arch/x86/mm/fault.c

   1 /*
   2  *  Copyright (C) 1995  Linus Torvalds
   3  *  Copyright (C) 2001,2002 Andi Kleen, SuSE Labs.
   4  */
   5
   6 #include <linux/signal.h>
   7 #include <linux/sched.h>
   8 #include <linux/kernel.h>
   9 #include <linux/errno.h>
  10 #include <linux/string.h>
  11 #include <linux/types.h>
  12 #include <linux/ptrace.h>
  13 #include <linux/mmiotrace.h>
  14 #include <linux/mman.h>
  15 #include <linux/mm.h>
  16 #include <linux/smp.h>
  17 #include <linux/interrupt.h>
  18 #include <linux/init.h>
  19 #include <linux/tty.h>
  20 #include <linux/vt_kern.h>              /* For unblank_screen() */
  21 #include <linux/compiler.h>
  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 #include <asm/pgalloc.h>
  34 #include <asm/smp.h>
  35 #include <asm/tlbflush.h>
  36 #include <asm/proto.h>
  37 #include <asm-generic/sections.h>
  38 #include <asm/traps.h>
  39
  40 /*
  41  * Page fault error code bits
  42  *      bit 0 == 0 means no page found, 1 means protection fault
  43  *      bit 1 == 0 means read, 1 means write
  44  *      bit 2 == 0 means kernel, 1 means user-mode
  45  *      bit 3 == 1 means use of reserved bit detected
  46  *      bit 4 == 1 means fault was an instruction fetch
  47  */
  48 #define PF_PROT         (1<<0)
  49 #define PF_WRITE        (1<<1)
  50 #define PF_USER         (1<<2)
  51 #define PF_RSVD         (1<<3)
  52 #define PF_INSTR        (1<<4)
  53
  54 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
  55 {
  56 #ifdef CONFIG_MMIOTRACE_HOOKS
  57         if (unlikely(is_kmmio_active()))
  58                 if (kmmio_handler(regs, addr) == 1)
  59                         return -1;
  60 #endif
  61         return 0;
  62 }
  63
  64 static inline int notify_page_fault(struct pt_regs *regs)
  65 {
  66 #ifdef CONFIG_KPROBES
  67         int ret = 0;
  68
  69         /* kprobe_running() needs smp_processor_id() */
  70         if (!user_mode_vm(regs)) {
  71                 preempt_disable();
  72                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  73                         ret = 1;
  74                 preempt_enable();
  75         }
  76
  77         return ret;
  78 #else
  79         return 0;
  80 #endif
  81 }
  82
  83 /*
  84  * X86_32
  85  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  86  * Check that here and ignore it.
  87  *
  88  * X86_64
  89  * Sometimes the CPU reports invalid exceptions on prefetch.
  90  * Check that here and ignore it.
  91  *
  92  * Opcode checker based on code by Richard Brunner
  93  */
  94 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
  95                        unsigned long error_code)
  96 {
  97         unsigned char *instr;
  98         int scan_more = 1;
  99         int prefetch = 0;
 100         unsigned char *max_instr;
 101
 102         /*
 103          * If it was a exec (instruction fetch) fault on NX page, then
 104          * do not ignore the fault:
 105          */
 106         if (error_code & PF_INSTR)
 107                 return 0;
 108
 109         instr = (unsigned char *)convert_ip_to_linear(current, regs);
 110         max_instr = instr + 15;
 111
 112         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 113                 return 0;
 114
 115         while (scan_more && instr < max_instr) {
 116                 unsigned char opcode;
 117                 unsigned char instr_hi;
 118                 unsigned char instr_lo;
 119
 120                 if (probe_kernel_address(instr, opcode))
 121                         break;
 122
 123                 instr_hi = opcode & 0xf0;
 124                 instr_lo = opcode & 0x0f;
 125                 instr++;
 126
 127                 switch (instr_hi) {
 128                 case 0x20:
 129                 case 0x30:
 130                         /*
 131                          * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
 132                          * In X86_64 long mode, the CPU will signal invalid
 133                          * opcode if some of these prefixes are present so
 134                          * X86_64 will never get here anyway
 135                          */
 136                         scan_more = ((instr_lo & 7) == 0x6);
 137                         break;
 138 #ifdef CONFIG_X86_64
 139                 case 0x40:
 140                         /*
 141                          * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 142                          * Need to figure out under what instruction mode the
 143                          * instruction was issued. Could check the LDT for lm,
 144                          * but for now it's good enough to assume that long
 145                          * mode only uses well known segments or kernel.
