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