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> /* NOKPROBE_SYMBOL, ... */
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_ADDR */
22 #include <asm/vsyscall.h> /* emulate_vsyscall */
24 #define CREATE_TRACE_POINTS
25 #include <asm/trace/exceptions.h>
28 * Page fault error code bits:
30 * bit 0 == 0: no page found 1: protection fault
31 * bit 1 == 0: read access 1: write access
32 * bit 2 == 0: kernel-mode access 1: user-mode access
33 * bit 3 == 1: use of reserved bit detected
34 * bit 4 == 1: fault was an instruction fetch
36 enum x86_pf_error_code {
46 * Returns 0 if mmiotrace is disabled, or if the fault is not
47 * handled by mmiotrace:
49 static nokprobe_inline int
50 kmmio_fault(struct pt_regs *regs, unsigned long addr)
52 if (unlikely(is_kmmio_active()))
53 if (kmmio_handler(regs, addr) == 1)
58 static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
62 /* kprobe_running() needs smp_processor_id() */
63 if (kprobes_built_in() && !user_mode_vm(regs)) {
65 if (kprobe_running() && kprobe_fault_handler(regs, 14))
78 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
79 * Check that here and ignore it.
83 * Sometimes the CPU reports invalid exceptions on prefetch.
84 * Check that here and ignore it.
86 * Opcode checker based on code by Richard Brunner.
89 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
90 unsigned char opcode, int *prefetch)
92 unsigned char instr_hi = opcode & 0xf0;
93 unsigned char instr_lo = opcode & 0x0f;
99 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
100 * In X86_64 long mode, the CPU will signal invalid
101 * opcode if some of these prefixes are present so
102 * X86_64 will never get here anyway
104 return ((instr_lo & 7) == 0x6);
108 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
109 * Need to figure out under what instruction mode the
110 * instruction was issued. Could check the LDT for lm,
111 * but for now it's good enough to assume that long
112 * mode only uses well known segments or kernel.
114 return (!user_mode(regs) || user_64bit_mode(regs));
117 /* 0x64 thru 0x67 are valid prefixes in all modes. */
118 return (instr_lo & 0xC) == 0x4;
120 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
121 return !instr_lo || (instr_lo>>1) == 1;
123 /* Prefetch instruction is 0x0F0D or 0x0F18 */
124 if (probe_kernel_address(instr, opcode))
127 *prefetch = (instr_lo == 0xF) &&
128 (opcode == 0x0D || opcode == 0x18);
136 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
138 unsigned char *max_instr;
139 unsigned char *instr;
143 * If it was a exec (instruction fetch) fault on NX page, then
144 * do not ignore the fault:
146 if (error_code & PF_INSTR)
149 instr = (void *)convert_ip_to_linear(current, regs);
150 max_instr = instr + 15;
152 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
155 while (instr < max_instr) {
156 unsigned char opcode;
158 if (probe_kernel_address(instr, opcode))
163 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
170 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
171 struct task_struct *tsk, int fault)
176 info.si_signo = si_signo;
178 info.si_code = si_code;
179 info.si_addr = (void __user *)address;
180 if (fault & VM_FAULT_HWPOISON_LARGE)
181 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
182 if (fault & VM_FAULT_HWPOISON)
184 info.si_addr_lsb = lsb;
186 force_sig_info(si_signo, &info, tsk);
189 DEFINE_SPINLOCK(pgd_lock);
193 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
195 unsigned index = pgd_index(address);
201 pgd_k = init_mm.pgd + index;
203 if (!pgd_present(*pgd_k))
207 * set_pgd(pgd, *pgd_k); here would be useless on PAE
208 * and redundant with the set_pmd() on non-PAE. As would
211 pud = pud_offset(pgd, address);
212 pud_k = pud_offset(pgd_k, address);
213 if (!pud_present(*pud_k))
216 pmd = pmd_offset(pud, address);
217 pmd_k = pmd_offset(pud_k, address);
218 if (!pmd_present(*pmd_k))
221 if (!pmd_present(*pmd))
222 set_pmd(pmd, *pmd_k);
224 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
229 void vmalloc_sync_all(void)
231 unsigned long address;
233 if (SHARED_KERNEL_PMD)
236 for (address = VMALLOC_START & PMD_MASK;
237 address >= TASK_SIZE && address < FIXADDR_TOP;
238 address += PMD_SIZE) {
241 spin_lock(&pgd_lock);
242 list_for_each_entry(page, &pgd_list, lru) {
243 spinlock_t *pgt_lock;
246 /* the pgt_lock only for Xen */
247 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
250 ret = vmalloc_sync_one(page_address(page), address);
251 spin_unlock(pgt_lock);
256 spin_unlock(&pgd_lock);
263 * Handle a fault on the vmalloc or module mapping area
265 static noinline int vmalloc_fault(unsigned long address)
267 unsigned long pgd_paddr;
271 /* Make sure we are in vmalloc area: */
272 if (!(address >= VMALLOC_START && address < VMALLOC_END))
275 WARN_ON_ONCE(in_nmi());
278 * Synchronize this task's top level page-table
279 * with the 'reference' page table.
