xref: /linux/arch/x86/kernel/traps.c (revision 7fc2cd2e4b398c57c9cf961cfea05eadbf34c05c)
1 /*
2  *  Copyright (C) 1991, 1992  Linus Torvalds
3  *  Copyright (C) 2000, 2001, 2002 Andi Kleen, SuSE Labs
4  *
5  *  Pentium III FXSR, SSE support
6  *	Gareth Hughes <gareth@valinux.com>, May 2000
7  */
8 
9 /*
10  * Handle hardware traps and faults.
11  */
12 
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 
15 #include <linux/context_tracking.h>
16 #include <linux/interrupt.h>
17 #include <linux/kallsyms.h>
18 #include <linux/kmsan.h>
19 #include <linux/spinlock.h>
20 #include <linux/kprobes.h>
21 #include <linux/uaccess.h>
22 #include <linux/kdebug.h>
23 #include <linux/kgdb.h>
24 #include <linux/kernel.h>
25 #include <linux/export.h>
26 #include <linux/ptrace.h>
27 #include <linux/uprobes.h>
28 #include <linux/string.h>
29 #include <linux/delay.h>
30 #include <linux/errno.h>
31 #include <linux/kexec.h>
32 #include <linux/sched.h>
33 #include <linux/sched/task_stack.h>
34 #include <linux/static_call.h>
35 #include <linux/timer.h>
36 #include <linux/init.h>
37 #include <linux/bug.h>
38 #include <linux/nmi.h>
39 #include <linux/mm.h>
40 #include <linux/smp.h>
41 #include <linux/cpu.h>
42 #include <linux/io.h>
43 #include <linux/hardirq.h>
44 #include <linux/atomic.h>
45 #include <linux/iommu.h>
46 #include <linux/ubsan.h>
47 
48 #include <asm/stacktrace.h>
49 #include <asm/processor.h>
50 #include <asm/debugreg.h>
51 #include <asm/realmode.h>
52 #include <asm/text-patching.h>
53 #include <asm/ftrace.h>
54 #include <asm/traps.h>
55 #include <asm/desc.h>
56 #include <asm/fred.h>
57 #include <asm/fpu/api.h>
58 #include <asm/cpu.h>
59 #include <asm/cpu_entry_area.h>
60 #include <asm/mce.h>
61 #include <asm/fixmap.h>
62 #include <asm/mach_traps.h>
63 #include <asm/alternative.h>
64 #include <asm/fpu/xstate.h>
65 #include <asm/vm86.h>
66 #include <asm/umip.h>
67 #include <asm/insn.h>
68 #include <asm/insn-eval.h>
69 #include <asm/vdso.h>
70 #include <asm/tdx.h>
71 #include <asm/cfi.h>
72 #include <asm/msr.h>
73 
74 #ifdef CONFIG_X86_64
75 #include <asm/x86_init.h>
76 #else
77 #include <asm/processor-flags.h>
78 #include <asm/setup.h>
79 #endif
80 
81 #include <asm/proto.h>
82 
83 DECLARE_BITMAP(system_vectors, NR_VECTORS);
84 
85 __always_inline int is_valid_bugaddr(unsigned long addr)
86 {
87 	if (addr < TASK_SIZE_MAX)
88 		return 0;
89 
90 	/*
91 	 * We got #UD, if the text isn't readable we'd have gotten
92 	 * a different exception.
93 	 */
94 	return *(unsigned short *)addr == INSN_UD2;
95 }
96 
97 /*
98  * Check for UD1 or UD2, accounting for Address Size Override Prefixes.
99  * If it's a UD1, further decode to determine its use:
100  *
101  * FineIBT:      d6                      udb
102  * FineIBT:      f0 75 f9                lock jne . - 6
103  * UBSan{0}:     67 0f b9 00             ud1    (%eax),%eax
104  * UBSan{10}:    67 0f b9 40 10          ud1    0x10(%eax),%eax
105  * static_call:  0f b9 cc                ud1    %esp,%ecx
106  * __WARN_trap:  67 48 0f b9 3a          ud1    (%edx),%reg
107  *
108  * Notable, since __WARN_trap can use all registers, the distinction between
109  * UD1 users is through R/M.
110  */
111 __always_inline int decode_bug(unsigned long addr, s32 *imm, int *len)
112 {
113 	unsigned long start = addr;
114 	u8 v, reg, rm, rex = 0;
115 	int type = BUG_UD1;
116 	bool lock = false;
117 
118 	if (addr < TASK_SIZE_MAX)
119 		return BUG_NONE;
120 
121 	for (;;) {
122 		v = *(u8 *)(addr++);
123 		if (v == INSN_ASOP)
124 			continue;
125 
126 		if (v == INSN_LOCK) {
127 			lock = true;
128 			continue;
129 		}
130 
131 		if ((v & 0xf0) == 0x40) {
132 			rex = v;
133 			continue;
134 		}
135 
136 		break;
137 	}
138 
139 	switch (v) {
140 	case 0x70 ... 0x7f: /* Jcc.d8 */
141 		addr += 1; /* d8 */
142 		*len = addr - start;
143 		WARN_ON_ONCE(!lock);
144 		return BUG_LOCK;
145 
146 	case 0xd6:
147 		*len = addr - start;
148 		return BUG_UDB;
149 
150 	case OPCODE_ESCAPE:
151 		break;
152 
153 	default:
154 		return BUG_NONE;
155 	}
156 
157 	v = *(u8 *)(addr++);
158 	if (v == SECOND_BYTE_OPCODE_UD2) {
159 		*len = addr - start;
160 		return BUG_UD2;
161 	}
162 
163 	if (v != SECOND_BYTE_OPCODE_UD1)
164 		return BUG_NONE;
165 
166 	*imm = 0;
167 	v = *(u8 *)(addr++);		/* ModRM */
168 
169 	if (X86_MODRM_MOD(v) != 3 && X86_MODRM_RM(v) == 4)
170 		addr++;			/* SIB */
171 
172 	reg = X86_MODRM_REG(v) + 8*!!X86_REX_R(rex);
173 	rm  = X86_MODRM_RM(v)  + 8*!!X86_REX_B(rex);
174 
175 	/* Decode immediate, if present */
176 	switch (X86_MODRM_MOD(v)) {
177 	case 0: if (X86_MODRM_RM(v) == 5)
178 			addr += 4;	/* RIP + disp32 */
179 
180 		if (rm == 0)		/* (%eax) */
181 			type = BUG_UD1_UBSAN;
182 
183 		if (rm == 2) {		/* (%edx) */
184 			*imm = reg;
185 			type = BUG_UD1_WARN;
186 		}
187 		break;
188 
189 	case 1: *imm = *(s8 *)addr;
190 		addr += 1;
191 		if (rm == 0)		/* (%eax) */
192 			type = BUG_UD1_UBSAN;
193 		break;
194 
195 	case 2: *imm = *(s32 *)addr;
196 		addr += 4;
197 		if (rm == 0)		/* (%eax) */
198 			type = BUG_UD1_UBSAN;
199 		break;
200 
201 	case 3: break;
202 	}
203 
204 	/* record instruction length */
205 	*len = addr - start;
206 
207 	return type;
208 }
209 
210 static inline unsigned long pt_regs_val(struct pt_regs *regs, int nr)
211 {
212 	int offset = pt_regs_offset(regs, nr);
213 	if (WARN_ON_ONCE(offset < -0))
214 		return 0;
215 	return *((unsigned long *)((void *)regs + offset));
216 }
217 
218 #ifdef HAVE_ARCH_BUG_FORMAT_ARGS
219 DEFINE_STATIC_CALL(WARN_trap, __WARN_trap);
220 EXPORT_STATIC_CALL_TRAMP(WARN_trap);
221 
222 /*
223  * Create a va_list from an exception context.
224  */
225 void *__warn_args(struct arch_va_list *args, struct pt_regs *regs)
226 {
227 	/*
228 	 * Register save area; populate with function call argument registers
229 	 */
230 	args->regs[0] = regs->di;
231 	args->regs[1] = regs->si;
232 	args->regs[2] = regs->dx;
233 	args->regs[3] = regs->cx;
234 	args->regs[4] = regs->r8;
235 	args->regs[5] = regs->r9;
236 
237 	/*
238 	 * From the ABI document:
239 	 *
240 	 * @gp_offset - the element holds the offset in bytes from
241 	 * reg_save_area to the place where the next available general purpose
242 	 * argument register is saved. In case all argument registers have
243 	 * been exhausted, it is set to the value 48 (6*8).
