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 */
32 #include <asm/pgtable_areas.h> /* VMALLOC_START, ... */
33 #include <asm/kvm_para.h> /* kvm_handle_async_pf */
35 #define CREATE_TRACE_POINTS
36 #include <asm/trace/exceptions.h>
39 * Returns 0 if mmiotrace is disabled, or if the fault is not
40 * handled by mmiotrace:
42 static nokprobe_inline
int
43 kmmio_fault(struct pt_regs
*regs
, unsigned long addr
)
45 if (unlikely(is_kmmio_active()))
46 if (kmmio_handler(regs
, addr
) == 1)
56 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
57 * Check that here and ignore it.
61 * Sometimes the CPU reports invalid exceptions on prefetch.
62 * Check that here and ignore it.
64 * Opcode checker based on code by Richard Brunner.
67 check_prefetch_opcode(struct pt_regs
*regs
, unsigned char *instr
,
68 unsigned char opcode
, int *prefetch
)
70 unsigned char instr_hi
= opcode
& 0xf0;
71 unsigned char instr_lo
= opcode
& 0x0f;
77 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
78 * In X86_64 long mode, the CPU will signal invalid
79 * opcode if some of these prefixes are present so
80 * X86_64 will never get here anyway
82 return ((instr_lo
& 7) == 0x6);
86 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
87 * Need to figure out under what instruction mode the
88 * instruction was issued. Could check the LDT for lm,
89 * but for now it's good enough to assume that long
90 * mode only uses well known segments or kernel.
92 return (!user_mode(regs
) || user_64bit_mode(regs
));
95 /* 0x64 thru 0x67 are valid prefixes in all modes. */
96 return (instr_lo
& 0xC) == 0x4;
98 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
99 return !instr_lo
|| (instr_lo
>>1) == 1;
101 /* Prefetch instruction is 0x0F0D or 0x0F18 */
102 if (probe_kernel_address(instr
, opcode
))
105 *prefetch
= (instr_lo
== 0xF) &&
106 (opcode
== 0x0D || opcode
== 0x18);
114 is_prefetch(struct pt_regs
*regs
, unsigned long error_code
, unsigned long addr
)
116 unsigned char *max_instr
;
117 unsigned char *instr
;
121 * If it was a exec (instruction fetch) fault on NX page, then
122 * do not ignore the fault:
124 if (error_code
& X86_PF_INSTR
)
127 instr
= (void *)convert_ip_to_linear(current
, regs
);
128 max_instr
= instr
+ 15;
130 if (user_mode(regs
) && instr
>= (unsigned char *)TASK_SIZE_MAX
)
133 while (instr
< max_instr
) {
134 unsigned char opcode
;
136 if (probe_kernel_address(instr
, opcode
))
141 if (!check_prefetch_opcode(regs
, instr
, opcode
, &prefetch
))
147 DEFINE_SPINLOCK(pgd_lock
);
151 static inline pmd_t
*vmalloc_sync_one(pgd_t
*pgd
, unsigned long address
)
153 unsigned index
= pgd_index(address
);
160 pgd_k
= init_mm
.pgd
+ index
;
162 if (!pgd_present(*pgd_k
))
166 * set_pgd(pgd, *pgd_k); here would be useless on PAE
167 * and redundant with the set_pmd() on non-PAE. As would
170 p4d
= p4d_offset(pgd
, address
);
171 p4d_k
= p4d_offset(pgd_k
, address
);
172 if (!p4d_present(*p4d_k
))
175 pud
= pud_offset(p4d
, address
);
176 pud_k
= pud_offset(p4d_k
, address
);
177 if (!pud_present(*pud_k
))
180 pmd
= pmd_offset(pud
, address
);
181 pmd_k
= pmd_offset(pud_k
, address
);
183 if (pmd_present(*pmd
) != pmd_present(*pmd_k
))
184 set_pmd(pmd
, *pmd_k
);
186 if (!pmd_present(*pmd_k
))
189 BUG_ON(pmd_pfn(*pmd
) != pmd_pfn(*pmd_k
));
194 void arch_sync_kernel_mappings(unsigned long start
, unsigned long end
)
198 for (addr
= start
& PMD_MASK
;
199 addr
>= TASK_SIZE_MAX
&& addr
< VMALLOC_END
;
203 spin_lock(&pgd_lock
);
204 list_for_each_entry(page
, &pgd_list
, lru
) {
205 spinlock_t
*pgt_lock
;
207 /* the pgt_lock only for Xen */
208 pgt_lock
= &pgd_page_get_mm(page
)->page_table_lock
;
211 vmalloc_sync_one(page_address(page
), addr
);
212 spin_unlock(pgt_lock
);
214 spin_unlock(&pgd_lock
);
219 * Did it hit the DOS screen memory VA from vm86 mode?
