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 #include <linux/magic.h>
  30
  31 #include <asm/system.h>
  32 #include <asm/desc.h>
  33 #include <asm/segment.h>
  34 #include <asm/pgalloc.h>
  35 #include <asm/smp.h>
  36 #include <asm/tlbflush.h>
  37 #include <asm/proto.h>
  38 #include <asm-generic/sections.h>
  39 #include <asm/traps.h>
  40
  41 /*
  42  * Page fault error code bits
  43  *      bit 0 == 0 means no page found, 1 means protection fault
  44  *      bit 1 == 0 means read, 1 means write
  45  *      bit 2 == 0 means kernel, 1 means user-mode
  46  *      bit 3 == 1 means use of reserved bit detected
  47  *      bit 4 == 1 means fault was an instruction fetch
  48  */
  49 #define PF_PROT         (1<<0)
  50 #define PF_WRITE        (1<<1)
  51 #define PF_USER         (1<<2)
  52 #define PF_RSVD         (1<<3)
  53 #define PF_INSTR        (1<<4)
  54
  55 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
  56 {
  57 #ifdef CONFIG_MMIOTRACE
  58         if (unlikely(is_kmmio_active()))
  59                 if (kmmio_handler(regs, addr) == 1)
  60                         return -1;
  61 #endif
  62         return 0;
  63 }
  64
  65 static inline int notify_page_fault(struct pt_regs *regs)
  66 {
  67 #ifdef CONFIG_KPROBES
  68         int ret = 0;
  69
  70         /* kprobe_running() needs smp_processor_id() */
  71         if (!user_mode_vm(regs)) {
  72                 preempt_disable();
  73                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  74                         ret = 1;
  75                 preempt_enable();
  76         }
  77
  78         return ret;
  79 #else
  80         return 0;
  81 #endif
  82 }
  83
  84 /*
  85  * X86_32
  86  * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
  87  * Check that here and ignore it.
  88  *
  89  * X86_64
  90  * Sometimes the CPU reports invalid exceptions on prefetch.
  91  * Check that here and ignore it.
  92  *
  93  * Opcode checker based on code by Richard Brunner
  94  */
  95 static int is_prefetch(struct pt_regs *regs, unsigned long error_code,
  96                         unsigned long addr)
  97 {
  98         unsigned char *instr;
  99         int scan_more = 1;
 100         int prefetch = 0;
 101         unsigned char *max_instr;
 102
 103         /*
 104          * If it was a exec (instruction fetch) fault on NX page, then
 105          * do not ignore the fault:
 106          */
 107         if (error_code & PF_INSTR)
 108                 return 0;
 109
 110         instr = (unsigned char *)convert_ip_to_linear(current, regs);
 111         max_instr = instr + 15;
 112
 113         if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
 114                 return 0;
 115
 116         while (scan_more && instr < max_instr) {
 117                 unsigned char opcode;
 118                 unsigned char instr_hi;
 119                 unsigned char instr_lo;
 120
 121                 if (probe_kernel_address(instr, opcode))
 122                         break;
 123
 124                 instr_hi = opcode & 0xf0;
 125                 instr_lo = opcode & 0x0f;
 126                 instr++;
 127
 128                 switch (instr_hi) {
 129                 case 0x20:
 130                 case 0x30:
 131                         /*
 132                          * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
 133                          * In X86_64 long mode, the CPU will signal invalid
 134                          * opcode if some of these prefixes are present so
 135                          * X86_64 will never get here anyway
 136                          */
 137                         scan_more = ((instr_lo & 7) == 0x6);
 138                         break;
 139 #ifdef CONFIG_X86_64
 140                 case 0x40:
 141                         /*
 142                          * In AMD64 long mode 0x40..0x4F are valid REX prefixes
 143                          * Need to figure out under what instruction mode the
 144                          * instruction was issued. Could check the LDT for lm,
 145                          * but for now it's good enough to assume that long
 146                          * mode only uses well known segments or kernel.
