2 * linux/arch/x86-64/mm/fault.c
4 * Copyright (C) 1995 Linus Torvalds
5 * Copyright (C) 2001,2002 Andi Kleen, SuSE Labs.
8 #include <linux/config.h>
9 #include <linux/signal.h>
10 #include <linux/sched.h>
11 #include <linux/kernel.h>
12 #include <linux/errno.h>
13 #include <linux/string.h>
14 #include <linux/types.h>
15 #include <linux/ptrace.h>
16 #include <linux/mman.h>
18 #include <linux/smp.h>
19 #include <linux/smp_lock.h>
20 #include <linux/interrupt.h>
21 #include <linux/init.h>
22 #include <linux/tty.h>
23 #include <linux/vt_kern.h> /* For unblank_screen() */
24 #include <linux/compiler.h>
25 #include <linux/module.h>
26 #include <linux/kprobes.h>
28 #include <asm/system.h>
29 #include <asm/uaccess.h>
30 #include <asm/pgalloc.h>
32 #include <asm/tlbflush.h>
33 #include <asm/proto.h>
34 #include <asm/kdebug.h>
35 #include <asm-generic/sections.h>
36 #include <asm/kdebug.h>
38 void bust_spinlocks(int yes)
40 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 /* Sometimes the CPU reports invalid exceptions on prefetch.
60 Check that here and ignore.
61 Opcode checker based on code by Richard Brunner */
62 static noinline int is_prefetch(struct pt_regs *regs, unsigned long addr,
63 unsigned long error_code)
65 unsigned char *instr = (unsigned char *)(regs->rip);
68 unsigned char *max_instr = instr + 15;
70 /* If it was a exec fault ignore */
71 if (error_code & (1<<4))
74 /* Code segments in LDT could have a non zero base. Don't check
75 when that's possible */
76 if (regs->cs & (1<<2))
79 if ((regs->cs & 3) != 0 && regs->rip >= TASK_SIZE)
82 while (scan_more && instr < max_instr) {
84 unsigned char instr_hi;
85 unsigned char instr_lo;
87 if (__get_user(opcode, instr))
90 instr_hi = opcode & 0xf0;
91 instr_lo = opcode & 0x0f;
97 /* Values 0x26,0x2E,0x36,0x3E are valid x86
98 prefixes. In long mode, the CPU will signal
99 invalid opcode if some of these prefixes are
100 present so we will never get here anyway */
101 scan_more = ((instr_lo & 7) == 0x6);
105 /* In AMD64 long mode, 0x40 to 0x4F are valid REX prefixes
106 Need to figure out under what instruction mode the
107 instruction was issued ... */
108 /* Could check the LDT for lm, but for now it's good
109 enough to assume that long mode only uses well known
110 segments or kernel. */
111 scan_more = ((regs->cs & 3) == 0) || (regs->cs == __USER_CS);
115 /* 0x64 thru 0x67 are valid prefixes in all modes. */
116 scan_more = (instr_lo & 0xC) == 0x4;
119 /* 0xF0, 0xF2, and 0xF3 are valid prefixes in all modes. */
120 scan_more = !instr_lo || (instr_lo>>1) == 1;
123 /* Prefetch instruction is 0x0F0D or 0x0F18 */
125 if (__get_user(opcode, instr))
127 prefetch = (instr_lo == 0xF) &&
128 (opcode == 0x0D || opcode == 0x18);
138 static int bad_address(void *p)
141 return __get_user(dummy, (unsigned long *)p);
144 void dump_pagetable(unsigned long address)
151 asm("movq %%cr3,%0" : "=r" (pml4));
153 pml4 = __va((unsigned long)pml4 & PHYSICAL_PAGE_MASK);
154 pml4 += pml4_index(address);
155 printk("PML4 %lx ", pml4_val(*pml4));
156 if (bad_address(pml4)) goto bad;
157 if (!pml4_present(*pml4)) goto ret;
159 pgd = __pgd_offset_k((pgd_t *)pml4_page(*pml4), address);
160 if (bad_address(pgd)) goto bad;
161 printk("PGD %lx ", pgd_val(*pgd));
162 if (!pgd_present(*pgd)) goto ret;
164 pmd = pmd_offset(pgd, address);
165 if (bad_address(pmd)) goto bad;
166 printk("PMD %lx ", pmd_val(*pmd));
167 if (!pmd_present(*pmd)) goto ret;
169 pte = pte_offset_kernel(pmd, address);
170 if (bad_address(pte)) goto bad;
171 printk("PTE %lx", pte_val(*pte));
179 static const char errata93_warning[] =
180 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
181 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
182 KERN_ERR "******* Please consider a BIOS update.\n"
183 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
185 /* Workaround for K8 erratum #93 & buggy BIOS.
