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/mman.h>
  14 #include <linux/mm.h>
  15 #include <linux/smp.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/init.h>
  18 #include <linux/tty.h>
  19 #include <linux/vt_kern.h>              /* For unblank_screen() */
  20 #include <linux/compiler.h>
  21 #include <linux/highmem.h>
  22 #include <linux/bootmem.h>              /* for max_low_pfn */
  23 #include <linux/vmalloc.h>
  24 #include <linux/module.h>
  25 #include <linux/kprobes.h>
  26 #include <linux/uaccess.h>
  27 #include <linux/kdebug.h>
  28
  29 #include <asm/system.h>
  30 #include <asm/desc.h>
  31 #include <asm/segment.h>
  32 #include <asm/pgalloc.h>
  33 #include <asm/smp.h>
  34 #include <asm/tlbflush.h>
  35 #include <asm/proto.h>
  36 #include <asm-generic/sections.h>
  37
  38 /*
  39  * Page fault error code bits
  40  *      bit 0 == 0 means no page found, 1 means protection fault
  41  *      bit 1 == 0 means read, 1 means write
  42  *      bit 2 == 0 means kernel, 1 means user-mode
  43  *      bit 3 == 1 means use of reserved bit detected
  44  *      bit 4 == 1 means fault was an instruction fetch
  45  */
  46 #define PF_PROT         (1<<0)
  47 #define PF_WRITE        (1<<1)
  48 #define PF_USER         (1<<2)
  49 #define PF_RSVD         (1<<3)
  50 #define PF_INSTR        (1<<4)
  51
  52 static inline int notify_page_fault(struct pt_regs *regs)
  53 {
  54 #ifdef CONFIG_KPROBES
  55         int ret = 0;
  56
  57         /* kprobe_running() needs smp_processor_id() */
  58 #ifdef CONFIG_X86_32
  59         if (!user_mode_vm(regs)) {
  60 #else
  61         if (!user_mode(regs)) {
  62 #endif
  63                 preempt_disable();
  64                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  65                         ret = 1;
  66                 preempt_enable();
  67         }
  68
  69         return ret;
  70 #else
  71         return 0;
  72 #endif
  73 }
  74
  75 /*
  76  * X86_32
  77  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  78  * Check that here and ignore it.
  79  *
  80  * X86_64
  81  * Sometimes the CPU reports invalid exceptions on prefetch.
  82  * Check that here and ignore it.
  83  *
  84  * Opcode checker based on code by Richard Brunner
  85  */
  86 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
  87                        unsigned long error_code)
  88 {
  89         unsigned char *instr;
  90         int scan_more = 1;
  91         int prefetch = 0;
  92         unsigned char *max_instr;
  93
  94         /*
  95          * If it was a exec (instruction fetch) fault on NX page, then
  96          * do not ignore the fault:
  97          */
  98         if (error_code & PF_INSTR)
  99                 return 0;
 100
 101         instr = (unsigned char *)convert_ip_to_linear(current, regs);
 102         max_instr = instr + 15;
 103
 104         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 105                 return 0;
 106
 107         while (scan_more && instr < max_instr) {
 108                 unsigned char opcode;
 109                 unsigned char instr_hi;
 110                 unsigned char instr_lo;
 111
 112                 if (probe_kernel_address(instr, opcode))
 113                         break;
 114
 115                 instr_hi = opcode & 0xf0;
 116                 instr_lo = opcode & 0x0f;
 117                 instr++;
 118
 119                 switch (instr_hi) {
 120                 case 0x20:
 121                 case 0x30:
 122                         /*
 123                          * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
 124                          * In X86_64 long mode, the CPU will signal invalid
 125                          * opcode if some of these prefixes are present so
 126                          * X86_64 will never get here anyway
 127                          */
 128                         scan_more = ((instr_lo & 7) == 0x6);
 129                         break;
 130 #ifdef CONFIG_X86_64
 131                 case 0x40:
 132                         /*
 133                          * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 134                          * Need to figure out under what instruction mode the
 135                          * instruction was issued. Could check the LDT for lm,
 136                          * but for now it's good enough to assume that long
 137                          * mode only uses well known segments or kernel.
