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/vmalloc.h>
  25 #include <linux/module.h>
  26 #include <linux/kprobes.h>
  27 #include <linux/uaccess.h>
  28 #include <linux/kdebug.h>
  29
  30 #include <asm/system.h>
  31 #include <asm/desc.h>
  32 #include <asm/segment.h>
  33
  34 extern void die(const char *,struct pt_regs *,long);
  35
  36 static ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
  37
  38 int register_page_fault_notifier(struct notifier_block *nb)
  39 {
  40         vmalloc_sync_all();
  41         return atomic_notifier_chain_register(&notify_page_fault_chain, nb);
  42 }
  43 EXPORT_SYMBOL_GPL(register_page_fault_notifier);
  44
  45 int unregister_page_fault_notifier(struct notifier_block *nb)
  46 {
  47         return atomic_notifier_chain_unregister(&notify_page_fault_chain, nb);
  48 }
  49 EXPORT_SYMBOL_GPL(unregister_page_fault_notifier);
  50
  51 static inline int notify_page_fault(struct pt_regs *regs, long err)
  52 {
  53         struct die_args args = {
  54                 .regs = regs,
  55                 .str = "page fault",
  56                 .err = err,
  57                 .trapnr = 14,
  58                 .signr = SIGSEGV
  59         };
  60         return atomic_notifier_call_chain(&notify_page_fault_chain,
  61                                           DIE_PAGE_FAULT, &args);
  62 }
  63
  64 /*
  65  * Return EIP plus the CS segment base.  The segment limit is also
  66  * adjusted, clamped to the kernel/user address space (whichever is
  67  * appropriate), and returned in *eip_limit.
  68  *
  69  * The segment is checked, because it might have been changed by another
  70  * task between the original faulting instruction and here.
  71  *
  72  * If CS is no longer a valid code segment, or if EIP is beyond the
  73  * limit, or if it is a kernel address when CS is not a kernel segment,
  74  * then the returned value will be greater than *eip_limit.
  75  *
  76  * This is slow, but is very rarely executed.
  77  */
  78 static inline unsigned long get_segment_eip(struct pt_regs *regs,
  79                                             unsigned long *eip_limit)
  80 {
  81         unsigned long eip = regs->eip;
  82         unsigned seg = regs->xcs & 0xffff;
  83         u32 seg_ar, seg_limit, base, *desc;
  84
  85         /* Unlikely, but must come before segment checks. */
  86         if (unlikely(regs->eflags & VM_MASK)) {
  87                 base = seg << 4;
  88                 *eip_limit = base + 0xffff;
  89                 return base + (eip & 0xffff);
  90         }
  91
  92         /* The standard kernel/user address space limit. */
  93         *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
  94
  95         /* By far the most common cases. */
  96         if (likely(SEGMENT_IS_FLAT_CODE(seg)))
  97                 return eip;
  98
  99         /* Check the segment exists, is within the current LDT/GDT size,
 100            that kernel/user (ring 0..3) has the appropriate privilege,
 101            that it's a code segment, and get the limit. */
 102         __asm__ ("larl %3,%0; lsll %3,%1"
 103                  : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
 104         if ((~seg_ar & 0x9800) || eip > seg_limit) {
 105                 *eip_limit = 0;
 106                 return 1;        /* So that returned eip > *eip_limit. */
 107         }
 108
 109         /* Get the GDT/LDT descriptor base.
 110            When you look for races in this code remember that
 111            LDT and other horrors are only used in user space. */
 112         if (seg & (1<<2)) {
 113                 /* Must lock the LDT while reading it. */
 114                 down(&current->mm->context.sem);
 115                 desc = current->mm->context.ldt;
 116                 desc = (void *)desc + (seg & ~7);
 117         } else {
 118                 /* Must disable preemption while reading the GDT. */
 119                 desc = (u32 *)get_cpu_gdt_table(get_cpu());
 120                 desc = (void *)desc + (seg & ~7);
 121         }
 122
 123         /* Decode the code segment base from the descriptor */
 124         base = get_desc_base((unsigned long *)desc);
 125
 126         if (seg & (1<<2)) {
 127                 up(&current->mm->context.sem);
 128         } else
 129                 put_cpu();
 130
 131         /* Adjust EIP and segment limit, and clamp at the kernel limit.
