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(), ... */
17 #include <asm/traps.h> /* dotraplinkage, ... */
18 #include <asm/pgalloc.h> /* pgd_*(), ... */
19 #include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
20 #include <asm/fixmap.h> /* VSYSCALL_ADDR */
21 #include <asm/vsyscall.h> /* emulate_vsyscall */
23 #define CREATE_TRACE_POINTS
24 #include <asm/trace/exceptions.h>
27 * Page fault error code bits:
29 * bit 0 == 0: no page found 1: protection fault
30 * bit 1 == 0: read access 1: write access
31 * bit 2 == 0: kernel-mode access 1: user-mode access
32 * bit 3 == 1: use of reserved bit detected
33 * bit 4 == 1: fault was an instruction fetch
35 enum x86_pf_error_code {
45 * Returns 0 if mmiotrace is disabled, or if the fault is not
46 * handled by mmiotrace:
48 static nokprobe_inline int
49 kmmio_fault(struct pt_regs *regs, unsigned long addr)
51 if (unlikely(is_kmmio_active()))
52 if (kmmio_handler(regs, addr) == 1)
57 static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
61 /* kprobe_running() needs smp_processor_id() */
62 if (kprobes_built_in() && !user_mode(regs)) {
64 if (kprobe_running() && kprobe_fault_handler(regs, 14))
77 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
78 * Check that here and ignore it.
82 * Sometimes the CPU reports invalid exceptions on prefetch.
83 * Check that here and ignore it.
85 * Opcode checker based on code by Richard Brunner.
88 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
89 unsigned char opcode, int *prefetch)
91 unsigned char instr_hi = opcode & 0xf0;
92 unsigned char instr_lo = opcode & 0x0f;
98 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
99 * In X86_64 long mode, the CPU will signal invalid
100 * opcode if some of these prefixes are present so
101 * X86_64 will never get here anyway
103 return ((instr_lo & 7) == 0x6);
107 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
108 * Need to figure out under what instruction mode the
109 * instruction was issued. Could check the LDT for lm,
110 * but for now it's good enough to assume that long
111 * mode only uses well known segments or kernel.
113 return (!user_mode(regs) || user_64bit_mode(regs));
116 /* 0x64 thru 0x67 are valid prefixes in all modes. */
117 return (instr_lo & 0xC) == 0x4;
119 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
120 return !instr_lo || (instr_lo>>1) == 1;
122 /* Prefetch instruction is 0x0F0D or 0x0F18 */
123 if (probe_kernel_address(instr, opcode))
126 *prefetch = (instr_lo == 0xF) &&
127 (opcode == 0x0D || opcode == 0x18);
135 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
137 unsigned char *max_instr;
138 unsigned char *instr;
142 * If it was a exec (instruction fetch) fault on NX page, then
143 * do not ignore the fault:
145 if (error_code & PF_INSTR)
148 instr = (void *)convert_ip_to_linear(current, regs);
149 max_instr = instr + 15;
151 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
154 while (instr < max_instr) {
155 unsigned char opcode;
157 if (probe_kernel_address(instr, opcode))
162 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
169 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
170 struct task_struct *tsk, int fault)
175 info.si_signo = si_signo;
177 info.si_code = si_code;
178 info.si_addr = (void __user *)address;
179 if (fault & VM_FAULT_HWPOISON_LARGE)
180 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
181 if (fault & VM_FAULT_HWPOISON)
183 info.si_addr_lsb = lsb;
185 force_sig_info(si_signo, &info, tsk);
188 DEFINE_SPINLOCK(pgd_lock);
192 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
194 unsigned index = pgd_index(address);
200 pgd_k = init_mm.pgd + index;
202 if (!pgd_present(*pgd_k))
206 * set_pgd(pgd, *pgd_k); here would be useless on PAE
207 * and redundant with the set_pmd() on non-PAE. As would
210 pud = pud_offset(pgd, address);
211 pud_k = pud_offset(pgd_k, address);
212 if (!pud_present(*pud_k))
215 pmd = pmd_offset(pud, address);
216 pmd_k = pmd_offset(pud_k, address);
217 if (!pmd_present(*pmd_k))
220 if (!pmd_present(*pmd))
221 set_pmd(pmd, *pmd_k);
223 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
228 void vmalloc_sync_all(void)
230 unsigned long address;
232 if (SHARED_KERNEL_PMD)
235 for (address = VMALLOC_START & PMD_MASK;
236 address >= TASK_SIZE && address < FIXADDR_TOP;
237 address += PMD_SIZE) {
240 spin_lock(&pgd_lock);
241 list_for_each_entry(page, &pgd_list, lru) {
242 spinlock_t *pgt_lock;
245 /* the pgt_lock only for Xen */
246 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
249 ret = vmalloc_sync_one(page_address(page), address);
250 spin_unlock(pgt_lock);
255 spin_unlock(&pgd_lock);
262 * Handle a fault on the vmalloc or module mapping area
264 static noinline int vmalloc_fault(unsigned long address)
266 unsigned long pgd_paddr;
270 /* Make sure we are in vmalloc area: */
271 if (!(address >= VMALLOC_START && address < VMALLOC_END))
274 WARN_ON_ONCE(in_nmi());
277 * Synchronize this task's top level page-table
278 * with the 'reference' page table.
