xref: /openbmc/linux/arch/x86/mm/fault.c (revision 2f4e3926)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  Copyright (C) 1995  Linus Torvalds
4  *  Copyright (C) 2001, 2002 Andi Kleen, SuSE Labs.
5  *  Copyright (C) 2008-2009, Red Hat Inc., Ingo Molnar
6  */
7 #include <linux/sched.h>		/* test_thread_flag(), ...	*/
8 #include <linux/sched/task_stack.h>	/* task_stack_*(), ...		*/
9 #include <linux/kdebug.h>		/* oops_begin/end, ...		*/
10 #include <linux/extable.h>		/* search_exception_tables	*/
11 #include <linux/memblock.h>		/* max_low_pfn			*/
12 #include <linux/kfence.h>		/* kfence_handle_page_fault	*/
13 #include <linux/kprobes.h>		/* NOKPROBE_SYMBOL, ...		*/
14 #include <linux/mmiotrace.h>		/* kmmio_handler, ...		*/
15 #include <linux/perf_event.h>		/* perf_sw_event		*/
16 #include <linux/hugetlb.h>		/* hstate_index_to_shift	*/
17 #include <linux/prefetch.h>		/* prefetchw			*/
18 #include <linux/context_tracking.h>	/* exception_enter(), ...	*/
19 #include <linux/uaccess.h>		/* faulthandler_disabled()	*/
20 #include <linux/efi.h>			/* efi_crash_gracefully_on_page_fault()*/
21 #include <linux/mm_types.h>
22 #include <linux/mm.h>			/* find_and_lock_vma() */
23 
24 #include <asm/cpufeature.h>		/* boot_cpu_has, ...		*/
25 #include <asm/traps.h>			/* dotraplinkage, ...		*/
26 #include <asm/fixmap.h>			/* VSYSCALL_ADDR		*/
27 #include <asm/vsyscall.h>		/* emulate_vsyscall		*/
28 #include <asm/vm86.h>			/* struct vm86			*/
29 #include <asm/mmu_context.h>		/* vma_pkey()			*/
30 #include <asm/efi.h>			/* efi_crash_gracefully_on_page_fault()*/
31 #include <asm/desc.h>			/* store_idt(), ...		*/
32 #include <asm/cpu_entry_area.h>		/* exception stack		*/
33 #include <asm/pgtable_areas.h>		/* VMALLOC_START, ...		*/
34 #include <asm/kvm_para.h>		/* kvm_handle_async_pf		*/
35 #include <asm/vdso.h>			/* fixup_vdso_exception()	*/
36 #include <asm/irq_stack.h>
37 
38 #define CREATE_TRACE_POINTS
39 #include <asm/trace/exceptions.h>
40 
41 /*
42  * Returns 0 if mmiotrace is disabled, or if the fault is not
43  * handled by mmiotrace:
44  */
45 static nokprobe_inline int
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 /*
55  * Prefetch quirks:
56  *
57  * 32-bit mode:
58  *
59  *   Sometimes AMD Athlon/Opteron CPUs report invalid exceptions on prefetch.
60  *   Check that here and ignore it.  This is AMD erratum #91.
61  *
62  * 64-bit mode:
63  *
64  *   Sometimes the CPU reports invalid exceptions on prefetch.
65  *   Check that here and ignore it.
66  *
67  * Opcode checker based on code by Richard Brunner.
68  */
69 static inline int
70 check_prefetch_opcode(struct pt_regs *regs, unsigned char *instr,
71 		      unsigned char opcode, int *prefetch)
72 {
73 	unsigned char instr_hi = opcode & 0xf0;
74 	unsigned char instr_lo = opcode & 0x0f;
75 
76 	switch (instr_hi) {
77 	case 0x20:
78 	case 0x30:
79 		/*
80 		 * Values 0x26,0x2E,0x36,0x3E are valid x86 prefixes.
81 		 * In X86_64 long mode, the CPU will signal invalid
82 		 * opcode if some of these prefixes are present so
83 		 * X86_64 will never get here anyway
84 		 */
85 		return ((instr_lo & 7) == 0x6);
86 #ifdef CONFIG_X86_64
87 	case 0x40:
88 		/*
89 		 * In 64-bit mode 0x40..0x4F are valid REX prefixes
90 		 */
91 		return (!user_mode(regs) || user_64bit_mode(regs));
92 #endif
93 	case 0x60:
94 		/* 0x64 thru 0x67 are valid prefixes in all modes. */
95 		return (instr_lo & 0xC) == 0x4;
96 	case 0xF0:
97 		/* 0xF0, 0xF2, 0xF3 are valid prefixes in all modes. */
98 		return !instr_lo || (instr_lo>>1) == 1;
99 	case 0x00:
100 		/* Prefetch instruction is 0x0F0D or 0x0F18 */
101 		if (get_kernel_nofault(opcode, instr))
102 			return 0;
103 
104 		*prefetch = (instr_lo == 0xF) &&
105 			(opcode == 0x0D || opcode == 0x18);
106 		return 0;
107 	default:
108 		return 0;
109 	}
110 }
111 
112 static bool is_amd_k8_pre_npt(void)
113 {
114 	struct cpuinfo_x86 *c = &boot_cpu_data;
115 
116 	return unlikely(IS_ENABLED(CONFIG_CPU_SUP_AMD) &&
117 			c->x86_vendor == X86_VENDOR_AMD &&
118 			c->x86 == 0xf && c->x86_model < 0x40);
119 }
120 
121 static int
122 is_prefetch(struct pt_regs *regs, unsigned long error_code, unsigned long addr)
123 {
124 	unsigned char *max_instr;
125 	unsigned char *instr;
126 	int prefetch = 0;
127 
128 	/* Erratum #91 affects AMD K8, pre-NPT CPUs */
129 	if (!is_amd_k8_pre_npt())
130 		return 0;
131 
132 	/*
133 	 * If it was a exec (instruction fetch) fault on NX page, then
134 	 * do not ignore the fault:
135 	 */
136 	if (error_code & X86_PF_INSTR)
137 		return 0;
138 
139 	instr = (void *)convert_ip_to_linear(current, regs);
140 	max_instr = instr + 15;
141 
142 	/*
143 	 * This code has historically always bailed out if IP points to a
144 	 * not-present page (e.g. due to a race).  No one has ever
145 	 * complained about this.
146 	 */
147 	pagefault_disable();
148 
149 	while (instr < max_instr) {
150 		unsigned char opcode;
151 
152 		if (user_mode(regs)) {
153 			if (get_user(opcode, (unsigned char __user *) instr))
154 				break;
155 		} else {
156 			if (get_kernel_nofault(opcode, instr))
157 				break;
158 		}
159 
160 		instr++;
161 
162 		if (!check_prefetch_opcode(regs, instr, opcode, &prefetch))
163 			break;
164 	}
165 
166 	pagefault_enable();
167 	return prefetch;
168 }
169 
170 DEFINE_SPINLOCK(pgd_lock);
171 LIST_HEAD(pgd_list);
172 
173 #ifdef CONFIG_X86_32
174 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
175 {
176 	unsigned index = pgd_index(address);
177 	pgd_t *pgd_k;
178 	p4d_t *p4d, *p4d_k;
179 	pud_t *pud, *pud_k;
180 	pmd_t *pmd, *pmd_k;
181 
182 	pgd += index;
183 	pgd_k = init_mm.pgd + index;
184 
185 	if (!pgd_present(*pgd_k))
186 		return NULL;
187 
188 	/*
189 	 * set_pgd(pgd, *pgd_k); here would be useless on PAE
190 	 * and redundant with the set_pmd() on non-PAE. As would
191 	 * set_p4d/set_pud.
