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
  25 #include <asm/system.h>
  26 #include <asm/uaccess.h>
  27 #include <asm/desc.h>
  28 #include <asm/kdebug.h>
  29
  30 extern void die(const char *,struct pt_regs *,long);
  31
  32 /*
  33  * Unlock any spinlocks which will prevent us from getting the
  34  * message out
  35  */
  36 void bust_spinlocks(int yes)
  37 {
  38         int loglevel_save = console_loglevel;
  39
  40         if (yes) {
  41                 oops_in_progress = 1;
  42                 return;
  43         }
  44 #ifdef CONFIG_VT
  45         unblank_screen();
  46 #endif
  47         oops_in_progress = 0;
  48         /*
  49          * OK, the message is on the console.  Now we call printk()
  50          * without oops_in_progress set so that printk will give klogd
  51          * a poke.  Hold onto your hats...
  52          */
  53         console_loglevel = 15;          /* NMI oopser may have shut the console up */
  54         printk(" ");
  55         console_loglevel = loglevel_save;
  56 }
  57
  58 /*
  59  * Return EIP plus the CS segment base.  The segment limit is also
  60  * adjusted, clamped to the kernel/user address space (whichever is
  61  * appropriate), and returned in *eip_limit.
  62  *
  63  * The segment is checked, because it might have been changed by another
  64  * task between the original faulting instruction and here.
  65  *
  66  * If CS is no longer a valid code segment, or if EIP is beyond the
  67  * limit, or if it is a kernel address when CS is not a kernel segment,
  68  * then the returned value will be greater than *eip_limit.
  69  *
  70  * This is slow, but is very rarely executed.
  71  */
  72 static inline unsigned long get_segment_eip(struct pt_regs *regs,
  73                                             unsigned long *eip_limit)
  74 {
  75         unsigned long eip = regs->eip;
  76         unsigned seg = regs->xcs & 0xffff;
  77         u32 seg_ar, seg_limit, base, *desc;
  78
  79         /* The standard kernel/user address space limit. */
  80         *eip_limit = (seg & 3) ? USER_DS.seg : KERNEL_DS.seg;
  81
  82         /* Unlikely, but must come before segment checks. */
  83         if (unlikely((regs->eflags & VM_MASK) != 0))
  84                 return eip + (seg << 4);
  85
  86         /* By far the most common cases. */
  87         if (likely(seg == __USER_CS || seg == __KERNEL_CS))
  88                 return eip;
  89
  90         /* Check the segment exists, is within the current LDT/GDT size,
  91            that kernel/user (ring 0..3) has the appropriate privilege,
  92            that it's a code segment, and get the limit. */
  93         __asm__ ("larl %3,%0; lsll %3,%1"
  94                  : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
  95         if ((~seg_ar & 0x9800) || eip > seg_limit) {
  96                 *eip_limit = 0;
  97                 return 1;        /* So that returned eip > *eip_limit. */
  98         }
  99
 100         /* Get the GDT/LDT descriptor base.
 101            When you look for races in this code remember that
 102            LDT and other horrors are only used in user space. */
 103         if (seg & (1<<2)) {
 104                 /* Must lock the LDT while reading it. */
 105                 down(&current->mm->context.sem);
 106                 desc = current->mm->context.ldt;
 107                 desc = (void *)desc + (seg & ~7);
 108         } else {
 109                 /* Must disable preemption while reading the GDT. */
 110                 desc = (u32 *)&per_cpu(cpu_gdt_table, get_cpu());
 111                 desc = (void *)desc + (seg & ~7);
 112         }
 113
 114         /* Decode the code segment base from the descriptor */
 115         base = get_desc_base((unsigned long *)desc);
 116
 117         if (seg & (1<<2)) {
 118                 up(&current->mm->context.sem);
 119         } else
 120                 put_cpu();
 121
 122         /* Adjust EIP and segment limit, and clamp at the kernel limit.
 123            It's legitimate for segments to wrap at 0xffffffff. */
 124         seg_limit += base;
 125         if (seg_limit < *eip_limit && seg_limit >= base)
 126                 *eip_limit = seg_limit;
 127         return eip + base;
 128 }
 129
 130 /*
 131  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
 132  * Check that here and ignore it.