222 check_v8086_mode(struct pt_regs
*regs
, unsigned long address
,
223 struct task_struct
*tsk
)
228 if (!v8086_mode(regs
) || !tsk
->thread
.vm86
)
231 bit
= (address
- 0xA0000) >> PAGE_SHIFT
;
233 tsk
->thread
.vm86
->screen_bitmap
|= 1 << bit
;
237 static bool low_pfn(unsigned long pfn
)
239 return pfn
< max_low_pfn
;
242 static void dump_pagetable(unsigned long address
)
244 pgd_t
*base
= __va(read_cr3_pa());
245 pgd_t
*pgd
= &base
[pgd_index(address
)];
251 #ifdef CONFIG_X86_PAE
252 pr_info("*pdpt = %016Lx ", pgd_val(*pgd
));
253 if (!low_pfn(pgd_val(*pgd
) >> PAGE_SHIFT
) || !pgd_present(*pgd
))
255 #define pr_pde pr_cont
257 #define pr_pde pr_info
259 p4d
= p4d_offset(pgd
, address
);
260 pud
= pud_offset(p4d
, address
);
261 pmd
= pmd_offset(pud
, address
);
262 pr_pde("*pde = %0*Lx ", sizeof(*pmd
) * 2, (u64
)pmd_val(*pmd
));
266 * We must not directly access the pte in the highpte
267 * case if the page table is located in highmem.
268 * And let's rather not kmap-atomic the pte, just in case
269 * it's allocated already:
271 if (!low_pfn(pmd_pfn(*pmd
)) || !pmd_present(*pmd
) || pmd_large(*pmd
))
274 pte
= pte_offset_kernel(pmd
, address
);
275 pr_cont("*pte = %0*Lx ", sizeof(*pte
) * 2, (u64
)pte_val(*pte
));
280 #else /* CONFIG_X86_64: */
282 #ifdef CONFIG_CPU_SUP_AMD
283 static const char errata93_warning
[] =
285 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
286 "******* Working around it, but it may cause SEGVs or burn power.\n"
287 "******* Please consider a BIOS update.\n"
288 "******* Disabling USB legacy in the BIOS may also help.\n";
292 * No vm86 mode in 64-bit mode:
295 check_v8086_mode(struct pt_regs
*regs
, unsigned long address
,
296 struct task_struct
*tsk
)
300 static int bad_address(void *p
)
304 return probe_kernel_address((unsigned long *)p
, dummy
);
307 static void dump_pagetable(unsigned long address
)
309 pgd_t
*base
= __va(read_cr3_pa());
310 pgd_t
*pgd
= base
+ pgd_index(address
);
316 if (bad_address(pgd
))
319 pr_info("PGD %lx ", pgd_val(*pgd
));
321 if (!pgd_present(*pgd
))
324 p4d
= p4d_offset(pgd
, address
);
325 if (bad_address(p4d
))
328 pr_cont("P4D %lx ", p4d_val(*p4d
));
329 if (!p4d_present(*p4d
) || p4d_large(*p4d
))
332 pud
= pud_offset(p4d
, address
);
333 if (bad_address(pud
))
336 pr_cont("PUD %lx ", pud_val(*pud
));
337 if (!pud_present(*pud
) || pud_large(*pud
))
340 pmd
= pmd_offset(pud
, address
);
341 if (bad_address(pmd
))
344 pr_cont("PMD %lx ", pmd_val(*pmd
));
345 if (!pmd_present(*pmd
) || pmd_large(*pmd
))
348 pte
= pte_offset_kernel(pmd
, address
);
349 if (bad_address(pte
))
352 pr_cont("PTE %lx", pte_val(*pte
));
360 #endif /* CONFIG_X86_64 */
363 * Workaround for K8 erratum #93 & buggy BIOS.
365 * BIOS SMM functions are required to use a specific workaround
366 * to avoid corruption of the 64bit RIP register on C stepping K8.
368 * A lot of BIOS that didn't get tested properly miss this.
370 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
371 * Try to work around it here.
373 * Note we only handle faults in kernel here.
374 * Does nothing on 32-bit.
376 static int is_errata93(struct pt_regs
*regs
, unsigned long address
)
378 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
379 if (boot_cpu_data
.x86_vendor
!= X86_VENDOR_AMD
380 || boot_cpu_data
.x86
!= 0xf)
383 if (address
!= regs
->ip
)
386 if ((address
>> 32) != 0)
389 address
|= 0xffffffffUL
<< 32;
390 if ((address
>= (u64
)_stext
&& address
<= (u64
)_etext
) ||
391 (address
>= MODULES_VADDR
&& address
<= MODULES_END
)) {
392 printk_once(errata93_warning
);
401 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
402 * to illegal addresses >4GB.
404 * We catch this in the page fault handler because these addresses
405 * are not reachable. Just detect this case and return. Any code
406 * segment in LDT is compatibility mode.
