xref: /linux/arch/x86/entry/entry_64.S (revision cf2f33a4e54096f90652cca3511fd6a456ea5abe)
1/*
2 *  linux/arch/x86_64/entry.S
3 *
4 *  Copyright (C) 1991, 1992  Linus Torvalds
5 *  Copyright (C) 2000, 2001, 2002  Andi Kleen SuSE Labs
6 *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
7 *
8 * entry.S contains the system-call and fault low-level handling routines.
9 *
10 * Some of this is documented in Documentation/x86/entry_64.txt
11 *
12 * A note on terminology:
13 * - iret frame:	Architecture defined interrupt frame from SS to RIP
14 *			at the top of the kernel process stack.
15 *
16 * Some macro usage:
17 * - ENTRY/END:		Define functions in the symbol table.
18 * - TRACE_IRQ_*:	Trace hardirq state for lock debugging.
19 * - idtentry:		Define exception entry points.
20 */
21#include <linux/linkage.h>
22#include <asm/segment.h>
23#include <asm/cache.h>
24#include <asm/errno.h>
25#include "calling.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/context_tracking.h>
37#include <asm/smap.h>
38#include <asm/pgtable_types.h>
39#include <linux/err.h>
40
41/* Avoid __ASSEMBLER__'ifying <linux/audit.h> just for this.  */
42#include <linux/elf-em.h>
43#define AUDIT_ARCH_X86_64			(EM_X86_64|__AUDIT_ARCH_64BIT|__AUDIT_ARCH_LE)
44#define __AUDIT_ARCH_64BIT			0x80000000
45#define __AUDIT_ARCH_LE				0x40000000
46
47.code64
48.section .entry.text, "ax"
49
50#ifdef CONFIG_PARAVIRT
51ENTRY(native_usergs_sysret64)
52	swapgs
53	sysretq
54ENDPROC(native_usergs_sysret64)
55#endif /* CONFIG_PARAVIRT */
56
57.macro TRACE_IRQS_IRETQ
58#ifdef CONFIG_TRACE_IRQFLAGS
59	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
60	jnc	1f
61	TRACE_IRQS_ON
621:
63#endif
64.endm
65
66/*
67 * When dynamic function tracer is enabled it will add a breakpoint
68 * to all locations that it is about to modify, sync CPUs, update
69 * all the code, sync CPUs, then remove the breakpoints. In this time
70 * if lockdep is enabled, it might jump back into the debug handler
71 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF).
72 *
73 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to
74 * make sure the stack pointer does not get reset back to the top
75 * of the debug stack, and instead just reuses the current stack.
76 */
77#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS)
78
79.macro TRACE_IRQS_OFF_DEBUG
80	call	debug_stack_set_zero
81	TRACE_IRQS_OFF
82	call	debug_stack_reset
83.endm
84
85.macro TRACE_IRQS_ON_DEBUG
86	call	debug_stack_set_zero
87	TRACE_IRQS_ON
88	call	debug_stack_reset
89.endm
90
91.macro TRACE_IRQS_IRETQ_DEBUG
92	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
93	jnc	1f
94	TRACE_IRQS_ON_DEBUG
951:
96.endm
97
98#else
99# define TRACE_IRQS_OFF_DEBUG			TRACE_IRQS_OFF
100# define TRACE_IRQS_ON_DEBUG			TRACE_IRQS_ON
101# define TRACE_IRQS_IRETQ_DEBUG			TRACE_IRQS_IRETQ
102#endif
103
104/*
105 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
106 *
107 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
108 * then loads new ss, cs, and rip from previously programmed MSRs.
109 * rflags gets masked by a value from another MSR (so CLD and CLAC
110 * are not needed). SYSCALL does not save anything on the stack
111 * and does not change rsp.
112 *
113 * Registers on entry:
114 * rax  system call number
115 * rcx  return address
116 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
117 * rdi  arg0
118 * rsi  arg1
119 * rdx  arg2
120 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
121 * r8   arg4
122 * r9   arg5
123 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
124 *
125 * Only called from user space.
126 *
127 * When user can change pt_regs->foo always force IRET. That is because
128 * it deals with uncanonical addresses better. SYSRET has trouble
129 * with them due to bugs in both AMD and Intel CPUs.
130 */
131
132ENTRY(entry_SYSCALL_64)
133	/*
134	 * Interrupts are off on entry.
135	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
136	 * it is too small to ever cause noticeable irq latency.
137	 */
138	SWAPGS_UNSAFE_STACK
139	/*
140	 * A hypervisor implementation might want to use a label
141	 * after the swapgs, so that it can do the swapgs
142	 * for the guest and jump here on syscall.
143	 */
144GLOBAL(entry_SYSCALL_64_after_swapgs)
145
146	movq	%rsp, PER_CPU_VAR(rsp_scratch)
147	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
148
149	/* Construct struct pt_regs on stack */
150	pushq	$__USER_DS			/* pt_regs->ss */
151	pushq	PER_CPU_VAR(rsp_scratch)	/* pt_regs->sp */
152	/*
153	 * Re-enable interrupts.
154	 * We use 'rsp_scratch' as a scratch space, hence irq-off block above
155	 * must execute atomically in the face of possible interrupt-driven
156	 * task preemption. We must enable interrupts only after we're done
157	 * with using rsp_scratch:
158	 */
159	ENABLE_INTERRUPTS(CLBR_NONE)
160	pushq	%r11				/* pt_regs->flags */
161	pushq	$__USER_CS			/* pt_regs->cs */
162	pushq	%rcx				/* pt_regs->ip */
163	pushq	%rax				/* pt_regs->orig_ax */
164	pushq	%rdi				/* pt_regs->di */
165	pushq	%rsi				/* pt_regs->si */
166	pushq	%rdx				/* pt_regs->dx */
167	pushq	%rcx				/* pt_regs->cx */
168	pushq	$-ENOSYS			/* pt_regs->ax */
169	pushq	%r8				/* pt_regs->r8 */
170	pushq	%r9				/* pt_regs->r9 */
171	pushq	%r10				/* pt_regs->r10 */
172	pushq	%r11				/* pt_regs->r11 */
173	sub	$(6*8), %rsp			/* pt_regs->bp, bx, r12-15 not saved */
174
175	testl	$_TIF_WORK_SYSCALL_ENTRY, ASM_THREAD_INFO(TI_flags, %rsp, SIZEOF_PTREGS)
176	jnz	tracesys
177entry_SYSCALL_64_fastpath:
178#if __SYSCALL_MASK == ~0
179	cmpq	$__NR_syscall_max, %rax
180#else
181	andl	$__SYSCALL_MASK, %eax
182	cmpl	$__NR_syscall_max, %eax
183#endif
184	ja	1f				/* return -ENOSYS (already in pt_regs->ax) */
185	movq	%r10, %rcx
186	call	*sys_call_table(, %rax, 8)
187	movq	%rax, RAX(%rsp)
1881:
189/*
190 * Syscall return path ending with SYSRET (fast path).
191 * Has incompletely filled pt_regs.
192 */
193	LOCKDEP_SYS_EXIT
194	/*
195	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
196	 * it is too small to ever cause noticeable irq latency.
