arch/x86/mm/fault_32.c

   1 /*
   2  *  Copyright (C) 1995  Linus Torvalds
   3  */
   4
   5 #include <linux/signal.h>
   6 #include <linux/sched.h>
   7 #include <linux/kernel.h>
   8 #include <linux/errno.h>
   9 #include <linux/string.h>
  10 #include <linux/types.h>
  11 #include <linux/ptrace.h>
  12 #include <linux/mman.h>
  13 #include <linux/mm.h>
  14 #include <linux/smp.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/init.h>
  17 #include <linux/tty.h>
  18 #include <linux/vt_kern.h>              /* For unblank_screen() */
  19 #include <linux/highmem.h>
  20 #include <linux/bootmem.h>              /* for max_low_pfn */
  21 #include <linux/vmalloc.h>
  22 #include <linux/module.h>
  23 #include <linux/kprobes.h>
  24 #include <linux/uaccess.h>
  25 #include <linux/kdebug.h>
  26
  27 #include <asm/system.h>
  28 #include <asm/desc.h>
  29 #include <asm/segment.h>
  30
  31 /*
  32  * Page fault error code bits
  33  *      bit 0 == 0 means no page found, 1 means protection fault
  34  *      bit 1 == 0 means read, 1 means write
  35  *      bit 2 == 0 means kernel, 1 means user-mode
  36  *      bit 3 == 1 means use of reserved bit detected
  37  *      bit 4 == 1 means fault was an instruction fetch
  38  */
  39 #define PF_PROT (1<<0)
  40 #define PF_WRITE        (1<<1)
  41 #define PF_USER (1<<2)
  42 #define PF_RSVD (1<<3)
  43 #define PF_INSTR        (1<<4)
  44
  45 static inline int notify_page_fault(struct pt_regs *regs)
  46 {
  47 #ifdef CONFIG_KPROBES
  48         int ret = 0;
  49
  50         /* kprobe_running() needs smp_processor_id() */
  51         if (!user_mode_vm(regs)) {
  52                 preempt_disable();
  53                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  54                         ret = 1;
  55                 preempt_enable();
  56         }
  57
  58         return ret;
  59 #else
  60         return 0;
  61 #endif
  62 }
  63
  64 /*
  65  * X86_32
  66  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  67  * Check that here and ignore it.
  68  *
  69  * X86_64
  70  * Sometimes the CPU reports invalid exceptions on prefetch.
  71  * Check that here and ignore it.
  72  *
  73  * Opcode checker based on code by Richard Brunner
  74  */
  75 static 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 #ifdef CONFIG_X86_32
  84         if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
  85                      boot_cpu_data.x86 >= 6)) {
  86                 /* Catch an obscure case of prefetch inside an NX page. */
  87                 if (nx_enabled && (error_code & PF_INSTR))
  88                         return 0;
  89         } else {
  90                 return 0;
  91         }
  92 #else
  93         /* If it was a exec fault ignore */
  94         if (error_code & PF_INSTR)
  95                 return 0;
  96 #endif
  97
  98         instr = (unsigned char *)convert_ip_to_linear(current, regs);
  99         max_instr = instr + 15;
 100
 101         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 102                 return 0;
 103
 104         while (scan_more && instr < max_instr) {
 105                 unsigned char opcode;
 106                 unsigned char instr_hi;
 107                 unsigned char instr_lo;
 108
 109                 if (probe_kernel_address(instr, opcode))
 110                         break;
 111
 112                 instr_hi = opcode & 0xf0;
 113                 instr_lo = opcode & 0x0f;
 114                 instr++;
 115
 116                 switch (instr_hi) {
 117                 case 0x20:
 118                 case 0x30:
 119                         /*
 120                          * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
 121                          * In X86_64 long mode, the CPU will signal invalid
 122                          * opcode if some of these prefixes are present so
 123                          * X86_64 will never get here anyway
 124                          */
 125                         scan_more = ((instr_lo & 7) == 0x6);
 126                         break;
 127 #ifdef CONFIG_X86_64
 128                 case 0x40:
 129                         /*
 130                          * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 131                          * Need to figure out under what instruction mode the
 132                          * instruction was issued. Could check the LDT for lm,
 133                          * but for now it's good enough to assume that long
 134                          * mode only uses well known segments or kernel.
