xref: /openbmc/linux/arch/x86/entry/entry_64.S (revision b240b419db5d624ce7a5a397d6f62a1a686009ec)
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	bt	$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
145	.pushsection .entry_trampoline, "ax"
146
147/*
148 * The code in here gets remapped into cpu_entry_area's trampoline.  This means
149 * that the assembler and linker have the wrong idea as to where this code
150 * lives (and, in fact, it's mapped more than once, so it's not even at a
151 * fixed address).  So we can't reference any symbols outside the entry
152 * trampoline and expect it to work.
153 *
154 * Instead, we carefully abuse %rip-relative addressing.
155 * _entry_trampoline(%rip) refers to the start of the remapped) entry
156 * trampoline.  We can thus find cpu_entry_area with this macro:
157 */
158
159#define CPU_ENTRY_AREA \
160	_entry_trampoline - CPU_ENTRY_AREA_entry_trampoline(%rip)
161
162/* The top word of the SYSENTER stack is hot and is usable as scratch space. */
163#define RSP_SCRATCH	CPU_ENTRY_AREA_entry_stack + \
164			SIZEOF_entry_stack - 8 + CPU_ENTRY_AREA
165
166ENTRY(entry_SYSCALL_64_trampoline)
167	UNWIND_HINT_EMPTY
168	swapgs
169
170	/* Stash the user RSP. */
171	movq	%rsp, RSP_SCRATCH
172
173	/* Note: using %rsp as a scratch reg. */
174	SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
175
176	/* Load the top of the task stack into RSP */
177	movq	CPU_ENTRY_AREA_tss + TSS_sp1 + CPU_ENTRY_AREA, %rsp
178
179	/* Start building the simulated IRET frame. */
180	pushq	$__USER_DS			/* pt_regs->ss */
181	pushq	RSP_SCRATCH			/* pt_regs->sp */
182	pushq	%r11				/* pt_regs->flags */
183	pushq	$__USER_CS			/* pt_regs->cs */
184	pushq	%rcx				/* pt_regs->ip */
185
186	/*
187	 * x86 lacks a near absolute jump, and we can't jump to the real
188	 * entry text with a relative jump.  We could push the target
189	 * address and then use retq, but this destroys the pipeline on
190	 * many CPUs (wasting over 20 cycles on Sandy Bridge).  Instead,
191	 * spill RDI and restore it in a second-stage trampoline.
192	 */
193	pushq	%rdi
194	movq	$entry_SYSCALL_64_stage2, %rdi
195	JMP_NOSPEC %rdi
196END(entry_SYSCALL_64_trampoline)
197
198	.popsection
199
200ENTRY(entry_SYSCALL_64_stage2)
201	UNWIND_HINT_EMPTY
202	popq	%rdi
203	jmp	entry_SYSCALL_64_after_hwframe
204END(entry_SYSCALL_64_stage2)
205
206ENTRY(entry_SYSCALL_64)
207	UNWIND_HINT_EMPTY
208	/*
209	 * Interrupts are off on entry.
210	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
211	 * it is too small to ever cause noticeable irq latency.
212	 */
213
214	swapgs
215	/*
216	 * This path is only taken when PAGE_TABLE_ISOLATION is disabled so it
217	 * is not required to switch CR3.
218	 */
219	movq	%rsp, PER_CPU_VAR(rsp_scratch)
220	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
221
222	/* Construct struct pt_regs on stack */
223	pushq	$__USER_DS			/* pt_regs->ss */
224	pushq	PER_CPU_VAR(rsp_scratch)	/* pt_regs->sp */
225	pushq	%r11				/* pt_regs->flags */
226	pushq	$__USER_CS			/* pt_regs->cs */
227	pushq	%rcx				/* pt_regs->ip */
228GLOBAL(entry_SYSCALL_64_after_hwframe)
229	pushq	%rax				/* pt_regs->orig_ax */
230
231	PUSH_AND_CLEAR_REGS rax=$-ENOSYS
232
233	TRACE_IRQS_OFF
234
235	/* IRQs are off. */
236	movq	%rsp, %rdi
237	call	do_syscall_64		/* returns with IRQs disabled */
238
239	TRACE_IRQS_IRETQ		/* we're about to change IF */
240
241	/*
242	 * Try to use SYSRET instead of IRET if we're returning to
243	 * a completely clean 64-bit userspace context.  If we're not,
244	 * go to the slow exit path.
245	 */
246	movq	RCX(%rsp), %rcx
247	movq	RIP(%rsp), %r11
248
249	cmpq	%rcx, %r11	/* SYSRET requires RCX == RIP */
250	jne	swapgs_restore_regs_and_return_to_usermode
251
252	/*
253	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
254	 * in kernel space.  This essentially lets the user take over
255	 * the kernel, since userspace controls RSP.
256	 *
257	 * If width of "canonical tail" ever becomes variable, this will need
258	 * to be updated to remain correct on both old and new CPUs.
259	 *
260	 * Change top bits to match most significant bit (47th or 56th bit
261	 * depending on paging mode) in the address.
262	 */
263#ifdef CONFIG_X86_5LEVEL
264	ALTERNATIVE "shl $(64 - 48), %rcx; sar $(64 - 48), %rcx", \
265		"shl $(64 - 57), %rcx; sar $(64 - 57), %rcx", X86_FEATURE_LA57
266#else
267	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
268	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
269#endif
270
271	/* If this changed %rcx, it was not canonical */
272	cmpq	%rcx, %r11
273	jne	swapgs_restore_regs_and_return_to_usermode
274
275	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
276	jne	swapgs_restore_regs_and_return_to_usermode
277
278	movq	R11(%rsp), %r11
279	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
280	jne	swapgs_restore_regs_and_return_to_usermode
281
282	/*
283	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
284	 * restore RF properly. If the slowpath sets it for whatever reason, we
285	 * need to restore it correctly.
286	 *
287	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
288	 * trap from userspace immediately after SYSRET.  This would cause an
289	 * infinite loop whenever #DB happens with register state that satisfies
290	 * the opportunistic SYSRET conditions.  For example, single-stepping
291	 * this user code:
292	 *
293	 *           movq	$stuck_here, %rcx
294	 *           pushfq
295	 *           popq %r11
296	 *   stuck_here:
297	 *
298	 * would never get past 'stuck_here'.
299	 */
300	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
301	jnz	swapgs_restore_regs_and_return_to_usermode
302
303	/* nothing to check for RSP */
304
305	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
306	jne	swapgs_restore_regs_and_return_to_usermode
307
308	/*
309	 * We win! This label is here just for ease of understanding
310	 * perf profiles. Nothing jumps here.
