Loading...
1/*
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
5 */
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 */
14#include <linux/hugetlb.h> /* hstate_index_to_shift */
15#include <linux/prefetch.h> /* prefetchw */
16
17#include <asm/traps.h> /* dotraplinkage, ... */
18#include <asm/pgalloc.h> /* pgd_*(), ... */
19#include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
20#include <asm/vsyscall.h>
21
22/*
23 * Page fault error code bits:
24 *
25 * bit 0 == 0: no page found 1: protection fault
26 * bit 1 == 0: read access 1: write access
27 * bit 2 == 0: kernel-mode access 1: user-mode access
28 * bit 3 == 1: use of reserved bit detected
29 * bit 4 == 1: fault was an instruction fetch
30 */
31enum x86_pf_error_code {
32
33 PF_PROT = 1 << 0,
34 PF_WRITE = 1 << 1,
35 PF_USER = 1 << 2,
36 PF_RSVD = 1 << 3,
37 PF_INSTR = 1 << 4,
38};
39
40/*
41 * Returns 0 if mmiotrace is disabled, or if the fault is not
42 * handled by mmiotrace:
43 */
44static inline int __kprobes
45kmmio_fault(struct pt_regs *regs, unsigned long addr)
46{
47 if (unlikely(is_kmmio_active()))
48 if (kmmio_handler(regs, addr) == 1)
49 return -1;
50 return 0;
51}
52
53static inline int __kprobes notify_page_fault(struct pt_regs *regs)
54{
55 int ret = 0;
56
57 /* kprobe_running() needs smp_processor_id() */
58 if (kprobes_built_in() && !user_mode_vm(regs)) {
59 preempt_disable();
60 if (kprobe_running() && kprobe_fault_handler(regs, 14))
61 ret = 1;
62 preempt_enable();
63 }
64
65 return ret;
66}
67
68/*
69 * Prefetch quirks:
70 *
71 * 32-bit mode:
72 *
73 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
74 * Check that here and ignore it.
75 *
76 * 64-bit mode:
77 *
78 * Sometimes the CPU reports invalid exceptions on prefetch.
79 * Check that here and ignore it.
80 *
81 * Opcode checker based on code by Richard Brunner.
82 */
83static inline int
84check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
85 unsigned char opcode, int *prefetch)
86{
87 unsigned char instr_hi = opcode & 0xf0;
88 unsigned char instr_lo = opcode & 0x0f;
89
90 switch (instr_hi) {
91 case 0x20:
92 case 0x30:
93 /*
94 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
95 * In X86_64 long mode, the CPU will signal invalid
96 * opcode if some of these prefixes are present so
97 * X86_64 will never get here anyway
98 */
99 return ((instr_lo & 7) == 0x6);
100#ifdef CONFIG_X86_64
101 case 0x40:
102 /*
103 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
104 * Need to figure out under what instruction mode the
105 * instruction was issued. Could check the LDT for lm,
106 * but for now it's good enough to assume that long
107 * mode only uses well known segments or kernel.
108 */
109 return (!user_mode(regs) || user_64bit_mode(regs));
110#endif
111 case 0x60:
112 /* 0x64 thru 0x67 are valid prefixes in all modes. */
113 return (instr_lo & 0xC) == 0x4;
114 case 0xF0:
115 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
116 return !instr_lo || (instr_lo>>1) == 1;
117 case 0x00:
118 /* Prefetch instruction is 0x0F0D or 0x0F18 */
119 if (probe_kernel_address(instr, opcode))
120 return 0;
121
122 *prefetch = (instr_lo == 0xF) &&
123 (opcode == 0x0D || opcode == 0x18);
124 return 0;
125 default:
126 return 0;
127 }
128}
129
130static int
131is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
132{
133 unsigned char *max_instr;
134 unsigned char *instr;
135 int prefetch = 0;
136
137 /*
138 * If it was a exec (instruction fetch) fault on NX page, then
139 * do not ignore the fault:
140 */
141 if (error_code & PF_INSTR)
142 return 0;
143
144 instr = (void *)convert_ip_to_linear(current, regs);
145 max_instr = instr + 15;
146
147 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
148 return 0;
149
150 while (instr < max_instr) {
151 unsigned char opcode;
152
153 if (probe_kernel_address(instr, opcode))
154 break;
155
156 instr++;
157
158 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
159 break;
160 }
161 return prefetch;
162}
163
164static void
165force_sig_info_fault(int si_signo, int si_code, unsigned long address,
166 struct task_struct *tsk, int fault)
167{
168 unsigned lsb = 0;
169 siginfo_t info;
170
171 info.si_signo = si_signo;
172 info.si_errno = 0;
173 info.si_code = si_code;
174 info.si_addr = (void __user *)address;
175 if (fault & VM_FAULT_HWPOISON_LARGE)
176 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
177 if (fault & VM_FAULT_HWPOISON)
178 lsb = PAGE_SHIFT;
179 info.si_addr_lsb = lsb;
180
181 force_sig_info(si_signo, &info, tsk);
182}
183
184DEFINE_SPINLOCK(pgd_lock);
185LIST_HEAD(pgd_list);
186
187#ifdef CONFIG_X86_32
188static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
189{
190 unsigned index = pgd_index(address);
191 pgd_t *pgd_k;
192 pud_t *pud, *pud_k;
193 pmd_t *pmd, *pmd_k;
194
195 pgd += index;
196 pgd_k = init_mm.pgd + index;
197
198 if (!pgd_present(*pgd_k))
199 return NULL;
200
201 /*
202 * set_pgd(pgd, *pgd_k); here would be useless on PAE
203 * and redundant with the set_pmd() on non-PAE. As would
204 * set_pud.
205 */
206 pud = pud_offset(pgd, address);
207 pud_k = pud_offset(pgd_k, address);
208 if (!pud_present(*pud_k))
209 return NULL;
210
211 pmd = pmd_offset(pud, address);
212 pmd_k = pmd_offset(pud_k, address);
213 if (!pmd_present(*pmd_k))
214 return NULL;
215
216 if (!pmd_present(*pmd))
217 set_pmd(pmd, *pmd_k);
218 else
219 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
220
221 return pmd_k;
222}
223
224void vmalloc_sync_all(void)
225{
226 unsigned long address;
227
228 if (SHARED_KERNEL_PMD)
229 return;
230
231 for (address = VMALLOC_START & PMD_MASK;
232 address >= TASK_SIZE && address < FIXADDR_TOP;
233 address += PMD_SIZE) {
234 struct page *page;
235
236 spin_lock(&pgd_lock);
237 list_for_each_entry(page, &pgd_list, lru) {
238 spinlock_t *pgt_lock;
239 pmd_t *ret;
240
241 /* the pgt_lock only for Xen */
242 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
243
244 spin_lock(pgt_lock);
245 ret = vmalloc_sync_one(page_address(page), address);
246 spin_unlock(pgt_lock);
247
248 if (!ret)
249 break;
250 }
251 spin_unlock(&pgd_lock);
252 }
253}
254
255/*
256 * 32-bit:
257 *
258 * Handle a fault on the vmalloc or module mapping area
259 */
260static noinline __kprobes int vmalloc_fault(unsigned long address)
261{
262 unsigned long pgd_paddr;
263 pmd_t *pmd_k;
264 pte_t *pte_k;
265
266 /* Make sure we are in vmalloc area: */
267 if (!(address >= VMALLOC_START && address < VMALLOC_END))
268 return -1;
269
270 WARN_ON_ONCE(in_nmi());
271
272 /*
273 * Synchronize this task's top level page-table
274 * with the 'reference' page table.
275 *
276 * Do _not_ use "current" here. We might be inside
277 * an interrupt in the middle of a task switch..
278 */
279 pgd_paddr = read_cr3();
280 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
281 if (!pmd_k)
282 return -1;
283
284 pte_k = pte_offset_kernel(pmd_k, address);
285 if (!pte_present(*pte_k))
286 return -1;
287
288 return 0;
289}
290
291/*
292 * Did it hit the DOS screen memory VA from vm86 mode?
293 */
294static inline void
295check_v8086_mode(struct pt_regs *regs, unsigned long address,
296 struct task_struct *tsk)
297{
298 unsigned long bit;
299
300 if (!v8086_mode(regs))
301 return;
302
303 bit = (address - 0xA0000) >> PAGE_SHIFT;
304 if (bit < 32)
305 tsk->thread.screen_bitmap |= 1 << bit;
306}
307
308static bool low_pfn(unsigned long pfn)
309{
310 return pfn < max_low_pfn;
311}
312
313static void dump_pagetable(unsigned long address)
314{
315 pgd_t *base = __va(read_cr3());
316 pgd_t *pgd = &base[pgd_index(address)];
317 pmd_t *pmd;
318 pte_t *pte;
319
320#ifdef CONFIG_X86_PAE
321 printk("*pdpt = %016Lx ", pgd_val(*pgd));
322 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
323 goto out;
324#endif
325 pmd = pmd_offset(pud_offset(pgd, address), address);
326 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
327
328 /*
329 * We must not directly access the pte in the highpte
330 * case if the page table is located in highmem.
331 * And let's rather not kmap-atomic the pte, just in case
332 * it's allocated already:
333 */
334 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
335 goto out;
336
337 pte = pte_offset_kernel(pmd, address);
338 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
339out:
340 printk("\n");
341}
342
343#else /* CONFIG_X86_64: */
344
345void vmalloc_sync_all(void)
346{
347 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
348}
349
350/*
351 * 64-bit:
352 *
353 * Handle a fault on the vmalloc area
354 *
355 * This assumes no large pages in there.