 135                          */
 136                         scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
 137                         break;
 138 #endif
 139                 case 0x60:
 140                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 141                         scan_more = (instr_lo & 0xC) == 0x4;
 142                         break;
 143                 case 0xF0:
 144                         /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 145                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 146                         break;
 147                 case 0x00:
 148                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 149                         scan_more = 0;
 150
 151                         if (probe_kernel_address(instr, opcode))
 152                                 break;
 153                         prefetch = (instr_lo == 0xF) &&
 154                                 (opcode == 0x0D || opcode == 0x18);
 155                         break;
 156                 default:
 157                         scan_more = 0;
 158                         break;
 159                 }
 160         }
 161         return prefetch;
 162 }
 163
 164 static void force_sig_info_fault(int si_signo, int si_code,
 165         unsigned long address, struct task_struct *tsk)
 166 {
 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         force_sig_info(si_signo, &info, tsk);
 174 }
 175
 176 void do_invalid_op(struct pt_regs *, unsigned long);
 177
 178 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 179 {
 180         unsigned index = pgd_index(address);
 181         pgd_t *pgd_k;
 182         pud_t *pud, *pud_k;
 183         pmd_t *pmd, *pmd_k;
 184
 185         pgd += index;
 186         pgd_k = init_mm.pgd + index;
 187
 188         if (!pgd_present(*pgd_k))
 189                 return NULL;
 190
 191         /*
 192          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 193          * and redundant with the set_pmd() on non-PAE. As would
 194          * set_pud.
 195          */
 196
 197         pud = pud_offset(pgd, address);
 198         pud_k = pud_offset(pgd_k, address);
 199         if (!pud_present(*pud_k))
 200                 return NULL;
 201
 202         pmd = pmd_offset(pud, address);
 203         pmd_k = pmd_offset(pud_k, address);
 204         if (!pmd_present(*pmd_k))
 205                 return NULL;
 206         if (!pmd_present(*pmd)) {
 207                 set_pmd(pmd, *pmd_k);
 208                 arch_flush_lazy_mmu_mode();
 209         } else
 210                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 211         return pmd_k;
 212 }
 213
 214 #ifdef CONFIG_X86_64
 215 static const char errata93_warning[] =
 216 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 217 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 218 KERN_ERR "******* Please consider a BIOS update.\n"
 219 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 220
 221 /* Workaround for K8 erratum #93 & buggy BIOS.
 222    BIOS SMM functions are required to use a specific workaround
 223    to avoid corruption of the 64bit RIP register on C stepping K8.
 224    A lot of BIOS that didn't get tested properly miss this.
 225    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 226    Try to work around it here.
 227    Note we only handle faults in kernel here. */
 228
 229 static int is_errata93(struct pt_regs *regs, unsigned long address)
 230 {
 231         static int warned;
 232         if (address != regs->ip)
 233                 return 0;
 234         if ((address >> 32) != 0)
 235                 return 0;
 236         address |= 0xffffffffUL << 32;
 237         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 238             (address >= MODULES_VADDR && address <= MODULES_END)) {
 239                 if (!warned) {
 240                         printk(errata93_warning);
 241                         warned = 1;
 242                 }
 243                 regs->ip = address;
 244                 return 1;
 245         }
 246         return 0;
 247 }
 248 #endif
 249
 250 /*
 251  * Handle a fault on the vmalloc or module mapping area
 252  *
 253  * This assumes no large pages in there.
 254  */
 255 static inline int vmalloc_fault(unsigned long address)
 256 {
 257         unsigned long pgd_paddr;
 258         pmd_t *pmd_k;
 259         pte_t *pte_k;
 260         /*
 261          * Synchronize this task's top level page-table
 262          * with the 'reference' page table.
 263          *
 264          * Do _not_ use "current" here. We might be inside
 265          * an interrupt in the middle of a task switch..
 266          */
 267         pgd_paddr = read_cr3();
 268         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 269         if (!pmd_k)
 270                 return -1;
 271         pte_k = pte_offset_kernel(pmd_k, address);
 272         if (!pte_present(*pte_k))
 273                 return -1;
 274         return 0;
 275 }
 276
 277 int show_unhandled_signals = 1;
 278
 279 /*
 280  * This routine handles page faults.  It determines the address,
 281  * and the problem, and then passes it off to one of the appropriate
 282  * routines.
