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 /*
 218  * This routine handles page faults.  It determines the address,
 219  * and the problem, and then passes it off to one of the appropriate
 220  * routines.
 221  *
 222  * error_code:
 223  *      bit 0 == 0 means no page found, 1 means protection fault
 224  *      bit 1 == 0 means read, 1 means write
 225  *      bit 2 == 0 means kernel, 1 means user-mode
 226  */
 227 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
 228                                       unsigned long error_code)
 229 {
 230         struct task_struct *tsk;
 231         struct mm_struct *mm;
 232         struct vm_area_struct * vma;
 233         unsigned long address;
 234         unsigned long page;
 235         int write, si_code;
 236
 237         /* get the address */
 238         address = read_cr2();
 239
 240         if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
 241                                         SIGSEGV) == NOTIFY_STOP)
 242                 return;
 243         /* It's safe to allow irq's after cr2 has been saved */
 244         if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
 245                 local_irq_enable();
 246
 247         tsk = current;
 248
 249         si_code = SEGV_MAPERR;
 250
 251         /*
 252          * We fault-in kernel-space virtual memory on-demand. The
 253          * 'reference' page table is init_mm.pgd.
 254          *
 255          * NOTE! We MUST NOT take any locks for this case. We may
 256          * be in an interrupt or a critical region, and should
 257          * only copy the information from the master page table,
 258          * nothing more.
 259          *
 260          * This verifies that the fault happens in kernel space
 261          * (error_code & 4) == 0, and that the fault was not a
 262          * protection error (error_code & 1) == 0.
 263          */
 264         if (unlikely(address >= TASK_SIZE)) {
 265                 if (!(error_code & 5))
 266                         goto vmalloc_fault;
 267                 /*
 268                  * Don't take the mm semaphore here. If we fixup a prefetch
 269                  * fault we could otherwise deadlock.
 270                  */
 271                 goto bad_area_nosemaphore;
 272         }
 273
 274         mm = tsk->mm;
 275
 276         /*
 277          * If we're in an interrupt, have no user context or are running in an
 278          * atomic region then we must not take the fault..
 279          */
 280         if (in_atomic() || !mm)
 281                 goto bad_area_nosemaphore;
 282
 283         /* When running in the kernel we expect faults to occur only to
 284          * addresses in user space.  All other faults represent errors in the
 285          * kernel and should generate an OOPS.  Unfortunatly, in the case of an
 286          * erroneous fault occuring in a code path which already holds mmap_sem
 287          * we will deadlock attempting to validate the fault against the
 288          * address space.  Luckily the kernel only validly references user
 289          * space from well defined areas of code, which are listed in the
 290          * exceptions table.
 291          *
 292          * As the vast majority of faults will be valid we will only perform
 293          * the source reference check when there is a possibilty of a deadlock.
 294          * Attempt to lock the address space, if we cannot we then validate the
 295          * source.  If this is invalid we can skip the address space check,
 296          * thus avoiding the deadlock.
 297          */
 298         if (!down_read_trylock(&mm->mmap_sem)) {
 299                 if ((error_code & 4) == 0 &&
 300                     !search_exception_tables(regs->eip))
 301                         goto bad_area_nosemaphore;
 302                 down_read(&mm->mmap_sem);
 303         }
 304
 305         vma = find_vma(mm, address);
 306         if (!vma)
 307                 goto bad_area;
 308         if (vma->vm_start <= address)
 309                 goto good_area;
 310         if (!(vma->vm_flags & VM_GROWSDOWN))
 311                 goto bad_area;
 312         if (error_code & 4) {
 313                 /*
 314                  * accessing the stack below %esp is always a bug.
 315                  * The "+ 32" is there due to some instructions (like
 316                  * pusha) doing post-decrement on the stack and that
 317                  * doesn't show up until later..
 318                  */
 319                 if (address + 32 < regs->esp)
 320                         goto bad_area;
 321         }
 322         if (expand_stack(vma, address))
 323                 goto bad_area;
 324 /*
 325  * Ok, we have a good vm_area for this memory access, so
 326  * we can handle it..
