xref: /linux/arch/x86/entry/entry_64.S (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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/smap.h>
37#include <asm/pgtable_types.h>
38#include <linux/err.h>
39
40/* Avoid __ASSEMBLER__'ifying <linux/audit.h> just for this.  */
41#include <linux/elf-em.h>
42#define AUDIT_ARCH_X86_64			(EM_X86_64|__AUDIT_ARCH_64BIT|__AUDIT_ARCH_LE)
43#define __AUDIT_ARCH_64BIT			0x80000000
44#define __AUDIT_ARCH_LE				0x40000000
45
46.code64
47.section .entry.text, "ax"
48
49#ifdef CONFIG_PARAVIRT
50ENTRY(native_usergs_sysret64)
51	swapgs
52	sysretq
53ENDPROC(native_usergs_sysret64)
54#endif /* CONFIG_PARAVIRT */
55
56.macro TRACE_IRQS_IRETQ
57#ifdef CONFIG_TRACE_IRQFLAGS
58	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
59	jnc	1f
60	TRACE_IRQS_ON
611:
62#endif
63.endm
64
65/*
66 * When dynamic function tracer is enabled it will add a breakpoint
67 * to all locations that it is about to modify, sync CPUs, update
68 * all the code, sync CPUs, then remove the breakpoints. In this time
69 * if lockdep is enabled, it might jump back into the debug handler
70 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF).
71 *
72 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to
73 * make sure the stack pointer does not get reset back to the top
74 * of the debug stack, and instead just reuses the current stack.
75 */
76#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS)
77
78.macro TRACE_IRQS_OFF_DEBUG
79	call	debug_stack_set_zero
80	TRACE_IRQS_OFF
81	call	debug_stack_reset
82.endm
83
84.macro TRACE_IRQS_ON_DEBUG
85	call	debug_stack_set_zero
86	TRACE_IRQS_ON
87	call	debug_stack_reset
88.endm
89
90.macro TRACE_IRQS_IRETQ_DEBUG
91	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
92	jnc	1f
93	TRACE_IRQS_ON_DEBUG
941:
95.endm
96
97#else
98# define TRACE_IRQS_OFF_DEBUG			TRACE_IRQS_OFF
99# define TRACE_IRQS_ON_DEBUG			TRACE_IRQS_ON
100# define TRACE_IRQS_IRETQ_DEBUG			TRACE_IRQS_IRETQ
101#endif
102
103/*
104 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
105 *
106 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
107 * then loads new ss, cs, and rip from previously programmed MSRs.
108 * rflags gets masked by a value from another MSR (so CLD and CLAC
109 * are not needed). SYSCALL does not save anything on the stack
110 * and does not change rsp.
111 *
112 * Registers on entry:
113 * rax  system call number
114 * rcx  return address
115 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
116 * rdi  arg0
117 * rsi  arg1
118 * rdx  arg2
119 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
120 * r8   arg4
121 * r9   arg5
122 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
123 *
124 * Only called from user space.
125 *
126 * When user can change pt_regs->foo always force IRET. That is because
127 * it deals with uncanonical addresses better. SYSRET has trouble
128 * with them due to bugs in both AMD and Intel CPUs.
129 */
130
131ENTRY(entry_SYSCALL_64)
132	/*
133	 * Interrupts are off on entry.
134	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
135	 * it is too small to ever cause noticeable irq latency.
136	 */
137	SWAPGS_UNSAFE_STACK
138	/*
139	 * A hypervisor implementation might want to use a label
140	 * after the swapgs, so that it can do the swapgs
141	 * for the guest and jump here on syscall.
142	 */
143GLOBAL(entry_SYSCALL_64_after_swapgs)
144
145	movq	%rsp, PER_CPU_VAR(rsp_scratch)
146	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
147
148	/* Construct struct pt_regs on stack */
149	pushq	$__USER_DS			/* pt_regs->ss */
150	pushq	PER_CPU_VAR(rsp_scratch)	/* pt_regs->sp */
151	/*
152	 * Re-enable interrupts.
153	 * We use 'rsp_scratch' as a scratch space, hence irq-off block above
154	 * must execute atomically in the face of possible interrupt-driven
155	 * task preemption. We must enable interrupts only after we're done
156	 * with using rsp_scratch:
157	 */
158	ENABLE_INTERRUPTS(CLBR_NONE)
159	pushq	%r11				/* pt_regs->flags */
160	pushq	$__USER_CS			/* pt_regs->cs */
161	pushq	%rcx				/* pt_regs->ip */
162	pushq	%rax				/* pt_regs->orig_ax */
163	pushq	%rdi				/* pt_regs->di */
164	pushq	%rsi				/* pt_regs->si */
165	pushq	%rdx				/* pt_regs->dx */
166	pushq	%rcx				/* pt_regs->cx */
167	pushq	$-ENOSYS			/* pt_regs->ax */
168	pushq	%r8				/* pt_regs->r8 */
169	pushq	%r9				/* pt_regs->r9 */
170	pushq	%r10				/* pt_regs->r10 */
171	pushq	%r11				/* pt_regs->r11 */
172	sub	$(6*8), %rsp			/* pt_regs->bp, bx, r12-15 not saved */
173
174	testl	$_TIF_WORK_SYSCALL_ENTRY, ASM_THREAD_INFO(TI_flags, %rsp, SIZEOF_PTREGS)
175	jnz	tracesys
176entry_SYSCALL_64_fastpath:
177#if __SYSCALL_MASK == ~0
178	cmpq	$__NR_syscall_max, %rax
179#else
180	andl	$__SYSCALL_MASK, %eax
181	cmpl	$__NR_syscall_max, %eax
182#endif
183	ja	1f				/* return -ENOSYS (already in pt_regs->ax) */
184	movq	%r10, %rcx
185	call	*sys_call_table(, %rax, 8)
186	movq	%rax, RAX(%rsp)
1871:
188/*
189 * Syscall return path ending with SYSRET (fast path).
190 * Has incompletely filled pt_regs.
191 */
192	LOCKDEP_SYS_EXIT
193	/*
194	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
195	 * it is too small to ever cause noticeable irq latency.
196	 */
197	DISABLE_INTERRUPTS(CLBR_NONE)
198
199	/*
200	 * We must check ti flags with interrupts (or at least preemption)
201	 * off because we must *never* return to userspace without
202	 * processing exit work that is enqueued if we're preempted here.
203	 * In particular, returning to userspace with any of the one-shot
204	 * flags (TIF_NOTIFY_RESUME, TIF_USER_RETURN_NOTIFY, etc) set is
205	 * very bad.
