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/smp_lock.h>
18 #include <linux/interrupt.h>
19 #include <linux/init.h>
20 #include <linux/tty.h>
21 #include <linux/vt_kern.h> /* For unblank_screen() */
22 #include <linux/highmem.h>
23 #include <linux/module.h>
25 #include <asm/system.h>
26 #include <asm/uaccess.h>
27 #include <asm/hardirq.h>
29 #include <asm/tlbflush.h>
31 extern void die(const char *,struct pt_regs *,long);
34 * Unlock any spinlocks which will prevent us from getting the
37 void bust_spinlocks(int yes)
39 int loglevel_save = console_loglevel;
50 * OK, the message is on the console. Now we call printk()
51 * without oops_in_progress set so that printk will give klogd
52 * a poke. Hold onto your hats...
54 console_loglevel = 15; /* NMI oopser may have shut the console up */
56 console_loglevel = loglevel_save;
60 * Return EIP plus the CS segment base. The segment limit is also
61 * adjusted, clamped to the kernel/user address space (whichever is
62 * appropriate), and returned in *eip_limit.
64 * The segment is checked, because it might have been changed by another
65 * task between the original faulting instruction and here.
67 * If CS is no longer a valid code segment, or if EIP is beyond the
68 * limit, or if it is a kernel address when CS is not a kernel segment,
69 * then the returned value will be greater than *eip_limit.
71 * This is slow, but is very rarely executed.
73 static inline unsigned long get_segment_eip(struct pt_regs *regs,
74 unsigned long *eip_limit)
76 unsigned long eip = regs->eip;
77 unsigned seg = regs->xcs & 0xffff;
78 u32 seg_ar, seg_limit, base, *desc;
80 /* The standard kernel/user address space limit. */
81 *eip_limit = (seg & 3) ? USER_DS.seg : KERNEL_DS.seg;
83 /* Unlikely, but must come before segment checks. */
84 if (unlikely((regs->eflags & VM_MASK) != 0))
85 return eip + (seg << 4);
87 /* By far the most common cases. */
88 if (likely(seg == __USER_CS || seg == __KERNEL_CS))
91 /* Check the segment exists, is within the current LDT/GDT size,
92 that kernel/user (ring 0..3) has the appropriate privilege,
93 that it's a code segment, and get the limit. */
94 __asm__ ("larl %3,%0; lsll %3,%1"
95 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
96 if ((~seg_ar & 0x9800) || eip > seg_limit) {
98 return 1; /* So that returned eip > *eip_limit. */
101 /* Get the GDT/LDT descriptor base.
102 When you look for races in this code remember that
103 LDT and other horrors are only used in user space. */
105 /* Must lock the LDT while reading it. */
106 down(¤t->mm->context.sem);
108 /* horrible hack for 4/4 disabled kernels.
109 I'm not quite sure what the TLB flush is good for,
110 it's mindlessly copied from the read_ldt code */
111 __flush_tlb_global();
112 desc = kmap(current->mm->context.ldt_pages[(seg&~7)/PAGE_SIZE]);
113 desc = (void *)desc + ((seg & ~7) % PAGE_SIZE);
115 desc = current->mm->context.ldt;
116 desc = (void *)desc + (seg & ~7);
119 /* Must disable preemption while reading the GDT. */
120 desc = (u32 *)&cpu_gdt_table[get_cpu()];
121 desc = (void *)desc + (seg & ~7);
124 /* Decode the code segment base from the descriptor */
125 base = (desc[0] >> 16) |
126 ((desc[1] & 0xff) << 16) |
127 (desc[1] & 0xff000000);
131 kunmap((void *)((unsigned long)desc & PAGE_MASK));
133 up(¤t->mm->context.sem);
137 /* Adjust EIP and segment limit, and clamp at the kernel limit.
138 It's legitimate for segments to wrap at 0xffffffff. */
140 if (seg_limit < *eip_limit && seg_limit >= base)
141 *eip_limit = seg_limit;
146 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
147 * Check that here and ignore it.
