xref: /openbmc/linux/arch/x86/entry/entry_64.S (revision 2c64e9cb)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 *  linux/arch/x86_64/entry.S
4 *
5 *  Copyright (C) 1991, 1992  Linus Torvalds
6 *  Copyright (C) 2000, 2001, 2002  Andi Kleen SuSE Labs
7 *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
8 *
9 * entry.S contains the system-call and fault low-level handling routines.
10 *
11 * Some of this is documented in Documentation/x86/entry_64.txt
12 *
13 * A note on terminology:
14 * - iret frame:	Architecture defined interrupt frame from SS to RIP
15 *			at the top of the kernel process stack.
16 *
17 * Some macro usage:
18 * - ENTRY/END:		Define functions in the symbol table.
19 * - TRACE_IRQ_*:	Trace hardirq state for lock debugging.
20 * - idtentry:		Define exception entry points.
21 */
22#include <linux/linkage.h>
23#include <asm/segment.h>
24#include <asm/cache.h>
25#include <asm/errno.h>
26#include <asm/asm-offsets.h>
27#include <asm/msr.h>
28#include <asm/unistd.h>
29#include <asm/thread_info.h>
30#include <asm/hw_irq.h>
31#include <asm/page_types.h>
32#include <asm/irqflags.h>
33#include <asm/paravirt.h>
34#include <asm/percpu.h>
35#include <asm/asm.h>
36#include <asm/smap.h>
37#include <asm/pgtable_types.h>
38#include <asm/export.h>
39#include <asm/frame.h>
40#include <asm/nospec-branch.h>
41#include <linux/err.h>
42
43#include "calling.h"
44
45.code64
46.section .entry.text, "ax"
47
48#ifdef CONFIG_PARAVIRT
49ENTRY(native_usergs_sysret64)
50	UNWIND_HINT_EMPTY
51	swapgs
52	sysretq
53END(native_usergs_sysret64)
54#endif /* CONFIG_PARAVIRT */
55
56.macro TRACE_IRQS_FLAGS flags:req
57#ifdef CONFIG_TRACE_IRQFLAGS
58	btl	$9, \flags		/* interrupts off? */
59	jnc	1f
60	TRACE_IRQS_ON
611:
62#endif
63.endm
64
65.macro TRACE_IRQS_IRETQ
66	TRACE_IRQS_FLAGS EFLAGS(%rsp)
67.endm
68
69/*
70 * When dynamic function tracer is enabled it will add a breakpoint
71 * to all locations that it is about to modify, sync CPUs, update
72 * all the code, sync CPUs, then remove the breakpoints. In this time
73 * if lockdep is enabled, it might jump back into the debug handler
74 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF).
75 *
76 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to
77 * make sure the stack pointer does not get reset back to the top
78 * of the debug stack, and instead just reuses the current stack.
79 */
80#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS)
81
82.macro TRACE_IRQS_OFF_DEBUG
83	call	debug_stack_set_zero
84	TRACE_IRQS_OFF
85	call	debug_stack_reset
86.endm
87
88.macro TRACE_IRQS_ON_DEBUG
89	call	debug_stack_set_zero
90	TRACE_IRQS_ON
91	call	debug_stack_reset
92.endm
93
94.macro TRACE_IRQS_IRETQ_DEBUG
95	btl	$9, EFLAGS(%rsp)		/* interrupts off? */
96	jnc	1f
97	TRACE_IRQS_ON_DEBUG
981:
99.endm
100
101#else
102# define TRACE_IRQS_OFF_DEBUG			TRACE_IRQS_OFF
103# define TRACE_IRQS_ON_DEBUG			TRACE_IRQS_ON
104# define TRACE_IRQS_IRETQ_DEBUG			TRACE_IRQS_IRETQ
105#endif
106
107/*
108 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
109 *
110 * This is the only entry point used for 64-bit system calls.  The
111 * hardware interface is reasonably well designed and the register to
112 * argument mapping Linux uses fits well with the registers that are
113 * available when SYSCALL is used.
114 *
115 * SYSCALL instructions can be found inlined in libc implementations as
116 * well as some other programs and libraries.  There are also a handful
117 * of SYSCALL instructions in the vDSO used, for example, as a
118 * clock_gettimeofday fallback.
119 *
120 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
121 * then loads new ss, cs, and rip from previously programmed MSRs.
122 * rflags gets masked by a value from another MSR (so CLD and CLAC
123 * are not needed). SYSCALL does not save anything on the stack
124 * and does not change rsp.
125 *
126 * Registers on entry:
127 * rax  system call number
128 * rcx  return address
129 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
130 * rdi  arg0
131 * rsi  arg1
132 * rdx  arg2
133 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
134 * r8   arg4
135 * r9   arg5
136 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
137 *
138 * Only called from user space.
139 *
140 * When user can change pt_regs->foo always force IRET. That is because
141 * it deals with uncanonical addresses better. SYSRET has trouble
142 * with them due to bugs in both AMD and Intel CPUs.
143 */
144
145ENTRY(entry_SYSCALL_64)
146	UNWIND_HINT_EMPTY
147	/*
148	 * Interrupts are off on entry.
149	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
150	 * it is too small to ever cause noticeable irq latency.
151	 */
152
153	swapgs
154	/* tss.sp2 is scratch space. */
155	movq	%rsp, PER_CPU_VAR(cpu_tss_rw + TSS_sp2)
156	SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
157	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
158
159	/* Construct struct pt_regs on stack */
160	pushq	$__USER_DS				/* pt_regs->ss */
161	pushq	PER_CPU_VAR(cpu_tss_rw + TSS_sp2)	/* pt_regs->sp */
162	pushq	%r11					/* pt_regs->flags */
163	pushq	$__USER_CS				/* pt_regs->cs */
164	pushq	%rcx					/* pt_regs->ip */
165GLOBAL(entry_SYSCALL_64_after_hwframe)
166	pushq	%rax					/* pt_regs->orig_ax */
167
168	PUSH_AND_CLEAR_REGS rax=$-ENOSYS
169
170	TRACE_IRQS_OFF
171
172	/* IRQs are off. */
173	movq	%rax, %rdi
174	movq	%rsp, %rsi
175	call	do_syscall_64		/* returns with IRQs disabled */
176
177	TRACE_IRQS_IRETQ		/* we're about to change IF */
178
179	/*
180	 * Try to use SYSRET instead of IRET if we're returning to
181	 * a completely clean 64-bit userspace context.  If we're not,
182	 * go to the slow exit path.
183	 */
184	movq	RCX(%rsp), %rcx
185	movq	RIP(%rsp), %r11
186
187	cmpq	%rcx, %r11	/* SYSRET requires RCX == RIP */
188	jne	swapgs_restore_regs_and_return_to_usermode
189
190	/*
191	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
192	 * in kernel space.  This essentially lets the user take over
193	 * the kernel, since userspace controls RSP.
194	 *
195	 * If width of "canonical tail" ever becomes variable, this will need
196	 * to be updated to remain correct on both old and new CPUs.
197	 *
198	 * Change top bits to match most significant bit (47th or 56th bit
199	 * depending on paging mode) in the address.
