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 extern void die(const char *, struct pt_regs *, long);
47 static inline int notify_page_fault(struct pt_regs *regs)
52 /* kprobe_running() needs smp_processor_id() */
53 if (!user_mode_vm(regs)) {
55 if (kprobe_running() && kprobe_fault_handler(regs, 14))
67 * Return EIP plus the CS segment base. The segment limit is also
68 * adjusted, clamped to the kernel/user address space (whichever is
69 * appropriate), and returned in *eip_limit.
71 * The segment is checked, because it might have been changed by another
72 * task between the original faulting instruction and here.
74 * If CS is no longer a valid code segment, or if EIP is beyond the
75 * limit, or if it is a kernel address when CS is not a kernel segment,
76 * then the returned value will be greater than *eip_limit.
78 * This is slow, but is very rarely executed.
80 static inline unsigned long get_segment_eip(struct pt_regs *regs,
81 unsigned long *eip_limit)
83 unsigned long ip = regs->ip;
84 unsigned seg = regs->cs & 0xffff;
85 u32 seg_ar, seg_limit, base, *desc;
87 /* Unlikely, but must come before segment checks. */
88 if (unlikely(regs->flags & VM_MASK)) {
90 *eip_limit = base + 0xffff;
91 return base + (ip & 0xffff);
94 /* The standard kernel/user address space limit. */
95 *eip_limit = user_mode(regs) ? USER_DS.seg : KERNEL_DS.seg;
97 /* By far the most common cases. */
98 if (likely(SEGMENT_IS_FLAT_CODE(seg)))
101 /* Check the segment exists, is within the current LDT/GDT size,
102 that kernel/user (ring 0..3) has the appropriate privilege,
103 that it's a code segment, and get the limit. */
104 __asm__ ("larl %3,%0; lsll %3,%1"
105 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
106 if ((~seg_ar & 0x9800) || ip > seg_limit) {
108 return 1; /* So that returned ip > *eip_limit. */
111 /* Get the GDT/LDT descriptor base.
112 When you look for races in this code remember that
113 LDT and other horrors are only used in user space. */
115 /* Must lock the LDT while reading it. */
116 mutex_lock(¤t->mm->context.lock);
117 desc = current->mm->context.ldt;
118 desc = (void *)desc + (seg & ~7);
120 /* Must disable preemption while reading the GDT. */
121 desc = (u32 *)get_cpu_gdt_table(get_cpu());
122 desc = (void *)desc + (seg & ~7);
125 /* Decode the code segment base from the descriptor */
126 base = get_desc_base((struct desc_struct *)desc);
129 mutex_unlock(¤t->mm->context.lock);
133 /* Adjust EIP and segment limit, and clamp at the kernel limit.
134 It's legitimate for segments to wrap at 0xffffffff. */
136 if (seg_limit < *eip_limit && seg_limit >= base)
137 *eip_limit = seg_limit;
142 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
143 * Check that here and ignore it.
145 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
148 unsigned char *instr = (unsigned char *)get_segment_eip(regs, &limit);
153 for (i = 0; scan_more && i < 15; i++) {
154 unsigned char opcode;
155 unsigned char instr_hi;
156 unsigned char instr_lo;
158 if (instr > (unsigned char *)limit)
160 if (probe_kernel_address(instr, opcode))
163 instr_hi = opcode & 0xf0;
164 instr_lo = opcode & 0x0f;
171 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
172 * In X86_64 long mode, the CPU will signal invalid
173 * opcode if some of these prefixes are present so
174 * X86_64 will never get here anyway
176 scan_more = ((instr_lo & 7) == 0x6);
181 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
182 * Need to figure out under what instruction mode the
183 * instruction was issued. Could check the LDT for lm,
184 * but for now it's good enough to assume that long
185 * mode only uses well known segments or kernel.
187 scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
191 /* 0x64 thru 0x67 are valid prefixes in all modes. */
192 scan_more = (instr_lo & 0xC) == 0x4;
195 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
196 scan_more = !instr_lo || (instr_lo>>1) == 1;
199 /* Prefetch instruction is 0x0F0D or 0x0F18 */
201 if (instr > (unsigned char *)limit)
203 if (probe_kernel_address(instr, opcode))
205 prefetch = (instr_lo == 0xF) &&
206 (opcode == 0x0D || opcode == 0x18);
216 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
217 unsigned long error_code)
219 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
220 boot_cpu_data.x86 >= 6)) {
221 /* Catch an obscure case of prefetch inside an NX page. */
222 if (nx_enabled && (error_code & 16))
224 return __is_prefetch(regs, addr);
229 static noinline void force_sig_info_fault(int si_signo, int si_code,
230 unsigned long address, struct task_struct *tsk)
234 info.si_signo = si_signo;
236 info.si_code = si_code;
237 info.si_addr = (void __user *)address;
238 force_sig_info(si_signo, &info, tsk);
241 void do_invalid_op(struct pt_regs *, unsigned long);
243 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
245 unsigned index = pgd_index(address);
251 pgd_k = init_mm.pgd + index;
253 if (!pgd_present(*pgd_k))
257 * set_pgd(pgd, *pgd_k); here would be useless on PAE
258 * and redundant with the set_pmd() on non-PAE. As would
262 pud = pud_offset(pgd, address);
263 pud_k = pud_offset(pgd_k, address);
264 if (!pud_present(*pud_k))
267 pmd = pmd_offset(pud, address);
268 pmd_k = pmd_offset(pud_k, address);
269 if (!pmd_present(*pmd_k))
271 if (!pmd_present(*pmd)) {
272 set_pmd(pmd, *pmd_k);
273 arch_flush_lazy_mmu_mode();
275 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
280 * Handle a fault on the vmalloc or module mapping area
282 * This assumes no large pages in there.
