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>
22 #include <linux/mm.h> /* find_and_lock_vma() */
24 #include <asm/cpufeature.h> /* boot_cpu_has, ... */
25 #include <asm/traps.h> /* dotraplinkage, ... */
26 #include <asm/fixmap.h> /* VSYSCALL_ADDR */
27 #include <asm/vsyscall.h> /* emulate_vsyscall */
28 #include <asm/vm86.h> /* struct vm86 */
29 #include <asm/mmu_context.h> /* vma_pkey() */
30 #include <asm/efi.h> /* efi_crash_gracefully_on_page_fault()*/
31 #include <asm/desc.h> /* store_idt(), ... */
32 #include <asm/cpu_entry_area.h> /* exception stack */
33 #include <asm/pgtable_areas.h> /* VMALLOC_START, ... */
34 #include <asm/kvm_para.h> /* kvm_handle_async_pf */
35 #include <asm/vdso.h> /* fixup_vdso_exception() */
36 #include <asm/irq_stack.h>
37 #include <asm/sev.h> /* snp_dump_hva_rmpentry() */
39 #define CREATE_TRACE_POINTS
40 #include <asm/trace/exceptions.h>
43 * Returns 0 if mmiotrace is disabled, or if the fault is not
44 * handled by mmiotrace:
46 static nokprobe_inline int
47 kmmio_fault(struct pt_regs *regs, unsigned long addr)
49 if (unlikely(is_kmmio_active()))
50 if (kmmio_handler(regs, addr) == 1)
60 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
61 * Check that here and ignore it. This is AMD erratum #91.
65 * Sometimes the CPU reports invalid exceptions on prefetch.
66 * Check that here and ignore it.
68 * Opcode checker based on code by Richard Brunner.
71 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
72 unsigned char opcode, int *prefetch)
74 unsigned char instr_hi = opcode & 0xf0;
75 unsigned char instr_lo = opcode & 0x0f;
81 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
82 * In X86_64 long mode, the CPU will signal invalid
83 * opcode if some of these prefixes are present so
84 * X86_64 will never get here anyway
86 return ((instr_lo & 7) == 0x6);
90 * In 64-bit mode 0x40..0x4F are valid REX prefixes
92 return (!user_mode(regs) || user_64bit_mode(regs));
95 /* 0x64 thru 0x67 are valid prefixes in all modes. */
96 return (instr_lo & 0xC) == 0x4;
98 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
99 return !instr_lo || (instr_lo>>1) == 1;
101 /* Prefetch instruction is 0x0F0D or 0x0F18 */
102 if (get_kernel_nofault(opcode, instr))
105 *prefetch = (instr_lo == 0xF) &&
106 (opcode == 0x0D || opcode == 0x18);
113 static bool is_amd_k8_pre_npt(void)
115 struct cpuinfo_x86 *c = &boot_cpu_data;
117 return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
118 c->x86_vendor == X86_VENDOR_AMD &&
119 c->x86 == 0xf && c->x86_model < 0x40);
123 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
125 unsigned char *max_instr;
126 unsigned char *instr;
129 /* Erratum #91 affects AMD K8, pre-NPT CPUs */
130 if (!is_amd_k8_pre_npt())
134 * If it was a exec (instruction fetch) fault on NX page, then
135 * do not ignore the fault:
137 if (error_code & X86_PF_INSTR)
140 instr = (void *)convert_ip_to_linear(current, regs);
141 max_instr = instr + 15;
144 * This code has historically always bailed out if IP points to a
145 * not-present page (e.g. due to a race). No one has ever
146 * complained about this.
150 while (instr < max_instr) {
151 unsigned char opcode;
153 if (user_mode(regs)) {
154 if (get_user(opcode, (unsigned char __user *) instr))
157 if (get_kernel_nofault(opcode, instr))
163 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
171 DEFINE_SPINLOCK(pgd_lock);
175 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
177 unsigned index = pgd_index(address);
184 pgd_k = init_mm.pgd + index;
186 if (!pgd_present(*pgd_k))
190 * set_pgd(pgd, *pgd_k); here would be useless on PAE
191 * and redundant with the set_pmd() on non-PAE. As would
194 p4d = p4d_offset(pgd, address);
195 p4d_k = p4d_offset(pgd_k, address);
196 if (!p4d_present(*p4d_k))
199 pud = pud_offset(p4d, address);
200 pud_k = pud_offset(p4d_k, address);
201 if (!pud_present(*pud_k))
204 pmd = pmd_offset(pud, address);
205 pmd_k = pmd_offset(pud_k, address);
207 if (pmd_present(*pmd) != pmd_present(*pmd_k))
208 set_pmd(pmd, *pmd_k);
210 if (!pmd_present(*pmd_k))
213 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
219 * Handle a fault on the vmalloc or module mapping area
221 * This is needed because there is a race condition between the time
222 * when the vmalloc mapping code updates the PMD to the point in time
223 * where it synchronizes this update with the other page-tables in the
226 * In this race window another thread/CPU can map an area on the same
227 * PMD, finds it already present and does not synchronize it with the
228 * rest of the system yet. As a result v[mz]alloc might return areas
229 * which are not mapped in every page-table in the system, causing an
230 * unhandled page-fault when they are accessed.
