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
  26 #include <asm/system.h>
  27 #include <asm/uaccess.h>
  28 #include <asm/desc.h>
  29 #include <asm/kdebug.h>
  30
  31 extern void die(const char *,struct pt_regs *,long);
  32
  33 #ifdef CONFIG_KPROBES
  34 ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
  35 int register_page_fault_notifier(struct notifier_block *nb)
  36 {
  37         vmalloc_sync_all();
  38         return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
  39 }
  40
  41 int unregister_page_fault_notifier(struct notifier_block *nb)
  42 {
  43         return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
  44 }
  45
  46 static inline int notify_page_fault(enum die_val val, const char *str,
  47                         struct pt_regs *regs, long err, int trap, int sig)
  48 {
  49         struct die_args args = {
  50                 .regs = regs,
  51                 .str = str,
  52                 .err = err,
  53                 .trapnr = trap,
  54                 .signr = sig
  55         };
  56         return atomic_notifier_call_chain(&notify_page_fault_chain, val, &args);
  57 }
  58 #else
  59 static inline int notify_page_fault(enum die_val val, const char *str,
  60                         struct pt_regs *regs, long err, int trap, int sig)
  61 {
  62         return NOTIFY_DONE;
  63 }
  64 #endif
  65
  66
  67 /*
  68  * Unlock any spinlocks which will prevent us from getting the
  69  * message out
  70  */
  71 void bust_spinlocks(int yes)
  72 {
  73         int loglevel_save = console_loglevel;
  74
  75         if (yes) {
  76                 oops_in_progress = 1;
  77                 return;
  78         }
  79 #ifdef CONFIG_VT
  80         unblank_screen();
  81 #endif
  82         oops_in_progress = 0;
  83         /*
  84          * OK, the message is on the console.  Now we call printk()
  85          * without oops_in_progress set so that printk will give klogd
  86          * a poke.  Hold onto your hats...
  87          */
  88         console_loglevel = 15;          /* NMI oopser may have shut the console up */
  89         printk(" ");
  90         console_loglevel = loglevel_save;
  91 }
  92
  93 /*
  94  * Return EIP plus the CS segment base.  The segment limit is also
  95  * adjusted, clamped to the kernel/user address space (whichever is
  96  * appropriate), and returned in *eip_limit.
  97  *
  98  * The segment is checked, because it might have been changed by another
  99  * task between the original faulting instruction and here.
 100  *
 101  * If CS is no longer a valid code segment, or if EIP is beyond the
 102  * limit, or if it is a kernel address when CS is not a kernel segment,
 103  * then the returned value will be greater than *eip_limit.
 104  *
 105  * This is slow, but is very rarely executed.
 106  */
 107 static inline unsigned long get_segment_eip(struct pt_regs *regs,
 108                                             unsigned long *eip_limit)
 109 {
 110         unsigned long eip = regs->eip;
 111         unsigned seg = regs->xcs & 0xffff;
 112         u32 seg_ar, seg_limit, base, *desc;
 113
 114         /* Unlikely, but must come before segment checks. */
 115         if (unlikely(regs->eflags & VM_MASK)) {
 116                 base = seg << 4;
 117                 *eip_limit = base + 0xffff;
 118                 return base + (eip & 0xffff);
 119         }
 120
 121         /* The standard kernel/user address space limit. */
 122         *eip_limit = (seg & 3) ? USER_DS.seg : KERNEL_DS.seg;
 123
 124         /* By far the most common cases. */
 125         if (likely(seg == __USER_CS || seg == __KERNEL_CS))
 126                 return eip;
 127
 128         /* Check the segment exists, is within the current LDT/GDT size,
 129            that kernel/user (ring 0..3) has the appropriate privilege,
 130            that it's a code segment, and get the limit. */
 131         __asm__ ("larl %3,%0; lsll %3,%1"
 132                  : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
 133         if ((~seg_ar & 0x9800) || eip > seg_limit) {
 134                 *eip_limit = 0;
 135                 return 1;        /* So that returned eip > *eip_limit. */
 136         }
 137
 138         /* Get the GDT/LDT descriptor base.
