2 * Copyright (C) 1995 Linus Torvalds
3 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
4 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
6 #include <linux/magic.h> /* STACK_END_MAGIC */
7 #include <linux/sched.h> /* test_thread_flag(), ... */
8 #include <linux/kdebug.h> /* oops_begin/end, ... */
9 #include <linux/module.h> /* search_exception_table */
10 #include <linux/bootmem.h> /* max_low_pfn */
11 #include <linux/kprobes.h> /* __kprobes, ... */
12 #include <linux/mmiotrace.h> /* kmmio_handler, ... */
13 #include <linux/perf_event.h> /* perf_sw_event */
15 #include <asm/traps.h> /* dotraplinkage, ... */
16 #include <asm/pgalloc.h> /* pgd_*(), ... */
17 #include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
20 * Page fault error code bits:
22 * bit 0 == 0: no page found 1: protection fault
23 * bit 1 == 0: read access 1: write access
24 * bit 2 == 0: kernel-mode access 1: user-mode access
25 * bit 3 == 1: use of reserved bit detected
26 * bit 4 == 1: fault was an instruction fetch
28 enum x86_pf_error_code {
38 * Returns 0 if mmiotrace is disabled, or if the fault is not
39 * handled by mmiotrace:
41 static inline int kmmio_fault(struct pt_regs *regs, unsigned long addr)
43 if (unlikely(is_kmmio_active()))
44 if (kmmio_handler(regs, addr) == 1)
49 static inline int notify_page_fault(struct pt_regs *regs)
53 /* kprobe_running() needs smp_processor_id() */
54 if (kprobes_built_in() && !user_mode_vm(regs)) {
56 if (kprobe_running() && kprobe_fault_handler(regs, 14))
69 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
70 * Check that here and ignore it.
74 * Sometimes the CPU reports invalid exceptions on prefetch.
75 * Check that here and ignore it.
77 * Opcode checker based on code by Richard Brunner.
80 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
81 unsigned char opcode, int *prefetch)
83 unsigned char instr_hi = opcode & 0xf0;
84 unsigned char instr_lo = opcode & 0x0f;
90 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
91 * In X86_64 long mode, the CPU will signal invalid
92 * opcode if some of these prefixes are present so
93 * X86_64 will never get here anyway
95 return ((instr_lo & 7) == 0x6);
99 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
100 * Need to figure out under what instruction mode the
101 * instruction was issued. Could check the LDT for lm,
102 * but for now it's good enough to assume that long
103 * mode only uses well known segments or kernel.
105 return (!user_mode(regs)) || (regs->cs == __USER_CS);
108 /* 0x64 thru 0x67 are valid prefixes in all modes. */
109 return (instr_lo & 0xC) == 0x4;
111 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
112 return !instr_lo || (instr_lo>>1) == 1;
114 /* Prefetch instruction is 0x0F0D or 0x0F18 */
115 if (probe_kernel_address(instr, opcode))
118 *prefetch = (instr_lo == 0xF) &&
119 (opcode == 0x0D || opcode == 0x18);
127 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
129 unsigned char *max_instr;
130 unsigned char *instr;
134 * If it was a exec (instruction fetch) fault on NX page, then
135 * do not ignore the fault:
137 if (error_code & PF_INSTR)
140 instr = (void *)convert_ip_to_linear(current, regs);
141 max_instr = instr + 15;
143 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
146 while (instr < max_instr) {
147 unsigned char opcode;
149 if (probe_kernel_address(instr, opcode))
154 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
161 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
162 struct task_struct *tsk)
166 info.si_signo = si_signo;
