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>
37 /* Page fault error code bits */
38 #define PF_PROT (1<<0) /* or no page found */
39 #define PF_WRITE (1<<1)
40 #define PF_USER (1<<2)
41 #define PF_RSVD (1<<3)
42 #define PF_INSTR (1<<4)
44 void bust_spinlocks(int yes)
46 int loglevel_save = console_loglevel;
55 * OK, the message is on the console. Now we call printk()
56 * without oops_in_progress set so that printk will give klogd
57 * a poke. Hold onto your hats...
59 console_loglevel = 15; /* NMI oopser may have shut the console up */
61 console_loglevel = loglevel_save;
65 /* Sometimes the CPU reports invalid exceptions on prefetch.
66 Check that here and ignore.
67 Opcode checker based on code by Richard Brunner */
68 static noinline int is_prefetch(struct pt_regs *regs, unsigned long addr,
69 unsigned long error_code)
74 unsigned char *max_instr;
76 /* If it was a exec fault ignore */
77 if (error_code & PF_INSTR)
80 instr = (unsigned char *)convert_rip_to_linear(current, regs);
81 max_instr = instr + 15;
83 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
86 while (scan_more && instr < max_instr) {
88 unsigned char instr_hi;
89 unsigned char instr_lo;
91 if (__get_user(opcode, instr))
94 instr_hi = opcode & 0xf0;
95 instr_lo = opcode & 0x0f;
101 /* Values 0x26,0x2E,0x36,0x3E are valid x86
102 prefixes. In long mode, the CPU will signal
103 invalid opcode if some of these prefixes are
104 present so we will never get here anyway */
105 scan_more = ((instr_lo & 7) == 0x6);
109 /* In AMD64 long mode, 0x40 to 0x4F are valid REX prefixes
110 Need to figure out under what instruction mode the
111 instruction was issued ... */
112 /* Could check the LDT for lm, but for now it's good
113 enough to assume that long mode only uses well known
114 segments or kernel. */
115 scan_more = (!user_mode(regs)) || (regs->cs == __USER_CS);
119 /* 0x64 thru 0x67 are valid prefixes in all modes. */
120 scan_more = (instr_lo & 0xC) == 0x4;
123 /* 0xF0, 0xF2, and 0xF3 are valid prefixes in all modes. */
124 scan_more = !instr_lo || (instr_lo>>1) == 1;
127 /* Prefetch instruction is 0x0F0D or 0x0F18 */
129 if (__get_user(opcode, instr))
131 prefetch = (instr_lo == 0xF) &&
132 (opcode == 0x0D || opcode == 0x18);
142 static int bad_address(void *p)
145 return __get_user(dummy, (unsigned long *)p);
148 void dump_pagetable(unsigned long address)
155 asm("movq %%cr3,%0" : "=r" (pgd));
157 pgd = __va((unsigned long)pgd & PHYSICAL_PAGE_MASK);
158 pgd += pgd_index(address);
159 if (bad_address(pgd)) goto bad;
160 printk("PGD %lx ", pgd_val(*pgd));
161 if (!pgd_present(*pgd)) goto ret;
163 pud = __pud_offset_k((pud_t *)pgd_page(*pgd), address);
164 if (bad_address(pud)) goto bad;
165 printk("PUD %lx ", pud_val(*pud));
166 if (!pud_present(*pud)) goto ret;
168 pmd = pmd_offset(pud, address);
169 if (bad_address(pmd)) goto bad;
170 printk("PMD %lx ", pmd_val(*pmd));
171 if (!pmd_present(*pmd)) goto ret;
173 pte = pte_offset_kernel(pmd, address);
174 if (bad_address(pte)) goto bad;
175 printk("PTE %lx", pte_val(*pte));
183 static const char errata93_warning[] =
184 KERN_ERR "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
185 KERN_ERR "******* Working around it, but it may cause SEGVs or burn power.\n"
186 KERN_ERR "******* Please consider a BIOS update.\n"
187 KERN_ERR "******* Disabling USB legacy in the BIOS may also help.\n";
189 /* Workaround for K8 erratum #93 & buggy BIOS.
