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_code, siginfo_t *info, u32 *pkey)
177 /* This is effectively an #ifdef */
178 if (!boot_cpu_has(X86_FEATURE_OSPKE))
181 /* Fault not from Protection Keys: nothing to do */
182 if (si_code != SEGV_PKUERR)
185 * force_sig_info_fault() is called from a number of
186 * contexts, some of which have a VMA and some of which
187 * do not. The X86_PF_PK handing happens after we have a
188 * valid VMA, so we should never reach this without a
192 WARN_ONCE(1, "PKU fault with no VMA passed in");
197 * si_pkey should be thought of as a strong hint, but not
198 * absolutely guranteed to be 100% accurate because of
199 * the race explained above.
201 info->si_pkey = *pkey;
205 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
206 struct task_struct *tsk, u32 *pkey, int fault)
211 info.si_signo = si_signo;
213 info.si_code = si_code;
214 info.si_addr = (void __user *)address;
215 if (fault & VM_FAULT_HWPOISON_LARGE)
216 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
217 if (fault & VM_FAULT_HWPOISON)
219 info.si_addr_lsb = lsb;
221 fill_sig_info_pkey(si_code, &info, pkey);
223 force_sig_info(si_signo, &info, tsk);
226 DEFINE_SPINLOCK(pgd_lock);
230 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
232 unsigned index = pgd_index(address);
239 pgd_k = init_mm.pgd + index;
241 if (!pgd_present(*pgd_k))
245 * set_pgd(pgd, *pgd_k); here would be useless on PAE
246 * and redundant with the set_pmd() on non-PAE. As would
249 p4d = p4d_offset(pgd, address);
250 p4d_k = p4d_offset(pgd_k, address);
251 if (!p4d_present(*p4d_k))
254 pud = pud_offset(p4d, address);
255 pud_k = pud_offset(p4d_k, address);
256 if (!pud_present(*pud_k))
259 pmd = pmd_offset(pud, address);
260 pmd_k = pmd_offset(pud_k, address);
261 if (!pmd_present(*pmd_k))
264 if (!pmd_present(*pmd))
265 set_pmd(pmd, *pmd_k);
267 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
272 void vmalloc_sync_all(void)
274 unsigned long address;
276 if (SHARED_KERNEL_PMD)
279 for (address = VMALLOC_START & PMD_MASK;
280 address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
281 address += PMD_SIZE) {
284 spin_lock(&pgd_lock);
285 list_for_each_entry(page, &pgd_list, lru) {
286 spinlock_t *pgt_lock;
289 /* the pgt_lock only for Xen */
290 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
293 ret = vmalloc_sync_one(page_address(page), address);
294 spin_unlock(pgt_lock);
299 spin_unlock(&pgd_lock);
306 * Handle a fault on the vmalloc or module mapping area
308 static noinline int vmalloc_fault(unsigned long address)
310 unsigned long pgd_paddr;
314 /* Make sure we are in vmalloc area: */
315 if (!(address >= VMALLOC_START && address < VMALLOC_END))
318 WARN_ON_ONCE(in_nmi());
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_huge(*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)
419 pgd_t *pgd, *pgd_ref;
420 p4d_t *p4d, *p4d_ref;
421 pud_t *pud, *pud_ref;
422 pmd_t *pmd, *pmd_ref;
423 pte_t *pte, *pte_ref;
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_ref = pgd_offset_k(address);
438 if (pgd_none(*pgd_ref))
441 if (pgd_none(*pgd)) {
442 set_pgd(pgd, *pgd_ref);
443 arch_flush_lazy_mmu_mode();
444 } else if (CONFIG_PGTABLE_LEVELS > 4) {
446 * With folded p4d, pgd_none() is always false, so the pgd may
447 * point to an empty page table entry and pgd_page_vaddr()
448 * will return garbage.
450 * We will do the correct sanity check on the p4d level.
