1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 1995 Linus Torvalds
4 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
5 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
7 #include <linux/sched.h> /* test_thread_flag(), ... */
8 #include <linux/sched/task_stack.h> /* task_stack_*(), ... */
9 #include <linux/kdebug.h> /* oops_begin/end, ... */
10 #include <linux/extable.h> /* search_exception_tables */
11 #include <linux/bootmem.h> /* max_low_pfn */
12 #include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */
13 #include <linux/mmiotrace.h> /* kmmio_handler, ... */
14 #include <linux/perf_event.h> /* perf_sw_event */
15 #include <linux/hugetlb.h> /* hstate_index_to_shift */
16 #include <linux/prefetch.h> /* prefetchw */
17 #include <linux/context_tracking.h> /* exception_enter(), ... */
18 #include <linux/uaccess.h> /* faulthandler_disabled() */
19 #include <linux/mm_types.h>
21 #include <asm/cpufeature.h> /* boot_cpu_has, ... */
22 #include <asm/traps.h> /* dotraplinkage, ... */
23 #include <asm/pgalloc.h> /* pgd_*(), ... */
24 #include <asm/fixmap.h> /* VSYSCALL_ADDR */
25 #include <asm/vsyscall.h> /* emulate_vsyscall */
26 #include <asm/vm86.h> /* struct vm86 */
27 #include <asm/mmu_context.h> /* vma_pkey() */
29 #define CREATE_TRACE_POINTS
30 #include <asm/trace/exceptions.h>
33 * Returns 0 if mmiotrace is disabled, or if the fault is not
34 * handled by mmiotrace:
36 static nokprobe_inline int
37 kmmio_fault(struct pt_regs *regs, unsigned long addr)
39 if (unlikely(is_kmmio_active()))
40 if (kmmio_handler(regs, addr) == 1)
45 static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
49 /* kprobe_running() needs smp_processor_id() */
50 if (kprobes_built_in() && !user_mode(regs)) {
52 if (kprobe_running() && kprobe_fault_handler(regs, 14))
65 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
66 * Check that here and ignore it.
70 * Sometimes the CPU reports invalid exceptions on prefetch.
71 * Check that here and ignore it.
73 * Opcode checker based on code by Richard Brunner.
76 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
77 unsigned char opcode, int *prefetch)
79 unsigned char instr_hi = opcode & 0xf0;
80 unsigned char instr_lo = opcode & 0x0f;
86 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
87 * In X86_64 long mode, the CPU will signal invalid
88 * opcode if some of these prefixes are present so
89 * X86_64 will never get here anyway
91 return ((instr_lo & 7) == 0x6);
95 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
96 * Need to figure out under what instruction mode the
97 * instruction was issued. Could check the LDT for lm,
98 * but for now it's good enough to assume that long
99 * mode only uses well known segments or kernel.
101 return (!user_mode(regs) || user_64bit_mode(regs));
104 /* 0x64 thru 0x67 are valid prefixes in all modes. */
105 return (instr_lo & 0xC) == 0x4;
107 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
108 return !instr_lo || (instr_lo>>1) == 1;
110 /* Prefetch instruction is 0x0F0D or 0x0F18 */
111 if (probe_kernel_address(instr, opcode))
114 *prefetch = (instr_lo == 0xF) &&
115 (opcode == 0x0D || opcode == 0x18);
123 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
125 unsigned char *max_instr;
126 unsigned char *instr;
130 * If it was a exec (instruction fetch) fault on NX page, then
131 * do not ignore the fault:
133 if (error_code & X86_PF_INSTR)
136 instr = (void *)convert_ip_to_linear(current, regs);
137 max_instr = instr + 15;
139 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
142 while (instr < max_instr) {
143 unsigned char opcode;
145 if (probe_kernel_address(instr, opcode))
150 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
157 * A protection key fault means that the PKRU value did not allow
158 * access to some PTE. Userspace can figure out what PKRU was
159 * from the XSAVE state, and this function fills out a field in
160 * siginfo so userspace can discover which protection key was set
163 * If we get here, we know that the hardware signaled a X86_PF_PK
164 * fault and that there was a VMA once we got in the fault
165 * handler. It does *not* guarantee that the VMA we find here
166 * was the one that we faulted on.
168 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
169 * 2. T1 : set PKRU to deny access to pkey=4, touches page
171 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
172 * 5. T1 : enters fault handler, takes mmap_sem, etc...
173 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
174 * faulted on a pte with its pkey=4.