 146                          */
 147                         scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
 148                         break;
 149 #endif
 150                 case 0x60:
 151                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 152                         scan_more = (instr_lo & 0xC) == 0x4;
 153                         break;
 154                 case 0xF0:
 155                         /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 156                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 157                         break;
 158                 case 0x00:
 159                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 160                         scan_more = 0;
 161
 162                         if (probe_kernel_address(instr, opcode))
 163                                 break;
 164                         prefetch = (instr_lo == 0xF) &&
 165                                 (opcode == 0x0D || opcode == 0x18);
 166                         break;
 167                 default:
 168                         scan_more = 0;
 169                         break;
 170                 }
 171         }
 172         return prefetch;
 173 }
 174
 175 static void force_sig_info_fault(int si_signo, int si_code,
 176         unsigned long address, struct task_struct *tsk)
 177 {
 178         siginfo_t info;
 179
 180         info.si_signo = si_signo;
 181         info.si_errno = 0;
 182         info.si_code = si_code;
 183         info.si_addr = (void __user *)address;
 184         force_sig_info(si_signo, &info, tsk);
 185 }
 186
 187 #ifdef CONFIG_X86_64
 188 static int bad_address(void *p)
 189 {
 190         unsigned long dummy;
 191         return probe_kernel_address((unsigned long *)p, dummy);
 192 }
 193 #endif
 194
 195 static void dump_pagetable(unsigned long address)
 196 {
 197 #ifdef CONFIG_X86_32
 198         __typeof__(pte_val(__pte(0))) page;
 199
 200         page = read_cr3();
 201         page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 202 #ifdef CONFIG_X86_PAE
 203         printk("*pdpt = %016Lx ", page);
 204         if ((page >> PAGE_SHIFT) < max_low_pfn
 205             && page & _PAGE_PRESENT) {
 206                 page &= PAGE_MASK;
 207                 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 208                                                          & (PTRS_PER_PMD - 1)];
 209                 printk(KERN_CONT "*pde = %016Lx ", page);
 210                 page &= ~_PAGE_NX;
 211         }
 212 #else
 213         printk("*pde = %08lx ", page);
 214 #endif
 215
 216         /*
 217          * We must not directly access the pte in the highpte
 218          * case if the page table is located in highmem.
 219          * And let's rather not kmap-atomic the pte, just in case
 220          * it's allocated already.
