2 * Copyright (C) 1995 Linus Torvalds
3 * Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
4 * Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
6 #include <linux/sched.h> /* test_thread_flag(), ... */
7 #include <linux/kdebug.h> /* oops_begin/end, ... */
8 #include <linux/module.h> /* search_exception_table */
9 #include <linux/bootmem.h> /* max_low_pfn */
10 #include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */
11 #include <linux/mmiotrace.h> /* kmmio_handler, ... */
12 #include <linux/perf_event.h> /* perf_sw_event */
13 #include <linux/hugetlb.h> /* hstate_index_to_shift */
14 #include <linux/prefetch.h> /* prefetchw */
15 #include <linux/context_tracking.h> /* exception_enter(), ... */
16 #include <linux/uaccess.h> /* faulthandler_disabled() */
18 #include <asm/cpufeature.h> /* boot_cpu_has, ... */
19 #include <asm/traps.h> /* dotraplinkage, ... */
20 #include <asm/pgalloc.h> /* pgd_*(), ... */
21 #include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
22 #include <asm/fixmap.h> /* VSYSCALL_ADDR */
23 #include <asm/vsyscall.h> /* emulate_vsyscall */
24 #include <asm/vm86.h> /* struct vm86 */
25 #include <asm/mmu_context.h> /* vma_pkey() */
27 #define CREATE_TRACE_POINTS
28 #include <asm/trace/exceptions.h>
31 * Page fault error code bits:
33 * bit 0 == 0: no page found 1: protection fault
34 * bit 1 == 0: read access 1: write access
35 * bit 2 == 0: kernel-mode access 1: user-mode access
36 * bit 3 == 1: use of reserved bit detected
37 * bit 4 == 1: fault was an instruction fetch
38 * bit 5 == 1: protection keys block access
40 enum x86_pf_error_code {
51 * Returns 0 if mmiotrace is disabled, or if the fault is not
52 * handled by mmiotrace:
54 static nokprobe_inline int
55 kmmio_fault(struct pt_regs *regs, unsigned long addr)
57 if (unlikely(is_kmmio_active()))
58 if (kmmio_handler(regs, addr) == 1)
63 static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
67 /* kprobe_running() needs smp_processor_id() */
68 if (kprobes_built_in() && !user_mode(regs)) {
70 if (kprobe_running() && kprobe_fault_handler(regs, 14))
83 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
84 * Check that here and ignore it.
88 * Sometimes the CPU reports invalid exceptions on prefetch.
89 * Check that here and ignore it.
91 * Opcode checker based on code by Richard Brunner.
94 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
95 unsigned char opcode, int *prefetch)
97 unsigned char instr_hi = opcode & 0xf0;
98 unsigned char instr_lo = opcode & 0x0f;
104 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
105 * In X86_64 long mode, the CPU will signal invalid
106 * opcode if some of these prefixes are present so
107 * X86_64 will never get here anyway
109 return ((instr_lo & 7) == 0x6);
113 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
114 * Need to figure out under what instruction mode the
115 * instruction was issued. Could check the LDT for lm,
116 * but for now it's good enough to assume that long
117 * mode only uses well known segments or kernel.
119 return (!user_mode(regs) || user_64bit_mode(regs));
122 /* 0x64 thru 0x67 are valid prefixes in all modes. */
123 return (instr_lo & 0xC) == 0x4;
125 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
126 return !instr_lo || (instr_lo>>1) == 1;
128 /* Prefetch instruction is 0x0F0D or 0x0F18 */
129 if (probe_kernel_address(instr, opcode))
132 *prefetch = (instr_lo == 0xF) &&
133 (opcode == 0x0D || opcode == 0x18);
141 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
143 unsigned char *max_instr;
144 unsigned char *instr;
148 * If it was a exec (instruction fetch) fault on NX page, then
149 * do not ignore the fault:
151 if (error_code & PF_INSTR)
154 instr = (void *)convert_ip_to_linear(current, regs);
155 max_instr = instr + 15;
157 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
160 while (instr < max_instr) {
161 unsigned char opcode;
163 if (probe_kernel_address(instr, opcode))
168 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
175 * A protection key fault means that the PKRU value did not allow
176 * access to some PTE. Userspace can figure out what PKRU was
177 * from the XSAVE state, and this function fills out a field in
178 * siginfo so userspace can discover which protection key was set
181 * If we get here, we know that the hardware signaled a PF_PK
182 * fault and that there was a VMA once we got in the fault
183 * handler. It does *not* guarantee that the VMA we find here
184 * was the one that we faulted on.
