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/mmiotrace.h>
14 #include <linux/mman.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>
31 #include <asm/system.h>
33 #include <asm/segment.h>
34 #include <asm/pgalloc.h>
36 #include <asm/tlbflush.h>
37 #include <asm/proto.h>
38 #include <asm-generic/sections.h>
39 #include <asm/traps.h>
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
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)
55 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
57 #ifdef CONFIG_MMIOTRACE
58 if (unlikely(is_kmmio_active()))
59 if (kmmio_handler(regs, addr) == 1)
65 static inline int notify_page_fault(struct pt_regs *regs)
70 /* kprobe_running() needs smp_processor_id() */
71 if (!user_mode_vm(regs)) {
73 if (kprobe_running() && kprobe_fault_handler(regs, 14))
86 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
87 * Check that here and ignore it.
90 * Sometimes the CPU reports invalid exceptions on prefetch.
91 * Check that here and ignore it.
93 * Opcode checker based on code by Richard Brunner
95 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
96 unsigned long error_code)
101 unsigned char *max_instr;
104 * If it was a exec (instruction fetch) fault on NX page, then
105 * do not ignore the fault:
107 if (error_code & PF_INSTR)
110 instr = (unsigned char *)convert_ip_to_linear(current, regs);
111 max_instr = instr + 15;
113 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
116 while (scan_more && instr < max_instr) {
117 unsigned char opcode;
118 unsigned char instr_hi;
119 unsigned char instr_lo;
121 if (probe_kernel_address(instr, opcode))
124 instr_hi = opcode & 0xf0;
125 instr_lo = opcode & 0x0f;
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
137 scan_more = ((instr_lo & 7) == 0x6);
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.
148 scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
152 /* 0x64 thru 0x67 are valid prefixes in all modes. */
153 scan_more = (instr_lo & 0xC) == 0x4;
156 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
157 scan_more = !instr_lo || (instr_lo>>1) == 1;
160 /* Prefetch instruction is 0x0F0D or 0x0F18 */
163 if (probe_kernel_address(instr, opcode))
165 prefetch = (instr_lo == 0xF) &&
166 (opcode == 0x0D || opcode == 0x18);
176 static void force_sig_info_fault(int si_signo, int si_code,
177 unsigned long address, struct task_struct *tsk)
181 info.si_signo = si_signo;
183 info.si_code = si_code;
184 info.si_addr = (void __user *)address;
185 force_sig_info(si_signo, &info, tsk);
189 static int bad_address(void *p)
192 return probe_kernel_address((unsigned long *)p, dummy);
196 static void dump_pagetable(unsigned long address)
199 __typeof__(pte_val(__pte(0))) page;
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) {
208 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
209 & (PTRS_PER_PMD - 1)];
210 printk(KERN_CONT "*pde = %016Lx ", page);
214 printk("*pde = %08lx ", page);
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.
223 if ((page >> PAGE_SHIFT) < max_low_pfn
224 && (page & _PAGE_PRESENT)
225 && !(page & _PAGE_PSE)) {
227 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
228 & (PTRS_PER_PTE - 1)];
229 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
233 #else /* CONFIG_X86_64 */
239 pgd = (pgd_t *)read_cr3();
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;
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))
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;
258 pte = pte_offset_kernel(pmd, address);
259 if (bad_address(pte)) goto bad;
260 printk("PTE %lx", pte_val(*pte));
270 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
272 unsigned index = pgd_index(address);
278 pgd_k = init_mm.pgd + index;
280 if (!pgd_present(*pgd_k))
284 * set_pgd(pgd, *pgd_k); here would be useless on PAE
285 * and redundant with the set_pmd() on non-PAE. As would
289 pud = pud_offset(pgd, address);
290 pud_k = pud_offset(pgd_k, address);
291 if (!pud_present(*pud_k))
294 pmd = pmd_offset(pud, address);
295 pmd_k = pmd_offset(pud_k, address);
296 if (!pmd_present(*pmd_k))
298 if (!pmd_present(*pmd)) {
299 set_pmd(pmd, *pmd_k);
300 arch_flush_lazy_mmu_mode();
302 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
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";
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
324 static int is_errata93(struct pt_regs *regs, unsigned long address)
328 if (address != regs->ip)
330 if ((address >> 32) != 0)
332 address |= 0xffffffffUL << 32;
333 if ((address >= (u64)_stext && address <= (u64)_etext) ||
334 (address >= MODULES_VADDR && address <= MODULES_END)) {
336 printk(errata93_warning);
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.