 327  */
 328 good_area:
 329         si_code = SEGV_ACCERR;
 330         write = 0;
 331         switch (error_code & 3) {
 332                 default:        /* 3: write, present */
 333 #ifdef TEST_VERIFY_AREA
 334                         if (regs->cs == KERNEL_CS)
 335                                 printk("WP fault at %08lx\n", regs->eip);
 336 #endif
 337                         /* fall through */
 338                 case 2:         /* write, not present */
 339                         if (!(vma->vm_flags & VM_WRITE))
 340                                 goto bad_area;
 341                         write++;
 342                         break;
 343                 case 1:         /* read, present */
 344                         goto bad_area;
 345                 case 0:         /* read, not present */
 346                         if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
 347                                 goto bad_area;
 348         }
 349
 350  survive:
 351         /*
 352          * If for any reason at all we couldn't handle the fault,
 353          * make sure we exit gracefully rather than endlessly redo
 354          * the fault.
 355          */
 356         switch (handle_mm_fault(mm, vma, address, write)) {
 357                 case VM_FAULT_MINOR:
 358                         tsk->min_flt++;
 359                         break;
 360                 case VM_FAULT_MAJOR:
 361                         tsk->maj_flt++;
 362                         break;
 363                 case VM_FAULT_SIGBUS:
 364                         goto do_sigbus;
 365                 case VM_FAULT_OOM:
 366                         goto out_of_memory;
 367                 default:
 368                         BUG();
 369         }
 370
 371         /*
 372          * Did it hit the DOS screen memory VA from vm86 mode?
 373          */
 374         if (regs->eflags & VM_MASK) {
 375                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 376                 if (bit < 32)
 377                         tsk->thread.screen_bitmap |= 1 << bit;
 378         }
 379         up_read(&mm->mmap_sem);
 380         return;
 381
 382 /*
 383  * Something tried to access memory that isn't in our memory map..
 384  * Fix it, but check if it's kernel or user first..
 385  */
 386 bad_area:
 387         up_read(&mm->mmap_sem);
 388
 389 bad_area_nosemaphore:
 390         /* User mode accesses just cause a SIGSEGV */
 391         if (error_code & 4) {
 392                 /*
 393                  * Valid to do another page fault here because this one came
 394                  * from user space.
 395                  */
 396                 if (is_prefetch(regs, address, error_code))
 397                         return;
 398
 399                 tsk->thread.cr2 = address;
 400                 /* Kernel addresses are always protection faults */
 401                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 402                 tsk->thread.trap_no = 14;
 403                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 404                 return;
 405         }
 406
 407 #ifdef CONFIG_X86_F00F_BUG
 408         /*
 409          * Pentium F0 0F C7 C8 bug workaround.
 410          */
 411         if (boot_cpu_data.f00f_bug) {
 412                 unsigned long nr;
 413
 414                 nr = (address - idt_descr.address) >> 3;
 415
 416                 if (nr == 6) {
 417                         do_invalid_op(regs, 0);
 418                         return;
 419                 }
 420         }
 421 #endif
 422
 423 no_context:
 424         /* Are we prepared to handle this kernel fault?  */
 425         if (fixup_exception(regs))
 426                 return;
 427
 428         /*
 429          * Valid to do another page fault here, because if this fault
 430          * had been triggered by is_prefetch fixup_exception would have
 431          * handled it.
 432          */
 433         if (is_prefetch(regs, address, error_code))
 434                 return;
 435
 436 /*
 437  * Oops. The kernel tried to access some bad page. We'll have to
 438  * terminate things with extreme prejudice.
