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/memblock.h> /* max_low_pfn */
12 #include <linux/kfence.h> /* kfence_handle_page_fault */
13 #include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */
14 #include <linux/mmiotrace.h> /* kmmio_handler, ... */
15 #include <linux/perf_event.h> /* perf_sw_event */
16 #include <linux/hugetlb.h> /* hstate_index_to_shift */
17 #include <linux/prefetch.h> /* prefetchw */
18 #include <linux/context_tracking.h> /* exception_enter(), ... */
19 #include <linux/uaccess.h> /* faulthandler_disabled() */
20 #include <linux/efi.h> /* efi_crash_gracefully_on_page_fault()*/
21 #include <linux/mm_types.h>
23 #include <asm/cpufeature.h> /* boot_cpu_has, ... */
24 #include <asm/traps.h> /* dotraplinkage, ... */
25 #include <asm/fixmap.h> /* VSYSCALL_ADDR */
26 #include <asm/vsyscall.h> /* emulate_vsyscall */
27 #include <asm/vm86.h> /* struct vm86 */
28 #include <asm/mmu_context.h> /* vma_pkey() */
29 #include <asm/efi.h> /* efi_crash_gracefully_on_page_fault()*/
30 #include <asm/desc.h> /* store_idt(), ... */
31 #include <asm/cpu_entry_area.h> /* exception stack */
32 #include <asm/pgtable_areas.h> /* VMALLOC_START, ... */
33 #include <asm/kvm_para.h> /* kvm_handle_async_pf */
34 #include <asm/vdso.h> /* fixup_vdso_exception() */
35 #include <asm/irq_stack.h>
37 #define CREATE_TRACE_POINTS
38 #include <asm/trace/exceptions.h>
41 * Returns 0 if mmiotrace is disabled, or if the fault is not
42 * handled by mmiotrace:
44 static nokprobe_inline int
45 kmmio_fault(struct pt_regs *regs, unsigned long addr)
47 if (unlikely(is_kmmio_active()))
48 if (kmmio_handler(regs, addr) == 1)
58 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
59 * Check that here and ignore it. This is AMD erratum #91.
63 * Sometimes the CPU reports invalid exceptions on prefetch.
64 * Check that here and ignore it.
66 * Opcode checker based on code by Richard Brunner.
69 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
70 unsigned char opcode, int *prefetch)
72 unsigned char instr_hi = opcode & 0xf0;
73 unsigned char instr_lo = opcode & 0x0f;
79 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
80 * In X86_64 long mode, the CPU will signal invalid
81 * opcode if some of these prefixes are present so
82 * X86_64 will never get here anyway
84 return ((instr_lo & 7) == 0x6);
88 * In 64-bit mode 0x40..0x4F are valid REX prefixes
90 return (!user_mode(regs) || user_64bit_mode(regs));
93 /* 0x64 thru 0x67 are valid prefixes in all modes. */
94 return (instr_lo & 0xC) == 0x4;
96 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
97 return !instr_lo || (instr_lo>>1) == 1;
99 /* Prefetch instruction is 0x0F0D or 0x0F18 */
100 if (get_kernel_nofault(opcode, instr))
103 *prefetch = (instr_lo == 0xF) &&
104 (opcode == 0x0D || opcode == 0x18);
111 static bool is_amd_k8_pre_npt(void)
113 struct cpuinfo_x86 *c = &boot_cpu_data;
115 return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
116 c->x86_vendor == X86_VENDOR_AMD &&
117 c->x86 == 0xf && c->x86_model < 0x40);
121 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
123 unsigned char *max_instr;
124 unsigned char *instr;
127 /* Erratum #91 affects AMD K8, pre-NPT CPUs */
128 if (!is_amd_k8_pre_npt())
132 * If it was a exec (instruction fetch) fault on NX page, then
133 * do not ignore the fault:
135 if (error_code & X86_PF_INSTR)
138 instr = (void *)convert_ip_to_linear(current, regs);
139 max_instr = instr + 15;
142 * This code has historically always bailed out if IP points to a
143 * not-present page (e.g. due to a race). No one has ever
144 * complained about this.