197	 */
198	DISABLE_INTERRUPTS(CLBR_NONE)
199
200	/*
201	 * We must check ti flags with interrupts (or at least preemption)
202	 * off because we must *never* return to userspace without
203	 * processing exit work that is enqueued if we're preempted here.
204	 * In particular, returning to userspace with any of the one-shot
205	 * flags (TIF_NOTIFY_RESUME, TIF_USER_RETURN_NOTIFY, etc) set is
206	 * very bad.
207	 */
208	testl	$_TIF_ALLWORK_MASK, ASM_THREAD_INFO(TI_flags, %rsp, SIZEOF_PTREGS)
209	jnz	int_ret_from_sys_call_irqs_off	/* Go to the slow path */
210
211	RESTORE_C_REGS_EXCEPT_RCX_R11
212	movq	RIP(%rsp), %rcx
213	movq	EFLAGS(%rsp), %r11
214	movq	RSP(%rsp), %rsp
215	/*
216	 * 64-bit SYSRET restores rip from rcx,
217	 * rflags from r11 (but RF and VM bits are forced to 0),
218	 * cs and ss are loaded from MSRs.
219	 * Restoration of rflags re-enables interrupts.
220	 *
221	 * NB: On AMD CPUs with the X86_BUG_SYSRET_SS_ATTRS bug, the ss
222	 * descriptor is not reinitialized.  This means that we should
223	 * avoid SYSRET with SS == NULL, which could happen if we schedule,
224	 * exit the kernel, and re-enter using an interrupt vector.  (All
225	 * interrupt entries on x86_64 set SS to NULL.)  We prevent that
226	 * from happening by reloading SS in __switch_to.  (Actually
227	 * detecting the failure in 64-bit userspace is tricky but can be
228	 * done.)
229	 */
230	USERGS_SYSRET64
231
232	/* Do syscall entry tracing */
233tracesys:
234	movq	%rsp, %rdi
235	movl	$AUDIT_ARCH_X86_64, %esi
236	call	syscall_trace_enter_phase1
237	test	%rax, %rax
238	jnz	tracesys_phase2			/* if needed, run the slow path */
239	RESTORE_C_REGS_EXCEPT_RAX		/* else restore clobbered regs */
240	movq	ORIG_RAX(%rsp), %rax
241	jmp	entry_SYSCALL_64_fastpath	/* and return to the fast path */
242
243tracesys_phase2:
244	SAVE_EXTRA_REGS
245	movq	%rsp, %rdi
246	movl	$AUDIT_ARCH_X86_64, %esi
247	movq	%rax, %rdx
248	call	syscall_trace_enter_phase2
249
250	/*
251	 * Reload registers from stack in case ptrace changed them.
252	 * We don't reload %rax because syscall_trace_entry_phase2() returned
253	 * the value it wants us to use in the table lookup.
254	 */
255	RESTORE_C_REGS_EXCEPT_RAX
256	RESTORE_EXTRA_REGS
257#if __SYSCALL_MASK == ~0
258	cmpq	$__NR_syscall_max, %rax
259#else
260	andl	$__SYSCALL_MASK, %eax
261	cmpl	$__NR_syscall_max, %eax
262#endif
263	ja	1f				/* return -ENOSYS (already in pt_regs->ax) */
264	movq	%r10, %rcx			/* fixup for C */
265	call	*sys_call_table(, %rax, 8)
266	movq	%rax, RAX(%rsp)
2671:
268	/* Use IRET because user could have changed pt_regs->foo */
269
270/*
271 * Syscall return path ending with IRET.
272 * Has correct iret frame.
273 */
274GLOBAL(int_ret_from_sys_call)
275	DISABLE_INTERRUPTS(CLBR_NONE)
276int_ret_from_sys_call_irqs_off: /* jumps come here from the irqs-off SYSRET path */
277	TRACE_IRQS_OFF
278	movl	$_TIF_ALLWORK_MASK, %edi
279	/* edi:	mask to check */
280GLOBAL(int_with_check)
281	LOCKDEP_SYS_EXIT_IRQ
282	GET_THREAD_INFO(%rcx)
283	movl	TI_flags(%rcx), %edx
284	andl	%edi, %edx
285	jnz	int_careful
286	andl	$~TS_COMPAT, TI_status(%rcx)
287	jmp	syscall_return
288
289	/*
290	 * Either reschedule or signal or syscall exit tracking needed.
291	 * First do a reschedule test.
292	 * edx:	work, edi: workmask
293	 */
294int_careful:
295	bt	$TIF_NEED_RESCHED, %edx
296	jnc	int_very_careful
297	TRACE_IRQS_ON
298	ENABLE_INTERRUPTS(CLBR_NONE)
299	pushq	%rdi
300	SCHEDULE_USER
301	popq	%rdi
302	DISABLE_INTERRUPTS(CLBR_NONE)
303	TRACE_IRQS_OFF
304	jmp	int_with_check
305
306	/* handle signals and tracing -- both require a full pt_regs */
307int_very_careful:
308	TRACE_IRQS_ON
309	ENABLE_INTERRUPTS(CLBR_NONE)
310	SAVE_EXTRA_REGS
311	/* Check for syscall exit trace */
312	testl	$_TIF_WORK_SYSCALL_EXIT, %edx
313	jz	int_signal
314	pushq	%rdi
315	leaq	8(%rsp), %rdi			/* &ptregs -> arg1 */
316	call	syscall_trace_leave
317	popq	%rdi
318	andl	$~(_TIF_WORK_SYSCALL_EXIT|_TIF_SYSCALL_EMU), %edi
319	jmp	int_restore_rest
320
321int_signal:
322	testl	$_TIF_DO_NOTIFY_MASK, %edx
323	jz	1f
324	movq	%rsp, %rdi			/* &ptregs -> arg1 */
325	xorl	%esi, %esi			/* oldset -> arg2 */
326	call	do_notify_resume
3271:	movl	$_TIF_WORK_MASK, %edi
328int_restore_rest:
329	RESTORE_EXTRA_REGS
330	DISABLE_INTERRUPTS(CLBR_NONE)
331	TRACE_IRQS_OFF
332	jmp	int_with_check
333
334syscall_return:
335	/* The IRETQ could re-enable interrupts: */
336	DISABLE_INTERRUPTS(CLBR_ANY)
337	TRACE_IRQS_IRETQ
338
339	/*
340	 * Try to use SYSRET instead of IRET if we're returning to
341	 * a completely clean 64-bit userspace context.
342	 */
343	movq	RCX(%rsp), %rcx
344	movq	RIP(%rsp), %r11
345	cmpq	%rcx, %r11			/* RCX == RIP */
346	jne	opportunistic_sysret_failed
347
348	/*
349	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
350	 * in kernel space.  This essentially lets the user take over
351	 * the kernel, since userspace controls RSP.
352	 *
353	 * If width of "canonical tail" ever becomes variable, this will need
354	 * to be updated to remain correct on both old and new CPUs.
