xref: /openbmc/linux/arch/x86/mm/fault.c (revision a977d045)
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/sched/task_stack.h>	/* task_stack_*(), ...		*/
8 #include <linux/kdebug.h>		/* oops_begin/end, ...		*/
9 #include <linux/extable.h>		/* search_exception_tables	*/
10 #include <linux/bootmem.h>		/* max_low_pfn			*/
11 #include <linux/kprobes.h>		/* NOKPROBE_SYMBOL, ...		*/
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 #include <linux/uaccess.h>		/* faulthandler_disabled()	*/
18 
19 #include <asm/cpufeature.h>		/* boot_cpu_has, ...		*/
20 #include <asm/traps.h>			/* dotraplinkage, ...		*/
21 #include <asm/pgalloc.h>		/* pgd_*(), ...			*/
22 #include <asm/kmemcheck.h>		/* kmemcheck_*(), ...		*/
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  * Page fault error code bits:
33  *
34  *   bit 0 ==	 0: no page found	1: protection fault
35  *   bit 1 ==	 0: read access		1: write access
36  *   bit 2 ==	 0: kernel-mode access	1: user-mode access
37  *   bit 3 ==				1: use of reserved bit detected
38  *   bit 4 ==				1: fault was an instruction fetch
39  *   bit 5 ==				1: protection keys block access
40  */
41 enum x86_pf_error_code {
42 
43 	PF_PROT		=		1 << 0,
44 	PF_WRITE	=		1 << 1,
45 	PF_USER		=		1 << 2,
46 	PF_RSVD		=		1 << 3,
47 	PF_INSTR	=		1 << 4,
48 	PF_PK		=		1 << 5,
49 };
50 
51 /*
52  * Returns 0 if mmiotrace is disabled, or if the fault is not
53  * handled by mmiotrace:
54  */
55 static nokprobe_inline int
56 kmmio_fault(struct pt_regs *regs, unsigned long addr)
57 {
58 	if (unlikely(is_kmmio_active()))
59 		if (kmmio_handler(regs, addr) == 1)
60 			return -1;
61 	return 0;
62 }
63 
64 static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
65 {
66 	int ret = 0;
67 
68 	/* kprobe_running() needs smp_processor_id() */
69 	if (kprobes_built_in() && !user_mode(regs)) {
70 		preempt_disable();
71 		if (kprobe_running() && kprobe_fault_handler(regs, 14))
72 			ret = 1;
73 		preempt_enable();
74 	}
75 
76 	return ret;
77 }
78 
79 /*
80  * Prefetch quirks:
81  *
82  * 32-bit mode:
83  *
84  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
85  *   Check that here and ignore it.
86  *
87  * 64-bit mode:
88  *
89  *   Sometimes the CPU reports invalid exceptions on prefetch.
90  *   Check that here and ignore it.
91  *
92  * Opcode checker based on code by Richard Brunner.
93  */
94 static inline int
95 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
96 		      unsigned char opcode, int *prefetch)
97 {
98 	unsigned char instr_hi = opcode & 0xf0;
99 	unsigned char instr_lo = opcode & 0x0f;
100 
101 	switch (instr_hi) {
102 	case 0x20:
103 	case 0x30:
104 		/*
105 		 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
106 		 * In X86_64 long mode, the CPU will signal invalid
107 		 * opcode if some of these prefixes are present so
108 		 * X86_64 will never get here anyway
109 		 */
110 		return ((instr_lo & 7) == 0x6);
111 #ifdef CONFIG_X86_64
112 	case 0x40:
113 		/*
114 		 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
115 		 * Need to figure out under what instruction mode the
116 		 * instruction was issued. Could check the LDT for lm,
117 		 * but for now it's good enough to assume that long
118 		 * mode only uses well known segments or kernel.
119 		 */
120 		return (!user_mode(regs) || user_64bit_mode(regs));
121 #endif
122 	case 0x60:
123 		/* 0x64 thru 0x67 are valid prefixes in all modes. */
124 		return (instr_lo & 0xC) == 0x4;
125 	case 0xF0:
126 		/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
127 		return !instr_lo || (instr_lo>>1) == 1;
128 	case 0x00:
129 		/* Prefetch instruction is 0x0F0D or 0x0F18 */
130 		if (probe_kernel_address(instr, opcode))
131 			return 0;
132 
133 		*prefetch = (instr_lo == 0xF) &&
134 			(opcode == 0x0D || opcode == 0x18);
135 		return 0;
136 	default:
137 		return 0;
138 	}
139 }
140 
141 static int
142 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
143 {
144 	unsigned char *max_instr;
145 	unsigned char *instr;
146 	int prefetch = 0;
147 
148 	/*
149 	 * If it was a exec (instruction fetch) fault on NX page, then
150 	 * do not ignore the fault:
151 	 */
152 	if (error_code & PF_INSTR)
153 		return 0;
154 
155 	instr = (void *)convert_ip_to_linear(current, regs);
156 	max_instr = instr + 15;
157 
158 	if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
159 		return 0;
160 
161 	while (instr < max_instr) {
162 		unsigned char opcode;
163 
164 		if (probe_kernel_address(instr, opcode))
165 			break;
166 
167 		instr++;
168 
169 		if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
170 			break;
171 	}
172 	return prefetch;
173 }
174 
175 /*
176  * A protection key fault means that the PKRU value did not allow
177  * access to some PTE.  Userspace can figure out what PKRU was
178  * from the XSAVE state, and this function fills out a field in
179  * siginfo so userspace can discover which protection key was set
180  * on the PTE.
181  *
182  * If we get here, we know that the hardware signaled a PF_PK
183  * fault and that there was a VMA once we got in the fault
184  * handler.  It does *not* guarantee that the VMA we find here
185  * was the one that we faulted on.
186  *
187  * 1. T1   : mprotect_key(foo, PAGE_SIZE, pkey=4);
188  * 2. T1   : set PKRU to deny access to pkey=4, touches page
189  * 3. T1   : faults...
190  * 4.    T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
191  * 5. T1   : enters fault handler, takes mmap_sem, etc...
192  * 6. T1   : reaches here, sees vma_pkey(vma)=5, when we really
193  *	     faulted on a pte with its pkey=4.