280 * Do _not_ use "current" here. We might be inside
281 * an interrupt in the middle of a task switch..
283 pgd_paddr = read_cr3();
284 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
288 pte_k = pte_offset_kernel(pmd_k, address);
289 if (!pte_present(*pte_k))
294 NOKPROBE_SYMBOL(vmalloc_fault);
297 * Did it hit the DOS screen memory VA from vm86 mode?
300 check_v8086_mode(struct pt_regs *regs, unsigned long address,
301 struct task_struct *tsk)
305 if (!v8086_mode(regs))
308 bit = (address - 0xA0000) >> PAGE_SHIFT;
310 tsk->thread.screen_bitmap |= 1 << bit;
313 static bool low_pfn(unsigned long pfn)
315 return pfn < max_low_pfn;
318 static void dump_pagetable(unsigned long address)
320 pgd_t *base = __va(read_cr3());
321 pgd_t *pgd = &base[pgd_index(address)];
325 #ifdef CONFIG_X86_PAE
326 printk("*pdpt = %016Lx ", pgd_val(*pgd));
327 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
330 pmd = pmd_offset(pud_offset(pgd, address), address);
331 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
334 * We must not directly access the pte in the highpte
335 * case if the page table is located in highmem.
336 * And let's rather not kmap-atomic the pte, just in case
337 * it's allocated already:
339 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
342 pte = pte_offset_kernel(pmd, address);
343 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
348 #else /* CONFIG_X86_64: */
350 void vmalloc_sync_all(void)
352 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END, 0);
358 * Handle a fault on the vmalloc area
360 * This assumes no large pages in there.
362 static noinline int vmalloc_fault(unsigned long address)
364 pgd_t *pgd, *pgd_ref;
365 pud_t *pud, *pud_ref;
366 pmd_t *pmd, *pmd_ref;
367 pte_t *pte, *pte_ref;
369 /* Make sure we are in vmalloc area: */
370 if (!(address >= VMALLOC_START && address < VMALLOC_END))
373 WARN_ON_ONCE(in_nmi());
376 * Copy kernel mappings over when needed. This can also
377 * happen within a race in page table update. In the later
380 pgd = pgd_offset(current->active_mm, address);
381 pgd_ref = pgd_offset_k(address);
382 if (pgd_none(*pgd_ref))
385 if (pgd_none(*pgd)) {
386 set_pgd(pgd, *pgd_ref);
387 arch_flush_lazy_mmu_mode();
389 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
393 * Below here mismatches are bugs because these lower tables
397 pud = pud_offset(pgd, address);
398 pud_ref = pud_offset(pgd_ref, address);
399 if (pud_none(*pud_ref))
402 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
405 pmd = pmd_offset(pud, address);
406 pmd_ref = pmd_offset(pud_ref, address);
407 if (pmd_none(*pmd_ref))
410 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
413 pte_ref = pte_offset_kernel(pmd_ref, address);
414 if (!pte_present(*pte_ref))
417 pte = pte_offset_kernel(pmd, address);
420 * Don't use pte_page here, because the mappings can point
421 * outside mem_map, and the NUMA hash lookup cannot handle
424 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
429 NOKPROBE_SYMBOL(vmalloc_fault);
431 #ifdef CONFIG_CPU_SUP_AMD
432 static const char errata93_warning[] =
434 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
435 "******* Working around it, but it may cause SEGVs or burn power.\n"
436 "******* Please consider a BIOS update.\n"
437 "******* Disabling USB legacy in the BIOS may also help.\n";
441 * No vm86 mode in 64-bit mode:
444 check_v8086_mode(struct pt_regs *regs, unsigned long address,
445 struct task_struct *tsk)
449 static int bad_address(void *p)
453 return probe_kernel_address((unsigned long *)p, dummy);
456 static void dump_pagetable(unsigned long address)
458 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
459 pgd_t *pgd = base + pgd_index(address);
464 if (bad_address(pgd))
467 printk("PGD %lx ", pgd_val(*pgd));
469 if (!pgd_present(*pgd))
472 pud = pud_offset(pgd, address);
473 if (bad_address(pud))