206	 */
207	testl	$_TIF_ALLWORK_MASK, ASM_THREAD_INFO(TI_flags, %rsp, SIZEOF_PTREGS)
208	jnz	int_ret_from_sys_call_irqs_off	/* Go to the slow path */
209
210	RESTORE_C_REGS_EXCEPT_RCX_R11
211	movq	RIP(%rsp), %rcx
212	movq	EFLAGS(%rsp), %r11
213	movq	RSP(%rsp), %rsp
214	/*
215	 * 64-bit SYSRET restores rip from rcx,
216	 * rflags from r11 (but RF and VM bits are forced to 0),
217	 * cs and ss are loaded from MSRs.
218	 * Restoration of rflags re-enables interrupts.
219	 *
220	 * NB: On AMD CPUs with the X86_BUG_SYSRET_SS_ATTRS bug, the ss
221	 * descriptor is not reinitialized.  This means that we should
222	 * avoid SYSRET with SS == NULL, which could happen if we schedule,
223	 * exit the kernel, and re-enter using an interrupt vector.  (All
224	 * interrupt entries on x86_64 set SS to NULL.)  We prevent that
225	 * from happening by reloading SS in __switch_to.  (Actually
226	 * detecting the failure in 64-bit userspace is tricky but can be
227	 * done.)
228	 */
229	USERGS_SYSRET64
230
231GLOBAL(int_ret_from_sys_call_irqs_off)
232	TRACE_IRQS_ON
233	ENABLE_INTERRUPTS(CLBR_NONE)
234	jmp int_ret_from_sys_call
235
236	/* Do syscall entry tracing */
237tracesys:
238	movq	%rsp, %rdi
239	movl	$AUDIT_ARCH_X86_64, %esi
240	call	syscall_trace_enter_phase1
241	test	%rax, %rax
242	jnz	tracesys_phase2			/* if needed, run the slow path */
243	RESTORE_C_REGS_EXCEPT_RAX		/* else restore clobbered regs */
244	movq	ORIG_RAX(%rsp), %rax
245	jmp	entry_SYSCALL_64_fastpath	/* and return to the fast path */
246
247tracesys_phase2:
248	SAVE_EXTRA_REGS
249	movq	%rsp, %rdi
250	movl	$AUDIT_ARCH_X86_64, %esi
251	movq	%rax, %rdx
252	call	syscall_trace_enter_phase2
253
254	/*
255	 * Reload registers from stack in case ptrace changed them.
256	 * We don't reload %rax because syscall_trace_entry_phase2() returned
257	 * the value it wants us to use in the table lookup.
258	 */
259	RESTORE_C_REGS_EXCEPT_RAX
260	RESTORE_EXTRA_REGS
261#if __SYSCALL_MASK == ~0
262	cmpq	$__NR_syscall_max, %rax
263#else
264	andl	$__SYSCALL_MASK, %eax
265	cmpl	$__NR_syscall_max, %eax
266#endif
267	ja	1f				/* return -ENOSYS (already in pt_regs->ax) */
268	movq	%r10, %rcx			/* fixup for C */
269	call	*sys_call_table(, %rax, 8)
270	movq	%rax, RAX(%rsp)
2711:
272	/* Use IRET because user could have changed pt_regs->foo */
273
274/*
275 * Syscall return path ending with IRET.
276 * Has correct iret frame.
277 */
278GLOBAL(int_ret_from_sys_call)
279	SAVE_EXTRA_REGS
280	movq	%rsp, %rdi
281	call	syscall_return_slowpath	/* returns with IRQs disabled */
282	RESTORE_EXTRA_REGS
283	TRACE_IRQS_IRETQ		/* we're about to change IF */
284
285	/*
286	 * Try to use SYSRET instead of IRET if we're returning to
287	 * a completely clean 64-bit userspace context.
288	 */
289	movq	RCX(%rsp), %rcx
290	movq	RIP(%rsp), %r11
291	cmpq	%rcx, %r11			/* RCX == RIP */
292	jne	opportunistic_sysret_failed
293
294	/*
295	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
296	 * in kernel space.  This essentially lets the user take over
297	 * the kernel, since userspace controls RSP.
298	 *
299	 * If width of "canonical tail" ever becomes variable, this will need
300	 * to be updated to remain correct on both old and new CPUs.
301	 */
302	.ifne __VIRTUAL_MASK_SHIFT - 47
303	.error "virtual address width changed -- SYSRET checks need update"
304	.endif
305
306	/* Change top 16 bits to be the sign-extension of 47th bit */
307	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
308	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
309
310	/* If this changed %rcx, it was not canonical */
311	cmpq	%rcx, %r11
312	jne	opportunistic_sysret_failed
313
314	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
315	jne	opportunistic_sysret_failed
316
317	movq	R11(%rsp), %r11
318	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
319	jne	opportunistic_sysret_failed
320
321	/*
322	 * SYSRET can't restore RF.  SYSRET can restore TF, but unlike IRET,
323	 * restoring TF results in a trap from userspace immediately after
324	 * SYSRET.  This would cause an infinite loop whenever #DB happens
325	 * with register state that satisfies the opportunistic SYSRET
326	 * conditions.  For example, single-stepping this user code:
327	 *
328	 *           movq	$stuck_here, %rcx
329	 *           pushfq
330	 *           popq %r11
331	 *   stuck_here:
332	 *
333	 * would never get past 'stuck_here'.
334	 */
335	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
336	jnz	opportunistic_sysret_failed
337
338	/* nothing to check for RSP */
339
340	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
341	jne	opportunistic_sysret_failed
342
343	/*
344	 * We win! This label is here just for ease of understanding
345	 * perf profiles. Nothing jumps here.
346	 */
347syscall_return_via_sysret:
348	/* rcx and r11 are already restored (see code above) */
349	RESTORE_C_REGS_EXCEPT_RCX_R11
350	movq	RSP(%rsp), %rsp
351	USERGS_SYSRET64
352
353opportunistic_sysret_failed:
354	SWAPGS
355	jmp	restore_c_regs_and_iret
356END(entry_SYSCALL_64)
357
358
359	.macro FORK_LIKE func
360ENTRY(stub_\func)
361	SAVE_EXTRA_REGS 8
362	jmp	sys_\func
363END(stub_\func)
364	.endm
365
366	FORK_LIKE  clone
367	FORK_LIKE  fork
368	FORK_LIKE  vfork
369
370ENTRY(stub_execve)
371	call	sys_execve
372return_from_execve:
373	testl	%eax, %eax
374	jz	1f
375	/* exec failed, can use fast SYSRET code path in this case */
376	ret
3771:
378	/* must use IRET code path (pt_regs->cs may have changed) */
379	addq	$8, %rsp
380	ZERO_EXTRA_REGS
381	movq	%rax, RAX(%rsp)
382	jmp	int_ret_from_sys_call
383END(stub_execve)
384/*
385 * Remaining execve stubs are only 7 bytes long.