192 	 */
193 	p4d = p4d_offset(pgd, address);
194 	p4d_k = p4d_offset(pgd_k, address);
195 	if (!p4d_present(*p4d_k))
196 		return NULL;
197 
198 	pud = pud_offset(p4d, address);
199 	pud_k = pud_offset(p4d_k, address);
200 	if (!pud_present(*pud_k))
201 		return NULL;
202 
203 	pmd = pmd_offset(pud, address);
204 	pmd_k = pmd_offset(pud_k, address);
205 
206 	if (pmd_present(*pmd) != pmd_present(*pmd_k))
207 		set_pmd(pmd, *pmd_k);
208 
209 	if (!pmd_present(*pmd_k))
210 		return NULL;
211 	else
212 		BUG_ON(pmd_pfn(*pmd) != pmd_pfn(*pmd_k));
213 
214 	return pmd_k;
215 }
216 
217 /*
218  *   Handle a fault on the vmalloc or module mapping area
219  *
220  *   This is needed because there is a race condition between the time
221  *   when the vmalloc mapping code updates the PMD to the point in time
222  *   where it synchronizes this update with the other page-tables in the
223  *   system.
224  *
225  *   In this race window another thread/CPU can map an area on the same
226  *   PMD, finds it already present and does not synchronize it with the
227  *   rest of the system yet. As a result v[mz]alloc might return areas
228  *   which are not mapped in every page-table in the system, causing an
229  *   unhandled page-fault when they are accessed.
230  */
231 static noinline int vmalloc_fault(unsigned long address)
232 {
233 	unsigned long pgd_paddr;
234 	pmd_t *pmd_k;
235 	pte_t *pte_k;
236 
237 	/* Make sure we are in vmalloc area: */
238 	if (!(address >= VMALLOC_START && address < VMALLOC_END))
239 		return -1;
240 
241 	/*
242 	 * Synchronize this task's top level page-table
243 	 * with the 'reference' page table.
244 	 *
245 	 * Do _not_ use "current" here. We might be inside
246 	 * an interrupt in the middle of a task switch..
247 	 */
248 	pgd_paddr = read_cr3_pa();
249 	pmd_k = vmalloc_sync_one(__va(pgd_paddr), address);
250 	if (!pmd_k)
251 		return -1;
252 
253 	if (pmd_large(*pmd_k))
254 		return 0;
255 
256 	pte_k = pte_offset_kernel(pmd_k, address);
257 	if (!pte_present(*pte_k))
258 		return -1;
259 
260 	return 0;
261 }
262 NOKPROBE_SYMBOL(vmalloc_fault);
263 
264 void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
265 {
266 	unsigned long addr;
267 
268 	for (addr = start & PMD_MASK;
269 	     addr >= TASK_SIZE_MAX && addr < VMALLOC_END;
270 	     addr += PMD_SIZE) {
271 		struct page *page;
272 
273 		spin_lock(&pgd_lock);
274 		list_for_each_entry(page, &pgd_list, lru) {
275 			spinlock_t *pgt_lock;
276 
277 			/* the pgt_lock only for Xen */
278 			pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
279 
280 			spin_lock(pgt_lock);
281 			vmalloc_sync_one(page_address(page), addr);
282 			spin_unlock(pgt_lock);
283 		}
284 		spin_unlock(&pgd_lock);
285 	}
286 }
287 
288 static bool low_pfn(unsigned long pfn)
289 {
290 	return pfn < max_low_pfn;
291 }
292 
293 static void dump_pagetable(unsigned long address)
294 {
295 	pgd_t *base = __va(read_cr3_pa());
296 	pgd_t *pgd = &base[pgd_index(address)];
297 	p4d_t *p4d;
298 	pud_t *pud;
299 	pmd_t *pmd;
300 	pte_t *pte;
301 
302 #ifdef CONFIG_X86_PAE
303 	pr_info("*pdpt = %016Lx ", pgd_val(*pgd));
304 	if (!low_pfn(pgd_val(*pgd) >> PAGE_SHIFT) || !pgd_present(*pgd))
305 		goto out;
306 #define pr_pde pr_cont
307 #else
308 #define pr_pde pr_info
309 #endif
310 	p4d = p4d_offset(pgd, address);
311 	pud = pud_offset(p4d, address);
312 	pmd = pmd_offset(pud, address);
313 	pr_pde("*pde = %0*Lx ", sizeof(*pmd) * 2, (u64)pmd_val(*pmd));
314 #undef pr_pde
315 
316 	/*
317 	 * We must not directly access the pte in the highpte
318 	 * case if the page table is located in highmem.
319 	 * And let's rather not kmap-atomic the pte, just in case
320 	 * it's allocated already:
321 	 */
322 	if (!low_pfn(pmd_pfn(*pmd)) || !pmd_present(*pmd) || pmd_large(*pmd))
323 		goto out;
324 
325 	pte = pte_offset_kernel(pmd, address);
326 	pr_cont("*pte = %0*Lx ", sizeof(*pte) * 2, (u64)pte_val(*pte));
327 out:
328 	pr_cont("\n");
329 }
330 
331 #else /* CONFIG_X86_64: */
332 
333 #ifdef CONFIG_CPU_SUP_AMD
334 static const char errata93_warning[] =
335 KERN_ERR
336 "******* Your BIOS seems to not contain a fix for K8 errata #93\n"
337 "******* Working around it, but it may cause SEGVs or burn power.\n"
338 "******* Please consider a BIOS update.\n"
339 "******* Disabling USB legacy in the BIOS may also help.\n";
340 #endif
341 
342 static int bad_address(void *p)
343 {
344 	unsigned long dummy;
345 
346 	return get_kernel_nofault(dummy, (unsigned long *)p);
347 }
348 
349 static void dump_pagetable(unsigned long address)
350 {
351 	pgd_t *base = __va(read_cr3_pa());
352 	pgd_t *pgd = base + pgd_index(address);
353 	p4d_t *p4d;
354 	pud_t *pud;
355 	pmd_t *pmd;
356 	pte_t *pte;
357 
358 	if (bad_address(pgd))
359 		goto bad;
360 
361 	pr_info("PGD %lx ", pgd_val(*pgd));
362 
363 	if (!pgd_present(*pgd))
364 		goto out;
365 
366 	p4d = p4d_offset(pgd, address);
367 	if (bad_address(p4d))
368 		goto bad;
369 
370 	pr_cont("P4D %lx ", p4d_val(*p4d));
371 	if (!p4d_present(*p4d) || p4d_large(*p4d))
372 		goto out;
373 
374 	pud = pud_offset(p4d, address);
375 	if (bad_address(pud))
376 		goto bad;
377 
378 	pr_cont("PUD %lx ", pud_val(*pud));
379 	if (!pud_present(*pud) || pud_leaf(*pud))
380 		goto out;
381 
382 	pmd = pmd_offset(pud, address);
383 	if (bad_address(pmd))
384 		goto bad;
385 
386 	pr_cont("PMD %lx ", pmd_val(*pmd));
387 	if (!pmd_present(*pmd) || pmd_large(*pmd))
388 		goto out;
389 
390 	pte = pte_offset_kernel(pmd, address);
391 	if (bad_address(pte))
392 		goto bad;
393 
394 	pr_cont("PTE %lx", pte_val(*pte));
395 out:
396 	pr_cont("\n");
397 	return;
398 bad:
399 	pr_info("BAD\n");
400 }
401 
402 #endif /* CONFIG_X86_64 */
403 
404 /*
405  * Workaround for K8 erratum #93 & buggy BIOS.
