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_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 nokprobe_inline int
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 nokprobe_inline int 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 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))
294 NOKPROBE_SYMBOL(vmalloc_fault);
297 * Did it hit the DOS screen memory VA from vm86 mode?
300 check_v8086_mode(struct pt_regs *regs, unsigned long address,
301 struct task_struct *tsk)
305 if (!v8086_mode(regs))
308 bit = (address - 0xA0000) >> PAGE_SHIFT;
310 tsk->thread.screen_bitmap |= 1 << bit;
313 static bool low_pfn(unsigned long pfn)
315 return pfn < max_low_pfn;
318 static void dump_pagetable(unsigned long address)
320 pgd_t *base = __va(read_cr3());
321 pgd_t *pgd = &base[pgd_index(address)];
325 #ifdef CONFIG_X86_PAE
326 printk("*pdpt = %016Lx ", pgd_val(*pgd));
327 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
330 pmd = pmd_offset(pud_offset(pgd, address), address);
331 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
334 * We must not directly access the pte in the highpte
335 * case if the page table is located in highmem.
336 * And let's rather not kmap-atomic the pte, just in case
337 * it's allocated already:
339 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
342 pte = pte_offset_kernel(pmd, address);
343 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
348 #else /* CONFIG_X86_64: */
350 void vmalloc_sync_all(void)
352 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
358 * Handle a fault on the vmalloc area
360 * This assumes no large pages in there.
362 static noinline int vmalloc_fault(unsigned long address)
364 pgd_t *pgd, *pgd_ref;
365 pud_t *pud, *pud_ref;
366 pmd_t *pmd, *pmd_ref;
367 pte_t *pte, *pte_ref;
369 /* Make sure we are in vmalloc area: */
370 if (!(address >= VMALLOC_START && address < VMALLOC_END))
373 WARN_ON_ONCE(in_nmi());
376 * Copy kernel mappings over when needed. This can also
377 * happen within a race in page table update. In the later
380 pgd = pgd_offset(current->active_mm, address);
381 pgd_ref = pgd_offset_k(address);
382 if (pgd_none(*pgd_ref))
385 if (pgd_none(*pgd)) {
386 set_pgd(pgd, *pgd_ref);
387 arch_flush_lazy_mmu_mode();
389 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
393 * Below here mismatches are bugs because these lower tables
397 pud = pud_offset(pgd, address);
398 pud_ref = pud_offset(pgd_ref, address);
399 if (pud_none(*pud_ref))
402 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
405 pmd = pmd_offset(pud, address);
406 pmd_ref = pmd_offset(pud_ref, address);
407 if (pmd_none(*pmd_ref))
410 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
413 pte_ref = pte_offset_kernel(pmd_ref, address);
414 if (!pte_present(*pte_ref))
417 pte = pte_offset_kernel(pmd, address);
420 * Don't use pte_page here, because the mappings can point
421 * outside mem_map, and the NUMA hash lookup cannot handle
424 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
429 NOKPROBE_SYMBOL(vmalloc_fault);
431 #ifdef CONFIG_CPU_SUP_AMD
432 static const char errata93_warning[] =
434 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
435 "******* Working around it, but it may cause SEGVs or burn power.\n"
436 "******* Please consider a BIOS update.\n"
437 "******* Disabling USB legacy in the BIOS may also help.\n";
441 * No vm86 mode in 64-bit mode:
444 check_v8086_mode(struct pt_regs *regs, unsigned long address,
445 struct task_struct *tsk)
449 static int bad_address(void *p)
453 return probe_kernel_address((unsigned long *)p, dummy);
456 static void dump_pagetable(unsigned long address)
458 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
459 pgd_t *pgd = base + pgd_index(address);
464 if (bad_address(pgd))
467 printk("PGD %lx ", pgd_val(*pgd));
469 if (!pgd_present(*pgd))
472 pud = pud_offset(pgd, address);
473 if (bad_address(pud))
476 printk("PUD %lx ", pud_val(*pud));
477 if (!pud_present(*pud) || pud_large(*pud))
480 pmd = pmd_offset(pud, address);
481 if (bad_address(pmd))
484 printk("PMD %lx ", pmd_val(*pmd));
485 if (!pmd_present(*pmd) || pmd_large(*pmd))
488 pte = pte_offset_kernel(pmd, address);
489 if (bad_address(pte))
492 printk("PTE %lx", pte_val(*pte));
500 #endif /* CONFIG_X86_64 */
503 * Workaround for K8 erratum #93 & buggy BIOS.