190 BIOS SMM functions are required to use a specific workaround
191 to avoid corruption of the 64bit RIP register on C stepping K8.
192 A lot of BIOS that didn't get tested properly miss this.
193 The OS sees this as a page fault with the upper 32bits of RIP cleared.
194 Try to work around it here.
195 Note we only handle faults in kernel here. */
197 static int is_errata93(struct pt_regs *regs, unsigned long address)
200 if (address != regs->rip)
202 if ((address >> 32) != 0)
204 address |= 0xffffffffUL << 32;
205 if ((address >= (u64)_stext && address <= (u64)_etext) ||
206 (address >= MODULES_VADDR && address <= MODULES_END)) {
208 printk(errata93_warning);
217 int unhandled_signal(struct task_struct *tsk, int sig)
221 if (tsk->ptrace & PT_PTRACED)
223 return (tsk->sighand->action[sig-1].sa.sa_handler == SIG_IGN) ||
224 (tsk->sighand->action[sig-1].sa.sa_handler == SIG_DFL);
227 static noinline void pgtable_bad(unsigned long address, struct pt_regs *regs,
228 unsigned long error_code)
230 unsigned long flags = oops_begin();
231 struct task_struct *tsk;
233 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
234 current->comm, address);
235 dump_pagetable(address);
237 tsk->thread.cr2 = address;
238 tsk->thread.trap_no = 14;
239 tsk->thread.error_code = error_code;
240 __die("Bad pagetable", regs, error_code);
246 * Handle a fault on the vmalloc area
248 * This assumes no large pages in there.
250 static int vmalloc_fault(unsigned long address)
252 pgd_t *pgd, *pgd_ref;
253 pud_t *pud, *pud_ref;
254 pmd_t *pmd, *pmd_ref;
255 pte_t *pte, *pte_ref;
257 /* Copy kernel mappings over when needed. This can also
258 happen within a race in page table update. In the later
261 pgd = pgd_offset(current->mm ?: &init_mm, address);
262 pgd_ref = pgd_offset_k(address);
263 if (pgd_none(*pgd_ref))
266 set_pgd(pgd, *pgd_ref);
268 BUG_ON(pgd_page(*pgd) != pgd_page(*pgd_ref));
270 /* Below here mismatches are bugs because these lower tables
273 pud = pud_offset(pgd, address);
274 pud_ref = pud_offset(pgd_ref, address);
275 if (pud_none(*pud_ref))
277 if (pud_none(*pud) || pud_page(*pud) != pud_page(*pud_ref))
279 pmd = pmd_offset(pud, address);
280 pmd_ref = pmd_offset(pud_ref, address);
281 if (pmd_none(*pmd_ref))
283 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
285 pte_ref = pte_offset_kernel(pmd_ref, address);
286 if (!pte_present(*pte_ref))
288 pte = pte_offset_kernel(pmd, address);
289 /* Don't use pte_page here, because the mappings can point
290 outside mem_map, and the NUMA hash lookup cannot handle
292 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
297 int page_fault_trace = 0;
298 int exception_trace = 1;
301 * This routine handles page faults. It determines the address,
302 * and the problem, and then passes it off to one of the appropriate
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;
319 prefetchw(&mm->mmap_sem);
321 /* get the address */
322 __asm__("movq %%cr2,%0":"=r" (address));
324 info.si_code = SEGV_MAPERR;
328 * We fault-in kernel-space virtual memory on-demand. The
329 * 'reference' page table is init_mm.pgd.
331 * NOTE! We MUST NOT take any locks for this case. We may
332 * be in an interrupt or a critical region, and should
333 * only copy the information from the master page table,
336 * This verifies that the fault happens in kernel space
337 * (error_code & 4) == 0, and that the fault was not a
338 * protection error (error_code & 9) == 0.
340 if (unlikely(address >= TASK_SIZE64)) {
342 * Don't check for the module range here: its PML4
343 * is always initialized because it's shared with the main
344 * kernel text. Only vmalloc may need PML4 syncups.