355	 */
356	.ifne __VIRTUAL_MASK_SHIFT - 47
357	.error "virtual address width changed -- SYSRET checks need update"
358	.endif
359
360	/* Change top 16 bits to be the sign-extension of 47th bit */
361	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
362	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
363
364	/* If this changed %rcx, it was not canonical */
365	cmpq	%rcx, %r11
366	jne	opportunistic_sysret_failed
367
368	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
369	jne	opportunistic_sysret_failed
370
371	movq	R11(%rsp), %r11
372	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
373	jne	opportunistic_sysret_failed
374
375	/*
376	 * SYSRET can't restore RF.  SYSRET can restore TF, but unlike IRET,
377	 * restoring TF results in a trap from userspace immediately after
378	 * SYSRET.  This would cause an infinite loop whenever #DB happens
379	 * with register state that satisfies the opportunistic SYSRET
380	 * conditions.  For example, single-stepping this user code:
381	 *
382	 *           movq	$stuck_here, %rcx
383	 *           pushfq
384	 *           popq %r11
385	 *   stuck_here:
386	 *
387	 * would never get past 'stuck_here'.
388	 */
389	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
390	jnz	opportunistic_sysret_failed
391
392	/* nothing to check for RSP */
393
394	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
395	jne	opportunistic_sysret_failed
396
397	/*
398	 * We win! This label is here just for ease of understanding
399	 * perf profiles. Nothing jumps here.
400	 */
401syscall_return_via_sysret:
402	/* rcx and r11 are already restored (see code above) */
403	RESTORE_C_REGS_EXCEPT_RCX_R11
404	movq	RSP(%rsp), %rsp
405	USERGS_SYSRET64
406
407opportunistic_sysret_failed:
408	SWAPGS
409	jmp	restore_c_regs_and_iret
410END(entry_SYSCALL_64)
411
412
413	.macro FORK_LIKE func
414ENTRY(stub_\func)
415	SAVE_EXTRA_REGS 8
416	jmp	sys_\func
417END(stub_\func)
418	.endm
419
420	FORK_LIKE  clone
421	FORK_LIKE  fork
422	FORK_LIKE  vfork
423
424ENTRY(stub_execve)
425	call	sys_execve
426return_from_execve:
427	testl	%eax, %eax
428	jz	1f
429	/* exec failed, can use fast SYSRET code path in this case */
430	ret
4311:
432	/* must use IRET code path (pt_regs->cs may have changed) */
433	addq	$8, %rsp
434	ZERO_EXTRA_REGS
435	movq	%rax, RAX(%rsp)
436	jmp	int_ret_from_sys_call
437END(stub_execve)
438/*
439 * Remaining execve stubs are only 7 bytes long.
440 * ENTRY() often aligns to 16 bytes, which in this case has no benefits.
441 */
442	.align	8
443GLOBAL(stub_execveat)
444	call	sys_execveat
445	jmp	return_from_execve
446END(stub_execveat)
447
448#if defined(CONFIG_X86_X32_ABI) || defined(CONFIG_IA32_EMULATION)
449	.align	8
450GLOBAL(stub_x32_execve)
451GLOBAL(stub32_execve)
452	call	compat_sys_execve
453	jmp	return_from_execve
454END(stub32_execve)
455END(stub_x32_execve)
456	.align	8
457GLOBAL(stub_x32_execveat)
458GLOBAL(stub32_execveat)
459	call	compat_sys_execveat
460	jmp	return_from_execve
461END(stub32_execveat)
462END(stub_x32_execveat)
463#endif
464
465/*
466 * sigreturn is special because it needs to restore all registers on return.
467 * This cannot be done with SYSRET, so use the IRET return path instead.
468 */
469ENTRY(stub_rt_sigreturn)
470	/*
471	 * SAVE_EXTRA_REGS result is not normally needed:
472	 * sigreturn overwrites all pt_regs->GPREGS.
473	 * But sigreturn can fail (!), and there is no easy way to detect that.
474	 * To make sure RESTORE_EXTRA_REGS doesn't restore garbage on error,
475	 * we SAVE_EXTRA_REGS here.
476	 */
477	SAVE_EXTRA_REGS 8
478	call	sys_rt_sigreturn
479return_from_stub:
480	addq	$8, %rsp
481	RESTORE_EXTRA_REGS
482	movq	%rax, RAX(%rsp)
483	jmp	int_ret_from_sys_call
484END(stub_rt_sigreturn)
485
486#ifdef CONFIG_X86_X32_ABI
487ENTRY(stub_x32_rt_sigreturn)
488	SAVE_EXTRA_REGS 8
489	call	sys32_x32_rt_sigreturn
490	jmp	return_from_stub
491END(stub_x32_rt_sigreturn)
492#endif
493
494/*
495 * A newly forked process directly context switches into this address.
496 *
497 * rdi: prev task we switched from
498 */
499ENTRY(ret_from_fork)
500
501	LOCK ; btr $TIF_FORK, TI_flags(%r8)
502
503	pushq	$0x0002
504	popfq					/* reset kernel eflags */
505
506	call	schedule_tail			/* rdi: 'prev' task parameter */
507
508	RESTORE_EXTRA_REGS
509
510	testb	$3, CS(%rsp)			/* from kernel_thread? */
511
512	/*
513	 * By the time we get here, we have no idea whether our pt_regs,
514	 * ti flags, and ti status came from the 64-bit SYSCALL fast path,
515	 * the slow path, or one of the 32-bit compat paths.
516	 * Use IRET code path to return, since it can safely handle
517	 * all of the above.
518	 */
519	jnz	int_ret_from_sys_call
520
521	/*
522	 * We came from kernel_thread
523	 * nb: we depend on RESTORE_EXTRA_REGS above
524	 */
525	movq	%rbp, %rdi
526	call	*%rbx
527	movl	$0, RAX(%rsp)
528	RESTORE_EXTRA_REGS
529	jmp	int_ret_from_sys_call
530END(ret_from_fork)
531
532/*
533 * Build the entry stubs with some assembler magic.
534 * We pack 1 stub into every 8-byte block.
535 */
536	.align 8
537ENTRY(irq_entries_start)
538    vector=FIRST_EXTERNAL_VECTOR
539    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
540	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
541    vector=vector+1
542	jmp	common_interrupt
543	.align	8
544    .endr
545END(irq_entries_start)
546
547/*
548 * Interrupt entry/exit.
549 *
550 * Interrupt entry points save only callee clobbered registers in fast path.
551 *
552 * Entry runs with interrupts off.
553 */
554
555/* 0(%rsp): ~(interrupt number) */
556	.macro interrupt func
557	cld
558	/*
559	 * Since nothing in interrupt handling code touches r12...r15 members
560	 * of "struct pt_regs", and since interrupts can nest, we can save
561	 * four stack slots and simultaneously provide
562	 * an unwind-friendly stack layout by saving "truncated" pt_regs
563	 * exactly up to rbp slot, without these members.
