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>
30 #include <asm/system.h>
32 #include <asm/segment.h>
33 #include <asm/pgalloc.h>
35 #include <asm/tlbflush.h>
36 #include <asm/proto.h>
37 #include <asm-generic/sections.h>
38 #include <asm/traps.h>
41 * Page fault error code bits
42 * bit 0 == 0 means no page found, 1 means protection fault
43 * bit 1 == 0 means read, 1 means write
44 * bit 2 == 0 means kernel, 1 means user-mode
45 * bit 3 == 1 means use of reserved bit detected
46 * bit 4 == 1 means fault was an instruction fetch
48 #define PF_PROT (1<<0)
49 #define PF_WRITE (1<<1)
50 #define PF_USER (1<<2)
51 #define PF_RSVD (1<<3)
52 #define PF_INSTR (1<<4)
54 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
56 #ifdef CONFIG_MMIOTRACE_HOOKS
57 if (unlikely(is_kmmio_active()))
58 if (kmmio_handler(regs, addr) == 1)
64 static inline int notify_page_fault(struct pt_regs *regs)
69 /* kprobe_running() needs smp_processor_id() */
70 if (!user_mode_vm(regs)) {
72 if (kprobe_running() && kprobe_fault_handler(regs, 14))
85 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
86 * Check that here and ignore it.
89 * Sometimes the CPU reports invalid exceptions on prefetch.
90 * Check that here and ignore it.
92 * Opcode checker based on code by Richard Brunner
94 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
95 unsigned long error_code)
100 unsigned char *max_instr;
103 * If it was a exec (instruction fetch) fault on NX page, then
104 * do not ignore the fault:
106 if (error_code & PF_INSTR)
109 instr = (unsigned char *)convert_ip_to_linear(current, regs);
110 max_instr = instr + 15;
112 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
115 while (scan_more && instr < max_instr) {
116 unsigned char opcode;
117 unsigned char instr_hi;
118 unsigned char instr_lo;
120 if (probe_kernel_address(instr, opcode))
123 instr_hi = opcode & 0xf0;
124 instr_lo = opcode & 0x0f;
131 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
132 * In X86_64 long mode, the CPU will signal invalid
133 * opcode if some of these prefixes are present so
134 * X86_64 will never get here anyway
136 scan_more = ((instr_lo & 7) == 0x6);
141 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
142 * Need to figure out under what instruction mode the
143 * instruction was issued. Could check the LDT for lm,
144 * but for now it's good enough to assume that long
145 * mode only uses well known segments or kernel.
147 scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
151 /* 0x64 thru 0x67 are valid prefixes in all modes. */
152 scan_more = (instr_lo & 0xC) == 0x4;
155 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
156 scan_more = !instr_lo || (instr_lo>>1) == 1;
159 /* Prefetch instruction is 0x0F0D or 0x0F18 */
162 if (probe_kernel_address(instr, opcode))
164 prefetch = (instr_lo == 0xF) &&
165 (opcode == 0x0D || opcode == 0x18);
175 static void force_sig_info_fault(int si_signo, int si_code,
176 unsigned long address, struct task_struct *tsk)
180 info.si_signo = si_signo;
182 info.si_code = si_code;
183 info.si_addr = (void __user *)address;
184 force_sig_info(si_signo, &info, tsk);
188 static int bad_address(void *p)
191 return probe_kernel_address((unsigned long *)p, dummy);
195 static void dump_pagetable(unsigned long address)
198 __typeof__(pte_val(__pte(0))) page;
201 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
202 #ifdef CONFIG_X86_PAE
203 printk("*pdpt = %016Lx ", page);
204 if ((page >> PAGE_SHIFT) < max_low_pfn
205 && page & _PAGE_PRESENT) {
207 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
208 & (PTRS_PER_PMD - 1)];
209 printk(KERN_CONT "*pde = %016Lx ", page);
213 printk("*pde = %08lx ", page);
217 * We must not directly access the pte in the highpte
218 * case if the page table is located in highmem.
