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 #include <linux/prefetch.h>             /* prefetchw                    */
  16
  17 #include <asm/traps.h>                  /* dotraplinkage, ...           */
  18 #include <asm/pgalloc.h>                /* pgd_*(), ...                 */
  19 #include <asm/kmemcheck.h>              /* kmemcheck_*(), ...           */
  20
  21 /*
  22  * Page fault error code bits:
  23  *
  24  *   bit 0 ==    0: no page found       1: protection fault
  25  *   bit 1 ==    0: read access         1: write access
  26  *   bit 2 ==    0: kernel-mode access  1: user-mode access
  27  *   bit 3 ==                           1: use of reserved bit detected
  28  *   bit 4 ==                           1: fault was an instruction fetch
  29  */
  30 enum x86_pf_error_code {
  31
  32         PF_PROT         =               1 << 0,
  33         PF_WRITE        =               1 << 1,
  34         PF_USER         =               1 << 2,
  35         PF_RSVD         =               1 << 3,
  36         PF_INSTR        =               1 << 4,
  37 };
  38
  39 /*
  40  * Returns 0 if mmiotrace is disabled, or if the fault is not
  41  * handled by mmiotrace:
  42  */
  43 static inline int __kprobes
  44 kmmio_fault(struct pt_regs *regs, unsigned long addr)
  45 {
  46         if (unlikely(is_kmmio_active()))
  47                 if (kmmio_handler(regs, addr) == 1)
  48                         return -1;
  49         return 0;
  50 }
  51
  52 static inline int __kprobes notify_page_fault(struct pt_regs *regs)
  53 {
  54         int ret = 0;
  55
  56         /* kprobe_running() needs smp_processor_id() */
  57         if (kprobes_built_in() && !user_mode_vm(regs)) {
  58                 preempt_disable();
  59                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  60                         ret = 1;
  61                 preempt_enable();
  62         }
  63
  64         return ret;
  65 }
  66
  67 /*
  68  * Prefetch quirks:
  69  *
  70  * 32-bit mode:
  71  *
  72  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  73  *   Check that here and ignore it.
  74  *
  75  * 64-bit mode:
  76  *
  77  *   Sometimes the CPU reports invalid exceptions on prefetch.
  78  *   Check that here and ignore it.
  79  *
  80  * Opcode checker based on code by Richard Brunner.
  81  */
  82 static inline int
  83 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
  84                       unsigned char opcode, int *prefetch)
  85 {
  86         unsigned char instr_hi = opcode & 0xf0;
  87         unsigned char instr_lo = opcode & 0x0f;
  88
  89         switch (instr_hi) {
  90         case 0x20:
  91         case 0x30:
  92                 /*
  93                  * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
  94                  * In X86_64 long mode, the CPU will signal invalid
  95                  * opcode if some of these prefixes are present so
  96                  * X86_64 will never get here anyway
  97                  */
  98                 return ((instr_lo & 7) == 0x6);
  99 #ifdef CONFIG_X86_64
 100         case 0x40:
 101                 /*
 102                  * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 103                  * Need to figure out under what instruction mode the
 104                  * instruction was issued. Could check the LDT for lm,
 105                  * but for now it's good enough to assume that long
 106                  * mode only uses well known segments or kernel.
