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 #include <linux/hugetlb.h>              /* hstate_index_to_shift        */
  15
  16 #include <asm/traps.h>                  /* dotraplinkage, ...           */
  17 #include <asm/pgalloc.h>                /* pgd_*(), ...                 */
  18 #include <asm/kmemcheck.h>              /* kmemcheck_*(), ...           */
  19
  20 /*
  21  * Page fault error code bits:
  22  *
  23  *   bit 0 ==    0: no page found       1: protection fault
  24  *   bit 1 ==    0: read access         1: write access
  25  *   bit 2 ==    0: kernel-mode access  1: user-mode access
  26  *   bit 3 ==                           1: use of reserved bit detected
  27  *   bit 4 ==                           1: fault was an instruction fetch
  28  */
  29 enum x86_pf_error_code {
  30
  31         PF_PROT         =               1 << 0,
  32         PF_WRITE        =               1 << 1,
  33         PF_USER         =               1 << 2,
  34         PF_RSVD         =               1 << 3,
  35         PF_INSTR        =               1 << 4,
  36 };
  37
  38 /*
  39  * Returns 0 if mmiotrace is disabled, or if the fault is not
  40  * handled by mmiotrace:
  41  */
  42 static inline int __kprobes
  43 kmmio_fault(struct pt_regs *regs, unsigned long addr)
  44 {
  45         if (unlikely(is_kmmio_active()))
  46                 if (kmmio_handler(regs, addr) == 1)
  47                         return -1;
  48         return 0;
  49 }
  50
  51 static inline int __kprobes notify_page_fault(struct pt_regs *regs)
  52 {
  53         int ret = 0;
  54
  55         /* kprobe_running() needs smp_processor_id() */
  56         if (kprobes_built_in() && !user_mode_vm(regs)) {
  57                 preempt_disable();
  58                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  59                         ret = 1;
  60                 preempt_enable();
  61         }
  62
  63         return ret;
  64 }
  65
  66 /*
  67  * Prefetch quirks:
  68  *
  69  * 32-bit mode:
  70  *
  71  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  72  *   Check that here and ignore it.
  73  *
  74  * 64-bit mode:
  75  *
  76  *   Sometimes the CPU reports invalid exceptions on prefetch.
  77  *   Check that here and ignore it.
  78  *
  79  * Opcode checker based on code by Richard Brunner.
  80  */
  81 static inline int
  82 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  83                       unsigned char opcode, int *prefetch)
  84 {
  85         unsigned char instr_hi = opcode & 0xf0;
  86         unsigned char instr_lo = opcode & 0x0f;
  87
  88         switch (instr_hi) {
  89         case 0x20:
  90         case 0x30:
  91                 /*
  92                  * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
  93                  * In X86_64 long mode, the CPU will signal invalid
  94                  * opcode if some of these prefixes are present so
  95                  * X86_64 will never get here anyway
  96                  */
  97                 return ((instr_lo & 7) == 0x6);
  98 #ifdef CONFIG_X86_64
  99         case 0x40:
 100                 /*
 101                  * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 102                  * Need to figure out under what instruction mode the
 103                  * instruction was issued. Could check the LDT for lm,
 104                  * but for now it's good enough to assume that long
 105                  * mode only uses well known segments or kernel.
