xref: /linux/arch/x86/entry/entry_64.S (revision 55f3538c4923e9dfca132e99ebec370e8094afda)
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	bt	$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 not 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	pushq	%rdi				/* pt_regs->di */
231	pushq	%rsi				/* pt_regs->si */
232	pushq	%rdx				/* pt_regs->dx */
233	pushq	%rcx				/* pt_regs->cx */
234	pushq	$-ENOSYS			/* pt_regs->ax */
235	pushq	%r8				/* pt_regs->r8 */
236	pushq	%r9				/* pt_regs->r9 */
237	pushq	%r10				/* pt_regs->r10 */
238	pushq	%r11				/* pt_regs->r11 */
239	pushq	%rbx				/* pt_regs->rbx */
240	pushq	%rbp				/* pt_regs->rbp */
241	pushq	%r12				/* pt_regs->r12 */
242	pushq	%r13				/* pt_regs->r13 */
243	pushq	%r14				/* pt_regs->r14 */
244	pushq	%r15				/* pt_regs->r15 */
245	UNWIND_HINT_REGS
246
247	TRACE_IRQS_OFF
248
249	/* IRQs are off. */
250	movq	%rsp, %rdi
251	call	do_syscall_64		/* returns with IRQs disabled */
252
253	TRACE_IRQS_IRETQ		/* we're about to change IF */
254
255	/*
256	 * Try to use SYSRET instead of IRET if we're returning to
257	 * a completely clean 64-bit userspace context.  If we're not,
258	 * go to the slow exit path.
259	 */
260	movq	RCX(%rsp), %rcx
261	movq	RIP(%rsp), %r11
262
263	cmpq	%rcx, %r11	/* SYSRET requires RCX == RIP */
264	jne	swapgs_restore_regs_and_return_to_usermode
265
266	/*
267	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
268	 * in kernel space.  This essentially lets the user take over
269	 * the kernel, since userspace controls RSP.
270	 *
271	 * If width of "canonical tail" ever becomes variable, this will need
272	 * to be updated to remain correct on both old and new CPUs.
273	 *
274	 * Change top bits to match most significant bit (47th or 56th bit
275	 * depending on paging mode) in the address.
276	 */
277	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
278	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
279
280	/* If this changed %rcx, it was not canonical */
281	cmpq	%rcx, %r11
282	jne	swapgs_restore_regs_and_return_to_usermode
283
284	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
285	jne	swapgs_restore_regs_and_return_to_usermode
286
287	movq	R11(%rsp), %r11
288	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
289	jne	swapgs_restore_regs_and_return_to_usermode
290
291	/*
292	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
293	 * restore RF properly. If the slowpath sets it for whatever reason, we
294	 * need to restore it correctly.
295	 *
296	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
297	 * trap from userspace immediately after SYSRET.  This would cause an
298	 * infinite loop whenever #DB happens with register state that satisfies
299	 * the opportunistic SYSRET conditions.  For example, single-stepping
300	 * this user code:
301	 *
302	 *           movq	$stuck_here, %rcx
303	 *           pushfq
304	 *           popq %r11
305	 *   stuck_here:
306	 *
307	 * would never get past 'stuck_here'.
308	 */
309	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
310	jnz	swapgs_restore_regs_and_return_to_usermode
311
312	/* nothing to check for RSP */
313
314	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
315	jne	swapgs_restore_regs_and_return_to_usermode
316
317	/*
318	 * We win! This label is here just for ease of understanding
319	 * perf profiles. Nothing jumps here.
320	 */
321syscall_return_via_sysret:
322	/* rcx and r11 are already restored (see code above) */
323	UNWIND_HINT_EMPTY
324	POP_EXTRA_REGS
325	popq	%rsi	/* skip r11 */
326	popq	%r10
327	popq	%r9
328	popq	%r8
329	popq	%rax
330	popq	%rsi	/* skip rcx */
331	popq	%rdx
332	popq	%rsi
333
334	/*
335	 * Now all regs are restored except RSP and RDI.
336	 * Save old stack pointer and switch to trampoline stack.
337	 */
338	movq	%rsp, %rdi
339	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
340
341	pushq	RSP-RDI(%rdi)	/* RSP */
342	pushq	(%rdi)		/* RDI */
343
344	/*
345	 * We are on the trampoline stack.  All regs except RDI are live.
346	 * We can do future final exit work right here.
347	 */
348	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
349
350	popq	%rdi
351	popq	%rsp
352	USERGS_SYSRET64
353END(entry_SYSCALL_64)
354
355/*
356 * %rdi: prev task
357 * %rsi: next task
358 */
359ENTRY(__switch_to_asm)
360	UNWIND_HINT_FUNC
361	/*
362	 * Save callee-saved registers
363	 * This must match the order in inactive_task_frame
364	 */
365	pushq	%rbp
366	pushq	%rbx
367	pushq	%r12
368	pushq	%r13
369	pushq	%r14
370	pushq	%r15
371
372	/* switch stack */
373	movq	%rsp, TASK_threadsp(%rdi)
374	movq	TASK_threadsp(%rsi), %rsp
375
376#ifdef CONFIG_CC_STACKPROTECTOR
377	movq	TASK_stack_canary(%rsi), %rbx
378	movq	%rbx, PER_CPU_VAR(irq_stack_union)+stack_canary_offset
379#endif
380
381#ifdef CONFIG_RETPOLINE
382	/*
383	 * When switching from a shallower to a deeper call stack
384	 * the RSB may either underflow or use entries populated
385	 * with userspace addresses. On CPUs where those concerns
386	 * exist, overwrite the RSB with entries which capture
387	 * speculative execution to prevent attack.
388	 */
389	/* Clobbers %rbx */
390	FILL_RETURN_BUFFER RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW
391#endif
392
393	/* restore callee-saved registers */
394	popq	%r15
395	popq	%r14
396	popq	%r13
397	popq	%r12
398	popq	%rbx
399	popq	%rbp
400
401	jmp	__switch_to
402END(__switch_to_asm)
403
404/*
405 * A newly forked process directly context switches into this address.
