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/interrupt.h>
18 #include <linux/init.h>
19 #include <linux/tty.h>
20 #include <linux/vt_kern.h> /* For unblank_screen() */
21 #include <linux/highmem.h>
22 #include <linux/bootmem.h> /* for max_low_pfn */
23 #include <linux/vmalloc.h>
24 #include <linux/module.h>
25 #include <linux/kprobes.h>
26 #include <linux/uaccess.h>
27 #include <linux/kdebug.h>
29 #include <asm/system.h>
31 #include <asm/segment.h>
33 extern void die(const char *,struct pt_regs
*,long);
35 static ATOMIC_NOTIFIER_HEAD(notify_page_fault_chain
);
37 int register_page_fault_notifier(struct notifier_block
*nb
)
40 return atomic_notifier_chain_register(¬ify_page_fault_chain
, nb
);
42 EXPORT_SYMBOL_GPL(register_page_fault_notifier
);
44 int unregister_page_fault_notifier(struct notifier_block
*nb
)
46 return atomic_notifier_chain_unregister(¬ify_page_fault_chain
, nb
);
48 EXPORT_SYMBOL_GPL(unregister_page_fault_notifier
);
50 static inline int notify_page_fault(struct pt_regs
*regs
, long err
)
52 struct die_args args
= {
59 return atomic_notifier_call_chain(¬ify_page_fault_chain
,
60 DIE_PAGE_FAULT
, &args
);
64 * Return EIP plus the CS segment base. The segment limit is also
65 * adjusted, clamped to the kernel/user address space (whichever is
66 * appropriate), and returned in *eip_limit.
68 * The segment is checked, because it might have been changed by another
69 * task between the original faulting instruction and here.
71 * If CS is no longer a valid code segment, or if EIP is beyond the
72 * limit, or if it is a kernel address when CS is not a kernel segment,
73 * then the returned value will be greater than *eip_limit.
75 * This is slow, but is very rarely executed.
77 static inline unsigned long get_segment_eip(struct pt_regs
*regs
,
78 unsigned long *eip_limit
)
80 unsigned long eip
= regs
->eip
;
81 unsigned seg
= regs
->xcs
& 0xffff;
82 u32 seg_ar
, seg_limit
, base
, *desc
;
84 /* Unlikely, but must come before segment checks. */
85 if (unlikely(regs
->eflags
& VM_MASK
)) {
87 *eip_limit
= base
+ 0xffff;
88 return base
+ (eip
& 0xffff);
91 /* The standard kernel/user address space limit. */
92 *eip_limit
= user_mode(regs
) ? USER_DS
.seg
: KERNEL_DS
.seg
;
94 /* By far the most common cases. */
95 if (likely(SEGMENT_IS_FLAT_CODE(seg
)))
98 /* Check the segment exists, is within the current LDT/GDT size,
99 that kernel/user (ring 0..3) has the appropriate privilege,
100 that it's a code segment, and get the limit. */
101 __asm__ ("larl %3,%0; lsll %3,%1"
102 : "=&r" (seg_ar
), "=r" (seg_limit
) : "0" (0), "rm" (seg
));
103 if ((~seg_ar
& 0x9800) || eip
> seg_limit
) {
105 return 1; /* So that returned eip > *eip_limit. */
108 /* Get the GDT/LDT descriptor base.
109 When you look for races in this code remember that
110 LDT and other horrors are only used in user space. */
112 /* Must lock the LDT while reading it. */
113 down(¤t
->mm
->context
.sem
);
114 desc
= current
->mm
->context
.ldt
;
115 desc
= (void *)desc
+ (seg
& ~7);
117 /* Must disable preemption while reading the GDT. */
118 desc
= (u32
*)get_cpu_gdt_table(get_cpu());
119 desc
= (void *)desc
+ (seg
& ~7);
122 /* Decode the code segment base from the descriptor */
123 base
= get_desc_base((unsigned long *)desc
);
126 up(¤t
->mm
->context
.sem
);
130 /* Adjust EIP and segment limit, and clamp at the kernel limit.
131 It's legitimate for segments to wrap at 0xffffffff. */
133 if (seg_limit
< *eip_limit
&& seg_limit
>= base
)
134 *eip_limit
= seg_limit
;
139 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
140 * Check that here and ignore it.
