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() */
20 #include <asm/cpufeature.h> /* boot_cpu_has, ... */
21 #include <asm/traps.h> /* dotraplinkage, ... */
22 #include <asm/pgalloc.h> /* pgd_*(), ... */
23 #include <asm/fixmap.h> /* VSYSCALL_ADDR */
24 #include <asm/vsyscall.h> /* emulate_vsyscall */
25 #include <asm/vm86.h> /* struct vm86 */
26 #include <asm/mmu_context.h> /* vma_pkey() */
28 #define CREATE_TRACE_POINTS
29 #include <asm/trace/exceptions.h>
32 * Returns 0 if mmiotrace is disabled, or if the fault is not
33 * handled by mmiotrace:
35 static nokprobe_inline int
36 kmmio_fault(struct pt_regs *regs, unsigned long addr)
38 if (unlikely(is_kmmio_active()))
39 if (kmmio_handler(regs, addr) == 1)
44 static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
48 /* kprobe_running() needs smp_processor_id() */
49 if (kprobes_built_in() && !user_mode(regs)) {
51 if (kprobe_running() && kprobe_fault_handler(regs, 14))
64 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
65 * Check that here and ignore it.
69 * Sometimes the CPU reports invalid exceptions on prefetch.
70 * Check that here and ignore it.
72 * Opcode checker based on code by Richard Brunner.
75 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
76 unsigned char opcode, int *prefetch)
78 unsigned char instr_hi = opcode & 0xf0;
79 unsigned char instr_lo = opcode & 0x0f;
85 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
86 * In X86_64 long mode, the CPU will signal invalid
87 * opcode if some of these prefixes are present so
88 * X86_64 will never get here anyway
90 return ((instr_lo & 7) == 0x6);
94 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
95 * Need to figure out under what instruction mode the
96 * instruction was issued. Could check the LDT for lm,
97 * but for now it's good enough to assume that long
98 * mode only uses well known segments or kernel.
100 return (!user_mode(regs) || user_64bit_mode(regs));
103 /* 0x64 thru 0x67 are valid prefixes in all modes. */
104 return (instr_lo & 0xC) == 0x4;
106 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
107 return !instr_lo || (instr_lo>>1) == 1;
109 /* Prefetch instruction is 0x0F0D or 0x0F18 */
110 if (probe_kernel_address(instr, opcode))
113 *prefetch = (instr_lo == 0xF) &&
114 (opcode == 0x0D || opcode == 0x18);
122 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
124 unsigned char *max_instr;
125 unsigned char *instr;
129 * If it was a exec (instruction fetch) fault on NX page, then
130 * do not ignore the fault:
132 if (error_code & X86_PF_INSTR)
135 instr = (void *)convert_ip_to_linear(current, regs);
136 max_instr = instr + 15;
138 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
141 while (instr < max_instr) {
142 unsigned char opcode;
144 if (probe_kernel_address(instr, opcode))
149 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
156 * A protection key fault means that the PKRU value did not allow
157 * access to some PTE. Userspace can figure out what PKRU was
158 * from the XSAVE state, and this function fills out a field in
159 * siginfo so userspace can discover which protection key was set
162 * If we get here, we know that the hardware signaled a X86_PF_PK
163 * fault and that there was a VMA once we got in the fault
164 * handler. It does *not* guarantee that the VMA we find here
165 * was the one that we faulted on.
167 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
168 * 2. T1 : set PKRU to deny access to pkey=4, touches page
170 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
171 * 5. T1 : enters fault handler, takes mmap_sem, etc...
172 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
173 * faulted on a pte with its pkey=4.
175 static void fill_sig_info_pkey(int si_signo, int si_code, siginfo_t *info,
178 /* This is effectively an #ifdef */
179 if (!boot_cpu_has(X86_FEATURE_OSPKE))
182 /* Fault not from Protection Keys: nothing to do */
183 if ((si_code != SEGV_PKUERR) || (si_signo != SIGSEGV))
186 * force_sig_info_fault() is called from a number of
187 * contexts, some of which have a VMA and some of which
188 * do not. The X86_PF_PK handing happens after we have a
189 * valid VMA, so we should never reach this without a
193 WARN_ONCE(1, "PKU fault with no VMA passed in");
198 * si_pkey should be thought of as a strong hint, but not
199 * absolutely guranteed to be 100% accurate because of
200 * the race explained above.
