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/config.h>
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/string.h>
  14 #include <linux/types.h>
  15 #include <linux/ptrace.h>
  16 #include <linux/mman.h>
  17 #include <linux/mm.h>
  18 #include <linux/smp.h>
  19 #include <linux/smp_lock.h>
  20 #include <linux/interrupt.h>
  21 #include <linux/init.h>
  22 #include <linux/tty.h>
  23 #include <linux/vt_kern.h>              /* For unblank_screen() */
  24 #include <linux/compiler.h>
  25 #include <linux/module.h>
  26 #include <linux/kprobes.h>
  27
  28 #include <asm/system.h>
  29 #include <asm/uaccess.h>
  30 #include <asm/pgalloc.h>
  31 #include <asm/smp.h>
  32 #include <asm/tlbflush.h>
  33 #include <asm/proto.h>
  34 #include <asm/kdebug.h>
  35 #include <asm-generic/sections.h>
  36 #include <asm/kdebug.h>
  37
  38 void bust_spinlocks(int yes)
  39 {
  40         int loglevel_save = console_loglevel;
  41         if (yes) {
  42                 oops_in_progress = 1;
  43         } else {
  44 #ifdef CONFIG_VT
  45                 unblank_screen();
  46 #endif
  47                 oops_in_progress = 0;
  48                 /*
  49                  * OK, the message is on the console.  Now we call printk()
  50                  * without oops_in_progress set so that printk will give klogd
  51                  * a poke.  Hold onto your hats...
  52                  */
  53                 console_loglevel = 15;          /* NMI oopser may have shut the console up */
  54                 printk(" ");
  55                 console_loglevel = loglevel_save;
  56         }
  57 }
  58
  59 /* Sometimes the CPU reports invalid exceptions on prefetch.
  60    Check that here and ignore.
  61    Opcode checker based on code by Richard Brunner */
  62 static noinline int is_prefetch(struct pt_regs *regs, unsigned long addr,
  63                                 unsigned long error_code)
  64 {
  65         unsigned char *instr = (unsigned char *)(regs->rip);
  66         int scan_more = 1;
  67         int prefetch = 0;
  68         unsigned char *max_instr = instr + 15;
  69
  70         /* If it was a exec fault ignore */
  71         if (error_code & (1<<4))
  72                 return 0;
  73
  74         /* Code segments in LDT could have a non zero base. Don't check
  75            when that's possible */
  76         if (regs->cs & (1<<2))
  77                 return 0;
  78
  79         if ((regs->cs & 3) != 0 && regs->rip >= TASK_SIZE)
  80                 return 0;
  81
  82         while (scan_more && instr < max_instr) {
  83                 unsigned char opcode;
  84                 unsigned char instr_hi;
  85                 unsigned char instr_lo;
  86
  87                 if (__get_user(opcode, instr))
  88                         break;
  89
  90                 instr_hi = opcode & 0xf0;
  91                 instr_lo = opcode & 0x0f;
  92                 instr++;
  93
  94                 switch (instr_hi) {
  95                 case 0x20:
  96                 case 0x30:
  97                         /* Values 0x26,0x2E,0x36,0x3E are valid x86
  98                            prefixes.  In long mode, the CPU will signal
  99                            invalid opcode if some of these prefixes are
 100                            present so we will never get here anyway */
 101                         scan_more = ((instr_lo & 7) == 0x6);
 102                         break;
 103
 104                 case 0x40:
 105                         /* In AMD64 long mode, 0x40 to 0x4F are valid REX prefixes
 106                            Need to figure out under what instruction mode the
 107                            instruction was issued ... */
 108                         /* Could check the LDT for lm, but for now it's good
 109                            enough to assume that long mode only uses well known
 110                            segments or kernel. */
 111                         scan_more = ((regs->cs & 3) == 0) || (regs->cs == __USER_CS);
 112                         break;
 113
 114                 case 0x60:
 115                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 116                         scan_more = (instr_lo & 0xC) == 0x4;
 117                         break;
 118                 case 0xF0:
