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#include <linux/context_tracking.h> /* exception_enter(), ... */
17
18#include <asm/traps.h> /* dotraplinkage, ... */
19#include <asm/pgalloc.h> /* pgd_*(), ... */
20#include <asm/kmemcheck.h> /* kmemcheck_*(), ... */
21#include <asm/fixmap.h> /* VSYSCALL_START */
22
23#define CREATE_TRACE_POINTS
24#include <asm/trace/exceptions.h>
25
26/*
27 * Page fault error code bits:
28 *
29 * bit 0 == 0: no page found 1: protection fault
30 * bit 1 == 0: read access 1: write access
31 * bit 2 == 0: kernel-mode access 1: user-mode access
32 * bit 3 == 1: use of reserved bit detected
33 * bit 4 == 1: fault was an instruction fetch
34 */
35enum x86_pf_error_code {
36
37 PF_PROT = 1 << 0,
38 PF_WRITE = 1 << 1,
39 PF_USER = 1 << 2,
40 PF_RSVD = 1 << 3,
41 PF_INSTR = 1 << 4,
42};
43
44/*
45 * Returns 0 if mmiotrace is disabled, or if the fault is not
46 * handled by mmiotrace:
47 */
48static inline int __kprobes
49kmmio_fault(struct pt_regs *regs, unsigned long addr)
50{
51 if (unlikely(is_kmmio_active()))
52 if (kmmio_handler(regs, addr) == 1)
53 return -1;
54 return 0;
55}
56
57static inline int __kprobes kprobes_fault(struct pt_regs *regs)
58{
59 int ret = 0;
60
61 /* kprobe_running() needs smp_processor_id() */
62 if (kprobes_built_in() && !user_mode_vm(regs)) {
63 preempt_disable();
64 if (kprobe_running() && kprobe_fault_handler(regs, 14))
65 ret = 1;
66 preempt_enable();
67 }
68
69 return ret;
70}
71
72/*
73 * Prefetch quirks:
74 *
75 * 32-bit mode:
76 *
77 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
78 * Check that here and ignore it.
79 *
80 * 64-bit mode:
81 *
82 * Sometimes the CPU reports invalid exceptions on prefetch.
83 * Check that here and ignore it.
84 *
85 * Opcode checker based on code by Richard Brunner.
86 */
87static inline int
88check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
89 unsigned char opcode, int *prefetch)
90{
91 unsigned char instr_hi = opcode & 0xf0;
92 unsigned char instr_lo = opcode & 0x0f;
93
94 switch (instr_hi) {
95 case 0x20:
96 case 0x30:
97 /*
98 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
99 * In X86_64 long mode, the CPU will signal invalid
100 * opcode if some of these prefixes are present so
101 * X86_64 will never get here anyway
102 */
103 return ((instr_lo & 7) == 0x6);
104#ifdef CONFIG_X86_64
105 case 0x40:
106 /*
107 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
108 * Need to figure out under what instruction mode the
109 * instruction was issued. Could check the LDT for lm,
110 * but for now it's good enough to assume that long
111 * mode only uses well known segments or kernel.
112 */
113 return (!user_mode(regs) || user_64bit_mode(regs));
114#endif
115 case 0x60:
116 /* 0x64 thru 0x67 are valid prefixes in all modes. */
117 return (instr_lo & 0xC) == 0x4;
118 case 0xF0:
119 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
120 return !instr_lo || (instr_lo>>1) == 1;
121 case 0x00:
122 /* Prefetch instruction is 0x0F0D or 0x0F18 */
123 if (probe_kernel_address(instr, opcode))
124 return 0;
125
126 *prefetch = (instr_lo == 0xF) &&
127 (opcode == 0x0D || opcode == 0x18);
128 return 0;
129 default:
130 return 0;
131 }
132}
133
134static int
135is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
136{
137 unsigned char *max_instr;
138 unsigned char *instr;
139 int prefetch = 0;
140
141 /*
142 * If it was a exec (instruction fetch) fault on NX page, then
143 * do not ignore the fault:
144 */
145 if (error_code & PF_INSTR)
146 return 0;
147
148 instr = (void *)convert_ip_to_linear(current, regs);
149 max_instr = instr + 15;
150
151 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
152 return 0;
153
154 while (instr < max_instr) {
155 unsigned char opcode;
156
157 if (probe_kernel_address(instr, opcode))
158 break;
159
160 instr++;
161
162 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
163 break;
164 }
165 return prefetch;
166}
167
168static void
169force_sig_info_fault(int si_signo, int si_code, unsigned long address,
170 struct task_struct *tsk, int fault)
171{
172 unsigned lsb = 0;
173 siginfo_t info;
174
175 info.si_signo = si_signo;
176 info.si_errno = 0;
177 info.si_code = si_code;
178 info.si_addr = (void __user *)address;
179 if (fault & VM_FAULT_HWPOISON_LARGE)
180 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
181 if (fault & VM_FAULT_HWPOISON)
182 lsb = PAGE_SHIFT;
183 info.si_addr_lsb = lsb;
184
185 force_sig_info(si_signo, &info, tsk);
186}
187
188DEFINE_SPINLOCK(pgd_lock);
189LIST_HEAD(pgd_list);
190
191#ifdef CONFIG_X86_32
192static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
193{
194 unsigned index = pgd_index(address);
195 pgd_t *pgd_k;
196 pud_t *pud, *pud_k;
197 pmd_t *pmd, *pmd_k;
198
199 pgd += index;
200 pgd_k = init_mm.pgd + index;
201
202 if (!pgd_present(*pgd_k))
203 return NULL;
204
205 /*
206 * set_pgd(pgd, *pgd_k); here would be useless on PAE
207 * and redundant with the set_pmd() on non-PAE. As would
208 * set_pud.
209 */
210 pud = pud_offset(pgd, address);
211 pud_k = pud_offset(pgd_k, address);
212 if (!pud_present(*pud_k))
213 return NULL;
214
215 pmd = pmd_offset(pud, address);
216 pmd_k = pmd_offset(pud_k, address);
217 if (!pmd_present(*pmd_k))
218 return NULL;
219
220 if (!pmd_present(*pmd))
221 set_pmd(pmd, *pmd_k);
222 else
223 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
224
225 return pmd_k;
226}
227
228void vmalloc_sync_all(void)
229{
230 unsigned long address;
231
232 if (SHARED_KERNEL_PMD)
233 return;
234
235 for (address = VMALLOC_START & PMD_MASK;
236 address >= TASK_SIZE && address < FIXADDR_TOP;
237 address += PMD_SIZE) {
238 struct page *page;
239
240 spin_lock(&pgd_lock);
241 list_for_each_entry(page, &pgd_list, lru) {
242 spinlock_t *pgt_lock;
243 pmd_t *ret;
244
245 /* the pgt_lock only for Xen */
246 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
247
248 spin_lock(pgt_lock);
249 ret = vmalloc_sync_one(page_address(page), address);
250 spin_unlock(pgt_lock);
251
252 if (!ret)
253 break;
254 }
255 spin_unlock(&pgd_lock);
256 }
257}
258
259/*
260 * 32-bit:
261 *
262 * Handle a fault on the vmalloc or module mapping area
263 */
264static noinline __kprobes int vmalloc_fault(unsigned long address)
265{
266 unsigned long pgd_paddr;
267 pmd_t *pmd_k;
268 pte_t *pte_k;
269
270 /* Make sure we are in vmalloc area: */
271 if (!(address >= VMALLOC_START && address < VMALLOC_END))
272 return -1;
273
274 WARN_ON_ONCE(in_nmi());
275
276 /*
277 * Synchronize this task's top level page-table
278 * with the 'reference' page table.
279 *
280 * Do _not_ use "current" here. We might be inside
281 * an interrupt in the middle of a task switch..
282 */
283 pgd_paddr = read_cr3();
284 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
285 if (!pmd_k)
286 return -1;
287
288 pte_k = pte_offset_kernel(pmd_k, address);
289 if (!pte_present(*pte_k))
290 return -1;
291
292 return 0;
293}
294
295/*
296 * Did it hit the DOS screen memory VA from vm86 mode?
297 */
298static inline void
299check_v8086_mode(struct pt_regs *regs, unsigned long address,
300 struct task_struct *tsk)
301{
302 unsigned long bit;
303
304 if (!v8086_mode(regs))
305 return;
306
307 bit = (address - 0xA0000) >> PAGE_SHIFT;
308 if (bit < 32)
309 tsk->thread.screen_bitmap |= 1 << bit;
310}
311
312static bool low_pfn(unsigned long pfn)
313{
314 return pfn < max_low_pfn;
315}
316
317static void dump_pagetable(unsigned long address)
318{
319 pgd_t *base = __va(read_cr3());
320 pgd_t *pgd = &base[pgd_index(address)];
321 pmd_t *pmd;
322 pte_t *pte;
323
324#ifdef CONFIG_X86_PAE
325 printk("*pdpt = %016Lx ", pgd_val(*pgd));
326 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
327 goto out;
328#endif
329 pmd = pmd_offset(pud_offset(pgd, address), address);
330 printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
331
332 /*
333 * We must not directly access the pte in the highpte
334 * case if the page table is located in highmem.
335 * And let's rather not kmap-atomic the pte, just in case
336 * it's allocated already:
337 */
338 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
339 goto out;
340
341 pte = pte_offset_kernel(pmd, address);
342 printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
343out:
344 printk("\n");
345}
346
347#else /* CONFIG_X86_64: */
348
349void vmalloc_sync_all(void)
350{
351 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
352}
353
354/*
355 * 64-bit:
356 *
357 * Handle a fault on the vmalloc area
358 *
359 * This assumes no large pages in there.