 106                  */
 107                 return (!user_mode(regs)) || (regs->cs == __USER_CS);
 108 #endif
 109         case 0x60:
 110                 /* 0x64 thru 0x67 are valid prefixes in all modes. */
 111                 return (instr_lo & 0xC) == 0x4;
 112         case 0xF0:
 113                 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 114                 return !instr_lo || (instr_lo>>1) == 1;
 115         case 0x00:
 116                 /* Prefetch instruction is 0x0F0D or 0x0F18 */
 117                 if (probe_kernel_address(instr, opcode))
 118                         return 0;
 119
 120                 *prefetch = (instr_lo == 0xF) &&
 121                         (opcode == 0x0D || opcode == 0x18);
 122                 return 0;
 123         default:
 124                 return 0;
 125         }
 126 }
 127
 128 static int
 129 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 130 {
 131         unsigned char *max_instr;
 132         unsigned char *instr;
 133         int prefetch = 0;
 134
 135         /*
 136          * If it was a exec (instruction fetch) fault on NX page, then
 137          * do not ignore the fault:
 138          */
 139         if (error_code & PF_INSTR)
 140                 return 0;
 141
 142         instr = (void *)convert_ip_to_linear(current, regs);
 143         max_instr = instr + 15;
 144
 145         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 146                 return 0;
 147
 148         while (instr < max_instr) {
 149                 unsigned char opcode;
 150
 151                 if (probe_kernel_address(instr, opcode))
 152                         break;
 153
 154                 instr++;
 155
 156                 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 157                         break;
 158         }
 159         return prefetch;
 160 }
 161
 162 static void
 163 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
 164                      struct task_struct *tsk, int fault)
 165 {
 166         unsigned lsb = 0;
 167         siginfo_t info;
 168
 169         info.si_signo   = si_signo;
 170         info.si_errno   = 0;
 171         info.si_code    = si_code;
 172         info.si_addr    = (void __user *)address;
 173         if (fault & VM_FAULT_HWPOISON_LARGE)
 174                 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
 175         if (fault & VM_FAULT_HWPOISON)
 176                 lsb = PAGE_SHIFT;
 177         info.si_addr_lsb = lsb;
 178
 179         force_sig_info(si_signo, &info, tsk);
 180 }
 181
 182 DEFINE_SPINLOCK(pgd_lock);
 183 LIST_HEAD(pgd_list);
 184
 185 #ifdef CONFIG_X86_32
 186 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 187 {
 188         unsigned index = pgd_index(address);
 189         pgd_t *pgd_k;
 190         pud_t *pud, *pud_k;
 191         pmd_t *pmd, *pmd_k;
 192
 193         pgd += index;
 194         pgd_k = init_mm.pgd + index;
 195
 196         if (!pgd_present(*pgd_k))
 197                 return NULL;
 198
 199         /*
 200          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 201          * and redundant with the set_pmd() on non-PAE. As would
 202          * set_pud.
 203          */
 204         pud = pud_offset(pgd, address);
 205         pud_k = pud_offset(pgd_k, address);
 206         if (!pud_present(*pud_k))
 207                 return NULL;
 208
 209         pmd = pmd_offset(pud, address);
 210         pmd_k = pmd_offset(pud_k, address);
 211         if (!pmd_present(*pmd_k))
 212                 return NULL;
 213
 214         if (!pmd_present(*pmd))
 215                 set_pmd(pmd, *pmd_k);
 216         else
 217                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 218
 219         return pmd_k;
 220 }
 221
 222 void vmalloc_sync_all(void)
 223 {
 224         unsigned long address;
 225
 226         if (SHARED_KERNEL_PMD)
 227                 return;
 228
 229         for (address = VMALLOC_START & PMD_MASK;
 230              address >= TASK_SIZE && address < FIXADDR_TOP;
 231              address += PMD_SIZE) {
 232                 struct page *page;
 233
 234                 spin_lock(&pgd_lock);
 235                 list_for_each_entry(page, &pgd_list, lru) {
 236                         spinlock_t *pgt_lock;
 237                         pmd_t *ret;
 238
 239                         /* the pgt_lock only for Xen */
 240                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 241
 242                         spin_lock(pgt_lock);
 243                         ret = vmalloc_sync_one(page_address(page), address);
 244                         spin_unlock(pgt_lock);
 245
 246                         if (!ret)
 247                                 break;
 248                 }
 249                 spin_unlock(&pgd_lock);
 250         }
 251 }
 252
 253 /*
 254  * 32-bit:
 255  *
 256  *   Handle a fault on the vmalloc or module mapping area
 257  */
 258 static noinline __kprobes int vmalloc_fault(unsigned long address)
 259 {
 260         unsigned long pgd_paddr;
 261         pmd_t *pmd_k;
 262         pte_t *pte_k;
 263
 264         /* Make sure we are in vmalloc area: */
 265         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 266                 return -1;
 267
 268         WARN_ON_ONCE(in_nmi());
 269
 270         /*
 271          * Synchronize this task's top level page-table
 272          * with the 'reference' page table.
 273          *
 274          * Do _not_ use "current" here. We might be inside
 275          * an interrupt in the middle of a task switch..
 276          */
 277         pgd_paddr = read_cr3();
 278         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 279         if (!pmd_k)
 280                 return -1;
 281
 282         pte_k = pte_offset_kernel(pmd_k, address);
 283         if (!pte_present(*pte_k))
 284                 return -1;
 285
 286         return 0;
 287 }
 288
 289 /*
 290  * Did it hit the DOS screen memory VA from vm86 mode?