202 info->si_pkey = *pkey;
206 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
207 struct task_struct *tsk, u32 *pkey, int fault)
212 clear_siginfo(&info);
213 info.si_signo = si_signo;
215 info.si_code = si_code;
216 info.si_addr = (void __user *)address;
217 if (fault & VM_FAULT_HWPOISON_LARGE)
218 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
219 if (fault & VM_FAULT_HWPOISON)
221 info.si_addr_lsb = lsb;
223 fill_sig_info_pkey(si_signo, si_code, &info, pkey);
225 force_sig_info(si_signo, &info, tsk);
228 DEFINE_SPINLOCK(pgd_lock);
232 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
234 unsigned index = pgd_index(address);
241 pgd_k = init_mm.pgd + index;
243 if (!pgd_present(*pgd_k))
247 * set_pgd(pgd, *pgd_k); here would be useless on PAE
248 * and redundant with the set_pmd() on non-PAE. As would
251 p4d = p4d_offset(pgd, address);
252 p4d_k = p4d_offset(pgd_k, address);
253 if (!p4d_present(*p4d_k))
256 pud = pud_offset(p4d, address);
257 pud_k = pud_offset(p4d_k, address);
258 if (!pud_present(*pud_k))
261 pmd = pmd_offset(pud, address);
262 pmd_k = pmd_offset(pud_k, address);
263 if (!pmd_present(*pmd_k))
266 if (!pmd_present(*pmd))
267 set_pmd(pmd, *pmd_k);
269 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
274 void vmalloc_sync_all(void)
276 unsigned long address;
278 if (SHARED_KERNEL_PMD)
281 for (address = VMALLOC_START & PMD_MASK;
282 address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
283 address += PMD_SIZE) {
286 spin_lock(&pgd_lock);
287 list_for_each_entry(page, &pgd_list, lru) {
288 spinlock_t *pgt_lock;
291 /* the pgt_lock only for Xen */
292 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
295 ret = vmalloc_sync_one(page_address(page), address);
296 spin_unlock(pgt_lock);
301 spin_unlock(&pgd_lock);
308 * Handle a fault on the vmalloc or module mapping area
310 static noinline int vmalloc_fault(unsigned long address)
312 unsigned long pgd_paddr;
316 /* Make sure we are in vmalloc area: */
317 if (!(address >= VMALLOC_START && address < VMALLOC_END))
321 * Synchronize this task's top level page-table
322 * with the 'reference' page table.
324 * Do _not_ use "current" here. We might be inside
325 * an interrupt in the middle of a task switch..
327 pgd_paddr = read_cr3_pa();
328 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
332 if (pmd_large(*pmd_k))
335 pte_k = pte_offset_kernel(pmd_k, address);
336 if (!pte_present(*pte_k))
341 NOKPROBE_SYMBOL(vmalloc_fault);
344 * Did it hit the DOS screen memory VA from vm86 mode?
347 check_v8086_mode(struct pt_regs *regs, unsigned long address,
348 struct task_struct *tsk)
353 if (!v8086_mode(regs) || !tsk->thread.vm86)
356 bit = (address - 0xA0000) >> PAGE_SHIFT;
358 tsk->thread.vm86->screen_bitmap |= 1 << bit;
362 static bool low_pfn(unsigned long pfn)
364 return pfn < max_low_pfn;
367 static void dump_pagetable(unsigned long address)
369 pgd_t *base = __va(read_cr3_pa());
370 pgd_t *pgd = &base[pgd_index(address)];
376 #ifdef CONFIG_X86_PAE
377 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
378 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
380 #define pr_pde pr_cont
382 #define pr_pde pr_info
384 p4d = p4d_offset(pgd, address);
385 pud = pud_offset(p4d, address);
386 pmd = pmd_offset(pud, address);
387 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
391 * We must not directly access the pte in the highpte
392 * case if the page table is located in highmem.
