xref: /linux/arch/x86/mm/fault.c (revision 0d456bad36d42d16022be045c8a53ddbb59ee478)
1 /*
2  *  Copyright (C) 1995  Linus Torvalds
3  *  Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
4  *  Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
5  */
6 #include <linux/magic.h>		/* STACK_END_MAGIC		*/
7 #include <linux/sched.h>		/* test_thread_flag(), ...	*/
8 #include <linux/kdebug.h>		/* oops_begin/end, ...		*/
9 #include <linux/module.h>		/* search_exception_table	*/
10 #include <linux/bootmem.h>		/* max_low_pfn			*/
11 #include <linux/kprobes.h>		/* __kprobes, ...		*/
12 #include <linux/mmiotrace.h>		/* kmmio_handler, ...		*/
13 #include <linux/perf_event.h>		/* perf_sw_event		*/
14 #include <linux/hugetlb.h>		/* hstate_index_to_shift	*/
15 #include <linux/prefetch.h>		/* prefetchw			*/
16 
17 #include <asm/traps.h>			/* dotraplinkage, ...		*/
18 #include <asm/pgalloc.h>		/* pgd_*(), ...			*/
19 #include <asm/kmemcheck.h>		/* kmemcheck_*(), ...		*/
20 #include <asm/fixmap.h>			/* VSYSCALL_START		*/
21 #include <asm/context_tracking.h>	/* exception_enter(), ...	*/
22 
23 /*
24  * Page fault error code bits:
25  *
26  *   bit 0 ==	 0: no page found	1: protection fault
27  *   bit 1 ==	 0: read access		1: write access
28  *   bit 2 ==	 0: kernel-mode access	1: user-mode access
29  *   bit 3 ==				1: use of reserved bit detected
30  *   bit 4 ==				1: fault was an instruction fetch
31  */
32 enum x86_pf_error_code {
33 
34 	PF_PROT		=		1 << 0,
35 	PF_WRITE	=		1 << 1,
36 	PF_USER		=		1 << 2,
37 	PF_RSVD		=		1 << 3,
38 	PF_INSTR	=		1 << 4,
39 };
40 
41 /*
42  * Returns 0 if mmiotrace is disabled, or if the fault is not
43  * handled by mmiotrace:
44  */
45 static inline int __kprobes
46 kmmio_fault(struct pt_regs *regs, unsigned long addr)
47 {
48 	if (unlikely(is_kmmio_active()))
49 		if (kmmio_handler(regs, addr) == 1)
50 			return -1;
51 	return 0;
52 }
53 
54 static inline int __kprobes notify_page_fault(struct pt_regs *regs)
55 {
56 	int ret = 0;
57 
58 	/* kprobe_running() needs smp_processor_id() */
59 	if (kprobes_built_in() && !user_mode_vm(regs)) {
60 		preempt_disable();
61 		if (kprobe_running() && kprobe_fault_handler(regs, 14))
62 			ret = 1;
63 		preempt_enable();
64 	}
65 
66 	return ret;
67 }
68 
69 /*
70  * Prefetch quirks:
71  *
72  * 32-bit mode:
73  *
74  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
75  *   Check that here and ignore it.
76  *
77  * 64-bit mode:
78  *
79  *   Sometimes the CPU reports invalid exceptions on prefetch.
80  *   Check that here and ignore it.
81  *
82  * Opcode checker based on code by Richard Brunner.
83  */
84 static inline int
85 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
86 		      unsigned char opcode, int *prefetch)
87 {
88 	unsigned char instr_hi = opcode & 0xf0;
89 	unsigned char instr_lo = opcode & 0x0f;
90 
91 	switch (instr_hi) {
92 	case 0x20:
93 	case 0x30:
94 		/*
95 		 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
96 		 * In X86_64 long mode, the CPU will signal invalid
97 		 * opcode if some of these prefixes are present so
98 		 * X86_64 will never get here anyway
99 		 */
100 		return ((instr_lo & 7) == 0x6);
101 #ifdef CONFIG_X86_64
102 	case 0x40:
103 		/*
104 		 * In AMD64 long mode 0x40..0x4F are valid REX prefixes
105 		 * Need to figure out under what instruction mode the
106 		 * instruction was issued. Could check the LDT for lm,
107 		 * but for now it's good enough to assume that long
108 		 * mode only uses well known segments or kernel.