564	 */
565	ALLOC_PT_GPREGS_ON_STACK -RBP
566	SAVE_C_REGS -RBP
567	/* this goes to 0(%rsp) for unwinder, not for saving the value: */
568	SAVE_EXTRA_REGS_RBP -RBP
569
570	leaq	-RBP(%rsp), %rdi		/* arg1 for \func (pointer to pt_regs) */
571
572	testb	$3, CS-RBP(%rsp)
573	jz	1f
574	SWAPGS
5751:
576	/*
577	 * Save previous stack pointer, optionally switch to interrupt stack.
578	 * irq_count is used to check if a CPU is already on an interrupt stack
579	 * or not. While this is essentially redundant with preempt_count it is
580	 * a little cheaper to use a separate counter in the PDA (short of
581	 * moving irq_enter into assembly, which would be too much work)
582	 */
583	movq	%rsp, %rsi
584	incl	PER_CPU_VAR(irq_count)
585	cmovzq	PER_CPU_VAR(irq_stack_ptr), %rsp
586	pushq	%rsi
587	/* We entered an interrupt context - irqs are off: */
588	TRACE_IRQS_OFF
589
590	call	\func
591	.endm
592
593	/*
594	 * The interrupt stubs push (~vector+0x80) onto the stack and
595	 * then jump to common_interrupt.
596	 */
597	.p2align CONFIG_X86_L1_CACHE_SHIFT
598common_interrupt:
599	ASM_CLAC
600	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
601	interrupt do_IRQ
602	/* 0(%rsp): old RSP */
603ret_from_intr:
604	DISABLE_INTERRUPTS(CLBR_NONE)
605	TRACE_IRQS_OFF
606	decl	PER_CPU_VAR(irq_count)
607
608	/* Restore saved previous stack */
609	popq	%rsi
610	/* return code expects complete pt_regs - adjust rsp accordingly: */
611	leaq	-RBP(%rsi), %rsp
612
613	testb	$3, CS(%rsp)
614	jz	retint_kernel
615	/* Interrupt came from user space */
616retint_user:
617	GET_THREAD_INFO(%rcx)
618
619	/* %rcx: thread info. Interrupts are off. */
620retint_with_reschedule:
621	movl	$_TIF_WORK_MASK, %edi
622retint_check:
623	LOCKDEP_SYS_EXIT_IRQ
624	movl	TI_flags(%rcx), %edx
625	andl	%edi, %edx
626	jnz	retint_careful
627
628retint_swapgs:					/* return to user-space */
629	/*
630	 * The iretq could re-enable interrupts:
631	 */
632	DISABLE_INTERRUPTS(CLBR_ANY)
633	TRACE_IRQS_IRETQ
634
635	SWAPGS
636	jmp	restore_c_regs_and_iret
637
638/* Returning to kernel space */
639retint_kernel:
640#ifdef CONFIG_PREEMPT
641	/* Interrupts are off */
642	/* Check if we need preemption */
643	bt	$9, EFLAGS(%rsp)		/* were interrupts off? */
644	jnc	1f
6450:	cmpl	$0, PER_CPU_VAR(__preempt_count)
646	jnz	1f
647	call	preempt_schedule_irq
648	jmp	0b
6491:
650#endif
651	/*
652	 * The iretq could re-enable interrupts:
653	 */
654	TRACE_IRQS_IRETQ
655
656/*
657 * At this label, code paths which return to kernel and to user,
658 * which come from interrupts/exception and from syscalls, merge.
659 */
660restore_c_regs_and_iret:
661	RESTORE_C_REGS
662	REMOVE_PT_GPREGS_FROM_STACK 8
663	INTERRUPT_RETURN
664
665ENTRY(native_iret)
666	/*
667	 * Are we returning to a stack segment from the LDT?  Note: in
668	 * 64-bit mode SS:RSP on the exception stack is always valid.
669	 */
670#ifdef CONFIG_X86_ESPFIX64
671	testb	$4, (SS-RIP)(%rsp)
672	jnz	native_irq_return_ldt
673#endif
674
675.global native_irq_return_iret
676native_irq_return_iret:
677	/*
678	 * This may fault.  Non-paranoid faults on return to userspace are
679	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
680	 * Double-faults due to espfix64 are handled in do_double_fault.
681	 * Other faults here are fatal.
682	 */
683	iretq
684
685#ifdef CONFIG_X86_ESPFIX64
686native_irq_return_ldt:
687	pushq	%rax
688	pushq	%rdi
689	SWAPGS
690	movq	PER_CPU_VAR(espfix_waddr), %rdi
691	movq	%rax, (0*8)(%rdi)		/* RAX */
692	movq	(2*8)(%rsp), %rax		/* RIP */
693	movq	%rax, (1*8)(%rdi)
694	movq	(3*8)(%rsp), %rax		/* CS */
695	movq	%rax, (2*8)(%rdi)
696	movq	(4*8)(%rsp), %rax		/* RFLAGS */
697	movq	%rax, (3*8)(%rdi)
698	movq	(6*8)(%rsp), %rax		/* SS */
699	movq	%rax, (5*8)(%rdi)
700	movq	(5*8)(%rsp), %rax		/* RSP */
701	movq	%rax, (4*8)(%rdi)
702	andl	$0xffff0000, %eax
703	popq	%rdi
704	orq	PER_CPU_VAR(espfix_stack), %rax
705	SWAPGS
706	movq	%rax, %rsp
707	popq	%rax
708	jmp	native_irq_return_iret
709#endif
710
711	/* edi: workmask, edx: work */
712retint_careful:
713	bt	$TIF_NEED_RESCHED, %edx
714	jnc	retint_signal
715	TRACE_IRQS_ON
716	ENABLE_INTERRUPTS(CLBR_NONE)
717	pushq	%rdi
718	SCHEDULE_USER
719	popq	%rdi
720	GET_THREAD_INFO(%rcx)
721	DISABLE_INTERRUPTS(CLBR_NONE)
722	TRACE_IRQS_OFF
723	jmp	retint_check
724
725retint_signal:
726	testl	$_TIF_DO_NOTIFY_MASK, %edx
727	jz	retint_swapgs
728	TRACE_IRQS_ON
729	ENABLE_INTERRUPTS(CLBR_NONE)
730	SAVE_EXTRA_REGS
731	movq	$-1, ORIG_RAX(%rsp)
732	xorl	%esi, %esi			/* oldset */
733	movq	%rsp, %rdi			/* &pt_regs */
734	call	do_notify_resume
735	RESTORE_EXTRA_REGS
736	DISABLE_INTERRUPTS(CLBR_NONE)
737	TRACE_IRQS_OFF
738	GET_THREAD_INFO(%rcx)
739	jmp	retint_with_reschedule
740
741END(common_interrupt)
742
743/*
744 * APIC interrupts.