356 */
357static noinline __kprobes int vmalloc_fault(unsigned long address)
358{
359 pgd_t *pgd, *pgd_ref;
360 pud_t *pud, *pud_ref;
361 pmd_t *pmd, *pmd_ref;
362 pte_t *pte, *pte_ref;
363
364 /* Make sure we are in vmalloc area: */
365 if (!(address >= VMALLOC_START && address < VMALLOC_END))
366 return -1;
367
368 WARN_ON_ONCE(in_nmi());
369
370 /*
371 * Copy kernel mappings over when needed. This can also
372 * happen within a race in page table update. In the later
373 * case just flush:
374 */
375 pgd = pgd_offset(current->active_mm, address);
376 pgd_ref = pgd_offset_k(address);
377 if (pgd_none(*pgd_ref))
378 return -1;
379
380 if (pgd_none(*pgd))
381 set_pgd(pgd, *pgd_ref);
382 else
383 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
384
385 /*
386 * Below here mismatches are bugs because these lower tables
387 * are shared:
388 */
389
390 pud = pud_offset(pgd, address);
391 pud_ref = pud_offset(pgd_ref, address);
392 if (pud_none(*pud_ref))
393 return -1;
394
395 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
396 BUG();
397
398 pmd = pmd_offset(pud, address);
399 pmd_ref = pmd_offset(pud_ref, address);
400 if (pmd_none(*pmd_ref))
401 return -1;
402
403 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
404 BUG();
405
406 pte_ref = pte_offset_kernel(pmd_ref, address);
407 if (!pte_present(*pte_ref))
408 return -1;
409
410 pte = pte_offset_kernel(pmd, address);
411
412 /*
413 * Don't use pte_page here, because the mappings can point
414 * outside mem_map, and the NUMA hash lookup cannot handle
415 * that:
416 */
417 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
418 BUG();
419
420 return 0;
421}
422
423static const char errata93_warning[] =
424KERN_ERR
425"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
426"******* Working around it, but it may cause SEGVs or burn power.\n"
427"******* Please consider a BIOS update.\n"
428"******* Disabling USB legacy in the BIOS may also help.\n";
429
430/*
431 * No vm86 mode in 64-bit mode:
432 */
433static inline void
434check_v8086_mode(struct pt_regs *regs, unsigned long address,
435 struct task_struct *tsk)
436{
437}
438
439static int bad_address(void *p)
440{
441 unsigned long dummy;
442
443 return probe_kernel_address((unsigned long *)p, dummy);
444}
445
446static void dump_pagetable(unsigned long address)
447{
448 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
449 pgd_t *pgd = base + pgd_index(address);
450 pud_t *pud;
451 pmd_t *pmd;
452 pte_t *pte;
453
454 if (bad_address(pgd))
455 goto bad;
456
457 printk("PGD %lx ", pgd_val(*pgd));
458
459 if (!pgd_present(*pgd))
460 goto out;
461
462 pud = pud_offset(pgd, address);
463 if (bad_address(pud))
464 goto bad;
465
466 printk("PUD %lx ", pud_val(*pud));
467 if (!pud_present(*pud) || pud_large(*pud))
468 goto out;
469
470 pmd = pmd_offset(pud, address);
471 if (bad_address(pmd))
472 goto bad;
473
474 printk("PMD %lx ", pmd_val(*pmd));
475 if (!pmd_present(*pmd) || pmd_large(*pmd))
476 goto out;
477
478 pte = pte_offset_kernel(pmd, address);
479 if (bad_address(pte))
480 goto bad;
481
482 printk("PTE %lx", pte_val(*pte));
483out:
484 printk("\n");
485 return;
486bad:
487 printk("BAD\n");
488}
489
490#endif /* CONFIG_X86_64 */
491
492/*
493 * Workaround for K8 erratum #93 & buggy BIOS.
494 *
495 * BIOS SMM functions are required to use a specific workaround
496 * to avoid corruption of the 64bit RIP register on C stepping K8.
497 *
498 * A lot of BIOS that didn't get tested properly miss this.
499 *
500 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
501 * Try to work around it here.
502 *
503 * Note we only handle faults in kernel here.
504 * Does nothing on 32-bit.
505 */
506static int is_errata93(struct pt_regs *regs, unsigned long address)
507{
508#ifdef CONFIG_X86_64
509 if (address != regs->ip)
510 return 0;
511
512 if ((address >> 32) != 0)
513 return 0;
514
515 address |= 0xffffffffUL << 32;
516 if ((address >= (u64)_stext && address <= (u64)_etext) ||
517 (address >= MODULES_VADDR && address <= MODULES_END)) {
518 printk_once(errata93_warning);
519 regs->ip = address;
520 return 1;
521 }
522#endif
523 return 0;
524}
525
526/*
527 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
528 * to illegal addresses >4GB.
529 *
530 * We catch this in the page fault handler because these addresses
531 * are not reachable. Just detect this case and return. Any code
532 * segment in LDT is compatibility mode.
533 */
534static int is_errata100(struct pt_regs *regs, unsigned long address)
535{
536#ifdef CONFIG_X86_64
537 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
538 return 1;
539#endif
540 return 0;
541}
542
543static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
544{
545#ifdef CONFIG_X86_F00F_BUG
546 unsigned long nr;
547
548 /*
549 * Pentium F0 0F C7 C8 bug workaround:
550 */
551 if (boot_cpu_data.f00f_bug) {
552 nr = (address - idt_descr.address) >> 3;
553
554 if (nr == 6) {
555 do_invalid_op(regs, 0);
556 return 1;
557 }
558 }
559#endif
560 return 0;
561}
562
563static const char nx_warning[] = KERN_CRIT
564"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
565
566static void
567show_fault_oops(struct pt_regs *regs, unsigned long error_code,
568 unsigned long address)
569{
570 if (!oops_may_print())
571 return;
572
573 if (error_code & PF_INSTR) {
574 unsigned int level;
575
576 pte_t *pte = lookup_address(address, &level);
577
578 if (pte && pte_present(*pte) && !pte_exec(*pte))
579 printk(nx_warning, current_uid());
580 }
581
582 printk(KERN_ALERT "BUG: unable to handle kernel ");
583 if (address < PAGE_SIZE)
584 printk(KERN_CONT "NULL pointer dereference");
585 else
586 printk(KERN_CONT "paging request");
587
588 printk(KERN_CONT " at %p\n", (void *) address);
589 printk(KERN_ALERT "IP:");
590 printk_address(regs->ip, 1);
591
592 dump_pagetable(address);
593}
594
595static noinline void
596pgtable_bad(struct pt_regs *regs, unsigned long error_code,
597 unsigned long address)
598{
599 struct task_struct *tsk;
600 unsigned long flags;
601 int sig;
602
603 flags = oops_begin();
604 tsk = current;
605 sig = SIGKILL;
606
607 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
608 tsk->comm, address);
609 dump_pagetable(address);
610
611 tsk->thread.cr2 = address;
612 tsk->thread.trap_no = 14;
613 tsk->thread.error_code = error_code;
614
615 if (__die("Bad pagetable", regs, error_code))
616 sig = 0;
617
618 oops_end(flags, regs, sig);
619}
620
621static noinline void
622no_context(struct pt_regs *regs, unsigned long error_code,
623 unsigned long address)
624{
625 struct task_struct *tsk = current;
626 unsigned long *stackend;
627 unsigned long flags;
628 int sig;
629
630 /* Are we prepared to handle this kernel fault? */
631 if (fixup_exception(regs))
632 return;
633
634 /*
635 * 32-bit:
636 *
637 * Valid to do another page fault here, because if this fault
638 * had been triggered by is_prefetch fixup_exception would have
639 * handled it.
640 *
641 * 64-bit:
642 *
643 * Hall of shame of CPU/BIOS bugs.
644 */
645 if (is_prefetch(regs, error_code, address))
646 return;
647
648 if (is_errata93(regs, address))
649 return;
650
651 /*
652 * Oops. The kernel tried to access some bad page. We'll have to
653 * terminate things with extreme prejudice:
654 */
655 flags = oops_begin();
656
657 show_fault_oops(regs, error_code, address);
658
659 stackend = end_of_stack(tsk);
660 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
661 printk(KERN_ALERT "Thread overran stack, or stack corrupted\n");
662
663 tsk->thread.cr2 = address;
664 tsk->thread.trap_no = 14;
665 tsk->thread.error_code = error_code;
666
667 sig = SIGKILL;
668 if (__die("Oops", regs, error_code))
669 sig = 0;
670
671 /* Executive summary in case the body of the oops scrolled away */
672 printk(KERN_EMERG "CR2: %016lx\n", address);
673
674 oops_end(flags, regs, sig);
675}
676
677/*
678 * Print out info about fatal segfaults, if the show_unhandled_signals
679 * sysctl is set:
680 */
681static inline void
682show_signal_msg(struct pt_regs *regs, unsigned long error_code,
683 unsigned long address, struct task_struct *tsk)
684{
685 if (!unhandled_signal(tsk, SIGSEGV))
686 return;
687
688 if (!printk_ratelimit())
689 return;
690
691 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
692 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
693 tsk->comm, task_pid_nr(tsk), address,
694 (void *)regs->ip, (void *)regs->sp, error_code);
695
696 print_vma_addr(KERN_CONT " in ", regs->ip);
697
698 printk(KERN_CONT "\n");
699}
700
701static void
702__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
703 unsigned long address, int si_code)
704{
705 struct task_struct *tsk = current;
706
707 /* User mode accesses just cause a SIGSEGV */
708 if (error_code & PF_USER) {
709 /*
710 * It's possible to have interrupts off here:
711 */
712 local_irq_enable();
713
714 /*
715 * Valid to do another page fault here because this one came
716 * from user space:
717 */
718 if (is_prefetch(regs, error_code, address))
719 return;
720
721 if (is_errata100(regs, address))
722 return;
723
724#ifdef CONFIG_X86_64
725 /*
726 * Instruction fetch faults in the vsyscall page might need
727 * emulation.
