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(enum die_val val, const char *str,
50 struct pt_regs *regs, long err, int trap, int sig)
52 struct die_args args = {
59 return atomic_notifier_call_chain(¬ify_page_fault_chain, val, &args);
63 * Unlock any spinlocks which will prevent us from getting the
66 void bust_spinlocks(int yes)
68 int loglevel_save = console_loglevel;
79 * OK, the message is on the console. Now we call printk()
80 * without oops_in_progress set so that printk will give klogd
81 * a poke. Hold onto your hats...
83 console_loglevel = 15; /* NMI oopser may have shut the console up */
85 console_loglevel = loglevel_save;
89 * Return EIP plus the CS segment base. The segment limit is also
90 * adjusted, clamped to the kernel/user address space (whichever is
91 * appropriate), and returned in *eip_limit.
93 * The segment is checked, because it might have been changed by another
94 * task between the original faulting instruction and here.
96 * If CS is no longer a valid code segment, or if EIP is beyond the
97 * limit, or if it is a kernel address when CS is not a kernel segment,
98 * then the returned value will be greater than *eip_limit.
100 * This is slow, but is very rarely executed.
102 static inline unsigned long get_segment_eip(struct pt_regs *regs,
103 unsigned long *eip_limit)
105 unsigned long eip = regs->eip;
106 unsigned seg = regs->xcs & 0xffff;
107 u32 seg_ar, seg_limit, base, *desc;
109 /* Unlikely, but must come before segment checks. */
110 if (unlikely(regs->eflags & VM_MASK)) {
112 *eip_limit = base + 0xffff;
113 return base + (eip & 0xffff);
116 /* The standard kernel/user address space limit. */
117 *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
119 /* By far the most common cases. */
120 if (likely(SEGMENT_IS_FLAT_CODE(seg)))
123 /* Check the segment exists, is within the current LDT/GDT size,
124 that kernel/user (ring 0..3) has the appropriate privilege,
125 that it's a code segment, and get the limit. */
126 __asm__ ("larl %3,%0; lsll %3,%1"
127 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
128 if ((~seg_ar & 0x9800) || eip > seg_limit) {
130 return 1; /* So that returned eip > *eip_limit. */
133 /* Get the GDT/LDT descriptor base.
134 When you look for races in this code remember that
135 LDT and other horrors are only used in user space. */
137 /* Must lock the LDT while reading it. */
138 down(¤t->mm->context.sem);
139 desc = current->mm->context.ldt;
140 desc = (void *)desc + (seg & ~7);
142 /* Must disable preemption while reading the GDT. */
143 desc = (u32 *)get_cpu_gdt_table(get_cpu());
144 desc = (void *)desc + (seg & ~7);
147 /* Decode the code segment base from the descriptor */
148 base = get_desc_base((unsigned long *)desc);
151 up(¤t->mm->context.sem);
155 /* Adjust EIP and segment limit, and clamp at the kernel limit.
156 It's legitimate for segments to wrap at 0xffffffff. */
158 if (seg_limit < *eip_limit && seg_limit >= base)
159 *eip_limit = seg_limit;
164 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
165 * Check that here and ignore it.