 147                          */
 148                         scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
 149                         break;
 150 #endif
 151                 case 0x60:
 152                         /* 0x64 thru 0x67 are valid prefixes in all modes. */
 153                         scan_more = (instr_lo & 0xC) == 0x4;
 154                         break;
 155                 case 0xF0:
 156                         /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
 157                         scan_more = !instr_lo || (instr_lo>>1) == 1;
 158                         break;
 159                 case 0x00:
 160                         /* Prefetch instruction is 0x0F0D or 0x0F18 */
 161                         scan_more = 0;
 162
 163                         if (probe_kernel_address(instr, opcode))
 164                                 break;
 165                         prefetch = (instr_lo == 0xF) &&
 166                                 (opcode == 0x0D || opcode == 0x18);
 167                         break;
 168                 default:
 169                         scan_more = 0;
 170                         break;
 171                 }
 172         }
 173         return prefetch;
 174 }
 175
 176 static void force_sig_info_fault(int si_signo, int si_code,
 177         unsigned long address, struct task_struct *tsk)
 178 {
 179         siginfo_t info;
 180
 181         info.si_signo = si_signo;
 182         info.si_errno = 0;
 183         info.si_code = si_code;
 184         info.si_addr = (void __user *)address;
 185         force_sig_info(si_signo, &info, tsk);
 186 }
 187
 188 #ifdef CONFIG_X86_64
 189 static int bad_address(void *p)
 190 {
 191         unsigned long dummy;
 192         return probe_kernel_address((unsigned long *)p, dummy);
 193 }
 194 #endif
 195
 196 static void dump_pagetable(unsigned long address)
 197 {
 198 #ifdef CONFIG_X86_32
 199         __typeof__(pte_val(__pte(0))) page;
 200
 201         page = read_cr3();
 202         page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
 203 #ifdef CONFIG_X86_PAE
 204         printk("*pdpt = %016Lx ", page);
 205         if ((page >> PAGE_SHIFT) < max_low_pfn
 206             && page & _PAGE_PRESENT) {
 207                 page &= PAGE_MASK;
 208                 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
 209                                                          & (PTRS_PER_PMD - 1)];
 210                 printk(KERN_CONT "*pde = %016Lx ", page);
 211                 page &= ~_PAGE_NX;
 212         }
 213 #else
 214         printk("*pde = %08lx ", page);
 215 #endif
 216
 217         /*
 218          * We must not directly access the pte in the highpte
 219          * case if the page table is located in highmem.
 220          * And let's rather not kmap-atomic the pte, just in case
 221          * it's allocated already.
 222          */
 223         if ((page >> PAGE_SHIFT) < max_low_pfn
 224             && (page & _PAGE_PRESENT)
 225             && !(page & _PAGE_PSE)) {
 226                 page &= PAGE_MASK;
 227                 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
 228                                                          & (PTRS_PER_PTE - 1)];
 229                 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
 230         }
 231
 232         printk("\n");
 233 #else /* CONFIG_X86_64 */
 234         pgd_t *pgd;
 235         pud_t *pud;
 236         pmd_t *pmd;
 237         pte_t *pte;
 238
 239         pgd = (pgd_t *)read_cr3();
 240
 241         pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
 242         pgd += pgd_index(address);
 243         if (bad_address(pgd)) goto bad;
 244         printk("PGD %lx ", pgd_val(*pgd));
 245         if (!pgd_present(*pgd)) goto ret;
 246
 247         pud = pud_offset(pgd, address);
 248         if (bad_address(pud)) goto bad;
 249         printk("PUD %lx ", pud_val(*pud));
 250         if (!pud_present(*pud) || pud_large(*pud))
 251                 goto ret;
 252
 253         pmd = pmd_offset(pud, address);
 254         if (bad_address(pmd)) goto bad;
 255         printk("PMD %lx ", pmd_val(*pmd));
 256         if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
 257
 258         pte = pte_offset_kernel(pmd, address);
 259         if (bad_address(pte)) goto bad;
 260         printk("PTE %lx", pte_val(*pte));
 261 ret:
 262         printk("\n");
 263         return;
 264 bad:
 265         printk("BAD\n");
 266 #endif
 267 }
 268
 269 #ifdef CONFIG_X86_32
 270 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 271 {
 272         unsigned index = pgd_index(address);
 273         pgd_t *pgd_k;
 274         pud_t *pud, *pud_k;
 275         pmd_t *pmd, *pmd_k;
 276
 277         pgd += index;
 278         pgd_k = init_mm.pgd + index;
 279
 280         if (!pgd_present(*pgd_k))
 281                 return NULL;
 282
 283         /*
 284          * set_pgd(pgd, *pgd_k); here would be useless on PAE
 285          * and redundant with the set_pmd() on non-PAE. As would
 286          * set_pud.