148 while (instr < max_instr) {
149 unsigned char opcode;
151 if (user_mode(regs)) {
152 if (get_user(opcode, instr))
155 if (get_kernel_nofault(opcode, instr))
161 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
169 DEFINE_SPINLOCK(pgd_lock);
173 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
175 unsigned index = pgd_index(address);
182 pgd_k = init_mm.pgd + index;
184 if (!pgd_present(*pgd_k))
188 * set_pgd(pgd, *pgd_k); here would be useless on PAE
189 * and redundant with the set_pmd() on non-PAE. As would
192 p4d = p4d_offset(pgd, address);
193 p4d_k = p4d_offset(pgd_k, address);
194 if (!p4d_present(*p4d_k))
197 pud = pud_offset(p4d, address);
198 pud_k = pud_offset(p4d_k, address);
199 if (!pud_present(*pud_k))
202 pmd = pmd_offset(pud, address);
203 pmd_k = pmd_offset(pud_k, address);
205 if (pmd_present(*pmd) != pmd_present(*pmd_k))
206 set_pmd(pmd, *pmd_k);
208 if (!pmd_present(*pmd_k))
211 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
217 * Handle a fault on the vmalloc or module mapping area
219 * This is needed because there is a race condition between the time
220 * when the vmalloc mapping code updates the PMD to the point in time
221 * where it synchronizes this update with the other page-tables in the
224 * In this race window another thread/CPU can map an area on the same
225 * PMD, finds it already present and does not synchronize it with the
226 * rest of the system yet. As a result v[mz]alloc might return areas
227 * which are not mapped in every page-table in the system, causing an
228 * unhandled page-fault when they are accessed.
230 static noinline int vmalloc_fault(unsigned long address)
232 unsigned long pgd_paddr;
236 /* Make sure we are in vmalloc area: */
237 if (!(address >= VMALLOC_START && address < VMALLOC_END))
241 * Synchronize this task's top level page-table
242 * with the 'reference' page table.
244 * Do _not_ use "current" here. We might be inside
245 * an interrupt in the middle of a task switch..
247 pgd_paddr = read_cr3_pa();
248 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
252 if (pmd_large(*pmd_k))
255 pte_k = pte_offset_kernel(pmd_k, address);
256 if (!pte_present(*pte_k))
261 NOKPROBE_SYMBOL(vmalloc_fault);
263 void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
267 for (addr = start & PMD_MASK;
268 addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
272 spin_lock(&pgd_lock);
273 list_for_each_entry(page, &pgd_list, lru) {
274 spinlock_t *pgt_lock;
276 /* the pgt_lock only for Xen */
277 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
280 vmalloc_sync_one(page_address(page), addr);
281 spin_unlock(pgt_lock);
283 spin_unlock(&pgd_lock);
287 static bool low_pfn(unsigned long pfn)
289 return pfn < max_low_pfn;
292 static void dump_pagetable(unsigned long address)
294 pgd_t *base = __va(read_cr3_pa());
295 pgd_t *pgd = &base[pgd_index(address)];
301 #ifdef CONFIG_X86_PAE
302 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
303 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
305 #define pr_pde pr_cont
307 #define pr_pde pr_info
309 p4d = p4d_offset(pgd, address);
310 pud = pud_offset(p4d, address);
311 pmd = pmd_offset(pud, address);
312 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
316 * We must not directly access the pte in the highpte
317 * case if the page table is located in highmem.
318 * And let's rather not kmap-atomic the pte, just in case
319 * it's allocated already:
321 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
324 pte = pte_offset_kernel(pmd, address);
325 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
330 #else /* CONFIG_X86_64: */
332 #ifdef CONFIG_CPU_SUP_AMD
333 static const char errata93_warning[] =
335 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
336 "******* Working around it, but it may cause SEGVs or burn power.\n"
337 "******* Please consider a BIOS update.\n"
338 "******* Disabling USB legacy in the BIOS may also help.\n";
341 static int bad_address(void *p)
345 return get_kernel_nofault(dummy, (unsigned long *)p);
348 static void dump_pagetable(unsigned long address)
350 pgd_t *base = __va(read_cr3_pa());
351 pgd_t *pgd = base + pgd_index(address);
357 if (bad_address(pgd))
360 pr_info("PGD %lx ", pgd_val(*pgd));
362 if (!pgd_present(*pgd))
365 p4d = p4d_offset(pgd, address);
366 if (bad_address(p4d))
369 pr_cont("P4D %lx ", p4d_val(*p4d));
370 if (!p4d_present(*p4d) || p4d_large(*p4d))
373 pud = pud_offset(p4d, address);
374 if (bad_address(pud))
377 pr_cont("PUD %lx ", pud_val(*pud));
378 if (!pud_present(*pud) || pud_large(*pud))
381 pmd = pmd_offset(pud, address);
382 if (bad_address(pmd))
385 pr_cont("PMD %lx ", pmd_val(*pmd));
386 if (!pmd_present(*pmd) || pmd_large(*pmd))
389 pte = pte_offset_kernel(pmd, address);
390 if (bad_address(pte))
393 pr_cont("PTE %lx", pte_val(*pte));
401 #endif /* CONFIG_X86_64 */
404 * Workaround for K8 erratum #93 & buggy BIOS.
