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>
27 #include <asm/system.h>
28 #include <asm/uaccess.h>
29 #include <asm/pgalloc.h>
31 #include <asm/tlbflush.h>
32 #include <asm/proto.h>
33 #include <asm/kdebug.h>
34 #include <asm-generic/sections.h>
36 void bust_spinlocks(int yes)
38 int loglevel_save = console_loglevel;
47 * OK, the message is on the console. Now we call printk()
48 * without oops_in_progress set so that printk will give klogd
49 * a poke. Hold onto your hats...
51 console_loglevel = 15; /* NMI oopser may have shut the console up */
53 console_loglevel = loglevel_save;
57 /* Sometimes the CPU reports invalid exceptions on prefetch.
58 Check that here and ignore.
59 Opcode checker based on code by Richard Brunner */
60 static noinline int is_prefetch(struct pt_regs *regs, unsigned long addr,
61 unsigned long error_code)
63 unsigned char *instr = (unsigned char *)(regs->rip);
66 unsigned char *max_instr = instr + 15;
68 /* If it was a exec fault ignore */
69 if (error_code & (1<<4))
72 /* Code segments in LDT could have a non zero base. Don't check
73 when that's possible */
74 if (regs->cs & (1<<2))
77 if ((regs->cs & 3) != 0 && regs->rip >= TASK_SIZE)
80 while (scan_more && instr < max_instr) {
82 unsigned char instr_hi;
83 unsigned char instr_lo;
85 if (__get_user(opcode, instr))
88 instr_hi = opcode & 0xf0;
89 instr_lo = opcode & 0x0f;
95 /* Values 0x26,0x2E,0x36,0x3E are valid x86
96 prefixes. In long mode, the CPU will signal
97 invalid opcode if some of these prefixes are
98 present so we will never get here anyway */
99 scan_more = ((instr_lo & 7) == 0x6);
103 /* In AMD64 long mode, 0x40 to 0x4F are valid REX prefixes
104 Need to figure out under what instruction mode the
105 instruction was issued ... */
106 /* Could check the LDT for lm, but for now it's good
107 enough to assume that long mode only uses well known
108 segments or kernel. */
109 scan_more = ((regs->cs & 3) == 0) || (regs->cs == __USER_CS);
113 /* 0x64 thru 0x67 are valid prefixes in all modes. */
114 scan_more = (instr_lo & 0xC) == 0x4;
117 /* 0xF0, 0xF2, and 0xF3 are valid prefixes in all modes. */
118 scan_more = !instr_lo || (instr_lo>>1) == 1;
121 /* Prefetch instruction is 0x0F0D or 0x0F18 */
123 if (__get_user(opcode, instr))
125 prefetch = (instr_lo == 0xF) &&
126 (opcode == 0x0D || opcode == 0x18);
136 static int bad_address(void *p)
139 return __get_user(dummy, (unsigned long *)p);
142 void dump_pagetable(unsigned long address)
149 asm("movq %%cr3,%0" : "=r" (pml4));
151 pml4 = __va((unsigned long)pml4 & PHYSICAL_PAGE_MASK);
152 pml4 += pml4_index(address);
153 printk("PML4 %lx ", pml4_val(*pml4));
154 if (bad_address(pml4)) goto bad;
155 if (!pml4_present(*pml4)) goto ret;
157 pgd = __pgd_offset_k((pgd_t *)pml4_page(*pml4), address);
158 if (bad_address(pgd)) goto bad;
159 printk("PGD %lx ", pgd_val(*pgd));
160 if (!pgd_present(*pgd)) goto ret;
162 pmd = pmd_offset(pgd, address);
163 if (bad_address(pmd)) goto bad;
164 printk("PMD %lx ", pmd_val(*pmd));
165 if (!pmd_present(*pmd)) goto ret;
167 pte = pte_offset_kernel(pmd, address);
168 if (bad_address(pte)) goto bad;
169 printk("PTE %lx", pte_val(*pte));
177 static const char errata93_warning[] =
178 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
179 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
180 KERN_ERR "******* Please consider a BIOS update.\n"
181 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
183 /* Workaround for K8 erratum #93 & buggy BIOS.
184 BIOS SMM functions are required to use a specific workaround
185 to avoid corruption of the 64bit RIP register on C stepping K8.
