xref: /openbmc/linux/arch/x86/entry/entry_64.S (revision 0bea2a65)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 *  linux/arch/x86_64/entry.S
4 *
5 *  Copyright (C) 1991, 1992  Linus Torvalds
6 *  Copyright (C) 2000, 2001, 2002  Andi Kleen SuSE Labs
7 *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
8 *
9 * entry.S contains the system-call and fault low-level handling routines.
10 *
11 * Some of this is documented in Documentation/x86/entry_64.txt
12 *
13 * A note on terminology:
14 * - iret frame:	Architecture defined interrupt frame from SS to RIP
15 *			at the top of the kernel process stack.
16 *
17 * Some macro usage:
18 * - ENTRY/END:		Define functions in the symbol table.
19 * - TRACE_IRQ_*:	Trace hardirq state for lock debugging.
20 * - idtentry:		Define exception entry points.
21 */
22#include <linux/linkage.h>
23#include <asm/segment.h>
24#include <asm/cache.h>
25#include <asm/errno.h>
26#include "calling.h"
27#include <asm/asm-offsets.h>
28#include <asm/msr.h>
29#include <asm/unistd.h>
30#include <asm/thread_info.h>
31#include <asm/hw_irq.h>
32#include <asm/page_types.h>
33#include <asm/irqflags.h>
34#include <asm/paravirt.h>
35#include <asm/percpu.h>
36#include <asm/asm.h>
37#include <asm/smap.h>
38#include <asm/pgtable_types.h>
39#include <asm/export.h>
40#include <asm/frame.h>
41#include <linux/err.h>
42
43.code64
44.section .entry.text, "ax"
45
46#ifdef CONFIG_PARAVIRT
47ENTRY(native_usergs_sysret64)
48	UNWIND_HINT_EMPTY
49	swapgs
50	sysretq
51END(native_usergs_sysret64)
52#endif /* CONFIG_PARAVIRT */
53
54.macro TRACE_IRQS_FLAGS flags:req
55#ifdef CONFIG_TRACE_IRQFLAGS
56	bt	$9, \flags		/* interrupts off? */
57	jnc	1f
58	TRACE_IRQS_ON
591:
60#endif
61.endm
62
63.macro TRACE_IRQS_IRETQ
64	TRACE_IRQS_FLAGS EFLAGS(%rsp)
65.endm
66
67/*
68 * When dynamic function tracer is enabled it will add a breakpoint
69 * to all locations that it is about to modify, sync CPUs, update
70 * all the code, sync CPUs, then remove the breakpoints. In this time
71 * if lockdep is enabled, it might jump back into the debug handler
72 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF).
73 *
74 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to
75 * make sure the stack pointer does not get reset back to the top
76 * of the debug stack, and instead just reuses the current stack.
77 */
78#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS)
79
80.macro TRACE_IRQS_OFF_DEBUG
81	call	debug_stack_set_zero
82	TRACE_IRQS_OFF
83	call	debug_stack_reset
84.endm
85
86.macro TRACE_IRQS_ON_DEBUG
87	call	debug_stack_set_zero
88	TRACE_IRQS_ON
89	call	debug_stack_reset
90.endm
91
92.macro TRACE_IRQS_IRETQ_DEBUG
93	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
94	jnc	1f
95	TRACE_IRQS_ON_DEBUG
961:
97.endm
98
99#else
100# define TRACE_IRQS_OFF_DEBUG			TRACE_IRQS_OFF
101# define TRACE_IRQS_ON_DEBUG			TRACE_IRQS_ON
102# define TRACE_IRQS_IRETQ_DEBUG			TRACE_IRQS_IRETQ
103#endif
104
105/*
106 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
107 *
108 * This is the only entry point used for 64-bit system calls.  The
109 * hardware interface is reasonably well designed and the register to
110 * argument mapping Linux uses fits well with the registers that are
111 * available when SYSCALL is used.
112 *
113 * SYSCALL instructions can be found inlined in libc implementations as
114 * well as some other programs and libraries.  There are also a handful
115 * of SYSCALL instructions in the vDSO used, for example, as a
116 * clock_gettimeofday fallback.
117 *
118 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
119 * then loads new ss, cs, and rip from previously programmed MSRs.
120 * rflags gets masked by a value from another MSR (so CLD and CLAC
121 * are not needed). SYSCALL does not save anything on the stack
122 * and does not change rsp.
123 *
124 * Registers on entry:
125 * rax  system call number
126 * rcx  return address
127 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
128 * rdi  arg0
129 * rsi  arg1
130 * rdx  arg2
131 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
132 * r8   arg4
133 * r9   arg5
134 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
135 *
136 * Only called from user space.
137 *
138 * When user can change pt_regs->foo always force IRET. That is because
139 * it deals with uncanonical addresses better. SYSRET has trouble
140 * with them due to bugs in both AMD and Intel CPUs.
141 */
142
143ENTRY(entry_SYSCALL_64)
144	UNWIND_HINT_EMPTY
145	/*
146	 * Interrupts are off on entry.
147	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
148	 * it is too small to ever cause noticeable irq latency.
149	 */
150
151	swapgs
152	movq	%rsp, PER_CPU_VAR(rsp_scratch)
153	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
154
155	/* Construct struct pt_regs on stack */
156	pushq	$__USER_DS			/* pt_regs->ss */
157	pushq	PER_CPU_VAR(rsp_scratch)	/* pt_regs->sp */
158	pushq	%r11				/* pt_regs->flags */
159	pushq	$__USER_CS			/* pt_regs->cs */
160	pushq	%rcx				/* pt_regs->ip */
161GLOBAL(entry_SYSCALL_64_after_hwframe)
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	UNWIND_HINT_REGS extra=0
174
175	TRACE_IRQS_OFF
176
177	/*
178	 * If we need to do entry work or if we guess we'll need to do
179	 * exit work, go straight to the slow path.
180	 */
181	movq	PER_CPU_VAR(current_task), %r11
182	testl	$_TIF_WORK_SYSCALL_ENTRY|_TIF_ALLWORK_MASK, TASK_TI_flags(%r11)
183	jnz	entry_SYSCALL64_slow_path
184
185entry_SYSCALL_64_fastpath:
186	/*
187	 * Easy case: enable interrupts and issue the syscall.  If the syscall
188	 * needs pt_regs, we'll call a stub that disables interrupts again
189	 * and jumps to the slow path.
190	 */
191	TRACE_IRQS_ON
192	ENABLE_INTERRUPTS(CLBR_NONE)
193#if __SYSCALL_MASK == ~0
194	cmpq	$__NR_syscall_max, %rax
195#else
196	andl	$__SYSCALL_MASK, %eax
197	cmpl	$__NR_syscall_max, %eax
198#endif
199	ja	1f				/* return -ENOSYS (already in pt_regs->ax) */
200	movq	%r10, %rcx
201
202	/*
203	 * This call instruction is handled specially in stub_ptregs_64.
204	 * It might end up jumping to the slow path.  If it jumps, RAX
205	 * and all argument registers are clobbered.