452 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
455 /* With 4-level paging, copying happens on the p4d level. */
456 p4d = p4d_offset(pgd, address);
457 p4d_ref = p4d_offset(pgd_ref, address);
458 if (p4d_none(*p4d_ref))
461 if (p4d_none(*p4d)) {
462 set_p4d(p4d, *p4d_ref);
463 arch_flush_lazy_mmu_mode();
465 BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_ref));
469 * Below here mismatches are bugs because these lower tables
473 pud = pud_offset(p4d, address);
474 pud_ref = pud_offset(p4d_ref, address);
475 if (pud_none(*pud_ref))
478 if (pud_none(*pud) || pud_pfn(*pud) != pud_pfn(*pud_ref))
484 pmd = pmd_offset(pud, address);
485 pmd_ref = pmd_offset(pud_ref, address);
486 if (pmd_none(*pmd_ref))
489 if (pmd_none(*pmd) || pmd_pfn(*pmd) != pmd_pfn(*pmd_ref))
495 pte_ref = pte_offset_kernel(pmd_ref, address);
496 if (!pte_present(*pte_ref))
499 pte = pte_offset_kernel(pmd, address);
502 * Don't use pte_page here, because the mappings can point
503 * outside mem_map, and the NUMA hash lookup cannot handle
506 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
511 NOKPROBE_SYMBOL(vmalloc_fault);
513 #ifdef CONFIG_CPU_SUP_AMD
514 static const char errata93_warning[] =
516 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
517 "******* Working around it, but it may cause SEGVs or burn power.\n"
518 "******* Please consider a BIOS update.\n"
519 "******* Disabling USB legacy in the BIOS may also help.\n";
523 * No vm86 mode in 64-bit mode:
526 check_v8086_mode(struct pt_regs *regs, unsigned long address,
527 struct task_struct *tsk)
531 static int bad_address(void *p)
535 return probe_kernel_address((unsigned long *)p, dummy);
538 static void dump_pagetable(unsigned long address)
540 pgd_t *base = __va(read_cr3_pa());
541 pgd_t *pgd = base + pgd_index(address);
547 if (bad_address(pgd))
550 pr_info("PGD %lx ", pgd_val(*pgd));
552 if (!pgd_present(*pgd))
555 p4d = p4d_offset(pgd, address);
556 if (bad_address(p4d))
559 pr_cont("P4D %lx ", p4d_val(*p4d));
560 if (!p4d_present(*p4d) || p4d_large(*p4d))
563 pud = pud_offset(p4d, address);
564 if (bad_address(pud))
567 pr_cont("PUD %lx ", pud_val(*pud));
568 if (!pud_present(*pud) || pud_large(*pud))
571 pmd = pmd_offset(pud, address);
572 if (bad_address(pmd))
575 pr_cont("PMD %lx ", pmd_val(*pmd));
576 if (!pmd_present(*pmd) || pmd_large(*pmd))
579 pte = pte_offset_kernel(pmd, address);
580 if (bad_address(pte))
583 pr_cont("PTE %lx", pte_val(*pte));
591 #endif /* CONFIG_X86_64 */
594 * Workaround for K8 erratum #93 & buggy BIOS.
596 * BIOS SMM functions are required to use a specific workaround
597 * to avoid corruption of the 64bit RIP register on C stepping K8.
599 * A lot of BIOS that didn't get tested properly miss this.
601 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
602 * Try to work around it here.
604 * Note we only handle faults in kernel here.
605 * Does nothing on 32-bit.
607 static int is_errata93(struct pt_regs *regs, unsigned long address)
609 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
610 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
611 || boot_cpu_data.x86 != 0xf)
614 if (address != regs->ip)
617 if ((address >> 32) != 0)
620 address |= 0xffffffffUL << 32;
621 if ((address >= (u64)_stext && address <= (u64)_etext) ||
622 (address >= MODULES_VADDR && address <= MODULES_END)) {
623 printk_once(errata93_warning);
632 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
633 * to illegal addresses >4GB.
635 * We catch this in the page fault handler because these addresses
636 * are not reachable. Just detect this case and return. Any code
637 * segment in LDT is compatibility mode.