 283  */
 284 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
 285 {
 286         struct task_struct *tsk;
 287         struct mm_struct *mm;
 288         struct vm_area_struct *vma;
 289         unsigned long address;
 290         int write, si_code;
 291         int fault;
 292
 293         /*
 294          * We can fault from pretty much anywhere, with unknown IRQ state.
 295          */
 296         trace_hardirqs_fixup();
 297
 298         /* get the address */
 299         address = read_cr2();
 300
 301         tsk = current;
 302
 303         si_code = SEGV_MAPERR;
 304
 305         /*
 306          * We fault-in kernel-space virtual memory on-demand. The
 307          * 'reference' page table is init_mm.pgd.
 308          *
 309          * NOTE! We MUST NOT take any locks for this case. We may
 310          * be in an interrupt or a critical region, and should
 311          * only copy the information from the master page table,
 312          * nothing more.
 313          *
 314          * This verifies that the fault happens in kernel space
 315          * (error_code & 4) == 0, and that the fault was not a
 316          * protection error (error_code & 9) == 0.
 317          */
 318         if (unlikely(address >= TASK_SIZE)) {
 319                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 320                     vmalloc_fault(address) >= 0)
 321                         return;
 322                 if (notify_page_fault(regs))
 323                         return;
 324                 /*
 325                  * Don't take the mm semaphore here. If we fixup a prefetch
 326                  * fault we could otherwise deadlock.
 327                  */
 328                 goto bad_area_nosemaphore;
 329         }
 330
 331         if (notify_page_fault(regs))
 332                 return;
 333
 334         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 335            fault has been handled. */
 336         if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
 337                 local_irq_enable();
 338
 339         mm = tsk->mm;
 340
 341         /*
 342          * If we're in an interrupt, have no user context or are running in an
 343          * atomic region then we must not take the fault.
 344          */
 345         if (in_atomic() || !mm)
 346                 goto bad_area_nosemaphore;
 347
 348         /* When running in the kernel we expect faults to occur only to
 349          * addresses in user space.  All other faults represent errors in the
 350          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 351          * erroneous fault occurring in a code path which already holds mmap_sem
 352          * we will deadlock attempting to validate the fault against the
 353          * address space.  Luckily the kernel only validly references user
 354          * space from well defined areas of code, which are listed in the
 355          * exceptions table.
 356          *
 357          * As the vast majority of faults will be valid we will only perform
 358          * the source reference check when there is a possibility of a deadlock.
 359          * Attempt to lock the address space, if we cannot we then validate the
 360          * source.  If this is invalid we can skip the address space check,
 361          * thus avoiding the deadlock.
 362          */
 363         if (!down_read_trylock(&mm->mmap_sem)) {
 364                 if ((error_code & PF_USER) == 0 &&
 365                     !search_exception_tables(regs->ip))
 366                         goto bad_area_nosemaphore;
 367                 down_read(&mm->mmap_sem);
 368         }
 369
 370         vma = find_vma(mm, address);
 371         if (!vma)
 372                 goto bad_area;
 373         if (vma->vm_start <= address)
 374                 goto good_area;
 375         if (!(vma->vm_flags & VM_GROWSDOWN))
 376                 goto bad_area;
 377         if (error_code & PF_USER) {
 378                 /*
 379                  * Accessing the stack below %sp is always a bug.
 380                  * The large cushion allows instructions like enter
 381                  * and pusha to work.  ("enter $65535,$31" pushes
 382                  * 32 pointers and then decrements %sp by 65535.)
 383                  */
 384                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
 385                         goto bad_area;
 386         }
 387         if (expand_stack(vma, address))
 388                 goto bad_area;
 389 /*
 390  * Ok, we have a good vm_area for this memory access, so
 391  * we can handle it..