 139            When you look for races in this code remember that
 140            LDT and other horrors are only used in user space. */
 141         if (seg & (1<<2)) {
 142                 /* Must lock the LDT while reading it. */
 143                 down(&current->mm->context.sem);
 144                 desc = current->mm->context.ldt;
 145                 desc = (void *)desc + (seg & ~7);
 146         } else {
 147                 /* Must disable preemption while reading the GDT. */
 148                 desc = (u32 *)get_cpu_gdt_table(get_cpu());
 149                 desc = (void *)desc + (seg & ~7);
 150         }
 151
 152         /* Decode the code segment base from the descriptor */
 153         base = get_desc_base((unsigned long *)desc);
 154
 155         if (seg & (1<<2)) {
 156                 up(&current->mm->context.sem);
 157         } else
 158                 put_cpu();
 159
 160         /* Adjust EIP and segment limit, and clamp at the kernel limit.
 161            It's legitimate for segments to wrap at 0xffffffff. */
 162         seg_limit += base;
 163         if (seg_limit < *eip_limit && seg_limit >= base)
 164                 *eip_limit = seg_limit;
 165         return eip + base;
 166 }
 167
 168 /*
 169  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
 170  * Check that here and ignore it.
 171  */
 172 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
 173 {
 174         unsigned long limit;
 175         unsigned long instr = get_segment_eip (regs, &limit);
 176         int scan_more = 1;
 177         int prefetch = 0;
 178         int i;
 179
 180         for (i = 0; scan_more && i < 15; i++) {
 181                 unsigned char opcode;
 182                 unsigned char instr_hi;
 183                 unsigned char instr_lo;
 184
 185                 if (instr > limit)
 186                         break;
 187                 if (__get_user(opcode, (unsigned char __user *) instr))
 188                         break;
 189
 190                 instr_hi = opcode & 0xf0;
 191                 instr_lo = opcode & 0x0f;
 192                 instr++;
 193
 194                 switch (instr_hi) {
 195                 case 0x20:
 196                 case 0x30:
 197                         /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
 198                         scan_more = ((instr_lo & 7) == 0x6);
 199                         break;
 200
 201                 case 0x60:
 202                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 203                         scan_more = (instr_lo & 0xC) == 0x4;
 204                         break;
 205                 case 0xF0:
 206                         /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
 207                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 208                         break;
 209                 case 0x00:
 210                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 211                         scan_more = 0;
 212                         if (instr > limit)
 213                                 break;
 214                         if (__get_user(opcode, (unsigned char __user *) instr))
 215                                 break;
 216                         prefetch = (instr_lo == 0xF) &&
 217                                 (opcode == 0x0D || opcode == 0x18);
 218                         break;
 219                 default:
 220                         scan_more = 0;
 221                         break;
 222                 }
 223         }
 224         return prefetch;
 225 }
 226
 227 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
 228                               unsigned long error_code)
 229 {
 230         if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 231                      boot_cpu_data.x86 >= 6)) {
 232                 /* Catch an obscure case of prefetch inside an NX page. */
 233                 if (nx_enabled && (error_code & 16))
 234                         return 0;
 235                 return __is_prefetch(regs, addr);
 236         }
 237         return 0;
 238 }
 239
 240 static noinline void force_sig_info_fault(int si_signo, int si_code,
 241         unsigned long address, struct task_struct *tsk)
 242 {
 243         siginfo_t info;
 244
 245         info.si_signo = si_signo;
 246         info.si_errno = 0;
 247         info.si_code = si_code;
 248         info.si_addr = (void __user *)address;
 249         force_sig_info(si_signo, &info, tsk);
 250 }
 251
 252 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
 253
 254 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 255 {
 256         unsigned index = pgd_index(address);
 257         pgd_t *pgd_k;
 258         pud_t *pud, *pud_k;
 259         pmd_t *pmd, *pmd_k;
 260
 261         pgd += index;
 262         pgd_k = init_mm.pgd + index;
 263
 264         if (!pgd_present(*pgd_k))
 265                 return NULL;
 266
 267         /*
 268          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 269          * and redundant with the set_pmd() on non-PAE. As would
 270          * set_pud.
 271          */
 272
 273         pud = pud_offset(pgd, address);
 274         pud_k = pud_offset(pgd_k, address);
 275         if (!pud_present(*pud_k))
 276                 return NULL;
 277
 278         pmd = pmd_offset(pud, address);
 279         pmd_k = pmd_offset(pud_k, address);
 280         if (!pmd_present(*pmd_k))
 281                 return NULL;
 282         if (!pmd_present(*pmd))
 283                 set_pmd(pmd, *pmd_k);
 284         else
 285                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 286         return pmd_k;
 287 }
 288
 289 /*
 290  * Handle a fault on the vmalloc or module mapping area
 291  *
 292  * This assumes no large pages in there.