406 *
407 * rax: prev task we switched from
408 * rbx: kernel thread func (NULL for user thread)
409 * r12: kernel thread arg
410 */
411ENTRY(ret_from_fork)
412	UNWIND_HINT_EMPTY
413	movq	%rax, %rdi
414	call	schedule_tail			/* rdi: 'prev' task parameter */
415
416	testq	%rbx, %rbx			/* from kernel_thread? */
417	jnz	1f				/* kernel threads are uncommon */
418
4192:
420	UNWIND_HINT_REGS
421	movq	%rsp, %rdi
422	call	syscall_return_slowpath	/* returns with IRQs disabled */
423	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
424	jmp	swapgs_restore_regs_and_return_to_usermode
425
4261:
427	/* kernel thread */
428	movq	%r12, %rdi
429	CALL_NOSPEC %rbx
430	/*
431	 * A kernel thread is allowed to return here after successfully
432	 * calling do_execve().  Exit to userspace to complete the execve()
433	 * syscall.
434	 */
435	movq	$0, RAX(%rsp)
436	jmp	2b
437END(ret_from_fork)
438
439/*
440 * Build the entry stubs with some assembler magic.
441 * We pack 1 stub into every 8-byte block.
442 */
443	.align 8
444ENTRY(irq_entries_start)
445    vector=FIRST_EXTERNAL_VECTOR
446    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
447	UNWIND_HINT_IRET_REGS
448	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
449	jmp	common_interrupt
450	.align	8
451	vector=vector+1
452    .endr
453END(irq_entries_start)
454
455.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
456#ifdef CONFIG_DEBUG_ENTRY
457	pushq %rax
458	SAVE_FLAGS(CLBR_RAX)
459	testl $X86_EFLAGS_IF, %eax
460	jz .Lokay_\@
461	ud2
462.Lokay_\@:
463	popq %rax
464#endif
465.endm
466
467/*
468 * Enters the IRQ stack if we're not already using it.  NMI-safe.  Clobbers
469 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone.
470 * Requires kernel GSBASE.
471 *
472 * The invariant is that, if irq_count != -1, then the IRQ stack is in use.
473 */
474.macro ENTER_IRQ_STACK regs=1 old_rsp
475	DEBUG_ENTRY_ASSERT_IRQS_OFF
476	movq	%rsp, \old_rsp
477
478	.if \regs
479	UNWIND_HINT_REGS base=\old_rsp
480	.endif
481
482	incl	PER_CPU_VAR(irq_count)
483	jnz	.Lirq_stack_push_old_rsp_\@
484
485	/*
486	 * Right now, if we just incremented irq_count to zero, we've
487	 * claimed the IRQ stack but we haven't switched to it yet.
488	 *
489	 * If anything is added that can interrupt us here without using IST,
490	 * it must be *extremely* careful to limit its stack usage.  This
491	 * could include kprobes and a hypothetical future IST-less #DB
492	 * handler.
493	 *
494	 * The OOPS unwinder relies on the word at the top of the IRQ
495	 * stack linking back to the previous RSP for the entire time we're
496	 * on the IRQ stack.  For this to work reliably, we need to write
497	 * it before we actually move ourselves to the IRQ stack.
498	 */
499
500	movq	\old_rsp, PER_CPU_VAR(irq_stack_union + IRQ_STACK_SIZE - 8)
501	movq	PER_CPU_VAR(irq_stack_ptr), %rsp
502
503#ifdef CONFIG_DEBUG_ENTRY
504	/*
505	 * If the first movq above becomes wrong due to IRQ stack layout
506	 * changes, the only way we'll notice is if we try to unwind right
507	 * here.  Assert that we set up the stack right to catch this type
508	 * of bug quickly.
509	 */
510	cmpq	-8(%rsp), \old_rsp
511	je	.Lirq_stack_okay\@
512	ud2
513	.Lirq_stack_okay\@:
514#endif
515
516.Lirq_stack_push_old_rsp_\@:
517	pushq	\old_rsp
518
519	.if \regs
520	UNWIND_HINT_REGS indirect=1
521	.endif
522.endm
523
524/*
525 * Undoes ENTER_IRQ_STACK.
526 */
527.macro LEAVE_IRQ_STACK regs=1
528	DEBUG_ENTRY_ASSERT_IRQS_OFF
529	/* We need to be off the IRQ stack before decrementing irq_count. */
530	popq	%rsp
531
532	.if \regs
533	UNWIND_HINT_REGS
534	.endif
535
536	/*
537	 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming
538	 * the irq stack but we're not on it.
539	 */
540
541	decl	PER_CPU_VAR(irq_count)
542.endm
543
544/*
545 * Interrupt entry/exit.
546 *
547 * Interrupt entry points save only callee clobbered registers in fast path.
548 *
549 * Entry runs with interrupts off.
550 */
551
552/* 0(%rsp): ~(interrupt number) */
553	.macro interrupt func
554	cld
555
556	testb	$3, CS-ORIG_RAX(%rsp)
557	jz	1f
558	SWAPGS
559	call	switch_to_thread_stack
5601:
561
562	ALLOC_PT_GPREGS_ON_STACK
563	SAVE_C_REGS
564	SAVE_EXTRA_REGS
565	ENCODE_FRAME_POINTER
566
567	testb	$3, CS(%rsp)
568	jz	1f
569
570	/*
571	 * IRQ from user mode.
572	 *
573	 * We need to tell lockdep that IRQs are off.  We can't do this until
574	 * we fix gsbase, and we should do it before enter_from_user_mode
575	 * (which can take locks).  Since TRACE_IRQS_OFF idempotent,
576	 * the simplest way to handle it is to just call it twice if
577	 * we enter from user mode.  There's no reason to optimize this since
578	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
579	 */
580	TRACE_IRQS_OFF
581
582	CALL_enter_from_user_mode
583
5841:
585	ENTER_IRQ_STACK old_rsp=%rdi
586	/* We entered an interrupt context - irqs are off: */
587	TRACE_IRQS_OFF
588
589	call	\func	/* rdi points to pt_regs */
590	.endm
591
592	/*
593	 * The interrupt stubs push (~vector+0x80) onto the stack and
594	 * then jump to common_interrupt.
