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_*(), ... */
20 #include <asm/fixmap.h> /* VSYSCALL_START */
23 * Page fault error code bits:
25 * bit 0 == 0: no page found 1: protection fault
26 * bit 1 == 0: read access 1: write access
27 * bit 2 == 0: kernel-mode access 1: user-mode access
28 * bit 3 == 1: use of reserved bit detected
29 * bit 4 == 1: fault was an instruction fetch
31 enum x86_pf_error_code {
41 * Returns 0 if mmiotrace is disabled, or if the fault is not
42 * handled by mmiotrace:
44 static inline int __kprobes
45 kmmio_fault(struct pt_regs *regs, unsigned long addr)
47 if (unlikely(is_kmmio_active()))
48 if (kmmio_handler(regs, addr) == 1)
53 static inline int __kprobes notify_page_fault(struct pt_regs *regs)
57 /* kprobe_running() needs smp_processor_id() */
58 if (kprobes_built_in() && !user_mode_vm(regs)) {
60 if (kprobe_running() && kprobe_fault_handler(regs, 14))
73 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
74 * Check that here and ignore it.
78 * Sometimes the CPU reports invalid exceptions on prefetch.
79 * Check that here and ignore it.
81 * Opcode checker based on code by Richard Brunner.
84 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
85 unsigned char opcode, int *prefetch)
87 unsigned char instr_hi = opcode & 0xf0;
88 unsigned char instr_lo = opcode & 0x0f;
94 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
95 * In X86_64 long mode, the CPU will signal invalid
96 * opcode if some of these prefixes are present so
97 * X86_64 will never get here anyway
99 return ((instr_lo & 7) == 0x6);
103 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
104 * Need to figure out under what instruction mode the
105 * instruction was issued. Could check the LDT for lm,
106 * but for now it's good enough to assume that long
107 * mode only uses well known segments or kernel.
109 return (!user_mode(regs) || user_64bit_mode(regs));
112 /* 0x64 thru 0x67 are valid prefixes in all modes. */
113 return (instr_lo & 0xC) == 0x4;
115 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
116 return !instr_lo || (instr_lo>>1) == 1;
118 /* Prefetch instruction is 0x0F0D or 0x0F18 */
119 if (probe_kernel_address(instr, opcode))
122 *prefetch = (instr_lo == 0xF) &&
123 (opcode == 0x0D || opcode == 0x18);
131 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
133 unsigned char *max_instr;
134 unsigned char *instr;
138 * If it was a exec (instruction fetch) fault on NX page, then
139 * do not ignore the fault:
141 if (error_code & PF_INSTR)
144 instr = (void *)convert_ip_to_linear(current, regs);
145 max_instr = instr + 15;
147 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
150 while (instr < max_instr) {
151 unsigned char opcode;
153 if (probe_kernel_address(instr, opcode))
158 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
165 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
166 struct task_struct *tsk, int fault)
171 info.si_signo = si_signo;
173 info.si_code = si_code;
174 info.si_addr = (void __user *)address;
175 if (fault & VM_FAULT_HWPOISON_LARGE)
176 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
177 if (fault & VM_FAULT_HWPOISON)
179 info.si_addr_lsb = lsb;
181 force_sig_info(si_signo, &info, tsk);
184 DEFINE_SPINLOCK(pgd_lock);
188 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
190 unsigned index = pgd_index(address);
196 pgd_k = init_mm.pgd + index;
198 if (!pgd_present(*pgd_k))
202 * set_pgd(pgd, *pgd_k); here would be useless on PAE
203 * and redundant with the set_pmd() on non-PAE. As would
206 pud = pud_offset(pgd, address);
207 pud_k = pud_offset(pgd_k, address);
208 if (!pud_present(*pud_k))
211 pmd = pmd_offset(pud, address);
212 pmd_k = pmd_offset(pud_k, address);
213 if (!pmd_present(*pmd_k))
216 if (!pmd_present(*pmd))
217 set_pmd(pmd, *pmd_k);
219 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
224 void vmalloc_sync_all(void)
226 unsigned long address;
228 if (SHARED_KERNEL_PMD)
231 for (address = VMALLOC_START & PMD_MASK;
232 address >= TASK_SIZE && address < FIXADDR_TOP;
233 address += PMD_SIZE) {
236 spin_lock(&pgd_lock);
237 list_for_each_entry(page, &pgd_list, lru) {
238 spinlock_t *pgt_lock;
241 /* the pgt_lock only for Xen */
242 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
245 ret = vmalloc_sync_one(page_address(page), address);
246 spin_unlock(pgt_lock);
251 spin_unlock(&pgd_lock);
258 * Handle a fault on the vmalloc or module mapping area
260 static noinline __kprobes int vmalloc_fault(unsigned long address)
262 unsigned long pgd_paddr;
266 /* Make sure we are in vmalloc area: */
267 if (!(address >= VMALLOC_START && address < VMALLOC_END))
270 WARN_ON_ONCE(in_nmi());
273 * Synchronize this task's top level page-table
274 * with the 'reference' page table.