360 */
361static noinline __kprobes int vmalloc_fault(unsigned long address)
362{
363 pgd_t *pgd, *pgd_ref;
364 pud_t *pud, *pud_ref;
365 pmd_t *pmd, *pmd_ref;
366 pte_t *pte, *pte_ref;
367
368 /* Make sure we are in vmalloc area: */
369 if (!(address >= VMALLOC_START && address < VMALLOC_END))
370 return -1;
371
372 WARN_ON_ONCE(in_nmi());
373
374 /*
375 * Copy kernel mappings over when needed. This can also
376 * happen within a race in page table update. In the later
377 * case just flush:
378 */
379 pgd = pgd_offset(current->active_mm, address);
380 pgd_ref = pgd_offset_k(address);
381 if (pgd_none(*pgd_ref))
382 return -1;
383
384 if (pgd_none(*pgd)) {
385 set_pgd(pgd, *pgd_ref);
386 arch_flush_lazy_mmu_mode();
387 } else {
388 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
389 }
390
391 /*
392 * Below here mismatches are bugs because these lower tables
393 * are shared:
394 */
395
396 pud = pud_offset(pgd, address);
397 pud_ref = pud_offset(pgd_ref, address);
398 if (pud_none(*pud_ref))
399 return -1;
400
401 if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
402 BUG();
403
404 pmd = pmd_offset(pud, address);
405 pmd_ref = pmd_offset(pud_ref, address);
406 if (pmd_none(*pmd_ref))
407 return -1;
408
409 if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
410 BUG();
411
412 pte_ref = pte_offset_kernel(pmd_ref, address);
413 if (!pte_present(*pte_ref))
414 return -1;
415
416 pte = pte_offset_kernel(pmd, address);
417
418 /*
419 * Don't use pte_page here, because the mappings can point
420 * outside mem_map, and the NUMA hash lookup cannot handle
421 * that:
422 */
423 if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
424 BUG();
425
426 return 0;
427}
428
429#ifdef CONFIG_CPU_SUP_AMD
430static const char errata93_warning[] =
431KERN_ERR
432"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
433"******* Working around it, but it may cause SEGVs or burn power.\n"
434"******* Please consider a BIOS update.\n"
435"******* Disabling USB legacy in the BIOS may also help.\n";
436#endif
437
438/*
439 * No vm86 mode in 64-bit mode:
440 */
441static inline void
442check_v8086_mode(struct pt_regs *regs, unsigned long address,
443 struct task_struct *tsk)
444{
445}
446
447static int bad_address(void *p)
448{
449 unsigned long dummy;
450
451 return probe_kernel_address((unsigned long *)p, dummy);
452}
453
454static void dump_pagetable(unsigned long address)
455{
456 pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
457 pgd_t *pgd = base + pgd_index(address);
458 pud_t *pud;
459 pmd_t *pmd;
460 pte_t *pte;
461
462 if (bad_address(pgd))
463 goto bad;
464
465 printk("PGD %lx ", pgd_val(*pgd));
466
467 if (!pgd_present(*pgd))
468 goto out;
469
470 pud = pud_offset(pgd, address);
471 if (bad_address(pud))
472 goto bad;
473
474 printk("PUD %lx ", pud_val(*pud));
475 if (!pud_present(*pud) || pud_large(*pud))
476 goto out;
477
478 pmd = pmd_offset(pud, address);
479 if (bad_address(pmd))
480 goto bad;
481
482 printk("PMD %lx ", pmd_val(*pmd));
483 if (!pmd_present(*pmd) || pmd_large(*pmd))
484 goto out;
485
486 pte = pte_offset_kernel(pmd, address);
487 if (bad_address(pte))
488 goto bad;
489
490 printk("PTE %lx", pte_val(*pte));
491out:
492 printk("\n");
493 return;
494bad:
495 printk("BAD\n");
496}
497
498#endif /* CONFIG_X86_64 */
499
500/*
501 * Workaround for K8 erratum #93 & buggy BIOS.
502 *
503 * BIOS SMM functions are required to use a specific workaround
504 * to avoid corruption of the 64bit RIP register on C stepping K8.
505 *
506 * A lot of BIOS that didn't get tested properly miss this.
507 *
508 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
509 * Try to work around it here.
510 *
511 * Note we only handle faults in kernel here.
512 * Does nothing on 32-bit.
513 */
514static int is_errata93(struct pt_regs *regs, unsigned long address)
515{
516#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
517 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
518 || boot_cpu_data.x86 != 0xf)
519 return 0;
520
521 if (address != regs->ip)
522 return 0;
523
524 if ((address >> 32) != 0)
525 return 0;
526
527 address |= 0xffffffffUL << 32;
528 if ((address >= (u64)_stext && address <= (u64)_etext) ||
529 (address >= MODULES_VADDR && address <= MODULES_END)) {
530 printk_once(errata93_warning);
531 regs->ip = address;
532 return 1;
533 }
534#endif
535 return 0;
536}
537
538/*
539 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
540 * to illegal addresses >4GB.
541 *
542 * We catch this in the page fault handler because these addresses
543 * are not reachable. Just detect this case and return. Any code
544 * segment in LDT is compatibility mode.
545 */
546static int is_errata100(struct pt_regs *regs, unsigned long address)
547{
548#ifdef CONFIG_X86_64
549 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
550 return 1;
551#endif
552 return 0;
553}
554
555static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
556{
557#ifdef CONFIG_X86_F00F_BUG
558 unsigned long nr;
559
560 /*
561 * Pentium F0 0F C7 C8 bug workaround:
562 */
563 if (boot_cpu_has_bug(X86_BUG_F00F)) {
564 nr = (address - idt_descr.address) >> 3;
565
566 if (nr == 6) {
567 do_invalid_op(regs, 0);
568 return 1;
569 }
570 }
571#endif
572 return 0;
573}
574
575static const char nx_warning[] = KERN_CRIT
576"kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
577
578static void
579show_fault_oops(struct pt_regs *regs, unsigned long error_code,
580 unsigned long address)
581{
582 if (!oops_may_print())
583 return;
584
585 if (error_code & PF_INSTR) {
586 unsigned int level;
587 pgd_t *pgd;
588 pte_t *pte;
589
590 pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
591 pgd += pgd_index(address);
592
593 pte = lookup_address_in_pgd(pgd, address, &level);
594
595 if (pte && pte_present(*pte) && !pte_exec(*pte))
596 printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
597 }
598
599 printk(KERN_ALERT "BUG: unable to handle kernel ");
600 if (address < PAGE_SIZE)
601 printk(KERN_CONT "NULL pointer dereference");
602 else
603 printk(KERN_CONT "paging request");
604
605 printk(KERN_CONT " at %p\n", (void *) address);
606 printk(KERN_ALERT "IP:");
607 printk_address(regs->ip);
608
609 dump_pagetable(address);
610}
611
612static noinline void
613pgtable_bad(struct pt_regs *regs, unsigned long error_code,
614 unsigned long address)
615{
616 struct task_struct *tsk;
617 unsigned long flags;
618 int sig;
619
620 flags = oops_begin();
621 tsk = current;
622 sig = SIGKILL;
623
624 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
625 tsk->comm, address);
626 dump_pagetable(address);
627
628 tsk->thread.cr2 = address;
629 tsk->thread.trap_nr = X86_TRAP_PF;
630 tsk->thread.error_code = error_code;
631
632 if (__die("Bad pagetable", regs, error_code))
633 sig = 0;
634
635 oops_end(flags, regs, sig);
636}
637
638static noinline void
639no_context(struct pt_regs *regs, unsigned long error_code,
640 unsigned long address, int signal, int si_code)
641{
642 struct task_struct *tsk = current;
643 unsigned long *stackend;
644 unsigned long flags;
645 int sig;
646
647 /* Are we prepared to handle this kernel fault? */
648 if (fixup_exception(regs)) {
649 /*
650 * Any interrupt that takes a fault gets the fixup. This makes
651 * the below recursive fault logic only apply to a faults from
652 * task context.
653 */
654 if (in_interrupt())
655 return;
656
657 /*
658 * Per the above we're !in_interrupt(), aka. task context.
659 *
660 * In this case we need to make sure we're not recursively
661 * faulting through the emulate_vsyscall() logic.
662 */
663 if (current_thread_info()->sig_on_uaccess_error && signal) {
664 tsk->thread.trap_nr = X86_TRAP_PF;
665 tsk->thread.error_code = error_code | PF_USER;
666 tsk->thread.cr2 = address;
667
668 /* XXX: hwpoison faults will set the wrong code. */
669 force_sig_info_fault(signal, si_code, address, tsk, 0);
670 }
671
672 /*
673 * Barring that, we can do the fixup and be happy.
674 */
675 return;
676 }
677
678 /*
679 * 32-bit:
680 *
681 * Valid to do another page fault here, because if this fault
682 * had been triggered by is_prefetch fixup_exception would have
683 * handled it.
684 *
685 * 64-bit:
686 *
687 * Hall of shame of CPU/BIOS bugs.
688 */
689 if (is_prefetch(regs, error_code, address))
690 return;
691
692 if (is_errata93(regs, address))
693 return;
694
695 /*
696 * Oops. The kernel tried to access some bad page. We'll have to
697 * terminate things with extreme prejudice:
698 */
699 flags = oops_begin();
700
701 show_fault_oops(regs, error_code, address);
702
703 stackend = end_of_stack(tsk);
704 if (tsk != &init_task && *stackend != STACK_END_MAGIC)
705 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
706
707 tsk->thread.cr2 = address;
708 tsk->thread.trap_nr = X86_TRAP_PF;
709 tsk->thread.error_code = error_code;
710
711 sig = SIGKILL;
712 if (__die("Oops", regs, error_code))
713 sig = 0;
714
715 /* Executive summary in case the body of the oops scrolled away */
716 printk(KERN_DEFAULT "CR2: %016lx\n", address);
717
718 oops_end(flags, regs, sig);
719}
720
721/*
722 * Print out info about fatal segfaults, if the show_unhandled_signals
723 * sysctl is set:
724 */
725static inline void
726show_signal_msg(struct pt_regs *regs, unsigned long error_code,
727 unsigned long address, struct task_struct *tsk)
728{
729 if (!unhandled_signal(tsk, SIGSEGV))
730 return;
731
732 if (!printk_ratelimit())
733 return;
734
735 printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
736 task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
737 tsk->comm, task_pid_nr(tsk), address,
738 (void *)regs->ip, (void *)regs->sp, error_code);
739
740 print_vma_addr(KERN_CONT " in ", regs->ip);
741
742 printk(KERN_CONT "\n");
743}
744
745static void
746__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
747 unsigned long address, int si_code)
748{
749 struct task_struct *tsk = current;
750
751 /* User mode accesses just cause a SIGSEGV */
752 if (error_code & PF_USER) {
753 /*
754 * It's possible to have interrupts off here:
755 */
756 local_irq_enable();
757
758 /*
759 * Valid to do another page fault here because this one came
760 * from user space:
761 */
762 if (is_prefetch(regs, error_code, address))
763 return;
764
765 if (is_errata100(regs, address))
766 return;
767
768#ifdef CONFIG_X86_64
769 /*
770 * Instruction fetch faults in the vsyscall page might need
771 * emulation.