200	 */
201#ifdef CONFIG_X86_5LEVEL
202	ALTERNATIVE "shl $(64 - 48), %rcx; sar $(64 - 48), %rcx", \
203		"shl $(64 - 57), %rcx; sar $(64 - 57), %rcx", X86_FEATURE_LA57
204#else
205	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
206	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
207#endif
208
209	/* If this changed %rcx, it was not canonical */
210	cmpq	%rcx, %r11
211	jne	swapgs_restore_regs_and_return_to_usermode
212
213	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
214	jne	swapgs_restore_regs_and_return_to_usermode
215
216	movq	R11(%rsp), %r11
217	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
218	jne	swapgs_restore_regs_and_return_to_usermode
219
220	/*
221	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
222	 * restore RF properly. If the slowpath sets it for whatever reason, we
223	 * need to restore it correctly.
224	 *
225	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
226	 * trap from userspace immediately after SYSRET.  This would cause an
227	 * infinite loop whenever #DB happens with register state that satisfies
228	 * the opportunistic SYSRET conditions.  For example, single-stepping
229	 * this user code:
230	 *
231	 *           movq	$stuck_here, %rcx
232	 *           pushfq
233	 *           popq %r11
234	 *   stuck_here:
235	 *
236	 * would never get past 'stuck_here'.
237	 */
238	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
239	jnz	swapgs_restore_regs_and_return_to_usermode
240
241	/* nothing to check for RSP */
242
243	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
244	jne	swapgs_restore_regs_and_return_to_usermode
245
246	/*
247	 * We win! This label is here just for ease of understanding
248	 * perf profiles. Nothing jumps here.
249	 */
250syscall_return_via_sysret:
251	/* rcx and r11 are already restored (see code above) */
252	UNWIND_HINT_EMPTY
253	POP_REGS pop_rdi=0 skip_r11rcx=1
254
255	/*
256	 * Now all regs are restored except RSP and RDI.
257	 * Save old stack pointer and switch to trampoline stack.
258	 */
259	movq	%rsp, %rdi
260	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
261
262	pushq	RSP-RDI(%rdi)	/* RSP */
263	pushq	(%rdi)		/* RDI */
264
265	/*
266	 * We are on the trampoline stack.  All regs except RDI are live.
267	 * We can do future final exit work right here.
268	 */
269	STACKLEAK_ERASE_NOCLOBBER
270
271	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
272
273	popq	%rdi
274	popq	%rsp
275	USERGS_SYSRET64
276END(entry_SYSCALL_64)
277
278/*
279 * %rdi: prev task
280 * %rsi: next task
281 */
282ENTRY(__switch_to_asm)
283	UNWIND_HINT_FUNC
284	/*
285	 * Save callee-saved registers
286	 * This must match the order in inactive_task_frame
287	 */
288	pushq	%rbp
289	pushq	%rbx
290	pushq	%r12
291	pushq	%r13
292	pushq	%r14
293	pushq	%r15
294
295	/* switch stack */
296	movq	%rsp, TASK_threadsp(%rdi)
297	movq	TASK_threadsp(%rsi), %rsp
298
299#ifdef CONFIG_STACKPROTECTOR
300	movq	TASK_stack_canary(%rsi), %rbx
301	movq	%rbx, PER_CPU_VAR(fixed_percpu_data) + stack_canary_offset
302#endif
303
304#ifdef CONFIG_RETPOLINE
305	/*
306	 * When switching from a shallower to a deeper call stack
307	 * the RSB may either underflow or use entries populated
308	 * with userspace addresses. On CPUs where those concerns
309	 * exist, overwrite the RSB with entries which capture
310	 * speculative execution to prevent attack.
311	 */
312	FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW
313#endif
314
315	/* restore callee-saved registers */
316	popq	%r15
317	popq	%r14
318	popq	%r13
319	popq	%r12
320	popq	%rbx
321	popq	%rbp
322
323	jmp	__switch_to
324END(__switch_to_asm)
325
326/*
327 * A newly forked process directly context switches into this address.
328 *
329 * rax: prev task we switched from
330 * rbx: kernel thread func (NULL for user thread)
331 * r12: kernel thread arg
332 */
333ENTRY(ret_from_fork)
334	UNWIND_HINT_EMPTY
335	movq	%rax, %rdi
336	call	schedule_tail			/* rdi: 'prev' task parameter */
337
338	testq	%rbx, %rbx			/* from kernel_thread? */
339	jnz	1f				/* kernel threads are uncommon */
340
3412:
342	UNWIND_HINT_REGS
343	movq	%rsp, %rdi
344	call	syscall_return_slowpath	/* returns with IRQs disabled */
345	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
346	jmp	swapgs_restore_regs_and_return_to_usermode
347
3481:
349	/* kernel thread */
350	UNWIND_HINT_EMPTY
351	movq	%r12, %rdi
352	CALL_NOSPEC %rbx
353	/*
354	 * A kernel thread is allowed to return here after successfully
355	 * calling do_execve().  Exit to userspace to complete the execve()
356	 * syscall.
357	 */
358	movq	$0, RAX(%rsp)
359	jmp	2b
360END(ret_from_fork)
361
362/*
363 * Build the entry stubs with some assembler magic.
364 * We pack 1 stub into every 8-byte block.
365 */
366	.align 8
367ENTRY(irq_entries_start)
368    vector=FIRST_EXTERNAL_VECTOR
369    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
370	UNWIND_HINT_IRET_REGS
371	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
372	jmp	common_interrupt
373	.align	8
374	vector=vector+1
375    .endr
376END(irq_entries_start)
377
378.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
379#ifdef CONFIG_DEBUG_ENTRY
380	pushq %rax
381	SAVE_FLAGS(CLBR_RAX)
382	testl $X86_EFLAGS_IF, %eax
383	jz .Lokay_\@
384	ud2
385.Lokay_\@:
386	popq %rax
387#endif
388.endm
389
390/*
391 * Enters the IRQ stack if we're not already using it.  NMI-safe.  Clobbers
392 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone.
393 * Requires kernel GSBASE.
394 *
395 * The invariant is that, if irq_count != -1, then the IRQ stack is in use.
396 */
397.macro ENTER_IRQ_STACK regs=1 old_rsp save_ret=0
398	DEBUG_ENTRY_ASSERT_IRQS_OFF
399
400	.if \save_ret
401	/*
402	 * If save_ret is set, the original stack contains one additional
403	 * entry -- the return address. Therefore, move the address one
404	 * entry below %rsp to \old_rsp.
405	 */
406	leaq	8(%rsp), \old_rsp
407	.else
408	movq	%rsp, \old_rsp
409	.endif
410
411	.if \regs
412	UNWIND_HINT_REGS base=\old_rsp
413	.endif
414
415	incl	PER_CPU_VAR(irq_count)
416	jnz	.Lirq_stack_push_old_rsp_\@
417
418	/*
419	 * Right now, if we just incremented irq_count to zero, we've
420	 * claimed the IRQ stack but we haven't switched to it yet.
421	 *
422	 * If anything is added that can interrupt us here without using IST,
423	 * it must be *extremely* careful to limit its stack usage.  This
424	 * could include kprobes and a hypothetical future IST-less #DB
425	 * handler.
426	 *
427	 * The OOPS unwinder relies on the word at the top of the IRQ
428	 * stack linking back to the previous RSP for the entire time we're
429	 * on the IRQ stack.  For this to work reliably, we need to write
430	 * it before we actually move ourselves to the IRQ stack.