406 * BIOS SMM functions are required to use a specific workaround
407 * to avoid corruption of the 64bit RIP register on C stepping K8.
409 * A lot of BIOS that didn't get tested properly miss this.
411 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
412 * Try to work around it here.
414 * Note we only handle faults in kernel here.
415 * Does nothing on 32-bit.
417 static int is_errata93(struct pt_regs *regs, unsigned long address)
419 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
420 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
421 || boot_cpu_data.x86 != 0xf)
427 if (address != regs->ip)
430 if ((address >> 32) != 0)
433 address |= 0xffffffffUL << 32;
434 if ((address >= (u64)_stext && address <= (u64)_etext) ||
435 (address >= MODULES_VADDR && address <= MODULES_END)) {
436 printk_once(errata93_warning);
445 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
446 * to illegal addresses >4GB.
448 * We catch this in the page fault handler because these addresses
449 * are not reachable. Just detect this case and return. Any code
450 * segment in LDT is compatibility mode.
452 static int is_errata100(struct pt_regs *regs, unsigned long address)
455 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
461 /* Pentium F0 0F C7 C8 bug workaround: */
462 static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
463 unsigned long address)
465 #ifdef CONFIG_X86_F00F_BUG
466 if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
467 idt_is_f00f_address(address)) {
468 handle_invalid_op(regs);
475 static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
477 u32 offset = (index >> 3) * sizeof(struct desc_struct);
479 struct ldttss_desc desc;
482 pr_alert("%s: NULL\n", name);
486 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
487 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
491 if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset),
492 sizeof(struct ldttss_desc))) {
493 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
498 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
500 addr |= ((u64)desc.base3 << 32);
502 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
503 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
507 show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
509 if (!oops_may_print())
512 if (error_code & X86_PF_INSTR) {
517 pgd = __va(read_cr3_pa());
518 pgd += pgd_index(address);
520 pte = lookup_address_in_pgd(pgd, address, &level);
522 if (pte && pte_present(*pte) && !pte_exec(*pte))
523 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
524 from_kuid(&init_user_ns, current_uid()));
525 if (pte && pte_present(*pte) && pte_exec(*pte) &&
526 (pgd_flags(*pgd) & _PAGE_USER) &&
527 (__read_cr4() & X86_CR4_SMEP))
528 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
529 from_kuid(&init_user_ns, current_uid()));
532 if (address < PAGE_SIZE && !user_mode(regs))
533 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
536 pr_alert("BUG: unable to handle page fault for address: %px\n",
539 pr_alert("#PF: %s %s in %s mode\n",
540 (error_code & X86_PF_USER) ? "user" : "supervisor",
541 (error_code & X86_PF_INSTR) ? "instruction fetch" :
542 (error_code & X86_PF_WRITE) ? "write access" :
544 user_mode(regs) ? "user" : "kernel");
545 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
546 !(error_code & X86_PF_PROT) ? "not-present page" :
547 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
548 (error_code & X86_PF_PK) ? "protection keys violation" :
549 "permissions violation");
551 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
552 struct desc_ptr idt, gdt;
556 * This can happen for quite a few reasons. The more obvious
557 * ones are faults accessing the GDT, or LDT. Perhaps
558 * surprisingly, if the CPU tries to deliver a benign or
559 * contributory exception from user code and gets a page fault
560 * during delivery, the page fault can be delivered as though
561 * it originated directly from user code. This could happen
562 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
563 * kernel or IST stack.
567 /* Usable even on Xen PV -- it's just slow. */
568 native_store_gdt(&gdt);
570 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
571 idt.address, idt.size, gdt.address, gdt.size);
574 show_ldttss(&gdt, "LDTR", ldtr);
577 show_ldttss(&gdt, "TR", tr);
580 dump_pagetable(address);
584 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
585 unsigned long address)
587 struct task_struct *tsk;
591 flags = oops_begin();
595 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
597 dump_pagetable(address);
599 if (__die("Bad pagetable", regs, error_code))
602 oops_end(flags, regs, sig);
605 static void sanitize_error_code(unsigned long address,
606 unsigned long *error_code)
609 * To avoid leaking information about the kernel page
610 * table layout, pretend that user-mode accesses to
611 * kernel addresses are always protection faults.
613 * NB: This means that failed vsyscalls with vsyscall=none
614 * will have the PROT bit. This doesn't leak any
615 * information and does not appear to cause any problems.
