arch/x86_64/mm/fault.c

   1 /*
   2  *  linux/arch/x86-64/mm/fault.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright (C) 2001,2002 Andi Kleen, SuSE Labs.
   6  */
   7
   8 #include <linux/signal.h>
   9 #include <linux/sched.h>
  10 #include <linux/kernel.h>
  11 #include <linux/errno.h>
  12 #include <linux/string.h>
  13 #include <linux/types.h>
  14 #include <linux/ptrace.h>
  15 #include <linux/mman.h>
  16 #include <linux/mm.h>
  17 #include <linux/smp.h>
  18 #include <linux/smp_lock.h>
  19 #include <linux/interrupt.h>
  20 #include <linux/init.h>
  21 #include <linux/tty.h>
  22 #include <linux/vt_kern.h>              /* For unblank_screen() */
  23 #include <linux/compiler.h>
  24 #include <linux/module.h>
  25 #include <linux/kprobes.h>
  26 #include <linux/uaccess.h>
  27
  28 #include <asm/system.h>
  29 #include <asm/pgalloc.h>
  30 #include <asm/smp.h>
  31 #include <asm/tlbflush.h>
  32 #include <asm/proto.h>
  33 #include <asm/kdebug.h>
  34 #include <asm-generic/sections.h>
  35
  36 /* Page fault error code bits */
  37 #define PF_PROT (1<<0)          /* or no page found */
  38 #define PF_WRITE        (1<<1)
  39 #define PF_USER (1<<2)
  40 #define PF_RSVD (1<<3)
  41 #define PF_INSTR        (1<<4)
  42
  43 static ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
  44
  45 /* Hook to register for page fault notifications */
  46 int register_page_fault_notifier(struct notifier_block *nb)
  47 {
  48         vmalloc_sync_all();
  49         return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
  50 }
  51 EXPORT_SYMBOL_GPL(register_page_fault_notifier);
  52
  53 int unregister_page_fault_notifier(struct notifier_block *nb)
  54 {
  55         return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
  56 }
  57 EXPORT_SYMBOL_GPL(unregister_page_fault_notifier);
  58
  59 static inline int notify_page_fault(enum die_val val, const char *str,
  60                         struct pt_regs *regs, long err, int trap, int sig)
  61 {
  62         struct die_args args = {
  63                 .regs = regs,
  64                 .str = str,
  65                 .err = err,
  66                 .trapnr = trap,
  67                 .signr = sig
  68         };
  69         return atomic_notifier_call_chain(&notify_page_fault_chain, val, &args);
  70 }
  71
  72 /* Sometimes the CPU reports invalid exceptions on prefetch.
  73    Check that here and ignore.
  74    Opcode checker based on code by Richard Brunner */
  75 static noinline int is_prefetch(struct pt_regs *regs, unsigned long addr,
  76                                 unsigned long error_code)
  77 {
  78         unsigned char *instr;
  79         int scan_more = 1;
  80         int prefetch = 0;
  81         unsigned char *max_instr;
  82
  83         /* If it was a exec fault ignore */
  84         if (error_code & PF_INSTR)
  85                 return 0;
  86
  87         instr = (unsigned char __user *)convert_rip_to_linear(current, regs);
  88         max_instr = instr + 15;
  89
  90         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
  91                 return 0;
  92
  93         while (scan_more && instr < max_instr) {
  94                 unsigned char opcode;
  95                 unsigned char instr_hi;
  96                 unsigned char instr_lo;
  97
  98                 if (probe_kernel_address(instr, opcode))
  99                         break;
 100
 101                 instr_hi = opcode & 0xf0;
 102                 instr_lo = opcode & 0x0f;
 103                 instr++;
 104
 105                 switch (instr_hi) {
 106                 case 0x20:
 107                 case 0x30:
 108                         /* Values 0x26,0x2E,0x36,0x3E are valid x86
 109                            prefixes.  In long mode, the CPU will signal
 110                            invalid opcode if some of these prefixes are
 111                            present so we will never get here anyway */
 112                         scan_more = ((instr_lo & 7) == 0x6);
 113                         break;
 114
 115                 case 0x40:
 116                         /* In AMD64 long mode, 0x40 to 0x4F are valid REX prefixes
 117                            Need to figure out under what instruction mode the
 118                            instruction was issued ... */
 119                         /* Could check the LDT for lm, but for now it's good
 120                            enough to assume that long mode only uses well known
 121                            segments or kernel. */
 122                         scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
 123                         break;
 124
 125                 case 0x60:
 126                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 127                         scan_more = (instr_lo & 0xC) == 0x4;
 128                         break;
 129                 case 0xF0:
 130                         /* 0xF0, 0xF2, and 0xF3 are valid prefixes in all modes. */
 131                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 132                         break;
 133                 case 0x00:
 134                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 135                         scan_more = 0;
 136                         if (probe_kernel_address(instr, opcode))
 137                                 break;
 138                         prefetch = (instr_lo == 0xF) &&
 139                                 (opcode == 0x0D || opcode == 0x18);
 140                         break;
 141                 default:
 142                         scan_more = 0;
 143                         break;
 144                 }
 145         }
 146         return prefetch;
 147 }
 148
 149 static int bad_address(void *p)
 150 {
 151         unsigned long dummy;
 152         return probe_kernel_address((unsigned long *)p, dummy);
 153 }
 154
 155 void dump_pagetable(unsigned long address)
 156 {
 157         pgd_t *pgd;
 158         pud_t *pud;
 159         pmd_t *pmd;
 160         pte_t *pte;
 161
 162         asm("movq %%cr3,%0" : "=r" (pgd));
 163
 164         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 165         pgd += pgd_index(address);
 166         if (bad_address(pgd)) goto bad;
 167         printk("PGD %lx ", pgd_val(*pgd));
 168         if (!pgd_present(*pgd)) goto ret;
 169
 170         pud = pud_offset(pgd, address);
 171         if (bad_address(pud)) goto bad;
 172         printk("PUD %lx ", pud_val(*pud));
 173         if (!pud_present(*pud)) goto ret;
 174
 175         pmd = pmd_offset(pud, address);
 176         if (bad_address(pmd)) goto bad;
 177         printk("PMD %lx ", pmd_val(*pmd));
 178         if (!pmd_present(*pmd)) goto ret;
 179
 180         pte = pte_offset_kernel(pmd, address);
 181         if (bad_address(pte)) goto bad;
 182         printk("PTE %lx", pte_val(*pte));
 183 ret:
 184         printk("\n");
 185         return;
 186 bad:
 187         printk("BAD\n");
 188 }
 189
 190 static const char errata93_warning[] =
 191 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 192 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 193 KERN_ERR "******* Please consider a BIOS update.\n"
 194 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 195
 196 /* Workaround for K8 erratum #93 & buggy BIOS.
 197    BIOS SMM functions are required to use a specific workaround
 198    to avoid corruption of the 64bit RIP register on C stepping K8.
 199    A lot of BIOS that didn't get tested properly miss this.
 200    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 201    Try to work around it here.
 202    Note we only handle faults in kernel here. */
 203
 204 static int is_errata93(struct pt_regs *regs, unsigned long address)
 205 {
 206         static int warned;
 207         if (address != regs->rip)
 208                 return 0;
 209         if ((address >> 32) != 0)
 210                 return 0;
 211         address |= 0xffffffffUL << 32;
 212         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 213             (address >= MODULES_VADDR && address <= MODULES_END)) {
 214                 if (!warned) {
 215                         printk(errata93_warning);
 216                         warned = 1;
 217                 }
 218                 regs->rip = address;
 219                 return 1;
 220         }
 221         return 0;
 222 }
 223
 224 int unhandled_signal(struct task_struct *tsk, int sig)
 225 {
 226         if (is_init(tsk))
 227                 return 1;
 228         if (tsk->ptrace & PT_PTRACED)
 229                 return 0;
 230         return (tsk->sighand->action[sig-1].sa.sa_handler == SIG_IGN) ||
 231                 (tsk->sighand->action[sig-1].sa.sa_handler == SIG_DFL);
 232 }
 233
 234 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
 235                                  unsigned long error_code)
 236 {
 237         unsigned long flags = oops_begin();
 238         struct task_struct *tsk;
 239
 240         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 241                current->comm, address);
 242         dump_pagetable(address);
 243         tsk = current;
 244         tsk->thread.cr2 = address;
 245         tsk->thread.trap_no = 14;
 246         tsk->thread.error_code = error_code;
 247         __die("Bad pagetable", regs, error_code);
 248         oops_end(flags);
 249         do_exit(SIGKILL);
 250 }
 251
 252 /*
 253  * Handle a fault on the vmalloc area
 254  *
 255  * This assumes no large pages in there.