 221          */
 222         if ((page >> PAGE_SHIFT) < max_low_pfn
 223             && (page & _PAGE_PRESENT)
 224             && !(page & _PAGE_PSE)) {
 225                 page &= PAGE_MASK;
 226                 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 227                                                          & (PTRS_PER_PTE - 1)];
 228                 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 229         }
 230
 231         printk("\n");
 232 #else /* CONFIG_X86_64 */
 233         pgd_t *pgd;
 234         pud_t *pud;
 235         pmd_t *pmd;
 236         pte_t *pte;
 237
 238         pgd = (pgd_t *)read_cr3();
 239
 240         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 241         pgd += pgd_index(address);
 242         if (bad_address(pgd)) goto bad;
 243         printk("PGD %lx ", pgd_val(*pgd));
 244         if (!pgd_present(*pgd)) goto ret;
 245
 246         pud = pud_offset(pgd, address);
 247         if (bad_address(pud)) goto bad;
 248         printk("PUD %lx ", pud_val(*pud));
 249         if (!pud_present(*pud) || pud_large(*pud))
 250                 goto ret;
 251
 252         pmd = pmd_offset(pud, address);
 253         if (bad_address(pmd)) goto bad;
 254         printk("PMD %lx ", pmd_val(*pmd));
 255         if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
 256
 257         pte = pte_offset_kernel(pmd, address);
 258         if (bad_address(pte)) goto bad;
 259         printk("PTE %lx", pte_val(*pte));
 260 ret:
 261         printk("\n");
 262         return;
 263 bad:
 264         printk("BAD\n");
 265 #endif
 266 }
 267
 268 #ifdef CONFIG_X86_32
 269 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 270 {
 271         unsigned index = pgd_index(address);
 272         pgd_t *pgd_k;
 273         pud_t *pud, *pud_k;
 274         pmd_t *pmd, *pmd_k;
 275
 276         pgd += index;
 277         pgd_k = init_mm.pgd + index;
 278
 279         if (!pgd_present(*pgd_k))
 280                 return NULL;
 281
 282         /*
 283          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 284          * and redundant with the set_pmd() on non-PAE. As would
 285          * set_pud.
 286          */
 287
 288         pud = pud_offset(pgd, address);
 289         pud_k = pud_offset(pgd_k, address);
 290         if (!pud_present(*pud_k))
 291                 return NULL;
 292
 293         pmd = pmd_offset(pud, address);
 294         pmd_k = pmd_offset(pud_k, address);
 295         if (!pmd_present(*pmd_k))
 296                 return NULL;
 297         if (!pmd_present(*pmd)) {
 298                 set_pmd(pmd, *pmd_k);
 299                 arch_flush_lazy_mmu_mode();
 300         } else
 301                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 302         return pmd_k;
 303 }
 304 #endif
 305
 306 #ifdef CONFIG_X86_64
 307 static const char errata93_warning[] =
 308 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 309 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 310 KERN_ERR "******* Please consider a BIOS update.\n"
 311 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 312 #endif
 313
 314 /* Workaround for K8 erratum #93 & buggy BIOS.
 315    BIOS SMM functions are required to use a specific workaround
 316    to avoid corruption of the 64bit RIP register on C stepping K8.
 317    A lot of BIOS that didn't get tested properly miss this.
 318    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 319    Try to work around it here.
 320    Note we only handle faults in kernel here.
 321    Does nothing for X86_32
 322  */
 323 static int is_errata93(struct pt_regs *regs, unsigned long address)
 324 {
 325 #ifdef CONFIG_X86_64
 326         static int warned;
 327         if (address != regs->ip)
 328                 return 0;
 329         if ((address >> 32) != 0)
 330                 return 0;
 331         address |= 0xffffffffUL << 32;
 332         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 333             (address >= MODULES_VADDR && address <= MODULES_END)) {
 334                 if (!warned) {
 335                         printk(errata93_warning);
 336                         warned = 1;
 337                 }
 338                 regs->ip = address;
 339                 return 1;
 340         }
 341 #endif
 342         return 0;
 343 }
 344
 345 /*
 346  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
 347  * addresses >4GB.  We catch this in the page fault handler because these
 348  * addresses are not reachable. Just detect this case and return.  Any code
 349  * segment in LDT is compatibility mode.
 350  */
 351 static int is_errata100(struct pt_regs *regs, unsigned long address)
 352 {
 353 #ifdef CONFIG_X86_64
 354         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
 355             (address >> 32))
 356                 return 1;
 357 #endif
 358         return 0;
 359 }
 360
 361 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 362 {
 363 #ifdef CONFIG_X86_F00F_BUG
 364         unsigned long nr;
 365         /*
 366          * Pentium F0 0F C7 C8 bug workaround.