617 if (address >= TASK_SIZE_MAX)
618 *error_code |= X86_PF_PROT;
621 static void set_signal_archinfo(unsigned long address,
622 unsigned long error_code)
624 struct task_struct *tsk = current;
626 tsk->thread.trap_nr = X86_TRAP_PF;
627 tsk->thread.error_code = error_code | X86_PF_USER;
628 tsk->thread.cr2 = address;
632 page_fault_oops(struct pt_regs *regs, unsigned long error_code,
633 unsigned long address)
635 #ifdef CONFIG_VMAP_STACK
636 struct stack_info info;
641 if (user_mode(regs)) {
643 * Implicit kernel access from user mode? Skip the stack
644 * overflow and EFI special cases.
649 #ifdef CONFIG_VMAP_STACK
651 * Stack overflow? During boot, we can fault near the initial
652 * stack in the direct map, but that's not an overflow -- check
653 * that we're in vmalloc space to avoid this.
655 if (is_vmalloc_addr((void *)address) &&
656 get_stack_guard_info((void *)address, &info)) {
658 * We're likely to be running with very little stack space
659 * left. It's plausible that we'd hit this condition but
660 * double-fault even before we get this far, in which case
661 * we're fine: the double-fault handler will deal with it.
663 * We don't want to make it all the way into the oops code
664 * and then double-fault, though, because we're likely to
665 * break the console driver and lose most of the stack dump.
667 call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*),
668 handle_stack_overflow,
670 , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info));
677 * Buggy firmware could access regions which might page fault. If
678 * this happens, EFI has a special OOPS path that will try to
679 * avoid hanging the system.
681 if (IS_ENABLED(CONFIG_EFI))
682 efi_crash_gracefully_on_page_fault(address);
684 /* Only not-present faults should be handled by KFENCE. */
685 if (!(error_code & X86_PF_PROT) &&
686 kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs))
691 * Oops. The kernel tried to access some bad page. We'll have to
692 * terminate things with extreme prejudice:
694 flags = oops_begin();
696 show_fault_oops(regs, error_code, address);
698 if (task_stack_end_corrupted(current))
699 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
702 if (__die("Oops", regs, error_code))
705 /* Executive summary in case the body of the oops scrolled away */
706 printk(KERN_DEFAULT "CR2: %016lx\n", address);
708 oops_end(flags, regs, sig);
712 kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
713 unsigned long address, int signal, int si_code,
716 WARN_ON_ONCE(user_mode(regs));
718 /* Are we prepared to handle this kernel fault? */
719 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
721 * Any interrupt that takes a fault gets the fixup. This makes
722 * the below recursive fault logic only apply to a faults from
729 * Per the above we're !in_interrupt(), aka. task context.
731 * In this case we need to make sure we're not recursively
732 * faulting through the emulate_vsyscall() logic.
734 if (current->thread.sig_on_uaccess_err && signal) {
735 sanitize_error_code(address, &error_code);
737 set_signal_archinfo(address, error_code);
739 if (si_code == SEGV_PKUERR) {
740 force_sig_pkuerr((void __user *)address, pkey);
742 /* XXX: hwpoison faults will set the wrong code. */
743 force_sig_fault(signal, si_code, (void __user *)address);
748 * Barring that, we can do the fixup and be happy.
754 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
757 if (is_prefetch(regs, error_code, address))
760 page_fault_oops(regs, error_code, address);
764 * Print out info about fatal segfaults, if the show_unhandled_signals
768 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
769 unsigned long address, struct task_struct *tsk)
771 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
773 if (!unhandled_signal(tsk, SIGSEGV))
776 if (!printk_ratelimit())
779 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
780 loglvl, tsk->comm, task_pid_nr(tsk), address,
781 (void *)regs->ip, (void *)regs->sp, error_code);
783 print_vma_addr(KERN_CONT " in ", regs->ip);
785 printk(KERN_CONT "\n");
787 show_opcodes(regs, loglvl);
791 * The (legacy) vsyscall page is the long page in the kernel portion
792 * of the address space that has user-accessible permissions.
794 static bool is_vsyscall_vaddr(unsigned long vaddr)
796 return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR);
800 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
801 unsigned long address, u32 pkey, int si_code)
803 struct task_struct *tsk = current;
805 if (!user_mode(regs)) {
806 kernelmode_fixup_or_oops(regs, error_code, address,
807 SIGSEGV, si_code, pkey);
811 if (!(error_code & X86_PF_USER)) {
812 /* Implicit user access to kernel memory -- just oops */
813 page_fault_oops(regs, error_code, address);
818 * User mode accesses just cause a SIGSEGV.