311	 */
312syscall_return_via_sysret:
313	/* rcx and r11 are already restored (see code above) */
314	UNWIND_HINT_EMPTY
315	POP_REGS pop_rdi=0 skip_r11rcx=1
316
317	/*
318	 * Now all regs are restored except RSP and RDI.
319	 * Save old stack pointer and switch to trampoline stack.
320	 */
321	movq	%rsp, %rdi
322	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
323
324	pushq	RSP-RDI(%rdi)	/* RSP */
325	pushq	(%rdi)		/* RDI */
326
327	/*
328	 * We are on the trampoline stack.  All regs except RDI are live.
329	 * We can do future final exit work right here.
330	 */
331	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
332
333	popq	%rdi
334	popq	%rsp
335	USERGS_SYSRET64
336END(entry_SYSCALL_64)
337
338/*
339 * %rdi: prev task
340 * %rsi: next task
341 */
342ENTRY(__switch_to_asm)
343	UNWIND_HINT_FUNC
344	/*
345	 * Save callee-saved registers
346	 * This must match the order in inactive_task_frame
347	 */
348	pushq	%rbp
349	pushq	%rbx
350	pushq	%r12
351	pushq	%r13
352	pushq	%r14
353	pushq	%r15
354
355	/* switch stack */
356	movq	%rsp, TASK_threadsp(%rdi)
357	movq	TASK_threadsp(%rsi), %rsp
358
359#ifdef CONFIG_CC_STACKPROTECTOR
360	movq	TASK_stack_canary(%rsi), %rbx
361	movq	%rbx, PER_CPU_VAR(irq_stack_union)+stack_canary_offset
362#endif
363
364#ifdef CONFIG_RETPOLINE
365	/*
366	 * When switching from a shallower to a deeper call stack
367	 * the RSB may either underflow or use entries populated
368	 * with userspace addresses. On CPUs where those concerns
369	 * exist, overwrite the RSB with entries which capture
370	 * speculative execution to prevent attack.
371	 */
372	FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW
373#endif
374
375	/* restore callee-saved registers */
376	popq	%r15
377	popq	%r14
378	popq	%r13
379	popq	%r12
380	popq	%rbx
381	popq	%rbp
382
383	jmp	__switch_to
384END(__switch_to_asm)
385
386/*
387 * A newly forked process directly context switches into this address.
388 *
389 * rax: prev task we switched from
390 * rbx: kernel thread func (NULL for user thread)
391 * r12: kernel thread arg
392 */
393ENTRY(ret_from_fork)
394	UNWIND_HINT_EMPTY
395	movq	%rax, %rdi
396	call	schedule_tail			/* rdi: 'prev' task parameter */
397
398	testq	%rbx, %rbx			/* from kernel_thread? */
399	jnz	1f				/* kernel threads are uncommon */
400
4012:
402	UNWIND_HINT_REGS
403	movq	%rsp, %rdi
404	call	syscall_return_slowpath	/* returns with IRQs disabled */
405	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
406	jmp	swapgs_restore_regs_and_return_to_usermode
407
4081:
409	/* kernel thread */
410	movq	%r12, %rdi
411	CALL_NOSPEC %rbx
412	/*
413	 * A kernel thread is allowed to return here after successfully
414	 * calling do_execve().  Exit to userspace to complete the execve()
415	 * syscall.
416	 */
417	movq	$0, RAX(%rsp)
418	jmp	2b
419END(ret_from_fork)
420
421/*
422 * Build the entry stubs with some assembler magic.
423 * We pack 1 stub into every 8-byte block.
424 */
425	.align 8
426ENTRY(irq_entries_start)
427    vector=FIRST_EXTERNAL_VECTOR
428    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
429	UNWIND_HINT_IRET_REGS
430	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
431	jmp	common_interrupt
432	.align	8
433	vector=vector+1
434    .endr
435END(irq_entries_start)
436
437.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
438#ifdef CONFIG_DEBUG_ENTRY
439	pushq %rax
440	SAVE_FLAGS(CLBR_RAX)
441	testl $X86_EFLAGS_IF, %eax
442	jz .Lokay_\@
443	ud2
444.Lokay_\@:
445	popq %rax
446#endif
447.endm
448
449/*
450 * Enters the IRQ stack if we're not already using it.  NMI-safe.  Clobbers
451 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone.
452 * Requires kernel GSBASE.
453 *
454 * The invariant is that, if irq_count != -1, then the IRQ stack is in use.
455 */
456.macro ENTER_IRQ_STACK regs=1 old_rsp save_ret=0
457	DEBUG_ENTRY_ASSERT_IRQS_OFF
458
459	.if \save_ret
460	/*
461	 * If save_ret is set, the original stack contains one additional
462	 * entry -- the return address. Therefore, move the address one
463	 * entry below %rsp to \old_rsp.
464	 */
465	leaq	8(%rsp), \old_rsp
466	.else
467	movq	%rsp, \old_rsp
468	.endif
469
470	.if \regs
471	UNWIND_HINT_REGS base=\old_rsp
472	.endif
473
474	incl	PER_CPU_VAR(irq_count)
475	jnz	.Lirq_stack_push_old_rsp_\@
476
477	/*
478	 * Right now, if we just incremented irq_count to zero, we've
479	 * claimed the IRQ stack but we haven't switched to it yet.
480	 *
481	 * If anything is added that can interrupt us here without using IST,
482	 * it must be *extremely* careful to limit its stack usage.  This
483	 * could include kprobes and a hypothetical future IST-less #DB
484	 * handler.
485	 *
486	 * The OOPS unwinder relies on the word at the top of the IRQ
487	 * stack linking back to the previous RSP for the entire time we're
488	 * on the IRQ stack.  For this to work reliably, we need to write
489	 * it before we actually move ourselves to the IRQ stack.
490	 */
491
492	movq	\old_rsp, PER_CPU_VAR(irq_stack_union + IRQ_STACK_SIZE - 8)
493	movq	PER_CPU_VAR(irq_stack_ptr), %rsp
494
495#ifdef CONFIG_DEBUG_ENTRY
496	/*
497	 * If the first movq above becomes wrong due to IRQ stack layout
498	 * changes, the only way we'll notice is if we try to unwind right
499	 * here.  Assert that we set up the stack right to catch this type
500	 * of bug quickly.
501	 */
502	cmpq	-8(%rsp), \old_rsp
503	je	.Lirq_stack_okay\@
504	ud2
505	.Lirq_stack_okay\@:
506#endif
507
508.Lirq_stack_push_old_rsp_\@:
509	pushq	\old_rsp
510
511	.if \regs
512	UNWIND_HINT_REGS indirect=1
513	.endif
514
515	.if \save_ret
516	/*
517	 * Push the return address to the stack. This return address can
518	 * be found at the "real" original RSP, which was offset by 8 at
519	 * the beginning of this macro.