476 printk("PUD %lx ", pud_val(*pud));
477 if (!pud_present(*pud) || pud_large(*pud))
480 pmd = pmd_offset(pud, address);
481 if (bad_address(pmd))
484 printk("PMD %lx ", pmd_val(*pmd));
485 if (!pmd_present(*pmd) || pmd_large(*pmd))
488 pte = pte_offset_kernel(pmd, address);
489 if (bad_address(pte))
492 printk("PTE %lx", pte_val(*pte));
500 #endif /* CONFIG_X86_64 */
503 * Workaround for K8 erratum #93 & buggy BIOS.
505 * BIOS SMM functions are required to use a specific workaround
506 * to avoid corruption of the 64bit RIP register on C stepping K8.
508 * A lot of BIOS that didn't get tested properly miss this.
510 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
511 * Try to work around it here.
513 * Note we only handle faults in kernel here.
514 * Does nothing on 32-bit.
516 static int is_errata93(struct pt_regs *regs, unsigned long address)
518 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
519 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
520 || boot_cpu_data.x86 != 0xf)
523 if (address != regs->ip)
526 if ((address >> 32) != 0)
529 address |= 0xffffffffUL << 32;
530 if ((address >= (u64)_stext && address <= (u64)_etext) ||
531 (address >= MODULES_VADDR && address <= MODULES_END)) {
532 printk_once(errata93_warning);
541 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
542 * to illegal addresses >4GB.
544 * We catch this in the page fault handler because these addresses
545 * are not reachable. Just detect this case and return. Any code
546 * segment in LDT is compatibility mode.
548 static int is_errata100(struct pt_regs *regs, unsigned long address)
551 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
557 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
559 #ifdef CONFIG_X86_F00F_BUG
563 * Pentium F0 0F C7 C8 bug workaround:
565 if (boot_cpu_has_bug(X86_BUG_F00F)) {
566 nr = (address - idt_descr.address) >> 3;
569 do_invalid_op(regs, 0);
577 static const char nx_warning[] = KERN_CRIT
578 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
579 static const char smep_warning[] = KERN_CRIT
580 "unable to execute userspace code (SMEP?) (uid: %d)\n";
583 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
584 unsigned long address)
586 if (!oops_may_print())
589 if (error_code & PF_INSTR) {
594 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
595 pgd += pgd_index(address);
597 pte = lookup_address_in_pgd(pgd, address, &level);
599 if (pte && pte_present(*pte) && !pte_exec(*pte))
600 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
601 if (pte && pte_present(*pte) && pte_exec(*pte) &&
602 (pgd_flags(*pgd) & _PAGE_USER) &&
603 (__read_cr4() & X86_CR4_SMEP))
604 printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
607 printk(KERN_ALERT "BUG: unable to handle kernel ");
608 if (address < PAGE_SIZE)
609 printk(KERN_CONT "NULL pointer dereference");
611 printk(KERN_CONT "paging request");
613 printk(KERN_CONT " at %p\n", (void *) address);
614 printk(KERN_ALERT "IP:");
615 printk_address(regs->ip);
617 dump_pagetable(address);
621 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
622 unsigned long address)
624 struct task_struct *tsk;
628 flags = oops_begin();
632 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
634 dump_pagetable(address);
636 tsk->thread.cr2 = address;
637 tsk->thread.trap_nr = X86_TRAP_PF;
638 tsk->thread.error_code = error_code;
640 if (__die("Bad pagetable", regs, error_code))
643 oops_end(flags, regs, sig);
647 no_context(struct pt_regs *regs, unsigned long error_code,
648 unsigned long address, int signal, int si_code)
650 struct task_struct *tsk = current;
654 /* Are we prepared to handle this kernel fault? */
655 if (fixup_exception(regs)) {
657 * Any interrupt that takes a fault gets the fixup. This makes
658 * the below recursive fault logic only apply to a faults from
665 * Per the above we're !in_interrupt(), aka. task context.