 119                         /* 0xF0, 0xF2, and 0xF3 are valid prefixes in all modes. */
 120                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 121                         break;
 122                 case 0x00:
 123                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 124                         scan_more = 0;
 125                         if (__get_user(opcode, instr))
 126                                 break;
 127                         prefetch = (instr_lo == 0xF) &&
 128                                 (opcode == 0x0D || opcode == 0x18);
 129                         break;
 130                 default:
 131                         scan_more = 0;
 132                         break;
 133                 }
 134         }
 135         return prefetch;
 136 }
 137
 138 static int bad_address(void *p)
 139 {
 140         unsigned long dummy;
 141         return __get_user(dummy, (unsigned long *)p);
 142 }
 143
 144 void dump_pagetable(unsigned long address)
 145 {
 146         pgd_t *pgd;
 147         pud_t *pud;
 148         pmd_t *pmd;
 149         pte_t *pte;
 150
 151         asm("movq %%cr3,%0" : "=r" (pgd));
 152
 153         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 154         pgd += pgd_index(address);
 155         printk("PGD %lx ", pgd_val(*pgd));
 156         if (bad_address(pgd)) goto bad;
 157         if (!pgd_present(*pgd)) goto ret;
 158
 159         pud = __pud_offset_k((pud_t *)pgd_page(*pgd), address);
 160         if (bad_address(pud)) goto bad;
 161         printk("PUD %lx ", pud_val(*pud));
 162         if (!pud_present(*pud)) goto ret;
 163
 164         pmd = pmd_offset(pud, address);
 165         if (bad_address(pmd)) goto bad;
 166         printk("PMD %lx ", pmd_val(*pmd));
 167         if (!pmd_present(*pmd)) goto ret;
 168
 169         pte = pte_offset_kernel(pmd, address);
 170         if (bad_address(pte)) goto bad;
 171         printk("PTE %lx", pte_val(*pte));
 172 ret:
 173         printk("\n");
 174         return;
 175 bad:
 176         printk("BAD\n");
 177 }
 178
 179 static const char errata93_warning[] =
 180 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 181 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 182 KERN_ERR "******* Please consider a BIOS update.\n"
 183 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 184
 185 /* Workaround for K8 erratum #93 & buggy BIOS.
 186    BIOS SMM functions are required to use a specific workaround
 187    to avoid corruption of the 64bit RIP register on C stepping K8.
 188    A lot of BIOS that didn't get tested properly miss this.
 189    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 190    Try to work around it here.
 191    Note we only handle faults in kernel here. */
 192
 193 static int is_errata93(struct pt_regs *regs, unsigned long address)
 194 {
 195         static int warned;
 196         if (address != regs->rip)
 197                 return 0;
 198         if ((address >> 32) != 0)
 199                 return 0;
 200         address |= 0xffffffffUL << 32;
 201         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 202             (address >= MODULES_VADDR && address <= MODULES_END)) {
 203                 if (!warned) {
 204                         printk(errata93_warning);
 205                         warned = 1;
 206                 }
 207                 regs->rip = address;
 208                 return 1;
 209         }
 210         return 0;
 211 }
 212
 213 int unhandled_signal(struct task_struct *tsk, int sig)
 214 {
 215         if (tsk->pid == 1)
 216                 return 1;
 217         /* Warn for strace, but not for gdb */
 218         if (!test_ti_thread_flag(tsk->thread_info, TIF_SYSCALL_TRACE) &&
 219             (tsk->ptrace & PT_PTRACED))
 220                 return 0;
 221         return (tsk->sighand->action[sig-1].sa.sa_handler == SIG_IGN) ||
 222                 (tsk->sighand->action[sig-1].sa.sa_handler == SIG_DFL);
 223 }
 224
 225 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
 226                                  unsigned long error_code)
 227 {
 228         oops_begin();
 229         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 230                current->comm, address);
 231         dump_pagetable(address);
 232         __die("Bad pagetable", regs, error_code);
 233         oops_end();
 234         do_exit(SIGKILL);