186 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
187 * 2. T1 : set PKRU to deny access to pkey=4, touches page
189 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
190 * 5. T1 : enters fault handler, takes mmap_sem, etc...
191 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
192 * faulted on a pte with its pkey=4.
194 static void fill_sig_info_pkey(int si_code, siginfo_t *info,
195 struct vm_area_struct *vma)
197 /* This is effectively an #ifdef */
198 if (!boot_cpu_has(X86_FEATURE_OSPKE))
201 /* Fault not from Protection Keys: nothing to do */
202 if (si_code != SEGV_PKUERR)
205 * force_sig_info_fault() is called from a number of
206 * contexts, some of which have a VMA and some of which
207 * do not. The PF_PK handing happens after we have a
208 * valid VMA, so we should never reach this without a
212 WARN_ONCE(1, "PKU fault with no VMA passed in");
217 * si_pkey should be thought of as a strong hint, but not
218 * absolutely guranteed to be 100% accurate because of
219 * the race explained above.
221 info->si_pkey = vma_pkey(vma);
225 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
226 struct task_struct *tsk, struct vm_area_struct *vma,
232 info.si_signo = si_signo;
234 info.si_code = si_code;
235 info.si_addr = (void __user *)address;
236 if (fault & VM_FAULT_HWPOISON_LARGE)
237 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
238 if (fault & VM_FAULT_HWPOISON)
240 info.si_addr_lsb = lsb;
242 fill_sig_info_pkey(si_code, &info, vma);
244 force_sig_info(si_signo, &info, tsk);
247 DEFINE_SPINLOCK(pgd_lock);
251 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
253 unsigned index = pgd_index(address);
259 pgd_k = init_mm.pgd + index;
261 if (!pgd_present(*pgd_k))
265 * set_pgd(pgd, *pgd_k); here would be useless on PAE
266 * and redundant with the set_pmd() on non-PAE. As would
269 pud = pud_offset(pgd, address);
270 pud_k = pud_offset(pgd_k, address);
271 if (!pud_present(*pud_k))
274 pmd = pmd_offset(pud, address);
275 pmd_k = pmd_offset(pud_k, address);
276 if (!pmd_present(*pmd_k))
279 if (!pmd_present(*pmd))
280 set_pmd(pmd, *pmd_k);
282 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
287 void vmalloc_sync_all(void)
289 unsigned long address;
291 if (SHARED_KERNEL_PMD)
294 for (address = VMALLOC_START & PMD_MASK;
295 address >= TASK_SIZE && address < FIXADDR_TOP;
296 address += PMD_SIZE) {
299 spin_lock(&pgd_lock);
300 list_for_each_entry(page, &pgd_list, lru) {
301 spinlock_t *pgt_lock;
304 /* the pgt_lock only for Xen */
305 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
308 ret = vmalloc_sync_one(page_address(page), address);
309 spin_unlock(pgt_lock);
314 spin_unlock(&pgd_lock);
321 * Handle a fault on the vmalloc or module mapping area
323 static noinline int vmalloc_fault(unsigned long address)
325 unsigned long pgd_paddr;
329 /* Make sure we are in vmalloc area: */
330 if (!(address >= VMALLOC_START && address < VMALLOC_END))
333 WARN_ON_ONCE(in_nmi());
336 * Synchronize this task's top level page-table
337 * with the 'reference' page table.
339 * Do _not_ use "current" here. We might be inside
340 * an interrupt in the middle of a task switch..
342 pgd_paddr = read_cr3();
343 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
347 pte_k = pte_offset_kernel(pmd_k, address);
348 if (!pte_present(*pte_k))
353 NOKPROBE_SYMBOL(vmalloc_fault);
356 * Did it hit the DOS screen memory VA from vm86 mode?