176 static void fill_sig_info_pkey(int si_signo, int si_code, siginfo_t *info,
179 /* This is effectively an #ifdef */
180 if (!boot_cpu_has(X86_FEATURE_OSPKE))
183 /* Fault not from Protection Keys: nothing to do */
184 if ((si_code != SEGV_PKUERR) || (si_signo != SIGSEGV))
187 * force_sig_info_fault() is called from a number of
188 * contexts, some of which have a VMA and some of which
189 * do not. The X86_PF_PK handing happens after we have a
190 * valid VMA, so we should never reach this without a
194 WARN_ONCE(1, "PKU fault with no VMA passed in");
199 * si_pkey should be thought of as a strong hint, but not
200 * absolutely guranteed to be 100% accurate because of
201 * the race explained above.
203 info->si_pkey = *pkey;
207 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
208 struct task_struct *tsk, u32 *pkey, int fault)
213 clear_siginfo(&info);
214 info.si_signo = si_signo;
216 info.si_code = si_code;
217 info.si_addr = (void __user *)address;
218 if (fault & VM_FAULT_HWPOISON_LARGE)
219 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
220 if (fault & VM_FAULT_HWPOISON)
222 info.si_addr_lsb = lsb;
224 fill_sig_info_pkey(si_signo, si_code, &info, pkey);
226 force_sig_info(si_signo, &info, tsk);
229 DEFINE_SPINLOCK(pgd_lock);
233 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
235 unsigned index = pgd_index(address);
242 pgd_k = init_mm.pgd + index;
244 if (!pgd_present(*pgd_k))
248 * set_pgd(pgd, *pgd_k); here would be useless on PAE
249 * and redundant with the set_pmd() on non-PAE. As would
252 p4d = p4d_offset(pgd, address);
253 p4d_k = p4d_offset(pgd_k, address);
254 if (!p4d_present(*p4d_k))
257 pud = pud_offset(p4d, address);
258 pud_k = pud_offset(p4d_k, address);
259 if (!pud_present(*pud_k))
262 pmd = pmd_offset(pud, address);
263 pmd_k = pmd_offset(pud_k, address);
264 if (!pmd_present(*pmd_k))
267 if (!pmd_present(*pmd))
268 set_pmd(pmd, *pmd_k);
270 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
275 void vmalloc_sync_all(void)
277 unsigned long address;
279 if (SHARED_KERNEL_PMD)
282 for (address = VMALLOC_START & PMD_MASK;
283 address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
284 address += PMD_SIZE) {
287 spin_lock(&pgd_lock);
288 list_for_each_entry(page, &pgd_list, lru) {
289 spinlock_t *pgt_lock;
292 /* the pgt_lock only for Xen */
293 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
296 ret = vmalloc_sync_one(page_address(page), address);
297 spin_unlock(pgt_lock);
302 spin_unlock(&pgd_lock);
309 * Handle a fault on the vmalloc or module mapping area
311 static noinline int vmalloc_fault(unsigned long address)
313 unsigned long pgd_paddr;
317 /* Make sure we are in vmalloc area: */
318 if (!(address >= VMALLOC_START && address < VMALLOC_END))
322 * Synchronize this task's top level page-table
323 * with the 'reference' page table.
325 * Do _not_ use "current" here. We might be inside
326 * an interrupt in the middle of a task switch..
328 pgd_paddr = read_cr3_pa();
329 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
333 if (pmd_large(*pmd_k))
336 pte_k = pte_offset_kernel(pmd_k, address);
337 if (!pte_present(*pte_k))
342 NOKPROBE_SYMBOL(vmalloc_fault);
345 * Did it hit the DOS screen memory VA from vm86 mode?
348 check_v8086_mode(struct pt_regs *regs, unsigned long address,
349 struct task_struct *tsk)
354 if (!v8086_mode(regs) || !tsk->thread.vm86)
357 bit = (address - 0xA0000) >> PAGE_SHIFT;
359 tsk->thread.vm86->screen_bitmap |= 1 << bit;
363 static bool low_pfn(unsigned long pfn)
365 return pfn < max_low_pfn;
368 static void dump_pagetable(unsigned long address)
370 pgd_t *base = __va(read_cr3_pa());
371 pgd_t *pgd = &base[pgd_index(address)];
377 #ifdef CONFIG_X86_PAE
378 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
379 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
381 #define pr_pde pr_cont
383 #define pr_pde pr_info
385 p4d = p4d_offset(pgd, address);
386 pud = pud_offset(p4d, address);
387 pmd = pmd_offset(pud, address);
388 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
392 * We must not directly access the pte in the highpte
393 * case if the page table is located in highmem.