520	 */
521	pushq	-8(\old_rsp)
522	.endif
523.endm
524
525/*
526 * Undoes ENTER_IRQ_STACK.
527 */
528.macro LEAVE_IRQ_STACK regs=1
529	DEBUG_ENTRY_ASSERT_IRQS_OFF
530	/* We need to be off the IRQ stack before decrementing irq_count. */
531	popq	%rsp
532
533	.if \regs
534	UNWIND_HINT_REGS
535	.endif
536
537	/*
538	 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming
539	 * the irq stack but we're not on it.
540	 */
541
542	decl	PER_CPU_VAR(irq_count)
543.endm
544
545/*
546 * Interrupt entry helper function.
547 *
548 * Entry runs with interrupts off. Stack layout at entry:
549 * +----------------------------------------------------+
550 * | regs->ss						|
551 * | regs->rsp						|
552 * | regs->eflags					|
553 * | regs->cs						|
554 * | regs->ip						|
555 * +----------------------------------------------------+
556 * | regs->orig_ax = ~(interrupt number)		|
557 * +----------------------------------------------------+
558 * | return address					|
559 * +----------------------------------------------------+
560 */
561ENTRY(interrupt_entry)
562	UNWIND_HINT_FUNC
563	ASM_CLAC
564	cld
565
566	testb	$3, CS-ORIG_RAX+8(%rsp)
567	jz	1f
568	SWAPGS
569
570	/*
571	 * Switch to the thread stack. The IRET frame and orig_ax are
572	 * on the stack, as well as the return address. RDI..R12 are
573	 * not (yet) on the stack and space has not (yet) been
574	 * allocated for them.
575	 */
576	pushq	%rdi
577
578	/* Need to switch before accessing the thread stack. */
579	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi
580	movq	%rsp, %rdi
581	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
582
583	 /*
584	  * We have RDI, return address, and orig_ax on the stack on
585	  * top of the IRET frame. That means offset=24
586	  */
587	UNWIND_HINT_IRET_REGS base=%rdi offset=24
588
589	pushq	7*8(%rdi)		/* regs->ss */
590	pushq	6*8(%rdi)		/* regs->rsp */
591	pushq	5*8(%rdi)		/* regs->eflags */
592	pushq	4*8(%rdi)		/* regs->cs */
593	pushq	3*8(%rdi)		/* regs->ip */
594	pushq	2*8(%rdi)		/* regs->orig_ax */
595	pushq	8(%rdi)			/* return address */
596	UNWIND_HINT_FUNC
597
598	movq	(%rdi), %rdi
5991:
600
601	PUSH_AND_CLEAR_REGS save_ret=1
602	ENCODE_FRAME_POINTER 8
603
604	testb	$3, CS+8(%rsp)
605	jz	1f
606
607	/*
608	 * IRQ from user mode.
609	 *
610	 * We need to tell lockdep that IRQs are off.  We can't do this until
611	 * we fix gsbase, and we should do it before enter_from_user_mode
612	 * (which can take locks).  Since TRACE_IRQS_OFF is idempotent,
613	 * the simplest way to handle it is to just call it twice if
614	 * we enter from user mode.  There's no reason to optimize this since
615	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
616	 */
617	TRACE_IRQS_OFF
618
619	CALL_enter_from_user_mode
620
6211:
622	ENTER_IRQ_STACK old_rsp=%rdi save_ret=1
623	/* We entered an interrupt context - irqs are off: */
624	TRACE_IRQS_OFF
625
626	ret
627END(interrupt_entry)
628
629
630/* Interrupt entry/exit. */
631
632	/*
633	 * The interrupt stubs push (~vector+0x80) onto the stack and
634	 * then jump to common_interrupt.
635	 */
636	.p2align CONFIG_X86_L1_CACHE_SHIFT
637common_interrupt:
638	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
639	call	interrupt_entry
640	UNWIND_HINT_REGS indirect=1
641	call	do_IRQ	/* rdi points to pt_regs */
642	/* 0(%rsp): old RSP */
643ret_from_intr:
644	DISABLE_INTERRUPTS(CLBR_ANY)
645	TRACE_IRQS_OFF
646
647	LEAVE_IRQ_STACK
648
649	testb	$3, CS(%rsp)
650	jz	retint_kernel
651
652	/* Interrupt came from user space */
653GLOBAL(retint_user)
654	mov	%rsp,%rdi
655	call	prepare_exit_to_usermode
656	TRACE_IRQS_IRETQ
657
658GLOBAL(swapgs_restore_regs_and_return_to_usermode)
659#ifdef CONFIG_DEBUG_ENTRY
660	/* Assert that pt_regs indicates user mode. */
661	testb	$3, CS(%rsp)
662	jnz	1f
663	ud2
6641:
665#endif
666	POP_REGS pop_rdi=0
667
668	/*
669	 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS.
670	 * Save old stack pointer and switch to trampoline stack.
671	 */
672	movq	%rsp, %rdi
673	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
674
675	/* Copy the IRET frame to the trampoline stack. */
676	pushq	6*8(%rdi)	/* SS */
677	pushq	5*8(%rdi)	/* RSP */
678	pushq	4*8(%rdi)	/* EFLAGS */
679	pushq	3*8(%rdi)	/* CS */
680	pushq	2*8(%rdi)	/* RIP */
681
682	/* Push user RDI on the trampoline stack. */
683	pushq	(%rdi)
684
685	/*
686	 * We are on the trampoline stack.  All regs except RDI are live.
687	 * We can do future final exit work right here.
688	 */
689
690	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
691
692	/* Restore RDI. */
693	popq	%rdi
694	SWAPGS
695	INTERRUPT_RETURN
696
697
698/* Returning to kernel space */
699retint_kernel:
700#ifdef CONFIG_PREEMPT
701	/* Interrupts are off */
702	/* Check if we need preemption */
703	bt	$9, EFLAGS(%rsp)		/* were interrupts off? */
704	jnc	1f
7050:	cmpl	$0, PER_CPU_VAR(__preempt_count)
706	jnz	1f
707	call	preempt_schedule_irq
708	jmp	0b
7091:
710#endif
711	/*
712	 * The iretq could re-enable interrupts:
713	 */
714	TRACE_IRQS_IRETQ
715
716GLOBAL(restore_regs_and_return_to_kernel)
717#ifdef CONFIG_DEBUG_ENTRY
718	/* Assert that pt_regs indicates kernel mode. */
719	testb	$3, CS(%rsp)
720	jz	1f
721	ud2
7221:
723#endif
724	POP_REGS
725	addq	$8, %rsp	/* skip regs->orig_ax */
726	/*
727	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
728	 * when returning from IPI handler.
