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/pgalloc.h>
28 #include <asm/hardirq.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);
107 desc = current->mm->context.ldt;
108 desc = (void *)desc + (seg & ~7);
110 /* Must disable preemption while reading the GDT. */
111 desc = (u32 *)&cpu_gdt_table[get_cpu()];
112 desc = (void *)desc + (seg & ~7);
115 /* Decode the code segment base from the descriptor */
116 base = (desc[0] >> 16) |
117 ((desc[1] & 0xff) << 16) |
118 (desc[1] & 0xff000000);
121 up(¤t->mm->context.sem);
125 /* Adjust EIP and segment limit, and clamp at the kernel limit.
126 It's legitimate for segments to wrap at 0xffffffff. */
128 if (seg_limit < *eip_limit && seg_limit >= base)
129 *eip_limit = seg_limit;
134 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
135 * Check that here and ignore it.
137 static int __is_prefetch(struct pt_regs *regs, unsigned long addr)
140 unsigned long instr = get_segment_eip (regs, &limit);
145 for (i = 0; scan_more && i < 15; i++) {
146 unsigned char opcode;
147 unsigned char instr_hi;
148 unsigned char instr_lo;
152 if (__get_user(opcode, (unsigned char *) instr))
155 instr_hi = opcode & 0xf0;
156 instr_lo = opcode & 0x0f;
162 /* Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes. */
163 scan_more = ((instr_lo & 7) == 0x6);
167 /* 0x64 thru 0x67 are valid prefixes in all modes. */
168 scan_more = (instr_lo & 0xC) == 0x4;
171 /* 0xF0, 0xF2, and 0xF3 are valid prefixes */
172 scan_more = !instr_lo || (instr_lo>>1) == 1;
175 /* Prefetch instruction is 0x0F0D or 0x0F18 */
179 if (__get_user(opcode, (unsigned char *) instr))
181 prefetch = (instr_lo == 0xF) &&
182 (opcode == 0x0D || opcode == 0x18);
192 static inline int is_prefetch(struct pt_regs *regs, unsigned long addr)
194 if (unlikely(boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
195 boot_cpu_data.x86 >= 6))
196 return __is_prefetch(regs, addr);
200 asmlinkage void do_invalid_op(struct pt_regs *, unsigned long);
203 * This routine handles page faults. It determines the address,
204 * and the problem, and then passes it off to one of the appropriate
208 * bit 0 == 0 means no page found, 1 means protection fault
209 * bit 1 == 0 means read, 1 means write
210 * bit 2 == 0 means kernel, 1 means user-mode
212 asmlinkage void do_page_fault(struct pt_regs *regs, unsigned long error_code)
214 struct task_struct *tsk;
215 struct mm_struct *mm;
216 struct vm_area_struct * vma;
217 unsigned long address;
222 /* get the address */
223 __asm__("movl %%cr2,%0":"=r" (address));
225 /* It's safe to allow irq's after cr2 has been saved */
226 if (regs->eflags & (X86_EFLAGS_IF|VM_MASK))
231 info.si_code = SEGV_MAPERR;
234 * We fault-in kernel-space virtual memory on-demand. The
235 * 'reference' page table is init_mm.pgd.
237 * NOTE! We MUST NOT take any locks for this case. We may
238 * be in an interrupt or a critical region, and should
239 * only copy the information from the master page table,
242 * This verifies that the fault happens in kernel space
243 * (error_code & 4) == 0, and that the fault was not a
244 * protection error (error_code & 1) == 0.
246 if (unlikely(address >= TASK_SIZE)) {
247 if (!(error_code & 5))
250 * Don't take the mm semaphore here. If we fixup a prefetch
251 * fault we could otherwise deadlock.
253 goto bad_area_nosemaphore;
259 * If we're in an interrupt, have no user context or are running in an
260 * atomic region then we must not take the fault..
262 if (in_atomic() || !mm)
263 goto bad_area_nosemaphore;
265 down_read(&mm->mmap_sem);
267 vma = find_vma(mm, address);
270 if (vma->vm_start <= address)
272 if (!(vma->vm_flags & VM_GROWSDOWN))
274 if (error_code & 4) {
276 * accessing the stack below %esp is always a bug.
277 * The "+ 32" is there due to some instructions (like
278 * pusha) doing post-decrement on the stack and that
279 * doesn't show up until later..