745 */
746.macro apicinterrupt3 num sym do_sym
747ENTRY(\sym)
748	ASM_CLAC
749	pushq	$~(\num)
750.Lcommon_\sym:
751	interrupt \do_sym
752	jmp	ret_from_intr
753END(\sym)
754.endm
755
756#ifdef CONFIG_TRACING
757#define trace(sym) trace_##sym
758#define smp_trace(sym) smp_trace_##sym
759
760.macro trace_apicinterrupt num sym
761apicinterrupt3 \num trace(\sym) smp_trace(\sym)
762.endm
763#else
764.macro trace_apicinterrupt num sym do_sym
765.endm
766#endif
767
768.macro apicinterrupt num sym do_sym
769apicinterrupt3 \num \sym \do_sym
770trace_apicinterrupt \num \sym
771.endm
772
773#ifdef CONFIG_SMP
774apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
775apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
776#endif
777
778#ifdef CONFIG_X86_UV
779apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
780#endif
781
782apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
783apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
784
785#ifdef CONFIG_HAVE_KVM
786apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
787apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
788#endif
789
790#ifdef CONFIG_X86_MCE_THRESHOLD
791apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
792#endif
793
794#ifdef CONFIG_X86_MCE_AMD
795apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
796#endif
797
798#ifdef CONFIG_X86_THERMAL_VECTOR
799apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
800#endif
801
802#ifdef CONFIG_SMP
803apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
804apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
805apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
806#endif
807
808apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
809apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
810
811#ifdef CONFIG_IRQ_WORK
812apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
813#endif
814
815/*
816 * Exception entry points.
817 */
818#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss) + (TSS_ist + ((x) - 1) * 8)
819
820.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
821ENTRY(\sym)
822	/* Sanity check */
823	.if \shift_ist != -1 && \paranoid == 0
824	.error "using shift_ist requires paranoid=1"
825	.endif
826
827	ASM_CLAC
828	PARAVIRT_ADJUST_EXCEPTION_FRAME
829
830	.ifeq \has_error_code
831	pushq	$-1				/* ORIG_RAX: no syscall to restart */
832	.endif
833
834	ALLOC_PT_GPREGS_ON_STACK
835
836	.if \paranoid
837	.if \paranoid == 1
838	testb	$3, CS(%rsp)			/* If coming from userspace, switch stacks */
839	jnz	1f
840	.endif
841	call	paranoid_entry
842	.else
843	call	error_entry
844	.endif
845	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
846
847	.if \paranoid
848	.if \shift_ist != -1
849	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
850	.else
851	TRACE_IRQS_OFF
852	.endif
853	.endif
854
855	movq	%rsp, %rdi			/* pt_regs pointer */
856
857	.if \has_error_code
858	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
859	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
860	.else
861	xorl	%esi, %esi			/* no error code */
862	.endif
863
864	.if \shift_ist != -1
865	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
866	.endif
867
868	call	\do_sym
869
870	.if \shift_ist != -1
871	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
872	.endif
873
874	/* these procedures expect "no swapgs" flag in ebx */
875	.if \paranoid
876	jmp	paranoid_exit
877	.else
878	jmp	error_exit
879	.endif
880
881	.if \paranoid == 1
882	/*
883	 * Paranoid entry from userspace.  Switch stacks and treat it
884	 * as a normal entry.  This means that paranoid handlers
885	 * run in real process context if user_mode(regs).
886	 */
8871:
888	call	error_entry
889
890
891	movq	%rsp, %rdi			/* pt_regs pointer */
892	call	sync_regs
893	movq	%rax, %rsp			/* switch stack */
894
895	movq	%rsp, %rdi			/* pt_regs pointer */
896
897	.if \has_error_code
898	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
899	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
900	.else
901	xorl	%esi, %esi			/* no error code */
902	.endif
903
904	call	\do_sym
905
906	jmp	error_exit			/* %ebx: no swapgs flag */
907	.endif
908END(\sym)
909.endm
910
911#ifdef CONFIG_TRACING
912.macro trace_idtentry sym do_sym has_error_code:req
913idtentry trace(\sym) trace(\do_sym) has_error_code=\has_error_code
914idtentry \sym \do_sym has_error_code=\has_error_code
915.endm
916#else
917.macro trace_idtentry sym do_sym has_error_code:req
918idtentry \sym \do_sym has_error_code=\has_error_code
919.endm
920#endif
921
922idtentry divide_error			do_divide_error			has_error_code=0
923idtentry overflow			do_overflow			has_error_code=0
924idtentry bounds				do_bounds			has_error_code=0
925idtentry invalid_op			do_invalid_op			has_error_code=0
926idtentry device_not_available		do_device_not_available		has_error_code=0
927idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
928idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
929idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
930idtentry segment_not_present		do_segment_not_present		has_error_code=1
931idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
932idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
933idtentry alignment_check		do_alignment_check		has_error_code=1
934idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
935
936
937	/*
938	 * Reload gs selector with exception handling
939	 * edi:  new selector
940	 */
941ENTRY(native_load_gs_index)
942	pushfq
943	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
944	SWAPGS
945gs_change:
946	movl	%edi, %gs
9472:	mfence					/* workaround */
948	SWAPGS
949	popfq
950	ret
951END(native_load_gs_index)
952
953	_ASM_EXTABLE(gs_change, bad_gs)
954	.section .fixup, "ax"
955	/* running with kernelgs */
956bad_gs:
957	SWAPGS					/* switch back to user gs */
958	xorl	%eax, %eax
959	movl	%eax, %gs
960	jmp	2b
961	.previous
962
963/* Call softirq on interrupt stack. Interrupts are off. */
964ENTRY(do_softirq_own_stack)
965	pushq	%rbp
966	mov	%rsp, %rbp
967	incl	PER_CPU_VAR(irq_count)
968	cmove	PER_CPU_VAR(irq_stack_ptr), %rsp
969	push	%rbp				/* frame pointer backlink */
970	call	__do_softirq
971	leaveq
972	decl	PER_CPU_VAR(irq_count)
973	ret
974END(do_softirq_own_stack)
975
976#ifdef CONFIG_XEN
977idtentry xen_hypervisor_callback xen_do_hypervisor_callback has_error_code=0
978
979/*
980 * A note on the "critical region" in our callback handler.
981 * We want to avoid stacking callback handlers due to events occurring
982 * during handling of the last event. To do this, we keep events disabled
983 * until we've done all processing. HOWEVER, we must enable events before
984 * popping the stack frame (can't be done atomically) and so it would still
985 * be possible to get enough handler activations to overflow the stack.
986 * Although unlikely, bugs of that kind are hard to track down, so we'd
987 * like to avoid the possibility.
988 * So, on entry to the handler we detect whether we interrupted an
989 * existing activation in its critical region -- if so, we pop the current
990 * activation and restart the handler using the previous one.
991 */
992ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
993
994/*
995 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
996 * see the correct pointer to the pt_regs
997 */
998	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
99911:	incl	PER_CPU_VAR(irq_count)
1000	movq	%rsp, %rbp
1001	cmovzq	PER_CPU_VAR(irq_stack_ptr), %rsp
1002	pushq	%rbp				/* frame pointer backlink */
1003	call	xen_evtchn_do_upcall
1004	popq	%rsp
1005	decl	PER_CPU_VAR(irq_count)
1006#ifndef CONFIG_PREEMPT
1007	call	xen_maybe_preempt_hcall
1008#endif
1009	jmp	error_exit
1010END(xen_do_hypervisor_callback)
1011
1012/*
1013 * Hypervisor uses this for application faults while it executes.