194  */
195 static void fill_sig_info_pkey(int si_code, siginfo_t *info,
196 		struct vm_area_struct *vma)
197 {
198 	/* This is effectively an #ifdef */
199 	if (!boot_cpu_has(X86_FEATURE_OSPKE))
200 		return;
201 
202 	/* Fault not from Protection Keys: nothing to do */
203 	if (si_code != SEGV_PKUERR)
204 		return;
205 	/*
206 	 * force_sig_info_fault() is called from a number of
207 	 * contexts, some of which have a VMA and some of which
208 	 * do not.  The PF_PK handing happens after we have a
209 	 * valid VMA, so we should never reach this without a
210 	 * valid VMA.
211 	 */
212 	if (!vma) {
213 		WARN_ONCE(1, "PKU fault with no VMA passed in");
214 		info->si_pkey = 0;
215 		return;
216 	}
217 	/*
218 	 * si_pkey should be thought of as a strong hint, but not
219 	 * absolutely guranteed to be 100% accurate because of
220 	 * the race explained above.
221 	 */
222 	info->si_pkey = vma_pkey(vma);
223 }
224 
225 static void
226 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
227 		     struct task_struct *tsk, struct vm_area_struct *vma,
228 		     int fault)
229 {
230 	unsigned lsb = 0;
231 	siginfo_t info;
232 
233 	info.si_signo	= si_signo;
234 	info.si_errno	= 0;
235 	info.si_code	= si_code;
236 	info.si_addr	= (void __user *)address;
237 	if (fault & VM_FAULT_HWPOISON_LARGE)
238 		lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
239 	if (fault & VM_FAULT_HWPOISON)
240 		lsb = PAGE_SHIFT;
241 	info.si_addr_lsb = lsb;
242 
243 	fill_sig_info_pkey(si_code, &info, vma);
244 
245 	force_sig_info(si_signo, &info, tsk);
246 }
247 
248 DEFINE_SPINLOCK(pgd_lock);
249 LIST_HEAD(pgd_list);
250 
251 #ifdef CONFIG_X86_32
252 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
253 {
254 	unsigned index = pgd_index(address);
255 	pgd_t *pgd_k;
256 	p4d_t *p4d, *p4d_k;
257 	pud_t *pud, *pud_k;
258 	pmd_t *pmd, *pmd_k;
259 
260 	pgd += index;
261 	pgd_k = init_mm.pgd + index;
262 
263 	if (!pgd_present(*pgd_k))
264 		return NULL;
265 
266 	/*
267 	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
268 	 * and redundant with the set_pmd() on non-PAE. As would
269 	 * set_p4d/set_pud.
270 	 */
271 	p4d = p4d_offset(pgd, address);
272 	p4d_k = p4d_offset(pgd_k, address);
273 	if (!p4d_present(*p4d_k))
274 		return NULL;
275 
276 	pud = pud_offset(p4d, address);
277 	pud_k = pud_offset(p4d_k, address);
278 	if (!pud_present(*pud_k))
279 		return NULL;
280 
281 	pmd = pmd_offset(pud, address);
282 	pmd_k = pmd_offset(pud_k, address);
283 	if (!pmd_present(*pmd_k))
284 		return NULL;
285 
286 	if (!pmd_present(*pmd))
287 		set_pmd(pmd, *pmd_k);
288 	else
289 		BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
290 
291 	return pmd_k;
292 }
293 
294 void vmalloc_sync_all(void)
295 {
296 	unsigned long address;
297 
298 	if (SHARED_KERNEL_PMD)
299 		return;
300 
301 	for (address = VMALLOC_START & PMD_MASK;
302 	     address >= TASK_SIZE_MAX && address < FIXADDR_TOP;
303 	     address += PMD_SIZE) {
304 		struct page *page;
305 
306 		spin_lock(&pgd_lock);
307 		list_for_each_entry(page, &pgd_list, lru) {
308 			spinlock_t *pgt_lock;
309 			pmd_t *ret;
310 
311 			/* the pgt_lock only for Xen */
312 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
313 
314 			spin_lock(pgt_lock);
315 			ret = vmalloc_sync_one(page_address(page), address);
316 			spin_unlock(pgt_lock);
317 
318 			if (!ret)
319 				break;
320 		}
321 		spin_unlock(&pgd_lock);
322 	}
323 }
324 
325 /*
326  * 32-bit:
327  *
328  *   Handle a fault on the vmalloc or module mapping area
329  */
330 static noinline int vmalloc_fault(unsigned long address)
331 {
332 	unsigned long pgd_paddr;
333 	pmd_t *pmd_k;
334 	pte_t *pte_k;
335 
336 	/* Make sure we are in vmalloc area: */
337 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
338 		return -1;
339 
340 	WARN_ON_ONCE(in_nmi());
341 
342 	/*
343 	 * Synchronize this task's top level page-table
344 	 * with the 'reference' page table.
345 	 *
346 	 * Do _not_ use "current" here. We might be inside
347 	 * an interrupt in the middle of a task switch..
348 	 */
349 	pgd_paddr = read_cr3_pa();
350 	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
351 	if (!pmd_k)
352 		return -1;
353 
354 	if (pmd_huge(*pmd_k))
355 		return 0;
356 
357 	pte_k = pte_offset_kernel(pmd_k, address);
358 	if (!pte_present(*pte_k))
359 		return -1;
360 
361 	return 0;
362 }
363 NOKPROBE_SYMBOL(vmalloc_fault);
364 
365 /*
366  * Did it hit the DOS screen memory VA from vm86 mode?
367  */
368 static inline void
369 check_v8086_mode(struct pt_regs *regs, unsigned long address,
370 		 struct task_struct *tsk)
371 {
372 #ifdef CONFIG_VM86
373 	unsigned long bit;
374 
375 	if (!v8086_mode(regs) || !tsk->thread.vm86)
376 		return;
377 
378 	bit = (address - 0xA0000) >> PAGE_SHIFT;
379 	if (bit < 32)
380 		tsk->thread.vm86->screen_bitmap |= 1 << bit;
381 #endif
382 }
383 
384 static bool low_pfn(unsigned long pfn)
385 {
386 	return pfn < max_low_pfn;
387 }
388 
389 static void dump_pagetable(unsigned long address)
390 {
391 	pgd_t *base = __va(read_cr3_pa());
392 	pgd_t *pgd = &base[pgd_index(address)];
393 	p4d_t *p4d;
394 	pud_t *pud;
395 	pmd_t *pmd;
396 	pte_t *pte;
397 
398 #ifdef CONFIG_X86_PAE
399 	printk("*pdpt = %016Lx ", pgd_val(*pgd));
400 	if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
401 		goto out;
402 #endif
403 	p4d = p4d_offset(pgd, address);
404 	pud = pud_offset(p4d, address);
405 	pmd = pmd_offset(pud, address);
406 	printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
407 
408 	/*
409 	 * We must not directly access the pte in the highpte
410 	 * case if the page table is located in highmem.