595	 */
596	.p2align CONFIG_X86_L1_CACHE_SHIFT
597common_interrupt:
598	ASM_CLAC
599	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
600	interrupt do_IRQ
601	/* 0(%rsp): old RSP */
602ret_from_intr:
603	DISABLE_INTERRUPTS(CLBR_ANY)
604	TRACE_IRQS_OFF
605
606	LEAVE_IRQ_STACK
607
608	testb	$3, CS(%rsp)
609	jz	retint_kernel
610
611	/* Interrupt came from user space */
612GLOBAL(retint_user)
613	mov	%rsp,%rdi
614	call	prepare_exit_to_usermode
615	TRACE_IRQS_IRETQ
616
617GLOBAL(swapgs_restore_regs_and_return_to_usermode)
618#ifdef CONFIG_DEBUG_ENTRY
619	/* Assert that pt_regs indicates user mode. */
620	testb	$3, CS(%rsp)
621	jnz	1f
622	ud2
6231:
624#endif
625	POP_EXTRA_REGS
626	popq	%r11
627	popq	%r10
628	popq	%r9
629	popq	%r8
630	popq	%rax
631	popq	%rcx
632	popq	%rdx
633	popq	%rsi
634
635	/*
636	 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS.
637	 * Save old stack pointer and switch to trampoline stack.
638	 */
639	movq	%rsp, %rdi
640	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
641
642	/* Copy the IRET frame to the trampoline stack. */
643	pushq	6*8(%rdi)	/* SS */
644	pushq	5*8(%rdi)	/* RSP */
645	pushq	4*8(%rdi)	/* EFLAGS */
646	pushq	3*8(%rdi)	/* CS */
647	pushq	2*8(%rdi)	/* RIP */
648
649	/* Push user RDI on the trampoline stack. */
650	pushq	(%rdi)
651
652	/*
653	 * We are on the trampoline stack.  All regs except RDI are live.
654	 * We can do future final exit work right here.
655	 */
656
657	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
658
659	/* Restore RDI. */
660	popq	%rdi
661	SWAPGS
662	INTERRUPT_RETURN
663
664
665/* Returning to kernel space */
666retint_kernel:
667#ifdef CONFIG_PREEMPT
668	/* Interrupts are off */
669	/* Check if we need preemption */
670	bt	$9, EFLAGS(%rsp)		/* were interrupts off? */
671	jnc	1f
6720:	cmpl	$0, PER_CPU_VAR(__preempt_count)
673	jnz	1f
674	call	preempt_schedule_irq
675	jmp	0b
6761:
677#endif
678	/*
679	 * The iretq could re-enable interrupts:
680	 */
681	TRACE_IRQS_IRETQ
682
683GLOBAL(restore_regs_and_return_to_kernel)
684#ifdef CONFIG_DEBUG_ENTRY
685	/* Assert that pt_regs indicates kernel mode. */
686	testb	$3, CS(%rsp)
687	jz	1f
688	ud2
6891:
690#endif
691	POP_EXTRA_REGS
692	POP_C_REGS
693	addq	$8, %rsp	/* skip regs->orig_ax */
694	INTERRUPT_RETURN
695
696ENTRY(native_iret)
697	UNWIND_HINT_IRET_REGS
698	/*
699	 * Are we returning to a stack segment from the LDT?  Note: in
700	 * 64-bit mode SS:RSP on the exception stack is always valid.
701	 */
702#ifdef CONFIG_X86_ESPFIX64
703	testb	$4, (SS-RIP)(%rsp)
704	jnz	native_irq_return_ldt
705#endif
706
707.global native_irq_return_iret
708native_irq_return_iret:
709	/*
710	 * This may fault.  Non-paranoid faults on return to userspace are
711	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
712	 * Double-faults due to espfix64 are handled in do_double_fault.
713	 * Other faults here are fatal.
714	 */
715	iretq
716
717#ifdef CONFIG_X86_ESPFIX64
718native_irq_return_ldt:
719	/*
720	 * We are running with user GSBASE.  All GPRs contain their user
721	 * values.  We have a percpu ESPFIX stack that is eight slots
722	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
723	 * of the ESPFIX stack.
724	 *
725	 * We clobber RAX and RDI in this code.  We stash RDI on the
726	 * normal stack and RAX on the ESPFIX stack.
727	 *
728	 * The ESPFIX stack layout we set up looks like this:
729	 *
730	 * --- top of ESPFIX stack ---
731	 * SS
732	 * RSP
733	 * RFLAGS
734	 * CS
735	 * RIP  <-- RSP points here when we're done
736	 * RAX  <-- espfix_waddr points here
737	 * --- bottom of ESPFIX stack ---
738	 */
739
740	pushq	%rdi				/* Stash user RDI */
741	SWAPGS					/* to kernel GS */
742	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi	/* to kernel CR3 */
743
744	movq	PER_CPU_VAR(espfix_waddr), %rdi
745	movq	%rax, (0*8)(%rdi)		/* user RAX */
746	movq	(1*8)(%rsp), %rax		/* user RIP */
747	movq	%rax, (1*8)(%rdi)
748	movq	(2*8)(%rsp), %rax		/* user CS */
749	movq	%rax, (2*8)(%rdi)
750	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
751	movq	%rax, (3*8)(%rdi)
752	movq	(5*8)(%rsp), %rax		/* user SS */
753	movq	%rax, (5*8)(%rdi)
754	movq	(4*8)(%rsp), %rax		/* user RSP */
755	movq	%rax, (4*8)(%rdi)
756	/* Now RAX == RSP. */
757
758	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
759
760	/*
761	 * espfix_stack[31:16] == 0.  The page tables are set up such that
762	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
763	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
764	 * the same page.  Set up RSP so that RSP[31:16] contains the
765	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
766	 * still points to an RO alias of the ESPFIX stack.
767	 */
768	orq	PER_CPU_VAR(espfix_stack), %rax
769
770	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
771	SWAPGS					/* to user GS */
772	popq	%rdi				/* Restore user RDI */
773
774	movq	%rax, %rsp
775	UNWIND_HINT_IRET_REGS offset=8
776
777	/*
778	 * At this point, we cannot write to the stack any more, but we can
779	 * still read.
780	 */
781	popq	%rax				/* Restore user RAX */
782
783	/*
784	 * RSP now points to an ordinary IRET frame, except that the page
785	 * is read-only and RSP[31:16] are preloaded with the userspace
786	 * values.  We can now IRET back to userspace.
787	 */
788	jmp	native_irq_return_iret
789#endif
790END(common_interrupt)
791
792/*
793 * APIC interrupts.
