arch/x86/mm/fault_32.c

   1 /*
   2  *  linux/arch/i386/mm/fault.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  */
   6
   7 #include <linux/signal.h>
   8 #include <linux/sched.h>
   9 #include <linux/kernel.h>
  10 #include <linux/errno.h>
  11 #include <linux/string.h>
  12 #include <linux/types.h>
  13 #include <linux/ptrace.h>
  14 #include <linux/mman.h>
  15 #include <linux/mm.h>
  16 #include <linux/smp.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/init.h>
  19 #include <linux/tty.h>
  20 #include <linux/vt_kern.h>              /* For unblank_screen() */
  21 #include <linux/highmem.h>
  22 #include <linux/bootmem.h>              /* for max_low_pfn */
  23 #include <linux/vmalloc.h>
  24 #include <linux/module.h>
  25 #include <linux/kprobes.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/kdebug.h>
  28 #include <linux/kprobes.h>
  29
  30 #include <asm/system.h>
  31 #include <asm/desc.h>
  32 #include <asm/segment.h>
  33
  34 extern void die(const char *,struct pt_regs *,long);
  35
  36 #ifdef CONFIG_KPROBES
  37 static inline int notify_page_fault(struct pt_regs *regs)
  38 {
  39         int ret = 0;
  40
  41         /* kprobe_running() needs smp_processor_id() */
  42         if (!user_mode_vm(regs)) {
  43                 preempt_disable();
  44                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  45                         ret = 1;
  46                 preempt_enable();
  47         }
  48
  49         return ret;
  50 }
  51 #else
  52 static inline int notify_page_fault(struct pt_regs *regs)
  53 {
  54         return 0;
  55 }
  56 #endif
  57
  58 /*
  59  * Return EIP plus the CS segment base.  The segment limit is also
  60  * adjusted, clamped to the kernel/user address space (whichever is
  61  * appropriate), and returned in *eip_limit.
  62  *
  63  * The segment is checked, because it might have been changed by another
  64  * task between the original faulting instruction and here.
  65  *
  66  * If CS is no longer a valid code segment, or if EIP is beyond the
  67  * limit, or if it is a kernel address when CS is not a kernel segment,
  68  * then the returned value will be greater than *eip_limit.
  69  *
  70  * This is slow, but is very rarely executed.
  71  */
  72 static inline unsigned long get_segment_eip(struct pt_regs *regs,
  73                                             unsigned long *eip_limit)
  74 {
  75         unsigned long eip = regs->eip;
  76         unsigned seg = regs->xcs & 0xffff;
  77         u32 seg_ar, seg_limit, base, *desc;
  78
  79         /* Unlikely, but must come before segment checks. */
  80         if (unlikely(regs->eflags & VM_MASK)) {
  81                 base = seg << 4;
  82                 *eip_limit = base + 0xffff;
  83                 return base + (eip & 0xffff);
  84         }
  85
  86         /* The standard kernel/user address space limit. */
  87         *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
  88
  89         /* By far the most common cases. */
  90         if (likely(SEGMENT_IS_FLAT_CODE(seg)))
  91                 return eip;
  92
  93         /* Check the segment exists, is within the current LDT/GDT size,
  94            that kernel/user (ring 0..3) has the appropriate privilege,
  95            that it's a code segment, and get the limit. */
  96         __asm__ ("larl %3,%0; lsll %3,%1"
  97                  : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
  98         if ((~seg_ar & 0x9800) || eip > seg_limit) {
  99                 *eip_limit = 0;
 100                 return 1;        /* So that returned eip > *eip_limit. */
 101         }
 102
 103         /* Get the GDT/LDT descriptor base.
 104            When you look for races in this code remember that
 105            LDT and other horrors are only used in user space. */
 106         if (seg & (1<<2)) {
 107                 /* Must lock the LDT while reading it. */
 108                 mutex_lock(&current->mm->context.lock);
 109                 desc = current->mm->context.ldt;
 110                 desc = (void *)desc + (seg & ~7);
 111         } else {
 112                 /* Must disable preemption while reading the GDT. */
 113                 desc = (u32 *)get_cpu_gdt_table(get_cpu());
 114                 desc = (void *)desc + (seg & ~7);
 115         }
 116
 117         /* Decode the code segment base from the descriptor */
 118         base = get_desc_base((unsigned long *)desc);
 119
 120         if (seg & (1<<2)) {
 121                 mutex_unlock(&current->mm->context.lock);
 122         } else
 123                 put_cpu();
 124
 125         /* Adjust EIP and segment limit, and clamp at the kernel limit.