359 check_v8086_mode(struct pt_regs *regs, unsigned long address,
360 struct task_struct *tsk)
365 if (!v8086_mode(regs) || !tsk->thread.vm86)
368 bit = (address - 0xA0000) >> PAGE_SHIFT;
370 tsk->thread.vm86->screen_bitmap |= 1 << bit;
374 static bool low_pfn(unsigned long pfn)
376 return pfn < max_low_pfn;
379 static void dump_pagetable(unsigned long address)
381 pgd_t *base = __va(read_cr3());
382 pgd_t *pgd = &base[pgd_index(address)];
386 #ifdef CONFIG_X86_PAE
387 printk("*pdpt = %016Lx ", pgd_val(*pgd));
388 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
391 pmd = pmd_offset(pud_offset(pgd, address), address);
392 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
395 * We must not directly access the pte in the highpte
396 * case if the page table is located in highmem.
397 * And let's rather not kmap-atomic the pte, just in case
398 * it's allocated already:
400 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
403 pte = pte_offset_kernel(pmd, address);
404 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
409 #else /* CONFIG_X86_64: */
411 void vmalloc_sync_all(void)
413 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END, 0);
419 * Handle a fault on the vmalloc area
421 * This assumes no large pages in there.
423 static noinline int vmalloc_fault(unsigned long address)
425 pgd_t *pgd, *pgd_ref;
426 pud_t *pud, *pud_ref;
427 pmd_t *pmd, *pmd_ref;
428 pte_t *pte, *pte_ref;
430 /* Make sure we are in vmalloc area: */
431 if (!(address >= VMALLOC_START && address < VMALLOC_END))
434 WARN_ON_ONCE(in_nmi());
437 * Copy kernel mappings over when needed. This can also
438 * happen within a race in page table update. In the later
441 pgd = pgd_offset(current->active_mm, address);
442 pgd_ref = pgd_offset_k(address);
443 if (pgd_none(*pgd_ref))
446 if (pgd_none(*pgd)) {
447 set_pgd(pgd, *pgd_ref);
448 arch_flush_lazy_mmu_mode();
450 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
454 * Below here mismatches are bugs because these lower tables
458 pud = pud_offset(pgd, address);
459 pud_ref = pud_offset(pgd_ref, address);
460 if (pud_none(*pud_ref))
463 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
466 pmd = pmd_offset(pud, address);
467 pmd_ref = pmd_offset(pud_ref, address);
468 if (pmd_none(*pmd_ref))
471 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
474 pte_ref = pte_offset_kernel(pmd_ref, address);
475 if (!pte_present(*pte_ref))
478 pte = pte_offset_kernel(pmd, address);
481 * Don't use pte_page here, because the mappings can point
482 * outside mem_map, and the NUMA hash lookup cannot handle
485 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
490 NOKPROBE_SYMBOL(vmalloc_fault);
492 #ifdef CONFIG_CPU_SUP_AMD
493 static const char errata93_warning[] =
495 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
496 "******* Working around it, but it may cause SEGVs or burn power.\n"
497 "******* Please consider a BIOS update.\n"
498 "******* Disabling USB legacy in the BIOS may also help.\n";
502 * No vm86 mode in 64-bit mode:
505 check_v8086_mode(struct pt_regs *regs, unsigned long address,
506 struct task_struct *tsk)
510 static int bad_address(void *p)
514 return probe_kernel_address((unsigned long *)p, dummy);
517 static void dump_pagetable(unsigned long address)
519 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
520 pgd_t *pgd = base + pgd_index(address);
525 if (bad_address(pgd))
528 printk("PGD %lx ", pgd_val(*pgd));
530 if (!pgd_present(*pgd))
533 pud = pud_offset(pgd, address);
534 if (bad_address(pud))
537 printk("PUD %lx ", pud_val(*pud));
538 if (!pud_present(*pud) || pud_large(*pud))
541 pmd = pmd_offset(pud, address);
542 if (bad_address(pmd))
545 printk("PMD %lx ", pmd_val(*pmd));
546 if (!pmd_present(*pmd) || pmd_large(*pmd))
549 pte = pte_offset_kernel(pmd, address);
550 if (bad_address(pte))
553 printk("PTE %lx", pte_val(*pte));
561 #endif /* CONFIG_X86_64 */
564 * Workaround for K8 erratum #93 & buggy BIOS.