819 * It's possible to have interrupts off here:
824 * Valid to do another page fault here because this one came
827 if (is_prefetch(regs, error_code, address))
830 if (is_errata100(regs, address))
833 sanitize_error_code(address, &error_code);
835 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
838 if (likely(show_unhandled_signals))
839 show_signal_msg(regs, error_code, address, tsk);
841 set_signal_archinfo(address, error_code);
843 if (si_code == SEGV_PKUERR)
844 force_sig_pkuerr((void __user *)address, pkey);
846 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
852 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
853 unsigned long address)
855 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
859 __bad_area(struct pt_regs *regs, unsigned long error_code,
860 unsigned long address, u32 pkey, int si_code)
862 struct mm_struct *mm = current->mm;
864 * Something tried to access memory that isn't in our memory map..
865 * Fix it, but check if it's kernel or user first..
867 mmap_read_unlock(mm);
869 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
873 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
875 __bad_area(regs, error_code, address, 0, SEGV_MAPERR);
878 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
879 struct vm_area_struct *vma)
881 /* This code is always called on the current mm */
882 bool foreign = false;
884 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
886 if (error_code & X86_PF_PK)
888 /* this checks permission keys on the VMA: */
889 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
890 (error_code & X86_PF_INSTR), foreign))
896 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
897 unsigned long address, struct vm_area_struct *vma)
900 * This OSPKE check is not strictly necessary at runtime.
901 * But, doing it this way allows compiler optimizations
902 * if pkeys are compiled out.
904 if (bad_area_access_from_pkeys(error_code, vma)) {
906 * A protection key fault means that the PKRU value did not allow
907 * access to some PTE. Userspace can figure out what PKRU was
908 * from the XSAVE state. This function captures the pkey from
909 * the vma and passes it to userspace so userspace can discover
910 * which protection key was set on the PTE.
912 * If we get here, we know that the hardware signaled a X86_PF_PK
913 * fault and that there was a VMA once we got in the fault
914 * handler. It does *not* guarantee that the VMA we find here
915 * was the one that we faulted on.
917 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
918 * 2. T1 : set PKRU to deny access to pkey=4, touches page
920 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
921 * 5. T1 : enters fault handler, takes mmap_lock, etc...
922 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
923 * faulted on a pte with its pkey=4.
925 u32 pkey = vma_pkey(vma);
927 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
929 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
934 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
937 /* Kernel mode? Handle exceptions or die: */
938 if (!user_mode(regs)) {
939 kernelmode_fixup_or_oops(regs, error_code, address,
940 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
944 /* User-space => ok to do another page fault: */
945 if (is_prefetch(regs, error_code, address))
948 sanitize_error_code(address, &error_code);
950 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
953 set_signal_archinfo(address, error_code);
955 #ifdef CONFIG_MEMORY_FAILURE
956 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
957 struct task_struct *tsk = current;
961 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
962 tsk->comm, tsk->pid, address);
963 if (fault & VM_FAULT_HWPOISON_LARGE)
964 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
965 if (fault & VM_FAULT_HWPOISON)
967 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
971 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
974 static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
976 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
979 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
986 * Handle a spurious fault caused by a stale TLB entry.
988 * This allows us to lazily refresh the TLB when increasing the
989 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
990 * eagerly is very expensive since that implies doing a full
991 * cross-processor TLB flush, even if no stale TLB entries exist
992 * on other processors.
994 * Spurious faults may only occur if the TLB contains an entry with
995 * fewer permission than the page table entry. Non-present (P = 0)
996 * and reserved bit (R = 1) faults are never spurious.
998 * There are no security implications to leaving a stale TLB when
999 * increasing the permissions on a page.
1001 * Returns non-zero if a spurious fault was handled, zero otherwise.
1003 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1004 * (Optional Invalidation).
1007 spurious_kernel_fault(unsigned long error_code, unsigned long address)
1017 * Only writes to RO or instruction fetches from NX may cause
1020 * These could be from user or supervisor accesses but the TLB
1021 * is only lazily flushed after a kernel mapping protection
1022 * change, so user accesses are not expected to cause spurious
1025 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1026 error_code != (X86_PF_INSTR | X86_PF_PROT))
1029 pgd = init_mm.pgd + pgd_index(address);
1030 if (!pgd_present(*pgd))
1033 p4d = p4d_offset(pgd, address);
1034 if (!p4d_present(*p4d))
1037 if (p4d_large(*p4d))
1038 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1040 pud = pud_offset(p4d, address);
1041 if (!pud_present(*pud))
1044 if (pud_large(*pud))
1045 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1047 pmd = pmd_offset(pud, address);
1048 if (!pmd_present(*pmd))
1051 if (pmd_large(*pmd))
1052 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1054 pte = pte_offset_kernel(pmd, address);
1055 if (!pte_present(*pte))
1058 ret = spurious_kernel_fault_check(error_code, pte);
1063 * Make sure we have permissions in PMD.