352 static int is_errata100(struct pt_regs *regs, unsigned long address)
355 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
362 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
364 #ifdef CONFIG_X86_F00F_BUG
367 * Pentium F0 0F C7 C8 bug workaround.
369 if (boot_cpu_data.f00f_bug) {
370 nr = (address - idt_descr.address) >> 3;
373 do_invalid_op(regs, 0);
381 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
382 unsigned long address)
385 if (!oops_may_print())
389 #ifdef CONFIG_X86_PAE
390 if (error_code & PF_INSTR) {
392 pte_t *pte = lookup_address(address, &level);
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());
401 printk(KERN_ALERT "BUG: unable to handle kernel ");
402 if (address < PAGE_SIZE)
403 printk(KERN_CONT "NULL pointer dereference");
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);
413 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
414 unsigned long error_code)
416 unsigned long flags = oops_begin();
418 struct task_struct *tsk;
420 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
421 current->comm, address);
422 dump_pagetable(address);
424 tsk->thread.cr2 = address;
425 tsk->thread.trap_no = 14;
426 tsk->thread.error_code = error_code;
427 if (__die("Bad pagetable", regs, error_code))
429 oops_end(flags, regs, sig);
433 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
435 if ((error_code & PF_WRITE) && !pte_write(*pte))
437 if ((error_code & PF_INSTR) && !pte_exec(*pte))
444 * Handle a spurious fault caused by a stale TLB entry. This allows
445 * us to lazily refresh the TLB when increasing the permissions of a
446 * kernel page (RO -> RW or NX -> X). Doing it eagerly is very
447 * expensive since that implies doing a full cross-processor TLB
448 * flush, even if no stale TLB entries exist on other processors.
449 * There are no security implications to leaving a stale TLB when
450 * increasing the permissions on a page.
452 static int spurious_fault(unsigned long address,
453 unsigned long error_code)
460 /* Reserved-bit violation or user access to kernel space? */
461 if (error_code & (PF_USER | PF_RSVD))
464 pgd = init_mm.pgd + pgd_index(address);
465 if (!pgd_present(*pgd))
468 pud = pud_offset(pgd, address);
469 if (!pud_present(*pud))
473 return spurious_fault_check(error_code, (pte_t *) pud);
475 pmd = pmd_offset(pud, address);
476 if (!pmd_present(*pmd))
480 return spurious_fault_check(error_code, (pte_t *) pmd);
482 pte = pte_offset_kernel(pmd, address);
483 if (!pte_present(*pte))
486 return spurious_fault_check(error_code, pte);
491 * Handle a fault on the vmalloc or module mapping area
494 * Handle a fault on the vmalloc area
496 * This assumes no large pages in there.
498 static int vmalloc_fault(unsigned long address)
501 unsigned long pgd_paddr;
505 /* Make sure we are in vmalloc area */
506 if (!(address >= VMALLOC_START && address < VMALLOC_END))
510 * Synchronize this task's top level page-table
511 * with the 'reference' page table.
513 * Do _not_ use "current" here. We might be inside
514 * an interrupt in the middle of a task switch..