505 * BIOS SMM functions are required to use a specific workaround
506 * to avoid corruption of the 64bit RIP register on C stepping K8.
508 * A lot of BIOS that didn't get tested properly miss this.
510 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
511 * Try to work around it here.
513 * Note we only handle faults in kernel here.
514 * Does nothing on 32-bit.
516 static int is_errata93(struct pt_regs *regs, unsigned long address)
518 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
519 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
520 || boot_cpu_data.x86 != 0xf)
523 if (address != regs->ip)
526 if ((address >> 32) != 0)
529 address |= 0xffffffffUL << 32;
530 if ((address >= (u64)_stext && address <= (u64)_etext) ||
531 (address >= MODULES_VADDR && address <= MODULES_END)) {
532 printk_once(errata93_warning);
541 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
542 * to illegal addresses >4GB.
544 * We catch this in the page fault handler because these addresses
545 * are not reachable. Just detect this case and return. Any code
546 * segment in LDT is compatibility mode.
548 static int is_errata100(struct pt_regs *regs, unsigned long address)
551 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
557 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
559 #ifdef CONFIG_X86_F00F_BUG
563 * Pentium F0 0F C7 C8 bug workaround:
565 if (boot_cpu_has_bug(X86_BUG_F00F)) {
566 nr = (address - idt_descr.address) >> 3;
569 do_invalid_op(regs, 0);
577 static const char nx_warning[] = KERN_CRIT
578 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
581 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
582 unsigned long address)
584 if (!oops_may_print())
587 if (error_code & PF_INSTR) {
592 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
593 pgd += pgd_index(address);
595 pte = lookup_address_in_pgd(pgd, address, &level);
597 if (pte && pte_present(*pte) && !pte_exec(*pte))
598 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
601 printk(KERN_ALERT "BUG: unable to handle kernel ");
602 if (address < PAGE_SIZE)
603 printk(KERN_CONT "NULL pointer dereference");
605 printk(KERN_CONT "paging request");
607 printk(KERN_CONT " at %p\n", (void *) address);
608 printk(KERN_ALERT "IP:");
609 printk_address(regs->ip);
611 dump_pagetable(address);
615 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
616 unsigned long address)
618 struct task_struct *tsk;
622 flags = oops_begin();
626 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
628 dump_pagetable(address);
630 tsk->thread.cr2 = address;
631 tsk->thread.trap_nr = X86_TRAP_PF;
632 tsk->thread.error_code = error_code;
634 if (__die("Bad pagetable", regs, error_code))
637 oops_end(flags, regs, sig);
641 no_context(struct pt_regs *regs, unsigned long error_code,
642 unsigned long address, int signal, int si_code)
644 struct task_struct *tsk = current;
645 unsigned long *stackend;
649 /* Are we prepared to handle this kernel fault? */
650 if (fixup_exception(regs)) {
652 * Any interrupt that takes a fault gets the fixup. This makes
653 * the below recursive fault logic only apply to a faults from
660 * Per the above we're !in_interrupt(), aka. task context.
662 * In this case we need to make sure we're not recursively
663 * faulting through the emulate_vsyscall() logic.
665 if (current_thread_info()->sig_on_uaccess_error && signal) {
666 tsk->thread.trap_nr = X86_TRAP_PF;
667 tsk->thread.error_code = error_code | PF_USER;
668 tsk->thread.cr2 = address;
670 /* XXX: hwpoison faults will set the wrong code. */
671 force_sig_info_fault(signal, si_code, address, tsk, 0);
675 * Barring that, we can do the fixup and be happy.
683 * Valid to do another page fault here, because if this fault
684 * had been triggered by is_prefetch fixup_exception would have
689 * Hall of shame of CPU/BIOS bugs.