 107                  */
 108                 return (!user_mode(regs)) || (regs->cs == __USER_CS);
 109 #endif
 110         case 0x60:
 111                 /* 0x64 thru 0x67 are valid prefixes in all modes. */
 112                 return (instr_lo & 0xC) == 0x4;
 113         case 0xF0:
 114                 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 115                 return !instr_lo || (instr_lo>>1) == 1;
 116         case 0x00:
 117                 /* Prefetch instruction is 0x0F0D or 0x0F18 */
 118                 if (probe_kernel_address(instr, opcode))
 119                         return 0;
 120
 121                 *prefetch = (instr_lo == 0xF) &&
 122                         (opcode == 0x0D || opcode == 0x18);
 123                 return 0;
 124         default:
 125                 return 0;
 126         }
 127 }
 128
 129 static int
 130 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
 131 {
 132         unsigned char *max_instr;
 133         unsigned char *instr;
 134         int prefetch = 0;
 135
 136         /*
 137          * If it was a exec (instruction fetch) fault on NX page, then
 138          * do not ignore the fault:
 139          */
 140         if (error_code & PF_INSTR)
 141                 return 0;
 142
 143         instr = (void *)convert_ip_to_linear(current, regs);
 144         max_instr = instr + 15;
 145
 146         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 147                 return 0;
 148
 149         while (instr < max_instr) {
 150                 unsigned char opcode;
 151
 152                 if (probe_kernel_address(instr, opcode))
 153                         break;
 154
 155                 instr++;
 156
 157                 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
 158                         break;
 159         }
 160         return prefetch;
 161 }
 162
 163 static void
 164 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
 165                      struct task_struct *tsk, int fault)
 166 {
 167         unsigned lsb = 0;
 168         siginfo_t info;
 169
 170         info.si_signo   = si_signo;
 171         info.si_errno   = 0;
 172         info.si_code    = si_code;
 173         info.si_addr    = (void __user *)address;
 174         if (fault & VM_FAULT_HWPOISON_LARGE)
 175                 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
 176         if (fault & VM_FAULT_HWPOISON)
 177                 lsb = PAGE_SHIFT;
 178         info.si_addr_lsb = lsb;
 179
 180         force_sig_info(si_signo, &info, tsk);
 181 }
 182
 183 DEFINE_SPINLOCK(pgd_lock);
 184 LIST_HEAD(pgd_list);
 185
 186 #ifdef CONFIG_X86_32
 187 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 188 {
 189         unsigned index = pgd_index(address);
 190         pgd_t *pgd_k;
 191         pud_t *pud, *pud_k;
 192         pmd_t *pmd, *pmd_k;
 193
 194         pgd += index;
 195         pgd_k = init_mm.pgd + index;
 196
 197         if (!pgd_present(*pgd_k))
 198                 return NULL;
 199
 200         /*
 201          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 202          * and redundant with the set_pmd() on non-PAE. As would
 203          * set_pud.
 204          */
 205         pud = pud_offset(pgd, address);
 206         pud_k = pud_offset(pgd_k, address);
 207         if (!pud_present(*pud_k))
 208                 return NULL;
 209
 210         pmd = pmd_offset(pud, address);
 211         pmd_k = pmd_offset(pud_k, address);
 212         if (!pmd_present(*pmd_k))
 213                 return NULL;
 214
 215         if (!pmd_present(*pmd))
 216                 set_pmd(pmd, *pmd_k);
 217         else
 218                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 219
 220         return pmd_k;
 221 }
 222
 223 void vmalloc_sync_all(void)
 224 {
 225         unsigned long address;
 226
 227         if (SHARED_KERNEL_PMD)
 228                 return;
 229
 230         for (address = VMALLOC_START & PMD_MASK;
 231              address >= TASK_SIZE && address < FIXADDR_TOP;
 232              address += PMD_SIZE) {
 233                 struct page *page;
 234
 235                 spin_lock(&pgd_lock);
 236                 list_for_each_entry(page, &pgd_list, lru) {
 237                         spinlock_t *pgt_lock;
 238                         pmd_t *ret;
 239
 240                         /* the pgt_lock only for Xen */
 241                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 242
 243                         spin_lock(pgt_lock);
 244                         ret = vmalloc_sync_one(page_address(page), address);
 245                         spin_unlock(pgt_lock);
 246
 247                         if (!ret)
 248                                 break;
 249                 }
 250                 spin_unlock(&pgd_lock);
 251         }
 252 }
 253
 254 /*
 255  * 32-bit:
 256  *
 257  *   Handle a fault on the vmalloc or module mapping area
 258  */
 259 static noinline __kprobes int vmalloc_fault(unsigned long address)
 260 {
 261         unsigned long pgd_paddr;
 262         pmd_t *pmd_k;
 263         pte_t *pte_k;
 264
 265         /* Make sure we are in vmalloc area: */
 266         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 267                 return -1;
 268
 269         WARN_ON_ONCE(in_nmi());
 270
 271         /*
 272          * Synchronize this task's top level page-table
 273          * with the 'reference' page table.
 274          *
 275          * Do _not_ use "current" here. We might be inside
 276          * an interrupt in the middle of a task switch..
 277          */
 278         pgd_paddr = read_cr3();
 279         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 280         if (!pmd_k)
 281                 return -1;
 282
 283         pte_k = pte_offset_kernel(pmd_k, address);
 284         if (!pte_present(*pte_k))
 285                 return -1;
 286
 287         return 0;
 288 }
 289
 290 /*
 291  * Did it hit the DOS screen memory VA from vm86 mode?