728 */
729 if (unlikely((error_code & PF_INSTR) &&
730 ((address & ~0xfff) == VSYSCALL_START))) {
731 if (emulate_vsyscall(regs, address))
732 return;
733 }
734#endif
735
736 if (unlikely(show_unhandled_signals))
737 show_signal_msg(regs, error_code, address, tsk);
738
739 /* Kernel addresses are always protection faults: */
740 tsk->thread.cr2 = address;
741 tsk->thread.error_code = error_code | (address >= TASK_SIZE);
742 tsk->thread.trap_no = 14;
743
744 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
745
746 return;
747 }
748
749 if (is_f00f_bug(regs, address))
750 return;
751
752 no_context(regs, error_code, address);
753}
754
755static noinline void
756bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
757 unsigned long address)
758{
759 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
760}
761
762static void
763__bad_area(struct pt_regs *regs, unsigned long error_code,
764 unsigned long address, int si_code)
765{
766 struct mm_struct *mm = current->mm;
767
768 /*
769 * Something tried to access memory that isn't in our memory map..
770 * Fix it, but check if it's kernel or user first..
771 */
772 up_read(&mm->mmap_sem);
773
774 __bad_area_nosemaphore(regs, error_code, address, si_code);
775}
776
777static noinline void
778bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
779{
780 __bad_area(regs, error_code, address, SEGV_MAPERR);
781}
782
783static noinline void
784bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
785 unsigned long address)
786{
787 __bad_area(regs, error_code, address, SEGV_ACCERR);
788}
789
790/* TODO: fixup for "mm-invoke-oom-killer-from-page-fault.patch" */
791static void
792out_of_memory(struct pt_regs *regs, unsigned long error_code,
793 unsigned long address)
794{
795 /*
796 * We ran out of memory, call the OOM killer, and return the userspace
797 * (which will retry the fault, or kill us if we got oom-killed):
798 */
799 up_read(¤t->mm->mmap_sem);
800
801 pagefault_out_of_memory();
802}
803
804static void
805do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
806 unsigned int fault)
807{
808 struct task_struct *tsk = current;
809 struct mm_struct *mm = tsk->mm;
810 int code = BUS_ADRERR;
811
812 up_read(&mm->mmap_sem);
813
814 /* Kernel mode? Handle exceptions or die: */
815 if (!(error_code & PF_USER)) {
816 no_context(regs, error_code, address);
817 return;
818 }
819
820 /* User-space => ok to do another page fault: */
821 if (is_prefetch(regs, error_code, address))
822 return;
823
824 tsk->thread.cr2 = address;
825 tsk->thread.error_code = error_code;
826 tsk->thread.trap_no = 14;
827
828#ifdef CONFIG_MEMORY_FAILURE
829 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
830 printk(KERN_ERR
831 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
832 tsk->comm, tsk->pid, address);
833 code = BUS_MCEERR_AR;
834 }
835#endif
836 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
837}
838
839static noinline int
840mm_fault_error(struct pt_regs *regs, unsigned long error_code,
841 unsigned long address, unsigned int fault)
842{
843 /*
844 * Pagefault was interrupted by SIGKILL. We have no reason to
845 * continue pagefault.
846 */
847 if (fatal_signal_pending(current)) {
848 if (!(fault & VM_FAULT_RETRY))
849 up_read(¤t->mm->mmap_sem);
850 if (!(error_code & PF_USER))
851 no_context(regs, error_code, address);
852 return 1;
853 }
854 if (!(fault & VM_FAULT_ERROR))
855 return 0;
856
857 if (fault & VM_FAULT_OOM) {
858 /* Kernel mode? Handle exceptions or die: */
859 if (!(error_code & PF_USER)) {
860 up_read(¤t->mm->mmap_sem);
861 no_context(regs, error_code, address);
862 return 1;
863 }
864
865 out_of_memory(regs, error_code, address);
866 } else {
867 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
868 VM_FAULT_HWPOISON_LARGE))
869 do_sigbus(regs, error_code, address, fault);
870 else
871 BUG();
872 }
873 return 1;
874}
875
876static int spurious_fault_check(unsigned long error_code, pte_t *pte)
877{
878 if ((error_code & PF_WRITE) && !pte_write(*pte))
879 return 0;
880
881 if ((error_code & PF_INSTR) && !pte_exec(*pte))
882 return 0;
883
884 return 1;
885}
886
887/*
888 * Handle a spurious fault caused by a stale TLB entry.
889 *
890 * This allows us to lazily refresh the TLB when increasing the
891 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
892 * eagerly is very expensive since that implies doing a full
893 * cross-processor TLB flush, even if no stale TLB entries exist
894 * on other processors.
895 *
896 * There are no security implications to leaving a stale TLB when
897 * increasing the permissions on a page.
898 */
899static noinline __kprobes int
900spurious_fault(unsigned long error_code, unsigned long address)
901{
902 pgd_t *pgd;
903 pud_t *pud;
904 pmd_t *pmd;
905 pte_t *pte;
906 int ret;
907
908 /* Reserved-bit violation or user access to kernel space? */
909 if (error_code & (PF_USER | PF_RSVD))
910 return 0;
911
912 pgd = init_mm.pgd + pgd_index(address);
913 if (!pgd_present(*pgd))
914 return 0;
915
916 pud = pud_offset(pgd, address);
917 if (!pud_present(*pud))
918 return 0;
919
920 if (pud_large(*pud))
921 return spurious_fault_check(error_code, (pte_t *) pud);
922
923 pmd = pmd_offset(pud, address);
924 if (!pmd_present(*pmd))
925 return 0;
926
927 if (pmd_large(*pmd))
928 return spurious_fault_check(error_code, (pte_t *) pmd);
929
930 /*
931 * Note: don't use pte_present() here, since it returns true
932 * if the _PAGE_PROTNONE bit is set. However, this aliases the
933 * _PAGE_GLOBAL bit, which for kernel pages give false positives
934 * when CONFIG_DEBUG_PAGEALLOC is used.
935 */
936 pte = pte_offset_kernel(pmd, address);
937 if (!(pte_flags(*pte) & _PAGE_PRESENT))
938 return 0;
939
940 ret = spurious_fault_check(error_code, pte);
941 if (!ret)
942 return 0;
943
944 /*
945 * Make sure we have permissions in PMD.
946 * If not, then there's a bug in the page tables:
947 */
948 ret = spurious_fault_check(error_code, (pte_t *) pmd);
949 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
950
951 return ret;
952}
953
954int show_unhandled_signals = 1;
955
956static inline int
957access_error(unsigned long error_code, struct vm_area_struct *vma)
958{
959 if (error_code & PF_WRITE) {
960 /* write, present and write, not present: */
961 if (unlikely(!(vma->vm_flags & VM_WRITE)))
962 return 1;
963 return 0;
964 }
965
966 /* read, present: */
967 if (unlikely(error_code & PF_PROT))
968 return 1;
969
970 /* read, not present: */
971 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
972 return 1;
973
974 return 0;
975}
976
977static int fault_in_kernel_space(unsigned long address)
978{
979 return address >= TASK_SIZE_MAX;
980}
981
982/*
983 * This routine handles page faults. It determines the address,
984 * and the problem, and then passes it off to one of the appropriate
985 * routines.
986 */
987dotraplinkage void __kprobes
988do_page_fault(struct pt_regs *regs, unsigned long error_code)
989{
990 struct vm_area_struct *vma;
991 struct task_struct *tsk;
992 unsigned long address;
993 struct mm_struct *mm;
994 int fault;
995 int write = error_code & PF_WRITE;
996 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE |
997 (write ? FAULT_FLAG_WRITE : 0);
998
999 tsk = current;
1000 mm = tsk->mm;
1001
1002 /* Get the faulting address: */
1003 address = read_cr2();
1004
1005 /*
1006 * Detect and handle instructions that would cause a page fault for
1007 * both a tracked kernel page and a userspace page.
1008 */
1009 if (kmemcheck_active(regs))
1010 kmemcheck_hide(regs);
1011 prefetchw(&mm->mmap_sem);
1012
1013 if (unlikely(kmmio_fault(regs, address)))
1014 return;
1015
1016 /*
1017 * We fault-in kernel-space virtual memory on-demand. The
1018 * 'reference' page table is init_mm.pgd.
1019 *
1020 * NOTE! We MUST NOT take any locks for this case. We may
1021 * be in an interrupt or a critical region, and should
1022 * only copy the information from the master page table,
1023 * nothing more.
1024 *
1025 * This verifies that the fault happens in kernel space
1026 * (error_code & 4) == 0, and that the fault was not a
1027 * protection error (error_code & 9) == 0.
1028 */
1029 if (unlikely(fault_in_kernel_space(address))) {
1030 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1031 if (vmalloc_fault(address) >= 0)
1032 return;
1033
1034 if (kmemcheck_fault(regs, address, error_code))
1035 return;
1036 }
1037
1038 /* Can handle a stale RO->RW TLB: */
1039 if (spurious_fault(error_code, address))
1040 return;
1041
1042 /* kprobes don't want to hook the spurious faults: */
1043 if (notify_page_fault(regs))
1044 return;
1045 /*
1046 * Don't take the mm semaphore here. If we fixup a prefetch
1047 * fault we could otherwise deadlock:
1048 */
1049 bad_area_nosemaphore(regs, error_code, address);
1050
1051 return;
1052 }
1053
1054 /* kprobes don't want to hook the spurious faults: */
1055 if (unlikely(notify_page_fault(regs)))
1056 return;
1057 /*
1058 * It's safe to allow irq's after cr2 has been saved and the
1059 * vmalloc fault has been handled.
