2 * Copyright (C) 1995 Linus Torvalds
5 #include <linux/signal.h>
6 #include <linux/sched.h>
7 #include <linux/kernel.h>
8 #include <linux/errno.h>
9 #include <linux/string.h>
10 #include <linux/types.h>
11 #include <linux/ptrace.h>
12 #include <linux/mman.h>
14 #include <linux/smp.h>
15 #include <linux/interrupt.h>
16 #include <linux/init.h>
17 #include <linux/tty.h>
18 #include <linux/vt_kern.h> /* For unblank_screen() */
19 #include <linux/highmem.h>
20 #include <linux/bootmem.h> /* for max_low_pfn */
21 #include <linux/vmalloc.h>
22 #include <linux/module.h>
23 #include <linux/kprobes.h>
24 #include <linux/uaccess.h>
25 #include <linux/kdebug.h>
27 #include <asm/system.h>
29 #include <asm/segment.h>
32 * Page fault error code bits
33 * bit 0 == 0 means no page found, 1 means protection fault
34 * bit 1 == 0 means read, 1 means write
35 * bit 2 == 0 means kernel, 1 means user-mode
36 * bit 3 == 1 means use of reserved bit detected
37 * bit 4 == 1 means fault was an instruction fetch
39 #define PF_PROT (1<<0)
40 #define PF_WRITE (1<<1)
41 #define PF_USER (1<<2)
42 #define PF_RSVD (1<<3)
43 #define PF_INSTR (1<<4)
45 static inline int notify_page_fault(struct pt_regs
*regs
)
50 /* kprobe_running() needs smp_processor_id() */
51 if (!user_mode_vm(regs
)) {
53 if (kprobe_running() && kprobe_fault_handler(regs
, 14))
66 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
67 * Check that here and ignore it.
70 * Sometimes the CPU reports invalid exceptions on prefetch.
71 * Check that here and ignore it.
73 * Opcode checker based on code by Richard Brunner
75 static int is_prefetch(struct pt_regs
*regs
, unsigned long addr
,
76 unsigned long error_code
)
81 unsigned char *max_instr
;
84 if (unlikely(boot_cpu_data
.x86_vendor
== X86_VENDOR_AMD
&&
85 boot_cpu_data
.x86
>= 6)) {
86 /* Catch an obscure case of prefetch inside an NX page. */
87 if (nx_enabled
&& (error_code
& PF_INSTR
))
93 /* If it was a exec fault ignore */
94 if (error_code
& PF_INSTR
)
98 instr
= (unsigned char *)convert_ip_to_linear(current
, regs
);
99 max_instr
= instr
+ 15;
101 if (user_mode(regs
) && instr
>= (unsigned char *)TASK_SIZE
)
104 while (scan_more
&& instr
< max_instr
) {
105 unsigned char opcode
;
106 unsigned char instr_hi
;
107 unsigned char instr_lo
;
109 if (probe_kernel_address(instr
, opcode
))
112 instr_hi
= opcode
& 0xf0;
113 instr_lo
= opcode
& 0x0f;
120 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
121 * In X86_64 long mode, the CPU will signal invalid
122 * opcode if some of these prefixes are present so
123 * X86_64 will never get here anyway
125 scan_more
= ((instr_lo
& 7) == 0x6);
130 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
131 * Need to figure out under what instruction mode the
132 * instruction was issued. Could check the LDT for lm,
133 * but for now it's good enough to assume that long
134 * mode only uses well known segments or kernel.