 292  */
 293 static inline void
 294 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 295                  struct task_struct *tsk)
 296 {
 297         unsigned long bit;
 298
 299         if (!v8086_mode(regs))
 300                 return;
 301
 302         bit = (address - 0xA0000) >> PAGE_SHIFT;
 303         if (bit < 32)
 304                 tsk->thread.screen_bitmap |= 1 << bit;
 305 }
 306
 307 static bool low_pfn(unsigned long pfn)
 308 {
 309         return pfn < max_low_pfn;
 310 }
 311
 312 static void dump_pagetable(unsigned long address)
 313 {
 314         pgd_t *base = __va(read_cr3());
 315         pgd_t *pgd = &base[pgd_index(address)];
 316         pmd_t *pmd;
 317         pte_t *pte;
 318
 319 #ifdef CONFIG_X86_PAE
 320         printk("*pdpt = %016Lx ", pgd_val(*pgd));
 321         if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
 322                 goto out;
 323 #endif
 324         pmd = pmd_offset(pud_offset(pgd, address), address);
 325         printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
 326
 327         /*
 328          * We must not directly access the pte in the highpte
 329          * case if the page table is located in highmem.
 330          * And let's rather not kmap-atomic the pte, just in case
 331          * it's allocated already:
 332          */
 333         if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
 334                 goto out;
 335
 336         pte = pte_offset_kernel(pmd, address);
 337         printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
 338 out:
 339         printk("\n");
 340 }
 341
 342 #else /* CONFIG_X86_64: */
 343
 344 void vmalloc_sync_all(void)
 345 {
 346         sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
 347 }
 348
 349 /*
 350  * 64-bit:
 351  *
 352  *   Handle a fault on the vmalloc area
 353  *
 354  * This assumes no large pages in there.
 355  */
 356 static noinline __kprobes int vmalloc_fault(unsigned long address)
 357 {
 358         pgd_t *pgd, *pgd_ref;
 359         pud_t *pud, *pud_ref;
 360         pmd_t *pmd, *pmd_ref;
 361         pte_t *pte, *pte_ref;
 362
 363         /* Make sure we are in vmalloc area: */
 364         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 365                 return -1;
 366
 367         WARN_ON_ONCE(in_nmi());
 368
 369         /*
 370          * Copy kernel mappings over when needed. This can also
 371          * happen within a race in page table update. In the later
 372          * case just flush:
 373          */
 374         pgd = pgd_offset(current->active_mm, address);
 375         pgd_ref = pgd_offset_k(address);
 376         if (pgd_none(*pgd_ref))
 377                 return -1;
 378
 379         if (pgd_none(*pgd))
 380                 set_pgd(pgd, *pgd_ref);
 381         else
 382                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 383
 384         /*
 385          * Below here mismatches are bugs because these lower tables
 386          * are shared:
 387          */
 388
 389         pud = pud_offset(pgd, address);
 390         pud_ref = pud_offset(pgd_ref, address);
 391         if (pud_none(*pud_ref))
 392                 return -1;
 393
 394         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 395                 BUG();
 396
 397         pmd = pmd_offset(pud, address);
 398         pmd_ref = pmd_offset(pud_ref, address);
 399         if (pmd_none(*pmd_ref))
 400                 return -1;
 401
 402         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 403                 BUG();
 404
 405         pte_ref = pte_offset_kernel(pmd_ref, address);
 406         if (!pte_present(*pte_ref))
 407                 return -1;
 408
 409         pte = pte_offset_kernel(pmd, address);
 410
 411         /*
 412          * Don't use pte_page here, because the mappings can point
 413          * outside mem_map, and the NUMA hash lookup cannot handle
 414          * that:
 415          */
 416         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 417                 BUG();
 418
 419         return 0;
 420 }
 421
 422 static const char errata93_warning[] =
 423 KERN_ERR
 424 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 425 "******* Working around it, but it may cause SEGVs or burn power.\n"
 426 "******* Please consider a BIOS update.\n"
 427 "******* Disabling USB legacy in the BIOS may also help.\n";
 428
 429 /*
 430  * No vm86 mode in 64-bit mode:
 431  */
 432 static inline void
 433 check_v8086_mode(struct pt_regs *regs, unsigned long address,
 434                  struct task_struct *tsk)
 435 {
 436 }
 437
 438 static int bad_address(void *p)
 439 {
 440         unsigned long dummy;
 441
 442         return probe_kernel_address((unsigned long *)p, dummy);
 443 }
 444
 445 static void dump_pagetable(unsigned long address)
 446 {
 447         pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
 448         pgd_t *pgd = base + pgd_index(address);
 449         pud_t *pud;
 450         pmd_t *pmd;
 451         pte_t *pte;
 452
 453         if (bad_address(pgd))
 454                 goto bad;
 455
 456         printk("PGD %lx ", pgd_val(*pgd));
 457
 458         if (!pgd_present(*pgd))
 459                 goto out;
 460
 461         pud = pud_offset(pgd, address);
 462         if (bad_address(pud))
 463                 goto bad;
 464
 465         printk("PUD %lx ", pud_val(*pud));