 287          */
 288
 289         pud = pud_offset(pgd, address);
 290         pud_k = pud_offset(pgd_k, address);
 291         if (!pud_present(*pud_k))
 292                 return NULL;
 293
 294         pmd = pmd_offset(pud, address);
 295         pmd_k = pmd_offset(pud_k, address);
 296         if (!pmd_present(*pmd_k))
 297                 return NULL;
 298         if (!pmd_present(*pmd)) {
 299                 set_pmd(pmd, *pmd_k);
 300                 arch_flush_lazy_mmu_mode();
 301         } else
 302                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 303         return pmd_k;
 304 }
 305 #endif
 306
 307 #ifdef CONFIG_X86_64
 308 static const char errata93_warning[] =
 309 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
 310 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
 311 KERN_ERR "******* Please consider a BIOS update.\n"
 312 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
 313 #endif
 314
 315 /* Workaround for K8 erratum #93 & buggy BIOS.
 316    BIOS SMM functions are required to use a specific workaround
 317    to avoid corruption of the 64bit RIP register on C stepping K8.
 318    A lot of BIOS that didn't get tested properly miss this.
 319    The OS sees this as a page fault with the upper 32bits of RIP cleared.
 320    Try to work around it here.
 321    Note we only handle faults in kernel here.
 322    Does nothing for X86_32
 323  */
 324 static int is_errata93(struct pt_regs *regs, unsigned long address)
 325 {
 326 #ifdef CONFIG_X86_64
 327         static int warned;
 328         if (address != regs->ip)
 329                 return 0;
 330         if ((address >> 32) != 0)
 331                 return 0;
 332         address |= 0xffffffffUL << 32;
 333         if ((address >= (u64)_stext && address <= (u64)_etext) ||
 334             (address >= MODULES_VADDR && address <= MODULES_END)) {
 335                 if (!warned) {
 336                         printk(errata93_warning);
 337                         warned = 1;
 338                 }
 339                 regs->ip = address;
 340                 return 1;
 341         }
 342 #endif
 343         return 0;
 344 }
 345
 346 /*
 347  * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
 348  * addresses >4GB.  We catch this in the page fault handler because these
 349  * addresses are not reachable. Just detect this case and return.  Any code
 350  * segment in LDT is compatibility mode.
 351  */
 352 static int is_errata100(struct pt_regs *regs, unsigned long address)
 353 {
 354 #ifdef CONFIG_X86_64
 355         if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
 356             (address >> 32))
 357                 return 1;
 358 #endif
 359         return 0;
 360 }
 361
 362 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
 363 {
 364 #ifdef CONFIG_X86_F00F_BUG
 365         unsigned long nr;
 366         /*
 367          * Pentium F0 0F C7 C8 bug workaround.