346 if (!(error_code & (PF_RSVD|PF_USER|PF_PROT)) &&
347 ((address >= VMALLOC_START && address < VMALLOC_END))) {
348 if (vmalloc_fault(address) >= 0)
351 if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
352 SIGSEGV) == NOTIFY_STOP)
355 * Don't take the mm semaphore here. If we fixup a prefetch
356 * fault we could otherwise deadlock.
358 goto bad_area_nosemaphore;
361 if (notify_die(DIE_PAGE_FAULT, "page fault", regs, error_code, 14,
362 SIGSEGV) == NOTIFY_STOP)
365 if (likely(regs->eflags & X86_EFLAGS_IF))
368 if (unlikely(page_fault_trace))
369 printk("pagefault rip:%lx rsp:%lx cs:%lu ss:%lu address %lx error %lx\n",
370 regs->rip,regs->rsp,regs->cs,regs->ss,address,error_code);
372 if (unlikely(error_code & PF_RSVD))
373 pgtable_bad(address, regs, error_code);
376 * If we're in an interrupt or have no user
377 * context, we must not take the fault..
379 if (unlikely(in_atomic() || !mm))
380 goto bad_area_nosemaphore;
383 /* When running in the kernel we expect faults to occur only to
384 * addresses in user space. All other faults represent errors in the
385 * kernel and should generate an OOPS. Unfortunatly, in the case of an
386 * erroneous fault occuring in a code path which already holds mmap_sem
387 * we will deadlock attempting to validate the fault against the
388 * address space. Luckily the kernel only validly references user
389 * space from well defined areas of code, which are listed in the
392 * As the vast majority of faults will be valid we will only perform
393 * the source reference check when there is a possibilty of a deadlock.
394 * Attempt to lock the address space, if we cannot we then validate the
395 * source. If this is invalid we can skip the address space check,
396 * thus avoiding the deadlock.
398 if (!down_read_trylock(&mm->mmap_sem)) {
399 if ((error_code & PF_USER) == 0 &&
400 !search_exception_tables(regs->rip))
401 goto bad_area_nosemaphore;
402 down_read(&mm->mmap_sem);
405 vma = find_vma(mm, address);
408 if (likely(vma->vm_start <= address))
410 if (!(vma->vm_flags & VM_GROWSDOWN))
412 if (error_code & 4) {
413 // XXX: align red zone size with ABI
414 if (address + 128 < regs->rsp)
417 if (expand_stack(vma, address))
420 * Ok, we have a good vm_area for this memory access, so
424 info.si_code = SEGV_ACCERR;
426 switch (error_code & (PF_PROT|PF_WRITE)) {
427 default: /* 3: write, present */
429 case PF_WRITE: /* write, not present */
430 if (!(vma->vm_flags & VM_WRITE))
434 case PF_PROT: /* read, present */
436 case 0: /* read, not present */
437 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
442 * If for any reason at all we couldn't handle the fault,
443 * make sure we exit gracefully rather than endlessly redo
446 switch (handle_mm_fault(mm, vma, address, write)) {
453 case VM_FAULT_SIGBUS:
459 up_read(&mm->mmap_sem);
463 * Something tried to access memory that isn't in our memory map..
464 * Fix it, but check if it's kernel or user first..