109 		 */
110 		return (!user_mode(regs) || user_64bit_mode(regs));
111 #endif
112 	case 0x60:
113 		/* 0x64 thru 0x67 are valid prefixes in all modes. */
114 		return (instr_lo & 0xC) == 0x4;
115 	case 0xF0:
116 		/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
117 		return !instr_lo || (instr_lo>>1) == 1;
118 	case 0x00:
119 		/* Prefetch instruction is 0x0F0D or 0x0F18 */
120 		if (probe_kernel_address(instr, opcode))
121 			return 0;
122 
123 		*prefetch = (instr_lo == 0xF) &&
124 			(opcode == 0x0D || opcode == 0x18);
125 		return 0;
126 	default:
127 		return 0;
128 	}
129 }
130 
131 static int
132 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
133 {
134 	unsigned char *max_instr;
135 	unsigned char *instr;
136 	int prefetch = 0;
137 
138 	/*
139 	 * If it was a exec (instruction fetch) fault on NX page, then
140 	 * do not ignore the fault:
141 	 */
142 	if (error_code & PF_INSTR)
143 		return 0;
144 
145 	instr = (void *)convert_ip_to_linear(current, regs);
146 	max_instr = instr + 15;
147 
148 	if (user_mode(regs) && instr >= (unsigned char *)TASK_SIZE)
149 		return 0;
150 
151 	while (instr < max_instr) {
152 		unsigned char opcode;
153 
154 		if (probe_kernel_address(instr, opcode))
155 			break;
156 
157 		instr++;
158 
159 		if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
160 			break;
161 	}
162 	return prefetch;
163 }
164 
165 static void
166 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
167 		     struct task_struct *tsk, int fault)
168 {
169 	unsigned lsb = 0;
170 	siginfo_t info;
171 
172 	info.si_signo	= si_signo;
173 	info.si_errno	= 0;
174 	info.si_code	= si_code;
175 	info.si_addr	= (void __user *)address;
176 	if (fault & VM_FAULT_HWPOISON_LARGE)
177 		lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
178 	if (fault & VM_FAULT_HWPOISON)
179 		lsb = PAGE_SHIFT;
180 	info.si_addr_lsb = lsb;
181 
182 	force_sig_info(si_signo, &info, tsk);
183 }
184 
185 DEFINE_SPINLOCK(pgd_lock);
186 LIST_HEAD(pgd_list);
187 
188 #ifdef CONFIG_X86_32
189 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
190 {
191 	unsigned index = pgd_index(address);
192 	pgd_t *pgd_k;
193 	pud_t *pud, *pud_k;
194 	pmd_t *pmd, *pmd_k;
195 
196 	pgd += index;
197 	pgd_k = init_mm.pgd + index;
198 
199 	if (!pgd_present(*pgd_k))
200 		return NULL;
201 
202 	/*
203 	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
204 	 * and redundant with the set_pmd() on non-PAE. As would
205 	 * set_pud.
206 	 */
207 	pud = pud_offset(pgd, address);
208 	pud_k = pud_offset(pgd_k, address);
209 	if (!pud_present(*pud_k))
210 		return NULL;
211 
212 	pmd = pmd_offset(pud, address);
213 	pmd_k = pmd_offset(pud_k, address);
214 	if (!pmd_present(*pmd_k))
215 		return NULL;
216 
217 	if (!pmd_present(*pmd))
218 		set_pmd(pmd, *pmd_k);
219 	else
220 		BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
221 
222 	return pmd_k;
223 }
224 
225 void vmalloc_sync_all(void)
226 {
227 	unsigned long address;
228 
229 	if (SHARED_KERNEL_PMD)
230 		return;
231 
232 	for (address = VMALLOC_START & PMD_MASK;
233 	     address >= TASK_SIZE && address < FIXADDR_TOP;
234 	     address += PMD_SIZE) {
235 		struct page *page;
236 
237 		spin_lock(&pgd_lock);
238 		list_for_each_entry(page, &pgd_list, lru) {
239 			spinlock_t *pgt_lock;
240 			pmd_t *ret;
241 
242 			/* the pgt_lock only for Xen */
243 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
244 
245 			spin_lock(pgt_lock);
246 			ret = vmalloc_sync_one(page_address(page), address);
247 			spin_unlock(pgt_lock);
248 
249 			if (!ret)
250 				break;
251 		}
252 		spin_unlock(&pgd_lock);
253 	}
254 }
255 
256 /*
257  * 32-bit:
258  *
259  *   Handle a fault on the vmalloc or module mapping area
260  */
261 static noinline __kprobes int vmalloc_fault(unsigned long address)
262 {
263 	unsigned long pgd_paddr;
264 	pmd_t *pmd_k;
265 	pte_t *pte_k;
266 
267 	/* Make sure we are in vmalloc area: */
268 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
269 		return -1;
270 
271 	WARN_ON_ONCE(in_nmi());
272 
273 	/*
274 	 * Synchronize this task's top level page-table
275 	 * with the 'reference' page table.
276 	 *
277 	 * Do _not_ use "current" here. We might be inside
278 	 * an interrupt in the middle of a task switch..
279 	 */
280 	pgd_paddr = read_cr3();
281 	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
282 	if (!pmd_k)
283 		return -1;
284 
285 	pte_k = pte_offset_kernel(pmd_k, address);
286 	if (!pte_present(*pte_k))
287 		return -1;
288 
289 	return 0;
290 }
291 
292 /*
293  * Did it hit the DOS screen memory VA from vm86 mode?
