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 */
14 #include <linux/hugetlb.h> /* hstate_index_to_shift */
15 #include <linux/prefetch.h> /* prefetchw */
17 #include <asm/traps.h> /* dotraplinkage, ... */
18 #include <asm/pgalloc.h> /* pgd_*(), ... */
19 #include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
22 * Page fault error code bits:
24 * bit 0 == 0: no page found 1: protection fault
25 * bit 1 == 0: read access 1: write access
26 * bit 2 == 0: kernel-mode access 1: user-mode access
27 * bit 3 == 1: use of reserved bit detected
28 * bit 4 == 1: fault was an instruction fetch
30 enum x86_pf_error_code {
40 * Returns 0 if mmiotrace is disabled, or if the fault is not
41 * handled by mmiotrace:
43 static inline int __kprobes
44 kmmio_fault(struct pt_regs *regs, unsigned long addr)
46 if (unlikely(is_kmmio_active()))
47 if (kmmio_handler(regs, addr) == 1)
52 static inline int __kprobes notify_page_fault(struct pt_regs *regs)
56 /* kprobe_running() needs smp_processor_id() */
57 if (kprobes_built_in() && !user_mode_vm(regs)) {
59 if (kprobe_running() && kprobe_fault_handler(regs, 14))
72 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
73 * Check that here and ignore it.
77 * Sometimes the CPU reports invalid exceptions on prefetch.
78 * Check that here and ignore it.
80 * Opcode checker based on code by Richard Brunner.
83 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
84 unsigned char opcode, int *prefetch)
86 unsigned char instr_hi = opcode & 0xf0;
87 unsigned char instr_lo = opcode & 0x0f;
93 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
94 * In X86_64 long mode, the CPU will signal invalid
95 * opcode if some of these prefixes are present so
96 * X86_64 will never get here anyway
98 return ((instr_lo & 7) == 0x6);
102 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
103 * Need to figure out under what instruction mode the
104 * instruction was issued. Could check the LDT for lm,
105 * but for now it's good enough to assume that long
106 * mode only uses well known segments or kernel.
108 return (!user_mode(regs) || user_64bit_mode(regs));
111 /* 0x64 thru 0x67 are valid prefixes in all modes. */
112 return (instr_lo & 0xC) == 0x4;
114 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
115 return !instr_lo || (instr_lo>>1) == 1;
117 /* Prefetch instruction is 0x0F0D or 0x0F18 */
118 if (probe_kernel_address(instr, opcode))
121 *prefetch = (instr_lo == 0xF) &&
122 (opcode == 0x0D || opcode == 0x18);
130 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
132 unsigned char *max_instr;
133 unsigned char *instr;
137 * If it was a exec (instruction fetch) fault on NX page, then
138 * do not ignore the fault:
140 if (error_code & PF_INSTR)
143 instr = (void *)convert_ip_to_linear(current, regs);
144 max_instr = instr + 15;
146 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
149 while (instr < max_instr) {
150 unsigned char opcode;
152 if (probe_kernel_address(instr, opcode))
157 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
164 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
165 struct task_struct *tsk, int fault)
170 info.si_signo = si_signo;
172 info.si_code = si_code;
173 info.si_addr = (void __user *)address;
174 if (fault & VM_FAULT_HWPOISON_LARGE)
175 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
176 if (fault & VM_FAULT_HWPOISON)
178 info.si_addr_lsb = lsb;
180 force_sig_info(si_signo, &info, tsk);
183 DEFINE_SPINLOCK(pgd_lock);
187 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
189 unsigned index = pgd_index(address);
195 pgd_k = init_mm.pgd + index;
197 if (!pgd_present(*pgd_k))
201 * set_pgd(pgd, *pgd_k); here would be useless on PAE
202 * and redundant with the set_pmd() on non-PAE. As would
205 pud = pud_offset(pgd, address);
206 pud_k = pud_offset(pgd_k, address);
207 if (!pud_present(*pud_k))
210 pmd = pmd_offset(pud, address);
211 pmd_k = pmd_offset(pud_k, address);
212 if (!pmd_present(*pmd_k))
215 if (!pmd_present(*pmd))
216 set_pmd(pmd, *pmd_k);
218 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
223 void vmalloc_sync_all(void)
225 unsigned long address;
227 if (SHARED_KERNEL_PMD)
230 for (address = VMALLOC_START & PMD_MASK;
231 address >= TASK_SIZE && address < FIXADDR_TOP;
232 address += PMD_SIZE) {
235 spin_lock(&pgd_lock);
236 list_for_each_entry(page, &pgd_list, lru) {
237 spinlock_t *pgt_lock;
240 /* the pgt_lock only for Xen */
241 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
244 ret = vmalloc_sync_one(page_address(page), address);
245 spin_unlock(pgt_lock);
250 spin_unlock(&pgd_lock);
257 * Handle a fault on the vmalloc or module mapping area
259 static noinline __kprobes int vmalloc_fault(unsigned long address)
261 unsigned long pgd_paddr;
265 /* Make sure we are in vmalloc area: */
266 if (!(address >= VMALLOC_START && address < VMALLOC_END))
269 WARN_ON_ONCE(in_nmi());
272 * Synchronize this task's top level page-table
273 * with the 'reference' page table.