 466         if (!pud_present(*pud) || pud_large(*pud))
 467                 goto out;
 468
 469         pmd = pmd_offset(pud, address);
 470         if (bad_address(pmd))
 471                 goto bad;
 472
 473         printk("PMD %lx ", pmd_val(*pmd));
 474         if (!pmd_present(*pmd) || pmd_large(*pmd))
 475                 goto out;
 476
 477         pte = pte_offset_kernel(pmd, address);
 478         if (bad_address(pte))
 479                 goto bad;
 480
 481         printk("PTE %lx", pte_val(*pte));
 482 out:
 483         printk("\n");
 484         return;
 485 bad:
 486         printk("BAD\n");
 487 }
 488
 489 #endif /* CONFIG_X86_64 */
 490
 491 /*
 492  * Workaround for K8 erratum #93 & buggy BIOS.
 493  *
 494  * BIOS SMM functions are required to use a specific workaround
 495  * to avoid corruption of the 64bit RIP register on C stepping K8.
 496  *
 497  * A lot of BIOS that didn't get tested properly miss this.
 498  *
 499  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
 500  * Try to work around it here.
 501  *
 502  * Note we only handle faults in kernel here.
 503  * Does nothing on 32-bit.
 504  */
 505 static int is_errata93(struct pt_regs *regs, unsigned long address)
 506 {
 507 #ifdef CONFIG_X86_64
 508         if (address != regs->ip)
 509                 return 0;
 510
 511         if ((address >> 32) != 0)
 512                 return 0;
 513
 514         address |= 0xffffffffUL << 32;
 515         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 516             (address >= MODULES_VADDR && address <= MODULES_END)) {
 517                 printk_once(errata93_warning);
 518                 regs->ip = address;
 519                 return 1;
 520         }
 521 #endif
 522         return 0;
 523 }
 524
 525 /*
 526  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
 527  * to illegal addresses >4GB.
 528  *
 529  * We catch this in the page fault handler because these addresses
 530  * are not reachable. Just detect this case and return.  Any code
 531  * segment in LDT is compatibility mode.
 532  */
 533 static int is_errata100(struct pt_regs *regs, unsigned long address)
 534 {
 535 #ifdef CONFIG_X86_64
 536         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
 537                 return 1;
 538 #endif
 539         return 0;
 540 }
 541
 542 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 543 {
 544 #ifdef CONFIG_X86_F00F_BUG
 545         unsigned long nr;
 546
 547         /*
 548          * Pentium F0 0F C7 C8 bug workaround:
 549          */
 550         if (boot_cpu_data.f00f_bug) {
 551                 nr = (address - idt_descr.address) >> 3;
 552
 553                 if (nr == 6) {
 554                         do_invalid_op(regs, 0);
 555                         return 1;
 556                 }
 557         }
 558 #endif
 559         return 0;
 560 }
 561
 562 static const char nx_warning[] = KERN_CRIT
 563 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
 564
 565 static void
 566 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 567                 unsigned long address)
 568 {
 569         if (!oops_may_print())
 570                 return;
 571
 572         if (error_code & PF_INSTR) {
 573                 unsigned int level;
 574
 575                 pte_t *pte = lookup_address(address, &level);
 576
 577                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 578                         printk(nx_warning, current_uid());
 579         }
 580
 581         printk(KERN_ALERT "BUG: unable to handle kernel ");
 582         if (address < PAGE_SIZE)
 583                 printk(KERN_CONT "NULL pointer dereference");
 584         else
 585                 printk(KERN_CONT "paging request");
 586
 587         printk(KERN_CONT " at %p\n", (void *) address);
 588         printk(KERN_ALERT "IP:");
 589         printk_address(regs->ip, 1);
 590
 591         dump_pagetable(address);
 592 }
 593
 594 static noinline void
 595 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
 596             unsigned long address)
 597 {
 598         struct task_struct *tsk;
 599         unsigned long flags;
 600         int sig;
 601
 602         flags = oops_begin();
 603         tsk = current;
 604         sig = SIGKILL;
 605
 606         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 607                tsk->comm, address);
 608         dump_pagetable(address);
 609
 610         tsk->thread.cr2         = address;
 611         tsk->thread.trap_no     = 14;
 612         tsk->thread.error_code  = error_code;
 613
 614         if (__die("Bad pagetable", regs, error_code))
 615                 sig = 0;
 616
 617         oops_end(flags, regs, sig);
 618 }
 619
 620 static noinline void
 621 no_context(struct pt_regs *regs, unsigned long error_code,
 622            unsigned long address)
 623 {
 624         struct task_struct *tsk = current;
 625         unsigned long *stackend;
 626         unsigned long flags;
 627         int sig;
 628
 629         /* Are we prepared to handle this kernel fault? */
 630         if (fixup_exception(regs))
 631                 return;
 632
 633         /*
 634          * 32-bit:
 635          *
 636          *   Valid to do another page fault here, because if this fault
 637          *   had been triggered by is_prefetch fixup_exception would have
 638          *   handled it.