244 	 *
245 	 * @fp_offset - the element holds the offset in bytes from
246 	 * reg_save_area to the place where the next available floating point
247 	 * argument is saved. In case all argument registers have been
248 	 * exhausted, it is set to the value 176 (6*8 + 8*16)
249 	 *
250 	 * @overflow_arg_area - this pointer is used to fetch arguments passed
251 	 * on the stack. It is initialized with the address of the first
252 	 * argument passed on the stack, if any, and then always updated to
253 	 * point to the start of the next argument on the stack.
254 	 *
255 	 * @reg_save_area - the element points to the start of the register
256 	 * save area.
257 	 *
258 	 * Notably the vararg starts with the second argument and there are no
259 	 * floating point arguments in the kernel.
260 	 */
261 	args->args.gp_offset = 1*8;
262 	args->args.fp_offset = 6*8 + 8*16;
263 	args->args.reg_save_area = &args->regs;
264 	args->args.overflow_arg_area = (void *)regs->sp;
265 
266 	/*
267 	 * If the exception came from __WARN_trap, there is a return
268 	 * address on the stack, skip that. This is why any __WARN_trap()
269 	 * caller must inhibit tail-call optimization.
270 	 */
271 	if ((void *)regs->ip == &__WARN_trap)
272 		args->args.overflow_arg_area += 8;
273 
274 	return &args->args;
275 }
276 #endif /* HAVE_ARCH_BUG_FORMAT */
277 
278 static nokprobe_inline int
279 do_trap_no_signal(struct task_struct *tsk, int trapnr, const char *str,
280 		  struct pt_regs *regs,	long error_code)
281 {
282 	if (v8086_mode(regs)) {
283 		/*
284 		 * Traps 0, 1, 3, 4, and 5 should be forwarded to vm86.
285 		 * On nmi (interrupt 2), do_trap should not be called.
286 		 */
287 		if (trapnr < X86_TRAP_UD) {
288 			if (!handle_vm86_trap((struct kernel_vm86_regs *) regs,
289 						error_code, trapnr))
290 				return 0;
291 		}
292 	} else if (!user_mode(regs)) {
293 		if (fixup_exception(regs, trapnr, error_code, 0))
294 			return 0;
295 
296 		tsk->thread.error_code = error_code;
297 		tsk->thread.trap_nr = trapnr;
298 		die(str, regs, error_code);
299 	} else {
300 		if (fixup_vdso_exception(regs, trapnr, error_code, 0))
301 			return 0;
302 	}
303 
304 	/*
305 	 * We want error_code and trap_nr set for userspace faults and
306 	 * kernelspace faults which result in die(), but not
307 	 * kernelspace faults which are fixed up.  die() gives the
308 	 * process no chance to handle the signal and notice the
309 	 * kernel fault information, so that won't result in polluting
310 	 * the information about previously queued, but not yet
311 	 * delivered, faults.  See also exc_general_protection below.
312 	 */
313 	tsk->thread.error_code = error_code;
314 	tsk->thread.trap_nr = trapnr;
315 
316 	return -1;
317 }
318 
319 static void show_signal(struct task_struct *tsk, int signr,
320 			const char *type, const char *desc,
321 			struct pt_regs *regs, long error_code)
322 {
323 	if (show_unhandled_signals && unhandled_signal(tsk, signr) &&
324 	    printk_ratelimit()) {
325 		pr_info("%s[%d] %s%s ip:%lx sp:%lx error:%lx",
326 			tsk->comm, task_pid_nr(tsk), type, desc,
327 			regs->ip, regs->sp, error_code);
328 		print_vma_addr(KERN_CONT " in ", regs->ip);
329 		pr_cont("\n");
330 	}
331 }
332 
333 static void
334 do_trap(int trapnr, int signr, char *str, struct pt_regs *regs,
335 	long error_code, int sicode, void __user *addr)
336 {
337 	struct task_struct *tsk = current;
338 
339 	if (!do_trap_no_signal(tsk, trapnr, str, regs, error_code))
340 		return;
341 
342 	show_signal(tsk, signr, "trap ", str, regs, error_code);
343 
344 	if (!sicode)
345 		force_sig(signr);
346 	else
347 		force_sig_fault(signr, sicode, addr);
348 }
349 NOKPROBE_SYMBOL(do_trap);
350 
351 static void do_error_trap(struct pt_regs *regs, long error_code, char *str,
352 	unsigned long trapnr, int signr, int sicode, void __user *addr)
353 {
354 	RCU_LOCKDEP_WARN(!rcu_is_watching(), "entry code didn't wake RCU");
355 
356 	if (notify_die(DIE_TRAP, str, regs, error_code, trapnr, signr) !=
357 			NOTIFY_STOP) {
358 		cond_local_irq_enable(regs);
359 		do_trap(trapnr, signr, str, regs, error_code, sicode, addr);
360 		cond_local_irq_disable(regs);
361 	}
362 }
363 
364 /*
365  * Posix requires to provide the address of the faulting instruction for
366  * SIGILL (#UD) and SIGFPE (#DE) in the si_addr member of siginfo_t.
367  *
368  * This address is usually regs->ip, but when an uprobe moved the code out
369  * of line then regs->ip points to the XOL code which would confuse
370  * anything which analyzes the fault address vs. the unmodified binary. If
371  * a trap happened in XOL code then uprobe maps regs->ip back to the
372  * original instruction address.
373  */
374 static __always_inline void __user *error_get_trap_addr(struct pt_regs *regs)
375 {
376 	return (void __user *)uprobe_get_trap_addr(regs);
377 }
378 
379 DEFINE_IDTENTRY(exc_divide_error)
380 {
381 	do_error_trap(regs, 0, "divide error", X86_TRAP_DE, SIGFPE,
382 		      FPE_INTDIV, error_get_trap_addr(regs));
383 }
384 
385 DEFINE_IDTENTRY(exc_overflow)
386 {
387 	do_error_trap(regs, 0, "overflow", X86_TRAP_OF, SIGSEGV, 0, NULL);
388 }
389 
390 #ifdef CONFIG_X86_F00F_BUG
391 void handle_invalid_op(struct pt_regs *regs)
392 #else
393 static inline void handle_invalid_op(struct pt_regs *regs)
394 #endif
395 {
396 	do_error_trap(regs, 0, "invalid opcode", X86_TRAP_UD, SIGILL,
397 		      ILL_ILLOPN, error_get_trap_addr(regs));
398 }
399 
400 static noinstr bool handle_bug(struct pt_regs *regs)
401 {
402 	unsigned long addr = regs->ip;
403 	bool handled = false;
404 	int ud_type, ud_len;
405 	s32 ud_imm;
406 
407 	ud_type = decode_bug(addr, &ud_imm, &ud_len);
408 	if (ud_type == BUG_NONE)
409 		return handled;
410 
411 	/*
412 	 * All lies, just get the WARN/BUG out.
413 	 */
414 	instrumentation_begin();
415 	/*
416 	 * Normally @regs are unpoisoned by irqentry_enter(), but handle_bug()
417 	 * is a rare case that uses @regs without passing them to
418 	 * irqentry_enter().
419 	 */
420 	kmsan_unpoison_entry_regs(regs);
421 	/*
422 	 * Since we're emulating a CALL with exceptions, restore the interrupt
423 	 * state to what it was at the exception site.
424 	 */
425 	if (regs->flags & X86_EFLAGS_IF)
426 		raw_local_irq_enable();
427 
428 	switch (ud_type) {
429 	case BUG_UD1_WARN:
430 		if (report_bug_entry((void *)pt_regs_val(regs, ud_imm), regs) == BUG_TRAP_TYPE_WARN)
431 			handled = true;
432 		break;
433 
434 	case BUG_UD2:
435 		if (report_bug(regs->ip, regs) == BUG_TRAP_TYPE_WARN) {
436 			handled = true;
437 			break;
438 		}
439 		fallthrough;
440 
441 	case BUG_UDB:
442 	case BUG_LOCK:
443 		if (handle_cfi_failure(regs) == BUG_TRAP_TYPE_WARN) {
444 			handled = true;
445 			break;
446 		}
447 		break;
448 
449 	case BUG_UD1_UBSAN:
450 		if (IS_ENABLED(CONFIG_UBSAN_TRAP)) {
451 			pr_crit("%s at %pS\n",
452 				report_ubsan_failure(ud_imm),
453 				(void *)regs->ip);
454 		}
455 		break;
456 
457 	default:
458 		break;
459 	}
460 
461 	/*
462 	 * When continuing, and regs->ip hasn't changed, move it to the next
463 	 * instruction. When not continuing execution, restore the instruction
464 	 * pointer.
