2 * Copyright (C) 1995 Linus Torvalds
5 #include <linux/signal.h>
6 #include <linux/sched.h>
7 #include <linux/kernel.h>
8 #include <linux/errno.h>
9 #include <linux/string.h>
10 #include <linux/types.h>
11 #include <linux/ptrace.h>
12 #include <linux/mman.h>
14 #include <linux/smp.h>
15 #include <linux/interrupt.h>
16 #include <linux/init.h>
17 #include <linux/tty.h>
18 #include <linux/vt_kern.h> /* For unblank_screen() */
19 #include <linux/highmem.h>
20 #include <linux/bootmem.h> /* for max_low_pfn */
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>
29 #include <asm/segment.h>
32 * Page fault error code bits
33 * bit 0 == 0 means no page found, 1 means protection fault
34 * bit 1 == 0 means read, 1 means write
35 * bit 2 == 0 means kernel, 1 means user-mode
36 * bit 3 == 1 means use of reserved bit detected
37 * bit 4 == 1 means fault was an instruction fetch
39 #define PF_PROT (1<<0)
40 #define PF_WRITE (1<<1)
41 #define PF_USER (1<<2)
42 #define PF_RSVD (1<<3)
43 #define PF_INSTR (1<<4)
45 static inline int notify_page_fault(struct pt_regs *regs)
50 /* kprobe_running() needs smp_processor_id() */
51 if (!user_mode_vm(regs)) {
53 if (kprobe_running() && kprobe_fault_handler(regs, 14))
66 * Return EIP plus the CS segment base. The segment limit is also
67 * adjusted, clamped to the kernel/user address space (whichever is
68 * appropriate), and returned in *eip_limit.
70 * The segment is checked, because it might have been changed by another
71 * task between the original faulting instruction and here.
73 * If CS is no longer a valid code segment, or if EIP is beyond the
74 * limit, or if it is a kernel address when CS is not a kernel segment,
75 * then the returned value will be greater than *eip_limit.
77 * This is slow, but is very rarely executed.
79 static inline unsigned long get_segment_eip(struct pt_regs *regs,
80 unsigned long *eip_limit)
82 unsigned long ip = regs->ip;
83 unsigned seg = regs->cs & 0xffff;
84 u32 seg_ar, seg_limit, base, *desc;
86 /* Unlikely, but must come before segment checks. */
87 if (unlikely(regs->flags & VM_MASK)) {
89 *eip_limit = base + 0xffff;
90 return base + (ip & 0xffff);
93 /* The standard kernel/user address space limit. */
94 *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
96 /* By far the most common cases. */
97 if (likely(SEGMENT_IS_FLAT_CODE(seg)))
100 /* Check the segment exists, is within the current LDT/GDT size,
101 that kernel/user (ring 0..3) has the appropriate privilege,
102 that it's a code segment, and get the limit. */
103 __asm__ ("larl %3,%0; lsll %3,%1"
104 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
105 if ((~seg_ar & 0x9800) || ip > seg_limit) {
107 return 1; /* So that returned ip > *eip_limit. */
110 /* Get the GDT/LDT descriptor base.
111 When you look for races in this code remember that
112 LDT and other horrors are only used in user space. */
114 /* Must lock the LDT while reading it. */
115 mutex_lock(¤t->mm->context.lock);
116 desc = current->mm->context.ldt;
117 desc = (void *)desc + (seg & ~7);
119 /* Must disable preemption while reading the GDT. */
120 desc = (u32 *)get_cpu_gdt_table(get_cpu());
121 desc = (void *)desc + (seg & ~7);
124 /* Decode the code segment base from the descriptor */
125 base = get_desc_base((struct desc_struct *)desc);
128 mutex_unlock(¤t->mm->context.lock);
132 /* Adjust EIP and segment limit, and clamp at the kernel limit.
133 It's legitimate for segments to wrap at 0xffffffff. */
135 if (seg_limit < *eip_limit && seg_limit >= base)
136 *eip_limit = seg_limit;
143 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
144 * Check that here and ignore it.
147 * Sometimes the CPU reports invalid exceptions on prefetch.
148 * Check that here and ignore it.