232 static noinline int vmalloc_fault(unsigned long address)
234 unsigned long pgd_paddr;
238 /* Make sure we are in vmalloc area: */
239 if (!(address >= VMALLOC_START && address < VMALLOC_END))
243 * Synchronize this task's top level page-table
244 * with the 'reference' page table.
246 * Do _not_ use "current" here. We might be inside
247 * an interrupt in the middle of a task switch..
249 pgd_paddr = read_cr3_pa();
250 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
254 if (pmd_large(*pmd_k))
257 pte_k = pte_offset_kernel(pmd_k, address);
258 if (!pte_present(*pte_k))
263 NOKPROBE_SYMBOL(vmalloc_fault);
265 void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
269 for (addr = start & PMD_MASK;
270 addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
274 spin_lock(&pgd_lock);
275 list_for_each_entry(page, &pgd_list, lru) {
276 spinlock_t *pgt_lock;
278 /* the pgt_lock only for Xen */
279 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
282 vmalloc_sync_one(page_address(page), addr);
283 spin_unlock(pgt_lock);
285 spin_unlock(&pgd_lock);
289 static bool low_pfn(unsigned long pfn)
291 return pfn < max_low_pfn;
294 static void dump_pagetable(unsigned long address)
296 pgd_t *base = __va(read_cr3_pa());
297 pgd_t *pgd = &base[pgd_index(address)];
303 #ifdef CONFIG_X86_PAE
304 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
305 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
307 #define pr_pde pr_cont
309 #define pr_pde pr_info
311 p4d = p4d_offset(pgd, address);
312 pud = pud_offset(p4d, address);
313 pmd = pmd_offset(pud, address);
314 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
318 * We must not directly access the pte in the highpte
319 * case if the page table is located in highmem.
320 * And let's rather not kmap-atomic the pte, just in case
321 * it's allocated already:
323 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
326 pte = pte_offset_kernel(pmd, address);
327 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
332 #else /* CONFIG_X86_64: */
334 #ifdef CONFIG_CPU_SUP_AMD
335 static const char errata93_warning[] =
337 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
338 "******* Working around it, but it may cause SEGVs or burn power.\n"
339 "******* Please consider a BIOS update.\n"
340 "******* Disabling USB legacy in the BIOS may also help.\n";
343 static int bad_address(void *p)
347 return get_kernel_nofault(dummy, (unsigned long *)p);
350 static void dump_pagetable(unsigned long address)
352 pgd_t *base = __va(read_cr3_pa());
353 pgd_t *pgd = base + pgd_index(address);
359 if (bad_address(pgd))
362 pr_info("PGD %lx ", pgd_val(*pgd));
364 if (!pgd_present(*pgd))
367 p4d = p4d_offset(pgd, address);
368 if (bad_address(p4d))
371 pr_cont("P4D %lx ", p4d_val(*p4d));
372 if (!p4d_present(*p4d) || p4d_large(*p4d))
375 pud = pud_offset(p4d, address);
376 if (bad_address(pud))
379 pr_cont("PUD %lx ", pud_val(*pud));
380 if (!pud_present(*pud) || pud_large(*pud))
383 pmd = pmd_offset(pud, address);
384 if (bad_address(pmd))
387 pr_cont("PMD %lx ", pmd_val(*pmd));
388 if (!pmd_present(*pmd) || pmd_large(*pmd))
391 pte = pte_offset_kernel(pmd, address);
392 if (bad_address(pte))
395 pr_cont("PTE %lx", pte_val(*pte));
403 #endif /* CONFIG_X86_64 */
406 * Workaround for K8 erratum #93 & buggy BIOS.
408 * BIOS SMM functions are required to use a specific workaround
409 * to avoid corruption of the 64bit RIP register on C stepping K8.
411 * A lot of BIOS that didn't get tested properly miss this.
413 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
414 * Try to work around it here.
416 * Note we only handle faults in kernel here.
417 * Does nothing on 32-bit.
