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/kprobes.h> /* NOKPROBE_SYMBOL, ... */
13 #include <linux/mmiotrace.h> /* kmmio_handler, ... */
14 #include <linux/perf_event.h> /* perf_sw_event */
15 #include <linux/hugetlb.h> /* hstate_index_to_shift */
16 #include <linux/prefetch.h> /* prefetchw */
17 #include <linux/context_tracking.h> /* exception_enter(), ... */
18 #include <linux/uaccess.h> /* faulthandler_disabled() */
19 #include <linux/efi.h> /* efi_recover_from_page_fault()*/
20 #include <linux/mm_types.h>
22 #include <asm/cpufeature.h> /* boot_cpu_has, ... */
23 #include <asm/traps.h> /* dotraplinkage, ... */
24 #include <asm/fixmap.h> /* VSYSCALL_ADDR */
25 #include <asm/vsyscall.h> /* emulate_vsyscall */
26 #include <asm/vm86.h> /* struct vm86 */
27 #include <asm/mmu_context.h> /* vma_pkey() */
28 #include <asm/efi.h> /* efi_recover_from_page_fault()*/
29 #include <asm/desc.h> /* store_idt(), ... */
30 #include <asm/cpu_entry_area.h> /* exception stack */
31 #include <asm/pgtable_areas.h> /* VMALLOC_START, ... */
32 #include <asm/kvm_para.h> /* kvm_handle_async_pf */
34 #define CREATE_TRACE_POINTS
35 #include <asm/trace/exceptions.h>
38 * Returns 0 if mmiotrace is disabled, or if the fault is not
39 * handled by mmiotrace:
41 static nokprobe_inline int
42 kmmio_fault(struct pt_regs *regs, unsigned long addr)
44 if (unlikely(is_kmmio_active()))
45 if (kmmio_handler(regs, addr) == 1)
55 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
56 * Check that here and ignore it.
60 * Sometimes the CPU reports invalid exceptions on prefetch.
61 * Check that here and ignore it.
63 * Opcode checker based on code by Richard Brunner.
66 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
67 unsigned char opcode, int *prefetch)
69 unsigned char instr_hi = opcode & 0xf0;
70 unsigned char instr_lo = opcode & 0x0f;
76 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
77 * In X86_64 long mode, the CPU will signal invalid
78 * opcode if some of these prefixes are present so
79 * X86_64 will never get here anyway
81 return ((instr_lo & 7) == 0x6);
85 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
86 * Need to figure out under what instruction mode the
87 * instruction was issued. Could check the LDT for lm,
88 * but for now it's good enough to assume that long
89 * mode only uses well known segments or kernel.
91 return (!user_mode(regs) || user_64bit_mode(regs));
94 /* 0x64 thru 0x67 are valid prefixes in all modes. */
95 return (instr_lo & 0xC) == 0x4;
97 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
98 return !instr_lo || (instr_lo>>1) == 1;
100 /* Prefetch instruction is 0x0F0D or 0x0F18 */
101 if (get_kernel_nofault(opcode, instr))
104 *prefetch = (instr_lo == 0xF) &&
105 (opcode == 0x0D || opcode == 0x18);
113 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
115 unsigned char *max_instr;
116 unsigned char *instr;
120 * If it was a exec (instruction fetch) fault on NX page, then
121 * do not ignore the fault:
123 if (error_code & X86_PF_INSTR)
126 instr = (void *)convert_ip_to_linear(current, regs);
127 max_instr = instr + 15;
129 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
132 while (instr < max_instr) {
133 unsigned char opcode;
135 if (get_kernel_nofault(opcode, instr))
140 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
146 DEFINE_SPINLOCK(pgd_lock);
150 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
152 unsigned index = pgd_index(address);
159 pgd_k = init_mm.pgd + index;
161 if (!pgd_present(*pgd_k))
165 * set_pgd(pgd, *pgd_k); here would be useless on PAE
166 * and redundant with the set_pmd() on non-PAE. As would
169 p4d = p4d_offset(pgd, address);
170 p4d_k = p4d_offset(pgd_k, address);
171 if (!p4d_present(*p4d_k))
174 pud = pud_offset(p4d, address);
175 pud_k = pud_offset(p4d_k, address);
176 if (!pud_present(*pud_k))
179 pmd = pmd_offset(pud, address);
180 pmd_k = pmd_offset(pud_k, address);
182 if (pmd_present(*pmd) != pmd_present(*pmd_k))
183 set_pmd(pmd, *pmd_k);
185 if (!pmd_present(*pmd_k))
188 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
194 * Handle a fault on the vmalloc or module mapping area
196 * This is needed because there is a race condition between the time
197 * when the vmalloc mapping code updates the PMD to the point in time
198 * where it synchronizes this update with the other page-tables in the
201 * In this race window another thread/CPU can map an area on the same
202 * PMD, finds it already present and does not synchronize it with the
203 * rest of the system yet. As a result v[mz]alloc might return areas
204 * which are not mapped in every page-table in the system, causing an
205 * unhandled page-fault when they are accessed.
