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