 235 }
 236
 237 /*
 238  * Handle a fault on the vmalloc or module mapping area
 239  */
 240 static int vmalloc_fault(unsigned long address)
 241 {
 242         pgd_t *pgd, *pgd_ref;
 243         pud_t *pud, *pud_ref;
 244         pmd_t *pmd, *pmd_ref;
 245         pte_t *pte, *pte_ref;
 246
 247         /* Copy kernel mappings over when needed. This can also
 248            happen within a race in page table update. In the later
 249            case just flush. */
 250
 251         pgd = pgd_offset(current->mm ?: &init_mm, address);
 252         pgd_ref = pgd_offset_k(address);
 253         if (pgd_none(*pgd_ref))
 254                 return -1;
 255         if (pgd_none(*pgd))
 256                 set_pgd(pgd, *pgd_ref);
 257
 258         /* Below here mismatches are bugs because these lower tables
 259            are shared */
 260
 261         pud = pud_offset(pgd, address);
 262         pud_ref = pud_offset(pgd_ref, address);
 263         if (pud_none(*pud_ref))
 264                 return -1;
 265         if (pud_none(*pud) || pud_page(*pud) != pud_page(*pud_ref))
 266                 BUG();
 267         pmd = pmd_offset(pud, address);
 268         pmd_ref = pmd_offset(pud_ref, address);
 269         if (pmd_none(*pmd_ref))
 270                 return -1;
 271         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 272                 BUG();
 273         pte_ref = pte_offset_kernel(pmd_ref, address);
 274         if (!pte_present(*pte_ref))
 275                 return -1;
 276         pte = pte_offset_kernel(pmd, address);
 277         if (!pte_present(*pte) || pte_page(*pte) != pte_page(*pte_ref))
 278                 BUG();
 279         __flush_tlb_all();
 280         return 0;
 281 }
 282
 283 int page_fault_trace = 0;
 284 int exception_trace = 1;
 285
 286 /*
 287  * This routine handles page faults.  It determines the address,
 288  * and the problem, and then passes it off to one of the appropriate
 289  * routines.
 290  *
 291  * error_code:
 292  *      bit 0 == 0 means no page found, 1 means protection fault
 293  *      bit 1 == 0 means read, 1 means write
 294  *      bit 2 == 0 means kernel, 1 means user-mode
 295  *      bit 3 == 1 means fault was an instruction fetch
 296  */
 297 asmlinkage void do_page_fault(struct pt_regs *regs, unsigned long error_code)
 298 {
 299         struct task_struct *tsk;
 300         struct mm_struct *mm;
 301         struct vm_area_struct * vma;
 302         unsigned long address;
 303         const struct exception_table_entry *fixup;
 304         int write;
 305         siginfo_t info;
 306
 307 #ifdef CONFIG_CHECKING
 308         {
 309                 unsigned long gs;
 310                 struct x8664_pda *pda = cpu_pda + stack_smp_processor_id();
 311                 rdmsrl(MSR_GS_BASE, gs);
 312                 if (gs != (unsigned long)pda) {
 313                         wrmsrl(MSR_GS_BASE, pda);
 314                         printk("page_fault: wrong gs %lx expected %p\n", gs, pda);
 315                 }
 316         }
 317 #endif
 318
 319         /* get the address */
 320         __asm__("movq %%cr2,%0":"=r" (address));
 321         if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 322                                         SIGSEGV) == NOTIFY_STOP)
 323                 return;
 324
 325         if (likely(regs->eflags & X86_EFLAGS_IF))
 326                 local_irq_enable();
 327
 328         if (unlikely(page_fault_trace))
 329                 printk("pagefault rip:%lx rsp:%lx cs:%lu ss:%lu address %lx error %lx\n",
 330                        regs->rip,regs->rsp,regs->cs,regs->ss,address,error_code);
 331
 332         tsk = current;
 333         mm = tsk->mm;
 334         info.si_code = SEGV_MAPERR;
 335
 336
 337         /*
 338          * We fault-in kernel-space virtual memory on-demand. The
 339          * 'reference' page table is init_mm.pgd.
 340          *
 341          * NOTE! We MUST NOT take any locks for this case. We may
 342          * be in an interrupt or a critical region, and should
 343          * only copy the information from the master page table,
 344          * nothing more.
 345          *
 346          * This verifies that the fault happens in kernel space
 347          * (error_code & 4) == 0, and that the fault was not a
 348          * protection error (error_code & 1) == 0.