406  *
407  * BIOS SMM functions are required to use a specific workaround
408  * to avoid corruption of the 64bit RIP register on C stepping K8.
409  *
410  * A lot of BIOS that didn't get tested properly miss this.
411  *
412  * The OS sees this as a page fault with the upper 32bits of RIP cleared.
413  * Try to work around it here.
414  *
415  * Note we only handle faults in kernel here.
416  * Does nothing on 32-bit.
417  */
418 static int is_errata93(struct pt_regs *regs, unsigned long address)
419 {
420 #if defined(CONFIG_X86_64) && defined(CONFIG_CPU_SUP_AMD)
421 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD
422 	    || boot_cpu_data.x86 != 0xf)
423 		return 0;
424 
425 	if (user_mode(regs))
426 		return 0;
427 
428 	if (address != regs->ip)
429 		return 0;
430 
431 	if ((address >> 32) != 0)
432 		return 0;
433 
434 	address |= 0xffffffffUL << 32;
435 	if ((address >= (u64)_stext && address <= (u64)_etext) ||
436 	    (address >= MODULES_VADDR && address <= MODULES_END)) {
437 		printk_once(errata93_warning);
438 		regs->ip = address;
439 		return 1;
440 	}
441 #endif
442 	return 0;
443 }
444 
445 /*
446  * Work around K8 erratum #100 K8 in compat mode occasionally jumps
447  * to illegal addresses >4GB.
448  *
449  * We catch this in the page fault handler because these addresses
450  * are not reachable. Just detect this case and return.  Any code
451  * segment in LDT is compatibility mode.
452  */
453 static int is_errata100(struct pt_regs *regs, unsigned long address)
454 {
455 #ifdef CONFIG_X86_64
456 	if ((regs->cs == __USER32_CS || (regs->cs & (1<<2))) && (address >> 32))
457 		return 1;
458 #endif
459 	return 0;
460 }
461 
462 /* Pentium F0 0F C7 C8 bug workaround: */
463 static int is_f00f_bug(struct pt_regs *regs, unsigned long error_code,
464 		       unsigned long address)
465 {
466 #ifdef CONFIG_X86_F00F_BUG
467 	if (boot_cpu_has_bug(X86_BUG_F00F) && !(error_code & X86_PF_USER) &&
468 	    idt_is_f00f_address(address)) {
469 		handle_invalid_op(regs);
470 		return 1;
471 	}
472 #endif
473 	return 0;
474 }
475 
476 static void show_ldttss(const struct desc_ptr *gdt, const char *name, u16 index)
477 {
478 	u32 offset = (index >> 3) * sizeof(struct desc_struct);
479 	unsigned long addr;
480 	struct ldttss_desc desc;
481 
482 	if (index == 0) {
483 		pr_alert("%s: NULL\n", name);
484 		return;
485 	}
486 
487 	if (offset + sizeof(struct ldttss_desc) >= gdt->size) {
488 		pr_alert("%s: 0x%hx -- out of bounds\n", name, index);
489 		return;
490 	}
491 
492 	if (copy_from_kernel_nofault(&desc, (void *)(gdt->address + offset),
493 			      sizeof(struct ldttss_desc))) {
494 		pr_alert("%s: 0x%hx -- GDT entry is not readable\n",
495 			 name, index);
496 		return;
497 	}
498 
499 	addr = desc.base0 | (desc.base1 << 16) | ((unsigned long)desc.base2 << 24);
500 #ifdef CONFIG_X86_64
501 	addr |= ((u64)desc.base3 << 32);
502 #endif
503 	pr_alert("%s: 0x%hx -- base=0x%lx limit=0x%x\n",
504 		 name, index, addr, (desc.limit0 | (desc.limit1 << 16)));
505 }
506 
507 static void
508 show_fault_oops(struct pt_regs *regs, unsigned long error_code, unsigned long address)
509 {
510 	if (!oops_may_print())
511 		return;
512 
513 	if (error_code & X86_PF_INSTR) {
514 		unsigned int level;
515 		pgd_t *pgd;
516 		pte_t *pte;
517 
518 		pgd = __va(read_cr3_pa());
519 		pgd += pgd_index(address);
520 
521 		pte = lookup_address_in_pgd(pgd, address, &level);
522 
523 		if (pte && pte_present(*pte) && !pte_exec(*pte))
524 			pr_crit("kernel tried to execute NX-protected page - exploit attempt? (uid: %d)\n",
525 				from_kuid(&init_user_ns, current_uid()));
526 		if (pte && pte_present(*pte) && pte_exec(*pte) &&
527 				(pgd_flags(*pgd) & _PAGE_USER) &&
528 				(__read_cr4() & X86_CR4_SMEP))
529 			pr_crit("unable to execute userspace code (SMEP?) (uid: %d)\n",
530 				from_kuid(&init_user_ns, current_uid()));
531 	}
532 
533 	if (address < PAGE_SIZE && !user_mode(regs))
534 		pr_alert("BUG: kernel NULL pointer dereference, address: %px\n",
535 			(void *)address);
536 	else
537 		pr_alert("BUG: unable to handle page fault for address: %px\n",
538 			(void *)address);
539 
540 	pr_alert("#PF: %s %s in %s mode\n",
541 		 (error_code & X86_PF_USER)  ? "user" : "supervisor",
542 		 (error_code & X86_PF_INSTR) ? "instruction fetch" :
543 		 (error_code & X86_PF_WRITE) ? "write access" :
544 					       "read access",
545 			     user_mode(regs) ? "user" : "kernel");
546 	pr_alert("#PF: error_code(0x%04lx) - %s\n", error_code,
547 		 !(error_code & X86_PF_PROT) ? "not-present page" :
548 		 (error_code & X86_PF_RSVD)  ? "reserved bit violation" :
549 		 (error_code & X86_PF_PK)    ? "protection keys violation" :
550 					       "permissions violation");
551 
552 	if (!(error_code & X86_PF_USER) && user_mode(regs)) {
553 		struct desc_ptr idt, gdt;
554 		u16 ldtr, tr;
555 
556 		/*
557 		 * This can happen for quite a few reasons.  The more obvious
558 		 * ones are faults accessing the GDT, or LDT.  Perhaps
559 		 * surprisingly, if the CPU tries to deliver a benign or
560 		 * contributory exception from user code and gets a page fault
561 		 * during delivery, the page fault can be delivered as though
562 		 * it originated directly from user code.  This could happen
563 		 * due to wrong permissions on the IDT, GDT, LDT, TSS, or
564 		 * kernel or IST stack.