276 * Do _not_ use "current" here. We might be inside
277 * an interrupt in the middle of a task switch..
279 pgd_paddr = read_cr3();
280 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
284 pte_k = pte_offset_kernel(pmd_k, address);
285 if (!pte_present(*pte_k))
292 * Did it hit the DOS screen memory VA from vm86 mode?
295 check_v8086_mode(struct pt_regs *regs, unsigned long address,
296 struct task_struct *tsk)
300 if (!v8086_mode(regs))
303 bit = (address - 0xA0000) >> PAGE_SHIFT;
305 tsk->thread.screen_bitmap |= 1 << bit;
308 static bool low_pfn(unsigned long pfn)
310 return pfn < max_low_pfn;
313 static void dump_pagetable(unsigned long address)
315 pgd_t *base = __va(read_cr3());
316 pgd_t *pgd = &base[pgd_index(address)];
320 #ifdef CONFIG_X86_PAE
321 printk("*pdpt = %016Lx ", pgd_val(*pgd));
322 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
325 pmd = pmd_offset(pud_offset(pgd, address), address);
326 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
329 * We must not directly access the pte in the highpte
330 * case if the page table is located in highmem.
331 * And let's rather not kmap-atomic the pte, just in case
332 * it's allocated already:
334 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
337 pte = pte_offset_kernel(pmd, address);
338 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
343 #else /* CONFIG_X86_64: */
345 void vmalloc_sync_all(void)
347 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
353 * Handle a fault on the vmalloc area
355 * This assumes no large pages in there.
357 static noinline __kprobes int vmalloc_fault(unsigned long address)
359 pgd_t *pgd, *pgd_ref;
360 pud_t *pud, *pud_ref;
361 pmd_t *pmd, *pmd_ref;
362 pte_t *pte, *pte_ref;
364 /* Make sure we are in vmalloc area: */
365 if (!(address >= VMALLOC_START && address < VMALLOC_END))
368 WARN_ON_ONCE(in_nmi());
371 * Copy kernel mappings over when needed. This can also
372 * happen within a race in page table update. In the later
375 pgd = pgd_offset(current->active_mm, address);
376 pgd_ref = pgd_offset_k(address);
377 if (pgd_none(*pgd_ref))
380 if (pgd_none(*pgd)) {
381 set_pgd(pgd, *pgd_ref);
382 arch_flush_lazy_mmu_mode();
384 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
388 * Below here mismatches are bugs because these lower tables
392 pud = pud_offset(pgd, address);
393 pud_ref = pud_offset(pgd_ref, address);
394 if (pud_none(*pud_ref))
397 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
400 pmd = pmd_offset(pud, address);
401 pmd_ref = pmd_offset(pud_ref, address);
402 if (pmd_none(*pmd_ref))
405 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
408 pte_ref = pte_offset_kernel(pmd_ref, address);
409 if (!pte_present(*pte_ref))
412 pte = pte_offset_kernel(pmd, address);
415 * Don't use pte_page here, because the mappings can point
416 * outside mem_map, and the NUMA hash lookup cannot handle
419 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
425 #ifdef CONFIG_CPU_SUP_AMD
426 static const char errata93_warning[] =
428 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
429 "******* Working around it, but it may cause SEGVs or burn power.\n"
430 "******* Please consider a BIOS update.\n"
431 "******* Disabling USB legacy in the BIOS may also help.\n";
435 * No vm86 mode in 64-bit mode:
438 check_v8086_mode(struct pt_regs *regs, unsigned long address,
439 struct task_struct *tsk)
443 static int bad_address(void *p)
447 return probe_kernel_address((unsigned long *)p, dummy);
450 static void dump_pagetable(unsigned long address)
452 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
