xref: /linux/arch/x86/entry/entry_64.S (revision 155681fcd7f82882a730240c2dde7eee76a46314)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 *  linux/arch/x86_64/entry.S
4 *
5 *  Copyright (C) 1991, 1992  Linus Torvalds
6 *  Copyright (C) 2000, 2001, 2002  Andi Kleen SuSE Labs
7 *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
8 *
9 * entry.S contains the system-call and fault low-level handling routines.
10 *
11 * Some of this is documented in Documentation/x86/entry_64.txt
12 *
13 * A note on terminology:
14 * - iret frame:	Architecture defined interrupt frame from SS to RIP
15 *			at the top of the kernel process stack.
16 *
17 * Some macro usage:
18 * - ENTRY/END:		Define functions in the symbol table.
19 * - TRACE_IRQ_*:	Trace hardirq state for lock debugging.
20 * - idtentry:		Define exception entry points.
21 */
22#include <linux/linkage.h>
23#include <asm/segment.h>
24#include <asm/cache.h>
25#include <asm/errno.h>
26#include <asm/asm-offsets.h>
27#include <asm/msr.h>
28#include <asm/unistd.h>
29#include <asm/thread_info.h>
30#include <asm/hw_irq.h>
31#include <asm/page_types.h>
32#include <asm/irqflags.h>
33#include <asm/paravirt.h>
34#include <asm/percpu.h>
35#include <asm/asm.h>
36#include <asm/smap.h>
37#include <asm/pgtable_types.h>
38#include <asm/export.h>
39#include <asm/frame.h>
40#include <asm/nospec-branch.h>
41#include <asm/fsgsbase.h>
42#include <linux/err.h>
43
44#include "calling.h"
45
46.code64
47.section .entry.text, "ax"
48
49#ifdef CONFIG_PARAVIRT
50ENTRY(native_usergs_sysret64)
51	UNWIND_HINT_EMPTY
52	swapgs
53	sysretq
54END(native_usergs_sysret64)
55#endif /* CONFIG_PARAVIRT */
56
57.macro TRACE_IRQS_FLAGS flags:req
58#ifdef CONFIG_TRACE_IRQFLAGS
59	btl	$9, \flags		/* interrupts off? */
60	jnc	1f
61	TRACE_IRQS_ON
621:
63#endif
64.endm
65
66.macro TRACE_IRQS_IRETQ
67	TRACE_IRQS_FLAGS EFLAGS(%rsp)
68.endm
69
70/*
71 * When dynamic function tracer is enabled it will add a breakpoint
72 * to all locations that it is about to modify, sync CPUs, update
73 * all the code, sync CPUs, then remove the breakpoints. In this time
74 * if lockdep is enabled, it might jump back into the debug handler
75 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF).
76 *
77 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to
78 * make sure the stack pointer does not get reset back to the top
79 * of the debug stack, and instead just reuses the current stack.
80 */
81#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS)
82
83.macro TRACE_IRQS_OFF_DEBUG
84	call	debug_stack_set_zero
85	TRACE_IRQS_OFF
86	call	debug_stack_reset
87.endm
88
89.macro TRACE_IRQS_ON_DEBUG
90	call	debug_stack_set_zero
91	TRACE_IRQS_ON
92	call	debug_stack_reset
93.endm
94
95.macro TRACE_IRQS_IRETQ_DEBUG
96	btl	$9, EFLAGS(%rsp)		/* interrupts off? */
97	jnc	1f
98	TRACE_IRQS_ON_DEBUG
991:
100.endm
101
102#else
103# define TRACE_IRQS_OFF_DEBUG			TRACE_IRQS_OFF
104# define TRACE_IRQS_ON_DEBUG			TRACE_IRQS_ON
105# define TRACE_IRQS_IRETQ_DEBUG			TRACE_IRQS_IRETQ
106#endif
107
108/*
109 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
110 *
111 * This is the only entry point used for 64-bit system calls.  The
112 * hardware interface is reasonably well designed and the register to
113 * argument mapping Linux uses fits well with the registers that are
114 * available when SYSCALL is used.
115 *
116 * SYSCALL instructions can be found inlined in libc implementations as
117 * well as some other programs and libraries.  There are also a handful
118 * of SYSCALL instructions in the vDSO used, for example, as a
119 * clock_gettimeofday fallback.
120 *
121 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
122 * then loads new ss, cs, and rip from previously programmed MSRs.
123 * rflags gets masked by a value from another MSR (so CLD and CLAC
124 * are not needed). SYSCALL does not save anything on the stack
125 * and does not change rsp.
126 *
127 * Registers on entry:
128 * rax  system call number
129 * rcx  return address
130 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
131 * rdi  arg0
132 * rsi  arg1
133 * rdx  arg2
134 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
135 * r8   arg4
136 * r9   arg5
137 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
138 *
139 * Only called from user space.
140 *
141 * When user can change pt_regs->foo always force IRET. That is because
142 * it deals with uncanonical addresses better. SYSRET has trouble
143 * with them due to bugs in both AMD and Intel CPUs.
144 */
145
146ENTRY(entry_SYSCALL_64)
147	UNWIND_HINT_EMPTY
148	/*
149	 * Interrupts are off on entry.
150	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
151	 * it is too small to ever cause noticeable irq latency.
152	 */
153
154	swapgs
155	/* tss.sp2 is scratch space. */
156	movq	%rsp, PER_CPU_VAR(cpu_tss_rw + TSS_sp2)
157	SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
158	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
159
160	/* Construct struct pt_regs on stack */
161	pushq	$__USER_DS				/* pt_regs->ss */
162	pushq	PER_CPU_VAR(cpu_tss_rw + TSS_sp2)	/* pt_regs->sp */
163	pushq	%r11					/* pt_regs->flags */
164	pushq	$__USER_CS				/* pt_regs->cs */
165	pushq	%rcx					/* pt_regs->ip */
166GLOBAL(entry_SYSCALL_64_after_hwframe)
167	pushq	%rax					/* pt_regs->orig_ax */
168
169	PUSH_AND_CLEAR_REGS rax=$-ENOSYS
170
171	TRACE_IRQS_OFF
172
173	/* IRQs are off. */
174	movq	%rax, %rdi
175	movq	%rsp, %rsi
176	call	do_syscall_64		/* returns with IRQs disabled */
177
178	TRACE_IRQS_IRETQ		/* we're about to change IF */
179
180	/*
181	 * Try to use SYSRET instead of IRET if we're returning to
182	 * a completely clean 64-bit userspace context.  If we're not,
183	 * go to the slow exit path.
