2 * linux/arch/i386/mm/fault.c
4 * Copyright (C) 1995 Linus Torvalds
7 #include <linux/signal.h>
8 #include <linux/sched.h>
9 #include <linux/kernel.h>
10 #include <linux/errno.h>
11 #include <linux/string.h>
12 #include <linux/types.h>
13 #include <linux/ptrace.h>
14 #include <linux/mman.h>
16 #include <linux/smp.h>
17 #include <linux/smp_lock.h>
18 #include <linux/interrupt.h>
19 #include <linux/init.h>
20 #include <linux/tty.h>
21 #include <linux/vt_kern.h> /* For unblank_screen() */
22 #include <linux/highmem.h>
23 #include <linux/module.h>
24 #include <linux/kprobes.h>
25 #include <linux/uaccess.h>
27 #include <asm/system.h>
29 #include <asm/kdebug.h>
30 #include <asm/segment.h>
32 extern void die(const char *,struct pt_regs *,long);
34 static ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
36 int register_page_fault_notifier(struct notifier_block *nb)
39 return atomic_notifier_chain_register(¬ify_page_fault_chain, nb);
41 EXPORT_SYMBOL_GPL(register_page_fault_notifier);
43 int unregister_page_fault_notifier(struct notifier_block *nb)
45 return atomic_notifier_chain_unregister(¬ify_page_fault_chain, nb);
47 EXPORT_SYMBOL_GPL(unregister_page_fault_notifier);
49 static inline int notify_page_fault(struct pt_regs *regs, long err)
51 struct die_args args = {
58 return atomic_notifier_call_chain(¬ify_page_fault_chain,
59 DIE_PAGE_FAULT, &args);
63 * Return EIP plus the CS segment base. The segment limit is also
64 * adjusted, clamped to the kernel/user address space (whichever is
65 * appropriate), and returned in *eip_limit.
67 * The segment is checked, because it might have been changed by another
68 * task between the original faulting instruction and here.
70 * If CS is no longer a valid code segment, or if EIP is beyond the
71 * limit, or if it is a kernel address when CS is not a kernel segment,
72 * then the returned value will be greater than *eip_limit.
74 * This is slow, but is very rarely executed.
76 static inline unsigned long get_segment_eip(struct pt_regs *regs,
77 unsigned long *eip_limit)
79 unsigned long eip = regs->eip;
80 unsigned seg = regs->xcs & 0xffff;
81 u32 seg_ar, seg_limit, base, *desc;
83 /* Unlikely, but must come before segment checks. */
84 if (unlikely(regs->eflags & VM_MASK)) {
86 *eip_limit = base + 0xffff;
87 return base + (eip & 0xffff);
90 /* The standard kernel/user address space limit. */
91 *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
93 /* By far the most common cases. */
94 if (likely(SEGMENT_IS_FLAT_CODE(seg)))
97 /* Check the segment exists, is within the current LDT/GDT size,
98 that kernel/user (ring 0..3) has the appropriate privilege,
99 that it's a code segment, and get the limit. */
100 __asm__ ("larl %3,%0; lsll %3,%1"
101 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
102 if ((~seg_ar & 0x9800) || eip > seg_limit) {
104 return 1; /* So that returned eip > *eip_limit. */
107 /* Get the GDT/LDT descriptor base.
108 When you look for races in this code remember that
109 LDT and other horrors are only used in user space. */
111 /* Must lock the LDT while reading it. */
112 down(¤t->mm->context.sem);
113 desc = current->mm->context.ldt;
114 desc = (void *)desc + (seg & ~7);
116 /* Must disable preemption while reading the GDT. */
117 desc = (u32 *)get_cpu_gdt_table(get_cpu());
118 desc = (void *)desc + (seg & ~7);
121 /* Decode the code segment base from the descriptor */
122 base = get_desc_base((unsigned long *)desc);
125 up(¤t->mm->context.sem);
129 /* Adjust EIP and segment limit, and clamp at the kernel limit.
130 It's legitimate for segments to wrap at 0xffffffff. */
132 if (seg_limit < *eip_limit && seg_limit >= base)
133 *eip_limit = seg_limit;
138 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
139 * Check that here and ignore it.
