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 */
16 #include <linux/context_tracking.h> /* exception_enter(), ... */
18 #include <asm/traps.h> /* dotraplinkage, ... */
19 #include <asm/pgalloc.h> /* pgd_*(), ... */
20 #include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
21 #include <asm/fixmap.h> /* VSYSCALL_START */
23 #define CREATE_TRACE_POINTS
24 #include <asm/trace/exceptions.h>
27 * Page fault error code bits:
29 * bit 0 == 0: no page found 1: protection fault
30 * bit 1 == 0: read access 1: write access
31 * bit 2 == 0: kernel-mode access 1: user-mode access
32 * bit 3 == 1: use of reserved bit detected
33 * bit 4 == 1: fault was an instruction fetch
35 enum x86_pf_error_code {
45 * Returns 0 if mmiotrace is disabled, or if the fault is not
46 * handled by mmiotrace:
48 static inline int __kprobes
49 kmmio_fault(struct pt_regs *regs, unsigned long addr)
51 if (unlikely(is_kmmio_active()))
52 if (kmmio_handler(regs, addr) == 1)
57 static inline int __kprobes kprobes_fault(struct pt_regs *regs)
61 /* kprobe_running() needs smp_processor_id() */
62 if (kprobes_built_in() && !user_mode_vm(regs)) {
64 if (kprobe_running() && kprobe_fault_handler(regs, 14))
77 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
78 * Check that here and ignore it.
82 * Sometimes the CPU reports invalid exceptions on prefetch.
83 * Check that here and ignore it.
85 * Opcode checker based on code by Richard Brunner.
88 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
89 unsigned char opcode, int *prefetch)
91 unsigned char instr_hi = opcode & 0xf0;
92 unsigned char instr_lo = opcode & 0x0f;
98 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
99 * In X86_64 long mode, the CPU will signal invalid
100 * opcode if some of these prefixes are present so
101 * X86_64 will never get here anyway
103 return ((instr_lo & 7) == 0x6);
107 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
108 * Need to figure out under what instruction mode the
109 * instruction was issued. Could check the LDT for lm,
110 * but for now it's good enough to assume that long
111 * mode only uses well known segments or kernel.
113 return (!user_mode(regs) || user_64bit_mode(regs));
116 /* 0x64 thru 0x67 are valid prefixes in all modes. */
117 return (instr_lo & 0xC) == 0x4;
119 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
120 return !instr_lo || (instr_lo>>1) == 1;
122 /* Prefetch instruction is 0x0F0D or 0x0F18 */
123 if (probe_kernel_address(instr, opcode))
126 *prefetch = (instr_lo == 0xF) &&
127 (opcode == 0x0D || opcode == 0x18);
135 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
137 unsigned char *max_instr;
138 unsigned char *instr;
142 * If it was a exec (instruction fetch) fault on NX page, then
143 * do not ignore the fault:
145 if (error_code & PF_INSTR)
148 instr = (void *)convert_ip_to_linear(current, regs);
149 max_instr = instr + 15;
151 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
154 while (instr < max_instr) {
155 unsigned char opcode;
157 if (probe_kernel_address(instr, opcode))
162 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
169 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
170 struct task_struct *tsk, int fault)
175 info.si_signo = si_signo;
177 info.si_code = si_code;
178 info.si_addr = (void __user *)address;