 291  */
 292 static inline void
 293 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 294                  struct task_struct *tsk)
 295 {
 296         unsigned long bit;
 297
 298         if (!v8086_mode(regs))
 299                 return;
 300
 301         bit = (address - 0xA0000) >> PAGE_SHIFT;
 302         if (bit < 32)
 303                 tsk->thread.screen_bitmap |= 1 << bit;
 304 }
 305
 306 static bool low_pfn(unsigned long pfn)
 307 {
 308         return pfn < max_low_pfn;
 309 }
 310
 311 static void dump_pagetable(unsigned long address)
 312 {
 313         pgd_t *base = __va(read_cr3());
 314         pgd_t *pgd = &base[pgd_index(address)];
 315         pmd_t *pmd;
 316         pte_t *pte;
 317
 318 #ifdef CONFIG_X86_PAE
 319         printk("*pdpt = %016Lx ", pgd_val(*pgd));
 320         if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
 321                 goto out;
 322 #endif
 323         pmd = pmd_offset(pud_offset(pgd, address), address);
 324         printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
 325
 326         /*
 327          * We must not directly access the pte in the highpte
 328          * case if the page table is located in highmem.
 329          * And let's rather not kmap-atomic the pte, just in case
 330          * it's allocated already:
 331          */
 332         if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
 333                 goto out;
 334
 335         pte = pte_offset_kernel(pmd, address);
 336         printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
 337 out:
 338         printk("\n");
 339 }
 340
 341 #else /* CONFIG_X86_64: */
 342
 343 void vmalloc_sync_all(void)
 344 {
 345         sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
 346 }
 347
 348 /*
 349  * 64-bit:
 350  *
 351  *   Handle a fault on the vmalloc area
 352  *
 353  * This assumes no large pages in there.
 354  */
 355 static noinline __kprobes int vmalloc_fault(unsigned long address)
 356 {
 357         pgd_t *pgd, *pgd_ref;
 358         pud_t *pud, *pud_ref;
 359         pmd_t *pmd, *pmd_ref;
 360         pte_t *pte, *pte_ref;
 361
 362         /* Make sure we are in vmalloc area: */
 363         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 364                 return -1;
 365
 366         WARN_ON_ONCE(in_nmi());
 367
 368         /*
 369          * Copy kernel mappings over when needed. This can also
 370          * happen within a race in page table update. In the later
 371          * case just flush:
 372          */
 373         pgd = pgd_offset(current->active_mm, address);
 374         pgd_ref = pgd_offset_k(address);
 375         if (pgd_none(*pgd_ref))
 376                 return -1;
 377
 378         if (pgd_none(*pgd))
 379                 set_pgd(pgd, *pgd_ref);
 380         else
 381                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 382
 383         /*
 384          * Below here mismatches are bugs because these lower tables
 385          * are shared:
 386          */
 387
 388         pud = pud_offset(pgd, address);
 389         pud_ref = pud_offset(pgd_ref, address);
 390         if (pud_none(*pud_ref))
 391                 return -1;
 392
 393         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 394                 BUG();
 395
 396         pmd = pmd_offset(pud, address);
 397         pmd_ref = pmd_offset(pud_ref, address);
 398         if (pmd_none(*pmd_ref))
 399                 return -1;
 400
 401         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 402                 BUG();
 403
 404         pte_ref = pte_offset_kernel(pmd_ref, address);
 405         if (!pte_present(*pte_ref))
 406                 return -1;
 407
 408         pte = pte_offset_kernel(pmd, address);
 409
 410         /*
 411          * Don't use pte_page here, because the mappings can point
 412          * outside mem_map, and the NUMA hash lookup cannot handle
 413          * that:
 414          */
 415         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 416                 BUG();
 417
 418         return 0;
 419 }
 420
 421 static const char errata93_warning[] =
 422 KERN_ERR
 423 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 424 "******* Working around it, but it may cause SEGVs or burn power.\n"
 425 "******* Please consider a BIOS update.\n"
 426 "******* Disabling USB legacy in the BIOS may also help.\n";
 427
 428 /*
 429  * No vm86 mode in 64-bit mode:
 430  */
 431 static inline void
 432 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 433                  struct task_struct *tsk)
 434 {
 435 }
 436
 437 static int bad_address(void *p)
 438 {
 439         unsigned long dummy;
 440
 441         return probe_kernel_address((unsigned long *)p, dummy);
 442 }
 443
 444 static void dump_pagetable(unsigned long address)
 445 {
 446         pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
 447         pgd_t *pgd = base + pgd_index(address);
 448         pud_t *pud;
 449         pmd_t *pmd;
 450         pte_t *pte;
 451
 452         if (bad_address(pgd))
 453                 goto bad;
 454
 455         printk("PGD %lx ", pgd_val(*pgd));
 456
 457         if (!pgd_present(*pgd))
 458                 goto out;
 459
 460         pud = pud_offset(pgd, address);
 461         if (bad_address(pud))
 462                 goto bad;
 463
 464         printk("PUD %lx ", pud_val(*pud));
