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/bootmem.h>              /* for max_low_pfn */
  24 #include <linux/module.h>
  25 #include <linux/kprobes.h>
  26 #include <linux/uaccess.h>
  27
  28 #include <asm/system.h>
  29 #include <asm/desc.h>
  30 #include <asm/kdebug.h>
  31 #include <asm/segment.h>
  32
  33 extern void die(const char *,struct pt_regs *,long);
  34
  35 static ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
  36
  37 int register_page_fault_notifier(struct notifier_block *nb)
  38 {
  39         vmalloc_sync_all();
  40         return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
  41 }
  42 EXPORT_SYMBOL_GPL(register_page_fault_notifier);
  43
  44 int unregister_page_fault_notifier(struct notifier_block *nb)
  45 {
  46         return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
  47 }
  48 EXPORT_SYMBOL_GPL(unregister_page_fault_notifier);
  49
  50 static inline int notify_page_fault(struct pt_regs *regs, long err)
  51 {
  52         struct die_args args = {
  53                 .regs = regs,
  54                 .str = "page fault",
  55                 .err = err,
  56                 .trapnr = 14,
  57                 .signr = SIGSEGV
  58         };
  59         return atomic_notifier_call_chain(&notify_page_fault_chain,
  60                                           DIE_PAGE_FAULT, &args);
  61 }
  62
  63 /*
  64  * Return EIP plus the CS segment base.  The segment limit is also
  65  * adjusted, clamped to the kernel/user address space (whichever is
  66  * appropriate), and returned in *eip_limit.
  67  *
  68  * The segment is checked, because it might have been changed by another
  69  * task between the original faulting instruction and here.
  70  *
  71  * If CS is no longer a valid code segment, or if EIP is beyond the
  72  * limit, or if it is a kernel address when CS is not a kernel segment,
  73  * then the returned value will be greater than *eip_limit.
  74  *
  75  * This is slow, but is very rarely executed.
  76  */
  77 static inline unsigned long get_segment_eip(struct pt_regs *regs,
  78                                             unsigned long *eip_limit)
  79 {
  80         unsigned long eip = regs->eip;
  81         unsigned seg = regs->xcs & 0xffff;
  82         u32 seg_ar, seg_limit, base, *desc;
  83
  84         /* Unlikely, but must come before segment checks. */
  85         if (unlikely(regs->eflags & VM_MASK)) {
  86                 base = seg << 4;
  87                 *eip_limit = base + 0xffff;
  88                 return base + (eip & 0xffff);
  89         }
  90
  91         /* The standard kernel/user address space limit. */
  92         *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
  93
  94         /* By far the most common cases. */
  95         if (likely(SEGMENT_IS_FLAT_CODE(seg)))
  96                 return eip;
  97
  98         /* Check the segment exists, is within the current LDT/GDT size,
  99            that kernel/user (ring 0..3) has the appropriate privilege,
 100            that it's a code segment, and get the limit. */
 101         __asm__ ("larl %3,%0; lsll %3,%1"
 102                  : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
 103         if ((~seg_ar & 0x9800) || eip > seg_limit) {
 104                 *eip_limit = 0;
 105                 return 1;        /* So that returned eip > *eip_limit. */
 106         }
 107
 108         /* Get the GDT/LDT descriptor base.
 109            When you look for races in this code remember that
 110            LDT and other horrors are only used in user space. */
 111         if (seg & (1<<2)) {
 112                 /* Must lock the LDT while reading it. */
 113                 down(&current->mm->context.sem);
 114                 desc = current->mm->context.ldt;
 115                 desc = (void *)desc + (seg & ~7);
 116         } else {
 117                 /* Must disable preemption while reading the GDT. */
 118                 desc = (u32 *)get_cpu_gdt_table(get_cpu());
 119                 desc = (void *)desc + (seg & ~7);
 120         }
 121
 122         /* Decode the code segment base from the descriptor */
 123         base = get_desc_base((unsigned long *)desc);
 124
 125         if (seg & (1<<2)) {
 126                 up(&current->mm->context.sem);
 127         } else
 128                 put_cpu();
 129
 130         /* Adjust EIP and segment limit, and clamp at the kernel limit.
 131            It's legitimate for segments to wrap at 0xffffffff. */
 132         seg_limit += base;
 133         if (seg_limit < *eip_limit && seg_limit >= base)
 134                 *eip_limit = seg_limit;
 135         return eip + base;
 136 }
 137
 138 /*
 139  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
 140  * Check that here and ignore it.