1064 * If not, then there's a bug in the page tables:
1066 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1067 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1071 NOKPROBE_SYMBOL(spurious_kernel_fault);
1073 int show_unhandled_signals = 1;
1076 access_error(unsigned long error_code, struct vm_area_struct *vma)
1078 /* This is only called for the current mm, so: */
1079 bool foreign = false;
1082 * Read or write was blocked by protection keys. This is
1083 * always an unconditional error and can never result in
1084 * a follow-up action to resolve the fault, like a COW.
1086 if (error_code & X86_PF_PK)
1090 * SGX hardware blocked the access. This usually happens
1091 * when the enclave memory contents have been destroyed, like
1092 * after a suspend/resume cycle. In any case, the kernel can't
1093 * fix the cause of the fault. Handle the fault as an access
1094 * error even in cases where no actual access violation
1095 * occurred. This allows userspace to rebuild the enclave in
1096 * response to the signal.
1098 if (unlikely(error_code & X86_PF_SGX))
1102 * Make sure to check the VMA so that we do not perform
1103 * faults just to hit a X86_PF_PK as soon as we fill in a
1106 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1107 (error_code & X86_PF_INSTR), foreign))
1110 if (error_code & X86_PF_WRITE) {
1111 /* write, present and write, not present: */
1112 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1117 /* read, present: */
1118 if (unlikely(error_code & X86_PF_PROT))
1121 /* read, not present: */
1122 if (unlikely(!vma_is_accessible(vma)))
1128 bool fault_in_kernel_space(unsigned long address)
1131 * On 64-bit systems, the vsyscall page is at an address above
1132 * TASK_SIZE_MAX, but is not considered part of the kernel
1135 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1138 return address >= TASK_SIZE_MAX;
1142 * Called for all faults where 'address' is part of the kernel address
1143 * space. Might get called for faults that originate from *code* that
1144 * ran in userspace or the kernel.
1147 do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1148 unsigned long address)
1151 * Protection keys exceptions only happen on user pages. We
1152 * have no user pages in the kernel portion of the address
1153 * space, so do not expect them here.
1155 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1157 #ifdef CONFIG_X86_32
1159 * We can fault-in kernel-space virtual memory on-demand. The
1160 * 'reference' page table is init_mm.pgd.
1162 * NOTE! We MUST NOT take any locks for this case. We may
1163 * be in an interrupt or a critical region, and should
1164 * only copy the information from the master page table,
1167 * Before doing this on-demand faulting, ensure that the
1168 * fault is not any of the following:
1169 * 1. A fault on a PTE with a reserved bit set.
1170 * 2. A fault caused by a user-mode access. (Do not demand-
1171 * fault kernel memory due to user-mode accesses).
1172 * 3. A fault caused by a page-level protection violation.
1173 * (A demand fault would be on a non-present page which
1174 * would have X86_PF_PROT==0).
1176 * This is only needed to close a race condition on x86-32 in
1177 * the vmalloc mapping/unmapping code. See the comment above
1178 * vmalloc_fault() for details. On x86-64 the race does not
1179 * exist as the vmalloc mappings don't need to be synchronized
1182 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1183 if (vmalloc_fault(address) >= 0)
1188 if (is_f00f_bug(regs, hw_error_code, address))
1191 /* Was the fault spurious, caused by lazy TLB invalidation? */
1192 if (spurious_kernel_fault(hw_error_code, address))
1195 /* kprobes don't want to hook the spurious faults: */
1196 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1200 * Note, despite being a "bad area", there are quite a few
1201 * acceptable reasons to get here, such as erratum fixups
1202 * and handling kernel code that can fault, like get_user().
1204 * Don't take the mm semaphore here. If we fixup a prefetch
1205 * fault we could otherwise deadlock:
1207 bad_area_nosemaphore(regs, hw_error_code, address);
1209 NOKPROBE_SYMBOL(do_kern_addr_fault);
1212 * Handle faults in the user portion of the address space. Nothing in here
1213 * should check X86_PF_USER without a specific justification: for almost
1214 * all purposes, we should treat a normal kernel access to user memory
1215 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1216 * The one exception is AC flag handling, which is, per the x86
1217 * architecture, special for WRUSS.