419 static int is_errata93(struct pt_regs *regs, unsigned long address)
421 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
422 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
423 || boot_cpu_data.x86 != 0xf)
429 if (address != regs->ip)
432 if ((address >> 32) != 0)
435 address |= 0xffffffffUL << 32;
436 if ((address >= (u64)_stext && address <= (u64)_etext) ||
437 (address >= MODULES_VADDR && address <= MODULES_END)) {
438 printk_once(errata93_warning);
447 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
448 * to illegal addresses >4GB.
450 * We catch this in the page fault handler because these addresses
451 * are not reachable. Just detect this case and return. Any code
452 * segment in LDT is compatibility mode.
454 static int is_errata100(struct pt_regs *regs, unsigned long address)
457 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
463 /* Pentium F0 0F C7 C8 bug workaround: */
464 static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
465 unsigned long address)
467 #ifdef CONFIG_X86_F00F_BUG
468 if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
469 idt_is_f00f_address(address)) {
470 handle_invalid_op(regs);
477 static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
479 u32 offset = (index >> 3) * sizeof(struct desc_struct);
481 struct ldttss_desc desc;
484 pr_alert("%s: NULL\n", name);
488 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
489 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
493 if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset),
494 sizeof(struct ldttss_desc))) {
495 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
500 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
502 addr |= ((u64)desc.base3 << 32);
504 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
505 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
509 show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
511 if (!oops_may_print())
514 if (error_code & X86_PF_INSTR) {
519 pgd = __va(read_cr3_pa());
520 pgd += pgd_index(address);
522 pte = lookup_address_in_pgd(pgd, address, &level);
524 if (pte && pte_present(*pte) && !pte_exec(*pte))
525 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
526 from_kuid(&init_user_ns, current_uid()));
527 if (pte && pte_present(*pte) && pte_exec(*pte) &&
528 (pgd_flags(*pgd) & _PAGE_USER) &&
529 (__read_cr4() & X86_CR4_SMEP))
530 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
531 from_kuid(&init_user_ns, current_uid()));
534 if (address < PAGE_SIZE && !user_mode(regs))
535 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
538 pr_alert("BUG: unable to handle page fault for address: %px\n",
541 pr_alert("#PF: %s %s in %s mode\n",
542 (error_code & X86_PF_USER) ? "user" : "supervisor",
543 (error_code & X86_PF_INSTR) ? "instruction fetch" :
544 (error_code & X86_PF_WRITE) ? "write access" :
546 user_mode(regs) ? "user" : "kernel");
547 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
548 !(error_code & X86_PF_PROT) ? "not-present page" :
549 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
550 (error_code & X86_PF_PK) ? "protection keys violation" :
551 (error_code & X86_PF_RMP) ? "RMP violation" :
552 "permissions violation");
554 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
555 struct desc_ptr idt, gdt;
559 * This can happen for quite a few reasons. The more obvious
560 * ones are faults accessing the GDT, or LDT. Perhaps
561 * surprisingly, if the CPU tries to deliver a benign or
562 * contributory exception from user code and gets a page fault
563 * during delivery, the page fault can be delivered as though
564 * it originated directly from user code. This could happen
565 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
566 * kernel or IST stack.
570 /* Usable even on Xen PV -- it's just slow. */
571 native_store_gdt(&gdt);
573 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
574 idt.address, idt.size, gdt.address, gdt.size);
577 show_ldttss(&gdt, "LDTR", ldtr);
580 show_ldttss(&gdt, "TR", tr);
583 dump_pagetable(address);
585 if (error_code & X86_PF_RMP)
586 snp_dump_hva_rmpentry(address);
590 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
591 unsigned long address)
593 struct task_struct *tsk;
597 flags = oops_begin();
601 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
603 dump_pagetable(address);
605 if (__die("Bad pagetable", regs, error_code))
608 oops_end(flags, regs, sig);
611 static void sanitize_error_code(unsigned long address,
612 unsigned long *error_code)
615 * To avoid leaking information about the kernel page
616 * table layout, pretend that user-mode accesses to
617 * kernel addresses are always protection faults.
619 * NB: This means that failed vsyscalls with vsyscall=none
620 * will have the PROT bit. This doesn't leak any
621 * information and does not appear to cause any problems.
623 if (address >= TASK_SIZE_MAX)
624 *error_code |= X86_PF_PROT;
627 static void set_signal_archinfo(unsigned long address,
628 unsigned long error_code)
630 struct task_struct *tsk = current;
632 tsk->thread.trap_nr = X86_TRAP_PF;
633 tsk->thread.error_code = error_code | X86_PF_USER;
634 tsk->thread.cr2 = address;
638 page_fault_oops(struct pt_regs *regs, unsigned long error_code,
639 unsigned long address)
641 #ifdef CONFIG_VMAP_STACK
642 struct stack_info info;
647 if (user_mode(regs)) {
649 * Implicit kernel access from user mode? Skip the stack
650 * overflow and EFI special cases.