136 scan_more
= (!user_mode(regs
)) || (regs
->cs
== __USER_CS
);
140 /* 0x64 thru 0x67 are valid prefixes in all modes. */
141 scan_more
= (instr_lo
& 0xC) == 0x4;
144 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
145 scan_more
= !instr_lo
|| (instr_lo
>>1) == 1;
148 /* Prefetch instruction is 0x0F0D or 0x0F18 */
151 if (probe_kernel_address(instr
, opcode
))
153 prefetch
= (instr_lo
== 0xF) &&
154 (opcode
== 0x0D || opcode
== 0x18);
164 static void force_sig_info_fault(int si_signo
, int si_code
,
165 unsigned long address
, struct task_struct
*tsk
)
169 info
.si_signo
= si_signo
;
171 info
.si_code
= si_code
;
172 info
.si_addr
= (void __user
*)address
;
173 force_sig_info(si_signo
, &info
, tsk
);
176 void do_invalid_op(struct pt_regs
*, unsigned long);
178 static inline pmd_t
*vmalloc_sync_one(pgd_t
*pgd
, unsigned long address
)
180 unsigned index
= pgd_index(address
);
186 pgd_k
= init_mm
.pgd
+ index
;
188 if (!pgd_present(*pgd_k
))
192 * set_pgd(pgd, *pgd_k); here would be useless on PAE
193 * and redundant with the set_pmd() on non-PAE. As would
197 pud
= pud_offset(pgd
, address
);
198 pud_k
= pud_offset(pgd_k
, address
);
199 if (!pud_present(*pud_k
))
202 pmd
= pmd_offset(pud
, address
);
203 pmd_k
= pmd_offset(pud_k
, address
);
204 if (!pmd_present(*pmd_k
))
206 if (!pmd_present(*pmd
)) {
207 set_pmd(pmd
, *pmd_k
);
208 arch_flush_lazy_mmu_mode();
210 BUG_ON(pmd_page(*pmd
) != pmd_page(*pmd_k
));
215 static const char errata93_warning
[] =
216 KERN_ERR
"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
217 KERN_ERR
"******* Working around it, but it may cause SEGVs or burn power.\n"
218 KERN_ERR
"******* Please consider a BIOS update.\n"
219 KERN_ERR
"******* Disabling USB legacy in the BIOS may also help.\n";
222 /* Workaround for K8 erratum #93 & buggy BIOS.
223 BIOS SMM functions are required to use a specific workaround
224 to avoid corruption of the 64bit RIP register on C stepping K8.
225 A lot of BIOS that didn't get tested properly miss this.
226 The OS sees this as a page fault with the upper 32bits of RIP cleared.
227 Try to work around it here.
228 Note we only handle faults in kernel here.
229 Does nothing for X86_32
231 static int is_errata93(struct pt_regs
*regs
, unsigned long address
)
235 if (address
!= regs
->ip
)
237 if ((address
>> 32) != 0)
239 address
|= 0xffffffffUL
<< 32;
240 if ((address
>= (u64
)_stext
&& address
<= (u64
)_etext
) ||
241 (address
>= MODULES_VADDR
&& address
<= MODULES_END
)) {
243 printk(errata93_warning
);
255 * Handle a fault on the vmalloc or module mapping area
257 * This assumes no large pages in there.
259 static inline int vmalloc_fault(unsigned long address
)
262 unsigned long pgd_paddr
;
266 * Synchronize this task's top level page-table
267 * with the 'reference' page table.
269 * Do _not_ use "current" here. We might be inside
270 * an interrupt in the middle of a task switch..