431	 */
432
433	movq	\old_rsp, PER_CPU_VAR(irq_stack_backing_store + IRQ_STACK_SIZE - 8)
434	movq	PER_CPU_VAR(hardirq_stack_ptr), %rsp
435
436#ifdef CONFIG_DEBUG_ENTRY
437	/*
438	 * If the first movq above becomes wrong due to IRQ stack layout
439	 * changes, the only way we'll notice is if we try to unwind right
440	 * here.  Assert that we set up the stack right to catch this type
441	 * of bug quickly.
442	 */
443	cmpq	-8(%rsp), \old_rsp
444	je	.Lirq_stack_okay\@
445	ud2
446	.Lirq_stack_okay\@:
447#endif
448
449.Lirq_stack_push_old_rsp_\@:
450	pushq	\old_rsp
451
452	.if \regs
453	UNWIND_HINT_REGS indirect=1
454	.endif
455
456	.if \save_ret
457	/*
458	 * Push the return address to the stack. This return address can
459	 * be found at the "real" original RSP, which was offset by 8 at
460	 * the beginning of this macro.
461	 */
462	pushq	-8(\old_rsp)
463	.endif
464.endm
465
466/*
467 * Undoes ENTER_IRQ_STACK.
468 */
469.macro LEAVE_IRQ_STACK regs=1
470	DEBUG_ENTRY_ASSERT_IRQS_OFF
471	/* We need to be off the IRQ stack before decrementing irq_count. */
472	popq	%rsp
473
474	.if \regs
475	UNWIND_HINT_REGS
476	.endif
477
478	/*
479	 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming
480	 * the irq stack but we're not on it.
481	 */
482
483	decl	PER_CPU_VAR(irq_count)
484.endm
485
486/*
487 * Interrupt entry helper function.
488 *
489 * Entry runs with interrupts off. Stack layout at entry:
490 * +----------------------------------------------------+
491 * | regs->ss						|
492 * | regs->rsp						|
493 * | regs->eflags					|
494 * | regs->cs						|
495 * | regs->ip						|
496 * +----------------------------------------------------+
497 * | regs->orig_ax = ~(interrupt number)		|
498 * +----------------------------------------------------+
499 * | return address					|
500 * +----------------------------------------------------+
501 */
502ENTRY(interrupt_entry)
503	UNWIND_HINT_FUNC
504	ASM_CLAC
505	cld
506
507	testb	$3, CS-ORIG_RAX+8(%rsp)
508	jz	1f
509	SWAPGS
510
511	/*
512	 * Switch to the thread stack. The IRET frame and orig_ax are
513	 * on the stack, as well as the return address. RDI..R12 are
514	 * not (yet) on the stack and space has not (yet) been
515	 * allocated for them.
516	 */
517	pushq	%rdi
518
519	/* Need to switch before accessing the thread stack. */
520	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi
521	movq	%rsp, %rdi
522	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
523
524	 /*
525	  * We have RDI, return address, and orig_ax on the stack on
526	  * top of the IRET frame. That means offset=24
527	  */
528	UNWIND_HINT_IRET_REGS base=%rdi offset=24
529
530	pushq	7*8(%rdi)		/* regs->ss */
531	pushq	6*8(%rdi)		/* regs->rsp */
532	pushq	5*8(%rdi)		/* regs->eflags */
533	pushq	4*8(%rdi)		/* regs->cs */
534	pushq	3*8(%rdi)		/* regs->ip */
535	pushq	2*8(%rdi)		/* regs->orig_ax */
536	pushq	8(%rdi)			/* return address */
537	UNWIND_HINT_FUNC
538
539	movq	(%rdi), %rdi
5401:
541
542	PUSH_AND_CLEAR_REGS save_ret=1
543	ENCODE_FRAME_POINTER 8
544
545	testb	$3, CS+8(%rsp)
546	jz	1f
547
548	/*
549	 * IRQ from user mode.
550	 *
551	 * We need to tell lockdep that IRQs are off.  We can't do this until
552	 * we fix gsbase, and we should do it before enter_from_user_mode
553	 * (which can take locks).  Since TRACE_IRQS_OFF is idempotent,
554	 * the simplest way to handle it is to just call it twice if
555	 * we enter from user mode.  There's no reason to optimize this since
556	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
557	 */
558	TRACE_IRQS_OFF
559
560	CALL_enter_from_user_mode
561
5621:
563	ENTER_IRQ_STACK old_rsp=%rdi save_ret=1
564	/* We entered an interrupt context - irqs are off: */
565	TRACE_IRQS_OFF
566
567	ret
568END(interrupt_entry)
569_ASM_NOKPROBE(interrupt_entry)
570
571
572/* Interrupt entry/exit. */
573
574	/*
575	 * The interrupt stubs push (~vector+0x80) onto the stack and
576	 * then jump to common_interrupt.
577	 */
578	.p2align CONFIG_X86_L1_CACHE_SHIFT
579common_interrupt:
580	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
581	call	interrupt_entry
582	UNWIND_HINT_REGS indirect=1
583	call	do_IRQ	/* rdi points to pt_regs */
584	/* 0(%rsp): old RSP */
585ret_from_intr:
586	DISABLE_INTERRUPTS(CLBR_ANY)
587	TRACE_IRQS_OFF
588
589	LEAVE_IRQ_STACK
590
591	testb	$3, CS(%rsp)
592	jz	retint_kernel
593
594	/* Interrupt came from user space */
595GLOBAL(retint_user)
596	mov	%rsp,%rdi
597	call	prepare_exit_to_usermode
598	TRACE_IRQS_IRETQ
599
600GLOBAL(swapgs_restore_regs_and_return_to_usermode)
601#ifdef CONFIG_DEBUG_ENTRY
602	/* Assert that pt_regs indicates user mode. */
603	testb	$3, CS(%rsp)
604	jnz	1f
605	ud2
6061:
607#endif
608	POP_REGS pop_rdi=0
609
610	/*
611	 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS.
612	 * Save old stack pointer and switch to trampoline stack.
613	 */
614	movq	%rsp, %rdi
615	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
616
617	/* Copy the IRET frame to the trampoline stack. */
618	pushq	6*8(%rdi)	/* SS */
619	pushq	5*8(%rdi)	/* RSP */
620	pushq	4*8(%rdi)	/* EFLAGS */
621	pushq	3*8(%rdi)	/* CS */
622	pushq	2*8(%rdi)	/* RIP */
623
624	/* Push user RDI on the trampoline stack. */
625	pushq	(%rdi)
626
627	/*
628	 * We are on the trampoline stack.  All regs except RDI are live.
629	 * We can do future final exit work right here.
630	 */
631	STACKLEAK_ERASE_NOCLOBBER
632
633	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
634
635	/* Restore RDI. */
636	popq	%rdi
637	SWAPGS
638	INTERRUPT_RETURN
639
640
641/* Returning to kernel space */
642retint_kernel:
643#ifdef CONFIG_PREEMPT
644	/* Interrupts are off */
645	/* Check if we need preemption */
646	btl	$9, EFLAGS(%rsp)		/* were interrupts off? */
647	jnc	1f
648	cmpl	$0, PER_CPU_VAR(__preempt_count)
649	jnz	1f
650	call	preempt_schedule_irq
6511:
652#endif
653	/*
654	 * The iretq could re-enable interrupts:
655	 */
656	TRACE_IRQS_IRETQ
657
658GLOBAL(restore_regs_and_return_to_kernel)
659#ifdef CONFIG_DEBUG_ENTRY
660	/* Assert that pt_regs indicates kernel mode. */
661	testb	$3, CS(%rsp)
662	jz	1f
663	ud2
6641:
665#endif
666	POP_REGS
667	addq	$8, %rsp	/* skip regs->orig_ax */
668	/*
669	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
670	 * when returning from IPI handler.