565 		 */
566 		store_idt(&idt);
567 
568 		/* Usable even on Xen PV -- it's just slow. */
569 		native_store_gdt(&gdt);
570 
571 		pr_alert("IDT: 0x%lx (limit=0x%hx) GDT: 0x%lx (limit=0x%hx)\n",
572 			 idt.address, idt.size, gdt.address, gdt.size);
573 
574 		store_ldt(ldtr);
575 		show_ldttss(&gdt, "LDTR", ldtr);
576 
577 		store_tr(tr);
578 		show_ldttss(&gdt, "TR", tr);
579 	}
580 
581 	dump_pagetable(address);
582 }
583 
584 static noinline void
585 pgtable_bad(struct pt_regs *regs, unsigned long error_code,
586 	    unsigned long address)
587 {
588 	struct task_struct *tsk;
589 	unsigned long flags;
590 	int sig;
591 
592 	flags = oops_begin();
593 	tsk = current;
594 	sig = SIGKILL;
595 
596 	printk(KERN_ALERT "%s: Corrupted page table at address %lx\n",
597 	       tsk->comm, address);
598 	dump_pagetable(address);
599 
600 	if (__die("Bad pagetable", regs, error_code))
601 		sig = 0;
602 
603 	oops_end(flags, regs, sig);
604 }
605 
606 static void sanitize_error_code(unsigned long address,
607 				unsigned long *error_code)
608 {
609 	/*
610 	 * To avoid leaking information about the kernel page
611 	 * table layout, pretend that user-mode accesses to
612 	 * kernel addresses are always protection faults.
613 	 *
614 	 * NB: This means that failed vsyscalls with vsyscall=none
615 	 * will have the PROT bit.  This doesn't leak any
616 	 * information and does not appear to cause any problems.
617 	 */
618 	if (address >= TASK_SIZE_MAX)
619 		*error_code |= X86_PF_PROT;
620 }
621 
622 static void set_signal_archinfo(unsigned long address,
623 				unsigned long error_code)
624 {
625 	struct task_struct *tsk = current;
626 
627 	tsk->thread.trap_nr = X86_TRAP_PF;
628 	tsk->thread.error_code = error_code | X86_PF_USER;
629 	tsk->thread.cr2 = address;
630 }
631 
632 static noinline void
633 page_fault_oops(struct pt_regs *regs, unsigned long error_code,
634 		unsigned long address)
635 {
636 #ifdef CONFIG_VMAP_STACK
637 	struct stack_info info;
638 #endif
639 	unsigned long flags;
640 	int sig;
641 
642 	if (user_mode(regs)) {
643 		/*
644 		 * Implicit kernel access from user mode?  Skip the stack
645 		 * overflow and EFI special cases.
646 		 */
647 		goto oops;
648 	}
649 
650 #ifdef CONFIG_VMAP_STACK
651 	/*
652 	 * Stack overflow?  During boot, we can fault near the initial
653 	 * stack in the direct map, but that's not an overflow -- check
654 	 * that we're in vmalloc space to avoid this.
655 	 */
656 	if (is_vmalloc_addr((void *)address) &&
657 	    get_stack_guard_info((void *)address, &info)) {
658 		/*
659 		 * We're likely to be running with very little stack space
660 		 * left.  It's plausible that we'd hit this condition but
661 		 * double-fault even before we get this far, in which case
662 		 * we're fine: the double-fault handler will deal with it.
663 		 *
664 		 * We don't want to make it all the way into the oops code
665 		 * and then double-fault, though, because we're likely to
666 		 * break the console driver and lose most of the stack dump.
667 		 */
668 		call_on_stack(__this_cpu_ist_top_va(DF) - sizeof(void*),
669 			      handle_stack_overflow,
670 			      ASM_CALL_ARG3,
671 			      , [arg1] "r" (regs), [arg2] "r" (address), [arg3] "r" (&info));
672 
673 		unreachable();
674 	}
675 #endif
676 
677 	/*
678 	 * Buggy firmware could access regions which might page fault.  If
679 	 * this happens, EFI has a special OOPS path that will try to
680 	 * avoid hanging the system.
681 	 */
682 	if (IS_ENABLED(CONFIG_EFI))
683 		efi_crash_gracefully_on_page_fault(address);
684 
685 	/* Only not-present faults should be handled by KFENCE. */
686 	if (!(error_code & X86_PF_PROT) &&
687 	    kfence_handle_page_fault(address, error_code & X86_PF_WRITE, regs))
688 		return;
689 
690 oops:
691 	/*
692 	 * Oops. The kernel tried to access some bad page. We'll have to
693 	 * terminate things with extreme prejudice:
694 	 */
695 	flags = oops_begin();
696 
697 	show_fault_oops(regs, error_code, address);
698 
699 	if (task_stack_end_corrupted(current))
700 		printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
701 
702 	sig = SIGKILL;
703 	if (__die("Oops", regs, error_code))
704 		sig = 0;
705 
706 	/* Executive summary in case the body of the oops scrolled away */
707 	printk(KERN_DEFAULT "CR2: %016lx\n", address);
708 
709 	oops_end(flags, regs, sig);
710 }
711 
712 static noinline void
713 kernelmode_fixup_or_oops(struct pt_regs *regs, unsigned long error_code,
714 			 unsigned long address, int signal, int si_code,
715 			 u32 pkey)
716 {
717 	WARN_ON_ONCE(user_mode(regs));
718 
719 	/* Are we prepared to handle this kernel fault? */
720 	if (fixup_exception(regs, X86_TRAP_PF, error_code, address))
721 		return;
722 
723 	/*
724 	 * AMD erratum #91 manifests as a spurious page fault on a PREFETCH
725 	 * instruction.
726 	 */
727 	if (is_prefetch(regs, error_code, address))
728 		return;
729 
730 	page_fault_oops(regs, error_code, address);
731 }
732 
733 /*
734  * Print out info about fatal segfaults, if the show_unhandled_signals
735  * sysctl is set:
736  */
737 static inline void
738 show_signal_msg(struct pt_regs *regs, unsigned long error_code,
739 		unsigned long address, struct task_struct *tsk)
740 {
741 	const char *loglvl = task_pid_nr(tsk) > 1 ? KERN_INFO : KERN_EMERG;
742 	/* This is a racy snapshot, but it's better than nothing. */
743 	int cpu = raw_smp_processor_id();
744 
745 	if (!unhandled_signal(tsk, SIGSEGV))
746 		return;
747 
748 	if (!printk_ratelimit())
749 		return;
750 
751 	printk("%s%s[%d]: segfault at %lx ip %px sp %px error %lx",
752 		loglvl, tsk->comm, task_pid_nr(tsk), address,
753 		(void *)regs->ip, (void *)regs->sp, error_code);
754 
755 	print_vma_addr(KERN_CONT " in ", regs->ip);
756 
757 	/*
758 	 * Dump the likely CPU where the fatal segfault happened.
759 	 * This can help identify faulty hardware.
760 	 */
761 	printk(KERN_CONT " likely on CPU %d (core %d, socket %d)", cpu,
762 	       topology_core_id(cpu), topology_physical_package_id(cpu));
763 
764 
765 	printk(KERN_CONT "\n");
766 
767 	show_opcodes(regs, loglvl);
768 }
769 
770 static void
771 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
772 		       unsigned long address, u32 pkey, int si_code)
773 {
774 	struct task_struct *tsk = current;
775 
776 	if (!user_mode(regs)) {
777 		kernelmode_fixup_or_oops(regs, error_code, address,
778 					 SIGSEGV, si_code, pkey);
779 		return;
780 	}
781 
782 	if (!(error_code & X86_PF_USER)) {
783 		/* Implicit user access to kernel memory -- just oops */
784 		page_fault_oops(regs, error_code, address);
785 		return;
786 	}
787 
788 	/*
789 	 * User mode accesses just cause a SIGSEGV.
