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>
26 #include <asm/system.h>
27 #include <asm/uaccess.h>
29 #include <asm/kdebug.h>
31 extern void die(const char *,struct pt_regs *,long);
34 * Unlock any spinlocks which will prevent us from getting the
37 void bust_spinlocks(int yes)
39 int loglevel_save = console_loglevel;
50 * OK, the message is on the console. Now we call printk()
51 * without oops_in_progress set so that printk will give klogd
52 * a poke. Hold onto your hats...
54 console_loglevel = 15; /* NMI oopser may have shut the console up */
56 console_loglevel = loglevel_save;
60 * Return EIP plus the CS segment base. The segment limit is also
61 * adjusted, clamped to the kernel/user address space (whichever is
62 * appropriate), and returned in *eip_limit.
64 * The segment is checked, because it might have been changed by another
65 * task between the original faulting instruction and here.
67 * If CS is no longer a valid code segment, or if EIP is beyond the
68 * limit, or if it is a kernel address when CS is not a kernel segment,
69 * then the returned value will be greater than *eip_limit.
71 * This is slow, but is very rarely executed.
73 static inline unsigned long get_segment_eip(struct pt_regs *regs,
74 unsigned long *eip_limit)
76 unsigned long eip = regs->eip;
77 unsigned seg = regs->xcs & 0xffff;
78 u32 seg_ar, seg_limit, base, *desc;
80 /* The standard kernel/user address space limit. */
81 *eip_limit = (seg & 2) ? USER_DS.seg : KERNEL_DS.seg;
83 /* Unlikely, but must come before segment checks. */
84 if (unlikely((regs->eflags & VM_MASK) != 0))
85 return eip + (seg << 4);
87 /* By far the most common cases. */
88 if (likely(seg == __USER_CS || seg == GET_KERNEL_CS()))
91 /* Check the segment exists, is within the current LDT/GDT size,
92 that kernel/user (ring 0..3) has the appropriate privilege,
93 that it's a code segment, and get the limit. */
94 __asm__ ("larl %3,%0; lsll %3,%1"
95 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
96 if ((~seg_ar & 0x9800) || eip > seg_limit) {
98 return 1; /* So that returned eip > *eip_limit. */
101 /* Get the GDT/LDT descriptor base.
102 When you look for races in this code remember that
103 LDT and other horrors are only used in user space. */
105 /* Must lock the LDT while reading it. */
106 down(¤t->mm->context.sem);
107 desc = current->mm->context.ldt;
108 desc = (void *)desc + (seg & ~7);
110 /* Must disable preemption while reading the GDT. */
111 desc = (u32 *)get_cpu_gdt_table(get_cpu());
112 desc = (void *)desc + (seg & ~7);
115 /* Decode the code segment base from the descriptor */
116 base = get_desc_base((unsigned long *)desc);
119 up(¤t->mm->context.sem);
123 /* Adjust EIP and segment limit, and clamp at the kernel limit.
124 It's legitimate for segments to wrap at 0xffffffff. */
126 if (seg_limit < *eip_limit && seg_limit >= base)
127 *eip_limit = seg_limit;
132 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
133 * Check that here and ignore it.