 368          */
 369         if (boot_cpu_data.f00f_bug) {
 370                 nr = (address - idt_descr.address) >> 3;
 371
 372                 if (nr == 6) {
 373                         do_invalid_op(regs, 0);
 374                         return 1;
 375                 }
 376         }
 377 #endif
 378         return 0;
 379 }
 380
 381 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
 382                             unsigned long address)
 383 {
 384 #ifdef CONFIG_X86_32
 385         if (!oops_may_print())
 386                 return;
 387 #endif
 388
 389 #ifdef CONFIG_X86_PAE
 390         if (error_code & PF_INSTR) {
 391                 unsigned int level;
 392                 pte_t *pte = lookup_address(address, &level);
 393
 394                 if (pte && pte_present(*pte) && !pte_exec(*pte))
 395                         printk(KERN_CRIT "kernel tried to execute "
 396                                 "NX-protected page - exploit attempt? "
 397                                 "(uid: %d)\n", current_uid());
 398         }
 399 #endif
 400
 401         printk(KERN_ALERT "BUG: unable to handle kernel ");
 402         if (address < PAGE_SIZE)
 403                 printk(KERN_CONT "NULL pointer dereference");
 404         else
 405                 printk(KERN_CONT "paging request");
 406         printk(KERN_CONT " at %p\n", (void *) address);
 407         printk(KERN_ALERT "IP:");
 408         printk_address(regs->ip, 1);
 409         dump_pagetable(address);
 410 }
 411
 412 #ifdef CONFIG_X86_64
 413 static noinline void pgtable_bad(struct pt_regs *regs,
 414                          unsigned long error_code, unsigned long address)
 415 {
 416         unsigned long flags = oops_begin();
 417         int sig = SIGKILL;
 418         struct task_struct *tsk = current;
 419
 420         printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
 421                tsk->comm, address);
 422         dump_pagetable(address);
 423         tsk->thread.cr2 = address;
 424         tsk->thread.trap_no = 14;
 425         tsk->thread.error_code = error_code;
 426         if (__die("Bad pagetable", regs, error_code))
 427                 sig = 0;
 428         oops_end(flags, regs, sig);
 429 }
 430 #endif
 431
 432 static noinline void no_context(struct pt_regs *regs,
 433                         unsigned long error_code, unsigned long address)
 434 {
 435         struct task_struct *tsk = current;
 436         unsigned long *stackend;
 437
 438 #ifdef CONFIG_X86_64
 439         unsigned long flags;
 440         int sig;
 441 #endif
 442
 443         /* Are we prepared to handle this kernel fault?  */
 444         if (fixup_exception(regs))
 445                 return;
 446
 447         /*
 448          * X86_32
 449          * Valid to do another page fault here, because if this fault
 450          * had been triggered by is_prefetch fixup_exception would have
 451          * handled it.
 452          *
 453          * X86_64
 454          * Hall of shame of CPU/BIOS bugs.
 455          */
 456         if (is_prefetch(regs, error_code, address))
 457                 return;
 458
 459         if (is_errata93(regs, address))
 460                 return;
 461
 462         /*
 463          * Oops. The kernel tried to access some bad page. We'll have to
 464          * terminate things with extreme prejudice.
 465          */
 466 #ifdef CONFIG_X86_32
 467         bust_spinlocks(1);
 468 #else
 469         flags = oops_begin();
 470 #endif
 471
 472         show_fault_oops(regs, error_code, address);
 473
 474         stackend = end_of_stack(tsk);
 475         if (*stackend != STACK_END_MAGIC)
 476                 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
 477
 478         tsk->thread.cr2 = address;
 479         tsk->thread.trap_no = 14;
 480         tsk->thread.error_code = error_code;
 481
 482 #ifdef CONFIG_X86_32
 483         die("Oops", regs, error_code);
 484         bust_spinlocks(0);
 485         do_exit(SIGKILL);
 486 #else
 487         sig = SIGKILL;
 488         if (__die("Oops", regs, error_code))
 489                 sig = 0;
 490         /* Executive summary in case the body of the oops scrolled away */
 491         printk(KERN_EMERG "CR2: %016lx\n", address);
 492         oops_end(flags, regs, sig);
 493 #endif
 494 }
 495
 496 static void __bad_area_nosemaphore(struct pt_regs *regs,
 497                         unsigned long error_code, unsigned long address,
 498                         int si_code)
 499 {
 500         struct task_struct *tsk = current;
 501
 502         /* User mode accesses just cause a SIGSEGV */
 503         if (error_code & PF_USER) {
 504                 /*
 505                  * It's possible to have interrupts off here.
 506                  */
 507                 local_irq_enable();
 508
 509                 /*
 510                  * Valid to do another page fault here because this one came
 511                  * from user space.