 126            It's legitimate for segments to wrap at 0xffffffff. */
 127         seg_limit += base;
 128         if (seg_limit < *eip_limit && seg_limit >= base)
 129                 *eip_limit = seg_limit;
 130         return eip + base;
 131 }
 132
 133 /*
 134  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
 135  * Check that here and ignore it.
 136  */
 137 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
 138 {
 139         unsigned long limit;
 140         unsigned char *instr = (unsigned char *)get_segment_eip (regs, &limit);
 141         int scan_more = 1;
 142         int prefetch = 0;
 143         int i;
 144
 145         for (i = 0; scan_more && i < 15; i++) {
 146                 unsigned char opcode;
 147                 unsigned char instr_hi;
 148                 unsigned char instr_lo;
 149
 150                 if (instr > (unsigned char *)limit)
 151                         break;
 152                 if (probe_kernel_address(instr, opcode))
 153                         break;
 154
 155                 instr_hi = opcode & 0xf0;
 156                 instr_lo = opcode & 0x0f;
 157                 instr++;
 158
 159                 switch (instr_hi) {
 160                 case 0x20:
 161                 case 0x30:
 162                         /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
 163                         scan_more = ((instr_lo & 7) == 0x6);
 164                         break;
 165
 166                 case 0x60:
 167                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 168                         scan_more = (instr_lo & 0xC) == 0x4;
 169                         break;
 170                 case 0xF0:
 171                         /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
 172                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 173                         break;
 174                 case 0x00:
 175                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 176                         scan_more = 0;
 177                         if (instr > (unsigned char *)limit)
 178                                 break;
 179                         if (probe_kernel_address(instr, opcode))
 180                                 break;
 181                         prefetch = (instr_lo == 0xF) &&
 182                                 (opcode == 0x0D || opcode == 0x18);
 183                         break;
 184                 default:
 185                         scan_more = 0;
 186                         break;
 187                 }
 188         }
 189         return prefetch;
 190 }
 191
 192 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
 193                               unsigned long error_code)
 194 {
 195         if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 196                      boot_cpu_data.x86 >= 6)) {
 197                 /* Catch an obscure case of prefetch inside an NX page. */
 198                 if (nx_enabled && (error_code & 16))
 199                         return 0;
 200                 return __is_prefetch(regs, addr);
 201         }
 202         return 0;
 203 }
 204
 205 static noinline void force_sig_info_fault(int si_signo, int si_code,
 206         unsigned long address, struct task_struct *tsk)
 207 {
 208         siginfo_t info;
 209
 210         info.si_signo = si_signo;
 211         info.si_errno = 0;
 212         info.si_code = si_code;
 213         info.si_addr = (void __user *)address;
 214         force_sig_info(si_signo, &info, tsk);
 215 }
 216
 217 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
 218
 219 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 220 {
 221         unsigned index = pgd_index(address);
 222         pgd_t *pgd_k;
 223         pud_t *pud, *pud_k;
 224         pmd_t *pmd, *pmd_k;
 225
 226         pgd += index;
 227         pgd_k = init_mm.pgd + index;
 228
 229         if (!pgd_present(*pgd_k))
 230                 return NULL;
 231
 232         /*
 233          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 234          * and redundant with the set_pmd() on non-PAE. As would
 235          * set_pud.
 236          */
 237
 238         pud = pud_offset(pgd, address);
 239         pud_k = pud_offset(pgd_k, address);
 240         if (!pud_present(*pud_k))
 241                 return NULL;
 242
 243         pmd = pmd_offset(pud, address);
 244         pmd_k = pmd_offset(pud_k, address);
 245         if (!pmd_present(*pmd_k))
 246                 return NULL;
 247         if (!pmd_present(*pmd)) {
 248                 set_pmd(pmd, *pmd_k);
 249                 arch_flush_lazy_mmu_mode();
 250         } else
 251                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 252         return pmd_k;
 253 }
 254
 255 /*
 256  * Handle a fault on the vmalloc or module mapping area
 257  *
 258  * This assumes no large pages in there.