179 if (fault & VM_FAULT_HWPOISON_LARGE)
180 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
181 if (fault & VM_FAULT_HWPOISON)
183 info.si_addr_lsb = lsb;
185 force_sig_info(si_signo, &info, tsk);
188 DEFINE_SPINLOCK(pgd_lock);
192 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
194 unsigned index = pgd_index(address);
200 pgd_k = init_mm.pgd + index;
202 if (!pgd_present(*pgd_k))
206 * set_pgd(pgd, *pgd_k); here would be useless on PAE
207 * and redundant with the set_pmd() on non-PAE. As would
210 pud = pud_offset(pgd, address);
211 pud_k = pud_offset(pgd_k, address);
212 if (!pud_present(*pud_k))
215 pmd = pmd_offset(pud, address);
216 pmd_k = pmd_offset(pud_k, address);
217 if (!pmd_present(*pmd_k))
220 if (!pmd_present(*pmd))
221 set_pmd(pmd, *pmd_k);
223 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
228 void vmalloc_sync_all(void)
230 unsigned long address;
232 if (SHARED_KERNEL_PMD)
235 for (address = VMALLOC_START & PMD_MASK;
236 address >= TASK_SIZE && address < FIXADDR_TOP;
237 address += PMD_SIZE) {
240 spin_lock(&pgd_lock);
241 list_for_each_entry(page, &pgd_list, lru) {
242 spinlock_t *pgt_lock;
245 /* the pgt_lock only for Xen */
246 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
249 ret = vmalloc_sync_one(page_address(page), address);
250 spin_unlock(pgt_lock);
255 spin_unlock(&pgd_lock);
262 * Handle a fault on the vmalloc or module mapping area
264 static noinline __kprobes int vmalloc_fault(unsigned long address)
266 unsigned long pgd_paddr;
270 /* Make sure we are in vmalloc area: */
271 if (!(address >= VMALLOC_START && address < VMALLOC_END))
274 WARN_ON_ONCE(in_nmi());
277 * Synchronize this task's top level page-table
278 * with the 'reference' page table.
280 * Do _not_ use "current" here. We might be inside
281 * an interrupt in the middle of a task switch..
283 pgd_paddr = read_cr3();
284 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
288 pte_k = pte_offset_kernel(pmd_k, address);
289 if (!pte_present(*pte_k))
296 * Did it hit the DOS screen memory VA from vm86 mode?
299 check_v8086_mode(struct pt_regs *regs, unsigned long address,
300 struct task_struct *tsk)
304 if (!v8086_mode(regs))
307 bit = (address - 0xA0000) >> PAGE_SHIFT;
309 tsk->thread.screen_bitmap |= 1 << bit;
312 static bool low_pfn(unsigned long pfn)
314 return pfn < max_low_pfn;
317 static void dump_pagetable(unsigned long address)
319 pgd_t *base = __va(read_cr3());
320 pgd_t *pgd = &base[pgd_index(address)];
324 #ifdef CONFIG_X86_PAE
325 printk("*pdpt = %016Lx ", pgd_val(*pgd));
326 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
329 pmd = pmd_offset(pud_offset(pgd, address), address);
330 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
333 * We must not directly access the pte in the highpte
334 * case if the page table is located in highmem.
335 * And let's rather not kmap-atomic the pte, just in case
336 * it's allocated already:
338 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
341 pte = pte_offset_kernel(pmd, address);
342 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
347 #else /* CONFIG_X86_64: */
349 void vmalloc_sync_all(void)
351 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
357 * Handle a fault on the vmalloc area
359 * This assumes no large pages in there.