 141  */
 142 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
 143 {
 144         unsigned long limit;
 145         unsigned char *instr = (unsigned char *)get_segment_eip (regs, &limit);
 146         int scan_more = 1;
 147         int prefetch = 0;
 148         int i;
 149
 150         for (i = 0; scan_more && i < 15; i++) {
 151                 unsigned char opcode;
 152                 unsigned char instr_hi;
 153                 unsigned char instr_lo;
 154
 155                 if (instr > (unsigned char *)limit)
 156                         break;
 157                 if (probe_kernel_address(instr, opcode))
 158                         break;
 159
 160                 instr_hi = opcode & 0xf0;
 161                 instr_lo = opcode & 0x0f;
 162                 instr++;
 163
 164                 switch (instr_hi) {
 165                 case 0x20:
 166                 case 0x30:
 167                         /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
 168                         scan_more = ((instr_lo & 7) == 0x6);
 169                         break;
 170
 171                 case 0x60:
 172                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 173                         scan_more = (instr_lo & 0xC) == 0x4;
 174                         break;
 175                 case 0xF0:
 176                         /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
 177                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 178                         break;
 179                 case 0x00:
 180                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 181                         scan_more = 0;
 182                         if (instr > (unsigned char *)limit)
 183                                 break;
 184                         if (probe_kernel_address(instr, opcode))
 185                                 break;
 186                         prefetch = (instr_lo == 0xF) &&
 187                                 (opcode == 0x0D || opcode == 0x18);
 188                         break;
 189                 default:
 190                         scan_more = 0;
 191                         break;
 192                 }
 193         }
 194         return prefetch;
 195 }
 196
 197 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
 198                               unsigned long error_code)
 199 {
 200         if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 201                      boot_cpu_data.x86 >= 6)) {
 202                 /* Catch an obscure case of prefetch inside an NX page. */
 203                 if (nx_enabled && (error_code & 16))
 204                         return 0;
 205                 return __is_prefetch(regs, addr);
 206         }
 207         return 0;
 208 }
 209
 210 static noinline void force_sig_info_fault(int si_signo, int si_code,
 211         unsigned long address, struct task_struct *tsk)
 212 {
 213         siginfo_t info;
 214
 215         info.si_signo = si_signo;
 216         info.si_errno = 0;
 217         info.si_code = si_code;
 218         info.si_addr = (void __user *)address;
 219         force_sig_info(si_signo, &info, tsk);
 220 }
 221
 222 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
 223
 224 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 225 {
 226         unsigned index = pgd_index(address);
 227         pgd_t *pgd_k;
 228         pud_t *pud, *pud_k;
 229         pmd_t *pmd, *pmd_k;
 230
 231         pgd += index;
 232         pgd_k = init_mm.pgd + index;
 233
 234         if (!pgd_present(*pgd_k))
 235                 return NULL;
 236
 237         /*
 238          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 239          * and redundant with the set_pmd() on non-PAE. As would
 240          * set_pud.
 241          */
 242
 243         pud = pud_offset(pgd, address);
 244         pud_k = pud_offset(pgd_k, address);
 245         if (!pud_present(*pud_k))
 246                 return NULL;
 247
 248         pmd = pmd_offset(pud, address);
 249         pmd_k = pmd_offset(pud_k, address);
 250         if (!pmd_present(*pmd_k))
 251                 return NULL;
 252         if (!pmd_present(*pmd))
 253                 set_pmd(pmd, *pmd_k);
 254         else
 255                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 256         return pmd_k;
 257 }
 258
 259 /*
 260  * Handle a fault on the vmalloc or module mapping area
 261  *
 262  * This assumes no large pages in there.
 263  */
 264 static inline int vmalloc_fault(unsigned long address)
 265 {
 266         unsigned long pgd_paddr;
 267         pmd_t *pmd_k;
 268         pte_t *pte_k;
 269         /*
 270          * Synchronize this task's top level page-table
 271          * with the 'reference' page table.
 272          *
 273          * Do _not_ use "current" here. We might be inside
 274          * an interrupt in the middle of a task switch..
 275          */
 276         pgd_paddr = read_cr3();
 277         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 278         if (!pmd_k)
 279                 return -1;
 280         pte_k = pte_offset_kernel(pmd_k, address);
 281         if (!pte_present(*pte_k))
 282                 return -1;
 283         return 0;
 284 }
 285
 286 /*
 287  * This routine handles page faults.  It determines the address,
 288  * and the problem, and then passes it off to one of the appropriate
 289  * routines.