465 	 */
466 	if (handled) {
467 		if (regs->ip == addr)
468 			regs->ip += ud_len;
469 	} else {
470 		regs->ip = addr;
471 	}
472 
473 	if (regs->flags & X86_EFLAGS_IF)
474 		raw_local_irq_disable();
475 	instrumentation_end();
476 
477 	return handled;
478 }
479 
480 DEFINE_IDTENTRY_RAW(exc_invalid_op)
481 {
482 	irqentry_state_t state;
483 
484 	/*
485 	 * We use UD2 as a short encoding for 'CALL __WARN', as such
486 	 * handle it before exception entry to avoid recursive WARN
487 	 * in case exception entry is the one triggering WARNs.
488 	 */
489 	if (!user_mode(regs) && handle_bug(regs))
490 		return;
491 
492 	state = irqentry_enter(regs);
493 	instrumentation_begin();
494 	handle_invalid_op(regs);
495 	instrumentation_end();
496 	irqentry_exit(regs, state);
497 }
498 
499 DEFINE_IDTENTRY(exc_coproc_segment_overrun)
500 {
501 	do_error_trap(regs, 0, "coprocessor segment overrun",
502 		      X86_TRAP_OLD_MF, SIGFPE, 0, NULL);
503 }
504 
505 DEFINE_IDTENTRY_ERRORCODE(exc_invalid_tss)
506 {
507 	do_error_trap(regs, error_code, "invalid TSS", X86_TRAP_TS, SIGSEGV,
508 		      0, NULL);
509 }
510 
511 DEFINE_IDTENTRY_ERRORCODE(exc_segment_not_present)
512 {
513 	do_error_trap(regs, error_code, "segment not present", X86_TRAP_NP,
514 		      SIGBUS, 0, NULL);
515 }
516 
517 DEFINE_IDTENTRY_ERRORCODE(exc_stack_segment)
518 {
519 	do_error_trap(regs, error_code, "stack segment", X86_TRAP_SS, SIGBUS,
520 		      0, NULL);
521 }
522 
523 DEFINE_IDTENTRY_ERRORCODE(exc_alignment_check)
524 {
525 	char *str = "alignment check";
526 
527 	if (notify_die(DIE_TRAP, str, regs, error_code, X86_TRAP_AC, SIGBUS) == NOTIFY_STOP)
528 		return;
529 
530 	if (!user_mode(regs))
531 		die("Split lock detected\n", regs, error_code);
532 
533 	local_irq_enable();
534 
535 	if (handle_user_split_lock(regs, error_code))
536 		goto out;
537 
538 	do_trap(X86_TRAP_AC, SIGBUS, "alignment check", regs,
539 		error_code, BUS_ADRALN, NULL);
540 
541 out:
542 	local_irq_disable();
543 }
544 
545 #ifdef CONFIG_VMAP_STACK
546 __visible void __noreturn handle_stack_overflow(struct pt_regs *regs,
547 						unsigned long fault_address,
548 						struct stack_info *info)
549 {
550 	const char *name = stack_type_name(info->type);
551 
552 	printk(KERN_EMERG "BUG: %s stack guard page was hit at %p (stack is %p..%p)\n",
553 	       name, (void *)fault_address, info->begin, info->end);
554 
555 	die("stack guard page", regs, 0);
556 
557 	/* Be absolutely certain we don't return. */
558 	panic("%s stack guard hit", name);
559 }
560 #endif
561 
562 /*
563  * Prevent the compiler and/or objtool from marking the !CONFIG_X86_ESPFIX64
564  * version of exc_double_fault() as noreturn.  Otherwise the noreturn mismatch
565  * between configs triggers objtool warnings.
566  *
567  * This is a temporary hack until we have compiler or plugin support for
568  * annotating noreturns.
569  */
570 #ifdef CONFIG_X86_ESPFIX64
571 #define always_true() true
572 #else
573 bool always_true(void);
574 bool __weak always_true(void) { return true; }
575 #endif
576 
577 /*
578  * Runs on an IST stack for x86_64 and on a special task stack for x86_32.
579  *
580  * On x86_64, this is more or less a normal kernel entry.  Notwithstanding the
581  * SDM's warnings about double faults being unrecoverable, returning works as
582  * expected.  Presumably what the SDM actually means is that the CPU may get
583  * the register state wrong on entry, so returning could be a bad idea.
584  *
585  * Various CPU engineers have promised that double faults due to an IRET fault
586  * while the stack is read-only are, in fact, recoverable.
587  *
588  * On x86_32, this is entered through a task gate, and regs are synthesized
589  * from the TSS.  Returning is, in principle, okay, but changes to regs will
590  * be lost.  If, for some reason, we need to return to a context with modified
591  * regs, the shim code could be adjusted to synchronize the registers.
592  *
593  * The 32bit #DF shim provides CR2 already as an argument. On 64bit it needs
594  * to be read before doing anything else.
595  */
596 DEFINE_IDTENTRY_DF(exc_double_fault)
597 {
598 	static const char str[] = "double fault";
599 	struct task_struct *tsk = current;
600 
601 #ifdef CONFIG_VMAP_STACK
602 	unsigned long address = read_cr2();
603 	struct stack_info info;
604 #endif
605 
606 #ifdef CONFIG_X86_ESPFIX64
607 	extern unsigned char native_irq_return_iret[];
608 
609 	/*
610 	 * If IRET takes a non-IST fault on the espfix64 stack, then we
611 	 * end up promoting it to a doublefault.  In that case, take
612 	 * advantage of the fact that we're not using the normal (TSS.sp0)
613 	 * stack right now.  We can write a fake #GP(0) frame at TSS.sp0
614 	 * and then modify our own IRET frame so that, when we return,
615 	 * we land directly at the #GP(0) vector with the stack already
616 	 * set up according to its expectations.
617 	 *
618 	 * The net result is that our #GP handler will think that we
619 	 * entered from usermode with the bad user context.
620 	 *
621 	 * No need for nmi_enter() here because we don't use RCU.
622 	 */
623 	if (((long)regs->sp >> P4D_SHIFT) == ESPFIX_PGD_ENTRY &&
624 		regs->cs == __KERNEL_CS &&
625 		regs->ip == (unsigned long)native_irq_return_iret)
626 	{
627 		struct pt_regs *gpregs = (struct pt_regs *)this_cpu_read(cpu_tss_rw.x86_tss.sp0) - 1;
628 		unsigned long *p = (unsigned long *)regs->sp;
629 
630 		/*
631 		 * regs->sp points to the failing IRET frame on the
632 		 * ESPFIX64 stack.  Copy it to the entry stack.  This fills
633 		 * in gpregs->ss through gpregs->ip.
634 		 *
635 		 */
636 		gpregs->ip	= p[0];
637 		gpregs->cs	= p[1];
638 		gpregs->flags	= p[2];
639 		gpregs->sp	= p[3];
640 		gpregs->ss	= p[4];
641 		gpregs->orig_ax = 0;  /* Missing (lost) #GP error code */
642 
643 		/*
644 		 * Adjust our frame so that we return straight to the #GP
645 		 * vector with the expected RSP value.  This is safe because
646 		 * we won't enable interrupts or schedule before we invoke
647 		 * general_protection, so nothing will clobber the stack
648 		 * frame we just set up.
649 		 *
650 		 * We will enter general_protection with kernel GSBASE,
651 		 * which is what the stub expects, given that the faulting
652 		 * RIP will be the IRET instruction.
653 		 */
654 		regs->ip = (unsigned long)asm_exc_general_protection;
655 		regs->sp = (unsigned long)&gpregs->orig_ax;
656 
657 		return;
658 	}
659 #endif
660 
661 	irqentry_nmi_enter(regs);
662 	instrumentation_begin();
663 	notify_die(DIE_TRAP, str, regs, error_code, X86_TRAP_DF, SIGSEGV);
664 
665 	tsk->thread.error_code = error_code;
666 	tsk->thread.trap_nr = X86_TRAP_DF;
667 
668 #ifdef CONFIG_VMAP_STACK
669 	/*
670 	 * If we overflow the stack into a guard page, the CPU will fail
671 	 * to deliver #PF and will send #DF instead.  Similarly, if we
672 	 * take any non-IST exception while too close to the bottom of
673 	 * the stack, the processor will get a page fault while
674 	 * delivering the exception and will generate a double fault.
