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" (pgd));
153 pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
154 pgd += pgd_index(address);
155 printk("PGD %lx ", pgd_val(*pgd));
156 if (bad_address(pgd)) goto bad;
157 if (!pgd_present(*pgd)) goto ret;
159 pud = __pud_offset_k((pud_t *)pgd_page(*pgd), address);
160 if (bad_address(pud)) goto bad;
161 printk("PUD %lx ", pud_val(*pud));
162 if (!pud_present(*pud)) goto ret;
164 pmd = pmd_offset(pud, 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)
217 /* Warn for strace, but not for gdb */
218 if (!test_ti_thread_flag(tsk->thread_info, TIF_SYSCALL_TRACE) &&
219 (tsk->ptrace & PT_PTRACED))
221 return (tsk->sighand->action[sig-1].sa.sa_handler == SIG_IGN) ||
222 (tsk->sighand->action[sig-1].sa.sa_handler == SIG_DFL);
225 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
226 unsigned long error_code)
229 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
230 current->comm, address);
231 dump_pagetable(address);
232 __die("Bad pagetable", regs, error_code);
238 * Handle a fault on the vmalloc or module mapping area
240 * This assumes no large pages in there.
242 static int vmalloc_fault(unsigned long address)
244 pgd_t *pgd, *pgd_ref;
245 pud_t *pud, *pud_ref;
246 pmd_t *pmd, *pmd_ref;
247 pte_t *pte, *pte_ref;
249 /* Copy kernel mappings over when needed. This can also
250 happen within a race in page table update. In the later
253 pgd = pgd_offset(current->mm ?: &init_mm, address);
254 pgd_ref = pgd_offset_k(address);
255 if (pgd_none(*pgd_ref))
258 set_pgd(pgd, *pgd_ref);
260 /* Below here mismatches are bugs because these lower tables
263 pud = pud_offset(pgd, address);
264 pud_ref = pud_offset(pgd_ref, address);
265 if (pud_none(*pud_ref))
267 if (pud_none(*pud) || pud_page(*pud) != pud_page(*pud_ref))
269 pmd = pmd_offset(pud, address);
270 pmd_ref = pmd_offset(pud_ref, address);
271 if (pmd_none(*pmd_ref))
273 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
275 pte_ref = pte_offset_kernel(pmd_ref, address);
276 if (!pte_present(*pte_ref))
278 pte = pte_offset_kernel(pmd, address);
279 /* Don't use pte_page here, because the mappings can point
280 outside mem_map, and the NUMA hash lookup cannot handle
282 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
288 int page_fault_trace = 0;
289 int exception_trace = 1;
292 * This routine handles page faults. It determines the address,
293 * and the problem, and then passes it off to one of the appropriate
297 * bit 0 == 0 means no page found, 1 means protection fault
298 * bit 1 == 0 means read, 1 means write
299 * bit 2 == 0 means kernel, 1 means user-mode
300 * bit 3 == 1 means fault was an instruction fetch
302 asmlinkage void do_page_fault(struct pt_regs *regs, unsigned long error_code)
304 struct task_struct *tsk;
305 struct mm_struct *mm;
306 struct vm_area_struct * vma;
307 unsigned long address;
308 const struct exception_table_entry *fixup;
312 #ifdef CONFIG_CHECKING
315 struct x8664_pda *pda = cpu_pda + stack_smp_processor_id();
316 rdmsrl(MSR_GS_BASE, gs);
317 if (gs != (unsigned long)pda) {
318 wrmsrl(MSR_GS_BASE, pda);
319 printk("page_fault: wrong gs %lx expected %p\n", gs, pda);
324 /* get the address */
325 __asm__("movq %%cr2,%0":"=r" (address));
326 if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
327 SIGSEGV) == NOTIFY_STOP)
330 if (likely(regs->eflags & X86_EFLAGS_IF))
333 if (unlikely(page_fault_trace))
334 printk("pagefault rip:%lx rsp:%lx cs:%lu ss:%lu address %lx error %lx\n",
335 regs->rip,regs->rsp,regs->cs,regs->ss,address,error_code);
339 info.si_code = SEGV_MAPERR;
343 * We fault-in kernel-space virtual memory on-demand. The
344 * 'reference' page table is init_mm.pgd.
346 * NOTE! We MUST NOT take any locks for this case. We may
347 * be in an interrupt or a critical region, and should
348 * only copy the information from the master page table,
351 * This verifies that the fault happens in kernel space
352 * (error_code & 4) == 0, and that the fault was not a
353 * protection error (error_code & 1) == 0.
355 if (unlikely(address >= TASK_SIZE)) {
356 if (!(error_code & 5) &&
357 ((address >= VMALLOC_START && address < VMALLOC_END) ||
358 (address >= MODULES_VADDR && address < MODULES_END))) {
359 if (vmalloc_fault(address) < 0)
360 goto bad_area_nosemaphore;
364 * Don't take the mm semaphore here. If we fixup a prefetch
365 * fault we could otherwise deadlock.
367 goto bad_area_nosemaphore;
370 if (unlikely(error_code & (1 << 3)))
371 pgtable_bad(address, regs, error_code);
374 * If we're in an interrupt or have no user
375 * context, we must not take the fault..