150 * Opcode checker based on code by Richard Brunner
152 static int is_prefetch(struct pt_regs *regs, unsigned long addr,
153 unsigned long error_code)
155 unsigned char *instr;
158 unsigned char *max_instr;
162 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
163 boot_cpu_data.x86 >= 6)) {
164 /* Catch an obscure case of prefetch inside an NX page. */
165 if (nx_enabled && (error_code & PF_INSTR))
170 instr = (unsigned char *)get_segment_eip(regs, &limit);
172 /* If it was a exec fault ignore */
173 if (error_code & PF_INSTR)
175 instr = (unsigned char __user *)convert_rip_to_linear(current, regs);
178 max_instr = instr + 15;
181 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
185 while (scan_more && instr < max_instr) {
186 unsigned char opcode;
187 unsigned char instr_hi;
188 unsigned char instr_lo;
191 if (instr > (unsigned char *)limit)
194 if (probe_kernel_address(instr, opcode))
197 instr_hi = opcode & 0xf0;
198 instr_lo = opcode & 0x0f;
205 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
206 * In X86_64 long mode, the CPU will signal invalid
207 * opcode if some of these prefixes are present so
208 * X86_64 will never get here anyway
210 scan_more = ((instr_lo & 7) == 0x6);
215 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
216 * Need to figure out under what instruction mode the
217 * instruction was issued. Could check the LDT for lm,
218 * but for now it's good enough to assume that long
219 * mode only uses well known segments or kernel.
221 scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
225 /* 0x64 thru 0x67 are valid prefixes in all modes. */
226 scan_more = (instr_lo & 0xC) == 0x4;
229 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
230 scan_more = !instr_lo || (instr_lo>>1) == 1;
233 /* Prefetch instruction is 0x0F0D or 0x0F18 */
236 if (instr > (unsigned char *)limit)
239 if (probe_kernel_address(instr, opcode))
241 prefetch = (instr_lo == 0xF) &&
242 (opcode == 0x0D || opcode == 0x18);
252 static noinline void force_sig_info_fault(int si_signo, int si_code,
253 unsigned long address, struct task_struct *tsk)
257 info.si_signo = si_signo;
259 info.si_code = si_code;
260 info.si_addr = (void __user *)address;
261 force_sig_info(si_signo, &info, tsk);
264 void do_invalid_op(struct pt_regs *, unsigned long);
266 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
268 unsigned index = pgd_index(address);
274 pgd_k = init_mm.pgd + index;
276 if (!pgd_present(*pgd_k))
280 * set_pgd(pgd, *pgd_k); here would be useless on PAE
281 * and redundant with the set_pmd() on non-PAE. As would
285 pud = pud_offset(pgd, address);
286 pud_k = pud_offset(pgd_k, address);
287 if (!pud_present(*pud_k))
290 pmd = pmd_offset(pud, address);
291 pmd_k = pmd_offset(pud_k, address);
292 if (!pmd_present(*pmd_k))
294 if (!pmd_present(*pmd)) {
295 set_pmd(pmd, *pmd_k);
296 arch_flush_lazy_mmu_mode();
298 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
303 static const char errata93_warning[] =
304 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
305 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
306 KERN_ERR "******* Please consider a BIOS update.\n"
307 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
309 /* Workaround for K8 erratum #93 & buggy BIOS.
310 BIOS SMM functions are required to use a specific workaround
311 to avoid corruption of the 64bit RIP register on C stepping K8.
312 A lot of BIOS that didn't get tested properly miss this.
313 The OS sees this as a page fault with the upper 32bits of RIP cleared.
314 Try to work around it here.
315 Note we only handle faults in kernel here. */
317 static int is_errata93(struct pt_regs *regs, unsigned long address)
320 if (address != regs->ip)
322 if ((address >> 32) != 0)
324 address |= 0xffffffffUL << 32;
325 if ((address >= (u64)_stext && address <= (u64)_etext) ||
326 (address >= MODULES_VADDR && address <= MODULES_END)) {
328 printk(errata93_warning);
339 * Handle a fault on the vmalloc or module mapping area
341 * This assumes no large pages in there.
343 static inline int vmalloc_fault(unsigned long address)
345 unsigned long pgd_paddr;
349 * Synchronize this task's top level page-table
350 * with the 'reference' page table.
352 * Do _not_ use "current" here. We might be inside
353 * an interrupt in the middle of a task switch..