275 * Do _not_ use "current" here. We might be inside
276 * an interrupt in the middle of a task switch..
278 pgd_paddr = read_cr3();
279 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
283 pte_k = pte_offset_kernel(pmd_k, address);
284 if (!pte_present(*pte_k))
291 * Did it hit the DOS screen memory VA from vm86 mode?
294 check_v8086_mode(struct pt_regs *regs, unsigned long address,
295 struct task_struct *tsk)
299 if (!v8086_mode(regs))
302 bit = (address - 0xA0000) >> PAGE_SHIFT;
304 tsk->thread.screen_bitmap |= 1 << bit;
307 static bool low_pfn(unsigned long pfn)
309 return pfn < max_low_pfn;
312 static void dump_pagetable(unsigned long address)
314 pgd_t *base = __va(read_cr3());
315 pgd_t *pgd = &base[pgd_index(address)];
319 #ifdef CONFIG_X86_PAE
320 printk("*pdpt = %016Lx ", pgd_val(*pgd));
321 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
324 pmd = pmd_offset(pud_offset(pgd, address), address);
325 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
328 * We must not directly access the pte in the highpte
329 * case if the page table is located in highmem.
330 * And let's rather not kmap-atomic the pte, just in case
331 * it's allocated already:
333 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
336 pte = pte_offset_kernel(pmd, address);
337 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
342 #else /* CONFIG_X86_64: */
344 void vmalloc_sync_all(void)
346 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
352 * Handle a fault on the vmalloc area
354 * This assumes no large pages in there.
356 static noinline __kprobes int vmalloc_fault(unsigned long address)
358 pgd_t *pgd, *pgd_ref;
359 pud_t *pud, *pud_ref;
360 pmd_t *pmd, *pmd_ref;
361 pte_t *pte, *pte_ref;
363 /* Make sure we are in vmalloc area: */
364 if (!(address >= VMALLOC_START && address < VMALLOC_END))
367 WARN_ON_ONCE(in_nmi());
370 * Copy kernel mappings over when needed. This can also
371 * happen within a race in page table update. In the later
374 pgd = pgd_offset(current->active_mm, address);
375 pgd_ref = pgd_offset_k(address);
376 if (pgd_none(*pgd_ref))
380 set_pgd(pgd, *pgd_ref);
382 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
385 * Below here mismatches are bugs because these lower tables
389 pud = pud_offset(pgd, address);
390 pud_ref = pud_offset(pgd_ref, address);
391 if (pud_none(*pud_ref))
394 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
397 pmd = pmd_offset(pud, address);
398 pmd_ref = pmd_offset(pud_ref, address);
399 if (pmd_none(*pmd_ref))
402 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
405 pte_ref = pte_offset_kernel(pmd_ref, address);
406 if (!pte_present(*pte_ref))
409 pte = pte_offset_kernel(pmd, address);
412 * Don't use pte_page here, because the mappings can point
413 * outside mem_map, and the NUMA hash lookup cannot handle
416 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
422 static const char errata93_warning[] =
424 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
425 "******* Working around it, but it may cause SEGVs or burn power.\n"
426 "******* Please consider a BIOS update.\n"
427 "******* Disabling USB legacy in the BIOS may also help.\n";
430 * No vm86 mode in 64-bit mode:
433 check_v8086_mode(struct pt_regs *regs, unsigned long address,
434 struct task_struct *tsk)
438 static int bad_address(void *p)
442 return probe_kernel_address((unsigned long *)p, dummy);
445 static void dump_pagetable(unsigned long address)
447 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
