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/magic.h> /* STACK_END_MAGIC */
7 #include <linux/sched.h> /* test_thread_flag(), ... */
8 #include <linux/kdebug.h> /* oops_begin/end, ... */
9 #include <linux/module.h> /* search_exception_table */
10 #include <linux/bootmem.h> /* max_low_pfn */
11 #include <linux/kprobes.h> /* __kprobes, ... */
12 #include <linux/mmiotrace.h> /* kmmio_handler, ... */
13 #include <linux/perf_event.h> /* perf_sw_event */
15 #include <asm/traps.h> /* dotraplinkage, ... */
16 #include <asm/pgalloc.h> /* pgd_*(), ... */
17 #include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
20 * Page fault error code bits:
22 * bit 0 == 0: no page found 1: protection fault
23 * bit 1 == 0: read access 1: write access
24 * bit 2 == 0: kernel-mode access 1: user-mode access
25 * bit 3 == 1: use of reserved bit detected
26 * bit 4 == 1: fault was an instruction fetch
28 enum x86_pf_error_code {
38 * Returns 0 if mmiotrace is disabled, or if the fault is not
39 * handled by mmiotrace:
41 static inline int __kprobes
42 kmmio_fault(struct pt_regs *regs, unsigned long addr)
44 if (unlikely(is_kmmio_active()))
45 if (kmmio_handler(regs, addr) == 1)
50 static inline int __kprobes notify_page_fault(struct pt_regs *regs)
54 /* kprobe_running() needs smp_processor_id() */
55 if (kprobes_built_in() && !user_mode_vm(regs)) {
57 if (kprobe_running() && kprobe_fault_handler(regs, 14))
70 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
71 * Check that here and ignore it.
75 * Sometimes the CPU reports invalid exceptions on prefetch.
76 * Check that here and ignore it.
78 * Opcode checker based on code by Richard Brunner.
81 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
82 unsigned char opcode, int *prefetch)
84 unsigned char instr_hi = opcode & 0xf0;
85 unsigned char instr_lo = opcode & 0x0f;
91 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
92 * In X86_64 long mode, the CPU will signal invalid
93 * opcode if some of these prefixes are present so
94 * X86_64 will never get here anyway
96 return ((instr_lo & 7) == 0x6);
100 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
101 * Need to figure out under what instruction mode the
102 * instruction was issued. Could check the LDT for lm,
103 * but for now it's good enough to assume that long
104 * mode only uses well known segments or kernel.
106 return (!user_mode(regs)) || (regs->cs == __USER_CS);
109 /* 0x64 thru 0x67 are valid prefixes in all modes. */
110 return (instr_lo & 0xC) == 0x4;
112 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
113 return !instr_lo || (instr_lo>>1) == 1;
115 /* Prefetch instruction is 0x0F0D or 0x0F18 */
116 if (probe_kernel_address(instr, opcode))
119 *prefetch = (instr_lo == 0xF) &&
120 (opcode == 0x0D || opcode == 0x18);
128 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
130 unsigned char *max_instr;
131 unsigned char *instr;
135 * If it was a exec (instruction fetch) fault on NX page, then
136 * do not ignore the fault:
138 if (error_code & PF_INSTR)
141 instr = (void *)convert_ip_to_linear(current, regs);
142 max_instr = instr + 15;
144 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
147 while (instr < max_instr) {
148 unsigned char opcode;
150 if (probe_kernel_address(instr, opcode))