566 * BIOS SMM functions are required to use a specific workaround
567 * to avoid corruption of the 64bit RIP register on C stepping K8.
569 * A lot of BIOS that didn't get tested properly miss this.
571 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
572 * Try to work around it here.
574 * Note we only handle faults in kernel here.
575 * Does nothing on 32-bit.
577 static int is_errata93(struct pt_regs *regs, unsigned long address)
579 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
580 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
581 || boot_cpu_data.x86 != 0xf)
584 if (address != regs->ip)
587 if ((address >> 32) != 0)
590 address |= 0xffffffffUL << 32;
591 if ((address >= (u64)_stext && address <= (u64)_etext) ||
592 (address >= MODULES_VADDR && address <= MODULES_END)) {
593 printk_once(errata93_warning);
602 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
603 * to illegal addresses >4GB.
605 * We catch this in the page fault handler because these addresses
606 * are not reachable. Just detect this case and return. Any code
607 * segment in LDT is compatibility mode.
609 static int is_errata100(struct pt_regs *regs, unsigned long address)
612 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
618 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
620 #ifdef CONFIG_X86_F00F_BUG
624 * Pentium F0 0F C7 C8 bug workaround:
626 if (boot_cpu_has_bug(X86_BUG_F00F)) {
627 nr = (address - idt_descr.address) >> 3;
630 do_invalid_op(regs, 0);
638 static const char nx_warning[] = KERN_CRIT
639 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
640 static const char smep_warning[] = KERN_CRIT
641 "unable to execute userspace code (SMEP?) (uid: %d)\n";
644 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
645 unsigned long address)
647 if (!oops_may_print())
650 if (error_code & PF_INSTR) {
655 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
656 pgd += pgd_index(address);
658 pte = lookup_address_in_pgd(pgd, address, &level);
660 if (pte && pte_present(*pte) && !pte_exec(*pte))
661 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
662 if (pte && pte_present(*pte) && pte_exec(*pte) &&
663 (pgd_flags(*pgd) & _PAGE_USER) &&
664 (__read_cr4() & X86_CR4_SMEP))
665 printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
668 printk(KERN_ALERT "BUG: unable to handle kernel ");
669 if (address < PAGE_SIZE)
670 printk(KERN_CONT "NULL pointer dereference");
672 printk(KERN_CONT "paging request");
674 printk(KERN_CONT " at %p\n", (void *) address);
675 printk(KERN_ALERT "IP:");
676 printk_address(regs->ip);
678 dump_pagetable(address);
682 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
683 unsigned long address)
685 struct task_struct *tsk;
689 flags = oops_begin();
693 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
695 dump_pagetable(address);
697 tsk->thread.cr2 = address;
698 tsk->thread.trap_nr = X86_TRAP_PF;
699 tsk->thread.error_code = error_code;
701 if (__die("Bad pagetable", regs, error_code))
704 oops_end(flags, regs, sig);
708 no_context(struct pt_regs *regs, unsigned long error_code,
709 unsigned long address, int signal, int si_code)
711 struct task_struct *tsk = current;
714 /* No context means no VMA to pass down */
715 struct vm_area_struct *vma = NULL;
717 /* Are we prepared to handle this kernel fault? */
718 if (fixup_exception(regs)) {
720 * Any interrupt that takes a fault gets the fixup. This makes
721 * the below recursive fault logic only apply to a faults from
728 * Per the above we're !in_interrupt(), aka. task context.
730 * In this case we need to make sure we're not recursively
731 * faulting through the emulate_vsyscall() logic.
733 if (current_thread_info()->sig_on_uaccess_error && signal) {
734 tsk->thread.trap_nr = X86_TRAP_PF;
735 tsk->thread.error_code = error_code | PF_USER;
736 tsk->thread.cr2 = address;
738 /* XXX: hwpoison faults will set the wrong code. */
739 force_sig_info_fault(signal, si_code, address,
744 * Barring that, we can do the fixup and be happy.
752 * Valid to do another page fault here, because if this fault
753 * had been triggered by is_prefetch fixup_exception would have
758 * Hall of shame of CPU/BIOS bugs.