207 static noinline int vmalloc_fault(unsigned long address)
209 unsigned long pgd_paddr;
213 /* Make sure we are in vmalloc area: */
214 if (!(address >= VMALLOC_START && address < VMALLOC_END))
218 * Synchronize this task's top level page-table
219 * with the 'reference' page table.
221 * Do _not_ use "current" here. We might be inside
222 * an interrupt in the middle of a task switch..
224 pgd_paddr = read_cr3_pa();
225 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
229 if (pmd_large(*pmd_k))
232 pte_k = pte_offset_kernel(pmd_k, address);
233 if (!pte_present(*pte_k))
238 NOKPROBE_SYMBOL(vmalloc_fault);
240 void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
244 for (addr = start & PMD_MASK;
245 addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
249 spin_lock(&pgd_lock);
250 list_for_each_entry(page, &pgd_list, lru) {
251 spinlock_t *pgt_lock;
253 /* the pgt_lock only for Xen */
254 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
257 vmalloc_sync_one(page_address(page), addr);
258 spin_unlock(pgt_lock);
260 spin_unlock(&pgd_lock);
265 * Did it hit the DOS screen memory VA from vm86 mode?
268 check_v8086_mode(struct pt_regs *regs, unsigned long address,
269 struct task_struct *tsk)
274 if (!v8086_mode(regs) || !tsk->thread.vm86)
277 bit = (address - 0xA0000) >> PAGE_SHIFT;
279 tsk->thread.vm86->screen_bitmap |= 1 << bit;
283 static bool low_pfn(unsigned long pfn)
285 return pfn < max_low_pfn;
288 static void dump_pagetable(unsigned long address)
290 pgd_t *base = __va(read_cr3_pa());
291 pgd_t *pgd = &base[pgd_index(address)];
297 #ifdef CONFIG_X86_PAE
298 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
299 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
301 #define pr_pde pr_cont
303 #define pr_pde pr_info
305 p4d = p4d_offset(pgd, address);
306 pud = pud_offset(p4d, address);
307 pmd = pmd_offset(pud, address);
308 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
312 * We must not directly access the pte in the highpte
313 * case if the page table is located in highmem.
