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/pgalloc.h> /* pgd_*(), ... */
25 #include <asm/fixmap.h> /* VSYSCALL_ADDR */
26 #include <asm/vsyscall.h> /* emulate_vsyscall */
27 #include <asm/vm86.h> /* struct vm86 */
28 #include <asm/mmu_context.h> /* vma_pkey() */
29 #include <asm/efi.h> /* efi_recover_from_page_fault()*/
30 #include <asm/desc.h> /* store_idt(), ... */
31 #include <asm/cpu_entry_area.h> /* exception stack */
33 #define CREATE_TRACE_POINTS
34 #include <asm/trace/exceptions.h>
37 * Returns 0 if mmiotrace is disabled, or if the fault is not
38 * handled by mmiotrace:
40 static nokprobe_inline int
41 kmmio_fault(struct pt_regs *regs, unsigned long addr)
43 if (unlikely(is_kmmio_active()))
44 if (kmmio_handler(regs, addr) == 1)
54 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
55 * Check that here and ignore it.
59 * Sometimes the CPU reports invalid exceptions on prefetch.
60 * Check that here and ignore it.
62 * Opcode checker based on code by Richard Brunner.
65 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
66 unsigned char opcode, int *prefetch)
68 unsigned char instr_hi = opcode & 0xf0;
69 unsigned char instr_lo = opcode & 0x0f;
75 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
76 * In X86_64 long mode, the CPU will signal invalid
77 * opcode if some of these prefixes are present so
78 * X86_64 will never get here anyway
80 return ((instr_lo & 7) == 0x6);
84 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
85 * Need to figure out under what instruction mode the
86 * instruction was issued. Could check the LDT for lm,
87 * but for now it's good enough to assume that long
88 * mode only uses well known segments or kernel.
90 return (!user_mode(regs) || user_64bit_mode(regs));
93 /* 0x64 thru 0x67 are valid prefixes in all modes. */
94 return (instr_lo & 0xC) == 0x4;
96 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
97 return !instr_lo || (instr_lo>>1) == 1;
99 /* Prefetch instruction is 0x0F0D or 0x0F18 */
100 if (probe_kernel_address(instr, opcode))
103 *prefetch = (instr_lo == 0xF) &&
104 (opcode == 0x0D || opcode == 0x18);
112 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
114 unsigned char *max_instr;
115 unsigned char *instr;
119 * If it was a exec (instruction fetch) fault on NX page, then
120 * do not ignore the fault:
122 if (error_code & X86_PF_INSTR)
125 instr = (void *)convert_ip_to_linear(current, regs);
126 max_instr = instr + 15;
128 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
131 while (instr < max_instr) {
132 unsigned char opcode;
134 if (probe_kernel_address(instr, opcode))
139 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
145 DEFINE_SPINLOCK(pgd_lock);
149 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
151 unsigned index = pgd_index(address);
158 pgd_k = init_mm.pgd + index;
160 if (!pgd_present(*pgd_k))
164 * set_pgd(pgd, *pgd_k); here would be useless on PAE
165 * and redundant with the set_pmd() on non-PAE. As would
168 p4d = p4d_offset(pgd, address);
169 p4d_k = p4d_offset(pgd_k, address);
170 if (!p4d_present(*p4d_k))
173 pud = pud_offset(p4d, address);
174 pud_k = pud_offset(p4d_k, address);
175 if (!pud_present(*pud_k))
178 pmd = pmd_offset(pud, address);
179 pmd_k = pmd_offset(pud_k, address);
181 if (pmd_present(*pmd) != pmd_present(*pmd_k))
182 set_pmd(pmd, *pmd_k);
184 if (!pmd_present(*pmd_k))
187 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
192 static void vmalloc_sync(void)
194 unsigned long address;
196 if (SHARED_KERNEL_PMD)
199 for (address = VMALLOC_START & PMD_MASK;
200 address >= TASK_SIZE_MAX && address < VMALLOC_END;
201 address += PMD_SIZE) {
204 spin_lock(&pgd_lock);
205 list_for_each_entry(page, &pgd_list, lru) {
206 spinlock_t *pgt_lock;
208 /* the pgt_lock only for Xen */
209 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
212 vmalloc_sync_one(page_address(page), address);
213 spin_unlock(pgt_lock);
215 spin_unlock(&pgd_lock);
219 void vmalloc_sync_mappings(void)
224 void vmalloc_sync_unmappings(void)
232 * Handle a fault on the vmalloc or module mapping area
234 static noinline int vmalloc_fault(unsigned long address)
236 unsigned long pgd_paddr;
240 /* Make sure we are in vmalloc area: */
241 if (!(address >= VMALLOC_START && address < VMALLOC_END))
245 * Synchronize this task's top level page-table
246 * with the 'reference' page table.