760 if (is_prefetch(regs, error_code, address))
763 if (is_errata93(regs, address))
767 * Oops. The kernel tried to access some bad page. We'll have to
768 * terminate things with extreme prejudice:
770 flags = oops_begin();
772 show_fault_oops(regs, error_code, address);
774 if (task_stack_end_corrupted(tsk))
775 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
777 tsk->thread.cr2 = address;
778 tsk->thread.trap_nr = X86_TRAP_PF;
779 tsk->thread.error_code = error_code;
782 if (__die("Oops", regs, error_code))
785 /* Executive summary in case the body of the oops scrolled away */
786 printk(KERN_DEFAULT "CR2: %016lx\n", address);
788 oops_end(flags, regs, sig);
792 * Print out info about fatal segfaults, if the show_unhandled_signals
796 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
797 unsigned long address, struct task_struct *tsk)
799 if (!unhandled_signal(tsk, SIGSEGV))
802 if (!printk_ratelimit())
805 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
806 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
807 tsk->comm, task_pid_nr(tsk), address,
808 (void *)regs->ip, (void *)regs->sp, error_code);
810 print_vma_addr(KERN_CONT " in ", regs->ip);
812 printk(KERN_CONT "\n");
816 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
817 unsigned long address, struct vm_area_struct *vma,
820 struct task_struct *tsk = current;
822 /* User mode accesses just cause a SIGSEGV */
823 if (error_code & PF_USER) {
825 * It's possible to have interrupts off here:
830 * Valid to do another page fault here because this one came
833 if (is_prefetch(regs, error_code, address))
836 if (is_errata100(regs, address))
841 * Instruction fetch faults in the vsyscall page might need
844 if (unlikely((error_code & PF_INSTR) &&
845 ((address & ~0xfff) == VSYSCALL_ADDR))) {
846 if (emulate_vsyscall(regs, address))
850 /* Kernel addresses are always protection faults: */
851 if (address >= TASK_SIZE)
852 error_code |= PF_PROT;
854 if (likely(show_unhandled_signals))
855 show_signal_msg(regs, error_code, address, tsk);
857 tsk->thread.cr2 = address;
858 tsk->thread.error_code = error_code;
859 tsk->thread.trap_nr = X86_TRAP_PF;
861 force_sig_info_fault(SIGSEGV, si_code, address, tsk, vma, 0);
866 if (is_f00f_bug(regs, address))
869 no_context(regs, error_code, address, SIGSEGV, si_code);
873 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
874 unsigned long address, struct vm_area_struct *vma)
876 __bad_area_nosemaphore(regs, error_code, address, vma, SEGV_MAPERR);
880 __bad_area(struct pt_regs *regs, unsigned long error_code,
881 unsigned long address, struct vm_area_struct *vma, int si_code)
883 struct mm_struct *mm = current->mm;
886 * Something tried to access memory that isn't in our memory map..
887 * Fix it, but check if it's kernel or user first..
889 up_read(&mm->mmap_sem);
891 __bad_area_nosemaphore(regs, error_code, address, vma, si_code);
895 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
897 __bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
900 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
901 struct vm_area_struct *vma)
903 /* This code is always called on the current mm */
904 bool foreign = false;
906 if (!boot_cpu_has(X86_FEATURE_OSPKE))
908 if (error_code & PF_PK)
910 /* this checks permission keys on the VMA: */
911 if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
912 (error_code & PF_INSTR), foreign))
918 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
919 unsigned long address, struct vm_area_struct *vma)
922 * This OSPKE check is not strictly necessary at runtime.
923 * But, doing it this way allows compiler optimizations
924 * if pkeys are compiled out.