394 * And let's rather not kmap-atomic the pte, just in case
395 * it's allocated already:
397 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
400 pte = pte_offset_kernel(pmd, address);
401 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
406 #else /* CONFIG_X86_64: */
408 void vmalloc_sync_all(void)
410 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
416 * Handle a fault on the vmalloc area
418 static noinline int vmalloc_fault(unsigned long address)
426 /* Make sure we are in vmalloc area: */
427 if (!(address >= VMALLOC_START && address < VMALLOC_END))
430 WARN_ON_ONCE(in_nmi());
433 * Copy kernel mappings over when needed. This can also
434 * happen within a race in page table update. In the later
437 pgd = (pgd_t *)__va(read_cr3_pa()) + pgd_index(address);
438 pgd_k = pgd_offset_k(address);
439 if (pgd_none(*pgd_k))
442 if (pgtable_l5_enabled()) {
443 if (pgd_none(*pgd)) {
444 set_pgd(pgd, *pgd_k);
445 arch_flush_lazy_mmu_mode();
447 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_k));
451 /* With 4-level paging, copying happens on the p4d level. */
452 p4d = p4d_offset(pgd, address);
453 p4d_k = p4d_offset(pgd_k, address);
454 if (p4d_none(*p4d_k))
457 if (p4d_none(*p4d) && !pgtable_l5_enabled()) {
458 set_p4d(p4d, *p4d_k);
459 arch_flush_lazy_mmu_mode();
461 BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_k));
464 BUILD_BUG_ON(CONFIG_PGTABLE_LEVELS < 4);
466 pud = pud_offset(p4d, address);
473 pmd = pmd_offset(pud, address);
480 pte = pte_offset_kernel(pmd, address);
481 if (!pte_present(*pte))
486 NOKPROBE_SYMBOL(vmalloc_fault);
488 #ifdef CONFIG_CPU_SUP_AMD
489 static const char errata93_warning[] =
491 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
492 "******* Working around it, but it may cause SEGVs or burn power.\n"
493 "******* Please consider a BIOS update.\n"
494 "******* Disabling USB legacy in the BIOS may also help.\n";
498 * No vm86 mode in 64-bit mode:
501 check_v8086_mode(struct pt_regs *regs, unsigned long address,
502 struct task_struct *tsk)
506 static int bad_address(void *p)
510 return probe_kernel_address((unsigned long *)p, dummy);
513 static void dump_pagetable(unsigned long address)
515 pgd_t *base = __va(read_cr3_pa());
516 pgd_t *pgd = base + pgd_index(address);
522 if (bad_address(pgd))
525 pr_info("PGD %lx ", pgd_val(*pgd));
527 if (!pgd_present(*pgd))
530 p4d = p4d_offset(pgd, address);
531 if (bad_address(p4d))
534 pr_cont("P4D %lx ", p4d_val(*p4d));
535 if (!p4d_present(*p4d) || p4d_large(*p4d))
538 pud = pud_offset(p4d, address);
539 if (bad_address(pud))
542 pr_cont("PUD %lx ", pud_val(*pud));
543 if (!pud_present(*pud) || pud_large(*pud))
546 pmd = pmd_offset(pud, address);
547 if (bad_address(pmd))
550 pr_cont("PMD %lx ", pmd_val(*pmd));
551 if (!pmd_present(*pmd) || pmd_large(*pmd))
554 pte = pte_offset_kernel(pmd, address);
555 if (bad_address(pte))
558 pr_cont("PTE %lx", pte_val(*pte));
566 #endif /* CONFIG_X86_64 */
569 * Workaround for K8 erratum #93 & buggy BIOS.
571 * BIOS SMM functions are required to use a specific workaround
572 * to avoid corruption of the 64bit RIP register on C stepping K8.
574 * A lot of BIOS that didn't get tested properly miss this.
576 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
577 * Try to work around it here.
579 * Note we only handle faults in kernel here.
580 * Does nothing on 32-bit.
582 static int is_errata93(struct pt_regs *regs, unsigned long address)
584 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
585 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
586 || boot_cpu_data.x86 != 0xf)
589 if (address != regs->ip)
592 if ((address >> 32) != 0)
595 address |= 0xffffffffUL << 32;
596 if ((address >= (u64)_stext && address <= (u64)_etext) ||
597 (address >= MODULES_VADDR && address <= MODULES_END)) {
598 printk_once(errata93_warning);
607 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
608 * to illegal addresses >4GB.