 465         if (!pud_present(*pud) || pud_large(*pud))
 466                 goto out;
 467
 468         pmd = pmd_offset(pud, address);
 469         if (bad_address(pmd))
 470                 goto bad;
 471
 472         printk("PMD %lx ", pmd_val(*pmd));
 473         if (!pmd_present(*pmd) || pmd_large(*pmd))
 474                 goto out;
 475
 476         pte = pte_offset_kernel(pmd, address);
 477         if (bad_address(pte))
 478                 goto bad;
 479
 480         printk("PTE %lx", pte_val(*pte));
 481 out:
 482         printk("\n");
 483         return;
 484 bad:
 485         printk("BAD\n");
 486 }
 487
 488 #endif /* CONFIG_X86_64 */
 489
 490 /*
 491  * Workaround for K8 erratum #93 & buggy BIOS.
 492  *
 493  * BIOS SMM functions are required to use a specific workaround
 494  * to avoid corruption of the 64bit RIP register on C stepping K8.
 495  *
 496  * A lot of BIOS that didn't get tested properly miss this.
 497  *
 498  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 499  * Try to work around it here.
 500  *
 501  * Note we only handle faults in kernel here.
 502  * Does nothing on 32-bit.
 503  */
 504 static int is_errata93(struct pt_regs *regs, unsigned long address)
 505 {
 506 #ifdef CONFIG_X86_64
 507         if (address != regs->ip)
 508                 return 0;
 509
 510         if ((address >> 32) != 0)
 511                 return 0;
 512
 513         address |= 0xffffffffUL << 32;
 514         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 515             (address >= MODULES_VADDR && address <= MODULES_END)) {
 516                 printk_once(errata93_warning);
 517                 regs->ip = address;
 518                 return 1;
 519         }
 520 #endif
 521         return 0;
 522 }
 523
 524 /*
 525  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 526  * to illegal addresses >4GB.
 527  *
 528  * We catch this in the page fault handler because these addresses
 529  * are not reachable. Just detect this case and return.  Any code
 530  * segment in LDT is compatibility mode.
 531  */
 532 static int is_errata100(struct pt_regs *regs, unsigned long address)
 533 {
 534 #ifdef CONFIG_X86_64
 535         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 536                 return 1;
 537 #endif
 538         return 0;
 539 }
 540
 541 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 542 {
 543 #ifdef CONFIG_X86_F00F_BUG
 544         unsigned long nr;
 545
 546         /*
 547          * Pentium F0 0F C7 C8 bug workaround:
 548          */
 549         if (boot_cpu_data.f00f_bug) {
 550                 nr = (address - idt_descr.address) >> 3;
 551
 552                 if (nr == 6) {
 553                         do_invalid_op(regs, 0);
 554                         return 1;
 555                 }
 556         }
 557 #endif
 558         return 0;
 559 }
 560
 561 static const char nx_warning[] = KERN_CRIT
 562 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
 563
 564 static void
 565 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 566                 unsigned long address)
 567 {
 568         if (!oops_may_print())
 569                 return;
 570
 571         if (error_code & PF_INSTR) {
 572                 unsigned int level;
 573
 574                 pte_t *pte = lookup_address(address, &level);
 575
 576                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 577                         printk(nx_warning, current_uid());
 578         }
 579
 580         printk(KERN_ALERT "BUG: unable to handle kernel ");
 581         if (address < PAGE_SIZE)
 582                 printk(KERN_CONT "NULL pointer dereference");
 583         else
 584                 printk(KERN_CONT "paging request");
 585
 586         printk(KERN_CONT " at %p\n", (void *) address);
 587         printk(KERN_ALERT "IP:");
 588         printk_address(regs->ip, 1);
 589
 590         dump_pagetable(address);
 591 }
 592
 593 static noinline void
 594 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 595             unsigned long address)
 596 {
 597         struct task_struct *tsk;
 598         unsigned long flags;
 599         int sig;
 600
 601         flags = oops_begin();
 602         tsk = current;
 603         sig = SIGKILL;
 604
 605         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 606                tsk->comm, address);
 607         dump_pagetable(address);
 608
 609         tsk->thread.cr2         = address;
 610         tsk->thread.trap_no     = 14;
 611         tsk->thread.error_code  = error_code;
 612
 613         if (__die("Bad pagetable", regs, error_code))
 614                 sig = 0;
 615
 616         oops_end(flags, regs, sig);
 617 }
 618
 619 static noinline void
 620 no_context(struct pt_regs *regs, unsigned long error_code,
 621            unsigned long address)
 622 {
 623         struct task_struct *tsk = current;
 624         unsigned long *stackend;
 625         unsigned long flags;
 626         int sig;
 627
 628         /* Are we prepared to handle this kernel fault? */
 629         if (fixup_exception(regs))
 630                 return;
 631
 632         /*
 633          * 32-bit:
 634          *
 635          *   Valid to do another page fault here, because if this fault
 636          *   had been triggered by is_prefetch fixup_exception would have
 637          *   handled it.