184	 */
185	movq	RCX(%rsp), %rcx
186	movq	RIP(%rsp), %r11
187
188	cmpq	%rcx, %r11	/* SYSRET requires RCX == RIP */
189	jne	swapgs_restore_regs_and_return_to_usermode
190
191	/*
192	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
193	 * in kernel space.  This essentially lets the user take over
194	 * the kernel, since userspace controls RSP.
195	 *
196	 * If width of "canonical tail" ever becomes variable, this will need
197	 * to be updated to remain correct on both old and new CPUs.
198	 *
199	 * Change top bits to match most significant bit (47th or 56th bit
200	 * depending on paging mode) in the address.
201	 */
202#ifdef CONFIG_X86_5LEVEL
203	ALTERNATIVE "shl $(64 - 48), %rcx; sar $(64 - 48), %rcx", \
204		"shl $(64 - 57), %rcx; sar $(64 - 57), %rcx", X86_FEATURE_LA57
205#else
206	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
207	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
208#endif
209
210	/* If this changed %rcx, it was not canonical */
211	cmpq	%rcx, %r11
212	jne	swapgs_restore_regs_and_return_to_usermode
213
214	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
215	jne	swapgs_restore_regs_and_return_to_usermode
216
217	movq	R11(%rsp), %r11
218	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
219	jne	swapgs_restore_regs_and_return_to_usermode
220
221	/*
222	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
223	 * restore RF properly. If the slowpath sets it for whatever reason, we
224	 * need to restore it correctly.
225	 *
226	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
227	 * trap from userspace immediately after SYSRET.  This would cause an
228	 * infinite loop whenever #DB happens with register state that satisfies
229	 * the opportunistic SYSRET conditions.  For example, single-stepping
230	 * this user code:
231	 *
232	 *           movq	$stuck_here, %rcx
233	 *           pushfq
234	 *           popq %r11
235	 *   stuck_here:
236	 *
237	 * would never get past 'stuck_here'.
238	 */
239	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
240	jnz	swapgs_restore_regs_and_return_to_usermode
241
242	/* nothing to check for RSP */
243
244	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
245	jne	swapgs_restore_regs_and_return_to_usermode
246
247	/*
248	 * We win! This label is here just for ease of understanding
249	 * perf profiles. Nothing jumps here.
250	 */
251syscall_return_via_sysret:
252	/* rcx and r11 are already restored (see code above) */
253	UNWIND_HINT_EMPTY
254	POP_REGS pop_rdi=0 skip_r11rcx=1
255
256	/*
257	 * Now all regs are restored except RSP and RDI.
258	 * Save old stack pointer and switch to trampoline stack.
259	 */
260	movq	%rsp, %rdi
261	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
262
263	pushq	RSP-RDI(%rdi)	/* RSP */
264	pushq	(%rdi)		/* RDI */
265
266	/*
267	 * We are on the trampoline stack.  All regs except RDI are live.
268	 * We can do future final exit work right here.
269	 */
270	STACKLEAK_ERASE_NOCLOBBER
271
272	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
273
274	popq	%rdi
275	popq	%rsp
276	USERGS_SYSRET64
277END(entry_SYSCALL_64)
278
279/*
280 * %rdi: prev task
281 * %rsi: next task
282 */
283ENTRY(__switch_to_asm)
284	UNWIND_HINT_FUNC
285	/*
286	 * Save callee-saved registers
287	 * This must match the order in inactive_task_frame
288	 */
289	pushq	%rbp
290	pushq	%rbx
291	pushq	%r12
292	pushq	%r13
293	pushq	%r14
294	pushq	%r15
295
296	/* switch stack */
297	movq	%rsp, TASK_threadsp(%rdi)
298	movq	TASK_threadsp(%rsi), %rsp
299
300#ifdef CONFIG_STACKPROTECTOR
301	movq	TASK_stack_canary(%rsi), %rbx
302	movq	%rbx, PER_CPU_VAR(fixed_percpu_data) + stack_canary_offset
303#endif
304
305#ifdef CONFIG_RETPOLINE
306	/*
307	 * When switching from a shallower to a deeper call stack
308	 * the RSB may either underflow or use entries populated
309	 * with userspace addresses. On CPUs where those concerns
310	 * exist, overwrite the RSB with entries which capture
311	 * speculative execution to prevent attack.
312	 */
313	FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW
314#endif
315
316	/* restore callee-saved registers */
317	popq	%r15
318	popq	%r14
319	popq	%r13
320	popq	%r12
321	popq	%rbx
322	popq	%rbp
323
324	jmp	__switch_to
325END(__switch_to_asm)
326
327/*
328 * A newly forked process directly context switches into this address.
329 *
330 * rax: prev task we switched from
331 * rbx: kernel thread func (NULL for user thread)
332 * r12: kernel thread arg
333 */
334ENTRY(ret_from_fork)
335	UNWIND_HINT_EMPTY
336	movq	%rax, %rdi
337	call	schedule_tail			/* rdi: 'prev' task parameter */
338
339	testq	%rbx, %rbx			/* from kernel_thread? */
340	jnz	1f				/* kernel threads are uncommon */
341
3422:
343	UNWIND_HINT_REGS
344	movq	%rsp, %rdi
345	call	syscall_return_slowpath	/* returns with IRQs disabled */
346	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
347	jmp	swapgs_restore_regs_and_return_to_usermode
348
3491:
350	/* kernel thread */
351	UNWIND_HINT_EMPTY
352	movq	%r12, %rdi
353	CALL_NOSPEC %rbx
354	/*
355	 * A kernel thread is allowed to return here after successfully
356	 * calling do_execve().  Exit to userspace to complete the execve()
357	 * syscall.