729	 */
730	INTERRUPT_RETURN
731
732ENTRY(native_iret)
733	UNWIND_HINT_IRET_REGS
734	/*
735	 * Are we returning to a stack segment from the LDT?  Note: in
736	 * 64-bit mode SS:RSP on the exception stack is always valid.
737	 */
738#ifdef CONFIG_X86_ESPFIX64
739	testb	$4, (SS-RIP)(%rsp)
740	jnz	native_irq_return_ldt
741#endif
742
743.global native_irq_return_iret
744native_irq_return_iret:
745	/*
746	 * This may fault.  Non-paranoid faults on return to userspace are
747	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
748	 * Double-faults due to espfix64 are handled in do_double_fault.
749	 * Other faults here are fatal.
750	 */
751	iretq
752
753#ifdef CONFIG_X86_ESPFIX64
754native_irq_return_ldt:
755	/*
756	 * We are running with user GSBASE.  All GPRs contain their user
757	 * values.  We have a percpu ESPFIX stack that is eight slots
758	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
759	 * of the ESPFIX stack.
760	 *
761	 * We clobber RAX and RDI in this code.  We stash RDI on the
762	 * normal stack and RAX on the ESPFIX stack.
763	 *
764	 * The ESPFIX stack layout we set up looks like this:
765	 *
766	 * --- top of ESPFIX stack ---
767	 * SS
768	 * RSP
769	 * RFLAGS
770	 * CS
771	 * RIP  <-- RSP points here when we're done
772	 * RAX  <-- espfix_waddr points here
773	 * --- bottom of ESPFIX stack ---
774	 */
775
776	pushq	%rdi				/* Stash user RDI */
777	SWAPGS					/* to kernel GS */
778	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi	/* to kernel CR3 */
779
780	movq	PER_CPU_VAR(espfix_waddr), %rdi
781	movq	%rax, (0*8)(%rdi)		/* user RAX */
782	movq	(1*8)(%rsp), %rax		/* user RIP */
783	movq	%rax, (1*8)(%rdi)
784	movq	(2*8)(%rsp), %rax		/* user CS */
785	movq	%rax, (2*8)(%rdi)
786	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
787	movq	%rax, (3*8)(%rdi)
788	movq	(5*8)(%rsp), %rax		/* user SS */
789	movq	%rax, (5*8)(%rdi)
790	movq	(4*8)(%rsp), %rax		/* user RSP */
791	movq	%rax, (4*8)(%rdi)
792	/* Now RAX == RSP. */
793
794	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
795
796	/*
797	 * espfix_stack[31:16] == 0.  The page tables are set up such that
798	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
799	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
800	 * the same page.  Set up RSP so that RSP[31:16] contains the
801	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
802	 * still points to an RO alias of the ESPFIX stack.
803	 */
804	orq	PER_CPU_VAR(espfix_stack), %rax
805
806	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
807	SWAPGS					/* to user GS */
808	popq	%rdi				/* Restore user RDI */
809
810	movq	%rax, %rsp
811	UNWIND_HINT_IRET_REGS offset=8
812
813	/*
814	 * At this point, we cannot write to the stack any more, but we can
815	 * still read.
816	 */
817	popq	%rax				/* Restore user RAX */
818
819	/*
820	 * RSP now points to an ordinary IRET frame, except that the page
821	 * is read-only and RSP[31:16] are preloaded with the userspace
822	 * values.  We can now IRET back to userspace.
823	 */
824	jmp	native_irq_return_iret
825#endif
826END(common_interrupt)
827
828/*
829 * APIC interrupts.
830 */
831.macro apicinterrupt3 num sym do_sym
832ENTRY(\sym)
833	UNWIND_HINT_IRET_REGS
834	pushq	$~(\num)
835.Lcommon_\sym:
836	call	interrupt_entry
837	UNWIND_HINT_REGS indirect=1
838	call	\do_sym	/* rdi points to pt_regs */
839	jmp	ret_from_intr
840END(\sym)
841.endm
842
843/* Make sure APIC interrupt handlers end up in the irqentry section: */
844#define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
845#define POP_SECTION_IRQENTRY	.popsection
846
847.macro apicinterrupt num sym do_sym
848PUSH_SECTION_IRQENTRY
849apicinterrupt3 \num \sym \do_sym
850POP_SECTION_IRQENTRY
851.endm
852
853#ifdef CONFIG_SMP
854apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
855apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
856#endif
857
858#ifdef CONFIG_X86_UV
859apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
860#endif
861
862apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
863apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
864
865#ifdef CONFIG_HAVE_KVM
866apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
867apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
868apicinterrupt3 POSTED_INTR_NESTED_VECTOR	kvm_posted_intr_nested_ipi	smp_kvm_posted_intr_nested_ipi
869#endif
870
871#ifdef CONFIG_X86_MCE_THRESHOLD
872apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
873#endif
874
875#ifdef CONFIG_X86_MCE_AMD
876apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
877#endif
878
879#ifdef CONFIG_X86_THERMAL_VECTOR
880apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
881#endif
882
883#ifdef CONFIG_SMP
884apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
885apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
886apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
887#endif
888
889apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
890apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
891
892#ifdef CONFIG_IRQ_WORK
893apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
894#endif
895
896/*
897 * Exception entry points.
898 */
899#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss_rw) + (TSS_ist + ((x) - 1) * 8)
900
901.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
902ENTRY(\sym)
903	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
904
905	/* Sanity check */
906	.if \shift_ist != -1 && \paranoid == 0
907	.error "using shift_ist requires paranoid=1"
908	.endif
909
910	ASM_CLAC
911
912	.if \has_error_code == 0
913	pushq	$-1				/* ORIG_RAX: no syscall to restart */
914	.endif
915
916	.if \paranoid < 2
917	testb	$3, CS-ORIG_RAX(%rsp)		/* If coming from userspace, switch stacks */
918	jnz	.Lfrom_usermode_switch_stack_\@
919	.endif
920
921	.if \paranoid
922	call	paranoid_entry
923	.else
924	call	error_entry
925	.endif
926	UNWIND_HINT_REGS
927	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
928
929	.if \paranoid
930	.if \shift_ist != -1
931	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
932	.else
933	TRACE_IRQS_OFF
934	.endif
935	.endif
936
937	movq	%rsp, %rdi			/* pt_regs pointer */
938
939	.if \has_error_code
940	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
941	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
942	.else
943	xorl	%esi, %esi			/* no error code */
944	.endif
945
946	.if \shift_ist != -1
947	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
948	.endif
949
950	call	\do_sym
951
952	.if \shift_ist != -1
953	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
954	.endif
955
956	/* these procedures expect "no swapgs" flag in ebx */
957	.if \paranoid
958	jmp	paranoid_exit
959	.else
960	jmp	error_exit
961	.endif
962
963	.if \paranoid < 2
964	/*
965	 * Entry from userspace.  Switch stacks and treat it
966	 * as a normal entry.  This means that paranoid handlers
967	 * run in real process context if user_mode(regs).