355 pgd_paddr = read_cr3();
356 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
359 pte_k = pte_offset_kernel(pmd_k, address);
360 if (!pte_present(*pte_k))
365 int show_unhandled_signals = 1;
368 * This routine handles page faults. It determines the address,
369 * and the problem, and then passes it off to one of the appropriate
372 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
374 struct task_struct *tsk;
375 struct mm_struct *mm;
376 struct vm_area_struct *vma;
377 unsigned long address;
382 * We can fault from pretty much anywhere, with unknown IRQ state.
384 trace_hardirqs_fixup();
386 /* get the address */
387 address = read_cr2();
391 si_code = SEGV_MAPERR;
394 * We fault-in kernel-space virtual memory on-demand. The
395 * 'reference' page table is init_mm.pgd.
397 * NOTE! We MUST NOT take any locks for this case. We may
398 * be in an interrupt or a critical region, and should
399 * only copy the information from the master page table,
402 * This verifies that the fault happens in kernel space
403 * (error_code & 4) == 0, and that the fault was not a
404 * protection error (error_code & 9) == 0.
406 if (unlikely(address >= TASK_SIZE)) {
407 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
409 if (notify_page_fault(regs))
412 * Don't take the mm semaphore here. If we fixup a prefetch
413 * fault we could otherwise deadlock.
415 goto bad_area_nosemaphore;
418 if (notify_page_fault(regs))
421 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
422 fault has been handled. */
423 if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
429 * If we're in an interrupt, have no user context or are running in an
430 * atomic region then we must not take the fault.
432 if (in_atomic() || !mm)
433 goto bad_area_nosemaphore;
435 /* When running in the kernel we expect faults to occur only to
436 * addresses in user space. All other faults represent errors in the
437 * kernel and should generate an OOPS. Unfortunately, in the case of an
438 * erroneous fault occurring in a code path which already holds mmap_sem
439 * we will deadlock attempting to validate the fault against the
440 * address space. Luckily the kernel only validly references user
441 * space from well defined areas of code, which are listed in the
444 * As the vast majority of faults will be valid we will only perform
445 * the source reference check when there is a possibility of a deadlock.
446 * Attempt to lock the address space, if we cannot we then validate the
447 * source. If this is invalid we can skip the address space check,
448 * thus avoiding the deadlock.
450 if (!down_read_trylock(&mm->mmap_sem)) {
451 if ((error_code & PF_USER) == 0 &&
452 !search_exception_tables(regs->ip))
453 goto bad_area_nosemaphore;
454 down_read(&mm->mmap_sem);
457 vma = find_vma(mm, address);
460 if (vma->vm_start <= address)
462 if (!(vma->vm_flags & VM_GROWSDOWN))
464 if (error_code & PF_USER) {
466 * Accessing the stack below %sp is always a bug.
467 * The large cushion allows instructions like enter
468 * and pusha to work. ("enter $65535,$31" pushes
469 * 32 pointers and then decrements %sp by 65535.)
471 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
474 if (expand_stack(vma, address))
477 * Ok, we have a good vm_area for this memory access, so
481 si_code = SEGV_ACCERR;
483 switch (error_code & (PF_PROT|PF_WRITE)) {
484 default: /* 3: write, present */
486 case PF_WRITE: /* write, not present */
487 if (!(vma->vm_flags & VM_WRITE))
491 case PF_PROT: /* read, present */
493 case 0: /* read, not present */
494 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
500 * If for any reason at all we couldn't handle the fault,
501 * make sure we exit gracefully rather than endlessly redo
504 fault = handle_mm_fault(mm, vma, address, write);
505 if (unlikely(fault & VM_FAULT_ERROR)) {
506 if (fault & VM_FAULT_OOM)
508 else if (fault & VM_FAULT_SIGBUS)
512 if (fault & VM_FAULT_MAJOR)
518 * Did it hit the DOS screen memory VA from vm86 mode?
520 if (regs->flags & VM_MASK) {
521 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
523 tsk->thread.screen_bitmap |= 1 << bit;
525 up_read(&mm->mmap_sem);
529 * Something tried to access memory that isn't in our memory map..
530 * Fix it, but check if it's kernel or user first..
533 up_read(&mm->mmap_sem);
535 bad_area_nosemaphore:
536 /* User mode accesses just cause a SIGSEGV */
537 if (error_code & PF_USER) {
539 * It's possible to have interrupts off here.