453 pgd_t *pgd = base + pgd_index(address);
458 if (bad_address(pgd))
461 printk("PGD %lx ", pgd_val(*pgd));
463 if (!pgd_present(*pgd))
466 pud = pud_offset(pgd, address);
467 if (bad_address(pud))
470 printk("PUD %lx ", pud_val(*pud));
471 if (!pud_present(*pud) || pud_large(*pud))
474 pmd = pmd_offset(pud, address);
475 if (bad_address(pmd))
478 printk("PMD %lx ", pmd_val(*pmd));
479 if (!pmd_present(*pmd) || pmd_large(*pmd))
482 pte = pte_offset_kernel(pmd, address);
483 if (bad_address(pte))
486 printk("PTE %lx", pte_val(*pte));
494 #endif /* CONFIG_X86_64 */
497 * Workaround for K8 erratum #93 & buggy BIOS.
499 * BIOS SMM functions are required to use a specific workaround
500 * to avoid corruption of the 64bit RIP register on C stepping K8.
502 * A lot of BIOS that didn't get tested properly miss this.
504 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
505 * Try to work around it here.
507 * Note we only handle faults in kernel here.
508 * Does nothing on 32-bit.
510 static int is_errata93(struct pt_regs *regs, unsigned long address)
512 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
513 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
514 || boot_cpu_data.x86 != 0xf)
517 if (address != regs->ip)
520 if ((address >> 32) != 0)
523 address |= 0xffffffffUL << 32;
524 if ((address >= (u64)_stext && address <= (u64)_etext) ||
525 (address >= MODULES_VADDR && address <= MODULES_END)) {
526 printk_once(errata93_warning);
535 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
536 * to illegal addresses >4GB.
538 * We catch this in the page fault handler because these addresses
539 * are not reachable. Just detect this case and return. Any code
540 * segment in LDT is compatibility mode.
542 static int is_errata100(struct pt_regs *regs, unsigned long address)
545 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
551 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
553 #ifdef CONFIG_X86_F00F_BUG
557 * Pentium F0 0F C7 C8 bug workaround:
559 if (boot_cpu_data.f00f_bug) {
560 nr = (address - idt_descr.address) >> 3;
563 do_invalid_op(regs, 0);
571 static const char nx_warning[] = KERN_CRIT
572 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
575 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
576 unsigned long address)
578 if (!oops_may_print())
581 if (error_code & PF_INSTR) {
584 pte_t *pte = lookup_address(address, &level);
586 if (pte && pte_present(*pte) && !pte_exec(*pte))
587 printk(nx_warning, current_uid());
590 printk(KERN_ALERT "BUG: unable to handle kernel ");
591 if (address < PAGE_SIZE)
592 printk(KERN_CONT "NULL pointer dereference");
594 printk(KERN_CONT "paging request");
596 printk(KERN_CONT " at %p\n", (void *) address);
597 printk(KERN_ALERT "IP:");
598 printk_address(regs->ip, 1);
600 dump_pagetable(address);
604 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
605 unsigned long address)
607 struct task_struct *tsk;
611 flags = oops_begin();
615 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
617 dump_pagetable(address);
619 tsk->thread.cr2 = address;
620 tsk->thread.trap_no = 14;
621 tsk->thread.error_code = error_code;
623 if (__die("Bad pagetable", regs, error_code))
626 oops_end(flags, regs, sig);
630 no_context(struct pt_regs *regs, unsigned long error_code,
631 unsigned long address)
633 struct task_struct *tsk = current;
634 unsigned long *stackend;
638 /* Are we prepared to handle this kernel fault? */
639 if (fixup_exception(regs))
645 * Valid to do another page fault here, because if this fault
646 * had been triggered by is_prefetch fixup_exception would have
651 * Hall of shame of CPU/BIOS bugs.