377 if (unlikely(in_atomic() || !mm))
378 goto bad_area_nosemaphore;
381 /* When running in the kernel we expect faults to occur only to
382 * addresses in user space. All other faults represent errors in the
383 * kernel and should generate an OOPS. Unfortunatly, in the case of an
384 * erroneous fault occuring in a code path which already holds mmap_sem
385 * we will deadlock attempting to validate the fault against the
386 * address space. Luckily the kernel only validly references user
387 * space from well defined areas of code, which are listed in the
390 * As the vast majority of faults will be valid we will only perform
391 * the source reference check when there is a possibilty of a deadlock.
392 * Attempt to lock the address space, if we cannot we then validate the
393 * source. If this is invalid we can skip the address space check,
394 * thus avoiding the deadlock.
396 if (!down_read_trylock(&mm->mmap_sem)) {
397 if ((error_code & 4) == 0 &&
398 !search_exception_tables(regs->rip))
399 goto bad_area_nosemaphore;
400 down_read(&mm->mmap_sem);
403 vma = find_vma(mm, address);
406 if (likely(vma->vm_start <= address))
408 if (!(vma->vm_flags & VM_GROWSDOWN))
410 if (error_code & 4) {
411 // XXX: align red zone size with ABI
412 if (address + 128 < regs->rsp)
415 if (expand_stack(vma, address))
418 * Ok, we have a good vm_area for this memory access, so
422 info.si_code = SEGV_ACCERR;
424 switch (error_code & 3) {
425 default: /* 3: write, present */
427 case 2: /* write, not present */
428 if (!(vma->vm_flags & VM_WRITE))
432 case 1: /* read, present */
434 case 0: /* read, not present */
435 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
440 * If for any reason at all we couldn't handle the fault,
441 * make sure we exit gracefully rather than endlessly redo
444 switch (handle_mm_fault(mm, vma, address, write)) {
457 up_read(&mm->mmap_sem);
461 * Something tried to access memory that isn't in our memory map..
462 * Fix it, but check if it's kernel or user first..
465 up_read(&mm->mmap_sem);
467 bad_area_nosemaphore:
469 #ifdef CONFIG_IA32_EMULATION
470 /* 32bit vsyscall. map on demand. */
471 if (test_thread_flag(TIF_IA32) &&
472 address >= VSYSCALL32_BASE && address < VSYSCALL32_END) {
473 if (map_syscall32(mm, address) < 0)
479 /* User mode accesses just cause a SIGSEGV */
480 if (error_code & 4) {
481 if (is_prefetch(regs, address, error_code))
484 /* Work around K8 erratum #100 K8 in compat mode
485 occasionally jumps to illegal addresses >4GB. We
486 catch this here in the page fault handler because
487 these addresses are not reachable. Just detect this
488 case and return. Any code segment in LDT is
489 compatibility mode. */
490 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
494 if (exception_trace && unhandled_signal(tsk, SIGSEGV)) {
496 "%s%s[%d]: segfault at %016lx rip %016lx rsp %016lx error %lx\n",
497 tsk->pid > 1 ? KERN_INFO : KERN_EMERG,
498 tsk->comm, tsk->pid, address, regs->rip,
499 regs->rsp, error_code);
502 tsk->thread.cr2 = address;
503 /* Kernel addresses are always protection faults */
504 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
505 tsk->thread.trap_no = 14;
506 info.si_signo = SIGSEGV;
508 /* info.si_code has been set above */
509 info.si_addr = (void __user *)address;
510 force_sig_info(SIGSEGV, &info, tsk);
516 /* Are we prepared to handle this kernel fault? */
517 fixup = search_exception_tables(regs->rip);
519 regs->rip = fixup->fixup;
524 * Hall of shame of CPU/BIOS bugs.
527 if (is_prefetch(regs, address, error_code))
530 if (is_errata93(regs, address))
534 * Oops. The kernel tried to access some bad page. We'll have to
535 * terminate things with extreme prejudice.
540 if (address < PAGE_SIZE)
541 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
543 printk(KERN_ALERT "Unable to handle kernel paging request");
544 printk(" at %016lx RIP: \n" KERN_ALERT,address);
545 printk_address(regs->rip);
547 dump_pagetable(address);
548 __die("Oops", regs, error_code);
549 /* Executive summary in case the body of the oops scrolled away */
550 printk(KERN_EMERG "CR2: %016lx\n", address);
555 * We ran out of memory, or some other thing happened to us that made
556 * us unable to handle the page fault gracefully.
559 up_read(&mm->mmap_sem);
561 if (current->pid == 1) {
565 printk("VM: killing process %s\n", tsk->comm);
571 up_read(&mm->mmap_sem);
573 /* Kernel mode? Handle exceptions or die */
574 if (!(error_code & 4))
577 tsk->thread.cr2 = address;
578 tsk->thread.error_code = error_code;
579 tsk->thread.trap_no = 14;
580 info.si_signo = SIGBUS;
582 info.si_code = BUS_ADRERR;
583 info.si_addr = (void __user *)address;
584 force_sig_info(SIGBUS, &info, tsk);