 138                          */
 139                         scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
 140                         break;
 141 #endif
 142                 case 0x60:
 143                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 144                         scan_more = (instr_lo & 0xC) == 0x4;
 145                         break;
 146                 case 0xF0:
 147                         /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 148                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 149                         break;
 150                 case 0x00:
 151                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 152                         scan_more = 0;
 153
 154                         if (probe_kernel_address(instr, opcode))
 155                                 break;
 156                         prefetch = (instr_lo == 0xF) &&
 157                                 (opcode == 0x0D || opcode == 0x18);
 158                         break;
 159                 default:
 160                         scan_more = 0;
 161                         break;
 162                 }
 163         }
 164         return prefetch;
 165 }
 166
 167 static void force_sig_info_fault(int si_signo, int si_code,
 168         unsigned long address, struct task_struct *tsk)
 169 {
 170         siginfo_t info;
 171
 172         info.si_signo = si_signo;
 173         info.si_errno = 0;
 174         info.si_code = si_code;
 175         info.si_addr = (void __user *)address;
 176         force_sig_info(si_signo, &info, tsk);
 177 }
 178
 179 #ifdef CONFIG_X86_64
 180 static int bad_address(void *p)
 181 {
 182         unsigned long dummy;
 183         return probe_kernel_address((unsigned long *)p, dummy);
 184 }
 185 #endif
 186
 187 static void dump_pagetable(unsigned long address)
 188 {
 189 #ifdef CONFIG_X86_32
 190         __typeof__(pte_val(__pte(0))) page;
 191
 192         page = read_cr3();
 193         page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 194 #ifdef CONFIG_X86_PAE
 195         printk("*pdpt = %016Lx ", page);
 196         if ((page >> PAGE_SHIFT) < max_low_pfn
 197             && page & _PAGE_PRESENT) {
 198                 page &= PAGE_MASK;
 199                 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 200                                                          & (PTRS_PER_PMD - 1)];
 201                 printk(KERN_CONT "*pde = %016Lx ", page);
 202                 page &= ~_PAGE_NX;
 203         }
 204 #else
 205         printk("*pde = %08lx ", page);
 206 #endif
 207
 208         /*
 209          * We must not directly access the pte in the highpte
 210          * case if the page table is located in highmem.
 211          * And let's rather not kmap-atomic the pte, just in case
 212          * it's allocated already.
 213          */
 214         if ((page >> PAGE_SHIFT) < max_low_pfn
 215             && (page & _PAGE_PRESENT)
 216             && !(page & _PAGE_PSE)) {
 217                 page &= PAGE_MASK;
 218                 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 219                                                          & (PTRS_PER_PTE - 1)];
 220                 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 221         }
 222
 223         printk("\n");
 224 #else /* CONFIG_X86_64 */
 225         pgd_t *pgd;
 226         pud_t *pud;
 227         pmd_t *pmd;
 228         pte_t *pte;
 229
 230         pgd = (pgd_t *)read_cr3();
 231
 232         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 233         pgd += pgd_index(address);
 234         if (bad_address(pgd)) goto bad;
 235         printk("PGD %lx ", pgd_val(*pgd));
 236         if (!pgd_present(*pgd)) goto ret;
 237
 238         pud = pud_offset(pgd, address);
 239         if (bad_address(pud)) goto bad;
 240         printk("PUD %lx ", pud_val(*pud));
 241         if (!pud_present(*pud) || pud_large(*pud))
 242                 goto ret;
 243
 244         pmd = pmd_offset(pud, address);
 245         if (bad_address(pmd)) goto bad;
 246         printk("PMD %lx ", pmd_val(*pmd));
 247         if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
 248
 249         pte = pte_offset_kernel(pmd, address);
 250         if (bad_address(pte)) goto bad;
 251         printk("PTE %lx", pte_val(*pte));
 252 ret:
 253         printk("\n");
 254         return;
 255 bad:
 256         printk("BAD\n");
 257 #endif
 258 }
 259
 260 #ifdef CONFIG_X86_32
 261 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 262 {
 263         unsigned index = pgd_index(address);
 264         pgd_t *pgd_k;
 265         pud_t *pud, *pud_k;
 266         pmd_t *pmd, *pmd_k;
 267
 268         pgd += index;
 269         pgd_k = init_mm.pgd + index;
 270
 271         if (!pgd_present(*pgd_k))
 272                 return NULL;
 273
 274         /*
 275          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 276          * and redundant with the set_pmd() on non-PAE. As would
 277          * set_pud.