206	 */
207	call	*sys_call_table(, %rax, 8)
208.Lentry_SYSCALL_64_after_fastpath_call:
209
210	movq	%rax, RAX(%rsp)
2111:
212
213	/*
214	 * If we get here, then we know that pt_regs is clean for SYSRET64.
215	 * If we see that no exit work is required (which we are required
216	 * to check with IRQs off), then we can go straight to SYSRET64.
217	 */
218	DISABLE_INTERRUPTS(CLBR_ANY)
219	TRACE_IRQS_OFF
220	movq	PER_CPU_VAR(current_task), %r11
221	testl	$_TIF_ALLWORK_MASK, TASK_TI_flags(%r11)
222	jnz	1f
223
224	LOCKDEP_SYS_EXIT
225	TRACE_IRQS_ON		/* user mode is traced as IRQs on */
226	movq	RIP(%rsp), %rcx
227	movq	EFLAGS(%rsp), %r11
228	addq	$6*8, %rsp	/* skip extra regs -- they were preserved */
229	UNWIND_HINT_EMPTY
230	jmp	.Lpop_c_regs_except_rcx_r11_and_sysret
231
2321:
233	/*
234	 * The fast path looked good when we started, but something changed
235	 * along the way and we need to switch to the slow path.  Calling
236	 * raise(3) will trigger this, for example.  IRQs are off.
237	 */
238	TRACE_IRQS_ON
239	ENABLE_INTERRUPTS(CLBR_ANY)
240	SAVE_EXTRA_REGS
241	movq	%rsp, %rdi
242	call	syscall_return_slowpath	/* returns with IRQs disabled */
243	jmp	return_from_SYSCALL_64
244
245entry_SYSCALL64_slow_path:
246	/* IRQs are off. */
247	SAVE_EXTRA_REGS
248	movq	%rsp, %rdi
249	call	do_syscall_64		/* returns with IRQs disabled */
250
251return_from_SYSCALL_64:
252	TRACE_IRQS_IRETQ		/* we're about to change IF */
253
254	/*
255	 * Try to use SYSRET instead of IRET if we're returning to
256	 * a completely clean 64-bit userspace context.  If we're not,
257	 * go to the slow exit path.
258	 */
259	movq	RCX(%rsp), %rcx
260	movq	RIP(%rsp), %r11
261
262	cmpq	%rcx, %r11	/* SYSRET requires RCX == RIP */
263	jne	swapgs_restore_regs_and_return_to_usermode
264
265	/*
266	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
267	 * in kernel space.  This essentially lets the user take over
268	 * the kernel, since userspace controls RSP.
269	 *
270	 * If width of "canonical tail" ever becomes variable, this will need
271	 * to be updated to remain correct on both old and new CPUs.
272	 *
273	 * Change top bits to match most significant bit (47th or 56th bit
274	 * depending on paging mode) in the address.
275	 */
276	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
277	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
278
279	/* If this changed %rcx, it was not canonical */
280	cmpq	%rcx, %r11
281	jne	swapgs_restore_regs_and_return_to_usermode
282
283	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
284	jne	swapgs_restore_regs_and_return_to_usermode
285
286	movq	R11(%rsp), %r11
287	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
288	jne	swapgs_restore_regs_and_return_to_usermode
289
290	/*
291	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
292	 * restore RF properly. If the slowpath sets it for whatever reason, we
293	 * need to restore it correctly.
294	 *
295	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
296	 * trap from userspace immediately after SYSRET.  This would cause an
297	 * infinite loop whenever #DB happens with register state that satisfies
298	 * the opportunistic SYSRET conditions.  For example, single-stepping
299	 * this user code:
300	 *
301	 *           movq	$stuck_here, %rcx
302	 *           pushfq
303	 *           popq %r11
304	 *   stuck_here:
305	 *
306	 * would never get past 'stuck_here'.
307	 */
308	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
309	jnz	swapgs_restore_regs_and_return_to_usermode
310
311	/* nothing to check for RSP */
312
313	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
314	jne	swapgs_restore_regs_and_return_to_usermode
315
316	/*
317	 * We win! This label is here just for ease of understanding
318	 * perf profiles. Nothing jumps here.
319	 */
320syscall_return_via_sysret:
321	/* rcx and r11 are already restored (see code above) */
322	UNWIND_HINT_EMPTY
323	POP_EXTRA_REGS
324.Lpop_c_regs_except_rcx_r11_and_sysret:
325	popq	%rsi	/* skip r11 */
326	popq	%r10
327	popq	%r9
328	popq	%r8
329	popq	%rax
330	popq	%rsi	/* skip rcx */
331	popq	%rdx
332	popq	%rsi
333	popq	%rdi
334	movq	RSP-ORIG_RAX(%rsp), %rsp
335	USERGS_SYSRET64
336END(entry_SYSCALL_64)
337
338ENTRY(stub_ptregs_64)
339	/*
340	 * Syscalls marked as needing ptregs land here.
341	 * If we are on the fast path, we need to save the extra regs,
342	 * which we achieve by trying again on the slow path.  If we are on
343	 * the slow path, the extra regs are already saved.
344	 *
345	 * RAX stores a pointer to the C function implementing the syscall.
346	 * IRQs are on.
347	 */
348	cmpq	$.Lentry_SYSCALL_64_after_fastpath_call, (%rsp)
349	jne	1f
350
351	/*
352	 * Called from fast path -- disable IRQs again, pop return address
353	 * and jump to slow path
354	 */
355	DISABLE_INTERRUPTS(CLBR_ANY)
356	TRACE_IRQS_OFF
357	popq	%rax
358	UNWIND_HINT_REGS extra=0
359	jmp	entry_SYSCALL64_slow_path
360
3611:
362	jmp	*%rax				/* Called from C */
363END(stub_ptregs_64)
364
365.macro ptregs_stub func
366ENTRY(ptregs_\func)
367	UNWIND_HINT_FUNC
368	leaq	\func(%rip), %rax
369	jmp	stub_ptregs_64
370END(ptregs_\func)
371.endm
372
373/* Instantiate ptregs_stub for each ptregs-using syscall */
374#define __SYSCALL_64_QUAL_(sym)
375#define __SYSCALL_64_QUAL_ptregs(sym) ptregs_stub sym
376#define __SYSCALL_64(nr, sym, qual) __SYSCALL_64_QUAL_##qual(sym)
377#include <asm/syscalls_64.h>
378
379/*
380 * %rdi: prev task
381 * %rsi: next task
382 */
383ENTRY(__switch_to_asm)
384	UNWIND_HINT_FUNC
385	/*
386	 * Save callee-saved registers
387	 * This must match the order in inactive_task_frame
388	 */
389	pushq	%rbp
390	pushq	%rbx
391	pushq	%r12
392	pushq	%r13
393	pushq	%r14
394	pushq	%r15
395
396	/* switch stack */
397	movq	%rsp, TASK_threadsp(%rdi)
398	movq	TASK_threadsp(%rsi), %rsp
399
400#ifdef CONFIG_CC_STACKPROTECTOR
401	movq	TASK_stack_canary(%rsi), %rbx
402	movq	%rbx, PER_CPU_VAR(irq_stack_union)+stack_canary_offset
403#endif
404
405	/* restore callee-saved registers */
406	popq	%r15
407	popq	%r14
408	popq	%r13
409	popq	%r12
410	popq	%rbx
411	popq	%rbp
412
413	jmp	__switch_to
414END(__switch_to_asm)
415
416/*
417 * A newly forked process directly context switches into this address.
