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 dump_pagetable(unsigned long address)
178 __typeof__(pte_val(__pte(0))) page;
181 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
182 #ifdef CONFIG_X86_PAE
183 printk("*pdpt = %016Lx ", page);
184 if ((page >> PAGE_SHIFT) < max_low_pfn
185 && page & _PAGE_PRESENT) {
187 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
188 & (PTRS_PER_PMD - 1)];
189 printk(KERN_CONT "*pde = %016Lx ", page);
193 printk("*pde = %08lx ", page);
197 * We must not directly access the pte in the highpte
198 * case if the page table is located in highmem.
199 * And let's rather not kmap-atomic the pte, just in case
200 * it's allocated already.
202 if ((page >> PAGE_SHIFT) < max_low_pfn
203 && (page & _PAGE_PRESENT)
204 && !(page & _PAGE_PSE)) {
206 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
207 & (PTRS_PER_PTE - 1)];
208 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
214 void do_invalid_op(struct pt_regs *, unsigned long);
216 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
218 unsigned index = pgd_index(address);
224 pgd_k = init_mm.pgd + index;
226 if (!pgd_present(*pgd_k))
230 * set_pgd(pgd, *pgd_k); here would be useless on PAE
231 * and redundant with the set_pmd() on non-PAE. As would
235 pud = pud_offset(pgd, address);
236 pud_k = pud_offset(pgd_k, address);
237 if (!pud_present(*pud_k))
240 pmd = pmd_offset(pud, address);
241 pmd_k = pmd_offset(pud_k, address);
242 if (!pmd_present(*pmd_k))
244 if (!pmd_present(*pmd)) {
245 set_pmd(pmd, *pmd_k);
246 arch_flush_lazy_mmu_mode();
248 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
253 static const char errata93_warning[] =
254 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
255 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
256 KERN_ERR "******* Please consider a BIOS update.\n"
257 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
260 /* Workaround for K8 erratum #93 & buggy BIOS.
261 BIOS SMM functions are required to use a specific workaround
262 to avoid corruption of the 64bit RIP register on C stepping K8.
263 A lot of BIOS that didn't get tested properly miss this.
264 The OS sees this as a page fault with the upper 32bits of RIP cleared.
265 Try to work around it here.
266 Note we only handle faults in kernel here.
267 Does nothing for X86_32
269 static int is_errata93(struct pt_regs *regs, unsigned long address)
273 if (address != regs->ip)
275 if ((address >> 32) != 0)
277 address |= 0xffffffffUL << 32;
278 if ((address >= (u64)_stext && address <= (u64)_etext) ||
279 (address >= MODULES_VADDR && address <= MODULES_END)) {
281 printk(errata93_warning);
293 * Handle a fault on the vmalloc or module mapping area
295 * This assumes no large pages in there.
297 static inline int vmalloc_fault(unsigned long address)
300 unsigned long pgd_paddr;
304 * Synchronize this task's top level page-table
305 * with the 'reference' page table.
307 * Do _not_ use "current" here. We might be inside
308 * an interrupt in the middle of a task switch..
310 pgd_paddr = read_cr3();
311 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
314 pte_k = pte_offset_kernel(pmd_k, address);
315 if (!pte_present(*pte_k))
319 pgd_t *pgd, *pgd_ref;
320 pud_t *pud, *pud_ref;
321 pmd_t *pmd, *pmd_ref;
322 pte_t *pte, *pte_ref;
324 /* Copy kernel mappings over when needed. This can also
325 happen within a race in page table update. In the later
328 pgd = pgd_offset(current->mm ?: &init_mm, address);
329 pgd_ref = pgd_offset_k(address);
330 if (pgd_none(*pgd_ref))
333 set_pgd(pgd, *pgd_ref);
335 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
337 /* Below here mismatches are bugs because these lower tables
340 pud = pud_offset(pgd, address);
341 pud_ref = pud_offset(pgd_ref, address);
342 if (pud_none(*pud_ref))
344 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
346 pmd = pmd_offset(pud, address);
347 pmd_ref = pmd_offset(pud_ref, address);
348 if (pmd_none(*pmd_ref))
350 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
352 pte_ref = pte_offset_kernel(pmd_ref, address);
353 if (!pte_present(*pte_ref))
355 pte = pte_offset_kernel(pmd, address);
356 /* Don't use pte_page here, because the mappings can point
357 outside mem_map, and the NUMA hash lookup cannot handle
359 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
365 int show_unhandled_signals = 1;
368 * This routine handles page faults. It determines the address,
369 * and the problem, and then passes it off to one of the appropriate
372 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
374 struct task_struct *tsk;
375 struct mm_struct *mm;
376 struct vm_area_struct *vma;
377 unsigned long address;
382 * We can fault from pretty much anywhere, with unknown IRQ state.