926 if (bad_area_access_from_pkeys(error_code, vma))
927 __bad_area(regs, error_code, address, vma, SEGV_PKUERR);
929 __bad_area(regs, error_code, address, vma, SEGV_ACCERR);
933 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
934 struct vm_area_struct *vma, unsigned int fault)
936 struct task_struct *tsk = current;
937 int code = BUS_ADRERR;
939 /* Kernel mode? Handle exceptions or die: */
940 if (!(error_code & PF_USER)) {
941 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
945 /* User-space => ok to do another page fault: */
946 if (is_prefetch(regs, error_code, address))
949 tsk->thread.cr2 = address;
950 tsk->thread.error_code = error_code;
951 tsk->thread.trap_nr = X86_TRAP_PF;
953 #ifdef CONFIG_MEMORY_FAILURE
954 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
956 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
957 tsk->comm, tsk->pid, address);
958 code = BUS_MCEERR_AR;
961 force_sig_info_fault(SIGBUS, code, address, tsk, vma, fault);
965 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
966 unsigned long address, struct vm_area_struct *vma,
969 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
970 no_context(regs, error_code, address, 0, 0);
974 if (fault & VM_FAULT_OOM) {
975 /* Kernel mode? Handle exceptions or die: */
976 if (!(error_code & PF_USER)) {
977 no_context(regs, error_code, address,
978 SIGSEGV, SEGV_MAPERR);
983 * We ran out of memory, call the OOM killer, and return the
984 * userspace (which will retry the fault, or kill us if we got
987 pagefault_out_of_memory();
989 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
990 VM_FAULT_HWPOISON_LARGE))
991 do_sigbus(regs, error_code, address, vma, fault);
992 else if (fault & VM_FAULT_SIGSEGV)
993 bad_area_nosemaphore(regs, error_code, address, vma);
999 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
1001 if ((error_code & PF_WRITE) && !pte_write(*pte))
1004 if ((error_code & PF_INSTR) && !pte_exec(*pte))
1007 * Note: We do not do lazy flushing on protection key
1008 * changes, so no spurious fault will ever set PF_PK.
1010 if ((error_code & PF_PK))
1017 * Handle a spurious fault caused by a stale TLB entry.
1019 * This allows us to lazily refresh the TLB when increasing the
1020 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1021 * eagerly is very expensive since that implies doing a full
1022 * cross-processor TLB flush, even if no stale TLB entries exist
1023 * on other processors.
1025 * Spurious faults may only occur if the TLB contains an entry with
1026 * fewer permission than the page table entry. Non-present (P = 0)
1027 * and reserved bit (R = 1) faults are never spurious.
1029 * There are no security implications to leaving a stale TLB when
1030 * increasing the permissions on a page.
1032 * Returns non-zero if a spurious fault was handled, zero otherwise.
1034 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1035 * (Optional Invalidation).
1038 spurious_fault(unsigned long error_code, unsigned long address)
1047 * Only writes to RO or instruction fetches from NX may cause
1050 * These could be from user or supervisor accesses but the TLB
1051 * is only lazily flushed after a kernel mapping protection
1052 * change, so user accesses are not expected to cause spurious
1055 if (error_code != (PF_WRITE | PF_PROT)
1056 && error_code != (PF_INSTR | PF_PROT))
1059 pgd = init_mm.pgd + pgd_index(address);
1060 if (!pgd_present(*pgd))
1063 pud = pud_offset(pgd, address);
1064 if (!pud_present(*pud))
1067 if (pud_large(*pud))
1068 return spurious_fault_check(error_code, (pte_t *) pud);
1070 pmd = pmd_offset(pud, address);
1071 if (!pmd_present(*pmd))
1074 if (pmd_large(*pmd))
1075 return spurious_fault_check(error_code, (pte_t *) pmd);
1077 pte = pte_offset_kernel(pmd, address);
1078 if (!pte_present(*pte))
1081 ret = spurious_fault_check(error_code, pte);
1086 * Make sure we have permissions in PMD.
1087 * If not, then there's a bug in the page tables:
1089 ret = spurious_fault_check(error_code, (pte_t *) pmd);
1090 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1094 NOKPROBE_SYMBOL(spurious_fault);
1096 int show_unhandled_signals = 1;
1099 access_error(unsigned long error_code, struct vm_area_struct *vma)
1101 /* This is only called for the current mm, so: */
1102 bool foreign = false;
1104 * Access or read was blocked by protection keys. We do
1105 * this check before any others because we do not want
1106 * to, for instance, confuse a protection-key-denied
1107 * write with one for which we should do a COW.