 259  */
 260 static inline int vmalloc_fault(unsigned long address)
 261 {
 262         unsigned long pgd_paddr;
 263         pmd_t *pmd_k;
 264         pte_t *pte_k;
 265         /*
 266          * Synchronize this task's top level page-table
 267          * with the 'reference' page table.
 268          *
 269          * Do _not_ use "current" here. We might be inside
 270          * an interrupt in the middle of a task switch..
 271          */
 272         pgd_paddr = read_cr3();
 273         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 274         if (!pmd_k)
 275                 return -1;
 276         pte_k = pte_offset_kernel(pmd_k, address);
 277         if (!pte_present(*pte_k))
 278                 return -1;
 279         return 0;
 280 }
 281
 282 int show_unhandled_signals = 1;
 283
 284 /*
 285  * This routine handles page faults.  It determines the address,
 286  * and the problem, and then passes it off to one of the appropriate
 287  * routines.
 288  *
 289  * error_code:
 290  *      bit 0 == 0 means no page found, 1 means protection fault
 291  *      bit 1 == 0 means read, 1 means write
 292  *      bit 2 == 0 means kernel, 1 means user-mode
 293  *      bit 3 == 1 means use of reserved bit detected
 294  *      bit 4 == 1 means fault was an instruction fetch
 295  */
 296 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
 297                                       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         int write, si_code;
 304         int fault;
 305
 306         /*
 307          * We can fault from pretty much anywhere, with unknown IRQ state.
 308          */
 309         trace_hardirqs_fixup();
 310
 311         /* get the address */
 312         address = read_cr2();
 313
 314         tsk = current;
 315
 316         si_code = SEGV_MAPERR;
 317
 318         /*
 319          * We fault-in kernel-space virtual memory on-demand. The
 320          * 'reference' page table is init_mm.pgd.
 321          *
 322          * NOTE! We MUST NOT take any locks for this case. We may
 323          * be in an interrupt or a critical region, and should
 324          * only copy the information from the master page table,
 325          * nothing more.
 326          *
 327          * This verifies that the fault happens in kernel space
 328          * (error_code & 4) == 0, and that the fault was not a
 329          * protection error (error_code & 9) == 0.
 330          */
 331         if (unlikely(address >= TASK_SIZE)) {
 332                 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
 333                         return;
 334                 if (notify_page_fault(regs))
 335                         return;
 336                 /*
 337                  * Don't take the mm semaphore here. If we fixup a prefetch
 338                  * fault we could otherwise deadlock.
 339                  */
 340                 goto bad_area_nosemaphore;
 341         }
 342
 343         if (notify_page_fault(regs))
 344                 return;
 345
 346         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 347            fault has been handled. */
 348         if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
 349                 local_irq_enable();
 350
 351         mm = tsk->mm;
 352
 353         /*
 354          * If we're in an interrupt, have no user context or are running in an
 355          * atomic region then we must not take the fault..
 356          */
 357         if (in_atomic() || !mm)
 358                 goto bad_area_nosemaphore;
 359
 360         /* When running in the kernel we expect faults to occur only to
 361          * addresses in user space.  All other faults represent errors in the
 362          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 363          * erroneous fault occurring in a code path which already holds mmap_sem
 364          * we will deadlock attempting to validate the fault against the
 365          * address space.  Luckily the kernel only validly references user
 366          * space from well defined areas of code, which are listed in the
 367          * exceptions table.
 368          *
 369          * As the vast majority of faults will be valid we will only perform
 370          * the source reference check when there is a possibility of a deadlock.
 371          * Attempt to lock the address space, if we cannot we then validate the
 372          * source.  If this is invalid we can skip the address space check,
 373          * thus avoiding the deadlock.
 374          */
 375         if (!down_read_trylock(&mm->mmap_sem)) {
 376                 if ((error_code & 4) == 0 &&
 377                     !search_exception_tables(regs->eip))
 378                         goto bad_area_nosemaphore;
 379                 down_read(&mm->mmap_sem);
 380         }
 381
 382         vma = find_vma(mm, address);
 383         if (!vma)
 384                 goto bad_area;
 385         if (vma->vm_start <= address)
 386                 goto good_area;
 387         if (!(vma->vm_flags & VM_GROWSDOWN))
 388                 goto bad_area;
 389         if (error_code & 4) {
 390                 /*
 391                  * Accessing the stack below %esp is always a bug.