186 A lot of BIOS that didn't get tested properly miss this.
187 The OS sees this as a page fault with the upper 32bits of RIP cleared.
188 Try to work around it here.
189 Note we only handle faults in kernel here. */
191 static int is_errata93(struct pt_regs *regs, unsigned long address)
194 if (address != regs->rip)
196 if ((address >> 32) != 0)
198 address |= 0xffffffffUL << 32;
199 if ((address >= (u64)_stext && address <= (u64)_etext) ||
200 (address >= MODULES_VADDR && address <= MODULES_END)) {
202 printk(errata93_warning);
211 int unhandled_signal(struct task_struct *tsk, int sig)
213 /* Warn for strace, but not for gdb */
214 if (!test_ti_thread_flag(tsk->thread_info, TIF_SYSCALL_TRACE) &&
215 (tsk->ptrace & PT_PTRACED))
217 return (tsk->sighand->action[sig-1].sa.sa_handler == SIG_IGN) ||
218 (tsk->sighand->action[sig-1].sa.sa_handler == SIG_DFL);
221 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
222 unsigned long error_code)
225 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
226 current->comm, address);
227 dump_pagetable(address);
228 __die("Bad pagetable", regs, error_code);
233 int page_fault_trace;
234 int exception_trace = 1;
237 * This routine handles page faults. It determines the address,
238 * and the problem, and then passes it off to one of the appropriate
242 * bit 0 == 0 means no page found, 1 means protection fault
243 * bit 1 == 0 means read, 1 means write
244 * bit 2 == 0 means kernel, 1 means user-mode
245 * bit 3 == 1 means fault was an instruction fetch
247 asmlinkage void do_page_fault(struct pt_regs *regs, unsigned long error_code)
249 struct task_struct *tsk;
250 struct mm_struct *mm;
251 struct vm_area_struct * vma;
252 unsigned long address;
253 const struct exception_table_entry *fixup;
257 #ifdef CONFIG_CHECKING
260 struct x8664_pda *pda = cpu_pda + stack_smp_processor_id();
261 rdmsrl(MSR_GS_BASE, gs);
262 if (gs != (unsigned long)pda) {
263 wrmsrl(MSR_GS_BASE, pda);
264 printk("page_fault: wrong gs %lx expected %p\n", gs, pda);
269 /* get the address */
270 __asm__("movq %%cr2,%0":"=r" (address));
272 if (likely(regs->eflags & X86_EFLAGS_IF))
275 if (unlikely(page_fault_trace))
276 printk("pagefault rip:%lx rsp:%lx cs:%lu ss:%lu address %lx error %lx\n",
277 regs->rip,regs->rsp,regs->cs,regs->ss,address,error_code);
281 info.si_code = SEGV_MAPERR;
285 * We fault-in kernel-space virtual memory on-demand. The
286 * 'reference' page table is init_mm.pgd.
288 * NOTE! We MUST NOT take any locks for this case. We may
289 * be in an interrupt or a critical region, and should
290 * only copy the information from the master page table,
293 * This verifies that the fault happens in kernel space
294 * (error_code & 4) == 0, and that the fault was not a
295 * protection error (error_code & 1) == 0.
297 if (unlikely(address >= TASK_SIZE)) {
298 if (!(error_code & 5))
301 * Don't take the mm semaphore here. If we fixup a prefetch
302 * fault we could otherwise deadlock.
304 goto bad_area_nosemaphore;
307 if (unlikely(error_code & (1 << 3)))
308 goto page_table_corruption;
311 * If we're in an interrupt or have no user
312 * context, we must not take the fault..
314 if (unlikely(in_atomic() || !mm))
315 goto bad_area_nosemaphore;
318 /* When running in the kernel we expect faults to occur only to
319 * addresses in user space. All other faults represent errors in the
320 * kernel and should generate an OOPS. Unfortunatly, in the case of an
321 * erroneous fault occuring in a code path which already holds mmap_sem
322 * we will deadlock attempting to validate the fault against the
323 * address space. Luckily the kernel only validly references user
324 * space from well defined areas of code, which are listed in the
327 * As the vast majority of faults will be valid we will only perform
328 * the source reference check when there is a possibilty of a deadlock.
329 * Attempt to lock the address space, if we cannot we then validate the
330 * source. If this is invalid we can skip the address space check,
331 * thus avoiding the deadlock.