772 */
773 if (unlikely((error_code & PF_INSTR) &&
774 ((address & ~0xfff) == VSYSCALL_START))) {
775 if (emulate_vsyscall(regs, address))
776 return;
777 }
778#endif
779 /* Kernel addresses are always protection faults: */
780 if (address >= TASK_SIZE)
781 error_code |= PF_PROT;
782
783 if (likely(show_unhandled_signals))
784 show_signal_msg(regs, error_code, address, tsk);
785
786 tsk->thread.cr2 = address;
787 tsk->thread.error_code = error_code;
788 tsk->thread.trap_nr = X86_TRAP_PF;
789
790 force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
791
792 return;
793 }
794
795 if (is_f00f_bug(regs, address))
796 return;
797
798 no_context(regs, error_code, address, SIGSEGV, si_code);
799}
800
801static noinline void
802bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
803 unsigned long address)
804{
805 __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
806}
807
808static void
809__bad_area(struct pt_regs *regs, unsigned long error_code,
810 unsigned long address, int si_code)
811{
812 struct mm_struct *mm = current->mm;
813
814 /*
815 * Something tried to access memory that isn't in our memory map..
816 * Fix it, but check if it's kernel or user first..
817 */
818 up_read(&mm->mmap_sem);
819
820 __bad_area_nosemaphore(regs, error_code, address, si_code);
821}
822
823static noinline void
824bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
825{
826 __bad_area(regs, error_code, address, SEGV_MAPERR);
827}
828
829static noinline void
830bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
831 unsigned long address)
832{
833 __bad_area(regs, error_code, address, SEGV_ACCERR);
834}
835
836static void
837do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
838 unsigned int fault)
839{
840 struct task_struct *tsk = current;
841 struct mm_struct *mm = tsk->mm;
842 int code = BUS_ADRERR;
843
844 up_read(&mm->mmap_sem);
845
846 /* Kernel mode? Handle exceptions or die: */
847 if (!(error_code & PF_USER)) {
848 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
849 return;
850 }
851
852 /* User-space => ok to do another page fault: */
853 if (is_prefetch(regs, error_code, address))
854 return;
855
856 tsk->thread.cr2 = address;
857 tsk->thread.error_code = error_code;
858 tsk->thread.trap_nr = X86_TRAP_PF;
859
860#ifdef CONFIG_MEMORY_FAILURE
861 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
862 printk(KERN_ERR
863 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
864 tsk->comm, tsk->pid, address);
865 code = BUS_MCEERR_AR;
866 }
867#endif
868 force_sig_info_fault(SIGBUS, code, address, tsk, fault);
869}
870
871static noinline void
872mm_fault_error(struct pt_regs *regs, unsigned long error_code,
873 unsigned long address, unsigned int fault)
874{
875 if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
876 up_read(¤t->mm->mmap_sem);
877 no_context(regs, error_code, address, 0, 0);
878 return;
879 }
880
881 if (fault & VM_FAULT_OOM) {
882 /* Kernel mode? Handle exceptions or die: */
883 if (!(error_code & PF_USER)) {
884 up_read(¤t->mm->mmap_sem);
885 no_context(regs, error_code, address,
886 SIGSEGV, SEGV_MAPERR);
887 return;
888 }
889
890 up_read(¤t->mm->mmap_sem);
891
892 /*
893 * We ran out of memory, call the OOM killer, and return the
894 * userspace (which will retry the fault, or kill us if we got
895 * oom-killed):
896 */
897 pagefault_out_of_memory();
898 } else {
899 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
900 VM_FAULT_HWPOISON_LARGE))
901 do_sigbus(regs, error_code, address, fault);
902 else
903 BUG();
904 }
905}
906
907static int spurious_fault_check(unsigned long error_code, pte_t *pte)
908{
909 if ((error_code & PF_WRITE) && !pte_write(*pte))
910 return 0;
911
912 if ((error_code & PF_INSTR) && !pte_exec(*pte))
913 return 0;
914
915 return 1;
916}
917
918/*
919 * Handle a spurious fault caused by a stale TLB entry.
920 *
921 * This allows us to lazily refresh the TLB when increasing the
922 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
923 * eagerly is very expensive since that implies doing a full
924 * cross-processor TLB flush, even if no stale TLB entries exist
925 * on other processors.
926 *
927 * There are no security implications to leaving a stale TLB when
928 * increasing the permissions on a page.
929 */
930static noinline __kprobes int
931spurious_fault(unsigned long error_code, unsigned long address)
932{
933 pgd_t *pgd;
934 pud_t *pud;
935 pmd_t *pmd;
936 pte_t *pte;
937 int ret;
938
939 /* Reserved-bit violation or user access to kernel space? */
940 if (error_code & (PF_USER | PF_RSVD))
941 return 0;
942
943 pgd = init_mm.pgd + pgd_index(address);
944 if (!pgd_present(*pgd))
945 return 0;
946
947 pud = pud_offset(pgd, address);
948 if (!pud_present(*pud))
949 return 0;
950
951 if (pud_large(*pud))
952 return spurious_fault_check(error_code, (pte_t *) pud);
953
954 pmd = pmd_offset(pud, address);
955 if (!pmd_present(*pmd))
956 return 0;
957
958 if (pmd_large(*pmd))
959 return spurious_fault_check(error_code, (pte_t *) pmd);
960
961 pte = pte_offset_kernel(pmd, address);
962 if (!pte_present(*pte))
963 return 0;
964
965 ret = spurious_fault_check(error_code, pte);
966 if (!ret)
967 return 0;
968
969 /*
970 * Make sure we have permissions in PMD.
971 * If not, then there's a bug in the page tables:
972 */
973 ret = spurious_fault_check(error_code, (pte_t *) pmd);
974 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
975
976 return ret;
977}
978
979int show_unhandled_signals = 1;
980
981static inline int
982access_error(unsigned long error_code, struct vm_area_struct *vma)
983{
984 if (error_code & PF_WRITE) {
985 /* write, present and write, not present: */
986 if (unlikely(!(vma->vm_flags & VM_WRITE)))
987 return 1;
988 return 0;
989 }
990
991 /* read, present: */
992 if (unlikely(error_code & PF_PROT))
993 return 1;
994
995 /* read, not present: */
996 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
997 return 1;
998
999 return 0;
1000}
1001
1002static int fault_in_kernel_space(unsigned long address)
1003{
1004 return address >= TASK_SIZE_MAX;
1005}
1006
1007static inline bool smap_violation(int error_code, struct pt_regs *regs)
1008{
1009 if (!IS_ENABLED(CONFIG_X86_SMAP))
1010 return false;
1011
1012 if (!static_cpu_has(X86_FEATURE_SMAP))
1013 return false;
1014
1015 if (error_code & PF_USER)
1016 return false;
1017
1018 if (!user_mode_vm(regs) && (regs->flags & X86_EFLAGS_AC))
1019 return false;
1020
1021 return true;
1022}
1023
1024/*
1025 * This routine handles page faults. It determines the address,
1026 * and the problem, and then passes it off to one of the appropriate
1027 * routines.
1028 *
1029 * This function must have noinline because both callers
1030 * {,trace_}do_page_fault() have notrace on. Having this an actual function
1031 * guarantees there's a function trace entry.
1032 */
1033static void __kprobes noinline
1034__do_page_fault(struct pt_regs *regs, unsigned long error_code,
1035 unsigned long address)
1036{
1037 struct vm_area_struct *vma;
1038 struct task_struct *tsk;
1039 struct mm_struct *mm;
1040 int fault;
1041 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1042
1043 tsk = current;
1044 mm = tsk->mm;
1045
1046 /*
1047 * Detect and handle instructions that would cause a page fault for
1048 * both a tracked kernel page and a userspace page.
1049 */
1050 if (kmemcheck_active(regs))
1051 kmemcheck_hide(regs);
1052 prefetchw(&mm->mmap_sem);
1053
1054 if (unlikely(kmmio_fault(regs, address)))
1055 return;
1056
1057 /*
1058 * We fault-in kernel-space virtual memory on-demand. The
1059 * 'reference' page table is init_mm.pgd.
1060 *
1061 * NOTE! We MUST NOT take any locks for this case. We may
1062 * be in an interrupt or a critical region, and should
1063 * only copy the information from the master page table,
1064 * nothing more.
1065 *
1066 * This verifies that the fault happens in kernel space
1067 * (error_code & 4) == 0, and that the fault was not a
1068 * protection error (error_code & 9) == 0.
1069 */
1070 if (unlikely(fault_in_kernel_space(address))) {
1071 if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1072 if (vmalloc_fault(address) >= 0)
1073 return;
1074
1075 if (kmemcheck_fault(regs, address, error_code))
1076 return;
1077 }
1078
1079 /* Can handle a stale RO->RW TLB: */
1080 if (spurious_fault(error_code, address))
1081 return;
1082
1083 /* kprobes don't want to hook the spurious faults: */
1084 if (kprobes_fault(regs))
1085 return;
1086 /*
1087 * Don't take the mm semaphore here. If we fixup a prefetch
1088 * fault we could otherwise deadlock:
1089 */
1090 bad_area_nosemaphore(regs, error_code, address);
1091
1092 return;
1093 }
1094
1095 /* kprobes don't want to hook the spurious faults: */
1096 if (unlikely(kprobes_fault(regs)))
1097 return;
1098
1099 if (unlikely(error_code & PF_RSVD))
1100 pgtable_bad(regs, error_code, address);
1101
1102 if (unlikely(smap_violation(error_code, regs))) {
1103 bad_area_nosemaphore(regs, error_code, address);
1104 return;
1105 }
1106
1107 /*
1108 * If we're in an interrupt, have no user context or are running
1109 * in an atomic region then we must not take the fault:
1110 */
1111 if (unlikely(in_atomic() || !mm)) {
1112 bad_area_nosemaphore(regs, error_code, address);
1113 return;
1114 }
1115
1116 /*
1117 * It's safe to allow irq's after cr2 has been saved and the
1118 * vmalloc fault has been handled.