467 up_read(&mm->mmap_sem);
469 bad_area_nosemaphore:
470 /* User mode accesses just cause a SIGSEGV */
471 if (error_code & PF_USER) {
472 if (is_prefetch(regs, address, error_code))
475 /* Work around K8 erratum #100 K8 in compat mode
476 occasionally jumps to illegal addresses >4GB. We
477 catch this here in the page fault handler because
478 these addresses are not reachable. Just detect this
479 case and return. Any code segment in LDT is
480 compatibility mode. */
481 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) &&
485 if (exception_trace && unhandled_signal(tsk, SIGSEGV)) {
487 "%s%s[%d]: segfault at %016lx rip %016lx rsp %016lx error %lx\n",
488 tsk->pid > 1 ? KERN_INFO : KERN_EMERG,
489 tsk->comm, tsk->pid, address, regs->rip,
490 regs->rsp, error_code);
493 tsk->thread.cr2 = address;
494 /* Kernel addresses are always protection faults */
495 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
496 tsk->thread.trap_no = 14;
497 info.si_signo = SIGSEGV;
499 /* info.si_code has been set above */
500 info.si_addr = (void __user *)address;
501 force_sig_info(SIGSEGV, &info, tsk);
507 /* Are we prepared to handle this kernel fault? */
508 fixup = search_exception_tables(regs->rip);
510 regs->rip = fixup->fixup;
515 * Hall of shame of CPU/BIOS bugs.
518 if (is_prefetch(regs, address, error_code))
521 if (is_errata93(regs, address))
525 * Oops. The kernel tried to access some bad page. We'll have to
526 * terminate things with extreme prejudice.
529 flags = oops_begin();
531 if (address < PAGE_SIZE)
532 printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference");
534 printk(KERN_ALERT "Unable to handle kernel paging request");
535 printk(" at %016lx RIP: \n" KERN_ALERT,address);
536 printk_address(regs->rip);
538 dump_pagetable(address);
539 tsk->thread.cr2 = address;
540 tsk->thread.trap_no = 14;
541 tsk->thread.error_code = error_code;
542 __die("Oops", regs, error_code);
543 /* Executive summary in case the body of the oops scrolled away */
544 printk(KERN_EMERG "CR2: %016lx\n", address);
549 * We ran out of memory, or some other thing happened to us that made
550 * us unable to handle the page fault gracefully.
553 up_read(&mm->mmap_sem);
554 if (current->pid == 1) {
558 printk("VM: killing process %s\n", tsk->comm);
564 up_read(&mm->mmap_sem);
566 /* Kernel mode? Handle exceptions or die */
567 if (!(error_code & PF_USER))
570 tsk->thread.cr2 = address;
571 tsk->thread.error_code = error_code;
572 tsk->thread.trap_no = 14;
573 info.si_signo = SIGBUS;
575 info.si_code = BUS_ADRERR;
576 info.si_addr = (void __user *)address;
577 force_sig_info(SIGBUS, &info, tsk);
581 DEFINE_SPINLOCK(pgd_lock);
582 struct page *pgd_list;
584 void vmalloc_sync_all(void)
586 /* Note that races in the updates of insync and start aren't
588 insync can only get set bits added, and updates to start are only
589 improving performance (without affecting correctness if undone). */
590 static DECLARE_BITMAP(insync, PTRS_PER_PGD);
591 static unsigned long start = VMALLOC_START & PGDIR_MASK;
592 unsigned long address;
594 for (address = start; address <= VMALLOC_END; address += PGDIR_SIZE) {
595 if (!test_bit(pgd_index(address), insync)) {
596 const pgd_t *pgd_ref = pgd_offset_k(address);
599 if (pgd_none(*pgd_ref))
601 spin_lock(&pgd_lock);
602 for (page = pgd_list; page;
603 page = (struct page *)page->index) {
605 pgd = (pgd_t *)page_address(page) + pgd_index(address);
607 set_pgd(pgd, *pgd_ref);
609 BUG_ON(pgd_page(*pgd) != pgd_page(*pgd_ref));
611 spin_unlock(&pgd_lock);
612 set_bit(pgd_index(address), insync);
614 if (address == start)
615 start = address + PGDIR_SIZE;
617 /* Check that there is no need to do the same for the modules area. */
618 BUILD_BUG_ON(!(MODULES_VADDR > __START_KERNEL));
619 BUILD_BUG_ON(!(((MODULES_END - 1) & PGDIR_MASK) ==
620 (__START_KERNEL & PGDIR_MASK)));
623 static int __init enable_pagefaulttrace(char *str)
625 page_fault_trace = 1;
628 __setup("pagefaulttrace", enable_pagefaulttrace);