155 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
162 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
163 struct task_struct *tsk)
167 info.si_signo = si_signo;
169 info.si_code = si_code;
170 info.si_addr = (void __user *)address;
171 info.si_addr_lsb = si_code == BUS_MCEERR_AR ? PAGE_SHIFT : 0;
173 force_sig_info(si_signo, &info, tsk);
176 DEFINE_SPINLOCK(pgd_lock);
180 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
182 unsigned index = pgd_index(address);
188 pgd_k = init_mm.pgd + index;
190 if (!pgd_present(*pgd_k))
194 * set_pgd(pgd, *pgd_k); here would be useless on PAE
195 * and redundant with the set_pmd() on non-PAE. As would
198 pud = pud_offset(pgd, address);
199 pud_k = pud_offset(pgd_k, address);
200 if (!pud_present(*pud_k))
203 pmd = pmd_offset(pud, address);
204 pmd_k = pmd_offset(pud_k, address);
205 if (!pmd_present(*pmd_k))
208 if (!pmd_present(*pmd))
209 set_pmd(pmd, *pmd_k);
211 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
216 void vmalloc_sync_all(void)
218 unsigned long address;
220 if (SHARED_KERNEL_PMD)
223 for (address = VMALLOC_START & PMD_MASK;
224 address >= TASK_SIZE && address < FIXADDR_TOP;
225 address += PMD_SIZE) {
230 spin_lock_irqsave(&pgd_lock, flags);
231 list_for_each_entry(page, &pgd_list, lru) {
232 if (!vmalloc_sync_one(page_address(page), address))
235 spin_unlock_irqrestore(&pgd_lock, flags);
242 * Handle a fault on the vmalloc or module mapping area
244 static noinline __kprobes int vmalloc_fault(unsigned long address)
246 unsigned long pgd_paddr;
250 /* Make sure we are in vmalloc area: */
251 if (!(address >= VMALLOC_START && address < VMALLOC_END))
255 * Synchronize this task's top level page-table
256 * with the 'reference' page table.
258 * Do _not_ use "current" here. We might be inside
259 * an interrupt in the middle of a task switch..
261 pgd_paddr = read_cr3();
262 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
266 pte_k = pte_offset_kernel(pmd_k, address);
267 if (!pte_present(*pte_k))
274 * Did it hit the DOS screen memory VA from vm86 mode?
277 check_v8086_mode(struct pt_regs *regs, unsigned long address,
278 struct task_struct *tsk)
282 if (!v8086_mode(regs))
285 bit = (address - 0xA0000) >> PAGE_SHIFT;
287 tsk->thread.screen_bitmap |= 1 << bit;
290 static bool low_pfn(unsigned long pfn)
292 return pfn < max_low_pfn;
295 static void dump_pagetable(unsigned long address)
297 pgd_t *base = __va(read_cr3());
298 pgd_t *pgd = &base[pgd_index(address)];
302 #ifdef CONFIG_X86_PAE
303 printk("*pdpt = %016Lx ", pgd_val(*pgd));
304 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
307 pmd = pmd_offset(pud_offset(pgd, address), address);
308 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
311 * We must not directly access the pte in the highpte
312 * case if the page table is located in highmem.
313 * And let's rather not kmap-atomic the pte, just in case
314 * it's allocated already:
316 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
319 pte = pte_offset_kernel(pmd, address);
320 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
325 #else /* CONFIG_X86_64: */
327 void vmalloc_sync_all(void)
329 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
335 * Handle a fault on the vmalloc area
337 * This assumes no large pages in there.
