2 * Copyright (C) 1995 Linus Torvalds
3 * Copyright (C) 2001,2002 Andi Kleen, SuSE Labs.
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>
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/vmalloc.h>
22 #include <linux/module.h>
23 #include <linux/kprobes.h>
24 #include <linux/uaccess.h>
25 #include <linux/kdebug.h>
27 #include <asm/system.h>
28 #include <asm/pgalloc.h>
30 #include <asm/tlbflush.h>
31 #include <asm/proto.h>
32 #include <asm-generic/sections.h>
35 * Page fault error code bits
36 * bit 0 == 0 means no page found, 1 means protection fault
37 * bit 1 == 0 means read, 1 means write
38 * bit 2 == 0 means kernel, 1 means user-mode
39 * bit 3 == 1 means use of reserved bit detected
40 * bit 4 == 1 means fault was an instruction fetch
42 #define PF_PROT (1<<0)
43 #define PF_WRITE (1<<1)
44 #define PF_USER (1<<2)
45 #define PF_RSVD (1<<3)
46 #define PF_INSTR (1<<4)
48 static inline int notify_page_fault(struct pt_regs *regs)
53 /* kprobe_running() needs smp_processor_id() */
54 if (!user_mode(regs)) {
56 if (kprobe_running() && kprobe_fault_handler(regs, 14))
69 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
70 * Check that here and ignore it.
73 * Sometimes the CPU reports invalid exceptions on prefetch.
74 * Check that here and ignore it.
76 * Opcode checker based on code by Richard Brunner
78 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
79 unsigned long error_code)
84 unsigned char *max_instr;
87 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
88 boot_cpu_data.x86 >= 6)) {
89 /* Catch an obscure case of prefetch inside an NX page. */
90 if (nx_enabled && (error_code & PF_INSTR))
96 /* If it was a exec fault ignore */
97 if (error_code & PF_INSTR)
101 instr = (unsigned char *)convert_ip_to_linear(current, regs);
102 max_instr = instr + 15;
104 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
107 while (scan_more && instr < max_instr) {
108 unsigned char opcode;
109 unsigned char instr_hi;
110 unsigned char instr_lo;
112 if (probe_kernel_address(instr, opcode))
115 instr_hi = opcode & 0xf0;
116 instr_lo = opcode & 0x0f;
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
128 scan_more = ((instr_lo & 7) == 0x6);
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.
139 scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
143 /* 0x64 thru 0x67 are valid prefixes in all modes. */
144 scan_more = (instr_lo & 0xC) == 0x4;
147 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
148 scan_more = !instr_lo || (instr_lo>>1) == 1;
151 /* Prefetch instruction is 0x0F0D or 0x0F18 */
154 if (probe_kernel_address(instr, opcode))
156 prefetch = (instr_lo == 0xF) &&
157 (opcode == 0x0D || opcode == 0x18);
167 static void force_sig_info_fault(int si_signo, int si_code,
168 unsigned long address, struct task_struct *tsk)
172 info.si_signo = si_signo;
174 info.si_code = si_code;
175 info.si_addr = (void __user *)address;
176 force_sig_info(si_signo, &info, tsk);
180 static int bad_address(void *p)
183 return probe_kernel_address((unsigned long *)p, dummy);
187 void dump_pagetable(unsigned long address)
190 __typeof__(pte_val(__pte(0))) page;
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) {
199 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
200 & (PTRS_PER_PMD - 1)];
201 printk(KERN_CONT "*pde = %016Lx ", page);
205 printk("*pde = %08lx ", page);
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.
214 if ((page >> PAGE_SHIFT) < max_low_pfn
215 && (page & _PAGE_PRESENT)
216 && !(page & _PAGE_PSE)) {
218 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
219 & (PTRS_PER_PTE - 1)];
220 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
224 #else /* CONFIG_X86_64 */
230 pgd = (pgd_t *)read_cr3();
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;
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)) goto ret;
243 pmd = pmd_offset(pud, address);
244 if (bad_address(pmd)) goto bad;
245 printk("PMD %lx ", pmd_val(*pmd));
246 if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
248 pte = pte_offset_kernel(pmd, address);
249 if (bad_address(pte)) goto bad;
250 printk("PTE %lx", pte_val(*pte));
260 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
262 unsigned index = pgd_index(address);
268 pgd_k = init_mm.pgd + index;
270 if (!pgd_present(*pgd_k))
274 * set_pgd(pgd, *pgd_k); here would be useless on PAE
275 * and redundant with the set_pmd() on non-PAE. As would
279 pud = pud_offset(pgd, address);
280 pud_k = pud_offset(pgd_k, address);
281 if (!pud_present(*pud_k))
284 pmd = pmd_offset(pud, address);
285 pmd_k = pmd_offset(pud_k, address);
286 if (!pmd_present(*pmd_k))
288 if (!pmd_present(*pmd)) {
289 set_pmd(pmd, *pmd_k);
290 arch_flush_lazy_mmu_mode();
292 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
298 static const char errata93_warning[] =
299 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
300 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
301 KERN_ERR "******* Please consider a BIOS update.\n"
302 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
305 /* Workaround for K8 erratum #93 & buggy BIOS.