610 * We catch this in the page fault handler because these addresses
611 * are not reachable. Just detect this case and return. Any code
612 * segment in LDT is compatibility mode.
614 static int is_errata100(struct pt_regs *regs, unsigned long address)
617 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
623 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
625 #ifdef CONFIG_X86_F00F_BUG
629 * Pentium F0 0F C7 C8 bug workaround:
631 if (boot_cpu_has_bug(X86_BUG_F00F)) {
632 nr = (address - idt_descr.address) >> 3;
635 do_invalid_op(regs, 0);
644 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
645 unsigned long address)
647 if (!oops_may_print())
650 if (error_code & X86_PF_INSTR) {
655 pgd = __va(read_cr3_pa());
656 pgd += pgd_index(address);
658 pte = lookup_address_in_pgd(pgd, address, &level);
660 if (pte && pte_present(*pte) && !pte_exec(*pte))
661 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
662 from_kuid(&init_user_ns, current_uid()));
663 if (pte && pte_present(*pte) && pte_exec(*pte) &&
664 (pgd_flags(*pgd) & _PAGE_USER) &&
665 (__read_cr4() & X86_CR4_SMEP))
666 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
667 from_kuid(&init_user_ns, current_uid()));
670 pr_alert("BUG: unable to handle kernel %s at %px\n",
671 address < PAGE_SIZE ? "NULL pointer dereference" : "paging request",
674 dump_pagetable(address);
678 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
679 unsigned long address)
681 struct task_struct *tsk;
685 flags = oops_begin();
689 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
691 dump_pagetable(address);
693 tsk->thread.cr2 = address;
694 tsk->thread.trap_nr = X86_TRAP_PF;
695 tsk->thread.error_code = error_code;
697 if (__die("Bad pagetable", regs, error_code))
700 oops_end(flags, regs, sig);
704 no_context(struct pt_regs *regs, unsigned long error_code,
705 unsigned long address, int signal, int si_code)
707 struct task_struct *tsk = current;
711 /* Are we prepared to handle this kernel fault? */
712 if (fixup_exception(regs, X86_TRAP_PF)) {
714 * Any interrupt that takes a fault gets the fixup. This makes
715 * the below recursive fault logic only apply to a faults from
722 * Per the above we're !in_interrupt(), aka. task context.
724 * In this case we need to make sure we're not recursively
725 * faulting through the emulate_vsyscall() logic.
727 if (current->thread.sig_on_uaccess_err && signal) {
728 tsk->thread.trap_nr = X86_TRAP_PF;
729 tsk->thread.error_code = error_code | X86_PF_USER;
730 tsk->thread.cr2 = address;
732 /* XXX: hwpoison faults will set the wrong code. */
733 force_sig_info_fault(signal, si_code, address,
738 * Barring that, we can do the fixup and be happy.
743 #ifdef CONFIG_VMAP_STACK
745 * Stack overflow? During boot, we can fault near the initial
746 * stack in the direct map, but that's not an overflow -- check
747 * that we're in vmalloc space to avoid this.
749 if (is_vmalloc_addr((void *)address) &&
750 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
751 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
752 unsigned long stack = this_cpu_read(orig_ist.ist[DOUBLEFAULT_STACK]) - sizeof(void *);
754 * We're likely to be running with very little stack space
755 * left. It's plausible that we'd hit this condition but
756 * double-fault even before we get this far, in which case
757 * we're fine: the double-fault handler will deal with it.
759 * We don't want to make it all the way into the oops code
760 * and then double-fault, though, because we're likely to
761 * break the console driver and lose most of the stack dump.
763 asm volatile ("movq %[stack], %%rsp\n\t"
764 "call handle_stack_overflow\n\t"
766 : ASM_CALL_CONSTRAINT
767 : "D" ("kernel stack overflow (page fault)"),
768 "S" (regs), "d" (address),
769 [stack] "rm" (stack));
777 * Valid to do another page fault here, because if this fault
778 * had been triggered by is_prefetch fixup_exception would have
783 * Hall of shame of CPU/BIOS bugs.