639 static int is_errata100(struct pt_regs *regs, unsigned long address)
642 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
648 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
650 #ifdef CONFIG_X86_F00F_BUG
654 * Pentium F0 0F C7 C8 bug workaround:
656 if (boot_cpu_has_bug(X86_BUG_F00F)) {
657 nr = (address - idt_descr.address) >> 3;
660 do_invalid_op(regs, 0);
668 static const char nx_warning[] = KERN_CRIT
669 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
670 static const char smep_warning[] = KERN_CRIT
671 "unable to execute userspace code (SMEP?) (uid: %d)\n";
674 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
675 unsigned long address)
677 if (!oops_may_print())
680 if (error_code & X86_PF_INSTR) {
685 pgd = __va(read_cr3_pa());
686 pgd += pgd_index(address);
688 pte = lookup_address_in_pgd(pgd, address, &level);
690 if (pte && pte_present(*pte) && !pte_exec(*pte))
691 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
692 if (pte && pte_present(*pte) && pte_exec(*pte) &&
693 (pgd_flags(*pgd) & _PAGE_USER) &&
694 (__read_cr4() & X86_CR4_SMEP))
695 printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
698 printk(KERN_ALERT "BUG: unable to handle kernel ");
699 if (address < PAGE_SIZE)
700 printk(KERN_CONT "NULL pointer dereference");
702 printk(KERN_CONT "paging request");
704 printk(KERN_CONT " at %px\n", (void *) address);
705 printk(KERN_ALERT "IP: %pS\n", (void *)regs->ip);
707 dump_pagetable(address);
711 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
712 unsigned long address)
714 struct task_struct *tsk;
718 flags = oops_begin();
722 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
724 dump_pagetable(address);
726 tsk->thread.cr2 = address;
727 tsk->thread.trap_nr = X86_TRAP_PF;
728 tsk->thread.error_code = error_code;
730 if (__die("Bad pagetable", regs, error_code))
733 oops_end(flags, regs, sig);
737 no_context(struct pt_regs *regs, unsigned long error_code,
738 unsigned long address, int signal, int si_code)
740 struct task_struct *tsk = current;
744 /* Are we prepared to handle this kernel fault? */
745 if (fixup_exception(regs, X86_TRAP_PF)) {
747 * Any interrupt that takes a fault gets the fixup. This makes
748 * the below recursive fault logic only apply to a faults from
755 * Per the above we're !in_interrupt(), aka. task context.
757 * In this case we need to make sure we're not recursively
758 * faulting through the emulate_vsyscall() logic.
760 if (current->thread.sig_on_uaccess_err && signal) {
761 tsk->thread.trap_nr = X86_TRAP_PF;
762 tsk->thread.error_code = error_code | X86_PF_USER;
763 tsk->thread.cr2 = address;
765 /* XXX: hwpoison faults will set the wrong code. */
766 force_sig_info_fault(signal, si_code, address,
771 * Barring that, we can do the fixup and be happy.
776 #ifdef CONFIG_VMAP_STACK
778 * Stack overflow? During boot, we can fault near the initial
779 * stack in the direct map, but that's not an overflow -- check
780 * that we're in vmalloc space to avoid this.
782 if (is_vmalloc_addr((void *)address) &&
783 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
784 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
785 unsigned long stack = this_cpu_read(orig_ist.ist[DOUBLEFAULT_STACK]) - sizeof(void *);
787 * We're likely to be running with very little stack space
788 * left. It's plausible that we'd hit this condition but
789 * double-fault even before we get this far, in which case
790 * we're fine: the double-fault handler will deal with it.
792 * We don't want to make it all the way into the oops code
793 * and then double-fault, though, because we're likely to
794 * break the console driver and lose most of the stack dump.
796 asm volatile ("movq %[stack], %%rsp\n\t"
797 "call handle_stack_overflow\n\t"
799 : ASM_CALL_CONSTRAINT
800 : "D" ("kernel stack overflow (page fault)"),
801 "S" (regs), "d" (address),
802 [stack] "rm" (stack));
810 * Valid to do another page fault here, because if this fault
811 * had been triggered by is_prefetch fixup_exception would have
816 * Hall of shame of CPU/BIOS bugs.