408 static int is_errata100(struct pt_regs
*regs
, unsigned long address
)
411 if ((regs
->cs
== __USER32_CS
|| (regs
->cs
& (1<<2))) && (address
>> 32))
417 static int is_f00f_bug(struct pt_regs
*regs
, unsigned long address
)
419 #ifdef CONFIG_X86_F00F_BUG
423 * Pentium F0 0F C7 C8 bug workaround:
425 if (boot_cpu_has_bug(X86_BUG_F00F
)) {
426 nr
= (address
- idt_descr
.address
) >> 3;
429 do_invalid_op(regs
, 0);
437 static void show_ldttss(const struct desc_ptr
*gdt
, const char *name
, u16 index
)
439 u32 offset
= (index
>> 3) * sizeof(struct desc_struct
);
441 struct ldttss_desc desc
;
444 pr_alert("%s: NULL\n", name
);
448 if (offset
+ sizeof(struct ldttss_desc
) >= gdt
->size
) {
449 pr_alert("%s: 0x%hx -- out of bounds\n", name
, index
);
453 if (probe_kernel_read(&desc
, (void *)(gdt
->address
+ offset
),
454 sizeof(struct ldttss_desc
))) {
455 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
460 addr
= desc
.base0
| (desc
.base1
<< 16) | ((unsigned long)desc
.base2
<< 24);
462 addr
|= ((u64
)desc
.base3
<< 32);
464 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
465 name
, index
, addr
, (desc
.limit0
| (desc
.limit1
<< 16)));
469 show_fault_oops(struct pt_regs
*regs
, unsigned long error_code
, unsigned long address
)
471 if (!oops_may_print())
474 if (error_code
& X86_PF_INSTR
) {
479 pgd
= __va(read_cr3_pa());
480 pgd
+= pgd_index(address
);
482 pte
= lookup_address_in_pgd(pgd
, address
, &level
);
484 if (pte
&& pte_present(*pte
) && !pte_exec(*pte
))
485 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
486 from_kuid(&init_user_ns
, current_uid()));
487 if (pte
&& pte_present(*pte
) && pte_exec(*pte
) &&
488 (pgd_flags(*pgd
) & _PAGE_USER
) &&
489 (__read_cr4() & X86_CR4_SMEP
))
490 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
491 from_kuid(&init_user_ns
, current_uid()));
494 if (address
< PAGE_SIZE
&& !user_mode(regs
))
495 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
498 pr_alert("BUG: unable to handle page fault for address: %px\n",
501 pr_alert("#PF: %s %s in %s mode\n",
502 (error_code
& X86_PF_USER
) ? "user" : "supervisor",
503 (error_code
& X86_PF_INSTR
) ? "instruction fetch" :
504 (error_code
& X86_PF_WRITE
) ? "write access" :
506 user_mode(regs
) ? "user" : "kernel");
507 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code
,
508 !(error_code
& X86_PF_PROT
) ? "not-present page" :
509 (error_code
& X86_PF_RSVD
) ? "reserved bit violation" :
510 (error_code
& X86_PF_PK
) ? "protection keys violation" :
511 "permissions violation");
513 if (!(error_code
& X86_PF_USER
) && user_mode(regs
)) {
514 struct desc_ptr idt
, gdt
;
518 * This can happen for quite a few reasons. The more obvious
519 * ones are faults accessing the GDT, or LDT. Perhaps
520 * surprisingly, if the CPU tries to deliver a benign or
521 * contributory exception from user code and gets a page fault
522 * during delivery, the page fault can be delivered as though
523 * it originated directly from user code. This could happen
524 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
525 * kernel or IST stack.
529 /* Usable even on Xen PV -- it's just slow. */
530 native_store_gdt(&gdt
);
532 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
533 idt
.address
, idt
.size
, gdt
.address
, gdt
.size
);
536 show_ldttss(&gdt
, "LDTR", ldtr
);
539 show_ldttss(&gdt
, "TR", tr
);
542 dump_pagetable(address
);
546 pgtable_bad(struct pt_regs
*regs
, unsigned long error_code
,
547 unsigned long address
)
549 struct task_struct
*tsk
;
553 flags
= oops_begin();
557 printk(KERN_ALERT
"%s: Corrupted page table at address %lx\n",
559 dump_pagetable(address
);
561 if (__die("Bad pagetable", regs
, error_code
))
564 oops_end(flags
, regs
, sig
);
567 static void set_signal_archinfo(unsigned long address
,
568 unsigned long error_code
)
570 struct task_struct
*tsk
= current
;
573 * To avoid leaking information about the kernel page
574 * table layout, pretend that user-mode accesses to
575 * kernel addresses are always protection faults.
577 * NB: This means that failed vsyscalls with vsyscall=none
578 * will have the PROT bit. This doesn't leak any
579 * information and does not appear to cause any problems.