790 	 * It's possible to have interrupts off here:
791 	 */
792 	local_irq_enable();
793 
794 	/*
795 	 * Valid to do another page fault here because this one came
796 	 * from user space:
797 	 */
798 	if (is_prefetch(regs, error_code, address))
799 		return;
800 
801 	if (is_errata100(regs, address))
802 		return;
803 
804 	sanitize_error_code(address, &error_code);
805 
806 	if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
807 		return;
808 
809 	if (likely(show_unhandled_signals))
810 		show_signal_msg(regs, error_code, address, tsk);
811 
812 	set_signal_archinfo(address, error_code);
813 
814 	if (si_code == SEGV_PKUERR)
815 		force_sig_pkuerr((void __user *)address, pkey);
816 	else
817 		force_sig_fault(SIGSEGV, si_code, (void __user *)address);
818 
819 	local_irq_disable();
820 }
821 
822 static noinline void
823 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
824 		     unsigned long address)
825 {
826 	__bad_area_nosemaphore(regs, error_code, address, 0, SEGV_MAPERR);
827 }
828 
829 static void
830 __bad_area(struct pt_regs *regs, unsigned long error_code,
831 	   unsigned long address, u32 pkey, int si_code)
832 {
833 	struct mm_struct *mm = current->mm;
834 	/*
835 	 * Something tried to access memory that isn't in our memory map..
836 	 * Fix it, but check if it's kernel or user first..
837 	 */
838 	mmap_read_unlock(mm);
839 
840 	__bad_area_nosemaphore(regs, error_code, address, pkey, si_code);
841 }
842 
843 static inline bool bad_area_access_from_pkeys(unsigned long error_code,
844 		struct vm_area_struct *vma)
845 {
846 	/* This code is always called on the current mm */
847 	bool foreign = false;
848 
849 	if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
850 		return false;
851 	if (error_code & X86_PF_PK)
852 		return true;
853 	/* this checks permission keys on the VMA: */
854 	if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
855 				       (error_code & X86_PF_INSTR), foreign))
856 		return true;
857 	return false;
858 }
859 
860 static noinline void
861 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
862 		      unsigned long address, struct vm_area_struct *vma)
863 {
864 	/*
865 	 * This OSPKE check is not strictly necessary at runtime.
866 	 * But, doing it this way allows compiler optimizations
867 	 * if pkeys are compiled out.
868 	 */
869 	if (bad_area_access_from_pkeys(error_code, vma)) {
870 		/*
871 		 * A protection key fault means that the PKRU value did not allow
872 		 * access to some PTE.  Userspace can figure out what PKRU was
873 		 * from the XSAVE state.  This function captures the pkey from
874 		 * the vma and passes it to userspace so userspace can discover
875 		 * which protection key was set on the PTE.
876 		 *
877 		 * If we get here, we know that the hardware signaled a X86_PF_PK
878 		 * fault and that there was a VMA once we got in the fault
879 		 * handler.  It does *not* guarantee that the VMA we find here
880 		 * was the one that we faulted on.
881 		 *
882 		 * 1. T1   : mprotect_key(foo, PAGE_SIZE, pkey=4);
883 		 * 2. T1   : set PKRU to deny access to pkey=4, touches page
884 		 * 3. T1   : faults...
885 		 * 4.    T2: mprotect_key(foo, PAGE_SIZE, pkey=5);
886 		 * 5. T1   : enters fault handler, takes mmap_lock, etc...
887 		 * 6. T1   : reaches here, sees vma_pkey(vma)=5, when we really
888 		 *	     faulted on a pte with its pkey=4.
889 		 */
890 		u32 pkey = vma_pkey(vma);
891 
892 		__bad_area(regs, error_code, address, pkey, SEGV_PKUERR);
893 	} else {
894 		__bad_area(regs, error_code, address, 0, SEGV_ACCERR);
895 	}
896 }
897 
898 static void
899 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address,
900 	  vm_fault_t fault)
901 {
902 	/* Kernel mode? Handle exceptions or die: */
903 	if (!user_mode(regs)) {
904 		kernelmode_fixup_or_oops(regs, error_code, address,
905 					 SIGBUS, BUS_ADRERR, ARCH_DEFAULT_PKEY);
906 		return;
907 	}
908 
909 	/* User-space => ok to do another page fault: */
910 	if (is_prefetch(regs, error_code, address))
911 		return;
912 
913 	sanitize_error_code(address, &error_code);
914 
915 	if (fixup_vdso_exception(regs, X86_TRAP_PF, error_code, address))
916 		return;
917 
918 	set_signal_archinfo(address, error_code);
919 
920 #ifdef CONFIG_MEMORY_FAILURE
921 	if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
922 		struct task_struct *tsk = current;
923 		unsigned lsb = 0;
924 
925 		pr_err(
926 	"MCE: Killing %s:%d due to hardware memory corruption fault at %lx\n",
927 			tsk->comm, tsk->pid, address);
928 		if (fault & VM_FAULT_HWPOISON_LARGE)
929 			lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
930 		if (fault & VM_FAULT_HWPOISON)
931 			lsb = PAGE_SHIFT;
932 		force_sig_mceerr(BUS_MCEERR_AR, (void __user *)address, lsb);
933 		return;
934 	}
935 #endif
936 	force_sig_fault(SIGBUS, BUS_ADRERR, (void __user *)address);
937 }
938 
939 static int spurious_kernel_fault_check(unsigned long error_code, pte_t *pte)
940 {
941 	if ((error_code & X86_PF_WRITE) && !pte_write(*pte))
942 		return 0;
943 
944 	if ((error_code & X86_PF_INSTR) && !pte_exec(*pte))
945 		return 0;
946 
947 	return 1;
948 }
949 
950 /*
951  * Handle a spurious fault caused by a stale TLB entry.
952  *
953  * This allows us to lazily refresh the TLB when increasing the
954  * permissions of a kernel page (RO -> RW or NX -> X).  Doing it
955  * eagerly is very expensive since that implies doing a full
956  * cross-processor TLB flush, even if no stale TLB entries exist
957  * on other processors.
958  *
959  * Spurious faults may only occur if the TLB contains an entry with
960  * fewer permission than the page table entry.  Non-present (P = 0)
961  * and reserved bit (R = 1) faults are never spurious.
962  *
963  * There are no security implications to leaving a stale TLB when
964  * increasing the permissions on a page.
965  *
966  * Returns non-zero if a spurious fault was handled, zero otherwise.
967  *
968  * See Intel Developer's Manual Vol 3 Section 4.10.4.3, bullet 3
969  * (Optional Invalidation).
970  */
971 static noinline int
972 spurious_kernel_fault(unsigned long error_code, unsigned long address)
973 {
974 	pgd_t *pgd;
975 	p4d_t *p4d;
976 	pud_t *pud;
977 	pmd_t *pmd;
978 	pte_t *pte;
979 	int ret;
980 
981 	/*
982 	 * Only writes to RO or instruction fetches from NX may cause
983 	 * spurious faults.
984 	 *
985 	 * These could be from user or supervisor accesses but the TLB
986 	 * is only lazily flushed after a kernel mapping protection
987 	 * change, so user accesses are not expected to cause spurious
988 	 * faults.