281 * Do _not_ use "current" here. We might be inside
282 * an interrupt in the middle of a task switch..
284 pgd_paddr = read_cr3();
285 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
289 pte_k = pte_offset_kernel(pmd_k, address);
290 if (!pte_present(*pte_k))
295 NOKPROBE_SYMBOL(vmalloc_fault);
298 * Did it hit the DOS screen memory VA from vm86 mode?
301 check_v8086_mode(struct pt_regs *regs, unsigned long address,
302 struct task_struct *tsk)
306 if (!v8086_mode(regs))
309 bit = (address - 0xA0000) >> PAGE_SHIFT;
311 tsk->thread.screen_bitmap |= 1 << bit;
314 static bool low_pfn(unsigned long pfn)
316 return pfn < max_low_pfn;
319 static void dump_pagetable(unsigned long address)
321 pgd_t *base = __va(read_cr3());
322 pgd_t *pgd = &base[pgd_index(address)];
326 #ifdef CONFIG_X86_PAE
327 printk("*pdpt = %016Lx ", pgd_val(*pgd));
328 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
331 pmd = pmd_offset(pud_offset(pgd, address), address);
332 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
335 * We must not directly access the pte in the highpte
336 * case if the page table is located in highmem.
337 * And let's rather not kmap-atomic the pte, just in case
338 * it's allocated already:
340 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
343 pte = pte_offset_kernel(pmd, address);
344 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
349 #else /* CONFIG_X86_64: */
351 void vmalloc_sync_all(void)
353 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
359 * Handle a fault on the vmalloc area
361 * This assumes no large pages in there.
363 static noinline int vmalloc_fault(unsigned long address)
365 pgd_t *pgd, *pgd_ref;
366 pud_t *pud, *pud_ref;
367 pmd_t *pmd, *pmd_ref;
368 pte_t *pte, *pte_ref;
370 /* Make sure we are in vmalloc area: */
371 if (!(address >= VMALLOC_START && address < VMALLOC_END))
374 WARN_ON_ONCE(in_nmi());
377 * Copy kernel mappings over when needed. This can also
378 * happen within a race in page table update. In the later
381 pgd = pgd_offset(current->active_mm, address);
382 pgd_ref = pgd_offset_k(address);
383 if (pgd_none(*pgd_ref))
386 if (pgd_none(*pgd)) {
387 set_pgd(pgd, *pgd_ref);
388 arch_flush_lazy_mmu_mode();
390 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
394 * Below here mismatches are bugs because these lower tables
398 pud = pud_offset(pgd, address);
399 pud_ref = pud_offset(pgd_ref, address);
400 if (pud_none(*pud_ref))
403 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
406 pmd = pmd_offset(pud, address);
407 pmd_ref = pmd_offset(pud_ref, address);
408 if (pmd_none(*pmd_ref))
411 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
414 pte_ref = pte_offset_kernel(pmd_ref, address);
415 if (!pte_present(*pte_ref))
418 pte = pte_offset_kernel(pmd, address);
421 * Don't use pte_page here, because the mappings can point
422 * outside mem_map, and the NUMA hash lookup cannot handle
425 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
430 NOKPROBE_SYMBOL(vmalloc_fault);
432 #ifdef CONFIG_CPU_SUP_AMD
433 static const char errata93_warning[] =
435 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
436 "******* Working around it, but it may cause SEGVs or burn power.\n"
437 "******* Please consider a BIOS update.\n"
438 "******* Disabling USB legacy in the BIOS may also help.\n";
442 * No vm86 mode in 64-bit mode:
445 check_v8086_mode(struct pt_regs *regs, unsigned long address,
446 struct task_struct *tsk)
450 static int bad_address(void *p)
454 return probe_kernel_address((unsigned long *)p, dummy);
457 static void dump_pagetable(unsigned long address)
459 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
460 pgd_t *pgd = base + pgd_index(address);
465 if (bad_address(pgd))