361 static noinline __kprobes int vmalloc_fault(unsigned long address)
363 pgd_t *pgd, *pgd_ref;
364 pud_t *pud, *pud_ref;
365 pmd_t *pmd, *pmd_ref;
366 pte_t *pte, *pte_ref;
368 /* Make sure we are in vmalloc area: */
369 if (!(address >= VMALLOC_START && address < VMALLOC_END))
372 WARN_ON_ONCE(in_nmi());
375 * Copy kernel mappings over when needed. This can also
376 * happen within a race in page table update. In the later
379 pgd = pgd_offset(current->active_mm, address);
380 pgd_ref = pgd_offset_k(address);
381 if (pgd_none(*pgd_ref))
384 if (pgd_none(*pgd)) {
385 set_pgd(pgd, *pgd_ref);
386 arch_flush_lazy_mmu_mode();
388 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
392 * Below here mismatches are bugs because these lower tables
396 pud = pud_offset(pgd, address);
397 pud_ref = pud_offset(pgd_ref, address);
398 if (pud_none(*pud_ref))
401 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
404 pmd = pmd_offset(pud, address);
405 pmd_ref = pmd_offset(pud_ref, address);
406 if (pmd_none(*pmd_ref))
409 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
412 pte_ref = pte_offset_kernel(pmd_ref, address);
413 if (!pte_present(*pte_ref))
416 pte = pte_offset_kernel(pmd, address);
419 * Don't use pte_page here, because the mappings can point
420 * outside mem_map, and the NUMA hash lookup cannot handle
423 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
429 #ifdef CONFIG_CPU_SUP_AMD
430 static const char errata93_warning[] =
432 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
433 "******* Working around it, but it may cause SEGVs or burn power.\n"
434 "******* Please consider a BIOS update.\n"
435 "******* Disabling USB legacy in the BIOS may also help.\n";
439 * No vm86 mode in 64-bit mode:
442 check_v8086_mode(struct pt_regs *regs, unsigned long address,
443 struct task_struct *tsk)
447 static int bad_address(void *p)
451 return probe_kernel_address((unsigned long *)p, dummy);
454 static void dump_pagetable(unsigned long address)
456 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
457 pgd_t *pgd = base + pgd_index(address);
462 if (bad_address(pgd))
465 printk("PGD %lx ", pgd_val(*pgd));
467 if (!pgd_present(*pgd))
470 pud = pud_offset(pgd, address);
471 if (bad_address(pud))
474 printk("PUD %lx ", pud_val(*pud));
475 if (!pud_present(*pud) || pud_large(*pud))
478 pmd = pmd_offset(pud, address);
479 if (bad_address(pmd))
482 printk("PMD %lx ", pmd_val(*pmd));
483 if (!pmd_present(*pmd) || pmd_large(*pmd))
486 pte = pte_offset_kernel(pmd, address);
487 if (bad_address(pte))
490 printk("PTE %lx", pte_val(*pte));
498 #endif /* CONFIG_X86_64 */
501 * Workaround for K8 erratum #93 & buggy BIOS.
503 * BIOS SMM functions are required to use a specific workaround
504 * to avoid corruption of the 64bit RIP register on C stepping K8.
506 * A lot of BIOS that didn't get tested properly miss this.
508 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
509 * Try to work around it here.
511 * Note we only handle faults in kernel here.
512 * Does nothing on 32-bit.
514 static int is_errata93(struct pt_regs *regs, unsigned long address)
516 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
517 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
518 || boot_cpu_data.x86 != 0xf)
521 if (address != regs->ip)
524 if ((address >> 32) != 0)
527 address |= 0xffffffffUL << 32;
528 if ((address >= (u64)_stext && address <= (u64)_etext) ||
529 (address >= MODULES_VADDR && address <= MODULES_END)) {
530 printk_once(errata93_warning);
539 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
540 * to illegal addresses >4GB.
542 * We catch this in the page fault handler because these addresses
543 * are not reachable. Just detect this case and return. Any code
544 * segment in LDT is compatibility mode.
