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