186 BIOS SMM functions are required to use a specific workaround
187 to avoid corruption of the 64bit RIP register on C stepping K8.
188 A lot of BIOS that didn't get tested properly miss this.
189 The OS sees this as a page fault with the upper 32bits of RIP cleared.
190 Try to work around it here.
191 Note we only handle faults in kernel here. */
193 static int is_errata93(struct pt_regs *regs, unsigned long address)
196 if (address != regs->rip)
198 if ((address >> 32) != 0)
200 address |= 0xffffffffUL << 32;
201 if ((address >= (u64)_stext && address <= (u64)_etext) ||
202 (address >= MODULES_VADDR && address <= MODULES_END)) {
204 printk(errata93_warning);
213 int unhandled_signal(struct task_struct *tsk, int sig)
215 /* Warn for strace, but not for gdb */
216 if (!test_ti_thread_flag(tsk->thread_info, TIF_SYSCALL_TRACE) &&
217 (tsk->ptrace & PT_PTRACED))
219 return (tsk->sighand->action[sig-1].sa.sa_handler == SIG_IGN) ||
220 (tsk->sighand->action[sig-1].sa.sa_handler == SIG_DFL);
223 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
224 unsigned long error_code)
227 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
228 current->comm, address);
229 dump_pagetable(address);
230 __die("Bad pagetable", regs, error_code);
235 int page_fault_trace;
236 int exception_trace = 1;
239 * This routine handles page faults. It determines the address,
240 * and the problem, and then passes it off to one of the appropriate
244 * bit 0 == 0 means no page found, 1 means protection fault
245 * bit 1 == 0 means read, 1 means write
246 * bit 2 == 0 means kernel, 1 means user-mode
247 * bit 3 == 1 means fault was an instruction fetch
249 asmlinkage void do_page_fault(struct pt_regs *regs, unsigned long error_code)
251 struct task_struct *tsk;
252 struct mm_struct *mm;
253 struct vm_area_struct * vma;
254 unsigned long address;
255 const struct exception_table_entry *fixup;
259 #ifdef CONFIG_CHECKING
262 struct x8664_pda *pda = cpu_pda + stack_smp_processor_id();
263 rdmsrl(MSR_GS_BASE, gs);
264 if (gs != (unsigned long)pda) {
265 wrmsrl(MSR_GS_BASE, pda);
266 printk("page_fault: wrong gs %lx expected %p\n", gs, pda);
271 /* get the address */
272 __asm__("movq %%cr2,%0":"=r" (address));
273 if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
274 SIGSEGV) == NOTIFY_STOP)
277 if (likely(regs->eflags & X86_EFLAGS_IF))
280 if (unlikely(page_fault_trace))
281 printk("pagefault rip:%lx rsp:%lx cs:%lu ss:%lu address %lx error %lx\n",
282 regs->rip,regs->rsp,regs->cs,regs->ss,address,error_code);
286 info.si_code = SEGV_MAPERR;
290 * We fault-in kernel-space virtual memory on-demand. The
291 * 'reference' page table is init_mm.pgd.
293 * NOTE! We MUST NOT take any locks for this case. We may
294 * be in an interrupt or a critical region, and should
295 * only copy the information from the master page table,
298 * This verifies that the fault happens in kernel space
299 * (error_code & 4) == 0, and that the fault was not a
300 * protection error (error_code & 1) == 0.
302 if (unlikely(address >= TASK_SIZE)) {
303 if (!(error_code & 5))
306 * Don't take the mm semaphore here. If we fixup a prefetch
307 * fault we could otherwise deadlock.
309 goto bad_area_nosemaphore;
312 if (unlikely(error_code & (1 << 3)))
313 goto page_table_corruption;
316 * If we're in an interrupt or have no user
317 * context, we must not take the fault..
319 if (unlikely(in_atomic() || !mm))
320 goto bad_area_nosemaphore;
323 /* When running in the kernel we expect faults to occur only to
324 * addresses in user space. All other faults represent errors in the
325 * kernel and should generate an OOPS. Unfortunatly, in the case of an
326 * erroneous fault occuring in a code path which already holds mmap_sem
327 * we will deadlock attempting to validate the fault against the
328 * address space. Luckily the kernel only validly references user
329 * space from well defined areas of code, which are listed in the
332 * As the vast majority of faults will be valid we will only perform
333 * the source reference check when there is a possibilty of a deadlock.
334 * Attempt to lock the address space, if we cannot we then validate the
335 * source. If this is invalid we can skip the address space check,
336 * thus avoiding the deadlock.