1119 *
1120 * User-mode registers count as a user access even for any
1121 * potential system fault or CPU buglet:
1122 */
1123 if (user_mode_vm(regs)) {
1124 local_irq_enable();
1125 error_code |= PF_USER;
1126 flags |= FAULT_FLAG_USER;
1127 } else {
1128 if (regs->flags & X86_EFLAGS_IF)
1129 local_irq_enable();
1130 }
1131
1132 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1133
1134 if (error_code & PF_WRITE)
1135 flags |= FAULT_FLAG_WRITE;
1136
1137 /*
1138 * When running in the kernel we expect faults to occur only to
1139 * addresses in user space. All other faults represent errors in
1140 * the kernel and should generate an OOPS. Unfortunately, in the
1141 * case of an erroneous fault occurring in a code path which already
1142 * holds mmap_sem we will deadlock attempting to validate the fault
1143 * against the address space. Luckily the kernel only validly
1144 * references user space from well defined areas of code, which are
1145 * listed in the exceptions table.
1146 *
1147 * As the vast majority of faults will be valid we will only perform
1148 * the source reference check when there is a possibility of a
1149 * deadlock. Attempt to lock the address space, if we cannot we then
1150 * validate the source. If this is invalid we can skip the address
1151 * space check, thus avoiding the deadlock:
1152 */
1153 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1154 if ((error_code & PF_USER) == 0 &&
1155 !search_exception_tables(regs->ip)) {
1156 bad_area_nosemaphore(regs, error_code, address);
1157 return;
1158 }
1159retry:
1160 down_read(&mm->mmap_sem);
1161 } else {
1162 /*
1163 * The above down_read_trylock() might have succeeded in
1164 * which case we'll have missed the might_sleep() from
1165 * down_read():
1166 */
1167 might_sleep();
1168 }
1169
1170 vma = find_vma(mm, address);
1171 if (unlikely(!vma)) {
1172 bad_area(regs, error_code, address);
1173 return;
1174 }
1175 if (likely(vma->vm_start <= address))
1176 goto good_area;
1177 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1178 bad_area(regs, error_code, address);
1179 return;
1180 }
1181 if (error_code & PF_USER) {
1182 /*
1183 * Accessing the stack below %sp is always a bug.
1184 * The large cushion allows instructions like enter
1185 * and pusha to work. ("enter $65535, $31" pushes
1186 * 32 pointers and then decrements %sp by 65535.)
1187 */
1188 if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1189 bad_area(regs, error_code, address);
1190 return;
1191 }
1192 }
1193 if (unlikely(expand_stack(vma, address))) {
1194 bad_area(regs, error_code, address);
1195 return;
1196 }
1197
1198 /*
1199 * Ok, we have a good vm_area for this memory access, so
1200 * we can handle it..
1201 */
1202good_area:
1203 if (unlikely(access_error(error_code, vma))) {
1204 bad_area_access_error(regs, error_code, address);
1205 return;
1206 }
1207
1208 /*
1209 * If for any reason at all we couldn't handle the fault,
1210 * make sure we exit gracefully rather than endlessly redo
1211 * the fault:
1212 */
1213 fault = handle_mm_fault(mm, vma, address, flags);
1214
1215 /*
1216 * If we need to retry but a fatal signal is pending, handle the
1217 * signal first. We do not need to release the mmap_sem because it
1218 * would already be released in __lock_page_or_retry in mm/filemap.c.
1219 */
1220 if (unlikely((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)))
1221 return;
1222
1223 if (unlikely(fault & VM_FAULT_ERROR)) {
1224 mm_fault_error(regs, error_code, address, fault);
1225 return;
1226 }
1227
1228 /*
1229 * Major/minor page fault accounting is only done on the
1230 * initial attempt. If we go through a retry, it is extremely
1231 * likely that the page will be found in page cache at that point.
1232 */
1233 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1234 if (fault & VM_FAULT_MAJOR) {
1235 tsk->maj_flt++;
1236 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1237 regs, address);
1238 } else {
1239 tsk->min_flt++;
1240 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1241 regs, address);
1242 }
1243 if (fault & VM_FAULT_RETRY) {
1244 /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1245 * of starvation. */
1246 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1247 flags |= FAULT_FLAG_TRIED;
1248 goto retry;
1249 }
1250 }
1251
1252 check_v8086_mode(regs, address, tsk);
1253
1254 up_read(&mm->mmap_sem);
1255}
1256
1257dotraplinkage void __kprobes notrace
1258do_page_fault(struct pt_regs *regs, unsigned long error_code)
1259{
1260 unsigned long address = read_cr2(); /* Get the faulting address */
1261 enum ctx_state prev_state;
1262
1263 /*
1264 * We must have this function tagged with __kprobes, notrace and call
1265 * read_cr2() before calling anything else. To avoid calling any kind
1266 * of tracing machinery before we've observed the CR2 value.
1267 *
1268 * exception_{enter,exit}() contain all sorts of tracepoints.
1269 */
1270
1271 prev_state = exception_enter();
1272 __do_page_fault(regs, error_code, address);
1273 exception_exit(prev_state);
1274}
1275
1276#ifdef CONFIG_TRACING
1277static void trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1278 unsigned long error_code)
1279{
1280 if (user_mode(regs))
1281 trace_page_fault_user(address, regs, error_code);
1282 else
1283 trace_page_fault_kernel(address, regs, error_code);
1284}
1285
1286dotraplinkage void __kprobes notrace
1287trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1288{
1289 /*
1290 * The exception_enter and tracepoint processing could
1291 * trigger another page faults (user space callchain
1292 * reading) and destroy the original cr2 value, so read
1293 * the faulting address now.
1294 */
1295 unsigned long address = read_cr2();
1296 enum ctx_state prev_state;
1297
1298 prev_state = exception_enter();
1299 trace_page_fault_entries(address, regs, error_code);
1300 __do_page_fault(regs, error_code, address);
1301 exception_exit(prev_state);
1302}
1303#endif /* CONFIG_TRACING */
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/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>
21
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
33#define CREATE_TRACE_POINTS
34#include <asm/trace/exceptions.h>
35
36/*
37 * Returns 0 if mmiotrace is disabled, or if the fault is not
38 * handled by mmiotrace:
39 */
40static nokprobe_inline int
41kmmio_fault(struct pt_regs *regs, unsigned long addr)
42{
43 if (unlikely(is_kmmio_active()))
44 if (kmmio_handler(regs, addr) == 1)
45 return -1;
46 return 0;
47}
48
49/*
50 * Prefetch quirks:
51 *
52 * 32-bit mode:
53 *
54 * Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
55 * Check that here and ignore it.
56 *
57 * 64-bit mode:
58 *
59 * Sometimes the CPU reports invalid exceptions on prefetch.
60 * Check that here and ignore it.
61 *
62 * Opcode checker based on code by Richard Brunner.
63 */
64static inline int
65check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
66 unsigned char opcode, int *prefetch)
67{
68 unsigned char instr_hi = opcode & 0xf0;
69 unsigned char instr_lo = opcode & 0x0f;
70
71 switch (instr_hi) {
72 case 0x20:
73 case 0x30:
74 /*
75 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
76 * In X86_64 long mode, the CPU will signal invalid
77 * opcode if some of these prefixes are present so
78 * X86_64 will never get here anyway
79 */
80 return ((instr_lo & 7) == 0x6);
81#ifdef CONFIG_X86_64
82 case 0x40:
83 /*
84 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
85 * Need to figure out under what instruction mode the
86 * instruction was issued. Could check the LDT for lm,
87 * but for now it's good enough to assume that long
88 * mode only uses well known segments or kernel.
89 */
90 return (!user_mode(regs) || user_64bit_mode(regs));
91#endif
92 case 0x60:
93 /* 0x64 thru 0x67 are valid prefixes in all modes. */
94 return (instr_lo & 0xC) == 0x4;
95 case 0xF0:
96 /* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
97 return !instr_lo || (instr_lo>>1) == 1;
98 case 0x00:
99 /* Prefetch instruction is 0x0F0D or 0x0F18 */
100 if (probe_kernel_address(instr, opcode))
101 return 0;
102
103 *prefetch = (instr_lo == 0xF) &&
104 (opcode == 0x0D || opcode == 0x18);
105 return 0;
106 default:
107 return 0;
108 }
109}
110
111static int
112is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
113{
114 unsigned char *max_instr;
115 unsigned char *instr;
116 int prefetch = 0;
117
118 /*
119 * If it was a exec (instruction fetch) fault on NX page, then
120 * do not ignore the fault:
121 */
122 if (error_code & X86_PF_INSTR)
123 return 0;
124
125 instr = (void *)convert_ip_to_linear(current, regs);
126 max_instr = instr + 15;
127
128 if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
129 return 0;
130
131 while (instr < max_instr) {
132 unsigned char opcode;
133
134 if (probe_kernel_address(instr, opcode))
135 break;
136
137 instr++;
138
139 if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
140 break;
141 }
142 return prefetch;
143}
144
145DEFINE_SPINLOCK(pgd_lock);
146LIST_HEAD(pgd_list);
147
148#ifdef CONFIG_X86_32
149static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
150{
151 unsigned index = pgd_index(address);
152 pgd_t *pgd_k;
153 p4d_t *p4d, *p4d_k;
154 pud_t *pud, *pud_k;
155 pmd_t *pmd, *pmd_k;
156
157 pgd += index;
158 pgd_k = init_mm.pgd + index;
159
160 if (!pgd_present(*pgd_k))
161 return NULL;
162
163 /*
164 * set_pgd(pgd, *pgd_k); here would be useless on PAE
165 * and redundant with the set_pmd() on non-PAE. As would
166 * set_p4d/set_pud.