 132            It's legitimate for segments to wrap at 0xffffffff. */
 133         seg_limit += base;
 134         if (seg_limit < *eip_limit && seg_limit >= base)
 135                 *eip_limit = seg_limit;
 136         return eip + base;
 137 }
 138
 139 /*
 140  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
 141  * Check that here and ignore it.
 142  */
 143 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
 144 {
 145         unsigned long limit;
 146         unsigned char *instr = (unsigned char *)get_segment_eip (regs, &limit);
 147         int scan_more = 1;
 148         int prefetch = 0;
 149         int i;
 150
 151         for (i = 0; scan_more && i < 15; i++) {
 152                 unsigned char opcode;
 153                 unsigned char instr_hi;
 154                 unsigned char instr_lo;
 155
 156                 if (instr > (unsigned char *)limit)
 157                         break;
 158                 if (probe_kernel_address(instr, opcode))
 159                         break;
 160
 161                 instr_hi = opcode & 0xf0;
 162                 instr_lo = opcode & 0x0f;
 163                 instr++;
 164
 165                 switch (instr_hi) {
 166                 case 0x20:
 167                 case 0x30:
 168                         /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
 169                         scan_more = ((instr_lo & 7) == 0x6);
 170                         break;
 171
 172                 case 0x60:
 173                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 174                         scan_more = (instr_lo & 0xC) == 0x4;
 175                         break;
 176                 case 0xF0:
 177                         /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
 178                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 179                         break;
 180                 case 0x00:
 181                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 182                         scan_more = 0;
 183                         if (instr > (unsigned char *)limit)
 184                                 break;
 185                         if (probe_kernel_address(instr, opcode))
 186                                 break;
 187                         prefetch = (instr_lo == 0xF) &&
 188                                 (opcode == 0x0D || opcode == 0x18);
 189                         break;
 190                 default:
 191                         scan_more = 0;
 192                         break;
 193                 }
 194         }
 195         return prefetch;
 196 }
 197
 198 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
 199                               unsigned long error_code)
 200 {
 201         if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 202                      boot_cpu_data.x86 >= 6)) {
 203                 /* Catch an obscure case of prefetch inside an NX page. */
 204                 if (nx_enabled && (error_code & 16))
 205                         return 0;
 206                 return __is_prefetch(regs, addr);
 207         }
 208         return 0;
 209 }
 210
 211 static noinline void force_sig_info_fault(int si_signo, int si_code,
 212         unsigned long address, struct task_struct *tsk)
 213 {
 214         siginfo_t info;
 215
 216         info.si_signo = si_signo;
 217         info.si_errno = 0;
 218         info.si_code = si_code;
 219         info.si_addr = (void __user *)address;
 220         force_sig_info(si_signo, &info, tsk);
 221 }
 222
 223 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
 224
 225 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 226 {
 227         unsigned index = pgd_index(address);
 228         pgd_t *pgd_k;
 229         pud_t *pud, *pud_k;
 230         pmd_t *pmd, *pmd_k;
 231
 232         pgd += index;
 233         pgd_k = init_mm.pgd + index;
 234
 235         if (!pgd_present(*pgd_k))
 236                 return NULL;
 237
 238         /*
 239          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 240          * and redundant with the set_pmd() on non-PAE. As would
 241          * set_pud.
 242          */
 243
 244         pud = pud_offset(pgd, address);
 245         pud_k = pud_offset(pgd_k, address);
 246         if (!pud_present(*pud_k))
 247                 return NULL;
 248
 249         pmd = pmd_offset(pud, address);
 250         pmd_k = pmd_offset(pud_k, address);
 251         if (!pmd_present(*pmd_k))
 252                 return NULL;
 253         if (!pmd_present(*pmd))
 254                 set_pmd(pmd, *pmd_k);
 255         else
 256                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 257         return pmd_k;
 258 }
 259
 260 /*
 261  * Handle a fault on the vmalloc or module mapping area
 262  *
 263  * This assumes no large pages in there.