339 static noinline __kprobes int vmalloc_fault(unsigned long address)
341 pgd_t *pgd, *pgd_ref;
342 pud_t *pud, *pud_ref;
343 pmd_t *pmd, *pmd_ref;
344 pte_t *pte, *pte_ref;
346 /* Make sure we are in vmalloc area: */
347 if (!(address >= VMALLOC_START && address < VMALLOC_END))
351 * Copy kernel mappings over when needed. This can also
352 * happen within a race in page table update. In the later
355 pgd = pgd_offset(current->active_mm, address);
356 pgd_ref = pgd_offset_k(address);
357 if (pgd_none(*pgd_ref))
361 set_pgd(pgd, *pgd_ref);
363 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
366 * Below here mismatches are bugs because these lower tables
370 pud = pud_offset(pgd, address);
371 pud_ref = pud_offset(pgd_ref, address);
372 if (pud_none(*pud_ref))
375 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
378 pmd = pmd_offset(pud, address);
379 pmd_ref = pmd_offset(pud_ref, address);
380 if (pmd_none(*pmd_ref))
383 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
386 pte_ref = pte_offset_kernel(pmd_ref, address);
387 if (!pte_present(*pte_ref))
390 pte = pte_offset_kernel(pmd, address);
393 * Don't use pte_page here, because the mappings can point
394 * outside mem_map, and the NUMA hash lookup cannot handle
397 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
403 static const char errata93_warning[] =
405 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
406 "******* Working around it, but it may cause SEGVs or burn power.\n"
407 "******* Please consider a BIOS update.\n"
408 "******* Disabling USB legacy in the BIOS may also help.\n";
411 * No vm86 mode in 64-bit mode:
414 check_v8086_mode(struct pt_regs *regs, unsigned long address,
415 struct task_struct *tsk)
419 static int bad_address(void *p)
423 return probe_kernel_address((unsigned long *)p, dummy);
426 static void dump_pagetable(unsigned long address)
428 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
429 pgd_t *pgd = base + pgd_index(address);
434 if (bad_address(pgd))
437 printk("PGD %lx ", pgd_val(*pgd));
439 if (!pgd_present(*pgd))
442 pud = pud_offset(pgd, address);
443 if (bad_address(pud))
446 printk("PUD %lx ", pud_val(*pud));
447 if (!pud_present(*pud) || pud_large(*pud))
450 pmd = pmd_offset(pud, address);
451 if (bad_address(pmd))
454 printk("PMD %lx ", pmd_val(*pmd));
455 if (!pmd_present(*pmd) || pmd_large(*pmd))
458 pte = pte_offset_kernel(pmd, address);
459 if (bad_address(pte))
462 printk("PTE %lx", pte_val(*pte));
470 #endif /* CONFIG_X86_64 */
473 * Workaround for K8 erratum #93 & buggy BIOS.
475 * BIOS SMM functions are required to use a specific workaround
476 * to avoid corruption of the 64bit RIP register on C stepping K8.
478 * A lot of BIOS that didn't get tested properly miss this.
480 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
481 * Try to work around it here.
483 * Note we only handle faults in kernel here.
484 * Does nothing on 32-bit.
486 static int is_errata93(struct pt_regs *regs, unsigned long address)
489 if (address != regs->ip)
492 if ((address >> 32) != 0)
495 address |= 0xffffffffUL << 32;
496 if ((address >= (u64)_stext && address <= (u64)_etext) ||
497 (address >= MODULES_VADDR && address <= MODULES_END)) {
498 printk_once(errata93_warning);
507 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
508 * to illegal addresses >4GB.
510 * We catch this in the page fault handler because these addresses
511 * are not reachable. Just detect this case and return. Any code
512 * segment in LDT is compatibility mode.
514 static int is_errata100(struct pt_regs *regs, unsigned long address)
517 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
523 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
525 #ifdef CONFIG_X86_F00F_BUG
529 * Pentium F0 0F C7 C8 bug workaround:
531 if (boot_cpu_data.f00f_bug) {
532 nr = (address - idt_descr.address) >> 3;
535 do_invalid_op(regs, 0);
543 static const char nx_warning[] = KERN_CRIT
544 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
547 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
548 unsigned long address)
550 if (!oops_may_print())
553 if (error_code & PF_INSTR) {
556 pte_t *pte = lookup_address(address, &level);
558 if (pte && pte_present(*pte) && !pte_exec(*pte))
559 printk(nx_warning, current_uid());
562 printk(KERN_ALERT "BUG: unable to handle kernel ");
563 if (address < PAGE_SIZE)
564 printk(KERN_CONT "NULL pointer dereference");
566 printk(KERN_CONT "paging request");
568 printk(KERN_CONT " at %p\n", (void *) address);
569 printk(KERN_ALERT "IP:");
570 printk_address(regs->ip, 1);
572 dump_pagetable(address);
576 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
577 unsigned long address)
579 struct task_struct *tsk;
583 flags = oops_begin();
587 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
589 dump_pagetable(address);
591 tsk->thread.cr2 = address;
592 tsk->thread.trap_no = 14;
593 tsk->thread.error_code = error_code;
595 if (__die("Bad pagetable", regs, error_code))
598 oops_end(flags, regs, sig);
602 no_context(struct pt_regs *regs, unsigned long error_code,
603 unsigned long address)
605 struct task_struct *tsk = current;
606 unsigned long *stackend;
610 /* Are we prepared to handle this kernel fault? */
611 if (fixup_exception(regs))
617 * Valid to do another page fault here, because if this fault
618 * had been triggered by is_prefetch fixup_exception would have
623 * Hall of shame of CPU/BIOS bugs.