 133  */
 134 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
 135 {
 136         unsigned long limit;
 137         unsigned long instr = get_segment_eip (regs, &limit);
 138         int scan_more = 1;
 139         int prefetch = 0;
 140         int i;
 141
 142         for (i = 0; scan_more && i < 15; i++) {
 143                 unsigned char opcode;
 144                 unsigned char instr_hi;
 145                 unsigned char instr_lo;
 146
 147                 if (instr > limit)
 148                         break;
 149                 if (__get_user(opcode, (unsigned char *) instr))
 150                         break;
 151
 152                 instr_hi = opcode & 0xf0;
 153                 instr_lo = opcode & 0x0f;
 154                 instr++;
 155
 156                 switch (instr_hi) {
 157                 case 0x20:
 158                 case 0x30:
 159                         /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
 160                         scan_more = ((instr_lo & 7) == 0x6);
 161                         break;
 162
 163                 case 0x60:
 164                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 165                         scan_more = (instr_lo & 0xC) == 0x4;
 166                         break;
 167                 case 0xF0:
 168                         /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
 169                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 170                         break;
 171                 case 0x00:
 172                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 173                         scan_more = 0;
 174                         if (instr > limit)
 175                                 break;
 176                         if (__get_user(opcode, (unsigned char *) instr))
 177                                 break;
 178                         prefetch = (instr_lo == 0xF) &&
 179                                 (opcode == 0x0D || opcode == 0x18);
 180                         break;
 181                 default:
 182                         scan_more = 0;
 183                         break;
 184                 }
 185         }
 186         return prefetch;
 187 }
 188
 189 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
 190                               unsigned long error_code)
 191 {
 192         if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 193                      boot_cpu_data.x86 >= 6)) {
 194                 /* Catch an obscure case of prefetch inside an NX page. */
 195                 if (nx_enabled && (error_code & 16))
 196                         return 0;
 197                 return __is_prefetch(regs, addr);
 198         }
 199         return 0;
 200 }
 201
 202 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
 203
 204 /*
 205  * This routine handles page faults.  It determines the address,
 206  * and the problem, and then passes it off to one of the appropriate
 207  * routines.
 208  *
 209  * error_code:
 210  *      bit 0 == 0 means no page found, 1 means protection fault
 211  *      bit 1 == 0 means read, 1 means write
 212  *      bit 2 == 0 means kernel, 1 means user-mode
 213  */
 214 fastcall void do_page_fault(struct pt_regs *regs, unsigned long error_code)
 215 {
 216         struct task_struct *tsk;
 217         struct mm_struct *mm;
 218         struct vm_area_struct * vma;
 219         unsigned long address;
 220         unsigned long page;
 221         int write;
 222         siginfo_t info;
 223
 224         /* get the address */
 225         __asm__("movl %%cr2,%0":"=r" (address));
 226
 227         if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 228                                         SIGSEGV) == NOTIFY_STOP)
 229                 return;
 230         /* It's safe to allow irq's after cr2 has been saved */
 231         if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
 232                 local_irq_enable();
 233
 234         tsk = current;
 235
 236         info.si_code = SEGV_MAPERR;
 237
 238         /*
 239          * We fault-in kernel-space virtual memory on-demand. The
 240          * 'reference' page table is init_mm.pgd.
 241          *
 242          * NOTE! We MUST NOT take any locks for this case. We may
 243          * be in an interrupt or a critical region, and should
 244          * only copy the information from the master page table,
 245          * nothing more.
 246          *
 247          * This verifies that the fault happens in kernel space
 248          * (error_code & 4) == 0, and that the fault was not a
 249          * protection error (error_code & 1) == 0.
 250          */
 251         if (unlikely(address >= TASK_SIZE)) {
 252                 if (!(error_code & 5))
 253                         goto vmalloc_fault;
 254                 /*
 255                  * Don't take the mm semaphore here. If we fixup a prefetch
 256                  * fault we could otherwise deadlock.
 257                  */
 258                 goto bad_area_nosemaphore;
 259         }
 260
 261         mm = tsk->mm;
 262
 263         /*
 264          * If we're in an interrupt, have no user context or are running in an
 265          * atomic region then we must not take the fault..
 266          */
 267         if (in_atomic() || !mm)
 268                 goto bad_area_nosemaphore;
 269
 270         /* When running in the kernel we expect faults to occur only to
 271          * addresses in user space.  All other faults represent errors in the
 272          * kernel and should generate an OOPS.  Unfortunatly, in the case of an
 273          * erroneous fault occuring in a code path which already holds mmap_sem
 274          * we will deadlock attempting to validate the fault against the
 275          * address space.  Luckily the kernel only validly references user
 276          * space from well defined areas of code, which are listed in the
 277          * exceptions table.
 278          *
 279          * As the vast majority of faults will be valid we will only perform
 280          * the source reference check when there is a possibilty of a deadlock.
 281          * Attempt to lock the address space, if we cannot we then validate the
 282          * source.  If this is invalid we can skip the address space check,
 283          * thus avoiding the deadlock.