142 static int __is_prefetch(struct pt_regs
*regs
, unsigned long addr
)
145 unsigned char *instr
= (unsigned char *)get_segment_eip (regs
, &limit
);
150 for (i
= 0; scan_more
&& i
< 15; i
++) {
151 unsigned char opcode
;
152 unsigned char instr_hi
;
153 unsigned char instr_lo
;
155 if (instr
> (unsigned char *)limit
)
157 if (probe_kernel_address(instr
, opcode
))
160 instr_hi
= opcode
& 0xf0;
161 instr_lo
= opcode
& 0x0f;
167 /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
168 scan_more
= ((instr_lo
& 7) == 0x6);
172 /* 0x64 thru 0x67 are valid prefixes in all modes. */
173 scan_more
= (instr_lo
& 0xC) == 0x4;
176 /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
177 scan_more
= !instr_lo
|| (instr_lo
>>1) == 1;
180 /* Prefetch instruction is 0x0F0D or 0x0F18 */
182 if (instr
> (unsigned char *)limit
)
184 if (probe_kernel_address(instr
, opcode
))
186 prefetch
= (instr_lo
== 0xF) &&
187 (opcode
== 0x0D || opcode
== 0x18);
197 static inline int is_prefetch(struct pt_regs
*regs
, unsigned long addr
,
198 unsigned long error_code
)
200 if (unlikely(boot_cpu_data
.x86_vendor
== X86_VENDOR_AMD
&&
201 boot_cpu_data
.x86
>= 6)) {
202 /* Catch an obscure case of prefetch inside an NX page. */
203 if (nx_enabled
&& (error_code
& 16))
205 return __is_prefetch(regs
, addr
);
210 static noinline
void force_sig_info_fault(int si_signo
, int si_code
,
211 unsigned long address
, struct task_struct
*tsk
)
215 info
.si_signo
= si_signo
;
217 info
.si_code
= si_code
;
218 info
.si_addr
= (void __user
*)address
;
219 force_sig_info(si_signo
, &info
, tsk
);
222 fastcall
void do_invalid_op(struct pt_regs
*, unsigned long);
224 static inline pmd_t
*vmalloc_sync_one(pgd_t
*pgd
, unsigned long address
)
226 unsigned index
= pgd_index(address
);
232 pgd_k
= init_mm
.pgd
+ index
;
234 if (!pgd_present(*pgd_k
))
238 * set_pgd(pgd, *pgd_k); here would be useless on PAE
239 * and redundant with the set_pmd() on non-PAE. As would
243 pud
= pud_offset(pgd
, address
);
244 pud_k
= pud_offset(pgd_k
, address
);
245 if (!pud_present(*pud_k
))
248 pmd
= pmd_offset(pud
, address
);
249 pmd_k
= pmd_offset(pud_k
, address
);
250 if (!pmd_present(*pmd_k
))
252 if (!pmd_present(*pmd
))
253 set_pmd(pmd
, *pmd_k
);
255 BUG_ON(pmd_page(*pmd
) != pmd_page(*pmd_k
));
260 * Handle a fault on the vmalloc or module mapping area
262 * This assumes no large pages in there.
264 static inline int vmalloc_fault(unsigned long address
)
266 unsigned long pgd_paddr
;
270 * Synchronize this task's top level page-table
271 * with the 'reference' page table.
273 * Do _not_ use "current" here. We might be inside
274 * an interrupt in the middle of a task switch..
276 pgd_paddr
= read_cr3();
277 pmd_k
= vmalloc_sync_one(__va(pgd_paddr
), address
);
280 pte_k
= pte_offset_kernel(pmd_k
, address
);
281 if (!pte_present(*pte_k
))
287 * This routine handles page faults. It determines the address,
288 * and the problem, and then passes it off to one of the appropriate
292 * bit 0 == 0 means no page found, 1 means protection fault
293 * bit 1 == 0 means read, 1 means write
294 * bit 2 == 0 means kernel, 1 means user-mode
295 * bit 3 == 1 means use of reserved bit detected
296 * bit 4 == 1 means fault was an instruction fetch
298 fastcall
void __kprobes
do_page_fault(struct pt_regs
*regs
,
299 unsigned long error_code
)
301 struct task_struct
*tsk
;
302 struct mm_struct
*mm
;
303 struct vm_area_struct
* vma
;
304 unsigned long address
;
308 /* get the address */
309 address
= read_cr2();
313 si_code
= SEGV_MAPERR
;
316 * We fault-in kernel-space virtual memory on-demand. The
317 * 'reference' page table is init_mm.pgd.
319 * NOTE! We MUST NOT take any locks for this case. We may
320 * be in an interrupt or a critical region, and should
321 * only copy the information from the master page table,
324 * This verifies that the fault happens in kernel space
325 * (error_code & 4) == 0, and that the fault was not a
326 * protection error (error_code & 9) == 0.