167 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
170 unsigned char *instr = (unsigned char *)get_segment_eip (regs, &limit);
175 for (i = 0; scan_more && i < 15; i++) {
176 unsigned char opcode;
177 unsigned char instr_hi;
178 unsigned char instr_lo;
180 if (instr > (unsigned char *)limit)
182 if (probe_kernel_address(instr, opcode))
185 instr_hi = opcode & 0xf0;
186 instr_lo = opcode & 0x0f;
192 /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
193 scan_more = ((instr_lo & 7) == 0x6);
197 /* 0x64 thru 0x67 are valid prefixes in all modes. */
198 scan_more = (instr_lo & 0xC) == 0x4;
201 /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
202 scan_more = !instr_lo || (instr_lo>>1) == 1;
205 /* Prefetch instruction is 0x0F0D or 0x0F18 */
207 if (instr > (unsigned char *)limit)
209 if (probe_kernel_address(instr, opcode))
211 prefetch = (instr_lo == 0xF) &&
212 (opcode == 0x0D || opcode == 0x18);
222 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
223 unsigned long error_code)
225 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
226 boot_cpu_data.x86 >= 6)) {
227 /* Catch an obscure case of prefetch inside an NX page. */
228 if (nx_enabled && (error_code & 16))
230 return __is_prefetch(regs, addr);
235 static noinline void force_sig_info_fault(int si_signo, int si_code,
236 unsigned long address, struct task_struct *tsk)
240 info.si_signo = si_signo;
242 info.si_code = si_code;
243 info.si_addr = (void __user *)address;
244 force_sig_info(si_signo, &info, tsk);
247 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
249 #ifdef CONFIG_X86_PAE
250 static void dump_fault_path(unsigned long address)
252 unsigned long *p, page;
256 p = (unsigned long *)__va(page);
257 p += (address >> 30) * 2;
258 printk(KERN_ALERT "%08lx -> *pde = %08lx:%08lx\n", page, p[1], p[0]);
260 mfn = (p[0] >> PAGE_SHIFT) | ((p[1] & 0x7) << 20);
261 page = mfn_to_pfn(mfn) << PAGE_SHIFT;
262 p = (unsigned long *)__va(page);
263 address &= 0x3fffffff;
264 p += (address >> 21) * 2;
265 printk(KERN_ALERT "%08lx -> *pme = %08lx:%08lx\n",
267 #ifndef CONFIG_HIGHPTE
269 mfn = (p[0] >> PAGE_SHIFT) | ((p[1] & 0x7) << 20);
270 page = mfn_to_pfn(mfn) << PAGE_SHIFT;
271 p = (unsigned long *) __va(page);
272 address &= 0x001fffff;
273 p += (address >> 12) * 2;
274 printk(KERN_ALERT "%08lx -> *pte = %08lx:%08lx\n",
281 static void dump_fault_path(unsigned long address)
286 page = ((unsigned long *) __va(page))[address >> 22];
287 if (oops_may_print())
288 printk(KERN_ALERT "*pde = ma %08lx pa %08lx\n", page,
289 machine_to_phys(page));
291 * We must not directly access the pte in the highpte
292 * case, the page table might be allocated in highmem.
293 * And lets rather not kmap-atomic the pte, just in case
294 * it's allocated already.
296 #ifndef CONFIG_HIGHPTE
297 if ((page & 1) && oops_may_print()) {
299 address &= 0x003ff000;
300 page = machine_to_phys(page);
301 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
302 printk(KERN_ALERT "*pte = ma %08lx pa %08lx\n", page,
303 machine_to_phys(page));
309 static int spurious_fault(struct pt_regs *regs,
310 unsigned long address,
311 unsigned long error_code)
319 /* Faults in hypervisor area are never spurious. */
320 if (address >= HYPERVISOR_VIRT_START)
324 /* Reserved-bit violation or user access to kernel space? */
325 if (error_code & 0x0c)
328 pgd = init_mm.pgd + pgd_index(address);
329 if (!pgd_present(*pgd))
332 pud = pud_offset(pgd, address);
333 if (!pud_present(*pud))
336 pmd = pmd_offset(pud, address);
337 if (!pmd_present(*pmd))
340 pte = pte_offset_kernel(pmd, address);
341 if (!pte_present(*pte))
343 if ((error_code & 0x02) && !pte_write(*pte))
345 #ifdef CONFIG_X86_PAE
346 if ((error_code & 0x10) && (pte_val(*pte) & _PAGE_NX))
353 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
355 unsigned index = pgd_index(address);
361 pgd_k = init_mm.pgd + index;
363 if (!pgd_present(*pgd_k))
367 * set_pgd(pgd, *pgd_k); here would be useless on PAE
368 * and redundant with the set_pmd() on non-PAE. As would
372 pud = pud_offset(pgd, address);
373 pud_k = pud_offset(pgd_k, address);
374 if (!pud_present(*pud_k))
377 pmd = pmd_offset(pud, address);
378 pmd_k = pmd_offset(pud_k, address);
379 if (!pmd_present(*pmd_k))
381 if (!pmd_present(*pmd))
383 set_pmd(pmd, *pmd_k);
386 * When running on Xen we must launder *pmd_k through
387 * pmd_val() to ensure that _PAGE_PRESENT is correctly set.
389 set_pmd(pmd, __pmd(pmd_val(*pmd_k)));
392 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
397 * Handle a fault on the vmalloc or module mapping area
399 * This assumes no large pages in there.
