 732
 733                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 734
 735                 return;
 736         }
 737
 738         if (is_f00f_bug(regs, address))
 739                 return;
 740
 741         no_context(regs, error_code, address);
 742 }
 743
 744 static noinline void
 745 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 746                      unsigned long address)
 747 {
 748         __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 749 }
 750
 751 static void
 752 __bad_area(struct pt_regs *regs, unsigned long error_code,
 753            unsigned long address, int si_code)
 754 {
 755         struct mm_struct *mm = current->mm;
 756
 757         /*
 758          * Something tried to access memory that isn't in our memory map..
 759          * Fix it, but check if it's kernel or user first..
 760          */
 761         up_read(&mm->mmap_sem);
 762
 763         __bad_area_nosemaphore(regs, error_code, address, si_code);
 764 }
 765
 766 static noinline void
 767 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 768 {
 769         __bad_area(regs, error_code, address, SEGV_MAPERR);
 770 }
 771
 772 static noinline void
 773 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 774                       unsigned long address)
 775 {
 776         __bad_area(regs, error_code, address, SEGV_ACCERR);
 777 }
 778
 779 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
 780 static void
 781 out_of_memory(struct pt_regs *regs, unsigned long error_code,
 782               unsigned long address)
 783 {
 784         /*
 785          * We ran out of memory, call the OOM killer, and return the userspace
 786          * (which will retry the fault, or kill us if we got oom-killed):
 787          */
 788         up_read(&current->mm->mmap_sem);
 789
 790         pagefault_out_of_memory();
 791 }
 792
 793 static void
 794 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
 795           unsigned int fault)
 796 {
 797         struct task_struct *tsk = current;
 798         struct mm_struct *mm = tsk->mm;
 799         int code = BUS_ADRERR;
 800
 801         up_read(&mm->mmap_sem);
 802
 803         /* Kernel mode? Handle exceptions or die: */
 804         if (!(error_code & PF_USER))
 805                 no_context(regs, error_code, address);
 806
 807         /* User-space => ok to do another page fault: */
 808         if (is_prefetch(regs, error_code, address))
 809                 return;
 810
 811         tsk->thread.cr2         = address;
 812         tsk->thread.error_code  = error_code;
 813         tsk->thread.trap_no     = 14;
 814
 815 #ifdef CONFIG_MEMORY_FAILURE
 816         if (fault & VM_FAULT_HWPOISON) {
 817                 printk(KERN_ERR
 818         "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
 819                         tsk->comm, tsk->pid, address);
 820                 code = BUS_MCEERR_AR;
 821         }
 822 #endif
 823         force_sig_info_fault(SIGBUS, code, address, tsk);
 824 }
 825
 826 static noinline void
 827 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 828                unsigned long address, unsigned int fault)
 829 {
 830         if (fault & VM_FAULT_OOM) {
 831                 out_of_memory(regs, error_code, address);
 832         } else {
 833                 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON))
 834                         do_sigbus(regs, error_code, address, fault);
 835                 else
 836                         BUG();
 837         }
 838 }
 839
 840 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 841 {
 842         if ((error_code & PF_WRITE) && !pte_write(*pte))
 843                 return 0;
 844
 845         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 846                 return 0;
 847
 848         return 1;
 849 }
 850
 851 /*
 852  * Handle a spurious fault caused by a stale TLB entry.
 853  *
 854  * This allows us to lazily refresh the TLB when increasing the
 855  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
 856  * eagerly is very expensive since that implies doing a full
 857  * cross-processor TLB flush, even if no stale TLB entries exist
 858  * on other processors.
 859  *
 860  * There are no security implications to leaving a stale TLB when
 861  * increasing the permissions on a page.