284 static inline int vmalloc_fault(unsigned long address)
286 unsigned long pgd_paddr;
290 * Synchronize this task's top level page-table
291 * with the 'reference' page table.
293 * Do _not_ use "current" here. We might be inside
294 * an interrupt in the middle of a task switch..
296 pgd_paddr = read_cr3();
297 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
300 pte_k = pte_offset_kernel(pmd_k, address);
301 if (!pte_present(*pte_k))
306 int show_unhandled_signals = 1;
309 * This routine handles page faults. It determines the address,
310 * and the problem, and then passes it off to one of the appropriate
313 void __kprobes do_page_fault(struct pt_regs *regs, unsigned long error_code)
315 struct task_struct *tsk;
316 struct mm_struct *mm;
317 struct vm_area_struct *vma;
318 unsigned long address;
323 * We can fault from pretty much anywhere, with unknown IRQ state.
325 trace_hardirqs_fixup();
327 /* get the address */
328 address = read_cr2();
332 si_code = SEGV_MAPERR;
335 * We fault-in kernel-space virtual memory on-demand. The
336 * 'reference' page table is init_mm.pgd.
338 * NOTE! We MUST NOT take any locks for this case. We may
339 * be in an interrupt or a critical region, and should
340 * only copy the information from the master page table,
343 * This verifies that the fault happens in kernel space
344 * (error_code & 4) == 0, and that the fault was not a
345 * protection error (error_code & 9) == 0.
347 if (unlikely(address >= TASK_SIZE)) {
348 if (!(error_code & 0x0000000d) && vmalloc_fault(address) >= 0)
350 if (notify_page_fault(regs))
353 * Don't take the mm semaphore here. If we fixup a prefetch
354 * fault we could otherwise deadlock.
356 goto bad_area_nosemaphore;
359 if (notify_page_fault(regs))
362 /* It's safe to allow irq's after cr2 has been saved and the vmalloc
363 fault has been handled. */
364 if (regs->flags & (X86_EFLAGS_IF|VM_MASK))
370 * If we're in an interrupt, have no user context or are running in an
371 * atomic region then we must not take the fault.
373 if (in_atomic() || !mm)
374 goto bad_area_nosemaphore;
376 /* When running in the kernel we expect faults to occur only to
377 * addresses in user space. All other faults represent errors in the
378 * kernel and should generate an OOPS. Unfortunately, in the case of an
379 * erroneous fault occurring in a code path which already holds mmap_sem
380 * we will deadlock attempting to validate the fault against the
381 * address space. Luckily the kernel only validly references user
382 * space from well defined areas of code, which are listed in the
385 * As the vast majority of faults will be valid we will only perform
386 * the source reference check when there is a possibility of a deadlock.
387 * Attempt to lock the address space, if we cannot we then validate the
388 * source. If this is invalid we can skip the address space check,
389 * thus avoiding the deadlock.
391 if (!down_read_trylock(&mm->mmap_sem)) {
392 if ((error_code & PF_USER) == 0 &&
393 !search_exception_tables(regs->ip))
394 goto bad_area_nosemaphore;
395 down_read(&mm->mmap_sem);
398 vma = find_vma(mm, address);
401 if (vma->vm_start <= address)
403 if (!(vma->vm_flags & VM_GROWSDOWN))
405 if (error_code & PF_USER) {
407 * Accessing the stack below %sp is always a bug.
408 * The large cushion allows instructions like enter
409 * and pusha to work. ("enter $65535,$31" pushes
410 * 32 pointers and then decrements %sp by 65535.)
412 if (address + 65536 + 32 * sizeof(unsigned long) < regs->sp)
415 if (expand_stack(vma, address))
418 * Ok, we have a good vm_area for this memory access, so
422 si_code = SEGV_ACCERR;
424 switch (error_code & (PF_PROT|PF_WRITE)) {
425 default: /* 3: write, present */
427 case PF_WRITE: /* write, not present */
428 if (!(vma->vm_flags & VM_WRITE))
432 case PF_PROT: /* read, present */
434 case 0: /* read, not present */
435 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
441 * If for any reason at all we couldn't handle the fault,
442 * make sure we exit gracefully rather than endlessly redo
445 fault = handle_mm_fault(mm, vma, address, write);
446 if (unlikely(fault & VM_FAULT_ERROR)) {
447 if (fault & VM_FAULT_OOM)
449 else if (fault & VM_FAULT_SIGBUS)
453 if (fault & VM_FAULT_MAJOR)
459 * Did it hit the DOS screen memory VA from vm86 mode?