333 if (!down_read_trylock(&mm->mmap_sem)) {
334 if ((error_code & 4) == 0 &&
335 !search_exception_tables(regs->rip))
336 goto bad_area_nosemaphore;
337 down_read(&mm->mmap_sem);
340 vma = find_vma(mm, address);
343 if (likely(vma->vm_start <= address))
345 if (!(vma->vm_flags & VM_GROWSDOWN))
347 if (error_code & 4) {
348 // XXX: align red zone size with ABI
349 if (address + 128 < regs->rsp)
352 if (expand_stack(vma, address))
355 * Ok, we have a good vm_area for this memory access, so
359 info.si_code = SEGV_ACCERR;
361 switch (error_code & 3) {
362 default: /* 3: write, present */
364 case 2: /* write, not present */
365 if (!(vma->vm_flags & VM_WRITE))
369 case 1: /* read, present */
371 case 0: /* read, not present */
372 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
377 * If for any reason at all we couldn't handle the fault,
378 * make sure we exit gracefully rather than endlessly redo
381 switch (handle_mm_fault(mm, vma, address, write)) {
394 up_read(&mm->mmap_sem);
398 * Something tried to access memory that isn't in our memory map..
399 * Fix it, but check if it's kernel or user first..
402 up_read(&mm->mmap_sem);
404 bad_area_nosemaphore:
406 #ifdef CONFIG_IA32_EMULATION
407 /* 32bit vsyscall. map on demand. */
408 if (test_thread_flag(TIF_IA32) &&
409 address >= 0xffffe000 && address < 0xffffe000 + PAGE_SIZE) {
410 if (map_syscall32(mm, address) < 0)
416 /* User mode accesses just cause a SIGSEGV */
417 if (error_code & 4) {
418 if (is_prefetch(regs, address, error_code))
421 /* Work around K8 erratum #100 K8 in compat mode
422 occasionally jumps to illegal addresses >4GB. We
423 catch this here in the page fault handler because
424 these addresses are not reachable. Just detect this
425 case and return. Any code segment in LDT is
426 compatibility mode. */
427 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
431 if (exception_trace && unhandled_signal(tsk, SIGSEGV)) {
433 "%s[%d]: segfault at %016lx rip %016lx rsp %016lx error %lx\n",
434 tsk->comm, tsk->pid, address, regs->rip,
435 regs->rsp, error_code);
438 tsk->thread.cr2 = address;
439 /* Kernel addresses are always protection faults */
440 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
441 tsk->thread.trap_no = 14;
442 info.si_signo = SIGSEGV;
444 /* info.si_code has been set above */
445 info.si_addr = (void __user *)address;
446 force_sig_info(SIGSEGV, &info, tsk);
452 /* Are we prepared to handle this kernel fault? */
453 fixup = search_exception_tables(regs->rip);
455 regs->rip = fixup->fixup;
460 * Hall of shame of CPU/BIOS bugs.
463 if (is_prefetch(regs, address, error_code))
466 if (is_errata93(regs, address))
470 * Oops. The kernel tried to access some bad page. We'll have to
471 * terminate things with extreme prejudice.
476 if (address < PAGE_SIZE)
477 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
479 printk(KERN_ALERT "Unable to handle kernel paging request");
480 printk(" at %016lx RIP: \n" KERN_ALERT,address);
481 printk_address(regs->rip);
483 dump_pagetable(address);
484 __die("Oops", regs, error_code);
485 /* Executive summary in case the body of the oops scrolled away */
486 printk(KERN_EMERG "CR2: %016lx\n", address);
491 * We ran out of memory, or some other thing happened to us that made
492 * us unable to handle the page fault gracefully.
495 up_read(&mm->mmap_sem);
497 if (current->pid == 1) {
501 printk("VM: killing process %s\n", tsk->comm);
507 up_read(&mm->mmap_sem);
509 /* Kernel mode? Handle exceptions or die */
510 if (!(error_code & 4))
513 tsk->thread.cr2 = address;
514 tsk->thread.error_code = error_code;
515 tsk->thread.trap_no = 14;
516 info.si_signo = SIGBUS;
518 info.si_code = BUS_ADRERR;
519 info.si_addr = (void __user *)address;
520 force_sig_info(SIGBUS, &info, tsk);
530 * x86-64 has the same kernel 3rd level pages for all CPUs.
531 * But for vmalloc/modules the TLB synchronization works lazily,
532 * so it can happen that we get a page fault for something
533 * that is really already in the page table. Just check if it
534 * is really there and when yes flush the local TLB.
536 pgd = pgd_offset_k(address);
537 if (!pgd_present(*pgd))
538 goto bad_area_nosemaphore;
539 pmd = pmd_offset(pgd, address);
540 if (!pmd_present(*pmd))
541 goto bad_area_nosemaphore;
542 pte = pte_offset_kernel(pmd, address);
543 if (!pte_present(*pte))
544 goto bad_area_nosemaphore;
550 page_table_corruption:
551 pgtable_bad(address, regs, error_code);