818 if (is_prefetch(regs, error_code, address))
821 if (is_errata93(regs, address))
825 * Oops. The kernel tried to access some bad page. We'll have to
826 * terminate things with extreme prejudice:
828 flags = oops_begin();
830 show_fault_oops(regs, error_code, address);
832 if (task_stack_end_corrupted(tsk))
833 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
835 tsk->thread.cr2 = address;
836 tsk->thread.trap_nr = X86_TRAP_PF;
837 tsk->thread.error_code = error_code;
840 if (__die("Oops", regs, error_code))
843 /* Executive summary in case the body of the oops scrolled away */
844 printk(KERN_DEFAULT "CR2: %016lx\n", address);
846 oops_end(flags, regs, sig);
850 * Print out info about fatal segfaults, if the show_unhandled_signals
854 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
855 unsigned long address, struct task_struct *tsk)
857 if (!unhandled_signal(tsk, SIGSEGV))
860 if (!printk_ratelimit())
863 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
864 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
865 tsk->comm, task_pid_nr(tsk), address,
866 (void *)regs->ip, (void *)regs->sp, error_code);
868 print_vma_addr(KERN_CONT " in ", regs->ip);
870 printk(KERN_CONT "\n");
874 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
875 unsigned long address, u32 *pkey, int si_code)
877 struct task_struct *tsk = current;
879 /* User mode accesses just cause a SIGSEGV */
880 if (error_code & X86_PF_USER) {
882 * It's possible to have interrupts off here:
887 * Valid to do another page fault here because this one came
890 if (is_prefetch(regs, error_code, address))
893 if (is_errata100(regs, address))
898 * Instruction fetch faults in the vsyscall page might need
901 if (unlikely((error_code & X86_PF_INSTR) &&
902 ((address & ~0xfff) == VSYSCALL_ADDR))) {
903 if (emulate_vsyscall(regs, address))
909 * To avoid leaking information about the kernel page table
910 * layout, pretend that user-mode accesses to kernel addresses
911 * are always protection faults.
913 if (address >= TASK_SIZE_MAX)
914 error_code |= X86_PF_PROT;
916 if (likely(show_unhandled_signals))
917 show_signal_msg(regs, error_code, address, tsk);
919 tsk->thread.cr2 = address;
920 tsk->thread.error_code = error_code;
921 tsk->thread.trap_nr = X86_TRAP_PF;
923 force_sig_info_fault(SIGSEGV, si_code, address, tsk, pkey, 0);
928 if (is_f00f_bug(regs, address))
931 no_context(regs, error_code, address, SIGSEGV, si_code);
935 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
936 unsigned long address, u32 *pkey)
938 __bad_area_nosemaphore(regs, error_code, address, pkey, SEGV_MAPERR);
942 __bad_area(struct pt_regs *regs, unsigned long error_code,
943 unsigned long address, struct vm_area_struct *vma, int si_code)
945 struct mm_struct *mm = current->mm;
949 pkey = vma_pkey(vma);
952 * Something tried to access memory that isn't in our memory map..
953 * Fix it, but check if it's kernel or user first..
955 up_read(&mm->mmap_sem);
957 __bad_area_nosemaphore(regs, error_code, address,
958 (vma) ? &pkey : NULL, si_code);
962 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
964 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
967 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
968 struct vm_area_struct *vma)
970 /* This code is always called on the current mm */
971 bool foreign = false;
973 if (!boot_cpu_has(X86_FEATURE_OSPKE))
975 if (error_code & X86_PF_PK)
977 /* this checks permission keys on the VMA: */
978 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
979 (error_code & X86_PF_INSTR), foreign))
985 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
986 unsigned long address, struct vm_area_struct *vma)
989 * This OSPKE check is not strictly necessary at runtime.
990 * But, doing it this way allows compiler optimizations
991 * if pkeys are compiled out.