141 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
144 unsigned char *instr = (unsigned char *)get_segment_eip (regs, &limit);
149 for (i = 0; scan_more && i < 15; i++) {
150 unsigned char opcode;
151 unsigned char instr_hi;
152 unsigned char instr_lo;
154 if (instr > (unsigned char *)limit)
156 if (probe_kernel_address(instr, opcode))
159 instr_hi = opcode & 0xf0;
160 instr_lo = opcode & 0x0f;
166 /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
167 scan_more = ((instr_lo & 7) == 0x6);
171 /* 0x64 thru 0x67 are valid prefixes in all modes. */
172 scan_more = (instr_lo & 0xC) == 0x4;
175 /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
176 scan_more = !instr_lo || (instr_lo>>1) == 1;
179 /* Prefetch instruction is 0x0F0D or 0x0F18 */
181 if (instr > (unsigned char *)limit)
183 if (probe_kernel_address(instr, opcode))
185 prefetch = (instr_lo == 0xF) &&
186 (opcode == 0x0D || opcode == 0x18);
196 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
197 unsigned long error_code)
199 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
200 boot_cpu_data.x86 >= 6)) {
201 /* Catch an obscure case of prefetch inside an NX page. */
202 if (nx_enabled && (error_code & 16))
204 return __is_prefetch(regs, addr);
209 static noinline void force_sig_info_fault(int si_signo, int si_code,
210 unsigned long address, struct task_struct *tsk)
214 info.si_signo = si_signo;
216 info.si_code = si_code;
217 info.si_addr = (void __user *)address;
218 force_sig_info(si_signo, &info, tsk);
221 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
223 #ifdef CONFIG_X86_PAE
224 static void dump_fault_path(unsigned long address)
226 unsigned long *p, page;
230 p = (unsigned long *)__va(page);
231 p += (address >> 30) * 2;
232 printk(KERN_ALERT "%08lx -> *pde = %08lx:%08lx\n", page, p[1], p[0]);
234 mfn = (p[0] >> PAGE_SHIFT) | (p[1] << 20);
235 page = mfn_to_pfn(mfn) << PAGE_SHIFT;
236 p = (unsigned long *)__va(page);
237 address &= 0x3fffffff;
238 p += (address >> 21) * 2;
239 printk(KERN_ALERT "%08lx -> *pme = %08lx:%08lx\n",
241 mfn = (p[0] >> PAGE_SHIFT) | (p[1] << 20);
242 #ifdef CONFIG_HIGHPTE
243 if (mfn_to_pfn(mfn) >= highstart_pfn)
247 page = mfn_to_pfn(mfn) << PAGE_SHIFT;
248 p = (unsigned long *) __va(page);
249 address &= 0x001fffff;
250 p += (address >> 12) * 2;
251 printk(KERN_ALERT "%08lx -> *pte = %08lx:%08lx\n",
257 static void dump_fault_path(unsigned long address)
262 page = ((unsigned long *) __va(page))[address >> 22];
263 if (oops_may_print())
264 printk(KERN_ALERT "*pde = ma %08lx pa %08lx\n", page,
265 machine_to_phys(page));
267 * We must not directly access the pte in the highpte
268 * case if the page table is located in highmem.
269 * And lets rather not kmap-atomic the pte, just in case
270 * it's allocated already.
272 #ifdef CONFIG_HIGHPTE
273 if ((page >> PAGE_SHIFT) >= highstart_pfn)
276 if ((page & 1) && oops_may_print()) {
278 address &= 0x003ff000;
279 page = machine_to_phys(page);
280 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
281 printk(KERN_ALERT "*pte = ma %08lx pa %08lx\n", page,
282 machine_to_phys(page));
287 static int spurious_fault(struct pt_regs *regs,
288 unsigned long address,
289 unsigned long error_code)
296 /* Reserved-bit violation or user access to kernel space? */
297 if (error_code & 0x0c)
300 pgd = init_mm.pgd + pgd_index(address);
301 if (!pgd_present(*pgd))
304 pud = pud_offset(pgd, address);
305 if (!pud_present(*pud))
308 pmd = pmd_offset(pud, address);
309 if (!pmd_present(*pmd))
312 pte = pte_offset_kernel(pmd, address);
313 if (!pte_present(*pte))
315 if ((error_code & 0x02) && !pte_write(*pte))
317 #ifdef CONFIG_X86_PAE
318 if ((error_code & 0x10) && (pte_val(*pte) & _PAGE_NX))
325 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
327 unsigned index = pgd_index(address);
333 pgd_k = init_mm.pgd + index;
335 if (!pgd_present(*pgd_k))
339 * set_pgd(pgd, *pgd_k); here would be useless on PAE
340 * and redundant with the set_pmd() on non-PAE. As would
344 pud = pud_offset(pgd, address);
345 pud_k = pud_offset(pgd_k, address);
346 if (!pud_present(*pud_k))
349 pmd = pmd_offset(pud, address);
350 pmd_k = pmd_offset(pud_k, address);
351 if (!pmd_present(*pmd_k))
353 if (!pmd_present(*pmd))
355 set_pmd(pmd, *pmd_k);
358 * When running on Xen we must launder *pmd_k through
359 * pmd_val() to ensure that _PAGE_PRESENT is correctly set.