794 */
795.macro apicinterrupt3 num sym do_sym
796ENTRY(\sym)
797	UNWIND_HINT_IRET_REGS
798	ASM_CLAC
799	pushq	$~(\num)
800.Lcommon_\sym:
801	interrupt \do_sym
802	jmp	ret_from_intr
803END(\sym)
804.endm
805
806/* Make sure APIC interrupt handlers end up in the irqentry section: */
807#define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
808#define POP_SECTION_IRQENTRY	.popsection
809
810.macro apicinterrupt num sym do_sym
811PUSH_SECTION_IRQENTRY
812apicinterrupt3 \num \sym \do_sym
813POP_SECTION_IRQENTRY
814.endm
815
816#ifdef CONFIG_SMP
817apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
818apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
819#endif
820
821#ifdef CONFIG_X86_UV
822apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
823#endif
824
825apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
826apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
827
828#ifdef CONFIG_HAVE_KVM
829apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
830apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
831apicinterrupt3 POSTED_INTR_NESTED_VECTOR	kvm_posted_intr_nested_ipi	smp_kvm_posted_intr_nested_ipi
832#endif
833
834#ifdef CONFIG_X86_MCE_THRESHOLD
835apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
836#endif
837
838#ifdef CONFIG_X86_MCE_AMD
839apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
840#endif
841
842#ifdef CONFIG_X86_THERMAL_VECTOR
843apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
844#endif
845
846#ifdef CONFIG_SMP
847apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
848apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
849apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
850#endif
851
852apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
853apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
854
855#ifdef CONFIG_IRQ_WORK
856apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
857#endif
858
859/*
860 * Exception entry points.
861 */
862#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss_rw) + (TSS_ist + ((x) - 1) * 8)
863
864/*
865 * Switch to the thread stack.  This is called with the IRET frame and
866 * orig_ax on the stack.  (That is, RDI..R12 are not on the stack and
867 * space has not been allocated for them.)
868 */
869ENTRY(switch_to_thread_stack)
870	UNWIND_HINT_FUNC
871
872	pushq	%rdi
873	/* Need to switch before accessing the thread stack. */
874	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi
875	movq	%rsp, %rdi
876	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
877	UNWIND_HINT sp_offset=16 sp_reg=ORC_REG_DI
878
879	pushq	7*8(%rdi)		/* regs->ss */
880	pushq	6*8(%rdi)		/* regs->rsp */
881	pushq	5*8(%rdi)		/* regs->eflags */
882	pushq	4*8(%rdi)		/* regs->cs */
883	pushq	3*8(%rdi)		/* regs->ip */
884	pushq	2*8(%rdi)		/* regs->orig_ax */
885	pushq	8(%rdi)			/* return address */
886	UNWIND_HINT_FUNC
887
888	movq	(%rdi), %rdi
889	ret
890END(switch_to_thread_stack)
891
892.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
893ENTRY(\sym)
894	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
895
896	/* Sanity check */
897	.if \shift_ist != -1 && \paranoid == 0
898	.error "using shift_ist requires paranoid=1"
899	.endif
900
901	ASM_CLAC
902
903	.if \has_error_code == 0
904	pushq	$-1				/* ORIG_RAX: no syscall to restart */
905	.endif
906
907	ALLOC_PT_GPREGS_ON_STACK
908
909	.if \paranoid < 2
910	testb	$3, CS(%rsp)			/* If coming from userspace, switch stacks */
911	jnz	.Lfrom_usermode_switch_stack_\@
912	.endif
913
914	.if \paranoid
915	call	paranoid_entry
916	.else
917	call	error_entry
918	.endif
919	UNWIND_HINT_REGS
920	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
921
922	.if \paranoid
923	.if \shift_ist != -1
924	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
925	.else
926	TRACE_IRQS_OFF
927	.endif
928	.endif
929
930	movq	%rsp, %rdi			/* pt_regs pointer */
931
932	.if \has_error_code
933	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
934	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
935	.else
936	xorl	%esi, %esi			/* no error code */
937	.endif
938
939	.if \shift_ist != -1
940	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
941	.endif
942
943	call	\do_sym
944
945	.if \shift_ist != -1
946	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
947	.endif
948
949	/* these procedures expect "no swapgs" flag in ebx */
950	.if \paranoid
951	jmp	paranoid_exit
952	.else
953	jmp	error_exit
954	.endif
955
956	.if \paranoid < 2
957	/*
958	 * Entry from userspace.  Switch stacks and treat it
959	 * as a normal entry.  This means that paranoid handlers
960	 * run in real process context if user_mode(regs).
961	 */
962.Lfrom_usermode_switch_stack_\@:
963	call	error_entry
964
965	movq	%rsp, %rdi			/* pt_regs pointer */
966
967	.if \has_error_code
968	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
969	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
970	.else
971	xorl	%esi, %esi			/* no error code */
972	.endif
973
974	call	\do_sym
975
976	jmp	error_exit			/* %ebx: no swapgs flag */
977	.endif
978END(\sym)
979.endm
980
981idtentry divide_error			do_divide_error			has_error_code=0
982idtentry overflow			do_overflow			has_error_code=0
983idtentry bounds				do_bounds			has_error_code=0
984idtentry invalid_op			do_invalid_op			has_error_code=0
985idtentry device_not_available		do_device_not_available		has_error_code=0
986idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
987idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
988idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
989idtentry segment_not_present		do_segment_not_present		has_error_code=1
990idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
991idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
992idtentry alignment_check		do_alignment_check		has_error_code=1
993idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
994
995
996	/*
997	 * Reload gs selector with exception handling
998	 * edi:  new selector
999	 */
1000ENTRY(native_load_gs_index)
1001	FRAME_BEGIN
1002	pushfq
1003	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
1004	TRACE_IRQS_OFF
1005	SWAPGS
1006.Lgs_change:
1007	movl	%edi, %gs
10082:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
1009	SWAPGS
1010	TRACE_IRQS_FLAGS (%rsp)
1011	popfq
1012	FRAME_END
1013	ret
1014ENDPROC(native_load_gs_index)
1015EXPORT_SYMBOL(native_load_gs_index)
1016
1017	_ASM_EXTABLE(.Lgs_change, bad_gs)
1018	.section .fixup, "ax"
1019	/* running with kernelgs */
1020bad_gs:
1021	SWAPGS					/* switch back to user gs */
1022.macro ZAP_GS
1023	/* This can't be a string because the preprocessor needs to see it. */
1024	movl $__USER_DS, %eax
1025	movl %eax, %gs
1026.endm
1027	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
1028	xorl	%eax, %eax
1029	movl	%eax, %gs
1030	jmp	2b
1031	.previous
1032
1033/* Call softirq on interrupt stack. Interrupts are off. */
1034ENTRY(do_softirq_own_stack)
1035	pushq	%rbp
1036	mov	%rsp, %rbp
1037	ENTER_IRQ_STACK regs=0 old_rsp=%r11
1038	call	__do_softirq
1039	LEAVE_IRQ_STACK regs=0
1040	leaveq
1041	ret
1042ENDPROC(do_softirq_own_stack)
1043
1044#ifdef CONFIG_XEN
1045idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0
1046
1047/*
1048 * A note on the "critical region" in our callback handler.