386 * ENTRY() often aligns to 16 bytes, which in this case has no benefits.
387 */
388	.align	8
389GLOBAL(stub_execveat)
390	call	sys_execveat
391	jmp	return_from_execve
392END(stub_execveat)
393
394#if defined(CONFIG_X86_X32_ABI) || defined(CONFIG_IA32_EMULATION)
395	.align	8
396GLOBAL(stub_x32_execve)
397GLOBAL(stub32_execve)
398	call	compat_sys_execve
399	jmp	return_from_execve
400END(stub32_execve)
401END(stub_x32_execve)
402	.align	8
403GLOBAL(stub_x32_execveat)
404GLOBAL(stub32_execveat)
405	call	compat_sys_execveat
406	jmp	return_from_execve
407END(stub32_execveat)
408END(stub_x32_execveat)
409#endif
410
411/*
412 * sigreturn is special because it needs to restore all registers on return.
413 * This cannot be done with SYSRET, so use the IRET return path instead.
414 */
415ENTRY(stub_rt_sigreturn)
416	/*
417	 * SAVE_EXTRA_REGS result is not normally needed:
418	 * sigreturn overwrites all pt_regs->GPREGS.
419	 * But sigreturn can fail (!), and there is no easy way to detect that.
420	 * To make sure RESTORE_EXTRA_REGS doesn't restore garbage on error,
421	 * we SAVE_EXTRA_REGS here.
422	 */
423	SAVE_EXTRA_REGS 8
424	call	sys_rt_sigreturn
425return_from_stub:
426	addq	$8, %rsp
427	RESTORE_EXTRA_REGS
428	movq	%rax, RAX(%rsp)
429	jmp	int_ret_from_sys_call
430END(stub_rt_sigreturn)
431
432#ifdef CONFIG_X86_X32_ABI
433ENTRY(stub_x32_rt_sigreturn)
434	SAVE_EXTRA_REGS 8
435	call	sys32_x32_rt_sigreturn
436	jmp	return_from_stub
437END(stub_x32_rt_sigreturn)
438#endif
439
440/*
441 * A newly forked process directly context switches into this address.
442 *
443 * rdi: prev task we switched from
444 */
445ENTRY(ret_from_fork)
446
447	LOCK ; btr $TIF_FORK, TI_flags(%r8)
448
449	pushq	$0x0002
450	popfq					/* reset kernel eflags */
451
452	call	schedule_tail			/* rdi: 'prev' task parameter */
453
454	RESTORE_EXTRA_REGS
455
456	testb	$3, CS(%rsp)			/* from kernel_thread? */
457
458	/*
459	 * By the time we get here, we have no idea whether our pt_regs,
460	 * ti flags, and ti status came from the 64-bit SYSCALL fast path,
461	 * the slow path, or one of the 32-bit compat paths.
462	 * Use IRET code path to return, since it can safely handle
463	 * all of the above.
464	 */
465	jnz	int_ret_from_sys_call
466
467	/*
468	 * We came from kernel_thread
469	 * nb: we depend on RESTORE_EXTRA_REGS above
470	 */
471	movq	%rbp, %rdi
472	call	*%rbx
473	movl	$0, RAX(%rsp)
474	RESTORE_EXTRA_REGS
475	jmp	int_ret_from_sys_call
476END(ret_from_fork)
477
478/*
479 * Build the entry stubs with some assembler magic.
480 * We pack 1 stub into every 8-byte block.
481 */
482	.align 8
483ENTRY(irq_entries_start)
484    vector=FIRST_EXTERNAL_VECTOR
485    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
486	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
487    vector=vector+1
488	jmp	common_interrupt
489	.align	8
490    .endr
491END(irq_entries_start)
492
493/*
494 * Interrupt entry/exit.
495 *
496 * Interrupt entry points save only callee clobbered registers in fast path.
497 *
498 * Entry runs with interrupts off.
499 */
500
501/* 0(%rsp): ~(interrupt number) */
502	.macro interrupt func
503	cld
504	ALLOC_PT_GPREGS_ON_STACK
505	SAVE_C_REGS
506	SAVE_EXTRA_REGS
507
508	testb	$3, CS(%rsp)
509	jz	1f
510
511	/*
512	 * IRQ from user mode.  Switch to kernel gsbase and inform context
513	 * tracking that we're in kernel mode.
514	 */
515	SWAPGS
516#ifdef CONFIG_CONTEXT_TRACKING
517	call enter_from_user_mode
518#endif
519
5201:
521	/*
522	 * Save previous stack pointer, optionally switch to interrupt stack.
523	 * irq_count is used to check if a CPU is already on an interrupt stack
524	 * or not. While this is essentially redundant with preempt_count it is
525	 * a little cheaper to use a separate counter in the PDA (short of
526	 * moving irq_enter into assembly, which would be too much work)
527	 */
528	movq	%rsp, %rdi
529	incl	PER_CPU_VAR(irq_count)
530	cmovzq	PER_CPU_VAR(irq_stack_ptr), %rsp
531	pushq	%rdi
532	/* We entered an interrupt context - irqs are off: */
533	TRACE_IRQS_OFF
534
535	call	\func	/* rdi points to pt_regs */
536	.endm
537
538	/*
539	 * The interrupt stubs push (~vector+0x80) onto the stack and
540	 * then jump to common_interrupt.
541	 */
542	.p2align CONFIG_X86_L1_CACHE_SHIFT
543common_interrupt:
544	ASM_CLAC
545	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
546	interrupt do_IRQ
547	/* 0(%rsp): old RSP */
548ret_from_intr:
549	DISABLE_INTERRUPTS(CLBR_NONE)
550	TRACE_IRQS_OFF
551	decl	PER_CPU_VAR(irq_count)
552
553	/* Restore saved previous stack */
554	popq	%rsp
555
556	testb	$3, CS(%rsp)
557	jz	retint_kernel
558
559	/* Interrupt came from user space */
560	LOCKDEP_SYS_EXIT_IRQ
561GLOBAL(retint_user)
562	mov	%rsp,%rdi
563	call	prepare_exit_to_usermode
564	TRACE_IRQS_IRETQ
565	SWAPGS
566	jmp	restore_regs_and_iret
567
568/* Returning to kernel space */
569retint_kernel:
570#ifdef CONFIG_PREEMPT
571	/* Interrupts are off */
572	/* Check if we need preemption */
573	bt	$9, EFLAGS(%rsp)		/* were interrupts off? */
574	jnc	1f
5750:	cmpl	$0, PER_CPU_VAR(__preempt_count)
576	jnz	1f
577	call	preempt_schedule_irq
578	jmp	0b
5791:
580#endif
581	/*
582	 * The iretq could re-enable interrupts:
583	 */
584	TRACE_IRQS_IRETQ
585
586/*
587 * At this label, code paths which return to kernel and to user,
588 * which come from interrupts/exception and from syscalls, merge.