294  */
295 static inline void
296 check_v8086_mode(struct pt_regs *regs, unsigned long address,
297 		 struct task_struct *tsk)
298 {
299 	unsigned long bit;
300 
301 	if (!v8086_mode(regs))
302 		return;
303 
304 	bit = (address - 0xA0000) >> PAGE_SHIFT;
305 	if (bit < 32)
306 		tsk->thread.screen_bitmap |= 1 << bit;
307 }
308 
309 static bool low_pfn(unsigned long pfn)
310 {
311 	return pfn < max_low_pfn;
312 }
313 
314 static void dump_pagetable(unsigned long address)
315 {
316 	pgd_t *base = __va(read_cr3());
317 	pgd_t *pgd = &base[pgd_index(address)];
318 	pmd_t *pmd;
319 	pte_t *pte;
320 
321 #ifdef CONFIG_X86_PAE
322 	printk("*pdpt = %016Lx ", pgd_val(*pgd));
323 	if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
324 		goto out;
325 #endif
326 	pmd = pmd_offset(pud_offset(pgd, address), address);
327 	printk(KERN_CONT "*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
328 
329 	/*
330 	 * We must not directly access the pte in the highpte
331 	 * case if the page table is located in highmem.
332 	 * And let's rather not kmap-atomic the pte, just in case
333 	 * it's allocated already:
334 	 */
335 	if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
336 		goto out;
337 
338 	pte = pte_offset_kernel(pmd, address);
339 	printk("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
340 out:
341 	printk("\n");
342 }
343 
344 #else /* CONFIG_X86_64: */
345 
346 void vmalloc_sync_all(void)
347 {
348 	sync_global_pgds(VMALLOC_START & PGDIR_MASK, VMALLOC_END);
349 }
350 
351 /*
352  * 64-bit:
353  *
354  *   Handle a fault on the vmalloc area
355  *
356  * This assumes no large pages in there.
357  */
358 static noinline __kprobes int vmalloc_fault(unsigned long address)
359 {
360 	pgd_t *pgd, *pgd_ref;
361 	pud_t *pud, *pud_ref;
362 	pmd_t *pmd, *pmd_ref;
363 	pte_t *pte, *pte_ref;
364 
365 	/* Make sure we are in vmalloc area: */
366 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
367 		return -1;
368 
369 	WARN_ON_ONCE(in_nmi());
370 
371 	/*
372 	 * Copy kernel mappings over when needed. This can also
373 	 * happen within a race in page table update. In the later
374 	 * case just flush:
375 	 */
376 	pgd = pgd_offset(current->active_mm, address);
377 	pgd_ref = pgd_offset_k(address);
378 	if (pgd_none(*pgd_ref))
379 		return -1;
380 
381 	if (pgd_none(*pgd))
382 		set_pgd(pgd, *pgd_ref);
383 	else
384 		BUG_ON(pgd_page_vaddr(*pgd) != pgd_page_vaddr(*pgd_ref));
385 
386 	/*
387 	 * Below here mismatches are bugs because these lower tables
388 	 * are shared:
389 	 */
390 
391 	pud = pud_offset(pgd, address);
392 	pud_ref = pud_offset(pgd_ref, address);
393 	if (pud_none(*pud_ref))
394 		return -1;
395 
396 	if (pud_none(*pud) || pud_page_vaddr(*pud) != pud_page_vaddr(*pud_ref))
397 		BUG();
398 
399 	pmd = pmd_offset(pud, address);
400 	pmd_ref = pmd_offset(pud_ref, address);
401 	if (pmd_none(*pmd_ref))
402 		return -1;
403 
404 	if (pmd_none(*pmd) || pmd_page(*pmd) != pmd_page(*pmd_ref))
405 		BUG();
406 
407 	pte_ref = pte_offset_kernel(pmd_ref, address);
408 	if (!pte_present(*pte_ref))
409 		return -1;
410 
411 	pte = pte_offset_kernel(pmd, address);
412 
413 	/*
414 	 * Don't use pte_page here, because the mappings can point
415 	 * outside mem_map, and the NUMA hash lookup cannot handle
416 	 * that:
417 	 */
418 	if (!pte_present(*pte) || pte_pfn(*pte) != pte_pfn(*pte_ref))
419 		BUG();
420 
421 	return 0;
422 }
423 
424 #ifdef CONFIG_CPU_SUP_AMD
425 static const char errata93_warning[] =
426 KERN_ERR
427 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
428 "******* Working around it, but it may cause SEGVs or burn power.\n"
429 "******* Please consider a BIOS update.\n"
430 "******* Disabling USB legacy in the BIOS may also help.\n";
431 #endif
432 
433 /*
434  * No vm86 mode in 64-bit mode:
435  */
436 static inline void
437 check_v8086_mode(struct pt_regs *regs, unsigned long address,
438 		 struct task_struct *tsk)
439 {
440 }
441 
442 static int bad_address(void *p)
443 {
444 	unsigned long dummy;
445 
446 	return probe_kernel_address((unsigned long *)p, dummy);
447 }
448 
449 static void dump_pagetable(unsigned long address)
450 {
451 	pgd_t *base = __va(read_cr3() & PHYSICAL_PAGE_MASK);
452 	pgd_t *pgd = base + pgd_index(address);
453 	pud_t *pud;
454 	pmd_t *pmd;
455 	pte_t *pte;
456 
457 	if (bad_address(pgd))
458 		goto bad;
459 
460 	printk("PGD %lx ", pgd_val(*pgd));
461 
462 	if (!pgd_present(*pgd))
463 		goto out;
464 
465 	pud = pud_offset(pgd, address);
466 	if (bad_address(pud))
467 		goto bad;
468 
469 	printk("PUD %lx ", pud_val(*pud));
470 	if (!pud_present(*pud) || pud_large(*pud))
471 		goto out;
472 
473 	pmd = pmd_offset(pud, address);
474 	if (bad_address(pmd))
475 		goto bad;
476 
477 	printk("PMD %lx ", pmd_val(*pmd));
478 	if (!pmd_present(*pmd) || pmd_large(*pmd))
479 		goto out;
480 
481 	pte = pte_offset_kernel(pmd, address);
482 	if (bad_address(pte))
483 		goto bad;
484 
485 	printk("PTE %lx", pte_val(*pte));
486 out:
487 	printk("\n");
488 	return;
489 bad:
490 	printk("BAD\n");
491 }
492 
493 #endif /* CONFIG_X86_64 */
494 
495 /*
496  * Workaround for K8 erratum #93 & buggy BIOS.