1049 * We want to avoid stacking callback handlers due to events occurring
1050 * during handling of the last event. To do this, we keep events disabled
1051 * until we've done all processing. HOWEVER, we must enable events before
1052 * popping the stack frame (can't be done atomically) and so it would still
1053 * be possible to get enough handler activations to overflow the stack.
1054 * Although unlikely, bugs of that kind are hard to track down, so we'd
1055 * like to avoid the possibility.
1056 * So, on entry to the handler we detect whether we interrupted an
1057 * existing activation in its critical region -- if so, we pop the current
1058 * activation and restart the handler using the previous one.
1059 */
1060ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
1061
1062/*
1063 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
1064 * see the correct pointer to the pt_regs
1065 */
1066	UNWIND_HINT_FUNC
1067	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
1068	UNWIND_HINT_REGS
1069
1070	ENTER_IRQ_STACK old_rsp=%r10
1071	call	xen_evtchn_do_upcall
1072	LEAVE_IRQ_STACK
1073
1074#ifndef CONFIG_PREEMPT
1075	call	xen_maybe_preempt_hcall
1076#endif
1077	jmp	error_exit
1078END(xen_do_hypervisor_callback)
1079
1080/*
1081 * Hypervisor uses this for application faults while it executes.
1082 * We get here for two reasons:
1083 *  1. Fault while reloading DS, ES, FS or GS
1084 *  2. Fault while executing IRET
1085 * Category 1 we do not need to fix up as Xen has already reloaded all segment
1086 * registers that could be reloaded and zeroed the others.
1087 * Category 2 we fix up by killing the current process. We cannot use the
1088 * normal Linux return path in this case because if we use the IRET hypercall
1089 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
1090 * We distinguish between categories by comparing each saved segment register
1091 * with its current contents: any discrepancy means we in category 1.
1092 */
1093ENTRY(xen_failsafe_callback)
1094	UNWIND_HINT_EMPTY
1095	movl	%ds, %ecx
1096	cmpw	%cx, 0x10(%rsp)
1097	jne	1f
1098	movl	%es, %ecx
1099	cmpw	%cx, 0x18(%rsp)
1100	jne	1f
1101	movl	%fs, %ecx
1102	cmpw	%cx, 0x20(%rsp)
1103	jne	1f
1104	movl	%gs, %ecx
1105	cmpw	%cx, 0x28(%rsp)
1106	jne	1f
1107	/* All segments match their saved values => Category 2 (Bad IRET). */
1108	movq	(%rsp), %rcx
1109	movq	8(%rsp), %r11
1110	addq	$0x30, %rsp
1111	pushq	$0				/* RIP */
1112	UNWIND_HINT_IRET_REGS offset=8
1113	jmp	general_protection
11141:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
1115	movq	(%rsp), %rcx
1116	movq	8(%rsp), %r11
1117	addq	$0x30, %rsp
1118	UNWIND_HINT_IRET_REGS
1119	pushq	$-1 /* orig_ax = -1 => not a system call */
1120	ALLOC_PT_GPREGS_ON_STACK
1121	SAVE_C_REGS
1122	SAVE_EXTRA_REGS
1123	ENCODE_FRAME_POINTER
1124	jmp	error_exit
1125END(xen_failsafe_callback)
1126
1127apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1128	xen_hvm_callback_vector xen_evtchn_do_upcall
1129
1130#endif /* CONFIG_XEN */
1131
1132#if IS_ENABLED(CONFIG_HYPERV)
1133apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1134	hyperv_callback_vector hyperv_vector_handler
1135#endif /* CONFIG_HYPERV */
1136
1137idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1138idtentry int3			do_int3			has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1139idtentry stack_segment		do_stack_segment	has_error_code=1
1140
1141#ifdef CONFIG_XEN
1142idtentry xennmi			do_nmi			has_error_code=0
1143idtentry xendebug		do_debug		has_error_code=0
1144idtentry xenint3		do_int3			has_error_code=0
1145#endif
1146
1147idtentry general_protection	do_general_protection	has_error_code=1
1148idtentry page_fault		do_page_fault		has_error_code=1
1149
1150#ifdef CONFIG_KVM_GUEST
1151idtentry async_page_fault	do_async_page_fault	has_error_code=1
1152#endif
1153
1154#ifdef CONFIG_X86_MCE
1155idtentry machine_check		do_mce			has_error_code=0	paranoid=1
1156#endif
1157
1158/*
1159 * Save all registers in pt_regs, and switch gs if needed.
1160 * Use slow, but surefire "are we in kernel?" check.
1161 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
1162 */
1163ENTRY(paranoid_entry)
1164	UNWIND_HINT_FUNC
1165	cld
1166	SAVE_C_REGS 8
1167	SAVE_EXTRA_REGS 8
1168	ENCODE_FRAME_POINTER 8
1169	movl	$1, %ebx
1170	movl	$MSR_GS_BASE, %ecx
1171	rdmsr
1172	testl	%edx, %edx
1173	js	1f				/* negative -> in kernel */
1174	SWAPGS
1175	xorl	%ebx, %ebx
1176
11771:
1178	SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
1179
1180	ret
1181END(paranoid_entry)
1182
1183/*
1184 * "Paranoid" exit path from exception stack.  This is invoked
1185 * only on return from non-NMI IST interrupts that came
1186 * from kernel space.
1187 *
1188 * We may be returning to very strange contexts (e.g. very early
1189 * in syscall entry), so checking for preemption here would
1190 * be complicated.  Fortunately, we there's no good reason
1191 * to try to handle preemption here.
