xref: /openbmc/linux/arch/x86/mm/fault.c (revision 2596e07a)
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/traps.h>			/* dotraplinkage, ...		*/
19 #include <asm/pgalloc.h>		/* pgd_*(), ...			*/
20 #include <asm/kmemcheck.h>		/* kmemcheck_*(), ...		*/
21 #include <asm/fixmap.h>			/* VSYSCALL_ADDR		*/
22 #include <asm/vsyscall.h>		/* emulate_vsyscall		*/
23 #include <asm/vm86.h>			/* struct vm86			*/
24 
25 #define CREATE_TRACE_POINTS
26 #include <asm/trace/exceptions.h>
27 
28 /*
29  * Page fault error code bits:
30  *
31  *   bit 0 ==	 0: no page found	1: protection fault
32  *   bit 1 ==	 0: read access		1: write access
33  *   bit 2 ==	 0: kernel-mode access	1: user-mode access
34  *   bit 3 ==				1: use of reserved bit detected
35  *   bit 4 ==				1: fault was an instruction fetch
36  */
37 enum x86_pf_error_code {
38 
39 	PF_PROT		=		1 << 0,
40 	PF_WRITE	=		1 << 1,
41 	PF_USER		=		1 << 2,
42 	PF_RSVD		=		1 << 3,
43 	PF_INSTR	=		1 << 4,
44 };
45 
46 /*
47  * Returns 0 if mmiotrace is disabled, or if the fault is not
48  * handled by mmiotrace:
49  */
50 static nokprobe_inline int
51 kmmio_fault(struct pt_regs *regs, unsigned long addr)
52 {
53 	if (unlikely(is_kmmio_active()))
54 		if (kmmio_handler(regs, addr) == 1)
55 			return -1;
56 	return 0;
57 }
58 
59 static nokprobe_inline int kprobes_fault(struct pt_regs *regs)
60 {
61 	int ret = 0;
62 
63 	/* kprobe_running() needs smp_processor_id() */
64 	if (kprobes_built_in() && !user_mode(regs)) {
65 		preempt_disable();
66 		if (kprobe_running() && kprobe_fault_handler(regs, 14))
67 			ret = 1;
68 		preempt_enable();
69 	}
70 
71 	return ret;
72 }
73 
74 /*
75  * Prefetch quirks:
76  *
77  * 32-bit mode:
78  *
79  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
80  *   Check that here and ignore it.
81  *
82  * 64-bit mode:
83  *
84  *   Sometimes the CPU reports invalid exceptions on prefetch.
85  *   Check that here and ignore it.
86  *
87  * Opcode checker based on code by Richard Brunner.
88  */
89 static inline int
90 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
91 		      unsigned char opcode, int *prefetch)
92 {
93 	unsigned char instr_hi = opcode & 0xf0;
94 	unsigned char instr_lo = opcode & 0x0f;
95 
96 	switch (instr_hi) {
97 	case 0x20:
98 	case 0x30:
99 		/*
100 		 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
101 		 * In X86_64 long mode, the CPU will signal invalid
102 		 * opcode if some of these prefixes are present so
103 		 * X86_64 will never get here anyway
104 		 */
105 		return ((instr_lo & 7) == 0x6);
106 #ifdef CONFIG_X86_64
107 	case 0x40:
108 		/*
109 		 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
110 		 * Need to figure out under what instruction mode the
111 		 * instruction was issued. Could check the LDT for lm,
112 		 * but for now it's good enough to assume that long
113 		 * mode only uses well known segments or kernel.
114 		 */
115 		return (!user_mode(regs) || user_64bit_mode(regs));
116 #endif
117 	case 0x60:
118 		/* 0x64 thru 0x67 are valid prefixes in all modes. */
119 		return (instr_lo & 0xC) == 0x4;
120 	case 0xF0:
121 		/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
122 		return !instr_lo || (instr_lo>>1) == 1;
123 	case 0x00:
124 		/* Prefetch instruction is 0x0F0D or 0x0F18 */
125 		if (probe_kernel_address(instr, opcode))
126 			return 0;
127 
128 		*prefetch = (instr_lo == 0xF) &&
129 			(opcode == 0x0D || opcode == 0x18);
130 		return 0;
131 	default:
132 		return 0;
133 	}
134 }
135 
136 static int
137 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
138 {
139 	unsigned char *max_instr;
140 	unsigned char *instr;
141 	int prefetch = 0;
142 
143 	/*
144 	 * If it was a exec (instruction fetch) fault on NX page, then
145 	 * do not ignore the fault:
146 	 */
147 	if (error_code & PF_INSTR)
148 		return 0;
149 
150 	instr = (void *)convert_ip_to_linear(current, regs);
151 	max_instr = instr + 15;
152 
153 	if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE_MAX)
154 		return 0;
155 
156 	while (instr < max_instr) {
157 		unsigned char opcode;
158 
159 		if (probe_kernel_address(instr, opcode))
160 			break;
161 
162 		instr++;
163 
164 		if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
165 			break;
166 	}
167 	return prefetch;
168 }
169 
170 static void
171 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
172 		     struct task_struct *tsk, int fault)
173 {
174 	unsigned lsb = 0;
175 	siginfo_t info;
176 
177 	info.si_signo	= si_signo;
178 	info.si_errno	= 0;
179 	info.si_code	= si_code;
180 	info.si_addr	= (void __user *)address;
181 	if (fault & VM_FAULT_HWPOISON_LARGE)
182 		lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
183 	if (fault & VM_FAULT_HWPOISON)
184 		lsb = PAGE_SHIFT;
185 	info.si_addr_lsb = lsb;
186 
187 	force_sig_info(si_signo, &info, tsk);
188 }
189 
190 DEFINE_SPINLOCK(pgd_lock);
191 LIST_HEAD(pgd_list);
192 
193 #ifdef CONFIG_X86_32
194 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
195 {
196 	unsigned index = pgd_index(address);
197 	pgd_t *pgd_k;
198 	pud_t *pud, *pud_k;
199 	pmd_t *pmd, *pmd_k;
200 
201 	pgd += index;
202 	pgd_k = init_mm.pgd + index;
203 
204 	if (!pgd_present(*pgd_k))
205 		return NULL;
206 
207 	/*
208 	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
209 	 * and redundant with the set_pmd() on non-PAE. As would
210 	 * set_pud.