625 if (is_prefetch(regs, error_code, address))
628 if (is_errata93(regs, address))
632 * Oops. The kernel tried to access some bad page. We'll have to
633 * terminate things with extreme prejudice:
635 flags = oops_begin();
637 show_fault_oops(regs, error_code, address);
639 stackend = end_of_stack(tsk);
640 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
641 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
643 tsk->thread.cr2 = address;
644 tsk->thread.trap_no = 14;
645 tsk->thread.error_code = error_code;
648 if (__die("Oops", regs, error_code))
651 /* Executive summary in case the body of the oops scrolled away */
652 printk(KERN_EMERG "CR2: %016lx\n", address);
654 oops_end(flags, regs, sig);
658 * Print out info about fatal segfaults, if the show_unhandled_signals
662 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
663 unsigned long address, struct task_struct *tsk)
665 if (!unhandled_signal(tsk, SIGSEGV))
668 if (!printk_ratelimit())
671 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
672 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
673 tsk->comm, task_pid_nr(tsk), address,
674 (void *)regs->ip, (void *)regs->sp, error_code);
676 print_vma_addr(KERN_CONT " in ", regs->ip);
678 printk(KERN_CONT "\n");
682 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
683 unsigned long address, int si_code)
685 struct task_struct *tsk = current;
687 /* User mode accesses just cause a SIGSEGV */
688 if (error_code & PF_USER) {
690 * It's possible to have interrupts off here:
695 * Valid to do another page fault here because this one came
698 if (is_prefetch(regs, error_code, address))
701 if (is_errata100(regs, address))
704 if (unlikely(show_unhandled_signals))
705 show_signal_msg(regs, error_code, address, tsk);
707 /* Kernel addresses are always protection faults: */
708 tsk->thread.cr2 = address;
709 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
710 tsk->thread.trap_no = 14;
712 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
717 if (is_f00f_bug(regs, address))
720 no_context(regs, error_code, address);
724 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
725 unsigned long address)
727 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
731 __bad_area(struct pt_regs *regs, unsigned long error_code,
732 unsigned long address, int si_code)
734 struct mm_struct *mm = current->mm;
737 * Something tried to access memory that isn't in our memory map..
738 * Fix it, but check if it's kernel or user first..