272 pgd_paddr
= read_cr3();
273 pmd_k
= vmalloc_sync_one(__va(pgd_paddr
), address
);
276 pte_k
= pte_offset_kernel(pmd_k
, address
);
277 if (!pte_present(*pte_k
))
281 pgd_t
*pgd
, *pgd_ref
;
282 pud_t
*pud
, *pud_ref
;
283 pmd_t
*pmd
, *pmd_ref
;
284 pte_t
*pte
, *pte_ref
;
286 /* Copy kernel mappings over when needed. This can also
287 happen within a race in page table update. In the later
290 pgd
= pgd_offset(current
->mm
?: &init_mm
, address
);
291 pgd_ref
= pgd_offset_k(address
);
292 if (pgd_none(*pgd_ref
))
295 set_pgd(pgd
, *pgd_ref
);
297 BUG_ON(pgd_page_vaddr(*pgd
) != pgd_page_vaddr(*pgd_ref
));
299 /* Below here mismatches are bugs because these lower tables
302 pud
= pud_offset(pgd
, address
);
303 pud_ref
= pud_offset(pgd_ref
, address
);
304 if (pud_none(*pud_ref
))
306 if (pud_none(*pud
) || pud_page_vaddr(*pud
) != pud_page_vaddr(*pud_ref
))
308 pmd
= pmd_offset(pud
, address
);
309 pmd_ref
= pmd_offset(pud_ref
, address
);
310 if (pmd_none(*pmd_ref
))
312 if (pmd_none(*pmd
) || pmd_page(*pmd
) != pmd_page(*pmd_ref
))
314 pte_ref
= pte_offset_kernel(pmd_ref
, address
);
315 if (!pte_present(*pte_ref
))
317 pte
= pte_offset_kernel(pmd
, address
);
318 /* Don't use pte_page here, because the mappings can point
319 outside mem_map, and the NUMA hash lookup cannot handle
321 if (!pte_present(*pte
) || pte_pfn(*pte
) != pte_pfn(*pte_ref
))
327 int show_unhandled_signals
= 1;
330 * This routine handles page faults. It determines the address,
331 * and the problem, and then passes it off to one of the appropriate
334 void __kprobes
do_page_fault(struct pt_regs
*regs
, unsigned long error_code
)
336 struct task_struct
*tsk
;
337 struct mm_struct
*mm
;
338 struct vm_area_struct
*vma
;
339 unsigned long address
;
344 * We can fault from pretty much anywhere, with unknown IRQ state.
346 trace_hardirqs_fixup();
350 prefetchw(&mm
->mmap_sem
);
352 /* get the address */
353 address
= read_cr2();
355 si_code
= SEGV_MAPERR
;
357 if (notify_page_fault(regs
))
361 * We fault-in kernel-space virtual memory on-demand. The
362 * 'reference' page table is init_mm.pgd.
364 * NOTE! We MUST NOT take any locks for this case. We may
365 * be in an interrupt or a critical region, and should
366 * only copy the information from the master page table,
369 * This verifies that the fault happens in kernel space
370 * (error_code & 4) == 0, and that the fault was not a
371 * protection error (error_code & 9) == 0.
373 if (unlikely(address
>= TASK_SIZE
)) {
374 if (!(error_code
& (PF_RSVD
|PF_USER
|PF_PROT
)) &&
375 vmalloc_fault(address
) >= 0)
378 * Don't take the mm semaphore here. If we fixup a prefetch
379 * fault we could otherwise deadlock.
381 goto bad_area_nosemaphore
;
384 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
385 fault has been handled. */
386 if (regs
->flags
& (X86_EFLAGS_IF
|VM_MASK
))
390 * If we're in an interrupt, have no user context or are running in an
391 * atomic region then we must not take the fault.
393 if (in_atomic() || !mm
)
394 goto bad_area_nosemaphore
;
396 /* When running in the kernel we expect faults to occur only to
397 * addresses in user space. All other faults represent errors in the
398 * kernel and should generate an OOPS. Unfortunately, in the case of an
399 * erroneous fault occurring in a code path which already holds mmap_sem
400 * we will deadlock attempting to validate the fault against the
401 * address space. Luckily the kernel only validly references user
402 * space from well defined areas of code, which are listed in the
405 * As the vast majority of faults will be valid we will only perform
406 * the source reference check when there is a possibility of a deadlock.
407 * Attempt to lock the address space, if we cannot we then validate the
408 * source. If this is invalid we can skip the address space check,
409 * thus avoiding the deadlock.
411 if (!down_read_trylock(&mm
->mmap_sem
)) {
412 if ((error_code
& PF_USER
) == 0 &&
413 !search_exception_tables(regs
->ip
))
414 goto bad_area_nosemaphore
;
415 down_read(&mm
->mmap_sem
);
418 vma
= find_vma(mm
, address
);
421 if (vma
->vm_start
<= address
)
423 if (!(vma
->vm_flags
& VM_GROWSDOWN
))
425 if (error_code
& PF_USER
) {
427 * Accessing the stack below %sp is always a bug.