1014 * We get here for two reasons:
1015 *  1. Fault while reloading DS, ES, FS or GS
1016 *  2. Fault while executing IRET
1017 * Category 1 we do not need to fix up as Xen has already reloaded all segment
1018 * registers that could be reloaded and zeroed the others.
1019 * Category 2 we fix up by killing the current process. We cannot use the
1020 * normal Linux return path in this case because if we use the IRET hypercall
1021 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
1022 * We distinguish between categories by comparing each saved segment register
1023 * with its current contents: any discrepancy means we in category 1.
1024 */
1025ENTRY(xen_failsafe_callback)
1026	movl	%ds, %ecx
1027	cmpw	%cx, 0x10(%rsp)
1028	jne	1f
1029	movl	%es, %ecx
1030	cmpw	%cx, 0x18(%rsp)
1031	jne	1f
1032	movl	%fs, %ecx
1033	cmpw	%cx, 0x20(%rsp)
1034	jne	1f
1035	movl	%gs, %ecx
1036	cmpw	%cx, 0x28(%rsp)
1037	jne	1f
1038	/* All segments match their saved values => Category 2 (Bad IRET). */
1039	movq	(%rsp), %rcx
1040	movq	8(%rsp), %r11
1041	addq	$0x30, %rsp
1042	pushq	$0				/* RIP */
1043	pushq	%r11
1044	pushq	%rcx
1045	jmp	general_protection
10461:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
1047	movq	(%rsp), %rcx
1048	movq	8(%rsp), %r11
1049	addq	$0x30, %rsp
1050	pushq	$-1 /* orig_ax = -1 => not a system call */
1051	ALLOC_PT_GPREGS_ON_STACK
1052	SAVE_C_REGS
1053	SAVE_EXTRA_REGS
1054	jmp	error_exit
1055END(xen_failsafe_callback)
1056
1057apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1058	xen_hvm_callback_vector xen_evtchn_do_upcall
1059
1060#endif /* CONFIG_XEN */
1061
1062#if IS_ENABLED(CONFIG_HYPERV)
1063apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1064	hyperv_callback_vector hyperv_vector_handler
1065#endif /* CONFIG_HYPERV */
1066
1067idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1068idtentry int3			do_int3			has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1069idtentry stack_segment		do_stack_segment	has_error_code=1
1070
1071#ifdef CONFIG_XEN
1072idtentry xen_debug		do_debug		has_error_code=0
1073idtentry xen_int3		do_int3			has_error_code=0
1074idtentry xen_stack_segment	do_stack_segment	has_error_code=1
1075#endif
1076
1077idtentry general_protection	do_general_protection	has_error_code=1
1078trace_idtentry page_fault	do_page_fault		has_error_code=1
1079
1080#ifdef CONFIG_KVM_GUEST
1081idtentry async_page_fault	do_async_page_fault	has_error_code=1
1082#endif
1083
1084#ifdef CONFIG_X86_MCE
1085idtentry machine_check					has_error_code=0	paranoid=1 do_sym=*machine_check_vector(%rip)
1086#endif
1087
1088/*
1089 * Save all registers in pt_regs, and switch gs if needed.
1090 * Use slow, but surefire "are we in kernel?" check.
1091 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
1092 */
1093ENTRY(paranoid_entry)
1094	cld
1095	SAVE_C_REGS 8
1096	SAVE_EXTRA_REGS 8
1097	movl	$1, %ebx
1098	movl	$MSR_GS_BASE, %ecx
1099	rdmsr
1100	testl	%edx, %edx
1101	js	1f				/* negative -> in kernel */
1102	SWAPGS
1103	xorl	%ebx, %ebx
11041:	ret
1105END(paranoid_entry)
1106
1107/*
1108 * "Paranoid" exit path from exception stack.  This is invoked
1109 * only on return from non-NMI IST interrupts that came
1110 * from kernel space.
1111 *
1112 * We may be returning to very strange contexts (e.g. very early
1113 * in syscall entry), so checking for preemption here would
1114 * be complicated.  Fortunately, we there's no good reason
1115 * to try to handle preemption here.
1116 *
1117 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1118 */
1119ENTRY(paranoid_exit)
1120	DISABLE_INTERRUPTS(CLBR_NONE)
1121	TRACE_IRQS_OFF_DEBUG
1122	testl	%ebx, %ebx			/* swapgs needed? */
1123	jnz	paranoid_exit_no_swapgs
1124	TRACE_IRQS_IRETQ
1125	SWAPGS_UNSAFE_STACK
1126	jmp	paranoid_exit_restore
1127paranoid_exit_no_swapgs:
1128	TRACE_IRQS_IRETQ_DEBUG
1129paranoid_exit_restore:
1130	RESTORE_EXTRA_REGS
1131	RESTORE_C_REGS
1132	REMOVE_PT_GPREGS_FROM_STACK 8
1133	INTERRUPT_RETURN
1134END(paranoid_exit)
1135
1136/*
1137 * Save all registers in pt_regs, and switch gs if needed.
1138 * Return: EBX=0: came from user mode; EBX=1: otherwise
1139 */
1140ENTRY(error_entry)
1141	cld
1142	SAVE_C_REGS 8
1143	SAVE_EXTRA_REGS 8
1144	xorl	%ebx, %ebx
1145	testb	$3, CS+8(%rsp)
1146	jz	error_kernelspace
1147
1148	/* We entered from user mode */
1149	SWAPGS
1150
1151error_entry_done:
1152	TRACE_IRQS_OFF
1153	ret
1154
1155	/*
1156	 * There are two places in the kernel that can potentially fault with
1157	 * usergs. Handle them here.  B stepping K8s sometimes report a
1158	 * truncated RIP for IRET exceptions returning to compat mode. Check
1159	 * for these here too.
1160	 */
1161error_kernelspace:
1162	incl	%ebx
1163	leaq	native_irq_return_iret(%rip), %rcx
1164	cmpq	%rcx, RIP+8(%rsp)
1165	je	error_bad_iret
1166	movl	%ecx, %eax			/* zero extend */
1167	cmpq	%rax, RIP+8(%rsp)
1168	je	bstep_iret
1169	cmpq	$gs_change, RIP+8(%rsp)
1170	jne	error_entry_done
1171
1172	/*
1173	 * hack: gs_change can fail with user gsbase.  If this happens, fix up
1174	 * gsbase and proceed.  We'll fix up the exception and land in
1175	 * gs_change's error handler with kernel gsbase.
1176	 */
1177	SWAPGS
1178	jmp	error_entry_done
1179
1180bstep_iret:
1181	/* Fix truncated RIP */
1182	movq	%rcx, RIP+8(%rsp)
1183	/* fall through */
1184
1185error_bad_iret:
1186	/*
1187	 * We came from an IRET to user mode, so we have user gsbase.
1188	 * Switch to kernel gsbase:
1189	 */
1190	SWAPGS
1191
1192	/*
1193	 * Pretend that the exception came from user mode: set up pt_regs
1194	 * as if we faulted immediately after IRET and clear EBX so that
1195	 * error_exit knows that we will be returning to user mode.