989 	 */
990 	if (error_code != (X86_PF_WRITE | X86_PF_PROT) &&
991 	    error_code != (X86_PF_INSTR | X86_PF_PROT))
992 		return 0;
993 
994 	pgd = init_mm.pgd + pgd_index(address);
995 	if (!pgd_present(*pgd))
996 		return 0;
997 
998 	p4d = p4d_offset(pgd, address);
999 	if (!p4d_present(*p4d))
1000 		return 0;
1001 
1002 	if (p4d_large(*p4d))
1003 		return spurious_kernel_fault_check(error_code, (pte_t *) p4d);
1004 
1005 	pud = pud_offset(p4d, address);
1006 	if (!pud_present(*pud))
1007 		return 0;
1008 
1009 	if (pud_leaf(*pud))
1010 		return spurious_kernel_fault_check(error_code, (pte_t *) pud);
1011 
1012 	pmd = pmd_offset(pud, address);
1013 	if (!pmd_present(*pmd))
1014 		return 0;
1015 
1016 	if (pmd_large(*pmd))
1017 		return spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1018 
1019 	pte = pte_offset_kernel(pmd, address);
1020 	if (!pte_present(*pte))
1021 		return 0;
1022 
1023 	ret = spurious_kernel_fault_check(error_code, pte);
1024 	if (!ret)
1025 		return 0;
1026 
1027 	/*
1028 	 * Make sure we have permissions in PMD.
1029 	 * If not, then there's a bug in the page tables:
1030 	 */
1031 	ret = spurious_kernel_fault_check(error_code, (pte_t *) pmd);
1032 	WARN_ONCE(!ret, "PMD has incorrect permission bits\n");
1033 
1034 	return ret;
1035 }
1036 NOKPROBE_SYMBOL(spurious_kernel_fault);
1037 
1038 int show_unhandled_signals = 1;
1039 
1040 static inline int
1041 access_error(unsigned long error_code, struct vm_area_struct *vma)
1042 {
1043 	/* This is only called for the current mm, so: */
1044 	bool foreign = false;
1045 
1046 	/*
1047 	 * Read or write was blocked by protection keys.  This is
1048 	 * always an unconditional error and can never result in
1049 	 * a follow-up action to resolve the fault, like a COW.
1050 	 */
1051 	if (error_code & X86_PF_PK)
1052 		return 1;
1053 
1054 	/*
1055 	 * SGX hardware blocked the access.  This usually happens
1056 	 * when the enclave memory contents have been destroyed, like
1057 	 * after a suspend/resume cycle. In any case, the kernel can't
1058 	 * fix the cause of the fault.  Handle the fault as an access
1059 	 * error even in cases where no actual access violation
1060 	 * occurred.  This allows userspace to rebuild the enclave in
1061 	 * response to the signal.
1062 	 */
1063 	if (unlikely(error_code & X86_PF_SGX))
1064 		return 1;
1065 
1066 	/*
1067 	 * Make sure to check the VMA so that we do not perform
1068 	 * faults just to hit a X86_PF_PK as soon as we fill in a
1069 	 * page.
1070 	 */
1071 	if (!arch_vma_access_permitted(vma, (error_code & X86_PF_WRITE),
1072 				       (error_code & X86_PF_INSTR), foreign))
1073 		return 1;
1074 
1075 	/*
1076 	 * Shadow stack accesses (PF_SHSTK=1) are only permitted to
1077 	 * shadow stack VMAs. All other accesses result in an error.
1078 	 */
1079 	if (error_code & X86_PF_SHSTK) {
1080 		if (unlikely(!(vma->vm_flags & VM_SHADOW_STACK)))
1081 			return 1;
1082 		if (unlikely(!(vma->vm_flags & VM_WRITE)))
1083 			return 1;
1084 		return 0;
1085 	}
1086 
1087 	if (error_code & X86_PF_WRITE) {
1088 		/* write, present and write, not present: */
1089 		if (unlikely(vma->vm_flags & VM_SHADOW_STACK))
1090 			return 1;
1091 		if (unlikely(!(vma->vm_flags & VM_WRITE)))
1092 			return 1;
1093 		return 0;
1094 	}
1095 
1096 	/* read, present: */
1097 	if (unlikely(error_code & X86_PF_PROT))
1098 		return 1;
1099 
1100 	/* read, not present: */
1101 	if (unlikely(!vma_is_accessible(vma)))
1102 		return 1;
1103 
1104 	return 0;
1105 }
1106 
1107 bool fault_in_kernel_space(unsigned long address)
1108 {
1109 	/*
1110 	 * On 64-bit systems, the vsyscall page is at an address above
1111 	 * TASK_SIZE_MAX, but is not considered part of the kernel
1112 	 * address space.
1113 	 */
1114 	if (IS_ENABLED(CONFIG_X86_64) && is_vsyscall_vaddr(address))
1115 		return false;
1116 
1117 	return address >= TASK_SIZE_MAX;
1118 }
1119 
1120 /*
1121  * Called for all faults where 'address' is part of the kernel address
1122  * space.  Might get called for faults that originate from *code* that
1123  * ran in userspace or the kernel.
1124  */
1125 static void
1126 do_kern_addr_fault(struct pt_regs *regs, unsigned long hw_error_code,
1127 		   unsigned long address)
1128 {
1129 	/*
1130 	 * Protection keys exceptions only happen on user pages.  We
1131 	 * have no user pages in the kernel portion of the address
1132 	 * space, so do not expect them here.
1133 	 */
1134 	WARN_ON_ONCE(hw_error_code & X86_PF_PK);
1135 
1136 #ifdef CONFIG_X86_32
1137 	/*
1138 	 * We can fault-in kernel-space virtual memory on-demand. The
1139 	 * 'reference' page table is init_mm.pgd.
1140 	 *
1141 	 * NOTE! We MUST NOT take any locks for this case. We may
1142 	 * be in an interrupt or a critical region, and should
1143 	 * only copy the information from the master page table,
1144 	 * nothing more.
1145 	 *
1146 	 * Before doing this on-demand faulting, ensure that the
1147 	 * fault is not any of the following:
1148 	 * 1. A fault on a PTE with a reserved bit set.
1149 	 * 2. A fault caused by a user-mode access.  (Do not demand-
1150 	 *    fault kernel memory due to user-mode accesses).
1151 	 * 3. A fault caused by a page-level protection violation.
1152 	 *    (A demand fault would be on a non-present page which
1153 	 *     would have X86_PF_PROT==0).
1154 	 *
1155 	 * This is only needed to close a race condition on x86-32 in
1156 	 * the vmalloc mapping/unmapping code. See the comment above
1157 	 * vmalloc_fault() for details. On x86-64 the race does not
1158 	 * exist as the vmalloc mappings don't need to be synchronized
1159 	 * there.
1160 	 */
1161 	if (!(hw_error_code & (X86_PF_RSVD | X86_PF_USER | X86_PF_PROT))) {
1162 		if (vmalloc_fault(address) >= 0)
1163 			return;
1164 	}
1165 #endif
1166 
1167 	if (is_f00f_bug(regs, hw_error_code, address))
1168 		return;
1169 
1170 	/* Was the fault spurious, caused by lazy TLB invalidation? */
1171 	if (spurious_kernel_fault(hw_error_code, address))
1172 		return;
1173 
1174 	/* kprobes don't want to hook the spurious faults: */
1175 	if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1176 		return;
1177 
1178 	/*
1179 	 * Note, despite being a "bad area", there are quite a few
1180 	 * acceptable reasons to get here, such as erratum fixups
1181 	 * and handling kernel code that can fault, like get_user().