 284          */
 285         if (!down_read_trylock(&mm->mmap_sem)) {
 286                 if ((error_code & 4) == 0 &&
 287                     !search_exception_tables(regs->eip))
 288                         goto bad_area_nosemaphore;
 289                 down_read(&mm->mmap_sem);
 290         }
 291
 292         vma = find_vma(mm, address);
 293         if (!vma)
 294                 goto bad_area;
 295         if (vma->vm_start <= address)
 296                 goto good_area;
 297         if (!(vma->vm_flags & VM_GROWSDOWN))
 298                 goto bad_area;
 299         if (error_code & 4) {
 300                 /*
 301                  * accessing the stack below %esp is always a bug.
 302                  * The "+ 32" is there due to some instructions (like
 303                  * pusha) doing post-decrement on the stack and that
 304                  * doesn't show up until later..
 305                  */
 306                 if (address + 32 < regs->esp)
 307                         goto bad_area;
 308         }
 309         if (expand_stack(vma, address))
 310                 goto bad_area;
 311 /*
 312  * Ok, we have a good vm_area for this memory access, so
 313  * we can handle it..
 314  */
 315 good_area:
 316         info.si_code = SEGV_ACCERR;
 317         write = 0;
 318         switch (error_code & 3) {
 319                 default:        /* 3: write, present */
 320 #ifdef TEST_VERIFY_AREA
 321                         if (regs->cs == KERNEL_CS)
 322                                 printk("WP fault at %08lx\n", regs->eip);
 323 #endif
 324                         /* fall through */
 325                 case 2:         /* write, not present */
 326                         if (!(vma->vm_flags & VM_WRITE))
 327                                 goto bad_area;
 328                         write++;
 329                         break;
 330                 case 1:         /* read, present */
 331                         goto bad_area;
 332                 case 0:         /* read, not present */
 333                         if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
 334                                 goto bad_area;
 335         }
 336
 337  survive:
 338         /*
 339          * If for any reason at all we couldn't handle the fault,
 340          * make sure we exit gracefully rather than endlessly redo
 341          * the fault.
 342          */
 343         switch (handle_mm_fault(mm, vma, address, write)) {
 344                 case VM_FAULT_MINOR:
 345                         tsk->min_flt++;
 346                         break;
 347                 case VM_FAULT_MAJOR:
 348                         tsk->maj_flt++;
 349                         break;
 350                 case VM_FAULT_SIGBUS:
 351                         goto do_sigbus;
 352                 case VM_FAULT_OOM:
 353                         goto out_of_memory;
 354                 default:
 355                         BUG();
 356         }
 357
 358         /*
 359          * Did it hit the DOS screen memory VA from vm86 mode?
 360          */
 361         if (regs->eflags & VM_MASK) {
 362                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 363                 if (bit < 32)
 364                         tsk->thread.screen_bitmap |= 1 << bit;
 365         }
 366         up_read(&mm->mmap_sem);
 367         return;
 368
 369 /*
 370  * Something tried to access memory that isn't in our memory map..
 371  * Fix it, but check if it's kernel or user first..
 372  */
 373 bad_area:
 374         up_read(&mm->mmap_sem);
 375
 376 bad_area_nosemaphore:
 377         /* User mode accesses just cause a SIGSEGV */
 378         if (error_code & 4) {
 379                 /*
 380                  * Valid to do another page fault here because this one came
 381                  * from user space.
 382                  */
 383                 if (is_prefetch(regs, address, error_code))
 384                         return;
 385
 386                 tsk->thread.cr2 = address;
 387                 /* Kernel addresses are always protection faults */
 388                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 389                 tsk->thread.trap_no = 14;
 390                 info.si_signo = SIGSEGV;
 391                 info.si_errno = 0;
 392                 /* info.si_code has been set above */
 393                 info.si_addr = (void __user *)address;
 394                 force_sig_info(SIGSEGV, &info, tsk);
 395                 return;
 396         }
 397
 398 #ifdef CONFIG_X86_F00F_BUG
 399         /*
 400          * Pentium F0 0F C7 C8 bug workaround.
 401          */
 402         if (boot_cpu_data.f00f_bug) {
 403                 unsigned long nr;
 404
 405                 nr = (address - idt_descr.address) >> 3;
 406
 407                 if (nr == 6) {
 408                         do_invalid_op(regs, 0);
 409                         return;
 410                 }
 411         }
 412 #endif
 413
 414 no_context:
 415         /* Are we prepared to handle this kernel fault?  */
 416         if (fixup_exception(regs))
 417                 return;
 418
 419         /*
 420          * Valid to do another page fault here, because if this fault
 421          * had been triggered by is_prefetch fixup_exception would have
 422          * handled it.