 264  */
 265 static inline int vmalloc_fault(unsigned long address)
 266 {
 267         unsigned long pgd_paddr;
 268         pmd_t *pmd_k;
 269         pte_t *pte_k;
 270         /*
 271          * Synchronize this task's top level page-table
 272          * with the 'reference' page table.
 273          *
 274          * Do _not_ use "current" here. We might be inside
 275          * an interrupt in the middle of a task switch..
 276          */
 277         pgd_paddr = read_cr3();
 278         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 279         if (!pmd_k)
 280                 return -1;
 281         pte_k = pte_offset_kernel(pmd_k, address);
 282         if (!pte_present(*pte_k))
 283                 return -1;
 284         return 0;
 285 }
 286
 287 /*
 288  * This routine handles page faults.  It determines the address,
 289  * and the problem, and then passes it off to one of the appropriate
 290  * routines.
 291  *
 292  * error_code:
 293  *      bit 0 == 0 means no page found, 1 means protection fault
 294  *      bit 1 == 0 means read, 1 means write
 295  *      bit 2 == 0 means kernel, 1 means user-mode
 296  *      bit 3 == 1 means use of reserved bit detected
 297  *      bit 4 == 1 means fault was an instruction fetch
 298  */
 299 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
 300                                       unsigned long error_code)
 301 {
 302         struct task_struct *tsk;
 303         struct mm_struct *mm;
 304         struct vm_area_struct * vma;
 305         unsigned long address;
 306         int write, si_code;
 307
 308         /* get the address */
 309         address = read_cr2();
 310
 311         tsk = current;
 312
 313         si_code = SEGV_MAPERR;
 314
 315         /*
 316          * We fault-in kernel-space virtual memory on-demand. The
 317          * 'reference' page table is init_mm.pgd.
 318          *
 319          * NOTE! We MUST NOT take any locks for this case. We may
 320          * be in an interrupt or a critical region, and should
 321          * only copy the information from the master page table,
 322          * nothing more.
 323          *
 324          * This verifies that the fault happens in kernel space
 325          * (error_code & 4) == 0, and that the fault was not a
 326          * protection error (error_code & 9) == 0.
 327          */
 328         if (unlikely(address >= TASK_SIZE)) {
 329                 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
 330                         return;
 331                 if (notify_page_fault(regs, error_code) == NOTIFY_STOP)
 332                         return;
 333                 /*
 334                  * Don't take the mm semaphore here. If we fixup a prefetch
 335                  * fault we could otherwise deadlock.
 336                  */
 337                 goto bad_area_nosemaphore;
 338         }
 339
 340         if (notify_page_fault(regs, error_code) == NOTIFY_STOP)
 341                 return;
 342
 343         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 344            fault has been handled. */
 345         if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
 346                 local_irq_enable();
 347
 348         mm = tsk->mm;
 349
 350         /*
 351          * If we're in an interrupt, have no user context or are running in an
 352          * atomic region then we must not take the fault..
 353          */
 354         if (in_atomic() || !mm)
 355                 goto bad_area_nosemaphore;
 356
 357         /* When running in the kernel we expect faults to occur only to
 358          * addresses in user space.  All other faults represent errors in the
 359          * kernel and should generate an OOPS.  Unfortunatly, in the case of an
 360          * erroneous fault occurring in a code path which already holds mmap_sem
 361          * we will deadlock attempting to validate the fault against the
 362          * address space.  Luckily the kernel only validly references user
 363          * space from well defined areas of code, which are listed in the
 364          * exceptions table.
 365          *
 366          * As the vast majority of faults will be valid we will only perform
 367          * the source reference check when there is a possibilty of a deadlock.
 368          * Attempt to lock the address space, if we cannot we then validate the
 369          * source.  If this is invalid we can skip the address space check,
 370          * thus avoiding the deadlock.