581 if (address
>= TASK_SIZE_MAX
)
582 error_code
|= X86_PF_PROT
;
584 tsk
->thread
.trap_nr
= X86_TRAP_PF
;
585 tsk
->thread
.error_code
= error_code
| X86_PF_USER
;
586 tsk
->thread
.cr2
= address
;
590 no_context(struct pt_regs
*regs
, unsigned long error_code
,
591 unsigned long address
, int signal
, int si_code
)
593 struct task_struct
*tsk
= current
;
597 if (user_mode(regs
)) {
599 * This is an implicit supervisor-mode access from user
600 * mode. Bypass all the kernel-mode recovery code and just
606 /* Are we prepared to handle this kernel fault? */
607 if (fixup_exception(regs
, X86_TRAP_PF
, error_code
, address
)) {
609 * Any interrupt that takes a fault gets the fixup. This makes
610 * the below recursive fault logic only apply to a faults from
617 * Per the above we're !in_interrupt(), aka. task context.
619 * In this case we need to make sure we're not recursively
620 * faulting through the emulate_vsyscall() logic.
622 if (current
->thread
.sig_on_uaccess_err
&& signal
) {
623 set_signal_archinfo(address
, error_code
);
625 /* XXX: hwpoison faults will set the wrong code. */
626 force_sig_fault(signal
, si_code
, (void __user
*)address
);
630 * Barring that, we can do the fixup and be happy.
635 #ifdef CONFIG_VMAP_STACK
637 * Stack overflow? During boot, we can fault near the initial
638 * stack in the direct map, but that's not an overflow -- check
639 * that we're in vmalloc space to avoid this.
641 if (is_vmalloc_addr((void *)address
) &&
642 (((unsigned long)tsk
->stack
- 1 - address
< PAGE_SIZE
) ||
643 address
- ((unsigned long)tsk
->stack
+ THREAD_SIZE
) < PAGE_SIZE
)) {
644 unsigned long stack
= __this_cpu_ist_top_va(DF
) - sizeof(void *);
646 * We're likely to be running with very little stack space
647 * left. It's plausible that we'd hit this condition but
648 * double-fault even before we get this far, in which case
649 * we're fine: the double-fault handler will deal with it.
651 * We don't want to make it all the way into the oops code
652 * and then double-fault, though, because we're likely to
653 * break the console driver and lose most of the stack dump.
655 asm volatile ("movq %[stack], %%rsp\n\t"
656 "call handle_stack_overflow\n\t"
658 : ASM_CALL_CONSTRAINT
659 : "D" ("kernel stack overflow (page fault)"),
660 "S" (regs
), "d" (address
),
661 [stack
] "rm" (stack
));
669 * Valid to do another page fault here, because if this fault
670 * had been triggered by is_prefetch fixup_exception would have
675 * Hall of shame of CPU/BIOS bugs.
677 if (is_prefetch(regs
, error_code
, address
))
680 if (is_errata93(regs
, address
))
684 * Buggy firmware could access regions which might page fault, try to
685 * recover from such faults.
687 if (IS_ENABLED(CONFIG_EFI
))
688 efi_recover_from_page_fault(address
);
692 * Oops. The kernel tried to access some bad page. We'll have to
693 * terminate things with extreme prejudice:
695 flags
= oops_begin();
697 show_fault_oops(regs
, error_code
, address
);
699 if (task_stack_end_corrupted(tsk
))
700 printk(KERN_EMERG
"Thread overran stack, or stack corrupted\n");
703 if (__die("Oops", regs
, error_code
))
706 /* Executive summary in case the body of the oops scrolled away */
707 printk(KERN_DEFAULT
"CR2: %016lx\n", address
);
709 oops_end(flags
, regs
, sig
);
713 * Print out info about fatal segfaults, if the show_unhandled_signals
717 show_signal_msg(struct pt_regs
*regs
, unsigned long error_code
,
718 unsigned long address
, struct task_struct
*tsk
)
720 const char *loglvl
= task_pid_nr(tsk
) > 1 ? KERN_INFO
: KERN_EMERG
;
722 if (!unhandled_signal(tsk
, SIGSEGV
))
725 if (!printk_ratelimit())
728 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
729 loglvl
, tsk
->comm
, task_pid_nr(tsk
), address
,
730 (void *)regs
->ip
, (void *)regs
->sp
, error_code
);
732 print_vma_addr(KERN_CONT
" in ", regs
->ip
);
734 printk(KERN_CONT
"\n");
736 show_opcodes(regs
, loglvl
);
740 * The (legacy) vsyscall page is the long page in the kernel portion
741 * of the address space that has user-accessible permissions.