1109 if (error_code & PF_PK)
1112 * Make sure to check the VMA so that we do not perform
1113 * faults just to hit a PF_PK as soon as we fill in a
1116 if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
1117 (error_code & PF_INSTR), foreign))
1120 if (error_code & PF_WRITE) {
1121 /* write, present and write, not present: */
1122 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1127 /* read, present: */
1128 if (unlikely(error_code & PF_PROT))
1131 /* read, not present: */
1132 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1138 static int fault_in_kernel_space(unsigned long address)
1140 return address >= TASK_SIZE_MAX;
1143 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1145 if (!IS_ENABLED(CONFIG_X86_SMAP))
1148 if (!static_cpu_has(X86_FEATURE_SMAP))
1151 if (error_code & PF_USER)
1154 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1161 * This routine handles page faults. It determines the address,
1162 * and the problem, and then passes it off to one of the appropriate
1165 * This function must have noinline because both callers
1166 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1167 * guarantees there's a function trace entry.
1169 static noinline void
1170 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1171 unsigned long address)
1173 struct vm_area_struct *vma;
1174 struct task_struct *tsk;
1175 struct mm_struct *mm;
1176 int fault, major = 0;
1177 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1183 * Detect and handle instructions that would cause a page fault for
1184 * both a tracked kernel page and a userspace page.
1186 if (kmemcheck_active(regs))
1187 kmemcheck_hide(regs);
1188 prefetchw(&mm->mmap_sem);
1190 if (unlikely(kmmio_fault(regs, address)))
1194 * We fault-in kernel-space virtual memory on-demand. The
1195 * 'reference' page table is init_mm.pgd.
1197 * NOTE! We MUST NOT take any locks for this case. We may
1198 * be in an interrupt or a critical region, and should
1199 * only copy the information from the master page table,
1202 * This verifies that the fault happens in kernel space
1203 * (error_code & 4) == 0, and that the fault was not a
1204 * protection error (error_code & 9) == 0.
1206 if (unlikely(fault_in_kernel_space(address))) {
1207 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1208 if (vmalloc_fault(address) >= 0)
1211 if (kmemcheck_fault(regs, address, error_code))
1215 /* Can handle a stale RO->RW TLB: */
1216 if (spurious_fault(error_code, address))
1219 /* kprobes don't want to hook the spurious faults: */
1220 if (kprobes_fault(regs))
1223 * Don't take the mm semaphore here. If we fixup a prefetch
1224 * fault we could otherwise deadlock:
1226 bad_area_nosemaphore(regs, error_code, address, NULL);
1231 /* kprobes don't want to hook the spurious faults: */
1232 if (unlikely(kprobes_fault(regs)))
1235 if (unlikely(error_code & PF_RSVD))
1236 pgtable_bad(regs, error_code, address);
1238 if (unlikely(smap_violation(error_code, regs))) {
1239 bad_area_nosemaphore(regs, error_code, address, NULL);
1244 * If we're in an interrupt, have no user context or are running
1245 * in a region with pagefaults disabled then we must not take the fault
1247 if (unlikely(faulthandler_disabled() || !mm)) {
1248 bad_area_nosemaphore(regs, error_code, address, NULL);
1253 * It's safe to allow irq's after cr2 has been saved and the
1254 * vmalloc fault has been handled.
1256 * User-mode registers count as a user access even for any
1257 * potential system fault or CPU buglet:
1259 if (user_mode(regs)) {
1261 error_code |= PF_USER;
1262 flags |= FAULT_FLAG_USER;
1264 if (regs->flags & X86_EFLAGS_IF)
1268 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1270 if (error_code & PF_WRITE)
1271 flags |= FAULT_FLAG_WRITE;
1272 if (error_code & PF_INSTR)
1273 flags |= FAULT_FLAG_INSTRUCTION;
1276 * When running in the kernel we expect faults to occur only to
1277 * addresses in user space. All other faults represent errors in
1278 * the kernel and should generate an OOPS. Unfortunately, in the
1279 * case of an erroneous fault occurring in a code path which already
1280 * holds mmap_sem we will deadlock attempting to validate the fault
1281 * against the address space. Luckily the kernel only validly
1282 * references user space from well defined areas of code, which are
1283 * listed in the exceptions table.