1182 	 *
1183 	 * Don't take the mm semaphore here. If we fixup a prefetch
1184 	 * fault we could otherwise deadlock:
1185 	 */
1186 	bad_area_nosemaphore(regs, hw_error_code, address);
1187 }
1188 NOKPROBE_SYMBOL(do_kern_addr_fault);
1189 
1190 /*
1191  * Handle faults in the user portion of the address space.  Nothing in here
1192  * should check X86_PF_USER without a specific justification: for almost
1193  * all purposes, we should treat a normal kernel access to user memory
1194  * (e.g. get_user(), put_user(), etc.) the same as the WRUSS instruction.
1195  * The one exception is AC flag handling, which is, per the x86
1196  * architecture, special for WRUSS.
1197  */
1198 static inline
1199 void do_user_addr_fault(struct pt_regs *regs,
1200 			unsigned long error_code,
1201 			unsigned long address)
1202 {
1203 	struct vm_area_struct *vma;
1204 	struct task_struct *tsk;
1205 	struct mm_struct *mm;
1206 	vm_fault_t fault;
1207 	unsigned int flags = FAULT_FLAG_DEFAULT;
1208 
1209 	tsk = current;
1210 	mm = tsk->mm;
1211 
1212 	if (unlikely((error_code & (X86_PF_USER | X86_PF_INSTR)) == X86_PF_INSTR)) {
1213 		/*
1214 		 * Whoops, this is kernel mode code trying to execute from
1215 		 * user memory.  Unless this is AMD erratum #93, which
1216 		 * corrupts RIP such that it looks like a user address,
1217 		 * this is unrecoverable.  Don't even try to look up the
1218 		 * VMA or look for extable entries.
1219 		 */
1220 		if (is_errata93(regs, address))
1221 			return;
1222 
1223 		page_fault_oops(regs, error_code, address);
1224 		return;
1225 	}
1226 
1227 	/* kprobes don't want to hook the spurious faults: */
1228 	if (WARN_ON_ONCE(kprobe_page_fault(regs, X86_TRAP_PF)))
1229 		return;
1230 
1231 	/*
1232 	 * Reserved bits are never expected to be set on
1233 	 * entries in the user portion of the page tables.
1234 	 */
1235 	if (unlikely(error_code & X86_PF_RSVD))
1236 		pgtable_bad(regs, error_code, address);
1237 
1238 	/*
1239 	 * If SMAP is on, check for invalid kernel (supervisor) access to user
1240 	 * pages in the user address space.  The odd case here is WRUSS,
1241 	 * which, according to the preliminary documentation, does not respect
1242 	 * SMAP and will have the USER bit set so, in all cases, SMAP
1243 	 * enforcement appears to be consistent with the USER bit.
1244 	 */
1245 	if (unlikely(cpu_feature_enabled(X86_FEATURE_SMAP) &&
1246 		     !(error_code & X86_PF_USER) &&
1247 		     !(regs->flags & X86_EFLAGS_AC))) {
1248 		/*
1249 		 * No extable entry here.  This was a kernel access to an
1250 		 * invalid pointer.  get_kernel_nofault() will not get here.
1251 		 */
1252 		page_fault_oops(regs, error_code, address);
1253 		return;
1254 	}
1255 
1256 	/*
1257 	 * If we're in an interrupt, have no user context or are running
1258 	 * in a region with pagefaults disabled then we must not take the fault
1259 	 */
1260 	if (unlikely(faulthandler_disabled() || !mm)) {
1261 		bad_area_nosemaphore(regs, error_code, address);
1262 		return;
1263 	}
1264 
1265 	/*
1266 	 * It's safe to allow irq's after cr2 has been saved and the
1267 	 * vmalloc fault has been handled.
1268 	 *
1269 	 * User-mode registers count as a user access even for any
1270 	 * potential system fault or CPU buglet:
1271 	 */
1272 	if (user_mode(regs)) {
1273 		local_irq_enable();
1274 		flags |= FAULT_FLAG_USER;
1275 	} else {
1276 		if (regs->flags & X86_EFLAGS_IF)
1277 			local_irq_enable();
1278 	}
1279 
1280 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
1281 
1282 	/*
1283 	 * Read-only permissions can not be expressed in shadow stack PTEs.
1284 	 * Treat all shadow stack accesses as WRITE faults. This ensures
1285 	 * that the MM will prepare everything (e.g., break COW) such that
1286 	 * maybe_mkwrite() can create a proper shadow stack PTE.
1287 	 */
1288 	if (error_code & X86_PF_SHSTK)
1289 		flags |= FAULT_FLAG_WRITE;
1290 	if (error_code & X86_PF_WRITE)
1291 		flags |= FAULT_FLAG_WRITE;
1292 	if (error_code & X86_PF_INSTR)
1293 		flags |= FAULT_FLAG_INSTRUCTION;
1294 
1295 #ifdef CONFIG_X86_64
1296 	/*
1297 	 * Faults in the vsyscall page might need emulation.  The
1298 	 * vsyscall page is at a high address (>PAGE_OFFSET), but is
1299 	 * considered to be part of the user address space.
1300 	 *
1301 	 * The vsyscall page does not have a "real" VMA, so do this
1302 	 * emulation before we go searching for VMAs.
1303 	 *
1304 	 * PKRU never rejects instruction fetches, so we don't need
1305 	 * to consider the PF_PK bit.
1306 	 */
1307 	if (is_vsyscall_vaddr(address)) {
1308 		if (emulate_vsyscall(error_code, regs, address))
1309 			return;
1310 	}
1311 #endif
1312 
1313 	if (!(flags & FAULT_FLAG_USER))
1314 		goto lock_mmap;
1315 
1316 	vma = lock_vma_under_rcu(mm, address);
1317 	if (!vma)
1318 		goto lock_mmap;
1319 
1320 	if (unlikely(access_error(error_code, vma))) {
1321 		vma_end_read(vma);
1322 		goto lock_mmap;
1323 	}
1324 	fault = handle_mm_fault(vma, address, flags | FAULT_FLAG_VMA_LOCK, regs);
1325 	if (!(fault & (VM_FAULT_RETRY | VM_FAULT_COMPLETED)))
1326 		vma_end_read(vma);
1327 
1328 	if (!(fault & VM_FAULT_RETRY)) {
1329 		count_vm_vma_lock_event(VMA_LOCK_SUCCESS);
1330 		goto done;
1331 	}
1332 	count_vm_vma_lock_event(VMA_LOCK_RETRY);
1333 
1334 	/* Quick path to respond to signals */
1335 	if (fault_signal_pending(fault, regs)) {
1336 		if (!user_mode(regs))
1337 			kernelmode_fixup_or_oops(regs, error_code, address,
1338 						 SIGBUS, BUS_ADRERR,
1339 						 ARCH_DEFAULT_PKEY);
1340 		return;
1341 	}
1342 lock_mmap:
1343 
1344 retry:
1345 	vma = lock_mm_and_find_vma(mm, address, regs);
1346 	if (unlikely(!vma)) {
1347 		bad_area_nosemaphore(regs, error_code, address);
1348 		return;
1349 	}
1350 
1351 	/*
1352 	 * Ok, we have a good vm_area for this memory access, so
1353 	 * we can handle it..