461 if (regs->flags & VM_MASK) {
462 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
464 tsk->thread.screen_bitmap |= 1 << bit;
466 up_read(&mm->mmap_sem);
470 * Something tried to access memory that isn't in our memory map..
471 * Fix it, but check if it's kernel or user first..
474 up_read(&mm->mmap_sem);
476 bad_area_nosemaphore:
477 /* User mode accesses just cause a SIGSEGV */
478 if (error_code & PF_USER) {
480 * It's possible to have interrupts off here.
485 * Valid to do another page fault here because this one came
488 if (is_prefetch(regs, address, error_code))
491 if (show_unhandled_signals && unhandled_signal(tsk, SIGSEGV) &&
492 printk_ratelimit()) {
493 printk("%s%s[%d]: segfault at %08lx ip %08lx "
494 "sp %08lx error %lx\n",
495 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
496 tsk->comm, task_pid_nr(tsk), address, regs->ip,
497 regs->sp, 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 __typeof__(pte_val(__pte(0))) page;
546 #ifdef CONFIG_X86_PAE
547 if (error_code & 16) {
548 pte_t *pte = lookup_address(address);
550 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
551 printk(KERN_CRIT "kernel tried to execute "
552 "NX-protected page - exploit attempt? "
553 "(uid: %d)\n", current->uid);
556 if (address < PAGE_SIZE)
557 printk(KERN_ALERT "BUG: unable to handle kernel NULL "
558 "pointer dereference");
560 printk(KERN_ALERT "BUG: unable to handle kernel paging"
562 printk(" at virtual address %08lx\n", address);
563 printk(KERN_ALERT "printing ip: %08lx ", regs->ip);
566 page = ((__typeof__(page) *) __va(page))[address >> PGDIR_SHIFT];
567 #ifdef CONFIG_X86_PAE
568 printk("*pdpt = %016Lx ", page);
569 if ((page >> PAGE_SHIFT) < max_low_pfn
570 && page & _PAGE_PRESENT) {
572 page = ((__typeof__(page) *) __va(page))[(address >> PMD_SHIFT)
573 & (PTRS_PER_PMD - 1)];
574 printk(KERN_CONT "*pde = %016Lx ", page);
578 printk("*pde = %08lx ", page);
582 * We must not directly access the pte in the highpte
583 * case if the page table is located in highmem.
584 * And let's rather not kmap-atomic the pte, just in case
585 * it's allocated already.
587 if ((page >> PAGE_SHIFT) < max_low_pfn
588 && (page & _PAGE_PRESENT)
589 && !(page & _PAGE_PSE)) {
591 page = ((__typeof__(page) *) __va(page))[(address >> PAGE_SHIFT)
592 & (PTRS_PER_PTE - 1)];
593 printk("*pte = %0*Lx ", sizeof(page)*2, (u64)page);
599 tsk->thread.cr2 = address;
600 tsk->thread.trap_no = 14;
601 tsk->thread.error_code = error_code;
602 die("Oops", regs, error_code);
607 * We ran out of memory, or some other thing happened to us that made
608 * us unable to handle the page fault gracefully.
611 up_read(&mm->mmap_sem);
612 if (is_global_init(tsk)) {
614 down_read(&mm->mmap_sem);
617 printk("VM: killing process %s\n", tsk->comm);
619 do_group_exit(SIGKILL);
623 up_read(&mm->mmap_sem);
625 /* Kernel mode? Handle exceptions or die */
626 if (!(error_code & PF_USER))
629 /* User space => ok to do another page fault */
630 if (is_prefetch(regs, address, error_code))
633 tsk->thread.cr2 = address;
634 tsk->thread.error_code = error_code;
635 tsk->thread.trap_no = 14;
636 force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
639 void vmalloc_sync_all(void)
642 * Note that races in the updates of insync and start aren't
643 * problematic: insync can only get set bits added, and updates to
644 * start are only improving performance (without affecting correctness
647 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
648 static unsigned long start = TASK_SIZE;
649 unsigned long address;
651 if (SHARED_KERNEL_PMD)
654 BUILD_BUG_ON(TASK_SIZE & ~PGDIR_MASK);
655 for (address = start; address >= TASK_SIZE; address += PGDIR_SIZE) {
656 if (!test_bit(pgd_index(address), insync)) {
660 spin_lock_irqsave(&pgd_lock, flags);
661 for (page = pgd_list; page; page =
662 (struct page *)page->index)
663 if (!vmalloc_sync_one(page_address(page),
665 BUG_ON(page != pgd_list);
668 spin_unlock_irqrestore(&pgd_lock, flags);
670 set_bit(pgd_index(address), insync);
672 if (address == start && test_bit(pgd_index(address), insync))
673 start = address + PGDIR_SIZE;