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);
36 static inline int notify_page_fault(struct pt_regs *regs)
40 /* kprobe_running() needs smp_processor_id() */
41 if (!user_mode_vm(regs)) {
43 if (kprobe_running() && kprobe_fault_handler(regs, 14))
51 static inline int notify_page_fault(struct pt_regs *regs)
58 * Return EIP plus the CS segment base. The segment limit is also
59 * adjusted, clamped to the kernel/user address space (whichever is
60 * appropriate), and returned in *eip_limit.
62 * The segment is checked, because it might have been changed by another
63 * task between the original faulting instruction and here.
65 * If CS is no longer a valid code segment, or if EIP is beyond the
66 * limit, or if it is a kernel address when CS is not a kernel segment,
67 * then the returned value will be greater than *eip_limit.
69 * This is slow, but is very rarely executed.
71 static inline unsigned long get_segment_eip(struct pt_regs *regs,
72 unsigned long *eip_limit)
74 unsigned long ip = regs->ip;
75 unsigned seg = regs->cs & 0xffff;
76 u32 seg_ar, seg_limit, base, *desc;
78 /* Unlikely, but must come before segment checks. */
79 if (unlikely(regs->flags & VM_MASK)) {
81 *eip_limit = base + 0xffff;
82 return base + (ip & 0xffff);
85 /* The standard kernel/user address space limit. */
86 *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
88 /* By far the most common cases. */
89 if (likely(SEGMENT_IS_FLAT_CODE(seg)))
92 /* Check the segment exists, is within the current LDT/GDT size,
93 that kernel/user (ring 0..3) has the appropriate privilege,
94 that it's a code segment, and get the limit. */
95 __asm__ ("larl %3,%0; lsll %3,%1"
96 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
97 if ((~seg_ar & 0x9800) || ip > seg_limit) {
99 return 1; /* So that returned ip > *eip_limit. */
102 /* Get the GDT/LDT descriptor base.
103 When you look for races in this code remember that
104 LDT and other horrors are only used in user space. */
106 /* Must lock the LDT while reading it. */
107 mutex_lock(¤t->mm->context.lock);
108 desc = current->mm->context.ldt;
109 desc = (void *)desc + (seg & ~7);
111 /* Must disable preemption while reading the GDT. */
112 desc = (u32 *)get_cpu_gdt_table(get_cpu());
113 desc = (void *)desc + (seg & ~7);
116 /* Decode the code segment base from the descriptor */
117 base = get_desc_base((struct desc_struct *)desc);
120 mutex_unlock(¤t->mm->context.lock);
124 /* Adjust EIP and segment limit, and clamp at the kernel limit.
125 It's legitimate for segments to wrap at 0xffffffff. */
127 if (seg_limit < *eip_limit && seg_limit >= base)
128 *eip_limit = seg_limit;
133 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
134 * Check that here and ignore it.
136 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
139 unsigned char *instr = (unsigned char *)get_segment_eip (regs, &limit);
144 for (i = 0; scan_more && i < 15; i++) {
145 unsigned char opcode;
146 unsigned char instr_hi;
147 unsigned char instr_lo;
149 if (instr > (unsigned char *)limit)
151 if (probe_kernel_address(instr, opcode))
154 instr_hi = opcode & 0xf0;
155 instr_lo = opcode & 0x0f;
161 /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
162 scan_more = ((instr_lo & 7) == 0x6);
166 /* 0x64 thru 0x67 are valid prefixes in all modes. */
167 scan_more = (instr_lo & 0xC) == 0x4;
170 /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
171 scan_more = !instr_lo || (instr_lo>>1) == 1;
174 /* Prefetch instruction is 0x0F0D or 0x0F18 */
176 if (instr > (unsigned char *)limit)
178 if (probe_kernel_address(instr, opcode))
180 prefetch = (instr_lo == 0xF) &&
181 (opcode == 0x0D || opcode == 0x18);
191 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
192 unsigned long error_code)
194 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
195 boot_cpu_data.x86 >= 6)) {
196 /* Catch an obscure case of prefetch inside an NX page. */
197 if (nx_enabled && (error_code & 16))
199 return __is_prefetch(regs, addr);
204 static noinline void force_sig_info_fault(int si_signo, int si_code,
205 unsigned long address, struct task_struct *tsk)
209 info.si_signo = si_signo;
211 info.si_code = si_code;
212 info.si_addr = (void __user *)address;
213 force_sig_info(si_signo, &info, tsk);
216 void do_invalid_op(struct pt_regs *, unsigned long);
218 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
220 unsigned index = pgd_index(address);
226 pgd_k = init_mm.pgd + index;
228 if (!pgd_present(*pgd_k))
232 * set_pgd(pgd, *pgd_k); here would be useless on PAE
233 * and redundant with the set_pmd() on non-PAE. As would
237 pud = pud_offset(pgd, address);
238 pud_k = pud_offset(pgd_k, address);
239 if (!pud_present(*pud_k))
242 pmd = pmd_offset(pud, address);
243 pmd_k = pmd_offset(pud_k, address);
244 if (!pmd_present(*pmd_k))
246 if (!pmd_present(*pmd)) {
247 set_pmd(pmd, *pmd_k);
248 arch_flush_lazy_mmu_mode();
250 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
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)
261 unsigned long pgd_paddr;
265 * Synchronize this task's top level page-table
266 * with the 'reference' page table.