 371          */
 372         if (!down_read_trylock(&mm->mmap_sem)) {
 373                 if ((error_code & 4) == 0 &&
 374                     !search_exception_tables(regs->eip))
 375                         goto bad_area_nosemaphore;
 376                 down_read(&mm->mmap_sem);
 377         }
 378
 379         vma = find_vma(mm, address);
 380         if (!vma)
 381                 goto bad_area;
 382         if (vma->vm_start <= address)
 383                 goto good_area;
 384         if (!(vma->vm_flags & VM_GROWSDOWN))
 385                 goto bad_area;
 386         if (error_code & 4) {
 387                 /*
 388                  * Accessing the stack below %esp is always a bug.
 389                  * The large cushion allows instructions like enter
 390                  * and pusha to work.  ("enter $65535,$31" pushes
 391                  * 32 pointers and then decrements %esp by 65535.)
 392                  */
 393                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->esp)
 394                         goto bad_area;
 395         }
 396         if (expand_stack(vma, address))
 397                 goto bad_area;
 398 /*
 399  * Ok, we have a good vm_area for this memory access, so
 400  * we can handle it..
 401  */
 402 good_area:
 403         si_code = SEGV_ACCERR;
 404         write = 0;
 405         switch (error_code & 3) {
 406                 default:        /* 3: write, present */
 407                                 /* fall through */
 408                 case 2:         /* write, not present */
 409                         if (!(vma->vm_flags & VM_WRITE))
 410                                 goto bad_area;
 411                         write++;
 412                         break;
 413                 case 1:         /* read, present */
 414                         goto bad_area;
 415                 case 0:         /* read, not present */
 416                         if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 417                                 goto bad_area;
 418         }
 419
 420  survive:
 421         /*
 422          * If for any reason at all we couldn't handle the fault,
 423          * make sure we exit gracefully rather than endlessly redo
 424          * the fault.
 425          */
 426         switch (handle_mm_fault(mm, vma, address, write)) {
 427                 case VM_FAULT_MINOR:
 428                         tsk->min_flt++;
 429                         break;
 430                 case VM_FAULT_MAJOR:
 431                         tsk->maj_flt++;
 432                         break;
 433                 case VM_FAULT_SIGBUS:
 434                         goto do_sigbus;
 435                 case VM_FAULT_OOM:
 436                         goto out_of_memory;
 437                 default:
 438                         BUG();
 439         }
 440
 441         /*
 442          * Did it hit the DOS screen memory VA from vm86 mode?
 443          */
 444         if (regs->eflags & VM_MASK) {
 445                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 446                 if (bit < 32)
 447                         tsk->thread.screen_bitmap |= 1 << bit;
 448         }
 449         up_read(&mm->mmap_sem);
 450         return;
 451
 452 /*
 453  * Something tried to access memory that isn't in our memory map..
 454  * Fix it, but check if it's kernel or user first..
 455  */
 456 bad_area:
 457         up_read(&mm->mmap_sem);
 458
 459 bad_area_nosemaphore:
 460         /* User mode accesses just cause a SIGSEGV */
 461         if (error_code & 4) {
 462                 /*
 463                  * Valid to do another page fault here because this one came
 464                  * from user space.
 465                  */
 466                 if (is_prefetch(regs, address, error_code))
 467                         return;
 468
 469                 tsk->thread.cr2 = address;
 470                 /* Kernel addresses are always protection faults */
 471                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 472                 tsk->thread.trap_no = 14;
 473                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 474                 return;
 475         }
 476
 477 #ifdef CONFIG_X86_F00F_BUG
 478         /*
 479          * Pentium F0 0F C7 C8 bug workaround.
 480          */
 481         if (boot_cpu_data.f00f_bug) {
 482                 unsigned long nr;
 483
 484                 nr = (address - idt_descr.address) >> 3;
 485
 486                 if (nr == 6) {
 487                         do_invalid_op(regs, 0);
 488                         return;
 489                 }
 490         }
 491 #endif
 492
 493 no_context:
 494         /* Are we prepared to handle this kernel fault?  */
 495         if (fixup_exception(regs))
 496                 return;
 497
 498         /*
 499          * Valid to do another page fault here, because if this fault
 500          * had been triggered by is_prefetch fixup_exception would have
 501          * handled it.
 502          */
 503         if (is_prefetch(regs, address, error_code))
 504                 return;
 505
 506 /*
 507  * Oops. The kernel tried to access some bad page. We'll have to
 508  * terminate things with extreme prejudice.