448 pgd_t *pgd = base + pgd_index(address);
453 if (bad_address(pgd))
456 printk("PGD %lx ", pgd_val(*pgd));
458 if (!pgd_present(*pgd))
461 pud = pud_offset(pgd, address);
462 if (bad_address(pud))
465 printk("PUD %lx ", pud_val(*pud));
466 if (!pud_present(*pud) || pud_large(*pud))
469 pmd = pmd_offset(pud, address);
470 if (bad_address(pmd))
473 printk("PMD %lx ", pmd_val(*pmd));
474 if (!pmd_present(*pmd) || pmd_large(*pmd))
477 pte = pte_offset_kernel(pmd, address);
478 if (bad_address(pte))
481 printk("PTE %lx", pte_val(*pte));
489 #endif /* CONFIG_X86_64 */
492 * Workaround for K8 erratum #93 & buggy BIOS.
494 * BIOS SMM functions are required to use a specific workaround
495 * to avoid corruption of the 64bit RIP register on C stepping K8.
497 * A lot of BIOS that didn't get tested properly miss this.
499 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
500 * Try to work around it here.
502 * Note we only handle faults in kernel here.
503 * Does nothing on 32-bit.
505 static int is_errata93(struct pt_regs *regs, unsigned long address)
508 if (address != regs->ip)
511 if ((address >> 32) != 0)
514 address |= 0xffffffffUL << 32;
515 if ((address >= (u64)_stext && address <= (u64)_etext) ||
516 (address >= MODULES_VADDR && address <= MODULES_END)) {
517 printk_once(errata93_warning);
526 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
527 * to illegal addresses >4GB.
529 * We catch this in the page fault handler because these addresses
530 * are not reachable. Just detect this case and return. Any code
531 * segment in LDT is compatibility mode.
533 static int is_errata100(struct pt_regs *regs, unsigned long address)
536 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
542 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
544 #ifdef CONFIG_X86_F00F_BUG
548 * Pentium F0 0F C7 C8 bug workaround:
550 if (boot_cpu_data.f00f_bug) {
551 nr = (address - idt_descr.address) >> 3;
554 do_invalid_op(regs, 0);
562 static const char nx_warning[] = KERN_CRIT
563 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
566 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
567 unsigned long address)
569 if (!oops_may_print())
572 if (error_code & PF_INSTR) {
575 pte_t *pte = lookup_address(address, &level);
577 if (pte && pte_present(*pte) && !pte_exec(*pte))
578 printk(nx_warning, current_uid());
581 printk(KERN_ALERT "BUG: unable to handle kernel ");
582 if (address < PAGE_SIZE)
583 printk(KERN_CONT "NULL pointer dereference");
585 printk(KERN_CONT "paging request");
587 printk(KERN_CONT " at %p\n", (void *) address);
588 printk(KERN_ALERT "IP:");
589 printk_address(regs->ip, 1);
591 dump_pagetable(address);
595 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
596 unsigned long address)
598 struct task_struct *tsk;
602 flags = oops_begin();
606 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
608 dump_pagetable(address);
610 tsk->thread.cr2 = address;
611 tsk->thread.trap_no = 14;
612 tsk->thread.error_code = error_code;
614 if (__die("Bad pagetable", regs, error_code))
617 oops_end(flags, regs, sig);
621 no_context(struct pt_regs *regs, unsigned long error_code,
622 unsigned long address)
624 struct task_struct *tsk = current;
625 unsigned long *stackend;
629 /* Are we prepared to handle this kernel fault? */
630 if (fixup_exception(regs))
636 * Valid to do another page fault here, because if this fault
637 * had been triggered by is_prefetch fixup_exception would have
642 * Hall of shame of CPU/BIOS bugs.