497  *
498  * BIOS SMM functions are required to use a specific workaround
499  * to avoid corruption of the 64bit RIP register on C stepping K8.
500  *
501  * A lot of BIOS that didn't get tested properly miss this.
502  *
503  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
504  * Try to work around it here.
505  *
506  * Note we only handle faults in kernel here.
507  * Does nothing on 32-bit.
508  */
509 static int is_errata93(struct pt_regs *regs, unsigned long address)
510 {
511 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
512 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
513 	    || boot_cpu_data.x86 != 0xf)
514 		return 0;
515 
516 	if (address != regs->ip)
517 		return 0;
518 
519 	if ((address >> 32) != 0)
520 		return 0;
521 
522 	address |= 0xffffffffUL << 32;
523 	if ((address >= (u64)_stext && address <= (u64)_etext) ||
524 	    (address >= MODULES_VADDR && address <= MODULES_END)) {
525 		printk_once(errata93_warning);
526 		regs->ip = address;
527 		return 1;
528 	}
529 #endif
530 	return 0;
531 }
532 
533 /*
534  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
535  * to illegal addresses >4GB.
536  *
537  * We catch this in the page fault handler because these addresses
538  * are not reachable. Just detect this case and return.  Any code
539  * segment in LDT is compatibility mode.
540  */
541 static int is_errata100(struct pt_regs *regs, unsigned long address)
542 {
543 #ifdef CONFIG_X86_64
544 	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
545 		return 1;
546 #endif
547 	return 0;
548 }
549 
550 static int is_f00f_bug(struct pt_regs *regs, unsigned long address)
551 {
552 #ifdef CONFIG_X86_F00F_BUG
553 	unsigned long nr;
554 
555 	/*
556 	 * Pentium F0 0F C7 C8 bug workaround:
557 	 */
558 	if (boot_cpu_data.f00f_bug) {
559 		nr = (address - idt_descr.address) >> 3;
560 
561 		if (nr == 6) {
562 			do_invalid_op(regs, 0);
563 			return 1;
564 		}
565 	}
566 #endif
567 	return 0;
568 }
569 
570 static const char nx_warning[] = KERN_CRIT
571 "kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n";
572 
573 static void
574 show_fault_oops(struct pt_regs *regs, unsigned long error_code,
575 		unsigned long address)
576 {
577 	if (!oops_may_print())
578 		return;
579 
580 	if (error_code & PF_INSTR) {
581 		unsigned int level;
582 
583 		pte_t *pte = lookup_address(address, &level);
584 
585 		if (pte && pte_present(*pte) && !pte_exec(*pte))
586 			printk(nx_warning, from_kuid(&init_user_ns, current_uid()));
587 	}
588 
589 	printk(KERN_ALERT "BUG: unable to handle kernel ");
590 	if (address < PAGE_SIZE)
591 		printk(KERN_CONT "NULL pointer dereference");
592 	else
593 		printk(KERN_CONT "paging request");
594 
595 	printk(KERN_CONT " at %p\n", (void *) address);
596 	printk(KERN_ALERT "IP:");
597 	printk_address(regs->ip, 1);
598 
599 	dump_pagetable(address);
600 }
601 
602 static noinline void
603 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
604 	    unsigned long address)
605 {
606 	struct task_struct *tsk;
607 	unsigned long flags;
608 	int sig;
609 
610 	flags = oops_begin();
611 	tsk = current;
612 	sig = SIGKILL;
613 
614 	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
615 	       tsk->comm, address);
616 	dump_pagetable(address);
617 
618 	tsk->thread.cr2		= address;
619 	tsk->thread.trap_nr	= X86_TRAP_PF;
620 	tsk->thread.error_code	= error_code;
621 
622 	if (__die("Bad pagetable", regs, error_code))
623 		sig = 0;
624 
625 	oops_end(flags, regs, sig);
626 }
627 
628 static noinline void
629 no_context(struct pt_regs *regs, unsigned long error_code,
630 	   unsigned long address, int signal, int si_code)
631 {
632 	struct task_struct *tsk = current;
633 	unsigned long *stackend;
634 	unsigned long flags;
635 	int sig;
636 
637 	/* Are we prepared to handle this kernel fault? */
638 	if (fixup_exception(regs)) {
639 		if (current_thread_info()->sig_on_uaccess_error && signal) {
640 			tsk->thread.trap_nr = X86_TRAP_PF;
641 			tsk->thread.error_code = error_code | PF_USER;
642 			tsk->thread.cr2 = address;
643 
644 			/* XXX: hwpoison faults will set the wrong code. */
645 			force_sig_info_fault(signal, si_code, address, tsk, 0);
646 		}
647 		return;
648 	}
649 
650 	/*
651 	 * 32-bit:
652 	 *
653 	 *   Valid to do another page fault here, because if this fault
654 	 *   had been triggered by is_prefetch fixup_exception would have
655 	 *   handled it.