248 * Do _not_ use "current" here. We might be inside
249 * an interrupt in the middle of a task switch..
251 pgd_paddr = read_cr3_pa();
252 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
256 if (pmd_large(*pmd_k))
259 pte_k = pte_offset_kernel(pmd_k, address);
260 if (!pte_present(*pte_k))
265 NOKPROBE_SYMBOL(vmalloc_fault);
268 * Did it hit the DOS screen memory VA from vm86 mode?
271 check_v8086_mode(struct pt_regs *regs, unsigned long address,
272 struct task_struct *tsk)
277 if (!v8086_mode(regs) || !tsk->thread.vm86)
280 bit = (address - 0xA0000) >> PAGE_SHIFT;
282 tsk->thread.vm86->screen_bitmap |= 1 << bit;
286 static bool low_pfn(unsigned long pfn)
288 return pfn < max_low_pfn;
291 static void dump_pagetable(unsigned long address)
293 pgd_t *base = __va(read_cr3_pa());
294 pgd_t *pgd = &base[pgd_index(address)];
300 #ifdef CONFIG_X86_PAE
301 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
302 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
304 #define pr_pde pr_cont
306 #define pr_pde pr_info
308 p4d = p4d_offset(pgd, address);
309 pud = pud_offset(p4d, address);
310 pmd = pmd_offset(pud, address);
311 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
315 * We must not directly access the pte in the highpte
316 * case if the page table is located in highmem.
317 * And let's rather not kmap-atomic the pte, just in case
318 * it's allocated already:
320 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
323 pte = pte_offset_kernel(pmd, address);
324 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
329 #else /* CONFIG_X86_64: */
331 void vmalloc_sync_mappings(void)
334 * 64-bit mappings might allocate new p4d/pud pages
335 * that need to be propagated to all tasks' PGDs.
337 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
340 void vmalloc_sync_unmappings(void)
343 * Unmappings never allocate or free p4d/pud pages.
344 * No work is required here.
351 * Handle a fault on the vmalloc area
353 static noinline int vmalloc_fault(unsigned long address)
361 /* Make sure we are in vmalloc area: */
362 if (!(address >= VMALLOC_START && address < VMALLOC_END))
366 * Copy kernel mappings over when needed. This can also
367 * happen within a race in page table update. In the later
370 pgd = (pgd_t *)__va(read_cr3_pa()) + pgd_index(address);
371 pgd_k = pgd_offset_k(address);
372 if (pgd_none(*pgd_k))
375 if (pgtable_l5_enabled()) {
376 if (pgd_none(*pgd)) {
377 set_pgd(pgd, *pgd_k);
378 arch_flush_lazy_mmu_mode();
380 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_k));
384 /* With 4-level paging, copying happens on the p4d level. */
385 p4d = p4d_offset(pgd, address);
386 p4d_k = p4d_offset(pgd_k, address);
387 if (p4d_none(*p4d_k))
390 if (p4d_none(*p4d) && !pgtable_l5_enabled()) {
391 set_p4d(p4d, *p4d_k);
392 arch_flush_lazy_mmu_mode();
394 BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_k));
397 BUILD_BUG_ON(CONFIG_PGTABLE_LEVELS < 4);
399 pud = pud_offset(p4d, address);
406 pmd = pmd_offset(pud, address);
413 pte = pte_offset_kernel(pmd, address);
414 if (!pte_present(*pte))
419 NOKPROBE_SYMBOL(vmalloc_fault);
421 #ifdef CONFIG_CPU_SUP_AMD
422 static const char errata93_warning[] =
424 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
425 "******* Working around it, but it may cause SEGVs or burn power.\n"
426 "******* Please consider a BIOS update.\n"
427 "******* Disabling USB legacy in the BIOS may also help.\n";
431 * No vm86 mode in 64-bit mode:
434 check_v8086_mode(struct pt_regs *regs, unsigned long address,
435 struct task_struct *tsk)
439 static int bad_address(void *p)
443 return probe_kernel_address((unsigned long *)p, dummy);
446 static void dump_pagetable(unsigned long address)
448 pgd_t *base = __va(read_cr3_pa());
449 pgd_t *pgd = base + pgd_index(address);
455 if (bad_address(pgd))
458 pr_info("PGD %lx ", pgd_val(*pgd));
460 if (!pgd_present(*pgd))
463 p4d = p4d_offset(pgd, address);
464 if (bad_address(p4d))
467 pr_cont("P4D %lx ", p4d_val(*p4d));
468 if (!p4d_present(*p4d) || p4d_large(*p4d))
471 pud = pud_offset(p4d, address);
472 if (bad_address(pud))
475 pr_cont("PUD %lx ", pud_val(*pud));
476 if (!pud_present(*pud) || pud_large(*pud))
479 pmd = pmd_offset(pud, address);
480 if (bad_address(pmd))
483 pr_cont("PMD %lx ", pmd_val(*pmd));
484 if (!pmd_present(*pmd) || pmd_large(*pmd))
487 pte = pte_offset_kernel(pmd, address);
488 if (bad_address(pte))
491 pr_cont("PTE %lx", pte_val(*pte));
499 #endif /* CONFIG_X86_64 */
502 * Workaround for K8 erratum #93 & buggy BIOS.