 512                  */
 513                 if (is_prefetch(regs, error_code, address))
 514                         return;
 515
 516                 if (is_errata100(regs, address))
 517                         return;
 518
 519                 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
 520                     printk_ratelimit()) {
 521                         printk(
 522                         "%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
 523                         task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
 524                         tsk->comm, task_pid_nr(tsk), address,
 525                         (void *) regs->ip, (void *) regs->sp, error_code);
 526                         print_vma_addr(" in ", regs->ip);
 527                         printk("\n");
 528                 }
 529
 530                 tsk->thread.cr2 = address;
 531                 /* Kernel addresses are always protection faults */
 532                 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
 533                 tsk->thread.trap_no = 14;
 534                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 535                 return;
 536         }
 537
 538         if (is_f00f_bug(regs, address))
 539                 return;
 540
 541         no_context(regs, error_code, address);
 542 }
 543
 544 static noinline void bad_area_nosemaphore(struct pt_regs *regs,
 545                         unsigned long error_code, unsigned long address)
 546 {
 547         __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 548 }
 549
 550 static void __bad_area(struct pt_regs *regs,
 551                         unsigned long error_code, unsigned long address,
 552                         int si_code)
 553 {
 554         struct mm_struct *mm = current->mm;
 555
 556         /*
 557          * Something tried to access memory that isn't in our memory map..
 558          * Fix it, but check if it's kernel or user first..
 559          */
 560         up_read(&mm->mmap_sem);
 561
 562         __bad_area_nosemaphore(regs, error_code, address, si_code);
 563 }
 564
 565 static noinline void bad_area(struct pt_regs *regs,
 566                         unsigned long error_code, unsigned long address)
 567 {
 568         __bad_area(regs, error_code, address, SEGV_MAPERR);
 569 }
 570
 571 static noinline void bad_area_access_error(struct pt_regs *regs,
 572                         unsigned long error_code, unsigned long address)
 573 {
 574         __bad_area(regs, error_code, address, SEGV_ACCERR);
 575 }
 576
 577 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
 578 static void out_of_memory(struct pt_regs *regs,
 579                         unsigned long error_code, unsigned long address)
 580 {
 581         /*
 582          * We ran out of memory, call the OOM killer, and return the userspace
 583          * (which will retry the fault, or kill us if we got oom-killed).
 584          */
 585         up_read(&current->mm->mmap_sem);
 586         pagefault_out_of_memory();
 587 }
 588
 589 static void do_sigbus(struct pt_regs *regs,
 590                         unsigned long error_code, unsigned long address)
 591 {
 592         struct task_struct *tsk = current;
 593         struct mm_struct *mm = tsk->mm;
 594
 595         up_read(&mm->mmap_sem);
 596
 597         /* Kernel mode? Handle exceptions or die */
 598         if (!(error_code & PF_USER))
 599                 no_context(regs, error_code, address);
 600 #ifdef CONFIG_X86_32
 601         /* User space => ok to do another page fault */
 602         if (is_prefetch(regs, error_code, address))
 603                 return;
 604 #endif
 605         tsk->thread.cr2 = address;
 606         tsk->thread.error_code = error_code;
 607         tsk->thread.trap_no = 14;
 608         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 609 }
 610
 611 static noinline void mm_fault_error(struct pt_regs *regs,
 612                 unsigned long error_code, unsigned long address, unsigned int fault)
 613 {
 614         if (fault & VM_FAULT_OOM)
 615                 out_of_memory(regs, error_code, address);
 616         else if (fault & VM_FAULT_SIGBUS)
 617                 do_sigbus(regs, error_code, address);
 618         else
 619                 BUG();
 620 }
 621
 622 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
 623 {
 624         if ((error_code & PF_WRITE) && !pte_write(*pte))
 625                 return 0;
 626         if ((error_code & PF_INSTR) && !pte_exec(*pte))
 627                 return 0;
 628
 629         return 1;
 630 }
 631
 632 /*
 633  * Handle a spurious fault caused by a stale TLB entry.  This allows
 634  * us to lazily refresh the TLB when increasing the permissions of a
 635  * kernel page (RO -> RW or NX -> X).  Doing it eagerly is very
 636  * expensive since that implies doing a full cross-processor TLB
 637  * flush, even if no stale TLB entries exist on other processors.
 638  * There are no security implications to leaving a stale TLB when
 639  * increasing the permissions on a page.