740 up_read(&mm->mmap_sem);
742 __bad_area_nosemaphore(regs, error_code, address, si_code);
746 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
748 __bad_area(regs, error_code, address, SEGV_MAPERR);
752 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
753 unsigned long address)
755 __bad_area(regs, error_code, address, SEGV_ACCERR);
758 /* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
760 out_of_memory(struct pt_regs *regs, unsigned long error_code,
761 unsigned long address)
764 * We ran out of memory, call the OOM killer, and return the userspace
765 * (which will retry the fault, or kill us if we got oom-killed):
767 up_read(¤t->mm->mmap_sem);
769 pagefault_out_of_memory();
773 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
776 struct task_struct *tsk = current;
777 struct mm_struct *mm = tsk->mm;
778 int code = BUS_ADRERR;
780 up_read(&mm->mmap_sem);
782 /* Kernel mode? Handle exceptions or die: */
783 if (!(error_code & PF_USER)) {
784 no_context(regs, error_code, address);
788 /* User-space => ok to do another page fault: */
789 if (is_prefetch(regs, error_code, address))
792 tsk->thread.cr2 = address;
793 tsk->thread.error_code = error_code;
794 tsk->thread.trap_no = 14;
796 #ifdef CONFIG_MEMORY_FAILURE
797 if (fault & VM_FAULT_HWPOISON) {
799 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
800 tsk->comm, tsk->pid, address);
801 code = BUS_MCEERR_AR;
804 force_sig_info_fault(SIGBUS, code, address, tsk);
808 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
809 unsigned long address, unsigned int fault)
811 if (fault & VM_FAULT_OOM) {
812 out_of_memory(regs, error_code, address);
814 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON))
815 do_sigbus(regs, error_code, address, fault);
821 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
823 if ((error_code & PF_WRITE) && !pte_write(*pte))
826 if ((error_code & PF_INSTR) && !pte_exec(*pte))
833 * Handle a spurious fault caused by a stale TLB entry.
835 * This allows us to lazily refresh the TLB when increasing the
836 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
837 * eagerly is very expensive since that implies doing a full
838 * cross-processor TLB flush, even if no stale TLB entries exist
839 * on other processors.
841 * There are no security implications to leaving a stale TLB when
842 * increasing the permissions on a page.
844 static noinline __kprobes int
845 spurious_fault(unsigned long error_code, unsigned long address)
853 /* Reserved-bit violation or user access to kernel space? */
854 if (error_code & (PF_USER | PF_RSVD))
857 pgd = init_mm.pgd + pgd_index(address);
858 if (!pgd_present(*pgd))
861 pud = pud_offset(pgd, address);
862 if (!pud_present(*pud))
866 return spurious_fault_check(error_code, (pte_t *) pud);
868 pmd = pmd_offset(pud, address);
869 if (!pmd_present(*pmd))
873 return spurious_fault_check(error_code, (pte_t *) pmd);
875 pte = pte_offset_kernel(pmd, address);
876 if (!pte_present(*pte))
879 ret = spurious_fault_check(error_code, pte);
884 * Make sure we have permissions in PMD.
885 * If not, then there's a bug in the page tables:
887 ret = spurious_fault_check(error_code, (pte_t *) pmd);
888 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
893 int show_unhandled_signals = 1;
896 access_error(unsigned long error_code, int write, struct vm_area_struct *vma)
899 /* write, present and write, not present: */
900 if (unlikely(!(vma->vm_flags & VM_WRITE)))
906 if (unlikely(error_code & PF_PROT))
909 /* read, not present: */
910 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
916 static int fault_in_kernel_space(unsigned long address)
918 return address >= TASK_SIZE_MAX;
922 * This routine handles page faults. It determines the address,
923 * and the problem, and then passes it off to one of the appropriate
926 dotraplinkage void __kprobes
927 do_page_fault(struct pt_regs *regs, unsigned long error_code)
929 struct vm_area_struct *vma;
930 struct task_struct *tsk;
931 unsigned long address;
932 struct mm_struct *mm;
939 /* Get the faulting address: */
940 address = read_cr2();
943 * Detect and handle instructions that would cause a page fault for
944 * both a tracked kernel page and a userspace page.
946 if (kmemcheck_active(regs))
947 kmemcheck_hide(regs);
948 prefetchw(&mm->mmap_sem);
950 if (unlikely(kmmio_fault(regs, address)))
954 * We fault-in kernel-space virtual memory on-demand. The
955 * 'reference' page table is init_mm.pgd.
957 * NOTE! We MUST NOT take any locks for this case. We may
958 * be in an interrupt or a critical region, and should
959 * only copy the information from the master page table,
962 * This verifies that the fault happens in kernel space
963 * (error_code & 4) == 0, and that the fault was not a
964 * protection error (error_code & 9) == 0.