655 #ifdef CONFIG_VMAP_STACK
657 * Stack overflow? During boot, we can fault near the initial
658 * stack in the direct map, but that's not an overflow -- check
659 * that we're in vmalloc space to avoid this.
661 if (is_vmalloc_addr((void *)address) &&
662 get_stack_guard_info((void *)address, &info)) {
664 * We're likely to be running with very little stack space
665 * left. It's plausible that we'd hit this condition but
666 * double-fault even before we get this far, in which case
667 * we're fine: the double-fault handler will deal with it.
669 * We don't want to make it all the way into the oops code
670 * and then double-fault, though, because we're likely to
671 * break the console driver and lose most of the stack dump.
673 call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*),
674 handle_stack_overflow,
676 , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info));
683 * Buggy firmware could access regions which might page fault. If
684 * this happens, EFI has a special OOPS path that will try to
685 * avoid hanging the system.
687 if (IS_ENABLED(CONFIG_EFI))
688 efi_crash_gracefully_on_page_fault(address);
690 /* Only not-present faults should be handled by KFENCE. */
691 if (!(error_code & X86_PF_PROT) &&
692 kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs))
697 * Oops. The kernel tried to access some bad page. We'll have to
698 * terminate things with extreme prejudice:
700 flags = oops_begin();
702 show_fault_oops(regs, error_code, address);
704 if (task_stack_end_corrupted(current))
705 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
708 if (__die("Oops", regs, error_code))
711 /* Executive summary in case the body of the oops scrolled away */
712 printk(KERN_DEFAULT "CR2: %016lx\n", address);
714 oops_end(flags, regs, sig);
718 kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
719 unsigned long address, int signal, int si_code,
722 WARN_ON_ONCE(user_mode(regs));
724 /* Are we prepared to handle this kernel fault? */
725 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
727 * Any interrupt that takes a fault gets the fixup. This makes
728 * the below recursive fault logic only apply to a faults from
735 * Per the above we're !in_interrupt(), aka. task context.
737 * In this case we need to make sure we're not recursively
738 * faulting through the emulate_vsyscall() logic.
740 if (current->thread.sig_on_uaccess_err && signal) {
741 sanitize_error_code(address, &error_code);
743 set_signal_archinfo(address, error_code);
745 if (si_code == SEGV_PKUERR) {
746 force_sig_pkuerr((void __user *)address, pkey);
748 /* XXX: hwpoison faults will set the wrong code. */
749 force_sig_fault(signal, si_code, (void __user *)address);
754 * Barring that, we can do the fixup and be happy.
760 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
763 if (is_prefetch(regs, error_code, address))
766 page_fault_oops(regs, error_code, address);
770 * Print out info about fatal segfaults, if the show_unhandled_signals
774 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
775 unsigned long address, struct task_struct *tsk)
777 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
778 /* This is a racy snapshot, but it's better than nothing. */
779 int cpu = raw_smp_processor_id();
781 if (!unhandled_signal(tsk, SIGSEGV))
784 if (!printk_ratelimit())
787 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
788 loglvl, tsk->comm, task_pid_nr(tsk), address,
789 (void *)regs->ip, (void *)regs->sp, error_code);
791 print_vma_addr(KERN_CONT " in ", regs->ip);
794 * Dump the likely CPU where the fatal segfault happened.
795 * This can help identify faulty hardware.
797 printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu,
798 topology_core_id(cpu), topology_physical_package_id(cpu));
801 printk(KERN_CONT "\n");
803 show_opcodes(regs, loglvl);
807 * The (legacy) vsyscall page is the long page in the kernel portion
808 * of the address space that has user-accessible permissions.
810 static bool is_vsyscall_vaddr(unsigned long vaddr)
812 return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR);
816 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
817 unsigned long address, u32 pkey, int si_code)
819 struct task_struct *tsk = current;
821 if (!user_mode(regs)) {
822 kernelmode_fixup_or_oops(regs, error_code, address,
823 SIGSEGV, si_code, pkey);
827 if (!(error_code & X86_PF_USER)) {
828 /* Implicit user access to kernel memory -- just oops */
829 page_fault_oops(regs, error_code, address);
834 * User mode accesses just cause a SIGSEGV.