 862  */
 863 static noinline int
 864 spurious_fault(unsigned long error_code, unsigned long address)
 865 {
 866         pgd_t *pgd;
 867         pud_t *pud;
 868         pmd_t *pmd;
 869         pte_t *pte;
 870         int ret;
 871
 872         /* Reserved-bit violation or user access to kernel space? */
 873         if (error_code & (PF_USER | PF_RSVD))
 874                 return 0;
 875
 876         pgd = init_mm.pgd + pgd_index(address);
 877         if (!pgd_present(*pgd))
 878                 return 0;
 879
 880         pud = pud_offset(pgd, address);
 881         if (!pud_present(*pud))
 882                 return 0;
 883
 884         if (pud_large(*pud))
 885                 return spurious_fault_check(error_code, (pte_t *) pud);
 886
 887         pmd = pmd_offset(pud, address);
 888         if (!pmd_present(*pmd))
 889                 return 0;
 890
 891         if (pmd_large(*pmd))
 892                 return spurious_fault_check(error_code, (pte_t *) pmd);
 893
 894         pte = pte_offset_kernel(pmd, address);
 895         if (!pte_present(*pte))
 896                 return 0;
 897
 898         ret = spurious_fault_check(error_code, pte);
 899         if (!ret)
 900                 return 0;
 901
 902         /*
 903          * Make sure we have permissions in PMD.
 904          * If not, then there's a bug in the page tables:
 905          */
 906         ret = spurious_fault_check(error_code, (pte_t *) pmd);
 907         WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
 908
 909         return ret;
 910 }
 911
 912 int show_unhandled_signals = 1;
 913
 914 static inline int
 915 access_error(unsigned long error_code, int write, struct vm_area_struct *vma)
 916 {
 917         if (write) {
 918                 /* write, present and write, not present: */
 919                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 920                         return 1;
 921                 return 0;
 922         }
 923
 924         /* read, present: */
 925         if (unlikely(error_code & PF_PROT))
 926                 return 1;
 927
 928         /* read, not present: */
 929         if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 930                 return 1;
 931
 932         return 0;
 933 }
 934
 935 static int fault_in_kernel_space(unsigned long address)
 936 {
 937         return address >= TASK_SIZE_MAX;
 938 }
 939
 940 /*
 941  * This routine handles page faults.  It determines the address,
 942  * and the problem, and then passes it off to one of the appropriate
 943  * routines.
 944  */
 945 dotraplinkage void __kprobes
 946 do_page_fault(struct pt_regs *regs, unsigned long error_code)
 947 {
 948         struct vm_area_struct *vma;
 949         struct task_struct *tsk;
 950         unsigned long address;
 951         struct mm_struct *mm;
 952         int write;
 953         int fault;
 954
 955         tsk = current;
 956         mm = tsk->mm;
 957
 958         /* Get the faulting address: */
 959         address = read_cr2();
 960
 961         /*
 962          * Detect and handle instructions that would cause a page fault for
 963          * both a tracked kernel page and a userspace page.
 964          */
 965         if (kmemcheck_active(regs))
 966                 kmemcheck_hide(regs);
 967         prefetchw(&mm->mmap_sem);
 968
 969         if (unlikely(kmmio_fault(regs, address)))
 970                 return;
 971
 972         /*
 973          * We fault-in kernel-space virtual memory on-demand. The
 974          * 'reference' page table is init_mm.pgd.
 975          *
 976          * NOTE! We MUST NOT take any locks for this case. We may
 977          * be in an interrupt or a critical region, and should
 978          * only copy the information from the master page table,
 979          * nothing more.
 980          *
 981          * This verifies that the fault happens in kernel space
 982          * (error_code & 4) == 0, and that the fault was not a
 983          * protection error (error_code & 9) == 0.
 984          */
 985         if (unlikely(fault_in_kernel_space(address))) {
 986                 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
 987                         if (vmalloc_fault(address) >= 0)
 988                                 return;
 989
 990                         if (kmemcheck_fault(regs, address, error_code))
 991                                 return;
 992                 }
 993
 994                 /* Can handle a stale RO->RW TLB: */
 995                 if (spurious_fault(error_code, address))
 996                         return;
 997
 998                 /* kprobes don't want to hook the spurious faults: */
 999                 if (notify_page_fault(regs))
1000                         return;
1001                 /*
1002                  * Don't take the mm semaphore here. If we fixup a prefetch
1003                  * fault we could otherwise deadlock:
1004                  */
1005                 bad_area_nosemaphore(regs, error_code, address);
1006
1007                 return;
1008         }
1009
1010         /* kprobes don't want to hook the spurious faults: */
1011         if (unlikely(notify_page_fault(regs)))
1012                 return;
1013         /*
1014          * It's safe to allow irq's after cr2 has been saved and the
1015          * vmalloc fault has been handled.