675 	 *
676 	 * According to the SDM (footnote in 6.15 under "Interrupt 14 -
677 	 * Page-Fault Exception (#PF):
678 	 *
679 	 *   Processors update CR2 whenever a page fault is detected. If a
680 	 *   second page fault occurs while an earlier page fault is being
681 	 *   delivered, the faulting linear address of the second fault will
682 	 *   overwrite the contents of CR2 (replacing the previous
683 	 *   address). These updates to CR2 occur even if the page fault
684 	 *   results in a double fault or occurs during the delivery of a
685 	 *   double fault.
686 	 *
687 	 * The logic below has a small possibility of incorrectly diagnosing
688 	 * some errors as stack overflows.  For example, if the IDT or GDT
689 	 * gets corrupted such that #GP delivery fails due to a bad descriptor
690 	 * causing #GP and we hit this condition while CR2 coincidentally
691 	 * points to the stack guard page, we'll think we overflowed the
692 	 * stack.  Given that we're going to panic one way or another
693 	 * if this happens, this isn't necessarily worth fixing.
694 	 *
695 	 * If necessary, we could improve the test by only diagnosing
696 	 * a stack overflow if the saved RSP points within 47 bytes of
697 	 * the bottom of the stack: if RSP == tsk_stack + 48 and we
698 	 * take an exception, the stack is already aligned and there
699 	 * will be enough room SS, RSP, RFLAGS, CS, RIP, and a
700 	 * possible error code, so a stack overflow would *not* double
701 	 * fault.  With any less space left, exception delivery could
702 	 * fail, and, as a practical matter, we've overflowed the
703 	 * stack even if the actual trigger for the double fault was
704 	 * something else.
705 	 */
706 	if (get_stack_guard_info((void *)address, &info))
707 		handle_stack_overflow(regs, address, &info);
708 #endif
709 
710 	pr_emerg("PANIC: double fault, error_code: 0x%lx\n", error_code);
711 	die("double fault", regs, error_code);
712 	if (always_true())
713 		panic("Machine halted.");
714 	instrumentation_end();
715 }
716 
717 DEFINE_IDTENTRY(exc_bounds)
718 {
719 	if (notify_die(DIE_TRAP, "bounds", regs, 0,
720 			X86_TRAP_BR, SIGSEGV) == NOTIFY_STOP)
721 		return;
722 	cond_local_irq_enable(regs);
723 
724 	if (!user_mode(regs))
725 		die("bounds", regs, 0);
726 
727 	do_trap(X86_TRAP_BR, SIGSEGV, "bounds", regs, 0, 0, NULL);
728 
729 	cond_local_irq_disable(regs);
730 }
731 
732 enum kernel_gp_hint {
733 	GP_NO_HINT,
734 	GP_NON_CANONICAL,
735 	GP_CANONICAL,
736 	GP_LASS_VIOLATION,
737 	GP_NULL_POINTER,
738 };
739 
740 static const char * const kernel_gp_hint_help[] = {
741 	[GP_NON_CANONICAL]	= "probably for non-canonical address",
742 	[GP_CANONICAL]		= "maybe for address",
743 	[GP_LASS_VIOLATION]	= "probably LASS violation for address",
744 	[GP_NULL_POINTER]	= "kernel NULL pointer dereference",
745 };
746 
747 /*
748  * When an uncaught #GP occurs, try to determine the memory address accessed by
749  * the instruction and return that address to the caller. Also, try to figure
750  * out whether any part of the access to that address was non-canonical or
751  * across privilege levels.
752  */
753 static enum kernel_gp_hint get_kernel_gp_address(struct pt_regs *regs,
754 						 unsigned long *addr)
755 {
756 	u8 insn_buf[MAX_INSN_SIZE];
757 	struct insn insn;
758 	int ret;
759 
760 	if (copy_from_kernel_nofault(insn_buf, (void *)regs->ip,
761 			MAX_INSN_SIZE))
762 		return GP_NO_HINT;
763 
764 	ret = insn_decode_kernel(&insn, insn_buf);
765 	if (ret < 0)
766 		return GP_NO_HINT;
767 
768 	*addr = (unsigned long)insn_get_addr_ref(&insn, regs);
769 	if (*addr == -1UL)
770 		return GP_NO_HINT;
771 
772 #ifdef CONFIG_X86_64
773 	/* Operand is in the kernel half */
774 	if (*addr >= ~__VIRTUAL_MASK)
775 		return GP_CANONICAL;
776 
777 	/* The last byte of the operand is not in the user canonical half */
778 	if (*addr + insn.opnd_bytes - 1 > __VIRTUAL_MASK)
779 		return GP_NON_CANONICAL;
780 
781 	/*
782 	 * A NULL pointer dereference usually causes a #PF. However, it
783 	 * can result in a #GP when LASS is active. Provide the same
784 	 * hint in the rare case that the condition is hit without LASS.
785 	 */
786 	if (*addr < PAGE_SIZE)
787 		return GP_NULL_POINTER;
788 
789 	/*
790 	 * Assume that LASS caused the exception, because the address is
791 	 * canonical and in the user half.
792 	 */
793 	if (cpu_feature_enabled(X86_FEATURE_LASS))
794 		return GP_LASS_VIOLATION;
795 #endif
796 
797 	return GP_CANONICAL;
798 }
799 
800 #define GPFSTR "general protection fault"
801 
802 static bool fixup_iopl_exception(struct pt_regs *regs)
803 {
804 	struct thread_struct *t = &current->thread;
805 	unsigned char byte;
806 	unsigned long ip;
807 
808 	if (!IS_ENABLED(CONFIG_X86_IOPL_IOPERM) || t->iopl_emul != 3)
809 		return false;
810 
811 	if (insn_get_effective_ip(regs, &ip))
812 		return false;
813 
814 	if (get_user(byte, (const char __user *)ip))
815 		return false;
816 
817 	if (byte != 0xfa && byte != 0xfb)
818 		return false;
819 
820 	if (!t->iopl_warn && printk_ratelimit()) {
821 		pr_err("%s[%d] attempts to use CLI/STI, pretending it's a NOP, ip:%lx",
822 		       current->comm, task_pid_nr(current), ip);
823 		print_vma_addr(KERN_CONT " in ", ip);
824 		pr_cont("\n");
825 		t->iopl_warn = 1;
826 	}
827 
828 	regs->ip += 1;
829 	return true;
830 }
831 
832 /*
833  * The unprivileged ENQCMD instruction generates #GPs if the
834  * IA32_PASID MSR has not been populated.  If possible, populate
835  * the MSR from a PASID previously allocated to the mm.
836  */
837 static bool try_fixup_enqcmd_gp(void)
838 {
839 #ifdef CONFIG_ARCH_HAS_CPU_PASID
840 	u32 pasid;
841 
842 	/*
843 	 * MSR_IA32_PASID is managed using XSAVE.  Directly
844 	 * writing to the MSR is only possible when fpregs
845 	 * are valid and the fpstate is not.  This is
846 	 * guaranteed when handling a userspace exception
847 	 * in *before* interrupts are re-enabled.
848 	 */
849 	lockdep_assert_irqs_disabled();
850 
851 	/*
852 	 * Hardware without ENQCMD will not generate
853 	 * #GPs that can be fixed up here.
854 	 */
855 	if (!cpu_feature_enabled(X86_FEATURE_ENQCMD))
856 		return false;
857 
858 	/*
859 	 * If the mm has not been allocated a
860 	 * PASID, the #GP can not be fixed up.
861 	 */
862 	if (!mm_valid_pasid(current->mm))
863 		return false;
864 
865 	pasid = mm_get_enqcmd_pasid(current->mm);
866 
867 	/*
868 	 * Did this thread already have its PASID activated?
869 	 * If so, the #GP must be from something else.