428 * The large cushion allows instructions like enter
429 * and pusha to work. ("enter $65535,$31" pushes
430 * 32 pointers and then decrements %sp by 65535.)
432 if (address
+ 65536 + 32 * sizeof(unsigned long) < regs
->sp
)
435 if (expand_stack(vma
, address
))
438 * Ok, we have a good vm_area for this memory access, so
442 si_code
= SEGV_ACCERR
;
444 switch (error_code
& (PF_PROT
|PF_WRITE
)) {
445 default: /* 3: write, present */
447 case PF_WRITE
: /* write, not present */
448 if (!(vma
->vm_flags
& VM_WRITE
))
452 case PF_PROT
: /* read, present */
454 case 0: /* read, not present */
455 if (!(vma
->vm_flags
& (VM_READ
| VM_EXEC
| VM_WRITE
)))
461 * If for any reason at all we couldn't handle the fault,
462 * make sure we exit gracefully rather than endlessly redo
465 fault
= handle_mm_fault(mm
, vma
, address
, write
);
466 if (unlikely(fault
& VM_FAULT_ERROR
)) {
467 if (fault
& VM_FAULT_OOM
)
469 else if (fault
& VM_FAULT_SIGBUS
)
473 if (fault
& VM_FAULT_MAJOR
)
479 * Did it hit the DOS screen memory VA from vm86 mode?
481 if (regs
->flags
& VM_MASK
) {
482 unsigned long bit
= (address
- 0xA0000) >> PAGE_SHIFT
;
484 tsk
->thread
.screen_bitmap
|= 1 << bit
;
486 up_read(&mm
->mmap_sem
);
490 * Something tried to access memory that isn't in our memory map..
491 * Fix it, but check if it's kernel or user first..
494 up_read(&mm
->mmap_sem
);
496 bad_area_nosemaphore
:
497 /* User mode accesses just cause a SIGSEGV */
498 if (error_code
& PF_USER
) {
500 * It's possible to have interrupts off here.
505 * Valid to do another page fault here because this one came
508 if (is_prefetch(regs
, address
, error_code
))
511 if (show_unhandled_signals
&& unhandled_signal(tsk
, SIGSEGV
) &&
512 printk_ratelimit()) {
515 "%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx\n",
517 "%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx\n",
519 task_pid_nr(tsk
) > 1 ? KERN_INFO
: KERN_EMERG
,
520 tsk
->comm
, task_pid_nr(tsk
), address
, regs
->ip
,
521 regs
->sp
, error_code
);
523 tsk
->thread
.cr2
= address
;
524 /* Kernel addresses are always protection faults */
525 tsk
->thread
.error_code
= error_code
| (address
>= TASK_SIZE
);
526 tsk
->thread
.trap_no
= 14;
527 force_sig_info_fault(SIGSEGV
, si_code
, address
, tsk
);
531 #ifdef CONFIG_X86_F00F_BUG
533 * Pentium F0 0F C7 C8 bug workaround.
535 if (boot_cpu_data
.f00f_bug
) {
538 nr
= (address
- idt_descr
.address
) >> 3;
541 do_invalid_op(regs
, 0);
548 /* Are we prepared to handle this kernel fault? */
549 if (fixup_exception(regs
))
553 * Valid to do another page fault here, because if this fault
554 * had been triggered by is_prefetch fixup_exception would have
557 if (is_prefetch(regs
, address
, error_code
))
560 if (is_errata93(regs
, address
))
564 * Oops. The kernel tried to access some bad page. We'll have to
565 * terminate things with extreme prejudice.
