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);
33 static ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain);
35 int register_page_fault_notifier(struct notifier_block *nb)
38 return atomic_notifier_chain_register(¬ify_page_fault_chain, nb);
40 EXPORT_SYMBOL_GPL(register_page_fault_notifier);
42 int unregister_page_fault_notifier(struct notifier_block *nb)
44 return atomic_notifier_chain_unregister(¬ify_page_fault_chain, nb);
46 EXPORT_SYMBOL_GPL(unregister_page_fault_notifier);
48 static inline int notify_page_fault(enum die_val val, const char *str,
49 struct pt_regs *regs, long err, int trap, int sig)
51 struct die_args args = {
58 return atomic_notifier_call_chain(¬ify_page_fault_chain, val, &args);
62 * Unlock any spinlocks which will prevent us from getting the
65 void bust_spinlocks(int yes)
67 int loglevel_save = console_loglevel;
78 * OK, the message is on the console. Now we call printk()
79 * without oops_in_progress set so that printk will give klogd
80 * a poke. Hold onto your hats...
82 console_loglevel = 15; /* NMI oopser may have shut the console up */
84 console_loglevel = loglevel_save;
88 * Return EIP plus the CS segment base. The segment limit is also
89 * adjusted, clamped to the kernel/user address space (whichever is
90 * appropriate), and returned in *eip_limit.
92 * The segment is checked, because it might have been changed by another
93 * task between the original faulting instruction and here.
95 * If CS is no longer a valid code segment, or if EIP is beyond the
96 * limit, or if it is a kernel address when CS is not a kernel segment,
97 * then the returned value will be greater than *eip_limit.
99 * This is slow, but is very rarely executed.
101 static inline unsigned long get_segment_eip(struct pt_regs *regs,
102 unsigned long *eip_limit)
104 unsigned long eip = regs->eip;
105 unsigned seg = regs->xcs & 0xffff;
106 u32 seg_ar, seg_limit, base, *desc;
108 /* Unlikely, but must come before segment checks. */
109 if (unlikely(regs->eflags & VM_MASK)) {
111 *eip_limit = base + 0xffff;
112 return base + (eip & 0xffff);
115 /* The standard kernel/user address space limit. */
116 *eip_limit = (seg & 3) ? USER_DS.seg : KERNEL_DS.seg;
118 /* By far the most common cases. */
119 if (likely(seg == __USER_CS || seg == __KERNEL_CS))
122 /* Check the segment exists, is within the current LDT/GDT size,
123 that kernel/user (ring 0..3) has the appropriate privilege,
124 that it's a code segment, and get the limit. */
125 __asm__ ("larl %3,%0; lsll %3,%1"
126 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
127 if ((~seg_ar & 0x9800) || eip > seg_limit) {
129 return 1; /* So that returned eip > *eip_limit. */
132 /* Get the GDT/LDT descriptor base.
133 When you look for races in this code remember that
134 LDT and other horrors are only used in user space. */
136 /* Must lock the LDT while reading it. */
137 down(¤t->mm->context.sem);
138 desc = current->mm->context.ldt;
139 desc = (void *)desc + (seg & ~7);
141 /* Must disable preemption while reading the GDT. */
142 desc = (u32 *)get_cpu_gdt_table(get_cpu());
143 desc = (void *)desc + (seg & ~7);
146 /* Decode the code segment base from the descriptor */
147 base = get_desc_base((unsigned long *)desc);
150 up(¤t->mm->context.sem);
154 /* Adjust EIP and segment limit, and clamp at the kernel limit.
155 It's legitimate for segments to wrap at 0xffffffff. */
157 if (seg_limit < *eip_limit && seg_limit >= base)
158 *eip_limit = seg_limit;
163 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
164 * Check that here and ignore it.
166 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
169 unsigned long instr = get_segment_eip (regs, &limit);
174 for (i = 0; scan_more && i < 15; i++) {
175 unsigned char opcode;
176 unsigned char instr_hi;
177 unsigned char instr_lo;
181 if (__get_user(opcode, (unsigned char __user *) instr))
184 instr_hi = opcode & 0xf0;
185 instr_lo = opcode & 0x0f;
191 /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
192 scan_more = ((instr_lo & 7) == 0x6);
196 /* 0x64 thru 0x67 are valid prefixes in all modes. */
197 scan_more = (instr_lo & 0xC) == 0x4;
200 /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
201 scan_more = !instr_lo || (instr_lo>>1) == 1;
204 /* Prefetch instruction is 0x0F0D or 0x0F18 */
208 if (__get_user(opcode, (unsigned char __user *) instr))
210 prefetch = (instr_lo == 0xF) &&
211 (opcode == 0x0D || opcode == 0x18);
221 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
222 unsigned long error_code)
224 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
225 boot_cpu_data.x86 >= 6)) {
226 /* Catch an obscure case of prefetch inside an NX page. */
227 if (nx_enabled && (error_code & 16))
229 return __is_prefetch(regs, addr);
234 static noinline void force_sig_info_fault(int si_signo, int si_code,
235 unsigned long address, struct task_struct *tsk)
239 info.si_signo = si_signo;
241 info.si_code = si_code;
242 info.si_addr = (void __user *)address;
243 force_sig_info(si_signo, &info, tsk);
246 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
248 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
250 unsigned index = pgd_index(address);
256 pgd_k = init_mm.pgd + index;
258 if (!pgd_present(*pgd_k))
262 * set_pgd(pgd, *pgd_k); here would be useless on PAE
263 * and redundant with the set_pmd() on non-PAE. As would
267 pud = pud_offset(pgd, address);
268 pud_k = pud_offset(pgd_k, address);
269 if (!pud_present(*pud_k))
272 pmd = pmd_offset(pud, address);
273 pmd_k = pmd_offset(pud_k, address);
274 if (!pmd_present(*pmd_k))
276 if (!pmd_present(*pmd))
277 set_pmd(pmd, *pmd_k);
279 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
284 * Handle a fault on the vmalloc or module mapping area
286 * This assumes no large pages in there.