306 BIOS SMM functions are required to use a specific workaround
307 to avoid corruption of the 64bit RIP register on C stepping K8.
308 A lot of BIOS that didn't get tested properly miss this.
309 The OS sees this as a page fault with the upper 32bits of RIP cleared.
310 Try to work around it here.
311 Note we only handle faults in kernel here.
312 Does nothing for X86_32
314 static int is_errata93(struct pt_regs *regs, unsigned long address)
318 if (address != regs->ip)
320 if ((address >> 32) != 0)
322 address |= 0xffffffffUL << 32;
323 if ((address >= (u64)_stext && address <= (u64)_etext) ||
324 (address >= MODULES_VADDR && address <= MODULES_END)) {
326 printk(errata93_warning);
337 * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
338 * addresses >4GB. We catch this in the page fault handler because these
339 * addresses are not reachable. Just detect this case and return. Any code
340 * segment in LDT is compatibility mode.
342 static int is_errata100(struct pt_regs *regs, unsigned long address)
345 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
352 void do_invalid_op(struct pt_regs *, unsigned long);
354 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
356 #ifdef CONFIG_X86_F00F_BUG
359 * Pentium F0 0F C7 C8 bug workaround.
361 if (boot_cpu_data.f00f_bug) {
362 nr = (address - idt_descr.address) >> 3;
365 do_invalid_op(regs, 0);
373 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
374 unsigned long address)
377 if (!oops_may_print())
380 #ifdef CONFIG_X86_PAE
381 if (error_code & PF_INSTR) {
383 pte_t *pte = lookup_address(address, &level);
385 if (pte && pte_present(*pte) && !pte_exec(*pte))
386 printk(KERN_CRIT "kernel tried to execute "
387 "NX-protected page - exploit attempt? "
388 "(uid: %d)\n", current->uid);
391 printk(KERN_ALERT "BUG: unable to handle kernel ");
392 if (address < PAGE_SIZE)
393 printk(KERN_CONT "NULL pointer dereference");
395 printk(KERN_CONT "paging request");
396 printk(KERN_CONT " at %08lx\n", address);
398 printk(KERN_ALERT "IP:");
399 printk_address(regs->ip, 1);
400 dump_pagetable(address);
401 #else /* CONFIG_X86_64 */
402 printk(KERN_ALERT "BUG: unable to handle kernel ");
403 if (address < PAGE_SIZE)
404 printk(KERN_CONT "NULL pointer dereference");
406 printk(KERN_CONT "paging request");
407 printk(KERN_CONT " at %016lx\n", address);
409 printk(KERN_ALERT "IP:");
410 printk_address(regs->ip, 1);
411 dump_pagetable(address);
416 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
417 unsigned long error_code)
419 unsigned long flags = oops_begin();
420 struct task_struct *tsk;
422 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
423 current->comm, address);
424 dump_pagetable(address);
426 tsk->thread.cr2 = address;
427 tsk->thread.trap_no = 14;
428 tsk->thread.error_code = error_code;
429 if (__die("Bad pagetable", regs, error_code))
431 oops_end(flags, regs, SIGKILL);
436 * Handle a fault on the vmalloc area
438 * This assumes no large pages in there.
440 static int vmalloc_fault(unsigned long address)
443 unsigned long pgd_paddr;
447 * Synchronize this task's top level page-table
448 * with the 'reference' page table.
450 * Do _not_ use "current" here. We might be inside
451 * an interrupt in the middle of a task switch..