314 * And let's rather not kmap-atomic the pte, just in case
315 * it's allocated already:
317 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
320 pte = pte_offset_kernel(pmd, address);
321 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
326 #else /* CONFIG_X86_64: */
328 #ifdef CONFIG_CPU_SUP_AMD
329 static const char errata93_warning[] =
331 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
332 "******* Working around it, but it may cause SEGVs or burn power.\n"
333 "******* Please consider a BIOS update.\n"
334 "******* Disabling USB legacy in the BIOS may also help.\n";
338 * No vm86 mode in 64-bit mode:
341 check_v8086_mode(struct pt_regs *regs, unsigned long address,
342 struct task_struct *tsk)
346 static int bad_address(void *p)
350 return get_kernel_nofault(dummy, (unsigned long *)p);
353 static void dump_pagetable(unsigned long address)
355 pgd_t *base = __va(read_cr3_pa());
356 pgd_t *pgd = base + pgd_index(address);
362 if (bad_address(pgd))
365 pr_info("PGD %lx ", pgd_val(*pgd));
367 if (!pgd_present(*pgd))
370 p4d = p4d_offset(pgd, address);
371 if (bad_address(p4d))
374 pr_cont("P4D %lx ", p4d_val(*p4d));
375 if (!p4d_present(*p4d) || p4d_large(*p4d))
378 pud = pud_offset(p4d, address);
379 if (bad_address(pud))
382 pr_cont("PUD %lx ", pud_val(*pud));
383 if (!pud_present(*pud) || pud_large(*pud))
386 pmd = pmd_offset(pud, address);
387 if (bad_address(pmd))
390 pr_cont("PMD %lx ", pmd_val(*pmd));
391 if (!pmd_present(*pmd) || pmd_large(*pmd))
394 pte = pte_offset_kernel(pmd, address);
395 if (bad_address(pte))
398 pr_cont("PTE %lx", pte_val(*pte));
406 #endif /* CONFIG_X86_64 */
409 * Workaround for K8 erratum #93 & buggy BIOS.
411 * BIOS SMM functions are required to use a specific workaround
412 * to avoid corruption of the 64bit RIP register on C stepping K8.
414 * A lot of BIOS that didn't get tested properly miss this.
416 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
417 * Try to work around it here.
419 * Note we only handle faults in kernel here.
420 * Does nothing on 32-bit.
422 static int is_errata93(struct pt_regs *regs, unsigned long address)
424 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
425 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
426 || 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 address)
466 #ifdef CONFIG_X86_F00F_BUG
467 if (boot_cpu_has_bug(X86_BUG_F00F) && 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 no_context(struct pt_regs *regs, unsigned long error_code,
633 unsigned long address, int signal, int si_code)
635 struct task_struct *tsk = current;
639 if (user_mode(regs)) {
641 * This is an implicit supervisor-mode access from user
642 * mode. Bypass all the kernel-mode recovery code and just
648 /* Are we prepared to handle this kernel fault? */
649 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
651 * Any interrupt that takes a fault gets the fixup. This makes
652 * the below recursive fault logic only apply to a faults from
659 * Per the above we're !in_interrupt(), aka. task context.
661 * In this case we need to make sure we're not recursively
662 * faulting through the emulate_vsyscall() logic.
664 if (current->thread.sig_on_uaccess_err && signal) {
665 sanitize_error_code(address, &error_code);
667 set_signal_archinfo(address, error_code);
669 /* XXX: hwpoison faults will set the wrong code. */
670 force_sig_fault(signal, si_code, (void __user *)address);
674 * Barring that, we can do the fixup and be happy.
679 #ifdef CONFIG_VMAP_STACK
681 * Stack overflow? During boot, we can fault near the initial
682 * stack in the direct map, but that's not an overflow -- check
683 * that we're in vmalloc space to avoid this.
685 if (is_vmalloc_addr((void *)address) &&
686 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
687 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
688 unsigned long stack = __this_cpu_ist_top_va(DF) - sizeof(void *);
690 * We're likely to be running with very little stack space
691 * left. It's plausible that we'd hit this condition but
692 * double-fault even before we get this far, in which case
693 * we're fine: the double-fault handler will deal with it.
695 * We don't want to make it all the way into the oops code
696 * and then double-fault, though, because we're likely to
697 * break the console driver and lose most of the stack dump.
699 asm volatile ("movq %[stack], %%rsp\n\t"
700 "call handle_stack_overflow\n\t"
702 : ASM_CALL_CONSTRAINT
703 : "D" ("kernel stack overflow (page fault)"),
704 "S" (regs), "d" (address),
705 [stack] "rm" (stack));
713 * Valid to do another page fault here, because if this fault
714 * had been triggered by is_prefetch fixup_exception would have
719 * Hall of shame of CPU/BIOS bugs.