785 if (is_prefetch(regs, error_code, address))
788 if (is_errata93(regs, address))
792 * Oops. The kernel tried to access some bad page. We'll have to
793 * terminate things with extreme prejudice:
795 flags = oops_begin();
797 show_fault_oops(regs, error_code, address);
799 if (task_stack_end_corrupted(tsk))
800 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
802 tsk->thread.cr2 = address;
803 tsk->thread.trap_nr = X86_TRAP_PF;
804 tsk->thread.error_code = error_code;
807 if (__die("Oops", regs, error_code))
810 /* Executive summary in case the body of the oops scrolled away */
811 printk(KERN_DEFAULT "CR2: %016lx\n", address);
813 oops_end(flags, regs, sig);
817 * Print out info about fatal segfaults, if the show_unhandled_signals
821 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
822 unsigned long address, struct task_struct *tsk)
824 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
826 if (!unhandled_signal(tsk, SIGSEGV))
829 if (!printk_ratelimit())
832 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
833 loglvl, tsk->comm, task_pid_nr(tsk), address,
834 (void *)regs->ip, (void *)regs->sp, error_code);
836 print_vma_addr(KERN_CONT " in ", regs->ip);
838 printk(KERN_CONT "\n");
840 show_opcodes(regs, loglvl);
844 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
845 unsigned long address, u32 *pkey, int si_code)
847 struct task_struct *tsk = current;
849 /* User mode accesses just cause a SIGSEGV */
850 if (error_code & X86_PF_USER) {
852 * It's possible to have interrupts off here:
857 * Valid to do another page fault here because this one came
860 if (is_prefetch(regs, error_code, address))
863 if (is_errata100(regs, address))
868 * Instruction fetch faults in the vsyscall page might need
871 if (unlikely((error_code & X86_PF_INSTR) &&
872 ((address & ~0xfff) == VSYSCALL_ADDR))) {
873 if (emulate_vsyscall(regs, address))
879 * To avoid leaking information about the kernel page table
880 * layout, pretend that user-mode accesses to kernel addresses
881 * are always protection faults.
883 if (address >= TASK_SIZE_MAX)
884 error_code |= X86_PF_PROT;
886 if (likely(show_unhandled_signals))
887 show_signal_msg(regs, error_code, address, tsk);
889 tsk->thread.cr2 = address;
890 tsk->thread.error_code = error_code;
891 tsk->thread.trap_nr = X86_TRAP_PF;
893 force_sig_info_fault(SIGSEGV, si_code, address, tsk, pkey, 0);
898 if (is_f00f_bug(regs, address))
901 no_context(regs, error_code, address, SIGSEGV, si_code);
905 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
906 unsigned long address, u32 *pkey)
908 __bad_area_nosemaphore(regs, error_code, address, pkey, SEGV_MAPERR);
912 __bad_area(struct pt_regs *regs, unsigned long error_code,
913 unsigned long address, struct vm_area_struct *vma, int si_code)
915 struct mm_struct *mm = current->mm;
919 pkey = vma_pkey(vma);
922 * Something tried to access memory that isn't in our memory map..
923 * Fix it, but check if it's kernel or user first..
925 up_read(&mm->mmap_sem);
927 __bad_area_nosemaphore(regs, error_code, address,
928 (vma) ? &pkey : NULL, si_code);
932 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
934 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
937 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
938 struct vm_area_struct *vma)
940 /* This code is always called on the current mm */
941 bool foreign = false;
943 if (!boot_cpu_has(X86_FEATURE_OSPKE))
945 if (error_code & X86_PF_PK)
947 /* this checks permission keys on the VMA: */
948 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
949 (error_code & X86_PF_INSTR), foreign))
955 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
956 unsigned long address, struct vm_area_struct *vma)
959 * This OSPKE check is not strictly necessary at runtime.
960 * But, doing it this way allows compiler optimizations
961 * if pkeys are compiled out.
963 if (bad_area_access_from_pkeys(error_code, vma))
964 __bad_area(regs, error_code, address, vma, SEGV_PKUERR);
966 __bad_area(regs, error_code, address, vma, SEGV_ACCERR);
970 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
971 u32 *pkey, unsigned int fault)
973 struct task_struct *tsk = current;
974 int code = BUS_ADRERR;
976 /* Kernel mode? Handle exceptions or die: */
977 if (!(error_code & X86_PF_USER)) {
978 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
982 /* User-space => ok to do another page fault: */
983 if (is_prefetch(regs, error_code, address))
986 tsk->thread.cr2 = address;
987 tsk->thread.error_code = error_code;
988 tsk->thread.trap_nr = X86_TRAP_PF;
990 #ifdef CONFIG_MEMORY_FAILURE
991 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
993 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
994 tsk->comm, tsk->pid, address);
995 code = BUS_MCEERR_AR;
998 force_sig_info_fault(SIGBUS, code, address, tsk, pkey, fault);
1001 static noinline void
1002 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
1003 unsigned long address, u32 *pkey, vm_fault_t fault)
1005 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) {
1006 no_context(regs, error_code, address, 0, 0);
1010 if (fault & VM_FAULT_OOM) {
1011 /* Kernel mode? Handle exceptions or die: */
1012 if (!(error_code & X86_PF_USER)) {
1013 no_context(regs, error_code, address,
1014 SIGSEGV, SEGV_MAPERR);
1019 * We ran out of memory, call the OOM killer, and return the
1020 * userspace (which will retry the fault, or kill us if we got
1023 pagefault_out_of_memory();
1025 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1026 VM_FAULT_HWPOISON_LARGE))
1027 do_sigbus(regs, error_code, address, pkey, fault);
1028 else if (fault & VM_FAULT_SIGSEGV)
1029 bad_area_nosemaphore(regs, error_code, address, pkey);
1035 static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
1037 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
1040 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1043 * Note: We do not do lazy flushing on protection key
1044 * changes, so no spurious fault will ever set X86_PF_PK.