393 * And let's rather not kmap-atomic the pte, just in case
394 * it's allocated already:
396 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
399 pte = pte_offset_kernel(pmd, address);
400 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
405 #else /* CONFIG_X86_64: */
407 void vmalloc_sync_all(void)
409 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
415 * Handle a fault on the vmalloc area
417 static noinline int vmalloc_fault(unsigned long address)
425 /* Make sure we are in vmalloc area: */
426 if (!(address >= VMALLOC_START && address < VMALLOC_END))
429 WARN_ON_ONCE(in_nmi());
432 * Copy kernel mappings over when needed. This can also
433 * happen within a race in page table update. In the later
436 pgd = (pgd_t *)__va(read_cr3_pa()) + pgd_index(address);
437 pgd_k = pgd_offset_k(address);
438 if (pgd_none(*pgd_k))
441 if (pgtable_l5_enabled()) {
442 if (pgd_none(*pgd)) {
443 set_pgd(pgd, *pgd_k);
444 arch_flush_lazy_mmu_mode();
446 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_k));
450 /* With 4-level paging, copying happens on the p4d level. */
451 p4d = p4d_offset(pgd, address);
452 p4d_k = p4d_offset(pgd_k, address);
453 if (p4d_none(*p4d_k))
456 if (p4d_none(*p4d) && !pgtable_l5_enabled()) {
457 set_p4d(p4d, *p4d_k);
458 arch_flush_lazy_mmu_mode();
460 BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_k));
463 BUILD_BUG_ON(CONFIG_PGTABLE_LEVELS < 4);
465 pud = pud_offset(p4d, address);
472 pmd = pmd_offset(pud, address);
479 pte = pte_offset_kernel(pmd, address);
480 if (!pte_present(*pte))
485 NOKPROBE_SYMBOL(vmalloc_fault);
487 #ifdef CONFIG_CPU_SUP_AMD
488 static const char errata93_warning[] =
490 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
491 "******* Working around it, but it may cause SEGVs or burn power.\n"
492 "******* Please consider a BIOS update.\n"
493 "******* Disabling USB legacy in the BIOS may also help.\n";
497 * No vm86 mode in 64-bit mode:
500 check_v8086_mode(struct pt_regs *regs, unsigned long address,
501 struct task_struct *tsk)
505 static int bad_address(void *p)
509 return probe_kernel_address((unsigned long *)p, dummy);
512 static void dump_pagetable(unsigned long address)
514 pgd_t *base = __va(read_cr3_pa());
515 pgd_t *pgd = base + pgd_index(address);
521 if (bad_address(pgd))
524 pr_info("PGD %lx ", pgd_val(*pgd));
526 if (!pgd_present(*pgd))
529 p4d = p4d_offset(pgd, address);
530 if (bad_address(p4d))
533 pr_cont("P4D %lx ", p4d_val(*p4d));
534 if (!p4d_present(*p4d) || p4d_large(*p4d))
537 pud = pud_offset(p4d, address);
538 if (bad_address(pud))
541 pr_cont("PUD %lx ", pud_val(*pud));
542 if (!pud_present(*pud) || pud_large(*pud))
545 pmd = pmd_offset(pud, address);
546 if (bad_address(pmd))
549 pr_cont("PMD %lx ", pmd_val(*pmd));
550 if (!pmd_present(*pmd) || pmd_large(*pmd))
553 pte = pte_offset_kernel(pmd, address);
554 if (bad_address(pte))
557 pr_cont("PTE %lx", pte_val(*pte));
565 #endif /* CONFIG_X86_64 */
568 * Workaround for K8 erratum #93 & buggy BIOS.
570 * BIOS SMM functions are required to use a specific workaround
571 * to avoid corruption of the 64bit RIP register on C stepping K8.
573 * A lot of BIOS that didn't get tested properly miss this.
575 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
576 * Try to work around it here.
578 * Note we only handle faults in kernel here.
579 * Does nothing on 32-bit.
581 static int is_errata93(struct pt_regs *regs, unsigned long address)
583 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
584 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
585 || boot_cpu_data.x86 != 0xf)
588 if (address != regs->ip)
591 if ((address >> 32) != 0)
594 address |= 0xffffffffUL << 32;
595 if ((address >= (u64)_stext && address <= (u64)_etext) ||
596 (address >= MODULES_VADDR && address <= MODULES_END)) {
597 printk_once(errata93_warning);
606 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
607 * to illegal addresses >4GB.