358	 */
359	movq	$0, RAX(%rsp)
360	jmp	2b
361END(ret_from_fork)
362
363/*
364 * Build the entry stubs with some assembler magic.
365 * We pack 1 stub into every 8-byte block.
366 */
367	.align 8
368ENTRY(irq_entries_start)
369    vector=FIRST_EXTERNAL_VECTOR
370    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
371	UNWIND_HINT_IRET_REGS
372	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
373	jmp	common_interrupt
374	.align	8
375	vector=vector+1
376    .endr
377END(irq_entries_start)
378
379.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
380#ifdef CONFIG_DEBUG_ENTRY
381	pushq %rax
382	SAVE_FLAGS(CLBR_RAX)
383	testl $X86_EFLAGS_IF, %eax
384	jz .Lokay_\@
385	ud2
386.Lokay_\@:
387	popq %rax
388#endif
389.endm
390
391/*
392 * Enters the IRQ stack if we're not already using it.  NMI-safe.  Clobbers
393 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone.
394 * Requires kernel GSBASE.
395 *
396 * The invariant is that, if irq_count != -1, then the IRQ stack is in use.
397 */
398.macro ENTER_IRQ_STACK regs=1 old_rsp save_ret=0
399	DEBUG_ENTRY_ASSERT_IRQS_OFF
400
401	.if \save_ret
402	/*
403	 * If save_ret is set, the original stack contains one additional
404	 * entry -- the return address. Therefore, move the address one
405	 * entry below %rsp to \old_rsp.
406	 */
407	leaq	8(%rsp), \old_rsp
408	.else
409	movq	%rsp, \old_rsp
410	.endif
411
412	.if \regs
413	UNWIND_HINT_REGS base=\old_rsp
414	.endif
415
416	incl	PER_CPU_VAR(irq_count)
417	jnz	.Lirq_stack_push_old_rsp_\@
418
419	/*
420	 * Right now, if we just incremented irq_count to zero, we've
421	 * claimed the IRQ stack but we haven't switched to it yet.
422	 *
423	 * If anything is added that can interrupt us here without using IST,
424	 * it must be *extremely* careful to limit its stack usage.  This
425	 * could include kprobes and a hypothetical future IST-less #DB
426	 * handler.
427	 *
428	 * The OOPS unwinder relies on the word at the top of the IRQ
429	 * stack linking back to the previous RSP for the entire time we're
430	 * on the IRQ stack.  For this to work reliably, we need to write
431	 * it before we actually move ourselves to the IRQ stack.
432	 */
433
434	movq	\old_rsp, PER_CPU_VAR(irq_stack_backing_store + IRQ_STACK_SIZE - 8)
435	movq	PER_CPU_VAR(hardirq_stack_ptr), %rsp
436
437#ifdef CONFIG_DEBUG_ENTRY
438	/*
439	 * If the first movq above becomes wrong due to IRQ stack layout
440	 * changes, the only way we'll notice is if we try to unwind right
441	 * here.  Assert that we set up the stack right to catch this type
442	 * of bug quickly.
443	 */
444	cmpq	-8(%rsp), \old_rsp
445	je	.Lirq_stack_okay\@
446	ud2
447	.Lirq_stack_okay\@:
448#endif
449
450.Lirq_stack_push_old_rsp_\@:
451	pushq	\old_rsp
452
453	.if \regs
454	UNWIND_HINT_REGS indirect=1
455	.endif
456
457	.if \save_ret
458	/*
459	 * Push the return address to the stack. This return address can
460	 * be found at the "real" original RSP, which was offset by 8 at
461	 * the beginning of this macro.
462	 */
463	pushq	-8(\old_rsp)
464	.endif
465.endm
466
467/*
468 * Undoes ENTER_IRQ_STACK.
469 */
470.macro LEAVE_IRQ_STACK regs=1
471	DEBUG_ENTRY_ASSERT_IRQS_OFF
472	/* We need to be off the IRQ stack before decrementing irq_count. */
473	popq	%rsp
474
475	.if \regs
476	UNWIND_HINT_REGS
477	.endif
478
479	/*
480	 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming
481	 * the irq stack but we're not on it.
482	 */
483
484	decl	PER_CPU_VAR(irq_count)
485.endm
486
487/*
488 * Interrupt entry helper function.
489 *
490 * Entry runs with interrupts off. Stack layout at entry:
491 * +----------------------------------------------------+
492 * | regs->ss						|
493 * | regs->rsp						|
494 * | regs->eflags					|
495 * | regs->cs						|
496 * | regs->ip						|
497 * +----------------------------------------------------+
498 * | regs->orig_ax = ~(interrupt number)		|
499 * +----------------------------------------------------+
500 * | return address					|
501 * +----------------------------------------------------+
502 */
503ENTRY(interrupt_entry)
504	UNWIND_HINT_FUNC
505	ASM_CLAC
506	cld
507
508	testb	$3, CS-ORIG_RAX+8(%rsp)
509	jz	1f
510	SWAPGS
511
512	/*
513	 * Switch to the thread stack. The IRET frame and orig_ax are
514	 * on the stack, as well as the return address. RDI..R12 are
515	 * not (yet) on the stack and space has not (yet) been
516	 * allocated for them.