671	 */
672	INTERRUPT_RETURN
673
674ENTRY(native_iret)
675	UNWIND_HINT_IRET_REGS
676	/*
677	 * Are we returning to a stack segment from the LDT?  Note: in
678	 * 64-bit mode SS:RSP on the exception stack is always valid.
679	 */
680#ifdef CONFIG_X86_ESPFIX64
681	testb	$4, (SS-RIP)(%rsp)
682	jnz	native_irq_return_ldt
683#endif
684
685.global native_irq_return_iret
686native_irq_return_iret:
687	/*
688	 * This may fault.  Non-paranoid faults on return to userspace are
689	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
690	 * Double-faults due to espfix64 are handled in do_double_fault.
691	 * Other faults here are fatal.
692	 */
693	iretq
694
695#ifdef CONFIG_X86_ESPFIX64
696native_irq_return_ldt:
697	/*
698	 * We are running with user GSBASE.  All GPRs contain their user
699	 * values.  We have a percpu ESPFIX stack that is eight slots
700	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
701	 * of the ESPFIX stack.
702	 *
703	 * We clobber RAX and RDI in this code.  We stash RDI on the
704	 * normal stack and RAX on the ESPFIX stack.
705	 *
706	 * The ESPFIX stack layout we set up looks like this:
707	 *
708	 * --- top of ESPFIX stack ---
709	 * SS
710	 * RSP
711	 * RFLAGS
712	 * CS
713	 * RIP  <-- RSP points here when we're done
714	 * RAX  <-- espfix_waddr points here
715	 * --- bottom of ESPFIX stack ---
716	 */
717
718	pushq	%rdi				/* Stash user RDI */
719	SWAPGS					/* to kernel GS */
720	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi	/* to kernel CR3 */
721
722	movq	PER_CPU_VAR(espfix_waddr), %rdi
723	movq	%rax, (0*8)(%rdi)		/* user RAX */
724	movq	(1*8)(%rsp), %rax		/* user RIP */
725	movq	%rax, (1*8)(%rdi)
726	movq	(2*8)(%rsp), %rax		/* user CS */
727	movq	%rax, (2*8)(%rdi)
728	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
729	movq	%rax, (3*8)(%rdi)
730	movq	(5*8)(%rsp), %rax		/* user SS */
731	movq	%rax, (5*8)(%rdi)
732	movq	(4*8)(%rsp), %rax		/* user RSP */
733	movq	%rax, (4*8)(%rdi)
734	/* Now RAX == RSP. */
735
736	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
737
738	/*
739	 * espfix_stack[31:16] == 0.  The page tables are set up such that
740	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
741	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
742	 * the same page.  Set up RSP so that RSP[31:16] contains the
743	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
744	 * still points to an RO alias of the ESPFIX stack.
745	 */
746	orq	PER_CPU_VAR(espfix_stack), %rax
747
748	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
749	SWAPGS					/* to user GS */
750	popq	%rdi				/* Restore user RDI */
751
752	movq	%rax, %rsp
753	UNWIND_HINT_IRET_REGS offset=8
754
755	/*
756	 * At this point, we cannot write to the stack any more, but we can
757	 * still read.
758	 */
759	popq	%rax				/* Restore user RAX */
760
761	/*
762	 * RSP now points to an ordinary IRET frame, except that the page
763	 * is read-only and RSP[31:16] are preloaded with the userspace
764	 * values.  We can now IRET back to userspace.
765	 */
766	jmp	native_irq_return_iret
767#endif
768END(common_interrupt)
769_ASM_NOKPROBE(common_interrupt)
770
771/*
772 * APIC interrupts.
773 */
774.macro apicinterrupt3 num sym do_sym
775ENTRY(\sym)
776	UNWIND_HINT_IRET_REGS
777	pushq	$~(\num)
778.Lcommon_\sym:
779	call	interrupt_entry
780	UNWIND_HINT_REGS indirect=1
781	call	\do_sym	/* rdi points to pt_regs */
782	jmp	ret_from_intr
783END(\sym)
784_ASM_NOKPROBE(\sym)
785.endm
786
787/* Make sure APIC interrupt handlers end up in the irqentry section: */
788#define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
789#define POP_SECTION_IRQENTRY	.popsection
790
791.macro apicinterrupt num sym do_sym
792PUSH_SECTION_IRQENTRY
793apicinterrupt3 \num \sym \do_sym
794POP_SECTION_IRQENTRY
795.endm
796
797#ifdef CONFIG_SMP
798apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
799apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
800#endif
801
802#ifdef CONFIG_X86_UV
803apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
804#endif
805
806apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
807apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
808
809#ifdef CONFIG_HAVE_KVM
810apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
811apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
812apicinterrupt3 POSTED_INTR_NESTED_VECTOR	kvm_posted_intr_nested_ipi	smp_kvm_posted_intr_nested_ipi
813#endif
814
815#ifdef CONFIG_X86_MCE_THRESHOLD
816apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
817#endif
818
819#ifdef CONFIG_X86_MCE_AMD
820apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
821#endif
822
823#ifdef CONFIG_X86_THERMAL_VECTOR
824apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
825#endif
826
827#ifdef CONFIG_SMP
828apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
829apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
830apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
831#endif
832
833apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
834apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
835
836#ifdef CONFIG_IRQ_WORK
837apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
838#endif
839
840/*
841 * Exception entry points.
842 */
843#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss_rw) + (TSS_ist + (x) * 8)
844
845/**
846 * idtentry - Generate an IDT entry stub
847 * @sym:		Name of the generated entry point
848 * @do_sym: 		C function to be called
849 * @has_error_code: 	True if this IDT vector has an error code on the stack
850 * @paranoid: 		non-zero means that this vector may be invoked from
851 *			kernel mode with user GSBASE and/or user CR3.
852 *			2 is special -- see below.
853 * @shift_ist:		Set to an IST index if entries from kernel mode should
854 *             		decrement the IST stack so that nested entries get a
855 *			fresh stack.  (This is for #DB, which has a nasty habit
856 *             		of recursing.)
857 *
858 * idtentry generates an IDT stub that sets up a usable kernel context,
859 * creates struct pt_regs, and calls @do_sym.  The stub has the following
860 * special behaviors:
861 *
862 * On an entry from user mode, the stub switches from the trampoline or
863 * IST stack to the normal thread stack.  On an exit to user mode, the
864 * normal exit-to-usermode path is invoked.
865 *
866 * On an exit to kernel mode, if @paranoid == 0, we check for preemption,
867 * whereas we omit the preemption check if @paranoid != 0.  This is purely
868 * because the implementation is simpler this way.  The kernel only needs
869 * to check for asynchronous kernel preemption when IRQ handlers return.
870 *
871 * If @paranoid == 0, then the stub will handle IRET faults by pretending
872 * that the fault came from user mode.  It will handle gs_change faults by
873 * pretending that the fault happened with kernel GSBASE.  Since this handling
874 * is omitted for @paranoid != 0, the #GP, #SS, and #NP stubs must have
875 * @paranoid == 0.  This special handling will do the wrong thing for
876 * espfix-induced #DF on IRET, so #DF must not use @paranoid == 0.