1016          *
1017          * User-mode registers count as a user access even for any
1018          * potential system fault or CPU buglet:
1019          */
1020         if (user_mode_vm(regs)) {
1021                 local_irq_enable();
1022                 error_code |= PF_USER;
1023         } else {
1024                 if (regs->flags & X86_EFLAGS_IF)
1025                         local_irq_enable();
1026         }
1027
1028         if (unlikely(error_code & PF_RSVD))
1029                 pgtable_bad(regs, error_code, address);
1030
1031         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1032
1033         /*
1034          * If we're in an interrupt, have no user context or are running
1035          * in an atomic region then we must not take the fault:
1036          */
1037         if (unlikely(in_atomic() || !mm)) {
1038                 bad_area_nosemaphore(regs, error_code, address);
1039                 return;
1040         }
1041
1042         /*
1043          * When running in the kernel we expect faults to occur only to
1044          * addresses in user space.  All other faults represent errors in
1045          * the kernel and should generate an OOPS.  Unfortunately, in the
1046          * case of an erroneous fault occurring in a code path which already
1047          * holds mmap_sem we will deadlock attempting to validate the fault
1048          * against the address space.  Luckily the kernel only validly
1049          * references user space from well defined areas of code, which are
1050          * listed in the exceptions table.
1051          *
1052          * As the vast majority of faults will be valid we will only perform
1053          * the source reference check when there is a possibility of a
1054          * deadlock. Attempt to lock the address space, if we cannot we then
1055          * validate the source. If this is invalid we can skip the address
1056          * space check, thus avoiding the deadlock:
1057          */
1058         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1059                 if ((error_code & PF_USER) == 0 &&
1060                     !search_exception_tables(regs->ip)) {
1061                         bad_area_nosemaphore(regs, error_code, address);
1062                         return;
1063                 }
1064                 down_read(&mm->mmap_sem);
1065         } else {
1066                 /*
1067                  * The above down_read_trylock() might have succeeded in
1068                  * which case we'll have missed the might_sleep() from
1069                  * down_read():
1070                  */
1071                 might_sleep();
1072         }
1073
1074         vma = find_vma(mm, address);
1075         if (unlikely(!vma)) {
1076                 bad_area(regs, error_code, address);
1077                 return;
1078         }
1079         if (likely(vma->vm_start <= address))
1080                 goto good_area;
1081         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1082                 bad_area(regs, error_code, address);
1083                 return;
1084         }
1085         if (error_code & PF_USER) {
1086                 /*
1087                  * Accessing the stack below %sp is always a bug.
1088                  * The large cushion allows instructions like enter
1089                  * and pusha to work. ("enter $65535, $31" pushes
1090                  * 32 pointers and then decrements %sp by 65535.)
1091                  */
1092                 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1093                         bad_area(regs, error_code, address);
1094                         return;
1095                 }
1096         }
1097         if (unlikely(expand_stack(vma, address))) {
1098                 bad_area(regs, error_code, address);
1099                 return;
1100         }
1101
1102         /*
1103          * Ok, we have a good vm_area for this memory access, so
1104          * we can handle it..
1105          */
1106 good_area:
1107         write = error_code & PF_WRITE;
1108
1109         if (unlikely(access_error(error_code, write, vma))) {
1110                 bad_area_access_error(regs, error_code, address);
1111                 return;
1112         }
1113
1114         /*
1115          * If for any reason at all we couldn't handle the fault,
1116          * make sure we exit gracefully rather than endlessly redo
1117          * the fault:
1118          */
1119         fault = handle_mm_fault(mm, vma, address, write ? FAULT_FLAG_WRITE : 0);
1120
1121         if (unlikely(fault & VM_FAULT_ERROR)) {
1122                 mm_fault_error(regs, error_code, address, fault);
1123                 return;
1124         }
1125
1126         if (fault & VM_FAULT_MAJOR) {
1127                 tsk->maj_flt++;
1128                 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1129                                      regs, address);
1130         } else {
1131                 tsk->min_flt++;
1132                 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1133                                      regs, address);
1134         }
1135
1136         check_v8086_mode(regs, address, tsk);
1137
1138         up_read(&mm->mmap_sem);
1139 }