268 * Do _not_ use "current" here. We might be inside
269 * an interrupt in the middle of a task switch..
271 pgd_paddr = read_cr3();
272 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
275 pte_k = pte_offset_kernel(pmd_k, address);
276 if (!pte_present(*pte_k))
281 int show_unhandled_signals = 1;
284 * This routine handles page faults. It determines the address,
285 * and the problem, and then passes it off to one of the appropriate
289 * bit 0 == 0 means no page found, 1 means protection fault
290 * bit 1 == 0 means read, 1 means write
291 * bit 2 == 0 means kernel, 1 means user-mode
292 * bit 3 == 1 means use of reserved bit detected
293 * bit 4 == 1 means fault was an instruction fetch
295 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
297 struct task_struct *tsk;
298 struct mm_struct *mm;
299 struct vm_area_struct * vma;
300 unsigned long address;
305 * We can fault from pretty much anywhere, with unknown IRQ state.
307 trace_hardirqs_fixup();
309 /* get the address */
310 address = read_cr2();
314 si_code = SEGV_MAPERR;
317 * We fault-in kernel-space virtual memory on-demand. The
318 * 'reference' page table is init_mm.pgd.
320 * NOTE! We MUST NOT take any locks for this case. We may
321 * be in an interrupt or a critical region, and should
322 * only copy the information from the master page table,
325 * This verifies that the fault happens in kernel space
326 * (error_code & 4) == 0, and that the fault was not a
327 * protection error (error_code & 9) == 0.
329 if (unlikely(address >= TASK_SIZE)) {
330 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
332 if (notify_page_fault(regs))
335 * Don't take the mm semaphore here. If we fixup a prefetch
336 * fault we could otherwise deadlock.
338 goto bad_area_nosemaphore;
341 if (notify_page_fault(regs))
344 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
345 fault has been handled. */
346 if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
352 * If we're in an interrupt, have no user context or are running in an
353 * atomic region then we must not take the fault..
355 if (in_atomic() || !mm)
356 goto bad_area_nosemaphore;
358 /* When running in the kernel we expect faults to occur only to
359 * addresses in user space. All other faults represent errors in the
360 * kernel and should generate an OOPS. Unfortunately, in the case of an
361 * erroneous fault occurring in a code path which already holds mmap_sem
362 * we will deadlock attempting to validate the fault against the
363 * address space. Luckily the kernel only validly references user
364 * space from well defined areas of code, which are listed in the
367 * As the vast majority of faults will be valid we will only perform
368 * the source reference check when there is a possibility of a deadlock.
369 * Attempt to lock the address space, if we cannot we then validate the
370 * source. If this is invalid we can skip the address space check,
371 * thus avoiding the deadlock.
373 if (!down_read_trylock(&mm->mmap_sem)) {
374 if ((error_code & 4) == 0 &&
375 !search_exception_tables(regs->ip))
376 goto bad_area_nosemaphore;
377 down_read(&mm->mmap_sem);
380 vma = find_vma(mm, address);
383 if (vma->vm_start <= address)
385 if (!(vma->vm_flags & VM_GROWSDOWN))
387 if (error_code & 4) {
389 * Accessing the stack below %sp is always a bug.
390 * The large cushion allows instructions like enter
391 * and pusha to work. ("enter $65535,$31" pushes
392 * 32 pointers and then decrements %sp by 65535.)
394 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
397 if (expand_stack(vma, address))
400 * Ok, we have a good vm_area for this memory access, so
404 si_code = SEGV_ACCERR;
406 switch (error_code & 3) {
407 default: /* 3: write, present */
409 case 2: /* write, not present */
410 if (!(vma->vm_flags & VM_WRITE))
414 case 1: /* read, present */
416 case 0: /* read, not present */
417 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
423 * If for any reason at all we couldn't handle the fault,
424 * make sure we exit gracefully rather than endlessly redo
427 fault = handle_mm_fault(mm, vma, address, write);
428 if (unlikely(fault & VM_FAULT_ERROR)) {
429 if (fault & VM_FAULT_OOM)
431 else if (fault & VM_FAULT_SIGBUS)
435 if (fault & VM_FAULT_MAJOR)
441 * Did it hit the DOS screen memory VA from vm86 mode?
443 if (regs->flags & VM_MASK) {
444 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
446 tsk->thread.screen_bitmap |= 1 << bit;
448 up_read(&mm->mmap_sem);
452 * Something tried to access memory that isn't in our memory map..