835 * It's possible to have interrupts off here:
840 * Valid to do another page fault here because this one came
843 if (is_prefetch(regs, error_code, address))
846 if (is_errata100(regs, address))
849 sanitize_error_code(address, &error_code);
851 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
854 if (likely(show_unhandled_signals))
855 show_signal_msg(regs, error_code, address, tsk);
857 set_signal_archinfo(address, error_code);
859 if (si_code == SEGV_PKUERR)
860 force_sig_pkuerr((void __user *)address, pkey);
862 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
868 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
869 unsigned long address)
871 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
875 __bad_area(struct pt_regs *regs, unsigned long error_code,
876 unsigned long address, u32 pkey, int si_code)
878 struct mm_struct *mm = current->mm;
880 * Something tried to access memory that isn't in our memory map..
881 * Fix it, but check if it's kernel or user first..
883 mmap_read_unlock(mm);
885 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
888 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
889 struct vm_area_struct *vma)
891 /* This code is always called on the current mm */
892 bool foreign = false;
894 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
896 if (error_code & X86_PF_PK)
898 /* this checks permission keys on the VMA: */
899 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
900 (error_code & X86_PF_INSTR), foreign))
906 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
907 unsigned long address, struct vm_area_struct *vma)
910 * This OSPKE check is not strictly necessary at runtime.
911 * But, doing it this way allows compiler optimizations
912 * if pkeys are compiled out.
914 if (bad_area_access_from_pkeys(error_code, vma)) {
916 * A protection key fault means that the PKRU value did not allow
917 * access to some PTE. Userspace can figure out what PKRU was
918 * from the XSAVE state. This function captures the pkey from
919 * the vma and passes it to userspace so userspace can discover
920 * which protection key was set on the PTE.
922 * If we get here, we know that the hardware signaled a X86_PF_PK
923 * fault and that there was a VMA once we got in the fault
924 * handler. It does *not* guarantee that the VMA we find here
925 * was the one that we faulted on.
927 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
928 * 2. T1 : set PKRU to deny access to pkey=4, touches page
930 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
931 * 5. T1 : enters fault handler, takes mmap_lock, etc...
932 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
933 * faulted on a pte with its pkey=4.
935 u32 pkey = vma_pkey(vma);
937 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
939 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
944 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
947 /* Kernel mode? Handle exceptions or die: */
948 if (!user_mode(regs)) {
949 kernelmode_fixup_or_oops(regs, error_code, address,
950 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
954 /* User-space => ok to do another page fault: */
955 if (is_prefetch(regs, error_code, address))
958 sanitize_error_code(address, &error_code);
960 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
963 set_signal_archinfo(address, error_code);
965 #ifdef CONFIG_MEMORY_FAILURE
966 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
967 struct task_struct *tsk = current;
971 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
972 tsk->comm, tsk->pid, address);
973 if (fault & VM_FAULT_HWPOISON_LARGE)
974 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
975 if (fault & VM_FAULT_HWPOISON)
977 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
981 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
984 static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
986 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
989 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
996 * Handle a spurious fault caused by a stale TLB entry.
998 * This allows us to lazily refresh the TLB when increasing the
999 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1000 * eagerly is very expensive since that implies doing a full
1001 * cross-processor TLB flush, even if no stale TLB entries exist
1002 * on other processors.
1004 * Spurious faults may only occur if the TLB contains an entry with
1005 * fewer permission than the page table entry. Non-present (P = 0)
1006 * and reserved bit (R = 1) faults are never spurious.
1008 * There are no security implications to leaving a stale TLB when
1009 * increasing the permissions on a page.
1011 * Returns non-zero if a spurious fault was handled, zero otherwise.
1013 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1014 * (Optional Invalidation).
1017 spurious_kernel_fault(unsigned long error_code, unsigned long address)
1027 * Only writes to RO or instruction fetches from NX may cause
1030 * These could be from user or supervisor accesses but the TLB
1031 * is only lazily flushed after a kernel mapping protection
1032 * change, so user accesses are not expected to cause spurious
1035 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1036 error_code != (X86_PF_INSTR | X86_PF_PROT))
1039 pgd = init_mm.pgd + pgd_index(address);
1040 if (!pgd_present(*pgd))
1043 p4d = p4d_offset(pgd, address);
1044 if (!p4d_present(*p4d))
1047 if (p4d_large(*p4d))
1048 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1050 pud = pud_offset(p4d, address);
1051 if (!pud_present(*pud))
1054 if (pud_large(*pud))
1055 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1057 pmd = pmd_offset(pud, address);
1058 if (!pmd_present(*pmd))
1061 if (pmd_large(*pmd))
1062 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1064 pte = pte_offset_kernel(pmd, address);
1065 if (!pte_present(*pte))
1068 ret = spurious_kernel_fault_check(error_code, pte);
1073 * Make sure we have permissions in PMD.