743 static bool is_vsyscall_vaddr(unsigned long vaddr
)
745 return unlikely((vaddr
& PAGE_MASK
) == VSYSCALL_ADDR
);
749 __bad_area_nosemaphore(struct pt_regs
*regs
, unsigned long error_code
,
750 unsigned long address
, u32 pkey
, int si_code
)
752 struct task_struct
*tsk
= current
;
754 /* User mode accesses just cause a SIGSEGV */
755 if (user_mode(regs
) && (error_code
& X86_PF_USER
)) {
757 * It's possible to have interrupts off here:
762 * Valid to do another page fault here because this one came
765 if (is_prefetch(regs
, error_code
, address
))
768 if (is_errata100(regs
, address
))
772 * To avoid leaking information about the kernel page table
773 * layout, pretend that user-mode accesses to kernel addresses
774 * are always protection faults.
776 if (address
>= TASK_SIZE_MAX
)
777 error_code
|= X86_PF_PROT
;
779 if (likely(show_unhandled_signals
))
780 show_signal_msg(regs
, error_code
, address
, tsk
);
782 set_signal_archinfo(address
, error_code
);
784 if (si_code
== SEGV_PKUERR
)
785 force_sig_pkuerr((void __user
*)address
, pkey
);
787 force_sig_fault(SIGSEGV
, si_code
, (void __user
*)address
);
792 if (is_f00f_bug(regs
, address
))
795 no_context(regs
, error_code
, address
, SIGSEGV
, si_code
);
799 bad_area_nosemaphore(struct pt_regs
*regs
, unsigned long error_code
,
800 unsigned long address
)
802 __bad_area_nosemaphore(regs
, error_code
, address
, 0, SEGV_MAPERR
);
806 __bad_area(struct pt_regs
*regs
, unsigned long error_code
,
807 unsigned long address
, u32 pkey
, int si_code
)
809 struct mm_struct
*mm
= current
->mm
;
811 * Something tried to access memory that isn't in our memory map..
812 * Fix it, but check if it's kernel or user first..
814 mmap_read_unlock(mm
);
816 __bad_area_nosemaphore(regs
, error_code
, address
, pkey
, si_code
);
820 bad_area(struct pt_regs
*regs
, unsigned long error_code
, unsigned long address
)
822 __bad_area(regs
, error_code
, address
, 0, SEGV_MAPERR
);
825 static inline bool bad_area_access_from_pkeys(unsigned long error_code
,
826 struct vm_area_struct
*vma
)
828 /* This code is always called on the current mm */
829 bool foreign
= false;
831 if (!boot_cpu_has(X86_FEATURE_OSPKE
))
833 if (error_code
& X86_PF_PK
)
835 /* this checks permission keys on the VMA: */
836 if (!arch_vma_access_permitted(vma
, (error_code
& X86_PF_WRITE
),
837 (error_code
& X86_PF_INSTR
), foreign
))
843 bad_area_access_error(struct pt_regs
*regs
, unsigned long error_code
,
844 unsigned long address
, struct vm_area_struct
*vma
)
847 * This OSPKE check is not strictly necessary at runtime.
848 * But, doing it this way allows compiler optimizations
849 * if pkeys are compiled out.
851 if (bad_area_access_from_pkeys(error_code
, vma
)) {
853 * A protection key fault means that the PKRU value did not allow
854 * access to some PTE. Userspace can figure out what PKRU was
855 * from the XSAVE state. This function captures the pkey from
856 * the vma and passes it to userspace so userspace can discover
857 * which protection key was set on the PTE.
859 * If we get here, we know that the hardware signaled a X86_PF_PK
860 * fault and that there was a VMA once we got in the fault
861 * handler. It does *not* guarantee that the VMA we find here
862 * was the one that we faulted on.
864 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
865 * 2. T1 : set PKRU to deny access to pkey=4, touches page
867 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
868 * 5. T1 : enters fault handler, takes mmap_lock, etc...
869 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
870 * faulted on a pte with its pkey=4.