1060 *
1061 * User-mode registers count as a user access even for any
1062 * potential system fault or CPU buglet:
1063 */
1064 if (user_mode_vm(regs)) {
1065 local_irq_enable();
1066 error_code |= PF_USER;
1067 } else {
1068 if (regs->flags & X86_EFLAGS_IF)
1069 local_irq_enable();
1070 }
1071
1072 if (unlikely(error_code & PF_RSVD))
1073 pgtable_bad(regs, error_code, address);
1074
1075 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1076
1077 /*
1078 * If we're in an interrupt, have no user context or are running
1079 * in an atomic region then we must not take the fault:
1080 */
1081 if (unlikely(in_atomic() || !mm)) {
1082 bad_area_nosemaphore(regs, error_code, address);
1083 return;
1084 }
1085
1086 /*
1087 * When running in the kernel we expect faults to occur only to
1088 * addresses in user space. All other faults represent errors in
1089 * the kernel and should generate an OOPS. Unfortunately, in the
1090 * case of an erroneous fault occurring in a code path which already
1091 * holds mmap_sem we will deadlock attempting to validate the fault
1092 * against the address space. Luckily the kernel only validly
1093 * references user space from well defined areas of code, which are
1094 * listed in the exceptions table.
1095 *
1096 * As the vast majority of faults will be valid we will only perform
1097 * the source reference check when there is a possibility of a
1098 * deadlock. Attempt to lock the address space, if we cannot we then
1099 * validate the source. If this is invalid we can skip the address
1100 * space check, thus avoiding the deadlock:
1101 */
1102 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1103 if ((error_code & PF_USER) == 0 &&
1104 !search_exception_tables(regs->ip)) {
1105 bad_area_nosemaphore(regs, error_code, address);
1106 return;
1107 }
1108retry:
1109 down_read(&mm->mmap_sem);
1110 } else {
1111 /*
1112 * The above down_read_trylock() might have succeeded in
1113 * which case we'll have missed the might_sleep() from
1114 * down_read():
1115 */
1116 might_sleep();
1117 }
1118
1119 vma = find_vma(mm, address);
1120 if (unlikely(!vma)) {
1121 bad_area(regs, error_code, address);
1122 return;
1123 }
1124 if (likely(vma->vm_start <= address))
1125 goto good_area;
1126 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1127 bad_area(regs, error_code, address);
1128 return;
1129 }
1130 if (error_code & PF_USER) {
1131 /*
1132 * Accessing the stack below %sp is always a bug.
1133 * The large cushion allows instructions like enter
1134 * and pusha to work. ("enter $65535, $31" pushes
1135 * 32 pointers and then decrements %sp by 65535.)
1136 */
1137 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1138 bad_area(regs, error_code, address);
1139 return;
1140 }
1141 }
1142 if (unlikely(expand_stack(vma, address))) {
1143 bad_area(regs, error_code, address);
1144 return;
1145 }
1146
1147 /*
1148 * Ok, we have a good vm_area for this memory access, so
1149 * we can handle it..
1150 */
1151good_area:
1152 if (unlikely(access_error(error_code, vma))) {
1153 bad_area_access_error(regs, error_code, address);
1154 return;
1155 }
1156
1157 /*
1158 * If for any reason at all we couldn't handle the fault,
1159 * make sure we exit gracefully rather than endlessly redo
1160 * the fault:
1161 */
1162 fault = handle_mm_fault(mm, vma, address, flags);
1163
1164 if (unlikely(fault & (VM_FAULT_RETRY|VM_FAULT_ERROR))) {
1165 if (mm_fault_error(regs, error_code, address, fault))
1166 return;
1167 }
1168
1169 /*
1170 * Major/minor page fault accounting is only done on the
1171 * initial attempt. If we go through a retry, it is extremely
1172 * likely that the page will be found in page cache at that point.
1173 */
1174 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1175 if (fault & VM_FAULT_MAJOR) {
1176 tsk->maj_flt++;
1177 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1178 regs, address);
1179 } else {
1180 tsk->min_flt++;
1181 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1182 regs, address);
1183 }
1184 if (fault & VM_FAULT_RETRY) {
1185 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1186 * of starvation. */
1187 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1188 goto retry;
1189 }
1190 }
1191
1192 check_v8086_mode(regs, address, tsk);
1193
1194 up_read(&mm->mmap_sem);
1195}
1// SPDX-License-Identifier: GPL-2.0
2/*
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
6 */
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/kfence.h> /* kfence_handle_page_fault */
13#include <linux/kprobes.h> /* NOKPROBE_SYMBOL, ... */
14#include <linux/mmiotrace.h> /* kmmio_handler, ... */
15#include <linux/perf_event.h> /* perf_sw_event */
16#include <linux/hugetlb.h> /* hstate_index_to_shift */
17#include <linux/prefetch.h> /* prefetchw */
18#include <linux/context_tracking.h> /* exception_enter(), ... */
19#include <linux/uaccess.h> /* faulthandler_disabled() */
20#include <linux/efi.h> /* efi_crash_gracefully_on_page_fault()*/
21#include <linux/mm_types.h>
22
23#include <asm/cpufeature.h> /* boot_cpu_has, ... */
24#include <asm/traps.h> /* dotraplinkage, ... */
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_crash_gracefully_on_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 */
34#include <asm/vdso.h> /* fixup_vdso_exception() */
35
36#define CREATE_TRACE_POINTS
37#include <asm/trace/exceptions.h>
38
39/*
40 * Returns 0 if mmiotrace is disabled, or if the fault is not
41 * handled by mmiotrace:
42 */
43static nokprobe_inline int
44kmmio_fault(struct pt_regs *regs, unsigned long addr)
45{
46 if (unlikely(is_kmmio_active()))
47 if (kmmio_handler(regs, addr) == 1)
48 return -1;
49 return 0;
50}
51
52/*
53 * Prefetch quirks:
54 *
55 * 32-bit mode:
56 *
57 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
58 * Check that here and ignore it. This is AMD erratum #91.
59 *
60 * 64-bit mode:
61 *
62 * Sometimes the CPU reports invalid exceptions on prefetch.
63 * Check that here and ignore it.
64 *
65 * Opcode checker based on code by Richard Brunner.
66 */
67static inline int
68check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
69 unsigned char opcode, int *prefetch)
70{
71 unsigned char instr_hi = opcode & 0xf0;
72 unsigned char instr_lo = opcode & 0x0f;
73
74 switch (instr_hi) {
75 case 0x20:
76 case 0x30:
77 /*
78 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
79 * In X86_64 long mode, the CPU will signal invalid
80 * opcode if some of these prefixes are present so
81 * X86_64 will never get here anyway
82 */
83 return ((instr_lo & 7) == 0x6);
84#ifdef CONFIG_X86_64
85 case 0x40:
86 /*
87 * In 64-bit mode 0x40..0x4F are valid REX prefixes
88 */
89 return (!user_mode(regs) || user_64bit_mode(regs));
90#endif
91 case 0x60:
92 /* 0x64 thru 0x67 are valid prefixes in all modes. */
93 return (instr_lo & 0xC) == 0x4;
94 case 0xF0:
95 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
96 return !instr_lo || (instr_lo>>1) == 1;
97 case 0x00:
98 /* Prefetch instruction is 0x0F0D or 0x0F18 */
99 if (get_kernel_nofault(opcode, instr))
100 return 0;
101
102 *prefetch = (instr_lo == 0xF) &&
103 (opcode == 0x0D || opcode == 0x18);
104 return 0;
105 default:
106 return 0;
107 }
108}
109
110static bool is_amd_k8_pre_npt(void)
111{
112 struct cpuinfo_x86 *c = &boot_cpu_data;
113
114 return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
115 c->x86_vendor == X86_VENDOR_AMD &&
116 c->x86 == 0xf && c->x86_model < 0x40);
117}
118
119static int
120is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
121{
122 unsigned char *max_instr;
123 unsigned char *instr;
124 int prefetch = 0;
125
126 /* Erratum #91 affects AMD K8, pre-NPT CPUs */
127 if (!is_amd_k8_pre_npt())
128 return 0;
129
130 /*
131 * If it was a exec (instruction fetch) fault on NX page, then
132 * do not ignore the fault:
133 */
134 if (error_code & X86_PF_INSTR)
135 return 0;
136
137 instr = (void *)convert_ip_to_linear(current, regs);
138 max_instr = instr + 15;
139
140 /*
141 * This code has historically always bailed out if IP points to a
142 * not-present page (e.g. due to a race). No one has ever
143 * complained about this.
144 */
145 pagefault_disable();
146
147 while (instr < max_instr) {
148 unsigned char opcode;
149
150 if (user_mode(regs)) {
151 if (get_user(opcode, instr))
152 break;
153 } else {
154 if (get_kernel_nofault(opcode, instr))
155 break;
156 }
157
158 instr++;
159
160 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
161 break;
162 }
163
164 pagefault_enable();
165 return prefetch;
166}
167
168DEFINE_SPINLOCK(pgd_lock);
169LIST_HEAD(pgd_list);
170
171#ifdef CONFIG_X86_32
172static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
173{
174 unsigned index = pgd_index(address);
175 pgd_t *pgd_k;
176 p4d_t *p4d, *p4d_k;
177 pud_t *pud, *pud_k;
178 pmd_t *pmd, *pmd_k;
179
180 pgd += index;
181 pgd_k = init_mm.pgd + index;
182
183 if (!pgd_present(*pgd_k))
184 return NULL;
185
186 /*
187 * set_pgd(pgd, *pgd_k); here would be useless on PAE
188 * and redundant with the set_pmd() on non-PAE. As would
189 * set_p4d/set_pud.
190 */
191 p4d = p4d_offset(pgd, address);
192 p4d_k = p4d_offset(pgd_k, address);
193 if (!p4d_present(*p4d_k))
194 return NULL;
195
196 pud = pud_offset(p4d, address);
197 pud_k = pud_offset(p4d_k, address);
198 if (!pud_present(*pud_k))
199 return NULL;
200
201 pmd = pmd_offset(pud, address);
202 pmd_k = pmd_offset(pud_k, address);
203
204 if (pmd_present(*pmd) != pmd_present(*pmd_k))
205 set_pmd(pmd, *pmd_k);
206
207 if (!pmd_present(*pmd_k))
208 return NULL;
209 else
210 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
211
212 return pmd_k;
213}
214
215/*
216 * Handle a fault on the vmalloc or module mapping area
217 *
218 * This is needed because there is a race condition between the time
219 * when the vmalloc mapping code updates the PMD to the point in time
220 * where it synchronizes this update with the other page-tables in the
221 * system.