 638          *
 639          * 64-bit:
 640          *
 641          *   Hall of shame of CPU/BIOS bugs.
 642          */
 643         if (is_prefetch(regs, error_code, address))
 644                 return;
 645
 646         if (is_errata93(regs, address))
 647                 return;
 648
 649         /*
 650          * Oops. The kernel tried to access some bad page. We'll have to
 651          * terminate things with extreme prejudice:
 652          */
 653         flags = oops_begin();
 654
 655         show_fault_oops(regs, error_code, address);
 656
 657         stackend = end_of_stack(tsk);
 658         if (tsk != &init_task && *stackend != STACK_END_MAGIC)
 659                 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
 660
 661         tsk->thread.cr2         = address;
 662         tsk->thread.trap_no     = 14;
 663         tsk->thread.error_code  = error_code;
 664
 665         sig = SIGKILL;
 666         if (__die("Oops", regs, error_code))
 667                 sig = 0;
 668
 669         /* Executive summary in case the body of the oops scrolled away */
 670         printk(KERN_EMERG "CR2: %016lx\n", address);
 671
 672         oops_end(flags, regs, sig);
 673 }
 674
 675 /*
 676  * Print out info about fatal segfaults, if the show_unhandled_signals
 677  * sysctl is set:
 678  */
 679 static inline void
 680 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 681                 unsigned long address, struct task_struct *tsk)
 682 {
 683         if (!unhandled_signal(tsk, SIGSEGV))
 684                 return;
 685
 686         if (!printk_ratelimit())
 687                 return;
 688
 689         printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 690                 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 691                 tsk->comm, task_pid_nr(tsk), address,
 692                 (void *)regs->ip, (void *)regs->sp, error_code);
 693
 694         print_vma_addr(KERN_CONT " in ", regs->ip);
 695
 696         printk(KERN_CONT "\n");
 697 }
 698
 699 static void
 700 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 701                        unsigned long address, int si_code)
 702 {
 703         struct task_struct *tsk = current;
 704
 705         /* User mode accesses just cause a SIGSEGV */
 706         if (error_code & PF_USER) {
 707                 /*
 708                  * It's possible to have interrupts off here:
 709                  */
 710                 local_irq_enable();
 711
 712                 /*
 713                  * Valid to do another page fault here because this one came
 714                  * from user space:
 715                  */
 716                 if (is_prefetch(regs, error_code, address))
 717                         return;
 718
 719                 if (is_errata100(regs, address))
 720                         return;
 721
 722                 if (unlikely(show_unhandled_signals))
 723                         show_signal_msg(regs, error_code, address, tsk);
 724
 725                 /* Kernel addresses are always protection faults: */
 726                 tsk->thread.cr2         = address;
 727                 tsk->thread.error_code  = error_code | (address >= TASK_SIZE);
 728                 tsk->thread.trap_no     = 14;
 729
 730                 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
 731
 732                 return;
 733         }
 734
 735         if (is_f00f_bug(regs, address))
 736                 return;
 737
 738         no_context(regs, error_code, address);
 739 }
 740
 741 static noinline void
 742 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 743                      unsigned long address)
 744 {
 745         __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 746 }
 747
 748 static void
 749 __bad_area(struct pt_regs *regs, unsigned long error_code,
 750            unsigned long address, int si_code)
 751 {
 752         struct mm_struct *mm = current->mm;
 753
 754         /*
 755          * Something tried to access memory that isn't in our memory map..