 392  */
 393 good_area:
 394         si_code = SEGV_ACCERR;
 395         write = 0;
 396         switch (error_code & (PF_PROT|PF_WRITE)) {
 397         default:        /* 3: write, present */
 398                 /* fall through */
 399         case PF_WRITE:          /* write, not present */
 400                 if (!(vma->vm_flags & VM_WRITE))
 401                         goto bad_area;
 402                 write++;
 403                 break;
 404         case PF_PROT:           /* read, present */
 405                 goto bad_area;
 406         case 0:                 /* read, not present */
 407                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 408                         goto bad_area;
 409         }
 410
 411  survive:
 412         /*
 413          * If for any reason at all we couldn't handle the fault,
 414          * make sure we exit gracefully rather than endlessly redo
 415          * the fault.
 416          */
 417         fault = handle_mm_fault(mm, vma, address, write);
 418         if (unlikely(fault & VM_FAULT_ERROR)) {
 419                 if (fault & VM_FAULT_OOM)
 420                         goto out_of_memory;
 421                 else if (fault & VM_FAULT_SIGBUS)
 422                         goto do_sigbus;
 423                 BUG();
 424         }
 425         if (fault & VM_FAULT_MAJOR)
 426                 tsk->maj_flt++;
 427         else
 428                 tsk->min_flt++;
 429
 430         /*
 431          * Did it hit the DOS screen memory VA from vm86 mode?
 432          */
 433         if (regs->flags & VM_MASK) {
 434                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 435                 if (bit < 32)
 436                         tsk->thread.screen_bitmap |= 1 << bit;
 437         }
 438         up_read(&mm->mmap_sem);
 439         return;
 440
 441 /*
 442  * Something tried to access memory that isn't in our memory map..
 443  * Fix it, but check if it's kernel or user first..
 444  */
 445 bad_area:
 446         up_read(&mm->mmap_sem);
 447
 448 bad_area_nosemaphore:
 449         /* User mode accesses just cause a SIGSEGV */
 450         if (error_code & PF_USER) {
 451                 /*
 452                  * It's possible to have interrupts off here.
 453                  */
 454                 local_irq_enable();
 455
 456                 /*
 457                  * Valid to do another page fault here because this one came
 458                  * from user space.
 459                  */
 460                 if (is_prefetch(regs, address, error_code))
 461                         return;
 462
 463                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 464                     printk_ratelimit()) {
 465                         printk("%s%s[%d]: segfault at %08lx ip %08lx "
 466                             "sp %08lx error %lx\n",
 467                             task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 468                             tsk->comm, task_pid_nr(tsk), address, regs->ip,
 469                             regs->sp, error_code);
 470                 }
 471                 tsk->thread.cr2 = address;
 472                 /* Kernel addresses are always protection faults */
 473                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 474                 tsk->thread.trap_no = 14;
 475                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 476                 return;
 477         }
 478
 479 #ifdef CONFIG_X86_F00F_BUG
 480         /*
 481          * Pentium F0 0F C7 C8 bug workaround.
 482          */
 483         if (boot_cpu_data.f00f_bug) {
 484                 unsigned long nr;
 485
 486                 nr = (address - idt_descr.address) >> 3;
 487
 488                 if (nr == 6) {
 489                         do_invalid_op(regs, 0);
 490                         return;
 491                 }
 492         }
 493 #endif
 494
 495 no_context:
 496         /* Are we prepared to handle this kernel fault?  */
 497         if (fixup_exception(regs))
 498                 return;
 499
 500         /*
 501          * Valid to do another page fault here, because if this fault
 502          * had been triggered by is_prefetch fixup_exception would have
 503          * handled it.
 504          */
 505         if (is_prefetch(regs, address, error_code))
 506                 return;
 507
 508 /*
 509  * Oops. The kernel tried to access some bad page. We'll have to
 510  * terminate things with extreme prejudice.