667 * In this case we need to make sure we're not recursively
668 * faulting through the emulate_vsyscall() logic.
670 if (current_thread_info()->sig_on_uaccess_error && signal) {
671 tsk->thread.trap_nr = X86_TRAP_PF;
672 tsk->thread.error_code = error_code | PF_USER;
673 tsk->thread.cr2 = address;
675 /* XXX: hwpoison faults will set the wrong code. */
676 force_sig_info_fault(signal, si_code, address, tsk, 0);
680 * Barring that, we can do the fixup and be happy.
688 * Valid to do another page fault here, because if this fault
689 * had been triggered by is_prefetch fixup_exception would have
694 * Hall of shame of CPU/BIOS bugs.
696 if (is_prefetch(regs, error_code, address))
699 if (is_errata93(regs, address))
703 * Oops. The kernel tried to access some bad page. We'll have to
704 * terminate things with extreme prejudice:
706 flags = oops_begin();
708 show_fault_oops(regs, error_code, address);
710 if (task_stack_end_corrupted(tsk))
711 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
713 tsk->thread.cr2 = address;
714 tsk->thread.trap_nr = X86_TRAP_PF;
715 tsk->thread.error_code = error_code;
718 if (__die("Oops", regs, error_code))
721 /* Executive summary in case the body of the oops scrolled away */
722 printk(KERN_DEFAULT "CR2: %016lx\n", address);
724 oops_end(flags, regs, sig);
728 * Print out info about fatal segfaults, if the show_unhandled_signals
732 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
733 unsigned long address, struct task_struct *tsk)
735 if (!unhandled_signal(tsk, SIGSEGV))
738 if (!printk_ratelimit())
741 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
742 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
743 tsk->comm, task_pid_nr(tsk), address,
744 (void *)regs->ip, (void *)regs->sp, error_code);
746 print_vma_addr(KERN_CONT " in ", regs->ip);
748 printk(KERN_CONT "\n");
752 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
753 unsigned long address, int si_code)
755 struct task_struct *tsk = current;
757 /* User mode accesses just cause a SIGSEGV */
758 if (error_code & PF_USER) {
760 * It's possible to have interrupts off here:
765 * Valid to do another page fault here because this one came
768 if (is_prefetch(regs, error_code, address))
771 if (is_errata100(regs, address))
776 * Instruction fetch faults in the vsyscall page might need
779 if (unlikely((error_code & PF_INSTR) &&
780 ((address & ~0xfff) == VSYSCALL_ADDR))) {
781 if (emulate_vsyscall(regs, address))
785 /* Kernel addresses are always protection faults: */
786 if (address >= TASK_SIZE)
787 error_code |= PF_PROT;
789 if (likely(show_unhandled_signals))
790 show_signal_msg(regs, error_code, address, tsk);
792 tsk->thread.cr2 = address;
793 tsk->thread.error_code = error_code;
794 tsk->thread.trap_nr = X86_TRAP_PF;
796 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
801 if (is_f00f_bug(regs, address))
804 no_context(regs, error_code, address, SIGSEGV, si_code);
808 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
809 unsigned long address)
811 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
815 __bad_area(struct pt_regs *regs, unsigned long error_code,
816 unsigned long address, int si_code)
818 struct mm_struct *mm = current->mm;
821 * Something tried to access memory that isn't in our memory map..
