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)
68 unsigned char *max_instr;
70 /* If it was a exec fault ignore */
71 if (error_code & (1<<4))
74 instr = (unsigned char *)convert_rip_to_linear(current, regs);
75 max_instr = instr + 15;
77 if (user_mode(regs) && instr >= (unsigned char *)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 = (!user_mode(regs)) || (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" (pgd));
151 pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
152 pgd += pgd_index(address);
153 printk("PGD %lx ", pgd_val(*pgd));
154 if (bad_address(pgd)) goto bad;
155 if (!pgd_present(*pgd)) goto ret;
157 pud = __pud_offset_k((pud_t *)pgd_page(*pgd), address);
158 if (bad_address(pud)) goto bad;
159 printk("PUD %lx ", pud_val(*pud));
160 if (!pud_present(*pud)) goto ret;
162 pmd = pmd_offset(pud, 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)
215 if (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)
224 unsigned long flags = oops_begin();
225 struct task_struct *tsk;
227 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
228 current->comm, address);
229 dump_pagetable(address);
231 tsk->thread.cr2 = address;
232 tsk->thread.trap_no = 14;
233 tsk->thread.error_code = error_code;
234 __die("Bad pagetable", regs, error_code);
240 * Handle a fault on the vmalloc or module mapping area
242 * This assumes no large pages in there.
244 static int vmalloc_fault(unsigned long address)
246 pgd_t *pgd, *pgd_ref;
247 pud_t *pud, *pud_ref;
248 pmd_t *pmd, *pmd_ref;
249 pte_t *pte, *pte_ref;
251 /* Copy kernel mappings over when needed. This can also
252 happen within a race in page table update. In the later
255 pgd = pgd_offset(current->mm ?: &init_mm, address);
256 pgd_ref = pgd_offset_k(address);
257 if (pgd_none(*pgd_ref))
260 set_pgd(pgd, *pgd_ref);
262 /* Below here mismatches are bugs because these lower tables
265 pud = pud_offset(pgd, address);
266 pud_ref = pud_offset(pgd_ref, address);
267 if (pud_none(*pud_ref))
269 if (pud_none(*pud) || pud_page(*pud) != pud_page(*pud_ref))
271 pmd = pmd_offset(pud, address);
272 pmd_ref = pmd_offset(pud_ref, address);
273 if (pmd_none(*pmd_ref))
275 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
277 pte_ref = pte_offset_kernel(pmd_ref, address);
278 if (!pte_present(*pte_ref))
280 pte = pte_offset_kernel(pmd, address);
281 /* Don't use pte_page here, because the mappings can point
282 outside mem_map, and the NUMA hash lookup cannot handle
284 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
290 int page_fault_trace = 0;
291 int exception_trace = 1;
294 * This routine handles page faults. It determines the address,
295 * and the problem, and then passes it off to one of the appropriate
299 * bit 0 == 0 means no page found, 1 means protection fault
300 * bit 1 == 0 means read, 1 means write
301 * bit 2 == 0 means kernel, 1 means user-mode
302 * bit 3 == 1 means use of reserved bit detected
303 * bit 4 == 1 means fault was an instruction fetch
305 asmlinkage void __kprobes do_page_fault(struct pt_regs *regs,
306 unsigned long error_code)
308 struct task_struct *tsk;
309 struct mm_struct *mm;
310 struct vm_area_struct * vma;
311 unsigned long address;
312 const struct exception_table_entry *fixup;
317 /* get the address */
318 __asm__("movq %%cr2,%0":"=r" (address));
319 if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
320 SIGSEGV) == NOTIFY_STOP)
323 if (likely(regs->eflags & X86_EFLAGS_IF))
326 if (unlikely(page_fault_trace))
327 printk("pagefault rip:%lx rsp:%lx cs:%lu ss:%lu address %lx error %lx\n",
328 regs->rip,regs->rsp,regs->cs,regs->ss,address,error_code);
332 info.si_code = SEGV_MAPERR;
336 * We fault-in kernel-space virtual memory on-demand. The
337 * 'reference' page table is init_mm.pgd.
339 * NOTE! We MUST NOT take any locks for this case. We may
340 * be in an interrupt or a critical region, and should
341 * only copy the information from the master page table,
344 * This verifies that the fault happens in kernel space
345 * (error_code & 4) == 0, and that the fault was not a
346 * protection error (error_code & 9) == 0.
348 if (unlikely(address >= TASK_SIZE64)) {
349 if (!(error_code & 0xd) &&
350 ((address >= VMALLOC_START && address < VMALLOC_END) ||
351 (address >= MODULES_VADDR && address < MODULES_END))) {
352 if (vmalloc_fault(address) < 0)
353 goto bad_area_nosemaphore;
357 * Don't take the mm semaphore here. If we fixup a prefetch
358 * fault we could otherwise deadlock.
360 goto bad_area_nosemaphore;
363 if (unlikely(error_code & (1 << 3)))
364 pgtable_bad(address, regs, error_code);
367 * If we're in an interrupt or have no user
368 * context, we must not take the fault..