453 pgd_paddr = read_cr3();
454 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
457 pte_k = pte_offset_kernel(pmd_k, address);
458 if (!pte_present(*pte_k))
462 pgd_t *pgd, *pgd_ref;
463 pud_t *pud, *pud_ref;
464 pmd_t *pmd, *pmd_ref;
465 pte_t *pte, *pte_ref;
467 /* Copy kernel mappings over when needed. This can also
468 happen within a race in page table update. In the later
471 pgd = pgd_offset(current->mm ?: &init_mm, address);
472 pgd_ref = pgd_offset_k(address);
473 if (pgd_none(*pgd_ref))
476 set_pgd(pgd, *pgd_ref);
478 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
480 /* Below here mismatches are bugs because these lower tables
483 pud = pud_offset(pgd, address);
484 pud_ref = pud_offset(pgd_ref, address);
485 if (pud_none(*pud_ref))
487 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
489 pmd = pmd_offset(pud, address);
490 pmd_ref = pmd_offset(pud_ref, address);
491 if (pmd_none(*pmd_ref))
493 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
495 pte_ref = pte_offset_kernel(pmd_ref, address);
496 if (!pte_present(*pte_ref))
498 pte = pte_offset_kernel(pmd, address);
499 /* Don't use pte_page here, because the mappings can point
500 outside mem_map, and the NUMA hash lookup cannot handle
502 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
508 int show_unhandled_signals = 1;
511 * This routine handles page faults. It determines the address,
512 * and the problem, and then passes it off to one of the appropriate
515 asmlinkage void __kprobes do_page_fault(struct pt_regs *regs,
516 unsigned long error_code)
518 struct task_struct *tsk;
519 struct mm_struct *mm;
520 struct vm_area_struct *vma;
521 unsigned long address;
527 * We can fault from pretty much anywhere, with unknown IRQ state.
529 trace_hardirqs_fixup();
533 prefetchw(&mm->mmap_sem);
535 /* get the address */
536 address = read_cr2();
538 si_code = SEGV_MAPERR;
540 if (notify_page_fault(regs))
544 * We fault-in kernel-space virtual memory on-demand. The
545 * 'reference' page table is init_mm.pgd.
547 * NOTE! We MUST NOT take any locks for this case. We may
548 * be in an interrupt or a critical region, and should
549 * only copy the information from the master page table,
552 * This verifies that the fault happens in kernel space
553 * (error_code & 4) == 0, and that the fault was not a
554 * protection error (error_code & 9) == 0.
556 if (unlikely(address >= TASK_SIZE64)) {
558 * Don't check for the module range here: its PML4
559 * is always initialized because it's shared with the main
560 * kernel text. Only vmalloc may need PML4 syncups.
562 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
563 ((address >= VMALLOC_START && address < VMALLOC_END))) {
564 if (vmalloc_fault(address) >= 0)
568 * Don't take the mm semaphore here. If we fixup a prefetch
569 * fault we could otherwise deadlock.
571 goto bad_area_nosemaphore;
574 if (likely(regs->flags & X86_EFLAGS_IF))
577 if (unlikely(error_code & PF_RSVD))
578 pgtable_bad(address, regs, error_code);
581 * If we're in an interrupt, have no user context or are running in an
582 * atomic region then we must not take the fault.
584 if (unlikely(in_atomic() || !mm))
585 goto bad_area_nosemaphore;
588 * User-mode registers count as a user access even for any
589 * potential system fault or CPU buglet.
591 if (user_mode_vm(regs))
592 error_code |= PF_USER;
595 /* When running in the kernel we expect faults to occur only to
596 * addresses in user space. All other faults represent errors in the
597 * kernel and should generate an OOPS. Unfortunately, in the case of an
598 * erroneous fault occurring in a code path which already holds mmap_sem
599 * we will deadlock attempting to validate the fault against the
600 * address space. Luckily the kernel only validly references user
601 * space from well defined areas of code, which are listed in the
604 * As the vast majority of faults will be valid we will only perform
605 * the source reference check when there is a possibility of a deadlock.
606 * Attempt to lock the address space, if we cannot we then validate the
607 * source. If this is invalid we can skip the address space check,
608 * thus avoiding the deadlock.
610 if (!down_read_trylock(&mm->mmap_sem)) {
611 if ((error_code & PF_USER) == 0 &&
612 !search_exception_tables(regs->ip))
613 goto bad_area_nosemaphore;
614 down_read(&mm->mmap_sem);
617 vma = find_vma(mm, address);
620 if (likely(vma->vm_start <= address))
622 if (!(vma->vm_flags & VM_GROWSDOWN))
624 if (error_code & PF_USER) {
626 * Accessing the stack below %sp is always a bug.
627 * The large cushion allows instructions like enter
628 * and pusha to work. ("enter $65535,$31" pushes
629 * 32 pointers and then decrements %sp by 65535.)
631 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
634 if (expand_stack(vma, address))
637 * Ok, we have a good vm_area for this memory access, so
641 si_code = SEGV_ACCERR;
643 switch (error_code & (PF_PROT|PF_WRITE)) {
644 default: /* 3: write, present */
646 case PF_WRITE: /* write, not present */
647 if (!(vma->vm_flags & VM_WRITE))
651 case PF_PROT: /* read, present */
653 case 0: /* read, not present */
654 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
659 * If for any reason at all we couldn't handle the fault,
660 * make sure we exit gracefully rather than endlessly redo
663 fault = handle_mm_fault(mm, vma, address, write);
664 if (unlikely(fault & VM_FAULT_ERROR)) {
665 if (fault & VM_FAULT_OOM)
667 else if (fault & VM_FAULT_SIGBUS)
671 if (fault & VM_FAULT_MAJOR)
678 * Did it hit the DOS screen memory VA from vm86 mode?