418 *
419 * rax: prev task we switched from
420 * rbx: kernel thread func (NULL for user thread)
421 * r12: kernel thread arg
422 */
423ENTRY(ret_from_fork)
424	UNWIND_HINT_EMPTY
425	movq	%rax, %rdi
426	call	schedule_tail			/* rdi: 'prev' task parameter */
427
428	testq	%rbx, %rbx			/* from kernel_thread? */
429	jnz	1f				/* kernel threads are uncommon */
430
4312:
432	UNWIND_HINT_REGS
433	movq	%rsp, %rdi
434	call	syscall_return_slowpath	/* returns with IRQs disabled */
435	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
436	jmp	swapgs_restore_regs_and_return_to_usermode
437
4381:
439	/* kernel thread */
440	movq	%r12, %rdi
441	call	*%rbx
442	/*
443	 * A kernel thread is allowed to return here after successfully
444	 * calling do_execve().  Exit to userspace to complete the execve()
445	 * syscall.
446	 */
447	movq	$0, RAX(%rsp)
448	jmp	2b
449END(ret_from_fork)
450
451/*
452 * Build the entry stubs with some assembler magic.
453 * We pack 1 stub into every 8-byte block.
454 */
455	.align 8
456ENTRY(irq_entries_start)
457    vector=FIRST_EXTERNAL_VECTOR
458    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
459	UNWIND_HINT_IRET_REGS
460	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
461	jmp	common_interrupt
462	.align	8
463	vector=vector+1
464    .endr
465END(irq_entries_start)
466
467.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
468#ifdef CONFIG_DEBUG_ENTRY
469	pushfq
470	testl $X86_EFLAGS_IF, (%rsp)
471	jz .Lokay_\@
472	ud2
473.Lokay_\@:
474	addq $8, %rsp
475#endif
476.endm
477
478/*
479 * Enters the IRQ stack if we're not already using it.  NMI-safe.  Clobbers
480 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone.
481 * Requires kernel GSBASE.
482 *
483 * The invariant is that, if irq_count != -1, then the IRQ stack is in use.
484 */
485.macro ENTER_IRQ_STACK regs=1 old_rsp
486	DEBUG_ENTRY_ASSERT_IRQS_OFF
487	movq	%rsp, \old_rsp
488
489	.if \regs
490	UNWIND_HINT_REGS base=\old_rsp
491	.endif
492
493	incl	PER_CPU_VAR(irq_count)
494	jnz	.Lirq_stack_push_old_rsp_\@
495
496	/*
497	 * Right now, if we just incremented irq_count to zero, we've
498	 * claimed the IRQ stack but we haven't switched to it yet.
499	 *
500	 * If anything is added that can interrupt us here without using IST,
501	 * it must be *extremely* careful to limit its stack usage.  This
502	 * could include kprobes and a hypothetical future IST-less #DB
503	 * handler.
504	 *
505	 * The OOPS unwinder relies on the word at the top of the IRQ
506	 * stack linking back to the previous RSP for the entire time we're
507	 * on the IRQ stack.  For this to work reliably, we need to write
508	 * it before we actually move ourselves to the IRQ stack.
509	 */
510
511	movq	\old_rsp, PER_CPU_VAR(irq_stack_union + IRQ_STACK_SIZE - 8)
512	movq	PER_CPU_VAR(irq_stack_ptr), %rsp
513
514#ifdef CONFIG_DEBUG_ENTRY
515	/*
516	 * If the first movq above becomes wrong due to IRQ stack layout
517	 * changes, the only way we'll notice is if we try to unwind right
518	 * here.  Assert that we set up the stack right to catch this type
519	 * of bug quickly.
520	 */
521	cmpq	-8(%rsp), \old_rsp
522	je	.Lirq_stack_okay\@
523	ud2
524	.Lirq_stack_okay\@:
525#endif
526
527.Lirq_stack_push_old_rsp_\@:
528	pushq	\old_rsp
529
530	.if \regs
531	UNWIND_HINT_REGS indirect=1
532	.endif
533.endm
534
535/*
536 * Undoes ENTER_IRQ_STACK.
537 */
538.macro LEAVE_IRQ_STACK regs=1
539	DEBUG_ENTRY_ASSERT_IRQS_OFF
540	/* We need to be off the IRQ stack before decrementing irq_count. */
541	popq	%rsp
542
543	.if \regs
544	UNWIND_HINT_REGS
545	.endif
546
547	/*
548	 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming
549	 * the irq stack but we're not on it.
550	 */
551
552	decl	PER_CPU_VAR(irq_count)
553.endm
554
555/*
556 * Interrupt entry/exit.
557 *
558 * Interrupt entry points save only callee clobbered registers in fast path.
559 *
560 * Entry runs with interrupts off.
561 */
562
563/* 0(%rsp): ~(interrupt number) */
564	.macro interrupt func
565	cld
566	ALLOC_PT_GPREGS_ON_STACK
567	SAVE_C_REGS
568	SAVE_EXTRA_REGS
569	ENCODE_FRAME_POINTER
570
571	testb	$3, CS(%rsp)
572	jz	1f
573
574	/*
575	 * IRQ from user mode.  Switch to kernel gsbase and inform context
576	 * tracking that we're in kernel mode.
577	 */
578	SWAPGS
579
580	/*
581	 * We need to tell lockdep that IRQs are off.  We can't do this until
582	 * we fix gsbase, and we should do it before enter_from_user_mode
583	 * (which can take locks).  Since TRACE_IRQS_OFF idempotent,
584	 * the simplest way to handle it is to just call it twice if
585	 * we enter from user mode.  There's no reason to optimize this since
586	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
587	 */
588	TRACE_IRQS_OFF
589
590	CALL_enter_from_user_mode
591
5921:
593	ENTER_IRQ_STACK old_rsp=%rdi
594	/* We entered an interrupt context - irqs are off: */
595	TRACE_IRQS_OFF
596
597	call	\func	/* rdi points to pt_regs */
598	.endm
599
600	/*
601	 * The interrupt stubs push (~vector+0x80) onto the stack and
602	 * then jump to common_interrupt.