504 * BIOS SMM functions are required to use a specific workaround
505 * to avoid corruption of the 64bit RIP register on C stepping K8.
507 * A lot of BIOS that didn't get tested properly miss this.
509 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
510 * Try to work around it here.
512 * Note we only handle faults in kernel here.
513 * Does nothing on 32-bit.
515 static int is_errata93(struct pt_regs *regs, unsigned long address)
517 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
518 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
519 || boot_cpu_data.x86 != 0xf)
522 if (address != regs->ip)
525 if ((address >> 32) != 0)
528 address |= 0xffffffffUL << 32;
529 if ((address >= (u64)_stext && address <= (u64)_etext) ||
530 (address >= MODULES_VADDR && address <= MODULES_END)) {
531 printk_once(errata93_warning);
540 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
541 * to illegal addresses >4GB.
543 * We catch this in the page fault handler because these addresses
544 * are not reachable. Just detect this case and return. Any code
545 * segment in LDT is compatibility mode.
547 static int is_errata100(struct pt_regs *regs, unsigned long address)
550 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
556 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
558 #ifdef CONFIG_X86_F00F_BUG
562 * Pentium F0 0F C7 C8 bug workaround:
564 if (boot_cpu_has_bug(X86_BUG_F00F)) {
565 nr = (address - idt_descr.address) >> 3;
568 do_invalid_op(regs, 0);
576 static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
578 u32 offset = (index >> 3) * sizeof(struct desc_struct);
580 struct ldttss_desc desc;
583 pr_alert("%s: NULL\n", name);
587 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
588 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
592 if (probe_kernel_read(&desc, (void *)(gdt->address + offset),
593 sizeof(struct ldttss_desc))) {
594 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
599 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
601 addr |= ((u64)desc.base3 << 32);
603 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
604 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
608 show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
610 if (!oops_may_print())
613 if (error_code & X86_PF_INSTR) {
618 pgd = __va(read_cr3_pa());
619 pgd += pgd_index(address);
621 pte = lookup_address_in_pgd(pgd, address, &level);
623 if (pte && pte_present(*pte) && !pte_exec(*pte))
624 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
625 from_kuid(&init_user_ns, current_uid()));
626 if (pte && pte_present(*pte) && pte_exec(*pte) &&
627 (pgd_flags(*pgd) & _PAGE_USER) &&
628 (__read_cr4() & X86_CR4_SMEP))
629 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
630 from_kuid(&init_user_ns, current_uid()));
633 if (address < PAGE_SIZE && !user_mode(regs))
634 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
637 pr_alert("BUG: unable to handle page fault for address: %px\n",
640 pr_alert("#PF: %s %s in %s mode\n",
641 (error_code & X86_PF_USER) ? "user" : "supervisor",
642 (error_code & X86_PF_INSTR) ? "instruction fetch" :
643 (error_code & X86_PF_WRITE) ? "write access" :
645 user_mode(regs) ? "user" : "kernel");
646 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
647 !(error_code & X86_PF_PROT) ? "not-present page" :
648 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
649 (error_code & X86_PF_PK) ? "protection keys violation" :
650 "permissions violation");
652 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
653 struct desc_ptr idt, gdt;
657 * This can happen for quite a few reasons. The more obvious
658 * ones are faults accessing the GDT, or LDT. Perhaps
659 * surprisingly, if the CPU tries to deliver a benign or
660 * contributory exception from user code and gets a page fault
661 * during delivery, the page fault can be delivered as though
662 * it originated directly from user code. This could happen
663 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
664 * kernel or IST stack.