 367          */
 368         if (boot_cpu_data.f00f_bug) {
 369                 nr = (address - idt_descr.address) >> 3;
 370
 371                 if (nr == 6) {
 372                         do_invalid_op(regs, 0);
 373                         return 1;
 374                 }
 375         }
 376 #endif
 377         return 0;
 378 }
 379
 380 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 381                             unsigned long address)
 382 {
 383 #ifdef CONFIG_X86_32
 384         if (!oops_may_print())
 385                 return;
 386 #endif
 387
 388 #ifdef CONFIG_X86_PAE
 389         if (error_code & PF_INSTR) {
 390                 unsigned int level;
 391                 pte_t *pte = lookup_address(address, &level);
 392
 393                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 394                         printk(KERN_CRIT "kernel tried to execute "
 395                                 "NX-protected page - exploit attempt? "
 396                                 "(uid: %d)\n", current->uid);
 397         }
 398 #endif
 399
 400         printk(KERN_ALERT "BUG: unable to handle kernel ");
 401         if (address < PAGE_SIZE)
 402                 printk(KERN_CONT "NULL pointer dereference");
 403         else
 404                 printk(KERN_CONT "paging request");
 405         printk(KERN_CONT " at %p\n", (void *) address);
 406         printk(KERN_ALERT "IP:");
 407         printk_address(regs->ip, 1);
 408         dump_pagetable(address);
 409 }
 410
 411 #ifdef CONFIG_X86_64
 412 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
 413                                  unsigned long error_code)
 414 {
 415         unsigned long flags = oops_begin();
 416         struct task_struct *tsk;
 417
 418         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 419                current->comm, address);
 420         dump_pagetable(address);
 421         tsk = current;
 422         tsk->thread.cr2 = address;
 423         tsk->thread.trap_no = 14;
 424         tsk->thread.error_code = error_code;
 425         if (__die("Bad pagetable", regs, error_code))
 426                 regs = NULL;
 427         oops_end(flags, regs, SIGKILL);
 428 }
 429 #endif
 430
 431 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 432 {
 433         if ((error_code & PF_WRITE) && !pte_write(*pte))
 434                 return 0;
 435         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 436                 return 0;
 437
 438         return 1;
 439 }
 440
 441 /*
 442  * Handle a spurious fault caused by a stale TLB entry.  This allows
 443  * us to lazily refresh the TLB when increasing the permissions of a
 444  * kernel page (RO -> RW or NX -> X).  Doing it eagerly is very
 445  * expensive since that implies doing a full cross-processor TLB
 446  * flush, even if no stale TLB entries exist on other processors.
 447  * There are no security implications to leaving a stale TLB when
 448  * increasing the permissions on a page.
 449  */
 450 static int spurious_fault(unsigned long address,
 451                           unsigned long error_code)
 452 {
 453         pgd_t *pgd;
 454         pud_t *pud;
 455         pmd_t *pmd;
 456         pte_t *pte;
 457
 458         /* Reserved-bit violation or user access to kernel space? */
 459         if (error_code & (PF_USER | PF_RSVD))
 460                 return 0;
 461
 462         pgd = init_mm.pgd + pgd_index(address);
 463         if (!pgd_present(*pgd))
 464                 return 0;
 465
 466         pud = pud_offset(pgd, address);
 467         if (!pud_present(*pud))
 468                 return 0;
 469
 470         if (pud_large(*pud))
 471                 return spurious_fault_check(error_code, (pte_t *) pud);
 472
 473         pmd = pmd_offset(pud, address);
 474         if (!pmd_present(*pmd))
 475                 return 0;
 476
 477         if (pmd_large(*pmd))
 478                 return spurious_fault_check(error_code, (pte_t *) pmd);
 479
 480         pte = pte_offset_kernel(pmd, address);
 481         if (!pte_present(*pte))
 482                 return 0;
 483
 484         return spurious_fault_check(error_code, pte);
 485 }
 486
 487 /*
 488  * X86_32
 489  * Handle a fault on the vmalloc or module mapping area
 490  *
 491  * X86_64
 492  * Handle a fault on the vmalloc area
 493  *
 494  * This assumes no large pages in there.