288 static inline int vmalloc_fault(unsigned long address)
290 unsigned long pgd_paddr;
294 * Synchronize this task's top level page-table
295 * with the 'reference' page table.
297 * Do _not_ use "current" here. We might be inside
298 * an interrupt in the middle of a task switch..
300 pgd_paddr = read_cr3();
301 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
304 pte_k = pte_offset_kernel(pmd_k, address);
305 if (!pte_present(*pte_k))
311 * This routine handles page faults. It determines the address,
312 * and the problem, and then passes it off to one of the appropriate
316 * bit 0 == 0 means no page found, 1 means protection fault
317 * bit 1 == 0 means read, 1 means write
318 * bit 2 == 0 means kernel, 1 means user-mode
319 * bit 3 == 1 means use of reserved bit detected
320 * bit 4 == 1 means fault was an instruction fetch
322 fastcall void __kprobes do_page_fault(struct pt_regs *regs,
323 unsigned long error_code)
325 struct task_struct *tsk;
326 struct mm_struct *mm;
327 struct vm_area_struct * vma;
328 unsigned long address;
332 /* get the address */
333 address = read_cr2();
337 si_code = SEGV_MAPERR;
340 * We fault-in kernel-space virtual memory on-demand. The
341 * 'reference' page table is init_mm.pgd.
343 * NOTE! We MUST NOT take any locks for this case. We may
344 * be in an interrupt or a critical region, and should
345 * only copy the information from the master page table,
348 * This verifies that the fault happens in kernel space
349 * (error_code & 4) == 0, and that the fault was not a
350 * protection error (error_code & 9) == 0.
352 if (unlikely(address >= TASK_SIZE)) {
353 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
355 if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
356 SIGSEGV) == NOTIFY_STOP)
359 * Don't take the mm semaphore here. If we fixup a prefetch
360 * fault we could otherwise deadlock.
362 goto bad_area_nosemaphore;
365 if (notify_page_fault(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
366 SIGSEGV) == NOTIFY_STOP)
369 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
370 fault has been handled. */
371 if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
377 * If we're in an interrupt, have no user context or are running in an
378 * atomic region then we must not take the fault..
380 if (in_atomic() || !mm)
381 goto bad_area_nosemaphore;
383 /* When running in the kernel we expect faults to occur only to
384 * addresses in user space. All other faults represent errors in the
385 * kernel and should generate an OOPS. Unfortunatly, in the case of an
386 * erroneous fault occurring in a code path which already holds mmap_sem
387 * we will deadlock attempting to validate the fault against the
388 * address space. Luckily the kernel only validly references user
389 * space from well defined areas of code, which are listed in the
392 * As the vast majority of faults will be valid we will only perform
393 * the source reference check when there is a possibilty of a deadlock.
394 * Attempt to lock the address space, if we cannot we then validate the
395 * source. If this is invalid we can skip the address space check,
396 * thus avoiding the deadlock.
398 if (!down_read_trylock(&mm->mmap_sem)) {
399 if ((error_code & 4) == 0 &&
400 !search_exception_tables(regs->eip))
401 goto bad_area_nosemaphore;
402 down_read(&mm->mmap_sem);
405 vma = find_vma(mm, address);
408 if (vma->vm_start <= address)
410 if (!(vma->vm_flags & VM_GROWSDOWN))
412 if (error_code & 4) {
414 * Accessing the stack below %esp is always a bug.
415 * The large cushion allows instructions like enter
416 * and pusha to work. ("enter $65535,$31" pushes
417 * 32 pointers and then decrements %esp by 65535.)