609 * We catch this in the page fault handler because these addresses
610 * are not reachable. Just detect this case and return. Any code
611 * segment in LDT is compatibility mode.
613 static int is_errata100(struct pt_regs *regs, unsigned long address)
616 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
622 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
624 #ifdef CONFIG_X86_F00F_BUG
628 * Pentium F0 0F C7 C8 bug workaround:
630 if (boot_cpu_has_bug(X86_BUG_F00F)) {
631 nr = (address - idt_descr.address) >> 3;
634 do_invalid_op(regs, 0);
643 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
644 unsigned long address)
646 if (!oops_may_print())
649 if (error_code & X86_PF_INSTR) {
654 pgd = __va(read_cr3_pa());
655 pgd += pgd_index(address);
657 pte = lookup_address_in_pgd(pgd, address, &level);
659 if (pte && pte_present(*pte) && !pte_exec(*pte))
660 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
661 from_kuid(&init_user_ns, current_uid()));
662 if (pte && pte_present(*pte) && pte_exec(*pte) &&
663 (pgd_flags(*pgd) & _PAGE_USER) &&
664 (__read_cr4() & X86_CR4_SMEP))
665 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
666 from_kuid(&init_user_ns, current_uid()));
669 pr_alert("BUG: unable to handle kernel %s at %px\n",
670 address < PAGE_SIZE ? "NULL pointer dereference" : "paging request",
673 dump_pagetable(address);
677 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
678 unsigned long address)
680 struct task_struct *tsk;
684 flags = oops_begin();
688 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
690 dump_pagetable(address);
692 tsk->thread.cr2 = address;
693 tsk->thread.trap_nr = X86_TRAP_PF;
694 tsk->thread.error_code = error_code;
696 if (__die("Bad pagetable", regs, error_code))
699 oops_end(flags, regs, sig);
703 no_context(struct pt_regs *regs, unsigned long error_code,
704 unsigned long address, int signal, int si_code)
706 struct task_struct *tsk = current;
710 /* Are we prepared to handle this kernel fault? */
711 if (fixup_exception(regs, X86_TRAP_PF)) {
713 * Any interrupt that takes a fault gets the fixup. This makes
714 * the below recursive fault logic only apply to a faults from
721 * Per the above we're !in_interrupt(), aka. task context.
723 * In this case we need to make sure we're not recursively
724 * faulting through the emulate_vsyscall() logic.
726 if (current->thread.sig_on_uaccess_err && signal) {
727 tsk->thread.trap_nr = X86_TRAP_PF;
728 tsk->thread.error_code = error_code | X86_PF_USER;
729 tsk->thread.cr2 = address;
731 /* XXX: hwpoison faults will set the wrong code. */
732 force_sig_info_fault(signal, si_code, address,
737 * Barring that, we can do the fixup and be happy.
742 #ifdef CONFIG_VMAP_STACK
744 * Stack overflow? During boot, we can fault near the initial
745 * stack in the direct map, but that's not an overflow -- check
746 * that we're in vmalloc space to avoid this.
748 if (is_vmalloc_addr((void *)address) &&
749 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
750 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
751 unsigned long stack = this_cpu_read(orig_ist.ist[DOUBLEFAULT_STACK]) - sizeof(void *);
753 * We're likely to be running with very little stack space
754 * left. It's plausible that we'd hit this condition but
755 * double-fault even before we get this far, in which case
756 * we're fine: the double-fault handler will deal with it.
758 * We don't want to make it all the way into the oops code
759 * and then double-fault, though, because we're likely to
760 * break the console driver and lose most of the stack dump.
762 asm volatile ("movq %[stack], %%rsp\n\t"
763 "call handle_stack_overflow\n\t"
765 : ASM_CALL_CONSTRAINT
766 : "D" ("kernel stack overflow (page fault)"),
767 "S" (regs), "d" (address),
768 [stack] "rm" (stack));
776 * Valid to do another page fault here, because if this fault
777 * had been triggered by is_prefetch fixup_exception would have
782 * Hall of shame of CPU/BIOS bugs.