211 	 */
212 	pud = pud_offset(pgd, address);
213 	pud_k = pud_offset(pgd_k, address);
214 	if (!pud_present(*pud_k))
215 		return NULL;
216 
217 	pmd = pmd_offset(pud, address);
218 	pmd_k = pmd_offset(pud_k, address);
219 	if (!pmd_present(*pmd_k))
220 		return NULL;
221 
222 	if (!pmd_present(*pmd))
223 		set_pmd(pmd, *pmd_k);
224 	else
225 		BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
226 
227 	return pmd_k;
228 }
229 
230 void vmalloc_sync_all(void)
231 {
232 	unsigned long address;
233 
234 	if (SHARED_KERNEL_PMD)
235 		return;
236 
237 	for (address = VMALLOC_START & PMD_MASK;
238 	     address >= TASK_SIZE && address < FIXADDR_TOP;
239 	     address += PMD_SIZE) {
240 		struct page *page;
241 
242 		spin_lock(&pgd_lock);
243 		list_for_each_entry(page, &pgd_list, lru) {
244 			spinlock_t *pgt_lock;
245 			pmd_t *ret;
246 
247 			/* the pgt_lock only for Xen */
248 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
249 
250 			spin_lock(pgt_lock);
251 			ret = vmalloc_sync_one(page_address(page), address);
252 			spin_unlock(pgt_lock);
253 
254 			if (!ret)
255 				break;
256 		}
257 		spin_unlock(&pgd_lock);
258 	}
259 }
260 
261 /*
262  * 32-bit:
263  *
264  *   Handle a fault on the vmalloc or module mapping area
265  */
266 static noinline int vmalloc_fault(unsigned long address)
267 {
268 	unsigned long pgd_paddr;
269 	pmd_t *pmd_k;
270 	pte_t *pte_k;
271 
272 	/* Make sure we are in vmalloc area: */
273 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
274 		return -1;
275 
276 	WARN_ON_ONCE(in_nmi());
277 
278 	/*
279 	 * Synchronize this task's top level page-table
280 	 * with the 'reference' page table.
281 	 *
282 	 * Do _not_ use "current" here. We might be inside
283 	 * an interrupt in the middle of a task switch..
284 	 */
285 	pgd_paddr = read_cr3();
286 	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
287 	if (!pmd_k)
288 		return -1;
289 
290 	if (pmd_huge(*pmd_k))
291 		return 0;
292 
293 	pte_k = pte_offset_kernel(pmd_k, address);
294 	if (!pte_present(*pte_k))
295 		return -1;
296 
297 	return 0;
298 }
299 NOKPROBE_SYMBOL(vmalloc_fault);
300 
301 /*
302  * Did it hit the DOS screen memory VA from vm86 mode?
303  */
304 static inline void
305 check_v8086_mode(struct pt_regs *regs, unsigned long address,
306 		 struct task_struct *tsk)
307 {
308 #ifdef CONFIG_VM86
309 	unsigned long bit;
310 
311 	if (!v8086_mode(regs) || !tsk->thread.vm86)
312 		return;
313 
314 	bit = (address - 0xA0000) >> PAGE_SHIFT;
315 	if (bit < 32)
316 		tsk->thread.vm86->screen_bitmap |= 1 << bit;
317 #endif
318 }
319 
320 static bool low_pfn(unsigned long pfn)
321 {
322 	return pfn < max_low_pfn;
323 }
324 
325 static void dump_pagetable(unsigned long address)
326 {
327 	pgd_t *base = __va(read_cr3());
328 	pgd_t *pgd = &base[pgd_index(address)];
329 	pmd_t *pmd;
330 	pte_t *pte;
331 
332 #ifdef CONFIG_X86_PAE
333 	printk("*pdpt = %016Lx ", pgd_val(*pgd));
334 	if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
335 		goto out;
336 #endif
337 	pmd = pmd_offset(pud_offset(pgd, address), address);
338 	printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
339 
340 	/*
341 	 * We must not directly access the pte in the highpte
342 	 * case if the page table is located in highmem.