419 if (address + 65536 + 32 * sizeof(unsigned long) < regs->esp)
422 if (expand_stack(vma, address))
425 * Ok, we have a good vm_area for this memory access, so
429 si_code = SEGV_ACCERR;
431 switch (error_code & 3) {
432 default: /* 3: write, present */
433 #ifdef TEST_VERIFY_AREA
434 if (regs->cs == KERNEL_CS)
435 printk("WP fault at %08lx\n", regs->eip);
438 case 2: /* write, not present */
439 if (!(vma->vm_flags & VM_WRITE))
443 case 1: /* read, present */
445 case 0: /* read, not present */
446 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
452 * If for any reason at all we couldn't handle the fault,
453 * make sure we exit gracefully rather than endlessly redo
456 switch (handle_mm_fault(mm, vma, address, write)) {
463 case VM_FAULT_SIGBUS:
472 * Did it hit the DOS screen memory VA from vm86 mode?
474 if (regs->eflags & VM_MASK) {
475 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
477 tsk->thread.screen_bitmap |= 1 << bit;
479 up_read(&mm->mmap_sem);
483 * Something tried to access memory that isn't in our memory map..
484 * Fix it, but check if it's kernel or user first..
487 up_read(&mm->mmap_sem);
489 bad_area_nosemaphore:
490 /* User mode accesses just cause a SIGSEGV */
491 if (error_code & 4) {
493 * Valid to do another page fault here because this one came
496 if (is_prefetch(regs, address, error_code))
499 tsk->thread.cr2 = address;
500 /* Kernel addresses are always protection faults */
501 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
502 tsk->thread.trap_no = 14;
503 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
507 #ifdef CONFIG_X86_F00F_BUG
509 * Pentium F0 0F C7 C8 bug workaround.
511 if (boot_cpu_data.f00f_bug) {
514 nr = (address - idt_descr.address) >> 3;
517 do_invalid_op(regs, 0);
524 /* Are we prepared to handle this kernel fault? */
525 if (fixup_exception(regs))
529 * Valid to do another page fault here, because if this fault
530 * had been triggered by is_prefetch fixup_exception would have
533 if (is_prefetch(regs, address, error_code))
537 * Oops. The kernel tried to access some bad page. We'll have to
538 * terminate things with extreme prejudice.
543 if (oops_may_print()) {
544 #ifdef CONFIG_X86_PAE
545 if (error_code & 16) {
546 pte_t *pte = lookup_address(address);
548 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
549 printk(KERN_CRIT "kernel tried to execute "
550 "NX-protected page - exploit attempt? "
551 "(uid: %d)\n", current->uid);
554 if (address < PAGE_SIZE)
555 printk(KERN_ALERT "BUG: unable to handle kernel NULL "
556 "pointer dereference");
558 printk(KERN_ALERT "BUG: unable to handle kernel paging"
560 printk(" at virtual address %08lx\n",address);
561 printk(KERN_ALERT " printing eip:\n");
562 printk("%08lx\n", regs->eip);
565 page = ((unsigned long *) __va(page))[address >> 22];
566 if (oops_may_print())
567 printk(KERN_ALERT "*pde = %08lx\n", page);
569 * We must not directly access the pte in the highpte
570 * case, the page table might be allocated in highmem.
571 * And lets rather not kmap-atomic the pte, just in case
572 * it's allocated already.
574 #ifndef CONFIG_HIGHPTE
575 if ((page & 1) && oops_may_print()) {
577 address &= 0x003ff000;
578 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
579 printk(KERN_ALERT "*pte = %08lx\n", page);
582 tsk->thread.cr2 = address;
583 tsk->thread.trap_no = 14;
584 tsk->thread.error_code = error_code;
585 die("Oops", regs, error_code);
590 * We ran out of memory, or some other thing happened to us that made
591 * us unable to handle the page fault gracefully.
594 up_read(&mm->mmap_sem);
597 down_read(&mm->mmap_sem);
600 printk("VM: killing process %s\n", tsk->comm);
606 up_read(&mm->mmap_sem);
608 /* Kernel mode? Handle exceptions or die */
609 if (!(error_code & 4))
612 /* User space => ok to do another page fault */
613 if (is_prefetch(regs, address, error_code))
616 tsk->thread.cr2 = address;
617 tsk->thread.error_code = error_code;
618 tsk->thread.trap_no = 14;
619 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
622 #ifndef CONFIG_X86_PAE
623 void vmalloc_sync_all(void)
626 * Note that races in the updates of insync and start aren't
627 * problematic: insync can only get set bits added, and updates to
628 * start are only improving performance (without affecting correctness
631 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
632 static unsigned long start = TASK_SIZE;
633 unsigned long address;
635 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
636 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
637 if (!test_bit(pgd_index(address), insync)) {
641 spin_lock_irqsave(&pgd_lock, flags);
642 for (page = pgd_list; page; page =
643 (struct page *)page->index)
644 if (!vmalloc_sync_one(page_address(page),
646 BUG_ON(page != pgd_list);
649 spin_unlock_irqrestore(&pgd_lock, flags);
651 set_bit(pgd_index(address), insync);
653 if (address == start && test_bit(pgd_index(address), insync))
654 start = address + PGDIR_SIZE;