 639          *
 640          * 64-bit:
 641          *
 642          *   Hall of shame of CPU/BIOS bugs.
 643          */
 644         if (is_prefetch(regs, error_code, address))
 645                 return;
 646
 647         if (is_errata93(regs, address))
 648                 return;
 649
 650         /*
 651          * Oops. The kernel tried to access some bad page. We'll have to
 652          * terminate things with extreme prejudice:
 653          */
 654         flags = oops_begin();
 655
 656         show_fault_oops(regs, error_code, address);
 657
 658         stackend = end_of_stack(tsk);
 659         if (tsk != &init_task && *stackend != STACK_END_MAGIC)
 660                 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
 661
 662         tsk->thread.cr2         = address;
 663         tsk->thread.trap_no     = 14;
 664         tsk->thread.error_code  = error_code;
 665
 666         sig = SIGKILL;
 667         if (__die("Oops", regs, error_code))
 668                 sig = 0;
 669
 670         /* Executive summary in case the body of the oops scrolled away */
 671         printk(KERN_EMERG "CR2: %016lx\n", address);
 672
 673         oops_end(flags, regs, sig);
 674 }
 675
 676 /*
 677  * Print out info about fatal segfaults, if the show_unhandled_signals
 678  * sysctl is set:
 679  */
 680 static inline void
 681 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
 682                 unsigned long address, struct task_struct *tsk)
 683 {
 684         if (!unhandled_signal(tsk, SIGSEGV))
 685                 return;
 686
 687         if (!printk_ratelimit())
 688                 return;
 689
 690         printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 691                 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 692                 tsk->comm, task_pid_nr(tsk), address,
 693                 (void *)regs->ip, (void *)regs->sp, error_code);
 694
 695         print_vma_addr(KERN_CONT " in ", regs->ip);
 696
 697         printk(KERN_CONT "\n");
 698 }
 699
 700 static void
 701 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 702                        unsigned long address, int si_code)
 703 {
 704         struct task_struct *tsk = current;
 705
 706         /* User mode accesses just cause a SIGSEGV */
 707         if (error_code & PF_USER) {
 708                 /*
 709                  * It's possible to have interrupts off here:
 710                  */
 711                 local_irq_enable();
 712
 713                 /*
 714                  * Valid to do another page fault here because this one came
 715                  * from user space:
 716                  */
 717                 if (is_prefetch(regs, error_code, address))
 718                         return;
 719
 720                 if (is_errata100(regs, address))
 721                         return;
 722
 723                 if (unlikely(show_unhandled_signals))
 724                         show_signal_msg(regs, error_code, address, tsk);
 725
 726                 /* Kernel addresses are always protection faults: */
 727                 tsk->thread.cr2         = address;
 728                 tsk->thread.error_code  = error_code | (address >= TASK_SIZE);
 729                 tsk->thread.trap_no     = 14;
 730
 731                 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
 732
 733                 return;
 734         }
 735
 736         if (is_f00f_bug(regs, address))
 737                 return;
 738
 739         no_context(regs, error_code, address);
 740 }
 741
 742 static noinline void
 743 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 744                      unsigned long address)
 745 {
 746         __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 747 }
 748
 749 static void
 750 __bad_area(struct pt_regs *regs, unsigned long error_code,
 751            unsigned long address, int si_code)
 752 {
 753         struct mm_struct *mm = current->mm;
 754
 755         /*
 756          * Something tried to access memory that isn't in our memory map..