343 	 * And let's rather not kmap-atomic the pte, just in case
344 	 * it's allocated already:
345 	 */
346 	if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
347 		goto out;
348 
349 	pte = pte_offset_kernel(pmd, address);
350 	printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
351 out:
352 	printk("\n");
353 }
354 
355 #else /* CONFIG_X86_64: */
356 
357 void vmalloc_sync_all(void)
358 {
359 	sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END, 0);
360 }
361 
362 /*
363  * 64-bit:
364  *
365  *   Handle a fault on the vmalloc area
366  */
367 static noinline int vmalloc_fault(unsigned long address)
368 {
369 	pgd_t *pgd, *pgd_ref;
370 	pud_t *pud, *pud_ref;
371 	pmd_t *pmd, *pmd_ref;
372 	pte_t *pte, *pte_ref;
373 
374 	/* Make sure we are in vmalloc area: */
375 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
376 		return -1;
377 
378 	WARN_ON_ONCE(in_nmi());
379 
380 	/*
381 	 * Copy kernel mappings over when needed. This can also
382 	 * happen within a race in page table update. In the later
383 	 * case just flush:
384 	 */
385 	pgd = pgd_offset(current->active_mm, address);
386 	pgd_ref = pgd_offset_k(address);
387 	if (pgd_none(*pgd_ref))
388 		return -1;
389 
390 	if (pgd_none(*pgd)) {
391 		set_pgd(pgd, *pgd_ref);
392 		arch_flush_lazy_mmu_mode();
393 	} else {
394 		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
395 	}
396 
397 	/*
398 	 * Below here mismatches are bugs because these lower tables
399 	 * are shared:
400 	 */
401 
402 	pud = pud_offset(pgd, address);
403 	pud_ref = pud_offset(pgd_ref, address);
404 	if (pud_none(*pud_ref))
405 		return -1;
406 
407 	if (pud_none(*pud) || pud_pfn(*pud) != pud_pfn(*pud_ref))
408 		BUG();
409 
410 	if (pud_huge(*pud))
411 		return 0;
412 
413 	pmd = pmd_offset(pud, address);
414 	pmd_ref = pmd_offset(pud_ref, address);
415 	if (pmd_none(*pmd_ref))
416 		return -1;
417 
418 	if (pmd_none(*pmd) || pmd_pfn(*pmd) != pmd_pfn(*pmd_ref))
419 		BUG();
420 
421 	if (pmd_huge(*pmd))
422 		return 0;
423 
424 	pte_ref = pte_offset_kernel(pmd_ref, address);
425 	if (!pte_present(*pte_ref))
426 		return -1;
427 
428 	pte = pte_offset_kernel(pmd, address);
429 
430 	/*
431 	 * Don't use pte_page here, because the mappings can point
432 	 * outside mem_map, and the NUMA hash lookup cannot handle
433 	 * that:
434 	 */
435 	if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
436 		BUG();
437 
438 	return 0;
439 }
440 NOKPROBE_SYMBOL(vmalloc_fault);
441 
442 #ifdef CONFIG_CPU_SUP_AMD
443 static const char errata93_warning[] =
444 KERN_ERR
445 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
446 "******* Working around it, but it may cause SEGVs or burn power.\n"
447 "******* Please consider a BIOS update.\n"
448 "******* Disabling USB legacy in the BIOS may also help.\n";
449 #endif
450 
451 /*
452  * No vm86 mode in 64-bit mode:
453  */
454 static inline void
455 check_v8086_mode(struct pt_regs *regs, unsigned long address,
456 		 struct task_struct *tsk)
457 {
458 }
459 
460 static int bad_address(void *p)
461 {
462 	unsigned long dummy;
463 
464 	return probe_kernel_address((unsigned long *)p, dummy);
465 }
466 
467 static void dump_pagetable(unsigned long address)
468 {
469 	pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
470 	pgd_t *pgd = base + pgd_index(address);
471 	pud_t *pud;
472 	pmd_t *pmd;
473 	pte_t *pte;
474 
475 	if (bad_address(pgd))
476 		goto bad;
477 
478 	printk("PGD %lx ", pgd_val(*pgd));
479 
480 	if (!pgd_present(*pgd))
481 		goto out;
482 
483 	pud = pud_offset(pgd, address);
484 	if (bad_address(pud))
485 		goto bad;
486 
487 	printk("PUD %lx ", pud_val(*pud));
488 	if (!pud_present(*pud) || pud_large(*pud))
489 		goto out;
490 
491 	pmd = pmd_offset(pud, address);
492 	if (bad_address(pmd))
493 		goto bad;
494 
495 	printk("PMD %lx ", pmd_val(*pmd));
496 	if (!pmd_present(*pmd) || pmd_large(*pmd))
497 		goto out;
498 
499 	pte = pte_offset_kernel(pmd, address);
500 	if (bad_address(pte))
501 		goto bad;
502 
503 	printk("PTE %lx", pte_val(*pte));
504 out:
505 	printk("\n");
506 	return;
507 bad:
508 	printk("BAD\n");
509 }
510 
511 #endif /* CONFIG_X86_64 */
512 
513 /*
514  * Workaround for K8 erratum #93 & buggy BIOS.
515  *
516  * BIOS SMM functions are required to use a specific workaround
517  * to avoid corruption of the 64bit RIP register on C stepping K8.
518  *
519  * A lot of BIOS that didn't get tested properly miss this.
520  *
521  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
522  * Try to work around it here.
523  *
524  * Note we only handle faults in kernel here.
525  * Does nothing on 32-bit.