517	 */
518	pushq	%rdi
519
520	/* Need to switch before accessing the thread stack. */
521	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi
522	movq	%rsp, %rdi
523	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
524
525	 /*
526	  * We have RDI, return address, and orig_ax on the stack on
527	  * top of the IRET frame. That means offset=24
528	  */
529	UNWIND_HINT_IRET_REGS base=%rdi offset=24
530
531	pushq	7*8(%rdi)		/* regs->ss */
532	pushq	6*8(%rdi)		/* regs->rsp */
533	pushq	5*8(%rdi)		/* regs->eflags */
534	pushq	4*8(%rdi)		/* regs->cs */
535	pushq	3*8(%rdi)		/* regs->ip */
536	pushq	2*8(%rdi)		/* regs->orig_ax */
537	pushq	8(%rdi)			/* return address */
538	UNWIND_HINT_FUNC
539
540	movq	(%rdi), %rdi
5411:
542
543	PUSH_AND_CLEAR_REGS save_ret=1
544	ENCODE_FRAME_POINTER 8
545
546	testb	$3, CS+8(%rsp)
547	jz	1f
548
549	/*
550	 * IRQ from user mode.
551	 *
552	 * We need to tell lockdep that IRQs are off.  We can't do this until
553	 * we fix gsbase, and we should do it before enter_from_user_mode
554	 * (which can take locks).  Since TRACE_IRQS_OFF is idempotent,
555	 * the simplest way to handle it is to just call it twice if
556	 * we enter from user mode.  There's no reason to optimize this since
557	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
558	 */
559	TRACE_IRQS_OFF
560
561	CALL_enter_from_user_mode
562
5631:
564	ENTER_IRQ_STACK old_rsp=%rdi save_ret=1
565	/* We entered an interrupt context - irqs are off: */
566	TRACE_IRQS_OFF
567
568	ret
569END(interrupt_entry)
570_ASM_NOKPROBE(interrupt_entry)
571
572
573/* Interrupt entry/exit. */
574
575	/*
576	 * The interrupt stubs push (~vector+0x80) onto the stack and
577	 * then jump to common_interrupt.
578	 */
579	.p2align CONFIG_X86_L1_CACHE_SHIFT
580common_interrupt:
581	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
582	call	interrupt_entry
583	UNWIND_HINT_REGS indirect=1
584	call	do_IRQ	/* rdi points to pt_regs */
585	/* 0(%rsp): old RSP */
586ret_from_intr:
587	DISABLE_INTERRUPTS(CLBR_ANY)
588	TRACE_IRQS_OFF
589
590	LEAVE_IRQ_STACK
591
592	testb	$3, CS(%rsp)
593	jz	retint_kernel
594
595	/* Interrupt came from user space */
596GLOBAL(retint_user)
597	mov	%rsp,%rdi
598	call	prepare_exit_to_usermode
599	TRACE_IRQS_IRETQ
600
601GLOBAL(swapgs_restore_regs_and_return_to_usermode)
602#ifdef CONFIG_DEBUG_ENTRY
603	/* Assert that pt_regs indicates user mode. */
604	testb	$3, CS(%rsp)
605	jnz	1f
606	ud2
6071:
608#endif
609	POP_REGS pop_rdi=0
610
611	/*
612	 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS.
613	 * Save old stack pointer and switch to trampoline stack.
614	 */
615	movq	%rsp, %rdi
616	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
617
618	/* Copy the IRET frame to the trampoline stack. */
619	pushq	6*8(%rdi)	/* SS */
620	pushq	5*8(%rdi)	/* RSP */
621	pushq	4*8(%rdi)	/* EFLAGS */
622	pushq	3*8(%rdi)	/* CS */
623	pushq	2*8(%rdi)	/* RIP */
624
625	/* Push user RDI on the trampoline stack. */
626	pushq	(%rdi)
627
628	/*
629	 * We are on the trampoline stack.  All regs except RDI are live.
630	 * We can do future final exit work right here.
631	 */
632	STACKLEAK_ERASE_NOCLOBBER
633
634	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
635
636	/* Restore RDI. */
637	popq	%rdi
638	SWAPGS
639	INTERRUPT_RETURN
640
641
642/* Returning to kernel space */
643retint_kernel:
644#ifdef CONFIG_PREEMPT
645	/* Interrupts are off */
646	/* Check if we need preemption */
647	btl	$9, EFLAGS(%rsp)		/* were interrupts off? */
648	jnc	1f
649	cmpl	$0, PER_CPU_VAR(__preempt_count)
650	jnz	1f
651	call	preempt_schedule_irq
6521:
653#endif
654	/*
655	 * The iretq could re-enable interrupts:
656	 */
657	TRACE_IRQS_IRETQ
658
659GLOBAL(restore_regs_and_return_to_kernel)
660#ifdef CONFIG_DEBUG_ENTRY
661	/* Assert that pt_regs indicates kernel mode. */
662	testb	$3, CS(%rsp)
663	jz	1f
664	ud2
6651:
666#endif
667	POP_REGS
668	addq	$8, %rsp	/* skip regs->orig_ax */
669	/*
670	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
671	 * when returning from IPI handler.
672	 */
673	INTERRUPT_RETURN
674
675ENTRY(native_iret)
676	UNWIND_HINT_IRET_REGS
677	/*
678	 * Are we returning to a stack segment from the LDT?  Note: in
679	 * 64-bit mode SS:RSP on the exception stack is always valid.
680	 */
681#ifdef CONFIG_X86_ESPFIX64
682	testb	$4, (SS-RIP)(%rsp)
683	jnz	native_irq_return_ldt
684#endif
685
686.global native_irq_return_iret
687native_irq_return_iret:
688	/*
689	 * This may fault.  Non-paranoid faults on return to userspace are
690	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
691	 * Double-faults due to espfix64 are handled in do_double_fault.
692	 * Other faults here are fatal.
693	 */
694	iretq
695
696#ifdef CONFIG_X86_ESPFIX64
697native_irq_return_ldt:
698	/*
699	 * We are running with user GSBASE.  All GPRs contain their user
700	 * values.  We have a percpu ESPFIX stack that is eight slots
701	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
702	 * of the ESPFIX stack.