546 static int is_errata100(struct pt_regs *regs, unsigned long address)
549 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
555 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
557 #ifdef CONFIG_X86_F00F_BUG
561 * Pentium F0 0F C7 C8 bug workaround:
563 if (boot_cpu_has_bug(X86_BUG_F00F)) {
564 nr = (address - idt_descr.address) >> 3;
567 do_invalid_op(regs, 0);
575 static const char nx_warning[] = KERN_CRIT
576 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
579 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
580 unsigned long address)
582 if (!oops_may_print())
585 if (error_code & PF_INSTR) {
588 pte_t *pte = lookup_address(address, &level);
590 if (pte && pte_present(*pte) && !pte_exec(*pte))
591 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
594 printk(KERN_ALERT "BUG: unable to handle kernel ");
595 if (address < PAGE_SIZE)
596 printk(KERN_CONT "NULL pointer dereference");
598 printk(KERN_CONT "paging request");
600 printk(KERN_CONT " at %p\n", (void *) address);
601 printk(KERN_ALERT "IP:");
602 printk_address(regs->ip);
604 dump_pagetable(address);
608 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
609 unsigned long address)
611 struct task_struct *tsk;
615 flags = oops_begin();
619 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
621 dump_pagetable(address);
623 tsk->thread.cr2 = address;
624 tsk->thread.trap_nr = X86_TRAP_PF;
625 tsk->thread.error_code = error_code;
627 if (__die("Bad pagetable", regs, error_code))
630 oops_end(flags, regs, sig);
634 no_context(struct pt_regs *regs, unsigned long error_code,
635 unsigned long address, int signal, int si_code)
637 struct task_struct *tsk = current;
638 unsigned long *stackend;
642 /* Are we prepared to handle this kernel fault? */
643 if (fixup_exception(regs)) {
645 * Any interrupt that takes a fault gets the fixup. This makes
646 * the below recursive fault logic only apply to a faults from
653 * Per the above we're !in_interrupt(), aka. task context.
655 * In this case we need to make sure we're not recursively
656 * faulting through the emulate_vsyscall() logic.
658 if (current_thread_info()->sig_on_uaccess_error && signal) {
659 tsk->thread.trap_nr = X86_TRAP_PF;
660 tsk->thread.error_code = error_code | PF_USER;
661 tsk->thread.cr2 = address;
663 /* XXX: hwpoison faults will set the wrong code. */
664 force_sig_info_fault(signal, si_code, address, tsk, 0);
668 * Barring that, we can do the fixup and be happy.
676 * Valid to do another page fault here, because if this fault
677 * had been triggered by is_prefetch fixup_exception would have
682 * Hall of shame of CPU/BIOS bugs.
684 if (is_prefetch(regs, error_code, address))
687 if (is_errata93(regs, address))
691 * Oops. The kernel tried to access some bad page. We'll have to
692 * terminate things with extreme prejudice:
694 flags = oops_begin();
696 show_fault_oops(regs, error_code, address);
698 stackend = end_of_stack(tsk);
699 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
700 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
702 tsk->thread.cr2 = address;
703 tsk->thread.trap_nr = X86_TRAP_PF;
704 tsk->thread.error_code = error_code;
707 if (__die("Oops", regs, error_code))
710 /* Executive summary in case the body of the oops scrolled away */
711 printk(KERN_DEFAULT "CR2: %016lx\n", address);
713 oops_end(flags, regs, sig);
717 * Print out info about fatal segfaults, if the show_unhandled_signals
721 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
722 unsigned long address, struct task_struct *tsk)
724 if (!unhandled_signal(tsk, SIGSEGV))
727 if (!printk_ratelimit())
730 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
731 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
732 tsk->comm, task_pid_nr(tsk), address,
733 (void *)regs->ip, (void *)regs->sp, error_code);
735 print_vma_addr(KERN_CONT " in ", regs->ip);
737 printk(KERN_CONT "\n");
741 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
742 unsigned long address, int si_code)
744 struct task_struct *tsk = current;
746 /* User mode accesses just cause a SIGSEGV */
747 if (error_code & PF_USER) {
749 * It's possible to have interrupts off here:
754 * Valid to do another page fault here because this one came
757 if (is_prefetch(regs, error_code, address))
760 if (is_errata100(regs, address))
765 * Instruction fetch faults in the vsyscall page might need
768 if (unlikely((error_code & PF_INSTR) &&
769 ((address & ~0xfff) == VSYSCALL_START))) {
770 if (emulate_vsyscall(regs, address))
774 /* Kernel addresses are always protection faults: */
775 if (address >= TASK_SIZE)
776 error_code |= PF_PROT;
778 if (likely(show_unhandled_signals))
779 show_signal_msg(regs, error_code, address, tsk);
781 tsk->thread.cr2 = address;
782 tsk->thread.error_code = error_code;
783 tsk->thread.trap_nr = X86_TRAP_PF;
785 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
790 if (is_f00f_bug(regs, address))
793 no_context(regs, error_code, address, SIGSEGV, si_code);
797 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
798 unsigned long address)
800 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
804 __bad_area(struct pt_regs *regs, unsigned long error_code,
805 unsigned long address, int si_code)
807 struct mm_struct *mm = current->mm;
810 * Something tried to access memory that isn't in our memory map..