966 if (unlikely(fault_in_kernel_space(address))) {
967 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
968 if (vmalloc_fault(address) >= 0)
971 if (kmemcheck_fault(regs, address, error_code))
975 /* Can handle a stale RO->RW TLB: */
976 if (spurious_fault(error_code, address))
979 /* kprobes don't want to hook the spurious faults: */
980 if (notify_page_fault(regs))
983 * Don't take the mm semaphore here. If we fixup a prefetch
984 * fault we could otherwise deadlock:
986 bad_area_nosemaphore(regs, error_code, address);
991 /* kprobes don't want to hook the spurious faults: */
992 if (unlikely(notify_page_fault(regs)))
995 * It's safe to allow irq's after cr2 has been saved and the
996 * vmalloc fault has been handled.
998 * User-mode registers count as a user access even for any
999 * potential system fault or CPU buglet:
1001 if (user_mode_vm(regs)) {
1003 error_code |= PF_USER;
1005 if (regs->flags & X86_EFLAGS_IF)
1009 if (unlikely(error_code & PF_RSVD))
1010 pgtable_bad(regs, error_code, address);
1012 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, 0, regs, address);
1015 * If we're in an interrupt, have no user context or are running
1016 * in an atomic region then we must not take the fault:
1018 if (unlikely(in_atomic() || !mm)) {
1019 bad_area_nosemaphore(regs, error_code, address);
1024 * When running in the kernel we expect faults to occur only to
1025 * addresses in user space. All other faults represent errors in
1026 * the kernel and should generate an OOPS. Unfortunately, in the
1027 * case of an erroneous fault occurring in a code path which already
1028 * holds mmap_sem we will deadlock attempting to validate the fault
1029 * against the address space. Luckily the kernel only validly
1030 * references user space from well defined areas of code, which are
1031 * listed in the exceptions table.
1033 * As the vast majority of faults will be valid we will only perform
1034 * the source reference check when there is a possibility of a
1035 * deadlock. Attempt to lock the address space, if we cannot we then
1036 * validate the source. If this is invalid we can skip the address
1037 * space check, thus avoiding the deadlock:
1039 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1040 if ((error_code & PF_USER) == 0 &&
1041 !search_exception_tables(regs->ip)) {
1042 bad_area_nosemaphore(regs, error_code, address);
1045 down_read(&mm->mmap_sem);
1048 * The above down_read_trylock() might have succeeded in
1049 * which case we'll have missed the might_sleep() from
1055 vma = find_vma(mm, address);
1056 if (unlikely(!vma)) {
1057 bad_area(regs, error_code, address);
1060 if (likely(vma->vm_start <= address))
1062 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1063 bad_area(regs, error_code, address);
1066 if (error_code & PF_USER) {
1068 * Accessing the stack below %sp is always a bug.
1069 * The large cushion allows instructions like enter
1070 * and pusha to work. ("enter $65535, $31" pushes
1071 * 32 pointers and then decrements %sp by 65535.)
1073 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1074 bad_area(regs, error_code, address);
1078 if (unlikely(expand_stack(vma, address))) {
1079 bad_area(regs, error_code, address);
1084 * Ok, we have a good vm_area for this memory access, so
1085 * we can handle it..
1088 write = error_code & PF_WRITE;
1090 if (unlikely(access_error(error_code, write, vma))) {
1091 bad_area_access_error(regs, error_code, address);
1096 * If for any reason at all we couldn't handle the fault,
1097 * make sure we exit gracefully rather than endlessly redo
1100 fault = handle_mm_fault(mm, vma, address, write ? FAULT_FLAG_WRITE : 0);
1102 if (unlikely(fault & VM_FAULT_ERROR)) {
1103 mm_fault_error(regs, error_code, address, fault);
1107 if (fault & VM_FAULT_MAJOR) {
1109 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, 0,
1113 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, 0,
1117 check_v8086_mode(regs, address, tsk);
1119 up_read(&mm->mmap_sem);