784 if (is_prefetch(regs, error_code, address))
787 if (is_errata93(regs, address))
791 * Oops. The kernel tried to access some bad page. We'll have to
792 * terminate things with extreme prejudice:
794 flags = oops_begin();
796 show_fault_oops(regs, error_code, address);
798 if (task_stack_end_corrupted(tsk))
799 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
801 tsk->thread.cr2 = address;
802 tsk->thread.trap_nr = X86_TRAP_PF;
803 tsk->thread.error_code = error_code;
806 if (__die("Oops", regs, error_code))
809 /* Executive summary in case the body of the oops scrolled away */
810 printk(KERN_DEFAULT "CR2: %016lx\n", address);
812 oops_end(flags, regs, sig);
816 * Print out info about fatal segfaults, if the show_unhandled_signals
820 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
821 unsigned long address, struct task_struct *tsk)
823 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
825 if (!unhandled_signal(tsk, SIGSEGV))
828 if (!printk_ratelimit())
831 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
832 loglvl, tsk->comm, task_pid_nr(tsk), address,
833 (void *)regs->ip, (void *)regs->sp, error_code);
835 print_vma_addr(KERN_CONT " in ", regs->ip);
837 printk(KERN_CONT "\n");
839 show_opcodes((u8 *)regs->ip, loglvl);
843 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
844 unsigned long address, u32 *pkey, int si_code)
846 struct task_struct *tsk = current;
848 /* User mode accesses just cause a SIGSEGV */
849 if (error_code & X86_PF_USER) {
851 * It's possible to have interrupts off here:
856 * Valid to do another page fault here because this one came
859 if (is_prefetch(regs, error_code, address))
862 if (is_errata100(regs, address))
867 * Instruction fetch faults in the vsyscall page might need
870 if (unlikely((error_code & X86_PF_INSTR) &&
871 ((address & ~0xfff) == VSYSCALL_ADDR))) {
872 if (emulate_vsyscall(regs, address))
878 * To avoid leaking information about the kernel page table
879 * layout, pretend that user-mode accesses to kernel addresses
880 * are always protection faults.
882 if (address >= TASK_SIZE_MAX)
883 error_code |= X86_PF_PROT;
885 if (likely(show_unhandled_signals))
886 show_signal_msg(regs, error_code, address, tsk);
888 tsk->thread.cr2 = address;
889 tsk->thread.error_code = error_code;
890 tsk->thread.trap_nr = X86_TRAP_PF;
892 force_sig_info_fault(SIGSEGV, si_code, address, tsk, pkey, 0);
897 if (is_f00f_bug(regs, address))
900 no_context(regs, error_code, address, SIGSEGV, si_code);
904 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
905 unsigned long address, u32 *pkey)
907 __bad_area_nosemaphore(regs, error_code, address, pkey, SEGV_MAPERR);
911 __bad_area(struct pt_regs *regs, unsigned long error_code,
912 unsigned long address, struct vm_area_struct *vma, int si_code)
914 struct mm_struct *mm = current->mm;
918 pkey = vma_pkey(vma);
921 * Something tried to access memory that isn't in our memory map..
922 * Fix it, but check if it's kernel or user first..
924 up_read(&mm->mmap_sem);
926 __bad_area_nosemaphore(regs, error_code, address,
927 (vma) ? &pkey : NULL, si_code);
931 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
933 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
936 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
937 struct vm_area_struct *vma)
939 /* This code is always called on the current mm */
940 bool foreign = false;
942 if (!boot_cpu_has(X86_FEATURE_OSPKE))
944 if (error_code & X86_PF_PK)
946 /* this checks permission keys on the VMA: */
947 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
948 (error_code & X86_PF_INSTR), foreign))
954 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
955 unsigned long address, struct vm_area_struct *vma)
958 * This OSPKE check is not strictly necessary at runtime.
959 * But, doing it this way allows compiler optimizations
960 * if pkeys are compiled out.