603	 */
604	.p2align CONFIG_X86_L1_CACHE_SHIFT
605common_interrupt:
606	ASM_CLAC
607	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
608	interrupt do_IRQ
609	/* 0(%rsp): old RSP */
610ret_from_intr:
611	DISABLE_INTERRUPTS(CLBR_ANY)
612	TRACE_IRQS_OFF
613
614	LEAVE_IRQ_STACK
615
616	testb	$3, CS(%rsp)
617	jz	retint_kernel
618
619	/* Interrupt came from user space */
620GLOBAL(retint_user)
621	mov	%rsp,%rdi
622	call	prepare_exit_to_usermode
623	TRACE_IRQS_IRETQ
624
625GLOBAL(swapgs_restore_regs_and_return_to_usermode)
626#ifdef CONFIG_DEBUG_ENTRY
627	/* Assert that pt_regs indicates user mode. */
628	testb	$3, CS(%rsp)
629	jnz	1f
630	ud2
6311:
632#endif
633	SWAPGS
634	POP_EXTRA_REGS
635	POP_C_REGS
636	addq	$8, %rsp	/* skip regs->orig_ax */
637	INTERRUPT_RETURN
638
639
640/* Returning to kernel space */
641retint_kernel:
642#ifdef CONFIG_PREEMPT
643	/* Interrupts are off */
644	/* Check if we need preemption */
645	bt	$9, EFLAGS(%rsp)		/* were interrupts off? */
646	jnc	1f
6470:	cmpl	$0, PER_CPU_VAR(__preempt_count)
648	jnz	1f
649	call	preempt_schedule_irq
650	jmp	0b
6511:
652#endif
653	/*
654	 * The iretq could re-enable interrupts:
655	 */
656	TRACE_IRQS_IRETQ
657
658GLOBAL(restore_regs_and_return_to_kernel)
659#ifdef CONFIG_DEBUG_ENTRY
660	/* Assert that pt_regs indicates kernel mode. */
661	testb	$3, CS(%rsp)
662	jz	1f
663	ud2
6641:
665#endif
666	POP_EXTRA_REGS
667	POP_C_REGS
668	addq	$8, %rsp	/* skip regs->orig_ax */
669	INTERRUPT_RETURN
670
671ENTRY(native_iret)
672	UNWIND_HINT_IRET_REGS
673	/*
674	 * Are we returning to a stack segment from the LDT?  Note: in
675	 * 64-bit mode SS:RSP on the exception stack is always valid.
676	 */
677#ifdef CONFIG_X86_ESPFIX64
678	testb	$4, (SS-RIP)(%rsp)
679	jnz	native_irq_return_ldt
680#endif
681
682.global native_irq_return_iret
683native_irq_return_iret:
684	/*
685	 * This may fault.  Non-paranoid faults on return to userspace are
686	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
687	 * Double-faults due to espfix64 are handled in do_double_fault.
688	 * Other faults here are fatal.
689	 */
690	iretq
691
692#ifdef CONFIG_X86_ESPFIX64
693native_irq_return_ldt:
694	/*
695	 * We are running with user GSBASE.  All GPRs contain their user
696	 * values.  We have a percpu ESPFIX stack that is eight slots
697	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
698	 * of the ESPFIX stack.
699	 *
700	 * We clobber RAX and RDI in this code.  We stash RDI on the
701	 * normal stack and RAX on the ESPFIX stack.
702	 *
703	 * The ESPFIX stack layout we set up looks like this:
704	 *
705	 * --- top of ESPFIX stack ---
706	 * SS
707	 * RSP
708	 * RFLAGS
709	 * CS
710	 * RIP  <-- RSP points here when we're done
711	 * RAX  <-- espfix_waddr points here
712	 * --- bottom of ESPFIX stack ---
713	 */
714
715	pushq	%rdi				/* Stash user RDI */
716	SWAPGS
717	movq	PER_CPU_VAR(espfix_waddr), %rdi
718	movq	%rax, (0*8)(%rdi)		/* user RAX */
719	movq	(1*8)(%rsp), %rax		/* user RIP */
720	movq	%rax, (1*8)(%rdi)
721	movq	(2*8)(%rsp), %rax		/* user CS */
722	movq	%rax, (2*8)(%rdi)
723	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
724	movq	%rax, (3*8)(%rdi)
725	movq	(5*8)(%rsp), %rax		/* user SS */
726	movq	%rax, (5*8)(%rdi)
727	movq	(4*8)(%rsp), %rax		/* user RSP */
728	movq	%rax, (4*8)(%rdi)
729	/* Now RAX == RSP. */
730
731	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
732	popq	%rdi				/* Restore user RDI */
733
734	/*
735	 * espfix_stack[31:16] == 0.  The page tables are set up such that
736	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
737	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
738	 * the same page.  Set up RSP so that RSP[31:16] contains the
739	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
740	 * still points to an RO alias of the ESPFIX stack.
741	 */
742	orq	PER_CPU_VAR(espfix_stack), %rax
743	SWAPGS
744	movq	%rax, %rsp
745	UNWIND_HINT_IRET_REGS offset=8
746
747	/*
748	 * At this point, we cannot write to the stack any more, but we can
749	 * still read.
750	 */
751	popq	%rax				/* Restore user RAX */
752
753	/*
754	 * RSP now points to an ordinary IRET frame, except that the page
755	 * is read-only and RSP[31:16] are preloaded with the userspace
756	 * values.  We can now IRET back to userspace.
757	 */
758	jmp	native_irq_return_iret
759#endif
760END(common_interrupt)
761
762/*
763 * APIC interrupts.
764 */
765.macro apicinterrupt3 num sym do_sym
766ENTRY(\sym)
767	UNWIND_HINT_IRET_REGS
768	ASM_CLAC
769	pushq	$~(\num)
770.Lcommon_\sym:
771	interrupt \do_sym
772	jmp	ret_from_intr
773END(\sym)
774.endm
775
776/* Make sure APIC interrupt handlers end up in the irqentry section: */
777#define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
778#define POP_SECTION_IRQENTRY	.popsection
779
780.macro apicinterrupt num sym do_sym
781PUSH_SECTION_IRQENTRY
782apicinterrupt3 \num \sym \do_sym
783POP_SECTION_IRQENTRY
784.endm
785
786#ifdef CONFIG_SMP
787apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
788apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
789#endif
790
791#ifdef CONFIG_X86_UV
792apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
793#endif
794
795apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
796apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
797
798#ifdef CONFIG_HAVE_KVM
799apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
800apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
801apicinterrupt3 POSTED_INTR_NESTED_VECTOR	kvm_posted_intr_nested_ipi	smp_kvm_posted_intr_nested_ipi
802#endif
803
804#ifdef CONFIG_X86_MCE_THRESHOLD
805apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
806#endif
807
808#ifdef CONFIG_X86_MCE_AMD
809apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
810#endif
811
812#ifdef CONFIG_X86_THERMAL_VECTOR
813apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
814#endif
815
816#ifdef CONFIG_SMP
817apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
818apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
819apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
820#endif
821
822apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
823apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
824
825#ifdef CONFIG_IRQ_WORK
826apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
827#endif
828
829/*
830 * Exception entry points.