589 */
590restore_regs_and_iret:
591	RESTORE_EXTRA_REGS
592restore_c_regs_and_iret:
593	RESTORE_C_REGS
594	REMOVE_PT_GPREGS_FROM_STACK 8
595	INTERRUPT_RETURN
596
597ENTRY(native_iret)
598	/*
599	 * Are we returning to a stack segment from the LDT?  Note: in
600	 * 64-bit mode SS:RSP on the exception stack is always valid.
601	 */
602#ifdef CONFIG_X86_ESPFIX64
603	testb	$4, (SS-RIP)(%rsp)
604	jnz	native_irq_return_ldt
605#endif
606
607.global native_irq_return_iret
608native_irq_return_iret:
609	/*
610	 * This may fault.  Non-paranoid faults on return to userspace are
611	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
612	 * Double-faults due to espfix64 are handled in do_double_fault.
613	 * Other faults here are fatal.
614	 */
615	iretq
616
617#ifdef CONFIG_X86_ESPFIX64
618native_irq_return_ldt:
619	pushq	%rax
620	pushq	%rdi
621	SWAPGS
622	movq	PER_CPU_VAR(espfix_waddr), %rdi
623	movq	%rax, (0*8)(%rdi)		/* RAX */
624	movq	(2*8)(%rsp), %rax		/* RIP */
625	movq	%rax, (1*8)(%rdi)
626	movq	(3*8)(%rsp), %rax		/* CS */
627	movq	%rax, (2*8)(%rdi)
628	movq	(4*8)(%rsp), %rax		/* RFLAGS */
629	movq	%rax, (3*8)(%rdi)
630	movq	(6*8)(%rsp), %rax		/* SS */
631	movq	%rax, (5*8)(%rdi)
632	movq	(5*8)(%rsp), %rax		/* RSP */
633	movq	%rax, (4*8)(%rdi)
634	andl	$0xffff0000, %eax
635	popq	%rdi
636	orq	PER_CPU_VAR(espfix_stack), %rax
637	SWAPGS
638	movq	%rax, %rsp
639	popq	%rax
640	jmp	native_irq_return_iret
641#endif
642END(common_interrupt)
643
644/*
645 * APIC interrupts.
646 */
647.macro apicinterrupt3 num sym do_sym
648ENTRY(\sym)
649	ASM_CLAC
650	pushq	$~(\num)
651.Lcommon_\sym:
652	interrupt \do_sym
653	jmp	ret_from_intr
654END(\sym)
655.endm
656
657#ifdef CONFIG_TRACING
658#define trace(sym) trace_##sym
659#define smp_trace(sym) smp_trace_##sym
660
661.macro trace_apicinterrupt num sym
662apicinterrupt3 \num trace(\sym) smp_trace(\sym)
663.endm
664#else
665.macro trace_apicinterrupt num sym do_sym
666.endm
667#endif
668
669.macro apicinterrupt num sym do_sym
670apicinterrupt3 \num \sym \do_sym
671trace_apicinterrupt \num \sym
672.endm
673
674#ifdef CONFIG_SMP
675apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
676apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
677#endif
678
679#ifdef CONFIG_X86_UV
680apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
681#endif
682
683apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
684apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
685
686#ifdef CONFIG_HAVE_KVM
687apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
688apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
689#endif
690
691#ifdef CONFIG_X86_MCE_THRESHOLD
692apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
693#endif
694
695#ifdef CONFIG_X86_MCE_AMD
696apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
697#endif
698
699#ifdef CONFIG_X86_THERMAL_VECTOR
700apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
701#endif
702
703#ifdef CONFIG_SMP
704apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
705apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
706apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
707#endif
708
709apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
710apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
711
712#ifdef CONFIG_IRQ_WORK
713apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
714#endif
715
716/*
717 * Exception entry points.
718 */
719#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss) + (TSS_ist + ((x) - 1) * 8)
720
721.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
722ENTRY(\sym)
723	/* Sanity check */
724	.if \shift_ist != -1 && \paranoid == 0
725	.error "using shift_ist requires paranoid=1"
726	.endif
727
728	ASM_CLAC
729	PARAVIRT_ADJUST_EXCEPTION_FRAME
730
731	.ifeq \has_error_code
732	pushq	$-1				/* ORIG_RAX: no syscall to restart */
733	.endif
734
735	ALLOC_PT_GPREGS_ON_STACK
736
737	.if \paranoid
738	.if \paranoid == 1
739	testb	$3, CS(%rsp)			/* If coming from userspace, switch stacks */
740	jnz	1f
741	.endif
742	call	paranoid_entry
743	.else
744	call	error_entry
745	.endif
746	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
747
748	.if \paranoid
749	.if \shift_ist != -1
750	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
751	.else
752	TRACE_IRQS_OFF
753	.endif
754	.endif
755
756	movq	%rsp, %rdi			/* pt_regs pointer */
757
758	.if \has_error_code
759	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
760	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
761	.else
762	xorl	%esi, %esi			/* no error code */
763	.endif
764
765	.if \shift_ist != -1
766	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
767	.endif
768
769	call	\do_sym
770
771	.if \shift_ist != -1
772	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
773	.endif
774
775	/* these procedures expect "no swapgs" flag in ebx */
776	.if \paranoid
777	jmp	paranoid_exit
778	.else
779	jmp	error_exit
780	.endif
781
782	.if \paranoid == 1
783	/*
784	 * Paranoid entry from userspace.  Switch stacks and treat it
785	 * as a normal entry.  This means that paranoid handlers
786	 * run in real process context if user_mode(regs).