411 	 * And let's rather not kmap-atomic the pte, just in case
412 	 * it's allocated already:
413 	 */
414 	if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
415 		goto out;
416 
417 	pte = pte_offset_kernel(pmd, address);
418 	printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
419 out:
420 	printk("\n");
421 }
422 
423 #else /* CONFIG_X86_64: */
424 
425 void vmalloc_sync_all(void)
426 {
427 	sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
428 }
429 
430 /*
431  * 64-bit:
432  *
433  *   Handle a fault on the vmalloc area
434  */
435 static noinline int vmalloc_fault(unsigned long address)
436 {
437 	pgd_t *pgd, *pgd_ref;
438 	p4d_t *p4d, *p4d_ref;
439 	pud_t *pud, *pud_ref;
440 	pmd_t *pmd, *pmd_ref;
441 	pte_t *pte, *pte_ref;
442 
443 	/* Make sure we are in vmalloc area: */
444 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
445 		return -1;
446 
447 	WARN_ON_ONCE(in_nmi());
448 
449 	/*
450 	 * Copy kernel mappings over when needed. This can also
451 	 * happen within a race in page table update. In the later
452 	 * case just flush:
453 	 */
454 	pgd = (pgd_t *)__va(read_cr3_pa()) + pgd_index(address);
455 	pgd_ref = pgd_offset_k(address);
456 	if (pgd_none(*pgd_ref))
457 		return -1;
458 
459 	if (pgd_none(*pgd)) {
460 		set_pgd(pgd, *pgd_ref);
461 		arch_flush_lazy_mmu_mode();
462 	} else if (CONFIG_PGTABLE_LEVELS > 4) {
463 		/*
464 		 * With folded p4d, pgd_none() is always false, so the pgd may
465 		 * point to an empty page table entry and pgd_page_vaddr()
466 		 * will return garbage.
467 		 *
468 		 * We will do the correct sanity check on the p4d level.
469 		 */
470 		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
471 	}
472 
473 	/* With 4-level paging, copying happens on the p4d level. */
474 	p4d = p4d_offset(pgd, address);
475 	p4d_ref = p4d_offset(pgd_ref, address);
476 	if (p4d_none(*p4d_ref))
477 		return -1;
478 
479 	if (p4d_none(*p4d)) {
480 		set_p4d(p4d, *p4d_ref);
481 		arch_flush_lazy_mmu_mode();
482 	} else {
483 		BUG_ON(p4d_pfn(*p4d) != p4d_pfn(*p4d_ref));
484 	}
485 
486 	/*
487 	 * Below here mismatches are bugs because these lower tables
488 	 * are shared:
489 	 */
490 
491 	pud = pud_offset(p4d, address);
492 	pud_ref = pud_offset(p4d_ref, address);
493 	if (pud_none(*pud_ref))
494 		return -1;
495 
496 	if (pud_none(*pud) || pud_pfn(*pud) != pud_pfn(*pud_ref))
497 		BUG();
498 
499 	if (pud_huge(*pud))
500 		return 0;
501 
502 	pmd = pmd_offset(pud, address);
503 	pmd_ref = pmd_offset(pud_ref, address);
504 	if (pmd_none(*pmd_ref))
505 		return -1;
506 
507 	if (pmd_none(*pmd) || pmd_pfn(*pmd) != pmd_pfn(*pmd_ref))
508 		BUG();
509 
510 	if (pmd_huge(*pmd))
511 		return 0;
512 
513 	pte_ref = pte_offset_kernel(pmd_ref, address);
514 	if (!pte_present(*pte_ref))
515 		return -1;
516 
517 	pte = pte_offset_kernel(pmd, address);
518 
519 	/*
520 	 * Don't use pte_page here, because the mappings can point
521 	 * outside mem_map, and the NUMA hash lookup cannot handle
522 	 * that:
523 	 */
524 	if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
525 		BUG();
526 
527 	return 0;
528 }
529 NOKPROBE_SYMBOL(vmalloc_fault);
530 
531 #ifdef CONFIG_CPU_SUP_AMD
532 static const char errata93_warning[] =
533 KERN_ERR
534 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
535 "******* Working around it, but it may cause SEGVs or burn power.\n"
536 "******* Please consider a BIOS update.\n"
537 "******* Disabling USB legacy in the BIOS may also help.\n";
538 #endif
539 
540 /*
541  * No vm86 mode in 64-bit mode:
542  */
543 static inline void
544 check_v8086_mode(struct pt_regs *regs, unsigned long address,
545 		 struct task_struct *tsk)
546 {
547 }
548 
549 static int bad_address(void *p)
550 {
551 	unsigned long dummy;
552 
553 	return probe_kernel_address((unsigned long *)p, dummy);
554 }
555 
556 static void dump_pagetable(unsigned long address)
557 {
558 	pgd_t *base = __va(read_cr3_pa());
559 	pgd_t *pgd = base + pgd_index(address);
560 	p4d_t *p4d;
561 	pud_t *pud;
562 	pmd_t *pmd;
563 	pte_t *pte;
564 
565 	if (bad_address(pgd))
566 		goto bad;
567 
568 	printk("PGD %lx ", pgd_val(*pgd));
569 
570 	if (!pgd_present(*pgd))
571 		goto out;
572 
573 	p4d = p4d_offset(pgd, address);
574 	if (bad_address(p4d))
575 		goto bad;
576 
577 	printk("P4D %lx ", p4d_val(*p4d));
578 	if (!p4d_present(*p4d) || p4d_large(*p4d))
579 		goto out;
580 
581 	pud = pud_offset(p4d, address);
582 	if (bad_address(pud))
583 		goto bad;
584 
585 	printk("PUD %lx ", pud_val(*pud));
586 	if (!pud_present(*pud) || pud_large(*pud))
587 		goto out;
588 
589 	pmd = pmd_offset(pud, address);
590 	if (bad_address(pmd))
591 		goto bad;
592 
593 	printk("PMD %lx ", pmd_val(*pmd));
594 	if (!pmd_present(*pmd) || pmd_large(*pmd))
595 		goto out;
596 
597 	pte = pte_offset_kernel(pmd, address);
598 	if (bad_address(pte))
599 		goto bad;
600 
601 	printk("PTE %lx", pte_val(*pte));
602 out:
603 	printk("\n");
604 	return;
605 bad:
606 	printk("BAD\n");
607 }
608 
609 #endif /* CONFIG_X86_64 */
610 
611 /*
612  * Workaround for K8 erratum #93 & buggy BIOS.