 278          */
 279
 280         pud = pud_offset(pgd, address);
 281         pud_k = pud_offset(pgd_k, address);
 282         if (!pud_present(*pud_k))
 283                 return NULL;
 284
 285         pmd = pmd_offset(pud, address);
 286         pmd_k = pmd_offset(pud_k, address);
 287         if (!pmd_present(*pmd_k))
 288                 return NULL;
 289         if (!pmd_present(*pmd)) {
 290                 set_pmd(pmd, *pmd_k);
 291                 arch_flush_lazy_mmu_mode();
 292         } else
 293                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 294         return pmd_k;
 295 }
 296 #endif
 297
 298 #ifdef CONFIG_X86_64
 299 static const char errata93_warning[] =
 300 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 301 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 302 KERN_ERR "******* Please consider a BIOS update.\n"
 303 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 304 #endif
 305
 306 /* Workaround for K8 erratum #93 & buggy BIOS.
 307    BIOS SMM functions are required to use a specific workaround
 308    to avoid corruption of the 64bit RIP register on C stepping K8.
 309    A lot of BIOS that didn't get tested properly miss this.
 310    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 311    Try to work around it here.
 312    Note we only handle faults in kernel here.
 313    Does nothing for X86_32
 314  */
 315 static int is_errata93(struct pt_regs *regs, unsigned long address)
 316 {
 317 #ifdef CONFIG_X86_64
 318         static int warned;
 319         if (address != regs->ip)
 320                 return 0;
 321         if ((address >> 32) != 0)
 322                 return 0;
 323         address |= 0xffffffffUL << 32;
 324         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 325             (address >= MODULES_VADDR && address <= MODULES_END)) {
 326                 if (!warned) {
 327                         printk(errata93_warning);
 328                         warned = 1;
 329                 }
 330                 regs->ip = address;
 331                 return 1;
 332         }
 333 #endif
 334         return 0;
 335 }
 336
 337 /*
 338  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
 339  * addresses >4GB.  We catch this in the page fault handler because these
 340  * addresses are not reachable. Just detect this case and return.  Any code
 341  * segment in LDT is compatibility mode.
 342  */
 343 static int is_errata100(struct pt_regs *regs, unsigned long address)
 344 {
 345 #ifdef CONFIG_X86_64
 346         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
 347             (address >> 32))
 348                 return 1;
 349 #endif
 350         return 0;
 351 }
 352
 353 void do_invalid_op(struct pt_regs *, unsigned long);
 354
 355 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 356 {
 357 #ifdef CONFIG_X86_F00F_BUG
 358         unsigned long nr;
 359         /*
 360          * Pentium F0 0F C7 C8 bug workaround.