787	 */
7881:
789	call	error_entry
790
791
792	movq	%rsp, %rdi			/* pt_regs pointer */
793	call	sync_regs
794	movq	%rax, %rsp			/* switch stack */
795
796	movq	%rsp, %rdi			/* pt_regs pointer */
797
798	.if \has_error_code
799	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
800	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
801	.else
802	xorl	%esi, %esi			/* no error code */
803	.endif
804
805	call	\do_sym
806
807	jmp	error_exit			/* %ebx: no swapgs flag */
808	.endif
809END(\sym)
810.endm
811
812#ifdef CONFIG_TRACING
813.macro trace_idtentry sym do_sym has_error_code:req
814idtentry trace(\sym) trace(\do_sym) has_error_code=\has_error_code
815idtentry \sym \do_sym has_error_code=\has_error_code
816.endm
817#else
818.macro trace_idtentry sym do_sym has_error_code:req
819idtentry \sym \do_sym has_error_code=\has_error_code
820.endm
821#endif
822
823idtentry divide_error			do_divide_error			has_error_code=0
824idtentry overflow			do_overflow			has_error_code=0
825idtentry bounds				do_bounds			has_error_code=0
826idtentry invalid_op			do_invalid_op			has_error_code=0
827idtentry device_not_available		do_device_not_available		has_error_code=0
828idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
829idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
830idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
831idtentry segment_not_present		do_segment_not_present		has_error_code=1
832idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
833idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
834idtentry alignment_check		do_alignment_check		has_error_code=1
835idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
836
837
838	/*
839	 * Reload gs selector with exception handling
840	 * edi:  new selector
841	 */
842ENTRY(native_load_gs_index)
843	pushfq
844	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
845	SWAPGS
846gs_change:
847	movl	%edi, %gs
8482:	mfence					/* workaround */
849	SWAPGS
850	popfq
851	ret
852END(native_load_gs_index)
853
854	_ASM_EXTABLE(gs_change, bad_gs)
855	.section .fixup, "ax"
856	/* running with kernelgs */
857bad_gs:
858	SWAPGS					/* switch back to user gs */
859	xorl	%eax, %eax
860	movl	%eax, %gs
861	jmp	2b
862	.previous
863
864/* Call softirq on interrupt stack. Interrupts are off. */
865ENTRY(do_softirq_own_stack)
866	pushq	%rbp
867	mov	%rsp, %rbp
868	incl	PER_CPU_VAR(irq_count)
869	cmove	PER_CPU_VAR(irq_stack_ptr), %rsp
870	push	%rbp				/* frame pointer backlink */
871	call	__do_softirq
872	leaveq
873	decl	PER_CPU_VAR(irq_count)
874	ret
875END(do_softirq_own_stack)
876
877#ifdef CONFIG_XEN
878idtentry xen_hypervisor_callback xen_do_hypervisor_callback has_error_code=0
879
880/*
881 * A note on the "critical region" in our callback handler.
882 * We want to avoid stacking callback handlers due to events occurring
883 * during handling of the last event. To do this, we keep events disabled
884 * until we've done all processing. HOWEVER, we must enable events before
885 * popping the stack frame (can't be done atomically) and so it would still
886 * be possible to get enough handler activations to overflow the stack.
887 * Although unlikely, bugs of that kind are hard to track down, so we'd
888 * like to avoid the possibility.
889 * So, on entry to the handler we detect whether we interrupted an
890 * existing activation in its critical region -- if so, we pop the current
891 * activation and restart the handler using the previous one.
892 */
893ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
894
895/*
896 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
897 * see the correct pointer to the pt_regs
898 */
899	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
90011:	incl	PER_CPU_VAR(irq_count)
901	movq	%rsp, %rbp
902	cmovzq	PER_CPU_VAR(irq_stack_ptr), %rsp
903	pushq	%rbp				/* frame pointer backlink */
904	call	xen_evtchn_do_upcall
905	popq	%rsp
906	decl	PER_CPU_VAR(irq_count)
907#ifndef CONFIG_PREEMPT
908	call	xen_maybe_preempt_hcall
909#endif
910	jmp	error_exit
911END(xen_do_hypervisor_callback)
912
913/*
914 * Hypervisor uses this for application faults while it executes.
915 * We get here for two reasons:
916 *  1. Fault while reloading DS, ES, FS or GS
917 *  2. Fault while executing IRET
918 * Category 1 we do not need to fix up as Xen has already reloaded all segment
919 * registers that could be reloaded and zeroed the others.
920 * Category 2 we fix up by killing the current process. We cannot use the
921 * normal Linux return path in this case because if we use the IRET hypercall
922 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
923 * We distinguish between categories by comparing each saved segment register
924 * with its current contents: any discrepancy means we in category 1.
925 */
926ENTRY(xen_failsafe_callback)
927	movl	%ds, %ecx
928	cmpw	%cx, 0x10(%rsp)
929	jne	1f
930	movl	%es, %ecx
931	cmpw	%cx, 0x18(%rsp)
932	jne	1f
933	movl	%fs, %ecx
934	cmpw	%cx, 0x20(%rsp)
935	jne	1f
936	movl	%gs, %ecx
937	cmpw	%cx, 0x28(%rsp)
938	jne	1f
939	/* All segments match their saved values => Category 2 (Bad IRET). */
940	movq	(%rsp), %rcx
941	movq	8(%rsp), %r11
942	addq	$0x30, %rsp
943	pushq	$0				/* RIP */
944	pushq	%r11
945	pushq	%rcx
946	jmp	general_protection
9471:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
948	movq	(%rsp), %rcx
949	movq	8(%rsp), %r11
950	addq	$0x30, %rsp
951	pushq	$-1 /* orig_ax = -1 => not a system call */
952	ALLOC_PT_GPREGS_ON_STACK
953	SAVE_C_REGS
954	SAVE_EXTRA_REGS
955	jmp	error_exit
956END(xen_failsafe_callback)
957
958apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
959	xen_hvm_callback_vector xen_evtchn_do_upcall
960
961#endif /* CONFIG_XEN */
962
963#if IS_ENABLED(CONFIG_HYPERV)
964apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
965	hyperv_callback_vector hyperv_vector_handler
966#endif /* CONFIG_HYPERV */
967
968idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
969idtentry int3			do_int3			has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
970idtentry stack_segment		do_stack_segment	has_error_code=1
971
972#ifdef CONFIG_XEN
973idtentry xen_debug		do_debug		has_error_code=0
974idtentry xen_int3		do_int3			has_error_code=0
975idtentry xen_stack_segment	do_stack_segment	has_error_code=1
976#endif
977
978idtentry general_protection	do_general_protection	has_error_code=1
979trace_idtentry page_fault	do_page_fault		has_error_code=1
980
981#ifdef CONFIG_KVM_GUEST
982idtentry async_page_fault	do_async_page_fault	has_error_code=1
983#endif
984
985#ifdef CONFIG_X86_MCE
986idtentry machine_check					has_error_code=0	paranoid=1 do_sym=*machine_check_vector(%rip)
987#endif
988
989/*
990 * Save all registers in pt_regs, and switch gs if needed.