 349          */
 350         if (unlikely(address >= TASK_SIZE)) {
 351                 if (!(error_code & 5)) {
 352                         if (vmalloc_fault(address) < 0)
 353                                 goto bad_area_nosemaphore;
 354                         return;
 355                 }
 356                 /*
 357                  * Don't take the mm semaphore here. If we fixup a prefetch
 358                  * fault we could otherwise deadlock.
 359                  */
 360                 goto bad_area_nosemaphore;
 361         }
 362
 363         if (unlikely(error_code & (1 << 3)))
 364                 pgtable_bad(address, regs, error_code);
 365
 366         /*
 367          * If we're in an interrupt or have no user
 368          * context, we must not take the fault..
 369          */
 370         if (unlikely(in_atomic() || !mm))
 371                 goto bad_area_nosemaphore;
 372
 373  again:
 374         /* When running in the kernel we expect faults to occur only to
 375          * addresses in user space.  All other faults represent errors in the
 376          * kernel and should generate an OOPS.  Unfortunatly, in the case of an
 377          * erroneous fault occuring in a code path which already holds mmap_sem
 378          * we will deadlock attempting to validate the fault against the
 379          * address space.  Luckily the kernel only validly references user
 380          * space from well defined areas of code, which are listed in the
 381          * exceptions table.
 382          *
 383          * As the vast majority of faults will be valid we will only perform
 384          * the source reference check when there is a possibilty of a deadlock.
 385          * Attempt to lock the address space, if we cannot we then validate the
 386          * source.  If this is invalid we can skip the address space check,
 387          * thus avoiding the deadlock.
 388          */
 389         if (!down_read_trylock(&mm->mmap_sem)) {
 390                 if ((error_code & 4) == 0 &&
 391                     !search_exception_tables(regs->rip))
 392                         goto bad_area_nosemaphore;
 393                 down_read(&mm->mmap_sem);
 394         }
 395
 396         vma = find_vma(mm, address);
 397         if (!vma)
 398                 goto bad_area;
 399         if (likely(vma->vm_start <= address))
 400                 goto good_area;
 401         if (!(vma->vm_flags & VM_GROWSDOWN))
 402                 goto bad_area;
 403         if (error_code & 4) {
 404                 // XXX: align red zone size with ABI
 405                 if (address + 128 < regs->rsp)
 406                         goto bad_area;
 407         }
 408         if (expand_stack(vma, address))
 409                 goto bad_area;
 410 /*
 411  * Ok, we have a good vm_area for this memory access, so
 412  * we can handle it..
 413  */
 414 good_area:
 415         info.si_code = SEGV_ACCERR;
 416         write = 0;
 417         switch (error_code & 3) {
 418                 default:        /* 3: write, present */
 419                         /* fall through */
 420                 case 2:         /* write, not present */
 421                         if (!(vma->vm_flags & VM_WRITE))
 422                                 goto bad_area;
 423                         write++;
 424                         break;
 425                 case 1:         /* read, present */
 426                         goto bad_area;
 427                 case 0:         /* read, not present */
 428                         if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
 429                                 goto bad_area;
 430         }
 431
 432         /*
 433          * If for any reason at all we couldn't handle the fault,
 434          * make sure we exit gracefully rather than endlessly redo
 435          * the fault.
 436          */
 437         switch (handle_mm_fault(mm, vma, address, write)) {
 438         case 1:
 439                 tsk->min_flt++;
 440                 break;
 441         case 2:
 442                 tsk->maj_flt++;
 443                 break;
 444         case 0:
 445                 goto do_sigbus;
 446         default:
 447                 goto out_of_memory;
 448         }
 449
 450         up_read(&mm->mmap_sem);
 451         return;
 452
 453 /*
 454  * Something tried to access memory that isn't in our memory map..
 455  * Fix it, but check if it's kernel or user first..