168 info.si_code = si_code;
169 info.si_addr = (void __user *)address;
170 info.si_addr_lsb = si_code == BUS_MCEERR_AR ? PAGE_SHIFT : 0;
172 force_sig_info(si_signo, &info, tsk);
175 DEFINE_SPINLOCK(pgd_lock);
179 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
181 unsigned index = pgd_index(address);
187 pgd_k = init_mm.pgd + index;
189 if (!pgd_present(*pgd_k))
193 * set_pgd(pgd, *pgd_k); here would be useless on PAE
194 * and redundant with the set_pmd() on non-PAE. As would
197 pud = pud_offset(pgd, address);
198 pud_k = pud_offset(pgd_k, address);
199 if (!pud_present(*pud_k))
202 pmd = pmd_offset(pud, address);
203 pmd_k = pmd_offset(pud_k, address);
204 if (!pmd_present(*pmd_k))
207 if (!pmd_present(*pmd))
208 set_pmd(pmd, *pmd_k);
210 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
215 void vmalloc_sync_all(void)
217 unsigned long address;
219 if (SHARED_KERNEL_PMD)
222 for (address = VMALLOC_START & PMD_MASK;
223 address >= TASK_SIZE && address < FIXADDR_TOP;
224 address += PMD_SIZE) {
229 spin_lock_irqsave(&pgd_lock, flags);
230 list_for_each_entry(page, &pgd_list, lru) {
231 if (!vmalloc_sync_one(page_address(page), address))
234 spin_unlock_irqrestore(&pgd_lock, flags);
241 * Handle a fault on the vmalloc or module mapping area
243 static noinline int vmalloc_fault(unsigned long address)
245 unsigned long pgd_paddr;
249 /* Make sure we are in vmalloc area: */
250 if (!(address >= VMALLOC_START && address < VMALLOC_END))
254 * Synchronize this task's top level page-table
255 * with the 'reference' page table.
257 * Do _not_ use "current" here. We might be inside
258 * an interrupt in the middle of a task switch..
260 pgd_paddr = read_cr3();
261 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
265 pte_k = pte_offset_kernel(pmd_k, address);
266 if (!pte_present(*pte_k))
273 * Did it hit the DOS screen memory VA from vm86 mode?
276 check_v8086_mode(struct pt_regs *regs, unsigned long address,
277 struct task_struct *tsk)
281 if (!v8086_mode(regs))
284 bit = (address - 0xA0000) >> PAGE_SHIFT;
286 tsk->thread.screen_bitmap |= 1 << bit;
289 static bool low_pfn(unsigned long pfn)
291 return pfn < max_low_pfn;
294 static void dump_pagetable(unsigned long address)
296 pgd_t *base = __va(read_cr3());
297 pgd_t *pgd = &base[pgd_index(address)];
301 #ifdef CONFIG_X86_PAE
302 printk("*pdpt = %016Lx ", pgd_val(*pgd));
303 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
306 pmd = pmd_offset(pud_offset(pgd, address), address);
307 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
310 * We must not directly access the pte in the highpte
311 * case if the page table is located in highmem.
312 * And let's rather not kmap-atomic the pte, just in case
313 * it's allocated already:
315 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
318 pte = pte_offset_kernel(pmd, address);
319 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
324 #else /* CONFIG_X86_64: */
326 void vmalloc_sync_all(void)
328 unsigned long address;
330 for (address = VMALLOC_START & PGDIR_MASK; address <= VMALLOC_END;
331 address += PGDIR_SIZE) {
333 const pgd_t *pgd_ref = pgd_offset_k(address);
337 if (pgd_none(*pgd_ref))
340 spin_lock_irqsave(&pgd_lock, flags);
341 list_for_each_entry(page, &pgd_list, lru) {
343 pgd = (pgd_t *)page_address(page) + pgd_index(address);
345 set_pgd(pgd, *pgd_ref);
347 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
349 spin_unlock_irqrestore(&pgd_lock, flags);
356 * Handle a fault on the vmalloc area
358 * This assumes no large pages in there.