468 printk("PGD %lx ", pgd_val(*pgd));
470 if (!pgd_present(*pgd))
473 pud = pud_offset(pgd, address);
474 if (bad_address(pud))
477 printk("PUD %lx ", pud_val(*pud));
478 if (!pud_present(*pud) || pud_large(*pud))
481 pmd = pmd_offset(pud, address);
482 if (bad_address(pmd))
485 printk("PMD %lx ", pmd_val(*pmd));
486 if (!pmd_present(*pmd) || pmd_large(*pmd))
489 pte = pte_offset_kernel(pmd, address);
490 if (bad_address(pte))
493 printk("PTE %lx", pte_val(*pte));
501 #endif /* CONFIG_X86_64 */
504 * Workaround for K8 erratum #93 & buggy BIOS.
506 * BIOS SMM functions are required to use a specific workaround
507 * to avoid corruption of the 64bit RIP register on C stepping K8.
509 * A lot of BIOS that didn't get tested properly miss this.
511 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
512 * Try to work around it here.
514 * Note we only handle faults in kernel here.
515 * Does nothing on 32-bit.
517 static int is_errata93(struct pt_regs *regs, unsigned long address)
519 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
520 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
521 || boot_cpu_data.x86 != 0xf)
524 if (address != regs->ip)
527 if ((address >> 32) != 0)
530 address |= 0xffffffffUL << 32;
531 if ((address >= (u64)_stext && address <= (u64)_etext) ||
532 (address >= MODULES_VADDR && address <= MODULES_END)) {
533 printk_once(errata93_warning);
542 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
543 * to illegal addresses >4GB.
545 * We catch this in the page fault handler because these addresses
546 * are not reachable. Just detect this case and return. Any code
547 * segment in LDT is compatibility mode.
549 static int is_errata100(struct pt_regs *regs, unsigned long address)
552 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
558 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
560 #ifdef CONFIG_X86_F00F_BUG
564 * Pentium F0 0F C7 C8 bug workaround:
566 if (boot_cpu_has_bug(X86_BUG_F00F)) {
567 nr = (address - idt_descr.address) >> 3;
570 do_invalid_op(regs, 0);
578 static const char nx_warning[] = KERN_CRIT
579 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
582 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
583 unsigned long address)
585 if (!oops_may_print())
588 if (error_code & PF_INSTR) {
593 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
594 pgd += pgd_index(address);
596 pte = lookup_address_in_pgd(pgd, address, &level);
598 if (pte && pte_present(*pte) && !pte_exec(*pte))
599 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
602 printk(KERN_ALERT "BUG: unable to handle kernel ");
603 if (address < PAGE_SIZE)
604 printk(KERN_CONT "NULL pointer dereference");
606 printk(KERN_CONT "paging request");
608 printk(KERN_CONT " at %p\n", (void *) address);
609 printk(KERN_ALERT "IP:");
610 printk_address(regs->ip);
612 dump_pagetable(address);
616 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
617 unsigned long address)
619 struct task_struct *tsk;
623 flags = oops_begin();
627 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
629 dump_pagetable(address);
631 tsk->thread.cr2 = address;
632 tsk->thread.trap_nr = X86_TRAP_PF;
633 tsk->thread.error_code = error_code;
635 if (__die("Bad pagetable", regs, error_code))
638 oops_end(flags, regs, sig);
642 no_context(struct pt_regs *regs, unsigned long error_code,
643 unsigned long address, int signal, int si_code)
645 struct task_struct *tsk = current;
646 unsigned long *stackend;
650 /* Are we prepared to handle this kernel fault? */
651 if (fixup_exception(regs)) {
653 * Any interrupt that takes a fault gets the fixup. This makes
654 * the below recursive fault logic only apply to a faults from
661 * Per the above we're !in_interrupt(), aka. task context.