613  *
614  * BIOS SMM functions are required to use a specific workaround
615  * to avoid corruption of the 64bit RIP register on C stepping K8.
616  *
617  * A lot of BIOS that didn't get tested properly miss this.
618  *
619  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
620  * Try to work around it here.
621  *
622  * Note we only handle faults in kernel here.
623  * Does nothing on 32-bit.
624  */
625 static int is_errata93(struct pt_regs *regs, unsigned long address)
626 {
627 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
628 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
629 	    || boot_cpu_data.x86 != 0xf)
630 		return 0;
631 
632 	if (address != regs->ip)
633 		return 0;
634 
635 	if ((address >> 32) != 0)
636 		return 0;
637 
638 	address |= 0xffffffffUL << 32;
639 	if ((address >= (u64)_stext && address <= (u64)_etext) ||
640 	    (address >= MODULES_VADDR && address <= MODULES_END)) {
641 		printk_once(errata93_warning);
642 		regs->ip = address;
643 		return 1;
644 	}
645 #endif
646 	return 0;
647 }
648 
649 /*
650  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
651  * to illegal addresses >4GB.
652  *
653  * We catch this in the page fault handler because these addresses
654  * are not reachable. Just detect this case and return.  Any code
655  * segment in LDT is compatibility mode.
656  */
657 static int is_errata100(struct pt_regs *regs, unsigned long address)
658 {
659 #ifdef CONFIG_X86_64
660 	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
661 		return 1;
662 #endif
663 	return 0;
664 }
665 
666 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
667 {
668 #ifdef CONFIG_X86_F00F_BUG
669 	unsigned long nr;
670 
671 	/*
672 	 * Pentium F0 0F C7 C8 bug workaround:
673 	 */
674 	if (boot_cpu_has_bug(X86_BUG_F00F)) {
675 		nr = (address - idt_descr.address) >> 3;
676 
677 		if (nr == 6) {
678 			do_invalid_op(regs, 0);
679 			return 1;
680 		}
681 	}
682 #endif
683 	return 0;
684 }
685 
686 static const char nx_warning[] = KERN_CRIT
687 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
688 static const char smep_warning[] = KERN_CRIT
689 "unable to execute userspace code (SMEP?) (uid: %d)\n";
690 
691 static void
692 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
693 		unsigned long address)
694 {
695 	if (!oops_may_print())
696 		return;
697 
698 	if (error_code & PF_INSTR) {
699 		unsigned int level;
700 		pgd_t *pgd;
701 		pte_t *pte;
702 
703 		pgd = __va(read_cr3_pa());
704 		pgd += pgd_index(address);
705 
706 		pte = lookup_address_in_pgd(pgd, address, &level);
707 
708 		if (pte && pte_present(*pte) && !pte_exec(*pte))
709 			printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
710 		if (pte && pte_present(*pte) && pte_exec(*pte) &&
711 				(pgd_flags(*pgd) & _PAGE_USER) &&
712 				(__read_cr4() & X86_CR4_SMEP))
713 			printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
714 	}
715 
716 	printk(KERN_ALERT "BUG: unable to handle kernel ");
717 	if (address < PAGE_SIZE)
718 		printk(KERN_CONT "NULL pointer dereference");
719 	else
720 		printk(KERN_CONT "paging request");
721 
722 	printk(KERN_CONT " at %p\n", (void *) address);
723 	printk(KERN_ALERT "IP: %pS\n", (void *)regs->ip);
724 
725 	dump_pagetable(address);
726 }
727 
728 static noinline void
729 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
730 	    unsigned long address)
731 {
732 	struct task_struct *tsk;
733 	unsigned long flags;
734 	int sig;
735 
736 	flags = oops_begin();
737 	tsk = current;
738 	sig = SIGKILL;
739 
740 	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
741 	       tsk->comm, address);
742 	dump_pagetable(address);
743 
744 	tsk->thread.cr2		= address;
745 	tsk->thread.trap_nr	= X86_TRAP_PF;
746 	tsk->thread.error_code	= error_code;
747 
748 	if (__die("Bad pagetable", regs, error_code))
749 		sig = 0;
750 
751 	oops_end(flags, regs, sig);
752 }
753 
754 static noinline void
755 no_context(struct pt_regs *regs, unsigned long error_code,
756 	   unsigned long address, int signal, int si_code)
757 {
758 	struct task_struct *tsk = current;
759 	unsigned long flags;
760 	int sig;
761 	/* No context means no VMA to pass down */
762 	struct vm_area_struct *vma = NULL;
763 
764 	/* Are we prepared to handle this kernel fault? */
765 	if (fixup_exception(regs, X86_TRAP_PF)) {
766 		/*
767 		 * Any interrupt that takes a fault gets the fixup. This makes
768 		 * the below recursive fault logic only apply to a faults from
769 		 * task context.
770 		 */
771 		if (in_interrupt())
772 			return;
773 
774 		/*
775 		 * Per the above we're !in_interrupt(), aka. task context.
776 		 *
777 		 * In this case we need to make sure we're not recursively
778 		 * faulting through the emulate_vsyscall() logic.
779 		 */
780 		if (current->thread.sig_on_uaccess_err && signal) {
781 			tsk->thread.trap_nr = X86_TRAP_PF;
782 			tsk->thread.error_code = error_code | PF_USER;
783 			tsk->thread.cr2 = address;
784 
785 			/* XXX: hwpoison faults will set the wrong code. */
786 			force_sig_info_fault(signal, si_code, address,
787 					     tsk, vma, 0);
788 		}
789 
790 		/*
791 		 * Barring that, we can do the fixup and be happy.
792 		 */
793 		return;
794 	}
795 
796 #ifdef CONFIG_VMAP_STACK
797 	/*
798 	 * Stack overflow?  During boot, we can fault near the initial
799 	 * stack in the direct map, but that's not an overflow -- check
800 	 * that we're in vmalloc space to avoid this.