 361          */
 362         if (boot_cpu_data.f00f_bug) {
 363                 nr = (address - idt_descr.address) >> 3;
 364
 365                 if (nr == 6) {
 366                         do_invalid_op(regs, 0);
 367                         return 1;
 368                 }
 369         }
 370 #endif
 371         return 0;
 372 }
 373
 374 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 375                             unsigned long address)
 376 {
 377 #ifdef CONFIG_X86_32
 378         if (!oops_may_print())
 379                 return;
 380 #endif
 381
 382 #ifdef CONFIG_X86_PAE
 383         if (error_code & PF_INSTR) {
 384                 unsigned int level;
 385                 pte_t *pte = lookup_address(address, &level);
 386
 387                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 388                         printk(KERN_CRIT "kernel tried to execute "
 389                                 "NX-protected page - exploit attempt? "
 390                                 "(uid: %d)\n", current->uid);
 391         }
 392 #endif
 393
 394         printk(KERN_ALERT "BUG: unable to handle kernel ");
 395         if (address < PAGE_SIZE)
 396                 printk(KERN_CONT "NULL pointer dereference");
 397         else
 398                 printk(KERN_CONT "paging request");
 399 #ifdef CONFIG_X86_32
 400         printk(KERN_CONT " at %08lx\n", address);
 401 #else
 402         printk(KERN_CONT " at %016lx\n", address);
 403 #endif
 404         printk(KERN_ALERT "IP:");
 405         printk_address(regs->ip, 1);
 406         dump_pagetable(address);
 407 }
 408
 409 #ifdef CONFIG_X86_64
 410 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
 411                                  unsigned long error_code)
 412 {
 413         unsigned long flags = oops_begin();
 414         struct task_struct *tsk;
 415
 416         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 417                current->comm, address);
 418         dump_pagetable(address);
 419         tsk = current;
 420         tsk->thread.cr2 = address;
 421         tsk->thread.trap_no = 14;
 422         tsk->thread.error_code = error_code;
 423         if (__die("Bad pagetable", regs, error_code))
 424                 regs = NULL;
 425         oops_end(flags, regs, SIGKILL);
 426 }
 427 #endif
 428
 429 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 430 {
 431         if ((error_code & PF_WRITE) && !pte_write(*pte))
 432                 return 0;
 433         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 434                 return 0;
 435
 436         return 1;
 437 }
 438
 439 /*
 440  * Handle a spurious fault caused by a stale TLB entry.  This allows
 441  * us to lazily refresh the TLB when increasing the permissions of a
 442  * kernel page (RO -> RW or NX -> X).  Doing it eagerly is very
 443  * expensive since that implies doing a full cross-processor TLB
 444  * flush, even if no stale TLB entries exist on other processors.
 445  * There are no security implications to leaving a stale TLB when
 446  * increasing the permissions on a page.
 447  */
 448 static int spurious_fault(unsigned long address,
 449                           unsigned long error_code)
 450 {
 451         pgd_t *pgd;
 452         pud_t *pud;
 453         pmd_t *pmd;
 454         pte_t *pte;
 455
 456         /* Reserved-bit violation or user access to kernel space? */
 457         if (error_code & (PF_USER | PF_RSVD))
 458                 return 0;
 459
 460         pgd = init_mm.pgd + pgd_index(address);
 461         if (!pgd_present(*pgd))
 462                 return 0;
 463
 464         pud = pud_offset(pgd, address);
 465         if (!pud_present(*pud))
 466                 return 0;
 467
 468         if (pud_large(*pud))
 469                 return spurious_fault_check(error_code, (pte_t *) pud);
 470
 471         pmd = pmd_offset(pud, address);
 472         if (!pmd_present(*pmd))
 473                 return 0;
 474
 475         if (pmd_large(*pmd))
 476                 return spurious_fault_check(error_code, (pte_t *) pmd);
 477
 478         pte = pte_offset_kernel(pmd, address);
 479         if (!pte_present(*pte))
 480                 return 0;
 481
 482         return spurious_fault_check(error_code, pte);
 483 }
 484
 485 /*
 486  * X86_32
 487  * Handle a fault on the vmalloc or module mapping area
 488  *
 489  * X86_64
 490  * Handle a fault on the vmalloc area
 491  *
 492  * This assumes no large pages in there.
 493  */
 494 static int vmalloc_fault(unsigned long address)
 495 {
 496 #ifdef CONFIG_X86_32
 497         unsigned long pgd_paddr;
 498         pmd_t *pmd_k;
 499         pte_t *pte_k;
 500         /*
 501          * Synchronize this task's top level page-table
 502          * with the 'reference' page table.
 503          *
 504          * Do _not_ use "current" here. We might be inside
 505          * an interrupt in the middle of a task switch..