656 	 *
657 	 * 64-bit:
658 	 *
659 	 *   Hall of shame of CPU/BIOS bugs.
660 	 */
661 	if (is_prefetch(regs, error_code, address))
662 		return;
663 
664 	if (is_errata93(regs, address))
665 		return;
666 
667 	/*
668 	 * Oops. The kernel tried to access some bad page. We'll have to
669 	 * terminate things with extreme prejudice:
670 	 */
671 	flags = oops_begin();
672 
673 	show_fault_oops(regs, error_code, address);
674 
675 	stackend = end_of_stack(tsk);
676 	if (tsk != &init_task && *stackend != STACK_END_MAGIC)
677 		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
678 
679 	tsk->thread.cr2		= address;
680 	tsk->thread.trap_nr	= X86_TRAP_PF;
681 	tsk->thread.error_code	= error_code;
682 
683 	sig = SIGKILL;
684 	if (__die("Oops", regs, error_code))
685 		sig = 0;
686 
687 	/* Executive summary in case the body of the oops scrolled away */
688 	printk(KERN_DEFAULT "CR2: %016lx\n", address);
689 
690 	oops_end(flags, regs, sig);
691 }
692 
693 /*
694  * Print out info about fatal segfaults, if the show_unhandled_signals
695  * sysctl is set:
696  */
697 static inline void
698 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
699 		unsigned long address, struct task_struct *tsk)
700 {
701 	if (!unhandled_signal(tsk, SIGSEGV))
702 		return;
703 
704 	if (!printk_ratelimit())
705 		return;
706 
707 	printk("%s%s[%d]: segfault at %lx ip %p sp %p error %lx",
708 		task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG,
709 		tsk->comm, task_pid_nr(tsk), address,
710 		(void *)regs->ip, (void *)regs->sp, error_code);
711 
712 	print_vma_addr(KERN_CONT " in ", regs->ip);
713 
714 	printk(KERN_CONT "\n");
715 }
716 
717 static void
718 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
719 		       unsigned long address, int si_code)
720 {
721 	struct task_struct *tsk = current;
722 
723 	/* User mode accesses just cause a SIGSEGV */
724 	if (error_code & PF_USER) {
725 		/*
726 		 * It's possible to have interrupts off here:
727 		 */
728 		local_irq_enable();
729 
730 		/*
731 		 * Valid to do another page fault here because this one came
732 		 * from user space:
733 		 */
734 		if (is_prefetch(regs, error_code, address))
735 			return;
736 
737 		if (is_errata100(regs, address))
738 			return;
739 
740 #ifdef CONFIG_X86_64
741 		/*
742 		 * Instruction fetch faults in the vsyscall page might need
743 		 * emulation.
744 		 */
745 		if (unlikely((error_code & PF_INSTR) &&
746 			     ((address & ~0xfff) == VSYSCALL_START))) {
747 			if (emulate_vsyscall(regs, address))
748 				return;
749 		}
750 #endif
751 
752 		if (unlikely(show_unhandled_signals))
753 			show_signal_msg(regs, error_code, address, tsk);
754 
755 		/* Kernel addresses are always protection faults: */
756 		tsk->thread.cr2		= address;
757 		tsk->thread.error_code	= error_code | (address >= TASK_SIZE);
758 		tsk->thread.trap_nr	= X86_TRAP_PF;
759 
760 		force_sig_info_fault(SIGSEGV, si_code, address, tsk, 0);
761 
762 		return;
763 	}
764 
765 	if (is_f00f_bug(regs, address))
766 		return;
767 
768 	no_context(regs, error_code, address, SIGSEGV, si_code);
769 }
770 
771 static noinline void
772 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
773 		     unsigned long address)
774 {
775 	__bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
776 }
777 
778 static void
779 __bad_area(struct pt_regs *regs, unsigned long error_code,
780 	   unsigned long address, int si_code)
781 {
782 	struct mm_struct *mm = current->mm;
783 
784 	/*
785 	 * Something tried to access memory that isn't in our memory map..
786 	 * Fix it, but check if it's kernel or user first..