822 * Fix it, but check if it's kernel or user first..
824 up_read(&mm->mmap_sem);
826 __bad_area_nosemaphore(regs, error_code, address, si_code);
830 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
832 __bad_area(regs, error_code, address, SEGV_MAPERR);
836 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
837 unsigned long address)
839 __bad_area(regs, error_code, address, SEGV_ACCERR);
843 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
846 struct task_struct *tsk = current;
847 int code = BUS_ADRERR;
849 /* Kernel mode? Handle exceptions or die: */
850 if (!(error_code & PF_USER)) {
851 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
855 /* User-space => ok to do another page fault: */
856 if (is_prefetch(regs, error_code, address))
859 tsk->thread.cr2 = address;
860 tsk->thread.error_code = error_code;
861 tsk->thread.trap_nr = X86_TRAP_PF;
863 #ifdef CONFIG_MEMORY_FAILURE
864 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
866 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
867 tsk->comm, tsk->pid, address);
868 code = BUS_MCEERR_AR;
871 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
875 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
876 unsigned long address, unsigned int fault)
878 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
879 no_context(regs, error_code, address, 0, 0);
883 if (fault & VM_FAULT_OOM) {
884 /* Kernel mode? Handle exceptions or die: */
885 if (!(error_code & PF_USER)) {
886 no_context(regs, error_code, address,
887 SIGSEGV, SEGV_MAPERR);
892 * We ran out of memory, call the OOM killer, and return the
893 * userspace (which will retry the fault, or kill us if we got
896 pagefault_out_of_memory();
898 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
899 VM_FAULT_HWPOISON_LARGE))
900 do_sigbus(regs, error_code, address, fault);
901 else if (fault & VM_FAULT_SIGSEGV)
902 bad_area_nosemaphore(regs, error_code, address);
908 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
910 if ((error_code & PF_WRITE) && !pte_write(*pte))
913 if ((error_code & PF_INSTR) && !pte_exec(*pte))
920 * Handle a spurious fault caused by a stale TLB entry.
922 * This allows us to lazily refresh the TLB when increasing the
923 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
924 * eagerly is very expensive since that implies doing a full
925 * cross-processor TLB flush, even if no stale TLB entries exist
926 * on other processors.
928 * Spurious faults may only occur if the TLB contains an entry with
929 * fewer permission than the page table entry. Non-present (P = 0)
930 * and reserved bit (R = 1) faults are never spurious.
932 * There are no security implications to leaving a stale TLB when
933 * increasing the permissions on a page.
935 * Returns non-zero if a spurious fault was handled, zero otherwise.
937 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
938 * (Optional Invalidation).
941 spurious_fault(unsigned long error_code, unsigned long address)
950 * Only writes to RO or instruction fetches from NX may cause
953 * These could be from user or supervisor accesses but the TLB
954 * is only lazily flushed after a kernel mapping protection
955 * change, so user accesses are not expected to cause spurious
958 if (error_code != (PF_WRITE | PF_PROT)
959 && error_code != (PF_INSTR | PF_PROT))
962 pgd = init_mm.pgd + pgd_index(address);
963 if (!pgd_present(*pgd))
966 pud = pud_offset(pgd, address);
967 if (!pud_present(*pud))
971 return spurious_fault_check(error_code, (pte_t *) pud);
973 pmd = pmd_offset(pud, address);
974 if (!pmd_present(*pmd))
978 return spurious_fault_check(error_code, (pte_t *) pmd);
980 pte = pte_offset_kernel(pmd, address);
981 if (!pte_present(*pte))
984 ret = spurious_fault_check(error_code, pte);
989 * Make sure we have permissions in PMD.