 392                  * The large cushion allows instructions like enter
 393                  * and pusha to work.  ("enter $65535,$31" pushes
 394                  * 32 pointers and then decrements %esp by 65535.)
 395                  */
 396                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->esp)
 397                         goto bad_area;
 398         }
 399         if (expand_stack(vma, address))
 400                 goto bad_area;
 401 /*
 402  * Ok, we have a good vm_area for this memory access, so
 403  * we can handle it..
 404  */
 405 good_area:
 406         si_code = SEGV_ACCERR;
 407         write = 0;
 408         switch (error_code & 3) {
 409                 default:        /* 3: write, present */
 410                                 /* fall through */
 411                 case 2:         /* write, not present */
 412                         if (!(vma->vm_flags & VM_WRITE))
 413                                 goto bad_area;
 414                         write++;
 415                         break;
 416                 case 1:         /* read, present */
 417                         goto bad_area;
 418                 case 0:         /* read, not present */
 419                         if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 420                                 goto bad_area;
 421         }
 422
 423  survive:
 424         /*
 425          * If for any reason at all we couldn't handle the fault,
 426          * make sure we exit gracefully rather than endlessly redo
 427          * the fault.
 428          */
 429         fault = handle_mm_fault(mm, vma, address, write);
 430         if (unlikely(fault & VM_FAULT_ERROR)) {
 431                 if (fault & VM_FAULT_OOM)
 432                         goto out_of_memory;
 433                 else if (fault & VM_FAULT_SIGBUS)
 434                         goto do_sigbus;
 435                 BUG();
 436         }
 437         if (fault & VM_FAULT_MAJOR)
 438                 tsk->maj_flt++;
 439         else
 440                 tsk->min_flt++;
 441
 442         /*
 443          * Did it hit the DOS screen memory VA from vm86 mode?
 444          */
 445         if (regs->eflags & VM_MASK) {
 446                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 447                 if (bit < 32)
 448                         tsk->thread.screen_bitmap |= 1 << bit;
 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         /* User mode accesses just cause a SIGSEGV */
 462         if (error_code & 4) {
 463                 /*
 464                  * It's possible to have interrupts off here.
 465                  */
 466                 local_irq_enable();
 467
 468                 /*
 469                  * Valid to do another page fault here because this one came
 470                  * from user space.
 471                  */
 472                 if (is_prefetch(regs, address, error_code))
 473                         return;
 474
 475                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 476                     printk_ratelimit()) {
 477                         printk("%s%s[%d]: segfault at %08lx eip %08lx "
 478                             "esp %08lx error %lx\n",
 479                             task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 480                             tsk->comm, task_pid_nr(tsk), address, regs->eip,
 481                             regs->esp, error_code);
 482                 }
 483                 tsk->thread.cr2 = address;
 484                 /* Kernel addresses are always protection faults */
 485                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 486                 tsk->thread.trap_no = 14;
 487                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 488                 return;
 489         }
 490
 491 #ifdef CONFIG_X86_F00F_BUG
 492         /*
 493          * Pentium F0 0F C7 C8 bug workaround.
 494          */
 495         if (boot_cpu_data.f00f_bug) {
 496                 unsigned long nr;
 497
 498                 nr = (address - idt_descr.address) >> 3;
 499
 500                 if (nr == 6) {
 501                         do_invalid_op(regs, 0);
 502                         return;
 503                 }
 504         }
 505 #endif
 506
 507 no_context:
 508         /* Are we prepared to handle this kernel fault?  */
 509         if (fixup_exception(regs))
 510                 return;
 511
 512         /*
 513          * Valid to do another page fault here, because if this fault
 514          * had been triggered by is_prefetch fixup_exception would have
 515          * handled it.
 516          */
 517         if (is_prefetch(regs, address, error_code))
 518                 return;
 519
 520 /*
 521  * Oops. The kernel tried to access some bad page. We'll have to
 522  * terminate things with extreme prejudice.