691 if (is_prefetch(regs, error_code, address))
694 if (is_errata93(regs, address))
698 * Oops. The kernel tried to access some bad page. We'll have to
699 * terminate things with extreme prejudice:
701 flags = oops_begin();
703 show_fault_oops(regs, error_code, address);
705 stackend = end_of_stack(tsk);
706 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
707 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
709 tsk->thread.cr2 = address;
710 tsk->thread.trap_nr = X86_TRAP_PF;
711 tsk->thread.error_code = error_code;
714 if (__die("Oops", regs, error_code))
717 /* Executive summary in case the body of the oops scrolled away */
718 printk(KERN_DEFAULT "CR2: %016lx\n", address);
720 oops_end(flags, regs, sig);
724 * Print out info about fatal segfaults, if the show_unhandled_signals
728 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
729 unsigned long address, struct task_struct *tsk)
731 if (!unhandled_signal(tsk, SIGSEGV))
734 if (!printk_ratelimit())
737 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
738 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
739 tsk->comm, task_pid_nr(tsk), address,
740 (void *)regs->ip, (void *)regs->sp, error_code);
742 print_vma_addr(KERN_CONT " in ", regs->ip);
744 printk(KERN_CONT "\n");
748 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
749 unsigned long address, int si_code)
751 struct task_struct *tsk = current;
753 /* User mode accesses just cause a SIGSEGV */
754 if (error_code & PF_USER) {
756 * It's possible to have interrupts off here:
761 * Valid to do another page fault here because this one came
764 if (is_prefetch(regs, error_code, address))
767 if (is_errata100(regs, address))
772 * Instruction fetch faults in the vsyscall page might need
775 if (unlikely((error_code & PF_INSTR) &&
776 ((address & ~0xfff) == VSYSCALL_START))) {
777 if (emulate_vsyscall(regs, address))
781 /* Kernel addresses are always protection faults: */
782 if (address >= TASK_SIZE)
783 error_code |= PF_PROT;
785 if (likely(show_unhandled_signals))
786 show_signal_msg(regs, error_code, address, tsk);
788 tsk->thread.cr2 = address;
789 tsk->thread.error_code = error_code;
790 tsk->thread.trap_nr = X86_TRAP_PF;
792 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
797 if (is_f00f_bug(regs, address))
800 no_context(regs, error_code, address, SIGSEGV, si_code);
804 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
805 unsigned long address)
807 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
811 __bad_area(struct pt_regs *regs, unsigned long error_code,
812 unsigned long address, int si_code)
814 struct mm_struct *mm = current->mm;
817 * Something tried to access memory that isn't in our memory map..
818 * Fix it, but check if it's kernel or user first..
820 up_read(&mm->mmap_sem);
822 __bad_area_nosemaphore(regs, error_code, address, si_code);
826 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
828 __bad_area(regs, error_code, address, SEGV_MAPERR);
832 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
833 unsigned long address)
835 __bad_area(regs, error_code, address, SEGV_ACCERR);
839 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
842 struct task_struct *tsk = current;
843 struct mm_struct *mm = tsk->mm;
844 int code = BUS_ADRERR;
846 up_read(&mm->mmap_sem);
848 /* Kernel mode? Handle exceptions or die: */
849 if (!(error_code & PF_USER)) {
850 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
854 /* User-space => ok to do another page fault: */
855 if (is_prefetch(regs, error_code, address))
858 tsk->thread.cr2 = address;
859 tsk->thread.error_code = error_code;
860 tsk->thread.trap_nr = X86_TRAP_PF;
862 #ifdef CONFIG_MEMORY_FAILURE
863 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
865 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
866 tsk->comm, tsk->pid, address);
867 code = BUS_MCEERR_AR;
870 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
874 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
875 unsigned long address, unsigned int fault)
877 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
878 up_read(¤t->mm->mmap_sem);
879 no_context(regs, error_code, address, 0, 0);
883 if (fault & VM_FAULT_OOM) {
884 /* Kernel mode? Handle exceptions or die: */
885 if (!(error_code & PF_USER)) {
886 up_read(¤t->mm->mmap_sem);
887 no_context(regs, error_code, address,
888 SIGSEGV, SEGV_MAPERR);
892 up_read(¤t->mm->mmap_sem);
895 * We ran out of memory, call the OOM killer, and return the
896 * userspace (which will retry the fault, or kill us if we got
899 pagefault_out_of_memory();
901 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
902 VM_FAULT_HWPOISON_LARGE))
903 do_sigbus(regs, error_code, address, fault);
909 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
911 if ((error_code & PF_WRITE) && !pte_write(*pte))
914 if ((error_code & PF_INSTR) && !pte_exec(*pte))
921 * Handle a spurious fault caused by a stale TLB entry.
923 * This allows us to lazily refresh the TLB when increasing the
924 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
925 * eagerly is very expensive since that implies doing a full
926 * cross-processor TLB flush, even if no stale TLB entries exist
927 * on other processors.
929 * There are no security implications to leaving a stale TLB when
930 * increasing the permissions on a page.