 756          * Fix it, but check if it's kernel or user first..
 757          */
 758         up_read(&mm->mmap_sem);
 759
 760         __bad_area_nosemaphore(regs, error_code, address, si_code);
 761 }
 762
 763 static noinline void
 764 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 765 {
 766         __bad_area(regs, error_code, address, SEGV_MAPERR);
 767 }
 768
 769 static noinline void
 770 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 771                       unsigned long address)
 772 {
 773         __bad_area(regs, error_code, address, SEGV_ACCERR);
 774 }
 775
 776 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
 777 static void
 778 out_of_memory(struct pt_regs *regs, unsigned long error_code,
 779               unsigned long address)
 780 {
 781         /*
 782          * We ran out of memory, call the OOM killer, and return the userspace
 783          * (which will retry the fault, or kill us if we got oom-killed):
 784          */
 785         up_read(&current->mm->mmap_sem);
 786
 787         pagefault_out_of_memory();
 788 }
 789
 790 static void
 791 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
 792           unsigned int fault)
 793 {
 794         struct task_struct *tsk = current;
 795         struct mm_struct *mm = tsk->mm;
 796         int code = BUS_ADRERR;
 797
 798         up_read(&mm->mmap_sem);
 799
 800         /* Kernel mode? Handle exceptions or die: */
 801         if (!(error_code & PF_USER)) {
 802                 no_context(regs, error_code, address);
 803                 return;
 804         }
 805
 806         /* User-space => ok to do another page fault: */
 807         if (is_prefetch(regs, error_code, address))
 808                 return;
 809
 810         tsk->thread.cr2         = address;
 811         tsk->thread.error_code  = error_code;
 812         tsk->thread.trap_no     = 14;
 813
 814 #ifdef CONFIG_MEMORY_FAILURE
 815         if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
 816                 printk(KERN_ERR
 817         "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
 818                         tsk->comm, tsk->pid, address);
 819                 code = BUS_MCEERR_AR;
 820         }
 821 #endif
 822         force_sig_info_fault(SIGBUS, code, address, tsk, fault);
 823 }
 824
 825 static noinline void
 826 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 827                unsigned long address, unsigned int fault)
 828 {
 829         if (fault & VM_FAULT_OOM) {
 830                 /* Kernel mode? Handle exceptions or die: */
 831                 if (!(error_code & PF_USER)) {
 832                         up_read(&current->mm->mmap_sem);
 833                         no_context(regs, error_code, address);
 834                         return;
 835                 }
 836
 837                 out_of_memory(regs, error_code, address);
 838         } else {
 839                 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
 840                              VM_FAULT_HWPOISON_LARGE))
 841                         do_sigbus(regs, error_code, address, fault);
 842                 else
 843                         BUG();
 844         }
 845 }
 846
 847 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 848 {
 849         if ((error_code & PF_WRITE) && !pte_write(*pte))
 850                 return 0;
 851
 852         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 853                 return 0;
 854
 855         return 1;
 856 }
 857
 858 /*
 859  * Handle a spurious fault caused by a stale TLB entry.
 860  *
 861  * This allows us to lazily refresh the TLB when increasing the
 862  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
 863  * eagerly is very expensive since that implies doing a full
 864  * cross-processor TLB flush, even if no stale TLB entries exist
 865  * on other processors.
 866  *
 867  * There are no security implications to leaving a stale TLB when
 868  * increasing the permissions on a page.
 869  */
 870 static noinline __kprobes int
 871 spurious_fault(unsigned long error_code, unsigned long address)
 872 {
 873         pgd_t *pgd;
 874         pud_t *pud;
 875         pmd_t *pmd;
 876         pte_t *pte;
 877         int ret;
 878
 879         /* Reserved-bit violation or user access to kernel space? */
 880         if (error_code & (PF_USER | PF_RSVD))
 881                 return 0;
 882
 883         pgd = init_mm.pgd + pgd_index(address);
 884         if (!pgd_present(*pgd))
 885                 return 0;
 886
 887         pud = pud_offset(pgd, address);
 888         if (!pud_present(*pud))
 889                 return 0;
 890
 891         if (pud_large(*pud))
 892                 return spurious_fault_check(error_code, (pte_t *) pud);
 893
 894         pmd = pmd_offset(pud, address);
 895         if (!pmd_present(*pmd))
 896                 return 0;
 897
 898         if (pmd_large(*pmd))
 899                 return spurious_fault_check(error_code, (pte_t *) pmd);
 900
 901         /*
 902          * Note: don't use pte_present() here, since it returns true
 903          * if the _PAGE_PROTNONE bit is set.  However, this aliases the
 904          * _PAGE_GLOBAL bit, which for kernel pages give false positives
 905          * when CONFIG_DEBUG_PAGEALLOC is used.