1046 if ((error_code & X86_PF_PK))
1053 * Handle a spurious fault caused by a stale TLB entry.
1055 * This allows us to lazily refresh the TLB when increasing the
1056 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1057 * eagerly is very expensive since that implies doing a full
1058 * cross-processor TLB flush, even if no stale TLB entries exist
1059 * on other processors.
1061 * Spurious faults may only occur if the TLB contains an entry with
1062 * fewer permission than the page table entry. Non-present (P = 0)
1063 * and reserved bit (R = 1) faults are never spurious.
1065 * There are no security implications to leaving a stale TLB when
1066 * increasing the permissions on a page.
1068 * Returns non-zero if a spurious fault was handled, zero otherwise.
1070 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1071 * (Optional Invalidation).
1074 spurious_kernel_fault(unsigned long error_code, unsigned long address)
1084 * Only writes to RO or instruction fetches from NX may cause
1087 * These could be from user or supervisor accesses but the TLB
1088 * is only lazily flushed after a kernel mapping protection
1089 * change, so user accesses are not expected to cause spurious
1092 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1093 error_code != (X86_PF_INSTR | X86_PF_PROT))
1096 pgd = init_mm.pgd + pgd_index(address);
1097 if (!pgd_present(*pgd))
1100 p4d = p4d_offset(pgd, address);
1101 if (!p4d_present(*p4d))
1104 if (p4d_large(*p4d))
1105 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1107 pud = pud_offset(p4d, address);
1108 if (!pud_present(*pud))
1111 if (pud_large(*pud))
1112 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1114 pmd = pmd_offset(pud, address);
1115 if (!pmd_present(*pmd))
1118 if (pmd_large(*pmd))
1119 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1121 pte = pte_offset_kernel(pmd, address);
1122 if (!pte_present(*pte))
1125 ret = spurious_kernel_fault_check(error_code, pte);
1130 * Make sure we have permissions in PMD.
1131 * If not, then there's a bug in the page tables:
1133 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1134 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1138 NOKPROBE_SYMBOL(spurious_kernel_fault);
1140 int show_unhandled_signals = 1;
1143 access_error(unsigned long error_code, struct vm_area_struct *vma)
1145 /* This is only called for the current mm, so: */
1146 bool foreign = false;
1149 * Read or write was blocked by protection keys. This is
1150 * always an unconditional error and can never result in
1151 * a follow-up action to resolve the fault, like a COW.
1153 if (error_code & X86_PF_PK)
1157 * Make sure to check the VMA so that we do not perform
1158 * faults just to hit a X86_PF_PK as soon as we fill in a
1161 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1162 (error_code & X86_PF_INSTR), foreign))
1165 if (error_code & X86_PF_WRITE) {
1166 /* write, present and write, not present: */
1167 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1172 /* read, present: */
1173 if (unlikely(error_code & X86_PF_PROT))
1176 /* read, not present: */
1177 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1183 static int fault_in_kernel_space(unsigned long address)
1185 return address >= TASK_SIZE_MAX;
1188 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1190 if (!IS_ENABLED(CONFIG_X86_SMAP))
1193 if (!static_cpu_has(X86_FEATURE_SMAP))
1196 if (error_code & X86_PF_USER)
1199 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1206 * Called for all faults where 'address' is part of the kernel address
1207 * space. Might get called for faults that originate from *code* that
1208 * ran in userspace or the kernel.
1211 do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1212 unsigned long address)
1215 * We can fault-in kernel-space virtual memory on-demand. The
1216 * 'reference' page table is init_mm.pgd.