877 *
878 * @paranoid == 2 is special: the stub will never switch stacks.  This is for
879 * #DF: if the thread stack is somehow unusable, we'll still get a useful OOPS.
880 */
881.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1 ist_offset=0
882ENTRY(\sym)
883	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
884
885	/* Sanity check */
886	.if \shift_ist != -1 && \paranoid == 0
887	.error "using shift_ist requires paranoid=1"
888	.endif
889
890	ASM_CLAC
891
892	.if \has_error_code == 0
893	pushq	$-1				/* ORIG_RAX: no syscall to restart */
894	.endif
895
896	.if \paranoid == 1
897	testb	$3, CS-ORIG_RAX(%rsp)		/* If coming from userspace, switch stacks */
898	jnz	.Lfrom_usermode_switch_stack_\@
899	.endif
900
901	.if \paranoid
902	call	paranoid_entry
903	.else
904	call	error_entry
905	.endif
906	UNWIND_HINT_REGS
907	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
908
909	.if \paranoid
910	.if \shift_ist != -1
911	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
912	.else
913	TRACE_IRQS_OFF
914	.endif
915	.endif
916
917	movq	%rsp, %rdi			/* pt_regs pointer */
918
919	.if \has_error_code
920	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
921	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
922	.else
923	xorl	%esi, %esi			/* no error code */
924	.endif
925
926	.if \shift_ist != -1
927	subq	$\ist_offset, CPU_TSS_IST(\shift_ist)
928	.endif
929
930	call	\do_sym
931
932	.if \shift_ist != -1
933	addq	$\ist_offset, CPU_TSS_IST(\shift_ist)
934	.endif
935
936	/* these procedures expect "no swapgs" flag in ebx */
937	.if \paranoid
938	jmp	paranoid_exit
939	.else
940	jmp	error_exit
941	.endif
942
943	.if \paranoid == 1
944	/*
945	 * Entry from userspace.  Switch stacks and treat it
946	 * as a normal entry.  This means that paranoid handlers
947	 * run in real process context if user_mode(regs).
948	 */
949.Lfrom_usermode_switch_stack_\@:
950	call	error_entry
951
952	movq	%rsp, %rdi			/* pt_regs pointer */
953
954	.if \has_error_code
955	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
956	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
957	.else
958	xorl	%esi, %esi			/* no error code */
959	.endif
960
961	call	\do_sym
962
963	jmp	error_exit
964	.endif
965_ASM_NOKPROBE(\sym)
966END(\sym)
967.endm
968
969idtentry divide_error			do_divide_error			has_error_code=0
970idtentry overflow			do_overflow			has_error_code=0
971idtentry bounds				do_bounds			has_error_code=0
972idtentry invalid_op			do_invalid_op			has_error_code=0
973idtentry device_not_available		do_device_not_available		has_error_code=0
974idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
975idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
976idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
977idtentry segment_not_present		do_segment_not_present		has_error_code=1
978idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
979idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
980idtentry alignment_check		do_alignment_check		has_error_code=1
981idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
982
983
984	/*
985	 * Reload gs selector with exception handling
986	 * edi:  new selector
987	 */
988ENTRY(native_load_gs_index)
989	FRAME_BEGIN
990	pushfq
991	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
992	TRACE_IRQS_OFF
993	SWAPGS
994.Lgs_change:
995	movl	%edi, %gs
9962:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
997	SWAPGS
998	TRACE_IRQS_FLAGS (%rsp)
999	popfq
1000	FRAME_END
1001	ret
1002ENDPROC(native_load_gs_index)
1003EXPORT_SYMBOL(native_load_gs_index)
1004
1005	_ASM_EXTABLE(.Lgs_change, bad_gs)
1006	.section .fixup, "ax"
1007	/* running with kernelgs */
1008bad_gs:
1009	SWAPGS					/* switch back to user gs */
1010.macro ZAP_GS
1011	/* This can't be a string because the preprocessor needs to see it. */
1012	movl $__USER_DS, %eax
1013	movl %eax, %gs
1014.endm
1015	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
1016	xorl	%eax, %eax
1017	movl	%eax, %gs
1018	jmp	2b
1019	.previous
1020
1021/* Call softirq on interrupt stack. Interrupts are off. */
1022ENTRY(do_softirq_own_stack)
1023	pushq	%rbp
1024	mov	%rsp, %rbp
1025	ENTER_IRQ_STACK regs=0 old_rsp=%r11
1026	call	__do_softirq
1027	LEAVE_IRQ_STACK regs=0
1028	leaveq
1029	ret
1030ENDPROC(do_softirq_own_stack)
1031
1032#ifdef CONFIG_XEN_PV
1033idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0
1034
1035/*
1036 * A note on the "critical region" in our callback handler.
1037 * We want to avoid stacking callback handlers due to events occurring
1038 * during handling of the last event. To do this, we keep events disabled
1039 * until we've done all processing. HOWEVER, we must enable events before
1040 * popping the stack frame (can't be done atomically) and so it would still
1041 * be possible to get enough handler activations to overflow the stack.
1042 * Although unlikely, bugs of that kind are hard to track down, so we'd
1043 * like to avoid the possibility.
1044 * So, on entry to the handler we detect whether we interrupted an
1045 * existing activation in its critical region -- if so, we pop the current
1046 * activation and restart the handler using the previous one.
1047 */
1048ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
1049
1050/*
1051 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
1052 * see the correct pointer to the pt_regs
1053 */
1054	UNWIND_HINT_FUNC
1055	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
1056	UNWIND_HINT_REGS
1057
1058	ENTER_IRQ_STACK old_rsp=%r10
1059	call	xen_evtchn_do_upcall
1060	LEAVE_IRQ_STACK
1061
1062#ifndef CONFIG_PREEMPT
1063	call	xen_maybe_preempt_hcall
1064#endif
1065	jmp	error_exit
1066END(xen_do_hypervisor_callback)
1067
1068/*
1069 * Hypervisor uses this for application faults while it executes.
1070 * We get here for two reasons:
1071 *  1. Fault while reloading DS, ES, FS or GS
1072 *  2. Fault while executing IRET
1073 * Category 1 we do not need to fix up as Xen has already reloaded all segment
1074 * registers that could be reloaded and zeroed the others.
1075 * Category 2 we fix up by killing the current process. We cannot use the
1076 * normal Linux return path in this case because if we use the IRET hypercall
1077 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
1078 * We distinguish between categories by comparing each saved segment register
1079 * with its current contents: any discrepancy means we in category 1.