801 	 */
802 	if (is_vmalloc_addr((void *)address) &&
803 	    (((unsigned long)tsk->stack - 1 - address < PAGE_SIZE) ||
804 	     address - ((unsigned long)tsk->stack + THREAD_SIZE) < PAGE_SIZE)) {
805 		register void *__sp asm("rsp");
806 		unsigned long stack = this_cpu_read(orig_ist.ist[DOUBLEFAULT_STACK]) - sizeof(void *);
807 		/*
808 		 * We're likely to be running with very little stack space
809 		 * left.  It's plausible that we'd hit this condition but
810 		 * double-fault even before we get this far, in which case
811 		 * we're fine: the double-fault handler will deal with it.
812 		 *
813 		 * We don't want to make it all the way into the oops code
814 		 * and then double-fault, though, because we're likely to
815 		 * break the console driver and lose most of the stack dump.
816 		 */
817 		asm volatile ("movq %[stack], %%rsp\n\t"
818 			      "call handle_stack_overflow\n\t"
819 			      "1: jmp 1b"
820 			      : "+r" (__sp)
821 			      : "D" ("kernel stack overflow (page fault)"),
822 				"S" (regs), "d" (address),
823 				[stack] "rm" (stack));
824 		unreachable();
825 	}
826 #endif
827 
828 	/*
829 	 * 32-bit:
830 	 *
831 	 *   Valid to do another page fault here, because if this fault
832 	 *   had been triggered by is_prefetch fixup_exception would have
833 	 *   handled it.
834 	 *
835 	 * 64-bit:
836 	 *
837 	 *   Hall of shame of CPU/BIOS bugs.
838 	 */
839 	if (is_prefetch(regs, error_code, address))
840 		return;
841 
842 	if (is_errata93(regs, address))
843 		return;
844 
845 	/*
846 	 * Oops. The kernel tried to access some bad page. We'll have to
847 	 * terminate things with extreme prejudice:
848 	 */
849 	flags = oops_begin();
850 
851 	show_fault_oops(regs, error_code, address);
852 
853 	if (task_stack_end_corrupted(tsk))
854 		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
855 
856 	tsk->thread.cr2		= address;
857 	tsk->thread.trap_nr	= X86_TRAP_PF;
858 	tsk->thread.error_code	= error_code;
859 
860 	sig = SIGKILL;
861 	if (__die("Oops", regs, error_code))
862 		sig = 0;
863 
864 	/* Executive summary in case the body of the oops scrolled away */
865 	printk(KERN_DEFAULT "CR2: %016lx\n", address);
866 
867 	oops_end(flags, regs, sig);
868 }
869 
870 /*
871  * Print out info about fatal segfaults, if the show_unhandled_signals
872  * sysctl is set:
873  */
874 static inline void
875 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
876 		unsigned long address, struct task_struct *tsk)
877 {
878 	if (!unhandled_signal(tsk, SIGSEGV))
879 		return;
880 
881 	if (!printk_ratelimit())
882 		return;
883 
884 	printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
885 		task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
886 		tsk->comm, task_pid_nr(tsk), address,
887 		(void *)regs->ip, (void *)regs->sp, error_code);
888 
889 	print_vma_addr(KERN_CONT " in ", regs->ip);
890 
891 	printk(KERN_CONT "\n");
892 }
893 
894 static void
895 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
896 		       unsigned long address, struct vm_area_struct *vma,
897 		       int si_code)
898 {
899 	struct task_struct *tsk = current;
900 
901 	/* User mode accesses just cause a SIGSEGV */
902 	if (error_code & PF_USER) {
903 		/*
904 		 * It's possible to have interrupts off here:
905 		 */
906 		local_irq_enable();
907 
908 		/*
909 		 * Valid to do another page fault here because this one came
910 		 * from user space:
911 		 */
912 		if (is_prefetch(regs, error_code, address))
913 			return;
914 
915 		if (is_errata100(regs, address))
916 			return;
917 
918 #ifdef CONFIG_X86_64
919 		/*
920 		 * Instruction fetch faults in the vsyscall page might need
921 		 * emulation.
922 		 */
923 		if (unlikely((error_code & PF_INSTR) &&
924 			     ((address & ~0xfff) == VSYSCALL_ADDR))) {
925 			if (emulate_vsyscall(regs, address))
926 				return;
927 		}
928 #endif
929 
930 		/*
931 		 * To avoid leaking information about the kernel page table
932 		 * layout, pretend that user-mode accesses to kernel addresses
933 		 * are always protection faults.
934 		 */
935 		if (address >= TASK_SIZE_MAX)
936 			error_code |= PF_PROT;
937 
938 		if (likely(show_unhandled_signals))
939 			show_signal_msg(regs, error_code, address, tsk);
940 
941 		tsk->thread.cr2		= address;
942 		tsk->thread.error_code	= error_code;
943 		tsk->thread.trap_nr	= X86_TRAP_PF;
944 
945 		force_sig_info_fault(SIGSEGV, si_code, address, tsk, vma, 0);
946 
947 		return;
948 	}
949 
950 	if (is_f00f_bug(regs, address))
951 		return;
952 
953 	no_context(regs, error_code, address, SIGSEGV, si_code);
954 }
955 
956 static noinline void
957 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
958 		     unsigned long address, struct vm_area_struct *vma)
959 {
960 	__bad_area_nosemaphore(regs, error_code, address, vma, SEGV_MAPERR);
961 }
962 
963 static void
964 __bad_area(struct pt_regs *regs, unsigned long error_code,
965 	   unsigned long address,  struct vm_area_struct *vma, int si_code)
966 {
967 	struct mm_struct *mm = current->mm;
968 
969 	/*
970 	 * Something tried to access memory that isn't in our memory map..
971 	 * Fix it, but check if it's kernel or user first..
972 	 */
973 	up_read(&mm->mmap_sem);
974 
975 	__bad_area_nosemaphore(regs, error_code, address, vma, si_code);
976 }
977 
978 static noinline void
979 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
980 {
981 	__bad_area(regs, error_code, address, NULL, SEGV_MAPERR);
982 }
983 
984 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
985 		struct vm_area_struct *vma)
986 {
987 	/* This code is always called on the current mm */
988 	bool foreign = false;
989 
990 	if (!boot_cpu_has(X86_FEATURE_OSPKE))
991 		return false;
992 	if (error_code & PF_PK)
993 		return true;
994 	/* this checks permission keys on the VMA: */
995 	if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
996 				(error_code & PF_INSTR), foreign))
997 		return true;
998 	return false;
999 }
1000 
1001 static noinline void
1002 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
1003 		      unsigned long address, struct vm_area_struct *vma)
1004 {
1005 	/*
1006 	 * This OSPKE check is not strictly necessary at runtime.