663 * In this case we need to make sure we're not recursively
664 * faulting through the emulate_vsyscall() logic.
666 if (current_thread_info()->sig_on_uaccess_error && signal) {
667 tsk->thread.trap_nr = X86_TRAP_PF;
668 tsk->thread.error_code = error_code | PF_USER;
669 tsk->thread.cr2 = address;
671 /* XXX: hwpoison faults will set the wrong code. */
672 force_sig_info_fault(signal, si_code, address, tsk, 0);
676 * Barring that, we can do the fixup and be happy.
684 * Valid to do another page fault here, because if this fault
685 * had been triggered by is_prefetch fixup_exception would have
690 * Hall of shame of CPU/BIOS bugs.
692 if (is_prefetch(regs, error_code, address))
695 if (is_errata93(regs, address))
699 * Oops. The kernel tried to access some bad page. We'll have to
700 * terminate things with extreme prejudice:
702 flags = oops_begin();
704 show_fault_oops(regs, error_code, address);
706 stackend = end_of_stack(tsk);
707 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
708 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
710 tsk->thread.cr2 = address;
711 tsk->thread.trap_nr = X86_TRAP_PF;
712 tsk->thread.error_code = error_code;
715 if (__die("Oops", regs, error_code))
718 /* Executive summary in case the body of the oops scrolled away */
719 printk(KERN_DEFAULT "CR2: %016lx\n", address);
721 oops_end(flags, regs, sig);
725 * Print out info about fatal segfaults, if the show_unhandled_signals
729 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
730 unsigned long address, struct task_struct *tsk)
732 if (!unhandled_signal(tsk, SIGSEGV))
735 if (!printk_ratelimit())
738 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
739 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
740 tsk->comm, task_pid_nr(tsk), address,
741 (void *)regs->ip, (void *)regs->sp, error_code);
743 print_vma_addr(KERN_CONT " in ", regs->ip);
745 printk(KERN_CONT "\n");
749 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
750 unsigned long address, int si_code)
752 struct task_struct *tsk = current;
754 /* User mode accesses just cause a SIGSEGV */
755 if (error_code & PF_USER) {
757 * It's possible to have interrupts off here:
762 * Valid to do another page fault here because this one came
765 if (is_prefetch(regs, error_code, address))
768 if (is_errata100(regs, address))
773 * Instruction fetch faults in the vsyscall page might need
776 if (unlikely((error_code & PF_INSTR) &&
777 ((address & ~0xfff) == VSYSCALL_ADDR))) {
778 if (emulate_vsyscall(regs, address))
782 /* Kernel addresses are always protection faults: */
783 if (address >= TASK_SIZE)
784 error_code |= PF_PROT;
786 if (likely(show_unhandled_signals))
787 show_signal_msg(regs, error_code, address, tsk);
789 tsk->thread.cr2 = address;
790 tsk->thread.error_code = error_code;
791 tsk->thread.trap_nr = X86_TRAP_PF;
793 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
798 if (is_f00f_bug(regs, address))
801 no_context(regs, error_code, address, SIGSEGV, si_code);
805 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
806 unsigned long address)
808 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
812 __bad_area(struct pt_regs *regs, unsigned long error_code,
813 unsigned long address, int si_code)
815 struct mm_struct *mm = current->mm;
818 * Something tried to access memory that isn't in our memory map..