 439  */
 440
 441         bust_spinlocks(1);
 442
 443 #ifdef CONFIG_X86_PAE
 444         if (error_code & 16) {
 445                 pte_t *pte = lookup_address(address);
 446
 447                 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
 448                         printk(KERN_CRIT "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n", current->uid);
 449         }
 450 #endif
 451         if (address < PAGE_SIZE)
 452                 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
 453         else
 454                 printk(KERN_ALERT "Unable to handle kernel paging request");
 455         printk(" at virtual address %08lx\n",address);
 456         printk(KERN_ALERT " printing eip:\n");
 457         printk("%08lx\n", regs->eip);
 458         page = read_cr3();
 459         page = ((unsigned long *) __va(page))[address >> 22];
 460         printk(KERN_ALERT "*pde = %08lx\n", page);
 461         /*
 462          * We must not directly access the pte in the highpte
 463          * case, the page table might be allocated in highmem.
 464          * And lets rather not kmap-atomic the pte, just in case
 465          * it's allocated already.
 466          */
 467 #ifndef CONFIG_HIGHPTE
 468         if (page & 1) {
 469                 page &= PAGE_MASK;
 470                 address &= 0x003ff000;
 471                 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
 472                 printk(KERN_ALERT "*pte = %08lx\n", page);
 473         }
 474 #endif
 475         tsk->thread.cr2 = address;
 476         tsk->thread.trap_no = 14;
 477         tsk->thread.error_code = error_code;
 478         die("Oops", regs, error_code);
 479         bust_spinlocks(0);
 480         do_exit(SIGKILL);
 481
 482 /*
 483  * We ran out of memory, or some other thing happened to us that made
 484  * us unable to handle the page fault gracefully.
 485  */
 486 out_of_memory:
 487         up_read(&mm->mmap_sem);
 488         if (tsk->pid == 1) {
 489                 yield();
 490                 down_read(&mm->mmap_sem);
 491                 goto survive;
 492         }
 493         printk("VM: killing process %s\n", tsk->comm);
 494         if (error_code & 4)
 495                 do_exit(SIGKILL);
 496         goto no_context;
 497
 498 do_sigbus:
 499         up_read(&mm->mmap_sem);
 500
 501         /* Kernel mode? Handle exceptions or die */
 502         if (!(error_code & 4))
 503                 goto no_context;
 504
 505         /* User space => ok to do another page fault */
 506         if (is_prefetch(regs, address, error_code))
 507                 return;
 508
 509         tsk->thread.cr2 = address;
 510         tsk->thread.error_code = error_code;
 511         tsk->thread.trap_no = 14;
 512         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 513         return;
 514
 515 vmalloc_fault:
 516         {
 517                 /*
 518                  * Synchronize this task's top level page-table
 519                  * with the 'reference' page table.
 520                  *
 521                  * Do _not_ use "tsk" here. We might be inside
 522                  * an interrupt in the middle of a task switch..
 523                  */
 524                 int index = pgd_index(address);
 525                 unsigned long pgd_paddr;
 526                 pgd_t *pgd, *pgd_k;
 527                 pud_t *pud, *pud_k;
 528                 pmd_t *pmd, *pmd_k;
 529                 pte_t *pte_k;
 530
 531                 pgd_paddr = read_cr3();
 532                 pgd = index + (pgd_t *)__va(pgd_paddr);
 533                 pgd_k = init_mm.pgd + index;
 534
 535                 if (!pgd_present(*pgd_k))
 536                         goto no_context;
 537
 538                 /*
 539                  * set_pgd(pgd, *pgd_k); here would be useless on PAE
 540                  * and redundant with the set_pmd() on non-PAE. As would
 541                  * set_pud.
 542                  */
 543
 544                 pud = pud_offset(pgd, address);
 545                 pud_k = pud_offset(pgd_k, address);
 546                 if (!pud_present(*pud_k))
 547                         goto no_context;
 548
 549                 pmd = pmd_offset(pud, address);
 550                 pmd_k = pmd_offset(pud_k, address);
 551                 if (!pmd_present(*pmd_k))
 552                         goto no_context;
 553                 set_pmd(pmd, *pmd_k);
 554
 555                 pte_k = pte_offset_kernel(pmd_k, address);
 556                 if (!pte_present(*pte_k))
 557                         goto no_context;
 558                 return;
 559         }
 560 }