1220 void do_user_addr_fault(struct pt_regs *regs,
1221 unsigned long error_code,
1222 unsigned long address)
1224 struct vm_area_struct *vma;
1225 struct task_struct *tsk;
1226 struct mm_struct *mm;
1228 unsigned int flags = FAULT_FLAG_DEFAULT;
1233 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1235 * Whoops, this is kernel mode code trying to execute from
1236 * user memory. Unless this is AMD erratum #93, which
1237 * corrupts RIP such that it looks like a user address,
1238 * this is unrecoverable. Don't even try to look up the
1239 * VMA or look for extable entries.
1241 if (is_errata93(regs, address))
1244 page_fault_oops(regs, error_code, address);
1248 /* kprobes don't want to hook the spurious faults: */
1249 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1253 * Reserved bits are never expected to be set on
1254 * entries in the user portion of the page tables.
1256 if (unlikely(error_code & X86_PF_RSVD))
1257 pgtable_bad(regs, error_code, address);
1260 * If SMAP is on, check for invalid kernel (supervisor) access to user
1261 * pages in the user address space. The odd case here is WRUSS,
1262 * which, according to the preliminary documentation, does not respect
1263 * SMAP and will have the USER bit set so, in all cases, SMAP
1264 * enforcement appears to be consistent with the USER bit.
1266 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1267 !(error_code & X86_PF_USER) &&
1268 !(regs->flags & X86_EFLAGS_AC))) {
1270 * No extable entry here. This was a kernel access to an
1271 * invalid pointer. get_kernel_nofault() will not get here.
1273 page_fault_oops(regs, error_code, address);
1278 * If we're in an interrupt, have no user context or are running
1279 * in a region with pagefaults disabled then we must not take the fault
1281 if (unlikely(faulthandler_disabled() || !mm)) {
1282 bad_area_nosemaphore(regs, error_code, address);
1287 * It's safe to allow irq's after cr2 has been saved and the
1288 * vmalloc fault has been handled.
1290 * User-mode registers count as a user access even for any
1291 * potential system fault or CPU buglet:
1293 if (user_mode(regs)) {
1295 flags |= FAULT_FLAG_USER;
1297 if (regs->flags & X86_EFLAGS_IF)
1301 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1303 if (error_code & X86_PF_WRITE)
1304 flags |= FAULT_FLAG_WRITE;
1305 if (error_code & X86_PF_INSTR)
1306 flags |= FAULT_FLAG_INSTRUCTION;
1308 #ifdef CONFIG_X86_64
1310 * Faults in the vsyscall page might need emulation. The
1311 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1312 * considered to be part of the user address space.
1314 * The vsyscall page does not have a "real" VMA, so do this
1315 * emulation before we go searching for VMAs.
1317 * PKRU never rejects instruction fetches, so we don't need
1318 * to consider the PF_PK bit.
1320 if (is_vsyscall_vaddr(address)) {
1321 if (emulate_vsyscall(error_code, regs, address))
1327 * Kernel-mode access to the user address space should only occur
1328 * on well-defined single instructions listed in the exception
1329 * tables. But, an erroneous kernel fault occurring outside one of
1330 * those areas which also holds mmap_lock might deadlock attempting
1331 * to validate the fault against the address space.
1333 * Only do the expensive exception table search when we might be at
1334 * risk of a deadlock. This happens if we
1335 * 1. Failed to acquire mmap_lock, and
1336 * 2. The access did not originate in userspace.
1338 if (unlikely(!mmap_read_trylock(mm))) {
1339 if (!user_mode(regs) && !search_exception_tables(regs->ip)) {
1341 * Fault from code in kernel from
1342 * which we do not expect faults.
1344 bad_area_nosemaphore(regs, error_code, address);
1351 * The above down_read_trylock() might have succeeded in
1352 * which case we'll have missed the might_sleep() from
1358 vma = find_vma(mm, address);
1359 if (unlikely(!vma)) {
1360 bad_area(regs, error_code, address);
1363 if (likely(vma->vm_start <= address))
1365 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1366 bad_area(regs, error_code, address);
1369 if (unlikely(expand_stack(vma, address))) {
1370 bad_area(regs, error_code, address);
1375 * Ok, we have a good vm_area for this memory access, so
1376 * we can handle it..