644 if (is_prefetch(regs, error_code, address))
647 if (is_errata93(regs, address))
651 * Oops. The kernel tried to access some bad page. We'll have to
652 * terminate things with extreme prejudice:
654 flags = oops_begin();
656 show_fault_oops(regs, error_code, address);
658 stackend = end_of_stack(tsk);
659 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
660 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
662 tsk->thread.cr2 = address;
663 tsk->thread.trap_no = 14;
664 tsk->thread.error_code = error_code;
667 if (__die("Oops", regs, error_code))
670 /* Executive summary in case the body of the oops scrolled away */
671 printk(KERN_EMERG "CR2: %016lx\n", address);
673 oops_end(flags, regs, sig);
677 * Print out info about fatal segfaults, if the show_unhandled_signals
681 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
682 unsigned long address, struct task_struct *tsk)
684 if (!unhandled_signal(tsk, SIGSEGV))
687 if (!printk_ratelimit())
690 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
691 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
692 tsk->comm, task_pid_nr(tsk), address,
693 (void *)regs->ip, (void *)regs->sp, error_code);
695 print_vma_addr(KERN_CONT " in ", regs->ip);
697 printk(KERN_CONT "\n");
701 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
702 unsigned long address, int si_code)
704 struct task_struct *tsk = current;
706 /* User mode accesses just cause a SIGSEGV */
707 if (error_code & PF_USER) {
709 * It's possible to have interrupts off here:
714 * Valid to do another page fault here because this one came
717 if (is_prefetch(regs, error_code, address))
720 if (is_errata100(regs, address))
725 * Instruction fetch faults in the vsyscall page might need
728 if (unlikely((error_code & PF_INSTR) &&
729 ((address & ~0xfff) == VSYSCALL_START))) {
730 if (emulate_vsyscall(regs, address))
735 if (unlikely(show_unhandled_signals))
736 show_signal_msg(regs, error_code, address, tsk);
738 /* Kernel addresses are always protection faults: */
739 tsk->thread.cr2 = address;
740 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
741 tsk->thread.trap_no = 14;
743 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
748 if (is_f00f_bug(regs, address))
751 no_context(regs, error_code, address);
755 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
756 unsigned long address)
758 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
762 __bad_area(struct pt_regs *regs, unsigned long error_code,
763 unsigned long address, int si_code)
765 struct mm_struct *mm = current->mm;
768 * Something tried to access memory that isn't in our memory map..
769 * Fix it, but check if it's kernel or user first..
771 up_read(&mm->mmap_sem);
773 __bad_area_nosemaphore(regs, error_code, address, si_code);
777 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
779 __bad_area(regs, error_code, address, SEGV_MAPERR);
783 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
784 unsigned long address)
786 __bad_area(regs, error_code, address, SEGV_ACCERR);
789 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
791 out_of_memory(struct pt_regs *regs, unsigned long error_code,
792 unsigned long address)
795 * We ran out of memory, call the OOM killer, and return the userspace
796 * (which will retry the fault, or kill us if we got oom-killed):
798 up_read(¤t->mm->mmap_sem);
800 pagefault_out_of_memory();
804 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
807 struct task_struct *tsk = current;
808 struct mm_struct *mm = tsk->mm;
809 int code = BUS_ADRERR;
811 up_read(&mm->mmap_sem);
813 /* Kernel mode? Handle exceptions or die: */
814 if (!(error_code & PF_USER)) {
815 no_context(regs, error_code, address);
819 /* User-space => ok to do another page fault: */
820 if (is_prefetch(regs, error_code, address))
823 tsk->thread.cr2 = address;
824 tsk->thread.error_code = error_code;
825 tsk->thread.trap_no = 14;
827 #ifdef CONFIG_MEMORY_FAILURE
828 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
830 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
831 tsk->comm, tsk->pid, address);
832 code = BUS_MCEERR_AR;
835 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
839 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
840 unsigned long address, unsigned int fault)
843 * Pagefault was interrupted by SIGKILL. We have no reason to
844 * continue pagefault.