870 	 */
871 	if (current->pasid_activated)
872 		return false;
873 
874 	wrmsrq(MSR_IA32_PASID, pasid | MSR_IA32_PASID_VALID);
875 	current->pasid_activated = 1;
876 
877 	return true;
878 #else
879 	return false;
880 #endif
881 }
882 
883 static bool gp_try_fixup_and_notify(struct pt_regs *regs, int trapnr,
884 				    unsigned long error_code, const char *str,
885 				    unsigned long address)
886 {
887 	if (fixup_exception(regs, trapnr, error_code, address))
888 		return true;
889 
890 	current->thread.error_code = error_code;
891 	current->thread.trap_nr = trapnr;
892 
893 	/*
894 	 * To be potentially processing a kprobe fault and to trust the result
895 	 * from kprobe_running(), we have to be non-preemptible.
896 	 */
897 	if (!preemptible() && kprobe_running() &&
898 	    kprobe_fault_handler(regs, trapnr))
899 		return true;
900 
901 	return notify_die(DIE_GPF, str, regs, error_code, trapnr, SIGSEGV) == NOTIFY_STOP;
902 }
903 
904 static void gp_user_force_sig_segv(struct pt_regs *regs, int trapnr,
905 				   unsigned long error_code, const char *str)
906 {
907 	current->thread.error_code = error_code;
908 	current->thread.trap_nr = trapnr;
909 	show_signal(current, SIGSEGV, "", str, regs, error_code);
910 	force_sig(SIGSEGV);
911 }
912 
913 DEFINE_IDTENTRY_ERRORCODE(exc_general_protection)
914 {
915 	char desc[sizeof(GPFSTR) + 50 + 2*sizeof(unsigned long) + 1] = GPFSTR;
916 	enum kernel_gp_hint hint = GP_NO_HINT;
917 	unsigned long gp_addr;
918 
919 	if (user_mode(regs) && try_fixup_enqcmd_gp())
920 		return;
921 
922 	cond_local_irq_enable(regs);
923 
924 	if (static_cpu_has(X86_FEATURE_UMIP)) {
925 		if (user_mode(regs) && fixup_umip_exception(regs))
926 			goto exit;
927 	}
928 
929 	if (v8086_mode(regs)) {
930 		local_irq_enable();
931 		handle_vm86_fault((struct kernel_vm86_regs *) regs, error_code);
932 		local_irq_disable();
933 		return;
934 	}
935 
936 	if (user_mode(regs)) {
937 		if (fixup_iopl_exception(regs))
938 			goto exit;
939 
940 		if (fixup_vdso_exception(regs, X86_TRAP_GP, error_code, 0))
941 			goto exit;
942 
943 		gp_user_force_sig_segv(regs, X86_TRAP_GP, error_code, desc);
944 		goto exit;
945 	}
946 
947 	if (gp_try_fixup_and_notify(regs, X86_TRAP_GP, error_code, desc, 0))
948 		goto exit;
949 
950 	if (error_code)
951 		snprintf(desc, sizeof(desc), "segment-related " GPFSTR);
952 	else
953 		hint = get_kernel_gp_address(regs, &gp_addr);
954 
955 	if (hint != GP_NO_HINT)
956 		snprintf(desc, sizeof(desc), GPFSTR ", %s 0x%lx",
957 			 kernel_gp_hint_help[hint], gp_addr);
958 
959 	/*
960 	 * KASAN is interested only in the non-canonical case, clear it
961 	 * otherwise.
962 	 */
963 	if (hint != GP_NON_CANONICAL)
964 		gp_addr = 0;
965 
966 	die_addr(desc, regs, error_code, gp_addr);
967 
968 exit:
969 	cond_local_irq_disable(regs);
970 }
971 
972 static bool do_int3(struct pt_regs *regs)
973 {
974 	int res;
975 
976 #ifdef CONFIG_KGDB_LOW_LEVEL_TRAP
977 	if (kgdb_ll_trap(DIE_INT3, "int3", regs, 0, X86_TRAP_BP,
978 			 SIGTRAP) == NOTIFY_STOP)
979 		return true;
980 #endif /* CONFIG_KGDB_LOW_LEVEL_TRAP */
981 
982 #ifdef CONFIG_KPROBES
983 	if (kprobe_int3_handler(regs))
984 		return true;
985 #endif
986 	res = notify_die(DIE_INT3, "int3", regs, 0, X86_TRAP_BP, SIGTRAP);
987 
988 	return res == NOTIFY_STOP;
989 }
990 NOKPROBE_SYMBOL(do_int3);
991 
992 static void do_int3_user(struct pt_regs *regs)
993 {
994 	if (do_int3(regs))
995 		return;
996 
997 	cond_local_irq_enable(regs);
998 	do_trap(X86_TRAP_BP, SIGTRAP, "int3", regs, 0, 0, NULL);
999 	cond_local_irq_disable(regs);
1000 }
1001 
1002 DEFINE_IDTENTRY_RAW(exc_int3)
1003 {
1004 	/*
1005 	 * smp_text_poke_int3_handler() is completely self contained code; it does (and
1006 	 * must) *NOT* call out to anything, lest it hits upon yet another
1007 	 * INT3.
1008 	 */
1009 	if (smp_text_poke_int3_handler(regs))
1010 		return;
1011 
1012 	/*
1013 	 * irqentry_enter_from_user_mode() uses static_branch_{,un}likely()
1014 	 * and therefore can trigger INT3, hence smp_text_poke_int3_handler() must
1015 	 * be done before. If the entry came from kernel mode, then use
1016 	 * nmi_enter() because the INT3 could have been hit in any context
1017 	 * including NMI.
1018 	 */
1019 	if (user_mode(regs)) {
1020 		irqentry_enter_from_user_mode(regs);
1021 		instrumentation_begin();
1022 		do_int3_user(regs);
1023 		instrumentation_end();
1024 		irqentry_exit_to_user_mode(regs);
1025 	} else {
1026 		irqentry_state_t irq_state = irqentry_nmi_enter(regs);
1027 
1028 		instrumentation_begin();
1029 		if (!do_int3(regs))
1030 			die("int3", regs, 0);
1031 		instrumentation_end();
1032 		irqentry_nmi_exit(regs, irq_state);
1033 	}
1034 }
1035 
1036 #ifdef CONFIG_X86_64
1037 /*
1038  * Help handler running on a per-cpu (IST or entry trampoline) stack
1039  * to switch to the normal thread stack if the interrupted code was in
1040  * user mode. The actual stack switch is done in entry_64.S
1041  */
1042 asmlinkage __visible noinstr struct pt_regs *sync_regs(struct pt_regs *eregs)
1043 {
1044 	struct pt_regs *regs = (struct pt_regs *)current_top_of_stack() - 1;
1045 	if (regs != eregs)
1046 		*regs = *eregs;
1047 	return regs;
1048 }
1049 
1050 #ifdef CONFIG_AMD_MEM_ENCRYPT
1051 asmlinkage __visible noinstr struct pt_regs *vc_switch_off_ist(struct pt_regs *regs)
1052 {
1053 	unsigned long sp, *stack;
1054 	struct stack_info info;
1055 	struct pt_regs *regs_ret;
1056 
1057 	/*
1058 	 * In the SYSCALL entry path the RSP value comes from user-space - don't
1059 	 * trust it and switch to the current kernel stack
1060 	 */
1061 	if (ip_within_syscall_gap(regs)) {
1062 		sp = current_top_of_stack();
1063 		goto sync;
1064 	}
1065 
1066 	/*
1067 	 * From here on the RSP value is trusted. Now check whether entry
1068 	 * happened from a safe stack. Not safe are the entry or unknown stacks,
1069 	 * use the fall-back stack instead in this case.
1070 	 */
1071 	sp    = regs->sp;
1072 	stack = (unsigned long *)sp;
1073 
1074 	if (!get_stack_info_noinstr(stack, current, &info) || info.type == STACK_TYPE_ENTRY ||
1075 	    info.type > STACK_TYPE_EXCEPTION_LAST)
1076 		sp = __this_cpu_ist_top_va(VC2);
1077 
1078 sync:
1079 	/*
1080 	 * Found a safe stack - switch to it as if the entry didn't happen via
1081 	 * IST stack. The code below only copies pt_regs, the real switch happens
1082 	 * in assembly code.
1083 	 */
1084 	sp = ALIGN_DOWN(sp, 8) - sizeof(*regs_ret);
1085 
1086 	regs_ret = (struct pt_regs *)sp;
1087 	*regs_ret = *regs;
1088 
1089 	return regs_ret;
1090 }
1091 #endif
1092 
1093 asmlinkage __visible noinstr struct pt_regs *fixup_bad_iret(struct pt_regs *bad_regs)
1094 {
1095 	struct pt_regs tmp, *new_stack;
1096 
1097 	/*
1098 	 * This is called from entry_64.S early in handling a fault
1099 	 * caused by a bad iret to user mode.  To handle the fault
1100 	 * correctly, we want to move our stack frame to where it would
1101 	 * be had we entered directly on the entry stack (rather than
1102 	 * just below the IRET frame) and we want to pretend that the
1103 	 * exception came from the IRET target.