 506          */
 507         pgd_paddr = read_cr3();
 508         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 509         if (!pmd_k)
 510                 return -1;
 511         pte_k = pte_offset_kernel(pmd_k, address);
 512         if (!pte_present(*pte_k))
 513                 return -1;
 514         return 0;
 515 #else
 516         pgd_t *pgd, *pgd_ref;
 517         pud_t *pud, *pud_ref;
 518         pmd_t *pmd, *pmd_ref;
 519         pte_t *pte, *pte_ref;
 520
 521         /* Make sure we are in vmalloc area */
 522         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 523                 return -1;
 524
 525         /* Copy kernel mappings over when needed. This can also
 526            happen within a race in page table update. In the later
 527            case just flush. */
 528
 529         pgd = pgd_offset(current->mm ?: &init_mm, address);
 530         pgd_ref = pgd_offset_k(address);
 531         if (pgd_none(*pgd_ref))
 532                 return -1;
 533         if (pgd_none(*pgd))
 534                 set_pgd(pgd, *pgd_ref);
 535         else
 536                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 537
 538         /* Below here mismatches are bugs because these lower tables
 539            are shared */
 540
 541         pud = pud_offset(pgd, address);
 542         pud_ref = pud_offset(pgd_ref, address);
 543         if (pud_none(*pud_ref))
 544                 return -1;
 545         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 546                 BUG();
 547         pmd = pmd_offset(pud, address);
 548         pmd_ref = pmd_offset(pud_ref, address);
 549         if (pmd_none(*pmd_ref))
 550                 return -1;
 551         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 552                 BUG();
 553         pte_ref = pte_offset_kernel(pmd_ref, address);
 554         if (!pte_present(*pte_ref))
 555                 return -1;
 556         pte = pte_offset_kernel(pmd, address);
 557         /* Don't use pte_page here, because the mappings can point
 558            outside mem_map, and the NUMA hash lookup cannot handle
 559            that. */
 560         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 561                 BUG();
 562         return 0;
 563 #endif
 564 }
 565
 566 int show_unhandled_signals = 1;
 567
 568 /*
 569  * This routine handles page faults.  It determines the address,
 570  * and the problem, and then passes it off to one of the appropriate
 571  * routines.
 572  */
 573 #ifdef CONFIG_X86_64
 574 asmlinkage
 575 #endif
 576 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
 577 {
 578         struct task_struct *tsk;
 579         struct mm_struct *mm;
 580         struct vm_area_struct *vma;
 581         unsigned long address;
 582         int write, si_code;
 583         int fault;
 584 #ifdef CONFIG_X86_64
 585         unsigned long flags;
 586 #endif
 587
 588         /*
 589          * We can fault from pretty much anywhere, with unknown IRQ state.
 590          */
 591         trace_hardirqs_fixup();
 592
 593         tsk = current;
 594         mm = tsk->mm;
 595         prefetchw(&mm->mmap_sem);
 596
 597         /* get the address */
 598         address = read_cr2();
 599
 600         si_code = SEGV_MAPERR;
 601
 602         if (notify_page_fault(regs))
 603                 return;
 604
 605         /*
 606          * We fault-in kernel-space virtual memory on-demand. The
 607          * 'reference' page table is init_mm.pgd.
 608          *
 609          * NOTE! We MUST NOT take any locks for this case. We may
 610          * be in an interrupt or a critical region, and should
 611          * only copy the information from the master page table,
 612          * nothing more.
 613          *
 614          * This verifies that the fault happens in kernel space
 615          * (error_code & 4) == 0, and that the fault was not a
 616          * protection error (error_code & 9) == 0.
 617          */
 618 #ifdef CONFIG_X86_32
 619         if (unlikely(address >= TASK_SIZE)) {
 620 #else
 621         if (unlikely(address >= TASK_SIZE64)) {
 622 #endif
 623                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 624                     vmalloc_fault(address) >= 0)
 625                         return;
 626
 627                 /* Can handle a stale RO->RW TLB */
 628                 if (spurious_fault(address, error_code))
 629                         return;
 630
 631                 /*
 632                  * Don't take the mm semaphore here. If we fixup a prefetch
 633                  * fault we could otherwise deadlock.