811 * Fix it, but check if it's kernel or user first..
813 up_read(&mm->mmap_sem);
815 __bad_area_nosemaphore(regs, error_code, address, si_code);
819 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
821 __bad_area(regs, error_code, address, SEGV_MAPERR);
825 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
826 unsigned long address)
828 __bad_area(regs, error_code, address, SEGV_ACCERR);
832 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
835 struct task_struct *tsk = current;
836 struct mm_struct *mm = tsk->mm;
837 int code = BUS_ADRERR;
839 up_read(&mm->mmap_sem);
841 /* Kernel mode? Handle exceptions or die: */
842 if (!(error_code & PF_USER)) {
843 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
847 /* User-space => ok to do another page fault: */
848 if (is_prefetch(regs, error_code, address))
851 tsk->thread.cr2 = address;
852 tsk->thread.error_code = error_code;
853 tsk->thread.trap_nr = X86_TRAP_PF;
855 #ifdef CONFIG_MEMORY_FAILURE
856 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
858 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
859 tsk->comm, tsk->pid, address);
860 code = BUS_MCEERR_AR;
863 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
867 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
868 unsigned long address, unsigned int fault)
870 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
871 up_read(¤t->mm->mmap_sem);
872 no_context(regs, error_code, address, 0, 0);
876 if (fault & VM_FAULT_OOM) {
877 /* Kernel mode? Handle exceptions or die: */
878 if (!(error_code & PF_USER)) {
879 up_read(¤t->mm->mmap_sem);
880 no_context(regs, error_code, address,
881 SIGSEGV, SEGV_MAPERR);
885 up_read(¤t->mm->mmap_sem);
888 * We ran out of memory, call the OOM killer, and return the
889 * userspace (which will retry the fault, or kill us if we got
892 pagefault_out_of_memory();
894 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
895 VM_FAULT_HWPOISON_LARGE))
896 do_sigbus(regs, error_code, address, fault);
902 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
904 if ((error_code & PF_WRITE) && !pte_write(*pte))
907 if ((error_code & PF_INSTR) && !pte_exec(*pte))
914 * Handle a spurious fault caused by a stale TLB entry.
916 * This allows us to lazily refresh the TLB when increasing the
917 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
918 * eagerly is very expensive since that implies doing a full
919 * cross-processor TLB flush, even if no stale TLB entries exist
920 * on other processors.
922 * There are no security implications to leaving a stale TLB when
923 * increasing the permissions on a page.
925 static noinline __kprobes int
926 spurious_fault(unsigned long error_code, unsigned long address)
934 /* Reserved-bit violation or user access to kernel space? */
935 if (error_code & (PF_USER | PF_RSVD))
938 pgd = init_mm.pgd + pgd_index(address);
939 if (!pgd_present(*pgd))
942 pud = pud_offset(pgd, address);
943 if (!pud_present(*pud))
947 return spurious_fault_check(error_code, (pte_t *) pud);
949 pmd = pmd_offset(pud, address);
950 if (!pmd_present(*pmd))
954 return spurious_fault_check(error_code, (pte_t *) pmd);
956 pte = pte_offset_kernel(pmd, address);
957 if (!pte_present(*pte))
960 ret = spurious_fault_check(error_code, pte);
965 * Make sure we have permissions in PMD.