680 if (v8086_mode(regs)) {
681 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
683 tsk->thread.screen_bitmap |= 1 << bit;
686 up_read(&mm->mmap_sem);
690 * Something tried to access memory that isn't in our memory map..
691 * Fix it, but check if it's kernel or user first..
694 up_read(&mm->mmap_sem);
696 bad_area_nosemaphore:
697 /* User mode accesses just cause a SIGSEGV */
698 if (error_code & PF_USER) {
700 * It's possible to have interrupts off here.
705 * Valid to do another page fault here because this one came
708 if (is_prefetch(regs, address, error_code))
711 if (is_errata100(regs, address))
714 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
715 printk_ratelimit()) {
718 "%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx",
720 "%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx",
722 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
723 tsk->comm, task_pid_nr(tsk), address, regs->ip,
724 regs->sp, error_code);
725 print_vma_addr(" in ", regs->ip);
729 tsk->thread.cr2 = address;
730 /* Kernel addresses are always protection faults */
731 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
732 tsk->thread.trap_no = 14;
734 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
738 if (is_f00f_bug(regs, address))
742 /* Are we prepared to handle this kernel fault? */
743 if (fixup_exception(regs))
747 * Hall of shame of CPU/BIOS bugs.
750 if (is_prefetch(regs, address, error_code))
753 if (is_errata93(regs, address))
757 * Oops. The kernel tried to access some bad page. We'll have to
758 * terminate things with extreme prejudice.
761 flags = oops_begin();
763 show_fault_oops(regs, error_code, address);
765 tsk->thread.cr2 = address;
766 tsk->thread.trap_no = 14;
767 tsk->thread.error_code = error_code;
768 if (__die("Oops", regs, error_code))
770 /* Executive summary in case the body of the oops scrolled away */
771 printk(KERN_EMERG "CR2: %016lx\n", address);
772 oops_end(flags, regs, SIGKILL);
775 * We ran out of memory, or some other thing happened to us that made
776 * us unable to handle the page fault gracefully.
779 up_read(&mm->mmap_sem);
780 if (is_global_init(current)) {
784 printk("VM: killing process %s\n", tsk->comm);
785 if (error_code & PF_USER)
786 do_group_exit(SIGKILL);
790 up_read(&mm->mmap_sem);
792 /* Kernel mode? Handle exceptions or die */
793 if (!(error_code & PF_USER))
796 tsk->thread.cr2 = address;
797 tsk->thread.error_code = error_code;
798 tsk->thread.trap_no = 14;
799 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
803 DEFINE_SPINLOCK(pgd_lock);
806 void vmalloc_sync_all(void)
810 * Note that races in the updates of insync and start aren't
811 * problematic: insync can only get set bits added, and updates to
812 * start are only improving performance (without affecting correctness
815 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
816 static unsigned long start = TASK_SIZE;
817 unsigned long address;
819 if (SHARED_KERNEL_PMD)
822 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
823 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
824 if (!test_bit(pgd_index(address), insync)) {
828 spin_lock_irqsave(&pgd_lock, flags);
829 for (page = pgd_list; page; page =
830 (struct page *)page->index)
831 if (!vmalloc_sync_one(page_address(page),
833 BUG_ON(page != pgd_list);
836 spin_unlock_irqrestore(&pgd_lock, flags);
838 set_bit(pgd_index(address), insync);
840 if (address == start && test_bit(pgd_index(address), insync))
841 start = address + PGDIR_SIZE;
843 #else /* CONFIG_X86_64 */
845 * Note that races in the updates of insync and start aren't
846 * problematic: insync can only get set bits added, and updates to
847 * start are only improving performance (without affecting correctness
850 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
851 static unsigned long start = VMALLOC_START & PGDIR_MASK;
852 unsigned long address;
854 for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
855 if (!test_bit(pgd_index(address), insync)) {
856 const pgd_t *pgd_ref = pgd_offset_k(address);
859 if (pgd_none(*pgd_ref))
861 spin_lock(&pgd_lock);
862 list_for_each_entry(page, &pgd_list, lru) {
864 pgd = (pgd_t *)page_address(page) + pgd_index(address);
866 set_pgd(pgd, *pgd_ref);
868 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
870 spin_unlock(&pgd_lock);
871 set_bit(pgd_index(address), insync);
873 if (address == start)
874 start = address + PGDIR_SIZE;
876 /* Check that there is no need to do the same for the modules area. */
877 BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
878 BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
879 (__START_KERNEL & PGDIR_MASK)));