1080 */
1081ENTRY(xen_failsafe_callback)
1082	UNWIND_HINT_EMPTY
1083	movl	%ds, %ecx
1084	cmpw	%cx, 0x10(%rsp)
1085	jne	1f
1086	movl	%es, %ecx
1087	cmpw	%cx, 0x18(%rsp)
1088	jne	1f
1089	movl	%fs, %ecx
1090	cmpw	%cx, 0x20(%rsp)
1091	jne	1f
1092	movl	%gs, %ecx
1093	cmpw	%cx, 0x28(%rsp)
1094	jne	1f
1095	/* All segments match their saved values => Category 2 (Bad IRET). */
1096	movq	(%rsp), %rcx
1097	movq	8(%rsp), %r11
1098	addq	$0x30, %rsp
1099	pushq	$0				/* RIP */
1100	UNWIND_HINT_IRET_REGS offset=8
1101	jmp	general_protection
11021:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
1103	movq	(%rsp), %rcx
1104	movq	8(%rsp), %r11
1105	addq	$0x30, %rsp
1106	UNWIND_HINT_IRET_REGS
1107	pushq	$-1 /* orig_ax = -1 => not a system call */
1108	PUSH_AND_CLEAR_REGS
1109	ENCODE_FRAME_POINTER
1110	jmp	error_exit
1111END(xen_failsafe_callback)
1112#endif /* CONFIG_XEN_PV */
1113
1114#ifdef CONFIG_XEN_PVHVM
1115apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1116	xen_hvm_callback_vector xen_evtchn_do_upcall
1117#endif
1118
1119
1120#if IS_ENABLED(CONFIG_HYPERV)
1121apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1122	hyperv_callback_vector hyperv_vector_handler
1123
1124apicinterrupt3 HYPERV_REENLIGHTENMENT_VECTOR \
1125	hyperv_reenlightenment_vector hyperv_reenlightenment_intr
1126
1127apicinterrupt3 HYPERV_STIMER0_VECTOR \
1128	hv_stimer0_callback_vector hv_stimer0_vector_handler
1129#endif /* CONFIG_HYPERV */
1130
1131idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=IST_INDEX_DB ist_offset=DB_STACK_OFFSET
1132idtentry int3			do_int3			has_error_code=0
1133idtentry stack_segment		do_stack_segment	has_error_code=1
1134
1135#ifdef CONFIG_XEN_PV
1136idtentry xennmi			do_nmi			has_error_code=0
1137idtentry xendebug		do_debug		has_error_code=0
1138idtentry xenint3		do_int3			has_error_code=0
1139#endif
1140
1141idtentry general_protection	do_general_protection	has_error_code=1
1142idtentry page_fault		do_page_fault		has_error_code=1
1143
1144#ifdef CONFIG_KVM_GUEST
1145idtentry async_page_fault	do_async_page_fault	has_error_code=1
1146#endif
1147
1148#ifdef CONFIG_X86_MCE
1149idtentry machine_check		do_mce			has_error_code=0	paranoid=1
1150#endif
1151
1152/*
1153 * Save all registers in pt_regs, and switch gs if needed.
1154 * Use slow, but surefire "are we in kernel?" check.
1155 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
1156 */
1157ENTRY(paranoid_entry)
1158	UNWIND_HINT_FUNC
1159	cld
1160	PUSH_AND_CLEAR_REGS save_ret=1
1161	ENCODE_FRAME_POINTER 8
1162	movl	$1, %ebx
1163	movl	$MSR_GS_BASE, %ecx
1164	rdmsr
1165	testl	%edx, %edx
1166	js	1f				/* negative -> in kernel */
1167	SWAPGS
1168	xorl	%ebx, %ebx
1169
11701:
1171	/*
1172	 * Always stash CR3 in %r14.  This value will be restored,
1173	 * verbatim, at exit.  Needed if paranoid_entry interrupted
1174	 * another entry that already switched to the user CR3 value
1175	 * but has not yet returned to userspace.
1176	 *
1177	 * This is also why CS (stashed in the "iret frame" by the
1178	 * hardware at entry) can not be used: this may be a return
1179	 * to kernel code, but with a user CR3 value.
1180	 */
1181	SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
1182
1183	ret
1184END(paranoid_entry)
1185
1186/*
1187 * "Paranoid" exit path from exception stack.  This is invoked
1188 * only on return from non-NMI IST interrupts that came
1189 * from kernel space.
1190 *
1191 * We may be returning to very strange contexts (e.g. very early
1192 * in syscall entry), so checking for preemption here would
1193 * be complicated.  Fortunately, we there's no good reason
1194 * to try to handle preemption here.
1195 *
1196 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1197 */
1198ENTRY(paranoid_exit)
1199	UNWIND_HINT_REGS
1200	DISABLE_INTERRUPTS(CLBR_ANY)
1201	TRACE_IRQS_OFF_DEBUG
1202	testl	%ebx, %ebx			/* swapgs needed? */
1203	jnz	.Lparanoid_exit_no_swapgs
1204	TRACE_IRQS_IRETQ
1205	/* Always restore stashed CR3 value (see paranoid_entry) */
1206	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1207	SWAPGS_UNSAFE_STACK
1208	jmp	.Lparanoid_exit_restore
1209.Lparanoid_exit_no_swapgs:
1210	TRACE_IRQS_IRETQ_DEBUG
1211	/* Always restore stashed CR3 value (see paranoid_entry) */
1212	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1213.Lparanoid_exit_restore:
1214	jmp restore_regs_and_return_to_kernel
1215END(paranoid_exit)
1216
1217/*
1218 * Save all registers in pt_regs, and switch GS if needed.
1219 */
1220ENTRY(error_entry)
1221	UNWIND_HINT_FUNC
1222	cld
1223	PUSH_AND_CLEAR_REGS save_ret=1
1224	ENCODE_FRAME_POINTER 8
1225	testb	$3, CS+8(%rsp)
1226	jz	.Lerror_kernelspace
1227
1228	/*
1229	 * We entered from user mode or we're pretending to have entered
1230	 * from user mode due to an IRET fault.
1231	 */
1232	SWAPGS
1233	/* We have user CR3.  Change to kernel CR3. */
1234	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1235
1236.Lerror_entry_from_usermode_after_swapgs:
1237	/* Put us onto the real thread stack. */
1238	popq	%r12				/* save return addr in %12 */
1239	movq	%rsp, %rdi			/* arg0 = pt_regs pointer */
1240	call	sync_regs
1241	movq	%rax, %rsp			/* switch stack */
1242	ENCODE_FRAME_POINTER
1243	pushq	%r12
1244
1245	/*
1246	 * We need to tell lockdep that IRQs are off.  We can't do this until
1247	 * we fix gsbase, and we should do it before enter_from_user_mode
1248	 * (which can take locks).
1249	 */
1250	TRACE_IRQS_OFF
1251	CALL_enter_from_user_mode
1252	ret
1253
1254.Lerror_entry_done:
1255	TRACE_IRQS_OFF
1256	ret
1257
1258	/*
1259	 * There are two places in the kernel that can potentially fault with
1260	 * usergs. Handle them here.  B stepping K8s sometimes report a
1261	 * truncated RIP for IRET exceptions returning to compat mode. Check
1262	 * for these here too.
1263	 */
1264.Lerror_kernelspace:
1265	leaq	native_irq_return_iret(%rip), %rcx
1266	cmpq	%rcx, RIP+8(%rsp)
1267	je	.Lerror_bad_iret
1268	movl	%ecx, %eax			/* zero extend */
1269	cmpq	%rax, RIP+8(%rsp)
1270	je	.Lbstep_iret
1271	cmpq	$.Lgs_change, RIP+8(%rsp)
1272	jne	.Lerror_entry_done
1273
1274	/*
1275	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1276	 * gsbase and proceed.  We'll fix up the exception and land in
1277	 * .Lgs_change's error handler with kernel gsbase.