668 /* Usable even on Xen PV -- it's just slow. */
669 native_store_gdt(&gdt);
671 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
672 idt.address, idt.size, gdt.address, gdt.size);
675 show_ldttss(&gdt, "LDTR", ldtr);
678 show_ldttss(&gdt, "TR", tr);
681 dump_pagetable(address);
685 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
686 unsigned long address)
688 struct task_struct *tsk;
692 flags = oops_begin();
696 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
698 dump_pagetable(address);
700 if (__die("Bad pagetable", regs, error_code))
703 oops_end(flags, regs, sig);
706 static void set_signal_archinfo(unsigned long address,
707 unsigned long error_code)
709 struct task_struct *tsk = current;
712 * To avoid leaking information about the kernel page
713 * table layout, pretend that user-mode accesses to
714 * kernel addresses are always protection faults.
716 * NB: This means that failed vsyscalls with vsyscall=none
717 * will have the PROT bit. This doesn't leak any
718 * information and does not appear to cause any problems.
720 if (address >= TASK_SIZE_MAX)
721 error_code |= X86_PF_PROT;
723 tsk->thread.trap_nr = X86_TRAP_PF;
724 tsk->thread.error_code = error_code | X86_PF_USER;
725 tsk->thread.cr2 = address;
729 no_context(struct pt_regs *regs, unsigned long error_code,
730 unsigned long address, int signal, int si_code)
732 struct task_struct *tsk = current;
736 if (user_mode(regs)) {
738 * This is an implicit supervisor-mode access from user
739 * mode. Bypass all the kernel-mode recovery code and just
745 /* Are we prepared to handle this kernel fault? */
746 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
748 * Any interrupt that takes a fault gets the fixup. This makes
749 * the below recursive fault logic only apply to a faults from
756 * Per the above we're !in_interrupt(), aka. task context.
758 * In this case we need to make sure we're not recursively
759 * faulting through the emulate_vsyscall() logic.
761 if (current->thread.sig_on_uaccess_err && signal) {
762 set_signal_archinfo(address, error_code);
764 /* XXX: hwpoison faults will set the wrong code. */
765 force_sig_fault(signal, si_code, (void __user *)address);
769 * Barring that, we can do the fixup and be happy.
774 #ifdef CONFIG_VMAP_STACK
776 * Stack overflow? During boot, we can fault near the initial
777 * stack in the direct map, but that's not an overflow -- check
778 * that we're in vmalloc space to avoid this.
780 if (is_vmalloc_addr((void *)address) &&
781 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
782 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
783 unsigned long stack = __this_cpu_ist_top_va(DF) - sizeof(void *);
785 * We're likely to be running with very little stack space
786 * left. It's plausible that we'd hit this condition but
787 * double-fault even before we get this far, in which case
788 * we're fine: the double-fault handler will deal with it.
790 * We don't want to make it all the way into the oops code
791 * and then double-fault, though, because we're likely to
792 * break the console driver and lose most of the stack dump.
794 asm volatile ("movq %[stack], %%rsp\n\t"
795 "call handle_stack_overflow\n\t"
797 : ASM_CALL_CONSTRAINT
798 : "D" ("kernel stack overflow (page fault)"),
799 "S" (regs), "d" (address),
800 [stack] "rm" (stack));
808 * Valid to do another page fault here, because if this fault
809 * had been triggered by is_prefetch fixup_exception would have
814 * Hall of shame of CPU/BIOS bugs.
816 if (is_prefetch(regs, error_code, address))
819 if (is_errata93(regs, address))
823 * Buggy firmware could access regions which might page fault, try to
824 * recover from such faults.