653 if (is_prefetch(regs, error_code, address))
656 if (is_errata93(regs, address))
660 * Oops. The kernel tried to access some bad page. We'll have to
661 * terminate things with extreme prejudice:
663 flags = oops_begin();
665 show_fault_oops(regs, error_code, address);
667 stackend = end_of_stack(tsk);
668 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
669 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
671 tsk->thread.cr2 = address;
672 tsk->thread.trap_no = 14;
673 tsk->thread.error_code = error_code;
676 if (__die("Oops", regs, error_code))
679 /* Executive summary in case the body of the oops scrolled away */
680 printk(KERN_EMERG "CR2: %016lx\n", address);
682 oops_end(flags, regs, sig);
686 * Print out info about fatal segfaults, if the show_unhandled_signals
690 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
691 unsigned long address, struct task_struct *tsk)
693 if (!unhandled_signal(tsk, SIGSEGV))
696 if (!printk_ratelimit())
699 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
700 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
701 tsk->comm, task_pid_nr(tsk), address,
702 (void *)regs->ip, (void *)regs->sp, error_code);
704 print_vma_addr(KERN_CONT " in ", regs->ip);
706 printk(KERN_CONT "\n");
710 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
711 unsigned long address, int si_code)
713 struct task_struct *tsk = current;
715 /* User mode accesses just cause a SIGSEGV */
716 if (error_code & PF_USER) {
718 * It's possible to have interrupts off here:
723 * Valid to do another page fault here because this one came
726 if (is_prefetch(regs, error_code, address))
729 if (is_errata100(regs, address))
734 * Instruction fetch faults in the vsyscall page might need
737 if (unlikely((error_code & PF_INSTR) &&
738 ((address & ~0xfff) == VSYSCALL_START))) {
739 if (emulate_vsyscall(regs, address))
743 /* Kernel addresses are always protection faults: */
744 if (address >= TASK_SIZE)
745 error_code |= PF_PROT;
747 if (likely(show_unhandled_signals))
748 show_signal_msg(regs, error_code, address, tsk);
750 tsk->thread.cr2 = address;
751 tsk->thread.error_code = error_code;
752 tsk->thread.trap_no = 14;
754 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
759 if (is_f00f_bug(regs, address))
762 no_context(regs, error_code, address);
766 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
767 unsigned long address)
769 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
773 __bad_area(struct pt_regs *regs, unsigned long error_code,
774 unsigned long address, int si_code)
776 struct mm_struct *mm = current->mm;
779 * Something tried to access memory that isn't in our memory map..
780 * Fix it, but check if it's kernel or user first..
782 up_read(&mm->mmap_sem);
784 __bad_area_nosemaphore(regs, error_code, address, si_code);
788 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
790 __bad_area(regs, error_code, address, SEGV_MAPERR);
794 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
795 unsigned long address)
797 __bad_area(regs, error_code, address, SEGV_ACCERR);
800 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
802 out_of_memory(struct pt_regs *regs, unsigned long error_code,
803 unsigned long address)
806 * We ran out of memory, call the OOM killer, and return the userspace
807 * (which will retry the fault, or kill us if we got oom-killed):
809 up_read(¤t->mm->mmap_sem);
811 pagefault_out_of_memory();
815 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
818 struct task_struct *tsk = current;
819 struct mm_struct *mm = tsk->mm;
820 int code = BUS_ADRERR;
822 up_read(&mm->mmap_sem);
824 /* Kernel mode? Handle exceptions or die: */
825 if (!(error_code & PF_USER)) {
826 no_context(regs, error_code, address);
830 /* User-space => ok to do another page fault: */
831 if (is_prefetch(regs, error_code, address))
834 tsk->thread.cr2 = address;
835 tsk->thread.error_code = error_code;
836 tsk->thread.trap_no = 14;
838 #ifdef CONFIG_MEMORY_FAILURE
839 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
841 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
842 tsk->comm, tsk->pid, address);
843 code = BUS_MCEERR_AR;
846 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
850 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
851 unsigned long address, unsigned int fault)
854 * Pagefault was interrupted by SIGKILL. We have no reason to
855 * continue pagefault.