962 if (bad_area_access_from_pkeys(error_code, vma))
963 __bad_area(regs, error_code, address, vma, SEGV_PKUERR);
965 __bad_area(regs, error_code, address, vma, SEGV_ACCERR);
969 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
970 u32 *pkey, unsigned int fault)
972 struct task_struct *tsk = current;
973 int code = BUS_ADRERR;
975 /* Kernel mode? Handle exceptions or die: */
976 if (!(error_code & X86_PF_USER)) {
977 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
981 /* User-space => ok to do another page fault: */
982 if (is_prefetch(regs, error_code, address))
985 tsk->thread.cr2 = address;
986 tsk->thread.error_code = error_code;
987 tsk->thread.trap_nr = X86_TRAP_PF;
989 #ifdef CONFIG_MEMORY_FAILURE
990 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
992 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
993 tsk->comm, tsk->pid, address);
994 code = BUS_MCEERR_AR;
997 force_sig_info_fault(SIGBUS, code, address, tsk, pkey, fault);
1000 static noinline void
1001 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
1002 unsigned long address, u32 *pkey, unsigned int fault)
1004 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) {
1005 no_context(regs, error_code, address, 0, 0);
1009 if (fault & VM_FAULT_OOM) {
1010 /* Kernel mode? Handle exceptions or die: */
1011 if (!(error_code & X86_PF_USER)) {
1012 no_context(regs, error_code, address,
1013 SIGSEGV, SEGV_MAPERR);
1018 * We ran out of memory, call the OOM killer, and return the
1019 * userspace (which will retry the fault, or kill us if we got
1022 pagefault_out_of_memory();
1024 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1025 VM_FAULT_HWPOISON_LARGE))
1026 do_sigbus(regs, error_code, address, pkey, fault);
1027 else if (fault & VM_FAULT_SIGSEGV)
1028 bad_area_nosemaphore(regs, error_code, address, pkey);
1034 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
1036 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
1039 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1042 * Note: We do not do lazy flushing on protection key
1043 * changes, so no spurious fault will ever set X86_PF_PK.
1045 if ((error_code & X86_PF_PK))
1052 * Handle a spurious fault caused by a stale TLB entry.
1054 * This allows us to lazily refresh the TLB when increasing the
1055 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1056 * eagerly is very expensive since that implies doing a full
1057 * cross-processor TLB flush, even if no stale TLB entries exist
1058 * on other processors.
1060 * Spurious faults may only occur if the TLB contains an entry with
1061 * fewer permission than the page table entry. Non-present (P = 0)
1062 * and reserved bit (R = 1) faults are never spurious.
1064 * There are no security implications to leaving a stale TLB when
1065 * increasing the permissions on a page.
1067 * Returns non-zero if a spurious fault was handled, zero otherwise.
1069 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1070 * (Optional Invalidation).
1073 spurious_fault(unsigned long error_code, unsigned long address)
1083 * Only writes to RO or instruction fetches from NX may cause
1086 * These could be from user or supervisor accesses but the TLB
1087 * is only lazily flushed after a kernel mapping protection
1088 * change, so user accesses are not expected to cause spurious
1091 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1092 error_code != (X86_PF_INSTR | X86_PF_PROT))
1095 pgd = init_mm.pgd + pgd_index(address);
1096 if (!pgd_present(*pgd))
1099 p4d = p4d_offset(pgd, address);
1100 if (!p4d_present(*p4d))
1103 if (p4d_large(*p4d))
1104 return spurious_fault_check(error_code, (pte_t *) p4d);
1106 pud = pud_offset(p4d, address);
1107 if (!pud_present(*pud))
1110 if (pud_large(*pud))
1111 return spurious_fault_check(error_code, (pte_t *) pud);
1113 pmd = pmd_offset(pud, address);
1114 if (!pmd_present(*pmd))
1117 if (pmd_large(*pmd))
1118 return spurious_fault_check(error_code, (pte_t *) pmd);
1120 pte = pte_offset_kernel(pmd, address);
1121 if (!pte_present(*pte))
1124 ret = spurious_fault_check(error_code, pte);
1129 * Make sure we have permissions in PMD.
1130 * If not, then there's a bug in the page tables:
1132 ret = spurious_fault_check(error_code, (pte_t *) pmd);
1133 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1137 NOKPROBE_SYMBOL(spurious_fault);
1139 int show_unhandled_signals = 1;
1142 access_error(unsigned long error_code, struct vm_area_struct *vma)
1144 /* This is only called for the current mm, so: */
1145 bool foreign = false;
1148 * Read or write was blocked by protection keys. This is
1149 * always an unconditional error and can never result in
1150 * a follow-up action to resolve the fault, like a COW.