1192 *
1193 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1194 */
1195ENTRY(paranoid_exit)
1196	UNWIND_HINT_REGS
1197	DISABLE_INTERRUPTS(CLBR_ANY)
1198	TRACE_IRQS_OFF_DEBUG
1199	testl	%ebx, %ebx			/* swapgs needed? */
1200	jnz	.Lparanoid_exit_no_swapgs
1201	TRACE_IRQS_IRETQ
1202	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1203	SWAPGS_UNSAFE_STACK
1204	jmp	.Lparanoid_exit_restore
1205.Lparanoid_exit_no_swapgs:
1206	TRACE_IRQS_IRETQ_DEBUG
1207.Lparanoid_exit_restore:
1208	jmp restore_regs_and_return_to_kernel
1209END(paranoid_exit)
1210
1211/*
1212 * Save all registers in pt_regs, and switch gs if needed.
1213 * Return: EBX=0: came from user mode; EBX=1: otherwise
1214 */
1215ENTRY(error_entry)
1216	UNWIND_HINT_FUNC
1217	cld
1218	SAVE_C_REGS 8
1219	SAVE_EXTRA_REGS 8
1220	ENCODE_FRAME_POINTER 8
1221	xorl	%ebx, %ebx
1222	testb	$3, CS+8(%rsp)
1223	jz	.Lerror_kernelspace
1224
1225	/*
1226	 * We entered from user mode or we're pretending to have entered
1227	 * from user mode due to an IRET fault.
1228	 */
1229	SWAPGS
1230	/* We have user CR3.  Change to kernel CR3. */
1231	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1232
1233.Lerror_entry_from_usermode_after_swapgs:
1234	/* Put us onto the real thread stack. */
1235	popq	%r12				/* save return addr in %12 */
1236	movq	%rsp, %rdi			/* arg0 = pt_regs pointer */
1237	call	sync_regs
1238	movq	%rax, %rsp			/* switch stack */
1239	ENCODE_FRAME_POINTER
1240	pushq	%r12
1241
1242	/*
1243	 * We need to tell lockdep that IRQs are off.  We can't do this until
1244	 * we fix gsbase, and we should do it before enter_from_user_mode
1245	 * (which can take locks).
1246	 */
1247	TRACE_IRQS_OFF
1248	CALL_enter_from_user_mode
1249	ret
1250
1251.Lerror_entry_done:
1252	TRACE_IRQS_OFF
1253	ret
1254
1255	/*
1256	 * There are two places in the kernel that can potentially fault with
1257	 * usergs. Handle them here.  B stepping K8s sometimes report a
1258	 * truncated RIP for IRET exceptions returning to compat mode. Check
1259	 * for these here too.
1260	 */
1261.Lerror_kernelspace:
1262	incl	%ebx
1263	leaq	native_irq_return_iret(%rip), %rcx
1264	cmpq	%rcx, RIP+8(%rsp)
1265	je	.Lerror_bad_iret
1266	movl	%ecx, %eax			/* zero extend */
1267	cmpq	%rax, RIP+8(%rsp)
1268	je	.Lbstep_iret
1269	cmpq	$.Lgs_change, RIP+8(%rsp)
1270	jne	.Lerror_entry_done
1271
1272	/*
1273	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1274	 * gsbase and proceed.  We'll fix up the exception and land in
1275	 * .Lgs_change's error handler with kernel gsbase.
1276	 */
1277	SWAPGS
1278	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1279	jmp .Lerror_entry_done
1280
1281.Lbstep_iret:
1282	/* Fix truncated RIP */
1283	movq	%rcx, RIP+8(%rsp)
1284	/* fall through */
1285
1286.Lerror_bad_iret:
1287	/*
1288	 * We came from an IRET to user mode, so we have user
1289	 * gsbase and CR3.  Switch to kernel gsbase and CR3:
1290	 */
1291	SWAPGS
1292	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1293
1294	/*
1295	 * Pretend that the exception came from user mode: set up pt_regs
1296	 * as if we faulted immediately after IRET and clear EBX so that
1297	 * error_exit knows that we will be returning to user mode.
1298	 */
1299	mov	%rsp, %rdi
1300	call	fixup_bad_iret
1301	mov	%rax, %rsp
1302	decl	%ebx
1303	jmp	.Lerror_entry_from_usermode_after_swapgs
1304END(error_entry)
1305
1306
1307/*
1308 * On entry, EBX is a "return to kernel mode" flag:
1309 *   1: already in kernel mode, don't need SWAPGS
1310 *   0: user gsbase is loaded, we need SWAPGS and standard preparation for return to usermode
1311 */
1312ENTRY(error_exit)
1313	UNWIND_HINT_REGS
1314	DISABLE_INTERRUPTS(CLBR_ANY)
1315	TRACE_IRQS_OFF
1316	testl	%ebx, %ebx
1317	jnz	retint_kernel
1318	jmp	retint_user
1319END(error_exit)
1320
1321/*
1322 * Runs on exception stack.  Xen PV does not go through this path at all,
1323 * so we can use real assembly here.
1324 *
1325 * Registers:
1326 *	%r14: Used to save/restore the CR3 of the interrupted context
1327 *	      when PAGE_TABLE_ISOLATION is in use.  Do not clobber.
1328 */
1329ENTRY(nmi)
1330	UNWIND_HINT_IRET_REGS
1331
1332	/*
1333	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1334	 * the iretq it performs will take us out of NMI context.
1335	 * This means that we can have nested NMIs where the next
1336	 * NMI is using the top of the stack of the previous NMI. We
1337	 * can't let it execute because the nested NMI will corrupt the
1338	 * stack of the previous NMI. NMI handlers are not re-entrant
1339	 * anyway.
1340	 *
1341	 * To handle this case we do the following:
1342	 *  Check the a special location on the stack that contains
1343	 *  a variable that is set when NMIs are executing.
1344	 *  The interrupted task's stack is also checked to see if it
1345	 *  is an NMI stack.
1346	 *  If the variable is not set and the stack is not the NMI
1347	 *  stack then:
1348	 *    o Set the special variable on the stack
1349	 *    o Copy the interrupt frame into an "outermost" location on the
1350	 *      stack
1351	 *    o Copy the interrupt frame into an "iret" location on the stack
1352	 *    o Continue processing the NMI
1353	 *  If the variable is set or the previous stack is the NMI stack:
1354	 *    o Modify the "iret" location to jump to the repeat_nmi
1355	 *    o return back to the first NMI
1356	 *
1357	 * Now on exit of the first NMI, we first clear the stack variable
1358	 * The NMI stack will tell any nested NMIs at that point that it is
1359	 * nested. Then we pop the stack normally with iret, and if there was
1360	 * a nested NMI that updated the copy interrupt stack frame, a
1361	 * jump will be made to the repeat_nmi code that will handle the second
1362	 * NMI.