872 u32 pkey
= vma_pkey(vma
);
874 __bad_area(regs
, error_code
, address
, pkey
, SEGV_PKUERR
);
876 __bad_area(regs
, error_code
, address
, 0, SEGV_ACCERR
);
881 do_sigbus(struct pt_regs
*regs
, unsigned long error_code
, unsigned long address
,
884 /* Kernel mode? Handle exceptions or die: */
885 if (!(error_code
& X86_PF_USER
)) {
886 no_context(regs
, error_code
, address
, SIGBUS
, BUS_ADRERR
);
890 /* User-space => ok to do another page fault: */
891 if (is_prefetch(regs
, error_code
, address
))
894 set_signal_archinfo(address
, error_code
);
896 #ifdef CONFIG_MEMORY_FAILURE
897 if (fault
& (VM_FAULT_HWPOISON
|VM_FAULT_HWPOISON_LARGE
)) {
898 struct task_struct
*tsk
= current
;
902 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
903 tsk
->comm
, tsk
->pid
, address
);
904 if (fault
& VM_FAULT_HWPOISON_LARGE
)
905 lsb
= hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault
));
906 if (fault
& VM_FAULT_HWPOISON
)
908 force_sig_mceerr(BUS_MCEERR_AR
, (void __user
*)address
, lsb
);
912 force_sig_fault(SIGBUS
, BUS_ADRERR
, (void __user
*)address
);
916 mm_fault_error(struct pt_regs
*regs
, unsigned long error_code
,
917 unsigned long address
, vm_fault_t fault
)
919 if (fatal_signal_pending(current
) && !(error_code
& X86_PF_USER
)) {
920 no_context(regs
, error_code
, address
, 0, 0);
924 if (fault
& VM_FAULT_OOM
) {
925 /* Kernel mode? Handle exceptions or die: */
926 if (!(error_code
& X86_PF_USER
)) {
927 no_context(regs
, error_code
, address
,
928 SIGSEGV
, SEGV_MAPERR
);
933 * We ran out of memory, call the OOM killer, and return the
934 * userspace (which will retry the fault, or kill us if we got
937 pagefault_out_of_memory();
939 if (fault
& (VM_FAULT_SIGBUS
|VM_FAULT_HWPOISON
|
940 VM_FAULT_HWPOISON_LARGE
))
941 do_sigbus(regs
, error_code
, address
, fault
);
942 else if (fault
& VM_FAULT_SIGSEGV
)
943 bad_area_nosemaphore(regs
, error_code
, address
);
949 static int spurious_kernel_fault_check(unsigned long error_code
, pte_t
*pte
)
951 if ((error_code
& X86_PF_WRITE
) && !pte_write(*pte
))
954 if ((error_code
& X86_PF_INSTR
) && !pte_exec(*pte
))
961 * Handle a spurious fault caused by a stale TLB entry.
963 * This allows us to lazily refresh the TLB when increasing the
964 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
965 * eagerly is very expensive since that implies doing a full
966 * cross-processor TLB flush, even if no stale TLB entries exist
967 * on other processors.
969 * Spurious faults may only occur if the TLB contains an entry with
970 * fewer permission than the page table entry. Non-present (P = 0)
971 * and reserved bit (R = 1) faults are never spurious.
973 * There are no security implications to leaving a stale TLB when
974 * increasing the permissions on a page.
976 * Returns non-zero if a spurious fault was handled, zero otherwise.
978 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
979 * (Optional Invalidation).
982 spurious_kernel_fault(unsigned long error_code
, unsigned long address
)
992 * Only writes to RO or instruction fetches from NX may cause
995 * These could be from user or supervisor accesses but the TLB
996 * is only lazily flushed after a kernel mapping protection
997 * change, so user accesses are not expected to cause spurious
1000 if (error_code
!= (X86_PF_WRITE
| X86_PF_PROT
) &&
1001 error_code
!= (X86_PF_INSTR
| X86_PF_PROT
))
1004 pgd
= init_mm
.pgd
+ pgd_index(address
);
1005 if (!pgd_present(*pgd
))
1008 p4d
= p4d_offset(pgd
, address
);
1009 if (!p4d_present(*p4d
))
1012 if (p4d_large(*p4d
))
1013 return spurious_kernel_fault_check(error_code
, (pte_t
*) p4d
);
1015 pud
= pud_offset(p4d
, address
);
1016 if (!pud_present(*pud
))
1019 if (pud_large(*pud
))
1020 return spurious_kernel_fault_check(error_code
, (pte_t
*) pud
);
1022 pmd
= pmd_offset(pud
, address
);
1023 if (!pmd_present(*pmd
))
1026 if (pmd_large(*pmd
))
1027 return spurious_kernel_fault_check(error_code
, (pte_t
*) pmd
);
1029 pte
= pte_offset_kernel(pmd
, address
);
1030 if (!pte_present(*pte
))
1033 ret
= spurious_kernel_fault_check(error_code
, pte
);
1038 * Make sure we have permissions in PMD.
1039 * If not, then there's a bug in the page tables:
1041 ret
= spurious_kernel_fault_check(error_code
, (pte_t
*) pmd
);
1042 WARN_ONCE(!ret
, "PMD has incorrect permission bits\n");
1046 NOKPROBE_SYMBOL(spurious_kernel_fault
);
1048 int show_unhandled_signals
= 1;
1051 access_error(unsigned long error_code
, struct vm_area_struct
*vma
)
1053 /* This is only called for the current mm, so: */
1054 bool foreign
= false;
1057 * Read or write was blocked by protection keys. This is
1058 * always an unconditional error and can never result in
1059 * a follow-up action to resolve the fault, like a COW.