135 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
138 unsigned long instr = get_segment_eip (regs, &limit);
143 for (i = 0; scan_more && i < 15; i++) {
144 unsigned char opcode;
145 unsigned char instr_hi;
146 unsigned char instr_lo;
150 if (__get_user(opcode, (unsigned char __user *) instr))
153 instr_hi = opcode & 0xf0;
154 instr_lo = opcode & 0x0f;
160 /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
161 scan_more = ((instr_lo & 7) == 0x6);
165 /* 0x64 thru 0x67 are valid prefixes in all modes. */
166 scan_more = (instr_lo & 0xC) == 0x4;
169 /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
170 scan_more = !instr_lo || (instr_lo>>1) == 1;
173 /* Prefetch instruction is 0x0F0D or 0x0F18 */
177 if (__get_user(opcode, (unsigned char __user *) instr))
179 prefetch = (instr_lo == 0xF) &&
180 (opcode == 0x0D || opcode == 0x18);
190 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
191 unsigned long error_code)
193 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
194 boot_cpu_data.x86 >= 6)) {
195 /* Catch an obscure case of prefetch inside an NX page. */
196 if (nx_enabled && (error_code & 16))
198 return __is_prefetch(regs, addr);
203 static noinline void force_sig_info_fault(int si_signo, int si_code,
204 unsigned long address, struct task_struct *tsk)
208 info.si_signo = si_signo;
210 info.si_code = si_code;
211 info.si_addr = (void __user *)address;
212 force_sig_info(si_signo, &info, tsk);
215 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
217 #ifdef CONFIG_X86_PAE
218 static void dump_fault_path(unsigned long address)
220 unsigned long *p, page;
224 p = (unsigned long *)__va(page);
225 p += (address >> 30) * 2;
226 printk(KERN_ALERT "%08lx -> *pde = %08lx:%08lx\n", page, p[1], p[0]);
228 mfn = (p[0] >> PAGE_SHIFT) | ((p[1] & 0x7) << 20);
229 page = mfn_to_pfn(mfn) << PAGE_SHIFT;
230 p = (unsigned long *)__va(page);
231 address &= 0x3fffffff;
232 p += (address >> 21) * 2;
233 printk(KERN_ALERT "%08lx -> *pme = %08lx:%08lx\n",
235 #ifndef CONFIG_HIGHPTE
237 mfn = (p[0] >> PAGE_SHIFT) | ((p[1] & 0x7) << 20);
238 page = mfn_to_pfn(mfn) << PAGE_SHIFT;
239 p = (unsigned long *) __va(page);
240 address &= 0x001fffff;
241 p += (address >> 12) * 2;
242 printk(KERN_ALERT "%08lx -> *pte = %08lx:%08lx\n",
249 static void dump_fault_path(unsigned long address)
254 page = ((unsigned long *) __va(page))[address >> 22];
255 if (oops_may_print())
256 printk(KERN_ALERT "*pde = ma %08lx pa %08lx\n", page,
257 machine_to_phys(page));
259 * We must not directly access the pte in the highpte
260 * case, the page table might be allocated in highmem.
261 * And lets rather not kmap-atomic the pte, just in case
262 * it's allocated already.
264 #ifndef CONFIG_HIGHPTE
265 if ((page & 1) && oops_may_print()) {
267 address &= 0x003ff000;
268 page = machine_to_phys(page);
269 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
270 printk(KERN_ALERT "*pte = ma %08lx pa %08lx\n", page,
271 machine_to_phys(page));
277 static int spurious_fault(struct pt_regs *regs,
278 unsigned long address,
279 unsigned long error_code)
287 /* Faults in hypervisor area are never spurious. */
288 if (address >= HYPERVISOR_VIRT_START)
292 /* Reserved-bit violation or user access to kernel space? */
293 if (error_code & 0x0c)
296 pgd = init_mm.pgd + pgd_index(address);
297 if (!pgd_present(*pgd))
300 pud = pud_offset(pgd, address);
301 if (!pud_present(*pud))
304 pmd = pmd_offset(pud, address);
305 if (!pmd_present(*pmd))
308 pte = pte_offset_kernel(pmd, address);
309 if (!pte_present(*pte))
311 if ((error_code & 0x02) && !pte_write(*pte))
313 #ifdef CONFIG_X86_PAE
314 if ((error_code & 0x10) && (pte_val(*pte) & _PAGE_NX))
321 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
323 unsigned index = pgd_index(address);
329 pgd_k = init_mm.pgd + index;
331 if (!pgd_present(*pgd_k))
335 * set_pgd(pgd, *pgd_k); here would be useless on PAE
336 * and redundant with the set_pmd() on non-PAE. As would
340 pud = pud_offset(pgd, address);
341 pud_k = pud_offset(pgd_k, address);
342 if (!pud_present(*pud_k))
345 pmd = pmd_offset(pud, address);
346 pmd_k = pmd_offset(pud_k, address);
347 if (!pmd_present(*pmd_k))
349 if (!pmd_present(*pmd))
351 set_pmd(pmd, *pmd_k);
354 * When running on Xen we must launder *pmd_k through
355 * pmd_val() to ensure that _PAGE_PRESENT is correctly set.
357 set_pmd(pmd, __pmd(pmd_val(*pmd_k)));
360 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
365 * Handle a fault on the vmalloc or module mapping area
367 * This assumes no large pages in there.