993 if (bad_area_access_from_pkeys(error_code, vma))
994 __bad_area(regs, error_code, address, vma, SEGV_PKUERR);
996 __bad_area(regs, error_code, address, vma, SEGV_ACCERR);
1000 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
1001 u32 *pkey, unsigned int fault)
1003 struct task_struct *tsk = current;
1004 int code = BUS_ADRERR;
1006 /* Kernel mode? Handle exceptions or die: */
1007 if (!(error_code & X86_PF_USER)) {
1008 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1012 /* User-space => ok to do another page fault: */
1013 if (is_prefetch(regs, error_code, address))
1016 tsk->thread.cr2 = address;
1017 tsk->thread.error_code = error_code;
1018 tsk->thread.trap_nr = X86_TRAP_PF;
1020 #ifdef CONFIG_MEMORY_FAILURE
1021 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
1023 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
1024 tsk->comm, tsk->pid, address);
1025 code = BUS_MCEERR_AR;
1028 force_sig_info_fault(SIGBUS, code, address, tsk, pkey, fault);
1031 static noinline void
1032 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
1033 unsigned long address, u32 *pkey, unsigned int fault)
1035 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) {
1036 no_context(regs, error_code, address, 0, 0);
1040 if (fault & VM_FAULT_OOM) {
1041 /* Kernel mode? Handle exceptions or die: */
1042 if (!(error_code & X86_PF_USER)) {
1043 no_context(regs, error_code, address,
1044 SIGSEGV, SEGV_MAPERR);
1049 * We ran out of memory, call the OOM killer, and return the
1050 * userspace (which will retry the fault, or kill us if we got
1053 pagefault_out_of_memory();
1055 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1056 VM_FAULT_HWPOISON_LARGE))
1057 do_sigbus(regs, error_code, address, pkey, fault);
1058 else if (fault & VM_FAULT_SIGSEGV)
1059 bad_area_nosemaphore(regs, error_code, address, pkey);
1065 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
1067 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
1070 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1073 * Note: We do not do lazy flushing on protection key
1074 * changes, so no spurious fault will ever set X86_PF_PK.
1076 if ((error_code & X86_PF_PK))
1083 * Handle a spurious fault caused by a stale TLB entry.
1085 * This allows us to lazily refresh the TLB when increasing the
1086 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1087 * eagerly is very expensive since that implies doing a full
1088 * cross-processor TLB flush, even if no stale TLB entries exist
1089 * on other processors.
1091 * Spurious faults may only occur if the TLB contains an entry with
1092 * fewer permission than the page table entry. Non-present (P = 0)
1093 * and reserved bit (R = 1) faults are never spurious.
1095 * There are no security implications to leaving a stale TLB when
1096 * increasing the permissions on a page.
1098 * Returns non-zero if a spurious fault was handled, zero otherwise.
1100 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1101 * (Optional Invalidation).
1104 spurious_fault(unsigned long error_code, unsigned long address)
1114 * Only writes to RO or instruction fetches from NX may cause
1117 * These could be from user or supervisor accesses but the TLB
1118 * is only lazily flushed after a kernel mapping protection
1119 * change, so user accesses are not expected to cause spurious
1122 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1123 error_code != (X86_PF_INSTR | X86_PF_PROT))
1126 pgd = init_mm.pgd + pgd_index(address);
1127 if (!pgd_present(*pgd))
1130 p4d = p4d_offset(pgd, address);
1131 if (!p4d_present(*p4d))
1134 if (p4d_large(*p4d))
1135 return spurious_fault_check(error_code, (pte_t *) p4d);
1137 pud = pud_offset(p4d, address);
1138 if (!pud_present(*pud))
1141 if (pud_large(*pud))
1142 return spurious_fault_check(error_code, (pte_t *) pud);
1144 pmd = pmd_offset(pud, address);
1145 if (!pmd_present(*pmd))
1148 if (pmd_large(*pmd))
1149 return spurious_fault_check(error_code, (pte_t *) pmd);
1151 pte = pte_offset_kernel(pmd, address);
1152 if (!pte_present(*pte))
1155 ret = spurious_fault_check(error_code, pte);
1160 * Make sure we have permissions in PMD.
1161 * If not, then there's a bug in the page tables:
1163 ret = spurious_fault_check(error_code, (pte_t *) pmd);
1164 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1168 NOKPROBE_SYMBOL(spurious_fault);
1170 int show_unhandled_signals = 1;
1173 access_error(unsigned long error_code, struct vm_area_struct *vma)
1175 /* This is only called for the current mm, so: */
1176 bool foreign = false;
1179 * Read or write was blocked by protection keys. This is
1180 * always an unconditional error and can never result in
1181 * a follow-up action to resolve the fault, like a COW.