831 */
832#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss) + (TSS_ist + ((x) - 1) * 8)
833
834.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
835ENTRY(\sym)
836	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
837
838	/* Sanity check */
839	.if \shift_ist != -1 && \paranoid == 0
840	.error "using shift_ist requires paranoid=1"
841	.endif
842
843	ASM_CLAC
844
845	.if \has_error_code == 0
846	pushq	$-1				/* ORIG_RAX: no syscall to restart */
847	.endif
848
849	ALLOC_PT_GPREGS_ON_STACK
850
851	.if \paranoid
852	.if \paranoid == 1
853	testb	$3, CS(%rsp)			/* If coming from userspace, switch stacks */
854	jnz	1f
855	.endif
856	call	paranoid_entry
857	.else
858	call	error_entry
859	.endif
860	UNWIND_HINT_REGS
861	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
862
863	.if \paranoid
864	.if \shift_ist != -1
865	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
866	.else
867	TRACE_IRQS_OFF
868	.endif
869	.endif
870
871	movq	%rsp, %rdi			/* pt_regs pointer */
872
873	.if \has_error_code
874	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
875	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
876	.else
877	xorl	%esi, %esi			/* no error code */
878	.endif
879
880	.if \shift_ist != -1
881	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
882	.endif
883
884	call	\do_sym
885
886	.if \shift_ist != -1
887	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
888	.endif
889
890	/* these procedures expect "no swapgs" flag in ebx */
891	.if \paranoid
892	jmp	paranoid_exit
893	.else
894	jmp	error_exit
895	.endif
896
897	.if \paranoid == 1
898	/*
899	 * Paranoid entry from userspace.  Switch stacks and treat it
900	 * as a normal entry.  This means that paranoid handlers
901	 * run in real process context if user_mode(regs).
902	 */
9031:
904	call	error_entry
905
906
907	movq	%rsp, %rdi			/* pt_regs pointer */
908	call	sync_regs
909	movq	%rax, %rsp			/* switch stack */
910
911	movq	%rsp, %rdi			/* pt_regs pointer */
912
913	.if \has_error_code
914	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
915	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
916	.else
917	xorl	%esi, %esi			/* no error code */
918	.endif
919
920	call	\do_sym
921
922	jmp	error_exit			/* %ebx: no swapgs flag */
923	.endif
924END(\sym)
925.endm
926
927idtentry divide_error			do_divide_error			has_error_code=0
928idtentry overflow			do_overflow			has_error_code=0
929idtentry bounds				do_bounds			has_error_code=0
930idtentry invalid_op			do_invalid_op			has_error_code=0
931idtentry device_not_available		do_device_not_available		has_error_code=0
932idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
933idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
934idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
935idtentry segment_not_present		do_segment_not_present		has_error_code=1
936idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
937idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
938idtentry alignment_check		do_alignment_check		has_error_code=1
939idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
940
941
942	/*
943	 * Reload gs selector with exception handling
944	 * edi:  new selector
945	 */
946ENTRY(native_load_gs_index)
947	FRAME_BEGIN
948	pushfq
949	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
950	TRACE_IRQS_OFF
951	SWAPGS
952.Lgs_change:
953	movl	%edi, %gs
9542:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
955	SWAPGS
956	TRACE_IRQS_FLAGS (%rsp)
957	popfq
958	FRAME_END
959	ret
960ENDPROC(native_load_gs_index)
961EXPORT_SYMBOL(native_load_gs_index)
962
963	_ASM_EXTABLE(.Lgs_change, bad_gs)
964	.section .fixup, "ax"
965	/* running with kernelgs */
966bad_gs:
967	SWAPGS					/* switch back to user gs */
968.macro ZAP_GS
969	/* This can't be a string because the preprocessor needs to see it. */
970	movl $__USER_DS, %eax
971	movl %eax, %gs
972.endm
973	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
974	xorl	%eax, %eax
975	movl	%eax, %gs
976	jmp	2b
977	.previous
978
979/* Call softirq on interrupt stack. Interrupts are off. */
980ENTRY(do_softirq_own_stack)
981	pushq	%rbp
982	mov	%rsp, %rbp
983	ENTER_IRQ_STACK regs=0 old_rsp=%r11
984	call	__do_softirq
985	LEAVE_IRQ_STACK regs=0
986	leaveq
987	ret
988ENDPROC(do_softirq_own_stack)
989
990#ifdef CONFIG_XEN
991idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0
992
993/*
994 * A note on the "critical region" in our callback handler.
995 * We want to avoid stacking callback handlers due to events occurring
996 * during handling of the last event. To do this, we keep events disabled
997 * until we've done all processing. HOWEVER, we must enable events before
998 * popping the stack frame (can't be done atomically) and so it would still
999 * be possible to get enough handler activations to overflow the stack.
1000 * Although unlikely, bugs of that kind are hard to track down, so we'd
1001 * like to avoid the possibility.
1002 * So, on entry to the handler we detect whether we interrupted an
1003 * existing activation in its critical region -- if so, we pop the current
1004 * activation and restart the handler using the previous one.
1005 */
1006ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
1007
1008/*
1009 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
1010 * see the correct pointer to the pt_regs
1011 */
1012	UNWIND_HINT_FUNC
1013	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
1014	UNWIND_HINT_REGS
1015
1016	ENTER_IRQ_STACK old_rsp=%r10
1017	call	xen_evtchn_do_upcall
1018	LEAVE_IRQ_STACK
1019
1020#ifndef CONFIG_PREEMPT
1021	call	xen_maybe_preempt_hcall
1022#endif
1023	jmp	error_exit
1024END(xen_do_hypervisor_callback)
1025
1026/*
1027 * Hypervisor uses this for application faults while it executes.
1028 * We get here for two reasons:
1029 *  1. Fault while reloading DS, ES, FS or GS
1030 *  2. Fault while executing IRET
1031 * Category 1 we do not need to fix up as Xen has already reloaded all segment
1032 * registers that could be reloaded and zeroed the others.
1033 * Category 2 we fix up by killing the current process. We cannot use the
1034 * normal Linux return path in this case because if we use the IRET hypercall
1035 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
1036 * We distinguish between categories by comparing each saved segment register
1037 * with its current contents: any discrepancy means we in category 1.