 640  */
 641 static noinline int spurious_fault(unsigned long error_code,
 642                                 unsigned long address)
 643 {
 644         pgd_t *pgd;
 645         pud_t *pud;
 646         pmd_t *pmd;
 647         pte_t *pte;
 648
 649         /* Reserved-bit violation or user access to kernel space? */
 650         if (error_code & (PF_USER | PF_RSVD))
 651                 return 0;
 652
 653         pgd = init_mm.pgd + pgd_index(address);
 654         if (!pgd_present(*pgd))
 655                 return 0;
 656
 657         pud = pud_offset(pgd, address);
 658         if (!pud_present(*pud))
 659                 return 0;
 660
 661         if (pud_large(*pud))
 662                 return spurious_fault_check(error_code, (pte_t *) pud);
 663
 664         pmd = pmd_offset(pud, address);
 665         if (!pmd_present(*pmd))
 666                 return 0;
 667
 668         if (pmd_large(*pmd))
 669                 return spurious_fault_check(error_code, (pte_t *) pmd);
 670
 671         pte = pte_offset_kernel(pmd, address);
 672         if (!pte_present(*pte))
 673                 return 0;
 674
 675         return spurious_fault_check(error_code, pte);
 676 }
 677
 678 /*
 679  * X86_32
 680  * Handle a fault on the vmalloc or module mapping area
 681  *
 682  * X86_64
 683  * Handle a fault on the vmalloc area
 684  *
 685  * This assumes no large pages in there.
 686  */
 687 static noinline int vmalloc_fault(unsigned long address)
 688 {
 689 #ifdef CONFIG_X86_32
 690         unsigned long pgd_paddr;
 691         pmd_t *pmd_k;
 692         pte_t *pte_k;
 693
 694         /* Make sure we are in vmalloc area */
 695         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 696                 return -1;
 697
 698         /*
 699          * Synchronize this task's top level page-table
 700          * with the 'reference' page table.
 701          *
 702          * Do _not_ use "current" here. We might be inside
 703          * an interrupt in the middle of a task switch..
 704          */
 705         pgd_paddr = read_cr3();
 706         pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
 707         if (!pmd_k)
 708                 return -1;
 709         pte_k = pte_offset_kernel(pmd_k, address);
 710         if (!pte_present(*pte_k))
 711                 return -1;
 712         return 0;
 713 #else
 714         pgd_t *pgd, *pgd_ref;
 715         pud_t *pud, *pud_ref;
 716         pmd_t *pmd, *pmd_ref;
 717         pte_t *pte, *pte_ref;
 718
 719         /* Make sure we are in vmalloc area */
 720         if (!(address >= VMALLOC_START && address < VMALLOC_END))
 721                 return -1;
 722
 723         /* Copy kernel mappings over when needed. This can also
 724            happen within a race in page table update. In the later
 725            case just flush. */
 726
 727         pgd = pgd_offset(current->active_mm, address);
 728         pgd_ref = pgd_offset_k(address);
 729         if (pgd_none(*pgd_ref))
 730                 return -1;
 731         if (pgd_none(*pgd))
 732                 set_pgd(pgd, *pgd_ref);
 733         else
 734                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
 735
 736         /* Below here mismatches are bugs because these lower tables
 737            are shared */
 738
 739         pud = pud_offset(pgd, address);
 740         pud_ref = pud_offset(pgd_ref, address);
 741         if (pud_none(*pud_ref))
 742                 return -1;
 743         if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
 744                 BUG();
 745         pmd = pmd_offset(pud, address);
 746         pmd_ref = pmd_offset(pud_ref, address);
 747         if (pmd_none(*pmd_ref))
 748                 return -1;
 749         if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
 750                 BUG();
 751         pte_ref = pte_offset_kernel(pmd_ref, address);
 752         if (!pte_present(*pte_ref))
 753                 return -1;
 754         pte = pte_offset_kernel(pmd, address);
 755         /* Don't use pte_page here, because the mappings can point
 756            outside mem_map, and the NUMA hash lookup cannot handle
 757            that. */
 758         if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
 759                 BUG();
 760         return 0;
 761 #endif
 762 }
 763
 764 int show_unhandled_signals = 1;
 765
 766 static inline int access_error(unsigned long error_code, int write,
 767                                 struct vm_area_struct *vma)
 768 {
 769         if (write) {
 770                 /* write, present and write, not present */
 771                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 772                         return 1;
 773         } else if (unlikely(error_code & PF_PROT)) {
 774                 /* read, present */
 775                 return 1;
 776         } else {
 777                 /* read, not present */
 778                 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 779                         return 1;
 780         }
 781
 782         return 0;
 783 }
 784
 785 /*
 786  * This routine handles page faults.  It determines the address,
 787  * and the problem, and then passes it off to one of the appropriate
 788  * routines.