1183 if (error_code & X86_PF_PK)
1187 * Make sure to check the VMA so that we do not perform
1188 * faults just to hit a X86_PF_PK as soon as we fill in a
1191 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1192 (error_code & X86_PF_INSTR), foreign))
1195 if (error_code & X86_PF_WRITE) {
1196 /* write, present and write, not present: */
1197 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1202 /* read, present: */
1203 if (unlikely(error_code & X86_PF_PROT))
1206 /* read, not present: */
1207 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1213 static int fault_in_kernel_space(unsigned long address)
1215 return address >= TASK_SIZE_MAX;
1218 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1220 if (!IS_ENABLED(CONFIG_X86_SMAP))
1223 if (!static_cpu_has(X86_FEATURE_SMAP))
1226 if (error_code & X86_PF_USER)
1229 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1236 * This routine handles page faults. It determines the address,
1237 * and the problem, and then passes it off to one of the appropriate
1240 static noinline void
1241 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1242 unsigned long address)
1244 struct vm_area_struct *vma;
1245 struct task_struct *tsk;
1246 struct mm_struct *mm;
1247 int fault, major = 0;
1248 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1255 * Detect and handle instructions that would cause a page fault for
1256 * both a tracked kernel page and a userspace page.
1258 prefetchw(&mm->mmap_sem);
1260 if (unlikely(kmmio_fault(regs, address)))
1264 * We fault-in kernel-space virtual memory on-demand. The
1265 * 'reference' page table is init_mm.pgd.
1267 * NOTE! We MUST NOT take any locks for this case. We may
1268 * be in an interrupt or a critical region, and should
1269 * only copy the information from the master page table,
1272 * This verifies that the fault happens in kernel space
1273 * (error_code & 4) == 0, and that the fault was not a
1274 * protection error (error_code & 9) == 0.
1276 if (unlikely(fault_in_kernel_space(address))) {
1277 if (!(error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1278 if (vmalloc_fault(address) >= 0)
1282 /* Can handle a stale RO->RW TLB: */
1283 if (spurious_fault(error_code, address))
1286 /* kprobes don't want to hook the spurious faults: */
1287 if (kprobes_fault(regs))
1290 * Don't take the mm semaphore here. If we fixup a prefetch
1291 * fault we could otherwise deadlock:
1293 bad_area_nosemaphore(regs, error_code, address, NULL);
1298 /* kprobes don't want to hook the spurious faults: */
1299 if (unlikely(kprobes_fault(regs)))
1302 if (unlikely(error_code & X86_PF_RSVD))
1303 pgtable_bad(regs, error_code, address);
1305 if (unlikely(smap_violation(error_code, regs))) {
1306 bad_area_nosemaphore(regs, error_code, address, NULL);
1311 * If we're in an interrupt, have no user context or are running
1312 * in a region with pagefaults disabled then we must not take the fault
1314 if (unlikely(faulthandler_disabled() || !mm)) {
1315 bad_area_nosemaphore(regs, error_code, address, NULL);
1320 * It's safe to allow irq's after cr2 has been saved and the
1321 * vmalloc fault has been handled.
1323 * User-mode registers count as a user access even for any
1324 * potential system fault or CPU buglet:
1326 if (user_mode(regs)) {
1328 error_code |= X86_PF_USER;
1329 flags |= FAULT_FLAG_USER;
1331 if (regs->flags & X86_EFLAGS_IF)
1335 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1337 if (error_code & X86_PF_WRITE)
1338 flags |= FAULT_FLAG_WRITE;
1339 if (error_code & X86_PF_INSTR)
1340 flags |= FAULT_FLAG_INSTRUCTION;
1343 * When running in the kernel we expect faults to occur only to
1344 * addresses in user space. All other faults represent errors in
1345 * the kernel and should generate an OOPS. Unfortunately, in the
1346 * case of an erroneous fault occurring in a code path which already
1347 * holds mmap_sem we will deadlock attempting to validate the fault
1348 * against the address space. Luckily the kernel only validly
1349 * references user space from well defined areas of code, which are
1350 * listed in the exceptions table.