1007 	 * But, doing it this way allows compiler optimizations
1008 	 * if pkeys are compiled out.
1009 	 */
1010 	if (bad_area_access_from_pkeys(error_code, vma))
1011 		__bad_area(regs, error_code, address, vma, SEGV_PKUERR);
1012 	else
1013 		__bad_area(regs, error_code, address, vma, SEGV_ACCERR);
1014 }
1015 
1016 static void
1017 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
1018 	  struct vm_area_struct *vma, unsigned int fault)
1019 {
1020 	struct task_struct *tsk = current;
1021 	int code = BUS_ADRERR;
1022 
1023 	/* Kernel mode? Handle exceptions or die: */
1024 	if (!(error_code & PF_USER)) {
1025 		no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1026 		return;
1027 	}
1028 
1029 	/* User-space => ok to do another page fault: */
1030 	if (is_prefetch(regs, error_code, address))
1031 		return;
1032 
1033 	tsk->thread.cr2		= address;
1034 	tsk->thread.error_code	= error_code;
1035 	tsk->thread.trap_nr	= X86_TRAP_PF;
1036 
1037 #ifdef CONFIG_MEMORY_FAILURE
1038 	if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
1039 		printk(KERN_ERR
1040 	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
1041 			tsk->comm, tsk->pid, address);
1042 		code = BUS_MCEERR_AR;
1043 	}
1044 #endif
1045 	force_sig_info_fault(SIGBUS, code, address, tsk, vma, fault);
1046 }
1047 
1048 static noinline void
1049 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
1050 	       unsigned long address, struct vm_area_struct *vma,
1051 	       unsigned int fault)
1052 {
1053 	if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
1054 		no_context(regs, error_code, address, 0, 0);
1055 		return;
1056 	}
1057 
1058 	if (fault & VM_FAULT_OOM) {
1059 		/* Kernel mode? Handle exceptions or die: */
1060 		if (!(error_code & PF_USER)) {
1061 			no_context(regs, error_code, address,
1062 				   SIGSEGV, SEGV_MAPERR);
1063 			return;
1064 		}
1065 
1066 		/*
1067 		 * We ran out of memory, call the OOM killer, and return the
1068 		 * userspace (which will retry the fault, or kill us if we got
1069 		 * oom-killed):
1070 		 */
1071 		pagefault_out_of_memory();
1072 	} else {
1073 		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1074 			     VM_FAULT_HWPOISON_LARGE))
1075 			do_sigbus(regs, error_code, address, vma, fault);
1076 		else if (fault & VM_FAULT_SIGSEGV)
1077 			bad_area_nosemaphore(regs, error_code, address, vma);
1078 		else
1079 			BUG();
1080 	}
1081 }
1082 
1083 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
1084 {
1085 	if ((error_code & PF_WRITE) && !pte_write(*pte))
1086 		return 0;
1087 
1088 	if ((error_code & PF_INSTR) && !pte_exec(*pte))
1089 		return 0;
1090 	/*
1091 	 * Note: We do not do lazy flushing on protection key
1092 	 * changes, so no spurious fault will ever set PF_PK.
1093 	 */
1094 	if ((error_code & PF_PK))
1095 		return 1;
1096 
1097 	return 1;
1098 }
1099 
1100 /*
1101  * Handle a spurious fault caused by a stale TLB entry.
1102  *
1103  * This allows us to lazily refresh the TLB when increasing the
1104  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
1105  * eagerly is very expensive since that implies doing a full
1106  * cross-processor TLB flush, even if no stale TLB entries exist
1107  * on other processors.
1108  *
1109  * Spurious faults may only occur if the TLB contains an entry with
1110  * fewer permission than the page table entry.  Non-present (P = 0)
1111  * and reserved bit (R = 1) faults are never spurious.
1112  *
1113  * There are no security implications to leaving a stale TLB when
1114  * increasing the permissions on a page.
1115  *
1116  * Returns non-zero if a spurious fault was handled, zero otherwise.
1117  *
1118  * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
1119  * (Optional Invalidation).
1120  */
1121 static noinline int
1122 spurious_fault(unsigned long error_code, unsigned long address)
1123 {
1124 	pgd_t *pgd;
1125 	p4d_t *p4d;
1126 	pud_t *pud;
1127 	pmd_t *pmd;
1128 	pte_t *pte;
1129 	int ret;
1130 
1131 	/*
1132 	 * Only writes to RO or instruction fetches from NX may cause
1133 	 * spurious faults.
1134 	 *
1135 	 * These could be from user or supervisor accesses but the TLB
1136 	 * is only lazily flushed after a kernel mapping protection
1137 	 * change, so user accesses are not expected to cause spurious
1138 	 * faults.
1139 	 */
1140 	if (error_code != (PF_WRITE | PF_PROT)
1141 	    && error_code != (PF_INSTR | PF_PROT))
1142 		return 0;
1143 
1144 	pgd = init_mm.pgd + pgd_index(address);
1145 	if (!pgd_present(*pgd))
1146 		return 0;
1147 
1148 	p4d = p4d_offset(pgd, address);
1149 	if (!p4d_present(*p4d))
1150 		return 0;
1151 
1152 	if (p4d_large(*p4d))
1153 		return spurious_fault_check(error_code, (pte_t *) p4d);
1154 
1155 	pud = pud_offset(p4d, address);
1156 	if (!pud_present(*pud))
1157 		return 0;
1158 
1159 	if (pud_large(*pud))
1160 		return spurious_fault_check(error_code, (pte_t *) pud);
1161 
1162 	pmd = pmd_offset(pud, address);
1163 	if (!pmd_present(*pmd))
1164 		return 0;
1165 
1166 	if (pmd_large(*pmd))
1167 		return spurious_fault_check(error_code, (pte_t *) pmd);
1168 
1169 	pte = pte_offset_kernel(pmd, address);
1170 	if (!pte_present(*pte))
1171 		return 0;
1172 
1173 	ret = spurious_fault_check(error_code, pte);
1174 	if (!ret)
1175 		return 0;
1176 
1177 	/*
1178 	 * Make sure we have permissions in PMD.