826 if (IS_ENABLED(CONFIG_EFI))
827 efi_recover_from_page_fault(address);
831 * Oops. The kernel tried to access some bad page. We'll have to
832 * terminate things with extreme prejudice:
834 flags = oops_begin();
836 show_fault_oops(regs, error_code, address);
838 if (task_stack_end_corrupted(tsk))
839 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
842 if (__die("Oops", regs, error_code))
845 /* Executive summary in case the body of the oops scrolled away */
846 printk(KERN_DEFAULT "CR2: %016lx\n", address);
848 oops_end(flags, regs, sig);
852 * Print out info about fatal segfaults, if the show_unhandled_signals
856 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
857 unsigned long address, struct task_struct *tsk)
859 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
861 if (!unhandled_signal(tsk, SIGSEGV))
864 if (!printk_ratelimit())
867 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
868 loglvl, tsk->comm, task_pid_nr(tsk), address,
869 (void *)regs->ip, (void *)regs->sp, error_code);
871 print_vma_addr(KERN_CONT " in ", regs->ip);
873 printk(KERN_CONT "\n");
875 show_opcodes(regs, loglvl);
879 * The (legacy) vsyscall page is the long page in the kernel portion
880 * of the address space that has user-accessible permissions.
882 static bool is_vsyscall_vaddr(unsigned long vaddr)
884 return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR);
888 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
889 unsigned long address, u32 pkey, int si_code)
891 struct task_struct *tsk = current;
893 /* User mode accesses just cause a SIGSEGV */
894 if (user_mode(regs) && (error_code & X86_PF_USER)) {
896 * It's possible to have interrupts off here:
901 * Valid to do another page fault here because this one came
904 if (is_prefetch(regs, error_code, address))
907 if (is_errata100(regs, address))
911 * To avoid leaking information about the kernel page table
912 * layout, pretend that user-mode accesses to kernel addresses
913 * are always protection faults.
915 if (address >= TASK_SIZE_MAX)
916 error_code |= X86_PF_PROT;
918 if (likely(show_unhandled_signals))
919 show_signal_msg(regs, error_code, address, tsk);
921 set_signal_archinfo(address, error_code);
923 if (si_code == SEGV_PKUERR)
924 force_sig_pkuerr((void __user *)address, pkey);
926 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
931 if (is_f00f_bug(regs, address))
934 no_context(regs, error_code, address, SIGSEGV, si_code);
938 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
939 unsigned long address)
941 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
945 __bad_area(struct pt_regs *regs, unsigned long error_code,
946 unsigned long address, u32 pkey, int si_code)
948 struct mm_struct *mm = current->mm;
950 * Something tried to access memory that isn't in our memory map..
951 * Fix it, but check if it's kernel or user first..
953 up_read(&mm->mmap_sem);
955 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
959 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
961 __bad_area(regs, error_code, address, 0, SEGV_MAPERR);
964 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
965 struct vm_area_struct *vma)
967 /* This code is always called on the current mm */
968 bool foreign = false;
970 if (!boot_cpu_has(X86_FEATURE_OSPKE))
972 if (error_code & X86_PF_PK)
974 /* this checks permission keys on the VMA: */
975 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
976 (error_code & X86_PF_INSTR), foreign))
982 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
983 unsigned long address, struct vm_area_struct *vma)
986 * This OSPKE check is not strictly necessary at runtime.
987 * But, doing it this way allows compiler optimizations
988 * if pkeys are compiled out.
990 if (bad_area_access_from_pkeys(error_code, vma)) {
992 * A protection key fault means that the PKRU value did not allow
993 * access to some PTE. Userspace can figure out what PKRU was
994 * from the XSAVE state. This function captures the pkey from
995 * the vma and passes it to userspace so userspace can discover
996 * which protection key was set on the PTE.
998 * If we get here, we know that the hardware signaled a X86_PF_PK
999 * fault and that there was a VMA once we got in the fault
1000 * handler. It does *not* guarantee that the VMA we find here
1001 * was the one that we faulted on.
1003 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
1004 * 2. T1 : set PKRU to deny access to pkey=4, touches page
1006 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
1007 * 5. T1 : enters fault handler, takes mmap_sem, etc...
1008 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
1009 * faulted on a pte with its pkey=4.
