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