1104 	 */
1105 	new_stack = (struct pt_regs *)__this_cpu_read(cpu_tss_rw.x86_tss.sp0) - 1;
1106 
1107 	/* Copy the IRET target to the temporary storage. */
1108 	__memcpy(&tmp.ip, (void *)bad_regs->sp, 5*8);
1109 
1110 	/* Copy the remainder of the stack from the current stack. */
1111 	__memcpy(&tmp, bad_regs, offsetof(struct pt_regs, ip));
1112 
1113 	/* Update the entry stack */
1114 	__memcpy(new_stack, &tmp, sizeof(tmp));
1115 
1116 	BUG_ON(!user_mode(new_stack));
1117 	return new_stack;
1118 }
1119 #endif
1120 
1121 static bool is_sysenter_singlestep(struct pt_regs *regs)
1122 {
1123 	/*
1124 	 * We don't try for precision here.  If we're anywhere in the region of
1125 	 * code that can be single-stepped in the SYSENTER entry path, then
1126 	 * assume that this is a useless single-step trap due to SYSENTER
1127 	 * being invoked with TF set.  (We don't know in advance exactly
1128 	 * which instructions will be hit because BTF could plausibly
1129 	 * be set.)
1130 	 */
1131 #ifdef CONFIG_X86_32
1132 	return (regs->ip - (unsigned long)__begin_SYSENTER_singlestep_region) <
1133 		(unsigned long)__end_SYSENTER_singlestep_region -
1134 		(unsigned long)__begin_SYSENTER_singlestep_region;
1135 #elif defined(CONFIG_IA32_EMULATION)
1136 	return (regs->ip - (unsigned long)entry_SYSENTER_compat) <
1137 		(unsigned long)__end_entry_SYSENTER_compat -
1138 		(unsigned long)entry_SYSENTER_compat;
1139 #else
1140 	return false;
1141 #endif
1142 }
1143 
1144 static __always_inline unsigned long debug_read_reset_dr6(void)
1145 {
1146 	unsigned long dr6;
1147 
1148 	get_debugreg(dr6, 6);
1149 	dr6 ^= DR6_RESERVED; /* Flip to positive polarity */
1150 
1151 	/*
1152 	 * The Intel SDM says:
1153 	 *
1154 	 *   Certain debug exceptions may clear bits 0-3 of DR6.
1155 	 *
1156 	 *   BLD induced #DB clears DR6.BLD and any other debug
1157 	 *   exception doesn't modify DR6.BLD.
1158 	 *
1159 	 *   RTM induced #DB clears DR6.RTM and any other debug
1160 	 *   exception sets DR6.RTM.
1161 	 *
1162 	 *   To avoid confusion in identifying debug exceptions,
1163 	 *   debug handlers should set DR6.BLD and DR6.RTM, and
1164 	 *   clear other DR6 bits before returning.
1165 	 *
1166 	 * Keep it simple: write DR6 with its architectural reset
1167 	 * value 0xFFFF0FF0, defined as DR6_RESERVED, immediately.
1168 	 */
1169 	set_debugreg(DR6_RESERVED, 6);
1170 
1171 	return dr6;
1172 }
1173 
1174 /*
1175  * Our handling of the processor debug registers is non-trivial.
1176  * We do not clear them on entry and exit from the kernel. Therefore
1177  * it is possible to get a watchpoint trap here from inside the kernel.
1178  * However, the code in ./ptrace.c has ensured that the user can
1179  * only set watchpoints on userspace addresses. Therefore the in-kernel
1180  * watchpoint trap can only occur in code which is reading/writing
1181  * from user space. Such code must not hold kernel locks (since it
1182  * can equally take a page fault), therefore it is safe to call
1183  * force_sig_info even though that claims and releases locks.
1184  *
1185  * Code in ./signal.c ensures that the debug control register
1186  * is restored before we deliver any signal, and therefore that
1187  * user code runs with the correct debug control register even though
1188  * we clear it here.
1189  *
1190  * Being careful here means that we don't have to be as careful in a
1191  * lot of more complicated places (task switching can be a bit lazy
1192  * about restoring all the debug state, and ptrace doesn't have to
1193  * find every occurrence of the TF bit that could be saved away even
1194  * by user code)
1195  *
1196  * May run on IST stack.
1197  */
1198 
1199 static bool notify_debug(struct pt_regs *regs, unsigned long *dr6)
1200 {
1201 	/*
1202 	 * Notifiers will clear bits in @dr6 to indicate the event has been
1203 	 * consumed - hw_breakpoint_handler(), single_stop_cont().
1204 	 *
1205 	 * Notifiers will set bits in @virtual_dr6 to indicate the desire
1206 	 * for signals - ptrace_triggered(), kgdb_hw_overflow_handler().
1207 	 */
1208 	if (notify_die(DIE_DEBUG, "debug", regs, (long)dr6, 0, SIGTRAP) == NOTIFY_STOP)
1209 		return true;
1210 
1211 	return false;
1212 }
1213 
1214 static noinstr void exc_debug_kernel(struct pt_regs *regs, unsigned long dr6)
1215 {
1216 	/*
1217 	 * Disable breakpoints during exception handling; recursive exceptions
1218 	 * are exceedingly 'fun'.
1219 	 *
1220 	 * Since this function is NOKPROBE, and that also applies to
1221 	 * HW_BREAKPOINT_X, we can't hit a breakpoint before this (XXX except a
1222 	 * HW_BREAKPOINT_W on our stack)
1223 	 *
1224 	 * Entry text is excluded for HW_BP_X and cpu_entry_area, which
1225 	 * includes the entry stack is excluded for everything.
1226 	 *
1227 	 * For FRED, nested #DB should just work fine. But when a watchpoint or
1228 	 * breakpoint is set in the code path which is executed by #DB handler,
1229 	 * it results in an endless recursion and stack overflow. Thus we stay
1230 	 * with the IDT approach, i.e., save DR7 and disable #DB.
1231 	 */
1232 	unsigned long dr7 = local_db_save();
1233 	irqentry_state_t irq_state = irqentry_nmi_enter(regs);
1234 	instrumentation_begin();
1235 
1236 	/*
1237 	 * If something gets miswired and we end up here for a user mode
1238 	 * #DB, we will malfunction.
1239 	 */
1240 	WARN_ON_ONCE(user_mode(regs));
1241 
1242 	if (test_thread_flag(TIF_BLOCKSTEP)) {
1243 		/*
1244 		 * The SDM says "The processor clears the BTF flag when it
1245 		 * generates a debug exception." but PTRACE_BLOCKSTEP requested
1246 		 * it for userspace, but we just took a kernel #DB, so re-set
1247 		 * BTF.
1248 		 */
1249 		unsigned long debugctl;
1250 
1251 		rdmsrq(MSR_IA32_DEBUGCTLMSR, debugctl);
1252 		debugctl |= DEBUGCTLMSR_BTF;
1253 		wrmsrq(MSR_IA32_DEBUGCTLMSR, debugctl);
1254 	}
1255 
1256 	/*
1257 	 * Catch SYSENTER with TF set and clear DR_STEP. If this hit a
1258 	 * watchpoint at the same time then that will still be handled.
1259 	 */
1260 	if (!cpu_feature_enabled(X86_FEATURE_FRED) &&
1261 	    (dr6 & DR_STEP) && is_sysenter_singlestep(regs))
1262 		dr6 &= ~DR_STEP;
1263 
1264 	/*
1265 	 * The kernel doesn't use INT1
1266 	 */
1267 	if (!dr6)
1268 		goto out;
1269 
1270 	if (notify_debug(regs, &dr6))
1271 		goto out;
1272 
1273 	/*
1274 	 * The kernel doesn't use TF single-step outside of:
1275 	 *
1276 	 *  - Kprobes, consumed through kprobe_debug_handler()
1277 	 *  - KGDB, consumed through notify_debug()
1278 	 *
1279 	 * So if we get here with DR_STEP set, something is wonky.
1280 	 *
1281 	 * A known way to trigger this is through QEMU's GDB stub,
1282 	 * which leaks #DB into the guest and causes IST recursion.