401 static inline int vmalloc_fault(unsigned long address)
403 unsigned long pgd_paddr;
407 * Synchronize this task's top level page-table
408 * with the 'reference' page table.
410 * Do _not_ use "current" here. We might be inside
411 * an interrupt in the middle of a task switch..
413 pgd_paddr = read_cr3();
414 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
417 pte_k = pte_offset_kernel(pmd_k, address);
418 if (!pte_present(*pte_k))
424 * This routine handles page faults. It determines the address,
425 * and the problem, and then passes it off to one of the appropriate
429 * bit 0 == 0 means no page found, 1 means protection fault
430 * bit 1 == 0 means read, 1 means write
431 * bit 2 == 0 means kernel, 1 means user-mode
432 * bit 3 == 1 means use of reserved bit detected
433 * bit 4 == 1 means fault was an instruction fetch
435 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
436 unsigned long error_code)
438 struct task_struct *tsk;
439 struct mm_struct *mm;
440 struct vm_area_struct * vma;
441 unsigned long address;
444 /* get the address */
445 address = read_cr2();
447 /* Set the "privileged fault" bit to something sane. */
449 error_code |= (regs->xcs & 2) << 1;
450 if (regs->eflags & X86_EFLAGS_VM)
455 si_code = SEGV_MAPERR;
458 * We fault-in kernel-space virtual memory on-demand. The
459 * 'reference' page table is init_mm.pgd.
461 * NOTE! We MUST NOT take any locks for this case. We may
462 * be in an interrupt or a critical region, and should
463 * only copy the information from the master page table,
466 * This verifies that the fault happens in kernel space
467 * (error_code & 4) == 0, and that the fault was not a
468 * protection error (error_code & 9) == 0.
470 if (unlikely(address >= TASK_SIZE)) {
472 /* Faults in hypervisor area can never be patched up. */
473 if (address >= HYPERVISOR_VIRT_START)
474 goto bad_area_nosemaphore;
476 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
478 /* Can take a spurious fault if mapping changes R/O -> R/W. */
479 if (spurious_fault(regs, address, error_code))
481 if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
482 SIGSEGV) == NOTIFY_STOP)
485 * Don't take the mm semaphore here. If we fixup a prefetch
486 * fault we could otherwise deadlock.
488 goto bad_area_nosemaphore;
491 if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
492 SIGSEGV) == NOTIFY_STOP)
495 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
496 fault has been handled. */
497 if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
503 * If we're in an interrupt, have no user context or are running in an
504 * atomic region then we must not take the fault..
506 if (in_atomic() || !mm)
507 goto bad_area_nosemaphore;
509 /* When running in the kernel we expect faults to occur only to
510 * addresses in user space. All other faults represent errors in the
511 * kernel and should generate an OOPS. Unfortunatly, in the case of an
512 * erroneous fault occurring in a code path which already holds mmap_sem
513 * we will deadlock attempting to validate the fault against the
514 * address space. Luckily the kernel only validly references user
515 * space from well defined areas of code, which are listed in the
518 * As the vast majority of faults will be valid we will only perform
519 * the source reference check when there is a possibilty of a deadlock.
520 * Attempt to lock the address space, if we cannot we then validate the
521 * source. If this is invalid we can skip the address space check,
522 * thus avoiding the deadlock.
524 if (!down_read_trylock(&mm->mmap_sem)) {
525 if ((error_code & 4) == 0 &&
526 !search_exception_tables(regs->eip))
527 goto bad_area_nosemaphore;
528 down_read(&mm->mmap_sem);
531 vma = find_vma(mm, address);
534 if (vma->vm_start <= address)
536 if (!(vma->vm_flags & VM_GROWSDOWN))
538 if (error_code & 4) {
540 * Accessing the stack below %esp is always a bug.
541 * The large cushion allows instructions like enter
542 * and pusha to work. ("enter $65535,$31" pushes
543 * 32 pointers and then decrements %esp by 65535.)