 495  */
 496 static int vmalloc_fault(unsigned long address)
 497 {
 498 #ifdef CONFIG_X86_32
 499         unsigned long pgd_paddr;
 500         pmd_t *pmd_k;
 501         pte_t *pte_k;
 502
 503         /* Make sure we are in vmalloc area */
 504         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 505                 return -1;
 506
 507         /*
 508          * Synchronize this task's top level page-table
 509          * with the 'reference' page table.
 510          *
 511          * Do _not_ use "current" here. We might be inside
 512          * an interrupt in the middle of a task switch..
 513          */
 514         pgd_paddr = read_cr3();
 515         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 516         if (!pmd_k)
 517                 return -1;
 518         pte_k = pte_offset_kernel(pmd_k, address);
 519         if (!pte_present(*pte_k))
 520                 return -1;
 521         return 0;
 522 #else
 523         pgd_t *pgd, *pgd_ref;
 524         pud_t *pud, *pud_ref;
 525         pmd_t *pmd, *pmd_ref;
 526         pte_t *pte, *pte_ref;
 527
 528         /* Make sure we are in vmalloc area */
 529         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 530                 return -1;
 531
 532         /* Copy kernel mappings over when needed. This can also
 533            happen within a race in page table update. In the later
 534            case just flush. */
 535
 536         pgd = pgd_offset(current->mm ?: &init_mm, address);
 537         pgd_ref = pgd_offset_k(address);
 538         if (pgd_none(*pgd_ref))
 539                 return -1;
 540         if (pgd_none(*pgd))
 541                 set_pgd(pgd, *pgd_ref);
 542         else
 543                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 544
 545         /* Below here mismatches are bugs because these lower tables
 546            are shared */
 547
 548         pud = pud_offset(pgd, address);
 549         pud_ref = pud_offset(pgd_ref, address);
 550         if (pud_none(*pud_ref))
 551                 return -1;
 552         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 553                 BUG();
 554         pmd = pmd_offset(pud, address);
 555         pmd_ref = pmd_offset(pud_ref, address);
 556         if (pmd_none(*pmd_ref))
 557                 return -1;
 558         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 559                 BUG();
 560         pte_ref = pte_offset_kernel(pmd_ref, address);
 561         if (!pte_present(*pte_ref))
 562                 return -1;
 563         pte = pte_offset_kernel(pmd, address);
 564         /* Don't use pte_page here, because the mappings can point
 565            outside mem_map, and the NUMA hash lookup cannot handle
 566            that. */
 567         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 568                 BUG();
 569         return 0;
 570 #endif
 571 }
 572
 573 int show_unhandled_signals = 1;
 574
 575 /*
 576  * This routine handles page faults.  It determines the address,
 577  * and the problem, and then passes it off to one of the appropriate
 578  * routines.
 579  */
 580 #ifdef CONFIG_X86_64
 581 asmlinkage
 582 #endif
 583 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
 584 {
 585         struct task_struct *tsk;
 586         struct mm_struct *mm;
 587         struct vm_area_struct *vma;
 588         unsigned long address;
 589         int write, si_code;
 590         int fault;
 591 #ifdef CONFIG_X86_64
 592         unsigned long flags;
 593 #endif
 594
 595         /*
 596          * We can fault from pretty much anywhere, with unknown IRQ state.