219 * And let's rather not kmap-atomic the pte, just in case
220 * it's allocated already.
222 if ((page >> PAGE_SHIFT) < max_low_pfn
223 && (page & _PAGE_PRESENT)
224 && !(page & _PAGE_PSE)) {
226 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
227 & (PTRS_PER_PTE - 1)];
228 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
232 #else /* CONFIG_X86_64 */
238 pgd = (pgd_t *)read_cr3();
240 pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
241 pgd += pgd_index(address);
242 if (bad_address(pgd)) goto bad;
243 printk("PGD %lx ", pgd_val(*pgd));
244 if (!pgd_present(*pgd)) goto ret;
246 pud = pud_offset(pgd, address);
247 if (bad_address(pud)) goto bad;
248 printk("PUD %lx ", pud_val(*pud));
249 if (!pud_present(*pud) || pud_large(*pud))
252 pmd = pmd_offset(pud, address);
253 if (bad_address(pmd)) goto bad;
254 printk("PMD %lx ", pmd_val(*pmd));
255 if (!pmd_present(*pmd) || pmd_large(*pmd)) goto ret;
257 pte = pte_offset_kernel(pmd, address);
258 if (bad_address(pte)) goto bad;
259 printk("PTE %lx", pte_val(*pte));
269 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
271 unsigned index = pgd_index(address);
277 pgd_k = init_mm.pgd + index;
279 if (!pgd_present(*pgd_k))
283 * set_pgd(pgd, *pgd_k); here would be useless on PAE
284 * and redundant with the set_pmd() on non-PAE. As would
288 pud = pud_offset(pgd, address);
289 pud_k = pud_offset(pgd_k, address);
290 if (!pud_present(*pud_k))
293 pmd = pmd_offset(pud, address);
294 pmd_k = pmd_offset(pud_k, address);
295 if (!pmd_present(*pmd_k))
297 if (!pmd_present(*pmd)) {
298 set_pmd(pmd, *pmd_k);
299 arch_flush_lazy_mmu_mode();
301 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
307 static const char errata93_warning[] =
308 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
309 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
310 KERN_ERR "******* Please consider a BIOS update.\n"
311 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
314 /* Workaround for K8 erratum #93 & buggy BIOS.
315 BIOS SMM functions are required to use a specific workaround
316 to avoid corruption of the 64bit RIP register on C stepping K8.
317 A lot of BIOS that didn't get tested properly miss this.
318 The OS sees this as a page fault with the upper 32bits of RIP cleared.
319 Try to work around it here.
320 Note we only handle faults in kernel here.
321 Does nothing for X86_32
323 static int is_errata93(struct pt_regs *regs, unsigned long address)
327 if (address != regs->ip)
329 if ((address >> 32) != 0)
331 address |= 0xffffffffUL << 32;
332 if ((address >= (u64)_stext && address <= (u64)_etext) ||
333 (address >= MODULES_VADDR && address <= MODULES_END)) {
335 printk(errata93_warning);
346 * Work around K8 erratum #100 K8 in compat mode occasionally jumps to illegal
347 * addresses >4GB. We catch this in the page fault handler because these
348 * addresses are not reachable. Just detect this case and return. Any code
349 * segment in LDT is compatibility mode.
351 static int is_errata100(struct pt_regs *regs, unsigned long address)
354 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
361 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
363 #ifdef CONFIG_X86_F00F_BUG
366 * Pentium F0 0F C7 C8 bug workaround.