 256  */
 257 static int vmalloc_fault(unsigned long address)
 258 {
 259         pgd_t *pgd, *pgd_ref;
 260         pud_t *pud, *pud_ref;
 261         pmd_t *pmd, *pmd_ref;
 262         pte_t *pte, *pte_ref;
 263
 264         /* Copy kernel mappings over when needed. This can also
 265            happen within a race in page table update. In the later
 266            case just flush. */
 267
 268         pgd = pgd_offset(current->mm ?: &init_mm, address);
 269         pgd_ref = pgd_offset_k(address);
 270         if (pgd_none(*pgd_ref))
 271                 return -1;
 272         if (pgd_none(*pgd))
 273                 set_pgd(pgd, *pgd_ref);
 274         else
 275                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 276
 277         /* Below here mismatches are bugs because these lower tables
 278            are shared */
 279
 280         pud = pud_offset(pgd, address);
 281         pud_ref = pud_offset(pgd_ref, address);
 282         if (pud_none(*pud_ref))
 283                 return -1;
 284         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 285                 BUG();
 286         pmd = pmd_offset(pud, address);
 287         pmd_ref = pmd_offset(pud_ref, address);
 288         if (pmd_none(*pmd_ref))
 289                 return -1;
 290         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 291                 BUG();
 292         pte_ref = pte_offset_kernel(pmd_ref, address);
 293         if (!pte_present(*pte_ref))
 294                 return -1;
 295         pte = pte_offset_kernel(pmd, address);
 296         /* Don't use pte_page here, because the mappings can point
 297            outside mem_map, and the NUMA hash lookup cannot handle
 298            that. */
 299         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 300                 BUG();
 301         return 0;
 302 }
 303
 304 int page_fault_trace = 0;
 305 int exception_trace = 1;
 306
 307 /*
 308  * This routine handles page faults.  It determines the address,
 309  * and the problem, and then passes it off to one of the appropriate
 310  * routines.
 311  */
 312 asmlinkage void __kprobes do_page_fault(struct pt_regs *regs,
 313                                         unsigned long error_code)
 314 {
 315         struct task_struct *tsk;
 316         struct mm_struct *mm;
 317         struct vm_area_struct * vma;
 318         unsigned long address;
 319         const struct exception_table_entry *fixup;
 320         int write;
 321         unsigned long flags;
 322         siginfo_t info;
 323
 324         tsk = current;
 325         mm = tsk->mm;
 326         prefetchw(&mm->mmap_sem);
 327
 328         /* get the address */
 329         __asm__("movq %%cr2,%0":"=r" (address));
 330
 331         info.si_code = SEGV_MAPERR;
 332
 333
 334         /*
 335          * We fault-in kernel-space virtual memory on-demand. The
 336          * 'reference' page table is init_mm.pgd.
 337          *
 338          * NOTE! We MUST NOT take any locks for this case. We may
 339          * be in an interrupt or a critical region, and should
 340          * only copy the information from the master page table,
 341          * nothing more.
 342          *
 343          * This verifies that the fault happens in kernel space
 344          * (error_code & 4) == 0, and that the fault was not a
 345          * protection error (error_code & 9) == 0.
 346          */
 347         if (unlikely(address >= TASK_SIZE64)) {
 348                 /*
 349                  * Don't check for the module range here: its PML4
 350                  * is always initialized because it's shared with the main
 351                  * kernel text. Only vmalloc may need PML4 syncups.
 352                  */
 353                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 354                       ((address >= VMALLOC_START && address < VMALLOC_END))) {
 355                         if (vmalloc_fault(address) >= 0)
 356                                 return;
 357                 }
 358                 if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 359                                                 SIGSEGV) == NOTIFY_STOP)
 360                         return;
 361                 /*
 362                  * Don't take the mm semaphore here. If we fixup a prefetch
 363                  * fault we could otherwise deadlock.
 364                  */
 365                 goto bad_area_nosemaphore;
 366         }
 367
 368         if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 369                                         SIGSEGV) == NOTIFY_STOP)
 370                 return;
 371
 372         if (likely(regs->eflags & X86_EFLAGS_IF))
 373                 local_irq_enable();
 374
 375         if (unlikely(page_fault_trace))
 376                 printk("pagefault rip:%lx rsp:%lx cs:%lu ss:%lu address %lx error %lx\n",
 377                        regs->rip,regs->rsp,regs->cs,regs->ss,address,error_code);
 378
 379         if (unlikely(error_code & PF_RSVD))
 380                 pgtable_bad(address, regs, error_code);
 381
 382         /*
 383          * If we're in an interrupt or have no user
 384          * context, we must not take the fault..