 511  */
 512
 513         bust_spinlocks(1);
 514
 515         if (oops_may_print()) {
 516                 __typeof__(pte_val(__pte(0))) page;
 517
 518 #ifdef CONFIG_X86_PAE
 519                 if (error_code & PF_INSTR) {
 520                         pte_t *pte = lookup_address(address);
 521
 522                         if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 523                                 printk(KERN_CRIT "kernel tried to execute "
 524                                         "NX-protected page - exploit attempt? "
 525                                         "(uid: %d)\n", current->uid);
 526                 }
 527 #endif
 528                 if (address < PAGE_SIZE)
 529                         printk(KERN_ALERT "BUG: unable to handle kernel NULL "
 530                                         "pointer dereference");
 531                 else
 532                         printk(KERN_ALERT "BUG: unable to handle kernel paging"
 533                                         " request");
 534                 printk(" at virtual address %08lx\n", address);
 535                 printk(KERN_ALERT "printing ip: %08lx ", regs->ip);
 536
 537                 page = read_cr3();
 538                 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 539 #ifdef CONFIG_X86_PAE
 540                 printk("*pdpt = %016Lx ", page);
 541                 if ((page >> PAGE_SHIFT) < max_low_pfn
 542                     && page & _PAGE_PRESENT) {
 543                         page &= PAGE_MASK;
 544                         page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 545                                                                  & (PTRS_PER_PMD - 1)];
 546                         printk(KERN_CONT "*pde = %016Lx ", page);
 547                         page &= ~_PAGE_NX;
 548                 }
 549 #else
 550                 printk("*pde = %08lx ", page);
 551 #endif
 552
 553                 /*
 554                  * We must not directly access the pte in the highpte
 555                  * case if the page table is located in highmem.
 556                  * And let's rather not kmap-atomic the pte, just in case
 557                  * it's allocated already.
 558                  */
 559                 if ((page >> PAGE_SHIFT) < max_low_pfn
 560                     && (page & _PAGE_PRESENT)
 561                     && !(page & _PAGE_PSE)) {
 562                         page &= PAGE_MASK;
 563                         page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 564                                                                  & (PTRS_PER_PTE - 1)];
 565                         printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 566                 }
 567
 568                 printk("\n");
 569         }
 570
 571         tsk->thread.cr2 = address;
 572         tsk->thread.trap_no = 14;
 573         tsk->thread.error_code = error_code;
 574         die("Oops", regs, error_code);
 575         bust_spinlocks(0);
 576         do_exit(SIGKILL);
 577
 578 /*
 579  * We ran out of memory, or some other thing happened to us that made
 580  * us unable to handle the page fault gracefully.
 581  */
 582 out_of_memory:
 583         up_read(&mm->mmap_sem);
 584         if (is_global_init(tsk)) {
 585                 yield();
 586                 down_read(&mm->mmap_sem);
 587                 goto survive;
 588         }
 589         printk("VM: killing process %s\n", tsk->comm);
 590         if (error_code & PF_USER)
 591                 do_group_exit(SIGKILL);
 592         goto no_context;
 593
 594 do_sigbus:
 595         up_read(&mm->mmap_sem);
 596
 597         /* Kernel mode? Handle exceptions or die */
 598         if (!(error_code & PF_USER))
 599                 goto no_context;
 600
 601         /* User space => ok to do another page fault */
 602         if (is_prefetch(regs, address, error_code))
 603                 return;
 604
 605         tsk->thread.cr2 = address;
 606         tsk->thread.error_code = error_code;
 607         tsk->thread.trap_no = 14;
 608         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 609 }
 610
 611 void vmalloc_sync_all(void)
 612 {
 613         /*
 614          * Note that races in the updates of insync and start aren't
 615          * problematic: insync can only get set bits added, and updates to
 616          * start are only improving performance (without affecting correctness
 617          * if undone).
 618          */
 619         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 620         static unsigned long start = TASK_SIZE;
 621         unsigned long address;
 622
 623         if (SHARED_KERNEL_PMD)
 624                 return;
 625
 626         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 627         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 628                 if (!test_bit(pgd_index(address), insync)) {
 629                         unsigned long flags;
 630                         struct page *page;
 631
 632                         spin_lock_irqsave(&pgd_lock, flags);
 633                         for (page = pgd_list; page; page =
 634                                         (struct page *)page->index)
 635                                 if (!vmalloc_sync_one(page_address(page),
 636                                                                 address)) {
 637                                         BUG_ON(page != pgd_list);
 638                                         break;
 639                                 }
 640                         spin_unlock_irqrestore(&pgd_lock, flags);
 641                         if (!page)
 642                                 set_bit(pgd_index(address), insync);
 643                 }
 644                 if (address == start && test_bit(pgd_index(address), insync))
 645                         start = address + PGDIR_SIZE;
 646         }
 647 }