819 * Fix it, but check if it's kernel or user first..
821 up_read(&mm->mmap_sem);
823 __bad_area_nosemaphore(regs, error_code, address, si_code);
827 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
829 __bad_area(regs, error_code, address, SEGV_MAPERR);
833 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
834 unsigned long address)
836 __bad_area(regs, error_code, address, SEGV_ACCERR);
840 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
843 struct task_struct *tsk = current;
844 struct mm_struct *mm = tsk->mm;
845 int code = BUS_ADRERR;
847 up_read(&mm->mmap_sem);
849 /* Kernel mode? Handle exceptions or die: */
850 if (!(error_code & PF_USER)) {
851 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
855 /* User-space => ok to do another page fault: */
856 if (is_prefetch(regs, error_code, address))
859 tsk->thread.cr2 = address;
860 tsk->thread.error_code = error_code;
861 tsk->thread.trap_nr = X86_TRAP_PF;
863 #ifdef CONFIG_MEMORY_FAILURE
864 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
866 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
867 tsk->comm, tsk->pid, address);
868 code = BUS_MCEERR_AR;
871 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
875 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
876 unsigned long address, unsigned int fault)
878 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
879 up_read(¤t->mm->mmap_sem);
880 no_context(regs, error_code, address, 0, 0);
884 if (fault & VM_FAULT_OOM) {
885 /* Kernel mode? Handle exceptions or die: */
886 if (!(error_code & PF_USER)) {
887 up_read(¤t->mm->mmap_sem);
888 no_context(regs, error_code, address,
889 SIGSEGV, SEGV_MAPERR);
893 up_read(¤t->mm->mmap_sem);
896 * We ran out of memory, call the OOM killer, and return the
897 * userspace (which will retry the fault, or kill us if we got
900 pagefault_out_of_memory();
902 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
903 VM_FAULT_HWPOISON_LARGE))
904 do_sigbus(regs, error_code, address, fault);
910 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
912 if ((error_code & PF_WRITE) && !pte_write(*pte))
915 if ((error_code & PF_INSTR) && !pte_exec(*pte))
922 * Handle a spurious fault caused by a stale TLB entry.
924 * This allows us to lazily refresh the TLB when increasing the
925 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
926 * eagerly is very expensive since that implies doing a full
927 * cross-processor TLB flush, even if no stale TLB entries exist
928 * on other processors.
930 * There are no security implications to leaving a stale TLB when
931 * increasing the permissions on a page.
934 spurious_fault(unsigned long error_code, unsigned long address)
942 /* Reserved-bit violation or user access to kernel space? */
943 if (error_code & (PF_USER | PF_RSVD))
946 pgd = init_mm.pgd + pgd_index(address);
947 if (!pgd_present(*pgd))
950 pud = pud_offset(pgd, address);
951 if (!pud_present(*pud))
955 return spurious_fault_check(error_code, (pte_t *) pud);
957 pmd = pmd_offset(pud, address);
958 if (!pmd_present(*pmd))
962 return spurious_fault_check(error_code, (pte_t *) pmd);
964 pte = pte_offset_kernel(pmd, address);
965 if (!pte_present(*pte))
968 ret = spurious_fault_check(error_code, pte);
973 * Make sure we have permissions in PMD.