368 if (boot_cpu_data.f00f_bug) {
369 nr = (address - idt_descr.address) >> 3;
372 do_invalid_op(regs, 0);
380 static void show_fault_oops(struct pt_regs *regs, unsigned long error_code,
381 unsigned long address)
384 if (!oops_may_print())
388 #ifdef CONFIG_X86_PAE
389 if (error_code & PF_INSTR) {
391 pte_t *pte = lookup_address(address, &level);
393 if (pte && pte_present(*pte) && !pte_exec(*pte))
394 printk(KERN_CRIT "kernel tried to execute "
395 "NX-protected page - exploit attempt? "
396 "(uid: %d)\n", current->uid);
400 printk(KERN_ALERT "BUG: unable to handle kernel ");
401 if (address < PAGE_SIZE)
402 printk(KERN_CONT "NULL pointer dereference");
404 printk(KERN_CONT "paging request");
405 printk(KERN_CONT " at %p\n", (void *) address);
406 printk(KERN_ALERT "IP:");
407 printk_address(regs->ip, 1);
408 dump_pagetable(address);
412 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
413 unsigned long error_code)
415 unsigned long flags = oops_begin();
416 struct task_struct *tsk;
418 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
419 current->comm, address);
420 dump_pagetable(address);
422 tsk->thread.cr2 = address;
423 tsk->thread.trap_no = 14;
424 tsk->thread.error_code = error_code;
425 if (__die("Bad pagetable", regs, error_code))
427 oops_end(flags, regs, SIGKILL);
431 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
433 if ((error_code & PF_WRITE) && !pte_write(*pte))
435 if ((error_code & PF_INSTR) && !pte_exec(*pte))
442 * Handle a spurious fault caused by a stale TLB entry. This allows
443 * us to lazily refresh the TLB when increasing the permissions of a
444 * kernel page (RO -> RW or NX -> X). Doing it eagerly is very
445 * expensive since that implies doing a full cross-processor TLB
446 * flush, even if no stale TLB entries exist on other processors.
447 * There are no security implications to leaving a stale TLB when
448 * increasing the permissions on a page.
450 static int spurious_fault(unsigned long address,
451 unsigned long error_code)
458 /* Reserved-bit violation or user access to kernel space? */
459 if (error_code & (PF_USER | PF_RSVD))
462 pgd = init_mm.pgd + pgd_index(address);
463 if (!pgd_present(*pgd))
466 pud = pud_offset(pgd, address);
467 if (!pud_present(*pud))
471 return spurious_fault_check(error_code, (pte_t *) pud);
473 pmd = pmd_offset(pud, address);
474 if (!pmd_present(*pmd))
478 return spurious_fault_check(error_code, (pte_t *) pmd);
480 pte = pte_offset_kernel(pmd, address);
481 if (!pte_present(*pte))
484 return spurious_fault_check(error_code, pte);
489 * Handle a fault on the vmalloc or module mapping area
492 * Handle a fault on the vmalloc area
494 * This assumes no large pages in there.
496 static int vmalloc_fault(unsigned long address)
499 unsigned long pgd_paddr;
503 /* Make sure we are in vmalloc area */
504 if (!(address >= VMALLOC_START && address < VMALLOC_END))
508 * Synchronize this task's top level page-table
509 * with the 'reference' page table.
511 * Do _not_ use "current" here. We might be inside
512 * an interrupt in the middle of a task switch..
514 pgd_paddr = read_cr3();
515 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
518 pte_k = pte_offset_kernel(pmd_k, address);
519 if (!pte_present(*pte_k))
523 pgd_t *pgd, *pgd_ref;
524 pud_t *pud, *pud_ref;
525 pmd_t *pmd, *pmd_ref;
526 pte_t *pte, *pte_ref;
528 /* Make sure we are in vmalloc area */
529 if (!(address >= VMALLOC_START && address < VMALLOC_END))
532 /* Copy kernel mappings over when needed. This can also
533 happen within a race in page table update. In the later
536 pgd = pgd_offset(current->mm ?: &init_mm, address);
537 pgd_ref = pgd_offset_k(address);
538 if (pgd_none(*pgd_ref))
541 set_pgd(pgd, *pgd_ref);
543 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
545 /* Below here mismatches are bugs because these lower tables
548 pud = pud_offset(pgd, address);
549 pud_ref = pud_offset(pgd_ref, address);
550 if (pud_none(*pud_ref))
552 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
554 pmd = pmd_offset(pud, address);
555 pmd_ref = pmd_offset(pud_ref, address);
556 if (pmd_none(*pmd_ref))
558 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
560 pte_ref = pte_offset_kernel(pmd_ref, address);
561 if (!pte_present(*pte_ref))
563 pte = pte_offset_kernel(pmd, address);
564 /* Don't use pte_page here, because the mappings can point
565 outside mem_map, and the NUMA hash lookup cannot handle
567 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
573 int show_unhandled_signals = 1;
576 * This routine handles page faults. It determines the address,
577 * and the problem, and then passes it off to one of the appropriate
583 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
585 struct task_struct *tsk;
586 struct mm_struct *mm;
587 struct vm_area_struct *vma;
588 unsigned long address;
596 * We can fault from pretty much anywhere, with unknown IRQ state.