167 */
168 p4d = p4d_offset(pgd, address);
169 p4d_k = p4d_offset(pgd_k, address);
170 if (!p4d_present(*p4d_k))
171 return NULL;
172
173 pud = pud_offset(p4d, address);
174 pud_k = pud_offset(p4d_k, address);
175 if (!pud_present(*pud_k))
176 return NULL;
177
178 pmd = pmd_offset(pud, address);
179 pmd_k = pmd_offset(pud_k, address);
180
181 if (pmd_present(*pmd) != pmd_present(*pmd_k))
182 set_pmd(pmd, *pmd_k);
183
184 if (!pmd_present(*pmd_k))
185 return NULL;
186 else
187 BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
188
189 return pmd_k;
190}
191
192void vmalloc_sync_all(void)
193{
194 unsigned long address;
195
196 if (SHARED_KERNEL_PMD)
197 return;
198
199 for (address = VMALLOC_START & PMD_MASK;
200 address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
201 address += PMD_SIZE) {
202 struct page *page;
203
204 spin_lock(&pgd_lock);
205 list_for_each_entry(page, &pgd_list, lru) {
206 spinlock_t *pgt_lock;
207
208 /* the pgt_lock only for Xen */
209 pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
210
211 spin_lock(pgt_lock);
212 vmalloc_sync_one(page_address(page), address);
213 spin_unlock(pgt_lock);
214 }
215 spin_unlock(&pgd_lock);
216 }
217}
218
219/*
220 * 32-bit:
221 *
222 * Handle a fault on the vmalloc or module mapping area
223 */
224static noinline int vmalloc_fault(unsigned long address)
225{
226 unsigned long pgd_paddr;
227 pmd_t *pmd_k;
228 pte_t *pte_k;
229
230 /* Make sure we are in vmalloc area: */
231 if (!(address >= VMALLOC_START && address < VMALLOC_END))
232 return -1;
233
234 /*
235 * Synchronize this task's top level page-table
236 * with the 'reference' page table.
237 *
238 * Do _not_ use "current" here. We might be inside
239 * an interrupt in the middle of a task switch..
240 */
241 pgd_paddr = read_cr3_pa();
242 pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
243 if (!pmd_k)
244 return -1;
245
246 if (pmd_large(*pmd_k))
247 return 0;
248
249 pte_k = pte_offset_kernel(pmd_k, address);
250 if (!pte_present(*pte_k))
251 return -1;
252
253 return 0;
254}
255NOKPROBE_SYMBOL(vmalloc_fault);
256
257/*
258 * Did it hit the DOS screen memory VA from vm86 mode?
259 */
260static inline void
261check_v8086_mode(struct pt_regs *regs, unsigned long address,
262 struct task_struct *tsk)
263{
264#ifdef CONFIG_VM86
265 unsigned long bit;
266
267 if (!v8086_mode(regs) || !tsk->thread.vm86)
268 return;
269
270 bit = (address - 0xA0000) >> PAGE_SHIFT;
271 if (bit < 32)
272 tsk->thread.vm86->screen_bitmap |= 1 << bit;
273#endif
274}
275
276static bool low_pfn(unsigned long pfn)
277{
278 return pfn < max_low_pfn;
279}
280
281static void dump_pagetable(unsigned long address)
282{
283 pgd_t *base = __va(read_cr3_pa());
284 pgd_t *pgd = &base[pgd_index(address)];
285 p4d_t *p4d;
286 pud_t *pud;
287 pmd_t *pmd;
288 pte_t *pte;
289
290#ifdef CONFIG_X86_PAE
291 pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
292 if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
293 goto out;
294#define pr_pde pr_cont
295#else
296#define pr_pde pr_info
297#endif
298 p4d = p4d_offset(pgd, address);
299 pud = pud_offset(p4d, address);
300 pmd = pmd_offset(pud, address);
301 pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
302#undef pr_pde
303
304 /*
305 * We must not directly access the pte in the highpte
306 * case if the page table is located in highmem.
307 * And let's rather not kmap-atomic the pte, just in case
308 * it's allocated already:
309 */
310 if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
311 goto out;
312
313 pte = pte_offset_kernel(pmd, address);
314 pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
315out:
316 pr_cont("\n");
317}
318
319#else /* CONFIG_X86_64: */
320
321void vmalloc_sync_all(void)
322{
323 sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
324}
325
326/*
327 * 64-bit:
328 *
329 * Handle a fault on the vmalloc area
330 */
331static noinline int vmalloc_fault(unsigned long address)
332{
333 pgd_t *pgd, *pgd_k;
334 p4d_t *p4d, *p4d_k;
335 pud_t *pud;
336 pmd_t *pmd;
337 pte_t *pte;
338
339 /* Make sure we are in vmalloc area: */
340 if (!(address >= VMALLOC_START && address < VMALLOC_END))
341 return -1;
342
343 /*
344 * Copy kernel mappings over when needed. This can also
345 * happen within a race in page table update. In the later
346 * case just flush:
347 */
348 pgd = (pgd_t *)__va(read_cr3_pa()) + pgd_index(address);
349 pgd_k = pgd_offset_k(address);
350 if (pgd_none(*pgd_k))
351 return -1;
352
353 if (pgtable_l5_enabled()) {
354 if (pgd_none(*pgd)) {
355 set_pgd(pgd, *pgd_k);
356 arch_flush_lazy_mmu_mode();
357 } else {
358 BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_k));
359 }
360 }
361
362 /* With 4-level paging, copying happens on the p4d level. */
363 p4d = p4d_offset(pgd, address);
364 p4d_k = p4d_offset(pgd_k, address);
365 if (p4d_none(*p4d_k))
366 return -1;
367
368 if (p4d_none(*p4d) && !pgtable_l5_enabled()) {
369 set_p4d(p4d, *p4d_k);
370 arch_flush_lazy_mmu_mode();
371 } else {
372 BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_k));
373 }
374
375 BUILD_BUG_ON(CONFIG_PGTABLE_LEVELS < 4);
376
377 pud = pud_offset(p4d, address);
378 if (pud_none(*pud))
379 return -1;
380
381 if (pud_large(*pud))
382 return 0;
383
384 pmd = pmd_offset(pud, address);
385 if (pmd_none(*pmd))
386 return -1;
387
388 if (pmd_large(*pmd))
389 return 0;
390
391 pte = pte_offset_kernel(pmd, address);
392 if (!pte_present(*pte))
393 return -1;
394
395 return 0;
396}
397NOKPROBE_SYMBOL(vmalloc_fault);
398
399#ifdef CONFIG_CPU_SUP_AMD
400static const char errata93_warning[] =
401KERN_ERR
402"******* Your BIOS seems to not contain a fix for K8 errata #93\n"
403"******* Working around it, but it may cause SEGVs or burn power.\n"
404"******* Please consider a BIOS update.\n"
405"******* Disabling USB legacy in the BIOS may also help.\n";
406#endif
407
408/*
409 * No vm86 mode in 64-bit mode:
410 */
411static inline void
412check_v8086_mode(struct pt_regs *regs, unsigned long address,
413 struct task_struct *tsk)
414{
415}
416
417static int bad_address(void *p)
418{
419 unsigned long dummy;
420
421 return probe_kernel_address((unsigned long *)p, dummy);
422}
423
424static void dump_pagetable(unsigned long address)
425{
426 pgd_t *base = __va(read_cr3_pa());
427 pgd_t *pgd = base + pgd_index(address);
428 p4d_t *p4d;
429 pud_t *pud;
430 pmd_t *pmd;
431 pte_t *pte;
432
433 if (bad_address(pgd))
434 goto bad;
435
436 pr_info("PGD %lx ", pgd_val(*pgd));
437
438 if (!pgd_present(*pgd))
439 goto out;
440
441 p4d = p4d_offset(pgd, address);
442 if (bad_address(p4d))
443 goto bad;
444
445 pr_cont("P4D %lx ", p4d_val(*p4d));
446 if (!p4d_present(*p4d) || p4d_large(*p4d))
447 goto out;
448
449 pud = pud_offset(p4d, address);
450 if (bad_address(pud))
451 goto bad;
452
453 pr_cont("PUD %lx ", pud_val(*pud));
454 if (!pud_present(*pud) || pud_large(*pud))
455 goto out;
456
457 pmd = pmd_offset(pud, address);
458 if (bad_address(pmd))
459 goto bad;
460
461 pr_cont("PMD %lx ", pmd_val(*pmd));
462 if (!pmd_present(*pmd) || pmd_large(*pmd))
463 goto out;
464
465 pte = pte_offset_kernel(pmd, address);
466 if (bad_address(pte))
467 goto bad;
468
469 pr_cont("PTE %lx", pte_val(*pte));
470out:
471 pr_cont("\n");
472 return;
473bad:
474 pr_info("BAD\n");
475}
476
477#endif /* CONFIG_X86_64 */
478
479/*
480 * Workaround for K8 erratum #93 & buggy BIOS.
481 *
482 * BIOS SMM functions are required to use a specific workaround
483 * to avoid corruption of the 64bit RIP register on C stepping K8.
484 *
485 * A lot of BIOS that didn't get tested properly miss this.
486 *
487 * The OS sees this as a page fault with the upper 32bits of RIP cleared.
488 * Try to work around it here.
489 *
490 * Note we only handle faults in kernel here.
491 * Does nothing on 32-bit.
492 */
493static int is_errata93(struct pt_regs *regs, unsigned long address)
494{
495#if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
496 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
497 || boot_cpu_data.x86 != 0xf)
498 return 0;
499
500 if (address != regs->ip)
501 return 0;
502
503 if ((address >> 32) != 0)
504 return 0;
505
506 address |= 0xffffffffUL << 32;
507 if ((address >= (u64)_stext && address <= (u64)_etext) ||
508 (address >= MODULES_VADDR && address <= MODULES_END)) {
509 printk_once(errata93_warning);
510 regs->ip = address;
511 return 1;
512 }
513#endif
514 return 0;
515}
516
517/*
518 * Work around K8 erratum #100 K8 in compat mode occasionally jumps
519 * to illegal addresses >4GB.
520 *
521 * We catch this in the page fault handler because these addresses
522 * are not reachable. Just detect this case and return. Any code
523 * segment in LDT is compatibility mode.