1283 	 */
1284 	if (WARN_ON_ONCE(dr6 & DR_STEP))
1285 		regs->flags &= ~X86_EFLAGS_TF;
1286 out:
1287 	instrumentation_end();
1288 	irqentry_nmi_exit(regs, irq_state);
1289 
1290 	local_db_restore(dr7);
1291 }
1292 
1293 static noinstr void exc_debug_user(struct pt_regs *regs, unsigned long dr6)
1294 {
1295 	bool icebp;
1296 
1297 	/*
1298 	 * If something gets miswired and we end up here for a kernel mode
1299 	 * #DB, we will malfunction.
1300 	 */
1301 	WARN_ON_ONCE(!user_mode(regs));
1302 
1303 	/*
1304 	 * NB: We can't easily clear DR7 here because
1305 	 * irqentry_exit_to_usermode() can invoke ptrace, schedule, access
1306 	 * user memory, etc.  This means that a recursive #DB is possible.  If
1307 	 * this happens, that #DB will hit exc_debug_kernel() and clear DR7.
1308 	 * Since we're not on the IST stack right now, everything will be
1309 	 * fine.
1310 	 */
1311 
1312 	irqentry_enter_from_user_mode(regs);
1313 	instrumentation_begin();
1314 
1315 	/*
1316 	 * Start the virtual/ptrace DR6 value with just the DR_STEP mask
1317 	 * of the real DR6. ptrace_triggered() will set the DR_TRAPn bits.
1318 	 *
1319 	 * Userspace expects DR_STEP to be visible in ptrace_get_debugreg(6)
1320 	 * even if it is not the result of PTRACE_SINGLESTEP.
1321 	 */
1322 	current->thread.virtual_dr6 = (dr6 & DR_STEP);
1323 
1324 	/*
1325 	 * The SDM says "The processor clears the BTF flag when it
1326 	 * generates a debug exception."  Clear TIF_BLOCKSTEP to keep
1327 	 * TIF_BLOCKSTEP in sync with the hardware BTF flag.
1328 	 */
1329 	clear_thread_flag(TIF_BLOCKSTEP);
1330 
1331 	/*
1332 	 * If dr6 has no reason to give us about the origin of this trap,
1333 	 * then it's very likely the result of an icebp/int01 trap.
1334 	 * User wants a sigtrap for that.
1335 	 */
1336 	icebp = !dr6;
1337 
1338 	if (notify_debug(regs, &dr6))
1339 		goto out;
1340 
1341 	/* It's safe to allow irq's after DR6 has been saved */
1342 	local_irq_enable();
1343 
1344 	if (v8086_mode(regs)) {
1345 		handle_vm86_trap((struct kernel_vm86_regs *)regs, 0, X86_TRAP_DB);
1346 		goto out_irq;
1347 	}
1348 
1349 	/* #DB for bus lock can only be triggered from userspace. */
1350 	if (dr6 & DR_BUS_LOCK)
1351 		handle_bus_lock(regs);
1352 
1353 	/* Add the virtual_dr6 bits for signals. */
1354 	dr6 |= current->thread.virtual_dr6;
1355 	if (dr6 & (DR_STEP | DR_TRAP_BITS) || icebp)
1356 		send_sigtrap(regs, 0, get_si_code(dr6));
1357 
1358 out_irq:
1359 	local_irq_disable();
1360 out:
1361 	instrumentation_end();
1362 	irqentry_exit_to_user_mode(regs);
1363 }
1364 
1365 #ifdef CONFIG_X86_64
1366 /* IST stack entry */
1367 DEFINE_IDTENTRY_DEBUG(exc_debug)
1368 {
1369 	exc_debug_kernel(regs, debug_read_reset_dr6());
1370 }
1371 
1372 /* User entry, runs on regular task stack */
1373 DEFINE_IDTENTRY_DEBUG_USER(exc_debug)
1374 {
1375 	exc_debug_user(regs, debug_read_reset_dr6());
1376 }
1377 
1378 #ifdef CONFIG_X86_FRED
1379 /*
1380  * When occurred on different ring level, i.e., from user or kernel
1381  * context, #DB needs to be handled on different stack: User #DB on
1382  * current task stack, while kernel #DB on a dedicated stack.
1383  *
1384  * This is exactly how FRED event delivery invokes an exception
1385  * handler: ring 3 event on level 0 stack, i.e., current task stack;
1386  * ring 0 event on the #DB dedicated stack specified in the
1387  * IA32_FRED_STKLVLS MSR. So unlike IDT, the FRED debug exception
1388  * entry stub doesn't do stack switch.
1389  */
1390 DEFINE_FREDENTRY_DEBUG(exc_debug)
1391 {
1392 	/*
1393 	 * FRED #DB stores DR6 on the stack in the format which
1394 	 * debug_read_reset_dr6() returns for the IDT entry points.
1395 	 */
1396 	unsigned long dr6 = fred_event_data(regs);
1397 
1398 	if (user_mode(regs))
1399 		exc_debug_user(regs, dr6);
1400 	else
1401 		exc_debug_kernel(regs, dr6);
1402 }
1403 #endif /* CONFIG_X86_FRED */
1404 
1405 #else
1406 /* 32 bit does not have separate entry points. */
1407 DEFINE_IDTENTRY_RAW(exc_debug)
1408 {
1409 	unsigned long dr6 = debug_read_reset_dr6();
1410 
1411 	if (user_mode(regs))
1412 		exc_debug_user(regs, dr6);
1413 	else
1414 		exc_debug_kernel(regs, dr6);
1415 }
1416 #endif
1417 
1418 /*
1419  * Note that we play around with the 'TS' bit in an attempt to get
1420  * the correct behaviour even in the presence of the asynchronous
1421  * IRQ13 behaviour
1422  */
1423 static void math_error(struct pt_regs *regs, int trapnr)
1424 {
1425 	struct task_struct *task = current;
1426 	struct fpu *fpu = x86_task_fpu(task);
1427 	int si_code;
1428 	char *str = (trapnr == X86_TRAP_MF) ? "fpu exception" :
1429 						"simd exception";
1430 
1431 	cond_local_irq_enable(regs);
1432 
1433 	if (!user_mode(regs)) {
1434 		if (fixup_exception(regs, trapnr, 0, 0))
1435 			goto exit;
1436 
1437 		task->thread.error_code = 0;
1438 		task->thread.trap_nr = trapnr;
1439 
1440 		if (notify_die(DIE_TRAP, str, regs, 0, trapnr,
1441 			       SIGFPE) != NOTIFY_STOP)
1442 			die(str, regs, 0);
1443 		goto exit;
1444 	}
1445 
1446 	/*
1447 	 * Synchronize the FPU register state to the memory register state
1448 	 * if necessary. This allows the exception handler to inspect it.
1449 	 */
1450 	fpu_sync_fpstate(fpu);
1451 
1452 	task->thread.trap_nr	= trapnr;
1453 	task->thread.error_code = 0;
1454 
1455 	si_code = fpu__exception_code(fpu, trapnr);
1456 	/* Retry when we get spurious exceptions: */
1457 	if (!si_code)
1458 		goto exit;
1459 
1460 	if (fixup_vdso_exception(regs, trapnr, 0, 0))
1461 		goto exit;
1462 
1463 	force_sig_fault(SIGFPE, si_code,
1464 			(void __user *)uprobe_get_trap_addr(regs));
1465 exit:
1466 	cond_local_irq_disable(regs);
1467 }
1468 
1469 DEFINE_IDTENTRY(exc_coprocessor_error)
1470 {
1471 	math_error(regs, X86_TRAP_MF);
1472 }
1473 
1474 DEFINE_IDTENTRY(exc_simd_coprocessor_error)
1475 {
1476 	if (IS_ENABLED(CONFIG_X86_INVD_BUG)) {
1477 		/* AMD 486 bug: INVD in CPL 0 raises #XF instead of #GP */
1478 		if (!static_cpu_has(X86_FEATURE_XMM)) {
1479 			__exc_general_protection(regs, 0);
1480 			return;
1481 		}
1482 	}
1483 	math_error(regs, X86_TRAP_XF);
1484 }
1485 
1486 DEFINE_IDTENTRY(exc_spurious_interrupt_bug)
1487 {
1488 	/*
1489 	 * This addresses a Pentium Pro Erratum:
1490 	 *
1491 	 * PROBLEM: If the APIC subsystem is configured in mixed mode with
1492 	 * Virtual Wire mode implemented through the local APIC, an
1493 	 * interrupt vector of 0Fh (Intel reserved encoding) may be
1494 	 * generated by the local APIC (Int 15).  This vector may be
1495 	 * generated upon receipt of a spurious interrupt (an interrupt
1496 	 * which is removed before the system receives the INTA sequence)
1497 	 * instead of the programmed 8259 spurious interrupt vector.