968	 */
969.Lfrom_usermode_switch_stack_\@:
970	call	error_entry
971
972	movq	%rsp, %rdi			/* pt_regs pointer */
973
974	.if \has_error_code
975	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
976	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
977	.else
978	xorl	%esi, %esi			/* no error code */
979	.endif
980
981	call	\do_sym
982
983	jmp	error_exit			/* %ebx: no swapgs flag */
984	.endif
985END(\sym)
986.endm
987
988idtentry divide_error			do_divide_error			has_error_code=0
989idtentry overflow			do_overflow			has_error_code=0
990idtentry bounds				do_bounds			has_error_code=0
991idtentry invalid_op			do_invalid_op			has_error_code=0
992idtentry device_not_available		do_device_not_available		has_error_code=0
993idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
994idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
995idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
996idtentry segment_not_present		do_segment_not_present		has_error_code=1
997idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
998idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
999idtentry alignment_check		do_alignment_check		has_error_code=1
1000idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
1001
1002
1003	/*
1004	 * Reload gs selector with exception handling
1005	 * edi:  new selector
1006	 */
1007ENTRY(native_load_gs_index)
1008	FRAME_BEGIN
1009	pushfq
1010	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
1011	TRACE_IRQS_OFF
1012	SWAPGS
1013.Lgs_change:
1014	movl	%edi, %gs
10152:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
1016	SWAPGS
1017	TRACE_IRQS_FLAGS (%rsp)
1018	popfq
1019	FRAME_END
1020	ret
1021ENDPROC(native_load_gs_index)
1022EXPORT_SYMBOL(native_load_gs_index)
1023
1024	_ASM_EXTABLE(.Lgs_change, bad_gs)
1025	.section .fixup, "ax"
1026	/* running with kernelgs */
1027bad_gs:
1028	SWAPGS					/* switch back to user gs */
1029.macro ZAP_GS
1030	/* This can't be a string because the preprocessor needs to see it. */
1031	movl $__USER_DS, %eax
1032	movl %eax, %gs
1033.endm
1034	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
1035	xorl	%eax, %eax
1036	movl	%eax, %gs
1037	jmp	2b
1038	.previous
1039
1040/* Call softirq on interrupt stack. Interrupts are off. */
1041ENTRY(do_softirq_own_stack)
1042	pushq	%rbp
1043	mov	%rsp, %rbp
1044	ENTER_IRQ_STACK regs=0 old_rsp=%r11
1045	call	__do_softirq
1046	LEAVE_IRQ_STACK regs=0
1047	leaveq
1048	ret
1049ENDPROC(do_softirq_own_stack)
1050
1051#ifdef CONFIG_XEN
1052idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0
1053
1054/*
1055 * A note on the "critical region" in our callback handler.
1056 * We want to avoid stacking callback handlers due to events occurring
1057 * during handling of the last event. To do this, we keep events disabled
1058 * until we've done all processing. HOWEVER, we must enable events before
1059 * popping the stack frame (can't be done atomically) and so it would still
1060 * be possible to get enough handler activations to overflow the stack.
1061 * Although unlikely, bugs of that kind are hard to track down, so we'd
1062 * like to avoid the possibility.
1063 * So, on entry to the handler we detect whether we interrupted an
1064 * existing activation in its critical region -- if so, we pop the current
1065 * activation and restart the handler using the previous one.
1066 */
1067ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
1068
1069/*
1070 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
1071 * see the correct pointer to the pt_regs
1072 */
1073	UNWIND_HINT_FUNC
1074	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
1075	UNWIND_HINT_REGS
1076
1077	ENTER_IRQ_STACK old_rsp=%r10
1078	call	xen_evtchn_do_upcall
1079	LEAVE_IRQ_STACK
1080
1081#ifndef CONFIG_PREEMPT
1082	call	xen_maybe_preempt_hcall
1083#endif
1084	jmp	error_exit
1085END(xen_do_hypervisor_callback)
1086
1087/*
1088 * Hypervisor uses this for application faults while it executes.
1089 * We get here for two reasons:
1090 *  1. Fault while reloading DS, ES, FS or GS
1091 *  2. Fault while executing IRET
1092 * Category 1 we do not need to fix up as Xen has already reloaded all segment
1093 * registers that could be reloaded and zeroed the others.
1094 * Category 2 we fix up by killing the current process. We cannot use the
1095 * normal Linux return path in this case because if we use the IRET hypercall
1096 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
1097 * We distinguish between categories by comparing each saved segment register
1098 * with its current contents: any discrepancy means we in category 1.
1099 */
1100ENTRY(xen_failsafe_callback)
1101	UNWIND_HINT_EMPTY
1102	movl	%ds, %ecx
1103	cmpw	%cx, 0x10(%rsp)
1104	jne	1f
1105	movl	%es, %ecx
1106	cmpw	%cx, 0x18(%rsp)
1107	jne	1f
1108	movl	%fs, %ecx
1109	cmpw	%cx, 0x20(%rsp)
1110	jne	1f
1111	movl	%gs, %ecx
1112	cmpw	%cx, 0x28(%rsp)
1113	jne	1f
1114	/* All segments match their saved values => Category 2 (Bad IRET). */
1115	movq	(%rsp), %rcx
1116	movq	8(%rsp), %r11
1117	addq	$0x30, %rsp
1118	pushq	$0				/* RIP */
1119	UNWIND_HINT_IRET_REGS offset=8
1120	jmp	general_protection
11211:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
1122	movq	(%rsp), %rcx
1123	movq	8(%rsp), %r11
1124	addq	$0x30, %rsp
1125	UNWIND_HINT_IRET_REGS
1126	pushq	$-1 /* orig_ax = -1 => not a system call */
1127	PUSH_AND_CLEAR_REGS
1128	ENCODE_FRAME_POINTER
1129	jmp	error_exit
1130END(xen_failsafe_callback)
1131
1132apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1133	xen_hvm_callback_vector xen_evtchn_do_upcall
1134
1135#endif /* CONFIG_XEN */
1136
1137#if IS_ENABLED(CONFIG_HYPERV)
1138apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1139	hyperv_callback_vector hyperv_vector_handler
1140
1141apicinterrupt3 HYPERV_REENLIGHTENMENT_VECTOR \
1142	hyperv_reenlightenment_vector hyperv_reenlightenment_intr
1143
1144apicinterrupt3 HYPERV_STIMER0_VECTOR \
1145	hv_stimer0_callback_vector hv_stimer0_vector_handler
1146#endif /* CONFIG_HYPERV */
1147
1148idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1149idtentry int3			do_int3			has_error_code=0
1150idtentry stack_segment		do_stack_segment	has_error_code=1
1151
1152#ifdef CONFIG_XEN
1153idtentry xennmi			do_nmi			has_error_code=0
1154idtentry xendebug		do_debug		has_error_code=0
1155idtentry xenint3		do_int3			has_error_code=0
1156#endif
1157
1158idtentry general_protection	do_general_protection	has_error_code=1
1159idtentry page_fault		do_page_fault		has_error_code=1
1160
1161#ifdef CONFIG_KVM_GUEST
1162idtentry async_page_fault	do_async_page_fault	has_error_code=1
1163#endif
1164
1165#ifdef CONFIG_X86_MCE
1166idtentry machine_check		do_mce			has_error_code=0	paranoid=1
1167#endif
1168
1169/*
1170 * Save all registers in pt_regs, and switch gs if needed.