570 if (oops_may_print()) {
571 __typeof__(pte_val(__pte(0))) page
;
573 #ifdef CONFIG_X86_PAE
574 if (error_code
& PF_INSTR
) {
575 pte_t
*pte
= lookup_address(address
);
577 if (pte
&& pte_present(*pte
) && !pte_exec_kernel(*pte
))
578 printk(KERN_CRIT
"kernel tried to execute "
579 "NX-protected page - exploit attempt? "
580 "(uid: %d)\n", current
->uid
);
583 if (address
< PAGE_SIZE
)
584 printk(KERN_ALERT
"BUG: unable to handle kernel NULL "
585 "pointer dereference");
587 printk(KERN_ALERT
"BUG: unable to handle kernel paging"
589 printk(" at virtual address %08lx\n", address
);
590 printk(KERN_ALERT
"printing ip: %08lx ", regs
->ip
);
593 page
= ((__typeof__(page
) *) __va(page
))[address
>> PGDIR_SHIFT
];
594 #ifdef CONFIG_X86_PAE
595 printk("*pdpt = %016Lx ", page
);
596 if ((page
>> PAGE_SHIFT
) < max_low_pfn
597 && page
& _PAGE_PRESENT
) {
599 page
= ((__typeof__(page
) *) __va(page
))[(address
>> PMD_SHIFT
)
600 & (PTRS_PER_PMD
- 1)];
601 printk(KERN_CONT
"*pde = %016Lx ", page
);
605 printk("*pde = %08lx ", page
);
609 * We must not directly access the pte in the highpte
610 * case if the page table is located in highmem.
611 * And let's rather not kmap-atomic the pte, just in case
612 * it's allocated already.
614 if ((page
>> PAGE_SHIFT
) < max_low_pfn
615 && (page
& _PAGE_PRESENT
)
616 && !(page
& _PAGE_PSE
)) {
618 page
= ((__typeof__(page
) *) __va(page
))[(address
>> PAGE_SHIFT
)
619 & (PTRS_PER_PTE
- 1)];
620 printk("*pte = %0*Lx ", sizeof(page
)*2, (u64
)page
);
626 tsk
->thread
.cr2
= address
;
627 tsk
->thread
.trap_no
= 14;
628 tsk
->thread
.error_code
= error_code
;
629 die("Oops", regs
, error_code
);
634 * We ran out of memory, or some other thing happened to us that made
635 * us unable to handle the page fault gracefully.
638 up_read(&mm
->mmap_sem
);
639 if (is_global_init(tsk
)) {
641 down_read(&mm
->mmap_sem
);
644 printk("VM: killing process %s\n", tsk
->comm
);
645 if (error_code
& PF_USER
)
646 do_group_exit(SIGKILL
);
650 up_read(&mm
->mmap_sem
);
652 /* Kernel mode? Handle exceptions or die */
653 if (!(error_code
& PF_USER
))
656 /* User space => ok to do another page fault */
657 if (is_prefetch(regs
, address
, error_code
))
660 tsk
->thread
.cr2
= address
;
661 tsk
->thread
.error_code
= error_code
;
662 tsk
->thread
.trap_no
= 14;
663 force_sig_info_fault(SIGBUS
, BUS_ADRERR
, address
, tsk
);
666 void vmalloc_sync_all(void)
669 * Note that races in the updates of insync and start aren't
670 * problematic: insync can only get set bits added, and updates to
671 * start are only improving performance (without affecting correctness
674 static DECLARE_BITMAP(insync
, PTRS_PER_PGD
);
675 static unsigned long start
= TASK_SIZE
;
676 unsigned long address
;
678 if (SHARED_KERNEL_PMD
)
681 BUILD_BUG_ON(TASK_SIZE
& ~PGDIR_MASK
);
682 for (address
= start
; address
>= TASK_SIZE
; address
+= PGDIR_SIZE
) {
683 if (!test_bit(pgd_index(address
), insync
)) {
687 spin_lock_irqsave(&pgd_lock
, flags
);
688 for (page
= pgd_list
; page
; page
=
689 (struct page
*)page
->index
)
690 if (!vmalloc_sync_one(page_address(page
),
692 BUG_ON(page
!= pgd_list
);
695 spin_unlock_irqrestore(&pgd_lock
, flags
);
697 set_bit(pgd_index(address
), insync
);
699 if (address
== start
&& test_bit(pgd_index(address
), insync
))
700 start
= address
+ PGDIR_SIZE
;