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>
28 #include <asm/kdebug.h>
30 extern void die(const char *,struct pt_regs *,long);
33 * Unlock any spinlocks which will prevent us from getting the
36 void bust_spinlocks(int yes)
38 int loglevel_save = console_loglevel;
49 * OK, the message is on the console. Now we call printk()
50 * without oops_in_progress set so that printk will give klogd
51 * a poke. Hold onto your hats...
53 console_loglevel = 15; /* NMI oopser may have shut the console up */
55 console_loglevel = loglevel_save;
59 * Return EIP plus the CS segment base. The segment limit is also
60 * adjusted, clamped to the kernel/user address space (whichever is
61 * appropriate), and returned in *eip_limit.
63 * The segment is checked, because it might have been changed by another
64 * task between the original faulting instruction and here.
66 * If CS is no longer a valid code segment, or if EIP is beyond the
67 * limit, or if it is a kernel address when CS is not a kernel segment,
68 * then the returned value will be greater than *eip_limit.
70 * This is slow, but is very rarely executed.
72 static inline unsigned long get_segment_eip(struct pt_regs *regs,
73 unsigned long *eip_limit)
75 unsigned long eip = regs->eip;
76 unsigned seg = regs->xcs & 0xffff;
77 u32 seg_ar, seg_limit, base, *desc;
79 /* The standard kernel/user address space limit. */
80 *eip_limit = (seg & 3) ? USER_DS.seg : KERNEL_DS.seg;
82 /* Unlikely, but must come before segment checks. */
83 if (unlikely((regs->eflags & VM_MASK) != 0))
84 return eip + (seg << 4);
86 /* By far the most common cases. */
87 if (likely(seg == __USER_CS || seg == __KERNEL_CS))
90 /* Check the segment exists, is within the current LDT/GDT size,
91 that kernel/user (ring 0..3) has the appropriate privilege,
92 that it's a code segment, and get the limit. */
93 __asm__ ("larl %3,%0; lsll %3,%1"
94 : "=&r" (seg_ar), "=r" (seg_limit) : "0" (0), "rm" (seg));
95 if ((~seg_ar & 0x9800) || eip > seg_limit) {
97 return 1; /* So that returned eip > *eip_limit. */
100 /* Get the GDT/LDT descriptor base.
101 When you look for races in this code remember that
102 LDT and other horrors are only used in user space. */
104 /* Must lock the LDT while reading it. */
105 down(¤t->mm->context.sem);
106 desc = current->mm->context.ldt;
107 desc = (void *)desc + (seg & ~7);
109 /* Must disable preemption while reading the GDT. */
110 desc = (u32 *)&per_cpu(cpu_gdt_table, get_cpu());
111 desc = (void *)desc + (seg & ~7);
114 /* Decode the code segment base from the descriptor */
115 base = (desc[0] >> 16) |
116 ((desc[1] & 0xff) << 16) |
117 (desc[1] & 0xff000000);
120 up(¤t->mm->context.sem);
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 long instr = 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;
151 if (__get_user(opcode, (unsigned char *) instr))
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 */
178 if (__get_user(opcode, (unsigned char *) instr))
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 fastcall void do_invalid_op(struct pt_regs *, unsigned long);
207 * This routine handles page faults. It determines the address,
208 * and the problem, and then passes it off to one of the appropriate
212 * bit 0 == 0 means no page found, 1 means protection fault
213 * bit 1 == 0 means read, 1 means write
214 * bit 2 == 0 means kernel, 1 means user-mode
216 fastcall void do_page_fault(struct pt_regs *regs, unsigned long error_code)
218 struct task_struct *tsk;
219 struct mm_struct *mm;
220 struct vm_area_struct * vma;
221 unsigned long address;
226 /* get the address */
227 __asm__("movl %%cr2,%0":"=r" (address));
229 if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
230 SIGSEGV) == NOTIFY_STOP)
232 /* It's safe to allow irq's after cr2 has been saved */
233 if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
238 info.si_code = SEGV_MAPERR;
241 * We fault-in kernel-space virtual memory on-demand. The
242 * 'reference' page table is init_mm.pgd.
244 * NOTE! We MUST NOT take any locks for this case. We may
245 * be in an interrupt or a critical region, and should
246 * only copy the information from the master page table,
249 * This verifies that the fault happens in kernel space
250 * (error_code & 4) == 0, and that the fault was not a
251 * protection error (error_code & 1) == 0.
253 if (unlikely(address >= TASK_SIZE)) {
254 if (!(error_code & 5))
257 * Don't take the mm semaphore here. If we fixup a prefetch
258 * fault we could otherwise deadlock.
260 goto bad_area_nosemaphore;
266 * If we're in an interrupt, have no user context or are running in an
267 * atomic region then we must not take the fault..
269 if (in_atomic() || !mm)
270 goto bad_area_nosemaphore;
272 /* When running in the kernel we expect faults to occur only to
273 * addresses in user space. All other faults represent errors in the
274 * kernel and should generate an OOPS. Unfortunatly, in the case of an
275 * erroneous fault occuring in a code path which already holds mmap_sem
276 * we will deadlock attempting to validate the fault against the
277 * address space. Luckily the kernel only validly references user
278 * space from well defined areas of code, which are listed in the
281 * As the vast majority of faults will be valid we will only perform
282 * the source reference check when there is a possibilty of a deadlock.
283 * Attempt to lock the address space, if we cannot we then validate the
284 * source. If this is invalid we can skip the address space check,
285 * thus avoiding the deadlock.