 385          */
 386         if (unlikely(in_atomic() || !mm))
 387                 goto bad_area_nosemaphore;
 388
 389  again:
 390         /* When running in the kernel we expect faults to occur only to
 391          * addresses in user space.  All other faults represent errors in the
 392          * kernel and should generate an OOPS.  Unfortunatly, in the case of an
 393          * erroneous fault occurring in a code path which already holds mmap_sem
 394          * we will deadlock attempting to validate the fault against the
 395          * address space.  Luckily the kernel only validly references user
 396          * space from well defined areas of code, which are listed in the
 397          * exceptions table.
 398          *
 399          * As the vast majority of faults will be valid we will only perform
 400          * the source reference check when there is a possibilty of a deadlock.
 401          * Attempt to lock the address space, if we cannot we then validate the
 402          * source.  If this is invalid we can skip the address space check,
 403          * thus avoiding the deadlock.
 404          */
 405         if (!down_read_trylock(&mm->mmap_sem)) {
 406                 if ((error_code & PF_USER) == 0 &&
 407                     !search_exception_tables(regs->rip))
 408                         goto bad_area_nosemaphore;
 409                 down_read(&mm->mmap_sem);
 410         }
 411
 412         vma = find_vma(mm, address);
 413         if (!vma)
 414                 goto bad_area;
 415         if (likely(vma->vm_start <= address))
 416                 goto good_area;
 417         if (!(vma->vm_flags & VM_GROWSDOWN))
 418                 goto bad_area;
 419         if (error_code & 4) {
 420                 /* Allow userspace just enough access below the stack pointer
 421                  * to let the 'enter' instruction work.
 422                  */
 423                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->rsp)
 424                         goto bad_area;
 425         }
 426         if (expand_stack(vma, address))
 427                 goto bad_area;
 428 /*
 429  * Ok, we have a good vm_area for this memory access, so
 430  * we can handle it..
 431  */
 432 good_area:
 433         info.si_code = SEGV_ACCERR;
 434         write = 0;
 435         switch (error_code & (PF_PROT|PF_WRITE)) {
 436                 default:        /* 3: write, present */
 437                         /* fall through */
 438                 case PF_WRITE:          /* write, not present */
 439                         if (!(vma->vm_flags & VM_WRITE))
 440                                 goto bad_area;
 441                         write++;
 442                         break;
 443                 case PF_PROT:           /* read, present */
 444                         goto bad_area;
 445                 case 0:                 /* read, not present */
 446                         if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 447                                 goto bad_area;
 448         }
 449
 450         /*
 451          * If for any reason at all we couldn't handle the fault,
 452          * make sure we exit gracefully rather than endlessly redo
 453          * the fault.
 454          */
 455         switch (handle_mm_fault(mm, vma, address, write)) {
 456         case VM_FAULT_MINOR:
 457                 tsk->min_flt++;
 458                 break;
 459         case VM_FAULT_MAJOR:
 460                 tsk->maj_flt++;
 461                 break;
 462         case VM_FAULT_SIGBUS:
 463                 goto do_sigbus;
 464         default:
 465                 goto out_of_memory;
 466         }
 467
 468         up_read(&mm->mmap_sem);
 469         return;
 470
 471 /*
 472  * Something tried to access memory that isn't in our memory map..
 473  * Fix it, but check if it's kernel or user first..
 474  */
 475 bad_area:
 476         up_read(&mm->mmap_sem);
 477
 478 bad_area_nosemaphore:
 479         /* User mode accesses just cause a SIGSEGV */
 480         if (error_code & PF_USER) {
 481                 if (is_prefetch(regs, address, error_code))
 482                         return;
 483
 484                 /* Work around K8 erratum #100 K8 in compat mode
 485                    occasionally jumps to illegal addresses >4GB.  We
 486                    catch this here in the page fault handler because
 487                    these addresses are not reachable. Just detect this
 488                    case and return.  Any code segment in LDT is
 489                    compatibility mode. */
 490                 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
 491                     (address >> 32))
 492                         return;
 493
 494                 if (exception_trace && unhandled_signal(tsk, SIGSEGV)) {
 495                         printk(
 496                        "%s%s[%d]: segfault at %016lx rip %016lx rsp %016lx error %lx\n",
 497                                         tsk->pid > 1 ? KERN_INFO : KERN_EMERG,
 498                                         tsk->comm, tsk->pid, address, regs->rip,
 499                                         regs->rsp, error_code);
 500                 }
 501
 502                 tsk->thread.cr2 = address;
 503                 /* Kernel addresses are always protection faults */
 504                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 505                 tsk->thread.trap_no = 14;
 506                 info.si_signo = SIGSEGV;
 507                 info.si_errno = 0;
 508                 /* info.si_code has been set above */
 509                 info.si_addr = (void __user *)address;
 510                 force_sig_info(SIGSEGV, &info, tsk);
 511                 return;
 512         }
 513
 514 no_context:
 515
 516         /* Are we prepared to handle this kernel fault?  */
 517         fixup = search_exception_tables(regs->rip);
 518         if (fixup) {
 519                 regs->rip = fixup->fixup;
 520                 return;
 521         }
 522
 523         /*
 524          * Hall of shame of CPU/BIOS bugs.