1171 * Use slow, but surefire "are we in kernel?" check.
1172 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
1173 */
1174ENTRY(paranoid_entry)
1175	UNWIND_HINT_FUNC
1176	cld
1177	PUSH_AND_CLEAR_REGS save_ret=1
1178	ENCODE_FRAME_POINTER 8
1179	movl	$1, %ebx
1180	movl	$MSR_GS_BASE, %ecx
1181	rdmsr
1182	testl	%edx, %edx
1183	js	1f				/* negative -> in kernel */
1184	SWAPGS
1185	xorl	%ebx, %ebx
1186
11871:
1188	SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
1189
1190	ret
1191END(paranoid_entry)
1192
1193/*
1194 * "Paranoid" exit path from exception stack.  This is invoked
1195 * only on return from non-NMI IST interrupts that came
1196 * from kernel space.
1197 *
1198 * We may be returning to very strange contexts (e.g. very early
1199 * in syscall entry), so checking for preemption here would
1200 * be complicated.  Fortunately, we there's no good reason
1201 * to try to handle preemption here.
1202 *
1203 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1204 */
1205ENTRY(paranoid_exit)
1206	UNWIND_HINT_REGS
1207	DISABLE_INTERRUPTS(CLBR_ANY)
1208	TRACE_IRQS_OFF_DEBUG
1209	testl	%ebx, %ebx			/* swapgs needed? */
1210	jnz	.Lparanoid_exit_no_swapgs
1211	TRACE_IRQS_IRETQ
1212	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1213	SWAPGS_UNSAFE_STACK
1214	jmp	.Lparanoid_exit_restore
1215.Lparanoid_exit_no_swapgs:
1216	TRACE_IRQS_IRETQ_DEBUG
1217	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1218.Lparanoid_exit_restore:
1219	jmp restore_regs_and_return_to_kernel
1220END(paranoid_exit)
1221
1222/*
1223 * Save all registers in pt_regs, and switch GS if needed.
1224 * Return: EBX=0: came from user mode; EBX=1: otherwise
1225 */
1226ENTRY(error_entry)
1227	UNWIND_HINT_FUNC
1228	cld
1229	PUSH_AND_CLEAR_REGS save_ret=1
1230	ENCODE_FRAME_POINTER 8
1231	testb	$3, CS+8(%rsp)
1232	jz	.Lerror_kernelspace
1233
1234	/*
1235	 * We entered from user mode or we're pretending to have entered
1236	 * from user mode due to an IRET fault.
1237	 */
1238	SWAPGS
1239	/* We have user CR3.  Change to kernel CR3. */
1240	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1241
1242.Lerror_entry_from_usermode_after_swapgs:
1243	/* Put us onto the real thread stack. */
1244	popq	%r12				/* save return addr in %12 */
1245	movq	%rsp, %rdi			/* arg0 = pt_regs pointer */
1246	call	sync_regs
1247	movq	%rax, %rsp			/* switch stack */
1248	ENCODE_FRAME_POINTER
1249	pushq	%r12
1250
1251	/*
1252	 * We need to tell lockdep that IRQs are off.  We can't do this until
1253	 * we fix gsbase, and we should do it before enter_from_user_mode
1254	 * (which can take locks).
1255	 */
1256	TRACE_IRQS_OFF
1257	CALL_enter_from_user_mode
1258	ret
1259
1260.Lerror_entry_done:
1261	TRACE_IRQS_OFF
1262	ret
1263
1264	/*
1265	 * There are two places in the kernel that can potentially fault with
1266	 * usergs. Handle them here.  B stepping K8s sometimes report a
1267	 * truncated RIP for IRET exceptions returning to compat mode. Check
1268	 * for these here too.
1269	 */
1270.Lerror_kernelspace:
1271	incl	%ebx
1272	leaq	native_irq_return_iret(%rip), %rcx
1273	cmpq	%rcx, RIP+8(%rsp)
1274	je	.Lerror_bad_iret
1275	movl	%ecx, %eax			/* zero extend */
1276	cmpq	%rax, RIP+8(%rsp)
1277	je	.Lbstep_iret
1278	cmpq	$.Lgs_change, RIP+8(%rsp)
1279	jne	.Lerror_entry_done
1280
1281	/*
1282	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1283	 * gsbase and proceed.  We'll fix up the exception and land in
1284	 * .Lgs_change's error handler with kernel gsbase.
1285	 */
1286	SWAPGS
1287	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1288	jmp .Lerror_entry_done
1289
1290.Lbstep_iret:
1291	/* Fix truncated RIP */
1292	movq	%rcx, RIP+8(%rsp)
1293	/* fall through */
1294
1295.Lerror_bad_iret:
1296	/*
1297	 * We came from an IRET to user mode, so we have user
1298	 * gsbase and CR3.  Switch to kernel gsbase and CR3:
1299	 */
1300	SWAPGS
1301	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1302
1303	/*
1304	 * Pretend that the exception came from user mode: set up pt_regs
1305	 * as if we faulted immediately after IRET and clear EBX so that
1306	 * error_exit knows that we will be returning to user mode.
1307	 */
1308	mov	%rsp, %rdi
1309	call	fixup_bad_iret
1310	mov	%rax, %rsp
1311	decl	%ebx
1312	jmp	.Lerror_entry_from_usermode_after_swapgs
1313END(error_entry)
1314
1315
1316/*
1317 * On entry, EBX is a "return to kernel mode" flag:
1318 *   1: already in kernel mode, don't need SWAPGS
1319 *   0: user gsbase is loaded, we need SWAPGS and standard preparation for return to usermode
1320 */
1321ENTRY(error_exit)
1322	UNWIND_HINT_REGS
1323	DISABLE_INTERRUPTS(CLBR_ANY)
1324	TRACE_IRQS_OFF
1325	testl	%ebx, %ebx
1326	jnz	retint_kernel
1327	jmp	retint_user
1328END(error_exit)
1329
1330/*
1331 * Runs on exception stack.  Xen PV does not go through this path at all,
1332 * so we can use real assembly here.