974 * If not, then there's a bug in the page tables:
976 ret = spurious_fault_check(error_code, (pte_t *) pmd);
977 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
981 NOKPROBE_SYMBOL(spurious_fault);
983 int show_unhandled_signals = 1;
986 access_error(unsigned long error_code, struct vm_area_struct *vma)
988 if (error_code & PF_WRITE) {
989 /* write, present and write, not present: */
990 if (unlikely(!(vma->vm_flags & VM_WRITE)))
996 if (unlikely(error_code & PF_PROT))
999 /* read, not present: */
1000 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1006 static int fault_in_kernel_space(unsigned long address)
1008 return address >= TASK_SIZE_MAX;
1011 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1013 if (!IS_ENABLED(CONFIG_X86_SMAP))
1016 if (!static_cpu_has(X86_FEATURE_SMAP))
1019 if (error_code & PF_USER)
1022 if (!user_mode_vm(regs) && (regs->flags & X86_EFLAGS_AC))
1029 * This routine handles page faults. It determines the address,
1030 * and the problem, and then passes it off to one of the appropriate
1033 * This function must have noinline because both callers
1034 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1035 * guarantees there's a function trace entry.
1037 static noinline void
1038 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1039 unsigned long address)
1041 struct vm_area_struct *vma;
1042 struct task_struct *tsk;
1043 struct mm_struct *mm;
1045 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1051 * Detect and handle instructions that would cause a page fault for
1052 * both a tracked kernel page and a userspace page.
1054 if (kmemcheck_active(regs))
1055 kmemcheck_hide(regs);
1056 prefetchw(&mm->mmap_sem);
1058 if (unlikely(kmmio_fault(regs, address)))
1062 * We fault-in kernel-space virtual memory on-demand. The
1063 * 'reference' page table is init_mm.pgd.
1065 * NOTE! We MUST NOT take any locks for this case. We may
1066 * be in an interrupt or a critical region, and should
1067 * only copy the information from the master page table,
1070 * This verifies that the fault happens in kernel space
1071 * (error_code & 4) == 0, and that the fault was not a
1072 * protection error (error_code & 9) == 0.
1074 if (unlikely(fault_in_kernel_space(address))) {
1075 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1076 if (vmalloc_fault(address) >= 0)
1079 if (kmemcheck_fault(regs, address, error_code))
1083 /* Can handle a stale RO->RW TLB: */
1084 if (spurious_fault(error_code, address))
1087 /* kprobes don't want to hook the spurious faults: */
1088 if (kprobes_fault(regs))
1091 * Don't take the mm semaphore here. If we fixup a prefetch
1092 * fault we could otherwise deadlock:
1094 bad_area_nosemaphore(regs, error_code, address);
1099 /* kprobes don't want to hook the spurious faults: */
1100 if (unlikely(kprobes_fault(regs)))
1103 if (unlikely(error_code & PF_RSVD))
1104 pgtable_bad(regs, error_code, address);
1106 if (unlikely(smap_violation(error_code, regs))) {
1107 bad_area_nosemaphore(regs, error_code, address);
1112 * If we're in an interrupt, have no user context or are running
1113 * in an atomic region then we must not take the fault:
1115 if (unlikely(in_atomic() || !mm)) {
1116 bad_area_nosemaphore(regs, error_code, address);
1121 * It's safe to allow irq's after cr2 has been saved and the
1122 * vmalloc fault has been handled.
1124 * User-mode registers count as a user access even for any
1125 * potential system fault or CPU buglet:
1127 if (user_mode_vm(regs)) {
1129 error_code |= PF_USER;
1130 flags |= FAULT_FLAG_USER;
1132 if (regs->flags & X86_EFLAGS_IF)
1136 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1138 if (error_code & PF_WRITE)
1139 flags |= FAULT_FLAG_WRITE;
1142 * When running in the kernel we expect faults to occur only to
1143 * addresses in user space. All other faults represent errors in
1144 * the kernel and should generate an OOPS. Unfortunately, in the
1145 * case of an erroneous fault occurring in a code path which already
1146 * holds mmap_sem we will deadlock attempting to validate the fault
1147 * against the address space. Luckily the kernel only validly
1148 * references user space from well defined areas of code, which are
1149 * listed in the exceptions table.