1363	 *
1364	 * However, espfix prevents us from directly returning to userspace
1365	 * with a single IRET instruction.  Similarly, IRET to user mode
1366	 * can fault.  We therefore handle NMIs from user space like
1367	 * other IST entries.
1368	 */
1369
1370	ASM_CLAC
1371
1372	/* Use %rdx as our temp variable throughout */
1373	pushq	%rdx
1374
1375	testb	$3, CS-RIP+8(%rsp)
1376	jz	.Lnmi_from_kernel
1377
1378	/*
1379	 * NMI from user mode.  We need to run on the thread stack, but we
1380	 * can't go through the normal entry paths: NMIs are masked, and
1381	 * we don't want to enable interrupts, because then we'll end
1382	 * up in an awkward situation in which IRQs are on but NMIs
1383	 * are off.
1384	 *
1385	 * We also must not push anything to the stack before switching
1386	 * stacks lest we corrupt the "NMI executing" variable.
1387	 */
1388
1389	swapgs
1390	cld
1391	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx
1392	movq	%rsp, %rdx
1393	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1394	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1395	pushq	5*8(%rdx)	/* pt_regs->ss */
1396	pushq	4*8(%rdx)	/* pt_regs->rsp */
1397	pushq	3*8(%rdx)	/* pt_regs->flags */
1398	pushq	2*8(%rdx)	/* pt_regs->cs */
1399	pushq	1*8(%rdx)	/* pt_regs->rip */
1400	UNWIND_HINT_IRET_REGS
1401	pushq   $-1		/* pt_regs->orig_ax */
1402	pushq   %rdi		/* pt_regs->di */
1403	pushq   %rsi		/* pt_regs->si */
1404	pushq   (%rdx)		/* pt_regs->dx */
1405	pushq   %rcx		/* pt_regs->cx */
1406	pushq   %rax		/* pt_regs->ax */
1407	pushq   %r8		/* pt_regs->r8 */
1408	pushq   %r9		/* pt_regs->r9 */
1409	pushq   %r10		/* pt_regs->r10 */
1410	pushq   %r11		/* pt_regs->r11 */
1411	pushq	%rbx		/* pt_regs->rbx */
1412	pushq	%rbp		/* pt_regs->rbp */
1413	pushq	%r12		/* pt_regs->r12 */
1414	pushq	%r13		/* pt_regs->r13 */
1415	pushq	%r14		/* pt_regs->r14 */
1416	pushq	%r15		/* pt_regs->r15 */
1417	UNWIND_HINT_REGS
1418	ENCODE_FRAME_POINTER
1419
1420	/*
1421	 * At this point we no longer need to worry about stack damage
1422	 * due to nesting -- we're on the normal thread stack and we're
1423	 * done with the NMI stack.
1424	 */
1425
1426	movq	%rsp, %rdi
1427	movq	$-1, %rsi
1428	call	do_nmi
1429
1430	/*
1431	 * Return back to user mode.  We must *not* do the normal exit
1432	 * work, because we don't want to enable interrupts.
1433	 */
1434	jmp	swapgs_restore_regs_and_return_to_usermode
1435
1436.Lnmi_from_kernel:
1437	/*
1438	 * Here's what our stack frame will look like:
1439	 * +---------------------------------------------------------+
1440	 * | original SS                                             |
1441	 * | original Return RSP                                     |
1442	 * | original RFLAGS                                         |
1443	 * | original CS                                             |
1444	 * | original RIP                                            |
1445	 * +---------------------------------------------------------+
1446	 * | temp storage for rdx                                    |
1447	 * +---------------------------------------------------------+
1448	 * | "NMI executing" variable                                |
1449	 * +---------------------------------------------------------+
1450	 * | iret SS          } Copied from "outermost" frame        |
1451	 * | iret Return RSP  } on each loop iteration; overwritten  |
1452	 * | iret RFLAGS      } by a nested NMI to force another     |
1453	 * | iret CS          } iteration if needed.                 |
1454	 * | iret RIP         }                                      |
1455	 * +---------------------------------------------------------+
1456	 * | outermost SS          } initialized in first_nmi;       |
1457	 * | outermost Return RSP  } will not be changed before      |
1458	 * | outermost RFLAGS      } NMI processing is done.         |
1459	 * | outermost CS          } Copied to "iret" frame on each  |
1460	 * | outermost RIP         } iteration.                      |
1461	 * +---------------------------------------------------------+
1462	 * | pt_regs                                                 |
1463	 * +---------------------------------------------------------+
1464	 *
1465	 * The "original" frame is used by hardware.  Before re-enabling
1466	 * NMIs, we need to be done with it, and we need to leave enough
1467	 * space for the asm code here.
1468	 *
1469	 * We return by executing IRET while RSP points to the "iret" frame.
1470	 * That will either return for real or it will loop back into NMI
1471	 * processing.
1472	 *
1473	 * The "outermost" frame is copied to the "iret" frame on each
1474	 * iteration of the loop, so each iteration starts with the "iret"
1475	 * frame pointing to the final return target.
1476	 */
1477
1478	/*
1479	 * Determine whether we're a nested NMI.
1480	 *
1481	 * If we interrupted kernel code between repeat_nmi and
1482	 * end_repeat_nmi, then we are a nested NMI.  We must not
1483	 * modify the "iret" frame because it's being written by
1484	 * the outer NMI.  That's okay; the outer NMI handler is
1485	 * about to about to call do_nmi anyway, so we can just
1486	 * resume the outer NMI.
1487	 */
1488
1489	movq	$repeat_nmi, %rdx
1490	cmpq	8(%rsp), %rdx
1491	ja	1f
1492	movq	$end_repeat_nmi, %rdx
1493	cmpq	8(%rsp), %rdx
1494	ja	nested_nmi_out
14951:
1496
1497	/*
1498	 * Now check "NMI executing".  If it's set, then we're nested.
1499	 * This will not detect if we interrupted an outer NMI just
1500	 * before IRET.