1352 * As the vast majority of faults will be valid we will only perform
1353 * the source reference check when there is a possibility of a
1354 * deadlock. Attempt to lock the address space, if we cannot we then
1355 * validate the source. If this is invalid we can skip the address
1356 * space check, thus avoiding the deadlock:
1358 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1359 if (!(error_code & X86_PF_USER) &&
1360 !search_exception_tables(regs->ip)) {
1361 bad_area_nosemaphore(regs, error_code, address, NULL);
1365 down_read(&mm->mmap_sem);
1368 * The above down_read_trylock() might have succeeded in
1369 * which case we'll have missed the might_sleep() from
1375 vma = find_vma(mm, address);
1376 if (unlikely(!vma)) {
1377 bad_area(regs, error_code, address);
1380 if (likely(vma->vm_start <= address))
1382 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1383 bad_area(regs, error_code, address);
1386 if (error_code & X86_PF_USER) {
1388 * Accessing the stack below %sp is always a bug.
1389 * The large cushion allows instructions like enter
1390 * and pusha to work. ("enter $65535, $31" pushes
1391 * 32 pointers and then decrements %sp by 65535.)
1393 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1394 bad_area(regs, error_code, address);
1398 if (unlikely(expand_stack(vma, address))) {
1399 bad_area(regs, error_code, address);
1404 * Ok, we have a good vm_area for this memory access, so
1405 * we can handle it..
1408 if (unlikely(access_error(error_code, vma))) {
1409 bad_area_access_error(regs, error_code, address, vma);
1414 * If for any reason at all we couldn't handle the fault,
1415 * make sure we exit gracefully rather than endlessly redo
1416 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1417 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1419 * Note that handle_userfault() may also release and reacquire mmap_sem
1420 * (and not return with VM_FAULT_RETRY), when returning to userland to
1421 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1422 * (potentially after handling any pending signal during the return to
1423 * userland). The return to userland is identified whenever
1424 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1425 * Thus we have to be careful about not touching vma after handling the
1426 * fault, so we read the pkey beforehand.
1428 pkey = vma_pkey(vma);
1429 fault = handle_mm_fault(vma, address, flags);
1430 major |= fault & VM_FAULT_MAJOR;
1433 * If we need to retry the mmap_sem has already been released,
1434 * and if there is a fatal signal pending there is no guarantee
1435 * that we made any progress. Handle this case first.
1437 if (unlikely(fault & VM_FAULT_RETRY)) {
1438 /* Retry at most once */
1439 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1440 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1441 flags |= FAULT_FLAG_TRIED;
1442 if (!fatal_signal_pending(tsk))
1446 /* User mode? Just return to handle the fatal exception */
1447 if (flags & FAULT_FLAG_USER)
1450 /* Not returning to user mode? Handle exceptions or die: */
1451 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1455 up_read(&mm->mmap_sem);
1456 if (unlikely(fault & VM_FAULT_ERROR)) {
1457 mm_fault_error(regs, error_code, address, &pkey, fault);
1462 * Major/minor page fault accounting. If any of the events
1463 * returned VM_FAULT_MAJOR, we account it as a major fault.
1467 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1470 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1473 check_v8086_mode(regs, address, tsk);
1475 NOKPROBE_SYMBOL(__do_page_fault);
1477 static nokprobe_inline void
1478 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1479 unsigned long error_code)
1481 if (user_mode(regs))
1482 trace_page_fault_user(address, regs, error_code);
1484 trace_page_fault_kernel(address, regs, error_code);
1488 * We must have this function blacklisted from kprobes, tagged with notrace
1489 * and call read_cr2() before calling anything else. To avoid calling any
1490 * kind of tracing machinery before we've observed the CR2 value.
1492 * exception_{enter,exit}() contains all sorts of tracepoints.
1494 dotraplinkage void notrace
1495 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1497 unsigned long address = read_cr2(); /* Get the faulting address */
1498 enum ctx_state prev_state;
1500 prev_state = exception_enter();
1501 if (trace_pagefault_enabled())
1502 trace_page_fault_entries(address, regs, error_code);
1504 __do_page_fault(regs, error_code, address);
1505 exception_exit(prev_state);
1507 NOKPROBE_SYMBOL(do_page_fault);