1218 * NOTE! We MUST NOT take any locks for this case. We may
1219 * be in an interrupt or a critical region, and should
1220 * only copy the information from the master page table,
1223 * Before doing this on-demand faulting, ensure that the
1224 * fault is not any of the following:
1225 * 1. A fault on a PTE with a reserved bit set.
1226 * 2. A fault caused by a user-mode access. (Do not demand-
1227 * fault kernel memory due to user-mode accesses).
1228 * 3. A fault caused by a page-level protection violation.
1229 * (A demand fault would be on a non-present page which
1230 * would have X86_PF_PROT==0).
1232 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1233 if (vmalloc_fault(address) >= 0)
1237 /* Was the fault spurious, caused by lazy TLB invalidation? */
1238 if (spurious_kernel_fault(hw_error_code, address))
1241 /* kprobes don't want to hook the spurious faults: */
1242 if (kprobes_fault(regs))
1246 * Note, despite being a "bad area", there are quite a few
1247 * acceptable reasons to get here, such as erratum fixups
1248 * and handling kernel code that can fault, like get_user().
1250 * Don't take the mm semaphore here. If we fixup a prefetch
1251 * fault we could otherwise deadlock:
1253 bad_area_nosemaphore(regs, hw_error_code, address, NULL);
1255 NOKPROBE_SYMBOL(do_kern_addr_fault);
1258 * This routine handles page faults. It determines the address,
1259 * and the problem, and then passes it off to one of the appropriate
1262 static noinline void
1263 __do_page_fault(struct pt_regs *regs, unsigned long hw_error_code,
1264 unsigned long address)
1266 unsigned long sw_error_code;
1267 struct vm_area_struct *vma;
1268 struct task_struct *tsk;
1269 struct mm_struct *mm;
1270 vm_fault_t fault, major = 0;
1271 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1277 prefetchw(&mm->mmap_sem);
1279 if (unlikely(kmmio_fault(regs, address)))
1282 /* Was the fault on kernel-controlled part of the address space? */
1283 if (unlikely(fault_in_kernel_space(address))) {
1284 do_kern_addr_fault(regs, hw_error_code, address);
1288 /* kprobes don't want to hook the spurious faults: */
1289 if (unlikely(kprobes_fault(regs)))
1292 if (unlikely(hw_error_code & X86_PF_RSVD))
1293 pgtable_bad(regs, hw_error_code, address);
1295 if (unlikely(smap_violation(hw_error_code, regs))) {
1296 bad_area_nosemaphore(regs, hw_error_code, address, NULL);
1301 * If we're in an interrupt, have no user context or are running
1302 * in a region with pagefaults disabled then we must not take the fault
1304 if (unlikely(faulthandler_disabled() || !mm)) {
1305 bad_area_nosemaphore(regs, hw_error_code, address, NULL);
1310 * hw_error_code is literally the "page fault error code" passed to
1311 * the kernel directly from the hardware. But, we will shortly be
1312 * modifying it in software, so give it a new name.
1314 sw_error_code = hw_error_code;
1317 * It's safe to allow irq's after cr2 has been saved and the
1318 * vmalloc fault has been handled.
1320 * User-mode registers count as a user access even for any
1321 * potential system fault or CPU buglet:
1323 if (user_mode(regs)) {
1326 * Up to this point, X86_PF_USER set in hw_error_code
1327 * indicated a user-mode access. But, after this,
1328 * X86_PF_USER in sw_error_code will indicate either
1329 * that, *or* an implicit kernel(supervisor)-mode access
1330 * which originated from user mode.
1332 if (!(hw_error_code & X86_PF_USER)) {
1334 * The CPU was in user mode, but the CPU says
1335 * the fault was not a user-mode access.
1336 * Must be an implicit kernel-mode access,
1337 * which we do not expect to happen in the
1338 * user address space.
1340 pr_warn_once("kernel-mode error from user-mode: %lx\n",
1343 sw_error_code |= X86_PF_USER;
1345 flags |= FAULT_FLAG_USER;
1347 if (regs->flags & X86_EFLAGS_IF)
1351 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1353 if (sw_error_code & X86_PF_WRITE)
1354 flags |= FAULT_FLAG_WRITE;
1355 if (sw_error_code & X86_PF_INSTR)
1356 flags |= FAULT_FLAG_INSTRUCTION;
1359 * When running in the kernel we expect faults to occur only to
1360 * addresses in user space. All other faults represent errors in
1361 * the kernel and should generate an OOPS. Unfortunately, in the
1362 * case of an erroneous fault occurring in a code path which already
1363 * holds mmap_sem we will deadlock attempting to validate the fault
1364 * against the address space. Luckily the kernel only validly
1365 * references user space from well defined areas of code, which are
1366 * listed in the exceptions table.