1061 if (error_code
& X86_PF_PK
)
1065 * Make sure to check the VMA so that we do not perform
1066 * faults just to hit a X86_PF_PK as soon as we fill in a
1069 if (!arch_vma_access_permitted(vma
, (error_code
& X86_PF_WRITE
),
1070 (error_code
& X86_PF_INSTR
), foreign
))
1073 if (error_code
& X86_PF_WRITE
) {
1074 /* write, present and write, not present: */
1075 if (unlikely(!(vma
->vm_flags
& VM_WRITE
)))
1080 /* read, present: */
1081 if (unlikely(error_code
& X86_PF_PROT
))
1084 /* read, not present: */
1085 if (unlikely(!vma_is_accessible(vma
)))
1091 static int fault_in_kernel_space(unsigned long address
)
1094 * On 64-bit systems, the vsyscall page is at an address above
1095 * TASK_SIZE_MAX, but is not considered part of the kernel
1098 if (IS_ENABLED(CONFIG_X86_64
) && is_vsyscall_vaddr(address
))
1101 return address
>= TASK_SIZE_MAX
;
1105 * Called for all faults where 'address' is part of the kernel address
1106 * space. Might get called for faults that originate from *code* that
1107 * ran in userspace or the kernel.
1110 do_kern_addr_fault(struct pt_regs
*regs
, unsigned long hw_error_code
,
1111 unsigned long address
)
1114 * Protection keys exceptions only happen on user pages. We
1115 * have no user pages in the kernel portion of the address
1116 * space, so do not expect them here.
1118 WARN_ON_ONCE(hw_error_code
& X86_PF_PK
);
1120 /* Was the fault spurious, caused by lazy TLB invalidation? */
1121 if (spurious_kernel_fault(hw_error_code
, address
))
1124 /* kprobes don't want to hook the spurious faults: */
1125 if (kprobe_page_fault(regs
, X86_TRAP_PF
))
1129 * Note, despite being a "bad area", there are quite a few
1130 * acceptable reasons to get here, such as erratum fixups
1131 * and handling kernel code that can fault, like get_user().
1133 * Don't take the mm semaphore here. If we fixup a prefetch
1134 * fault we could otherwise deadlock:
1136 bad_area_nosemaphore(regs
, hw_error_code
, address
);
1138 NOKPROBE_SYMBOL(do_kern_addr_fault
);
1140 /* Handle faults in the user portion of the address space */
1142 void do_user_addr_fault(struct pt_regs
*regs
,
1143 unsigned long hw_error_code
,
1144 unsigned long address
)
1146 struct vm_area_struct
*vma
;
1147 struct task_struct
*tsk
;
1148 struct mm_struct
*mm
;
1149 vm_fault_t fault
, major
= 0;
1150 unsigned int flags
= FAULT_FLAG_DEFAULT
;
1155 /* kprobes don't want to hook the spurious faults: */
1156 if (unlikely(kprobe_page_fault(regs
, X86_TRAP_PF
)))
1160 * Reserved bits are never expected to be set on
1161 * entries in the user portion of the page tables.
1163 if (unlikely(hw_error_code
& X86_PF_RSVD
))
1164 pgtable_bad(regs
, hw_error_code
, address
);
1167 * If SMAP is on, check for invalid kernel (supervisor) access to user
1168 * pages in the user address space. The odd case here is WRUSS,
1169 * which, according to the preliminary documentation, does not respect
1170 * SMAP and will have the USER bit set so, in all cases, SMAP
1171 * enforcement appears to be consistent with the USER bit.
1173 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP
) &&
1174 !(hw_error_code
& X86_PF_USER
) &&
1175 !(regs
->flags
& X86_EFLAGS_AC
)))
1177 bad_area_nosemaphore(regs
, hw_error_code
, address
);
1182 * If we're in an interrupt, have no user context or are running
1183 * in a region with pagefaults disabled then we must not take the fault
1185 if (unlikely(faulthandler_disabled() || !mm
)) {
1186 bad_area_nosemaphore(regs
, hw_error_code
, address
);
1191 * It's safe to allow irq's after cr2 has been saved and the
1192 * vmalloc fault has been handled.
1194 * User-mode registers count as a user access even for any
1195 * potential system fault or CPU buglet:
1197 if (user_mode(regs
)) {
1199 flags
|= FAULT_FLAG_USER
;
1201 if (regs
->flags
& X86_EFLAGS_IF
)
1205 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS
, 1, regs
, address
);
1207 if (hw_error_code
& X86_PF_WRITE
)
1208 flags
|= FAULT_FLAG_WRITE
;
1209 if (hw_error_code
& X86_PF_INSTR
)
1210 flags
|= FAULT_FLAG_INSTRUCTION
;
1212 #ifdef CONFIG_X86_64
1214 * Faults in the vsyscall page might need emulation. The
1215 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1216 * considered to be part of the user address space.
1218 * The vsyscall page does not have a "real" VMA, so do this
1219 * emulation before we go searching for VMAs.
1221 * PKRU never rejects instruction fetches, so we don't need
1222 * to consider the PF_PK bit.
1224 if (is_vsyscall_vaddr(address
)) {
1225 if (emulate_vsyscall(hw_error_code
, regs
, address
))
1231 * Kernel-mode access to the user address space should only occur
1232 * on well-defined single instructions listed in the exception
1233 * tables. But, an erroneous kernel fault occurring outside one of
1234 * those areas which also holds mmap_lock might deadlock attempting
1235 * to validate the fault against the address space.