 757          * Fix it, but check if it's kernel or user first..
 758          */
 759         up_read(&mm->mmap_sem);
 760
 761         __bad_area_nosemaphore(regs, error_code, address, si_code);
 762 }
 763
 764 static noinline void
 765 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 766 {
 767         __bad_area(regs, error_code, address, SEGV_MAPERR);
 768 }
 769
 770 static noinline void
 771 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 772                       unsigned long address)
 773 {
 774         __bad_area(regs, error_code, address, SEGV_ACCERR);
 775 }
 776
 777 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
 778 static void
 779 out_of_memory(struct pt_regs *regs, unsigned long error_code,
 780               unsigned long address)
 781 {
 782         /*
 783          * We ran out of memory, call the OOM killer, and return the userspace
 784          * (which will retry the fault, or kill us if we got oom-killed):
 785          */
 786         up_read(&current->mm->mmap_sem);
 787
 788         pagefault_out_of_memory();
 789 }
 790
 791 static void
 792 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
 793           unsigned int fault)
 794 {
 795         struct task_struct *tsk = current;
 796         struct mm_struct *mm = tsk->mm;
 797         int code = BUS_ADRERR;
 798
 799         up_read(&mm->mmap_sem);
 800
 801         /* Kernel mode? Handle exceptions or die: */
 802         if (!(error_code & PF_USER)) {
 803                 no_context(regs, error_code, address);
 804                 return;
 805         }
 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|VM_FAULT_HWPOISON_LARGE)) {
 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, fault);
 824 }
 825
 826 static noinline int
 827 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 828                unsigned long address, unsigned int fault)
 829 {
 830         /*
 831          * Pagefault was interrupted by SIGKILL. We have no reason to
 832          * continue pagefault.
 833          */
 834         if (fatal_signal_pending(current)) {
 835                 if (!(fault & VM_FAULT_RETRY))
 836                         up_read(&current->mm->mmap_sem);
 837                 if (!(error_code & PF_USER))
 838                         no_context(regs, error_code, address);
 839                 return 1;
 840         }
 841         if (!(fault & VM_FAULT_ERROR))
 842                 return 0;
 843
 844         if (fault & VM_FAULT_OOM) {
 845                 /* Kernel mode? Handle exceptions or die: */
 846                 if (!(error_code & PF_USER)) {
 847                         up_read(&current->mm->mmap_sem);
 848                         no_context(regs, error_code, address);
 849                         return 1;
 850                 }
 851
 852                 out_of_memory(regs, error_code, address);
 853         } else {
 854                 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
 855                              VM_FAULT_HWPOISON_LARGE))
 856                         do_sigbus(regs, error_code, address, fault);
 857                 else
 858                         BUG();
 859         }
 860         return 1;
 861 }
 862
 863 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 864 {
 865         if ((error_code & PF_WRITE) && !pte_write(*pte))
 866                 return 0;
 867
 868         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 869                 return 0;
 870
 871         return 1;
 872 }
 873
 874 /*
 875  * Handle a spurious fault caused by a stale TLB entry.
 876  *
 877  * This allows us to lazily refresh the TLB when increasing the
 878  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
 879  * eagerly is very expensive since that implies doing a full
 880  * cross-processor TLB flush, even if no stale TLB entries exist
 881  * on other processors.
 882  *
 883  * There are no security implications to leaving a stale TLB when
 884  * increasing the permissions on a page.
 885  */
 886 static noinline __kprobes int
 887 spurious_fault(unsigned long error_code, unsigned long address)
 888 {
 889         pgd_t *pgd;
 890         pud_t *pud;
 891         pmd_t *pmd;
 892         pte_t *pte;
 893         int ret;
 894
 895         /* Reserved-bit violation or user access to kernel space? */
 896         if (error_code & (PF_USER | PF_RSVD))
 897                 return 0;
 898
 899         pgd = init_mm.pgd + pgd_index(address);
 900         if (!pgd_present(*pgd))
 901                 return 0;
 902
 903         pud = pud_offset(pgd, address);
 904         if (!pud_present(*pud))
 905                 return 0;
 906
 907         if (pud_large(*pud))
 908                 return spurious_fault_check(error_code, (pte_t *) pud);
 909
 910         pmd = pmd_offset(pud, address);
 911         if (!pmd_present(*pmd))
 912                 return 0;
 913
 914         if (pmd_large(*pmd))
 915                 return spurious_fault_check(error_code, (pte_t *) pmd);
 916
 917         /*
 918          * Note: don't use pte_present() here, since it returns true
 919          * if the _PAGE_PROTNONE bit is set.  However, this aliases the
 920          * _PAGE_GLOBAL bit, which for kernel pages give false positives
 921          * when CONFIG_DEBUG_PAGEALLOC is used.