149 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
152 unsigned long instr = get_segment_eip (regs, &limit);
157 for (i = 0; scan_more && i < 15; i++) {
158 unsigned char opcode;
159 unsigned char instr_hi;
160 unsigned char instr_lo;
164 if (__get_user(opcode, (unsigned char *) instr))
167 instr_hi = opcode & 0xf0;
168 instr_lo = opcode & 0x0f;
174 /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
175 scan_more = ((instr_lo & 7) == 0x6);
179 /* 0x64 thru 0x67 are valid prefixes in all modes. */
180 scan_more = (instr_lo & 0xC) == 0x4;
183 /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
184 scan_more = !instr_lo || (instr_lo>>1) == 1;
187 /* Prefetch instruction is 0x0F0D or 0x0F18 */
191 if (__get_user(opcode, (unsigned char *) instr))
193 prefetch = (instr_lo == 0xF) &&
194 (opcode == 0x0D || opcode == 0x18);
204 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr,
205 unsigned long error_code)
207 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
208 boot_cpu_data.x86 >= 6)) {
209 /* Catch an obscure case of prefetch inside an NX page. */
210 if (nx_enabled && (error_code & 16))
212 return __is_prefetch(regs, addr);
217 asmlinkage void do_invalid_op(struct pt_regs *, unsigned long);
220 * This routine handles page faults. It determines the address,
221 * and the problem, and then passes it off to one of the appropriate
225 * bit 0 == 0 means no page found, 1 means protection fault
226 * bit 1 == 0 means read, 1 means write
227 * bit 2 == 0 means kernel, 1 means user-mode
229 asmlinkage void do_page_fault(struct pt_regs *regs, unsigned long error_code)
231 struct task_struct *tsk;
232 struct mm_struct *mm;
233 struct vm_area_struct * vma;
234 unsigned long address;
239 /* get the address */
240 __asm__("movl %%cr2,%0":"=r" (address));
242 /* It's safe to allow irq's after cr2 has been saved */
243 if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
248 info.si_code = SEGV_MAPERR;
251 * We fault-in kernel-space virtual memory on-demand. The
252 * 'reference' page table is init_mm.pgd.
254 * NOTE! We MUST NOT take any locks for this case. We may
255 * be in an interrupt or a critical region, and should
256 * only copy the information from the master page table,
259 * This verifies that the fault happens in kernel space
260 * (error_code & 4) == 0, and that the fault was not a
261 * protection error (error_code & 1) == 0.
265 * On 4/4 all kernels faults are either bugs, vmalloc or prefetch
267 /* If it's vm86 fall through */
268 if (unlikely(!(regs->eflags & VM_MASK) && ((regs->xcs & 3) == 0))) {
270 goto bad_area_nosemaphore;
274 if (unlikely(address >= TASK_SIZE)) {
275 if (!(error_code & 5))
278 * Don't take the mm semaphore here. If we fixup a prefetch
279 * fault we could otherwise deadlock.
281 goto bad_area_nosemaphore;
288 * If we're in an interrupt, have no user context or are running in an
289 * atomic region then we must not take the fault..
291 if (in_atomic() || !mm)
292 goto bad_area_nosemaphore;
294 /* When running in the kernel we expect faults to occur only to
295 * addresses in user space. All other faults represent errors in the
296 * kernel and should generate an OOPS. Unfortunatly, in the case of an
297 * erroneous fault occuring in a code path which already holds mmap_sem
298 * we will deadlock attempting to validate the fault against the
299 * address space. Luckily the kernel only validly references user
300 * space from well defined areas of code, which are listed in the
303 * As the vast majority of faults will be valid we will only perform
304 * the source reference check when there is a possibilty of a deadlock.
305 * Attempt to lock the address space, if we cannot we then validate the
306 * source. If this is invalid we can skip the address space check,
307 * thus avoiding the deadlock.
309 if (!down_read_trylock(&mm->mmap_sem)) {
310 if ((error_code & 4) == 0 &&
311 !search_exception_tables(regs->eip))
312 goto bad_area_nosemaphore;
313 down_read(&mm->mmap_sem);
316 vma = find_vma(mm, address);
319 if (vma->vm_start <= address)
321 if (!(vma->vm_flags & VM_GROWSDOWN))
323 if (error_code & 4) {
325 * accessing the stack below %esp is always a bug.
326 * The "+ 32" is there due to some instructions (like
327 * pusha) doing post-decrement on the stack and that
328 * doesn't show up until later..
330 if (address + 32 < regs->esp)
333 if (expand_stack(vma, address))
336 * Ok, we have a good vm_area for this memory access, so
340 info.si_code = SEGV_ACCERR;
342 switch (error_code & 3) {
343 default: /* 3: write, present */
344 #ifdef TEST_VERIFY_AREA
345 if (regs->cs == KERNEL_CS)
346 printk("WP fault at %08lx\n", regs->eip);
349 case 2: /* write, not present */
350 if (!(vma->vm_flags & VM_WRITE))
354 case 1: /* read, present */
356 case 0: /* read, not present */
357 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
363 * If for any reason at all we couldn't handle the fault,
364 * make sure we exit gracefully rather than endlessly redo
367 switch (handle_mm_fault(mm, vma, address, write)) {
374 case VM_FAULT_SIGBUS:
383 * Did it hit the DOS screen memory VA from vm86 mode?