857 if (fatal_signal_pending(current)) {
858 if (!(fault & VM_FAULT_RETRY))
859 up_read(¤t->mm->mmap_sem);
860 if (!(error_code & PF_USER))
861 no_context(regs, error_code, address);
864 if (!(fault & VM_FAULT_ERROR))
867 if (fault & VM_FAULT_OOM) {
868 /* Kernel mode? Handle exceptions or die: */
869 if (!(error_code & PF_USER)) {
870 up_read(¤t->mm->mmap_sem);
871 no_context(regs, error_code, address);
875 out_of_memory(regs, error_code, address);
877 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
878 VM_FAULT_HWPOISON_LARGE))
879 do_sigbus(regs, error_code, address, fault);
886 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
888 if ((error_code & PF_WRITE) && !pte_write(*pte))
891 if ((error_code & PF_INSTR) && !pte_exec(*pte))
898 * Handle a spurious fault caused by a stale TLB entry.
900 * This allows us to lazily refresh the TLB when increasing the
901 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
902 * eagerly is very expensive since that implies doing a full
903 * cross-processor TLB flush, even if no stale TLB entries exist
904 * on other processors.
906 * There are no security implications to leaving a stale TLB when
907 * increasing the permissions on a page.
909 static noinline __kprobes int
910 spurious_fault(unsigned long error_code, unsigned long address)
918 /* Reserved-bit violation or user access to kernel space? */
919 if (error_code & (PF_USER | PF_RSVD))
922 pgd = init_mm.pgd + pgd_index(address);
923 if (!pgd_present(*pgd))
926 pud = pud_offset(pgd, address);
927 if (!pud_present(*pud))
931 return spurious_fault_check(error_code, (pte_t *) pud);
933 pmd = pmd_offset(pud, address);
934 if (!pmd_present(*pmd))
938 return spurious_fault_check(error_code, (pte_t *) pmd);
941 * Note: don't use pte_present() here, since it returns true
942 * if the _PAGE_PROTNONE bit is set. However, this aliases the
943 * _PAGE_GLOBAL bit, which for kernel pages give false positives
944 * when CONFIG_DEBUG_PAGEALLOC is used.
946 pte = pte_offset_kernel(pmd, address);
947 if (!(pte_flags(*pte) & _PAGE_PRESENT))
950 ret = spurious_fault_check(error_code, pte);
955 * Make sure we have permissions in PMD.
956 * If not, then there's a bug in the page tables:
958 ret = spurious_fault_check(error_code, (pte_t *) pmd);
959 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
964 int show_unhandled_signals = 1;
967 access_error(unsigned long error_code, struct vm_area_struct *vma)
969 if (error_code & PF_WRITE) {
970 /* write, present and write, not present: */
971 if (unlikely(!(vma->vm_flags & VM_WRITE)))
977 if (unlikely(error_code & PF_PROT))
980 /* read, not present: */
981 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
987 static int fault_in_kernel_space(unsigned long address)
989 return address >= TASK_SIZE_MAX;
993 * This routine handles page faults. It determines the address,
994 * and the problem, and then passes it off to one of the appropriate
997 dotraplinkage void __kprobes
998 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1000 struct vm_area_struct *vma;
1001 struct task_struct *tsk;
1002 unsigned long address;
1003 struct mm_struct *mm;
1005 int write = error_code & PF_WRITE;
1006 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE |
1007 (write ? FAULT_FLAG_WRITE : 0);
1012 /* Get the faulting address: */
1013 address = read_cr2();
1016 * Detect and handle instructions that would cause a page fault for
1017 * both a tracked kernel page and a userspace page.
1019 if (kmemcheck_active(regs))
1020 kmemcheck_hide(regs);
1021 prefetchw(&mm->mmap_sem);
1023 if (unlikely(kmmio_fault(regs, address)))
1027 * We fault-in kernel-space virtual memory on-demand. The
1028 * 'reference' page table is init_mm.pgd.
1030 * NOTE! We MUST NOT take any locks for this case. We may
1031 * be in an interrupt or a critical region, and should
1032 * only copy the information from the master page table,
1035 * This verifies that the fault happens in kernel space
1036 * (error_code & 4) == 0, and that the fault was not a
1037 * protection error (error_code & 9) == 0.