1379 if (unlikely(access_error(error_code, vma))) {
1380 bad_area_access_error(regs, error_code, address, vma);
1385 * If for any reason at all we couldn't handle the fault,
1386 * make sure we exit gracefully rather than endlessly redo
1387 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1388 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1390 * Note that handle_userfault() may also release and reacquire mmap_lock
1391 * (and not return with VM_FAULT_RETRY), when returning to userland to
1392 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1393 * (potentially after handling any pending signal during the return to
1394 * userland). The return to userland is identified whenever
1395 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1397 fault = handle_mm_fault(vma, address, flags, regs);
1399 if (fault_signal_pending(fault, regs)) {
1401 * Quick path to respond to signals. The core mm code
1402 * has unlocked the mm for us if we get here.
1404 if (!user_mode(regs))
1405 kernelmode_fixup_or_oops(regs, error_code, address,
1412 * If we need to retry the mmap_lock has already been released,
1413 * and if there is a fatal signal pending there is no guarantee
1414 * that we made any progress. Handle this case first.
1416 if (unlikely((fault & VM_FAULT_RETRY) &&
1417 (flags & FAULT_FLAG_ALLOW_RETRY))) {
1418 flags |= FAULT_FLAG_TRIED;
1422 mmap_read_unlock(mm);
1423 if (likely(!(fault & VM_FAULT_ERROR)))
1426 if (fatal_signal_pending(current) && !user_mode(regs)) {
1427 kernelmode_fixup_or_oops(regs, error_code, address,
1428 0, 0, ARCH_DEFAULT_PKEY);
1432 if (fault & VM_FAULT_OOM) {
1433 /* Kernel mode? Handle exceptions or die: */
1434 if (!user_mode(regs)) {
1435 kernelmode_fixup_or_oops(regs, error_code, address,
1436 SIGSEGV, SEGV_MAPERR,
1442 * We ran out of memory, call the OOM killer, and return the
1443 * userspace (which will retry the fault, or kill us if we got
1446 pagefault_out_of_memory();
1448 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1449 VM_FAULT_HWPOISON_LARGE))
1450 do_sigbus(regs, error_code, address, fault);
1451 else if (fault & VM_FAULT_SIGSEGV)
1452 bad_area_nosemaphore(regs, error_code, address);
1457 NOKPROBE_SYMBOL(do_user_addr_fault);
1459 static __always_inline void
1460 trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1461 unsigned long address)
1463 if (!trace_pagefault_enabled())
1466 if (user_mode(regs))
1467 trace_page_fault_user(address, regs, error_code);
1469 trace_page_fault_kernel(address, regs, error_code);
1472 static __always_inline void
1473 handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1474 unsigned long address)
1476 trace_page_fault_entries(regs, error_code, address);
1478 if (unlikely(kmmio_fault(regs, address)))
1481 /* Was the fault on kernel-controlled part of the address space? */
1482 if (unlikely(fault_in_kernel_space(address))) {
1483 do_kern_addr_fault(regs, error_code, address);
1485 do_user_addr_fault(regs, error_code, address);
1487 * User address page fault handling might have reenabled
1488 * interrupts. Fixing up all potential exit points of
1489 * do_user_addr_fault() and its leaf functions is just not
1490 * doable w/o creating an unholy mess or turning the code
1493 local_irq_disable();
1497 DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1499 unsigned long address = read_cr2();
1500 irqentry_state_t state;
1502 prefetchw(¤t->mm->mmap_lock);
1505 * KVM uses #PF vector to deliver 'page not present' events to guests
1506 * (asynchronous page fault mechanism). The event happens when a
1507 * userspace task is trying to access some valid (from guest's point of
1508 * view) memory which is not currently mapped by the host (e.g. the
1509 * memory is swapped out). Note, the corresponding "page ready" event
1510 * which is injected when the memory becomes available, is delivered via
1511 * an interrupt mechanism and not a #PF exception
1512 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1514 * We are relying on the interrupted context being sane (valid RSP,
1515 * relevant locks not held, etc.), which is fine as long as the
1516 * interrupted context had IF=1. We are also relying on the KVM
1517 * async pf type field and CR2 being read consistently instead of
1518 * getting values from real and async page faults mixed up.
1522 * The async #PF handling code takes care of idtentry handling
1525 if (kvm_handle_async_pf(regs, (u32)address))
1529 * Entry handling for valid #PF from kernel mode is slightly
1530 * different: RCU is already watching and rcu_irq_enter() must not
1531 * be invoked because a kernel fault on a user space address might
1534 * In case the fault hit a RCU idle region the conditional entry
1535 * code reenabled RCU to avoid subsequent wreckage which helps
1538 state = irqentry_enter(regs);
1540 instrumentation_begin();
1541 handle_page_fault(regs, error_code, address);
1542 instrumentation_end();
1544 irqentry_exit(regs, state);