360 static noinline int vmalloc_fault(unsigned long address)
362 pgd_t *pgd, *pgd_ref;
363 pud_t *pud, *pud_ref;
364 pmd_t *pmd, *pmd_ref;
365 pte_t *pte, *pte_ref;
367 /* Make sure we are in vmalloc area: */
368 if (!(address >= VMALLOC_START && address < VMALLOC_END))
372 * Copy kernel mappings over when needed. This can also
373 * happen within a race in page table update. In the later
376 pgd = pgd_offset(current->active_mm, address);
377 pgd_ref = pgd_offset_k(address);
378 if (pgd_none(*pgd_ref))
382 set_pgd(pgd, *pgd_ref);
384 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
387 * Below here mismatches are bugs because these lower tables
391 pud = pud_offset(pgd, address);
392 pud_ref = pud_offset(pgd_ref, address);
393 if (pud_none(*pud_ref))
396 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
399 pmd = pmd_offset(pud, address);
400 pmd_ref = pmd_offset(pud_ref, address);
401 if (pmd_none(*pmd_ref))
404 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
407 pte_ref = pte_offset_kernel(pmd_ref, address);
408 if (!pte_present(*pte_ref))
411 pte = pte_offset_kernel(pmd, address);
414 * Don't use pte_page here, because the mappings can point
415 * outside mem_map, and the NUMA hash lookup cannot handle
418 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
424 static const char errata93_warning[] =
426 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
427 "******* Working around it, but it may cause SEGVs or burn power.\n"
428 "******* Please consider a BIOS update.\n"
429 "******* Disabling USB legacy in the BIOS may also help.\n";
432 * No vm86 mode in 64-bit mode:
435 check_v8086_mode(struct pt_regs *regs, unsigned long address,
436 struct task_struct *tsk)
440 static int bad_address(void *p)
444 return probe_kernel_address((unsigned long *)p, dummy);
447 static void dump_pagetable(unsigned long address)
449 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
450 pgd_t *pgd = base + pgd_index(address);
455 if (bad_address(pgd))
458 printk("PGD %lx ", pgd_val(*pgd));
460 if (!pgd_present(*pgd))
463 pud = pud_offset(pgd, address);
464 if (bad_address(pud))
467 printk("PUD %lx ", pud_val(*pud));
468 if (!pud_present(*pud) || pud_large(*pud))
471 pmd = pmd_offset(pud, address);
472 if (bad_address(pmd))
475 printk("PMD %lx ", pmd_val(*pmd));
476 if (!pmd_present(*pmd) || pmd_large(*pmd))
479 pte = pte_offset_kernel(pmd, address);
480 if (bad_address(pte))
483 printk("PTE %lx", pte_val(*pte));
491 #endif /* CONFIG_X86_64 */
494 * Workaround for K8 erratum #93 & buggy BIOS.
496 * BIOS SMM functions are required to use a specific workaround
497 * to avoid corruption of the 64bit RIP register on C stepping K8.
499 * A lot of BIOS that didn't get tested properly miss this.
501 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
502 * Try to work around it here.
504 * Note we only handle faults in kernel here.
505 * Does nothing on 32-bit.
507 static int is_errata93(struct pt_regs *regs, unsigned long address)
510 if (address != regs->ip)
513 if ((address >> 32) != 0)
516 address |= 0xffffffffUL << 32;
517 if ((address >= (u64)_stext && address <= (u64)_etext) ||
518 (address >= MODULES_VADDR && address <= MODULES_END)) {
519 printk_once(errata93_warning);
528 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
529 * to illegal addresses >4GB.
531 * We catch this in the page fault handler because these addresses
532 * are not reachable. Just detect this case and return. Any code
533 * segment in LDT is compatibility mode.
