arch/i386/mm/fault.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/smp_lock.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/init.h>
  20 #include <linux/tty.h>
  21 #include <linux/vt_kern.h>              /* For unblank_screen() */
  22 #include <linux/highmem.h>
  23 #include <linux/module.h>
  24 #include <linux/kprobes.h>
  25 #include <linux/uaccess.h>
  26
  27 #include <asm/system.h>
  28 #include <asm/desc.h>
  29 #include <asm/kdebug.h>
  30 #include <asm/segment.h>
  31
  32 extern void die(const char *,struct pt_regs *,long);
  33
  34 static ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
  35
  36 int register_page_fault_notifier(struct notifier_block *nb)
  37 {
  38         vmalloc_sync_all();
  39         return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
  40 }
  41 EXPORT_SYMBOL_GPL(register_page_fault_notifier);
  42
  43 int unregister_page_fault_notifier(struct notifier_block *nb)
  44 {
  45         return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
  46 }
  47 EXPORT_SYMBOL_GPL(unregister_page_fault_notifier);
  48
  49 static inline int notify_page_fault(enum die_val val, const char *str,
  50                         struct pt_regs *regs, long err, int trap, int sig)
  51 {
  52         struct die_args args = {
  53                 .regs = regs,
  54                 .str = str,
  55                 .err = err,
  56                 .trapnr = trap,
  57                 .signr = sig
  58         };
  59         return atomic_notifier_call_chain(&notify_page_fault_chain, val, &args);
  60 }
  61
  62 /*
  63  * Unlock any spinlocks which will prevent us from getting the
  64  * message out
  65  */
  66 void bust_spinlocks(int yes)
  67 {
  68         int loglevel_save = console_loglevel;
  69
  70         if (yes) {
  71                 oops_in_progress = 1;
  72                 return;
  73         }
  74 #ifdef CONFIG_VT
  75         unblank_screen();
  76 #endif
  77         oops_in_progress = 0;
  78         /*
  79          * OK, the message is on the console.  Now we call printk()
  80          * without oops_in_progress set so that printk will give klogd
  81          * a poke.  Hold onto your hats...
  82          */
  83         console_loglevel = 15;          /* NMI oopser may have shut the console up */
  84         printk(" ");
  85         console_loglevel = loglevel_save;
  86 }
  87
  88 /*
  89  * Return EIP plus the CS segment base.  The segment limit is also
  90  * adjusted, clamped to the kernel/user address space (whichever is
  91  * appropriate), and returned in *eip_limit.
  92  *
  93  * The segment is checked, because it might have been changed by another
  94  * task between the original faulting instruction and here.
  95  *
  96  * If CS is no longer a valid code segment, or if EIP is beyond the
  97  * limit, or if it is a kernel address when CS is not a kernel segment,
  98  * then the returned value will be greater than *eip_limit.
  99  *
 100  * This is slow, but is very rarely executed.
 101  */
 102 static inline unsigned long get_segment_eip(struct pt_regs *regs,
 103                                             unsigned long *eip_limit)
 104 {
 105         unsigned long eip = regs->eip;
 106         unsigned seg = regs->xcs & 0xffff;
 107         u32 seg_ar, seg_limit, base, *desc;
 108
 109         /* Unlikely, but must come before segment checks. */
 110         if (unlikely(regs->eflags & VM_MASK)) {
 111                 base = seg << 4;
 112                 *eip_limit = base + 0xffff;
 113                 return base + (eip & 0xffff);
 114         }
 115
 116         /* The standard kernel/user address space limit. */
 117         *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
 118
 119         /* By far the most common cases. */
 120         if (likely(SEGMENT_IS_FLAT_CODE(seg)))
 121                 return eip;
 122
 123         /* Check the segment exists, is within the current LDT/GDT size,
 124            that kernel/user (ring 0..3) has the appropriate privilege,
 125            that it's a code segment, and get the limit. */
 126         __asm__ ("larl %3,%0; lsll %3,%1"
 127                  : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
 128         if ((~seg_ar & 0x9800) || eip > seg_limit) {
 129                 *eip_limit = 0;
 130                 return 1;        /* So that returned eip > *eip_limit. */
 131         }
 132
 133         /* Get the GDT/LDT descriptor base.