524 */
525static int is_errata100(struct pt_regs *regs, unsigned long address)
526{
527#ifdef CONFIG_X86_64
528 if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
529 return 1;
530#endif
531 return 0;
532}
533
534static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
535{
536#ifdef CONFIG_X86_F00F_BUG
537 unsigned long nr;
538
539 /*
540 * Pentium F0 0F C7 C8 bug workaround:
541 */
542 if (boot_cpu_has_bug(X86_BUG_F00F)) {
543 nr = (address - idt_descr.address) >> 3;
544
545 if (nr == 6) {
546 do_invalid_op(regs, 0);
547 return 1;
548 }
549 }
550#endif
551 return 0;
552}
553
554static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
555{
556 u32 offset = (index >> 3) * sizeof(struct desc_struct);
557 unsigned long addr;
558 struct ldttss_desc desc;
559
560 if (index == 0) {
561 pr_alert("%s: NULL\n", name);
562 return;
563 }
564
565 if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
566 pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
567 return;
568 }
569
570 if (probe_kernel_read(&desc, (void *)(gdt->address + offset),
571 sizeof(struct ldttss_desc))) {
572 pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
573 name, index);
574 return;
575 }
576
577 addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
578#ifdef CONFIG_X86_64
579 addr |= ((u64)desc.base3 << 32);
580#endif
581 pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
582 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
583}
584
585static void
586show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
587{
588 if (!oops_may_print())
589 return;
590
591 if (error_code & X86_PF_INSTR) {
592 unsigned int level;
593 pgd_t *pgd;
594 pte_t *pte;
595
596 pgd = __va(read_cr3_pa());
597 pgd += pgd_index(address);
598
599 pte = lookup_address_in_pgd(pgd, address, &level);
600
601 if (pte && pte_present(*pte) && !pte_exec(*pte))
602 pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
603 from_kuid(&init_user_ns, current_uid()));
604 if (pte && pte_present(*pte) && pte_exec(*pte) &&
605 (pgd_flags(*pgd) & _PAGE_USER) &&
606 (__read_cr4() & X86_CR4_SMEP))
607 pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
608 from_kuid(&init_user_ns, current_uid()));
609 }
610
611 if (address < PAGE_SIZE && !user_mode(regs))
612 pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
613 (void *)address);
614 else
615 pr_alert("BUG: unable to handle page fault for address: %px\n",
616 (void *)address);
617
618 pr_alert("#PF: %s %s in %s mode\n",
619 (error_code & X86_PF_USER) ? "user" : "supervisor",
620 (error_code & X86_PF_INSTR) ? "instruction fetch" :
621 (error_code & X86_PF_WRITE) ? "write access" :
622 "read access",
623 user_mode(regs) ? "user" : "kernel");
624 pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
625 !(error_code & X86_PF_PROT) ? "not-present page" :
626 (error_code & X86_PF_RSVD) ? "reserved bit violation" :
627 (error_code & X86_PF_PK) ? "protection keys violation" :
628 "permissions violation");
629
630 if (!(error_code & X86_PF_USER) && user_mode(regs)) {
631 struct desc_ptr idt, gdt;
632 u16 ldtr, tr;
633
634 /*
635 * This can happen for quite a few reasons. The more obvious
636 * ones are faults accessing the GDT, or LDT. Perhaps
637 * surprisingly, if the CPU tries to deliver a benign or
638 * contributory exception from user code and gets a page fault
639 * during delivery, the page fault can be delivered as though
640 * it originated directly from user code. This could happen
641 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
642 * kernel or IST stack.
643 */
644 store_idt(&idt);
645
646 /* Usable even on Xen PV -- it's just slow. */
647 native_store_gdt(&gdt);
648
649 pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
650 idt.address, idt.size, gdt.address, gdt.size);
651
652 store_ldt(ldtr);
653 show_ldttss(&gdt, "LDTR", ldtr);
654
655 store_tr(tr);
656 show_ldttss(&gdt, "TR", tr);
657 }
658
659 dump_pagetable(address);
660}
661
662static noinline void
663pgtable_bad(struct pt_regs *regs, unsigned long error_code,
664 unsigned long address)
665{
666 struct task_struct *tsk;
667 unsigned long flags;
668 int sig;
669
670 flags = oops_begin();
671 tsk = current;
672 sig = SIGKILL;
673
674 printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
675 tsk->comm, address);
676 dump_pagetable(address);
677
678 if (__die("Bad pagetable", regs, error_code))
679 sig = 0;
680
681 oops_end(flags, regs, sig);
682}
683
684static void set_signal_archinfo(unsigned long address,
685 unsigned long error_code)
686{
687 struct task_struct *tsk = current;
688
689 /*
690 * To avoid leaking information about the kernel page
691 * table layout, pretend that user-mode accesses to
692 * kernel addresses are always protection faults.
693 *
694 * NB: This means that failed vsyscalls with vsyscall=none
695 * will have the PROT bit. This doesn't leak any
696 * information and does not appear to cause any problems.
697 */
698 if (address >= TASK_SIZE_MAX)
699 error_code |= X86_PF_PROT;
700
701 tsk->thread.trap_nr = X86_TRAP_PF;
702 tsk->thread.error_code = error_code | X86_PF_USER;
703 tsk->thread.cr2 = address;
704}
705
706static noinline void
707no_context(struct pt_regs *regs, unsigned long error_code,
708 unsigned long address, int signal, int si_code)
709{
710 struct task_struct *tsk = current;
711 unsigned long flags;
712 int sig;
713
714 if (user_mode(regs)) {
715 /*
716 * This is an implicit supervisor-mode access from user
717 * mode. Bypass all the kernel-mode recovery code and just
718 * OOPS.
719 */
720 goto oops;
721 }
722
723 /* Are we prepared to handle this kernel fault? */
724 if (fixup_exception(regs, X86_TRAP_PF, error_code, address)) {
725 /*
726 * Any interrupt that takes a fault gets the fixup. This makes
727 * the below recursive fault logic only apply to a faults from
728 * task context.
729 */
730 if (in_interrupt())
731 return;
732
733 /*
734 * Per the above we're !in_interrupt(), aka. task context.
735 *
736 * In this case we need to make sure we're not recursively
737 * faulting through the emulate_vsyscall() logic.
738 */
739 if (current->thread.sig_on_uaccess_err && signal) {
740 set_signal_archinfo(address, error_code);
741
742 /* XXX: hwpoison faults will set the wrong code. */
743 force_sig_fault(signal, si_code, (void __user *)address);
744 }
745
746 /*
747 * Barring that, we can do the fixup and be happy.
748 */
749 return;
750 }
751
752#ifdef CONFIG_VMAP_STACK
753 /*
754 * Stack overflow? During boot, we can fault near the initial
755 * stack in the direct map, but that's not an overflow -- check
756 * that we're in vmalloc space to avoid this.
757 */
758 if (is_vmalloc_addr((void *)address) &&
759 (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
760 address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
761 unsigned long stack = __this_cpu_ist_top_va(DF) - sizeof(void *);
762 /*
763 * We're likely to be running with very little stack space
764 * left. It's plausible that we'd hit this condition but
765 * double-fault even before we get this far, in which case
766 * we're fine: the double-fault handler will deal with it.
767 *
768 * We don't want to make it all the way into the oops code
769 * and then double-fault, though, because we're likely to
770 * break the console driver and lose most of the stack dump.
771 */
772 asm volatile ("movq %[stack], %%rsp\n\t"
773 "call handle_stack_overflow\n\t"
774 "1: jmp 1b"
775 : ASM_CALL_CONSTRAINT
776 : "D" ("kernel stack overflow (page fault)"),
777 "S" (regs), "d" (address),
778 [stack] "rm" (stack));
779 unreachable();
780 }
781#endif
782
783 /*
784 * 32-bit:
785 *
786 * Valid to do another page fault here, because if this fault
787 * had been triggered by is_prefetch fixup_exception would have
788 * handled it.
789 *
790 * 64-bit:
791 *
792 * Hall of shame of CPU/BIOS bugs.
793 */
794 if (is_prefetch(regs, error_code, address))
795 return;
796
797 if (is_errata93(regs, address))
798 return;
799
800 /*
801 * Buggy firmware could access regions which might page fault, try to
802 * recover from such faults.
803 */
804 if (IS_ENABLED(CONFIG_EFI))
805 efi_recover_from_page_fault(address);
806
807oops:
808 /*
809 * Oops. The kernel tried to access some bad page. We'll have to
810 * terminate things with extreme prejudice:
811 */
812 flags = oops_begin();
813
814 show_fault_oops(regs, error_code, address);
815
816 if (task_stack_end_corrupted(tsk))
817 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
818
819 sig = SIGKILL;
820 if (__die("Oops", regs, error_code))
821 sig = 0;
822
823 /* Executive summary in case the body of the oops scrolled away */
824 printk(KERN_DEFAULT "CR2: %016lx\n", address);
825
826 oops_end(flags, regs, sig);
827}
828
829/*
830 * Print out info about fatal segfaults, if the show_unhandled_signals
831 * sysctl is set:
832 */
833static inline void
834show_signal_msg(struct pt_regs *regs, unsigned long error_code,
835 unsigned long address, struct task_struct *tsk)
836{
837 const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
838
839 if (!unhandled_signal(tsk, SIGSEGV))
840 return;
841
842 if (!printk_ratelimit())
843 return;
844
845 printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
846 loglvl, tsk->comm, task_pid_nr(tsk), address,
847 (void *)regs->ip, (void *)regs->sp, error_code);
848
849 print_vma_addr(KERN_CONT " in ", regs->ip);
850
851 printk(KERN_CONT "\n");
852
853 show_opcodes(regs, loglvl);
854}
855
856/*
857 * The (legacy) vsyscall page is the long page in the kernel portion
858 * of the address space that has user-accessible permissions.
859 */
860static bool is_vsyscall_vaddr(unsigned long vaddr)
861{
862 return unlikely((vaddr & PAGE_MASK) == VSYSCALL_ADDR);
863}
864
865static void
866__bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
867 unsigned long address, u32 pkey, int si_code)
868{
869 struct task_struct *tsk = current;
870
871 /* User mode accesses just cause a SIGSEGV */
872 if (user_mode(regs) && (error_code & X86_PF_USER)) {
873 /*
874 * It's possible to have interrupts off here:
875 */
876 local_irq_enable();
877
878 /*
879 * Valid to do another page fault here because this one came
880 * from user space:
881 */
882 if (is_prefetch(regs, error_code, address))
883 return;
884
885 if (is_errata100(regs, address))
886 return;
887
888 /*
889 * To avoid leaking information about the kernel page table
890 * layout, pretend that user-mode accesses to kernel addresses
891 * are always protection faults.