 597          */
 598         trace_hardirqs_fixup();
 599
 600         tsk = current;
 601         mm = tsk->mm;
 602         prefetchw(&mm->mmap_sem);
 603
 604         /* get the address */
 605         address = read_cr2();
 606
 607         si_code = SEGV_MAPERR;
 608
 609         if (notify_page_fault(regs))
 610                 return;
 611         if (unlikely(kmmio_fault(regs, address)))
 612                 return;
 613
 614         /*
 615          * We fault-in kernel-space virtual memory on-demand. The
 616          * 'reference' page table is init_mm.pgd.
 617          *
 618          * NOTE! We MUST NOT take any locks for this case. We may
 619          * be in an interrupt or a critical region, and should
 620          * only copy the information from the master page table,
 621          * nothing more.
 622          *
 623          * This verifies that the fault happens in kernel space
 624          * (error_code & 4) == 0, and that the fault was not a
 625          * protection error (error_code & 9) == 0.
 626          */
 627 #ifdef CONFIG_X86_32
 628         if (unlikely(address >= TASK_SIZE)) {
 629 #else
 630         if (unlikely(address >= TASK_SIZE64)) {
 631 #endif
 632                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 633                     vmalloc_fault(address) >= 0)
 634                         return;
 635
 636                 /* Can handle a stale RO->RW TLB */
 637                 if (spurious_fault(address, error_code))
 638                         return;
 639
 640                 /*
 641                  * Don't take the mm semaphore here. If we fixup a prefetch
 642                  * fault we could otherwise deadlock.
 643                  */
 644                 goto bad_area_nosemaphore;
 645         }
 646
 647
 648         /*
 649          * It's safe to allow irq's after cr2 has been saved and the
 650          * vmalloc fault has been handled.
 651          *
 652          * User-mode registers count as a user access even for any
 653          * potential system fault or CPU buglet.
 654          */
 655         if (user_mode_vm(regs)) {
 656                 local_irq_enable();
 657                 error_code |= PF_USER;
 658         } else if (regs->flags & X86_EFLAGS_IF)
 659                 local_irq_enable();
 660
 661 #ifdef CONFIG_X86_64
 662         if (unlikely(error_code & PF_RSVD))
 663                 pgtable_bad(address, regs, error_code);
 664 #endif
 665
 666         /*
 667          * If we're in an interrupt, have no user context or are running in an
 668          * atomic region then we must not take the fault.
 669          */
 670         if (unlikely(in_atomic() || !mm))
 671                 goto bad_area_nosemaphore;
 672
 673 again:
 674         /*
 675          * When running in the kernel we expect faults to occur only to
 676          * addresses in user space.  All other faults represent errors in the
 677          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 678          * erroneous fault occurring in a code path which already holds mmap_sem
 679          * we will deadlock attempting to validate the fault against the
 680          * address space.  Luckily the kernel only validly references user
 681          * space from well defined areas of code, which are listed in the
 682          * exceptions table.
 683          *
 684          * As the vast majority of faults will be valid we will only perform
 685          * the source reference check when there is a possibility of a deadlock.
 686          * Attempt to lock the address space, if we cannot we then validate the
 687          * source.  If this is invalid we can skip the address space check,
 688          * thus avoiding the deadlock.
 689          */
 690         if (!down_read_trylock(&mm->mmap_sem)) {
 691                 if ((error_code & PF_USER) == 0 &&
 692                     !search_exception_tables(regs->ip))
 693                         goto bad_area_nosemaphore;
 694                 down_read(&mm->mmap_sem);
 695         }
 696
 697         vma = find_vma(mm, address);
 698         if (!vma)
 699                 goto bad_area;
 700         if (vma->vm_start <= address)
 701                 goto good_area;
 702         if (!(vma->vm_flags & VM_GROWSDOWN))
 703                 goto bad_area;
 704         if (error_code & PF_USER) {
 705                 /*
 706                  * Accessing the stack below %sp is always a bug.
 707                  * The large cushion allows instructions like enter
 708                  * and pusha to work.  ("enter $65535,$31" pushes
 709                  * 32 pointers and then decrements %sp by 65535.)