1074 * If not, then there's a bug in the page tables:
1076 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1077 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1081 NOKPROBE_SYMBOL(spurious_kernel_fault);
1083 int show_unhandled_signals = 1;
1086 access_error(unsigned long error_code, struct vm_area_struct *vma)
1088 /* This is only called for the current mm, so: */
1089 bool foreign = false;
1092 * Read or write was blocked by protection keys. This is
1093 * always an unconditional error and can never result in
1094 * a follow-up action to resolve the fault, like a COW.
1096 if (error_code & X86_PF_PK)
1100 * SGX hardware blocked the access. This usually happens
1101 * when the enclave memory contents have been destroyed, like
1102 * after a suspend/resume cycle. In any case, the kernel can't
1103 * fix the cause of the fault. Handle the fault as an access
1104 * error even in cases where no actual access violation
1105 * occurred. This allows userspace to rebuild the enclave in
1106 * response to the signal.
1108 if (unlikely(error_code & X86_PF_SGX))
1112 * Make sure to check the VMA so that we do not perform
1113 * faults just to hit a X86_PF_PK as soon as we fill in a
1116 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1117 (error_code & X86_PF_INSTR), foreign))
1121 * Shadow stack accesses (PF_SHSTK=1) are only permitted to
1122 * shadow stack VMAs. All other accesses result in an error.
1124 if (error_code & X86_PF_SHSTK) {
1125 if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK)))
1127 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1132 if (error_code & X86_PF_WRITE) {
1133 /* write, present and write, not present: */
1134 if (unlikely(vma->vm_flags & VM_SHADOW_STACK))
1136 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1141 /* read, present: */
1142 if (unlikely(error_code & X86_PF_PROT))
1145 /* read, not present: */
1146 if (unlikely(!vma_is_accessible(vma)))
1152 bool fault_in_kernel_space(unsigned long address)
1155 * On 64-bit systems, the vsyscall page is at an address above
1156 * TASK_SIZE_MAX, but is not considered part of the kernel
1159 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1162 return address >= TASK_SIZE_MAX;
1166 * Called for all faults where 'address' is part of the kernel address
1167 * space. Might get called for faults that originate from *code* that
1168 * ran in userspace or the kernel.
1171 do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1172 unsigned long address)
1175 * Protection keys exceptions only happen on user pages. We
1176 * have no user pages in the kernel portion of the address
1177 * space, so do not expect them here.
1179 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1181 #ifdef CONFIG_X86_32
1183 * We can fault-in kernel-space virtual memory on-demand. The
1184 * 'reference' page table is init_mm.pgd.
1186 * NOTE! We MUST NOT take any locks for this case. We may
1187 * be in an interrupt or a critical region, and should
1188 * only copy the information from the master page table,
1191 * Before doing this on-demand faulting, ensure that the
1192 * fault is not any of the following:
1193 * 1. A fault on a PTE with a reserved bit set.
1194 * 2. A fault caused by a user-mode access. (Do not demand-
1195 * fault kernel memory due to user-mode accesses).
1196 * 3. A fault caused by a page-level protection violation.
1197 * (A demand fault would be on a non-present page which
1198 * would have X86_PF_PROT==0).
1200 * This is only needed to close a race condition on x86-32 in
1201 * the vmalloc mapping/unmapping code. See the comment above
1202 * vmalloc_fault() for details. On x86-64 the race does not
1203 * exist as the vmalloc mappings don't need to be synchronized
1206 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1207 if (vmalloc_fault(address) >= 0)
1212 if (is_f00f_bug(regs, hw_error_code, address))
1215 /* Was the fault spurious, caused by lazy TLB invalidation? */
1216 if (spurious_kernel_fault(hw_error_code, address))
1219 /* kprobes don't want to hook the spurious faults: */
1220 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1224 * Note, despite being a "bad area", there are quite a few
1225 * acceptable reasons to get here, such as erratum fixups
1226 * and handling kernel code that can fault, like get_user().
1228 * Don't take the mm semaphore here. If we fixup a prefetch
1229 * fault we could otherwise deadlock:
1231 bad_area_nosemaphore(regs, hw_error_code, address);
1233 NOKPROBE_SYMBOL(do_kern_addr_fault);
1236 * Handle faults in the user portion of the address space. Nothing in here
1237 * should check X86_PF_USER without a specific justification: for almost
1238 * all purposes, we should treat a normal kernel access to user memory
1239 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1240 * The one exception is AC flag handling, which is, per the x86
1241 * architecture, special for WRUSS.