1196	 */
1197	mov	%rsp, %rdi
1198	call	fixup_bad_iret
1199	mov	%rax, %rsp
1200	decl	%ebx
1201	jmp	error_entry_done
1202END(error_entry)
1203
1204
1205/*
1206 * On entry, EBS is a "return to kernel mode" flag:
1207 *   1: already in kernel mode, don't need SWAPGS
1208 *   0: user gsbase is loaded, we need SWAPGS and standard preparation for return to usermode
1209 */
1210ENTRY(error_exit)
1211	movl	%ebx, %eax
1212	RESTORE_EXTRA_REGS
1213	DISABLE_INTERRUPTS(CLBR_NONE)
1214	TRACE_IRQS_OFF
1215	testl	%eax, %eax
1216	jnz	retint_kernel
1217	jmp	retint_user
1218END(error_exit)
1219
1220/* Runs on exception stack */
1221ENTRY(nmi)
1222	PARAVIRT_ADJUST_EXCEPTION_FRAME
1223	/*
1224	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1225	 * the iretq it performs will take us out of NMI context.
1226	 * This means that we can have nested NMIs where the next
1227	 * NMI is using the top of the stack of the previous NMI. We
1228	 * can't let it execute because the nested NMI will corrupt the
1229	 * stack of the previous NMI. NMI handlers are not re-entrant
1230	 * anyway.
1231	 *
1232	 * To handle this case we do the following:
1233	 *  Check the a special location on the stack that contains
1234	 *  a variable that is set when NMIs are executing.
1235	 *  The interrupted task's stack is also checked to see if it
1236	 *  is an NMI stack.
1237	 *  If the variable is not set and the stack is not the NMI
1238	 *  stack then:
1239	 *    o Set the special variable on the stack
1240	 *    o Copy the interrupt frame into an "outermost" location on the
1241	 *      stack
1242	 *    o Copy the interrupt frame into an "iret" location on the stack
1243	 *    o Continue processing the NMI
1244	 *  If the variable is set or the previous stack is the NMI stack:
1245	 *    o Modify the "iret" location to jump to the repeat_nmi
1246	 *    o return back to the first NMI
1247	 *
1248	 * Now on exit of the first NMI, we first clear the stack variable
1249	 * The NMI stack will tell any nested NMIs at that point that it is
1250	 * nested. Then we pop the stack normally with iret, and if there was
1251	 * a nested NMI that updated the copy interrupt stack frame, a
1252	 * jump will be made to the repeat_nmi code that will handle the second
1253	 * NMI.
1254	 *
1255	 * However, espfix prevents us from directly returning to userspace
1256	 * with a single IRET instruction.  Similarly, IRET to user mode
1257	 * can fault.  We therefore handle NMIs from user space like
1258	 * other IST entries.
1259	 */
1260
1261	/* Use %rdx as our temp variable throughout */
1262	pushq	%rdx
1263
1264	testb	$3, CS-RIP+8(%rsp)
1265	jz	.Lnmi_from_kernel
1266
1267	/*
1268	 * NMI from user mode.  We need to run on the thread stack, but we
1269	 * can't go through the normal entry paths: NMIs are masked, and
1270	 * we don't want to enable interrupts, because then we'll end
1271	 * up in an awkward situation in which IRQs are on but NMIs
1272	 * are off.
1273	 */
1274
1275	SWAPGS
1276	cld
1277	movq	%rsp, %rdx
1278	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1279	pushq	5*8(%rdx)	/* pt_regs->ss */
1280	pushq	4*8(%rdx)	/* pt_regs->rsp */
1281	pushq	3*8(%rdx)	/* pt_regs->flags */
1282	pushq	2*8(%rdx)	/* pt_regs->cs */
1283	pushq	1*8(%rdx)	/* pt_regs->rip */
1284	pushq   $-1		/* pt_regs->orig_ax */
1285	pushq   %rdi		/* pt_regs->di */
1286	pushq   %rsi		/* pt_regs->si */
1287	pushq   (%rdx)		/* pt_regs->dx */
1288	pushq   %rcx		/* pt_regs->cx */
1289	pushq   %rax		/* pt_regs->ax */
1290	pushq   %r8		/* pt_regs->r8 */
1291	pushq   %r9		/* pt_regs->r9 */
1292	pushq   %r10		/* pt_regs->r10 */
1293	pushq   %r11		/* pt_regs->r11 */
1294	pushq	%rbx		/* pt_regs->rbx */
1295	pushq	%rbp		/* pt_regs->rbp */
1296	pushq	%r12		/* pt_regs->r12 */
1297	pushq	%r13		/* pt_regs->r13 */
1298	pushq	%r14		/* pt_regs->r14 */
1299	pushq	%r15		/* pt_regs->r15 */
1300
1301	/*
1302	 * At this point we no longer need to worry about stack damage
1303	 * due to nesting -- we're on the normal thread stack and we're
1304	 * done with the NMI stack.
1305	 */
1306
1307	movq	%rsp, %rdi
1308	movq	$-1, %rsi
1309	call	do_nmi
1310
1311	/*
1312	 * Return back to user mode.  We must *not* do the normal exit
1313	 * work, because we don't want to enable interrupts.  Fortunately,
1314	 * do_nmi doesn't modify pt_regs.
1315	 */
1316	SWAPGS
1317	jmp	restore_c_regs_and_iret
1318
1319.Lnmi_from_kernel:
1320	/*
1321	 * Here's what our stack frame will look like:
1322	 * +---------------------------------------------------------+
1323	 * | original SS                                             |
1324	 * | original Return RSP                                     |
1325	 * | original RFLAGS                                         |
1326	 * | original CS                                             |
1327	 * | original RIP                                            |
1328	 * +---------------------------------------------------------+
1329	 * | temp storage for rdx                                    |
1330	 * +---------------------------------------------------------+
1331	 * | "NMI executing" variable                                |
1332	 * +---------------------------------------------------------+
1333	 * | iret SS          } Copied from "outermost" frame        |
1334	 * | iret Return RSP  } on each loop iteration; overwritten  |
1335	 * | iret RFLAGS      } by a nested NMI to force another     |
1336	 * | iret CS          } iteration if needed.                 |
1337	 * | iret RIP         }                                      |
1338	 * +---------------------------------------------------------+
1339	 * | outermost SS          } initialized in first_nmi;       |
1340	 * | outermost Return RSP  } will not be changed before      |
1341	 * | outermost RFLAGS      } NMI processing is done.         |
1342	 * | outermost CS          } Copied to "iret" frame on each  |
1343	 * | outermost RIP         } iteration.                      |
1344	 * +---------------------------------------------------------+
1345	 * | pt_regs                                                 |
1346	 * +---------------------------------------------------------+
1347	 *
1348	 * The "original" frame is used by hardware.  Before re-enabling
1349	 * NMIs, we need to be done with it, and we need to leave enough
1350	 * space for the asm code here.
1351	 *
1352	 * We return by executing IRET while RSP points to the "iret" frame.
1353	 * That will either return for real or it will loop back into NMI
1354	 * processing.