892 */
893 if (address >= TASK_SIZE_MAX)
894 error_code |= X86_PF_PROT;
895
896 if (likely(show_unhandled_signals))
897 show_signal_msg(regs, error_code, address, tsk);
898
899 set_signal_archinfo(address, error_code);
900
901 if (si_code == SEGV_PKUERR)
902 force_sig_pkuerr((void __user *)address, pkey);
903
904 force_sig_fault(SIGSEGV, si_code, (void __user *)address);
905
906 return;
907 }
908
909 if (is_f00f_bug(regs, address))
910 return;
911
912 no_context(regs, error_code, address, SIGSEGV, si_code);
913}
914
915static noinline void
916bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
917 unsigned long address)
918{
919 __bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
920}
921
922static void
923__bad_area(struct pt_regs *regs, unsigned long error_code,
924 unsigned long address, u32 pkey, int si_code)
925{
926 struct mm_struct *mm = current->mm;
927 /*
928 * Something tried to access memory that isn't in our memory map..
929 * Fix it, but check if it's kernel or user first..
930 */
931 up_read(&mm->mmap_sem);
932
933 __bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
934}
935
936static noinline void
937bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
938{
939 __bad_area(regs, error_code, address, 0, SEGV_MAPERR);
940}
941
942static inline bool bad_area_access_from_pkeys(unsigned long error_code,
943 struct vm_area_struct *vma)
944{
945 /* This code is always called on the current mm */
946 bool foreign = false;
947
948 if (!boot_cpu_has(X86_FEATURE_OSPKE))
949 return false;
950 if (error_code & X86_PF_PK)
951 return true;
952 /* this checks permission keys on the VMA: */
953 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
954 (error_code & X86_PF_INSTR), foreign))
955 return true;
956 return false;
957}
958
959static noinline void
960bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
961 unsigned long address, struct vm_area_struct *vma)
962{
963 /*
964 * This OSPKE check is not strictly necessary at runtime.
965 * But, doing it this way allows compiler optimizations
966 * if pkeys are compiled out.
967 */
968 if (bad_area_access_from_pkeys(error_code, vma)) {
969 /*
970 * A protection key fault means that the PKRU value did not allow
971 * access to some PTE. Userspace can figure out what PKRU was
972 * from the XSAVE state. This function captures the pkey from
973 * the vma and passes it to userspace so userspace can discover
974 * which protection key was set on the PTE.
975 *
976 * If we get here, we know that the hardware signaled a X86_PF_PK
977 * fault and that there was a VMA once we got in the fault
978 * handler. It does *not* guarantee that the VMA we find here
979 * was the one that we faulted on.
980 *
981 * 1. T1 : mprotect_key(foo, PAGE_SIZE, pkey=4);
982 * 2. T1 : set PKRU to deny access to pkey=4, touches page
983 * 3. T1 : faults...
984 * 4. T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
985 * 5. T1 : enters fault handler, takes mmap_sem, etc...
986 * 6. T1 : reaches here, sees vma_pkey(vma)=5, when we really
987 * faulted on a pte with its pkey=4.
988 */
989 u32 pkey = vma_pkey(vma);
990
991 __bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
992 } else {
993 __bad_area(regs, error_code, address, 0, SEGV_ACCERR);
994 }
995}
996
997static void
998do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
999 vm_fault_t fault)
1000{
1001 /* Kernel mode? Handle exceptions or die: */
1002 if (!(error_code & X86_PF_USER)) {
1003 no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1004 return;
1005 }
1006
1007 /* User-space => ok to do another page fault: */
1008 if (is_prefetch(regs, error_code, address))
1009 return;
1010
1011 set_signal_archinfo(address, error_code);
1012
1013#ifdef CONFIG_MEMORY_FAILURE
1014 if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
1015 struct task_struct *tsk = current;
1016 unsigned lsb = 0;
1017
1018 pr_err(
1019 "MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
1020 tsk->comm, tsk->pid, address);
1021 if (fault & VM_FAULT_HWPOISON_LARGE)
1022 lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
1023 if (fault & VM_FAULT_HWPOISON)
1024 lsb = PAGE_SHIFT;
1025 force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
1026 return;
1027 }
1028#endif
1029 force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
1030}
1031
1032static noinline void
1033mm_fault_error(struct pt_regs *regs, unsigned long error_code,
1034 unsigned long address, vm_fault_t fault)
1035{
1036 if (fatal_signal_pending(current) && !(error_code & X86_PF_USER)) {
1037 no_context(regs, error_code, address, 0, 0);
1038 return;
1039 }
1040
1041 if (fault & VM_FAULT_OOM) {
1042 /* Kernel mode? Handle exceptions or die: */
1043 if (!(error_code & X86_PF_USER)) {
1044 no_context(regs, error_code, address,
1045 SIGSEGV, SEGV_MAPERR);
1046 return;
1047 }
1048
1049 /*
1050 * We ran out of memory, call the OOM killer, and return the
1051 * userspace (which will retry the fault, or kill us if we got
1052 * oom-killed):
1053 */
1054 pagefault_out_of_memory();
1055 } else {
1056 if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1057 VM_FAULT_HWPOISON_LARGE))
1058 do_sigbus(regs, error_code, address, fault);
1059 else if (fault & VM_FAULT_SIGSEGV)
1060 bad_area_nosemaphore(regs, error_code, address);
1061 else
1062 BUG();
1063 }
1064}
1065
1066static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
1067{
1068 if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
1069 return 0;
1070
1071 if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
1072 return 0;
1073
1074 return 1;
1075}
1076
1077/*
1078 * Handle a spurious fault caused by a stale TLB entry.
1079 *
1080 * This allows us to lazily refresh the TLB when increasing the
1081 * permissions of a kernel page (RO -> RW or NX -> X). Doing it
1082 * eagerly is very expensive since that implies doing a full
1083 * cross-processor TLB flush, even if no stale TLB entries exist
1084 * on other processors.
1085 *
1086 * Spurious faults may only occur if the TLB contains an entry with
1087 * fewer permission than the page table entry. Non-present (P = 0)
1088 * and reserved bit (R = 1) faults are never spurious.
1089 *
1090 * There are no security implications to leaving a stale TLB when
1091 * increasing the permissions on a page.
1092 *
1093 * Returns non-zero if a spurious fault was handled, zero otherwise.
1094 *
1095 * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1096 * (Optional Invalidation).
1097 */
1098static noinline int
1099spurious_kernel_fault(unsigned long error_code, unsigned long address)
1100{
1101 pgd_t *pgd;
1102 p4d_t *p4d;
1103 pud_t *pud;
1104 pmd_t *pmd;
1105 pte_t *pte;
1106 int ret;
1107
1108 /*
1109 * Only writes to RO or instruction fetches from NX may cause
1110 * spurious faults.
1111 *
1112 * These could be from user or supervisor accesses but the TLB
1113 * is only lazily flushed after a kernel mapping protection
1114 * change, so user accesses are not expected to cause spurious
1115 * faults.
1116 */
1117 if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
1118 error_code != (X86_PF_INSTR | X86_PF_PROT))
1119 return 0;
1120
1121 pgd = init_mm.pgd + pgd_index(address);
1122 if (!pgd_present(*pgd))
1123 return 0;
1124
1125 p4d = p4d_offset(pgd, address);
1126 if (!p4d_present(*p4d))
1127 return 0;
1128
1129 if (p4d_large(*p4d))
1130 return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1131
1132 pud = pud_offset(p4d, address);
1133 if (!pud_present(*pud))
1134 return 0;
1135
1136 if (pud_large(*pud))
1137 return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1138
1139 pmd = pmd_offset(pud, address);
1140 if (!pmd_present(*pmd))
1141 return 0;
1142
1143 if (pmd_large(*pmd))
1144 return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1145
1146 pte = pte_offset_kernel(pmd, address);
1147 if (!pte_present(*pte))
1148 return 0;
1149
1150 ret = spurious_kernel_fault_check(error_code, pte);
1151 if (!ret)
1152 return 0;
1153
1154 /*
1155 * Make sure we have permissions in PMD.
1156 * If not, then there's a bug in the page tables:
1157 */
1158 ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1159 WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1160
1161 return ret;
1162}
1163NOKPROBE_SYMBOL(spurious_kernel_fault);
1164
1165int show_unhandled_signals = 1;
1166
1167static inline int
1168access_error(unsigned long error_code, struct vm_area_struct *vma)
1169{
1170 /* This is only called for the current mm, so: */
1171 bool foreign = false;
1172
1173 /*
1174 * Read or write was blocked by protection keys. This is
1175 * always an unconditional error and can never result in
1176 * a follow-up action to resolve the fault, like a COW.
1177 */
1178 if (error_code & X86_PF_PK)
1179 return 1;
1180
1181 /*
1182 * Make sure to check the VMA so that we do not perform
1183 * faults just to hit a X86_PF_PK as soon as we fill in a
1184 * page.
1185 */
1186 if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1187 (error_code & X86_PF_INSTR), foreign))
1188 return 1;
1189
1190 if (error_code & X86_PF_WRITE) {
1191 /* write, present and write, not present: */
1192 if (unlikely(!(vma->vm_flags & VM_WRITE)))
1193 return 1;
1194 return 0;
1195 }
1196
1197 /* read, present: */
1198 if (unlikely(error_code & X86_PF_PROT))
1199 return 1;
1200
1201 /* read, not present: */
1202 if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1203 return 1;
1204
1205 return 0;
1206}
1207
1208static int fault_in_kernel_space(unsigned long address)
1209{
1210 /*
1211 * On 64-bit systems, the vsyscall page is at an address above
1212 * TASK_SIZE_MAX, but is not considered part of the kernel
1213 * address space.
1214 */
1215 if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1216 return false;
1217
1218 return address >= TASK_SIZE_MAX;
1219}
1220
1221/*
1222 * Called for all faults where 'address' is part of the kernel address
1223 * space. Might get called for faults that originate from *code* that
1224 * ran in userspace or the kernel.
1225 */
1226static void
1227do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1228 unsigned long address)
1229{
1230 /*
1231 * Protection keys exceptions only happen on user pages. We
1232 * have no user pages in the kernel portion of the address
1233 * space, so do not expect them here.
1234 */
1235 WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1236
1237 /*
1238 * We can fault-in kernel-space virtual memory on-demand. The
1239 * 'reference' page table is init_mm.pgd.