 523  */
 524
 525         bust_spinlocks(1);
 526
 527         if (oops_may_print()) {
 528                 __typeof__(pte_val(__pte(0))) page;
 529
 530 #ifdef CONFIG_X86_PAE
 531                 if (error_code & 16) {
 532                         pte_t *pte = lookup_address(address);
 533
 534                         if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 535                                 printk(KERN_CRIT "kernel tried to execute "
 536                                         "NX-protected page - exploit attempt? "
 537                                         "(uid: %d)\n", current->uid);
 538                 }
 539 #endif
 540                 if (address < PAGE_SIZE)
 541                         printk(KERN_ALERT "BUG: unable to handle kernel NULL "
 542                                         "pointer dereference");
 543                 else
 544                         printk(KERN_ALERT "BUG: unable to handle kernel paging"
 545                                         " request");
 546                 printk(" at virtual address %08lx\n",address);
 547                 printk(KERN_ALERT "printing eip: %08lx ", regs->eip);
 548
 549                 page = read_cr3();
 550                 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 551 #ifdef CONFIG_X86_PAE
 552                 printk("*pdpt = %016Lx ", page);
 553                 if ((page >> PAGE_SHIFT) < max_low_pfn
 554                     && page & _PAGE_PRESENT) {
 555                         page &= PAGE_MASK;
 556                         page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 557                                                                  & (PTRS_PER_PMD - 1)];
 558                         printk(KERN_CONT "*pde = %016Lx ", page);
 559                         page &= ~_PAGE_NX;
 560                 }
 561 #else
 562                 printk("*pde = %08lx ", page);
 563 #endif
 564
 565                 /*
 566                  * We must not directly access the pte in the highpte
 567                  * case if the page table is located in highmem.
 568                  * And let's rather not kmap-atomic the pte, just in case
 569                  * it's allocated already.
 570                  */
 571                 if ((page >> PAGE_SHIFT) < max_low_pfn
 572                     && (page & _PAGE_PRESENT)
 573                     && !(page & _PAGE_PSE)) {
 574                         page &= PAGE_MASK;
 575                         page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 576                                                                  & (PTRS_PER_PTE - 1)];
 577                         printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 578                 }
 579
 580                 printk("\n");
 581         }
 582
 583         tsk->thread.cr2 = address;
 584         tsk->thread.trap_no = 14;
 585         tsk->thread.error_code = error_code;
 586         die("Oops", regs, error_code);
 587         bust_spinlocks(0);
 588         do_exit(SIGKILL);
 589
 590 /*
 591  * We ran out of memory, or some other thing happened to us that made
 592  * us unable to handle the page fault gracefully.
 593  */
 594 out_of_memory:
 595         up_read(&mm->mmap_sem);
 596         if (is_global_init(tsk)) {
 597                 yield();
 598                 down_read(&mm->mmap_sem);
 599                 goto survive;
 600         }
 601         printk("VM: killing process %s\n", tsk->comm);
 602         if (error_code & 4)
 603                 do_group_exit(SIGKILL);
 604         goto no_context;
 605
 606 do_sigbus:
 607         up_read(&mm->mmap_sem);
 608
 609         /* Kernel mode? Handle exceptions or die */
 610         if (!(error_code & 4))
 611                 goto no_context;
 612
 613         /* User space => ok to do another page fault */
 614         if (is_prefetch(regs, address, error_code))
 615                 return;
 616
 617         tsk->thread.cr2 = address;
 618         tsk->thread.error_code = error_code;
 619         tsk->thread.trap_no = 14;
 620         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 621 }
 622
 623 void vmalloc_sync_all(void)
 624 {
 625         /*
 626          * Note that races in the updates of insync and start aren't
 627          * problematic: insync can only get set bits added, and updates to
 628          * start are only improving performance (without affecting correctness
 629          * if undone).
 630          */
 631         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 632         static unsigned long start = TASK_SIZE;
 633         unsigned long address;
 634
 635         if (SHARED_KERNEL_PMD)
 636                 return;
 637
 638         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 639         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 640                 if (!test_bit(pgd_index(address), insync)) {
 641                         unsigned long flags;
 642                         struct page *page;
 643
 644                         spin_lock_irqsave(&pgd_lock, flags);
 645                         for (page = pgd_list; page; page =
 646                                         (struct page *)page->index)
 647                                 if (!vmalloc_sync_one(page_address(page),
 648                                                                 address)) {
 649                                         BUG_ON(page != pgd_list);
 650                                         break;
 651                                 }
 652                         spin_unlock_irqrestore(&pgd_lock, flags);
 653                         if (!page)
 654                                 set_bit(pgd_index(address), insync);
 655                 }
 656                 if (address == start && test_bit(pgd_index(address), insync))
 657                         start = address + PGDIR_SIZE;
 658         }
 659 }