516 pgd_paddr = read_cr3();
517 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
520 pte_k = pte_offset_kernel(pmd_k, address);
521 if (!pte_present(*pte_k))
525 pgd_t *pgd, *pgd_ref;
526 pud_t *pud, *pud_ref;
527 pmd_t *pmd, *pmd_ref;
528 pte_t *pte, *pte_ref;
530 /* Make sure we are in vmalloc area */
531 if (!(address >= VMALLOC_START && address < VMALLOC_END))
534 /* Copy kernel mappings over when needed. This can also
535 happen within a race in page table update. In the later
538 pgd = pgd_offset(current->active_mm, address);
539 pgd_ref = pgd_offset_k(address);
540 if (pgd_none(*pgd_ref))
543 set_pgd(pgd, *pgd_ref);
545 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
547 /* Below here mismatches are bugs because these lower tables
550 pud = pud_offset(pgd, address);
551 pud_ref = pud_offset(pgd_ref, address);
552 if (pud_none(*pud_ref))
554 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
556 pmd = pmd_offset(pud, address);
557 pmd_ref = pmd_offset(pud_ref, address);
558 if (pmd_none(*pmd_ref))
560 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
562 pte_ref = pte_offset_kernel(pmd_ref, address);
563 if (!pte_present(*pte_ref))
565 pte = pte_offset_kernel(pmd, address);
566 /* Don't use pte_page here, because the mappings can point
567 outside mem_map, and the NUMA hash lookup cannot handle
569 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
575 int show_unhandled_signals = 1;
578 * This routine handles page faults. It determines the address,
579 * and the problem, and then passes it off to one of the appropriate
585 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
587 struct task_struct *tsk;
588 struct mm_struct *mm;
589 struct vm_area_struct *vma;
590 unsigned long address;
593 unsigned long *stackend;
602 prefetchw(&mm->mmap_sem);
604 /* get the address */
605 address = read_cr2();
607 si_code = SEGV_MAPERR;
609 if (notify_page_fault(regs))
611 if (unlikely(kmmio_fault(regs, address)))
615 * We fault-in kernel-space virtual memory on-demand. The
616 * 'reference' page table is init_mm.pgd.
618 * NOTE! We MUST NOT take any locks for this case. We may
619 * be in an interrupt or a critical region, and should
620 * only copy the information from the master page table,
623 * This verifies that the fault happens in kernel space
624 * (error_code & 4) == 0, and that the fault was not a
625 * protection error (error_code & 9) == 0.
628 if (unlikely(address >= TASK_SIZE)) {
630 if (unlikely(address >= TASK_SIZE64)) {
632 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
633 vmalloc_fault(address) >= 0)
636 /* Can handle a stale RO->RW TLB */
637 if (spurious_fault(address, error_code))
641 * Don't take the mm semaphore here. If we fixup a prefetch
642 * fault we could otherwise deadlock.
644 goto bad_area_nosemaphore;
649 * It's safe to allow irq's after cr2 has been saved and the
650 * vmalloc fault has been handled.
652 * User-mode registers count as a user access even for any
653 * potential system fault or CPU buglet.
655 if (user_mode_vm(regs)) {
657 error_code |= PF_USER;
658 } else if (regs->flags & X86_EFLAGS_IF)
662 if (unlikely(error_code & PF_RSVD))
663 pgtable_bad(address, regs, error_code);
667 * If we're in an interrupt, have no user context or are running in an
668 * atomic region then we must not take the fault.
670 if (unlikely(in_atomic() || !mm))
671 goto bad_area_nosemaphore;
674 * When running in the kernel we expect faults to occur only to
675 * addresses in user space. All other faults represent errors in the
676 * kernel and should generate an OOPS. Unfortunately, in the case of an
677 * erroneous fault occurring in a code path which already holds mmap_sem
678 * we will deadlock attempting to validate the fault against the
679 * address space. Luckily the kernel only validly references user
680 * space from well defined areas of code, which are listed in the
683 * As the vast majority of faults will be valid we will only perform
684 * the source reference check when there is a possibility of a deadlock.
685 * Attempt to lock the address space, if we cannot we then validate the
686 * source. If this is invalid we can skip the address space check,
687 * thus avoiding the deadlock.
689 if (!down_read_trylock(&mm->mmap_sem)) {
690 if ((error_code & PF_USER) == 0 &&
691 !search_exception_tables(regs->ip))
692 goto bad_area_nosemaphore;
693 down_read(&mm->mmap_sem);
696 vma = find_vma(mm, address);
699 if (vma->vm_start <= address)
701 if (!(vma->vm_flags & VM_GROWSDOWN))
703 if (error_code & PF_USER) {
705 * Accessing the stack below %sp is always a bug.
706 * The large cushion allows instructions like enter
707 * and pusha to work. ("enter $65535,$31" pushes
708 * 32 pointers and then decrements %sp by 65535.)
710 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
713 if (expand_stack(vma, address))
716 * Ok, we have a good vm_area for this memory access, so
720 si_code = SEGV_ACCERR;
722 switch (error_code & (PF_PROT|PF_WRITE)) {
723 default: /* 3: write, present */
725 case PF_WRITE: /* write, not present */
726 if (!(vma->vm_flags & VM_WRITE))
730 case PF_PROT: /* read, present */
732 case 0: /* read, not present */
733 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
738 * If for any reason at all we couldn't handle the fault,
739 * make sure we exit gracefully rather than endlessly redo
742 fault = handle_mm_fault(mm, vma, address, write);
743 if (unlikely(fault & VM_FAULT_ERROR)) {
744 if (fault & VM_FAULT_OOM)
746 else if (fault & VM_FAULT_SIGBUS)
750 if (fault & VM_FAULT_MAJOR)
757 * Did it hit the DOS screen memory VA from vm86 mode?