 906          */
 907         pte = pte_offset_kernel(pmd, address);
 908         if (!(pte_flags(*pte) & _PAGE_PRESENT))
 909                 return 0;
 910
 911         ret = spurious_fault_check(error_code, pte);
 912         if (!ret)
 913                 return 0;
 914
 915         /*
 916          * Make sure we have permissions in PMD.
 917          * If not, then there's a bug in the page tables:
 918          */
 919         ret = spurious_fault_check(error_code, (pte_t *) pmd);
 920         WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
 921
 922         return ret;
 923 }
 924
 925 int show_unhandled_signals = 1;
 926
 927 static inline int
 928 access_error(unsigned long error_code, struct vm_area_struct *vma)
 929 {
 930         if (error_code & PF_WRITE) {
 931                 /* write, present and write, not present: */
 932                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 933                         return 1;
 934                 return 0;
 935         }
 936
 937         /* read, present: */
 938         if (unlikely(error_code & PF_PROT))
 939                 return 1;
 940
 941         /* read, not present: */
 942         if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 943                 return 1;
 944
 945         return 0;
 946 }
 947
 948 static int fault_in_kernel_space(unsigned long address)
 949 {
 950         return address >= TASK_SIZE_MAX;
 951 }
 952
 953 /*
 954  * This routine handles page faults.  It determines the address,
 955  * and the problem, and then passes it off to one of the appropriate
 956  * routines.
 957  */
 958 dotraplinkage void __kprobes
 959 do_page_fault(struct pt_regs *regs, unsigned long error_code)
 960 {
 961         struct vm_area_struct *vma;
 962         struct task_struct *tsk;
 963         unsigned long address;
 964         struct mm_struct *mm;
 965         int fault;
 966         int write = error_code & PF_WRITE;
 967         unsigned int flags = FAULT_FLAG_ALLOW_RETRY |
 968                                         (write ? FAULT_FLAG_WRITE : 0);
 969
 970         tsk = current;
 971         mm = tsk->mm;
 972
 973         /* Get the faulting address: */
 974         address = read_cr2();
 975
 976         /*
 977          * Detect and handle instructions that would cause a page fault for
 978          * both a tracked kernel page and a userspace page.
 979          */
 980         if (kmemcheck_active(regs))
 981                 kmemcheck_hide(regs);
 982         prefetchw(&mm->mmap_sem);
 983
 984         if (unlikely(kmmio_fault(regs, address)))
 985                 return;
 986
 987         /*
 988          * We fault-in kernel-space virtual memory on-demand. The
 989          * 'reference' page table is init_mm.pgd.
 990          *
 991          * NOTE! We MUST NOT take any locks for this case. We may
 992          * be in an interrupt or a critical region, and should
 993          * only copy the information from the master page table,
 994          * nothing more.
 995          *
 996          * This verifies that the fault happens in kernel space
 997          * (error_code & 4) == 0, and that the fault was not a
 998          * protection error (error_code & 9) == 0.
 999          */
1000         if (unlikely(fault_in_kernel_space(address))) {
1001                 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1002                         if (vmalloc_fault(address) >= 0)
1003                                 return;
1004
1005                         if (kmemcheck_fault(regs, address, error_code))
1006                                 return;
1007                 }
1008
1009                 /* Can handle a stale RO->RW TLB: */
1010                 if (spurious_fault(error_code, address))
1011                         return;
1012
1013                 /* kprobes don't want to hook the spurious faults: */
1014                 if (notify_page_fault(regs))
1015                         return;
1016                 /*
1017                  * Don't take the mm semaphore here. If we fixup a prefetch
1018                  * fault we could otherwise deadlock:
1019                  */
1020                 bad_area_nosemaphore(regs, error_code, address);
1021
1022                 return;
1023         }
1024
1025         /* kprobes don't want to hook the spurious faults: */
1026         if (unlikely(notify_page_fault(regs)))
1027                 return;
1028         /*
1029          * It's safe to allow irq's after cr2 has been saved and the
1030          * vmalloc fault has been handled.