361 set_pmd(pmd, __pmd(pmd_val(*pmd_k)));
364 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
369 * Handle a fault on the vmalloc or module mapping area
371 * This assumes no large pages in there.
373 static inline int vmalloc_fault(unsigned long address)
375 unsigned long pgd_paddr;
379 * Synchronize this task's top level page-table
380 * with the 'reference' page table.
382 * Do _not_ use "current" here. We might be inside
383 * an interrupt in the middle of a task switch..
385 pgd_paddr = read_cr3();
386 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
389 pte_k = pte_offset_kernel(pmd_k, address);
390 if (!pte_present(*pte_k))
396 * This routine handles page faults. It determines the address,
397 * and the problem, and then passes it off to one of the appropriate
401 * bit 0 == 0 means no page found, 1 means protection fault
402 * bit 1 == 0 means read, 1 means write
403 * bit 2 == 0 means kernel, 1 means user-mode
404 * bit 3 == 1 means use of reserved bit detected
405 * bit 4 == 1 means fault was an instruction fetch
407 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
408 unsigned long error_code)
410 struct task_struct *tsk;
411 struct mm_struct *mm;
412 struct vm_area_struct * vma;
413 unsigned long address;
416 /* get the address */
417 address = read_cr2();
419 /* Set the "privileged fault" bit to something sane. */
421 error_code |= (regs->xcs & 2) << 1;
422 if (regs->eflags & X86_EFLAGS_VM)
427 si_code = SEGV_MAPERR;
430 * We fault-in kernel-space virtual memory on-demand. The
431 * 'reference' page table is init_mm.pgd.
433 * NOTE! We MUST NOT take any locks for this case. We may
434 * be in an interrupt or a critical region, and should
435 * only copy the information from the master page table,
438 * This verifies that the fault happens in kernel space
439 * (error_code & 4) == 0, and that the fault was not a
440 * protection error (error_code & 9) == 0.
442 if (unlikely(address >= TASK_SIZE)) {
444 /* Faults in hypervisor area can never be patched up. */
445 if (address >= hypervisor_virt_start)
446 goto bad_area_nosemaphore;
448 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
450 /* Can take a spurious fault if mapping changes R/O -> R/W. */
451 if (spurious_fault(regs, address, error_code))
453 if (notify_page_fault(regs, error_code) == NOTIFY_STOP)
456 * Don't take the mm semaphore here. If we fixup a prefetch
457 * fault we could otherwise deadlock.
459 goto bad_area_nosemaphore;
462 if (notify_page_fault(regs, error_code) == NOTIFY_STOP)
465 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
466 fault has been handled. */
467 if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
473 * If we're in an interrupt, have no user context or are running in an
474 * atomic region then we must not take the fault..
476 if (in_atomic() || !mm)
477 goto bad_area_nosemaphore;
479 /* When running in the kernel we expect faults to occur only to
480 * addresses in user space. All other faults represent errors in the
481 * kernel and should generate an OOPS. Unfortunatly, in the case of an
482 * erroneous fault occurring in a code path which already holds mmap_sem
483 * we will deadlock attempting to validate the fault against the
484 * address space. Luckily the kernel only validly references user
485 * space from well defined areas of code, which are listed in the
488 * As the vast majority of faults will be valid we will only perform
489 * the source reference check when there is a possibilty of a deadlock.
490 * Attempt to lock the address space, if we cannot we then validate the
491 * source. If this is invalid we can skip the address space check,
492 * thus avoiding the deadlock.
494 if (!down_read_trylock(&mm->mmap_sem)) {
495 if ((error_code & 4) == 0 &&
496 !search_exception_tables(regs->eip))
497 goto bad_area_nosemaphore;
498 down_read(&mm->mmap_sem);
501 vma = find_vma(mm, address);
504 if (vma->vm_start <= address)
506 if (!(vma->vm_flags & VM_GROWSDOWN))
508 if (error_code & 4) {
510 * Accessing the stack below %esp is always a bug.