1039 if (unlikely(fault_in_kernel_space(address))) {
1040 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1041 if (vmalloc_fault(address) >= 0)
1044 if (kmemcheck_fault(regs, address, error_code))
1048 /* Can handle a stale RO->RW TLB: */
1049 if (spurious_fault(error_code, address))
1052 /* kprobes don't want to hook the spurious faults: */
1053 if (notify_page_fault(regs))
1056 * Don't take the mm semaphore here. If we fixup a prefetch
1057 * fault we could otherwise deadlock:
1059 bad_area_nosemaphore(regs, error_code, address);
1064 /* kprobes don't want to hook the spurious faults: */
1065 if (unlikely(notify_page_fault(regs)))
1068 * It's safe to allow irq's after cr2 has been saved and the
1069 * vmalloc fault has been handled.
1071 * User-mode registers count as a user access even for any
1072 * potential system fault or CPU buglet:
1074 if (user_mode_vm(regs)) {
1076 error_code |= PF_USER;
1078 if (regs->flags & X86_EFLAGS_IF)
1082 if (unlikely(error_code & PF_RSVD))
1083 pgtable_bad(regs, error_code, address);
1085 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1088 * If we're in an interrupt, have no user context or are running
1089 * in an atomic region then we must not take the fault:
1091 if (unlikely(in_atomic() || !mm)) {
1092 bad_area_nosemaphore(regs, error_code, address);
1097 * When running in the kernel we expect faults to occur only to
1098 * addresses in user space. All other faults represent errors in
1099 * the kernel and should generate an OOPS. Unfortunately, in the
1100 * case of an erroneous fault occurring in a code path which already
1101 * holds mmap_sem we will deadlock attempting to validate the fault
1102 * against the address space. Luckily the kernel only validly
1103 * references user space from well defined areas of code, which are
1104 * listed in the exceptions table.
1106 * As the vast majority of faults will be valid we will only perform
1107 * the source reference check when there is a possibility of a
1108 * deadlock. Attempt to lock the address space, if we cannot we then
1109 * validate the source. If this is invalid we can skip the address
1110 * space check, thus avoiding the deadlock:
1112 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1113 if ((error_code & PF_USER) == 0 &&
1114 !search_exception_tables(regs->ip)) {
1115 bad_area_nosemaphore(regs, error_code, address);
1119 down_read(&mm->mmap_sem);
1122 * The above down_read_trylock() might have succeeded in
1123 * which case we'll have missed the might_sleep() from
1129 vma = find_vma(mm, address);
1130 if (unlikely(!vma)) {
1131 bad_area(regs, error_code, address);
1134 if (likely(vma->vm_start <= address))
1136 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1137 bad_area(regs, error_code, address);
1140 if (error_code & PF_USER) {
1142 * Accessing the stack below %sp is always a bug.
1143 * The large cushion allows instructions like enter
1144 * and pusha to work. ("enter $65535, $31" pushes
1145 * 32 pointers and then decrements %sp by 65535.)
1147 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1148 bad_area(regs, error_code, address);
1152 if (unlikely(expand_stack(vma, address))) {
1153 bad_area(regs, error_code, address);
1158 * Ok, we have a good vm_area for this memory access, so
1159 * we can handle it..
1162 if (unlikely(access_error(error_code, vma))) {
1163 bad_area_access_error(regs, error_code, address);
1168 * If for any reason at all we couldn't handle the fault,
1169 * make sure we exit gracefully rather than endlessly redo
1172 fault = handle_mm_fault(mm, vma, address, flags);
1174 if (unlikely(fault & (VM_FAULT_RETRY|VM_FAULT_ERROR))) {
1175 if (mm_fault_error(regs, error_code, address, fault))
1180 * Major/minor page fault accounting is only done on the
1181 * initial attempt. If we go through a retry, it is extremely
1182 * likely that the page will be found in page cache at that point.
1184 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1185 if (fault & VM_FAULT_MAJOR) {
1187 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1191 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1194 if (fault & VM_FAULT_RETRY) {
1195 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1197 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1202 check_v8086_mode(regs, address, tsk);
1204 up_read(&mm->mmap_sem);