 290  *
 291  * error_code:
 292  *      bit 0 == 0 means no page found, 1 means protection fault
 293  *      bit 1 == 0 means read, 1 means write
 294  *      bit 2 == 0 means kernel, 1 means user-mode
 295  *      bit 3 == 1 means use of reserved bit detected
 296  *      bit 4 == 1 means fault was an instruction fetch
 297  */
 298 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
 299                                       unsigned long error_code)
 300 {
 301         struct task_struct *tsk;
 302         struct mm_struct *mm;
 303         struct vm_area_struct * vma;
 304         unsigned long address;
 305         int write, si_code;
 306
 307         /* get the address */
 308         address = read_cr2();
 309
 310         tsk = current;
 311
 312         si_code = SEGV_MAPERR;
 313
 314         /*
 315          * We fault-in kernel-space virtual memory on-demand. The
 316          * 'reference' page table is init_mm.pgd.
 317          *
 318          * NOTE! We MUST NOT take any locks for this case. We may
 319          * be in an interrupt or a critical region, and should
 320          * only copy the information from the master page table,
 321          * nothing more.
 322          *
 323          * This verifies that the fault happens in kernel space
 324          * (error_code & 4) == 0, and that the fault was not a
 325          * protection error (error_code & 9) == 0.
 326          */
 327         if (unlikely(address >= TASK_SIZE)) {
 328                 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
 329                         return;
 330                 if (notify_page_fault(regs, error_code) == NOTIFY_STOP)
 331                         return;
 332                 /*
 333                  * Don't take the mm semaphore here. If we fixup a prefetch
 334                  * fault we could otherwise deadlock.
 335                  */
 336                 goto bad_area_nosemaphore;
 337         }
 338
 339         if (notify_page_fault(regs, error_code) == NOTIFY_STOP)
 340                 return;
 341
 342         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 343            fault has been handled. */
 344         if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
 345                 local_irq_enable();
 346
 347         mm = tsk->mm;
 348
 349         /*
 350          * If we're in an interrupt, have no user context or are running in an
 351          * atomic region then we must not take the fault..
 352          */
 353         if (in_atomic() || !mm)
 354                 goto bad_area_nosemaphore;
 355
 356         /* When running in the kernel we expect faults to occur only to
 357          * addresses in user space.  All other faults represent errors in the
 358          * kernel and should generate an OOPS.  Unfortunatly, in the case of an
 359          * erroneous fault occurring in a code path which already holds mmap_sem
 360          * we will deadlock attempting to validate the fault against the
 361          * address space.  Luckily the kernel only validly references user
 362          * space from well defined areas of code, which are listed in the
 363          * exceptions table.
 364          *
 365          * As the vast majority of faults will be valid we will only perform
 366          * the source reference check when there is a possibilty of a deadlock.
 367          * Attempt to lock the address space, if we cannot we then validate the
 368          * source.  If this is invalid we can skip the address space check,
 369          * thus avoiding the deadlock.
 370          */
 371         if (!down_read_trylock(&mm->mmap_sem)) {
 372                 if ((error_code & 4) == 0 &&
 373                     !search_exception_tables(regs->eip))
 374                         goto bad_area_nosemaphore;
 375                 down_read(&mm->mmap_sem);
 376         }
 377
 378         vma = find_vma(mm, address);
 379         if (!vma)
 380                 goto bad_area;
 381         if (vma->vm_start <= address)
 382                 goto good_area;
 383         if (!(vma->vm_flags & VM_GROWSDOWN))
 384                 goto bad_area;
 385         if (error_code & 4) {
 386                 /*
 387                  * Accessing the stack below %esp is always a bug.
 388                  * The large cushion allows instructions like enter
 389                  * and pusha to work.  ("enter $65535,$31" pushes
 390                  * 32 pointers and then decrements %esp by 65535.)
 391                  */
 392                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->esp)
 393                         goto bad_area;
 394         }
 395         if (expand_stack(vma, address))
 396                 goto bad_area;
 397 /*
 398  * Ok, we have a good vm_area for this memory access, so
 399  * we can handle it..