 423          */
 424         if (is_prefetch(regs, address, error_code))
 425                 return;
 426
 427 /*
 428  * Oops. The kernel tried to access some bad page. We'll have to
 429  * terminate things with extreme prejudice.
 430  */
 431
 432         bust_spinlocks(1);
 433
 434 #ifdef CONFIG_X86_PAE
 435         if (error_code & 16) {
 436                 pte_t *pte = lookup_address(address);
 437
 438                 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 439                         printk(KERN_CRIT "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n", current->uid);
 440         }
 441 #endif
 442         if (address < PAGE_SIZE)
 443                 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
 444         else
 445                 printk(KERN_ALERT "Unable to handle kernel paging request");
 446         printk(" at virtual address %08lx\n",address);
 447         printk(KERN_ALERT " printing eip:\n");
 448         printk("%08lx\n", regs->eip);
 449         asm("movl %%cr3,%0":"=r" (page));
 450         page = ((unsigned long *) __va(page))[address >> 22];
 451         printk(KERN_ALERT "*pde = %08lx\n", page);
 452         /*
 453          * We must not directly access the pte in the highpte
 454          * case, the page table might be allocated in highmem.
 455          * And lets rather not kmap-atomic the pte, just in case
 456          * it's allocated already.
 457          */
 458 #ifndef CONFIG_HIGHPTE
 459         if (page & 1) {
 460                 page &= PAGE_MASK;
 461                 address &= 0x003ff000;
 462                 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
 463                 printk(KERN_ALERT "*pte = %08lx\n", page);
 464         }
 465 #endif
 466         die("Oops", regs, error_code);
 467         bust_spinlocks(0);
 468         do_exit(SIGKILL);
 469
 470 /*
 471  * We ran out of memory, or some other thing happened to us that made
 472  * us unable to handle the page fault gracefully.
 473  */
 474 out_of_memory:
 475         up_read(&mm->mmap_sem);
 476         if (tsk->pid == 1) {
 477                 yield();
 478                 down_read(&mm->mmap_sem);
 479                 goto survive;
 480         }
 481         printk("VM: killing process %s\n", tsk->comm);
 482         if (error_code & 4)
 483                 do_exit(SIGKILL);
 484         goto no_context;
 485
 486 do_sigbus:
 487         up_read(&mm->mmap_sem);
 488
 489         /* Kernel mode? Handle exceptions or die */
 490         if (!(error_code & 4))
 491                 goto no_context;
 492
 493         /* User space => ok to do another page fault */
 494         if (is_prefetch(regs, address, error_code))
 495                 return;
 496
 497         tsk->thread.cr2 = address;
 498         tsk->thread.error_code = error_code;
 499         tsk->thread.trap_no = 14;
 500         info.si_signo = SIGBUS;
 501         info.si_errno = 0;
 502         info.si_code = BUS_ADRERR;
 503         info.si_addr = (void __user *)address;
 504         force_sig_info(SIGBUS, &info, tsk);
 505         return;
 506
 507 vmalloc_fault:
 508         {
 509                 /*
 510                  * Synchronize this task's top level page-table
 511                  * with the 'reference' page table.
 512                  *
 513                  * Do _not_ use "tsk" here. We might be inside
 514                  * an interrupt in the middle of a task switch..
 515                  */
 516                 int index = pgd_index(address);
 517                 unsigned long pgd_paddr;
 518                 pgd_t *pgd, *pgd_k;
 519                 pud_t *pud, *pud_k;
 520                 pmd_t *pmd, *pmd_k;
 521                 pte_t *pte_k;
 522
 523                 asm("movl %%cr3,%0":"=r" (pgd_paddr));
 524                 pgd = index + (pgd_t *)__va(pgd_paddr);
 525                 pgd_k = init_mm.pgd + index;
 526
 527                 if (!pgd_present(*pgd_k))
 528                         goto no_context;
 529
 530                 /*
 531                  * set_pgd(pgd, *pgd_k); here would be useless on PAE
 532                  * and redundant with the set_pmd() on non-PAE. As would
 533                  * set_pud.
 534                  */
 535
 536                 pud = pud_offset(pgd, address);
 537                 pud_k = pud_offset(pgd_k, address);
 538                 if (!pud_present(*pud_k))
 539                         goto no_context;
 540
 541                 pmd = pmd_offset(pud, address);
 542                 pmd_k = pmd_offset(pud_k, address);
 543                 if (!pmd_present(*pmd_k))
 544                         goto no_context;
 545                 set_pmd(pmd, *pmd_k);
 546
 547                 pte_k = pte_offset_kernel(pmd_k, address);
 548                 if (!pte_present(*pte_k))
 549                         goto no_context;
 550                 return;
 551         }
 552 }