 134            When you look for races in this code remember that
 135            LDT and other horrors are only used in user space. */
 136         if (seg & (1<<2)) {
 137                 /* Must lock the LDT while reading it. */
 138                 down(&current->mm->context.sem);
 139                 desc = current->mm->context.ldt;
 140                 desc = (void *)desc + (seg & ~7);
 141         } else {
 142                 /* Must disable preemption while reading the GDT. */
 143                 desc = (u32 *)get_cpu_gdt_table(get_cpu());
 144                 desc = (void *)desc + (seg & ~7);
 145         }
 146
 147         /* Decode the code segment base from the descriptor */
 148         base = get_desc_base((unsigned long *)desc);
 149
 150         if (seg & (1<<2)) {
 151                 up(&current->mm->context.sem);
 152         } else
 153                 put_cpu();
 154
 155         /* Adjust EIP and segment limit, and clamp at the kernel limit.
 156            It's legitimate for segments to wrap at 0xffffffff. */
 157         seg_limit += base;
 158         if (seg_limit < *eip_limit && seg_limit >= base)
 159                 *eip_limit = seg_limit;
 160         return eip + base;
 161 }
 162
 163 /*
 164  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
 165  * Check that here and ignore it.
 166  */
 167 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
 168 {
 169         unsigned long limit;
 170         unsigned char *instr = (unsigned char *)get_segment_eip (regs, &limit);
 171         int scan_more = 1;
 172         int prefetch = 0;
 173         int i;
 174
 175         for (i = 0; scan_more && i < 15; i++) {
 176                 unsigned char opcode;
 177                 unsigned char instr_hi;
 178                 unsigned char instr_lo;
 179
 180                 if (instr > (unsigned char *)limit)
 181                         break;
 182                 if (probe_kernel_address(instr, opcode))
 183                         break;
 184
 185                 instr_hi = opcode & 0xf0;
 186                 instr_lo = opcode & 0x0f;
 187                 instr++;
 188
 189                 switch (instr_hi) {
 190                 case 0x20:
 191                 case 0x30:
 192                         /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
 193                         scan_more = ((instr_lo & 7) == 0x6);
 194                         break;
 195
 196                 case 0x60:
 197                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 198                         scan_more = (instr_lo & 0xC) == 0x4;
 199                         break;
 200                 case 0xF0:
 201                         /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
 202                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 203                         break;
 204                 case 0x00:
 205                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 206                         scan_more = 0;
 207                         if (instr > (unsigned char *)limit)
 208                                 break;
 209                         if (probe_kernel_address(instr, opcode))
 210                                 break;
 211                         prefetch = (instr_lo == 0xF) &&
 212                                 (opcode == 0x0D || opcode == 0x18);
 213                         break;
 214                 default:
 215                         scan_more = 0;
 216                         break;
 217                 }
 218         }
 219         return prefetch;
 220 }
 221
 222 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
 223                               unsigned long error_code)
 224 {
 225         if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 226                      boot_cpu_data.x86 >= 6)) {
 227                 /* Catch an obscure case of prefetch inside an NX page. */
 228                 if (nx_enabled && (error_code & 16))
 229                         return 0;
 230                 return __is_prefetch(regs, addr);
 231         }
 232         return 0;
 233 }
 234
 235 static noinline void force_sig_info_fault(int si_signo, int si_code,
 236         unsigned long address, struct task_struct *tsk)
 237 {
 238         siginfo_t info;
 239
 240         info.si_signo = si_signo;
 241         info.si_errno = 0;
 242         info.si_code = si_code;
 243         info.si_addr = (void __user *)address;
 244         force_sig_info(si_signo, &info, tsk);
 245 }
 246
 247 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
 248
 249 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 250 {
 251         unsigned index = pgd_index(address);
 252         pgd_t *pgd_k;
 253         pud_t *pud, *pud_k;
 254         pmd_t *pmd, *pmd_k;
 255
 256         pgd += index;
 257         pgd_k = init_mm.pgd + index;
 258
 259         if (!pgd_present(*pgd_k))
 260                 return NULL;
 261
 262         /*
 263          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 264          * and redundant with the set_pmd() on non-PAE. As would
 265          * set_pud.