1355	 *
1356	 * The "outermost" frame is copied to the "iret" frame on each
1357	 * iteration of the loop, so each iteration starts with the "iret"
1358	 * frame pointing to the final return target.
1359	 */
1360
1361	/*
1362	 * Determine whether we're a nested NMI.
1363	 *
1364	 * If we interrupted kernel code between repeat_nmi and
1365	 * end_repeat_nmi, then we are a nested NMI.  We must not
1366	 * modify the "iret" frame because it's being written by
1367	 * the outer NMI.  That's okay; the outer NMI handler is
1368	 * about to about to call do_nmi anyway, so we can just
1369	 * resume the outer NMI.
1370	 */
1371
1372	movq	$repeat_nmi, %rdx
1373	cmpq	8(%rsp), %rdx
1374	ja	1f
1375	movq	$end_repeat_nmi, %rdx
1376	cmpq	8(%rsp), %rdx
1377	ja	nested_nmi_out
13781:
1379
1380	/*
1381	 * Now check "NMI executing".  If it's set, then we're nested.
1382	 * This will not detect if we interrupted an outer NMI just
1383	 * before IRET.
1384	 */
1385	cmpl	$1, -8(%rsp)
1386	je	nested_nmi
1387
1388	/*
1389	 * Now test if the previous stack was an NMI stack.  This covers
1390	 * the case where we interrupt an outer NMI after it clears
1391	 * "NMI executing" but before IRET.  We need to be careful, though:
1392	 * there is one case in which RSP could point to the NMI stack
1393	 * despite there being no NMI active: naughty userspace controls
1394	 * RSP at the very beginning of the SYSCALL targets.  We can
1395	 * pull a fast one on naughty userspace, though: we program
1396	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1397	 * if it controls the kernel's RSP.  We set DF before we clear
1398	 * "NMI executing".
1399	 */
1400	lea	6*8(%rsp), %rdx
1401	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1402	cmpq	%rdx, 4*8(%rsp)
1403	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1404	ja	first_nmi
1405
1406	subq	$EXCEPTION_STKSZ, %rdx
1407	cmpq	%rdx, 4*8(%rsp)
1408	/* If it is below the NMI stack, it is a normal NMI */
1409	jb	first_nmi
1410
1411	/* Ah, it is within the NMI stack. */
1412
1413	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1414	jz	first_nmi	/* RSP was user controlled. */
1415
1416	/* This is a nested NMI. */
1417
1418nested_nmi:
1419	/*
1420	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1421	 * iteration of NMI handling.
1422	 */
1423	subq	$8, %rsp
1424	leaq	-10*8(%rsp), %rdx
1425	pushq	$__KERNEL_DS
1426	pushq	%rdx
1427	pushfq
1428	pushq	$__KERNEL_CS
1429	pushq	$repeat_nmi
1430
1431	/* Put stack back */
1432	addq	$(6*8), %rsp
1433
1434nested_nmi_out:
1435	popq	%rdx
1436
1437	/* We are returning to kernel mode, so this cannot result in a fault. */
1438	INTERRUPT_RETURN
1439
1440first_nmi:
1441	/* Restore rdx. */
1442	movq	(%rsp), %rdx
1443
1444	/* Make room for "NMI executing". */
1445	pushq	$0
1446
1447	/* Leave room for the "iret" frame */
1448	subq	$(5*8), %rsp
1449
1450	/* Copy the "original" frame to the "outermost" frame */
1451	.rept 5
1452	pushq	11*8(%rsp)
1453	.endr
1454
1455	/* Everything up to here is safe from nested NMIs */
1456
1457#ifdef CONFIG_DEBUG_ENTRY
1458	/*
1459	 * For ease of testing, unmask NMIs right away.  Disabled by
1460	 * default because IRET is very expensive.
1461	 */
1462	pushq	$0		/* SS */
1463	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1464	addq	$8, (%rsp)	/* Fix up RSP */
1465	pushfq			/* RFLAGS */
1466	pushq	$__KERNEL_CS	/* CS */
1467	pushq	$1f		/* RIP */
1468	INTERRUPT_RETURN	/* continues at repeat_nmi below */
14691:
1470#endif
1471
1472repeat_nmi:
1473	/*
1474	 * If there was a nested NMI, the first NMI's iret will return
1475	 * here. But NMIs are still enabled and we can take another
1476	 * nested NMI. The nested NMI checks the interrupted RIP to see
1477	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1478	 * it will just return, as we are about to repeat an NMI anyway.
1479	 * This makes it safe to copy to the stack frame that a nested
1480	 * NMI will update.
1481	 *
1482	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1483	 * we're repeating an NMI, gsbase has the same value that it had on
1484	 * the first iteration.  paranoid_entry will load the kernel
1485	 * gsbase if needed before we call do_nmi.  "NMI executing"
1486	 * is zero.
1487	 */
1488	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1489
1490	/*
1491	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1492	 * here must not modify the "iret" frame while we're writing to
1493	 * it or it will end up containing garbage.
1494	 */
1495	addq	$(10*8), %rsp
1496	.rept 5
1497	pushq	-6*8(%rsp)
1498	.endr
1499	subq	$(5*8), %rsp
1500end_repeat_nmi:
1501
1502	/*
1503	 * Everything below this point can be preempted by a nested NMI.
1504	 * If this happens, then the inner NMI will change the "iret"
1505	 * frame to point back to repeat_nmi.
1506	 */
1507	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1508	ALLOC_PT_GPREGS_ON_STACK
1509
1510	/*
1511	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1512	 * as we should not be calling schedule in NMI context.
1513	 * Even with normal interrupts enabled. An NMI should not be
1514	 * setting NEED_RESCHED or anything that normal interrupts and
1515	 * exceptions might do.
1516	 */
1517	call	paranoid_entry
1518
1519	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1520	movq	%rsp, %rdi
1521	movq	$-1, %rsi
1522	call	do_nmi
1523
1524	testl	%ebx, %ebx			/* swapgs needed? */
1525	jnz	nmi_restore
1526nmi_swapgs:
1527	SWAPGS_UNSAFE_STACK
1528nmi_restore:
1529	RESTORE_EXTRA_REGS
1530	RESTORE_C_REGS
1531
1532	/* Point RSP at the "iret" frame. */
1533	REMOVE_PT_GPREGS_FROM_STACK 6*8
1534
1535	/*
1536	 * Clear "NMI executing".  Set DF first so that we can easily
1537	 * distinguish the remaining code between here and IRET from
1538	 * the SYSCALL entry and exit paths.  On a native kernel, we
1539	 * could just inspect RIP, but, on paravirt kernels,
1540	 * INTERRUPT_RETURN can translate into a jump into a
1541	 * hypercall page.
1542	 */
1543	std
1544	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1545
1546	/*
1547	 * INTERRUPT_RETURN reads the "iret" frame and exits the NMI
1548	 * stack in a single instruction.  We are returning to kernel
1549	 * mode, so this cannot result in a fault.
1550	 */
1551	INTERRUPT_RETURN
1552END(nmi)
1553
1554ENTRY(ignore_sysret)
1555	mov	$-ENOSYS, %eax
1556	sysret
1557END(ignore_sysret)
1558