721 if (is_prefetch(regs, error_code, address))
724 if (is_errata93(regs, address))
728 * Buggy firmware could access regions which might page fault, try to
729 * recover from such faults.
731 if (IS_ENABLED(CONFIG_EFI))
732 efi_recover_from_page_fault(address);
736 * Oops. The kernel tried to access some bad page. We'll have to
737 * terminate things with extreme prejudice:
739 flags = oops_begin();
741 show_fault_oops(regs, error_code, address);
743 if (task_stack_end_corrupted(tsk))
744 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
747 if (__die("Oops", regs, error_code))
750 /* Executive summary in case the body of the oops scrolled away */
751 printk(KERN_DEFAULT "CR2: %016lx\n", address);
753 oops_end(flags, regs, sig);
757 * Print out info about fatal segfaults, if the show_unhandled_signals
761 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
762 unsigned long address, struct task_struct *tsk)
764 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
766 if (!unhandled_signal(tsk, SIGSEGV))
769 if (!printk_ratelimit())
772 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
773 loglvl, tsk->comm, task_pid_nr(tsk), address,
774 (void *)regs->ip, (void *)regs->sp, error_code);
776 print_vma_addr(KERN_CONT " in ", regs->ip);
778 printk(KERN_CONT "\n");
780 show_opcodes(regs, loglvl);
784 * The (legacy) vsyscall page is the long page in the kernel portion
785 * of the address space that has user-accessible permissions.
787 static bool is_vsyscall_vaddr(unsigned long vaddr)
789 return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR);
793 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
794 unsigned long address, u32 pkey, int si_code)
796 struct task_struct *tsk = current;
798 /* User mode accesses just cause a SIGSEGV */
799 if (user_mode(regs) && (error_code & X86_PF_USER)) {
801 * It's possible to have interrupts off here:
806 * Valid to do another page fault here because this one came
809 if (is_prefetch(regs, error_code, address))
812 if (is_errata100(regs, address))
815 sanitize_error_code(address, &error_code);
817 if (likely(show_unhandled_signals))
818 show_signal_msg(regs, error_code, address, tsk);
820 set_signal_archinfo(address, error_code);
822 if (si_code == SEGV_PKUERR)
823 force_sig_pkuerr((void __user *)address, pkey);
825 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
832 if (is_f00f_bug(regs, address))
835 no_context(regs, error_code, address, SIGSEGV, si_code);
839 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
840 unsigned long address)
842 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
846 __bad_area(struct pt_regs *regs, unsigned long error_code,
847 unsigned long address, u32 pkey, int si_code)
849 struct mm_struct *mm = current->mm;
851 * Something tried to access memory that isn't in our memory map..
852 * Fix it, but check if it's kernel or user first..
854 mmap_read_unlock(mm);
856 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
860 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
862 __bad_area(regs, error_code, address, 0, SEGV_MAPERR);
865 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
866 struct vm_area_struct *vma)
868 /* This code is always called on the current mm */
869 bool foreign = false;
871 if (!boot_cpu_has(X86_FEATURE_OSPKE))
873 if (error_code & X86_PF_PK)
875 /* this checks permission keys on the VMA: */
876 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
877 (error_code & X86_PF_INSTR), foreign))
883 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
884 unsigned long address, struct vm_area_struct *vma)
887 * This OSPKE check is not strictly necessary at runtime.
888 * But, doing it this way allows compiler optimizations
889 * if pkeys are compiled out.
891 if (bad_area_access_from_pkeys(error_code, vma)) {
893 * A protection key fault means that the PKRU value did not allow
894 * access to some PTE. Userspace can figure out what PKRU was
895 * from the XSAVE state. This function captures the pkey from
896 * the vma and passes it to userspace so userspace can discover
897 * which protection key was set on the PTE.