526  */
527 static int is_errata93(struct pt_regs *regs, unsigned long address)
528 {
529 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
530 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
531 	    || boot_cpu_data.x86 != 0xf)
532 		return 0;
533 
534 	if (address != regs->ip)
535 		return 0;
536 
537 	if ((address >> 32) != 0)
538 		return 0;
539 
540 	address |= 0xffffffffUL << 32;
541 	if ((address >= (u64)_stext && address <= (u64)_etext) ||
542 	    (address >= MODULES_VADDR && address <= MODULES_END)) {
543 		printk_once(errata93_warning);
544 		regs->ip = address;
545 		return 1;
546 	}
547 #endif
548 	return 0;
549 }
550 
551 /*
552  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
553  * to illegal addresses >4GB.
554  *
555  * We catch this in the page fault handler because these addresses
556  * are not reachable. Just detect this case and return.  Any code
557  * segment in LDT is compatibility mode.
558  */
559 static int is_errata100(struct pt_regs *regs, unsigned long address)
560 {
561 #ifdef CONFIG_X86_64
562 	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
563 		return 1;
564 #endif
565 	return 0;
566 }
567 
568 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
569 {
570 #ifdef CONFIG_X86_F00F_BUG
571 	unsigned long nr;
572 
573 	/*
574 	 * Pentium F0 0F C7 C8 bug workaround:
575 	 */
576 	if (boot_cpu_has_bug(X86_BUG_F00F)) {
577 		nr = (address - idt_descr.address) >> 3;
578 
579 		if (nr == 6) {
580 			do_invalid_op(regs, 0);
581 			return 1;
582 		}
583 	}
584 #endif
585 	return 0;
586 }
587 
588 static const char nx_warning[] = KERN_CRIT
589 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
590 static const char smep_warning[] = KERN_CRIT
591 "unable to execute userspace code (SMEP?) (uid: %d)\n";
592 
593 static void
594 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
595 		unsigned long address)
596 {
597 	if (!oops_may_print())
598 		return;
599 
600 	if (error_code & PF_INSTR) {
601 		unsigned int level;
602 		pgd_t *pgd;
603 		pte_t *pte;
604 
605 		pgd = __va(read_cr3() & PHYSICAL_PAGE_MASK);
606 		pgd += pgd_index(address);
607 
608 		pte = lookup_address_in_pgd(pgd, address, &level);
609 
610 		if (pte && pte_present(*pte) && !pte_exec(*pte))
611 			printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
612 		if (pte && pte_present(*pte) && pte_exec(*pte) &&
613 				(pgd_flags(*pgd) & _PAGE_USER) &&
614 				(__read_cr4() & X86_CR4_SMEP))
615 			printk(smep_warning, from_kuid(&init_user_ns, current_uid()));
616 	}
617 
618 	printk(KERN_ALERT "BUG: unable to handle kernel ");
619 	if (address < PAGE_SIZE)
620 		printk(KERN_CONT "NULL pointer dereference");
621 	else
622 		printk(KERN_CONT "paging request");
623 
624 	printk(KERN_CONT " at %p\n", (void *) address);
625 	printk(KERN_ALERT "IP:");
626 	printk_address(regs->ip);
627 
628 	dump_pagetable(address);
629 }
630 
631 static noinline void
632 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
633 	    unsigned long address)
634 {
635 	struct task_struct *tsk;
636 	unsigned long flags;
637 	int sig;
638 
639 	flags = oops_begin();
640 	tsk = current;
641 	sig = SIGKILL;
642 
643 	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
644 	       tsk->comm, address);
645 	dump_pagetable(address);
646 
647 	tsk->thread.cr2		= address;
648 	tsk->thread.trap_nr	= X86_TRAP_PF;
649 	tsk->thread.error_code	= error_code;
650 
651 	if (__die("Bad pagetable", regs, error_code))
652 		sig = 0;
653 
654 	oops_end(flags, regs, sig);
655 }
656 
657 static noinline void
658 no_context(struct pt_regs *regs, unsigned long error_code,
659 	   unsigned long address, int signal, int si_code)
660 {
661 	struct task_struct *tsk = current;
662 	unsigned long flags;
663 	int sig;
664 
665 	/* Are we prepared to handle this kernel fault? */
666 	if (fixup_exception(regs)) {
667 		/*
668 		 * Any interrupt that takes a fault gets the fixup. This makes
669 		 * the below recursive fault logic only apply to a faults from
670 		 * task context.
671 		 */
672 		if (in_interrupt())
673 			return;
674 
675 		/*
676 		 * Per the above we're !in_interrupt(), aka. task context.
677 		 *
678 		 * In this case we need to make sure we're not recursively
679 		 * faulting through the emulate_vsyscall() logic.
680 		 */
681 		if (current_thread_info()->sig_on_uaccess_error && signal) {
682 			tsk->thread.trap_nr = X86_TRAP_PF;
683 			tsk->thread.error_code = error_code | PF_USER;
684 			tsk->thread.cr2 = address;
685 
686 			/* XXX: hwpoison faults will set the wrong code. */
687 			force_sig_info_fault(signal, si_code, address, tsk, 0);
688 		}
689 
690 		/*
691 		 * Barring that, we can do the fixup and be happy.
692 		 */
693 		return;
694 	}
695 
696 	/*
697 	 * 32-bit:
698 	 *
699 	 *   Valid to do another page fault here, because if this fault
700 	 *   had been triggered by is_prefetch fixup_exception would have
701 	 *   handled it.