384 trace_hardirqs_fixup();
388 prefetchw(&mm->mmap_sem);
390 /* get the address */
391 address = read_cr2();
393 si_code = SEGV_MAPERR;
395 if (notify_page_fault(regs))
399 * We fault-in kernel-space virtual memory on-demand. The
400 * 'reference' page table is init_mm.pgd.
402 * NOTE! We MUST NOT take any locks for this case. We may
403 * be in an interrupt or a critical region, and should
404 * only copy the information from the master page table,
407 * This verifies that the fault happens in kernel space
408 * (error_code & 4) == 0, and that the fault was not a
409 * protection error (error_code & 9) == 0.
411 if (unlikely(address >= TASK_SIZE)) {
412 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
413 vmalloc_fault(address) >= 0)
416 * Don't take the mm semaphore here. If we fixup a prefetch
417 * fault we could otherwise deadlock.
419 goto bad_area_nosemaphore;
422 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
423 fault has been handled. */
424 if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
428 * If we're in an interrupt, have no user context or are running in an
429 * atomic region then we must not take the fault.
431 if (in_atomic() || !mm)
432 goto bad_area_nosemaphore;
434 /* When running in the kernel we expect faults to occur only to
435 * addresses in user space. All other faults represent errors in the
436 * kernel and should generate an OOPS. Unfortunately, in the case of an
437 * erroneous fault occurring in a code path which already holds mmap_sem
438 * we will deadlock attempting to validate the fault against the
439 * address space. Luckily the kernel only validly references user
440 * space from well defined areas of code, which are listed in the
443 * As the vast majority of faults will be valid we will only perform
444 * the source reference check when there is a possibility of a deadlock.
445 * Attempt to lock the address space, if we cannot we then validate the
446 * source. If this is invalid we can skip the address space check,
447 * thus avoiding the deadlock.
449 if (!down_read_trylock(&mm->mmap_sem)) {
450 if ((error_code & PF_USER) == 0 &&
451 !search_exception_tables(regs->ip))
452 goto bad_area_nosemaphore;
453 down_read(&mm->mmap_sem);
456 vma = find_vma(mm, address);
459 if (vma->vm_start <= address)
461 if (!(vma->vm_flags & VM_GROWSDOWN))
463 if (error_code & PF_USER) {
465 * Accessing the stack below %sp is always a bug.
466 * The large cushion allows instructions like enter
467 * and pusha to work. ("enter $65535,$31" pushes
468 * 32 pointers and then decrements %sp by 65535.)
470 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
473 if (expand_stack(vma, address))
476 * Ok, we have a good vm_area for this memory access, so
480 si_code = SEGV_ACCERR;
482 switch (error_code & (PF_PROT|PF_WRITE)) {
483 default: /* 3: write, present */
485 case PF_WRITE: /* write, not present */
486 if (!(vma->vm_flags & VM_WRITE))
490 case PF_PROT: /* read, present */
492 case 0: /* read, not present */
493 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
499 * If for any reason at all we couldn't handle the fault,
500 * make sure we exit gracefully rather than endlessly redo
503 fault = handle_mm_fault(mm, vma, address, write);
504 if (unlikely(fault & VM_FAULT_ERROR)) {
505 if (fault & VM_FAULT_OOM)
507 else if (fault & VM_FAULT_SIGBUS)
511 if (fault & VM_FAULT_MAJOR)
518 * Did it hit the DOS screen memory VA from vm86 mode?
520 if (v8086_mode(regs)) {
521 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
523 tsk->thread.screen_bitmap |= 1 << bit;
526 up_read(&mm->mmap_sem);
530 * Something tried to access memory that isn't in our memory map..