703	 *
704	 * We clobber RAX and RDI in this code.  We stash RDI on the
705	 * normal stack and RAX on the ESPFIX stack.
706	 *
707	 * The ESPFIX stack layout we set up looks like this:
708	 *
709	 * --- top of ESPFIX stack ---
710	 * SS
711	 * RSP
712	 * RFLAGS
713	 * CS
714	 * RIP  <-- RSP points here when we're done
715	 * RAX  <-- espfix_waddr points here
716	 * --- bottom of ESPFIX stack ---
717	 */
718
719	pushq	%rdi				/* Stash user RDI */
720	SWAPGS					/* to kernel GS */
721	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi	/* to kernel CR3 */
722
723	movq	PER_CPU_VAR(espfix_waddr), %rdi
724	movq	%rax, (0*8)(%rdi)		/* user RAX */
725	movq	(1*8)(%rsp), %rax		/* user RIP */
726	movq	%rax, (1*8)(%rdi)
727	movq	(2*8)(%rsp), %rax		/* user CS */
728	movq	%rax, (2*8)(%rdi)
729	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
730	movq	%rax, (3*8)(%rdi)
731	movq	(5*8)(%rsp), %rax		/* user SS */
732	movq	%rax, (5*8)(%rdi)
733	movq	(4*8)(%rsp), %rax		/* user RSP */
734	movq	%rax, (4*8)(%rdi)
735	/* Now RAX == RSP. */
736
737	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
738
739	/*
740	 * espfix_stack[31:16] == 0.  The page tables are set up such that
741	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
742	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
743	 * the same page.  Set up RSP so that RSP[31:16] contains the
744	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
745	 * still points to an RO alias of the ESPFIX stack.
746	 */
747	orq	PER_CPU_VAR(espfix_stack), %rax
748
749	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
750	SWAPGS					/* to user GS */
751	popq	%rdi				/* Restore user RDI */
752
753	movq	%rax, %rsp
754	UNWIND_HINT_IRET_REGS offset=8
755
756	/*
757	 * At this point, we cannot write to the stack any more, but we can
758	 * still read.
759	 */
760	popq	%rax				/* Restore user RAX */
761
762	/*
763	 * RSP now points to an ordinary IRET frame, except that the page
764	 * is read-only and RSP[31:16] are preloaded with the userspace
765	 * values.  We can now IRET back to userspace.
766	 */
767	jmp	native_irq_return_iret
768#endif
769END(common_interrupt)
770_ASM_NOKPROBE(common_interrupt)
771
772/*
773 * APIC interrupts.
774 */
775.macro apicinterrupt3 num sym do_sym
776ENTRY(\sym)
777	UNWIND_HINT_IRET_REGS
778	pushq	$~(\num)
779.Lcommon_\sym:
780	call	interrupt_entry
781	UNWIND_HINT_REGS indirect=1
782	call	\do_sym	/* rdi points to pt_regs */
783	jmp	ret_from_intr
784END(\sym)
785_ASM_NOKPROBE(\sym)
786.endm
787
788/* Make sure APIC interrupt handlers end up in the irqentry section: */
789#define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
790#define POP_SECTION_IRQENTRY	.popsection
791
792.macro apicinterrupt num sym do_sym
793PUSH_SECTION_IRQENTRY
794apicinterrupt3 \num \sym \do_sym
795POP_SECTION_IRQENTRY
796.endm
797
798#ifdef CONFIG_SMP
799apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
800apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
801#endif
802
803#ifdef CONFIG_X86_UV
804apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
805#endif
806
807apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
808apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
809
810#ifdef CONFIG_HAVE_KVM
811apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
812apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
813apicinterrupt3 POSTED_INTR_NESTED_VECTOR	kvm_posted_intr_nested_ipi	smp_kvm_posted_intr_nested_ipi
814#endif
815
816#ifdef CONFIG_X86_MCE_THRESHOLD
817apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
818#endif
819
820#ifdef CONFIG_X86_MCE_AMD
821apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
822#endif
823
824#ifdef CONFIG_X86_THERMAL_VECTOR
825apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
826#endif
827
828#ifdef CONFIG_SMP
829apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
830apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
831apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
832#endif
833
834apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
835apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
836
837#ifdef CONFIG_IRQ_WORK
838apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
839#endif
840
841/*
842 * Exception entry points.
843 */
844#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss_rw) + (TSS_ist + (x) * 8)
845
846/**
847 * idtentry - Generate an IDT entry stub
848 * @sym:		Name of the generated entry point
849 * @do_sym: 		C function to be called
850 * @has_error_code: 	True if this IDT vector has an error code on the stack
851 * @paranoid: 		non-zero means that this vector may be invoked from
852 *			kernel mode with user GSBASE and/or user CR3.
853 *			2 is special -- see below.
854 * @shift_ist:		Set to an IST index if entries from kernel mode should
855 *             		decrement the IST stack so that nested entries get a
856 *			fresh stack.  (This is for #DB, which has a nasty habit
857 *             		of recursing.)
858 *
859 * idtentry generates an IDT stub that sets up a usable kernel context,
860 * creates struct pt_regs, and calls @do_sym.  The stub has the following
861 * special behaviors:
862 *
863 * On an entry from user mode, the stub switches from the trampoline or
864 * IST stack to the normal thread stack.  On an exit to user mode, the
865 * normal exit-to-usermode path is invoked.
866 *
867 * On an exit to kernel mode, if @paranoid == 0, we check for preemption,
868 * whereas we omit the preemption check if @paranoid != 0.  This is purely
869 * because the implementation is simpler this way.  The kernel only needs
870 * to check for asynchronous kernel preemption when IRQ handlers return.
871 *
872 * If @paranoid == 0, then the stub will handle IRET faults by pretending
873 * that the fault came from user mode.  It will handle gs_change faults by
874 * pretending that the fault happened with kernel GSBASE.  Since this handling
875 * is omitted for @paranoid != 0, the #GP, #SS, and #NP stubs must have
876 * @paranoid == 0.  This special handling will do the wrong thing for
877 * espfix-induced #DF on IRET, so #DF must not use @paranoid == 0.