 400  */
 401 good_area:
 402         si_code = SEGV_ACCERR;
 403         write = 0;
 404         switch (error_code & 3) {
 405                 default:        /* 3: write, present */
 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 | VM_WRITE)))
 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                 __typeof__(pte_val(__pte(0))) page;
 514
 515 #ifdef CONFIG_X86_PAE
 516                 if (error_code & 16) {
 517                         pte_t *pte = lookup_address(address);
 518
 519                         if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 520                                 printk(KERN_CRIT "kernel tried to execute "
 521                                         "NX-protected page - exploit attempt? "
 522                                         "(uid: %d)\n", current->uid);
 523                 }
 524 #endif
 525                 if (address < PAGE_SIZE)
 526                         printk(KERN_ALERT "BUG: unable to handle kernel NULL "
 527                                         "pointer dereference");
 528                 else
 529                         printk(KERN_ALERT "BUG: unable to handle kernel paging"
 530                                         " request");
 531                 printk(" at virtual address %08lx\n",address);
 532                 printk(KERN_ALERT " printing eip:\n");
 533                 printk("%08lx\n", regs->eip);
 534
 535                 page = read_cr3();
 536                 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 537 #ifdef CONFIG_X86_PAE
 538                 printk(KERN_ALERT "*pdpt = %016Lx\n", page);
 539                 if ((page >> PAGE_SHIFT) < max_low_pfn
 540                     && page & _PAGE_PRESENT) {
 541                         page &= PAGE_MASK;
 542                         page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 543                                                                  & (PTRS_PER_PMD - 1)];
 544                         printk(KERN_ALERT "*pde = %016Lx\n", page);
 545                         page &= ~_PAGE_NX;
 546                 }
 547 #else
 548                 printk(KERN_ALERT "*pde = %08lx\n", page);
 549 #endif
 550
 551                 /*
 552                  * We must not directly access the pte in the highpte
 553                  * case if the page table is located in highmem.
 554                  * And let's rather not kmap-atomic the pte, just in case
 555                  * it's allocated already.
 556                  */
 557                 if ((page >> PAGE_SHIFT) < max_low_pfn
 558                     && (page & _PAGE_PRESENT)) {
 559                         page &= PAGE_MASK;
 560                         page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 561                                                                  & (PTRS_PER_PTE - 1)];
 562                         printk(KERN_ALERT "*pte = %0*Lx\n", sizeof(page)*2, (u64)page);
 563                 }
 564         }
 565
 566         tsk->thread.cr2 = address;
 567         tsk->thread.trap_no = 14;
 568         tsk->thread.error_code = error_code;
 569         die("Oops", regs, error_code);
 570         bust_spinlocks(0);
 571         do_exit(SIGKILL);
 572
 573 /*
 574  * We ran out of memory, or some other thing happened to us that made
 575  * us unable to handle the page fault gracefully.
 576  */
 577 out_of_memory:
 578         up_read(&mm->mmap_sem);
 579         if (is_init(tsk)) {
 580                 yield();
 581                 down_read(&mm->mmap_sem);
 582                 goto survive;
 583         }
 584         printk("VM: killing process %s\n", tsk->comm);
 585         if (error_code & 4)
 586                 do_exit(SIGKILL);
 587         goto no_context;
 588
 589 do_sigbus:
 590         up_read(&mm->mmap_sem);
 591
 592         /* Kernel mode? Handle exceptions or die */
 593         if (!(error_code & 4))
 594                 goto no_context;
 595
 596         /* User space => ok to do another page fault */
 597         if (is_prefetch(regs, address, error_code))
 598                 return;
 599
 600         tsk->thread.cr2 = address;
 601         tsk->thread.error_code = error_code;
 602         tsk->thread.trap_no = 14;
 603         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 604 }
 605
 606 void vmalloc_sync_all(void)
 607 {
 608         /*
 609          * Note that races in the updates of insync and start aren't
 610          * problematic: insync can only get set bits added, and updates to
 611          * start are only improving performance (without affecting correctness
 612          * if undone).
 613          */
 614         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 615         static unsigned long start = TASK_SIZE;
 616         unsigned long address;
 617
 618         if (SHARED_KERNEL_PMD)
 619                 return;
 620
 621         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 622         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 623                 if (!test_bit(pgd_index(address), insync)) {
 624                         unsigned long flags;
 625                         struct page *page;
 626
 627                         spin_lock_irqsave(&pgd_lock, flags);
 628                         for (page = pgd_list; page; page =
 629                                         (struct page *)page->index)
 630                                 if (!vmalloc_sync_one(page_address(page),
 631                                                                 address)) {
 632                                         BUG_ON(page != pgd_list);
 633                                         break;
 634                                 }
 635                         spin_unlock_irqrestore(&pgd_lock, flags);
 636                         if (!page)
 637                                 set_bit(pgd_index(address), insync);
 638                 }
 639                 if (address == start && test_bit(pgd_index(address), insync))
 640                         start = address + PGDIR_SIZE;
 641         }
 642 }