369 static inline int vmalloc_fault(unsigned long address)
371 unsigned long pgd_paddr;
375 * Synchronize this task's top level page-table
376 * with the 'reference' page table.
378 * Do _not_ use "current" here. We might be inside
379 * an interrupt in the middle of a task switch..
381 pgd_paddr = read_cr3();
382 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
385 pte_k = pte_offset_kernel(pmd_k, address);
386 if (!pte_present(*pte_k))
392 * This routine handles page faults. It determines the address,
393 * and the problem, and then passes it off to one of the appropriate
397 * bit 0 == 0 means no page found, 1 means protection fault
398 * bit 1 == 0 means read, 1 means write
399 * bit 2 == 0 means kernel, 1 means user-mode
400 * bit 3 == 1 means use of reserved bit detected
401 * bit 4 == 1 means fault was an instruction fetch
403 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
404 unsigned long error_code)
406 struct task_struct *tsk;
407 struct mm_struct *mm;
408 struct vm_area_struct * vma;
409 unsigned long address;
412 /* get the address */
413 address = read_cr2();
415 /* Set the "privileged fault" bit to something sane. */
417 error_code |= (regs->xcs & 2) << 1;
418 if (regs->eflags & X86_EFLAGS_VM)
423 si_code = SEGV_MAPERR;
426 * We fault-in kernel-space virtual memory on-demand. The
427 * 'reference' page table is init_mm.pgd.
429 * NOTE! We MUST NOT take any locks for this case. We may
430 * be in an interrupt or a critical region, and should
431 * only copy the information from the master page table,
434 * This verifies that the fault happens in kernel space
435 * (error_code & 4) == 0, and that the fault was not a
436 * protection error (error_code & 9) == 0.
438 if (unlikely(address >= TASK_SIZE)) {
440 /* Faults in hypervisor area can never be patched up. */
441 if (address >= HYPERVISOR_VIRT_START)
442 goto bad_area_nosemaphore;
444 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
446 /* Can take a spurious fault if mapping changes R/O -> R/W. */
447 if (spurious_fault(regs, address, error_code))
449 if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
450 SIGSEGV) == NOTIFY_STOP)
453 * Don't take the mm semaphore here. If we fixup a prefetch
454 * fault we could otherwise deadlock.
456 goto bad_area_nosemaphore;
459 if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
460 SIGSEGV) == NOTIFY_STOP)
463 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
464 fault has been handled. */
465 if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
471 * If we're in an interrupt, have no user context or are running in an
472 * atomic region then we must not take the fault..
474 if (in_atomic() || !mm)
475 goto bad_area_nosemaphore;
477 /* When running in the kernel we expect faults to occur only to
478 * addresses in user space. All other faults represent errors in the
479 * kernel and should generate an OOPS. Unfortunatly, in the case of an
480 * erroneous fault occuring in a code path which already holds mmap_sem
481 * we will deadlock attempting to validate the fault against the
482 * address space. Luckily the kernel only validly references user
483 * space from well defined areas of code, which are listed in the
486 * As the vast majority of faults will be valid we will only perform
487 * the source reference check when there is a possibilty of a deadlock.
488 * Attempt to lock the address space, if we cannot we then validate the
489 * source. If this is invalid we can skip the address space check,
490 * thus avoiding the deadlock.
492 if (!down_read_trylock(&mm->mmap_sem)) {
493 if ((error_code & 4) == 0 &&
494 !search_exception_tables(regs->eip))
495 goto bad_area_nosemaphore;
496 down_read(&mm->mmap_sem);
499 vma = find_vma(mm, address);
502 if (vma->vm_start <= address)
504 if (!(vma->vm_flags & VM_GROWSDOWN))
506 if (error_code & 4) {
508 * accessing the stack below %esp is always a bug.
509 * The "+ 32" is there due to some instructions (like
510 * pusha) doing post-decrement on the stack and that
511 * doesn't show up until later..