370 if (unlikely(in_atomic() || !mm))
371 goto bad_area_nosemaphore;
374 /* When running in the kernel we expect faults to occur only to
375 * addresses in user space. All other faults represent errors in the
376 * kernel and should generate an OOPS. Unfortunatly, in the case of an
377 * erroneous fault occuring in a code path which already holds mmap_sem
378 * we will deadlock attempting to validate the fault against the
379 * address space. Luckily the kernel only validly references user
380 * space from well defined areas of code, which are listed in the
383 * As the vast majority of faults will be valid we will only perform
384 * the source reference check when there is a possibilty of a deadlock.
385 * Attempt to lock the address space, if we cannot we then validate the
386 * source. If this is invalid we can skip the address space check,
387 * thus avoiding the deadlock.
389 if (!down_read_trylock(&mm->mmap_sem)) {
390 if ((error_code & 4) == 0 &&
391 !search_exception_tables(regs->rip))
392 goto bad_area_nosemaphore;
393 down_read(&mm->mmap_sem);
396 vma = find_vma(mm, address);
399 if (likely(vma->vm_start <= address))
401 if (!(vma->vm_flags & VM_GROWSDOWN))
403 if (error_code & 4) {
404 // XXX: align red zone size with ABI
405 if (address + 128 < regs->rsp)
408 if (expand_stack(vma, address))
411 * Ok, we have a good vm_area for this memory access, so
415 info.si_code = SEGV_ACCERR;
417 switch (error_code & 3) {
418 default: /* 3: write, present */
420 case 2: /* write, not present */
421 if (!(vma->vm_flags & VM_WRITE))
425 case 1: /* read, present */
427 case 0: /* read, not present */
428 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
433 * If for any reason at all we couldn't handle the fault,
434 * make sure we exit gracefully rather than endlessly redo
437 switch (handle_mm_fault(mm, vma, address, write)) {
444 case VM_FAULT_SIGBUS:
450 up_read(&mm->mmap_sem);
454 * Something tried to access memory that isn't in our memory map..
455 * Fix it, but check if it's kernel or user first..
458 up_read(&mm->mmap_sem);
460 bad_area_nosemaphore:
461 /* User mode accesses just cause a SIGSEGV */
462 if (error_code & 4) {
463 if (is_prefetch(regs, address, error_code))
466 /* Work around K8 erratum #100 K8 in compat mode
467 occasionally jumps to illegal addresses >4GB. We
468 catch this here in the page fault handler because
469 these addresses are not reachable. Just detect this
470 case and return. Any code segment in LDT is
471 compatibility mode. */
472 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
476 if (exception_trace && unhandled_signal(tsk, SIGSEGV)) {
478 "%s%s[%d]: segfault at %016lx rip %016lx rsp %016lx error %lx\n",
479 tsk->pid > 1 ? KERN_INFO : KERN_EMERG,
480 tsk->comm, tsk->pid, address, regs->rip,
481 regs->rsp, error_code);
484 tsk->thread.cr2 = address;
485 /* Kernel addresses are always protection faults */
486 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
487 tsk->thread.trap_no = 14;
488 info.si_signo = SIGSEGV;
490 /* info.si_code has been set above */
491 info.si_addr = (void __user *)address;
492 force_sig_info(SIGSEGV, &info, tsk);
498 /* Are we prepared to handle this kernel fault? */
499 fixup = search_exception_tables(regs->rip);
501 regs->rip = fixup->fixup;
506 * Hall of shame of CPU/BIOS bugs.
509 if (is_prefetch(regs, address, error_code))
512 if (is_errata93(regs, address))
516 * Oops. The kernel tried to access some bad page. We'll have to
517 * terminate things with extreme prejudice.
520 flags = oops_begin();
522 if (address < PAGE_SIZE)
523 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
525 printk(KERN_ALERT "Unable to handle kernel paging request");
526 printk(" at %016lx RIP: \n" KERN_ALERT,address);
527 printk_address(regs->rip);
529 dump_pagetable(address);
530 tsk->thread.cr2 = address;
531 tsk->thread.trap_no = 14;
532 tsk->thread.error_code = error_code;
533 __die("Oops", regs, error_code);
534 /* Executive summary in case the body of the oops scrolled away */
535 printk(KERN_EMERG "CR2: %016lx\n", address);
540 * We ran out of memory, or some other thing happened to us that made
541 * us unable to handle the page fault gracefully.
544 up_read(&mm->mmap_sem);
545 if (current->pid == 1) {
549 printk("VM: killing process %s\n", tsk->comm);
555 up_read(&mm->mmap_sem);
557 /* Kernel mode? Handle exceptions or die */
558 if (!(error_code & 4))
561 tsk->thread.cr2 = address;
562 tsk->thread.error_code = error_code;
563 tsk->thread.trap_no = 14;
564 info.si_signo = SIGBUS;
566 info.si_code = BUS_ADRERR;
567 info.si_addr = (void __user *)address;
568 force_sig_info(SIGBUS, &info, tsk);
572 static int __init enable_pagefaulttrace(char *str)
574 page_fault_trace = 1;
577 __setup("pagefaulttrace", enable_pagefaulttrace);