991 * Use slow, but surefire "are we in kernel?" check.
992 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
993 */
994ENTRY(paranoid_entry)
995	cld
996	SAVE_C_REGS 8
997	SAVE_EXTRA_REGS 8
998	movl	$1, %ebx
999	movl	$MSR_GS_BASE, %ecx
1000	rdmsr
1001	testl	%edx, %edx
1002	js	1f				/* negative -> in kernel */
1003	SWAPGS
1004	xorl	%ebx, %ebx
10051:	ret
1006END(paranoid_entry)
1007
1008/*
1009 * "Paranoid" exit path from exception stack.  This is invoked
1010 * only on return from non-NMI IST interrupts that came
1011 * from kernel space.
1012 *
1013 * We may be returning to very strange contexts (e.g. very early
1014 * in syscall entry), so checking for preemption here would
1015 * be complicated.  Fortunately, we there's no good reason
1016 * to try to handle preemption here.
1017 *
1018 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1019 */
1020ENTRY(paranoid_exit)
1021	DISABLE_INTERRUPTS(CLBR_NONE)
1022	TRACE_IRQS_OFF_DEBUG
1023	testl	%ebx, %ebx			/* swapgs needed? */
1024	jnz	paranoid_exit_no_swapgs
1025	TRACE_IRQS_IRETQ
1026	SWAPGS_UNSAFE_STACK
1027	jmp	paranoid_exit_restore
1028paranoid_exit_no_swapgs:
1029	TRACE_IRQS_IRETQ_DEBUG
1030paranoid_exit_restore:
1031	RESTORE_EXTRA_REGS
1032	RESTORE_C_REGS
1033	REMOVE_PT_GPREGS_FROM_STACK 8
1034	INTERRUPT_RETURN
1035END(paranoid_exit)
1036
1037/*
1038 * Save all registers in pt_regs, and switch gs if needed.
1039 * Return: EBX=0: came from user mode; EBX=1: otherwise
1040 */
1041ENTRY(error_entry)
1042	cld
1043	SAVE_C_REGS 8
1044	SAVE_EXTRA_REGS 8
1045	xorl	%ebx, %ebx
1046	testb	$3, CS+8(%rsp)
1047	jz	.Lerror_kernelspace
1048
1049.Lerror_entry_from_usermode_swapgs:
1050	/*
1051	 * We entered from user mode or we're pretending to have entered
1052	 * from user mode due to an IRET fault.
1053	 */
1054	SWAPGS
1055
1056.Lerror_entry_from_usermode_after_swapgs:
1057#ifdef CONFIG_CONTEXT_TRACKING
1058	call enter_from_user_mode
1059#endif
1060
1061.Lerror_entry_done:
1062
1063	TRACE_IRQS_OFF
1064	ret
1065
1066	/*
1067	 * There are two places in the kernel that can potentially fault with
1068	 * usergs. Handle them here.  B stepping K8s sometimes report a
1069	 * truncated RIP for IRET exceptions returning to compat mode. Check
1070	 * for these here too.
1071	 */
1072.Lerror_kernelspace:
1073	incl	%ebx
1074	leaq	native_irq_return_iret(%rip), %rcx
1075	cmpq	%rcx, RIP+8(%rsp)
1076	je	.Lerror_bad_iret
1077	movl	%ecx, %eax			/* zero extend */
1078	cmpq	%rax, RIP+8(%rsp)
1079	je	.Lbstep_iret
1080	cmpq	$gs_change, RIP+8(%rsp)
1081	jne	.Lerror_entry_done
1082
1083	/*
1084	 * hack: gs_change can fail with user gsbase.  If this happens, fix up
1085	 * gsbase and proceed.  We'll fix up the exception and land in
1086	 * gs_change's error handler with kernel gsbase.
1087	 */
1088	jmp	.Lerror_entry_from_usermode_swapgs
1089
1090.Lbstep_iret:
1091	/* Fix truncated RIP */
1092	movq	%rcx, RIP+8(%rsp)
1093	/* fall through */
1094
1095.Lerror_bad_iret:
1096	/*
1097	 * We came from an IRET to user mode, so we have user gsbase.
1098	 * Switch to kernel gsbase:
1099	 */
1100	SWAPGS
1101
1102	/*
1103	 * Pretend that the exception came from user mode: set up pt_regs
1104	 * as if we faulted immediately after IRET and clear EBX so that
1105	 * error_exit knows that we will be returning to user mode.
1106	 */
1107	mov	%rsp, %rdi
1108	call	fixup_bad_iret
1109	mov	%rax, %rsp
1110	decl	%ebx
1111	jmp	.Lerror_entry_from_usermode_after_swapgs
1112END(error_entry)
1113
1114
1115/*
1116 * On entry, EBS is a "return to kernel mode" flag:
1117 *   1: already in kernel mode, don't need SWAPGS
1118 *   0: user gsbase is loaded, we need SWAPGS and standard preparation for return to usermode
1119 */
1120ENTRY(error_exit)
1121	movl	%ebx, %eax
1122	DISABLE_INTERRUPTS(CLBR_NONE)
1123	TRACE_IRQS_OFF
1124	testl	%eax, %eax
1125	jnz	retint_kernel
1126	jmp	retint_user
1127END(error_exit)
1128
1129/* Runs on exception stack */
1130ENTRY(nmi)
1131	/*
1132	 * Fix up the exception frame if we're on Xen.
1133	 * PARAVIRT_ADJUST_EXCEPTION_FRAME is guaranteed to push at most
1134	 * one value to the stack on native, so it may clobber the rdx
1135	 * scratch slot, but it won't clobber any of the important
1136	 * slots past it.
1137	 *
1138	 * Xen is a different story, because the Xen frame itself overlaps
1139	 * the "NMI executing" variable.
1140	 */
1141	PARAVIRT_ADJUST_EXCEPTION_FRAME
1142
1143	/*
1144	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1145	 * the iretq it performs will take us out of NMI context.
1146	 * This means that we can have nested NMIs where the next
1147	 * NMI is using the top of the stack of the previous NMI. We
1148	 * can't let it execute because the nested NMI will corrupt the
1149	 * stack of the previous NMI. NMI handlers are not re-entrant
1150	 * anyway.