385 if (regs->eflags & VM_MASK) {
386 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
388 tsk->thread.screen_bitmap |= 1 << bit;
390 up_read(&mm->mmap_sem);
394 * Something tried to access memory that isn't in our memory map..
395 * Fix it, but check if it's kernel or user first..
398 up_read(&mm->mmap_sem);
400 bad_area_nosemaphore:
401 /* User mode accesses just cause a SIGSEGV */
402 if (error_code & 4) {
404 * Valid to do another page fault here because this one came
407 if (is_prefetch(regs, address, error_code))
410 tsk->thread.cr2 = address;
411 /* Kernel addresses are always protection faults */
412 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
413 tsk->thread.trap_no = 14;
414 info.si_signo = SIGSEGV;
416 /* info.si_code has been set above */
417 info.si_addr = (void __user *)address;
418 force_sig_info(SIGSEGV, &info, tsk);
422 #ifdef CONFIG_X86_F00F_BUG
424 * Pentium F0 0F C7 C8 bug workaround.
426 if (boot_cpu_data.f00f_bug) {
429 nr = (address - idt_descr.address) >> 3;
432 do_invalid_op(regs, 0);
439 /* Are we prepared to handle this kernel fault? */
440 if (fixup_exception(regs))
444 * Valid to do another page fault here, because if this fault
445 * had been triggered by is_prefetch fixup_exception would have
448 if (is_prefetch(regs, address, error_code))
452 * Oops. The kernel tried to access some bad page. We'll have to
453 * terminate things with extreme prejudice.
458 #ifdef CONFIG_X86_PAE
459 if (error_code & 16) {
460 pte_t *pte = lookup_address(address);
462 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
463 printk(KERN_CRIT "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n", current->uid);
466 if (address < PAGE_SIZE)
467 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
469 printk(KERN_ALERT "Unable to handle kernel paging request");
470 printk(" at virtual address %08lx\n",address);
471 printk(KERN_ALERT " printing eip:\n");
472 printk("%08lx\n", regs->eip);
473 asm("movl %%cr3,%0":"=r" (page));
474 page = ((unsigned long *) __va(page))[address >> 22];
475 printk(KERN_ALERT "*pde = %08lx\n", page);
477 * We must not directly access the pte in the highpte
478 * case, the page table might be allocated in highmem.
479 * And lets rather not kmap-atomic the pte, just in case
480 * it's allocated already.
482 #ifndef CONFIG_HIGHPTE
485 address &= 0x003ff000;
486 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
487 printk(KERN_ALERT "*pte = %08lx\n", page);
490 die("Oops", regs, error_code);
495 * We ran out of memory, or some other thing happened to us that made
496 * us unable to handle the page fault gracefully.
499 up_read(&mm->mmap_sem);
502 down_read(&mm->mmap_sem);
505 printk("VM: killing process %s\n", tsk->comm);
511 up_read(&mm->mmap_sem);
513 /* Kernel mode? Handle exceptions or die */
514 if (!(error_code & 4))
517 /* User space => ok to do another page fault */
518 if (is_prefetch(regs, address, error_code))
521 tsk->thread.cr2 = address;
522 tsk->thread.error_code = error_code;
523 tsk->thread.trap_no = 14;
524 info.si_signo = SIGBUS;
526 info.si_code = BUS_ADRERR;
527 info.si_addr = (void __user *)address;
528 force_sig_info(SIGBUS, &info, tsk);
534 * Synchronize this task's top level page-table
535 * with the 'reference' page table.
537 * Do _not_ use "tsk" here. We might be inside
538 * an interrupt in the middle of a task switch..
540 int index = pgd_index(address);
545 asm("movl %%cr3,%0":"=r" (pgd));
546 pgd = index + (pgd_t *)__va(pgd);
547 pgd_k = init_mm.pgd + index;
549 if (!pgd_present(*pgd_k))
553 * set_pgd(pgd, *pgd_k); here would be useless on PAE
554 * and redundant with the set_pmd() on non-PAE.
557 pmd = pmd_offset(pgd, address);
558 pmd_k = pmd_offset(pgd_k, address);
559 if (!pmd_present(*pmd_k))
561 set_pmd(pmd, *pmd_k);
563 pte_k = pte_offset_kernel(pmd_k, address);
564 if (!pte_present(*pte_k))