 266          */
 267
 268         pud = pud_offset(pgd, address);
 269         pud_k = pud_offset(pgd_k, address);
 270         if (!pud_present(*pud_k))
 271                 return NULL;
 272
 273         pmd = pmd_offset(pud, address);
 274         pmd_k = pmd_offset(pud_k, address);
 275         if (!pmd_present(*pmd_k))
 276                 return NULL;
 277         if (!pmd_present(*pmd))
 278                 set_pmd(pmd, *pmd_k);
 279         else
 280                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 281         return pmd_k;
 282 }
 283
 284 /*
 285  * Handle a fault on the vmalloc or module mapping area
 286  *
 287  * This assumes no large pages in there.
 288  */
 289 static inline int vmalloc_fault(unsigned long address)
 290 {
 291         unsigned long pgd_paddr;
 292         pmd_t *pmd_k;
 293         pte_t *pte_k;
 294         /*
 295          * Synchronize this task's top level page-table
 296          * with the 'reference' page table.
 297          *
 298          * Do _not_ use "current" here. We might be inside
 299          * an interrupt in the middle of a task switch..
 300          */
 301         pgd_paddr = read_cr3();
 302         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 303         if (!pmd_k)
 304                 return -1;
 305         pte_k = pte_offset_kernel(pmd_k, address);
 306         if (!pte_present(*pte_k))
 307                 return -1;
 308         return 0;
 309 }
 310
 311 /*
 312  * This routine handles page faults.  It determines the address,
 313  * and the problem, and then passes it off to one of the appropriate
 314  * routines.
 315  *
 316  * error_code:
 317  *      bit 0 == 0 means no page found, 1 means protection fault
 318  *      bit 1 == 0 means read, 1 means write
 319  *      bit 2 == 0 means kernel, 1 means user-mode
 320  *      bit 3 == 1 means use of reserved bit detected
 321  *      bit 4 == 1 means fault was an instruction fetch
 322  */
 323 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
 324                                       unsigned long error_code)
 325 {
 326         struct task_struct *tsk;
 327         struct mm_struct *mm;
 328         struct vm_area_struct * vma;
 329         unsigned long address;
 330         unsigned long page;
 331         int write, si_code;
 332
 333         /* get the address */
 334         address = read_cr2();
 335
 336         tsk = current;
 337
 338         si_code = SEGV_MAPERR;
 339
 340         /*
 341          * We fault-in kernel-space virtual memory on-demand. The
 342          * 'reference' page table is init_mm.pgd.
 343          *
 344          * NOTE! We MUST NOT take any locks for this case. We may
 345          * be in an interrupt or a critical region, and should
 346          * only copy the information from the master page table,
 347          * nothing more.
 348          *
 349          * This verifies that the fault happens in kernel space
 350          * (error_code & 4) == 0, and that the fault was not a
 351          * protection error (error_code & 9) == 0.
 352          */
 353         if (unlikely(address >= TASK_SIZE)) {
 354                 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
 355                         return;
 356                 if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 357                                                 SIGSEGV) == NOTIFY_STOP)
 358                         return;
 359                 /*
 360                  * Don't take the mm semaphore here. If we fixup a prefetch
 361                  * fault we could otherwise deadlock.
 362                  */
 363                 goto bad_area_nosemaphore;
 364         }
 365
 366         if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 367                                         SIGSEGV) == NOTIFY_STOP)
 368                 return;
 369
 370         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 371            fault has been handled. */
 372         if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
 373                 local_irq_enable();
 374
 375         mm = tsk->mm;
 376
 377         /*
 378          * If we're in an interrupt, have no user context or are running in an
 379          * atomic region then we must not take the fault..
 380          */
 381         if (in_atomic() || !mm)
 382                 goto bad_area_nosemaphore;
 383
 384         /* When running in the kernel we expect faults to occur only to
 385          * addresses in user space.  All other faults represent errors in the
 386          * kernel and should generate an OOPS.  Unfortunatly, in the case of an
 387          * erroneous fault occurring in a code path which already holds mmap_sem
 388          * we will deadlock attempting to validate the fault against the
 389          * address space.  Luckily the kernel only validly references user
 390          * space from well defined areas of code, which are listed in the
 391          * exceptions table.
 392          *
 393          * As the vast majority of faults will be valid we will only perform
 394          * the source reference check when there is a possibilty of a deadlock.
 395          * Attempt to lock the address space, if we cannot we then validate the
 396          * source.  If this is invalid we can skip the address space check,
 397          * thus avoiding the deadlock.