966 * If not, then there's a bug in the page tables:
968 ret = spurious_fault_check(error_code, (pte_t *) pmd);
969 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
974 int show_unhandled_signals = 1;
977 access_error(unsigned long error_code, struct vm_area_struct *vma)
979 if (error_code & PF_WRITE) {
980 /* write, present and write, not present: */
981 if (unlikely(!(vma->vm_flags & VM_WRITE)))
987 if (unlikely(error_code & PF_PROT))
990 /* read, not present: */
991 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
997 static int fault_in_kernel_space(unsigned long address)
999 return address >= TASK_SIZE_MAX;
1002 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1004 if (!IS_ENABLED(CONFIG_X86_SMAP))
1007 if (!static_cpu_has(X86_FEATURE_SMAP))
1010 if (error_code & PF_USER)
1013 if (!user_mode_vm(regs) && (regs->flags & X86_EFLAGS_AC))
1020 * This routine handles page faults. It determines the address,
1021 * and the problem, and then passes it off to one of the appropriate
1024 * This function must have noinline because both callers
1025 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1026 * guarantees there's a function trace entry.
1028 static void __kprobes noinline
1029 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1030 unsigned long address)
1032 struct vm_area_struct *vma;
1033 struct task_struct *tsk;
1034 struct mm_struct *mm;
1036 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1042 * Detect and handle instructions that would cause a page fault for
1043 * both a tracked kernel page and a userspace page.
1045 if (kmemcheck_active(regs))
1046 kmemcheck_hide(regs);
1047 prefetchw(&mm->mmap_sem);
1049 if (unlikely(kmmio_fault(regs, address)))
1053 * We fault-in kernel-space virtual memory on-demand. The
1054 * 'reference' page table is init_mm.pgd.
1056 * NOTE! We MUST NOT take any locks for this case. We may
1057 * be in an interrupt or a critical region, and should
1058 * only copy the information from the master page table,
1061 * This verifies that the fault happens in kernel space
1062 * (error_code & 4) == 0, and that the fault was not a
1063 * protection error (error_code & 9) == 0.
1065 if (unlikely(fault_in_kernel_space(address))) {
1066 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1067 if (vmalloc_fault(address) >= 0)
1070 if (kmemcheck_fault(regs, address, error_code))
1074 /* Can handle a stale RO->RW TLB: */
1075 if (spurious_fault(error_code, address))
1078 /* kprobes don't want to hook the spurious faults: */
1079 if (kprobes_fault(regs))
1082 * Don't take the mm semaphore here. If we fixup a prefetch
1083 * fault we could otherwise deadlock:
1085 bad_area_nosemaphore(regs, error_code, address);
1090 /* kprobes don't want to hook the spurious faults: */
1091 if (unlikely(kprobes_fault(regs)))
1094 if (unlikely(error_code & PF_RSVD))
1095 pgtable_bad(regs, error_code, address);
1097 if (unlikely(smap_violation(error_code, regs))) {
1098 bad_area_nosemaphore(regs, error_code, address);
1103 * If we're in an interrupt, have no user context or are running
1104 * in an atomic region then we must not take the fault:
1106 if (unlikely(in_atomic() || !mm)) {
1107 bad_area_nosemaphore(regs, error_code, address);
1112 * It's safe to allow irq's after cr2 has been saved and the
1113 * vmalloc fault has been handled.
1115 * User-mode registers count as a user access even for any
1116 * potential system fault or CPU buglet:
1118 if (user_mode_vm(regs)) {
1120 error_code |= PF_USER;
1121 flags |= FAULT_FLAG_USER;
1123 if (regs->flags & X86_EFLAGS_IF)
1127 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1129 if (error_code & PF_WRITE)
1130 flags |= FAULT_FLAG_WRITE;
1133 * When running in the kernel we expect faults to occur only to
1134 * addresses in user space. All other faults represent errors in
1135 * the kernel and should generate an OOPS. Unfortunately, in the
1136 * case of an erroneous fault occurring in a code path which already
1137 * holds mmap_sem we will deadlock attempting to validate the fault
1138 * against the address space. Luckily the kernel only validly
1139 * references user space from well defined areas of code, which are
1140 * listed in the exceptions table.