598 trace_hardirqs_fixup();
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 vmalloc
650 fault has been handled. */
651 if (regs->flags & (X86_EFLAGS_IF | X86_VM_MASK))
655 * If we're in an interrupt, have no user context or are running in an
656 * atomic region then we must not take the fault.
658 if (in_atomic() || !mm)
659 goto bad_area_nosemaphore;
660 #else /* CONFIG_X86_64 */
661 if (likely(regs->flags & X86_EFLAGS_IF))
664 if (unlikely(error_code & PF_RSVD))
665 pgtable_bad(address, regs, error_code);
668 * If we're in an interrupt, have no user context or are running in an
669 * atomic region then we must not take the fault.
671 if (unlikely(in_atomic() || !mm))
672 goto bad_area_nosemaphore;
675 * User-mode registers count as a user access even for any
676 * potential system fault or CPU buglet.
678 if (user_mode_vm(regs))
679 error_code |= PF_USER;
682 /* When running in the kernel we expect faults to occur only to
683 * addresses in user space. All other faults represent errors in the
684 * kernel and should generate an OOPS. Unfortunately, in the case of an
685 * erroneous fault occurring in a code path which already holds mmap_sem
686 * we will deadlock attempting to validate the fault against the
687 * address space. Luckily the kernel only validly references user
688 * space from well defined areas of code, which are listed in the
691 * As the vast majority of faults will be valid we will only perform
692 * the source reference check when there is a possibility of a deadlock.
693 * Attempt to lock the address space, if we cannot we then validate the
694 * source. If this is invalid we can skip the address space check,
695 * thus avoiding the deadlock.
697 if (!down_read_trylock(&mm->mmap_sem)) {
698 if ((error_code & PF_USER) == 0 &&
699 !search_exception_tables(regs->ip))
700 goto bad_area_nosemaphore;
701 down_read(&mm->mmap_sem);
704 vma = find_vma(mm, address);
707 if (vma->vm_start <= address)
709 if (!(vma->vm_flags & VM_GROWSDOWN))
711 if (error_code & PF_USER) {
713 * Accessing the stack below %sp is always a bug.
714 * The large cushion allows instructions like enter
715 * and pusha to work. ("enter $65535,$31" pushes
716 * 32 pointers and then decrements %sp by 65535.)
718 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
721 if (expand_stack(vma, address))
724 * Ok, we have a good vm_area for this memory access, so
728 si_code = SEGV_ACCERR;
730 switch (error_code & (PF_PROT|PF_WRITE)) {
731 default: /* 3: write, present */
733 case PF_WRITE: /* write, not present */
734 if (!(vma->vm_flags & VM_WRITE))
738 case PF_PROT: /* read, present */
740 case 0: /* read, not present */
741 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
749 * If for any reason at all we couldn't handle the fault,
750 * make sure we exit gracefully rather than endlessly redo
753 fault = handle_mm_fault(mm, vma, address, write);
754 if (unlikely(fault & VM_FAULT_ERROR)) {
755 if (fault & VM_FAULT_OOM)
757 else if (fault & VM_FAULT_SIGBUS)
761 if (fault & VM_FAULT_MAJOR)
768 * Did it hit the DOS screen memory VA from vm86 mode?