 509  */
 510
 511         bust_spinlocks(1);
 512
 513         if (oops_may_print()) {
 514                 __typeof__(pte_val(__pte(0))) page;
 515
 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 = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 538 #ifdef CONFIG_X86_PAE
 539                 printk(KERN_ALERT "*pdpt = %016Lx\n", page);
 540                 if ((page >> PAGE_SHIFT) < max_low_pfn
 541                     && page & _PAGE_PRESENT) {
 542                         page &= PAGE_MASK;
 543                         page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 544                                                                  & (PTRS_PER_PMD - 1)];
 545                         printk(KERN_ALERT "*pde = %016Lx\n", page);
 546                         page &= ~_PAGE_NX;
 547                 }
 548 #else
 549                 printk(KERN_ALERT "*pde = %08lx\n", page);
 550 #endif
 551
 552                 /*
 553                  * We must not directly access the pte in the highpte
 554                  * case if the page table is located in highmem.
 555                  * And let's rather not kmap-atomic the pte, just in case
 556                  * it's allocated already.
 557                  */
 558                 if ((page >> PAGE_SHIFT) < max_low_pfn
 559                     && (page & _PAGE_PRESENT)) {
 560                         page &= PAGE_MASK;
 561                         page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 562                                                                  & (PTRS_PER_PTE - 1)];
 563                         printk(KERN_ALERT "*pte = %0*Lx\n", sizeof(page)*2, (u64)page);
 564                 }
 565         }
 566
 567         tsk->thread.cr2 = address;
 568         tsk->thread.trap_no = 14;
 569         tsk->thread.error_code = error_code;
 570         die("Oops", regs, error_code);
 571         bust_spinlocks(0);
 572         do_exit(SIGKILL);
 573
 574 /*
 575  * We ran out of memory, or some other thing happened to us that made
 576  * us unable to handle the page fault gracefully.
 577  */
 578 out_of_memory:
 579         up_read(&mm->mmap_sem);
 580         if (is_init(tsk)) {
 581                 yield();
 582                 down_read(&mm->mmap_sem);
 583                 goto survive;
 584         }
 585         printk("VM: killing process %s\n", tsk->comm);
 586         if (error_code & 4)
 587                 do_exit(SIGKILL);
 588         goto no_context;
 589
 590 do_sigbus:
 591         up_read(&mm->mmap_sem);
 592
 593         /* Kernel mode? Handle exceptions or die */
 594         if (!(error_code & 4))
 595                 goto no_context;
 596
 597         /* User space => ok to do another page fault */
 598         if (is_prefetch(regs, address, error_code))
 599                 return;
 600
 601         tsk->thread.cr2 = address;
 602         tsk->thread.error_code = error_code;
 603         tsk->thread.trap_no = 14;
 604         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 605 }
 606
 607 void vmalloc_sync_all(void)
 608 {
 609         /*
 610          * Note that races in the updates of insync and start aren't
 611          * problematic: insync can only get set bits added, and updates to
 612          * start are only improving performance (without affecting correctness
 613          * if undone).
 614          */
 615         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 616         static unsigned long start = TASK_SIZE;
 617         unsigned long address;
 618
 619         if (SHARED_KERNEL_PMD)
 620                 return;
 621
 622         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 623         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 624                 if (!test_bit(pgd_index(address), insync)) {
 625                         unsigned long flags;
 626                         struct page *page;
 627
 628                         spin_lock_irqsave(&pgd_lock, flags);
 629                         for (page = pgd_list; page; page =
 630                                         (struct page *)page->index)
 631                                 if (!vmalloc_sync_one(page_address(page),
 632                                                                 address)) {
 633                                         BUG_ON(page != pgd_list);
 634                                         break;
 635                                 }
 636                         spin_unlock_irqrestore(&pgd_lock, flags);
 637                         if (!page)
 638                                 set_bit(pgd_index(address), insync);
 639                 }
 640                 if (address == start && test_bit(pgd_index(address), insync))
 641                         start = address + PGDIR_SIZE;
 642         }
 643 }