 789  */
 790 #ifdef CONFIG_X86_64
 791 asmlinkage
 792 #endif
 793 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
 794 {
 795         unsigned long address;
 796         struct task_struct *tsk;
 797         struct mm_struct *mm;
 798         struct vm_area_struct *vma;
 799         int write;
 800         int fault;
 801
 802         tsk = current;
 803         mm = tsk->mm;
 804         prefetchw(&mm->mmap_sem);
 805
 806         /* get the address */
 807         address = read_cr2();
 808
 809         if (unlikely(notify_page_fault(regs)))
 810                 return;
 811         if (unlikely(kmmio_fault(regs, address)))
 812                 return;
 813
 814         /*
 815          * We fault-in kernel-space virtual memory on-demand. The
 816          * 'reference' page table is init_mm.pgd.
 817          *
 818          * NOTE! We MUST NOT take any locks for this case. We may
 819          * be in an interrupt or a critical region, and should
 820          * only copy the information from the master page table,
 821          * nothing more.
 822          *
 823          * This verifies that the fault happens in kernel space
 824          * (error_code & 4) == 0, and that the fault was not a
 825          * protection error (error_code & 9) == 0.
 826          */
 827 #ifdef CONFIG_X86_32
 828         if (unlikely(address >= TASK_SIZE)) {
 829 #else
 830         if (unlikely(address >= TASK_SIZE64)) {
 831 #endif
 832                 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
 833                     vmalloc_fault(address) >= 0)
 834                         return;
 835
 836                 /* Can handle a stale RO->RW TLB */
 837                 if (spurious_fault(error_code, address))
 838                         return;
 839
 840                 /*
 841                  * Don't take the mm semaphore here. If we fixup a prefetch
 842                  * fault we could otherwise deadlock.
 843                  */
 844                 bad_area_nosemaphore(regs, error_code, address);
 845                 return;
 846         }
 847
 848         /*
 849          * It's safe to allow irq's after cr2 has been saved and the
 850          * vmalloc fault has been handled.
 851          *
 852          * User-mode registers count as a user access even for any
 853          * potential system fault or CPU buglet.
 854          */
 855         if (user_mode_vm(regs)) {
 856                 local_irq_enable();
 857                 error_code |= PF_USER;
 858         } else if (regs->flags & X86_EFLAGS_IF)
 859                 local_irq_enable();
 860
 861 #ifdef CONFIG_X86_64
 862         if (unlikely(error_code & PF_RSVD))
 863                 pgtable_bad(regs, error_code, address);
 864 #endif
 865
 866         /*
 867          * If we're in an interrupt, have no user context or are running in an
 868          * atomic region then we must not take the fault.
 869          */
 870         if (unlikely(in_atomic() || !mm)) {
 871                 bad_area_nosemaphore(regs, error_code, address);
 872                 return;
 873         }
 874
 875         /*
 876          * When running in the kernel we expect faults to occur only to
 877          * addresses in user space.  All other faults represent errors in the
 878          * kernel and should generate an OOPS.  Unfortunately, in the case of an
 879          * erroneous fault occurring in a code path which already holds mmap_sem
 880          * we will deadlock attempting to validate the fault against the
 881          * address space.  Luckily the kernel only validly references user
 882          * space from well defined areas of code, which are listed in the
 883          * exceptions table.
 884          *
 885          * As the vast majority of faults will be valid we will only perform
 886          * the source reference check when there is a possibility of a deadlock.
 887          * Attempt to lock the address space, if we cannot we then validate the
 888          * source.  If this is invalid we can skip the address space check,
 889          * thus avoiding the deadlock.