545 if (address + 65536 + 32 * sizeof(unsigned long) < regs->esp)
548 if (expand_stack(vma, address))
551 * Ok, we have a good vm_area for this memory access, so
555 si_code = SEGV_ACCERR;
557 switch (error_code & 3) {
558 default: /* 3: write, present */
560 case 2: /* write, not present */
561 if (!(vma->vm_flags & VM_WRITE))
565 case 1: /* read, present */
567 case 0: /* read, not present */
568 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
574 * If for any reason at all we couldn't handle the fault,
575 * make sure we exit gracefully rather than endlessly redo
578 switch (handle_mm_fault(mm, vma, address, write)) {
585 case VM_FAULT_SIGBUS:
594 * Did it hit the DOS screen memory VA from vm86 mode?
596 if (regs->eflags & VM_MASK) {
597 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
599 tsk->thread.screen_bitmap |= 1 << bit;
601 up_read(&mm->mmap_sem);
605 * Something tried to access memory that isn't in our memory map..
606 * Fix it, but check if it's kernel or user first..
609 up_read(&mm->mmap_sem);
611 bad_area_nosemaphore:
612 /* User mode accesses just cause a SIGSEGV */
613 if (error_code & 4) {
615 * Valid to do another page fault here because this one came
618 if (is_prefetch(regs, address, error_code))
621 tsk->thread.cr2 = address;
622 /* Kernel addresses are always protection faults */
623 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
624 tsk->thread.trap_no = 14;
625 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
629 #ifdef CONFIG_X86_F00F_BUG
631 * Pentium F0 0F C7 C8 bug workaround.
633 if (boot_cpu_data.f00f_bug) {
636 nr = (address - idt_descr.address) >> 3;
639 do_invalid_op(regs, 0);
646 /* Are we prepared to handle this kernel fault? */
647 if (fixup_exception(regs))
651 * Valid to do another page fault here, because if this fault
652 * had been triggered by is_prefetch fixup_exception would have
655 if (is_prefetch(regs, address, error_code))
659 * Oops. The kernel tried to access some bad page. We'll have to
660 * terminate things with extreme prejudice.
665 if (oops_may_print()) {
666 #ifdef CONFIG_X86_PAE
667 if (error_code & 16) {
668 pte_t *pte = lookup_address(address);
670 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
671 printk(KERN_CRIT "kernel tried to execute "
672 "NX-protected page - exploit attempt? "
673 "(uid: %d)\n", current->uid);
676 if (address < PAGE_SIZE)
677 printk(KERN_ALERT "BUG: unable to handle kernel NULL "
678 "pointer dereference");
680 printk(KERN_ALERT "BUG: unable to handle kernel paging"
682 printk(" at virtual address %08lx\n",address);
683 printk(KERN_ALERT " printing eip:\n");
684 printk("%08lx\n", regs->eip);
685 dump_fault_path(address);
687 tsk->thread.cr2 = address;
688 tsk->thread.trap_no = 14;
689 tsk->thread.error_code = error_code;
690 die("Oops", regs, error_code);
695 * We ran out of memory, or some other thing happened to us that made
696 * us unable to handle the page fault gracefully.
699 up_read(&mm->mmap_sem);
702 down_read(&mm->mmap_sem);
705 printk("VM: killing process %s(%d:#%u)\n",
706 tsk->comm, tsk->pid, tsk->xid);
712 up_read(&mm->mmap_sem);
714 /* Kernel mode? Handle exceptions or die */
715 if (!(error_code & 4))
718 /* User space => ok to do another page fault */
719 if (is_prefetch(regs, address, error_code))
722 tsk->thread.cr2 = address;
723 tsk->thread.error_code = error_code;
724 tsk->thread.trap_no = 14;
725 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
728 #ifndef CONFIG_X86_PAE
729 void vmalloc_sync_all(void)
732 * Note that races in the updates of insync and start aren't
733 * problematic: insync can only get set bits added, and updates to
734 * start are only improving performance (without affecting correctness
737 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
738 static unsigned long start = TASK_SIZE;
739 unsigned long address;
741 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
742 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
743 if (!test_bit(pgd_index(address), insync)) {
747 spin_lock_irqsave(&pgd_lock, flags);
748 for (page = pgd_list; page; page =
749 (struct page *)page->index)
750 if (!vmalloc_sync_one(page_address(page),
752 BUG_ON(page != pgd_list);
755 spin_unlock_irqrestore(&pgd_lock, flags);
757 set_bit(pgd_index(address), insync);
759 if (address == start && test_bit(pgd_index(address), insync))
760 start = address + PGDIR_SIZE;