1011 u32 pkey = vma_pkey(vma);
1013 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
1015 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
1020 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
1023 /* Kernel mode? Handle exceptions or die: */
1024 if (!(error_code & X86_PF_USER)) {
1025 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1029 /* User-space => ok to do another page fault: */
1030 if (is_prefetch(regs, error_code, address))
1033 set_signal_archinfo(address, error_code);
1035 #ifdef CONFIG_MEMORY_FAILURE
1036 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
1037 struct task_struct *tsk = current;
1041 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
1042 tsk->comm, tsk->pid, address);
1043 if (fault & VM_FAULT_HWPOISON_LARGE)
1044 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
1045 if (fault & VM_FAULT_HWPOISON)
1047 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
1051 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
1054 static noinline void
1055 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
1056 unsigned long address, vm_fault_t fault)
1058 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) {
1059 no_context(regs, error_code, address, 0, 0);
1063 if (fault & VM_FAULT_OOM) {
1064 /* Kernel mode? Handle exceptions or die: */
1065 if (!(error_code & X86_PF_USER)) {
1066 no_context(regs, error_code, address,
1067 SIGSEGV, SEGV_MAPERR);
1072 * We ran out of memory, call the OOM killer, and return the
1073 * userspace (which will retry the fault, or kill us if we got
1076 pagefault_out_of_memory();
1078 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1079 VM_FAULT_HWPOISON_LARGE))
1080 do_sigbus(regs, error_code, address, fault);
1081 else if (fault & VM_FAULT_SIGSEGV)
1082 bad_area_nosemaphore(regs, error_code, address);
1088 static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
1090 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
1093 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1100 * Handle a spurious fault caused by a stale TLB entry.
1102 * This allows us to lazily refresh the TLB when increasing the
1103 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1104 * eagerly is very expensive since that implies doing a full
1105 * cross-processor TLB flush, even if no stale TLB entries exist
1106 * on other processors.
1108 * Spurious faults may only occur if the TLB contains an entry with
1109 * fewer permission than the page table entry. Non-present (P = 0)
1110 * and reserved bit (R = 1) faults are never spurious.
1112 * There are no security implications to leaving a stale TLB when
1113 * increasing the permissions on a page.
1115 * Returns non-zero if a spurious fault was handled, zero otherwise.
1117 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1118 * (Optional Invalidation).
1121 spurious_kernel_fault(unsigned long error_code, unsigned long address)
1131 * Only writes to RO or instruction fetches from NX may cause
1134 * These could be from user or supervisor accesses but the TLB
1135 * is only lazily flushed after a kernel mapping protection
1136 * change, so user accesses are not expected to cause spurious
1139 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1140 error_code != (X86_PF_INSTR | X86_PF_PROT))
1143 pgd = init_mm.pgd + pgd_index(address);
1144 if (!pgd_present(*pgd))
1147 p4d = p4d_offset(pgd, address);
1148 if (!p4d_present(*p4d))
1151 if (p4d_large(*p4d))
1152 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1154 pud = pud_offset(p4d, address);
1155 if (!pud_present(*pud))
1158 if (pud_large(*pud))
1159 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1161 pmd = pmd_offset(pud, address);
1162 if (!pmd_present(*pmd))
1165 if (pmd_large(*pmd))
1166 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1168 pte = pte_offset_kernel(pmd, address);
1169 if (!pte_present(*pte))
1172 ret = spurious_kernel_fault_check(error_code, pte);
1177 * Make sure we have permissions in PMD.
1178 * If not, then there's a bug in the page tables:
1180 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1181 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1185 NOKPROBE_SYMBOL(spurious_kernel_fault);
1187 int show_unhandled_signals = 1;
1190 access_error(unsigned long error_code, struct vm_area_struct *vma)
1192 /* This is only called for the current mm, so: */
1193 bool foreign = false;
1196 * Read or write was blocked by protection keys. This is
1197 * always an unconditional error and can never result in
1198 * a follow-up action to resolve the fault, like a COW.
1200 if (error_code & X86_PF_PK)
1204 * Make sure to check the VMA so that we do not perform
1205 * faults just to hit a X86_PF_PK as soon as we fill in a
1208 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1209 (error_code & X86_PF_INSTR), foreign))
1212 if (error_code & X86_PF_WRITE) {
1213 /* write, present and write, not present: */
1214 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1219 /* read, present: */
1220 if (unlikely(error_code & X86_PF_PROT))
1223 /* read, not present: */
1224 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1230 static int fault_in_kernel_space(unsigned long address)
1233 * On 64-bit systems, the vsyscall page is at an address above
1234 * TASK_SIZE_MAX, but is not considered part of the kernel
1237 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1240 return address >= TASK_SIZE_MAX;
1244 * Called for all faults where 'address' is part of the kernel address
1245 * space. Might get called for faults that originate from *code* that
1246 * ran in userspace or the kernel.