990 * If not, then there's a bug in the page tables:
992 ret = spurious_fault_check(error_code, (pte_t *) pmd);
993 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
997 NOKPROBE_SYMBOL(spurious_fault);
999 int show_unhandled_signals = 1;
1002 access_error(unsigned long error_code, struct vm_area_struct *vma)
1004 if (error_code & PF_WRITE) {
1005 /* write, present and write, not present: */
1006 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1011 /* read, present: */
1012 if (unlikely(error_code & PF_PROT))
1015 /* read, not present: */
1016 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1022 static int fault_in_kernel_space(unsigned long address)
1024 return address >= TASK_SIZE_MAX;
1027 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1029 if (!IS_ENABLED(CONFIG_X86_SMAP))
1032 if (!static_cpu_has(X86_FEATURE_SMAP))
1035 if (error_code & PF_USER)
1038 if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1045 * This routine handles page faults. It determines the address,
1046 * and the problem, and then passes it off to one of the appropriate
1049 * This function must have noinline because both callers
1050 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1051 * guarantees there's a function trace entry.
1053 static noinline void
1054 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1055 unsigned long address)
1057 struct vm_area_struct *vma;
1058 struct task_struct *tsk;
1059 struct mm_struct *mm;
1060 int fault, major = 0;
1061 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1067 * Detect and handle instructions that would cause a page fault for
1068 * both a tracked kernel page and a userspace page.
1070 if (kmemcheck_active(regs))
1071 kmemcheck_hide(regs);
1072 prefetchw(&mm->mmap_sem);
1074 if (unlikely(kmmio_fault(regs, address)))
1078 * We fault-in kernel-space virtual memory on-demand. The
1079 * 'reference' page table is init_mm.pgd.
1081 * NOTE! We MUST NOT take any locks for this case. We may
1082 * be in an interrupt or a critical region, and should
1083 * only copy the information from the master page table,
1086 * This verifies that the fault happens in kernel space
1087 * (error_code & 4) == 0, and that the fault was not a
1088 * protection error (error_code & 9) == 0.
1090 if (unlikely(fault_in_kernel_space(address))) {
1091 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1092 if (vmalloc_fault(address) >= 0)
1095 if (kmemcheck_fault(regs, address, error_code))
1099 /* Can handle a stale RO->RW TLB: */
1100 if (spurious_fault(error_code, address))
1103 /* kprobes don't want to hook the spurious faults: */
1104 if (kprobes_fault(regs))
1107 * Don't take the mm semaphore here. If we fixup a prefetch
1108 * fault we could otherwise deadlock:
1110 bad_area_nosemaphore(regs, error_code, address);
1115 /* kprobes don't want to hook the spurious faults: */
1116 if (unlikely(kprobes_fault(regs)))
1119 if (unlikely(error_code & PF_RSVD))
1120 pgtable_bad(regs, error_code, address);
1122 if (unlikely(smap_violation(error_code, regs))) {
1123 bad_area_nosemaphore(regs, error_code, address);
1128 * If we're in an interrupt, have no user context or are running
1129 * in an atomic region then we must not take the fault:
1131 if (unlikely(in_atomic() || !mm)) {
1132 bad_area_nosemaphore(regs, error_code, address);
1137 * It's safe to allow irq's after cr2 has been saved and the
1138 * vmalloc fault has been handled.
1140 * User-mode registers count as a user access even for any
1141 * potential system fault or CPU buglet:
1143 if (user_mode(regs)) {
1145 error_code |= PF_USER;
1146 flags |= FAULT_FLAG_USER;
1148 if (regs->flags & X86_EFLAGS_IF)
1152 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1154 if (error_code & PF_WRITE)
1155 flags |= FAULT_FLAG_WRITE;
1158 * When running in the kernel we expect faults to occur only to
1159 * addresses in user space. All other faults represent errors in
1160 * the kernel and should generate an OOPS. Unfortunately, in the
1161 * case of an erroneous fault occurring in a code path which already
1162 * holds mmap_sem we will deadlock attempting to validate the fault
1163 * against the address space. Luckily the kernel only validly
1164 * references user space from well defined areas of code, which are
1165 * listed in the exceptions table.