1244 void do_user_addr_fault(struct pt_regs *regs,
1245 unsigned long error_code,
1246 unsigned long address)
1248 struct vm_area_struct *vma;
1249 struct task_struct *tsk;
1250 struct mm_struct *mm;
1252 unsigned int flags = FAULT_FLAG_DEFAULT;
1257 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1259 * Whoops, this is kernel mode code trying to execute from
1260 * user memory. Unless this is AMD erratum #93, which
1261 * corrupts RIP such that it looks like a user address,
1262 * this is unrecoverable. Don't even try to look up the
1263 * VMA or look for extable entries.
1265 if (is_errata93(regs, address))
1268 page_fault_oops(regs, error_code, address);
1272 /* kprobes don't want to hook the spurious faults: */
1273 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1277 * Reserved bits are never expected to be set on
1278 * entries in the user portion of the page tables.
1280 if (unlikely(error_code & X86_PF_RSVD))
1281 pgtable_bad(regs, error_code, address);
1284 * If SMAP is on, check for invalid kernel (supervisor) access to user
1285 * pages in the user address space. The odd case here is WRUSS,
1286 * which, according to the preliminary documentation, does not respect
1287 * SMAP and will have the USER bit set so, in all cases, SMAP
1288 * enforcement appears to be consistent with the USER bit.
1290 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1291 !(error_code & X86_PF_USER) &&
1292 !(regs->flags & X86_EFLAGS_AC))) {
1294 * No extable entry here. This was a kernel access to an
1295 * invalid pointer. get_kernel_nofault() will not get here.
1297 page_fault_oops(regs, error_code, address);
1302 * If we're in an interrupt, have no user context or are running
1303 * in a region with pagefaults disabled then we must not take the fault
1305 if (unlikely(faulthandler_disabled() || !mm)) {
1306 bad_area_nosemaphore(regs, error_code, address);
1311 * It's safe to allow irq's after cr2 has been saved and the
1312 * vmalloc fault has been handled.
1314 * User-mode registers count as a user access even for any
1315 * potential system fault or CPU buglet:
1317 if (user_mode(regs)) {
1319 flags |= FAULT_FLAG_USER;
1321 if (regs->flags & X86_EFLAGS_IF)
1325 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1328 * Read-only permissions can not be expressed in shadow stack PTEs.
1329 * Treat all shadow stack accesses as WRITE faults. This ensures
1330 * that the MM will prepare everything (e.g., break COW) such that
1331 * maybe_mkwrite() can create a proper shadow stack PTE.
1333 if (error_code & X86_PF_SHSTK)
1334 flags |= FAULT_FLAG_WRITE;
1335 if (error_code & X86_PF_WRITE)
1336 flags |= FAULT_FLAG_WRITE;
1337 if (error_code & X86_PF_INSTR)
1338 flags |= FAULT_FLAG_INSTRUCTION;
1340 #ifdef CONFIG_X86_64
1342 * Faults in the vsyscall page might need emulation. The
1343 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1344 * considered to be part of the user address space.
1346 * The vsyscall page does not have a "real" VMA, so do this
1347 * emulation before we go searching for VMAs.
1349 * PKRU never rejects instruction fetches, so we don't need
1350 * to consider the PF_PK bit.
1352 if (is_vsyscall_vaddr(address)) {
1353 if (emulate_vsyscall(error_code, regs, address))
1358 if (!(flags & FAULT_FLAG_USER))
1361 vma = lock_vma_under_rcu(mm, address);
1365 if (unlikely(access_error(error_code, vma))) {
1369 fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
1370 if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
1373 if (!(fault & VM_FAULT_RETRY)) {
1374 count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1377 count_vm_vma_lock_event(VMA_LOCK_RETRY);
1378 if (fault & VM_FAULT_MAJOR)
1379 flags |= FAULT_FLAG_TRIED;
1381 /* Quick path to respond to signals */
1382 if (fault_signal_pending(fault, regs)) {
1383 if (!user_mode(regs))
1384 kernelmode_fixup_or_oops(regs, error_code, address,
1392 vma = lock_mm_and_find_vma(mm, address, regs);
1393 if (unlikely(!vma)) {
1394 bad_area_nosemaphore(regs, error_code, address);
1399 * Ok, we have a good vm_area for this memory access, so
1400 * we can handle it..