846 if (fatal_signal_pending(current)) {
847 if (!(fault & VM_FAULT_RETRY))
848 up_read(¤t->mm->mmap_sem);
849 if (!(error_code & PF_USER))
850 no_context(regs, error_code, address);
853 if (!(fault & VM_FAULT_ERROR))
856 if (fault & VM_FAULT_OOM) {
857 /* Kernel mode? Handle exceptions or die: */
858 if (!(error_code & PF_USER)) {
859 up_read(¤t->mm->mmap_sem);
860 no_context(regs, error_code, address);
864 out_of_memory(regs, error_code, address);
866 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
867 VM_FAULT_HWPOISON_LARGE))
868 do_sigbus(regs, error_code, address, fault);
875 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
877 if ((error_code & PF_WRITE) && !pte_write(*pte))
880 if ((error_code & PF_INSTR) && !pte_exec(*pte))
887 * Handle a spurious fault caused by a stale TLB entry.
889 * This allows us to lazily refresh the TLB when increasing the
890 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
891 * eagerly is very expensive since that implies doing a full
892 * cross-processor TLB flush, even if no stale TLB entries exist
893 * on other processors.
895 * There are no security implications to leaving a stale TLB when
896 * increasing the permissions on a page.
898 static noinline __kprobes int
899 spurious_fault(unsigned long error_code, unsigned long address)
907 /* Reserved-bit violation or user access to kernel space? */
908 if (error_code & (PF_USER | PF_RSVD))
911 pgd = init_mm.pgd + pgd_index(address);
912 if (!pgd_present(*pgd))
915 pud = pud_offset(pgd, address);
916 if (!pud_present(*pud))
920 return spurious_fault_check(error_code, (pte_t *) pud);
922 pmd = pmd_offset(pud, address);
923 if (!pmd_present(*pmd))
927 return spurious_fault_check(error_code, (pte_t *) pmd);
930 * Note: don't use pte_present() here, since it returns true
931 * if the _PAGE_PROTNONE bit is set. However, this aliases the
932 * _PAGE_GLOBAL bit, which for kernel pages give false positives
933 * when CONFIG_DEBUG_PAGEALLOC is used.
935 pte = pte_offset_kernel(pmd, address);
936 if (!(pte_flags(*pte) & _PAGE_PRESENT))
939 ret = spurious_fault_check(error_code, pte);
944 * Make sure we have permissions in PMD.
945 * If not, then there's a bug in the page tables:
947 ret = spurious_fault_check(error_code, (pte_t *) pmd);
948 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
953 int show_unhandled_signals = 1;
956 access_error(unsigned long error_code, struct vm_area_struct *vma)
958 if (error_code & PF_WRITE) {
959 /* write, present and write, not present: */
960 if (unlikely(!(vma->vm_flags & VM_WRITE)))
966 if (unlikely(error_code & PF_PROT))
969 /* read, not present: */
970 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
976 static int fault_in_kernel_space(unsigned long address)
978 return address >= TASK_SIZE_MAX;
982 * This routine handles page faults. It determines the address,
983 * and the problem, and then passes it off to one of the appropriate
986 dotraplinkage void __kprobes
987 do_page_fault(struct pt_regs *regs, unsigned long error_code)
989 struct vm_area_struct *vma;
990 struct task_struct *tsk;
991 unsigned long address;
992 struct mm_struct *mm;
994 int write = error_code & PF_WRITE;
995 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE |
996 (write ? FAULT_FLAG_WRITE : 0);
1001 /* Get the faulting address: */
1002 address = read_cr2();
1005 * Detect and handle instructions that would cause a page fault for
1006 * both a tracked kernel page and a userspace page.
1008 if (kmemcheck_active(regs))
1009 kmemcheck_hide(regs);
1010 prefetchw(&mm->mmap_sem);
1012 if (unlikely(kmmio_fault(regs, address)))
1016 * We fault-in kernel-space virtual memory on-demand. The
1017 * 'reference' page table is init_mm.pgd.
1019 * NOTE! We MUST NOT take any locks for this case. We may
1020 * be in an interrupt or a critical region, and should
1021 * only copy the information from the master page table,
1024 * This verifies that the fault happens in kernel space
1025 * (error_code & 4) == 0, and that the fault was not a
1026 * protection error (error_code & 9) == 0.