1249 do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1250 unsigned long address)
1253 * Protection keys exceptions only happen on user pages. We
1254 * have no user pages in the kernel portion of the address
1255 * space, so do not expect them here.
1257 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1260 * We can fault-in kernel-space virtual memory on-demand. The
1261 * 'reference' page table is init_mm.pgd.
1263 * NOTE! We MUST NOT take any locks for this case. We may
1264 * be in an interrupt or a critical region, and should
1265 * only copy the information from the master page table,
1268 * Before doing this on-demand faulting, ensure that the
1269 * fault is not any of the following:
1270 * 1. A fault on a PTE with a reserved bit set.
1271 * 2. A fault caused by a user-mode access. (Do not demand-
1272 * fault kernel memory due to user-mode accesses).
1273 * 3. A fault caused by a page-level protection violation.
1274 * (A demand fault would be on a non-present page which
1275 * would have X86_PF_PROT==0).
1277 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1278 if (vmalloc_fault(address) >= 0)
1282 /* Was the fault spurious, caused by lazy TLB invalidation? */
1283 if (spurious_kernel_fault(hw_error_code, address))
1286 /* kprobes don't want to hook the spurious faults: */
1287 if (kprobe_page_fault(regs, X86_TRAP_PF))
1291 * Note, despite being a "bad area", there are quite a few
1292 * acceptable reasons to get here, such as erratum fixups
1293 * and handling kernel code that can fault, like get_user().
1295 * Don't take the mm semaphore here. If we fixup a prefetch
1296 * fault we could otherwise deadlock:
1298 bad_area_nosemaphore(regs, hw_error_code, address);
1300 NOKPROBE_SYMBOL(do_kern_addr_fault);
1302 /* Handle faults in the user portion of the address space */
1304 void do_user_addr_fault(struct pt_regs *regs,
1305 unsigned long hw_error_code,
1306 unsigned long address)
1308 struct vm_area_struct *vma;
1309 struct task_struct *tsk;
1310 struct mm_struct *mm;
1311 vm_fault_t fault, major = 0;
1312 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1317 /* kprobes don't want to hook the spurious faults: */
1318 if (unlikely(kprobe_page_fault(regs, X86_TRAP_PF)))
1322 * Reserved bits are never expected to be set on
1323 * entries in the user portion of the page tables.
1325 if (unlikely(hw_error_code & X86_PF_RSVD))
1326 pgtable_bad(regs, hw_error_code, address);
1329 * If SMAP is on, check for invalid kernel (supervisor) access to user
1330 * pages in the user address space. The odd case here is WRUSS,
1331 * which, according to the preliminary documentation, does not respect
1332 * SMAP and will have the USER bit set so, in all cases, SMAP
1333 * enforcement appears to be consistent with the USER bit.
1335 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1336 !(hw_error_code & X86_PF_USER) &&
1337 !(regs->flags & X86_EFLAGS_AC)))
1339 bad_area_nosemaphore(regs, hw_error_code, address);
1344 * If we're in an interrupt, have no user context or are running
1345 * in a region with pagefaults disabled then we must not take the fault
1347 if (unlikely(faulthandler_disabled() || !mm)) {
1348 bad_area_nosemaphore(regs, hw_error_code, address);
1353 * It's safe to allow irq's after cr2 has been saved and the
1354 * vmalloc fault has been handled.
1356 * User-mode registers count as a user access even for any
1357 * potential system fault or CPU buglet:
1359 if (user_mode(regs)) {
1361 flags |= FAULT_FLAG_USER;
1363 if (regs->flags & X86_EFLAGS_IF)
1367 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1369 if (hw_error_code & X86_PF_WRITE)
1370 flags |= FAULT_FLAG_WRITE;
1371 if (hw_error_code & X86_PF_INSTR)
1372 flags |= FAULT_FLAG_INSTRUCTION;
1374 #ifdef CONFIG_X86_64
1376 * Faults in the vsyscall page might need emulation. The
1377 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1378 * considered to be part of the user address space.
1380 * The vsyscall page does not have a "real" VMA, so do this
1381 * emulation before we go searching for VMAs.
1383 * PKRU never rejects instruction fetches, so we don't need
1384 * to consider the PF_PK bit.
1386 if (is_vsyscall_vaddr(address)) {
1387 if (emulate_vsyscall(hw_error_code, regs, address))
1393 * Kernel-mode access to the user address space should only occur
1394 * on well-defined single instructions listed in the exception
1395 * tables. But, an erroneous kernel fault occurring outside one of
1396 * those areas which also holds mmap_sem might deadlock attempting
1397 * to validate the fault against the address space.