287 if (!down_read_trylock(&mm->mmap_sem)) {
288 if ((error_code & 4) == 0 &&
289 !search_exception_tables(regs->eip))
290 goto bad_area_nosemaphore;
291 down_read(&mm->mmap_sem);
294 vma = find_vma(mm, address);
297 if (vma->vm_start <= address)
299 if (!(vma->vm_flags & VM_GROWSDOWN))
301 if (error_code & 4) {
303 * accessing the stack below %esp is always a bug.
304 * The "+ 32" is there due to some instructions (like
305 * pusha) doing post-decrement on the stack and that
306 * doesn't show up until later..
308 if (address + 32 < regs->esp)
311 if (expand_stack(vma, address))
314 * Ok, we have a good vm_area for this memory access, so
318 info.si_code = SEGV_ACCERR;
320 switch (error_code & 3) {
321 default: /* 3: write, present */
322 #ifdef TEST_VERIFY_AREA
323 if (regs->cs == KERNEL_CS)
324 printk("WP fault at %08lx\n", regs->eip);
327 case 2: /* write, not present */
328 if (!(vma->vm_flags & VM_WRITE))
332 case 1: /* read, present */
334 case 0: /* read, not present */
335 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
341 * If for any reason at all we couldn't handle the fault,
342 * make sure we exit gracefully rather than endlessly redo
345 switch (handle_mm_fault(mm, vma, address, write)) {
352 case VM_FAULT_SIGBUS:
361 * Did it hit the DOS screen memory VA from vm86 mode?
363 if (regs->eflags & VM_MASK) {
364 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
366 tsk->thread.screen_bitmap |= 1 << bit;
368 up_read(&mm->mmap_sem);
372 * Something tried to access memory that isn't in our memory map..
373 * Fix it, but check if it's kernel or user first..
376 up_read(&mm->mmap_sem);
378 bad_area_nosemaphore:
379 /* User mode accesses just cause a SIGSEGV */
380 if (error_code & 4) {
382 * Valid to do another page fault here because this one came
385 if (is_prefetch(regs, address, error_code))
388 tsk->thread.cr2 = address;
389 /* Kernel addresses are always protection faults */
390 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
391 tsk->thread.trap_no = 14;
392 info.si_signo = SIGSEGV;
394 /* info.si_code has been set above */
395 info.si_addr = (void __user *)address;
396 force_sig_info(SIGSEGV, &info, tsk);
400 #ifdef CONFIG_X86_F00F_BUG
402 * Pentium F0 0F C7 C8 bug workaround.
404 if (boot_cpu_data.f00f_bug) {
407 nr = (address - idt_descr.address) >> 3;
410 do_invalid_op(regs, 0);
417 /* Are we prepared to handle this kernel fault? */
418 if (fixup_exception(regs))
422 * Valid to do another page fault here, because if this fault
423 * had been triggered by is_prefetch fixup_exception would have
426 if (is_prefetch(regs, address, error_code))
430 * Oops. The kernel tried to access some bad page. We'll have to
431 * terminate things with extreme prejudice.
436 #ifdef CONFIG_X86_PAE
437 if (error_code & 16) {
438 pte_t *pte = lookup_address(address);
440 if (pte && pte_present(*pte) && !pte_exec_kernel(*pte))
441 printk(KERN_CRIT "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n", current->uid);
444 if (address < PAGE_SIZE)
445 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
447 printk(KERN_ALERT "Unable to handle kernel paging request");
448 printk(" at virtual address %08lx\n",address);
449 printk(KERN_ALERT " printing eip:\n");
450 printk("%08lx\n", regs->eip);
451 asm("movl %%cr3,%0":"=r" (page));
452 page = ((unsigned long *) __va(page))[address >> 22];
453 printk(KERN_ALERT "*pde = %08lx\n", page);
455 * We must not directly access the pte in the highpte
456 * case, the page table might be allocated in highmem.
457 * And lets rather not kmap-atomic the pte, just in case
458 * it's allocated already.
460 #ifndef CONFIG_HIGHPTE
463 address &= 0x003ff000;
464 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
465 printk(KERN_ALERT "*pte = %08lx\n", page);
468 die("Oops", regs, error_code);
473 * We ran out of memory, or some other thing happened to us that made
474 * us unable to handle the page fault gracefully.
477 up_read(&mm->mmap_sem);
480 down_read(&mm->mmap_sem);
483 printk("VM: killing process %s\n", tsk->comm);
489 up_read(&mm->mmap_sem);
491 /* Kernel mode? Handle exceptions or die */
492 if (!(error_code & 4))
495 /* User space => ok to do another page fault */
496 if (is_prefetch(regs, address, error_code))
499 tsk->thread.cr2 = address;
500 tsk->thread.error_code = error_code;
501 tsk->thread.trap_no = 14;
502 info.si_signo = SIGBUS;
504 info.si_code = BUS_ADRERR;
505 info.si_addr = (void __user *)address;
506 force_sig_info(SIGBUS, &info, tsk);
512 * Synchronize this task's top level page-table
513 * with the 'reference' page table.
515 * Do _not_ use "tsk" here. We might be inside
516 * an interrupt in the middle of a task switch..
518 int index = pgd_index(address);
519 unsigned long pgd_paddr;
524 asm("movl %%cr3,%0":"=r" (pgd_paddr));
525 pgd = index + (pgd_t *)__va(pgd_paddr);
526 pgd_k = init_mm.pgd + index;
528 if (!pgd_present(*pgd_k))
532 * set_pgd(pgd, *pgd_k); here would be useless on PAE
533 * and redundant with the set_pmd() on non-PAE.
536 pmd = pmd_offset(pgd, address);
537 pmd_k = pmd_offset(pgd_k, address);
538 if (!pmd_present(*pmd_k))
540 set_pmd(pmd, *pmd_k);
542 pte_k = pte_offset_kernel(pmd_k, address);
543 if (!pte_present(*pte_k))