 293  */
 294 static inline int vmalloc_fault(unsigned long address)
 295 {
 296         unsigned long pgd_paddr;
 297         pmd_t *pmd_k;
 298         pte_t *pte_k;
 299         /*
 300          * Synchronize this task's top level page-table
 301          * with the 'reference' page table.
 302          *
 303          * Do _not_ use "current" here. We might be inside
 304          * an interrupt in the middle of a task switch..
 305          */
 306         pgd_paddr = read_cr3();
 307         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 308         if (!pmd_k)
 309                 return -1;
 310         pte_k = pte_offset_kernel(pmd_k, address);
 311         if (!pte_present(*pte_k))
 312                 return -1;
 313         return 0;
 314 }
 315
 316 /*
 317  * This routine handles page faults.  It determines the address,
 318  * and the problem, and then passes it off to one of the appropriate
 319  * routines.
 320  *
 321  * error_code:
 322  *      bit 0 == 0 means no page found, 1 means protection fault
 323  *      bit 1 == 0 means read, 1 means write
 324  *      bit 2 == 0 means kernel, 1 means user-mode
 325  *      bit 3 == 1 means use of reserved bit detected
 326  *      bit 4 == 1 means fault was an instruction fetch
 327  */
 328 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
 329                                       unsigned long error_code)
 330 {
 331         struct task_struct *tsk;
 332         struct mm_struct *mm;
 333         struct vm_area_struct * vma;
 334         unsigned long address;
 335         unsigned long page;
 336         int write, si_code;
 337
 338         /* get the address */
 339         address = read_cr2();
 340
 341         tsk = current;
 342
 343         si_code = SEGV_MAPERR;
 344
 345         /*
 346          * We fault-in kernel-space virtual memory on-demand. The
 347          * 'reference' page table is init_mm.pgd.
 348          *
 349          * NOTE! We MUST NOT take any locks for this case. We may
 350          * be in an interrupt or a critical region, and should
 351          * only copy the information from the master page table,
 352          * nothing more.
 353          *
 354          * This verifies that the fault happens in kernel space
 355          * (error_code & 4) == 0, and that the fault was not a
 356          * protection error (error_code & 9) == 0.
 357          */
 358         if (unlikely(address >= TASK_SIZE)) {
 359                 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
 360                         return;
 361                 if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 362                                                 SIGSEGV) == NOTIFY_STOP)
 363                         return;
 364                 /*
 365                  * Don't take the mm semaphore here. If we fixup a prefetch
 366                  * fault we could otherwise deadlock.
 367                  */
 368                 goto bad_area_nosemaphore;
 369         }
 370
 371         if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 372                                         SIGSEGV) == NOTIFY_STOP)
 373                 return;
 374
 375         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 376            fault has been handled. */
 377         if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
 378                 local_irq_enable();
 379
 380         mm = tsk->mm;
 381
 382         /*
 383          * If we're in an interrupt, have no user context or are running in an
 384          * atomic region then we must not take the fault..
 385          */
 386         if (in_atomic() || !mm)
 387                 goto bad_area_nosemaphore;
 388
 389         /* When running in the kernel we expect faults to occur only to
 390          * addresses in user space.  All other faults represent errors in the
 391          * kernel and should generate an OOPS.  Unfortunatly, in the case of an
 392          * erroneous fault occuring in a code path which already holds mmap_sem
 393          * we will deadlock attempting to validate the fault against the
 394          * address space.  Luckily the kernel only validly references user
 395          * space from well defined areas of code, which are listed in the
 396          * exceptions table.
 397          *
 398          * As the vast majority of faults will be valid we will only perform
 399          * the source reference check when there is a possibilty of a deadlock.
 400          * Attempt to lock the address space, if we cannot we then validate the
 401          * source.  If this is invalid we can skip the address space check,
 402          * thus avoiding the deadlock.