787 	 */
788 	up_read(&mm->mmap_sem);
789 
790 	__bad_area_nosemaphore(regs, error_code, address, si_code);
791 }
792 
793 static noinline void
794 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
795 {
796 	__bad_area(regs, error_code, address, SEGV_MAPERR);
797 }
798 
799 static noinline void
800 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
801 		      unsigned long address)
802 {
803 	__bad_area(regs, error_code, address, SEGV_ACCERR);
804 }
805 
806 static void
807 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
808 	  unsigned int fault)
809 {
810 	struct task_struct *tsk = current;
811 	struct mm_struct *mm = tsk->mm;
812 	int code = BUS_ADRERR;
813 
814 	up_read(&mm->mmap_sem);
815 
816 	/* Kernel mode? Handle exceptions or die: */
817 	if (!(error_code & PF_USER)) {
818 		no_context(regs, error_code, address, SIGBUS, BUS_ADRERR);
819 		return;
820 	}
821 
822 	/* User-space => ok to do another page fault: */
823 	if (is_prefetch(regs, error_code, address))
824 		return;
825 
826 	tsk->thread.cr2		= address;
827 	tsk->thread.error_code	= error_code;
828 	tsk->thread.trap_nr	= X86_TRAP_PF;
829 
830 #ifdef CONFIG_MEMORY_FAILURE
831 	if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
832 		printk(KERN_ERR
833 	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
834 			tsk->comm, tsk->pid, address);
835 		code = BUS_MCEERR_AR;
836 	}
837 #endif
838 	force_sig_info_fault(SIGBUS, code, address, tsk, fault);
839 }
840 
841 static noinline int
842 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
843 	       unsigned long address, unsigned int fault)
844 {
845 	/*
846 	 * Pagefault was interrupted by SIGKILL. We have no reason to
847 	 * continue pagefault.
848 	 */
849 	if (fatal_signal_pending(current)) {
850 		if (!(fault & VM_FAULT_RETRY))
851 			up_read(&current->mm->mmap_sem);
852 		if (!(error_code & PF_USER))
853 			no_context(regs, error_code, address, 0, 0);
854 		return 1;
855 	}
856 	if (!(fault & VM_FAULT_ERROR))
857 		return 0;
858 
859 	if (fault & VM_FAULT_OOM) {
860 		/* Kernel mode? Handle exceptions or die: */
861 		if (!(error_code & PF_USER)) {
862 			up_read(&current->mm->mmap_sem);
863 			no_context(regs, error_code, address,
864 				   SIGSEGV, SEGV_MAPERR);
865 			return 1;
866 		}
867 
868 		up_read(&current->mm->mmap_sem);
869 
870 		/*
871 		 * We ran out of memory, call the OOM killer, and return the
872 		 * userspace (which will retry the fault, or kill us if we got
873 		 * oom-killed):
874 		 */
875 		pagefault_out_of_memory();
876 	} else {
877 		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
878 			     VM_FAULT_HWPOISON_LARGE))
879 			do_sigbus(regs, error_code, address, fault);
880 		else
881 			BUG();
882 	}
883 	return 1;
884 }
885 
886 static int spurious_fault_check(unsigned long error_code, pte_t *pte)
887 {
888 	if ((error_code & PF_WRITE) && !pte_write(*pte))
889 		return 0;
890 
891 	if ((error_code & PF_INSTR) && !pte_exec(*pte))
892 		return 0;
893 
894 	return 1;
895 }
896 
897 /*
898  * Handle a spurious fault caused by a stale TLB entry.
899  *
900  * This allows us to lazily refresh the TLB when increasing the
901  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
902  * eagerly is very expensive since that implies doing a full
903  * cross-processor TLB flush, even if no stale TLB entries exist
904  * on other processors.
905  *
906  * There are no security implications to leaving a stale TLB when
907  * increasing the permissions on a page.
908  */
909 static noinline __kprobes int
910 spurious_fault(unsigned long error_code, unsigned long address)
911 {
912 	pgd_t *pgd;
913 	pud_t *pud;
914 	pmd_t *pmd;
915 	pte_t *pte;
916 	int ret;
917 
918 	/* Reserved-bit violation or user access to kernel space? */
919 	if (error_code & (PF_USER | PF_RSVD))
920 		return 0;
921 
922 	pgd = init_mm.pgd + pgd_index(address);
923 	if (!pgd_present(*pgd))
924 		return 0;
925 
926 	pud = pud_offset(pgd, address);
927 	if (!pud_present(*pud))
928 		return 0;
929 
930 	if (pud_large(*pud))
931 		return spurious_fault_check(error_code, (pte_t *) pud);
932 
933 	pmd = pmd_offset(pud, address);
934 	if (!pmd_present(*pmd))
935 		return 0;
936 
937 	if (pmd_large(*pmd))
938 		return spurious_fault_check(error_code, (pte_t *) pmd);
939 
940 	/*
941 	 * Note: don't use pte_present() here, since it returns true
942 	 * if the _PAGE_PROTNONE bit is set.  However, this aliases the
943 	 * _PAGE_GLOBAL bit, which for kernel pages give false positives
944 	 * when CONFIG_DEBUG_PAGEALLOC is used.