1333 *
1334 * Registers:
1335 *	%r14: Used to save/restore the CR3 of the interrupted context
1336 *	      when PAGE_TABLE_ISOLATION is in use.  Do not clobber.
1337 */
1338ENTRY(nmi)
1339	UNWIND_HINT_IRET_REGS
1340
1341	/*
1342	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1343	 * the iretq it performs will take us out of NMI context.
1344	 * This means that we can have nested NMIs where the next
1345	 * NMI is using the top of the stack of the previous NMI. We
1346	 * can't let it execute because the nested NMI will corrupt the
1347	 * stack of the previous NMI. NMI handlers are not re-entrant
1348	 * anyway.
1349	 *
1350	 * To handle this case we do the following:
1351	 *  Check the a special location on the stack that contains
1352	 *  a variable that is set when NMIs are executing.
1353	 *  The interrupted task's stack is also checked to see if it
1354	 *  is an NMI stack.
1355	 *  If the variable is not set and the stack is not the NMI
1356	 *  stack then:
1357	 *    o Set the special variable on the stack
1358	 *    o Copy the interrupt frame into an "outermost" location on the
1359	 *      stack
1360	 *    o Copy the interrupt frame into an "iret" location on the stack
1361	 *    o Continue processing the NMI
1362	 *  If the variable is set or the previous stack is the NMI stack:
1363	 *    o Modify the "iret" location to jump to the repeat_nmi
1364	 *    o return back to the first NMI
1365	 *
1366	 * Now on exit of the first NMI, we first clear the stack variable
1367	 * The NMI stack will tell any nested NMIs at that point that it is
1368	 * nested. Then we pop the stack normally with iret, and if there was
1369	 * a nested NMI that updated the copy interrupt stack frame, a
1370	 * jump will be made to the repeat_nmi code that will handle the second
1371	 * NMI.
1372	 *
1373	 * However, espfix prevents us from directly returning to userspace
1374	 * with a single IRET instruction.  Similarly, IRET to user mode
1375	 * can fault.  We therefore handle NMIs from user space like
1376	 * other IST entries.
1377	 */
1378
1379	ASM_CLAC
1380
1381	/* Use %rdx as our temp variable throughout */
1382	pushq	%rdx
1383
1384	testb	$3, CS-RIP+8(%rsp)
1385	jz	.Lnmi_from_kernel
1386
1387	/*
1388	 * NMI from user mode.  We need to run on the thread stack, but we
1389	 * can't go through the normal entry paths: NMIs are masked, and
1390	 * we don't want to enable interrupts, because then we'll end
1391	 * up in an awkward situation in which IRQs are on but NMIs
1392	 * are off.
1393	 *
1394	 * We also must not push anything to the stack before switching
1395	 * stacks lest we corrupt the "NMI executing" variable.
1396	 */
1397
1398	swapgs
1399	cld
1400	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx
1401	movq	%rsp, %rdx
1402	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1403	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1404	pushq	5*8(%rdx)	/* pt_regs->ss */
1405	pushq	4*8(%rdx)	/* pt_regs->rsp */
1406	pushq	3*8(%rdx)	/* pt_regs->flags */
1407	pushq	2*8(%rdx)	/* pt_regs->cs */
1408	pushq	1*8(%rdx)	/* pt_regs->rip */
1409	UNWIND_HINT_IRET_REGS
1410	pushq   $-1		/* pt_regs->orig_ax */
1411	PUSH_AND_CLEAR_REGS rdx=(%rdx)
1412	ENCODE_FRAME_POINTER
1413
1414	/*
1415	 * At this point we no longer need to worry about stack damage
1416	 * due to nesting -- we're on the normal thread stack and we're
1417	 * done with the NMI stack.
1418	 */
1419
1420	movq	%rsp, %rdi
1421	movq	$-1, %rsi
1422	call	do_nmi
1423
1424	/*
1425	 * Return back to user mode.  We must *not* do the normal exit
1426	 * work, because we don't want to enable interrupts.
1427	 */
1428	jmp	swapgs_restore_regs_and_return_to_usermode
1429
1430.Lnmi_from_kernel:
1431	/*
1432	 * Here's what our stack frame will look like:
1433	 * +---------------------------------------------------------+
1434	 * | original SS                                             |
1435	 * | original Return RSP                                     |
1436	 * | original RFLAGS                                         |
1437	 * | original CS                                             |
1438	 * | original RIP                                            |
1439	 * +---------------------------------------------------------+
1440	 * | temp storage for rdx                                    |
1441	 * +---------------------------------------------------------+
1442	 * | "NMI executing" variable                                |
1443	 * +---------------------------------------------------------+
1444	 * | iret SS          } Copied from "outermost" frame        |
1445	 * | iret Return RSP  } on each loop iteration; overwritten  |
1446	 * | iret RFLAGS      } by a nested NMI to force another     |
1447	 * | iret CS          } iteration if needed.                 |
1448	 * | iret RIP         }                                      |
1449	 * +---------------------------------------------------------+
1450	 * | outermost SS          } initialized in first_nmi;       |
1451	 * | outermost Return RSP  } will not be changed before      |
1452	 * | outermost RFLAGS      } NMI processing is done.         |
1453	 * | outermost CS          } Copied to "iret" frame on each  |
1454	 * | outermost RIP         } iteration.                      |
1455	 * +---------------------------------------------------------+
1456	 * | pt_regs                                                 |
1457	 * +---------------------------------------------------------+
1458	 *
1459	 * The "original" frame is used by hardware.  Before re-enabling
1460	 * NMIs, we need to be done with it, and we need to leave enough
1461	 * space for the asm code here.
1462	 *
1463	 * We return by executing IRET while RSP points to the "iret" frame.
1464	 * That will either return for real or it will loop back into NMI
1465	 * processing.
1466	 *
1467	 * The "outermost" frame is copied to the "iret" frame on each
1468	 * iteration of the loop, so each iteration starts with the "iret"
1469	 * frame pointing to the final return target.
1470	 */
1471
1472	/*
1473	 * Determine whether we're a nested NMI.
1474	 *
1475	 * If we interrupted kernel code between repeat_nmi and
1476	 * end_repeat_nmi, then we are a nested NMI.  We must not
1477	 * modify the "iret" frame because it's being written by
1478	 * the outer NMI.  That's okay; the outer NMI handler is
1479	 * about to about to call do_nmi anyway, so we can just
1480	 * resume the outer NMI.
1481	 */
1482
1483	movq	$repeat_nmi, %rdx
1484	cmpq	8(%rsp), %rdx
1485	ja	1f
1486	movq	$end_repeat_nmi, %rdx
1487	cmpq	8(%rsp), %rdx
1488	ja	nested_nmi_out
14891:
1490
1491	/*
1492	 * Now check "NMI executing".  If it's set, then we're nested.