511 * The large cushion allows instructions like enter
512 * and pusha to work. ("enter $65535,$31" pushes
513 * 32 pointers and then decrements %esp by 65535.)
515 if (address + 65536 + 32 * sizeof(unsigned long) < regs->esp)
518 if (expand_stack(vma, address))
521 * Ok, we have a good vm_area for this memory access, so
525 si_code = SEGV_ACCERR;
527 switch (error_code & 3) {
528 default: /* 3: write, present */
530 case 2: /* write, not present */
531 if (!(vma->vm_flags & VM_WRITE))
535 case 1: /* read, present */
537 case 0: /* read, not present */
538 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
544 * If for any reason at all we couldn't handle the fault,
545 * make sure we exit gracefully rather than endlessly redo
548 switch (handle_mm_fault(mm, vma, address, write)) {
555 case VM_FAULT_SIGBUS:
564 * Did it hit the DOS screen memory VA from vm86 mode?
566 if (regs->eflags & VM_MASK) {
567 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
569 tsk->thread.screen_bitmap |= 1 << bit;
571 up_read(&mm->mmap_sem);
575 * Something tried to access memory that isn't in our memory map..
576 * Fix it, but check if it's kernel or user first..
579 up_read(&mm->mmap_sem);
581 bad_area_nosemaphore:
582 /* User mode accesses just cause a SIGSEGV */
583 if (error_code & 4) {
585 * Valid to do another page fault here because this one came
588 if (is_prefetch(regs, address, error_code))
591 tsk->thread.cr2 = address;
592 /* Kernel addresses are always protection faults */
593 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
594 tsk->thread.trap_no = 14;
595 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
599 #ifdef CONFIG_X86_F00F_BUG
601 * Pentium F0 0F C7 C8 bug workaround.
603 if (boot_cpu_data.f00f_bug) {
606 nr = (address - idt_descr.address) >> 3;
609 do_invalid_op(regs, 0);
616 /* Are we prepared to handle this kernel fault? */
617 if (fixup_exception(regs))
621 * Valid to do another page fault here, because if this fault
622 * had been triggered by is_prefetch fixup_exception would have
625 if (is_prefetch(regs, address, error_code))
629 * Oops. The kernel tried to access some bad page. We'll have to
630 * terminate things with extreme prejudice.
635 if (oops_may_print()) {
636 #ifdef CONFIG_X86_PAE
637 if (error_code & 16) {
638 pte_t *pte = lookup_address(address);
640 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
641 printk(KERN_CRIT "kernel tried to execute "
642 "NX-protected page - exploit attempt? "
643 "(uid: %d)\n", current->uid);
646 if (address < PAGE_SIZE)
647 printk(KERN_ALERT "BUG: unable to handle kernel NULL "
648 "pointer dereference");
650 printk(KERN_ALERT "BUG: unable to handle kernel paging"
652 printk(" at virtual address %08lx\n",address);
653 printk(KERN_ALERT " printing eip:\n");
654 printk("%08lx\n", regs->eip);
656 dump_fault_path(address);
657 tsk->thread.cr2 = address;
658 tsk->thread.trap_no = 14;
659 tsk->thread.error_code = error_code;
660 die("Oops", regs, error_code);
665 * We ran out of memory, or some other thing happened to us that made
666 * us unable to handle the page fault gracefully.
669 up_read(&mm->mmap_sem);
672 down_read(&mm->mmap_sem);
675 printk("VM: killing process %s\n", tsk->comm);
681 up_read(&mm->mmap_sem);
683 /* Kernel mode? Handle exceptions or die */
684 if (!(error_code & 4))
687 /* User space => ok to do another page fault */
688 if (is_prefetch(regs, address, error_code))
691 tsk->thread.cr2 = address;
692 tsk->thread.error_code = error_code;
693 tsk->thread.trap_no = 14;
694 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
697 #if !HAVE_SHARED_KERNEL_PMD
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 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
711 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
712 if (!test_bit(pgd_index(address), insync)) {
716 spin_lock_irqsave(&pgd_lock, flags);
717 for (page = pgd_list; page; page =
718 (struct page *)page->index)
719 if (!vmalloc_sync_one(page_address(page),
721 BUG_ON(page != pgd_list);
724 spin_unlock_irqrestore(&pgd_lock, flags);
726 set_bit(pgd_index(address), insync);
728 if (address == start && test_bit(pgd_index(address), insync))
729 start = address + PGDIR_SIZE;