899 * If we get here, we know that the hardware signaled a X86_PF_PK
900 * fault and that there was a VMA once we got in the fault
901 * handler. It does *not* guarantee that the VMA we find here
902 * was the one that we faulted on.
904 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
905 * 2. T1 : set PKRU to deny access to pkey=4, touches page
907 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
908 * 5. T1 : enters fault handler, takes mmap_lock, etc...
909 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
910 * faulted on a pte with its pkey=4.
912 u32 pkey = vma_pkey(vma);
914 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
916 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
921 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
924 /* Kernel mode? Handle exceptions or die: */
925 if (!(error_code & X86_PF_USER)) {
926 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
930 /* User-space => ok to do another page fault: */
931 if (is_prefetch(regs, error_code, address))
934 sanitize_error_code(address, &error_code);
936 set_signal_archinfo(address, error_code);
938 #ifdef CONFIG_MEMORY_FAILURE
939 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
940 struct task_struct *tsk = current;
944 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
945 tsk->comm, tsk->pid, address);
946 if (fault & VM_FAULT_HWPOISON_LARGE)
947 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
948 if (fault & VM_FAULT_HWPOISON)
950 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
954 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
958 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
959 unsigned long address, vm_fault_t fault)
961 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) {
962 no_context(regs, error_code, address, 0, 0);
966 if (fault & VM_FAULT_OOM) {
967 /* Kernel mode? Handle exceptions or die: */
968 if (!(error_code & X86_PF_USER)) {
969 no_context(regs, error_code, address,
970 SIGSEGV, SEGV_MAPERR);
975 * We ran out of memory, call the OOM killer, and return the
976 * userspace (which will retry the fault, or kill us if we got
979 pagefault_out_of_memory();
981 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
982 VM_FAULT_HWPOISON_LARGE))
983 do_sigbus(regs, error_code, address, fault);
984 else if (fault & VM_FAULT_SIGSEGV)
985 bad_area_nosemaphore(regs, error_code, address);
991 static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
993 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
996 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1003 * Handle a spurious fault caused by a stale TLB entry.
1005 * This allows us to lazily refresh the TLB when increasing the
1006 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1007 * eagerly is very expensive since that implies doing a full
1008 * cross-processor TLB flush, even if no stale TLB entries exist
1009 * on other processors.
1011 * Spurious faults may only occur if the TLB contains an entry with
1012 * fewer permission than the page table entry. Non-present (P = 0)
1013 * and reserved bit (R = 1) faults are never spurious.
1015 * There are no security implications to leaving a stale TLB when
1016 * increasing the permissions on a page.
1018 * Returns non-zero if a spurious fault was handled, zero otherwise.
1020 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1021 * (Optional Invalidation).
1024 spurious_kernel_fault(unsigned long error_code, unsigned long address)
1034 * Only writes to RO or instruction fetches from NX may cause
1037 * These could be from user or supervisor accesses but the TLB
1038 * is only lazily flushed after a kernel mapping protection
1039 * change, so user accesses are not expected to cause spurious
1042 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1043 error_code != (X86_PF_INSTR | X86_PF_PROT))
1046 pgd = init_mm.pgd + pgd_index(address);
1047 if (!pgd_present(*pgd))
1050 p4d = p4d_offset(pgd, address);
1051 if (!p4d_present(*p4d))
1054 if (p4d_large(*p4d))
1055 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1057 pud = pud_offset(p4d, address);
1058 if (!pud_present(*pud))
1061 if (pud_large(*pud))
1062 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1064 pmd = pmd_offset(pud, address);
1065 if (!pmd_present(*pmd))
1068 if (pmd_large(*pmd))
1069 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1071 pte = pte_offset_kernel(pmd, address);
1072 if (!pte_present(*pte))
1075 ret = spurious_kernel_fault_check(error_code, pte);
1080 * Make sure we have permissions in PMD.