 398          */
 399         if (!down_read_trylock(&mm->mmap_sem)) {
 400                 if ((error_code & 4) == 0 &&
 401                     !search_exception_tables(regs->eip))
 402                         goto bad_area_nosemaphore;
 403                 down_read(&mm->mmap_sem);
 404         }
 405
 406         vma = find_vma(mm, address);
 407         if (!vma)
 408                 goto bad_area;
 409         if (vma->vm_start <= address)
 410                 goto good_area;
 411         if (!(vma->vm_flags & VM_GROWSDOWN))
 412                 goto bad_area;
 413         if (error_code & 4) {
 414                 /*
 415                  * Accessing the stack below %esp is always a bug.
 416                  * The large cushion allows instructions like enter
 417                  * and pusha to work.  ("enter $65535,$31" pushes
 418                  * 32 pointers and then decrements %esp by 65535.)
 419                  */
 420                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->esp)
 421                         goto bad_area;
 422         }
 423         if (expand_stack(vma, address))
 424                 goto bad_area;
 425 /*
 426  * Ok, we have a good vm_area for this memory access, so
 427  * we can handle it..
 428  */
 429 good_area:
 430         si_code = SEGV_ACCERR;
 431         write = 0;
 432         switch (error_code & 3) {
 433                 default:        /* 3: write, present */
 434                                 /* fall through */
 435                 case 2:         /* write, not present */
 436                         if (!(vma->vm_flags & VM_WRITE))
 437                                 goto bad_area;
 438                         write++;
 439                         break;
 440                 case 1:         /* read, present */
 441                         goto bad_area;
 442                 case 0:         /* read, not present */
 443                         if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 444                                 goto bad_area;
 445         }
 446
 447  survive:
 448         /*
 449          * If for any reason at all we couldn't handle the fault,
 450          * make sure we exit gracefully rather than endlessly redo
 451          * the fault.
 452          */
 453         switch (handle_mm_fault(mm, vma, address, write)) {
 454                 case VM_FAULT_MINOR:
 455                         tsk->min_flt++;
 456                         break;
 457                 case VM_FAULT_MAJOR:
 458                         tsk->maj_flt++;
 459                         break;
 460                 case VM_FAULT_SIGBUS:
 461                         goto do_sigbus;
 462                 case VM_FAULT_OOM:
 463                         goto out_of_memory;
 464                 default:
 465                         BUG();
 466         }
 467
 468         /*
 469          * Did it hit the DOS screen memory VA from vm86 mode?
 470          */
 471         if (regs->eflags & VM_MASK) {
 472                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 473                 if (bit < 32)
 474                         tsk->thread.screen_bitmap |= 1 << bit;
 475         }
 476         up_read(&mm->mmap_sem);
 477         return;
 478
 479 /*
 480  * Something tried to access memory that isn't in our memory map..
 481  * Fix it, but check if it's kernel or user first..
 482  */
 483 bad_area:
 484         up_read(&mm->mmap_sem);
 485
 486 bad_area_nosemaphore:
 487         /* User mode accesses just cause a SIGSEGV */
 488         if (error_code & 4) {
 489                 /*
 490                  * Valid to do another page fault here because this one came
 491                  * from user space.
 492                  */
 493                 if (is_prefetch(regs, address, error_code))
 494                         return;
 495
 496                 tsk->thread.cr2 = address;
 497                 /* Kernel addresses are always protection faults */
 498                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 499                 tsk->thread.trap_no = 14;
 500                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 501                 return;
 502         }
 503
 504 #ifdef CONFIG_X86_F00F_BUG
 505         /*
 506          * Pentium F0 0F C7 C8 bug workaround.
 507          */
 508         if (boot_cpu_data.f00f_bug) {
 509                 unsigned long nr;
 510
 511                 nr = (address - idt_descr.address) >> 3;
 512
 513                 if (nr == 6) {
 514                         do_invalid_op(regs, 0);
 515                         return;
 516                 }
 517         }
 518 #endif
 519
 520 no_context:
 521         /* Are we prepared to handle this kernel fault?  */
 522         if (fixup_exception(regs))
 523                 return;
 524
 525         /*
 526          * Valid to do another page fault here, because if this fault
 527          * had been triggered by is_prefetch fixup_exception would have
 528          * handled it.