1240 *
1241 * NOTE! We MUST NOT take any locks for this case. We may
1242 * be in an interrupt or a critical region, and should
1243 * only copy the information from the master page table,
1244 * nothing more.
1245 *
1246 * Before doing this on-demand faulting, ensure that the
1247 * fault is not any of the following:
1248 * 1. A fault on a PTE with a reserved bit set.
1249 * 2. A fault caused by a user-mode access. (Do not demand-
1250 * fault kernel memory due to user-mode accesses).
1251 * 3. A fault caused by a page-level protection violation.
1252 * (A demand fault would be on a non-present page which
1253 * would have X86_PF_PROT==0).
1254 */
1255 if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1256 if (vmalloc_fault(address) >= 0)
1257 return;
1258 }
1259
1260 /* Was the fault spurious, caused by lazy TLB invalidation? */
1261 if (spurious_kernel_fault(hw_error_code, address))
1262 return;
1263
1264 /* kprobes don't want to hook the spurious faults: */
1265 if (kprobe_page_fault(regs, X86_TRAP_PF))
1266 return;
1267
1268 /*
1269 * Note, despite being a "bad area", there are quite a few
1270 * acceptable reasons to get here, such as erratum fixups
1271 * and handling kernel code that can fault, like get_user().
1272 *
1273 * Don't take the mm semaphore here. If we fixup a prefetch
1274 * fault we could otherwise deadlock:
1275 */
1276 bad_area_nosemaphore(regs, hw_error_code, address);
1277}
1278NOKPROBE_SYMBOL(do_kern_addr_fault);
1279
1280/* Handle faults in the user portion of the address space */
1281static inline
1282void do_user_addr_fault(struct pt_regs *regs,
1283 unsigned long hw_error_code,
1284 unsigned long address)
1285{
1286 struct vm_area_struct *vma;
1287 struct task_struct *tsk;
1288 struct mm_struct *mm;
1289 vm_fault_t fault, major = 0;
1290 unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1291
1292 tsk = current;
1293 mm = tsk->mm;
1294
1295 /* kprobes don't want to hook the spurious faults: */
1296 if (unlikely(kprobe_page_fault(regs, X86_TRAP_PF)))
1297 return;
1298
1299 /*
1300 * Reserved bits are never expected to be set on
1301 * entries in the user portion of the page tables.
1302 */
1303 if (unlikely(hw_error_code & X86_PF_RSVD))
1304 pgtable_bad(regs, hw_error_code, address);
1305
1306 /*
1307 * If SMAP is on, check for invalid kernel (supervisor) access to user
1308 * pages in the user address space. The odd case here is WRUSS,
1309 * which, according to the preliminary documentation, does not respect
1310 * SMAP and will have the USER bit set so, in all cases, SMAP
1311 * enforcement appears to be consistent with the USER bit.
1312 */
1313 if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1314 !(hw_error_code & X86_PF_USER) &&
1315 !(regs->flags & X86_EFLAGS_AC)))
1316 {
1317 bad_area_nosemaphore(regs, hw_error_code, address);
1318 return;
1319 }
1320
1321 /*
1322 * If we're in an interrupt, have no user context or are running
1323 * in a region with pagefaults disabled then we must not take the fault
1324 */
1325 if (unlikely(faulthandler_disabled() || !mm)) {
1326 bad_area_nosemaphore(regs, hw_error_code, address);
1327 return;
1328 }
1329
1330 /*
1331 * It's safe to allow irq's after cr2 has been saved and the
1332 * vmalloc fault has been handled.
1333 *
1334 * User-mode registers count as a user access even for any
1335 * potential system fault or CPU buglet:
1336 */
1337 if (user_mode(regs)) {
1338 local_irq_enable();
1339 flags |= FAULT_FLAG_USER;
1340 } else {
1341 if (regs->flags & X86_EFLAGS_IF)
1342 local_irq_enable();
1343 }
1344
1345 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1346
1347 if (hw_error_code & X86_PF_WRITE)
1348 flags |= FAULT_FLAG_WRITE;
1349 if (hw_error_code & X86_PF_INSTR)
1350 flags |= FAULT_FLAG_INSTRUCTION;
1351
1352#ifdef CONFIG_X86_64
1353 /*
1354 * Faults in the vsyscall page might need emulation. The
1355 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1356 * considered to be part of the user address space.
1357 *
1358 * The vsyscall page does not have a "real" VMA, so do this
1359 * emulation before we go searching for VMAs.
1360 *
1361 * PKRU never rejects instruction fetches, so we don't need
1362 * to consider the PF_PK bit.
1363 */
1364 if (is_vsyscall_vaddr(address)) {
1365 if (emulate_vsyscall(hw_error_code, regs, address))
1366 return;
1367 }
1368#endif
1369
1370 /*
1371 * Kernel-mode access to the user address space should only occur
1372 * on well-defined single instructions listed in the exception
1373 * tables. But, an erroneous kernel fault occurring outside one of
1374 * those areas which also holds mmap_sem might deadlock attempting
1375 * to validate the fault against the address space.
1376 *
1377 * Only do the expensive exception table search when we might be at
1378 * risk of a deadlock. This happens if we
1379 * 1. Failed to acquire mmap_sem, and
1380 * 2. The access did not originate in userspace.
1381 */
1382 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1383 if (!user_mode(regs) && !search_exception_tables(regs->ip)) {
1384 /*
1385 * Fault from code in kernel from
1386 * which we do not expect faults.
1387 */
1388 bad_area_nosemaphore(regs, hw_error_code, address);
1389 return;
1390 }
1391retry:
1392 down_read(&mm->mmap_sem);
1393 } else {
1394 /*
1395 * The above down_read_trylock() might have succeeded in
1396 * which case we'll have missed the might_sleep() from
1397 * down_read():
1398 */
1399 might_sleep();
1400 }
1401
1402 vma = find_vma(mm, address);
1403 if (unlikely(!vma)) {
1404 bad_area(regs, hw_error_code, address);
1405 return;
1406 }
1407 if (likely(vma->vm_start <= address))
1408 goto good_area;
1409 if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1410 bad_area(regs, hw_error_code, address);
1411 return;
1412 }
1413 if (unlikely(expand_stack(vma, address))) {
1414 bad_area(regs, hw_error_code, address);
1415 return;
1416 }
1417
1418 /*
1419 * Ok, we have a good vm_area for this memory access, so
1420 * we can handle it..
1421 */
1422good_area:
1423 if (unlikely(access_error(hw_error_code, vma))) {
1424 bad_area_access_error(regs, hw_error_code, address, vma);
1425 return;
1426 }
1427
1428 /*
1429 * If for any reason at all we couldn't handle the fault,
1430 * make sure we exit gracefully rather than endlessly redo
1431 * the fault. Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1432 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1433 *
1434 * Note that handle_userfault() may also release and reacquire mmap_sem
1435 * (and not return with VM_FAULT_RETRY), when returning to userland to
1436 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1437 * (potentially after handling any pending signal during the return to
1438 * userland). The return to userland is identified whenever
1439 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1440 */
1441 fault = handle_mm_fault(vma, address, flags);
1442 major |= fault & VM_FAULT_MAJOR;
1443
1444 /*
1445 * If we need to retry the mmap_sem has already been released,
1446 * and if there is a fatal signal pending there is no guarantee
1447 * that we made any progress. Handle this case first.
1448 */
1449 if (unlikely(fault & VM_FAULT_RETRY)) {
1450 /* Retry at most once */
1451 if (flags & FAULT_FLAG_ALLOW_RETRY) {
1452 flags &= ~FAULT_FLAG_ALLOW_RETRY;
1453 flags |= FAULT_FLAG_TRIED;
1454 if (!fatal_signal_pending(tsk))
1455 goto retry;
1456 }
1457
1458 /* User mode? Just return to handle the fatal exception */
1459 if (flags & FAULT_FLAG_USER)
1460 return;
1461
1462 /* Not returning to user mode? Handle exceptions or die: */
1463 no_context(regs, hw_error_code, address, SIGBUS, BUS_ADRERR);
1464 return;
1465 }
1466
1467 up_read(&mm->mmap_sem);
1468 if (unlikely(fault & VM_FAULT_ERROR)) {
1469 mm_fault_error(regs, hw_error_code, address, fault);
1470 return;
1471 }
1472
1473 /*
1474 * Major/minor page fault accounting. If any of the events
1475 * returned VM_FAULT_MAJOR, we account it as a major fault.
1476 */
1477 if (major) {
1478 tsk->maj_flt++;
1479 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1480 } else {
1481 tsk->min_flt++;
1482 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1483 }
1484
1485 check_v8086_mode(regs, address, tsk);
1486}
1487NOKPROBE_SYMBOL(do_user_addr_fault);
1488
1489/*
1490 * Explicitly marked noinline such that the function tracer sees this as the
1491 * page_fault entry point.
1492 */
1493static noinline void
1494__do_page_fault(struct pt_regs *regs, unsigned long hw_error_code,
1495 unsigned long address)
1496{
1497 prefetchw(¤t->mm->mmap_sem);
1498
1499 if (unlikely(kmmio_fault(regs, address)))
1500 return;
1501
1502 /* Was the fault on kernel-controlled part of the address space? */
1503 if (unlikely(fault_in_kernel_space(address)))
1504 do_kern_addr_fault(regs, hw_error_code, address);
1505 else
1506 do_user_addr_fault(regs, hw_error_code, address);
1507}
1508NOKPROBE_SYMBOL(__do_page_fault);
1509
1510static __always_inline void
1511trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1512 unsigned long address)
1513{
1514 if (!trace_pagefault_enabled())
1515 return;
1516
1517 if (user_mode(regs))
1518 trace_page_fault_user(address, regs, error_code);
1519 else
1520 trace_page_fault_kernel(address, regs, error_code);
1521}
1522
1523dotraplinkage void
1524do_page_fault(struct pt_regs *regs, unsigned long error_code, unsigned long address)
1525{
1526 enum ctx_state prev_state;
1527
1528 prev_state = exception_enter();
1529 trace_page_fault_entries(regs, error_code, address);
1530 __do_page_fault(regs, error_code, address);
1531 exception_exit(prev_state);
1532}
1533NOKPROBE_SYMBOL(do_page_fault);