 922          */
 923         pte = pte_offset_kernel(pmd, address);
 924         if (!(pte_flags(*pte) & _PAGE_PRESENT))
 925                 return 0;
 926
 927         ret = spurious_fault_check(error_code, pte);
 928         if (!ret)
 929                 return 0;
 930
 931         /*
 932          * Make sure we have permissions in PMD.
 933          * If not, then there's a bug in the page tables:
 934          */
 935         ret = spurious_fault_check(error_code, (pte_t *) pmd);
 936         WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
 937
 938         return ret;
 939 }
 940
 941 int show_unhandled_signals = 1;
 942
 943 static inline int
 944 access_error(unsigned long error_code, struct vm_area_struct *vma)
 945 {
 946         if (error_code & PF_WRITE) {
 947                 /* write, present and write, not present: */
 948                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 949                         return 1;
 950                 return 0;
 951         }
 952
 953         /* read, present: */
 954         if (unlikely(error_code & PF_PROT))
 955                 return 1;
 956
 957         /* read, not present: */
 958         if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 959                 return 1;
 960
 961         return 0;
 962 }
 963
 964 static int fault_in_kernel_space(unsigned long address)
 965 {
 966         return address >= TASK_SIZE_MAX;
 967 }
 968
 969 /*
 970  * This routine handles page faults.  It determines the address,
 971  * and the problem, and then passes it off to one of the appropriate
 972  * routines.
 973  */
 974 dotraplinkage void __kprobes
 975 do_page_fault(struct pt_regs *regs, unsigned long error_code)
 976 {
 977         struct vm_area_struct *vma;
 978         struct task_struct *tsk;
 979         unsigned long address;
 980         struct mm_struct *mm;
 981         int fault;
 982         int write = error_code & PF_WRITE;
 983         unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE |
 984                                         (write ? FAULT_FLAG_WRITE : 0);
 985
 986         tsk = current;
 987         mm = tsk->mm;
 988
 989         /* Get the faulting address: */
 990         address = read_cr2();
 991
 992         /*
 993          * Detect and handle instructions that would cause a page fault for
 994          * both a tracked kernel page and a userspace page.
 995          */
 996         if (kmemcheck_active(regs))
 997                 kmemcheck_hide(regs);
 998         prefetchw(&mm->mmap_sem);
 999
1000         if (unlikely(kmmio_fault(regs, address)))
1001                 return;
1002
1003         /*
1004          * We fault-in kernel-space virtual memory on-demand. The
1005          * 'reference' page table is init_mm.pgd.
1006          *
1007          * NOTE! We MUST NOT take any locks for this case. We may
1008          * be in an interrupt or a critical region, and should
1009          * only copy the information from the master page table,
1010          * nothing more.
1011          *
1012          * This verifies that the fault happens in kernel space
1013          * (error_code & 4) == 0, and that the fault was not a
1014          * protection error (error_code & 9) == 0.
1015          */
1016         if (unlikely(fault_in_kernel_space(address))) {
1017                 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1018                         if (vmalloc_fault(address) >= 0)
1019                                 return;
1020
1021                         if (kmemcheck_fault(regs, address, error_code))
1022                                 return;
1023                 }
1024
1025                 /* Can handle a stale RO->RW TLB: */
1026                 if (spurious_fault(error_code, address))
1027                         return;
1028
1029                 /* kprobes don't want to hook the spurious faults: */
1030                 if (notify_page_fault(regs))
1031                         return;
1032                 /*
1033                  * Don't take the mm semaphore here. If we fixup a prefetch
1034                  * fault we could otherwise deadlock:
1035                  */
1036                 bad_area_nosemaphore(regs, error_code, address);
1037
1038                 return;
1039         }
1040
1041         /* kprobes don't want to hook the spurious faults: */
1042         if (unlikely(notify_page_fault(regs)))
1043                 return;
1044         /*
1045          * It's safe to allow irq's after cr2 has been saved and the
1046          * vmalloc fault has been handled.