1402 if (unlikely(access_error(error_code, vma))) {
1403 bad_area_access_error(regs, error_code, address, vma);
1408 * If for any reason at all we couldn't handle the fault,
1409 * make sure we exit gracefully rather than endlessly redo
1410 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1411 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1413 * Note that handle_userfault() may also release and reacquire mmap_lock
1414 * (and not return with VM_FAULT_RETRY), when returning to userland to
1415 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1416 * (potentially after handling any pending signal during the return to
1417 * userland). The return to userland is identified whenever
1418 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1420 fault = handle_mm_fault(vma, address, flags, regs);
1422 if (fault_signal_pending(fault, regs)) {
1424 * Quick path to respond to signals. The core mm code
1425 * has unlocked the mm for us if we get here.
1427 if (!user_mode(regs))
1428 kernelmode_fixup_or_oops(regs, error_code, address,
1434 /* The fault is fully completed (including releasing mmap lock) */
1435 if (fault & VM_FAULT_COMPLETED)
1439 * If we need to retry the mmap_lock has already been released,
1440 * and if there is a fatal signal pending there is no guarantee
1441 * that we made any progress. Handle this case first.
1443 if (unlikely(fault & VM_FAULT_RETRY)) {
1444 flags |= FAULT_FLAG_TRIED;
1448 mmap_read_unlock(mm);
1450 if (likely(!(fault & VM_FAULT_ERROR)))
1453 if (fatal_signal_pending(current) && !user_mode(regs)) {
1454 kernelmode_fixup_or_oops(regs, error_code, address,
1455 0, 0, ARCH_DEFAULT_PKEY);
1459 if (fault & VM_FAULT_OOM) {
1460 /* Kernel mode? Handle exceptions or die: */
1461 if (!user_mode(regs)) {
1462 kernelmode_fixup_or_oops(regs, error_code, address,
1463 SIGSEGV, SEGV_MAPERR,
1469 * We ran out of memory, call the OOM killer, and return the
1470 * userspace (which will retry the fault, or kill us if we got
1473 pagefault_out_of_memory();
1475 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1476 VM_FAULT_HWPOISON_LARGE))
1477 do_sigbus(regs, error_code, address, fault);
1478 else if (fault & VM_FAULT_SIGSEGV)
1479 bad_area_nosemaphore(regs, error_code, address);
1484 NOKPROBE_SYMBOL(do_user_addr_fault);
1486 static __always_inline void
1487 trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1488 unsigned long address)
1490 if (!trace_pagefault_enabled())
1493 if (user_mode(regs))
1494 trace_page_fault_user(address, regs, error_code);
1496 trace_page_fault_kernel(address, regs, error_code);
1499 static __always_inline void
1500 handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1501 unsigned long address)
1503 trace_page_fault_entries(regs, error_code, address);
1505 if (unlikely(kmmio_fault(regs, address)))
1508 /* Was the fault on kernel-controlled part of the address space? */
1509 if (unlikely(fault_in_kernel_space(address))) {
1510 do_kern_addr_fault(regs, error_code, address);
1512 do_user_addr_fault(regs, error_code, address);
1514 * User address page fault handling might have reenabled
1515 * interrupts. Fixing up all potential exit points of
1516 * do_user_addr_fault() and its leaf functions is just not
1517 * doable w/o creating an unholy mess or turning the code
1520 local_irq_disable();
1524 DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1526 unsigned long address = read_cr2();
1527 irqentry_state_t state;
1529 prefetchw(¤t->mm->mmap_lock);
1532 * KVM uses #PF vector to deliver 'page not present' events to guests
1533 * (asynchronous page fault mechanism). The event happens when a
1534 * userspace task is trying to access some valid (from guest's point of
1535 * view) memory which is not currently mapped by the host (e.g. the
1536 * memory is swapped out). Note, the corresponding "page ready" event
1537 * which is injected when the memory becomes available, is delivered via
1538 * an interrupt mechanism and not a #PF exception
1539 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1541 * We are relying on the interrupted context being sane (valid RSP,
1542 * relevant locks not held, etc.), which is fine as long as the
1543 * interrupted context had IF=1. We are also relying on the KVM
1544 * async pf type field and CR2 being read consistently instead of
1545 * getting values from real and async page faults mixed up.
1549 * The async #PF handling code takes care of idtentry handling
1552 if (kvm_handle_async_pf(regs, (u32)address))
1556 * Entry handling for valid #PF from kernel mode is slightly
1557 * different: RCU is already watching and ct_irq_enter() must not
1558 * be invoked because a kernel fault on a user space address might
1561 * In case the fault hit a RCU idle region the conditional entry
1562 * code reenabled RCU to avoid subsequent wreckage which helps
1565 state = irqentry_enter(regs);
1567 instrumentation_begin();
1568 handle_page_fault(regs, error_code, address);
1569 instrumentation_end();
1571 irqentry_exit(regs, state);