1237 * Only do the expensive exception table search when we might be at
1238 * risk of a deadlock. This happens if we
1239 * 1. Failed to acquire mmap_lock, and
1240 * 2. The access did not originate in userspace.
1242 if (unlikely(!mmap_read_trylock(mm
))) {
1243 if (!user_mode(regs
) && !search_exception_tables(regs
->ip
)) {
1245 * Fault from code in kernel from
1246 * which we do not expect faults.
1248 bad_area_nosemaphore(regs
, hw_error_code
, address
);
1255 * The above down_read_trylock() might have succeeded in
1256 * which case we'll have missed the might_sleep() from
1262 vma
= find_vma(mm
, address
);
1263 if (unlikely(!vma
)) {
1264 bad_area(regs
, hw_error_code
, address
);
1267 if (likely(vma
->vm_start
<= address
))
1269 if (unlikely(!(vma
->vm_flags
& VM_GROWSDOWN
))) {
1270 bad_area(regs
, hw_error_code
, address
);
1273 if (unlikely(expand_stack(vma
, address
))) {
1274 bad_area(regs
, hw_error_code
, address
);
1279 * Ok, we have a good vm_area for this memory access, so
1280 * we can handle it..
1283 if (unlikely(access_error(hw_error_code
, vma
))) {
1284 bad_area_access_error(regs
, hw_error_code
, address
, vma
);
1289 * If for any reason at all we couldn't handle the fault,
1290 * make sure we exit gracefully rather than endlessly redo
1291 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1292 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1294 * Note that handle_userfault() may also release and reacquire mmap_lock
1295 * (and not return with VM_FAULT_RETRY), when returning to userland to
1296 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1297 * (potentially after handling any pending signal during the return to
1298 * userland). The return to userland is identified whenever
1299 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1301 fault
= handle_mm_fault(vma
, address
, flags
);
1302 major
|= fault
& VM_FAULT_MAJOR
;
1304 /* Quick path to respond to signals */
1305 if (fault_signal_pending(fault
, regs
)) {
1306 if (!user_mode(regs
))
1307 no_context(regs
, hw_error_code
, address
, SIGBUS
,
1313 * If we need to retry the mmap_lock has already been released,
1314 * and if there is a fatal signal pending there is no guarantee
1315 * that we made any progress. Handle this case first.
1317 if (unlikely((fault
& VM_FAULT_RETRY
) &&
1318 (flags
& FAULT_FLAG_ALLOW_RETRY
))) {
1319 flags
|= FAULT_FLAG_TRIED
;
1323 mmap_read_unlock(mm
);
1324 if (unlikely(fault
& VM_FAULT_ERROR
)) {
1325 mm_fault_error(regs
, hw_error_code
, address
, fault
);
1330 * Major/minor page fault accounting. If any of the events
1331 * returned VM_FAULT_MAJOR, we account it as a major fault.
1335 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ
, 1, regs
, address
);
1338 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN
, 1, regs
, address
);
1341 check_v8086_mode(regs
, address
, tsk
);
1343 NOKPROBE_SYMBOL(do_user_addr_fault
);
1345 static __always_inline
void
1346 trace_page_fault_entries(struct pt_regs
*regs
, unsigned long error_code
,
1347 unsigned long address
)
1349 if (!trace_pagefault_enabled())
1352 if (user_mode(regs
))
1353 trace_page_fault_user(address
, regs
, error_code
);
1355 trace_page_fault_kernel(address
, regs
, error_code
);
1359 do_page_fault(struct pt_regs
*regs
, unsigned long hw_error_code
,
1360 unsigned long address
)
1362 prefetchw(¤t
->mm
->mmap_lock
);
1364 * KVM has two types of events that are, logically, interrupts, but
1365 * are unfortunately delivered using the #PF vector. These events are
1366 * "you just accessed valid memory, but the host doesn't have it right
1367 * now, so I'll put you to sleep if you continue" and "that memory
1368 * you tried to access earlier is available now."
1370 * We are relying on the interrupted context being sane (valid RSP,
1371 * relevant locks not held, etc.), which is fine as long as the
1372 * interrupted context had IF=1. We are also relying on the KVM
1373 * async pf type field and CR2 being read consistently instead of
1374 * getting values from real and async page faults mixed up.
1378 if (kvm_handle_async_pf(regs
, (u32
)address
))
1381 trace_page_fault_entries(regs
, hw_error_code
, address
);
1383 if (unlikely(kmmio_fault(regs
, address
)))
1386 /* Was the fault on kernel-controlled part of the address space? */
1387 if (unlikely(fault_in_kernel_space(address
)))
1388 do_kern_addr_fault(regs
, hw_error_code
, address
);
1390 do_user_addr_fault(regs
, hw_error_code
, address
);
1392 NOKPROBE_SYMBOL(do_page_fault
);