702 	 *
703 	 * 64-bit:
704 	 *
705 	 *   Hall of shame of CPU/BIOS bugs.
706 	 */
707 	if (is_prefetch(regs, error_code, address))
708 		return;
709 
710 	if (is_errata93(regs, address))
711 		return;
712 
713 	/*
714 	 * Oops. The kernel tried to access some bad page. We'll have to
715 	 * terminate things with extreme prejudice:
716 	 */
717 	flags = oops_begin();
718 
719 	show_fault_oops(regs, error_code, address);
720 
721 	if (task_stack_end_corrupted(tsk))
722 		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
723 
724 	tsk->thread.cr2		= address;
725 	tsk->thread.trap_nr	= X86_TRAP_PF;
726 	tsk->thread.error_code	= error_code;
727 
728 	sig = SIGKILL;
729 	if (__die("Oops", regs, error_code))
730 		sig = 0;
731 
732 	/* Executive summary in case the body of the oops scrolled away */
733 	printk(KERN_DEFAULT "CR2: %016lx\n", address);
734 
735 	oops_end(flags, regs, sig);
736 }
737 
738 /*
739  * Print out info about fatal segfaults, if the show_unhandled_signals
740  * sysctl is set:
741  */
742 static inline void
743 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
744 		unsigned long address, struct task_struct *tsk)
745 {
746 	if (!unhandled_signal(tsk, SIGSEGV))
747 		return;
748 
749 	if (!printk_ratelimit())
750 		return;
751 
752 	printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
753 		task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
754 		tsk->comm, task_pid_nr(tsk), address,
755 		(void *)regs->ip, (void *)regs->sp, error_code);
756 
757 	print_vma_addr(KERN_CONT " in ", regs->ip);
758 
759 	printk(KERN_CONT "\n");
760 }
761 
762 static void
763 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
764 		       unsigned long address, int si_code)
765 {
766 	struct task_struct *tsk = current;
767 
768 	/* User mode accesses just cause a SIGSEGV */
769 	if (error_code & PF_USER) {
770 		/*
771 		 * It's possible to have interrupts off here:
772 		 */
773 		local_irq_enable();
774 
775 		/*
776 		 * Valid to do another page fault here because this one came
777 		 * from user space:
778 		 */
779 		if (is_prefetch(regs, error_code, address))
780 			return;
781 
782 		if (is_errata100(regs, address))
783 			return;
784 
785 #ifdef CONFIG_X86_64
786 		/*
787 		 * Instruction fetch faults in the vsyscall page might need
788 		 * emulation.
789 		 */
790 		if (unlikely((error_code & PF_INSTR) &&
791 			     ((address & ~0xfff) == VSYSCALL_ADDR))) {
792 			if (emulate_vsyscall(regs, address))
793 				return;
794 		}
795 #endif
796 		/* Kernel addresses are always protection faults: */
797 		if (address >= TASK_SIZE)
798 			error_code |= PF_PROT;
799 
800 		if (likely(show_unhandled_signals))
801 			show_signal_msg(regs, error_code, address, tsk);
802 
803 		tsk->thread.cr2		= address;
804 		tsk->thread.error_code	= error_code;
805 		tsk->thread.trap_nr	= X86_TRAP_PF;
806 
807 		force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
808 
809 		return;
810 	}
811 
812 	if (is_f00f_bug(regs, address))
813 		return;
814 
815 	no_context(regs, error_code, address, SIGSEGV, si_code);
816 }
817 
818 static noinline void
819 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
820 		     unsigned long address)
821 {
822 	__bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
823 }
824 
825 static void
826 __bad_area(struct pt_regs *regs, unsigned long error_code,
827 	   unsigned long address, int si_code)
828 {
829 	struct mm_struct *mm = current->mm;
830 
831 	/*
832 	 * Something tried to access memory that isn't in our memory map..
833 	 * Fix it, but check if it's kernel or user first..
834 	 */
835 	up_read(&mm->mmap_sem);
836 
837 	__bad_area_nosemaphore(regs, error_code, address, si_code);
838 }
839 
840 static noinline void
841 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
842 {
843 	__bad_area(regs, error_code, address, SEGV_MAPERR);
844 }
845 
846 static noinline void
847 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
848 		      unsigned long address)
849 {
850 	__bad_area(regs, error_code, address, SEGV_ACCERR);
851 }
852 
853 static void
854 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
855 	  unsigned int fault)
856 {
857 	struct task_struct *tsk = current;
858 	int code = BUS_ADRERR;
859 
860 	/* Kernel mode? Handle exceptions or die: */
861 	if (!(error_code & PF_USER)) {
862 		no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
863 		return;
864 	}
865 
866 	/* User-space => ok to do another page fault: */
867 	if (is_prefetch(regs, error_code, address))
868 		return;
869 
870 	tsk->thread.cr2		= address;
871 	tsk->thread.error_code	= error_code;
872 	tsk->thread.trap_nr	= X86_TRAP_PF;
873 
874 #ifdef CONFIG_MEMORY_FAILURE
875 	if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
876 		printk(KERN_ERR
877 	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
878 			tsk->comm, tsk->pid, address);
879 		code = BUS_MCEERR_AR;
880 	}
881 #endif
882 	force_sig_info_fault(SIGBUS, code, address, tsk, fault);
883 }
884 
885 static noinline void
886 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
887 	       unsigned long address, unsigned int fault)
888 {
889 	if (fatal_signal_pending(current) && !(error_code & PF_USER)) {
890 		no_context(regs, error_code, address, 0, 0);
891 		return;
892 	}
893 
894 	if (fault & VM_FAULT_OOM) {
895 		/* Kernel mode? Handle exceptions or die: */
896 		if (!(error_code & PF_USER)) {
897 			no_context(regs, error_code, address,
898 				   SIGSEGV, SEGV_MAPERR);
899 			return;
900 		}
901 
902 		/*
903 		 * We ran out of memory, call the OOM killer, and return the
904 		 * userspace (which will retry the fault, or kill us if we got
905 		 * oom-killed):
906 		 */
907 		pagefault_out_of_memory();
908 	} else {
909 		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
910 			     VM_FAULT_HWPOISON_LARGE))
911 			do_sigbus(regs, error_code, address, fault);
912 		else if (fault & VM_FAULT_SIGSEGV)
913 			bad_area_nosemaphore(regs, error_code, address);
914 		else
915 			BUG();
916 	}
917 }
918 
919 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
920 {
921 	if ((error_code & PF_WRITE) && !pte_write(*pte))
922 		return 0;
923 
924 	if ((error_code & PF_INSTR) && !pte_exec(*pte))
925 		return 0;
926 
927 	return 1;
928 }
929 
930 /*
931  * Handle a spurious fault caused by a stale TLB entry.