1038 */
1039ENTRY(xen_failsafe_callback)
1040	UNWIND_HINT_EMPTY
1041	movl	%ds, %ecx
1042	cmpw	%cx, 0x10(%rsp)
1043	jne	1f
1044	movl	%es, %ecx
1045	cmpw	%cx, 0x18(%rsp)
1046	jne	1f
1047	movl	%fs, %ecx
1048	cmpw	%cx, 0x20(%rsp)
1049	jne	1f
1050	movl	%gs, %ecx
1051	cmpw	%cx, 0x28(%rsp)
1052	jne	1f
1053	/* All segments match their saved values => Category 2 (Bad IRET). */
1054	movq	(%rsp), %rcx
1055	movq	8(%rsp), %r11
1056	addq	$0x30, %rsp
1057	pushq	$0				/* RIP */
1058	UNWIND_HINT_IRET_REGS offset=8
1059	jmp	general_protection
10601:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
1061	movq	(%rsp), %rcx
1062	movq	8(%rsp), %r11
1063	addq	$0x30, %rsp
1064	UNWIND_HINT_IRET_REGS
1065	pushq	$-1 /* orig_ax = -1 => not a system call */
1066	ALLOC_PT_GPREGS_ON_STACK
1067	SAVE_C_REGS
1068	SAVE_EXTRA_REGS
1069	ENCODE_FRAME_POINTER
1070	jmp	error_exit
1071END(xen_failsafe_callback)
1072
1073apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1074	xen_hvm_callback_vector xen_evtchn_do_upcall
1075
1076#endif /* CONFIG_XEN */
1077
1078#if IS_ENABLED(CONFIG_HYPERV)
1079apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1080	hyperv_callback_vector hyperv_vector_handler
1081#endif /* CONFIG_HYPERV */
1082
1083idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1084idtentry int3			do_int3			has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1085idtentry stack_segment		do_stack_segment	has_error_code=1
1086
1087#ifdef CONFIG_XEN
1088idtentry xennmi			do_nmi			has_error_code=0
1089idtentry xendebug		do_debug		has_error_code=0
1090idtentry xenint3		do_int3			has_error_code=0
1091#endif
1092
1093idtentry general_protection	do_general_protection	has_error_code=1
1094idtentry page_fault		do_page_fault		has_error_code=1
1095
1096#ifdef CONFIG_KVM_GUEST
1097idtentry async_page_fault	do_async_page_fault	has_error_code=1
1098#endif
1099
1100#ifdef CONFIG_X86_MCE
1101idtentry machine_check					has_error_code=0	paranoid=1 do_sym=*machine_check_vector(%rip)
1102#endif
1103
1104/*
1105 * Save all registers in pt_regs, and switch gs if needed.
1106 * Use slow, but surefire "are we in kernel?" check.
1107 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
1108 */
1109ENTRY(paranoid_entry)
1110	UNWIND_HINT_FUNC
1111	cld
1112	SAVE_C_REGS 8
1113	SAVE_EXTRA_REGS 8
1114	ENCODE_FRAME_POINTER 8
1115	movl	$1, %ebx
1116	movl	$MSR_GS_BASE, %ecx
1117	rdmsr
1118	testl	%edx, %edx
1119	js	1f				/* negative -> in kernel */
1120	SWAPGS
1121	xorl	%ebx, %ebx
11221:	ret
1123END(paranoid_entry)
1124
1125/*
1126 * "Paranoid" exit path from exception stack.  This is invoked
1127 * only on return from non-NMI IST interrupts that came
1128 * from kernel space.
1129 *
1130 * We may be returning to very strange contexts (e.g. very early
1131 * in syscall entry), so checking for preemption here would
1132 * be complicated.  Fortunately, we there's no good reason
1133 * to try to handle preemption here.
1134 *
1135 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1136 */
1137ENTRY(paranoid_exit)
1138	UNWIND_HINT_REGS
1139	DISABLE_INTERRUPTS(CLBR_ANY)
1140	TRACE_IRQS_OFF_DEBUG
1141	testl	%ebx, %ebx			/* swapgs needed? */
1142	jnz	.Lparanoid_exit_no_swapgs
1143	TRACE_IRQS_IRETQ
1144	SWAPGS_UNSAFE_STACK
1145	jmp	.Lparanoid_exit_restore
1146.Lparanoid_exit_no_swapgs:
1147	TRACE_IRQS_IRETQ_DEBUG
1148.Lparanoid_exit_restore:
1149	jmp restore_regs_and_return_to_kernel
1150END(paranoid_exit)
1151
1152/*
1153 * Save all registers in pt_regs, and switch gs if needed.
1154 * Return: EBX=0: came from user mode; EBX=1: otherwise
1155 */
1156ENTRY(error_entry)
1157	UNWIND_HINT_FUNC
1158	cld
1159	SAVE_C_REGS 8
1160	SAVE_EXTRA_REGS 8
1161	ENCODE_FRAME_POINTER 8
1162	xorl	%ebx, %ebx
1163	testb	$3, CS+8(%rsp)
1164	jz	.Lerror_kernelspace
1165
1166	/*
1167	 * We entered from user mode or we're pretending to have entered
1168	 * from user mode due to an IRET fault.
1169	 */
1170	SWAPGS
1171
1172.Lerror_entry_from_usermode_after_swapgs:
1173	/*
1174	 * We need to tell lockdep that IRQs are off.  We can't do this until
1175	 * we fix gsbase, and we should do it before enter_from_user_mode
1176	 * (which can take locks).
1177	 */
1178	TRACE_IRQS_OFF
1179	CALL_enter_from_user_mode
1180	ret
1181
1182.Lerror_entry_done:
1183	TRACE_IRQS_OFF
1184	ret
1185
1186	/*
1187	 * There are two places in the kernel that can potentially fault with
1188	 * usergs. Handle them here.  B stepping K8s sometimes report a
1189	 * truncated RIP for IRET exceptions returning to compat mode. Check
1190	 * for these here too.
1191	 */
1192.Lerror_kernelspace:
1193	incl	%ebx
1194	leaq	native_irq_return_iret(%rip), %rcx
1195	cmpq	%rcx, RIP+8(%rsp)
1196	je	.Lerror_bad_iret
1197	movl	%ecx, %eax			/* zero extend */
1198	cmpq	%rax, RIP+8(%rsp)
1199	je	.Lbstep_iret
1200	cmpq	$.Lgs_change, RIP+8(%rsp)
1201	jne	.Lerror_entry_done
1202
1203	/*
1204	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1205	 * gsbase and proceed.  We'll fix up the exception and land in
1206	 * .Lgs_change's error handler with kernel gsbase.
1207	 */
1208	SWAPGS
1209	jmp .Lerror_entry_done
1210
1211.Lbstep_iret:
1212	/* Fix truncated RIP */
1213	movq	%rcx, RIP+8(%rsp)
1214	/* fall through */
1215
1216.Lerror_bad_iret:
1217	/*
1218	 * We came from an IRET to user mode, so we have user gsbase.
1219	 * Switch to kernel gsbase:
1220	 */
1221	SWAPGS
1222
1223	/*
1224	 * Pretend that the exception came from user mode: set up pt_regs
1225	 * as if we faulted immediately after IRET and clear EBX so that
1226	 * error_exit knows that we will be returning to user mode.
1227	 */
1228	mov	%rsp, %rdi
1229	call	fixup_bad_iret
1230	mov	%rax, %rsp
1231	decl	%ebx
1232	jmp	.Lerror_entry_from_usermode_after_swapgs
1233END(error_entry)
1234
1235
1236/*
1237 * On entry, EBX is a "return to kernel mode" flag:
1238 *   1: already in kernel mode, don't need SWAPGS
1239 *   0: user gsbase is loaded, we need SWAPGS and standard preparation for return to usermode
1240 */
1241ENTRY(error_exit)
1242	UNWIND_HINT_REGS
1243	DISABLE_INTERRUPTS(CLBR_ANY)
1244	TRACE_IRQS_OFF
1245	testl	%ebx, %ebx
1246	jnz	retint_kernel
1247	jmp	retint_user
1248END(error_exit)
1249
1250/*
1251 * Runs on exception stack.  Xen PV does not go through this path at all,
1252 * so we can use real assembly here.