 456  */
 457 bad_area:
 458         up_read(&mm->mmap_sem);
 459
 460 bad_area_nosemaphore:
 461
 462 #ifdef CONFIG_IA32_EMULATION
 463         /* 32bit vsyscall. map on demand. */
 464         if (test_thread_flag(TIF_IA32) &&
 465             address >= VSYSCALL32_BASE && address < VSYSCALL32_END) {
 466                 if (map_syscall32(mm, address) < 0)
 467                         goto out_of_memory2;
 468                 return;
 469         }
 470 #endif
 471
 472         /* User mode accesses just cause a SIGSEGV */
 473         if (error_code & 4) {
 474                 if (is_prefetch(regs, address, error_code))
 475                         return;
 476
 477                 /* Work around K8 erratum #100 K8 in compat mode
 478                    occasionally jumps to illegal addresses >4GB.  We
 479                    catch this here in the page fault handler because
 480                    these addresses are not reachable. Just detect this
 481                    case and return.  Any code segment in LDT is
 482                    compatibility mode. */
 483                 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
 484                     (address >> 32))
 485                         return;
 486
 487                 if (exception_trace && unhandled_signal(tsk, SIGSEGV)) {
 488                         printk(
 489                        "%s%s[%d]: segfault at %016lx rip %016lx rsp %016lx error %lx\n",
 490                                         tsk->pid > 1 ? KERN_INFO : KERN_EMERG,
 491                                         tsk->comm, tsk->pid, address, regs->rip,
 492                                         regs->rsp, error_code);
 493                 }
 494
 495                 tsk->thread.cr2 = address;
 496                 /* Kernel addresses are always protection faults */
 497                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 498                 tsk->thread.trap_no = 14;
 499                 info.si_signo = SIGSEGV;
 500                 info.si_errno = 0;
 501                 /* info.si_code has been set above */
 502                 info.si_addr = (void __user *)address;
 503                 force_sig_info(SIGSEGV, &info, tsk);
 504                 return;
 505         }
 506
 507 no_context:
 508
 509         /* Are we prepared to handle this kernel fault?  */
 510         fixup = search_exception_tables(regs->rip);
 511         if (fixup) {
 512                 regs->rip = fixup->fixup;
 513                 return;
 514         }
 515
 516         /*
 517          * Hall of shame of CPU/BIOS bugs.
 518          */
 519
 520         if (is_prefetch(regs, address, error_code))
 521                 return;
 522
 523         if (is_errata93(regs, address))
 524                 return;
 525
 526 /*
 527  * Oops. The kernel tried to access some bad page. We'll have to
 528  * terminate things with extreme prejudice.
 529  */
 530
 531         oops_begin();
 532
 533         if (address < PAGE_SIZE)
 534                 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
 535         else
 536                 printk(KERN_ALERT "Unable to handle kernel paging request");
 537         printk(" at %016lx RIP: \n" KERN_ALERT,address);
 538         printk_address(regs->rip);
 539         printk("\n");
 540         dump_pagetable(address);
 541         __die("Oops", regs, error_code);
 542         /* Executive summary in case the body of the oops scrolled away */
 543         printk(KERN_EMERG "CR2: %016lx\n", address);
 544         oops_end();
 545         do_exit(SIGKILL);
 546
 547 /*
 548  * We ran out of memory, or some other thing happened to us that made
 549  * us unable to handle the page fault gracefully.
 550  */
 551 out_of_memory:
 552         up_read(&mm->mmap_sem);
 553 out_of_memory2:
 554         if (current->pid == 1) {
 555                 yield();
 556                 goto again;
 557         }
 558         printk("VM: killing process %s\n", tsk->comm);
 559         if (error_code & 4)
 560                 do_exit(SIGKILL);
 561         goto no_context;
 562
 563 do_sigbus:
 564         up_read(&mm->mmap_sem);
 565
 566         /* Kernel mode? Handle exceptions or die */
 567         if (!(error_code & 4))
 568                 goto no_context;
 569
 570         tsk->thread.cr2 = address;
 571         tsk->thread.error_code = error_code;
 572         tsk->thread.trap_no = 14;
 573         info.si_signo = SIGBUS;
 574         info.si_errno = 0;
 575         info.si_code = BUS_ADRERR;
 576         info.si_addr = (void __user *)address;
 577         force_sig_info(SIGBUS, &info, tsk);
 578         return;
 579 }