 529          */
 530         if (is_prefetch(regs, address, error_code))
 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         bust_spinlocks(1);
 539
 540         if (oops_may_print()) {
 541         #ifdef CONFIG_X86_PAE
 542                 if (error_code & 16) {
 543                         pte_t *pte = lookup_address(address);
 544
 545                         if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 546                                 printk(KERN_CRIT "kernel tried to execute "
 547                                         "NX-protected page - exploit attempt? "
 548                                         "(uid: %d)\n", current->uid);
 549                 }
 550         #endif
 551                 if (address < PAGE_SIZE)
 552                         printk(KERN_ALERT "BUG: unable to handle kernel NULL "
 553                                         "pointer dereference");
 554                 else
 555                         printk(KERN_ALERT "BUG: unable to handle kernel paging"
 556                                         " request");
 557                 printk(" at virtual address %08lx\n",address);
 558                 printk(KERN_ALERT " printing eip:\n");
 559                 printk("%08lx\n", regs->eip);
 560         }
 561         page = read_cr3();
 562         page = ((unsigned long *) __va(page))[address >> 22];
 563         if (oops_may_print())
 564                 printk(KERN_ALERT "*pde = %08lx\n", page);
 565         /*
 566          * We must not directly access the pte in the highpte
 567          * case, the page table might be allocated in highmem.
 568          * And lets rather not kmap-atomic the pte, just in case
 569          * it's allocated already.
 570          */
 571 #ifndef CONFIG_HIGHPTE
 572         if ((page & 1) && oops_may_print()) {
 573                 page &= PAGE_MASK;
 574                 address &= 0x003ff000;
 575                 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
 576                 printk(KERN_ALERT "*pte = %08lx\n", page);
 577         }
 578 #endif
 579         tsk->thread.cr2 = address;
 580         tsk->thread.trap_no = 14;
 581         tsk->thread.error_code = error_code;
 582         die("Oops", regs, error_code);
 583         bust_spinlocks(0);
 584         do_exit(SIGKILL);
 585
 586 /*
 587  * We ran out of memory, or some other thing happened to us that made
 588  * us unable to handle the page fault gracefully.
 589  */
 590 out_of_memory:
 591         up_read(&mm->mmap_sem);
 592         if (is_init(tsk)) {
 593                 yield();
 594                 down_read(&mm->mmap_sem);
 595                 goto survive;
 596         }
 597         printk("VM: killing process %s\n", tsk->comm);
 598         if (error_code & 4)
 599                 do_exit(SIGKILL);
 600         goto no_context;
 601
 602 do_sigbus:
 603         up_read(&mm->mmap_sem);
 604
 605         /* Kernel mode? Handle exceptions or die */
 606         if (!(error_code & 4))
 607                 goto no_context;
 608
 609         /* User space => ok to do another page fault */
 610         if (is_prefetch(regs, address, error_code))
 611                 return;
 612
 613         tsk->thread.cr2 = address;
 614         tsk->thread.error_code = error_code;
 615         tsk->thread.trap_no = 14;
 616         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 617 }
 618
 619 #ifndef CONFIG_X86_PAE
 620 void vmalloc_sync_all(void)
 621 {
 622         /*
 623          * Note that races in the updates of insync and start aren't
 624          * problematic: insync can only get set bits added, and updates to
 625          * start are only improving performance (without affecting correctness
 626          * if undone).
 627          */
 628         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 629         static unsigned long start = TASK_SIZE;
 630         unsigned long address;
 631
 632         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 633         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 634                 if (!test_bit(pgd_index(address), insync)) {
 635                         unsigned long flags;
 636                         struct page *page;
 637
 638                         spin_lock_irqsave(&pgd_lock, flags);
 639                         for (page = pgd_list; page; page =
 640                                         (struct page *)page->index)
 641                                 if (!vmalloc_sync_one(page_address(page),
 642                                                                 address)) {
 643                                         BUG_ON(page != pgd_list);
 644                                         break;
 645                                 }
 646                         spin_unlock_irqrestore(&pgd_lock, flags);
 647                         if (!page)
 648                                 set_bit(pgd_index(address), insync);
 649                 }
 650                 if (address == start && test_bit(pgd_index(address), insync))
 651                         start = address + PGDIR_SIZE;
 652         }
 653 }
 654 #endif