1501	 */
1502	cmpl	$1, -8(%rsp)
1503	je	nested_nmi
1504
1505	/*
1506	 * Now test if the previous stack was an NMI stack.  This covers
1507	 * the case where we interrupt an outer NMI after it clears
1508	 * "NMI executing" but before IRET.  We need to be careful, though:
1509	 * there is one case in which RSP could point to the NMI stack
1510	 * despite there being no NMI active: naughty userspace controls
1511	 * RSP at the very beginning of the SYSCALL targets.  We can
1512	 * pull a fast one on naughty userspace, though: we program
1513	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1514	 * if it controls the kernel's RSP.  We set DF before we clear
1515	 * "NMI executing".
1516	 */
1517	lea	6*8(%rsp), %rdx
1518	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1519	cmpq	%rdx, 4*8(%rsp)
1520	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1521	ja	first_nmi
1522
1523	subq	$EXCEPTION_STKSZ, %rdx
1524	cmpq	%rdx, 4*8(%rsp)
1525	/* If it is below the NMI stack, it is a normal NMI */
1526	jb	first_nmi
1527
1528	/* Ah, it is within the NMI stack. */
1529
1530	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1531	jz	first_nmi	/* RSP was user controlled. */
1532
1533	/* This is a nested NMI. */
1534
1535nested_nmi:
1536	/*
1537	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1538	 * iteration of NMI handling.
1539	 */
1540	subq	$8, %rsp
1541	leaq	-10*8(%rsp), %rdx
1542	pushq	$__KERNEL_DS
1543	pushq	%rdx
1544	pushfq
1545	pushq	$__KERNEL_CS
1546	pushq	$repeat_nmi
1547
1548	/* Put stack back */
1549	addq	$(6*8), %rsp
1550
1551nested_nmi_out:
1552	popq	%rdx
1553
1554	/* We are returning to kernel mode, so this cannot result in a fault. */
1555	iretq
1556
1557first_nmi:
1558	/* Restore rdx. */
1559	movq	(%rsp), %rdx
1560
1561	/* Make room for "NMI executing". */
1562	pushq	$0
1563
1564	/* Leave room for the "iret" frame */
1565	subq	$(5*8), %rsp
1566
1567	/* Copy the "original" frame to the "outermost" frame */
1568	.rept 5
1569	pushq	11*8(%rsp)
1570	.endr
1571	UNWIND_HINT_IRET_REGS
1572
1573	/* Everything up to here is safe from nested NMIs */
1574
1575#ifdef CONFIG_DEBUG_ENTRY
1576	/*
1577	 * For ease of testing, unmask NMIs right away.  Disabled by
1578	 * default because IRET is very expensive.
1579	 */
1580	pushq	$0		/* SS */
1581	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1582	addq	$8, (%rsp)	/* Fix up RSP */
1583	pushfq			/* RFLAGS */
1584	pushq	$__KERNEL_CS	/* CS */
1585	pushq	$1f		/* RIP */
1586	iretq			/* continues at repeat_nmi below */
1587	UNWIND_HINT_IRET_REGS
15881:
1589#endif
1590
1591repeat_nmi:
1592	/*
1593	 * If there was a nested NMI, the first NMI's iret will return
1594	 * here. But NMIs are still enabled and we can take another
1595	 * nested NMI. The nested NMI checks the interrupted RIP to see
1596	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1597	 * it will just return, as we are about to repeat an NMI anyway.
1598	 * This makes it safe to copy to the stack frame that a nested
1599	 * NMI will update.
1600	 *
1601	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1602	 * we're repeating an NMI, gsbase has the same value that it had on
1603	 * the first iteration.  paranoid_entry will load the kernel
1604	 * gsbase if needed before we call do_nmi.  "NMI executing"
1605	 * is zero.
1606	 */
1607	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1608
1609	/*
1610	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1611	 * here must not modify the "iret" frame while we're writing to
1612	 * it or it will end up containing garbage.
1613	 */
1614	addq	$(10*8), %rsp
1615	.rept 5
1616	pushq	-6*8(%rsp)
1617	.endr
1618	subq	$(5*8), %rsp
1619end_repeat_nmi:
1620
1621	/*
1622	 * Everything below this point can be preempted by a nested NMI.
1623	 * If this happens, then the inner NMI will change the "iret"
1624	 * frame to point back to repeat_nmi.
1625	 */
1626	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1627	ALLOC_PT_GPREGS_ON_STACK
1628
1629	/*
1630	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1631	 * as we should not be calling schedule in NMI context.
1632	 * Even with normal interrupts enabled. An NMI should not be
1633	 * setting NEED_RESCHED or anything that normal interrupts and
1634	 * exceptions might do.
1635	 */
1636	call	paranoid_entry
1637	UNWIND_HINT_REGS
1638
1639	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1640	movq	%rsp, %rdi
1641	movq	$-1, %rsi
1642	call	do_nmi
1643
1644	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
1645
1646	testl	%ebx, %ebx			/* swapgs needed? */
1647	jnz	nmi_restore
1648nmi_swapgs:
1649	SWAPGS_UNSAFE_STACK
1650nmi_restore:
1651	POP_EXTRA_REGS
1652	POP_C_REGS
1653
1654	/*
1655	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1656	 * at the "iret" frame.
1657	 */
1658	addq	$6*8, %rsp
1659
1660	/*
1661	 * Clear "NMI executing".  Set DF first so that we can easily
1662	 * distinguish the remaining code between here and IRET from
1663	 * the SYSCALL entry and exit paths.
1664	 *
1665	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1666	 * this code when I had the misapprehension that Xen PV supported
1667	 * NMIs, and Xen PV would break that approach.
1668	 */
1669	std
1670	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1671
1672	/*
1673	 * iretq reads the "iret" frame and exits the NMI stack in a
1674	 * single instruction.  We are returning to kernel mode, so this
1675	 * cannot result in a fault.  Similarly, we don't need to worry
1676	 * about espfix64 on the way back to kernel mode.
1677	 */
1678	iretq
1679END(nmi)
1680
1681ENTRY(ignore_sysret)
1682	UNWIND_HINT_EMPTY
1683	mov	$-ENOSYS, %eax
1684	sysret
1685END(ignore_sysret)
1686
1687ENTRY(rewind_stack_do_exit)
1688	UNWIND_HINT_FUNC
1689	/* Prevent any naive code from trying to unwind to our caller. */
1690	xorl	%ebp, %ebp
1691
1692	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1693	leaq	-PTREGS_SIZE(%rax), %rsp
1694	UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE
1695
1696	call	do_exit
1697END(rewind_stack_do_exit)
1698