1278	 */
1279	SWAPGS
1280	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1281	jmp .Lerror_entry_done
1282
1283.Lbstep_iret:
1284	/* Fix truncated RIP */
1285	movq	%rcx, RIP+8(%rsp)
1286	/* fall through */
1287
1288.Lerror_bad_iret:
1289	/*
1290	 * We came from an IRET to user mode, so we have user
1291	 * gsbase and CR3.  Switch to kernel gsbase and CR3:
1292	 */
1293	SWAPGS
1294	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1295
1296	/*
1297	 * Pretend that the exception came from user mode: set up pt_regs
1298	 * as if we faulted immediately after IRET.
1299	 */
1300	mov	%rsp, %rdi
1301	call	fixup_bad_iret
1302	mov	%rax, %rsp
1303	jmp	.Lerror_entry_from_usermode_after_swapgs
1304END(error_entry)
1305
1306ENTRY(error_exit)
1307	UNWIND_HINT_REGS
1308	DISABLE_INTERRUPTS(CLBR_ANY)
1309	TRACE_IRQS_OFF
1310	testb	$3, CS(%rsp)
1311	jz	retint_kernel
1312	jmp	retint_user
1313END(error_exit)
1314
1315/*
1316 * Runs on exception stack.  Xen PV does not go through this path at all,
1317 * so we can use real assembly here.
1318 *
1319 * Registers:
1320 *	%r14: Used to save/restore the CR3 of the interrupted context
1321 *	      when PAGE_TABLE_ISOLATION is in use.  Do not clobber.
1322 */
1323ENTRY(nmi)
1324	UNWIND_HINT_IRET_REGS
1325
1326	/*
1327	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1328	 * the iretq it performs will take us out of NMI context.
1329	 * This means that we can have nested NMIs where the next
1330	 * NMI is using the top of the stack of the previous NMI. We
1331	 * can't let it execute because the nested NMI will corrupt the
1332	 * stack of the previous NMI. NMI handlers are not re-entrant
1333	 * anyway.
1334	 *
1335	 * To handle this case we do the following:
1336	 *  Check the a special location on the stack that contains
1337	 *  a variable that is set when NMIs are executing.
1338	 *  The interrupted task's stack is also checked to see if it
1339	 *  is an NMI stack.
1340	 *  If the variable is not set and the stack is not the NMI
1341	 *  stack then:
1342	 *    o Set the special variable on the stack
1343	 *    o Copy the interrupt frame into an "outermost" location on the
1344	 *      stack
1345	 *    o Copy the interrupt frame into an "iret" location on the stack
1346	 *    o Continue processing the NMI
1347	 *  If the variable is set or the previous stack is the NMI stack:
1348	 *    o Modify the "iret" location to jump to the repeat_nmi
1349	 *    o return back to the first NMI
1350	 *
1351	 * Now on exit of the first NMI, we first clear the stack variable
1352	 * The NMI stack will tell any nested NMIs at that point that it is
1353	 * nested. Then we pop the stack normally with iret, and if there was
1354	 * a nested NMI that updated the copy interrupt stack frame, a
1355	 * jump will be made to the repeat_nmi code that will handle the second
1356	 * NMI.
1357	 *
1358	 * However, espfix prevents us from directly returning to userspace
1359	 * with a single IRET instruction.  Similarly, IRET to user mode
1360	 * can fault.  We therefore handle NMIs from user space like
1361	 * other IST entries.
1362	 */
1363
1364	ASM_CLAC
1365
1366	/* Use %rdx as our temp variable throughout */
1367	pushq	%rdx
1368
1369	testb	$3, CS-RIP+8(%rsp)
1370	jz	.Lnmi_from_kernel
1371
1372	/*
1373	 * NMI from user mode.  We need to run on the thread stack, but we
1374	 * can't go through the normal entry paths: NMIs are masked, and
1375	 * we don't want to enable interrupts, because then we'll end
1376	 * up in an awkward situation in which IRQs are on but NMIs
1377	 * are off.
1378	 *
1379	 * We also must not push anything to the stack before switching
1380	 * stacks lest we corrupt the "NMI executing" variable.
1381	 */
1382
1383	swapgs
1384	cld
1385	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx
1386	movq	%rsp, %rdx
1387	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1388	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1389	pushq	5*8(%rdx)	/* pt_regs->ss */
1390	pushq	4*8(%rdx)	/* pt_regs->rsp */
1391	pushq	3*8(%rdx)	/* pt_regs->flags */
1392	pushq	2*8(%rdx)	/* pt_regs->cs */
1393	pushq	1*8(%rdx)	/* pt_regs->rip */
1394	UNWIND_HINT_IRET_REGS
1395	pushq   $-1		/* pt_regs->orig_ax */
1396	PUSH_AND_CLEAR_REGS rdx=(%rdx)
1397	ENCODE_FRAME_POINTER
1398
1399	/*
1400	 * At this point we no longer need to worry about stack damage
1401	 * due to nesting -- we're on the normal thread stack and we're
1402	 * done with the NMI stack.
1403	 */
1404
1405	movq	%rsp, %rdi
1406	movq	$-1, %rsi
1407	call	do_nmi
1408
1409	/*
1410	 * Return back to user mode.  We must *not* do the normal exit
1411	 * work, because we don't want to enable interrupts.
1412	 */
1413	jmp	swapgs_restore_regs_and_return_to_usermode
1414
1415.Lnmi_from_kernel:
1416	/*
1417	 * Here's what our stack frame will look like:
1418	 * +---------------------------------------------------------+
1419	 * | original SS                                             |
1420	 * | original Return RSP                                     |
1421	 * | original RFLAGS                                         |
1422	 * | original CS                                             |
1423	 * | original RIP                                            |
1424	 * +---------------------------------------------------------+
1425	 * | temp storage for rdx                                    |
1426	 * +---------------------------------------------------------+
1427	 * | "NMI executing" variable                                |
1428	 * +---------------------------------------------------------+
1429	 * | iret SS          } Copied from "outermost" frame        |
1430	 * | iret Return RSP  } on each loop iteration; overwritten  |
1431	 * | iret RFLAGS      } by a nested NMI to force another     |
1432	 * | iret CS          } iteration if needed.                 |
1433	 * | iret RIP         }                                      |
1434	 * +---------------------------------------------------------+
1435	 * | outermost SS          } initialized in first_nmi;       |
1436	 * | outermost Return RSP  } will not be changed before      |
1437	 * | outermost RFLAGS      } NMI processing is done.         |
1438	 * | outermost CS          } Copied to "iret" frame on each  |
1439	 * | outermost RIP         } iteration.                      |
1440	 * +---------------------------------------------------------+
1441	 * | pt_regs                                                 |
1442	 * +---------------------------------------------------------+
1443	 *
1444	 * The "original" frame is used by hardware.  Before re-enabling
1445	 * NMIs, we need to be done with it, and we need to leave enough
1446	 * space for the asm code here.
1447	 *
1448	 * We return by executing IRET while RSP points to the "iret" frame.
1449	 * That will either return for real or it will loop back into NMI
1450	 * processing.
1451	 *
1452	 * The "outermost" frame is copied to the "iret" frame on each
1453	 * iteration of the loop, so each iteration starts with the "iret"
1454	 * frame pointing to the final return target.
1455	 */
1456
1457	/*
1458	 * Determine whether we're a nested NMI.