222 *
223 * In this race window another thread/CPU can map an area on the same
224 * PMD, finds it already present and does not synchronize it with the
225 * rest of the system yet. As a result v[mz]alloc might return areas
226 * which are not mapped in every page-table in the system, causing an
227 * unhandled page-fault when they are accessed.
228 */
229static noinline int vmalloc_fault(unsigned long address)
230{
231 unsigned long pgd_paddr;
232 pmd_t *pmd_k;
233 pte_t *pte_k;
234
235 /* Make sure we are in vmalloc area: */
236 if (!(address >= VMALLOC_START && address < VMALLOC_END))
237 return -1;
238
239 /*
240 * Synchronize this task's top level page-table
241 * with the 'reference' page table.
242 *
243 * Do _not_ use "current" here. We might be inside
244 * an interrupt in the middle of a task switch..
245 */
246 pgd_paddr = read_cr3_pa();
247 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
248 if (!pmd_k)
249 return -1;
250
251 if (pmd_large(*pmd_k))
252 return 0;
253
254 pte_k = pte_offset_kernel(pmd_k, address);
255 if (!pte_present(*pte_k))
256 return -1;
257
258 return 0;
259}
260NOKPROBE_SYMBOL(vmalloc_fault);
261
262void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
263{
264 unsigned long addr;
265
266 for (addr = start & PMD_MASK;
267 addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
268 addr += PMD_SIZE) {
269 struct page *page;
270
271 spin_lock(&pgd_lock);
272 list_for_each_entry(page, &pgd_list, lru) {
273 spinlock_t *pgt_lock;
274
275 /* the pgt_lock only for Xen */
276 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
277
278 spin_lock(pgt_lock);
279 vmalloc_sync_one(page_address(page), addr);
280 spin_unlock(pgt_lock);
281 }
282 spin_unlock(&pgd_lock);
283 }
284}
285
286static bool low_pfn(unsigned long pfn)
287{
288 return pfn < max_low_pfn;
289}
290
291static void dump_pagetable(unsigned long address)
292{
293 pgd_t *base = __va(read_cr3_pa());
294 pgd_t *pgd = &base[pgd_index(address)];
295 p4d_t *p4d;
296 pud_t *pud;
297 pmd_t *pmd;
298 pte_t *pte;
299
300#ifdef CONFIG_X86_PAE
301 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
302 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
303 goto out;
304#define pr_pde pr_cont
305#else
306#define pr_pde pr_info
307#endif
308 p4d = p4d_offset(pgd, address);
309 pud = pud_offset(p4d, address);
310 pmd = pmd_offset(pud, address);
311 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
312#undef pr_pde
313
314 /*
315 * We must not directly access the pte in the highpte
316 * case if the page table is located in highmem.
317 * And let's rather not kmap-atomic the pte, just in case
318 * it's allocated already:
319 */
320 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
321 goto out;
322
323 pte = pte_offset_kernel(pmd, address);
324 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
325out:
326 pr_cont("\n");
327}
328
329#else /* CONFIG_X86_64: */
330
331#ifdef CONFIG_CPU_SUP_AMD
332static const char errata93_warning[] =
333KERN_ERR
334"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
335"******* Working around it, but it may cause SEGVs or burn power.\n"
336"******* Please consider a BIOS update.\n"
337"******* Disabling USB legacy in the BIOS may also help.\n";
338#endif
339
340static int bad_address(void *p)
341{
342 unsigned long dummy;
343
344 return get_kernel_nofault(dummy, (unsigned long *)p);
345}
346
347static void dump_pagetable(unsigned long address)
348{
349 pgd_t *base = __va(read_cr3_pa());
350 pgd_t *pgd = base + pgd_index(address);
351 p4d_t *p4d;
352 pud_t *pud;
353 pmd_t *pmd;
354 pte_t *pte;
355
356 if (bad_address(pgd))
357 goto bad;
358
359 pr_info("PGD %lx ", pgd_val(*pgd));
360
361 if (!pgd_present(*pgd))
362 goto out;
363
364 p4d = p4d_offset(pgd, address);
365 if (bad_address(p4d))
366 goto bad;
367
368 pr_cont("P4D %lx ", p4d_val(*p4d));
369 if (!p4d_present(*p4d) || p4d_large(*p4d))
370 goto out;
371
372 pud = pud_offset(p4d, address);
373 if (bad_address(pud))
374 goto bad;
375
376 pr_cont("PUD %lx ", pud_val(*pud));
377 if (!pud_present(*pud) || pud_large(*pud))
378 goto out;
379
380 pmd = pmd_offset(pud, address);
381 if (bad_address(pmd))
382 goto bad;
383
384 pr_cont("PMD %lx ", pmd_val(*pmd));
385 if (!pmd_present(*pmd) || pmd_large(*pmd))
386 goto out;
387
388 pte = pte_offset_kernel(pmd, address);
389 if (bad_address(pte))
390 goto bad;
391
392 pr_cont("PTE %lx", pte_val(*pte));
393out:
394 pr_cont("\n");
395 return;
396bad:
397 pr_info("BAD\n");
398}
399
400#endif /* CONFIG_X86_64 */
401
402/*
403 * Workaround for K8 erratum #93 & buggy BIOS.
404 *
405 * BIOS SMM functions are required to use a specific workaround
406 * to avoid corruption of the 64bit RIP register on C stepping K8.
407 *
408 * A lot of BIOS that didn't get tested properly miss this.
409 *
410 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
411 * Try to work around it here.
412 *
413 * Note we only handle faults in kernel here.
414 * Does nothing on 32-bit.
415 */
416static int is_errata93(struct pt_regs *regs, unsigned long address)
417{
418#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
419 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
420 || boot_cpu_data.x86 != 0xf)
421 return 0;
422
423 if (user_mode(regs))
424 return 0;
425
426 if (address != regs->ip)
427 return 0;
428
429 if ((address >> 32) != 0)
430 return 0;
431
432 address |= 0xffffffffUL << 32;
433 if ((address >= (u64)_stext && address <= (u64)_etext) ||
434 (address >= MODULES_VADDR && address <= MODULES_END)) {
435 printk_once(errata93_warning);
436 regs->ip = address;
437 return 1;
438 }
439#endif
440 return 0;
441}
442
443/*
444 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
445 * to illegal addresses >4GB.
446 *
447 * We catch this in the page fault handler because these addresses
448 * are not reachable. Just detect this case and return. Any code
449 * segment in LDT is compatibility mode.
450 */
451static int is_errata100(struct pt_regs *regs, unsigned long address)
452{
453#ifdef CONFIG_X86_64
454 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
455 return 1;
456#endif
457 return 0;
458}
459
460/* Pentium F0 0F C7 C8 bug workaround: */
461static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
462 unsigned long address)
463{
464#ifdef CONFIG_X86_F00F_BUG
465 if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
466 idt_is_f00f_address(address)) {
467 handle_invalid_op(regs);
468 return 1;
469 }
470#endif
471 return 0;
472}
473
474static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
475{
476 u32 offset = (index >> 3) * sizeof(struct desc_struct);
477 unsigned long addr;
478 struct ldttss_desc desc;
479
480 if (index == 0) {
481 pr_alert("%s: NULL\n", name);
482 return;
483 }
484
485 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
486 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
487 return;
488 }
489
490 if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset),
491 sizeof(struct ldttss_desc))) {
492 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
493 name, index);
494 return;
495 }
496
497 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
498#ifdef CONFIG_X86_64
499 addr |= ((u64)desc.base3 << 32);
500#endif
501 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
502 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
503}
504
505static void
506show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
507{
508 if (!oops_may_print())
509 return;
510
511 if (error_code & X86_PF_INSTR) {
512 unsigned int level;
513 pgd_t *pgd;
514 pte_t *pte;
515
516 pgd = __va(read_cr3_pa());
517 pgd += pgd_index(address);
518
519 pte = lookup_address_in_pgd(pgd, address, &level);
520
521 if (pte && pte_present(*pte) && !pte_exec(*pte))
522 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
523 from_kuid(&init_user_ns, current_uid()));
524 if (pte && pte_present(*pte) && pte_exec(*pte) &&
525 (pgd_flags(*pgd) & _PAGE_USER) &&
526 (__read_cr4() & X86_CR4_SMEP))
527 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
528 from_kuid(&init_user_ns, current_uid()));
529 }
530
531 if (address < PAGE_SIZE && !user_mode(regs))
532 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
533 (void *)address);
534 else
535 pr_alert("BUG: unable to handle page fault for address: %px\n",
536 (void *)address);
537
538 pr_alert("#PF: %s %s in %s mode\n",
539 (error_code & X86_PF_USER) ? "user" : "supervisor",
540 (error_code & X86_PF_INSTR) ? "instruction fetch" :
541 (error_code & X86_PF_WRITE) ? "write access" :
542 "read access",
543 user_mode(regs) ? "user" : "kernel");
544 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
545 !(error_code & X86_PF_PROT) ? "not-present page" :
546 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
547 (error_code & X86_PF_PK) ? "protection keys violation" :
548 "permissions violation");
549
550 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
551 struct desc_ptr idt, gdt;
552 u16 ldtr, tr;
553
554 /*
555 * This can happen for quite a few reasons. The more obvious
556 * ones are faults accessing the GDT, or LDT. Perhaps
557 * surprisingly, if the CPU tries to deliver a benign or
558 * contributory exception from user code and gets a page fault
559 * during delivery, the page fault can be delivered as though
560 * it originated directly from user code. This could happen
561 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
562 * kernel or IST stack.