531 * Fix it, but check if it's kernel or user first..
534 up_read(&mm->mmap_sem);
536 bad_area_nosemaphore:
537 /* User mode accesses just cause a SIGSEGV */
538 if (error_code & PF_USER) {
540 * It's possible to have interrupts off here.
545 * Valid to do another page fault here because this one came
548 if (is_prefetch(regs, address, error_code))
551 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
552 printk_ratelimit()) {
555 "%s%s[%d]: segfault at %lx ip %08lx sp %08lx error %lx",
557 "%s%s[%d]: segfault at %lx ip %lx sp %lx error %lx",
559 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
560 tsk->comm, task_pid_nr(tsk), address, regs->ip,
561 regs->sp, error_code);
562 print_vma_addr(" in ", regs->ip);
565 tsk->thread.cr2 = address;
566 /* Kernel addresses are always protection faults */
567 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
568 tsk->thread.trap_no = 14;
569 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
573 #ifdef CONFIG_X86_F00F_BUG
575 * Pentium F0 0F C7 C8 bug workaround.
577 if (boot_cpu_data.f00f_bug) {
580 nr = (address - idt_descr.address) >> 3;
583 do_invalid_op(regs, 0);
590 /* Are we prepared to handle this kernel fault? */
591 if (fixup_exception(regs))
595 * Valid to do another page fault here, because if this fault
596 * had been triggered by is_prefetch fixup_exception would have
599 if (is_prefetch(regs, address, error_code))
602 if (is_errata93(regs, address))
606 * Oops. The kernel tried to access some bad page. We'll have to
607 * terminate things with extreme prejudice.
612 if (oops_may_print()) {
614 #ifdef CONFIG_X86_PAE
615 if (error_code & PF_INSTR) {
616 pte_t *pte = lookup_address(address);
618 if (pte && pte_present(*pte) && !pte_exec(*pte))
619 printk(KERN_CRIT "kernel tried to execute "
620 "NX-protected page - exploit attempt? "
621 "(uid: %d)\n", current->uid);
624 if (address < PAGE_SIZE)
625 printk(KERN_ALERT "BUG: unable to handle kernel NULL "
626 "pointer dereference");
628 printk(KERN_ALERT "BUG: unable to handle kernel paging"
630 printk(" at virtual address %08lx\n", address);
631 printk(KERN_ALERT "printing ip: %08lx ", regs->ip);
633 dump_pagetable(address);
636 tsk->thread.cr2 = address;
637 tsk->thread.trap_no = 14;
638 tsk->thread.error_code = error_code;
639 die("Oops", regs, error_code);
644 * We ran out of memory, or some other thing happened to us that made
645 * us unable to handle the page fault gracefully.
648 up_read(&mm->mmap_sem);
649 if (is_global_init(tsk)) {
651 down_read(&mm->mmap_sem);
654 printk("VM: killing process %s\n", tsk->comm);
655 if (error_code & PF_USER)
656 do_group_exit(SIGKILL);
660 up_read(&mm->mmap_sem);
662 /* Kernel mode? Handle exceptions or die */
663 if (!(error_code & PF_USER))
666 /* User space => ok to do another page fault */
667 if (is_prefetch(regs, address, error_code))
670 tsk->thread.cr2 = address;
671 tsk->thread.error_code = error_code;
672 tsk->thread.trap_no = 14;
673 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
676 void vmalloc_sync_all(void)
679 * Note that races in the updates of insync and start aren't
680 * problematic: insync can only get set bits added, and updates to
681 * start are only improving performance (without affecting correctness
684 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
685 static unsigned long start = TASK_SIZE;
686 unsigned long address;
688 if (SHARED_KERNEL_PMD)
691 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
692 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
693 if (!test_bit(pgd_index(address), insync)) {
697 spin_lock_irqsave(&pgd_lock, flags);
698 for (page = pgd_list; page; page =
699 (struct page *)page->index)
700 if (!vmalloc_sync_one(page_address(page),
702 BUG_ON(page != pgd_list);
705 spin_unlock_irqrestore(&pgd_lock, flags);
707 set_bit(pgd_index(address), insync);
709 if (address == start && test_bit(pgd_index(address), insync))
710 start = address + PGDIR_SIZE;