 525          */
 526
 527         if (is_prefetch(regs, address, error_code))
 528                 return;
 529
 530         if (is_errata93(regs, address))
 531                 return;
 532
 533 /*
 534  * Oops. The kernel tried to access some bad page. We'll have to
 535  * terminate things with extreme prejudice.
 536  */
 537
 538         flags = oops_begin();
 539
 540         if (address < PAGE_SIZE)
 541                 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
 542         else
 543                 printk(KERN_ALERT "Unable to handle kernel paging request");
 544         printk(" at %016lx RIP: \n" KERN_ALERT,address);
 545         printk_address(regs->rip);
 546         dump_pagetable(address);
 547         tsk->thread.cr2 = address;
 548         tsk->thread.trap_no = 14;
 549         tsk->thread.error_code = error_code;
 550         __die("Oops", regs, error_code);
 551         /* Executive summary in case the body of the oops scrolled away */
 552         printk(KERN_EMERG "CR2: %016lx\n", address);
 553         oops_end(flags);
 554         do_exit(SIGKILL);
 555
 556 /*
 557  * We ran out of memory, or some other thing happened to us that made
 558  * us unable to handle the page fault gracefully.
 559  */
 560 out_of_memory:
 561         up_read(&mm->mmap_sem);
 562         if (is_init(current)) {
 563                 yield();
 564                 goto again;
 565         }
 566         printk("VM: killing process %s\n", tsk->comm);
 567         if (error_code & 4)
 568                 do_exit(SIGKILL);
 569         goto no_context;
 570
 571 do_sigbus:
 572         up_read(&mm->mmap_sem);
 573
 574         /* Kernel mode? Handle exceptions or die */
 575         if (!(error_code & PF_USER))
 576                 goto no_context;
 577
 578         tsk->thread.cr2 = address;
 579         tsk->thread.error_code = error_code;
 580         tsk->thread.trap_no = 14;
 581         info.si_signo = SIGBUS;
 582         info.si_errno = 0;
 583         info.si_code = BUS_ADRERR;
 584         info.si_addr = (void __user *)address;
 585         force_sig_info(SIGBUS, &info, tsk);
 586         return;
 587 }
 588
 589 DEFINE_SPINLOCK(pgd_lock);
 590 struct page *pgd_list;
 591
 592 void vmalloc_sync_all(void)
 593 {
 594         /* Note that races in the updates of insync and start aren't
 595            problematic:
 596            insync can only get set bits added, and updates to start are only
 597            improving performance (without affecting correctness if undone). */
 598         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 599         static unsigned long start = VMALLOC_START & PGDIR_MASK;
 600         unsigned long address;
 601
 602         for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
 603                 if (!test_bit(pgd_index(address), insync)) {
 604                         const pgd_t *pgd_ref = pgd_offset_k(address);
 605                         struct page *page;
 606
 607                         if (pgd_none(*pgd_ref))
 608                                 continue;
 609                         spin_lock(&pgd_lock);
 610                         for (page = pgd_list; page;
 611                              page = (struct page *)page->index) {
 612                                 pgd_t *pgd;
 613                                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 614                                 if (pgd_none(*pgd))
 615                                         set_pgd(pgd, *pgd_ref);
 616                                 else
 617                                         BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 618                         }
 619                         spin_unlock(&pgd_lock);
 620                         set_bit(pgd_index(address), insync);
 621                 }
 622                 if (address == start)
 623                         start = address + PGDIR_SIZE;
 624         }
 625         /* Check that there is no need to do the same for the modules area. */
 626         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 627         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 628                                 (__START_KERNEL & PGDIR_MASK)));
 629 }
 630
 631 static int __init enable_pagefaulttrace(char *str)
 632 {
 633         page_fault_trace = 1;
 634         return 1;
 635 }
 636 __setup("pagefaulttrace", enable_pagefaulttrace);