 890          */
 891         if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
 892                 if ((error_code & PF_USER) == 0 &&
 893                     !search_exception_tables(regs->ip)) {
 894                         bad_area_nosemaphore(regs, error_code, address);
 895                         return;
 896                 }
 897                 down_read(&mm->mmap_sem);
 898         }
 899
 900         vma = find_vma(mm, address);
 901         if (unlikely(!vma)) {
 902                 bad_area(regs, error_code, address);
 903                 return;
 904         }
 905         if (likely(vma->vm_start <= address))
 906                 goto good_area;
 907         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
 908                 bad_area(regs, error_code, address);
 909                 return;
 910         }
 911         if (error_code & PF_USER) {
 912                 /*
 913                  * Accessing the stack below %sp is always a bug.
 914                  * The large cushion allows instructions like enter
 915                  * and pusha to work.  ("enter $65535,$31" pushes
 916                  * 32 pointers and then decrements %sp by 65535.)
 917                  */
 918                 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
 919                         bad_area(regs, error_code, address);
 920                         return;
 921                 }
 922         }
 923         if (unlikely(expand_stack(vma, address))) {
 924                 bad_area(regs, error_code, address);
 925                 return;
 926         }
 927
 928         /*
 929          * Ok, we have a good vm_area for this memory access, so
 930          * we can handle it..
 931          */
 932 good_area:
 933         write = error_code & PF_WRITE;
 934         if (unlikely(access_error(error_code, write, vma))) {
 935                 bad_area_access_error(regs, error_code, address);
 936                 return;
 937         }
 938
 939         /*
 940          * If for any reason at all we couldn't handle the fault,
 941          * make sure we exit gracefully rather than endlessly redo
 942          * the fault.
 943          */
 944         fault = handle_mm_fault(mm, vma, address, write);
 945         if (unlikely(fault & VM_FAULT_ERROR)) {
 946                 mm_fault_error(regs, error_code, address, fault);
 947                 return;
 948         }
 949         if (fault & VM_FAULT_MAJOR)
 950                 tsk->maj_flt++;
 951         else
 952                 tsk->min_flt++;
 953
 954 #ifdef CONFIG_X86_32
 955         /*
 956          * Did it hit the DOS screen memory VA from vm86 mode?
 957          */
 958         if (v8086_mode(regs)) {
 959                 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
 960                 if (bit < 32)
 961                         tsk->thread.screen_bitmap |= 1 << bit;
 962         }
 963 #endif
 964         up_read(&mm->mmap_sem);
 965 }
 966
 967 DEFINE_SPINLOCK(pgd_lock);
 968 LIST_HEAD(pgd_list);
 969
 970 void vmalloc_sync_all(void)
 971 {
 972         unsigned long address;
 973
 974 #ifdef CONFIG_X86_32
 975         if (SHARED_KERNEL_PMD)
 976                 return;
 977
 978         for (address = VMALLOC_START & PMD_MASK;
 979              address >= TASK_SIZE && address < FIXADDR_TOP;
 980              address += PMD_SIZE) {
 981                 unsigned long flags;
 982                 struct page *page;
 983
 984                 spin_lock_irqsave(&pgd_lock, flags);
 985                 list_for_each_entry(page, &pgd_list, lru) {
 986                         if (!vmalloc_sync_one(page_address(page),
 987                                               address))
 988                                 break;
 989                 }
 990                 spin_unlock_irqrestore(&pgd_lock, flags);
 991         }
 992 #else /* CONFIG_X86_64 */
 993         for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
 994              address += PGDIR_SIZE) {
 995                 const pgd_t *pgd_ref = pgd_offset_k(address);
 996                 unsigned long flags;
 997                 struct page *page;
 998
 999                 if (pgd_none(*pgd_ref))
1000                         continue;
1001                 spin_lock_irqsave(&pgd_lock, flags);
1002                 list_for_each_entry(page, &pgd_list, lru) {
1003                         pgd_t *pgd;
1004                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
1005                         if (pgd_none(*pgd))
1006                                 set_pgd(pgd, *pgd_ref);
1007                         else
1008                                 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
1009                 }
1010                 spin_unlock_irqrestore(&pgd_lock, flags);
1011         }
1012 #endif
1013 }