932  *
933  * This allows us to lazily refresh the TLB when increasing the
934  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
935  * eagerly is very expensive since that implies doing a full
936  * cross-processor TLB flush, even if no stale TLB entries exist
937  * on other processors.
938  *
939  * Spurious faults may only occur if the TLB contains an entry with
940  * fewer permission than the page table entry.  Non-present (P = 0)
941  * and reserved bit (R = 1) faults are never spurious.
942  *
943  * There are no security implications to leaving a stale TLB when
944  * increasing the permissions on a page.
945  *
946  * Returns non-zero if a spurious fault was handled, zero otherwise.
947  *
948  * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
949  * (Optional Invalidation).
950  */
951 static noinline int
952 spurious_fault(unsigned long error_code, unsigned long address)
953 {
954 	pgd_t *pgd;
955 	pud_t *pud;
956 	pmd_t *pmd;
957 	pte_t *pte;
958 	int ret;
959 
960 	/*
961 	 * Only writes to RO or instruction fetches from NX may cause
962 	 * spurious faults.
963 	 *
964 	 * These could be from user or supervisor accesses but the TLB
965 	 * is only lazily flushed after a kernel mapping protection
966 	 * change, so user accesses are not expected to cause spurious
967 	 * faults.
968 	 */
969 	if (error_code != (PF_WRITE | PF_PROT)
970 	    && error_code != (PF_INSTR | PF_PROT))
971 		return 0;
972 
973 	pgd = init_mm.pgd + pgd_index(address);
974 	if (!pgd_present(*pgd))
975 		return 0;
976 
977 	pud = pud_offset(pgd, address);
978 	if (!pud_present(*pud))
979 		return 0;
980 
981 	if (pud_large(*pud))
982 		return spurious_fault_check(error_code, (pte_t *) pud);
983 
984 	pmd = pmd_offset(pud, address);
985 	if (!pmd_present(*pmd))
986 		return 0;
987 
988 	if (pmd_large(*pmd))
989 		return spurious_fault_check(error_code, (pte_t *) pmd);
990 
991 	pte = pte_offset_kernel(pmd, address);
992 	if (!pte_present(*pte))
993 		return 0;
994 
995 	ret = spurious_fault_check(error_code, pte);
996 	if (!ret)
997 		return 0;
998 
999 	/*
1000 	 * Make sure we have permissions in PMD.
1001 	 * If not, then there's a bug in the page tables:
1002 	 */
1003 	ret = spurious_fault_check(error_code, (pte_t *) pmd);
1004 	WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1005 
1006 	return ret;
1007 }
1008 NOKPROBE_SYMBOL(spurious_fault);
1009 
1010 int show_unhandled_signals = 1;
1011 
1012 static inline int
1013 access_error(unsigned long error_code, struct vm_area_struct *vma)
1014 {
1015 	if (error_code & PF_WRITE) {
1016 		/* write, present and write, not present: */
1017 		if (unlikely(!(vma->vm_flags & VM_WRITE)))
1018 			return 1;
1019 		return 0;
1020 	}
1021 
1022 	/* read, present: */
1023 	if (unlikely(error_code & PF_PROT))
1024 		return 1;
1025 
1026 	/* read, not present: */
1027 	if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
1028 		return 1;
1029 
1030 	return 0;
1031 }
1032 
1033 static int fault_in_kernel_space(unsigned long address)
1034 {
1035 	return address >= TASK_SIZE_MAX;
1036 }
1037 
1038 static inline bool smap_violation(int error_code, struct pt_regs *regs)
1039 {
1040 	if (!IS_ENABLED(CONFIG_X86_SMAP))
1041 		return false;
1042 
1043 	if (!static_cpu_has(X86_FEATURE_SMAP))
1044 		return false;
1045 
1046 	if (error_code & PF_USER)
1047 		return false;
1048 
1049 	if (!user_mode(regs) && (regs->flags & X86_EFLAGS_AC))
1050 		return false;
1051 
1052 	return true;
1053 }
1054 
1055 /*
1056  * This routine handles page faults.  It determines the address,
1057  * and the problem, and then passes it off to one of the appropriate
1058  * routines.