1031          *
1032          * User-mode registers count as a user access even for any
1033          * potential system fault or CPU buglet:
1034          */
1035         if (user_mode_vm(regs)) {
1036                 local_irq_enable();
1037                 error_code |= PF_USER;
1038         } else {
1039                 if (regs->flags & X86_EFLAGS_IF)
1040                         local_irq_enable();
1041         }
1042
1043         if (unlikely(error_code & PF_RSVD))
1044                 pgtable_bad(regs, error_code, address);
1045
1046         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1047
1048         /*
1049          * If we're in an interrupt, have no user context or are running
1050          * in an atomic region then we must not take the fault:
1051          */
1052         if (unlikely(in_atomic() || !mm)) {
1053                 bad_area_nosemaphore(regs, error_code, address);
1054                 return;
1055         }
1056
1057         /*
1058          * When running in the kernel we expect faults to occur only to
1059          * addresses in user space.  All other faults represent errors in
1060          * the kernel and should generate an OOPS.  Unfortunately, in the
1061          * case of an erroneous fault occurring in a code path which already
1062          * holds mmap_sem we will deadlock attempting to validate the fault
1063          * against the address space.  Luckily the kernel only validly
1064          * references user space from well defined areas of code, which are
1065          * listed in the exceptions table.
1066          *
1067          * As the vast majority of faults will be valid we will only perform
1068          * the source reference check when there is a possibility of a
1069          * deadlock. Attempt to lock the address space, if we cannot we then
1070          * validate the source. If this is invalid we can skip the address
1071          * space check, thus avoiding the deadlock:
1072          */
1073         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1074                 if ((error_code & PF_USER) == 0 &&
1075                     !search_exception_tables(regs->ip)) {
1076                         bad_area_nosemaphore(regs, error_code, address);
1077                         return;
1078                 }
1079 retry:
1080                 down_read(&mm->mmap_sem);
1081         } else {
1082                 /*
1083                  * The above down_read_trylock() might have succeeded in
1084                  * which case we'll have missed the might_sleep() from
1085                  * down_read():
1086                  */
1087                 might_sleep();
1088         }
1089
1090         vma = find_vma(mm, address);
1091         if (unlikely(!vma)) {
1092                 bad_area(regs, error_code, address);
1093                 return;
1094         }
1095         if (likely(vma->vm_start <= address))
1096                 goto good_area;
1097         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1098                 bad_area(regs, error_code, address);
1099                 return;
1100         }
1101         if (error_code & PF_USER) {
1102                 /*
1103                  * Accessing the stack below %sp is always a bug.
1104                  * The large cushion allows instructions like enter
1105                  * and pusha to work. ("enter $65535, $31" pushes
1106                  * 32 pointers and then decrements %sp by 65535.)
1107                  */
1108                 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1109                         bad_area(regs, error_code, address);
1110                         return;
1111                 }
1112         }
1113         if (unlikely(expand_stack(vma, address))) {
1114                 bad_area(regs, error_code, address);
1115                 return;
1116         }
1117
1118         /*
1119          * Ok, we have a good vm_area for this memory access, so
1120          * we can handle it..
1121          */
1122 good_area:
1123         if (unlikely(access_error(error_code, vma))) {
1124                 bad_area_access_error(regs, error_code, address);
1125                 return;
1126         }
1127
1128         /*
1129          * If for any reason at all we couldn't handle the fault,
1130          * make sure we exit gracefully rather than endlessly redo
1131          * the fault:
1132          */
1133         fault = handle_mm_fault(mm, vma, address, flags);
1134
1135         if (unlikely(fault & VM_FAULT_ERROR)) {
1136                 mm_fault_error(regs, error_code, address, fault);
1137                 return;
1138         }
1139
1140         /*
1141          * Major/minor page fault accounting is only done on the
1142          * initial attempt. If we go through a retry, it is extremely
1143          * likely that the page will be found in page cache at that point.
1144          */
1145         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1146                 if (fault & VM_FAULT_MAJOR) {
1147                         tsk->maj_flt++;
1148                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1149                                       regs, address);
1150                 } else {
1151                         tsk->min_flt++;
1152                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1153                                       regs, address);
1154                 }
1155                 if (fault & VM_FAULT_RETRY) {
1156                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1157                          * of starvation. */
1158                         flags &= ~FAULT_FLAG_ALLOW_RETRY;
1159                         goto retry;
1160                 }
1161         }
1162
1163         check_v8086_mode(regs, address, tsk);
1164
1165         up_read(&mm->mmap_sem);
1166 }