945 	 */
946 	pte = pte_offset_kernel(pmd, address);
947 	if (!(pte_flags(*pte) & _PAGE_PRESENT))
948 		return 0;
949 
950 	ret = spurious_fault_check(error_code, pte);
951 	if (!ret)
952 		return 0;
953 
954 	/*
955 	 * Make sure we have permissions in PMD.
956 	 * If not, then there's a bug in the page tables:
957 	 */
958 	ret = spurious_fault_check(error_code, (pte_t *) pmd);
959 	WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
960 
961 	return ret;
962 }
963 
964 int show_unhandled_signals = 1;
965 
966 static inline int
967 access_error(unsigned long error_code, struct vm_area_struct *vma)
968 {
969 	if (error_code & PF_WRITE) {
970 		/* write, present and write, not present: */
971 		if (unlikely(!(vma->vm_flags & VM_WRITE)))
972 			return 1;
973 		return 0;
974 	}
975 
976 	/* read, present: */
977 	if (unlikely(error_code & PF_PROT))
978 		return 1;
979 
980 	/* read, not present: */
981 	if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
982 		return 1;
983 
984 	return 0;
985 }
986 
987 static int fault_in_kernel_space(unsigned long address)
988 {
989 	return address >= TASK_SIZE_MAX;
990 }
991 
992 static inline bool smap_violation(int error_code, struct pt_regs *regs)
993 {
994 	if (error_code & PF_USER)
995 		return false;
996 
997 	if (!user_mode_vm(regs) && (regs->flags & X86_EFLAGS_AC))
998 		return false;
999 
1000 	return true;
1001 }
1002 
1003 /*
1004  * This routine handles page faults.  It determines the address,
1005  * and the problem, and then passes it off to one of the appropriate
1006  * routines.
1007  */
1008 static void __kprobes
1009 __do_page_fault(struct pt_regs *regs, unsigned long error_code)
1010 {
1011 	struct vm_area_struct *vma;
1012 	struct task_struct *tsk;
1013 	unsigned long address;
1014 	struct mm_struct *mm;
1015 	int fault;
1016 	int write = error_code & PF_WRITE;
1017 	unsigned int flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE |
1018 					(write ? FAULT_FLAG_WRITE : 0);
1019 
1020 	tsk = current;
1021 	mm = tsk->mm;
1022 
1023 	/* Get the faulting address: */
1024 	address = read_cr2();
1025 
1026 	/*
1027 	 * Detect and handle instructions that would cause a page fault for
1028 	 * both a tracked kernel page and a userspace page.
1029 	 */
1030 	if (kmemcheck_active(regs))
1031 		kmemcheck_hide(regs);
1032 	prefetchw(&mm->mmap_sem);
1033 
1034 	if (unlikely(kmmio_fault(regs, address)))
1035 		return;
1036 
1037 	/*
1038 	 * We fault-in kernel-space virtual memory on-demand. The
1039 	 * 'reference' page table is init_mm.pgd.
1040 	 *
1041 	 * NOTE! We MUST NOT take any locks for this case. We may
1042 	 * be in an interrupt or a critical region, and should
1043 	 * only copy the information from the master page table,
1044 	 * nothing more.
1045 	 *
1046 	 * This verifies that the fault happens in kernel space
1047 	 * (error_code & 4) == 0, and that the fault was not a
1048 	 * protection error (error_code & 9) == 0.
1049 	 */
1050 	if (unlikely(fault_in_kernel_space(address))) {
1051 		if (!(error_code & (PF_RSVD | PF_USER | PF_PROT))) {
1052 			if (vmalloc_fault(address) >= 0)
1053 				return;
1054 
1055 			if (kmemcheck_fault(regs, address, error_code))
1056 				return;
1057 		}
1058 
1059 		/* Can handle a stale RO->RW TLB: */
1060 		if (spurious_fault(error_code, address))
1061 			return;
1062 
1063 		/* kprobes don't want to hook the spurious faults: */
1064 		if (notify_page_fault(regs))
1065 			return;
1066 		/*
1067 		 * Don't take the mm semaphore here. If we fixup a prefetch
1068 		 * fault we could otherwise deadlock:
1069 		 */
1070 		bad_area_nosemaphore(regs, error_code, address);
1071 
1072 		return;
1073 	}
1074 
1075 	/* kprobes don't want to hook the spurious faults: */
1076 	if (unlikely(notify_page_fault(regs)))
1077 		return;
1078 	/*
1079 	 * It's safe to allow irq's after cr2 has been saved and the
1080 	 * vmalloc fault has been handled.