 710                  */
 711                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
 712                         goto bad_area;
 713         }
 714         if (expand_stack(vma, address))
 715                 goto bad_area;
 716 /*
 717  * Ok, we have a good vm_area for this memory access, so
 718  * we can handle it..
 719  */
 720 good_area:
 721         si_code = SEGV_ACCERR;
 722         write = 0;
 723         switch (error_code & (PF_PROT|PF_WRITE)) {
 724         default:        /* 3: write, present */
 725                 /* fall through */
 726         case PF_WRITE:          /* write, not present */
 727                 if (!(vma->vm_flags & VM_WRITE))
 728                         goto bad_area;
 729                 write++;
 730                 break;
 731         case PF_PROT:           /* read, present */
 732                 goto bad_area;
 733         case 0:                 /* read, not present */
 734                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 735                         goto bad_area;
 736         }
 737
 738         /*
 739          * If for any reason at all we couldn't handle the fault,
 740          * make sure we exit gracefully rather than endlessly redo
 741          * the fault.
 742          */
 743         fault = handle_mm_fault(mm, vma, address, write);
 744         if (unlikely(fault & VM_FAULT_ERROR)) {
 745                 if (fault & VM_FAULT_OOM)
 746                         goto out_of_memory;
 747                 else if (fault & VM_FAULT_SIGBUS)
 748                         goto do_sigbus;
 749                 BUG();
 750         }
 751         if (fault & VM_FAULT_MAJOR)
 752                 tsk->maj_flt++;
 753         else
 754                 tsk->min_flt++;
 755
 756 #ifdef CONFIG_X86_32
 757         /*
 758          * Did it hit the DOS screen memory VA from vm86 mode?
 759          */
 760         if (v8086_mode(regs)) {
 761                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 762                 if (bit < 32)
 763                         tsk->thread.screen_bitmap |= 1 << bit;
 764         }
 765 #endif
 766         up_read(&mm->mmap_sem);
 767         return;
 768
 769 /*
 770  * Something tried to access memory that isn't in our memory map..
 771  * Fix it, but check if it's kernel or user first..
 772  */
 773 bad_area:
 774         up_read(&mm->mmap_sem);
 775
 776 bad_area_nosemaphore:
 777         /* User mode accesses just cause a SIGSEGV */
 778         if (error_code & PF_USER) {
 779                 /*
 780                  * It's possible to have interrupts off here.
 781                  */
 782                 local_irq_enable();
 783
 784                 /*
 785                  * Valid to do another page fault here because this one came
 786                  * from user space.
 787                  */
 788                 if (is_prefetch(regs, address, error_code))
 789                         return;
 790
 791                 if (is_errata100(regs, address))
 792                         return;
 793
 794                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 795                     printk_ratelimit()) {
 796                         printk(
 797                         "%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 798                         task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 799                         tsk->comm, task_pid_nr(tsk), address,
 800                         (void *) regs->ip, (void *) regs->sp, error_code);
 801                         print_vma_addr(" in ", regs->ip);
 802                         printk("\n");
 803                 }
 804
 805                 tsk->thread.cr2 = address;
 806                 /* Kernel addresses are always protection faults */
 807                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 808                 tsk->thread.trap_no = 14;
 809                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 810                 return;
 811         }
 812
 813         if (is_f00f_bug(regs, address))
 814                 return;
 815
 816 no_context:
 817         /* Are we prepared to handle this kernel fault?  */
 818         if (fixup_exception(regs))
 819                 return;
 820
 821         /*
 822          * X86_32
 823          * Valid to do another page fault here, because if this fault
 824          * had been triggered by is_prefetch fixup_exception would have
 825          * handled it.
 826          *
 827          * X86_64
 828          * Hall of shame of CPU/BIOS bugs.
 829          */
 830         if (is_prefetch(regs, address, error_code))
 831                 return;
 832
 833         if (is_errata93(regs, address))
 834                 return;
 835
 836 /*
 837  * Oops. The kernel tried to access some bad page. We'll have to
 838  * terminate things with extreme prejudice.