1179 	 * If not, then there's a bug in the page tables:
1180 	 */
1181 	ret = spurious_fault_check(error_code, (pte_t *) pmd);
1182 	WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1183 
1184 	return ret;
1185 }
1186 NOKPROBE_SYMBOL(spurious_fault);
1187 
1188 int show_unhandled_signals = 1;
1189 
1190 static inline int
1191 access_error(unsigned long error_code, struct vm_area_struct *vma)
1192 {
1193 	/* This is only called for the current mm, so: */
1194 	bool foreign = false;
1195 
1196 	/*
1197 	 * Read or write was blocked by protection keys.  This is
1198 	 * always an unconditional error and can never result in
1199 	 * a follow-up action to resolve the fault, like a COW.
1200 	 */
1201 	if (error_code & PF_PK)
1202 		return 1;
1203 
1204 	/*
1205 	 * Make sure to check the VMA so that we do not perform
1206 	 * faults just to hit a PF_PK as soon as we fill in a
1207 	 * page.
1208 	 */
1209 	if (!arch_vma_access_permitted(vma, (error_code & PF_WRITE),
1210 				(error_code & PF_INSTR), foreign))
1211 		return 1;
1212 
1213 	if (error_code & PF_WRITE) {
1214 		/* write, present and write, not present: */
1215 		if (unlikely(!(vma->vm_flags & VM_WRITE)))
1216 			return 1;
1217 		return 0;
1218 	}
1219 
1220 	/* read, present: */
1221 	if (unlikely(error_code & PF_PROT))
1222 		return 1;
1223 
1224 	/* read, not present: */
1225 	if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1226 		return 1;
1227 
1228 	return 0;
1229 }
1230 
1231 static int fault_in_kernel_space(unsigned long address)
1232 {
1233 	return address >= TASK_SIZE_MAX;
1234 }
1235 
1236 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1237 {
1238 	if (!IS_ENABLED(CONFIG_X86_SMAP))
1239 		return false;
1240 
1241 	if (!static_cpu_has(X86_FEATURE_SMAP))
1242 		return false;
1243 
1244 	if (error_code & PF_USER)
1245 		return false;
1246 
1247 	if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1248 		return false;
1249 
1250 	return true;
1251 }
1252 
1253 /*
1254  * This routine handles page faults.  It determines the address,
1255  * and the problem, and then passes it off to one of the appropriate
1256  * routines.
1257  *
1258  * This function must have noinline because both callers
1259  * {,trace_}do_page_fault() have notrace on. Having this an actual function
1260  * guarantees there's a function trace entry.
1261  */
1262 static noinline void
1263 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1264 		unsigned long address)
1265 {
1266 	struct vm_area_struct *vma;
1267 	struct task_struct *tsk;
1268 	struct mm_struct *mm;
1269 	int fault, major = 0;
1270 	unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1271 
1272 	tsk = current;
1273 	mm = tsk->mm;
1274 
1275 	/*
1276 	 * Detect and handle instructions that would cause a page fault for
1277 	 * both a tracked kernel page and a userspace page.
1278 	 */
1279 	if (kmemcheck_active(regs))
1280 		kmemcheck_hide(regs);
1281 	prefetchw(&mm->mmap_sem);
1282 
1283 	if (unlikely(kmmio_fault(regs, address)))
1284 		return;
1285 
1286 	/*
1287 	 * We fault-in kernel-space virtual memory on-demand. The
1288 	 * 'reference' page table is init_mm.pgd.
1289 	 *
1290 	 * NOTE! We MUST NOT take any locks for this case. We may
1291 	 * be in an interrupt or a critical region, and should
1292 	 * only copy the information from the master page table,
1293 	 * nothing more.
1294 	 *
1295 	 * This verifies that the fault happens in kernel space
1296 	 * (error_code & 4) == 0, and that the fault was not a
1297 	 * protection error (error_code & 9) == 0.
1298 	 */
1299 	if (unlikely(fault_in_kernel_space(address))) {
1300 		if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1301 			if (vmalloc_fault(address) >= 0)
1302 				return;
1303 
1304 			if (kmemcheck_fault(regs, address, error_code))
1305 				return;
1306 		}
1307 
1308 		/* Can handle a stale RO->RW TLB: */
1309 		if (spurious_fault(error_code, address))
1310 			return;
1311 
1312 		/* kprobes don't want to hook the spurious faults: */
1313 		if (kprobes_fault(regs))
1314 			return;
1315 		/*
1316 		 * Don't take the mm semaphore here. If we fixup a prefetch
1317 		 * fault we could otherwise deadlock:
1318 		 */
1319 		bad_area_nosemaphore(regs, error_code, address, NULL);
1320 
1321 		return;
1322 	}
1323 
1324 	/* kprobes don't want to hook the spurious faults: */
1325 	if (unlikely(kprobes_fault(regs)))
1326 		return;
1327 
1328 	if (unlikely(error_code & PF_RSVD))
1329 		pgtable_bad(regs, error_code, address);
1330 
1331 	if (unlikely(smap_violation(error_code, regs))) {
1332 		bad_area_nosemaphore(regs, error_code, address, NULL);
1333 		return;
1334 	}
1335 
1336 	/*
1337 	 * If we're in an interrupt, have no user context or are running
1338 	 * in a region with pagefaults disabled then we must not take the fault
1339 	 */
1340 	if (unlikely(faulthandler_disabled() || !mm)) {
1341 		bad_area_nosemaphore(regs, error_code, address, NULL);
1342 		return;
1343 	}
1344 
1345 	/*
1346 	 * It's safe to allow irq's after cr2 has been saved and the
1347 	 * vmalloc fault has been handled.
1348 	 *
1349 	 * User-mode registers count as a user access even for any
1350 	 * potential system fault or CPU buglet:
1351 	 */
1352 	if (user_mode(regs)) {
1353 		local_irq_enable();
1354 		error_code |= PF_USER;
1355 		flags |= FAULT_FLAG_USER;
1356 	} else {
1357 		if (regs->flags & X86_EFLAGS_IF)
1358 			local_irq_enable();
1359 	}
1360 
1361 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1362 
1363 	if (error_code & PF_WRITE)
1364 		flags |= FAULT_FLAG_WRITE;
1365 	if (error_code & PF_INSTR)
1366 		flags |= FAULT_FLAG_INSTRUCTION;
1367 
1368 	/*
1369 	 * When running in the kernel we expect faults to occur only to
1370 	 * addresses in user space.  All other faults represent errors in
1371 	 * the kernel and should generate an OOPS.  Unfortunately, in the
1372 	 * case of an erroneous fault occurring in a code path which already
1373 	 * holds mmap_sem we will deadlock attempting to validate the fault
1374 	 * against the address space.  Luckily the kernel only validly
1375 	 * references user space from well defined areas of code, which are
1376 	 * listed in the exceptions table.