1081 	 *
1082 	 * User-mode registers count as a user access even for any
1083 	 * potential system fault or CPU buglet:
1084 	 */
1085 	if (user_mode_vm(regs)) {
1086 		local_irq_enable();
1087 		error_code |= PF_USER;
1088 	} else {
1089 		if (regs->flags & X86_EFLAGS_IF)
1090 			local_irq_enable();
1091 	}
1092 
1093 	if (unlikely(error_code & PF_RSVD))
1094 		pgtable_bad(regs, error_code, address);
1095 
1096 	if (static_cpu_has(X86_FEATURE_SMAP)) {
1097 		if (unlikely(smap_violation(error_code, regs))) {
1098 			bad_area_nosemaphore(regs, error_code, address);
1099 			return;
1100 		}
1101 	}
1102 
1103 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1104 
1105 	/*
1106 	 * If we're in an interrupt, have no user context or are running
1107 	 * in an atomic region then we must not take the fault:
1108 	 */
1109 	if (unlikely(in_atomic() || !mm)) {
1110 		bad_area_nosemaphore(regs, error_code, address);
1111 		return;
1112 	}
1113 
1114 	/*
1115 	 * When running in the kernel we expect faults to occur only to
1116 	 * addresses in user space.  All other faults represent errors in
1117 	 * the kernel and should generate an OOPS.  Unfortunately, in the
1118 	 * case of an erroneous fault occurring in a code path which already
1119 	 * holds mmap_sem we will deadlock attempting to validate the fault
1120 	 * against the address space.  Luckily the kernel only validly
1121 	 * references user space from well defined areas of code, which are
1122 	 * listed in the exceptions table.
1123 	 *
1124 	 * As the vast majority of faults will be valid we will only perform
1125 	 * the source reference check when there is a possibility of a
1126 	 * deadlock. Attempt to lock the address space, if we cannot we then
1127 	 * validate the source. If this is invalid we can skip the address
1128 	 * space check, thus avoiding the deadlock:
1129 	 */
1130 	if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
1131 		if ((error_code & PF_USER) == 0 &&
1132 		    !search_exception_tables(regs->ip)) {
1133 			bad_area_nosemaphore(regs, error_code, address);
1134 			return;
1135 		}
1136 retry:
1137 		down_read(&mm->mmap_sem);
1138 	} else {
1139 		/*
1140 		 * The above down_read_trylock() might have succeeded in
1141 		 * which case we'll have missed the might_sleep() from
1142 		 * down_read():
1143 		 */
1144 		might_sleep();
1145 	}
1146 
1147 	vma = find_vma(mm, address);
1148 	if (unlikely(!vma)) {
1149 		bad_area(regs, error_code, address);
1150 		return;
1151 	}
1152 	if (likely(vma->vm_start <= address))
1153 		goto good_area;
1154 	if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
1155 		bad_area(regs, error_code, address);
1156 		return;
1157 	}
1158 	if (error_code & PF_USER) {
1159 		/*
1160 		 * Accessing the stack below %sp is always a bug.
1161 		 * The large cushion allows instructions like enter
1162 		 * and pusha to work. ("enter $65535, $31" pushes
1163 		 * 32 pointers and then decrements %sp by 65535.)
1164 		 */
1165 		if (unlikely(address + 65536 + 32 * sizeof(unsigned long) < regs->sp)) {
1166 			bad_area(regs, error_code, address);
1167 			return;
1168 		}
1169 	}
1170 	if (unlikely(expand_stack(vma, address))) {
1171 		bad_area(regs, error_code, address);
1172 		return;
1173 	}
1174 
1175 	/*
1176 	 * Ok, we have a good vm_area for this memory access, so
1177 	 * we can handle it..
1178 	 */
1179 good_area:
1180 	if (unlikely(access_error(error_code, vma))) {
1181 		bad_area_access_error(regs, error_code, address);
1182 		return;
1183 	}
1184 
1185 	/*
1186 	 * If for any reason at all we couldn't handle the fault,
1187 	 * make sure we exit gracefully rather than endlessly redo
1188 	 * the fault:
1189 	 */
1190 	fault = handle_mm_fault(mm, vma, address, flags);
1191 
1192 	if (unlikely(fault & (VM_FAULT_RETRY|VM_FAULT_ERROR))) {
1193 		if (mm_fault_error(regs, error_code, address, fault))
1194 			return;
1195 	}
1196 
1197 	/*
1198 	 * Major/minor page fault accounting is only done on the
1199 	 * initial attempt. If we go through a retry, it is extremely
1200 	 * likely that the page will be found in page cache at that point.
1201 	 */
1202 	if (flags & FAULT_FLAG_ALLOW_RETRY) {
1203 		if (fault & VM_FAULT_MAJOR) {
1204 			tsk->maj_flt++;
1205 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
1206 				      regs, address);
1207 		} else {
1208 			tsk->min_flt++;
1209 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
1210 				      regs, address);
1211 		}
1212 		if (fault & VM_FAULT_RETRY) {
1213 			/* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
1214 			 * of starvation. */
1215 			flags &= ~FAULT_FLAG_ALLOW_RETRY;
1216 			flags |= FAULT_FLAG_TRIED;
1217 			goto retry;
1218 		}
1219 	}
1220 
1221 	check_v8086_mode(regs, address, tsk);
1222 
1223 	up_read(&mm->mmap_sem);
1224 }
1225 
1226 dotraplinkage void __kprobes
1227 do_page_fault(struct pt_regs *regs, unsigned long error_code)
1228 {
1229 	exception_enter(regs);
1230 	__do_page_fault(regs, error_code);
1231 	exception_exit(regs);
1232 }
1233