1047          *
1048          * User-mode registers count as a user access even for any
1049          * potential system fault or CPU buglet:
1050          */
1051         if (user_mode_vm(regs)) {
1052                 local_irq_enable();
1053                 error_code |= PF_USER;
1054         } else {
1055                 if (regs->flags & X86_EFLAGS_IF)
1056                         local_irq_enable();
1057         }
1058
1059         if (unlikely(error_code & PF_RSVD))
1060                 pgtable_bad(regs, error_code, address);
1061
1062         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1063
1064         /*
1065          * If we're in an interrupt, have no user context or are running
1066          * in an atomic region then we must not take the fault:
1067          */
1068         if (unlikely(in_atomic() || !mm)) {
1069                 bad_area_nosemaphore(regs, error_code, address);
1070                 return;
1071         }
1072
1073         /*
1074          * When running in the kernel we expect faults to occur only to
1075          * addresses in user space.  All other faults represent errors in
1076          * the kernel and should generate an OOPS.  Unfortunately, in the
1077          * case of an erroneous fault occurring in a code path which already
1078          * holds mmap_sem we will deadlock attempting to validate the fault
1079          * against the address space.  Luckily the kernel only validly
1080          * references user space from well defined areas of code, which are
1081          * listed in the exceptions table.
1082          *
1083          * As the vast majority of faults will be valid we will only perform
1084          * the source reference check when there is a possibility of a
1085          * deadlock. Attempt to lock the address space, if we cannot we then
1086          * validate the source. If this is invalid we can skip the address
1087          * space check, thus avoiding the deadlock:
1088          */
1089         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1090                 if ((error_code & PF_USER) == 0 &&
1091                     !search_exception_tables(regs->ip)) {
1092                         bad_area_nosemaphore(regs, error_code, address);
1093                         return;
1094                 }
1095 retry:
1096                 down_read(&mm->mmap_sem);
1097         } else {
1098                 /*
1099                  * The above down_read_trylock() might have succeeded in
1100                  * which case we'll have missed the might_sleep() from
1101                  * down_read():
1102                  */
1103                 might_sleep();
1104         }
1105
1106         vma = find_vma(mm, address);
1107         if (unlikely(!vma)) {
1108                 bad_area(regs, error_code, address);
1109                 return;
1110         }
1111         if (likely(vma->vm_start <= address))
1112                 goto good_area;
1113         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1114                 bad_area(regs, error_code, address);
1115                 return;
1116         }
1117         if (error_code & PF_USER) {
1118                 /*
1119                  * Accessing the stack below %sp is always a bug.
1120                  * The large cushion allows instructions like enter
1121                  * and pusha to work. ("enter $65535, $31" pushes
1122                  * 32 pointers and then decrements %sp by 65535.)
1123                  */
1124                 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1125                         bad_area(regs, error_code, address);
1126                         return;
1127                 }
1128         }
1129         if (unlikely(expand_stack(vma, address))) {
1130                 bad_area(regs, error_code, address);
1131                 return;
1132         }
1133
1134         /*
1135          * Ok, we have a good vm_area for this memory access, so
1136          * we can handle it..
1137          */
1138 good_area:
1139         if (unlikely(access_error(error_code, vma))) {
1140                 bad_area_access_error(regs, error_code, address);
1141                 return;
1142         }
1143
1144         /*
1145          * If for any reason at all we couldn't handle the fault,
1146          * make sure we exit gracefully rather than endlessly redo
1147          * the fault:
1148          */
1149         fault = handle_mm_fault(mm, vma, address, flags);
1150
1151         if (unlikely(fault & (VM_FAULT_RETRY|VM_FAULT_ERROR))) {
1152                 if (mm_fault_error(regs, error_code, address, fault))
1153                         return;
1154         }
1155
1156         /*
1157          * Major/minor page fault accounting is only done on the
1158          * initial attempt. If we go through a retry, it is extremely
1159          * likely that the page will be found in page cache at that point.
1160          */
1161         if (flags & FAULT_FLAG_ALLOW_RETRY) {
1162                 if (fault & VM_FAULT_MAJOR) {
1163                         tsk->maj_flt++;
1164                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1165                                       regs, address);
1166                 } else {
1167                         tsk->min_flt++;
1168                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1169                                       regs, address);
1170                 }
1171                 if (fault & VM_FAULT_RETRY) {
1172                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1173                          * of starvation. */
1174                         flags &= ~FAULT_FLAG_ALLOW_RETRY;
1175                         goto retry;
1176                 }
1177         }
1178
1179         check_v8086_mode(regs, address, tsk);
1180
1181         up_read(&mm->mmap_sem);
1182 }