1377 	 *
1378 	 * As the vast majority of faults will be valid we will only perform
1379 	 * the source reference check when there is a possibility of a
1380 	 * deadlock. Attempt to lock the address space, if we cannot we then
1381 	 * validate the source. If this is invalid we can skip the address
1382 	 * space check, thus avoiding the deadlock:
1383 	 */
1384 	if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1385 		if ((error_code & PF_USER) == 0 &&
1386 		    !search_exception_tables(regs->ip)) {
1387 			bad_area_nosemaphore(regs, error_code, address, NULL);
1388 			return;
1389 		}
1390 retry:
1391 		down_read(&mm->mmap_sem);
1392 	} else {
1393 		/*
1394 		 * The above down_read_trylock() might have succeeded in
1395 		 * which case we'll have missed the might_sleep() from
1396 		 * down_read():
1397 		 */
1398 		might_sleep();
1399 	}
1400 
1401 	vma = find_vma(mm, address);
1402 	if (unlikely(!vma)) {
1403 		bad_area(regs, error_code, address);
1404 		return;
1405 	}
1406 	if (likely(vma->vm_start <= address))
1407 		goto good_area;
1408 	if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1409 		bad_area(regs, error_code, address);
1410 		return;
1411 	}
1412 	if (error_code & PF_USER) {
1413 		/*
1414 		 * Accessing the stack below %sp is always a bug.
1415 		 * The large cushion allows instructions like enter
1416 		 * and pusha to work. ("enter $65535, $31" pushes
1417 		 * 32 pointers and then decrements %sp by 65535.)
1418 		 */
1419 		if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1420 			bad_area(regs, error_code, address);
1421 			return;
1422 		}
1423 	}
1424 	if (unlikely(expand_stack(vma, address))) {
1425 		bad_area(regs, error_code, address);
1426 		return;
1427 	}
1428 
1429 	/*
1430 	 * Ok, we have a good vm_area for this memory access, so
1431 	 * we can handle it..
1432 	 */
1433 good_area:
1434 	if (unlikely(access_error(error_code, vma))) {
1435 		bad_area_access_error(regs, error_code, address, vma);
1436 		return;
1437 	}
1438 
1439 	/*
1440 	 * If for any reason at all we couldn't handle the fault,
1441 	 * make sure we exit gracefully rather than endlessly redo
1442 	 * the fault.  Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1443 	 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1444 	 */
1445 	fault = handle_mm_fault(vma, address, flags);
1446 	major |= fault & VM_FAULT_MAJOR;
1447 
1448 	/*
1449 	 * If we need to retry the mmap_sem has already been released,
1450 	 * and if there is a fatal signal pending there is no guarantee
1451 	 * that we made any progress. Handle this case first.
1452 	 */
1453 	if (unlikely(fault & VM_FAULT_RETRY)) {
1454 		/* Retry at most once */
1455 		if (flags & FAULT_FLAG_ALLOW_RETRY) {
1456 			flags &= ~FAULT_FLAG_ALLOW_RETRY;
1457 			flags |= FAULT_FLAG_TRIED;
1458 			if (!fatal_signal_pending(tsk))
1459 				goto retry;
1460 		}
1461 
1462 		/* User mode? Just return to handle the fatal exception */
1463 		if (flags & FAULT_FLAG_USER)
1464 			return;
1465 
1466 		/* Not returning to user mode? Handle exceptions or die: */
1467 		no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1468 		return;
1469 	}
1470 
1471 	up_read(&mm->mmap_sem);
1472 	if (unlikely(fault & VM_FAULT_ERROR)) {
1473 		mm_fault_error(regs, error_code, address, vma, fault);
1474 		return;
1475 	}
1476 
1477 	/*
1478 	 * Major/minor page fault accounting. If any of the events
1479 	 * returned VM_FAULT_MAJOR, we account it as a major fault.
1480 	 */
1481 	if (major) {
1482 		tsk->maj_flt++;
1483 		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1484 	} else {
1485 		tsk->min_flt++;
1486 		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1487 	}
1488 
1489 	check_v8086_mode(regs, address, tsk);
1490 }
1491 NOKPROBE_SYMBOL(__do_page_fault);
1492 
1493 dotraplinkage void notrace
1494 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1495 {
1496 	unsigned long address = read_cr2(); /* Get the faulting address */
1497 	enum ctx_state prev_state;
1498 
1499 	/*
1500 	 * We must have this function tagged with __kprobes, notrace and call
1501 	 * read_cr2() before calling anything else. To avoid calling any kind
1502 	 * of tracing machinery before we've observed the CR2 value.
1503 	 *
1504 	 * exception_{enter,exit}() contain all sorts of tracepoints.
1505 	 */
1506 
1507 	prev_state = exception_enter();
1508 	__do_page_fault(regs, error_code, address);
1509 	exception_exit(prev_state);
1510 }
1511 NOKPROBE_SYMBOL(do_page_fault);
1512 
1513 #ifdef CONFIG_TRACING
1514 static nokprobe_inline void
1515 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1516 			 unsigned long error_code)
1517 {
1518 	if (user_mode(regs))
1519 		trace_page_fault_user(address, regs, error_code);
1520 	else
1521 		trace_page_fault_kernel(address, regs, error_code);
1522 }
1523 
1524 dotraplinkage void notrace
1525 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1526 {
1527 	/*
1528 	 * The exception_enter and tracepoint processing could
1529 	 * trigger another page faults (user space callchain
1530 	 * reading) and destroy the original cr2 value, so read
1531 	 * the faulting address now.
1532 	 */
1533 	unsigned long address = read_cr2();
1534 	enum ctx_state prev_state;
1535 
1536 	prev_state = exception_enter();
1537 	trace_page_fault_entries(address, regs, error_code);
1538 	__do_page_fault(regs, error_code, address);
1539 	exception_exit(prev_state);
1540 }
1541 NOKPROBE_SYMBOL(trace_do_page_fault);
1542 #endif /* CONFIG_TRACING */
1543