933 spurious_fault(unsigned long error_code, unsigned long address)
941 /* Reserved-bit violation or user access to kernel space? */
942 if (error_code & (PF_USER | PF_RSVD))
945 pgd = init_mm.pgd + pgd_index(address);
946 if (!pgd_present(*pgd))
949 pud = pud_offset(pgd, address);
950 if (!pud_present(*pud))
954 return spurious_fault_check(error_code, (pte_t *) pud);
956 pmd = pmd_offset(pud, address);
957 if (!pmd_present(*pmd))
961 return spurious_fault_check(error_code, (pte_t *) pmd);
963 pte = pte_offset_kernel(pmd, address);
964 if (!pte_present(*pte))
967 ret = spurious_fault_check(error_code, pte);
972 * Make sure we have permissions in PMD.
973 * If not, then there's a bug in the page tables:
975 ret = spurious_fault_check(error_code, (pte_t *) pmd);
976 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
980 NOKPROBE_SYMBOL(spurious_fault);
982 int show_unhandled_signals = 1;
985 access_error(unsigned long error_code, struct vm_area_struct *vma)
987 if (error_code & PF_WRITE) {
988 /* write, present and write, not present: */
989 if (unlikely(!(vma->vm_flags & VM_WRITE)))
995 if (unlikely(error_code & PF_PROT))
998 /* read, not present: */
999 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1005 static int fault_in_kernel_space(unsigned long address)
1007 return address >= TASK_SIZE_MAX;
1010 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1012 if (!IS_ENABLED(CONFIG_X86_SMAP))
1015 if (!static_cpu_has(X86_FEATURE_SMAP))
1018 if (error_code & PF_USER)
1021 if (!user_mode_vm(regs) && (regs->flags & X86_EFLAGS_AC))
1028 * This routine handles page faults. It determines the address,
1029 * and the problem, and then passes it off to one of the appropriate
1032 * This function must have noinline because both callers
1033 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1034 * guarantees there's a function trace entry.
1036 static noinline void
1037 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1038 unsigned long address)
1040 struct vm_area_struct *vma;
1041 struct task_struct *tsk;
1042 struct mm_struct *mm;
1044 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1050 * Detect and handle instructions that would cause a page fault for
1051 * both a tracked kernel page and a userspace page.
1053 if (kmemcheck_active(regs))
1054 kmemcheck_hide(regs);
1055 prefetchw(&mm->mmap_sem);
1057 if (unlikely(kmmio_fault(regs, address)))
1061 * We fault-in kernel-space virtual memory on-demand. The
1062 * 'reference' page table is init_mm.pgd.
1064 * NOTE! We MUST NOT take any locks for this case. We may
1065 * be in an interrupt or a critical region, and should
1066 * only copy the information from the master page table,
1069 * This verifies that the fault happens in kernel space
1070 * (error_code & 4) == 0, and that the fault was not a
1071 * protection error (error_code & 9) == 0.
1073 if (unlikely(fault_in_kernel_space(address))) {
1074 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1075 if (vmalloc_fault(address) >= 0)
1078 if (kmemcheck_fault(regs, address, error_code))
1082 /* Can handle a stale RO->RW TLB: */
1083 if (spurious_fault(error_code, address))
1086 /* kprobes don't want to hook the spurious faults: */
1087 if (kprobes_fault(regs))
1090 * Don't take the mm semaphore here. If we fixup a prefetch
1091 * fault we could otherwise deadlock:
1093 bad_area_nosemaphore(regs, error_code, address);
1098 /* kprobes don't want to hook the spurious faults: */
1099 if (unlikely(kprobes_fault(regs)))
1102 if (unlikely(error_code & PF_RSVD))
1103 pgtable_bad(regs, error_code, address);
1105 if (unlikely(smap_violation(error_code, regs))) {
1106 bad_area_nosemaphore(regs, error_code, address);
1111 * If we're in an interrupt, have no user context or are running
1112 * in an atomic region then we must not take the fault:
1114 if (unlikely(in_atomic() || !mm)) {
1115 bad_area_nosemaphore(regs, error_code, address);
1120 * It's safe to allow irq's after cr2 has been saved and the
1121 * vmalloc fault has been handled.
1123 * User-mode registers count as a user access even for any
1124 * potential system fault or CPU buglet:
1126 if (user_mode_vm(regs)) {
1128 error_code |= PF_USER;
1129 flags |= FAULT_FLAG_USER;
1131 if (regs->flags & X86_EFLAGS_IF)
1135 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1137 if (error_code & PF_WRITE)
1138 flags |= FAULT_FLAG_WRITE;
1141 * When running in the kernel we expect faults to occur only to
1142 * addresses in user space. All other faults represent errors in
1143 * the kernel and should generate an OOPS. Unfortunately, in the
1144 * case of an erroneous fault occurring in a code path which already
1145 * holds mmap_sem we will deadlock attempting to validate the fault
1146 * against the address space. Luckily the kernel only validly
1147 * references user space from well defined areas of code, which are
1148 * listed in the exceptions table.