878 *
879 * @paranoid == 2 is special: the stub will never switch stacks.  This is for
880 * #DF: if the thread stack is somehow unusable, we'll still get a useful OOPS.
881 */
882.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1 ist_offset=0 create_gap=0
883ENTRY(\sym)
884	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
885
886	/* Sanity check */
887	.if \shift_ist != -1 && \paranoid == 0
888	.error "using shift_ist requires paranoid=1"
889	.endif
890
891	ASM_CLAC
892
893	.if \has_error_code == 0
894	pushq	$-1				/* ORIG_RAX: no syscall to restart */
895	.endif
896
897	.if \paranoid == 1
898	testb	$3, CS-ORIG_RAX(%rsp)		/* If coming from userspace, switch stacks */
899	jnz	.Lfrom_usermode_switch_stack_\@
900	.endif
901
902	.if \create_gap == 1
903	/*
904	 * If coming from kernel space, create a 6-word gap to allow the
905	 * int3 handler to emulate a call instruction.
906	 */
907	testb	$3, CS-ORIG_RAX(%rsp)
908	jnz	.Lfrom_usermode_no_gap_\@
909	.rept	6
910	pushq	5*8(%rsp)
911	.endr
912	UNWIND_HINT_IRET_REGS offset=8
913.Lfrom_usermode_no_gap_\@:
914	.endif
915
916	.if \paranoid
917	call	paranoid_entry
918	.else
919	call	error_entry
920	.endif
921	UNWIND_HINT_REGS
922	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
923
924	.if \paranoid
925	.if \shift_ist != -1
926	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
927	.else
928	TRACE_IRQS_OFF
929	.endif
930	.endif
931
932	movq	%rsp, %rdi			/* pt_regs pointer */
933
934	.if \has_error_code
935	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
936	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
937	.else
938	xorl	%esi, %esi			/* no error code */
939	.endif
940
941	.if \shift_ist != -1
942	subq	$\ist_offset, CPU_TSS_IST(\shift_ist)
943	.endif
944
945	call	\do_sym
946
947	.if \shift_ist != -1
948	addq	$\ist_offset, CPU_TSS_IST(\shift_ist)
949	.endif
950
951	.if \paranoid
952	jmp	paranoid_exit
953	.else
954	jmp	error_exit
955	.endif
956
957	.if \paranoid == 1
958	/*
959	 * Entry from userspace.  Switch stacks and treat it
960	 * as a normal entry.  This means that paranoid handlers
961	 * run in real process context if user_mode(regs).
962	 */
963.Lfrom_usermode_switch_stack_\@:
964	call	error_entry
965
966	movq	%rsp, %rdi			/* pt_regs pointer */
967
968	.if \has_error_code
969	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
970	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
971	.else
972	xorl	%esi, %esi			/* no error code */
973	.endif
974
975	call	\do_sym
976
977	jmp	error_exit
978	.endif
979_ASM_NOKPROBE(\sym)
980END(\sym)
981.endm
982
983idtentry divide_error			do_divide_error			has_error_code=0
984idtentry overflow			do_overflow			has_error_code=0
985idtentry bounds				do_bounds			has_error_code=0
986idtentry invalid_op			do_invalid_op			has_error_code=0
987idtentry device_not_available		do_device_not_available		has_error_code=0
988idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
989idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
990idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
991idtentry segment_not_present		do_segment_not_present		has_error_code=1
992idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
993idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
994idtentry alignment_check		do_alignment_check		has_error_code=1
995idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
996
997
998	/*
999	 * Reload gs selector with exception handling
1000	 * edi:  new selector
1001	 */
1002ENTRY(native_load_gs_index)
1003	FRAME_BEGIN
1004	pushfq
1005	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
1006	TRACE_IRQS_OFF
1007	SWAPGS
1008.Lgs_change:
1009	movl	%edi, %gs
10102:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
1011	SWAPGS
1012	TRACE_IRQS_FLAGS (%rsp)
1013	popfq
1014	FRAME_END
1015	ret
1016ENDPROC(native_load_gs_index)
1017EXPORT_SYMBOL(native_load_gs_index)
1018
1019	_ASM_EXTABLE(.Lgs_change, bad_gs)
1020	.section .fixup, "ax"
1021	/* running with kernelgs */
1022bad_gs:
1023	SWAPGS					/* switch back to user gs */
1024.macro ZAP_GS
1025	/* This can't be a string because the preprocessor needs to see it. */
1026	movl $__USER_DS, %eax
1027	movl %eax, %gs
1028.endm
1029	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
1030	xorl	%eax, %eax
1031	movl	%eax, %gs
1032	jmp	2b
1033	.previous
1034
1035/* Call softirq on interrupt stack. Interrupts are off. */
1036ENTRY(do_softirq_own_stack)
1037	pushq	%rbp
1038	mov	%rsp, %rbp
1039	ENTER_IRQ_STACK regs=0 old_rsp=%r11
1040	call	__do_softirq
1041	LEAVE_IRQ_STACK regs=0
1042	leaveq
1043	ret
1044ENDPROC(do_softirq_own_stack)
1045
1046#ifdef CONFIG_XEN_PV
1047idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0
1048
1049/*
1050 * A note on the "critical region" in our callback handler.
1051 * We want to avoid stacking callback handlers due to events occurring
1052 * during handling of the last event. To do this, we keep events disabled
1053 * until we've done all processing. HOWEVER, we must enable events before
1054 * popping the stack frame (can't be done atomically) and so it would still
1055 * be possible to get enough handler activations to overflow the stack.
1056 * Although unlikely, bugs of that kind are hard to track down, so we'd
1057 * like to avoid the possibility.