1285 * As the vast majority of faults will be valid we will only perform
1286 * the source reference check when there is a possibility of a
1287 * deadlock. Attempt to lock the address space, if we cannot we then
1288 * validate the source. If this is invalid we can skip the address
1289 * space check, thus avoiding the deadlock:
1291 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1292 if ((error_code & PF_USER) == 0 &&
1293 !search_exception_tables(regs->ip)) {
1294 bad_area_nosemaphore(regs, error_code, address, NULL);
1298 down_read(&mm->mmap_sem);
1301 * The above down_read_trylock() might have succeeded in
1302 * which case we'll have missed the might_sleep() from
1308 vma = find_vma(mm, address);
1309 if (unlikely(!vma)) {
1310 bad_area(regs, error_code, address);
1313 if (likely(vma->vm_start <= address))
1315 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1316 bad_area(regs, error_code, address);
1319 if (error_code & PF_USER) {
1321 * Accessing the stack below %sp is always a bug.
1322 * The large cushion allows instructions like enter
1323 * and pusha to work. ("enter $65535, $31" pushes
1324 * 32 pointers and then decrements %sp by 65535.)
1326 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1327 bad_area(regs, error_code, address);
1331 if (unlikely(expand_stack(vma, address))) {
1332 bad_area(regs, error_code, address);
1337 * Ok, we have a good vm_area for this memory access, so
1338 * we can handle it..
1341 if (unlikely(access_error(error_code, vma))) {
1342 bad_area_access_error(regs, error_code, address, vma);
1347 * If for any reason at all we couldn't handle the fault,
1348 * make sure we exit gracefully rather than endlessly redo
1349 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1350 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1352 fault = handle_mm_fault(mm, vma, address, flags);
1353 major |= fault & VM_FAULT_MAJOR;
1356 * If we need to retry the mmap_sem has already been released,
1357 * and if there is a fatal signal pending there is no guarantee
1358 * that we made any progress. Handle this case first.
1360 if (unlikely(fault & VM_FAULT_RETRY)) {
1361 /* Retry at most once */
1362 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1363 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1364 flags |= FAULT_FLAG_TRIED;
1365 if (!fatal_signal_pending(tsk))
1369 /* User mode? Just return to handle the fatal exception */
1370 if (flags & FAULT_FLAG_USER)
1373 /* Not returning to user mode? Handle exceptions or die: */
1374 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1378 up_read(&mm->mmap_sem);
1379 if (unlikely(fault & VM_FAULT_ERROR)) {
1380 mm_fault_error(regs, error_code, address, vma, fault);
1385 * Major/minor page fault accounting. If any of the events
1386 * returned VM_FAULT_MAJOR, we account it as a major fault.
1390 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1393 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1396 check_v8086_mode(regs, address, tsk);
1398 NOKPROBE_SYMBOL(__do_page_fault);
1400 dotraplinkage void notrace
1401 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1403 unsigned long address = read_cr2(); /* Get the faulting address */
1404 enum ctx_state prev_state;
1407 * We must have this function tagged with __kprobes, notrace and call
1408 * read_cr2() before calling anything else. To avoid calling any kind
1409 * of tracing machinery before we've observed the CR2 value.
1411 * exception_{enter,exit}() contain all sorts of tracepoints.
1414 prev_state = exception_enter();
1415 __do_page_fault(regs, error_code, address);
1416 exception_exit(prev_state);
1418 NOKPROBE_SYMBOL(do_page_fault);
1420 #ifdef CONFIG_TRACING
1421 static nokprobe_inline void
1422 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1423 unsigned long error_code)
1425 if (user_mode(regs))
1426 trace_page_fault_user(address, regs, error_code);
1428 trace_page_fault_kernel(address, regs, error_code);
1431 dotraplinkage void notrace
1432 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1435 * The exception_enter and tracepoint processing could
1436 * trigger another page faults (user space callchain
1437 * reading) and destroy the original cr2 value, so read
1438 * the faulting address now.
1440 unsigned long address = read_cr2();
1441 enum ctx_state prev_state;
1443 prev_state = exception_enter();
1444 trace_page_fault_entries(address, regs, error_code);
1445 __do_page_fault(regs, error_code, address);
1446 exception_exit(prev_state);
1448 NOKPROBE_SYMBOL(trace_do_page_fault);
1449 #endif /* CONFIG_TRACING */