 839  */
 840 #ifdef CONFIG_X86_32
 841         bust_spinlocks(1);
 842 #else
 843         flags = oops_begin();
 844 #endif
 845
 846         show_fault_oops(regs, error_code, address);
 847
 848         tsk->thread.cr2 = address;
 849         tsk->thread.trap_no = 14;
 850         tsk->thread.error_code = error_code;
 851
 852 #ifdef CONFIG_X86_32
 853         die("Oops", regs, error_code);
 854         bust_spinlocks(0);
 855         do_exit(SIGKILL);
 856 #else
 857         if (__die("Oops", regs, error_code))
 858                 regs = NULL;
 859         /* Executive summary in case the body of the oops scrolled away */
 860         printk(KERN_EMERG "CR2: %016lx\n", address);
 861         oops_end(flags, regs, SIGKILL);
 862 #endif
 863
 864 /*
 865  * We ran out of memory, or some other thing happened to us that made
 866  * us unable to handle the page fault gracefully.
 867  */
 868 out_of_memory:
 869         up_read(&mm->mmap_sem);
 870         if (is_global_init(tsk)) {
 871                 yield();
 872                 /*
 873                  * Re-lookup the vma - in theory the vma tree might
 874                  * have changed:
 875                  */
 876                 goto again;
 877         }
 878
 879         printk("VM: killing process %s\n", tsk->comm);
 880         if (error_code & PF_USER)
 881                 do_group_exit(SIGKILL);
 882         goto no_context;
 883
 884 do_sigbus:
 885         up_read(&mm->mmap_sem);
 886
 887         /* Kernel mode? Handle exceptions or die */
 888         if (!(error_code & PF_USER))
 889                 goto no_context;
 890 #ifdef CONFIG_X86_32
 891         /* User space => ok to do another page fault */
 892         if (is_prefetch(regs, address, error_code))
 893                 return;
 894 #endif
 895         tsk->thread.cr2 = address;
 896         tsk->thread.error_code = error_code;
 897         tsk->thread.trap_no = 14;
 898         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 899 }
 900
 901 DEFINE_SPINLOCK(pgd_lock);
 902 LIST_HEAD(pgd_list);
 903
 904 void vmalloc_sync_all(void)
 905 {
 906         unsigned long address;
 907
 908 #ifdef CONFIG_X86_32
 909         if (SHARED_KERNEL_PMD)
 910                 return;
 911
 912         for (address = VMALLOC_START & PMD_MASK;
 913              address >= TASK_SIZE && address < FIXADDR_TOP;
 914              address += PMD_SIZE) {
 915                 unsigned long flags;
 916                 struct page *page;
 917
 918                 spin_lock_irqsave(&pgd_lock, flags);
 919                 list_for_each_entry(page, &pgd_list, lru) {
 920                         if (!vmalloc_sync_one(page_address(page),
 921                                               address))
 922                                 break;
 923                 }
 924                 spin_unlock_irqrestore(&pgd_lock, flags);
 925         }
 926 #else /* CONFIG_X86_64 */
 927         for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
 928              address += PGDIR_SIZE) {
 929                 const pgd_t *pgd_ref = pgd_offset_k(address);
 930                 unsigned long flags;
 931                 struct page *page;
 932
 933                 if (pgd_none(*pgd_ref))
 934                         continue;
 935                 spin_lock_irqsave(&pgd_lock, flags);
 936                 list_for_each_entry(page, &pgd_list, lru) {
 937                         pgd_t *pgd;
 938                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 939                         if (pgd_none(*pgd))
 940                                 set_pgd(pgd, *pgd_ref);
 941                         else
 942                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 943                 }
 944                 spin_unlock_irqrestore(&pgd_lock, flags);
 945         }
 946 #endif
 947 }