1059  *
1060  * This function must have noinline because both callers
1061  * {,trace_}do_page_fault() have notrace on. Having this an actual function
1062  * guarantees there's a function trace entry.
1063  */
1064 static noinline void
1065 __do_page_fault(struct pt_regs *regs, unsigned long error_code,
1066 		unsigned long address)
1067 {
1068 	struct vm_area_struct *vma;
1069 	struct task_struct *tsk;
1070 	struct mm_struct *mm;
1071 	int fault, major = 0;
1072 	unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
1073 
1074 	tsk = current;
1075 	mm = tsk->mm;
1076 
1077 	/*
1078 	 * Detect and handle instructions that would cause a page fault for
1079 	 * both a tracked kernel page and a userspace page.
1080 	 */
1081 	if (kmemcheck_active(regs))
1082 		kmemcheck_hide(regs);
1083 	prefetchw(&mm->mmap_sem);
1084 
1085 	if (unlikely(kmmio_fault(regs, address)))
1086 		return;
1087 
1088 	/*
1089 	 * We fault-in kernel-space virtual memory on-demand. The
1090 	 * 'reference' page table is init_mm.pgd.
1091 	 *
1092 	 * NOTE! We MUST NOT take any locks for this case. We may
1093 	 * be in an interrupt or a critical region, and should
1094 	 * only copy the information from the master page table,
1095 	 * nothing more.
1096 	 *
1097 	 * This verifies that the fault happens in kernel space
1098 	 * (error_code & 4) == 0, and that the fault was not a
1099 	 * protection error (error_code & 9) == 0.
1100 	 */
1101 	if (unlikely(fault_in_kernel_space(address))) {
1102 		if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1103 			if (vmalloc_fault(address) >= 0)
1104 				return;
1105 
1106 			if (kmemcheck_fault(regs, address, error_code))
1107 				return;
1108 		}
1109 
1110 		/* Can handle a stale RO->RW TLB: */
1111 		if (spurious_fault(error_code, address))
1112 			return;
1113 
1114 		/* kprobes don't want to hook the spurious faults: */
1115 		if (kprobes_fault(regs))
1116 			return;
1117 		/*
1118 		 * Don't take the mm semaphore here. If we fixup a prefetch
1119 		 * fault we could otherwise deadlock:
1120 		 */
1121 		bad_area_nosemaphore(regs, error_code, address);
1122 
1123 		return;
1124 	}
1125 
1126 	/* kprobes don't want to hook the spurious faults: */
1127 	if (unlikely(kprobes_fault(regs)))
1128 		return;
1129 
1130 	if (unlikely(error_code & PF_RSVD))
1131 		pgtable_bad(regs, error_code, address);
1132 
1133 	if (unlikely(smap_violation(error_code, regs))) {
1134 		bad_area_nosemaphore(regs, error_code, address);
1135 		return;
1136 	}
1137 
1138 	/*
1139 	 * If we're in an interrupt, have no user context or are running
1140 	 * in a region with pagefaults disabled then we must not take the fault
1141 	 */
1142 	if (unlikely(faulthandler_disabled() || !mm)) {
1143 		bad_area_nosemaphore(regs, error_code, address);
1144 		return;
1145 	}
1146 
1147 	/*
1148 	 * It's safe to allow irq's after cr2 has been saved and the
1149 	 * vmalloc fault has been handled.
1150 	 *
1151 	 * User-mode registers count as a user access even for any
1152 	 * potential system fault or CPU buglet:
1153 	 */
1154 	if (user_mode(regs)) {
1155 		local_irq_enable();
1156 		error_code |= PF_USER;
1157 		flags |= FAULT_FLAG_USER;
1158 	} else {
1159 		if (regs->flags & X86_EFLAGS_IF)
1160 			local_irq_enable();
1161 	}
1162 
1163 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1164 
1165 	if (error_code & PF_WRITE)
1166 		flags |= FAULT_FLAG_WRITE;
1167 
1168 	/*
1169 	 * When running in the kernel we expect faults to occur only to
1170 	 * addresses in user space.  All other faults represent errors in
1171 	 * the kernel and should generate an OOPS.  Unfortunately, in the
1172 	 * case of an erroneous fault occurring in a code path which already
1173 	 * holds mmap_sem we will deadlock attempting to validate the fault
1174 	 * against the address space.  Luckily the kernel only validly
1175 	 * references user space from well defined areas of code, which are
1176 	 * listed in the exceptions table.
1177 	 *
1178 	 * As the vast majority of faults will be valid we will only perform
1179 	 * the source reference check when there is a possibility of a
1180 	 * deadlock. Attempt to lock the address space, if we cannot we then
1181 	 * validate the source. If this is invalid we can skip the address
1182 	 * space check, thus avoiding the deadlock:
1183 	 */
1184 	if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1185 		if ((error_code & PF_USER) == 0 &&
1186 		    !search_exception_tables(regs->ip)) {
1187 			bad_area_nosemaphore(regs, error_code, address);
1188 			return;
1189 		}
1190 retry:
1191 		down_read(&mm->mmap_sem);
1192 	} else {
1193 		/*
1194 		 * The above down_read_trylock() might have succeeded in
1195 		 * which case we'll have missed the might_sleep() from
1196 		 * down_read():
1197 		 */
1198 		might_sleep();
1199 	}
1200 
1201 	vma = find_vma(mm, address);
1202 	if (unlikely(!vma)) {
1203 		bad_area(regs, error_code, address);
1204 		return;
1205 	}
1206 	if (likely(vma->vm_start <= address))
1207 		goto good_area;
1208 	if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1209 		bad_area(regs, error_code, address);
1210 		return;
1211 	}
1212 	if (error_code & PF_USER) {
1213 		/*
1214 		 * Accessing the stack below %sp is always a bug.