1081 * If not, then there's a bug in the page tables:
1083 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1084 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1088 NOKPROBE_SYMBOL(spurious_kernel_fault);
1090 int show_unhandled_signals = 1;
1093 access_error(unsigned long error_code, struct vm_area_struct *vma)
1095 /* This is only called for the current mm, so: */
1096 bool foreign = false;
1099 * Read or write was blocked by protection keys. This is
1100 * always an unconditional error and can never result in
1101 * a follow-up action to resolve the fault, like a COW.
1103 if (error_code & X86_PF_PK)
1107 * SGX hardware blocked the access. This usually happens
1108 * when the enclave memory contents have been destroyed, like
1109 * after a suspend/resume cycle. In any case, the kernel can't
1110 * fix the cause of the fault. Handle the fault as an access
1111 * error even in cases where no actual access violation
1112 * occurred. This allows userspace to rebuild the enclave in
1113 * response to the signal.
1115 if (unlikely(error_code & X86_PF_SGX))
1119 * Make sure to check the VMA so that we do not perform
1120 * faults just to hit a X86_PF_PK as soon as we fill in a
1123 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1124 (error_code & X86_PF_INSTR), foreign))
1127 if (error_code & X86_PF_WRITE) {
1128 /* write, present and write, not present: */
1129 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1134 /* read, present: */
1135 if (unlikely(error_code & X86_PF_PROT))
1138 /* read, not present: */
1139 if (unlikely(!vma_is_accessible(vma)))
1145 bool fault_in_kernel_space(unsigned long address)
1148 * On 64-bit systems, the vsyscall page is at an address above
1149 * TASK_SIZE_MAX, but is not considered part of the kernel
1152 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1155 return address >= TASK_SIZE_MAX;
1159 * Called for all faults where 'address' is part of the kernel address
1160 * space. Might get called for faults that originate from *code* that
1161 * ran in userspace or the kernel.
1164 do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1165 unsigned long address)
1168 * Protection keys exceptions only happen on user pages. We
1169 * have no user pages in the kernel portion of the address
1170 * space, so do not expect them here.
1172 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1174 #ifdef CONFIG_X86_32
1176 * We can fault-in kernel-space virtual memory on-demand. The
1177 * 'reference' page table is init_mm.pgd.
1179 * NOTE! We MUST NOT take any locks for this case. We may
1180 * be in an interrupt or a critical region, and should
1181 * only copy the information from the master page table,
1184 * Before doing this on-demand faulting, ensure that the
1185 * fault is not any of the following:
1186 * 1. A fault on a PTE with a reserved bit set.
1187 * 2. A fault caused by a user-mode access. (Do not demand-
1188 * fault kernel memory due to user-mode accesses).
1189 * 3. A fault caused by a page-level protection violation.
1190 * (A demand fault would be on a non-present page which
1191 * would have X86_PF_PROT==0).
1193 * This is only needed to close a race condition on x86-32 in
1194 * the vmalloc mapping/unmapping code. See the comment above
1195 * vmalloc_fault() for details. On x86-64 the race does not
1196 * exist as the vmalloc mappings don't need to be synchronized
1199 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1200 if (vmalloc_fault(address) >= 0)
1205 /* Was the fault spurious, caused by lazy TLB invalidation? */
1206 if (spurious_kernel_fault(hw_error_code, address))
1209 /* kprobes don't want to hook the spurious faults: */
1210 if (kprobe_page_fault(regs, X86_TRAP_PF))
1214 * Note, despite being a "bad area", there are quite a few
1215 * acceptable reasons to get here, such as erratum fixups
1216 * and handling kernel code that can fault, like get_user().