 403          */
 404         if (!down_read_trylock(&mm->mmap_sem)) {
 405                 if ((error_code & 4) == 0 &&
 406                     !search_exception_tables(regs->eip))
 407                         goto bad_area_nosemaphore;
 408                 down_read(&mm->mmap_sem);
 409         }
 410
 411         vma = find_vma(mm, address);
 412         if (!vma)
 413                 goto bad_area;
 414         if (vma->vm_start <= address)
 415                 goto good_area;
 416         if (!(vma->vm_flags & VM_GROWSDOWN))
 417                 goto bad_area;
 418         if (error_code & 4) {
 419                 /*
 420                  * Accessing the stack below %esp is always a bug.
 421                  * The large cushion allows instructions like enter
 422                  * and pusha to work.  ("enter $65535,$31" pushes
 423                  * 32 pointers and then decrements %esp by 65535.)
 424                  */
 425                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->esp)
 426                         goto bad_area;
 427         }
 428         if (expand_stack(vma, address))
 429                 goto bad_area;
 430 /*
 431  * Ok, we have a good vm_area for this memory access, so
 432  * we can handle it..
 433  */
 434 good_area:
 435         si_code = SEGV_ACCERR;
 436         write = 0;
 437         switch (error_code & 3) {
 438                 default:        /* 3: write, present */
 439 #ifdef TEST_VERIFY_AREA
 440                         if (regs->cs == KERNEL_CS)
 441                                 printk("WP fault at %08lx\n", regs->eip);
 442 #endif
 443                         /* fall through */
 444                 case 2:         /* write, not present */
 445                         if (!(vma->vm_flags & VM_WRITE))
 446                                 goto bad_area;
 447                         write++;
 448                         break;
 449                 case 1:         /* read, present */
 450                         goto bad_area;
 451                 case 0:         /* read, not present */
 452                         if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
 453                                 goto bad_area;
 454         }
 455
 456  survive:
 457         /*
 458          * If for any reason at all we couldn't handle the fault,
 459          * make sure we exit gracefully rather than endlessly redo
 460          * the fault.
 461          */
 462         switch (handle_mm_fault(mm, vma, address, write)) {
 463                 case VM_FAULT_MINOR:
 464                         tsk->min_flt++;
 465                         break;
 466                 case VM_FAULT_MAJOR:
 467                         tsk->maj_flt++;
 468                         break;
 469                 case VM_FAULT_SIGBUS:
 470                         goto do_sigbus;
 471                 case VM_FAULT_OOM:
 472                         goto out_of_memory;
 473                 default:
 474                         BUG();
 475         }
 476
 477         /*
 478          * Did it hit the DOS screen memory VA from vm86 mode?
 479          */
 480         if (regs->eflags & VM_MASK) {
 481                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 482                 if (bit < 32)
 483                         tsk->thread.screen_bitmap |= 1 << bit;
 484         }
 485         up_read(&mm->mmap_sem);
 486         return;
 487
 488 /*
 489  * Something tried to access memory that isn't in our memory map..
 490  * Fix it, but check if it's kernel or user first..
 491  */
 492 bad_area:
 493         up_read(&mm->mmap_sem);
 494
 495 bad_area_nosemaphore:
 496         /* User mode accesses just cause a SIGSEGV */
 497         if (error_code & 4) {
 498                 /*
 499                  * Valid to do another page fault here because this one came
 500                  * from user space.
 501                  */
 502                 if (is_prefetch(regs, address, error_code))
 503                         return;
 504
 505                 tsk->thread.cr2 = address;
 506                 /* Kernel addresses are always protection faults */
 507                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 508                 tsk->thread.trap_no = 14;
 509                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 510                 return;
 511         }
 512
 513 #ifdef CONFIG_X86_F00F_BUG
 514         /*
 515          * Pentium F0 0F C7 C8 bug workaround.
 516          */
 517         if (boot_cpu_data.f00f_bug) {
 518                 unsigned long nr;
 519
 520                 nr = (address - idt_descr.address) >> 3;
 521
 522                 if (nr == 6) {
 523                         do_invalid_op(regs, 0);
 524                         return;
 525                 }
 526         }
 527 #endif
 528
 529 no_context:
 530         /* Are we prepared to handle this kernel fault?  */
 531         if (fixup_exception(regs))
 532                 return;
 533
 534         /*
 535          * Valid to do another page fault here, because if this fault
 536          * had been triggered by is_prefetch fixup_exception would have
 537          * handled it.
 538          */
 539         if (is_prefetch(regs, address, error_code))
 540                 return;
 541
 542 /*
 543  * Oops. The kernel tried to access some bad page. We'll have to
 544  * terminate things with extreme prejudice.