 634                  */
 635                 goto bad_area_nosemaphore;
 636         }
 637
 638
 639 #ifdef CONFIG_X86_32
 640         /* It's safe to allow irq's after cr2 has been saved and the vmalloc
 641            fault has been handled. */
 642         if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
 643                 local_irq_enable();
 644
 645         /*
 646          * If we're in an interrupt, have no user context or are running in an
 647          * atomic region then we must not take the fault.
 648          */
 649         if (in_atomic() || !mm)
 650                 goto bad_area_nosemaphore;
 651 #else /* CONFIG_X86_64 */
 652         if (likely(regs->flags & X86_EFLAGS_IF))
 653                 local_irq_enable();
 654
 655         if (unlikely(error_code & PF_RSVD))
 656                 pgtable_bad(address, regs, error_code);
 657
 658         /*
 659          * If we're in an interrupt, have no user context or are running in an
 660          * atomic region then we must not take the fault.
 661          */
 662         if (unlikely(in_atomic() || !mm))
 663                 goto bad_area_nosemaphore;
 664
 665         /*
 666          * User-mode registers count as a user access even for any
 667          * potential system fault or CPU buglet.
 668          */
 669         if (user_mode_vm(regs))
 670                 error_code |= PF_USER;
 671 again:
 672 #endif
 673         /* When running in the kernel we expect faults to occur only to
 674          * addresses in user space.  All other faults represent errors in the
 675          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 676          * erroneous fault occurring in a code path which already holds mmap_sem
 677          * we will deadlock attempting to validate the fault against the
 678          * address space.  Luckily the kernel only validly references user
 679          * space from well defined areas of code, which are listed in the
 680          * exceptions table.
 681          *
 682          * As the vast majority of faults will be valid we will only perform
 683          * the source reference check when there is a possibility of a deadlock.
 684          * Attempt to lock the address space, if we cannot we then validate the
 685          * source.  If this is invalid we can skip the address space check,
 686          * thus avoiding the deadlock.
 687          */
 688         if (!down_read_trylock(&mm->mmap_sem)) {
 689                 if ((error_code & PF_USER) == 0 &&
 690                     !search_exception_tables(regs->ip))
 691                         goto bad_area_nosemaphore;
 692                 down_read(&mm->mmap_sem);
 693         }
 694
 695         vma = find_vma(mm, address);
 696         if (!vma)
 697                 goto bad_area;
 698         if (vma->vm_start <= address)
 699                 goto good_area;
 700         if (!(vma->vm_flags & VM_GROWSDOWN))
 701                 goto bad_area;
 702         if (error_code & PF_USER) {
 703                 /*
 704                  * Accessing the stack below %sp is always a bug.
 705                  * The large cushion allows instructions like enter
 706                  * and pusha to work.  ("enter $65535,$31" pushes
 707                  * 32 pointers and then decrements %sp by 65535.)
 708                  */
 709                 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
 710                         goto bad_area;
 711         }
 712         if (expand_stack(vma, address))
 713                 goto bad_area;
 714 /*
 715  * Ok, we have a good vm_area for this memory access, so
 716  * we can handle it..
 717  */
 718 good_area:
 719         si_code = SEGV_ACCERR;
 720         write = 0;
 721         switch (error_code & (PF_PROT|PF_WRITE)) {
 722         default:        /* 3: write, present */
 723                 /* fall through */
 724         case PF_WRITE:          /* write, not present */
 725                 if (!(vma->vm_flags & VM_WRITE))
 726                         goto bad_area;
 727                 write++;
 728                 break;
 729         case PF_PROT:           /* read, present */
 730                 goto bad_area;
 731         case 0:                 /* read, not present */
 732                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
 733                         goto bad_area;
 734         }
 735
 736 #ifdef CONFIG_X86_32
 737 survive:
 738 #endif
 739         /*
 740          * If for any reason at all we couldn't handle the fault,
 741          * make sure we exit gracefully rather than endlessly redo
 742          * the fault.