1167 * As the vast majority of faults will be valid we will only perform
1168 * the source reference check when there is a possibility of a
1169 * deadlock. Attempt to lock the address space, if we cannot we then
1170 * validate the source. If this is invalid we can skip the address
1171 * space check, thus avoiding the deadlock:
1173 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1174 if ((error_code & PF_USER) == 0 &&
1175 !search_exception_tables(regs->ip)) {
1176 bad_area_nosemaphore(regs, error_code, address);
1180 down_read(&mm->mmap_sem);
1183 * The above down_read_trylock() might have succeeded in
1184 * which case we'll have missed the might_sleep() from
1190 vma = find_vma(mm, address);
1191 if (unlikely(!vma)) {
1192 bad_area(regs, error_code, address);
1195 if (likely(vma->vm_start <= address))
1197 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1198 bad_area(regs, error_code, address);
1201 if (error_code & PF_USER) {
1203 * Accessing the stack below %sp is always a bug.
1204 * The large cushion allows instructions like enter
1205 * and pusha to work. ("enter $65535, $31" pushes
1206 * 32 pointers and then decrements %sp by 65535.)
1208 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1209 bad_area(regs, error_code, address);
1213 if (unlikely(expand_stack(vma, address))) {
1214 bad_area(regs, error_code, address);
1219 * Ok, we have a good vm_area for this memory access, so
1220 * we can handle it..
1223 if (unlikely(access_error(error_code, vma))) {
1224 bad_area_access_error(regs, error_code, address);
1229 * If for any reason at all we couldn't handle the fault,
1230 * make sure we exit gracefully rather than endlessly redo
1231 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1232 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1234 fault = handle_mm_fault(mm, vma, address, flags);
1235 major |= fault & VM_FAULT_MAJOR;
1238 * If we need to retry the mmap_sem has already been released,
1239 * and if there is a fatal signal pending there is no guarantee
1240 * that we made any progress. Handle this case first.
1242 if (unlikely(fault & VM_FAULT_RETRY)) {
1243 /* Retry at most once */
1244 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1245 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1246 flags |= FAULT_FLAG_TRIED;
1247 if (!fatal_signal_pending(tsk))
1251 /* User mode? Just return to handle the fatal exception */
1252 if (flags & FAULT_FLAG_USER)
1255 /* Not returning to user mode? Handle exceptions or die: */
1256 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1260 up_read(&mm->mmap_sem);
1261 if (unlikely(fault & VM_FAULT_ERROR)) {
1262 mm_fault_error(regs, error_code, address, fault);
1267 * Major/minor page fault accounting. If any of the events
1268 * returned VM_FAULT_MAJOR, we account it as a major fault.
1272 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1275 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1278 check_v8086_mode(regs, address, tsk);
1280 NOKPROBE_SYMBOL(__do_page_fault);
1282 dotraplinkage void notrace
1283 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1285 unsigned long address = read_cr2(); /* Get the faulting address */
1286 enum ctx_state prev_state;
1289 * We must have this function tagged with __kprobes, notrace and call
1290 * read_cr2() before calling anything else. To avoid calling any kind
1291 * of tracing machinery before we've observed the CR2 value.
1293 * exception_{enter,exit}() contain all sorts of tracepoints.
1296 prev_state = exception_enter();
1297 __do_page_fault(regs, error_code, address);
1298 exception_exit(prev_state);
1300 NOKPROBE_SYMBOL(do_page_fault);
1302 #ifdef CONFIG_TRACING
1303 static nokprobe_inline void
1304 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1305 unsigned long error_code)
1307 if (user_mode(regs))
1308 trace_page_fault_user(address, regs, error_code);
1310 trace_page_fault_kernel(address, regs, error_code);
1313 dotraplinkage void notrace
1314 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1317 * The exception_enter and tracepoint processing could
1318 * trigger another page faults (user space callchain
1319 * reading) and destroy the original cr2 value, so read
1320 * the faulting address now.
1322 unsigned long address = read_cr2();
1323 enum ctx_state prev_state;
1325 prev_state = exception_enter();
1326 trace_page_fault_entries(address, regs, error_code);
1327 __do_page_fault(regs, error_code, address);
1328 exception_exit(prev_state);
1330 NOKPROBE_SYMBOL(trace_do_page_fault);
1331 #endif /* CONFIG_TRACING */