1493	 * This will not detect if we interrupted an outer NMI just
1494	 * before IRET.
1495	 */
1496	cmpl	$1, -8(%rsp)
1497	je	nested_nmi
1498
1499	/*
1500	 * Now test if the previous stack was an NMI stack.  This covers
1501	 * the case where we interrupt an outer NMI after it clears
1502	 * "NMI executing" but before IRET.  We need to be careful, though:
1503	 * there is one case in which RSP could point to the NMI stack
1504	 * despite there being no NMI active: naughty userspace controls
1505	 * RSP at the very beginning of the SYSCALL targets.  We can
1506	 * pull a fast one on naughty userspace, though: we program
1507	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1508	 * if it controls the kernel's RSP.  We set DF before we clear
1509	 * "NMI executing".
1510	 */
1511	lea	6*8(%rsp), %rdx
1512	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1513	cmpq	%rdx, 4*8(%rsp)
1514	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1515	ja	first_nmi
1516
1517	subq	$EXCEPTION_STKSZ, %rdx
1518	cmpq	%rdx, 4*8(%rsp)
1519	/* If it is below the NMI stack, it is a normal NMI */
1520	jb	first_nmi
1521
1522	/* Ah, it is within the NMI stack. */
1523
1524	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1525	jz	first_nmi	/* RSP was user controlled. */
1526
1527	/* This is a nested NMI. */
1528
1529nested_nmi:
1530	/*
1531	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1532	 * iteration of NMI handling.
1533	 */
1534	subq	$8, %rsp
1535	leaq	-10*8(%rsp), %rdx
1536	pushq	$__KERNEL_DS
1537	pushq	%rdx
1538	pushfq
1539	pushq	$__KERNEL_CS
1540	pushq	$repeat_nmi
1541
1542	/* Put stack back */
1543	addq	$(6*8), %rsp
1544
1545nested_nmi_out:
1546	popq	%rdx
1547
1548	/* We are returning to kernel mode, so this cannot result in a fault. */
1549	iretq
1550
1551first_nmi:
1552	/* Restore rdx. */
1553	movq	(%rsp), %rdx
1554
1555	/* Make room for "NMI executing". */
1556	pushq	$0
1557
1558	/* Leave room for the "iret" frame */
1559	subq	$(5*8), %rsp
1560
1561	/* Copy the "original" frame to the "outermost" frame */
1562	.rept 5
1563	pushq	11*8(%rsp)
1564	.endr
1565	UNWIND_HINT_IRET_REGS
1566
1567	/* Everything up to here is safe from nested NMIs */
1568
1569#ifdef CONFIG_DEBUG_ENTRY
1570	/*
1571	 * For ease of testing, unmask NMIs right away.  Disabled by
1572	 * default because IRET is very expensive.
1573	 */
1574	pushq	$0		/* SS */
1575	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1576	addq	$8, (%rsp)	/* Fix up RSP */
1577	pushfq			/* RFLAGS */
1578	pushq	$__KERNEL_CS	/* CS */
1579	pushq	$1f		/* RIP */
1580	iretq			/* continues at repeat_nmi below */
1581	UNWIND_HINT_IRET_REGS
15821:
1583#endif
1584
1585repeat_nmi:
1586	/*
1587	 * If there was a nested NMI, the first NMI's iret will return
1588	 * here. But NMIs are still enabled and we can take another
1589	 * nested NMI. The nested NMI checks the interrupted RIP to see
1590	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1591	 * it will just return, as we are about to repeat an NMI anyway.
1592	 * This makes it safe to copy to the stack frame that a nested
1593	 * NMI will update.
1594	 *
1595	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1596	 * we're repeating an NMI, gsbase has the same value that it had on
1597	 * the first iteration.  paranoid_entry will load the kernel
1598	 * gsbase if needed before we call do_nmi.  "NMI executing"
1599	 * is zero.
1600	 */
1601	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1602
1603	/*
1604	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1605	 * here must not modify the "iret" frame while we're writing to
1606	 * it or it will end up containing garbage.
1607	 */
1608	addq	$(10*8), %rsp
1609	.rept 5
1610	pushq	-6*8(%rsp)
1611	.endr
1612	subq	$(5*8), %rsp
1613end_repeat_nmi:
1614
1615	/*
1616	 * Everything below this point can be preempted by a nested NMI.
1617	 * If this happens, then the inner NMI will change the "iret"
1618	 * frame to point back to repeat_nmi.
1619	 */
1620	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1621
1622	/*
1623	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1624	 * as we should not be calling schedule in NMI context.
1625	 * Even with normal interrupts enabled. An NMI should not be
1626	 * setting NEED_RESCHED or anything that normal interrupts and
1627	 * exceptions might do.
1628	 */
1629	call	paranoid_entry
1630	UNWIND_HINT_REGS
1631
1632	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1633	movq	%rsp, %rdi
1634	movq	$-1, %rsi
1635	call	do_nmi
1636
1637	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
1638
1639	testl	%ebx, %ebx			/* swapgs needed? */
1640	jnz	nmi_restore
1641nmi_swapgs:
1642	SWAPGS_UNSAFE_STACK
1643nmi_restore:
1644	POP_REGS
1645
1646	/*
1647	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1648	 * at the "iret" frame.
1649	 */
1650	addq	$6*8, %rsp
1651
1652	/*
1653	 * Clear "NMI executing".  Set DF first so that we can easily
1654	 * distinguish the remaining code between here and IRET from
1655	 * the SYSCALL entry and exit paths.
1656	 *
1657	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1658	 * this code when I had the misapprehension that Xen PV supported
1659	 * NMIs, and Xen PV would break that approach.
1660	 */
1661	std
1662	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1663
1664	/*
1665	 * iretq reads the "iret" frame and exits the NMI stack in a
1666	 * single instruction.  We are returning to kernel mode, so this
1667	 * cannot result in a fault.  Similarly, we don't need to worry
1668	 * about espfix64 on the way back to kernel mode.
1669	 */
1670	iretq
1671END(nmi)
1672
1673ENTRY(ignore_sysret)
1674	UNWIND_HINT_EMPTY
1675	mov	$-ENOSYS, %eax
1676	sysret
1677END(ignore_sysret)
1678
1679ENTRY(rewind_stack_do_exit)
1680	UNWIND_HINT_FUNC
1681	/* Prevent any naive code from trying to unwind to our caller. */
1682	xorl	%ebp, %ebp
1683
1684	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1685	leaq	-PTREGS_SIZE(%rax), %rsp
1686	UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE
1687
1688	call	do_exit
1689END(rewind_stack_do_exit)
1690