 743          */
 744         fault = handle_mm_fault(mm, vma, address, write);
 745         if (unlikely(fault & VM_FAULT_ERROR)) {
 746                 if (fault & VM_FAULT_OOM)
 747                         goto out_of_memory;
 748                 else if (fault & VM_FAULT_SIGBUS)
 749                         goto do_sigbus;
 750                 BUG();
 751         }
 752         if (fault & VM_FAULT_MAJOR)
 753                 tsk->maj_flt++;
 754         else
 755                 tsk->min_flt++;
 756
 757 #ifdef CONFIG_X86_32
 758         /*
 759          * Did it hit the DOS screen memory VA from vm86 mode?
 760          */
 761         if (v8086_mode(regs)) {
 762                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 763                 if (bit < 32)
 764                         tsk->thread.screen_bitmap |= 1 << bit;
 765         }
 766 #endif
 767         up_read(&mm->mmap_sem);
 768         return;
 769
 770 /*
 771  * Something tried to access memory that isn't in our memory map..
 772  * Fix it, but check if it's kernel or user first..
 773  */
 774 bad_area:
 775         up_read(&mm->mmap_sem);
 776
 777 bad_area_nosemaphore:
 778         /* User mode accesses just cause a SIGSEGV */
 779         if (error_code & PF_USER) {
 780                 /*
 781                  * It's possible to have interrupts off here.
 782                  */
 783                 local_irq_enable();
 784
 785                 /*
 786                  * Valid to do another page fault here because this one came
 787                  * from user space.
 788                  */
 789                 if (is_prefetch(regs, address, error_code))
 790                         return;
 791
 792                 if (is_errata100(regs, address))
 793                         return;
 794
 795                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 796                     printk_ratelimit()) {
 797                         printk(
 798 #ifdef CONFIG_X86_32
 799                         "%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx",
 800 #else
 801                         "%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx",
 802 #endif
 803                         task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 804                         tsk->comm, task_pid_nr(tsk), address, regs->ip,
 805                         regs->sp, error_code);
 806                         print_vma_addr(" in ", regs->ip);
 807                         printk("\n");
 808                 }
 809
 810                 tsk->thread.cr2 = address;
 811                 /* Kernel addresses are always protection faults */
 812                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 813                 tsk->thread.trap_no = 14;
 814                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 815                 return;
 816         }
 817
 818         if (is_f00f_bug(regs, address))
 819                 return;
 820
 821 no_context:
 822         /* Are we prepared to handle this kernel fault?  */
 823         if (fixup_exception(regs))
 824                 return;
 825
 826         /*
 827          * X86_32
 828          * Valid to do another page fault here, because if this fault
 829          * had been triggered by is_prefetch fixup_exception would have
 830          * handled it.
 831          *
 832          * X86_64
 833          * Hall of shame of CPU/BIOS bugs.
 834          */
 835         if (is_prefetch(regs, address, error_code))
 836                 return;
 837
 838         if (is_errata93(regs, address))
 839                 return;
 840
 841 /*
 842  * Oops. The kernel tried to access some bad page. We'll have to
 843  * terminate things with extreme prejudice.
 844  */
 845 #ifdef CONFIG_X86_32
 846         bust_spinlocks(1);
 847 #else
 848         flags = oops_begin();
 849 #endif
 850
 851         show_fault_oops(regs, error_code, address);
 852
 853         tsk->thread.cr2 = address;
 854         tsk->thread.trap_no = 14;
 855         tsk->thread.error_code = error_code;
 856
 857 #ifdef CONFIG_X86_32
 858         die("Oops", regs, error_code);
 859         bust_spinlocks(0);
 860         do_exit(SIGKILL);
 861 #else
 862         if (__die("Oops", regs, error_code))
 863                 regs = NULL;
 864         /* Executive summary in case the body of the oops scrolled away */
 865         printk(KERN_EMERG "CR2: %016lx\n", address);
 866         oops_end(flags, regs, SIGKILL);
 867 #endif
 868
 869 /*
 870  * We ran out of memory, or some other thing happened to us that made
 871  * us unable to handle the page fault gracefully.