563 */
564 store_idt(&idt);
565
566 /* Usable even on Xen PV -- it's just slow. */
567 native_store_gdt(&gdt);
568
569 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
570 idt.address, idt.size, gdt.address, gdt.size);
571
572 store_ldt(ldtr);
573 show_ldttss(&gdt, "LDTR", ldtr);
574
575 store_tr(tr);
576 show_ldttss(&gdt, "TR", tr);
577 }
578
579 dump_pagetable(address);
580}
581
582static noinline void
583pgtable_bad(struct pt_regs *regs, unsigned long error_code,
584 unsigned long address)
585{
586 struct task_struct *tsk;
587 unsigned long flags;
588 int sig;
589
590 flags = oops_begin();
591 tsk = current;
592 sig = SIGKILL;
593
594 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
595 tsk->comm, address);
596 dump_pagetable(address);
597
598 if (__die("Bad pagetable", regs, error_code))
599 sig = 0;
600
601 oops_end(flags, regs, sig);
602}
603
604static void sanitize_error_code(unsigned long address,
605 unsigned long *error_code)
606{
607 /*
608 * To avoid leaking information about the kernel page
609 * table layout, pretend that user-mode accesses to
610 * kernel addresses are always protection faults.
611 *
612 * NB: This means that failed vsyscalls with vsyscall=none
613 * will have the PROT bit. This doesn't leak any
614 * information and does not appear to cause any problems.
615 */
616 if (address >= TASK_SIZE_MAX)
617 *error_code |= X86_PF_PROT;
618}
619
620static void set_signal_archinfo(unsigned long address,
621 unsigned long error_code)
622{
623 struct task_struct *tsk = current;
624
625 tsk->thread.trap_nr = X86_TRAP_PF;
626 tsk->thread.error_code = error_code | X86_PF_USER;
627 tsk->thread.cr2 = address;
628}
629
630static noinline void
631page_fault_oops(struct pt_regs *regs, unsigned long error_code,
632 unsigned long address)
633{
634 unsigned long flags;
635 int sig;
636
637 if (user_mode(regs)) {
638 /*
639 * Implicit kernel access from user mode? Skip the stack
640 * overflow and EFI special cases.
641 */
642 goto oops;
643 }
644
645#ifdef CONFIG_VMAP_STACK
646 /*
647 * Stack overflow? During boot, we can fault near the initial
648 * stack in the direct map, but that's not an overflow -- check
649 * that we're in vmalloc space to avoid this.
650 */
651 if (is_vmalloc_addr((void *)address) &&
652 (((unsigned long)current->stack - 1 - address < PAGE_SIZE) ||
653 address - ((unsigned long)current->stack + THREAD_SIZE) < PAGE_SIZE)) {
654 unsigned long stack = __this_cpu_ist_top_va(DF) - sizeof(void *);
655 /*
656 * We're likely to be running with very little stack space
657 * left. It's plausible that we'd hit this condition but
658 * double-fault even before we get this far, in which case
659 * we're fine: the double-fault handler will deal with it.
660 *
661 * We don't want to make it all the way into the oops code
662 * and then double-fault, though, because we're likely to
663 * break the console driver and lose most of the stack dump.
664 */
665 asm volatile ("movq %[stack], %%rsp\n\t"
666 "call handle_stack_overflow\n\t"
667 "1: jmp 1b"
668 : ASM_CALL_CONSTRAINT
669 : "D" ("kernel stack overflow (page fault)"),
670 "S" (regs), "d" (address),
671 [stack] "rm" (stack));
672 unreachable();
673 }
674#endif
675
676 /*
677 * Buggy firmware could access regions which might page fault. If
678 * this happens, EFI has a special OOPS path that will try to
679 * avoid hanging the system.
680 */
681 if (IS_ENABLED(CONFIG_EFI))
682 efi_crash_gracefully_on_page_fault(address);
683
684 /* Only not-present faults should be handled by KFENCE. */
685 if (!(error_code & X86_PF_PROT) &&
686 kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs))
687 return;
688
689oops:
690 /*
691 * Oops. The kernel tried to access some bad page. We'll have to
692 * terminate things with extreme prejudice:
693 */
694 flags = oops_begin();
695
696 show_fault_oops(regs, error_code, address);
697
698 if (task_stack_end_corrupted(current))
699 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
700
701 sig = SIGKILL;
702 if (__die("Oops", regs, error_code))
703 sig = 0;
704
705 /* Executive summary in case the body of the oops scrolled away */
706 printk(KERN_DEFAULT "CR2: %016lx\n", address);
707
708 oops_end(flags, regs, sig);
709}
710
711static noinline void
712kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
713 unsigned long address, int signal, int si_code,
714 u32 pkey)
715{
716 WARN_ON_ONCE(user_mode(regs));
717
718 /* Are we prepared to handle this kernel fault? */
719 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
720 /*
721 * Any interrupt that takes a fault gets the fixup. This makes
722 * the below recursive fault logic only apply to a faults from
723 * task context.
724 */
725 if (in_interrupt())
726 return;
727
728 /*
729 * Per the above we're !in_interrupt(), aka. task context.
730 *
731 * In this case we need to make sure we're not recursively
732 * faulting through the emulate_vsyscall() logic.
733 */
734 if (current->thread.sig_on_uaccess_err && signal) {
735 sanitize_error_code(address, &error_code);
736
737 set_signal_archinfo(address, error_code);
738
739 if (si_code == SEGV_PKUERR) {
740 force_sig_pkuerr((void __user *)address, pkey);
741 } else {
742 /* XXX: hwpoison faults will set the wrong code. */
743 force_sig_fault(signal, si_code, (void __user *)address);
744 }
745 }
746
747 /*
748 * Barring that, we can do the fixup and be happy.
749 */
750 return;
751 }
752
753 /*
754 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
755 * instruction.
756 */
757 if (is_prefetch(regs, error_code, address))
758 return;
759
760 page_fault_oops(regs, error_code, address);
761}
762
763/*
764 * Print out info about fatal segfaults, if the show_unhandled_signals
765 * sysctl is set:
766 */
767static inline void
768show_signal_msg(struct pt_regs *regs, unsigned long error_code,
769 unsigned long address, struct task_struct *tsk)
770{
771 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
772
773 if (!unhandled_signal(tsk, SIGSEGV))
774 return;
775
776 if (!printk_ratelimit())
777 return;
778
779 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
780 loglvl, tsk->comm, task_pid_nr(tsk), address,
781 (void *)regs->ip, (void *)regs->sp, error_code);
782
783 print_vma_addr(KERN_CONT " in ", regs->ip);
784
785 printk(KERN_CONT "\n");
786
787 show_opcodes(regs, loglvl);
788}
789
790/*
791 * The (legacy) vsyscall page is the long page in the kernel portion
792 * of the address space that has user-accessible permissions.
793 */
794static bool is_vsyscall_vaddr(unsigned long vaddr)
795{
796 return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR);
797}
798
799static void
800__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
801 unsigned long address, u32 pkey, int si_code)
802{
803 struct task_struct *tsk = current;
804
805 if (!user_mode(regs)) {
806 kernelmode_fixup_or_oops(regs, error_code, address,
807 SIGSEGV, si_code, pkey);
808 return;
809 }
810
811 if (!(error_code & X86_PF_USER)) {
812 /* Implicit user access to kernel memory -- just oops */
813 page_fault_oops(regs, error_code, address);
814 return;
815 }
816
817 /*
818 * User mode accesses just cause a SIGSEGV.
819 * It's possible to have interrupts off here:
820 */
821 local_irq_enable();
822
823 /*
824 * Valid to do another page fault here because this one came
825 * from user space:
826 */
827 if (is_prefetch(regs, error_code, address))
828 return;
829
830 if (is_errata100(regs, address))
831 return;
832
833 sanitize_error_code(address, &error_code);
834
835 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
836 return;
837
838 if (likely(show_unhandled_signals))
839 show_signal_msg(regs, error_code, address, tsk);
840
841 set_signal_archinfo(address, error_code);
842
843 if (si_code == SEGV_PKUERR)
844 force_sig_pkuerr((void __user *)address, pkey);
845 else
846 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
847
848 local_irq_disable();
849}
850
851static noinline void
852bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
853 unsigned long address)
854{
855 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
856}
857
858static void
859__bad_area(struct pt_regs *regs, unsigned long error_code,
860 unsigned long address, u32 pkey, int si_code)
861{
862 struct mm_struct *mm = current->mm;
863 /*
864 * Something tried to access memory that isn't in our memory map..
865 * Fix it, but check if it's kernel or user first..
866 */
867 mmap_read_unlock(mm);
868
869 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
870}
871
872static noinline void
873bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
874{
875 __bad_area(regs, error_code, address, 0, SEGV_MAPERR);
876}
877
878static inline bool bad_area_access_from_pkeys(unsigned long error_code,
879 struct vm_area_struct *vma)
880{
881 /* This code is always called on the current mm */
882 bool foreign = false;
883
884 if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
885 return false;
886 if (error_code & X86_PF_PK)
887 return true;
888 /* this checks permission keys on the VMA: */
889 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
890 (error_code & X86_PF_INSTR), foreign))
891 return true;
892 return false;
893}
894
895static noinline void
896bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
897 unsigned long address, struct vm_area_struct *vma)
898{
899 /*
900 * This OSPKE check is not strictly necessary at runtime.
901 * But, doing it this way allows compiler optimizations
902 * if pkeys are compiled out.
903 */
904 if (bad_area_access_from_pkeys(error_code, vma)) {
905 /*
906 * A protection key fault means that the PKRU value did not allow
907 * access to some PTE. Userspace can figure out what PKRU was
908 * from the XSAVE state. This function captures the pkey from
909 * the vma and passes it to userspace so userspace can discover
910 * which protection key was set on the PTE.
911 *
912 * If we get here, we know that the hardware signaled a X86_PF_PK
913 * fault and that there was a VMA once we got in the fault
914 * handler. It does *not* guarantee that the VMA we find here
915 * was the one that we faulted on.
916 *
917 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
918 * 2. T1 : set PKRU to deny access to pkey=4, touches page
919 * 3. T1 : faults...
920 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
921 * 5. T1 : enters fault handler, takes mmap_lock, etc...