1218 * Don't take the mm semaphore here. If we fixup a prefetch
1219 * fault we could otherwise deadlock:
1221 bad_area_nosemaphore(regs, hw_error_code, address);
1223 NOKPROBE_SYMBOL(do_kern_addr_fault);
1225 /* Handle faults in the user portion of the address space */
1227 void do_user_addr_fault(struct pt_regs *regs,
1228 unsigned long hw_error_code,
1229 unsigned long address)
1231 struct vm_area_struct *vma;
1232 struct task_struct *tsk;
1233 struct mm_struct *mm;
1235 unsigned int flags = FAULT_FLAG_DEFAULT;
1240 /* kprobes don't want to hook the spurious faults: */
1241 if (unlikely(kprobe_page_fault(regs, X86_TRAP_PF)))
1245 * Reserved bits are never expected to be set on
1246 * entries in the user portion of the page tables.
1248 if (unlikely(hw_error_code & X86_PF_RSVD))
1249 pgtable_bad(regs, hw_error_code, address);
1252 * If SMAP is on, check for invalid kernel (supervisor) access to user
1253 * pages in the user address space. The odd case here is WRUSS,
1254 * which, according to the preliminary documentation, does not respect
1255 * SMAP and will have the USER bit set so, in all cases, SMAP
1256 * enforcement appears to be consistent with the USER bit.
1258 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1259 !(hw_error_code & X86_PF_USER) &&
1260 !(regs->flags & X86_EFLAGS_AC)))
1262 bad_area_nosemaphore(regs, hw_error_code, address);
1267 * If we're in an interrupt, have no user context or are running
1268 * in a region with pagefaults disabled then we must not take the fault
1270 if (unlikely(faulthandler_disabled() || !mm)) {
1271 bad_area_nosemaphore(regs, hw_error_code, address);
1276 * It's safe to allow irq's after cr2 has been saved and the
1277 * vmalloc fault has been handled.
1279 * User-mode registers count as a user access even for any
1280 * potential system fault or CPU buglet:
1282 if (user_mode(regs)) {
1284 flags |= FAULT_FLAG_USER;
1286 if (regs->flags & X86_EFLAGS_IF)
1290 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1292 if (hw_error_code & X86_PF_WRITE)
1293 flags |= FAULT_FLAG_WRITE;
1294 if (hw_error_code & X86_PF_INSTR)
1295 flags |= FAULT_FLAG_INSTRUCTION;
1297 #ifdef CONFIG_X86_64
1299 * Faults in the vsyscall page might need emulation. The
1300 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1301 * considered to be part of the user address space.
1303 * The vsyscall page does not have a "real" VMA, so do this
1304 * emulation before we go searching for VMAs.
1306 * PKRU never rejects instruction fetches, so we don't need
1307 * to consider the PF_PK bit.
1309 if (is_vsyscall_vaddr(address)) {
1310 if (emulate_vsyscall(hw_error_code, regs, address))
1316 * Kernel-mode access to the user address space should only occur
1317 * on well-defined single instructions listed in the exception
1318 * tables. But, an erroneous kernel fault occurring outside one of
1319 * those areas which also holds mmap_lock might deadlock attempting
1320 * to validate the fault against the address space.
1322 * Only do the expensive exception table search when we might be at
1323 * risk of a deadlock. This happens if we
1324 * 1. Failed to acquire mmap_lock, and
1325 * 2. The access did not originate in userspace.
1327 if (unlikely(!mmap_read_trylock(mm))) {
1328 if (!user_mode(regs) && !search_exception_tables(regs->ip)) {
1330 * Fault from code in kernel from
1331 * which we do not expect faults.
1333 bad_area_nosemaphore(regs, hw_error_code, address);
1340 * The above down_read_trylock() might have succeeded in
1341 * which case we'll have missed the might_sleep() from
1347 vma = find_vma(mm, address);
1348 if (unlikely(!vma)) {
1349 bad_area(regs, hw_error_code, address);
1352 if (likely(vma->vm_start <= address))
1354 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1355 bad_area(regs, hw_error_code, address);
1358 if (unlikely(expand_stack(vma, address))) {
1359 bad_area(regs, hw_error_code, address);
1364 * Ok, we have a good vm_area for this memory access, so
1365 * we can handle it..