1498 	 *
1499 	 * IMPLICATION: The spurious interrupt vector programmed in the
1500 	 * 8259 is normally handled by an operating system's spurious
1501 	 * interrupt handler. However, a vector of 0Fh is unknown to some
1502 	 * operating systems, which would crash if this erratum occurred.
1503 	 *
1504 	 * In theory this could be limited to 32bit, but the handler is not
1505 	 * hurting and who knows which other CPUs suffer from this.
1506 	 */
1507 }
1508 
1509 static bool handle_xfd_event(struct pt_regs *regs)
1510 {
1511 	u64 xfd_err;
1512 	int err;
1513 
1514 	if (!IS_ENABLED(CONFIG_X86_64) || !cpu_feature_enabled(X86_FEATURE_XFD))
1515 		return false;
1516 
1517 	rdmsrq(MSR_IA32_XFD_ERR, xfd_err);
1518 	if (!xfd_err)
1519 		return false;
1520 
1521 	wrmsrq(MSR_IA32_XFD_ERR, 0);
1522 
1523 	/* Die if that happens in kernel space */
1524 	if (WARN_ON(!user_mode(regs)))
1525 		return false;
1526 
1527 	local_irq_enable();
1528 
1529 	err = xfd_enable_feature(xfd_err);
1530 
1531 	switch (err) {
1532 	case -EPERM:
1533 		force_sig_fault(SIGILL, ILL_ILLOPC, error_get_trap_addr(regs));
1534 		break;
1535 	case -EFAULT:
1536 		force_sig(SIGSEGV);
1537 		break;
1538 	}
1539 
1540 	local_irq_disable();
1541 	return true;
1542 }
1543 
1544 DEFINE_IDTENTRY(exc_device_not_available)
1545 {
1546 	unsigned long cr0 = read_cr0();
1547 
1548 	if (handle_xfd_event(regs))
1549 		return;
1550 
1551 #ifdef CONFIG_MATH_EMULATION
1552 	if (!boot_cpu_has(X86_FEATURE_FPU) && (cr0 & X86_CR0_EM)) {
1553 		struct math_emu_info info = { };
1554 
1555 		cond_local_irq_enable(regs);
1556 
1557 		info.regs = regs;
1558 		math_emulate(&info);
1559 
1560 		cond_local_irq_disable(regs);
1561 		return;
1562 	}
1563 #endif
1564 
1565 	/* This should not happen. */
1566 	if (WARN(cr0 & X86_CR0_TS, "CR0.TS was set")) {
1567 		/* Try to fix it up and carry on. */
1568 		write_cr0(cr0 & ~X86_CR0_TS);
1569 	} else {
1570 		/*
1571 		 * Something terrible happened, and we're better off trying
1572 		 * to kill the task than getting stuck in a never-ending
1573 		 * loop of #NM faults.
1574 		 */
1575 		die("unexpected #NM exception", regs, 0);
1576 	}
1577 }
1578 
1579 #ifdef CONFIG_INTEL_TDX_GUEST
1580 
1581 #define VE_FAULT_STR "VE fault"
1582 
1583 static void ve_raise_fault(struct pt_regs *regs, long error_code,
1584 			   unsigned long address)
1585 {
1586 	if (user_mode(regs)) {
1587 		gp_user_force_sig_segv(regs, X86_TRAP_VE, error_code, VE_FAULT_STR);
1588 		return;
1589 	}
1590 
1591 	if (gp_try_fixup_and_notify(regs, X86_TRAP_VE, error_code,
1592 				    VE_FAULT_STR, address)) {
1593 		return;
1594 	}
1595 
1596 	die_addr(VE_FAULT_STR, regs, error_code, address);
1597 }
1598 
1599 /*
1600  * Virtualization Exceptions (#VE) are delivered to TDX guests due to
1601  * specific guest actions which may happen in either user space or the
1602  * kernel:
1603  *
1604  *  * Specific instructions (WBINVD, for example)
1605  *  * Specific MSR accesses
1606  *  * Specific CPUID leaf accesses
1607  *  * Access to specific guest physical addresses
1608  *
1609  * In the settings that Linux will run in, virtualization exceptions are
1610  * never generated on accesses to normal, TD-private memory that has been
1611  * accepted (by BIOS or with tdx_enc_status_changed()).
1612  *
1613  * Syscall entry code has a critical window where the kernel stack is not
1614  * yet set up. Any exception in this window leads to hard to debug issues
1615  * and can be exploited for privilege escalation. Exceptions in the NMI
1616  * entry code also cause issues. Returning from the exception handler with
1617  * IRET will re-enable NMIs and nested NMI will corrupt the NMI stack.
1618  *
1619  * For these reasons, the kernel avoids #VEs during the syscall gap and
1620  * the NMI entry code. Entry code paths do not access TD-shared memory,
1621  * MMIO regions, use #VE triggering MSRs, instructions, or CPUID leaves
1622  * that might generate #VE. VMM can remove memory from TD at any point,
1623  * but access to unaccepted (or missing) private memory leads to VM
1624  * termination, not to #VE.
1625  *
1626  * Similarly to page faults and breakpoints, #VEs are allowed in NMI
1627  * handlers once the kernel is ready to deal with nested NMIs.
1628  *
1629  * During #VE delivery, all interrupts, including NMIs, are blocked until
1630  * TDGETVEINFO is called. It prevents #VE nesting until the kernel reads
1631  * the VE info.
1632  *
1633  * If a guest kernel action which would normally cause a #VE occurs in
1634  * the interrupt-disabled region before TDGETVEINFO, a #DF (fault
1635  * exception) is delivered to the guest which will result in an oops.
1636  *
1637  * The entry code has been audited carefully for following these expectations.
1638  * Changes in the entry code have to be audited for correctness vs. this
1639  * aspect. Similarly to #PF, #VE in these places will expose kernel to
1640  * privilege escalation or may lead to random crashes.
1641  */
1642 DEFINE_IDTENTRY(exc_virtualization_exception)
1643 {
1644 	struct ve_info ve;
1645 
1646 	/*
1647 	 * NMIs/Machine-checks/Interrupts will be in a disabled state
1648 	 * till TDGETVEINFO TDCALL is executed. This ensures that VE
1649 	 * info cannot be overwritten by a nested #VE.
1650 	 */
1651 	tdx_get_ve_info(&ve);
1652 
1653 	cond_local_irq_enable(regs);
1654 
1655 	/*
1656 	 * If tdx_handle_virt_exception() could not process
1657 	 * it successfully, treat it as #GP(0) and handle it.
1658 	 */
1659 	if (!tdx_handle_virt_exception(regs, &ve))
1660 		ve_raise_fault(regs, 0, ve.gla);
1661 
1662 	cond_local_irq_disable(regs);
1663 }
1664 
1665 #endif
1666 
1667 #ifdef CONFIG_X86_32
1668 DEFINE_IDTENTRY_SW(iret_error)
1669 {
1670 	local_irq_enable();
1671 	if (notify_die(DIE_TRAP, "iret exception", regs, 0,
1672 			X86_TRAP_IRET, SIGILL) != NOTIFY_STOP) {
1673 		do_trap(X86_TRAP_IRET, SIGILL, "iret exception", regs, 0,
1674 			ILL_BADSTK, (void __user *)NULL);
1675 	}
1676 	local_irq_disable();
1677 }
1678 #endif
1679 
1680 void __init trap_init(void)
1681 {
1682 	/* Init cpu_entry_area before IST entries are set up */
1683 	setup_cpu_entry_areas();
1684 
1685 	/* Init GHCB memory pages when running as an SEV-ES guest */
1686 	sev_es_init_vc_handling();
1687 
1688 	/* Initialize TSS before setting up traps so ISTs work */
1689 	cpu_init_exception_handling(true);
1690 
1691 	/* Setup traps as cpu_init() might #GP */
1692 	if (!cpu_feature_enabled(X86_FEATURE_FRED))
1693 		idt_setup_traps();
1694 
1695 	cpu_init();
1696 }
1697