1399 * Only do the expensive exception table search when we might be at
1400 * risk of a deadlock. This happens if we
1401 * 1. Failed to acquire mmap_sem, and
1402 * 2. The access did not originate in userspace.
1404 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1405 if (!user_mode(regs) && !search_exception_tables(regs->ip)) {
1407 * Fault from code in kernel from
1408 * which we do not expect faults.
1410 bad_area_nosemaphore(regs, hw_error_code, address);
1414 down_read(&mm->mmap_sem);
1417 * The above down_read_trylock() might have succeeded in
1418 * which case we'll have missed the might_sleep() from
1424 vma = find_vma(mm, address);
1425 if (unlikely(!vma)) {
1426 bad_area(regs, hw_error_code, address);
1429 if (likely(vma->vm_start <= address))
1431 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1432 bad_area(regs, hw_error_code, address);
1435 if (unlikely(expand_stack(vma, address))) {
1436 bad_area(regs, hw_error_code, address);
1441 * Ok, we have a good vm_area for this memory access, so
1442 * we can handle it..
1445 if (unlikely(access_error(hw_error_code, vma))) {
1446 bad_area_access_error(regs, hw_error_code, address, vma);
1451 * If for any reason at all we couldn't handle the fault,
1452 * make sure we exit gracefully rather than endlessly redo
1453 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1454 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1456 * Note that handle_userfault() may also release and reacquire mmap_sem
1457 * (and not return with VM_FAULT_RETRY), when returning to userland to
1458 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1459 * (potentially after handling any pending signal during the return to
1460 * userland). The return to userland is identified whenever
1461 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1463 fault = handle_mm_fault(vma, address, flags);
1464 major |= fault & VM_FAULT_MAJOR;
1467 * If we need to retry the mmap_sem has already been released,
1468 * and if there is a fatal signal pending there is no guarantee
1469 * that we made any progress. Handle this case first.
1471 if (unlikely(fault & VM_FAULT_RETRY)) {
1472 /* Retry at most once */
1473 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1474 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1475 flags |= FAULT_FLAG_TRIED;
1476 if (!fatal_signal_pending(tsk))
1480 /* User mode? Just return to handle the fatal exception */
1481 if (flags & FAULT_FLAG_USER)
1484 /* Not returning to user mode? Handle exceptions or die: */
1485 no_context(regs, hw_error_code, address, SIGBUS, BUS_ADRERR);
1489 up_read(&mm->mmap_sem);
1490 if (unlikely(fault & VM_FAULT_ERROR)) {
1491 mm_fault_error(regs, hw_error_code, address, fault);
1496 * Major/minor page fault accounting. If any of the events
1497 * returned VM_FAULT_MAJOR, we account it as a major fault.
1501 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1504 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1507 check_v8086_mode(regs, address, tsk);
1509 NOKPROBE_SYMBOL(do_user_addr_fault);
1512 * Explicitly marked noinline such that the function tracer sees this as the
1513 * page_fault entry point.
1515 static noinline void
1516 __do_page_fault(struct pt_regs *regs, unsigned long hw_error_code,
1517 unsigned long address)
1519 prefetchw(¤t->mm->mmap_sem);
1521 if (unlikely(kmmio_fault(regs, address)))
1524 /* Was the fault on kernel-controlled part of the address space? */
1525 if (unlikely(fault_in_kernel_space(address)))
1526 do_kern_addr_fault(regs, hw_error_code, address);
1528 do_user_addr_fault(regs, hw_error_code, address);
1530 NOKPROBE_SYMBOL(__do_page_fault);
1532 static __always_inline void
1533 trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1534 unsigned long address)
1536 if (!trace_pagefault_enabled())
1539 if (user_mode(regs))
1540 trace_page_fault_user(address, regs, error_code);
1542 trace_page_fault_kernel(address, regs, error_code);
1546 do_page_fault(struct pt_regs *regs, unsigned long error_code, unsigned long address)
1548 enum ctx_state prev_state;
1550 prev_state = exception_enter();
1551 trace_page_fault_entries(regs, error_code, address);
1552 __do_page_fault(regs, error_code, address);
1553 exception_exit(prev_state);
1555 NOKPROBE_SYMBOL(do_page_fault);