 872  */
 873 out_of_memory:
 874         up_read(&mm->mmap_sem);
 875         if (is_global_init(tsk)) {
 876                 yield();
 877 #ifdef CONFIG_X86_32
 878                 down_read(&mm->mmap_sem);
 879                 goto survive;
 880 #else
 881                 goto again;
 882 #endif
 883         }
 884
 885         printk("VM: killing process %s\n", tsk->comm);
 886         if (error_code & PF_USER)
 887                 do_group_exit(SIGKILL);
 888         goto no_context;
 889
 890 do_sigbus:
 891         up_read(&mm->mmap_sem);
 892
 893         /* Kernel mode? Handle exceptions or die */
 894         if (!(error_code & PF_USER))
 895                 goto no_context;
 896 #ifdef CONFIG_X86_32
 897         /* User space => ok to do another page fault */
 898         if (is_prefetch(regs, address, error_code))
 899                 return;
 900 #endif
 901         tsk->thread.cr2 = address;
 902         tsk->thread.error_code = error_code;
 903         tsk->thread.trap_no = 14;
 904         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 905 }
 906
 907 DEFINE_SPINLOCK(pgd_lock);
 908 LIST_HEAD(pgd_list);
 909
 910 void vmalloc_sync_all(void)
 911 {
 912 #ifdef CONFIG_X86_32
 913         /*
 914          * Note that races in the updates of insync and start aren't
 915          * problematic: insync can only get set bits added, and updates to
 916          * start are only improving performance (without affecting correctness
 917          * if undone).
 918          */
 919         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 920         static unsigned long start = TASK_SIZE;
 921         unsigned long address;
 922
 923         if (SHARED_KERNEL_PMD)
 924                 return;
 925
 926         BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
 927         for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
 928                 if (!test_bit(pgd_index(address), insync)) {
 929                         unsigned long flags;
 930                         struct page *page;
 931
 932                         spin_lock_irqsave(&pgd_lock, flags);
 933                         list_for_each_entry(page, &pgd_list, lru) {
 934                                 if (!vmalloc_sync_one(page_address(page),
 935                                                       address))
 936                                         break;
 937                         }
 938                         spin_unlock_irqrestore(&pgd_lock, flags);
 939                         if (!page)
 940                                 set_bit(pgd_index(address), insync);
 941                 }
 942                 if (address == start && test_bit(pgd_index(address), insync))
 943                         start = address + PGDIR_SIZE;
 944         }
 945 #else /* CONFIG_X86_64 */
 946         /*
 947          * Note that races in the updates of insync and start aren't
 948          * problematic: insync can only get set bits added, and updates to
 949          * start are only improving performance (without affecting correctness
 950          * if undone).
 951          */
 952         static DECLARE_BITMAP(insync, PTRS_PER_PGD);
 953         static unsigned long start = VMALLOC_START & PGDIR_MASK;
 954         unsigned long address;
 955
 956         for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
 957                 if (!test_bit(pgd_index(address), insync)) {
 958                         const pgd_t *pgd_ref = pgd_offset_k(address);
 959                         unsigned long flags;
 960                         struct page *page;
 961
 962                         if (pgd_none(*pgd_ref))
 963                                 continue;
 964                         spin_lock_irqsave(&pgd_lock, flags);
 965                         list_for_each_entry(page, &pgd_list, lru) {
 966                                 pgd_t *pgd;
 967                                 pgd = (pgd_t *)page_address(page) + pgd_index(address);
 968                                 if (pgd_none(*pgd))
 969                                         set_pgd(pgd, *pgd_ref);
 970                                 else
 971                                         BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 972                         }
 973                         spin_unlock_irqrestore(&pgd_lock, flags);
 974                         set_bit(pgd_index(address), insync);
 975                 }
 976                 if (address == start)
 977                         start = address + PGDIR_SIZE;
 978         }
 979         /* Check that there is no need to do the same for the modules area. */
 980         BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
 981         BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
 982                                 (__START_KERNEL & PGDIR_MASK)));
 983 #endif
 984 }