1151 * As the vast majority of faults will be valid we will only perform
1152 * the source reference check when there is a possibility of a
1153 * deadlock. Attempt to lock the address space, if we cannot we then
1154 * validate the source. If this is invalid we can skip the address
1155 * space check, thus avoiding the deadlock:
1157 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1158 if ((error_code & PF_USER) == 0 &&
1159 !search_exception_tables(regs->ip)) {
1160 bad_area_nosemaphore(regs, error_code, address);
1164 down_read(&mm->mmap_sem);
1167 * The above down_read_trylock() might have succeeded in
1168 * which case we'll have missed the might_sleep() from
1174 vma = find_vma(mm, address);
1175 if (unlikely(!vma)) {
1176 bad_area(regs, error_code, address);
1179 if (likely(vma->vm_start <= address))
1181 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1182 bad_area(regs, error_code, address);
1185 if (error_code & PF_USER) {
1187 * Accessing the stack below %sp is always a bug.
1188 * The large cushion allows instructions like enter
1189 * and pusha to work. ("enter $65535, $31" pushes
1190 * 32 pointers and then decrements %sp by 65535.)
1192 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1193 bad_area(regs, error_code, address);
1197 if (unlikely(expand_stack(vma, address))) {
1198 bad_area(regs, error_code, address);
1203 * Ok, we have a good vm_area for this memory access, so
1204 * we can handle it..
1207 if (unlikely(access_error(error_code, vma))) {
1208 bad_area_access_error(regs, error_code, address);
1213 * If for any reason at all we couldn't handle the fault,
1214 * make sure we exit gracefully rather than endlessly redo
1217 fault = handle_mm_fault(mm, vma, address, flags);
1220 * If we need to retry but a fatal signal is pending, handle the
1221 * signal first. We do not need to release the mmap_sem because it
1222 * would already be released in __lock_page_or_retry in mm/filemap.c.
1224 if (unlikely((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)))
1227 if (unlikely(fault & VM_FAULT_ERROR)) {
1228 mm_fault_error(regs, error_code, address, fault);
1233 * Major/minor page fault accounting is only done on the
1234 * initial attempt. If we go through a retry, it is extremely
1235 * likely that the page will be found in page cache at that point.
1237 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1238 if (fault & VM_FAULT_MAJOR) {
1240 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1244 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1247 if (fault & VM_FAULT_RETRY) {
1248 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1250 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1251 flags |= FAULT_FLAG_TRIED;
1256 check_v8086_mode(regs, address, tsk);
1258 up_read(&mm->mmap_sem);
1260 NOKPROBE_SYMBOL(__do_page_fault);
1262 dotraplinkage void notrace
1263 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1265 unsigned long address = read_cr2(); /* Get the faulting address */
1266 enum ctx_state prev_state;
1269 * We must have this function tagged with __kprobes, notrace and call
1270 * read_cr2() before calling anything else. To avoid calling any kind
1271 * of tracing machinery before we've observed the CR2 value.
1273 * exception_{enter,exit}() contain all sorts of tracepoints.
1276 prev_state = exception_enter();
1277 __do_page_fault(regs, error_code, address);
1278 exception_exit(prev_state);
1280 NOKPROBE_SYMBOL(do_page_fault);
1282 #ifdef CONFIG_TRACING
1283 static nokprobe_inline void
1284 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1285 unsigned long error_code)
1287 if (user_mode(regs))
1288 trace_page_fault_user(address, regs, error_code);
1290 trace_page_fault_kernel(address, regs, error_code);
1293 dotraplinkage void notrace
1294 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1297 * The exception_enter and tracepoint processing could
1298 * trigger another page faults (user space callchain
1299 * reading) and destroy the original cr2 value, so read
1300 * the faulting address now.
1302 unsigned long address = read_cr2();
1303 enum ctx_state prev_state;
1305 prev_state = exception_enter();
1306 trace_page_fault_entries(address, regs, error_code);
1307 __do_page_fault(regs, error_code, address);
1308 exception_exit(prev_state);
1310 NOKPROBE_SYMBOL(trace_do_page_fault);
1311 #endif /* CONFIG_TRACING */