1028 if (unlikely(fault_in_kernel_space(address))) {
1029 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1030 if (vmalloc_fault(address) >= 0)
1033 if (kmemcheck_fault(regs, address, error_code))
1037 /* Can handle a stale RO->RW TLB: */
1038 if (spurious_fault(error_code, address))
1041 /* kprobes don't want to hook the spurious faults: */
1042 if (notify_page_fault(regs))
1045 * Don't take the mm semaphore here. If we fixup a prefetch
1046 * fault we could otherwise deadlock:
1048 bad_area_nosemaphore(regs, error_code, address);
1053 /* kprobes don't want to hook the spurious faults: */
1054 if (unlikely(notify_page_fault(regs)))
1057 * It's safe to allow irq's after cr2 has been saved and the
1058 * vmalloc fault has been handled.
1060 * User-mode registers count as a user access even for any
1061 * potential system fault or CPU buglet:
1063 if (user_mode_vm(regs)) {
1065 error_code |= PF_USER;
1067 if (regs->flags & X86_EFLAGS_IF)
1071 if (unlikely(error_code & PF_RSVD))
1072 pgtable_bad(regs, error_code, address);
1074 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1077 * If we're in an interrupt, have no user context or are running
1078 * in an atomic region then we must not take the fault:
1080 if (unlikely(in_atomic() || !mm)) {
1081 bad_area_nosemaphore(regs, error_code, address);
1086 * When running in the kernel we expect faults to occur only to
1087 * addresses in user space. All other faults represent errors in
1088 * the kernel and should generate an OOPS. Unfortunately, in the
1089 * case of an erroneous fault occurring in a code path which already
1090 * holds mmap_sem we will deadlock attempting to validate the fault
1091 * against the address space. Luckily the kernel only validly
1092 * references user space from well defined areas of code, which are
1093 * listed in the exceptions table.
1095 * As the vast majority of faults will be valid we will only perform
1096 * the source reference check when there is a possibility of a
1097 * deadlock. Attempt to lock the address space, if we cannot we then
1098 * validate the source. If this is invalid we can skip the address
1099 * space check, thus avoiding the deadlock:
1101 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1102 if ((error_code & PF_USER) == 0 &&
1103 !search_exception_tables(regs->ip)) {
1104 bad_area_nosemaphore(regs, error_code, address);
1108 down_read(&mm->mmap_sem);
1111 * The above down_read_trylock() might have succeeded in
1112 * which case we'll have missed the might_sleep() from
1118 vma = find_vma(mm, address);
1119 if (unlikely(!vma)) {
1120 bad_area(regs, error_code, address);
1123 if (likely(vma->vm_start <= address))
1125 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1126 bad_area(regs, error_code, address);
1129 if (error_code & PF_USER) {
1131 * Accessing the stack below %sp is always a bug.
1132 * The large cushion allows instructions like enter
1133 * and pusha to work. ("enter $65535, $31" pushes
1134 * 32 pointers and then decrements %sp by 65535.)
1136 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1137 bad_area(regs, error_code, address);
1141 if (unlikely(expand_stack(vma, address))) {
1142 bad_area(regs, error_code, address);
1147 * Ok, we have a good vm_area for this memory access, so
1148 * we can handle it..
1151 if (unlikely(access_error(error_code, vma))) {
1152 bad_area_access_error(regs, error_code, address);
1157 * If for any reason at all we couldn't handle the fault,
1158 * make sure we exit gracefully rather than endlessly redo
1161 fault = handle_mm_fault(mm, vma, address, flags);
1163 if (unlikely(fault & (VM_FAULT_RETRY|VM_FAULT_ERROR))) {
1164 if (mm_fault_error(regs, error_code, address, fault))
1169 * Major/minor page fault accounting is only done on the
1170 * initial attempt. If we go through a retry, it is extremely
1171 * likely that the page will be found in page cache at that point.
1173 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1174 if (fault & VM_FAULT_MAJOR) {
1176 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1180 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1183 if (fault & VM_FAULT_RETRY) {
1184 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1186 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1191 check_v8086_mode(regs, address, tsk);
1193 up_read(&mm->mmap_sem);