1151	 *
1152	 * To handle this case we do the following:
1153	 *  Check the a special location on the stack that contains
1154	 *  a variable that is set when NMIs are executing.
1155	 *  The interrupted task's stack is also checked to see if it
1156	 *  is an NMI stack.
1157	 *  If the variable is not set and the stack is not the NMI
1158	 *  stack then:
1159	 *    o Set the special variable on the stack
1160	 *    o Copy the interrupt frame into an "outermost" location on the
1161	 *      stack
1162	 *    o Copy the interrupt frame into an "iret" location on the stack
1163	 *    o Continue processing the NMI
1164	 *  If the variable is set or the previous stack is the NMI stack:
1165	 *    o Modify the "iret" location to jump to the repeat_nmi
1166	 *    o return back to the first NMI
1167	 *
1168	 * Now on exit of the first NMI, we first clear the stack variable
1169	 * The NMI stack will tell any nested NMIs at that point that it is
1170	 * nested. Then we pop the stack normally with iret, and if there was
1171	 * a nested NMI that updated the copy interrupt stack frame, a
1172	 * jump will be made to the repeat_nmi code that will handle the second
1173	 * NMI.
1174	 *
1175	 * However, espfix prevents us from directly returning to userspace
1176	 * with a single IRET instruction.  Similarly, IRET to user mode
1177	 * can fault.  We therefore handle NMIs from user space like
1178	 * other IST entries.
1179	 */
1180
1181	/* Use %rdx as our temp variable throughout */
1182	pushq	%rdx
1183
1184	testb	$3, CS-RIP+8(%rsp)
1185	jz	.Lnmi_from_kernel
1186
1187	/*
1188	 * NMI from user mode.  We need to run on the thread stack, but we
1189	 * can't go through the normal entry paths: NMIs are masked, and
1190	 * we don't want to enable interrupts, because then we'll end
1191	 * up in an awkward situation in which IRQs are on but NMIs
1192	 * are off.
1193	 *
1194	 * We also must not push anything to the stack before switching
1195	 * stacks lest we corrupt the "NMI executing" variable.
1196	 */
1197
1198	SWAPGS_UNSAFE_STACK
1199	cld
1200	movq	%rsp, %rdx
1201	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1202	pushq	5*8(%rdx)	/* pt_regs->ss */
1203	pushq	4*8(%rdx)	/* pt_regs->rsp */
1204	pushq	3*8(%rdx)	/* pt_regs->flags */
1205	pushq	2*8(%rdx)	/* pt_regs->cs */
1206	pushq	1*8(%rdx)	/* pt_regs->rip */
1207	pushq   $-1		/* pt_regs->orig_ax */
1208	pushq   %rdi		/* pt_regs->di */
1209	pushq   %rsi		/* pt_regs->si */
1210	pushq   (%rdx)		/* pt_regs->dx */
1211	pushq   %rcx		/* pt_regs->cx */
1212	pushq   %rax		/* pt_regs->ax */
1213	pushq   %r8		/* pt_regs->r8 */
1214	pushq   %r9		/* pt_regs->r9 */
1215	pushq   %r10		/* pt_regs->r10 */
1216	pushq   %r11		/* pt_regs->r11 */
1217	pushq	%rbx		/* pt_regs->rbx */
1218	pushq	%rbp		/* pt_regs->rbp */
1219	pushq	%r12		/* pt_regs->r12 */
1220	pushq	%r13		/* pt_regs->r13 */
1221	pushq	%r14		/* pt_regs->r14 */
1222	pushq	%r15		/* pt_regs->r15 */
1223
1224	/*
1225	 * At this point we no longer need to worry about stack damage
1226	 * due to nesting -- we're on the normal thread stack and we're
1227	 * done with the NMI stack.
1228	 */
1229
1230	movq	%rsp, %rdi
1231	movq	$-1, %rsi
1232	call	do_nmi
1233
1234	/*
1235	 * Return back to user mode.  We must *not* do the normal exit
1236	 * work, because we don't want to enable interrupts.  Fortunately,
1237	 * do_nmi doesn't modify pt_regs.
1238	 */
1239	SWAPGS
1240	jmp	restore_c_regs_and_iret
1241
1242.Lnmi_from_kernel:
1243	/*
1244	 * Here's what our stack frame will look like:
1245	 * +---------------------------------------------------------+
1246	 * | original SS                                             |
1247	 * | original Return RSP                                     |
1248	 * | original RFLAGS                                         |
1249	 * | original CS                                             |
1250	 * | original RIP                                            |
1251	 * +---------------------------------------------------------+
1252	 * | temp storage for rdx                                    |
1253	 * +---------------------------------------------------------+
1254	 * | "NMI executing" variable                                |
1255	 * +---------------------------------------------------------+
1256	 * | iret SS          } Copied from "outermost" frame        |
1257	 * | iret Return RSP  } on each loop iteration; overwritten  |
1258	 * | iret RFLAGS      } by a nested NMI to force another     |
1259	 * | iret CS          } iteration if needed.                 |
1260	 * | iret RIP         }                                      |
1261	 * +---------------------------------------------------------+
1262	 * | outermost SS          } initialized in first_nmi;       |
1263	 * | outermost Return RSP  } will not be changed before      |
1264	 * | outermost RFLAGS      } NMI processing is done.         |
1265	 * | outermost CS          } Copied to "iret" frame on each  |
1266	 * | outermost RIP         } iteration.                      |
1267	 * +---------------------------------------------------------+
1268	 * | pt_regs                                                 |
1269	 * +---------------------------------------------------------+
1270	 *
1271	 * The "original" frame is used by hardware.  Before re-enabling
1272	 * NMIs, we need to be done with it, and we need to leave enough
1273	 * space for the asm code here.
1274	 *
1275	 * We return by executing IRET while RSP points to the "iret" frame.
1276	 * That will either return for real or it will loop back into NMI
1277	 * processing.
1278	 *
1279	 * The "outermost" frame is copied to the "iret" frame on each
1280	 * iteration of the loop, so each iteration starts with the "iret"
1281	 * frame pointing to the final return target.
1282	 */
1283
1284	/*
1285	 * Determine whether we're a nested NMI.
1286	 *
1287	 * If we interrupted kernel code between repeat_nmi and
1288	 * end_repeat_nmi, then we are a nested NMI.  We must not
1289	 * modify the "iret" frame because it's being written by
1290	 * the outer NMI.  That's okay; the outer NMI handler is
1291	 * about to about to call do_nmi anyway, so we can just
1292	 * resume the outer NMI.