513 if (address + 32 < regs->esp)
516 if (expand_stack(vma, address))
519 * Ok, we have a good vm_area for this memory access, so
523 si_code = SEGV_ACCERR;
525 switch (error_code & 3) {
526 default: /* 3: write, present */
527 #ifdef TEST_VERIFY_AREA
528 if (regs->cs == GET_KERNEL_CS())
529 printk("WP fault at %08lx\n", regs->eip);
532 case 2: /* write, not present */
533 if (!(vma->vm_flags & VM_WRITE))
537 case 1: /* read, present */
539 case 0: /* read, not present */
540 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
546 * If for any reason at all we couldn't handle the fault,
547 * make sure we exit gracefully rather than endlessly redo
550 switch (handle_mm_fault(mm, vma, address, write)) {
557 case VM_FAULT_SIGBUS:
566 * Did it hit the DOS screen memory VA from vm86 mode?
568 if (regs->eflags & VM_MASK) {
569 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
571 tsk->thread.screen_bitmap |= 1 << bit;
573 up_read(&mm->mmap_sem);
577 * Something tried to access memory that isn't in our memory map..
578 * Fix it, but check if it's kernel or user first..
581 up_read(&mm->mmap_sem);
583 bad_area_nosemaphore:
584 /* User mode accesses just cause a SIGSEGV */
585 if (error_code & 4) {
587 * Valid to do another page fault here because this one came
590 if (is_prefetch(regs, address, error_code))
593 tsk->thread.cr2 = address;
594 /* Kernel addresses are always protection faults */
595 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
596 tsk->thread.trap_no = 14;
597 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
601 #ifdef CONFIG_X86_F00F_BUG
603 * Pentium F0 0F C7 C8 bug workaround.
605 if (boot_cpu_data.f00f_bug) {
608 nr = (address - idt_descr.address) >> 3;
611 do_invalid_op(regs, 0);
618 /* Are we prepared to handle this kernel fault? */
619 if (fixup_exception(regs))
623 * Valid to do another page fault here, because if this fault
624 * had been triggered by is_prefetch fixup_exception would have
627 if (is_prefetch(regs, address, error_code))
631 * Oops. The kernel tried to access some bad page. We'll have to
632 * terminate things with extreme prejudice.
637 if (oops_may_print()) {
638 #ifdef CONFIG_X86_PAE
639 if (error_code & 16) {
640 pte_t *pte = lookup_address(address);
642 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
643 printk(KERN_CRIT "kernel tried to execute "
644 "NX-protected page - exploit attempt? "
645 "(uid: %d)\n", current->uid);
648 if (address < PAGE_SIZE)
649 printk(KERN_ALERT "BUG: unable to handle kernel NULL "
650 "pointer dereference");
652 printk(KERN_ALERT "BUG: unable to handle kernel paging"
654 printk(" at virtual address %08lx\n",address);
655 printk(KERN_ALERT " printing eip:\n");
656 printk("%08lx\n", regs->eip);
657 dump_fault_path(address);
659 tsk->thread.cr2 = address;
660 tsk->thread.trap_no = 14;
661 tsk->thread.error_code = error_code;
662 die("Oops", regs, error_code);
667 * We ran out of memory, or some other thing happened to us that made
668 * us unable to handle the page fault gracefully.
671 up_read(&mm->mmap_sem);
674 down_read(&mm->mmap_sem);
677 printk("VM: killing process %s\n", tsk->comm);
683 up_read(&mm->mmap_sem);
685 /* Kernel mode? Handle exceptions or die */
686 if (!(error_code & 4))
689 /* User space => ok to do another page fault */
690 if (is_prefetch(regs, address, error_code))
693 tsk->thread.cr2 = address;
694 tsk->thread.error_code = error_code;
695 tsk->thread.trap_no = 14;
696 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
699 #ifndef CONFIG_X86_PAE
700 void vmalloc_sync_all(void)
703 * Note that races in the updates of insync and start aren't
704 * problematic: insync can only get set bits added, and updates to
705 * start are only improving performance (without affecting correctness
708 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
709 static unsigned long start = TASK_SIZE;
710 unsigned long address;
712 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
713 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
714 if (!test_bit(pgd_index(address), insync)) {
718 spin_lock_irqsave(&pgd_lock, flags);
719 for (page = pgd_list; page; page =
720 (struct page *)page->index)
721 if (!vmalloc_sync_one(page_address(page),
723 BUG_ON(page != pgd_list);
726 spin_unlock_irqrestore(&pgd_lock, flags);
728 set_bit(pgd_index(address), insync);
730 if (address == start && test_bit(pgd_index(address), insync))
731 start = address + PGDIR_SIZE;