1150 * As the vast majority of faults will be valid we will only perform
1151 * the source reference check when there is a possibility of a
1152 * deadlock. Attempt to lock the address space, if we cannot we then
1153 * validate the source. If this is invalid we can skip the address
1154 * space check, thus avoiding the deadlock:
1156 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1157 if ((error_code & PF_USER) == 0 &&
1158 !search_exception_tables(regs->ip)) {
1159 bad_area_nosemaphore(regs, error_code, address);
1163 down_read(&mm->mmap_sem);
1166 * The above down_read_trylock() might have succeeded in
1167 * which case we'll have missed the might_sleep() from
1173 vma = find_vma(mm, address);
1174 if (unlikely(!vma)) {
1175 bad_area(regs, error_code, address);
1178 if (likely(vma->vm_start <= address))
1180 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1181 bad_area(regs, error_code, address);
1184 if (error_code & PF_USER) {
1186 * Accessing the stack below %sp is always a bug.
1187 * The large cushion allows instructions like enter
1188 * and pusha to work. ("enter $65535, $31" pushes
1189 * 32 pointers and then decrements %sp by 65535.)
1191 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1192 bad_area(regs, error_code, address);
1196 if (unlikely(expand_stack(vma, address))) {
1197 bad_area(regs, error_code, address);
1202 * Ok, we have a good vm_area for this memory access, so
1203 * we can handle it..
1206 if (unlikely(access_error(error_code, vma))) {
1207 bad_area_access_error(regs, error_code, address);
1212 * If for any reason at all we couldn't handle the fault,
1213 * make sure we exit gracefully rather than endlessly redo
1216 fault = handle_mm_fault(mm, vma, address, flags);
1219 * If we need to retry but a fatal signal is pending, handle the
1220 * signal first. We do not need to release the mmap_sem because it
1221 * would already be released in __lock_page_or_retry in mm/filemap.c.
1223 if (unlikely((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)))
1226 if (unlikely(fault & VM_FAULT_ERROR)) {
1227 mm_fault_error(regs, error_code, address, fault);
1232 * Major/minor page fault accounting is only done on the
1233 * initial attempt. If we go through a retry, it is extremely
1234 * likely that the page will be found in page cache at that point.
1236 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1237 if (fault & VM_FAULT_MAJOR) {
1239 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1243 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1246 if (fault & VM_FAULT_RETRY) {
1247 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1249 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1250 flags |= FAULT_FLAG_TRIED;
1255 check_v8086_mode(regs, address, tsk);
1257 up_read(&mm->mmap_sem);
1259 NOKPROBE_SYMBOL(__do_page_fault);
1261 dotraplinkage void notrace
1262 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1264 unsigned long address = read_cr2(); /* Get the faulting address */
1265 enum ctx_state prev_state;
1268 * We must have this function tagged with __kprobes, notrace and call
1269 * read_cr2() before calling anything else. To avoid calling any kind
1270 * of tracing machinery before we've observed the CR2 value.
1272 * exception_{enter,exit}() contain all sorts of tracepoints.
1275 prev_state = exception_enter();
1276 __do_page_fault(regs, error_code, address);
1277 exception_exit(prev_state);
1279 NOKPROBE_SYMBOL(do_page_fault);
1281 #ifdef CONFIG_TRACING
1282 static nokprobe_inline void
1283 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1284 unsigned long error_code)
1286 if (user_mode(regs))
1287 trace_page_fault_user(address, regs, error_code);
1289 trace_page_fault_kernel(address, regs, error_code);
1292 dotraplinkage void notrace
1293 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1296 * The exception_enter and tracepoint processing could
1297 * trigger another page faults (user space callchain
1298 * reading) and destroy the original cr2 value, so read
1299 * the faulting address now.
1301 unsigned long address = read_cr2();
1302 enum ctx_state prev_state;
1304 prev_state = exception_enter();
1305 trace_page_fault_entries(address, regs, error_code);
1306 __do_page_fault(regs, error_code, address);
1307 exception_exit(prev_state);
1309 NOKPROBE_SYMBOL(trace_do_page_fault);
1310 #endif /* CONFIG_TRACING */