 545  */
 546
 547         bust_spinlocks(1);
 548
 549         if (oops_may_print()) {
 550         #ifdef CONFIG_X86_PAE
 551                 if (error_code & 16) {
 552                         pte_t *pte = lookup_address(address);
 553
 554                         if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 555                                 printk(KERN_CRIT "kernel tried to execute "
 556                                         "NX-protected page - exploit attempt? "
 557                                         "(uid: %d)\n", current->uid);
 558                 }
 559         #endif
 560                 if (address < PAGE_SIZE)
 561                         printk(KERN_ALERT "BUG: unable to handle kernel NULL "
 562                                         "pointer dereference");
 563                 else
 564                         printk(KERN_ALERT "BUG: unable to handle kernel paging"
 565                                         " request");
 566                 printk(" at virtual address %08lx\n",address);
 567                 printk(KERN_ALERT " printing eip:\n");
 568                 printk("%08lx\n", regs->eip);
 569         }
 570         page = read_cr3();
 571         page = ((unsigned long *) __va(page))[address >> 22];
 572         if (oops_may_print())
 573                 printk(KERN_ALERT "*pde = %08lx\n", page);
 574         /*
 575          * We must not directly access the pte in the highpte
 576          * case, the page table might be allocated in highmem.
 577          * And lets rather not kmap-atomic the pte, just in case
 578          * it's allocated already.
 579          */
 580 #ifndef CONFIG_HIGHPTE
 581         if ((page & 1) && oops_may_print()) {
 582                 page &= PAGE_MASK;
 583                 address &= 0x003ff000;
 584                 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
 585                 printk(KERN_ALERT "*pte = %08lx\n", page);
 586         }
 587 #endif
 588         tsk->thread.cr2 = address;
 589         tsk->thread.trap_no = 14;
 590         tsk->thread.error_code = error_code;
 591         die("Oops", regs, error_code);
 592         bust_spinlocks(0);
 593         do_exit(SIGKILL);
 594
 595 /*
 596  * We ran out of memory, or some other thing happened to us that made
 597  * us unable to handle the page fault gracefully.
 598  */
 599 out_of_memory:
 600         up_read(&mm->mmap_sem);
 601         if (tsk->pid == 1) {
 602                 yield();
 603                 down_read(&mm->mmap_sem);
 604                 goto survive;
 605         }
 606         printk("VM: killing process %s\n", tsk->comm);
 607         if (error_code & 4)
 608                 do_exit(SIGKILL);
 609         goto no_context;
 610
 611 do_sigbus:
 612         up_read(&mm->mmap_sem);
 613
 614         /* Kernel mode? Handle exceptions or die */
 615         if (!(error_code & 4))
 616                 goto no_context;
 617
 618         /* User space => ok to do another page fault */
 619         if (is_prefetch(regs, address, error_code))
 620                 return;
 621
 622         tsk->thread.cr2 = address;
 623         tsk->thread.error_code = error_code;
 624         tsk->thread.trap_no = 14;
 625         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 626 }
 627
 628 #ifndef CONFIG_X86_PAE
 629 void vmalloc_sync_all(void)
 630 {
 631         /*
 632          * Note that races in the updates of insync and start aren't
 633          * problematic: insync can only get set bits added, and updates to
 634          * start are only improving performance (without affecting correctness
 635          * if undone).
 636          */
 637         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 638         static unsigned long start = TASK_SIZE;
 639         unsigned long address;
 640
 641         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 642         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 643                 if (!test_bit(pgd_index(address), insync)) {
 644                         unsigned long flags;
 645                         struct page *page;
 646
 647                         spin_lock_irqsave(&pgd_lock, flags);
 648                         for (page = pgd_list; page; page =
 649                                         (struct page *)page->index)
 650                                 if (!vmalloc_sync_one(page_address(page),
 651                                                                 address)) {
 652                                         BUG_ON(page != pgd_list);
 653                                         break;
 654                                 }
 655                         spin_unlock_irqrestore(&pgd_lock, flags);
 656                         if (!page)
 657                                 set_bit(pgd_index(address), insync);
 658                 }
 659                 if (address == start && test_bit(pgd_index(address), insync))
 660                         start = address + PGDIR_SIZE;
 661         }
 662 }
 663 #endif