535 static int is_errata100(struct pt_regs *regs, unsigned long address)
538 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
544 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
546 #ifdef CONFIG_X86_F00F_BUG
550 * Pentium F0 0F C7 C8 bug workaround:
552 if (boot_cpu_data.f00f_bug) {
553 nr = (address - idt_descr.address) >> 3;
556 do_invalid_op(regs, 0);
564 static const char nx_warning[] = KERN_CRIT
565 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
568 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
569 unsigned long address)
571 if (!oops_may_print())
574 if (error_code & PF_INSTR) {
577 pte_t *pte = lookup_address(address, &level);
579 if (pte && pte_present(*pte) && !pte_exec(*pte))
580 printk(nx_warning, current_uid());
583 printk(KERN_ALERT "BUG: unable to handle kernel ");
584 if (address < PAGE_SIZE)
585 printk(KERN_CONT "NULL pointer dereference");
587 printk(KERN_CONT "paging request");
589 printk(KERN_CONT " at %p\n", (void *) address);
590 printk(KERN_ALERT "IP:");
591 printk_address(regs->ip, 1);
593 dump_pagetable(address);
597 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
598 unsigned long address)
600 struct task_struct *tsk;
604 flags = oops_begin();
608 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
610 dump_pagetable(address);
612 tsk->thread.cr2 = address;
613 tsk->thread.trap_no = 14;
614 tsk->thread.error_code = error_code;
616 if (__die("Bad pagetable", regs, error_code))
619 oops_end(flags, regs, sig);
623 no_context(struct pt_regs *regs, unsigned long error_code,
624 unsigned long address)
626 struct task_struct *tsk = current;
627 unsigned long *stackend;
631 /* Are we prepared to handle this kernel fault? */
632 if (fixup_exception(regs))
638 * Valid to do another page fault here, because if this fault
639 * had been triggered by is_prefetch fixup_exception would have
644 * Hall of shame of CPU/BIOS bugs.
646 if (is_prefetch(regs, error_code, address))
649 if (is_errata93(regs, address))
653 * Oops. The kernel tried to access some bad page. We'll have to
654 * terminate things with extreme prejudice:
656 flags = oops_begin();
658 show_fault_oops(regs, error_code, address);
660 stackend = end_of_stack(tsk);
661 if (*stackend != STACK_END_MAGIC)
662 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
664 tsk->thread.cr2 = address;
665 tsk->thread.trap_no = 14;
666 tsk->thread.error_code = error_code;
669 if (__die("Oops", regs, error_code))
672 /* Executive summary in case the body of the oops scrolled away */
673 printk(KERN_EMERG "CR2: %016lx\n", address);
675 oops_end(flags, regs, sig);
679 * Print out info about fatal segfaults, if the show_unhandled_signals
683 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
684 unsigned long address, struct task_struct *tsk)
686 if (!unhandled_signal(tsk, SIGSEGV))
689 if (!printk_ratelimit())
692 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
693 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
694 tsk->comm, task_pid_nr(tsk), address,
695 (void *)regs->ip, (void *)regs->sp, error_code);
697 print_vma_addr(KERN_CONT " in ", regs->ip);
699 printk(KERN_CONT "\n");
703 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
704 unsigned long address, int si_code)
706 struct task_struct *tsk = current;
708 /* User mode accesses just cause a SIGSEGV */
709 if (error_code & PF_USER) {
711 * It's possible to have interrupts off here:
716 * Valid to do another page fault here because this one came
719 if (is_prefetch(regs, error_code, address))
722 if (is_errata100(regs, address))
725 if (unlikely(show_unhandled_signals))
726 show_signal_msg(regs, error_code, address, tsk);
728 /* Kernel addresses are always protection faults: */
729 tsk->thread.cr2 = address;
730 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
731 tsk->thread.trap_no = 14;
733 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
738 if (is_f00f_bug(regs, address))
741 no_context(regs, error_code, address);
745 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
746 unsigned long address)
748 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
752 __bad_area(struct pt_regs *regs, unsigned long error_code,
753 unsigned long address, int si_code)
755 struct mm_struct *mm = current->mm;
758 * Something tried to access memory that isn't in our memory map..
759 * Fix it, but check if it's kernel or user first..