1058 * So, on entry to the handler we detect whether we interrupted an
1059 * existing activation in its critical region -- if so, we pop the current
1060 * activation and restart the handler using the previous one.
1061 */
1062ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
1063
1064/*
1065 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
1066 * see the correct pointer to the pt_regs
1067 */
1068	UNWIND_HINT_FUNC
1069	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
1070	UNWIND_HINT_REGS
1071
1072	ENTER_IRQ_STACK old_rsp=%r10
1073	call	xen_evtchn_do_upcall
1074	LEAVE_IRQ_STACK
1075
1076#ifndef CONFIG_PREEMPT
1077	call	xen_maybe_preempt_hcall
1078#endif
1079	jmp	error_exit
1080END(xen_do_hypervisor_callback)
1081
1082/*
1083 * Hypervisor uses this for application faults while it executes.
1084 * We get here for two reasons:
1085 *  1. Fault while reloading DS, ES, FS or GS
1086 *  2. Fault while executing IRET
1087 * Category 1 we do not need to fix up as Xen has already reloaded all segment
1088 * registers that could be reloaded and zeroed the others.
1089 * Category 2 we fix up by killing the current process. We cannot use the
1090 * normal Linux return path in this case because if we use the IRET hypercall
1091 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
1092 * We distinguish between categories by comparing each saved segment register
1093 * with its current contents: any discrepancy means we in category 1.
1094 */
1095ENTRY(xen_failsafe_callback)
1096	UNWIND_HINT_EMPTY
1097	movl	%ds, %ecx
1098	cmpw	%cx, 0x10(%rsp)
1099	jne	1f
1100	movl	%es, %ecx
1101	cmpw	%cx, 0x18(%rsp)
1102	jne	1f
1103	movl	%fs, %ecx
1104	cmpw	%cx, 0x20(%rsp)
1105	jne	1f
1106	movl	%gs, %ecx
1107	cmpw	%cx, 0x28(%rsp)
1108	jne	1f
1109	/* All segments match their saved values => Category 2 (Bad IRET). */
1110	movq	(%rsp), %rcx
1111	movq	8(%rsp), %r11
1112	addq	$0x30, %rsp
1113	pushq	$0				/* RIP */
1114	UNWIND_HINT_IRET_REGS offset=8
1115	jmp	general_protection
11161:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
1117	movq	(%rsp), %rcx
1118	movq	8(%rsp), %r11
1119	addq	$0x30, %rsp
1120	UNWIND_HINT_IRET_REGS
1121	pushq	$-1 /* orig_ax = -1 => not a system call */
1122	PUSH_AND_CLEAR_REGS
1123	ENCODE_FRAME_POINTER
1124	jmp	error_exit
1125END(xen_failsafe_callback)
1126#endif /* CONFIG_XEN_PV */
1127
1128#ifdef CONFIG_XEN_PVHVM
1129apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1130	xen_hvm_callback_vector xen_evtchn_do_upcall
1131#endif
1132
1133
1134#if IS_ENABLED(CONFIG_HYPERV)
1135apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1136	hyperv_callback_vector hyperv_vector_handler
1137
1138apicinterrupt3 HYPERV_REENLIGHTENMENT_VECTOR \
1139	hyperv_reenlightenment_vector hyperv_reenlightenment_intr
1140
1141apicinterrupt3 HYPERV_STIMER0_VECTOR \
1142	hv_stimer0_callback_vector hv_stimer0_vector_handler
1143#endif /* CONFIG_HYPERV */
1144
1145idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=IST_INDEX_DB ist_offset=DB_STACK_OFFSET
1146idtentry int3			do_int3			has_error_code=0	create_gap=1
1147idtentry stack_segment		do_stack_segment	has_error_code=1
1148
1149#ifdef CONFIG_XEN_PV
1150idtentry xennmi			do_nmi			has_error_code=0
1151idtentry xendebug		do_debug		has_error_code=0
1152idtentry xenint3		do_int3			has_error_code=0
1153#endif
1154
1155idtentry general_protection	do_general_protection	has_error_code=1
1156idtentry page_fault		do_page_fault		has_error_code=1
1157
1158#ifdef CONFIG_KVM_GUEST
1159idtentry async_page_fault	do_async_page_fault	has_error_code=1
1160#endif
1161
1162#ifdef CONFIG_X86_MCE
1163idtentry machine_check		do_mce			has_error_code=0	paranoid=1
1164#endif
1165
1166/*
1167 * Save all registers in pt_regs. Return GSBASE related information
1168 * in EBX depending on the availability of the FSGSBASE instructions:
1169 *
1170 * FSGSBASE	R/EBX
1171 *     N        0 -> SWAPGS on exit
1172 *              1 -> no SWAPGS on exit
1173 *
1174 *     Y        GSBASE value at entry, must be restored in paranoid_exit
1175 */
1176ENTRY(paranoid_entry)
1177	UNWIND_HINT_FUNC
1178	cld
1179	PUSH_AND_CLEAR_REGS save_ret=1
1180	ENCODE_FRAME_POINTER 8
1181
1182	/*
1183	 * Always stash CR3 in %r14.  This value will be restored,
1184	 * verbatim, at exit.  Needed if paranoid_entry interrupted
1185	 * another entry that already switched to the user CR3 value
1186	 * but has not yet returned to userspace.
1187	 *
1188	 * This is also why CS (stashed in the "iret frame" by the
1189	 * hardware at entry) can not be used: this may be a return
1190	 * to kernel code, but with a user CR3 value.
1191	 *
1192	 * Switching CR3 does not depend on kernel GSBASE so it can
1193	 * be done before switching to the kernel GSBASE. This is
1194	 * required for FSGSBASE because the kernel GSBASE has to
1195	 * be retrieved from a kernel internal table.
1196	 */
1197	SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
1198
1199        /*
1200	 * Handling GSBASE depends on the availability of FSGSBASE.