1293	 */
1294
1295	movq	$repeat_nmi, %rdx
1296	cmpq	8(%rsp), %rdx
1297	ja	1f
1298	movq	$end_repeat_nmi, %rdx
1299	cmpq	8(%rsp), %rdx
1300	ja	nested_nmi_out
13011:
1302
1303	/*
1304	 * Now check "NMI executing".  If it's set, then we're nested.
1305	 * This will not detect if we interrupted an outer NMI just
1306	 * before IRET.
1307	 */
1308	cmpl	$1, -8(%rsp)
1309	je	nested_nmi
1310
1311	/*
1312	 * Now test if the previous stack was an NMI stack.  This covers
1313	 * the case where we interrupt an outer NMI after it clears
1314	 * "NMI executing" but before IRET.  We need to be careful, though:
1315	 * there is one case in which RSP could point to the NMI stack
1316	 * despite there being no NMI active: naughty userspace controls
1317	 * RSP at the very beginning of the SYSCALL targets.  We can
1318	 * pull a fast one on naughty userspace, though: we program
1319	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1320	 * if it controls the kernel's RSP.  We set DF before we clear
1321	 * "NMI executing".
1322	 */
1323	lea	6*8(%rsp), %rdx
1324	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1325	cmpq	%rdx, 4*8(%rsp)
1326	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1327	ja	first_nmi
1328
1329	subq	$EXCEPTION_STKSZ, %rdx
1330	cmpq	%rdx, 4*8(%rsp)
1331	/* If it is below the NMI stack, it is a normal NMI */
1332	jb	first_nmi
1333
1334	/* Ah, it is within the NMI stack. */
1335
1336	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1337	jz	first_nmi	/* RSP was user controlled. */
1338
1339	/* This is a nested NMI. */
1340
1341nested_nmi:
1342	/*
1343	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1344	 * iteration of NMI handling.
1345	 */
1346	subq	$8, %rsp
1347	leaq	-10*8(%rsp), %rdx
1348	pushq	$__KERNEL_DS
1349	pushq	%rdx
1350	pushfq
1351	pushq	$__KERNEL_CS
1352	pushq	$repeat_nmi
1353
1354	/* Put stack back */
1355	addq	$(6*8), %rsp
1356
1357nested_nmi_out:
1358	popq	%rdx
1359
1360	/* We are returning to kernel mode, so this cannot result in a fault. */
1361	INTERRUPT_RETURN
1362
1363first_nmi:
1364	/* Restore rdx. */
1365	movq	(%rsp), %rdx
1366
1367	/* Make room for "NMI executing". */
1368	pushq	$0
1369
1370	/* Leave room for the "iret" frame */
1371	subq	$(5*8), %rsp
1372
1373	/* Copy the "original" frame to the "outermost" frame */
1374	.rept 5
1375	pushq	11*8(%rsp)
1376	.endr
1377
1378	/* Everything up to here is safe from nested NMIs */
1379
1380#ifdef CONFIG_DEBUG_ENTRY
1381	/*
1382	 * For ease of testing, unmask NMIs right away.  Disabled by
1383	 * default because IRET is very expensive.
1384	 */
1385	pushq	$0		/* SS */
1386	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1387	addq	$8, (%rsp)	/* Fix up RSP */
1388	pushfq			/* RFLAGS */
1389	pushq	$__KERNEL_CS	/* CS */
1390	pushq	$1f		/* RIP */
1391	INTERRUPT_RETURN	/* continues at repeat_nmi below */
13921:
1393#endif
1394
1395repeat_nmi:
1396	/*
1397	 * If there was a nested NMI, the first NMI's iret will return
1398	 * here. But NMIs are still enabled and we can take another
1399	 * nested NMI. The nested NMI checks the interrupted RIP to see
1400	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1401	 * it will just return, as we are about to repeat an NMI anyway.
1402	 * This makes it safe to copy to the stack frame that a nested
1403	 * NMI will update.
1404	 *
1405	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1406	 * we're repeating an NMI, gsbase has the same value that it had on
1407	 * the first iteration.  paranoid_entry will load the kernel
1408	 * gsbase if needed before we call do_nmi.  "NMI executing"
1409	 * is zero.
1410	 */
1411	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1412
1413	/*
1414	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1415	 * here must not modify the "iret" frame while we're writing to
1416	 * it or it will end up containing garbage.
1417	 */
1418	addq	$(10*8), %rsp
1419	.rept 5
1420	pushq	-6*8(%rsp)
1421	.endr
1422	subq	$(5*8), %rsp
1423end_repeat_nmi:
1424
1425	/*
1426	 * Everything below this point can be preempted by a nested NMI.
1427	 * If this happens, then the inner NMI will change the "iret"
1428	 * frame to point back to repeat_nmi.
1429	 */
1430	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1431	ALLOC_PT_GPREGS_ON_STACK
1432
1433	/*
1434	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1435	 * as we should not be calling schedule in NMI context.
1436	 * Even with normal interrupts enabled. An NMI should not be
1437	 * setting NEED_RESCHED or anything that normal interrupts and
1438	 * exceptions might do.
1439	 */
1440	call	paranoid_entry
1441
1442	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1443	movq	%rsp, %rdi
1444	movq	$-1, %rsi
1445	call	do_nmi
1446
1447	testl	%ebx, %ebx			/* swapgs needed? */
1448	jnz	nmi_restore
1449nmi_swapgs:
1450	SWAPGS_UNSAFE_STACK
1451nmi_restore:
1452	RESTORE_EXTRA_REGS
1453	RESTORE_C_REGS
1454
1455	/* Point RSP at the "iret" frame. */
1456	REMOVE_PT_GPREGS_FROM_STACK 6*8
1457
1458	/*
1459	 * Clear "NMI executing".  Set DF first so that we can easily
1460	 * distinguish the remaining code between here and IRET from
1461	 * the SYSCALL entry and exit paths.  On a native kernel, we
1462	 * could just inspect RIP, but, on paravirt kernels,
1463	 * INTERRUPT_RETURN can translate into a jump into a
1464	 * hypercall page.
1465	 */
1466	std
1467	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1468
1469	/*
1470	 * INTERRUPT_RETURN reads the "iret" frame and exits the NMI
1471	 * stack in a single instruction.  We are returning to kernel
1472	 * mode, so this cannot result in a fault.
1473	 */
1474	INTERRUPT_RETURN
1475END(nmi)
1476
1477ENTRY(ignore_sysret)
1478	mov	$-ENOSYS, %eax
1479	sysret
1480END(ignore_sysret)
1481