281 if (address + 32 < regs->esp)
284 if (expand_stack(vma, address))
287 * Ok, we have a good vm_area for this memory access, so
291 info.si_code = SEGV_ACCERR;
293 switch (error_code & 3) {
294 default: /* 3: write, present */
295 #ifdef TEST_VERIFY_AREA
296 if (regs->cs == KERNEL_CS)
297 printk("WP fault at %08lx\n", regs->eip);
300 case 2: /* write, not present */
301 if (!(vma->vm_flags & VM_WRITE))
305 case 1: /* read, present */
307 case 0: /* read, not present */
308 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
314 * If for any reason at all we couldn't handle the fault,
315 * make sure we exit gracefully rather than endlessly redo
318 switch (handle_mm_fault(mm, vma, address, write)) {
325 case VM_FAULT_SIGBUS:
334 * Did it hit the DOS screen memory VA from vm86 mode?
336 if (regs->eflags & VM_MASK) {
337 unsigned long bit = (address - 0xA0000) >> PAGE_SHIFT;
339 tsk->thread.screen_bitmap |= 1 << bit;
341 up_read(&mm->mmap_sem);
345 * Something tried to access memory that isn't in our memory map..
346 * Fix it, but check if it's kernel or user first..
349 up_read(&mm->mmap_sem);
351 bad_area_nosemaphore:
352 /* User mode accesses just cause a SIGSEGV */
353 if (error_code & 4) {
355 * Valid to do another page fault here because this one came
358 if (is_prefetch(regs, address))
361 tsk->thread.cr2 = address;
362 /* Kernel addresses are always protection faults */
363 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
364 tsk->thread.trap_no = 14;
365 info.si_signo = SIGSEGV;
367 /* info.si_code has been set above */
368 info.si_addr = (void *)address;
369 force_sig_info(SIGSEGV, &info, tsk);
373 #ifdef CONFIG_X86_F00F_BUG
375 * Pentium F0 0F C7 C8 bug workaround.
377 if (boot_cpu_data.f00f_bug) {
380 nr = (address - idt_descr.address) >> 3;
383 do_invalid_op(regs, 0);
390 /* Are we prepared to handle this kernel fault? */
391 if (fixup_exception(regs))
395 * Valid to do another page fault here, because if this fault
396 * had been triggered by is_prefetch fixup_exception would have
399 if (is_prefetch(regs, address))
403 * Oops. The kernel tried to access some bad page. We'll have to
404 * terminate things with extreme prejudice.
409 if (address < PAGE_SIZE)
410 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
412 printk(KERN_ALERT "Unable to handle kernel paging request");
413 printk(" at virtual address %08lx\n",address);
414 printk(KERN_ALERT " printing eip:\n");
415 printk("%08lx\n", regs->eip);
416 asm("movl %%cr3,%0":"=r" (page));
417 page = ((unsigned long *) __va(page))[address >> 22];
418 printk(KERN_ALERT "*pde = %08lx\n", page);
420 * We must not directly access the pte in the highpte
421 * case, the page table might be allocated in highmem.
422 * And lets rather not kmap-atomic the pte, just in case
423 * it's allocated already.
425 #ifndef CONFIG_HIGHPTE
428 address &= 0x003ff000;
429 page = ((unsigned long *) __va(page))[address >> PAGE_SHIFT];
430 printk(KERN_ALERT "*pte = %08lx\n", page);
433 die("Oops", regs, error_code);
438 * We ran out of memory, or some other thing happened to us that made
439 * us unable to handle the page fault gracefully.
442 up_read(&mm->mmap_sem);
445 down_read(&mm->mmap_sem);
448 printk("VM: killing process %s\n", tsk->comm);
454 up_read(&mm->mmap_sem);
456 /* Kernel mode? Handle exceptions or die */
457 if (!(error_code & 4))
460 /* User space => ok to do another page fault */
461 if (is_prefetch(regs, address))
464 tsk->thread.cr2 = address;
465 tsk->thread.error_code = error_code;
466 tsk->thread.trap_no = 14;
467 info.si_signo = SIGBUS;
469 info.si_code = BUS_ADRERR;
470 info.si_addr = (void *)address;
471 force_sig_info(SIGBUS, &info, tsk);
477 * Synchronize this task's top level page-table
478 * with the 'reference' page table.
480 * Do _not_ use "tsk" here. We might be inside
481 * an interrupt in the middle of a task switch..
483 int index = pgd_index(address);
488 asm("movl %%cr3,%0":"=r" (pgd));
489 pgd = index + (pgd_t *)__va(pgd);
490 pgd_k = init_mm.pgd + index;
492 if (!pgd_present(*pgd_k))
496 * set_pgd(pgd, *pgd_k); here would be useless on PAE
497 * and redundant with the set_pmd() on non-PAE.
500 pmd = pmd_offset(pgd, address);
501 pmd_k = pmd_offset(pgd_k, address);
502 if (!pmd_present(*pmd_k))
504 set_pmd(pmd, *pmd_k);
506 pte_k = pte_offset_kernel(pmd_k, address);
507 if (!pte_present(*pte_k))