1368 if (unlikely(access_error(hw_error_code, vma))) {
1369 bad_area_access_error(regs, hw_error_code, address, vma);
1374 * If for any reason at all we couldn't handle the fault,
1375 * make sure we exit gracefully rather than endlessly redo
1376 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1377 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1379 * Note that handle_userfault() may also release and reacquire mmap_lock
1380 * (and not return with VM_FAULT_RETRY), when returning to userland to
1381 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1382 * (potentially after handling any pending signal during the return to
1383 * userland). The return to userland is identified whenever
1384 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1386 fault = handle_mm_fault(vma, address, flags, regs);
1388 /* Quick path to respond to signals */
1389 if (fault_signal_pending(fault, regs)) {
1390 if (!user_mode(regs))
1391 no_context(regs, hw_error_code, address, SIGBUS,
1397 * If we need to retry the mmap_lock has already been released,
1398 * and if there is a fatal signal pending there is no guarantee
1399 * that we made any progress. Handle this case first.
1401 if (unlikely((fault & VM_FAULT_RETRY) &&
1402 (flags & FAULT_FLAG_ALLOW_RETRY))) {
1403 flags |= FAULT_FLAG_TRIED;
1407 mmap_read_unlock(mm);
1408 if (unlikely(fault & VM_FAULT_ERROR)) {
1409 mm_fault_error(regs, hw_error_code, address, fault);
1413 check_v8086_mode(regs, address, tsk);
1415 NOKPROBE_SYMBOL(do_user_addr_fault);
1417 static __always_inline void
1418 trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1419 unsigned long address)
1421 if (!trace_pagefault_enabled())
1424 if (user_mode(regs))
1425 trace_page_fault_user(address, regs, error_code);
1427 trace_page_fault_kernel(address, regs, error_code);
1430 static __always_inline void
1431 handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1432 unsigned long address)
1434 trace_page_fault_entries(regs, error_code, address);
1436 if (unlikely(kmmio_fault(regs, address)))
1439 /* Was the fault on kernel-controlled part of the address space? */
1440 if (unlikely(fault_in_kernel_space(address))) {
1441 do_kern_addr_fault(regs, error_code, address);
1443 do_user_addr_fault(regs, error_code, address);
1445 * User address page fault handling might have reenabled
1446 * interrupts. Fixing up all potential exit points of
1447 * do_user_addr_fault() and its leaf functions is just not
1448 * doable w/o creating an unholy mess or turning the code
1451 local_irq_disable();
1455 DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1457 unsigned long address = read_cr2();
1458 irqentry_state_t state;
1460 prefetchw(¤t->mm->mmap_lock);
1463 * KVM uses #PF vector to deliver 'page not present' events to guests
1464 * (asynchronous page fault mechanism). The event happens when a
1465 * userspace task is trying to access some valid (from guest's point of
1466 * view) memory which is not currently mapped by the host (e.g. the
1467 * memory is swapped out). Note, the corresponding "page ready" event
1468 * which is injected when the memory becomes available, is delived via
1469 * an interrupt mechanism and not a #PF exception
1470 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1472 * We are relying on the interrupted context being sane (valid RSP,
1473 * relevant locks not held, etc.), which is fine as long as the
1474 * interrupted context had IF=1. We are also relying on the KVM
1475 * async pf type field and CR2 being read consistently instead of
1476 * getting values from real and async page faults mixed up.
1480 * The async #PF handling code takes care of idtentry handling
1483 if (kvm_handle_async_pf(regs, (u32)address))
1487 * Entry handling for valid #PF from kernel mode is slightly
1488 * different: RCU is already watching and rcu_irq_enter() must not
1489 * be invoked because a kernel fault on a user space address might
1492 * In case the fault hit a RCU idle region the conditional entry
1493 * code reenabled RCU to avoid subsequent wreckage which helps
1496 state = irqentry_enter(regs);
1498 instrumentation_begin();
1499 handle_page_fault(regs, error_code, address);
1500 instrumentation_end();
1502 irqentry_exit(regs, state);