1354 	 */
1355 	if (unlikely(access_error(error_code, vma))) {
1356 		bad_area_access_error(regs, error_code, address, vma);
1357 		return;
1358 	}
1359 
1360 	/*
1361 	 * If for any reason at all we couldn't handle the fault,
1362 	 * make sure we exit gracefully rather than endlessly redo
1363 	 * the fault.  Since we never set FAULT_FLAG_RETRY_NOWAIT, if
1364 	 * we get VM_FAULT_RETRY back, the mmap_lock has been unlocked.
1365 	 *
1366 	 * Note that handle_userfault() may also release and reacquire mmap_lock
1367 	 * (and not return with VM_FAULT_RETRY), when returning to userland to
1368 	 * repeat the page fault later with a VM_FAULT_NOPAGE retval
1369 	 * (potentially after handling any pending signal during the return to
1370 	 * userland). The return to userland is identified whenever
1371 	 * FAULT_FLAG_USER|FAULT_FLAG_KILLABLE are both set in flags.
1372 	 */
1373 	fault = handle_mm_fault(vma, address, flags, regs);
1374 
1375 	if (fault_signal_pending(fault, regs)) {
1376 		/*
1377 		 * Quick path to respond to signals.  The core mm code
1378 		 * has unlocked the mm for us if we get here.
1379 		 */
1380 		if (!user_mode(regs))
1381 			kernelmode_fixup_or_oops(regs, error_code, address,
1382 						 SIGBUS, BUS_ADRERR,
1383 						 ARCH_DEFAULT_PKEY);
1384 		return;
1385 	}
1386 
1387 	/* The fault is fully completed (including releasing mmap lock) */
1388 	if (fault & VM_FAULT_COMPLETED)
1389 		return;
1390 
1391 	/*
1392 	 * If we need to retry the mmap_lock has already been released,
1393 	 * and if there is a fatal signal pending there is no guarantee
1394 	 * that we made any progress. Handle this case first.
1395 	 */
1396 	if (unlikely(fault & VM_FAULT_RETRY)) {
1397 		flags |= FAULT_FLAG_TRIED;
1398 		goto retry;
1399 	}
1400 
1401 	mmap_read_unlock(mm);
1402 done:
1403 	if (likely(!(fault & VM_FAULT_ERROR)))
1404 		return;
1405 
1406 	if (fatal_signal_pending(current) && !user_mode(regs)) {
1407 		kernelmode_fixup_or_oops(regs, error_code, address,
1408 					 0, 0, ARCH_DEFAULT_PKEY);
1409 		return;
1410 	}
1411 
1412 	if (fault & VM_FAULT_OOM) {
1413 		/* Kernel mode? Handle exceptions or die: */
1414 		if (!user_mode(regs)) {
1415 			kernelmode_fixup_or_oops(regs, error_code, address,
1416 						 SIGSEGV, SEGV_MAPERR,
1417 						 ARCH_DEFAULT_PKEY);
1418 			return;
1419 		}
1420 
1421 		/*
1422 		 * We ran out of memory, call the OOM killer, and return the
1423 		 * userspace (which will retry the fault, or kill us if we got
1424 		 * oom-killed):
1425 		 */
1426 		pagefault_out_of_memory();
1427 	} else {
1428 		if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
1429 			     VM_FAULT_HWPOISON_LARGE))
1430 			do_sigbus(regs, error_code, address, fault);
1431 		else if (fault & VM_FAULT_SIGSEGV)
1432 			bad_area_nosemaphore(regs, error_code, address);
1433 		else
1434 			BUG();
1435 	}
1436 }
1437 NOKPROBE_SYMBOL(do_user_addr_fault);
1438 
1439 static __always_inline void
1440 trace_page_fault_entries(struct pt_regs *regs, unsigned long error_code,
1441 			 unsigned long address)
1442 {
1443 	if (!trace_pagefault_enabled())
1444 		return;
1445 
1446 	if (user_mode(regs))
1447 		trace_page_fault_user(address, regs, error_code);
1448 	else
1449 		trace_page_fault_kernel(address, regs, error_code);
1450 }
1451 
1452 static __always_inline void
1453 handle_page_fault(struct pt_regs *regs, unsigned long error_code,
1454 			      unsigned long address)
1455 {
1456 	trace_page_fault_entries(regs, error_code, address);
1457 
1458 	if (unlikely(kmmio_fault(regs, address)))
1459 		return;
1460 
1461 	/* Was the fault on kernel-controlled part of the address space? */
1462 	if (unlikely(fault_in_kernel_space(address))) {
1463 		do_kern_addr_fault(regs, error_code, address);
1464 	} else {
1465 		do_user_addr_fault(regs, error_code, address);
1466 		/*
1467 		 * User address page fault handling might have reenabled
1468 		 * interrupts. Fixing up all potential exit points of
1469 		 * do_user_addr_fault() and its leaf functions is just not
1470 		 * doable w/o creating an unholy mess or turning the code
1471 		 * upside down.
1472 		 */
1473 		local_irq_disable();
1474 	}
1475 }
1476 
1477 DEFINE_IDTENTRY_RAW_ERRORCODE(exc_page_fault)
1478 {
1479 	unsigned long address = read_cr2();
1480 	irqentry_state_t state;
1481 
1482 	prefetchw(&current->mm->mmap_lock);
1483 
1484 	/*
1485 	 * KVM uses #PF vector to deliver 'page not present' events to guests
1486 	 * (asynchronous page fault mechanism). The event happens when a
1487 	 * userspace task is trying to access some valid (from guest's point of
1488 	 * view) memory which is not currently mapped by the host (e.g. the
1489 	 * memory is swapped out). Note, the corresponding "page ready" event
1490 	 * which is injected when the memory becomes available, is delivered via
1491 	 * an interrupt mechanism and not a #PF exception
1492 	 * (see arch/x86/kernel/kvm.c: sysvec_kvm_asyncpf_interrupt()).
1493 	 *
1494 	 * We are relying on the interrupted context being sane (valid RSP,
1495 	 * relevant locks not held, etc.), which is fine as long as the
1496 	 * interrupted context had IF=1.  We are also relying on the KVM
1497 	 * async pf type field and CR2 being read consistently instead of
1498 	 * getting values from real and async page faults mixed up.
1499 	 *
1500 	 * Fingers crossed.
1501 	 *
1502 	 * The async #PF handling code takes care of idtentry handling
1503 	 * itself.
1504 	 */
1505 	if (kvm_handle_async_pf(regs, (u32)address))
1506 		return;
1507 
1508 	/*
1509 	 * Entry handling for valid #PF from kernel mode is slightly
1510 	 * different: RCU is already watching and ct_irq_enter() must not
1511 	 * be invoked because a kernel fault on a user space address might
1512 	 * sleep.
1513 	 *
1514 	 * In case the fault hit a RCU idle region the conditional entry
1515 	 * code reenabled RCU to avoid subsequent wreckage which helps
1516 	 * debuggability.
1517 	 */
1518 	state = irqentry_enter(regs);
1519 
1520 	instrumentation_begin();
1521 	handle_page_fault(regs, error_code, address);
1522 	instrumentation_end();
1523 
1524 	irqentry_exit(regs, state);
1525 }
1526