1152 if (error_code & X86_PF_PK)
1156 * Make sure to check the VMA so that we do not perform
1157 * faults just to hit a X86_PF_PK as soon as we fill in a
1160 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1161 (error_code & X86_PF_INSTR), foreign))
1164 if (error_code & X86_PF_WRITE) {
1165 /* write, present and write, not present: */
1166 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1171 /* read, present: */
1172 if (unlikely(error_code & X86_PF_PROT))
1175 /* read, not present: */
1176 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1182 static int fault_in_kernel_space(unsigned long address)
1184 return address >= TASK_SIZE_MAX;
1187 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1189 if (!IS_ENABLED(CONFIG_X86_SMAP))
1192 if (!static_cpu_has(X86_FEATURE_SMAP))
1195 if (error_code & X86_PF_USER)
1198 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1205 * This routine handles page faults. It determines the address,
1206 * and the problem, and then passes it off to one of the appropriate
1209 static noinline void
1210 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1211 unsigned long address)
1213 struct vm_area_struct *vma;
1214 struct task_struct *tsk;
1215 struct mm_struct *mm;
1216 int fault, major = 0;
1217 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1223 prefetchw(&mm->mmap_sem);
1225 if (unlikely(kmmio_fault(regs, address)))
1229 * We fault-in kernel-space virtual memory on-demand. The
1230 * 'reference' page table is init_mm.pgd.
1232 * NOTE! We MUST NOT take any locks for this case. We may
1233 * be in an interrupt or a critical region, and should
1234 * only copy the information from the master page table,
1237 * This verifies that the fault happens in kernel space
1238 * (error_code & 4) == 0, and that the fault was not a
1239 * protection error (error_code & 9) == 0.
1241 if (unlikely(fault_in_kernel_space(address))) {
1242 if (!(error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1243 if (vmalloc_fault(address) >= 0)
1247 /* Can handle a stale RO->RW TLB: */
1248 if (spurious_fault(error_code, address))
1251 /* kprobes don't want to hook the spurious faults: */
1252 if (kprobes_fault(regs))
1255 * Don't take the mm semaphore here. If we fixup a prefetch
1256 * fault we could otherwise deadlock:
1258 bad_area_nosemaphore(regs, error_code, address, NULL);
1263 /* kprobes don't want to hook the spurious faults: */
1264 if (unlikely(kprobes_fault(regs)))
1267 if (unlikely(error_code & X86_PF_RSVD))
1268 pgtable_bad(regs, error_code, address);
1270 if (unlikely(smap_violation(error_code, regs))) {
1271 bad_area_nosemaphore(regs, error_code, address, NULL);
1276 * If we're in an interrupt, have no user context or are running
1277 * in a region with pagefaults disabled then we must not take the fault
1279 if (unlikely(faulthandler_disabled() || !mm)) {
1280 bad_area_nosemaphore(regs, error_code, address, NULL);
1285 * It's safe to allow irq's after cr2 has been saved and the
1286 * vmalloc fault has been handled.
1288 * User-mode registers count as a user access even for any
1289 * potential system fault or CPU buglet:
1291 if (user_mode(regs)) {
1293 error_code |= X86_PF_USER;
1294 flags |= FAULT_FLAG_USER;
1296 if (regs->flags & X86_EFLAGS_IF)
1300 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1302 if (error_code & X86_PF_WRITE)
1303 flags |= FAULT_FLAG_WRITE;
1304 if (error_code & X86_PF_INSTR)
1305 flags |= FAULT_FLAG_INSTRUCTION;
1308 * When running in the kernel we expect faults to occur only to
1309 * addresses in user space. All other faults represent errors in
1310 * the kernel and should generate an OOPS. Unfortunately, in the
1311 * case of an erroneous fault occurring in a code path which already
1312 * holds mmap_sem we will deadlock attempting to validate the fault
1313 * against the address space. Luckily the kernel only validly
1314 * references user space from well defined areas of code, which are
1315 * listed in the exceptions table.