328 if (unlikely(address
>= TASK_SIZE
)) {
329 if (!(error_code
& 0x0000000d) && vmalloc_fault(address
) >= 0)
331 if (notify_page_fault(regs
, error_code
) == NOTIFY_STOP
)
334 * Don't take the mm semaphore here. If we fixup a prefetch
335 * fault we could otherwise deadlock.
337 goto bad_area_nosemaphore
;
340 if (notify_page_fault(regs
, error_code
) == NOTIFY_STOP
)
343 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
344 fault has been handled. */
345 if (regs
->eflags
& (X86_EFLAGS_IF
|VM_MASK
))
351 * If we're in an interrupt, have no user context or are running in an
352 * atomic region then we must not take the fault..
354 if (in_atomic() || !mm
)
355 goto bad_area_nosemaphore
;
357 /* When running in the kernel we expect faults to occur only to
358 * addresses in user space. All other faults represent errors in the
359 * kernel and should generate an OOPS. Unfortunatly, in the case of an
360 * erroneous fault occurring in a code path which already holds mmap_sem
361 * we will deadlock attempting to validate the fault against the
362 * address space. Luckily the kernel only validly references user
363 * space from well defined areas of code, which are listed in the
366 * As the vast majority of faults will be valid we will only perform
367 * the source reference check when there is a possibilty of a deadlock.
368 * Attempt to lock the address space, if we cannot we then validate the
369 * source. If this is invalid we can skip the address space check,
370 * thus avoiding the deadlock.
372 if (!down_read_trylock(&mm
->mmap_sem
)) {
373 if ((error_code
& 4) == 0 &&
374 !search_exception_tables(regs
->eip
))
375 goto bad_area_nosemaphore
;
376 down_read(&mm
->mmap_sem
);
379 vma
= find_vma(mm
, address
);
382 if (vma
->vm_start
<= address
)
384 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
386 if (error_code
& 4) {
388 * Accessing the stack below %esp is always a bug.
389 * The large cushion allows instructions like enter
390 * and pusha to work. ("enter $65535,$31" pushes
391 * 32 pointers and then decrements %esp by 65535.)
393 if (address
+ 65536 + 32 * sizeof(unsigned long) < regs
->esp
)
396 if (expand_stack(vma
, address
))
399 * Ok, we have a good vm_area for this memory access, so
403 si_code
= SEGV_ACCERR
;
405 switch (error_code
& 3) {
406 default: /* 3: write, present */
408 case 2: /* write, not present */
409 if (!(vma
->vm_flags
& VM_WRITE
))
413 case 1: /* read, present */
415 case 0: /* read, not present */
416 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
422 * If for any reason at all we couldn't handle the fault,
423 * make sure we exit gracefully rather than endlessly redo
426 fault
= handle_mm_fault(mm
, vma
, address
, write
);
427 if (unlikely(fault
& VM_FAULT_ERROR
)) {
428 if (fault
& VM_FAULT_OOM
)
430 else if (fault
& VM_FAULT_SIGBUS
)
434 if (fault
& VM_FAULT_MAJOR
)
440 * Did it hit the DOS screen memory VA from vm86 mode?
442 if (regs
->eflags
& VM_MASK
) {
443 unsigned long bit
= (address
- 0xA0000) >> PAGE_SHIFT
;
445 tsk
->thread
.screen_bitmap
|= 1 << bit
;
447 up_read(&mm
->mmap_sem
);
451 * Something tried to access memory that isn't in our memory map..
452 * Fix it, but check if it's kernel or user first..
455 up_read(&mm
->mmap_sem
);
457 bad_area_nosemaphore
:
458 /* User mode accesses just cause a SIGSEGV */
459 if (error_code
& 4) {
461 * It's possible to have interrupts off here.
466 * Valid to do another page fault here because this one came
469 if (is_prefetch(regs
, address
, error_code
))
472 tsk
->thread
.cr2
= address
;
473 /* Kernel addresses are always protection faults */
474 tsk
->thread
.error_code
= error_code
| (address
>= TASK_SIZE
);
475 tsk
->thread
.trap_no
= 14;
476 force_sig_info_fault(SIGSEGV
, si_code
, address
, tsk
);
480 #ifdef CONFIG_X86_F00F_BUG
482 * Pentium F0 0F C7 C8 bug workaround.
484 if (boot_cpu_data
.f00f_bug
) {
487 nr
= (address
- idt_descr
.address
) >> 3;
490 do_invalid_op(regs
, 0);
497 /* Are we prepared to handle this kernel fault? */
498 if (fixup_exception(regs
))
502 * Valid to do another page fault here, because if this fault
503 * had been triggered by is_prefetch fixup_exception would have
506 if (is_prefetch(regs
, address
, error_code
))
510 * Oops. The kernel tried to access some bad page. We'll have to
511 * terminate things with extreme prejudice.