759 if (v8086_mode(regs)) {
760 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
762 tsk->thread.screen_bitmap |= 1 << bit;
765 up_read(&mm->mmap_sem);
769 * Something tried to access memory that isn't in our memory map..
770 * Fix it, but check if it's kernel or user first..
773 up_read(&mm->mmap_sem);
775 bad_area_nosemaphore:
776 /* User mode accesses just cause a SIGSEGV */
777 if (error_code & PF_USER) {
779 * It's possible to have interrupts off here.
784 * Valid to do another page fault here because this one came
787 if (is_prefetch(regs, address, error_code))
790 if (is_errata100(regs, address))
793 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
794 printk_ratelimit()) {
796 "%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
797 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
798 tsk->comm, task_pid_nr(tsk), address,
799 (void *) regs->ip, (void *) regs->sp, error_code);
800 print_vma_addr(" in ", regs->ip);
804 tsk->thread.cr2 = address;
805 /* Kernel addresses are always protection faults */
806 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
807 tsk->thread.trap_no = 14;
808 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
812 if (is_f00f_bug(regs, address))
816 /* Are we prepared to handle this kernel fault? */
817 if (fixup_exception(regs))
822 * Valid to do another page fault here, because if this fault
823 * had been triggered by is_prefetch fixup_exception would have
827 * Hall of shame of CPU/BIOS bugs.
829 if (is_prefetch(regs, address, error_code))
832 if (is_errata93(regs, address))
836 * Oops. The kernel tried to access some bad page. We'll have to
837 * terminate things with extreme prejudice.
842 flags = oops_begin();
845 show_fault_oops(regs, error_code, address);
847 stackend = end_of_stack(tsk);
848 if (*stackend != STACK_END_MAGIC)
849 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
851 tsk->thread.cr2 = address;
852 tsk->thread.trap_no = 14;
853 tsk->thread.error_code = error_code;
856 die("Oops", regs, error_code);
861 if (__die("Oops", regs, error_code))
863 /* Executive summary in case the body of the oops scrolled away */
864 printk(KERN_EMERG "CR2: %016lx\n", address);
865 oops_end(flags, regs, sig);
870 * We ran out of memory, call the OOM killer, and return the userspace
871 * (which will retry the fault, or kill us if we got oom-killed).
873 up_read(&mm->mmap_sem);
874 pagefault_out_of_memory();
878 up_read(&mm->mmap_sem);
880 /* Kernel mode? Handle exceptions or die */
881 if (!(error_code & PF_USER))
884 /* User space => ok to do another page fault */
885 if (is_prefetch(regs, address, error_code))
888 tsk->thread.cr2 = address;
889 tsk->thread.error_code = error_code;
890 tsk->thread.trap_no = 14;
891 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
894 DEFINE_SPINLOCK(pgd_lock);
897 void vmalloc_sync_all(void)
899 unsigned long address;
902 if (SHARED_KERNEL_PMD)
905 for (address = VMALLOC_START & PMD_MASK;
906 address >= TASK_SIZE && address < FIXADDR_TOP;
907 address += PMD_SIZE) {
911 spin_lock_irqsave(&pgd_lock, flags);
912 list_for_each_entry(page, &pgd_list, lru) {
913 if (!vmalloc_sync_one(page_address(page),
917 spin_unlock_irqrestore(&pgd_lock, flags);
919 #else /* CONFIG_X86_64 */
920 for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
921 address += PGDIR_SIZE) {
922 const pgd_t *pgd_ref = pgd_offset_k(address);
926 if (pgd_none(*pgd_ref))
928 spin_lock_irqsave(&pgd_lock, flags);
929 list_for_each_entry(page, &pgd_list, lru) {
931 pgd = (pgd_t *)page_address(page) + pgd_index(address);
933 set_pgd(pgd, *pgd_ref);
935 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
937 spin_unlock_irqrestore(&pgd_lock, flags);