761 up_read(&mm->mmap_sem);
763 __bad_area_nosemaphore(regs, error_code, address, si_code);
767 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
769 __bad_area(regs, error_code, address, SEGV_MAPERR);
773 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
774 unsigned long address)
776 __bad_area(regs, error_code, address, SEGV_ACCERR);
779 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
781 out_of_memory(struct pt_regs *regs, unsigned long error_code,
782 unsigned long address)
785 * We ran out of memory, call the OOM killer, and return the userspace
786 * (which will retry the fault, or kill us if we got oom-killed):
788 up_read(¤t->mm->mmap_sem);
790 pagefault_out_of_memory();
794 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
797 struct task_struct *tsk = current;
798 struct mm_struct *mm = tsk->mm;
799 int code = BUS_ADRERR;
801 up_read(&mm->mmap_sem);
803 /* Kernel mode? Handle exceptions or die: */
804 if (!(error_code & PF_USER))
805 no_context(regs, error_code, address);
807 /* User-space => ok to do another page fault: */
808 if (is_prefetch(regs, error_code, address))
811 tsk->thread.cr2 = address;
812 tsk->thread.error_code = error_code;
813 tsk->thread.trap_no = 14;
815 #ifdef CONFIG_MEMORY_FAILURE
816 if (fault & VM_FAULT_HWPOISON) {
818 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
819 tsk->comm, tsk->pid, address);
820 code = BUS_MCEERR_AR;
823 force_sig_info_fault(SIGBUS, code, address, tsk);
827 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
828 unsigned long address, unsigned int fault)
830 if (fault & VM_FAULT_OOM) {
831 out_of_memory(regs, error_code, address);
833 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON))
834 do_sigbus(regs, error_code, address, fault);
840 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
842 if ((error_code & PF_WRITE) && !pte_write(*pte))
845 if ((error_code & PF_INSTR) && !pte_exec(*pte))
852 * Handle a spurious fault caused by a stale TLB entry.
854 * This allows us to lazily refresh the TLB when increasing the
855 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
856 * eagerly is very expensive since that implies doing a full
857 * cross-processor TLB flush, even if no stale TLB entries exist
858 * on other processors.
860 * There are no security implications to leaving a stale TLB when
861 * increasing the permissions on a page.
864 spurious_fault(unsigned long error_code, unsigned long address)
872 /* Reserved-bit violation or user access to kernel space? */
873 if (error_code & (PF_USER | PF_RSVD))
876 pgd = init_mm.pgd + pgd_index(address);
877 if (!pgd_present(*pgd))
880 pud = pud_offset(pgd, address);
881 if (!pud_present(*pud))
885 return spurious_fault_check(error_code, (pte_t *) pud);
887 pmd = pmd_offset(pud, address);
888 if (!pmd_present(*pmd))
892 return spurious_fault_check(error_code, (pte_t *) pmd);
894 pte = pte_offset_kernel(pmd, address);
895 if (!pte_present(*pte))
898 ret = spurious_fault_check(error_code, pte);
903 * Make sure we have permissions in PMD.
904 * If not, then there's a bug in the page tables:
906 ret = spurious_fault_check(error_code, (pte_t *) pmd);
907 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
912 int show_unhandled_signals = 1;
915 access_error(unsigned long error_code, int write, struct vm_area_struct *vma)
918 /* write, present and write, not present: */
919 if (unlikely(!(vma->vm_flags & VM_WRITE)))
925 if (unlikely(error_code & PF_PROT))
928 /* read, not present: */
929 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
935 static int fault_in_kernel_space(unsigned long address)
937 return address >= TASK_SIZE_MAX;
941 * This routine handles page faults. It determines the address,
942 * and the problem, and then passes it off to one of the appropriate
945 dotraplinkage void __kprobes
946 do_page_fault(struct pt_regs *regs, unsigned long error_code)
948 struct vm_area_struct *vma;
949 struct task_struct *tsk;
950 unsigned long address;
951 struct mm_struct *mm;
958 /* Get the faulting address: */
959 address = read_cr2();
962 * Detect and handle instructions that would cause a page fault for
963 * both a tracked kernel page and a userspace page.
965 if (kmemcheck_active(regs))
966 kmemcheck_hide(regs);
967 prefetchw(&mm->mmap_sem);
969 if (unlikely(kmmio_fault(regs, address)))
973 * We fault-in kernel-space virtual memory on-demand. The
974 * 'reference' page table is init_mm.pgd.
976 * NOTE! We MUST NOT take any locks for this case. We may
977 * be in an interrupt or a critical region, and should
978 * only copy the information from the master page table,
981 * This verifies that the fault happens in kernel space
982 * (error_code & 4) == 0, and that the fault was not a
983 * protection error (error_code & 9) == 0.