1253 */
1254ENTRY(nmi)
1255	UNWIND_HINT_IRET_REGS
1256
1257	/*
1258	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1259	 * the iretq it performs will take us out of NMI context.
1260	 * This means that we can have nested NMIs where the next
1261	 * NMI is using the top of the stack of the previous NMI. We
1262	 * can't let it execute because the nested NMI will corrupt the
1263	 * stack of the previous NMI. NMI handlers are not re-entrant
1264	 * anyway.
1265	 *
1266	 * To handle this case we do the following:
1267	 *  Check the a special location on the stack that contains
1268	 *  a variable that is set when NMIs are executing.
1269	 *  The interrupted task's stack is also checked to see if it
1270	 *  is an NMI stack.
1271	 *  If the variable is not set and the stack is not the NMI
1272	 *  stack then:
1273	 *    o Set the special variable on the stack
1274	 *    o Copy the interrupt frame into an "outermost" location on the
1275	 *      stack
1276	 *    o Copy the interrupt frame into an "iret" location on the stack
1277	 *    o Continue processing the NMI
1278	 *  If the variable is set or the previous stack is the NMI stack:
1279	 *    o Modify the "iret" location to jump to the repeat_nmi
1280	 *    o return back to the first NMI
1281	 *
1282	 * Now on exit of the first NMI, we first clear the stack variable
1283	 * The NMI stack will tell any nested NMIs at that point that it is
1284	 * nested. Then we pop the stack normally with iret, and if there was
1285	 * a nested NMI that updated the copy interrupt stack frame, a
1286	 * jump will be made to the repeat_nmi code that will handle the second
1287	 * NMI.
1288	 *
1289	 * However, espfix prevents us from directly returning to userspace
1290	 * with a single IRET instruction.  Similarly, IRET to user mode
1291	 * can fault.  We therefore handle NMIs from user space like
1292	 * other IST entries.
1293	 */
1294
1295	ASM_CLAC
1296
1297	/* Use %rdx as our temp variable throughout */
1298	pushq	%rdx
1299
1300	testb	$3, CS-RIP+8(%rsp)
1301	jz	.Lnmi_from_kernel
1302
1303	/*
1304	 * NMI from user mode.  We need to run on the thread stack, but we
1305	 * can't go through the normal entry paths: NMIs are masked, and
1306	 * we don't want to enable interrupts, because then we'll end
1307	 * up in an awkward situation in which IRQs are on but NMIs
1308	 * are off.
1309	 *
1310	 * We also must not push anything to the stack before switching
1311	 * stacks lest we corrupt the "NMI executing" variable.
1312	 */
1313
1314	swapgs
1315	cld
1316	movq	%rsp, %rdx
1317	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1318	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1319	pushq	5*8(%rdx)	/* pt_regs->ss */
1320	pushq	4*8(%rdx)	/* pt_regs->rsp */
1321	pushq	3*8(%rdx)	/* pt_regs->flags */
1322	pushq	2*8(%rdx)	/* pt_regs->cs */
1323	pushq	1*8(%rdx)	/* pt_regs->rip */
1324	UNWIND_HINT_IRET_REGS
1325	pushq   $-1		/* pt_regs->orig_ax */
1326	pushq   %rdi		/* pt_regs->di */
1327	pushq   %rsi		/* pt_regs->si */
1328	pushq   (%rdx)		/* pt_regs->dx */
1329	pushq   %rcx		/* pt_regs->cx */
1330	pushq   %rax		/* pt_regs->ax */
1331	pushq   %r8		/* pt_regs->r8 */
1332	pushq   %r9		/* pt_regs->r9 */
1333	pushq   %r10		/* pt_regs->r10 */
1334	pushq   %r11		/* pt_regs->r11 */
1335	pushq	%rbx		/* pt_regs->rbx */
1336	pushq	%rbp		/* pt_regs->rbp */
1337	pushq	%r12		/* pt_regs->r12 */
1338	pushq	%r13		/* pt_regs->r13 */
1339	pushq	%r14		/* pt_regs->r14 */
1340	pushq	%r15		/* pt_regs->r15 */
1341	UNWIND_HINT_REGS
1342	ENCODE_FRAME_POINTER
1343
1344	/*
1345	 * At this point we no longer need to worry about stack damage
1346	 * due to nesting -- we're on the normal thread stack and we're
1347	 * done with the NMI stack.
1348	 */
1349
1350	movq	%rsp, %rdi
1351	movq	$-1, %rsi
1352	call	do_nmi
1353
1354	/*
1355	 * Return back to user mode.  We must *not* do the normal exit
1356	 * work, because we don't want to enable interrupts.
1357	 */
1358	jmp	swapgs_restore_regs_and_return_to_usermode
1359
1360.Lnmi_from_kernel:
1361	/*
1362	 * Here's what our stack frame will look like:
1363	 * +---------------------------------------------------------+
1364	 * | original SS                                             |
1365	 * | original Return RSP                                     |
1366	 * | original RFLAGS                                         |
1367	 * | original CS                                             |
1368	 * | original RIP                                            |
1369	 * +---------------------------------------------------------+
1370	 * | temp storage for rdx                                    |
1371	 * +---------------------------------------------------------+
1372	 * | "NMI executing" variable                                |
1373	 * +---------------------------------------------------------+
1374	 * | iret SS          } Copied from "outermost" frame        |
1375	 * | iret Return RSP  } on each loop iteration; overwritten  |
1376	 * | iret RFLAGS      } by a nested NMI to force another     |
1377	 * | iret CS          } iteration if needed.                 |
1378	 * | iret RIP         }                                      |
1379	 * +---------------------------------------------------------+
1380	 * | outermost SS          } initialized in first_nmi;       |
1381	 * | outermost Return RSP  } will not be changed before      |
1382	 * | outermost RFLAGS      } NMI processing is done.         |
1383	 * | outermost CS          } Copied to "iret" frame on each  |
1384	 * | outermost RIP         } iteration.                      |
1385	 * +---------------------------------------------------------+
1386	 * | pt_regs                                                 |
1387	 * +---------------------------------------------------------+
1388	 *
1389	 * The "original" frame is used by hardware.  Before re-enabling
1390	 * NMIs, we need to be done with it, and we need to leave enough
1391	 * space for the asm code here.
1392	 *
1393	 * We return by executing IRET while RSP points to the "iret" frame.
1394	 * That will either return for real or it will loop back into NMI
1395	 * processing.
1396	 *
1397	 * The "outermost" frame is copied to the "iret" frame on each
1398	 * iteration of the loop, so each iteration starts with the "iret"
1399	 * frame pointing to the final return target.
1400	 */
1401
1402	/*
1403	 * Determine whether we're a nested NMI.
1404	 *
1405	 * If we interrupted kernel code between repeat_nmi and
1406	 * end_repeat_nmi, then we are a nested NMI.  We must not
1407	 * modify the "iret" frame because it's being written by
1408	 * the outer NMI.  That's okay; the outer NMI handler is
1409	 * about to about to call do_nmi anyway, so we can just
1410	 * resume the outer NMI.