453 * Fix it, but check if it's kernel or user first..
456 up_read(&mm->mmap_sem);
458 bad_area_nosemaphore:
459 /* User mode accesses just cause a SIGSEGV */
460 if (error_code & 4) {
462 * It's possible to have interrupts off here.
467 * Valid to do another page fault here because this one came
470 if (is_prefetch(regs, address, error_code))
473 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
474 printk_ratelimit()) {
475 printk("%s%s[%d]: segfault at %08lx ip %08lx "
476 "sp %08lx error %lx\n",
477 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
478 tsk->comm, task_pid_nr(tsk), address, regs->ip,
479 regs->sp, error_code);
481 tsk->thread.cr2 = address;
482 /* Kernel addresses are always protection faults */
483 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
484 tsk->thread.trap_no = 14;
485 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
489 #ifdef CONFIG_X86_F00F_BUG
491 * Pentium F0 0F C7 C8 bug workaround.
493 if (boot_cpu_data.f00f_bug) {
496 nr = (address - idt_descr.address) >> 3;
499 do_invalid_op(regs, 0);
506 /* Are we prepared to handle this kernel fault? */
507 if (fixup_exception(regs))
511 * Valid to do another page fault here, because if this fault
512 * had been triggered by is_prefetch fixup_exception would have
515 if (is_prefetch(regs, address, error_code))
519 * Oops. The kernel tried to access some bad page. We'll have to
520 * terminate things with extreme prejudice.
525 if (oops_may_print()) {
526 __typeof__(pte_val(__pte(0))) page;
528 #ifdef CONFIG_X86_PAE
529 if (error_code & 16) {
530 pte_t *pte = lookup_address(address);
532 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
533 printk(KERN_CRIT "kernel tried to execute "
534 "NX-protected page - exploit attempt? "
535 "(uid: %d)\n", current->uid);
538 if (address < PAGE_SIZE)
539 printk(KERN_ALERT "BUG: unable to handle kernel NULL "
540 "pointer dereference");
542 printk(KERN_ALERT "BUG: unable to handle kernel paging"
544 printk(" at virtual address %08lx\n",address);
545 printk(KERN_ALERT "printing ip: %08lx ", regs->ip);
548 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
549 #ifdef CONFIG_X86_PAE
550 printk("*pdpt = %016Lx ", page);
551 if ((page >> PAGE_SHIFT) < max_low_pfn
552 && page & _PAGE_PRESENT) {
554 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
555 & (PTRS_PER_PMD - 1)];
556 printk(KERN_CONT "*pde = %016Lx ", page);
560 printk("*pde = %08lx ", page);
564 * We must not directly access the pte in the highpte
565 * case if the page table is located in highmem.
566 * And let's rather not kmap-atomic the pte, just in case
567 * it's allocated already.
569 if ((page >> PAGE_SHIFT) < max_low_pfn
570 && (page & _PAGE_PRESENT)
571 && !(page & _PAGE_PSE)) {
573 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
574 & (PTRS_PER_PTE - 1)];
575 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
581 tsk->thread.cr2 = address;
582 tsk->thread.trap_no = 14;
583 tsk->thread.error_code = error_code;
584 die("Oops", regs, error_code);
589 * We ran out of memory, or some other thing happened to us that made
590 * us unable to handle the page fault gracefully.
593 up_read(&mm->mmap_sem);
594 if (is_global_init(tsk)) {
596 down_read(&mm->mmap_sem);
599 printk("VM: killing process %s\n", tsk->comm);
601 do_group_exit(SIGKILL);
605 up_read(&mm->mmap_sem);
607 /* Kernel mode? Handle exceptions or die */
608 if (!(error_code & 4))
611 /* User space => ok to do another page fault */
612 if (is_prefetch(regs, address, error_code))
615 tsk->thread.cr2 = address;
616 tsk->thread.error_code = error_code;
617 tsk->thread.trap_no = 14;
618 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
621 void vmalloc_sync_all(void)
624 * Note that races in the updates of insync and start aren't
625 * problematic: insync can only get set bits added, and updates to
626 * start are only improving performance (without affecting correctness
629 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
630 static unsigned long start = TASK_SIZE;
631 unsigned long address;
633 if (SHARED_KERNEL_PMD)
636 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
637 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
638 if (!test_bit(pgd_index(address), insync)) {
642 spin_lock_irqsave(&pgd_lock, flags);
643 for (page = pgd_list; page; page =
644 (struct page *)page->index)
645 if (!vmalloc_sync_one(page_address(page),
647 BUG_ON(page != pgd_list);
650 spin_unlock_irqrestore(&pgd_lock, flags);
652 set_bit(pgd_index(address), insync);
654 if (address == start && test_bit(pgd_index(address), insync))
655 start = address + PGDIR_SIZE;