985 if (unlikely(fault_in_kernel_space(address))) {
986 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
987 if (vmalloc_fault(address) >= 0)
990 if (kmemcheck_fault(regs, address, error_code))
994 /* Can handle a stale RO->RW TLB: */
995 if (spurious_fault(error_code, address))
998 /* kprobes don't want to hook the spurious faults: */
999 if (notify_page_fault(regs))
1002 * Don't take the mm semaphore here. If we fixup a prefetch
1003 * fault we could otherwise deadlock:
1005 bad_area_nosemaphore(regs, error_code, address);
1010 /* kprobes don't want to hook the spurious faults: */
1011 if (unlikely(notify_page_fault(regs)))
1014 * It's safe to allow irq's after cr2 has been saved and the
1015 * vmalloc fault has been handled.
1017 * User-mode registers count as a user access even for any
1018 * potential system fault or CPU buglet:
1020 if (user_mode_vm(regs)) {
1022 error_code |= PF_USER;
1024 if (regs->flags & X86_EFLAGS_IF)
1028 if (unlikely(error_code & PF_RSVD))
1029 pgtable_bad(regs, error_code, address);
1031 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1034 * If we're in an interrupt, have no user context or are running
1035 * in an atomic region then we must not take the fault:
1037 if (unlikely(in_atomic() || !mm)) {
1038 bad_area_nosemaphore(regs, error_code, address);
1043 * When running in the kernel we expect faults to occur only to
1044 * addresses in user space. All other faults represent errors in
1045 * the kernel and should generate an OOPS. Unfortunately, in the
1046 * case of an erroneous fault occurring in a code path which already
1047 * holds mmap_sem we will deadlock attempting to validate the fault
1048 * against the address space. Luckily the kernel only validly
1049 * references user space from well defined areas of code, which are
1050 * listed in the exceptions table.
1052 * As the vast majority of faults will be valid we will only perform
1053 * the source reference check when there is a possibility of a
1054 * deadlock. Attempt to lock the address space, if we cannot we then
1055 * validate the source. If this is invalid we can skip the address
1056 * space check, thus avoiding the deadlock:
1058 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1059 if ((error_code & PF_USER) == 0 &&
1060 !search_exception_tables(regs->ip)) {
1061 bad_area_nosemaphore(regs, error_code, address);
1064 down_read(&mm->mmap_sem);
1067 * The above down_read_trylock() might have succeeded in
1068 * which case we'll have missed the might_sleep() from
1074 vma = find_vma(mm, address);
1075 if (unlikely(!vma)) {
1076 bad_area(regs, error_code, address);
1079 if (likely(vma->vm_start <= address))
1081 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1082 bad_area(regs, error_code, address);
1085 if (error_code & PF_USER) {
1087 * Accessing the stack below %sp is always a bug.
1088 * The large cushion allows instructions like enter
1089 * and pusha to work. ("enter $65535, $31" pushes
1090 * 32 pointers and then decrements %sp by 65535.)
1092 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1093 bad_area(regs, error_code, address);
1097 if (unlikely(expand_stack(vma, address))) {
1098 bad_area(regs, error_code, address);
1103 * Ok, we have a good vm_area for this memory access, so
1104 * we can handle it..
1107 write = error_code & PF_WRITE;
1109 if (unlikely(access_error(error_code, write, vma))) {
1110 bad_area_access_error(regs, error_code, address);
1115 * If for any reason at all we couldn't handle the fault,
1116 * make sure we exit gracefully rather than endlessly redo
1119 fault = handle_mm_fault(mm, vma, address, write ? FAULT_FLAG_WRITE : 0);
1121 if (unlikely(fault & VM_FAULT_ERROR)) {
1122 mm_fault_error(regs, error_code, address, fault);
1126 if (fault & VM_FAULT_MAJOR) {
1128 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1132 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1136 check_v8086_mode(regs, address, tsk);
1138 up_read(&mm->mmap_sem);