1411	 */
1412
1413	movq	$repeat_nmi, %rdx
1414	cmpq	8(%rsp), %rdx
1415	ja	1f
1416	movq	$end_repeat_nmi, %rdx
1417	cmpq	8(%rsp), %rdx
1418	ja	nested_nmi_out
14191:
1420
1421	/*
1422	 * Now check "NMI executing".  If it's set, then we're nested.
1423	 * This will not detect if we interrupted an outer NMI just
1424	 * before IRET.
1425	 */
1426	cmpl	$1, -8(%rsp)
1427	je	nested_nmi
1428
1429	/*
1430	 * Now test if the previous stack was an NMI stack.  This covers
1431	 * the case where we interrupt an outer NMI after it clears
1432	 * "NMI executing" but before IRET.  We need to be careful, though:
1433	 * there is one case in which RSP could point to the NMI stack
1434	 * despite there being no NMI active: naughty userspace controls
1435	 * RSP at the very beginning of the SYSCALL targets.  We can
1436	 * pull a fast one on naughty userspace, though: we program
1437	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1438	 * if it controls the kernel's RSP.  We set DF before we clear
1439	 * "NMI executing".
1440	 */
1441	lea	6*8(%rsp), %rdx
1442	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1443	cmpq	%rdx, 4*8(%rsp)
1444	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1445	ja	first_nmi
1446
1447	subq	$EXCEPTION_STKSZ, %rdx
1448	cmpq	%rdx, 4*8(%rsp)
1449	/* If it is below the NMI stack, it is a normal NMI */
1450	jb	first_nmi
1451
1452	/* Ah, it is within the NMI stack. */
1453
1454	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1455	jz	first_nmi	/* RSP was user controlled. */
1456
1457	/* This is a nested NMI. */
1458
1459nested_nmi:
1460	/*
1461	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1462	 * iteration of NMI handling.
1463	 */
1464	subq	$8, %rsp
1465	leaq	-10*8(%rsp), %rdx
1466	pushq	$__KERNEL_DS
1467	pushq	%rdx
1468	pushfq
1469	pushq	$__KERNEL_CS
1470	pushq	$repeat_nmi
1471
1472	/* Put stack back */
1473	addq	$(6*8), %rsp
1474
1475nested_nmi_out:
1476	popq	%rdx
1477
1478	/* We are returning to kernel mode, so this cannot result in a fault. */
1479	iretq
1480
1481first_nmi:
1482	/* Restore rdx. */
1483	movq	(%rsp), %rdx
1484
1485	/* Make room for "NMI executing". */
1486	pushq	$0
1487
1488	/* Leave room for the "iret" frame */
1489	subq	$(5*8), %rsp
1490
1491	/* Copy the "original" frame to the "outermost" frame */
1492	.rept 5
1493	pushq	11*8(%rsp)
1494	.endr
1495	UNWIND_HINT_IRET_REGS
1496
1497	/* Everything up to here is safe from nested NMIs */
1498
1499#ifdef CONFIG_DEBUG_ENTRY
1500	/*
1501	 * For ease of testing, unmask NMIs right away.  Disabled by
1502	 * default because IRET is very expensive.
1503	 */
1504	pushq	$0		/* SS */
1505	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1506	addq	$8, (%rsp)	/* Fix up RSP */
1507	pushfq			/* RFLAGS */
1508	pushq	$__KERNEL_CS	/* CS */
1509	pushq	$1f		/* RIP */
1510	iretq			/* continues at repeat_nmi below */
1511	UNWIND_HINT_IRET_REGS
15121:
1513#endif
1514
1515repeat_nmi:
1516	/*
1517	 * If there was a nested NMI, the first NMI's iret will return
1518	 * here. But NMIs are still enabled and we can take another
1519	 * nested NMI. The nested NMI checks the interrupted RIP to see
1520	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1521	 * it will just return, as we are about to repeat an NMI anyway.
1522	 * This makes it safe to copy to the stack frame that a nested
1523	 * NMI will update.
1524	 *
1525	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1526	 * we're repeating an NMI, gsbase has the same value that it had on
1527	 * the first iteration.  paranoid_entry will load the kernel
1528	 * gsbase if needed before we call do_nmi.  "NMI executing"
1529	 * is zero.
1530	 */
1531	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1532
1533	/*
1534	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1535	 * here must not modify the "iret" frame while we're writing to
1536	 * it or it will end up containing garbage.
1537	 */
1538	addq	$(10*8), %rsp
1539	.rept 5
1540	pushq	-6*8(%rsp)
1541	.endr
1542	subq	$(5*8), %rsp
1543end_repeat_nmi:
1544
1545	/*
1546	 * Everything below this point can be preempted by a nested NMI.
1547	 * If this happens, then the inner NMI will change the "iret"
1548	 * frame to point back to repeat_nmi.
1549	 */
1550	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1551	ALLOC_PT_GPREGS_ON_STACK
1552
1553	/*
1554	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1555	 * as we should not be calling schedule in NMI context.
1556	 * Even with normal interrupts enabled. An NMI should not be
1557	 * setting NEED_RESCHED or anything that normal interrupts and
1558	 * exceptions might do.
1559	 */
1560	call	paranoid_entry
1561	UNWIND_HINT_REGS
1562
1563	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1564	movq	%rsp, %rdi
1565	movq	$-1, %rsi
1566	call	do_nmi
1567
1568	testl	%ebx, %ebx			/* swapgs needed? */
1569	jnz	nmi_restore
1570nmi_swapgs:
1571	SWAPGS_UNSAFE_STACK
1572nmi_restore:
1573	POP_EXTRA_REGS
1574	POP_C_REGS
1575
1576	/*
1577	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1578	 * at the "iret" frame.
1579	 */
1580	addq	$6*8, %rsp
1581
1582	/*
1583	 * Clear "NMI executing".  Set DF first so that we can easily
1584	 * distinguish the remaining code between here and IRET from
1585	 * the SYSCALL entry and exit paths.
1586	 *
1587	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1588	 * this code when I had the misapprehension that Xen PV supported
1589	 * NMIs, and Xen PV would break that approach.
1590	 */
1591	std
1592	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1593
1594	/*
1595	 * iretq reads the "iret" frame and exits the NMI stack in a
1596	 * single instruction.  We are returning to kernel mode, so this
1597	 * cannot result in a fault.  Similarly, we don't need to worry
1598	 * about espfix64 on the way back to kernel mode.
1599	 */
1600	iretq
1601END(nmi)
1602
1603ENTRY(ignore_sysret)
1604	UNWIND_HINT_EMPTY
1605	mov	$-ENOSYS, %eax
1606	sysret
1607END(ignore_sysret)
1608
1609ENTRY(rewind_stack_do_exit)
1610	UNWIND_HINT_FUNC
1611	/* Prevent any naive code from trying to unwind to our caller. */
1612	xorl	%ebp, %ebp
1613
1614	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1615	leaq	-PTREGS_SIZE(%rax), %rsp
1616	UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE
1617
1618	call	do_exit
1619END(rewind_stack_do_exit)
1620