1368 * As the vast majority of faults will be valid we will only perform
1369 * the source reference check when there is a possibility of a
1370 * deadlock. Attempt to lock the address space, if we cannot we then
1371 * validate the source. If this is invalid we can skip the address
1372 * space check, thus avoiding the deadlock:
1374 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1375 if (!(sw_error_code & X86_PF_USER) &&
1376 !search_exception_tables(regs->ip)) {
1377 bad_area_nosemaphore(regs, sw_error_code, address, NULL);
1381 down_read(&mm->mmap_sem);
1384 * The above down_read_trylock() might have succeeded in
1385 * which case we'll have missed the might_sleep() from
1391 vma = find_vma(mm, address);
1392 if (unlikely(!vma)) {
1393 bad_area(regs, sw_error_code, address);
1396 if (likely(vma->vm_start <= address))
1398 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1399 bad_area(regs, sw_error_code, address);
1402 if (sw_error_code & X86_PF_USER) {
1404 * Accessing the stack below %sp is always a bug.
1405 * The large cushion allows instructions like enter
1406 * and pusha to work. ("enter $65535, $31" pushes
1407 * 32 pointers and then decrements %sp by 65535.)
1409 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1410 bad_area(regs, sw_error_code, address);
1414 if (unlikely(expand_stack(vma, address))) {
1415 bad_area(regs, sw_error_code, address);
1420 * Ok, we have a good vm_area for this memory access, so
1421 * we can handle it..
1424 if (unlikely(access_error(sw_error_code, vma))) {
1425 bad_area_access_error(regs, sw_error_code, address, vma);
1430 * If for any reason at all we couldn't handle the fault,
1431 * make sure we exit gracefully rather than endlessly redo
1432 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1433 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1435 * Note that handle_userfault() may also release and reacquire mmap_sem
1436 * (and not return with VM_FAULT_RETRY), when returning to userland to
1437 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1438 * (potentially after handling any pending signal during the return to
1439 * userland). The return to userland is identified whenever
1440 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1441 * Thus we have to be careful about not touching vma after handling the
1442 * fault, so we read the pkey beforehand.
1444 pkey = vma_pkey(vma);
1445 fault = handle_mm_fault(vma, address, flags);
1446 major |= fault & VM_FAULT_MAJOR;
1449 * If we need to retry the mmap_sem has already been released,
1450 * and if there is a fatal signal pending there is no guarantee
1451 * that we made any progress. Handle this case first.
1453 if (unlikely(fault & VM_FAULT_RETRY)) {
1454 /* Retry at most once */
1455 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1456 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1457 flags |= FAULT_FLAG_TRIED;
1458 if (!fatal_signal_pending(tsk))
1462 /* User mode? Just return to handle the fatal exception */
1463 if (flags & FAULT_FLAG_USER)
1466 /* Not returning to user mode? Handle exceptions or die: */
1467 no_context(regs, sw_error_code, address, SIGBUS, BUS_ADRERR);
1471 up_read(&mm->mmap_sem);
1472 if (unlikely(fault & VM_FAULT_ERROR)) {
1473 mm_fault_error(regs, sw_error_code, address, &pkey, fault);
1478 * Major/minor page fault accounting. If any of the events
1479 * returned VM_FAULT_MAJOR, we account it as a major fault.
1483 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1486 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1489 check_v8086_mode(regs, address, tsk);
1491 NOKPROBE_SYMBOL(__do_page_fault);
1493 static nokprobe_inline void
1494 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1495 unsigned long error_code)
1497 if (user_mode(regs))
1498 trace_page_fault_user(address, regs, error_code);
1500 trace_page_fault_kernel(address, regs, error_code);
1504 * We must have this function blacklisted from kprobes, tagged with notrace
1505 * and call read_cr2() before calling anything else. To avoid calling any
1506 * kind of tracing machinery before we've observed the CR2 value.
1508 * exception_{enter,exit}() contains all sorts of tracepoints.
1510 dotraplinkage void notrace
1511 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1513 unsigned long address = read_cr2(); /* Get the faulting address */
1514 enum ctx_state prev_state;
1516 prev_state = exception_enter();
1517 if (trace_pagefault_enabled())
1518 trace_page_fault_entries(address, regs, error_code);
1520 __do_page_fault(regs, error_code, address);
1521 exception_exit(prev_state);
1523 NOKPROBE_SYMBOL(do_page_fault);