1459	 *
1460	 * If we interrupted kernel code between repeat_nmi and
1461	 * end_repeat_nmi, then we are a nested NMI.  We must not
1462	 * modify the "iret" frame because it's being written by
1463	 * the outer NMI.  That's okay; the outer NMI handler is
1464	 * about to about to call do_nmi anyway, so we can just
1465	 * resume the outer NMI.
1466	 */
1467
1468	movq	$repeat_nmi, %rdx
1469	cmpq	8(%rsp), %rdx
1470	ja	1f
1471	movq	$end_repeat_nmi, %rdx
1472	cmpq	8(%rsp), %rdx
1473	ja	nested_nmi_out
14741:
1475
1476	/*
1477	 * Now check "NMI executing".  If it's set, then we're nested.
1478	 * This will not detect if we interrupted an outer NMI just
1479	 * before IRET.
1480	 */
1481	cmpl	$1, -8(%rsp)
1482	je	nested_nmi
1483
1484	/*
1485	 * Now test if the previous stack was an NMI stack.  This covers
1486	 * the case where we interrupt an outer NMI after it clears
1487	 * "NMI executing" but before IRET.  We need to be careful, though:
1488	 * there is one case in which RSP could point to the NMI stack
1489	 * despite there being no NMI active: naughty userspace controls
1490	 * RSP at the very beginning of the SYSCALL targets.  We can
1491	 * pull a fast one on naughty userspace, though: we program
1492	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1493	 * if it controls the kernel's RSP.  We set DF before we clear
1494	 * "NMI executing".
1495	 */
1496	lea	6*8(%rsp), %rdx
1497	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1498	cmpq	%rdx, 4*8(%rsp)
1499	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1500	ja	first_nmi
1501
1502	subq	$EXCEPTION_STKSZ, %rdx
1503	cmpq	%rdx, 4*8(%rsp)
1504	/* If it is below the NMI stack, it is a normal NMI */
1505	jb	first_nmi
1506
1507	/* Ah, it is within the NMI stack. */
1508
1509	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1510	jz	first_nmi	/* RSP was user controlled. */
1511
1512	/* This is a nested NMI. */
1513
1514nested_nmi:
1515	/*
1516	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1517	 * iteration of NMI handling.
1518	 */
1519	subq	$8, %rsp
1520	leaq	-10*8(%rsp), %rdx
1521	pushq	$__KERNEL_DS
1522	pushq	%rdx
1523	pushfq
1524	pushq	$__KERNEL_CS
1525	pushq	$repeat_nmi
1526
1527	/* Put stack back */
1528	addq	$(6*8), %rsp
1529
1530nested_nmi_out:
1531	popq	%rdx
1532
1533	/* We are returning to kernel mode, so this cannot result in a fault. */
1534	iretq
1535
1536first_nmi:
1537	/* Restore rdx. */
1538	movq	(%rsp), %rdx
1539
1540	/* Make room for "NMI executing". */
1541	pushq	$0
1542
1543	/* Leave room for the "iret" frame */
1544	subq	$(5*8), %rsp
1545
1546	/* Copy the "original" frame to the "outermost" frame */
1547	.rept 5
1548	pushq	11*8(%rsp)
1549	.endr
1550	UNWIND_HINT_IRET_REGS
1551
1552	/* Everything up to here is safe from nested NMIs */
1553
1554#ifdef CONFIG_DEBUG_ENTRY
1555	/*
1556	 * For ease of testing, unmask NMIs right away.  Disabled by
1557	 * default because IRET is very expensive.
1558	 */
1559	pushq	$0		/* SS */
1560	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1561	addq	$8, (%rsp)	/* Fix up RSP */
1562	pushfq			/* RFLAGS */
1563	pushq	$__KERNEL_CS	/* CS */
1564	pushq	$1f		/* RIP */
1565	iretq			/* continues at repeat_nmi below */
1566	UNWIND_HINT_IRET_REGS
15671:
1568#endif
1569
1570repeat_nmi:
1571	/*
1572	 * If there was a nested NMI, the first NMI's iret will return
1573	 * here. But NMIs are still enabled and we can take another
1574	 * nested NMI. The nested NMI checks the interrupted RIP to see
1575	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1576	 * it will just return, as we are about to repeat an NMI anyway.
1577	 * This makes it safe to copy to the stack frame that a nested
1578	 * NMI will update.
1579	 *
1580	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1581	 * we're repeating an NMI, gsbase has the same value that it had on
1582	 * the first iteration.  paranoid_entry will load the kernel
1583	 * gsbase if needed before we call do_nmi.  "NMI executing"
1584	 * is zero.
1585	 */
1586	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1587
1588	/*
1589	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1590	 * here must not modify the "iret" frame while we're writing to
1591	 * it or it will end up containing garbage.
1592	 */
1593	addq	$(10*8), %rsp
1594	.rept 5
1595	pushq	-6*8(%rsp)
1596	.endr
1597	subq	$(5*8), %rsp
1598end_repeat_nmi:
1599
1600	/*
1601	 * Everything below this point can be preempted by a nested NMI.
1602	 * If this happens, then the inner NMI will change the "iret"
1603	 * frame to point back to repeat_nmi.
1604	 */
1605	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1606
1607	/*
1608	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1609	 * as we should not be calling schedule in NMI context.
1610	 * Even with normal interrupts enabled. An NMI should not be
1611	 * setting NEED_RESCHED or anything that normal interrupts and
1612	 * exceptions might do.
1613	 */
1614	call	paranoid_entry
1615	UNWIND_HINT_REGS
1616
1617	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1618	movq	%rsp, %rdi
1619	movq	$-1, %rsi
1620	call	do_nmi
1621
1622	/* Always restore stashed CR3 value (see paranoid_entry) */
1623	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
1624
1625	testl	%ebx, %ebx			/* swapgs needed? */
1626	jnz	nmi_restore
1627nmi_swapgs:
1628	SWAPGS_UNSAFE_STACK
1629nmi_restore:
1630	POP_REGS
1631
1632	/*
1633	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1634	 * at the "iret" frame.
1635	 */
1636	addq	$6*8, %rsp
1637
1638	/*
1639	 * Clear "NMI executing".  Set DF first so that we can easily
1640	 * distinguish the remaining code between here and IRET from
1641	 * the SYSCALL entry and exit paths.
1642	 *
1643	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1644	 * this code when I had the misapprehension that Xen PV supported
1645	 * NMIs, and Xen PV would break that approach.
1646	 */
1647	std
1648	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1649
1650	/*
1651	 * iretq reads the "iret" frame and exits the NMI stack in a
1652	 * single instruction.  We are returning to kernel mode, so this
1653	 * cannot result in a fault.  Similarly, we don't need to worry
1654	 * about espfix64 on the way back to kernel mode.
1655	 */
1656	iretq
1657END(nmi)
1658
1659ENTRY(ignore_sysret)
1660	UNWIND_HINT_EMPTY
1661	mov	$-ENOSYS, %eax
1662	sysret
1663END(ignore_sysret)
1664
1665ENTRY(rewind_stack_do_exit)
1666	UNWIND_HINT_FUNC
1667	/* Prevent any naive code from trying to unwind to our caller. */
1668	xorl	%ebp, %ebp
1669
1670	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1671	leaq	-PTREGS_SIZE(%rax), %rsp
1672	UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE
1673
1674	call	do_exit
1675END(rewind_stack_do_exit)
1676