770 if (v8086_mode(regs)) {
771 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
773 tsk->thread.screen_bitmap |= 1 << bit;
776 up_read(&mm->mmap_sem);
780 * Something tried to access memory that isn't in our memory map..
781 * Fix it, but check if it's kernel or user first..
784 up_read(&mm->mmap_sem);
786 bad_area_nosemaphore:
787 /* User mode accesses just cause a SIGSEGV */
788 if (error_code & PF_USER) {
790 * It's possible to have interrupts off here.
795 * Valid to do another page fault here because this one came
798 if (is_prefetch(regs, address, error_code))
801 if (is_errata100(regs, address))
804 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
805 printk_ratelimit()) {
807 "%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
808 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
809 tsk->comm, task_pid_nr(tsk), address,
810 (void *) regs->ip, (void *) regs->sp, error_code);
811 print_vma_addr(" in ", regs->ip);
815 tsk->thread.cr2 = address;
816 /* Kernel addresses are always protection faults */
817 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
818 tsk->thread.trap_no = 14;
819 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
823 if (is_f00f_bug(regs, address))
827 /* Are we prepared to handle this kernel fault? */
828 if (fixup_exception(regs))
833 * Valid to do another page fault here, because if this fault
834 * had been triggered by is_prefetch fixup_exception would have
838 * Hall of shame of CPU/BIOS bugs.
840 if (is_prefetch(regs, address, error_code))
843 if (is_errata93(regs, address))
847 * Oops. The kernel tried to access some bad page. We'll have to
848 * terminate things with extreme prejudice.
853 flags = oops_begin();
856 show_fault_oops(regs, error_code, address);
858 tsk->thread.cr2 = address;
859 tsk->thread.trap_no = 14;
860 tsk->thread.error_code = error_code;
863 die("Oops", regs, error_code);
867 if (__die("Oops", regs, error_code))
869 /* Executive summary in case the body of the oops scrolled away */
870 printk(KERN_EMERG "CR2: %016lx\n", address);
871 oops_end(flags, regs, SIGKILL);
875 * We ran out of memory, or some other thing happened to us that made
876 * us unable to handle the page fault gracefully.
879 up_read(&mm->mmap_sem);
880 if (is_global_init(tsk)) {
883 down_read(&mm->mmap_sem);
890 printk("VM: killing process %s\n", tsk->comm);
891 if (error_code & PF_USER)
892 do_group_exit(SIGKILL);
896 up_read(&mm->mmap_sem);
898 /* Kernel mode? Handle exceptions or die */
899 if (!(error_code & PF_USER))
902 /* User space => ok to do another page fault */
903 if (is_prefetch(regs, address, error_code))
906 tsk->thread.cr2 = address;
907 tsk->thread.error_code = error_code;
908 tsk->thread.trap_no = 14;
909 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
912 DEFINE_SPINLOCK(pgd_lock);
915 void vmalloc_sync_all(void)
918 unsigned long start = VMALLOC_START & PGDIR_MASK;
919 unsigned long address;
921 if (SHARED_KERNEL_PMD)
924 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
925 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
929 spin_lock_irqsave(&pgd_lock, flags);
930 list_for_each_entry(page, &pgd_list, lru) {
931 if (!vmalloc_sync_one(page_address(page),
935 spin_unlock_irqrestore(&pgd_lock, flags);
937 #else /* CONFIG_X86_64 */
938 unsigned long start = VMALLOC_START & PGDIR_MASK;
939 unsigned long address;
941 for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
942 const pgd_t *pgd_ref = pgd_offset_k(address);
946 if (pgd_none(*pgd_ref))
948 spin_lock_irqsave(&pgd_lock, flags);
949 list_for_each_entry(page, &pgd_list, lru) {
951 pgd = (pgd_t *)page_address(page) + pgd_index(address);
953 set_pgd(pgd, *pgd_ref);
955 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
957 spin_unlock_irqrestore(&pgd_lock, flags);