1215 		 * The large cushion allows instructions like enter
1216 		 * and pusha to work. ("enter $65535, $31" pushes
1217 		 * 32 pointers and then decrements %sp by 65535.)
1218 		 */
1219 		if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1220 			bad_area(regs, error_code, address);
1221 			return;
1222 		}
1223 	}
1224 	if (unlikely(expand_stack(vma, address))) {
1225 		bad_area(regs, error_code, address);
1226 		return;
1227 	}
1228 
1229 	/*
1230 	 * Ok, we have a good vm_area for this memory access, so
1231 	 * we can handle it..
1232 	 */
1233 good_area:
1234 	if (unlikely(access_error(error_code, vma))) {
1235 		bad_area_access_error(regs, error_code, address);
1236 		return;
1237 	}
1238 
1239 	/*
1240 	 * If for any reason at all we couldn't handle the fault,
1241 	 * make sure we exit gracefully rather than endlessly redo
1242 	 * the fault.  Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1243 	 * we get VM_FAULT_RETRY back, the mmap_sem has been unlocked.
1244 	 */
1245 	fault = handle_mm_fault(mm, vma, address, flags);
1246 	major |= fault & VM_FAULT_MAJOR;
1247 
1248 	/*
1249 	 * If we need to retry the mmap_sem has already been released,
1250 	 * and if there is a fatal signal pending there is no guarantee
1251 	 * that we made any progress. Handle this case first.
1252 	 */
1253 	if (unlikely(fault & VM_FAULT_RETRY)) {
1254 		/* Retry at most once */
1255 		if (flags & FAULT_FLAG_ALLOW_RETRY) {
1256 			flags &= ~FAULT_FLAG_ALLOW_RETRY;
1257 			flags |= FAULT_FLAG_TRIED;
1258 			if (!fatal_signal_pending(tsk))
1259 				goto retry;
1260 		}
1261 
1262 		/* User mode? Just return to handle the fatal exception */
1263 		if (flags & FAULT_FLAG_USER)
1264 			return;
1265 
1266 		/* Not returning to user mode? Handle exceptions or die: */
1267 		no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
1268 		return;
1269 	}
1270 
1271 	up_read(&mm->mmap_sem);
1272 	if (unlikely(fault & VM_FAULT_ERROR)) {
1273 		mm_fault_error(regs, error_code, address, fault);
1274 		return;
1275 	}
1276 
1277 	/*
1278 	 * Major/minor page fault accounting. If any of the events
1279 	 * returned VM_FAULT_MAJOR, we account it as a major fault.
1280 	 */
1281 	if (major) {
1282 		tsk->maj_flt++;
1283 		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
1284 	} else {
1285 		tsk->min_flt++;
1286 		perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
1287 	}
1288 
1289 	check_v8086_mode(regs, address, tsk);
1290 }
1291 NOKPROBE_SYMBOL(__do_page_fault);
1292 
1293 dotraplinkage void notrace
1294 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1295 {
1296 	unsigned long address = read_cr2(); /* Get the faulting address */
1297 	enum ctx_state prev_state;
1298 
1299 	/*
1300 	 * We must have this function tagged with __kprobes, notrace and call
1301 	 * read_cr2() before calling anything else. To avoid calling any kind
1302 	 * of tracing machinery before we've observed the CR2 value.
1303 	 *
1304 	 * exception_{enter,exit}() contain all sorts of tracepoints.
1305 	 */
1306 
1307 	prev_state = exception_enter();
1308 	__do_page_fault(regs, error_code, address);
1309 	exception_exit(prev_state);
1310 }
1311 NOKPROBE_SYMBOL(do_page_fault);
1312 
1313 #ifdef CONFIG_TRACING
1314 static nokprobe_inline void
1315 trace_page_fault_entries(unsigned long address, struct pt_regs *regs,
1316 			 unsigned long error_code)
1317 {
1318 	if (user_mode(regs))
1319 		trace_page_fault_user(address, regs, error_code);
1320 	else
1321 		trace_page_fault_kernel(address, regs, error_code);
1322 }
1323 
1324 dotraplinkage void notrace
1325 trace_do_page_fault(struct pt_regs *regs, unsigned long error_code)
1326 {
1327 	/*
1328 	 * The exception_enter and tracepoint processing could
1329 	 * trigger another page faults (user space callchain
1330 	 * reading) and destroy the original cr2 value, so read
1331 	 * the faulting address now.
1332 	 */
1333 	unsigned long address = read_cr2();
1334 	enum ctx_state prev_state;
1335 
1336 	prev_state = exception_enter();
1337 	trace_page_fault_entries(address, regs, error_code);
1338 	__do_page_fault(regs, error_code, address);
1339 	exception_exit(prev_state);
1340 }
1341 NOKPROBE_SYMBOL(trace_do_page_fault);
1342 #endif /* CONFIG_TRACING */
1343