516 if (oops_may_print()) {
517 __typeof__(pte_val(__pte(0))) page
;
519 #ifdef CONFIG_X86_PAE
520 if (error_code
& 16) {
521 pte_t
*pte
= lookup_address(address
);
523 if (pte
&& pte_present(*pte
) && !pte_exec_kernel(*pte
))
524 printk(KERN_CRIT
"kernel tried to execute "
525 "NX-protected page - exploit attempt? "
526 "(uid: %d)\n", current
->uid
);
529 if (address
< PAGE_SIZE
)
530 printk(KERN_ALERT
"BUG: unable to handle kernel NULL "
531 "pointer dereference");
533 printk(KERN_ALERT
"BUG: unable to handle kernel paging"
535 printk(" at virtual address %08lx\n",address
);
536 printk(KERN_ALERT
" printing eip:\n");
537 printk("%08lx\n", regs
->eip
);
540 page
= ((__typeof__(page
) *) __va(page
))[address
>> PGDIR_SHIFT
];
541 #ifdef CONFIG_X86_PAE
542 printk(KERN_ALERT
"*pdpt = %016Lx\n", page
);
543 if ((page
>> PAGE_SHIFT
) < max_low_pfn
544 && page
& _PAGE_PRESENT
) {
546 page
= ((__typeof__(page
) *) __va(page
))[(address
>> PMD_SHIFT
)
547 & (PTRS_PER_PMD
- 1)];
548 printk(KERN_ALERT
"*pde = %016Lx\n", page
);
552 printk(KERN_ALERT
"*pde = %08lx\n", page
);
556 * We must not directly access the pte in the highpte
557 * case if the page table is located in highmem.
558 * And let's rather not kmap-atomic the pte, just in case
559 * it's allocated already.
561 if ((page
>> PAGE_SHIFT
) < max_low_pfn
562 && (page
& _PAGE_PRESENT
)) {
564 page
= ((__typeof__(page
) *) __va(page
))[(address
>> PAGE_SHIFT
)
565 & (PTRS_PER_PTE
- 1)];
566 printk(KERN_ALERT
"*pte = %0*Lx\n", sizeof(page
)*2, (u64
)page
);
570 tsk
->thread
.cr2
= address
;
571 tsk
->thread
.trap_no
= 14;
572 tsk
->thread
.error_code
= error_code
;
573 die("Oops", regs
, error_code
);
578 * We ran out of memory, or some other thing happened to us that made
579 * us unable to handle the page fault gracefully.
582 up_read(&mm
->mmap_sem
);
585 down_read(&mm
->mmap_sem
);
588 printk("VM: killing process %s\n", tsk
->comm
);
594 up_read(&mm
->mmap_sem
);
596 /* Kernel mode? Handle exceptions or die */
597 if (!(error_code
& 4))
600 /* User space => ok to do another page fault */
601 if (is_prefetch(regs
, address
, error_code
))
604 tsk
->thread
.cr2
= address
;
605 tsk
->thread
.error_code
= error_code
;
606 tsk
->thread
.trap_no
= 14;
607 force_sig_info_fault(SIGBUS
, BUS_ADRERR
, address
, tsk
);
610 void vmalloc_sync_all(void)
613 * Note that races in the updates of insync and start aren't
614 * problematic: insync can only get set bits added, and updates to
615 * start are only improving performance (without affecting correctness
618 static DECLARE_BITMAP(insync
, PTRS_PER_PGD
);
619 static unsigned long start
= TASK_SIZE
;
620 unsigned long address
;
622 if (SHARED_KERNEL_PMD
)
625 BUILD_BUG_ON(TASK_SIZE
& ~PGDIR_MASK
);
626 for (address
= start
; address
>= TASK_SIZE
; address
+= PGDIR_SIZE
) {
627 if (!test_bit(pgd_index(address
), insync
)) {
631 spin_lock_irqsave(&pgd_lock
, flags
);
632 for (page
= pgd_list
; page
; page
=
633 (struct page
*)page
->index
)
634 if (!vmalloc_sync_one(page_address(page
),
636 BUG_ON(page
!= pgd_list
);
639 spin_unlock_irqrestore(&pgd_lock
, flags
);
641 set_bit(pgd_index(address
), insync
);
643 if (address
== start
&& test_bit(pgd_index(address
), insync
))
644 start
= address
+ PGDIR_SIZE
;