1142 * As the vast majority of faults will be valid we will only perform
1143 * the source reference check when there is a possibility of a
1144 * deadlock. Attempt to lock the address space, if we cannot we then
1145 * validate the source. If this is invalid we can skip the address
1146 * space check, thus avoiding the deadlock:
1148 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1149 if ((error_code & PF_USER) == 0 &&
1150 !search_exception_tables(regs->ip)) {
1151 bad_area_nosemaphore(regs, error_code, address);
1155 down_read(&mm->mmap_sem);
1158 * The above down_read_trylock() might have succeeded in
1159 * which case we'll have missed the might_sleep() from
1165 vma = find_vma(mm, address);
1166 if (unlikely(!vma)) {
1167 bad_area(regs, error_code, address);
1170 if (likely(vma->vm_start <= address))
1172 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1173 bad_area(regs, error_code, address);
1176 if (error_code & PF_USER) {
1178 * Accessing the stack below %sp is always a bug.
1179 * The large cushion allows instructions like enter
1180 * and pusha to work. ("enter $65535, $31" pushes
1181 * 32 pointers and then decrements %sp by 65535.)
1183 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1184 bad_area(regs, error_code, address);
1188 if (unlikely(expand_stack(vma, address))) {
1189 bad_area(regs, error_code, address);
1194 * Ok, we have a good vm_area for this memory access, so
1195 * we can handle it..
1198 if (unlikely(access_error(error_code, vma))) {
1199 bad_area_access_error(regs, error_code, address);
1204 * If for any reason at all we couldn't handle the fault,
1205 * make sure we exit gracefully rather than endlessly redo
1208 fault = handle_mm_fault(mm, vma, address, flags);
1211 * If we need to retry but a fatal signal is pending, handle the
1212 * signal first. We do not need to release the mmap_sem because it
1213 * would already be released in __lock_page_or_retry in mm/filemap.c.
1215 if (unlikely((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)))
1218 if (unlikely(fault & VM_FAULT_ERROR)) {
1219 mm_fault_error(regs, error_code, address, fault);
1224 * Major/minor page fault accounting is only done on the
1225 * initial attempt. If we go through a retry, it is extremely
1226 * likely that the page will be found in page cache at that point.
1228 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1229 if (fault & VM_FAULT_MAJOR) {
1231 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1235 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1238 if (fault & VM_FAULT_RETRY) {
1239 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1241 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1242 flags |= FAULT_FLAG_TRIED;
1247 check_v8086_mode(regs, address, tsk);
1249 up_read(&mm->mmap_sem);
1252 dotraplinkage void __kprobes notrace
1253 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1255 unsigned long address = read_cr2(); /* Get the faulting address */
1256 enum ctx_state prev_state;
1259 * We must have this function tagged with __kprobes, notrace and call
1260 * read_cr2() before calling anything else. To avoid calling any kind
1261 * of tracing machinery before we've observed the CR2 value.
1263 * exception_{enter,exit}() contain all sorts of tracepoints.
1266 prev_state = exception_enter();
1267 __do_page_fault(regs, error_code, address);
1268 exception_exit(prev_state);
1271 #ifdef CONFIG_TRACING
1272 static void trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1273 unsigned long error_code)
1275 if (user_mode(regs))
1276 trace_page_fault_user(address, regs, error_code);
1278 trace_page_fault_kernel(address, regs, error_code);
1281 dotraplinkage void __kprobes notrace
1282 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1285 * The exception_enter and tracepoint processing could
1286 * trigger another page faults (user space callchain
1287 * reading) and destroy the original cr2 value, so read
1288 * the faulting address now.
1290 unsigned long address = read_cr2();
1291 enum ctx_state prev_state;
1293 prev_state = exception_enter();
1294 trace_page_fault_entries(address, regs, error_code);
1295 __do_page_fault(regs, error_code, address);
1296 exception_exit(prev_state);
1298 #endif /* CONFIG_TRACING */