922 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
923 * faulted on a pte with its pkey=4.
924 */
925 u32 pkey = vma_pkey(vma);
926
927 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
928 } else {
929 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
930 }
931}
932
933static void
934do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
935 vm_fault_t fault)
936{
937 /* Kernel mode? Handle exceptions or die: */
938 if (!user_mode(regs)) {
939 kernelmode_fixup_or_oops(regs, error_code, address,
940 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
941 return;
942 }
943
944 /* User-space => ok to do another page fault: */
945 if (is_prefetch(regs, error_code, address))
946 return;
947
948 sanitize_error_code(address, &error_code);
949
950 if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
951 return;
952
953 set_signal_archinfo(address, error_code);
954
955#ifdef CONFIG_MEMORY_FAILURE
956 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
957 struct task_struct *tsk = current;
958 unsigned lsb = 0;
959
960 pr_err(
961 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
962 tsk->comm, tsk->pid, address);
963 if (fault & VM_FAULT_HWPOISON_LARGE)
964 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
965 if (fault & VM_FAULT_HWPOISON)
966 lsb = PAGE_SHIFT;
967 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
968 return;
969 }
970#endif
971 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
972}
973
974static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
975{
976 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
977 return 0;
978
979 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
980 return 0;
981
982 return 1;
983}
984
985/*
986 * Handle a spurious fault caused by a stale TLB entry.
987 *
988 * This allows us to lazily refresh the TLB when increasing the
989 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
990 * eagerly is very expensive since that implies doing a full
991 * cross-processor TLB flush, even if no stale TLB entries exist
992 * on other processors.
993 *
994 * Spurious faults may only occur if the TLB contains an entry with
995 * fewer permission than the page table entry. Non-present (P = 0)
996 * and reserved bit (R = 1) faults are never spurious.
997 *
998 * There are no security implications to leaving a stale TLB when
999 * increasing the permissions on a page.
1000 *
1001 * Returns non-zero if a spurious fault was handled, zero otherwise.
1002 *
1003 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1004 * (Optional Invalidation).
1005 */
1006static noinline int
1007spurious_kernel_fault(unsigned long error_code, unsigned long address)
1008{
1009 pgd_t *pgd;
1010 p4d_t *p4d;
1011 pud_t *pud;
1012 pmd_t *pmd;
1013 pte_t *pte;
1014 int ret;
1015
1016 /*
1017 * Only writes to RO or instruction fetches from NX may cause
1018 * spurious faults.
1019 *
1020 * These could be from user or supervisor accesses but the TLB
1021 * is only lazily flushed after a kernel mapping protection
1022 * change, so user accesses are not expected to cause spurious
1023 * faults.
1024 */
1025 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1026 error_code != (X86_PF_INSTR | X86_PF_PROT))
1027 return 0;
1028
1029 pgd = init_mm.pgd + pgd_index(address);
1030 if (!pgd_present(*pgd))
1031 return 0;
1032
1033 p4d = p4d_offset(pgd, address);
1034 if (!p4d_present(*p4d))
1035 return 0;
1036
1037 if (p4d_large(*p4d))
1038 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1039
1040 pud = pud_offset(p4d, address);
1041 if (!pud_present(*pud))
1042 return 0;
1043
1044 if (pud_large(*pud))
1045 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1046
1047 pmd = pmd_offset(pud, address);
1048 if (!pmd_present(*pmd))
1049 return 0;
1050
1051 if (pmd_large(*pmd))
1052 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1053
1054 pte = pte_offset_kernel(pmd, address);
1055 if (!pte_present(*pte))
1056 return 0;
1057
1058 ret = spurious_kernel_fault_check(error_code, pte);
1059 if (!ret)
1060 return 0;
1061
1062 /*
1063 * Make sure we have permissions in PMD.
1064 * If not, then there's a bug in the page tables:
1065 */
1066 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1067 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1068
1069 return ret;
1070}
1071NOKPROBE_SYMBOL(spurious_kernel_fault);
1072
1073int show_unhandled_signals = 1;
1074
1075static inline int
1076access_error(unsigned long error_code, struct vm_area_struct *vma)
1077{
1078 /* This is only called for the current mm, so: */
1079 bool foreign = false;
1080
1081 /*
1082 * Read or write was blocked by protection keys. This is
1083 * always an unconditional error and can never result in
1084 * a follow-up action to resolve the fault, like a COW.
1085 */
1086 if (error_code & X86_PF_PK)
1087 return 1;
1088
1089 /*
1090 * SGX hardware blocked the access. This usually happens
1091 * when the enclave memory contents have been destroyed, like
1092 * after a suspend/resume cycle. In any case, the kernel can't
1093 * fix the cause of the fault. Handle the fault as an access
1094 * error even in cases where no actual access violation
1095 * occurred. This allows userspace to rebuild the enclave in
1096 * response to the signal.
1097 */
1098 if (unlikely(error_code & X86_PF_SGX))
1099 return 1;
1100
1101 /*
1102 * Make sure to check the VMA so that we do not perform
1103 * faults just to hit a X86_PF_PK as soon as we fill in a
1104 * page.
1105 */
1106 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1107 (error_code & X86_PF_INSTR), foreign))
1108 return 1;
1109
1110 if (error_code & X86_PF_WRITE) {
1111 /* write, present and write, not present: */
1112 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1113 return 1;
1114 return 0;
1115 }
1116
1117 /* read, present: */
1118 if (unlikely(error_code & X86_PF_PROT))
1119 return 1;
1120
1121 /* read, not present: */
1122 if (unlikely(!vma_is_accessible(vma)))
1123 return 1;
1124
1125 return 0;
1126}
1127
1128bool fault_in_kernel_space(unsigned long address)
1129{
1130 /*
1131 * On 64-bit systems, the vsyscall page is at an address above
1132 * TASK_SIZE_MAX, but is not considered part of the kernel
1133 * address space.
1134 */
1135 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1136 return false;
1137
1138 return address >= TASK_SIZE_MAX;
1139}
1140
1141/*
1142 * Called for all faults where 'address' is part of the kernel address
1143 * space. Might get called for faults that originate from *code* that
1144 * ran in userspace or the kernel.
1145 */
1146static void
1147do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1148 unsigned long address)
1149{
1150 /*
1151 * Protection keys exceptions only happen on user pages. We
1152 * have no user pages in the kernel portion of the address
1153 * space, so do not expect them here.
1154 */
1155 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1156
1157#ifdef CONFIG_X86_32
1158 /*
1159 * We can fault-in kernel-space virtual memory on-demand. The
1160 * 'reference' page table is init_mm.pgd.
1161 *
1162 * NOTE! We MUST NOT take any locks for this case. We may
1163 * be in an interrupt or a critical region, and should
1164 * only copy the information from the master page table,
1165 * nothing more.
1166 *
1167 * Before doing this on-demand faulting, ensure that the
1168 * fault is not any of the following:
1169 * 1. A fault on a PTE with a reserved bit set.
1170 * 2. A fault caused by a user-mode access. (Do not demand-
1171 * fault kernel memory due to user-mode accesses).
1172 * 3. A fault caused by a page-level protection violation.
1173 * (A demand fault would be on a non-present page which
1174 * would have X86_PF_PROT==0).
1175 *
1176 * This is only needed to close a race condition on x86-32 in
1177 * the vmalloc mapping/unmapping code. See the comment above
1178 * vmalloc_fault() for details. On x86-64 the race does not
1179 * exist as the vmalloc mappings don't need to be synchronized
1180 * there.
1181 */
1182 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1183 if (vmalloc_fault(address) >= 0)
1184 return;
1185 }
1186#endif
1187
1188 if (is_f00f_bug(regs, hw_error_code, address))
1189 return;
1190
1191 /* Was the fault spurious, caused by lazy TLB invalidation? */
1192 if (spurious_kernel_fault(hw_error_code, address))
1193 return;
1194
1195 /* kprobes don't want to hook the spurious faults: */
1196 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1197 return;
1198
1199 /*
1200 * Note, despite being a "bad area", there are quite a few
1201 * acceptable reasons to get here, such as erratum fixups
1202 * and handling kernel code that can fault, like get_user().
1203 *
1204 * Don't take the mm semaphore here. If we fixup a prefetch
1205 * fault we could otherwise deadlock:
1206 */
1207 bad_area_nosemaphore(regs, hw_error_code, address);
1208}
1209NOKPROBE_SYMBOL(do_kern_addr_fault);
1210
1211/*
1212 * Handle faults in the user portion of the address space. Nothing in here
1213 * should check X86_PF_USER without a specific justification: for almost
1214 * all purposes, we should treat a normal kernel access to user memory
1215 * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1216 * The one exception is AC flag handling, which is, per the x86
1217 * architecture, special for WRUSS.
1218 */
1219static inline
1220void do_user_addr_fault(struct pt_regs *regs,
1221 unsigned long error_code,
1222 unsigned long address)
1223{
1224 struct vm_area_struct *vma;
1225 struct task_struct *tsk;
1226 struct mm_struct *mm;
1227 vm_fault_t fault;
1228 unsigned int flags = FAULT_FLAG_DEFAULT;
1229
1230 tsk = current;
1231 mm = tsk->mm;
1232
1233 if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1234 /*
1235 * Whoops, this is kernel mode code trying to execute from
1236 * user memory. Unless this is AMD erratum #93, which
1237 * corrupts RIP such that it looks like a user address,
1238 * this is unrecoverable. Don't even try to look up the
1239 * VMA or look for extable entries.
1240 */
1241 if (is_errata93(regs, address))
1242 return;
1243
1244 page_fault_oops(regs, error_code, address);
1245 return;
1246 }
1247
1248 /* kprobes don't want to hook the spurious faults: */
1249 if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1250 return;
1251
1252 /*
1253 * Reserved bits are never expected to be set on
1254 * entries in the user portion of the page tables.