338 if (!down_read_trylock(&mm->mmap_sem)) {
339 if ((error_code & 4) == 0 &&
340 !search_exception_tables(regs->rip))
341 goto bad_area_nosemaphore;
342 down_read(&mm->mmap_sem);
345 vma = find_vma(mm, address);
348 if (likely(vma->vm_start <= address))
350 if (!(vma->vm_flags & VM_GROWSDOWN))
352 if (error_code & 4) {
353 // XXX: align red zone size with ABI
354 if (address + 128 < regs->rsp)
357 if (expand_stack(vma, address))
360 * Ok, we have a good vm_area for this memory access, so
364 info.si_code = SEGV_ACCERR;
366 switch (error_code & 3) {
367 default: /* 3: write, present */
369 case 2: /* write, not present */
370 if (!(vma->vm_flags & VM_WRITE))
374 case 1: /* read, present */
376 case 0: /* read, not present */
377 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
382 * If for any reason at all we couldn't handle the fault,
383 * make sure we exit gracefully rather than endlessly redo
386 switch (handle_mm_fault(mm, vma, address, write)) {
399 up_read(&mm->mmap_sem);
403 * Something tried to access memory that isn't in our memory map..
404 * Fix it, but check if it's kernel or user first..
407 up_read(&mm->mmap_sem);
409 bad_area_nosemaphore:
411 #ifdef CONFIG_IA32_EMULATION
412 /* 32bit vsyscall. map on demand. */
413 if (test_thread_flag(TIF_IA32) &&
414 address >= VSYSCALL32_BASE && address < VSYSCALL32_END) {
415 if (map_syscall32(mm, address) < 0)
421 /* User mode accesses just cause a SIGSEGV */
422 if (error_code & 4) {
423 if (is_prefetch(regs, address, error_code))
426 /* Work around K8 erratum #100 K8 in compat mode
427 occasionally jumps to illegal addresses >4GB. We
428 catch this here in the page fault handler because
429 these addresses are not reachable. Just detect this
430 case and return. Any code segment in LDT is
431 compatibility mode. */
432 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
436 if (exception_trace && unhandled_signal(tsk, SIGSEGV)) {
438 "%s[%d]: segfault at %016lx rip %016lx rsp %016lx error %lx\n",
439 tsk->comm, tsk->pid, address, regs->rip,
440 regs->rsp, error_code);
443 tsk->thread.cr2 = address;
444 /* Kernel addresses are always protection faults */
445 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
446 tsk->thread.trap_no = 14;
447 info.si_signo = SIGSEGV;
449 /* info.si_code has been set above */
450 info.si_addr = (void __user *)address;
451 force_sig_info(SIGSEGV, &info, tsk);
457 /* Are we prepared to handle this kernel fault? */
458 fixup = search_exception_tables(regs->rip);
460 regs->rip = fixup->fixup;
465 * Hall of shame of CPU/BIOS bugs.
468 if (is_prefetch(regs, address, error_code))
471 if (is_errata93(regs, address))
475 * Oops. The kernel tried to access some bad page. We'll have to
476 * terminate things with extreme prejudice.
481 if (address < PAGE_SIZE)
482 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
484 printk(KERN_ALERT "Unable to handle kernel paging request");
485 printk(" at %016lx RIP: \n" KERN_ALERT,address);
486 printk_address(regs->rip);
488 dump_pagetable(address);
489 __die("Oops", regs, error_code);
490 /* Executive summary in case the body of the oops scrolled away */
491 printk(KERN_EMERG "CR2: %016lx\n", address);
497 * We ran out of memory, or some other thing happened to us that made
498 * us unable to handle the page fault gracefully.
501 up_read(&mm->mmap_sem);
503 if (current->pid == 1) {
507 printk("VM: killing process %s\n", tsk->comm);
513 up_read(&mm->mmap_sem);
515 /* Kernel mode? Handle exceptions or die */
516 if (!(error_code & 4))
519 tsk->thread.cr2 = address;
520 tsk->thread.error_code = error_code;
521 tsk->thread.trap_no = 14;
522 info.si_signo = SIGBUS;
524 info.si_code = BUS_ADRERR;
525 info.si_addr = (void __user *)address;
526 force_sig_info(SIGBUS, &info, tsk);
536 * x86-64 has the same kernel 3rd level pages for all CPUs.
537 * But for vmalloc/modules the TLB synchronization works lazily,
538 * so it can happen that we get a page fault for something
539 * that is really already in the page table. Just check if it
540 * is really there and when yes flush the local TLB.
542 pgd = pgd_offset_k(address);
543 if (!pgd_present(*pgd))
544 goto bad_area_nosemaphore;
545 pmd = pmd_offset(pgd, address);
546 if (!pmd_present(*pmd))
547 goto bad_area_nosemaphore;
548 pte = pte_offset_kernel(pmd, address);
549 if (!pte_present(*pte))
550 goto bad_area_nosemaphore;
556 page_table_corruption:
557 pgtable_bad(address, regs, error_code);