544 * Valid to do another page fault here because this one came
547 if (is_prefetch(regs, address, error_code))
550 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
551 printk_ratelimit()) {
552 printk("%s%s[%d]: segfault at %08lx ip %08lx "
553 "sp %08lx error %lx\n",
554 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
555 tsk->comm, task_pid_nr(tsk), address, regs->ip,
556 regs->sp, error_code);
558 tsk->thread.cr2 = address;
559 /* Kernel addresses are always protection faults */
560 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
561 tsk->thread.trap_no = 14;
562 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
566 #ifdef CONFIG_X86_F00F_BUG
568 * Pentium F0 0F C7 C8 bug workaround.
570 if (boot_cpu_data.f00f_bug) {
573 nr = (address - idt_descr.address) >> 3;
576 do_invalid_op(regs, 0);
583 /* Are we prepared to handle this kernel fault? */
584 if (fixup_exception(regs))
588 * Valid to do another page fault here, because if this fault
589 * had been triggered by is_prefetch fixup_exception would have
592 if (is_prefetch(regs, address, error_code))
596 * Oops. The kernel tried to access some bad page. We'll have to
597 * terminate things with extreme prejudice.
602 if (oops_may_print()) {
603 __typeof__(pte_val(__pte(0))) page;
605 #ifdef CONFIG_X86_PAE
606 if (error_code & 16) {
607 pte_t *pte = lookup_address(address);
609 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
610 printk(KERN_CRIT "kernel tried to execute "
611 "NX-protected page - exploit attempt? "
612 "(uid: %d)\n", current->uid);
615 if (address < PAGE_SIZE)
616 printk(KERN_ALERT "BUG: unable to handle kernel NULL "
617 "pointer dereference");
619 printk(KERN_ALERT "BUG: unable to handle kernel paging"
621 printk(" at virtual address %08lx\n", address);
622 printk(KERN_ALERT "printing ip: %08lx ", regs->ip);
625 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
626 #ifdef CONFIG_X86_PAE
627 printk("*pdpt = %016Lx ", page);
628 if ((page >> PAGE_SHIFT) < max_low_pfn
629 && page & _PAGE_PRESENT) {
631 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
632 & (PTRS_PER_PMD - 1)];
633 printk(KERN_CONT "*pde = %016Lx ", page);
637 printk("*pde = %08lx ", page);
641 * We must not directly access the pte in the highpte
642 * case if the page table is located in highmem.
643 * And let's rather not kmap-atomic the pte, just in case
644 * it's allocated already.
646 if ((page >> PAGE_SHIFT) < max_low_pfn
647 && (page & _PAGE_PRESENT)
648 && !(page & _PAGE_PSE)) {
650 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
651 & (PTRS_PER_PTE - 1)];
652 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
658 tsk->thread.cr2 = address;
659 tsk->thread.trap_no = 14;
660 tsk->thread.error_code = error_code;
661 die("Oops", regs, error_code);
666 * We ran out of memory, or some other thing happened to us that made
667 * us unable to handle the page fault gracefully.
670 up_read(&mm->mmap_sem);
671 if (is_global_init(tsk)) {
673 down_read(&mm->mmap_sem);
676 printk("VM: killing process %s\n", tsk->comm);
678 do_group_exit(SIGKILL);
682 up_read(&mm->mmap_sem);
684 /* Kernel mode? Handle exceptions or die */
685 if (!(error_code & PF_USER))
688 /* User space => ok to do another page fault */
689 if (is_prefetch(regs, address, error_code))
692 tsk->thread.cr2 = address;
693 tsk->thread.error_code = error_code;
694 tsk->thread.trap_no = 14;
695 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
698 void vmalloc_sync_all(void)
701 * Note that races in the updates of insync and start aren't
702 * problematic: insync can only get set bits added, and updates to
703 * start are only improving performance (without affecting correctness
706 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
707 static unsigned long start = TASK_SIZE;
708 unsigned long address;
710 if (SHARED_KERNEL_PMD)
713 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
714 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
715 if (!test_bit(pgd_index(address), insync)) {
719 spin_lock_irqsave(&pgd_lock, flags);
720 for (page = pgd_list; page; page =
721 (struct page *)page->index)
722 if (!vmalloc_sync_one(page_address(page),
724 BUG_ON(page != pgd_list);
727 spin_unlock_irqrestore(&pgd_lock, flags);
729 set_bit(pgd_index(address), insync);
731 if (address == start && test_bit(pgd_index(address), insync))
732 start = address + PGDIR_SIZE;