1317 * As the vast majority of faults will be valid we will only perform
1318 * the source reference check when there is a possibility of a
1319 * deadlock. Attempt to lock the address space, if we cannot we then
1320 * validate the source. If this is invalid we can skip the address
1321 * space check, thus avoiding the deadlock:
1323 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1324 if (!(error_code & X86_PF_USER) &&
1325 !search_exception_tables(regs->ip)) {
1326 bad_area_nosemaphore(regs, error_code, address, NULL);
1330 down_read(&mm->mmap_sem);
1333 * The above down_read_trylock() might have succeeded in
1334 * which case we'll have missed the might_sleep() from
1340 vma = find_vma(mm, address);
1341 if (unlikely(!vma)) {
1342 bad_area(regs, error_code, address);
1345 if (likely(vma->vm_start <= address))
1347 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1348 bad_area(regs, error_code, address);
1351 if (error_code & X86_PF_USER) {
1353 * Accessing the stack below %sp is always a bug.
1354 * The large cushion allows instructions like enter
1355 * and pusha to work. ("enter $65535, $31" pushes
1356 * 32 pointers and then decrements %sp by 65535.)
1358 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1359 bad_area(regs, error_code, address);
1363 if (unlikely(expand_stack(vma, address))) {
1364 bad_area(regs, error_code, address);
1369 * Ok, we have a good vm_area for this memory access, so
1370 * we can handle it..
1373 if (unlikely(access_error(error_code, vma))) {
1374 bad_area_access_error(regs, error_code, address, vma);
1379 * If for any reason at all we couldn't handle the fault,
1380 * make sure we exit gracefully rather than endlessly redo
1381 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1382 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1384 * Note that handle_userfault() may also release and reacquire mmap_sem
1385 * (and not return with VM_FAULT_RETRY), when returning to userland to
1386 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1387 * (potentially after handling any pending signal during the return to
1388 * userland). The return to userland is identified whenever
1389 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1390 * Thus we have to be careful about not touching vma after handling the
1391 * fault, so we read the pkey beforehand.
1393 pkey = vma_pkey(vma);
1394 fault = handle_mm_fault(vma, address, flags);
1395 major |= fault & VM_FAULT_MAJOR;
1398 * If we need to retry the mmap_sem has already been released,
1399 * and if there is a fatal signal pending there is no guarantee
1400 * that we made any progress. Handle this case first.
1402 if (unlikely(fault & VM_FAULT_RETRY)) {
1403 /* Retry at most once */
1404 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1405 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1406 flags |= FAULT_FLAG_TRIED;
1407 if (!fatal_signal_pending(tsk))
1411 /* User mode? Just return to handle the fatal exception */
1412 if (flags & FAULT_FLAG_USER)
1415 /* Not returning to user mode? Handle exceptions or die: */
1416 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1420 up_read(&mm->mmap_sem);
1421 if (unlikely(fault & VM_FAULT_ERROR)) {
1422 mm_fault_error(regs, error_code, address, &pkey, fault);
1427 * Major/minor page fault accounting. If any of the events
1428 * returned VM_FAULT_MAJOR, we account it as a major fault.
1432 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1435 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1438 check_v8086_mode(regs, address, tsk);
1440 NOKPROBE_SYMBOL(__do_page_fault);
1442 static nokprobe_inline void
1443 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1444 unsigned long error_code)
1446 if (user_mode(regs))
1447 trace_page_fault_user(address, regs, error_code);
1449 trace_page_fault_kernel(address, regs, error_code);
1453 * We must have this function blacklisted from kprobes, tagged with notrace
1454 * and call read_cr2() before calling anything else. To avoid calling any
1455 * kind of tracing machinery before we've observed the CR2 value.
1457 * exception_{enter,exit}() contains all sorts of tracepoints.
1459 dotraplinkage void notrace
1460 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1462 unsigned long address = read_cr2(); /* Get the faulting address */
1463 enum ctx_state prev_state;
1465 prev_state = exception_enter();
1466 if (trace_pagefault_enabled())
1467 trace_page_fault_entries(address, regs, error_code);
1469 __do_page_fault(regs, error_code, address);
1470 exception_exit(prev_state);
1472 NOKPROBE_SYMBOL(do_page_fault);