1255 */
1256 if (unlikely(error_code & X86_PF_RSVD))
1257 pgtable_bad(regs, error_code, address);
1258
1259 /*
1260 * If SMAP is on, check for invalid kernel (supervisor) access to user
1261 * pages in the user address space. The odd case here is WRUSS,
1262 * which, according to the preliminary documentation, does not respect
1263 * SMAP and will have the USER bit set so, in all cases, SMAP
1264 * enforcement appears to be consistent with the USER bit.
1265 */
1266 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1267 !(error_code & X86_PF_USER) &&
1268 !(regs->flags & X86_EFLAGS_AC))) {
1269 /*
1270 * No extable entry here. This was a kernel access to an
1271 * invalid pointer. get_kernel_nofault() will not get here.
1272 */
1273 page_fault_oops(regs, error_code, address);
1274 return;
1275 }
1276
1277 /*
1278 * If we're in an interrupt, have no user context or are running
1279 * in a region with pagefaults disabled then we must not take the fault
1280 */
1281 if (unlikely(faulthandler_disabled() || !mm)) {
1282 bad_area_nosemaphore(regs, error_code, address);
1283 return;
1284 }
1285
1286 /*
1287 * It's safe to allow irq's after cr2 has been saved and the
1288 * vmalloc fault has been handled.
1289 *
1290 * User-mode registers count as a user access even for any
1291 * potential system fault or CPU buglet:
1292 */
1293 if (user_mode(regs)) {
1294 local_irq_enable();
1295 flags |= FAULT_FLAG_USER;
1296 } else {
1297 if (regs->flags & X86_EFLAGS_IF)
1298 local_irq_enable();
1299 }
1300
1301 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1302
1303 if (error_code & X86_PF_WRITE)
1304 flags |= FAULT_FLAG_WRITE;
1305 if (error_code & X86_PF_INSTR)
1306 flags |= FAULT_FLAG_INSTRUCTION;
1307
1308#ifdef CONFIG_X86_64
1309 /*
1310 * Faults in the vsyscall page might need emulation. The
1311 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1312 * considered to be part of the user address space.
1313 *
1314 * The vsyscall page does not have a "real" VMA, so do this
1315 * emulation before we go searching for VMAs.
1316 *
1317 * PKRU never rejects instruction fetches, so we don't need
1318 * to consider the PF_PK bit.
1319 */
1320 if (is_vsyscall_vaddr(address)) {
1321 if (emulate_vsyscall(error_code, regs, address))
1322 return;
1323 }
1324#endif
1325
1326 /*
1327 * Kernel-mode access to the user address space should only occur
1328 * on well-defined single instructions listed in the exception
1329 * tables. But, an erroneous kernel fault occurring outside one of
1330 * those areas which also holds mmap_lock might deadlock attempting
1331 * to validate the fault against the address space.
1332 *
1333 * Only do the expensive exception table search when we might be at
1334 * risk of a deadlock. This happens if we
1335 * 1. Failed to acquire mmap_lock, and
1336 * 2. The access did not originate in userspace.
1337 */
1338 if (unlikely(!mmap_read_trylock(mm))) {
1339 if (!user_mode(regs) && !search_exception_tables(regs->ip)) {
1340 /*
1341 * Fault from code in kernel from
1342 * which we do not expect faults.
1343 */
1344 bad_area_nosemaphore(regs, error_code, address);
1345 return;
1346 }
1347retry:
1348 mmap_read_lock(mm);
1349 } else {
1350 /*
1351 * The above down_read_trylock() might have succeeded in
1352 * which case we'll have missed the might_sleep() from
1353 * down_read():
1354 */
1355 might_sleep();
1356 }
1357
1358 vma = find_vma(mm, address);
1359 if (unlikely(!vma)) {
1360 bad_area(regs, error_code, address);
1361 return;
1362 }
1363 if (likely(vma->vm_start <= address))
1364 goto good_area;
1365 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1366 bad_area(regs, error_code, address);
1367 return;
1368 }
1369 if (unlikely(expand_stack(vma, address))) {
1370 bad_area(regs, error_code, address);
1371 return;
1372 }
1373
1374 /*
1375 * Ok, we have a good vm_area for this memory access, so
1376 * we can handle it..
1377 */
1378good_area:
1379 if (unlikely(access_error(error_code, vma))) {
1380 bad_area_access_error(regs, error_code, address, vma);
1381 return;
1382 }
1383
1384 /*
1385 * If for any reason at all we couldn't handle the fault,
1386 * make sure we exit gracefully rather than endlessly redo
1387 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1388 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1389 *
1390 * Note that handle_userfault() may also release and reacquire mmap_lock
1391 * (and not return with VM_FAULT_RETRY), when returning to userland to
1392 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1393 * (potentially after handling any pending signal during the return to
1394 * userland). The return to userland is identified whenever
1395 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1396 */
1397 fault = handle_mm_fault(vma, address, flags, regs);
1398
1399 if (fault_signal_pending(fault, regs)) {
1400 /*
1401 * Quick path to respond to signals. The core mm code
1402 * has unlocked the mm for us if we get here.
1403 */
1404 if (!user_mode(regs))
1405 kernelmode_fixup_or_oops(regs, error_code, address,
1406 SIGBUS, BUS_ADRERR,
1407 ARCH_DEFAULT_PKEY);
1408 return;
1409 }
1410
1411 /*
1412 * If we need to retry the mmap_lock has already been released,
1413 * and if there is a fatal signal pending there is no guarantee
1414 * that we made any progress. Handle this case first.
1415 */
1416 if (unlikely((fault & VM_FAULT_RETRY) &&
1417 (flags & FAULT_FLAG_ALLOW_RETRY))) {
1418 flags |= FAULT_FLAG_TRIED;
1419 goto retry;
1420 }
1421
1422 mmap_read_unlock(mm);
1423 if (likely(!(fault & VM_FAULT_ERROR)))
1424 return;
1425
1426 if (fatal_signal_pending(current) && !user_mode(regs)) {
1427 kernelmode_fixup_or_oops(regs, error_code, address,
1428 0, 0, ARCH_DEFAULT_PKEY);
1429 return;
1430 }
1431
1432 if (fault & VM_FAULT_OOM) {
1433 /* Kernel mode? Handle exceptions or die: */
1434 if (!user_mode(regs)) {
1435 kernelmode_fixup_or_oops(regs, error_code, address,
1436 SIGSEGV, SEGV_MAPERR,
1437 ARCH_DEFAULT_PKEY);
1438 return;
1439 }
1440
1441 /*
1442 * We ran out of memory, call the OOM killer, and return the
1443 * userspace (which will retry the fault, or kill us if we got
1444 * oom-killed):
1445 */
1446 pagefault_out_of_memory();
1447 } else {
1448 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1449 VM_FAULT_HWPOISON_LARGE))
1450 do_sigbus(regs, error_code, address, fault);
1451 else if (fault & VM_FAULT_SIGSEGV)
1452 bad_area_nosemaphore(regs, error_code, address);
1453 else
1454 BUG();
1455 }
1456}
1457NOKPROBE_SYMBOL(do_user_addr_fault);
1458
1459static __always_inline void
1460trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1461 unsigned long address)
1462{
1463 if (!trace_pagefault_enabled())
1464 return;
1465
1466 if (user_mode(regs))
1467 trace_page_fault_user(address, regs, error_code);
1468 else
1469 trace_page_fault_kernel(address, regs, error_code);
1470}
1471
1472static __always_inline void
1473handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1474 unsigned long address)
1475{
1476 trace_page_fault_entries(regs, error_code, address);
1477
1478 if (unlikely(kmmio_fault(regs, address)))
1479 return;
1480
1481 /* Was the fault on kernel-controlled part of the address space? */
1482 if (unlikely(fault_in_kernel_space(address))) {
1483 do_kern_addr_fault(regs, error_code, address);
1484 } else {
1485 do_user_addr_fault(regs, error_code, address);
1486 /*
1487 * User address page fault handling might have reenabled
1488 * interrupts. Fixing up all potential exit points of
1489 * do_user_addr_fault() and its leaf functions is just not
1490 * doable w/o creating an unholy mess or turning the code
1491 * upside down.
1492 */
1493 local_irq_disable();
1494 }
1495}
1496
1497DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1498{
1499 unsigned long address = read_cr2();
1500 irqentry_state_t state;
1501
1502 prefetchw(¤t->mm->mmap_lock);
1503
1504 /*
1505 * KVM uses #PF vector to deliver 'page not present' events to guests
1506 * (asynchronous page fault mechanism). The event happens when a
1507 * userspace task is trying to access some valid (from guest's point of
1508 * view) memory which is not currently mapped by the host (e.g. the
1509 * memory is swapped out). Note, the corresponding "page ready" event
1510 * which is injected when the memory becomes available, is delivered via
1511 * an interrupt mechanism and not a #PF exception
1512 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1513 *
1514 * We are relying on the interrupted context being sane (valid RSP,
1515 * relevant locks not held, etc.), which is fine as long as the
1516 * interrupted context had IF=1. We are also relying on the KVM
1517 * async pf type field and CR2 being read consistently instead of
1518 * getting values from real and async page faults mixed up.
1519 *
1520 * Fingers crossed.
1521 *
1522 * The async #PF handling code takes care of idtentry handling
1523 * itself.
1524 */
1525 if (kvm_handle_async_pf(regs, (u32)address))
1526 return;
1527
1528 /*
1529 * Entry handling for valid #PF from kernel mode is slightly
1530 * different: RCU is already watching and rcu_irq_enter() must not
1531 * be invoked because a kernel fault on a user space address might
1532 * sleep.
1533 *
1534 * In case the fault hit a RCU idle region the conditional entry
1535 * code reenabled RCU to avoid subsequent wreckage which helps
1536 * debuggability.
1537 */
1538 state = irqentry_enter(regs);
1539
1540 instrumentation_begin();
1541 handle_page_fault(regs, error_code, address);
1542 instrumentation_end();
1543
1544 irqentry_exit(regs, state);
1545}