1201	 *
1202	 * Without FSGSBASE the kernel enforces that negative GSBASE
1203	 * values indicate kernel GSBASE. With FSGSBASE no assumptions
1204	 * can be made about the GSBASE value when entering from user
1205	 * space.
1206	*/
1207	ALTERNATIVE "jmp .Lparanoid_entry_checkgs", "", X86_FEATURE_FSGSBASE
1208
1209	/*
1210	 * Read the current GSBASE and store it in in %rbx unconditionally,
1211	 * retrieve and set the current CPUs kernel GSBASE. The stored value
1212	 * has to be restored in paranoid_exit unconditionally.
1213	 */
1214	SAVE_AND_SET_GSBASE scratch_reg=%rax save_reg=%rbx
1215	ret
1216
1217.Lparanoid_entry_checkgs:
1218	/* EBX = 1 -> kernel GSBASE active, no restore required */
1219	movl	$1, %ebx
1220	/*
1221	 * The kernel-enforced convention is a negative GSBASE indicates
1222	 * a kernel value. No SWAPGS needed on entry and exit.
1223	 */
1224	movl	$MSR_GS_BASE, %ecx
1225	rdmsr
1226	testl	%edx, %edx
1227	jns	.Lparanoid_entry_swapgs
1228	ret
1229
1230.Lparanoid_entry_swapgs:
1231	SWAPGS
1232	/* EBX = 0 -> SWAPGS required on exit */
1233	xorl	%ebx, %ebx
1234	ret
1235END(paranoid_entry)
1236
1237/*
1238 * "Paranoid" exit path from exception stack.  This is invoked
1239 * only on return from non-NMI IST interrupts that came
1240 * from kernel space.
1241 *
1242 * We may be returning to very strange contexts (e.g. very early
1243 * in syscall entry), so checking for preemption here would
1244 * be complicated.  Fortunately, there's no good reason to try
1245 * to handle preemption here.
1246 *
1247 * R/EBX contains the GSBASE related information depending on the
1248 * availability of the FSGSBASE instructions:
1249 *
1250 * FSGSBASE	R/EBX
1251 *     N        0 -> SWAPGS on exit
1252 *              1 -> no SWAPGS on exit
1253 *
1254 *     Y        User space GSBASE, must be restored unconditionally
1255 */
1256ENTRY(paranoid_exit)
1257	UNWIND_HINT_REGS
1258	DISABLE_INTERRUPTS(CLBR_ANY)
1259	TRACE_IRQS_OFF_DEBUG
1260
1261	/* Handle GS depending on FSGSBASE availability */
1262	ALTERNATIVE "jmp .Lparanoid_exit_checkgs", "nop",X86_FEATURE_FSGSBASE
1263
1264	/* With FSGSBASE enabled, unconditionally restore GSBASE */
1265	wrgsbase	%rbx
1266	jmp	.Lparanoid_exit_no_swapgs;
1267
1268.Lparanoid_exit_checkgs:
1269	/* On non-FSGSBASE systems, conditionally do SWAPGS */
1270	testl	%ebx, %ebx
1271	jnz	.Lparanoid_exit_no_swapgs
1272	TRACE_IRQS_IRETQ
1273	/* Always restore stashed CR3 value (see paranoid_entry) */
1274	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1275	SWAPGS_UNSAFE_STACK
1276	jmp	.Lparanoid_exit_restore
1277
1278.Lparanoid_exit_no_swapgs:
1279	TRACE_IRQS_IRETQ_DEBUG
1280	/* Always restore stashed CR3 value (see paranoid_entry) */
1281	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1282
1283.Lparanoid_exit_restore:
1284	jmp	restore_regs_and_return_to_kernel
1285END(paranoid_exit)
1286
1287/*
1288 * Save all registers in pt_regs, and switch GS if needed.
1289 */
1290ENTRY(error_entry)
1291	UNWIND_HINT_FUNC
1292	cld
1293	PUSH_AND_CLEAR_REGS save_ret=1
1294	ENCODE_FRAME_POINTER 8
1295	testb	$3, CS+8(%rsp)
1296	jz	.Lerror_kernelspace
1297
1298	/*
1299	 * We entered from user mode or we're pretending to have entered
1300	 * from user mode due to an IRET fault.
1301	 */
1302	SWAPGS
1303	/* We have user CR3.  Change to kernel CR3. */
1304	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1305
1306.Lerror_entry_from_usermode_after_swapgs:
1307	/* Put us onto the real thread stack. */
1308	popq	%r12				/* save return addr in %12 */
1309	movq	%rsp, %rdi			/* arg0 = pt_regs pointer */
1310	call	sync_regs
1311	movq	%rax, %rsp			/* switch stack */
1312	ENCODE_FRAME_POINTER
1313	pushq	%r12
1314
1315	/*
1316	 * We need to tell lockdep that IRQs are off.  We can't do this until
1317	 * we fix gsbase, and we should do it before enter_from_user_mode
1318	 * (which can take locks).
1319	 */
1320	TRACE_IRQS_OFF
1321	CALL_enter_from_user_mode
1322	ret
1323
1324.Lerror_entry_done:
1325	TRACE_IRQS_OFF
1326	ret
1327
1328	/*
1329	 * There are two places in the kernel that can potentially fault with
1330	 * usergs. Handle them here.  B stepping K8s sometimes report a
1331	 * truncated RIP for IRET exceptions returning to compat mode. Check
1332	 * for these here too.
1333	 */
1334.Lerror_kernelspace:
1335	leaq	native_irq_return_iret(%rip), %rcx
1336	cmpq	%rcx, RIP+8(%rsp)
1337	je	.Lerror_bad_iret
1338	movl	%ecx, %eax			/* zero extend */
1339	cmpq	%rax, RIP+8(%rsp)
1340	je	.Lbstep_iret
1341	cmpq	$.Lgs_change, RIP+8(%rsp)
1342	jne	.Lerror_entry_done
1343
1344	/*
1345	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1346	 * gsbase and proceed.  We'll fix up the exception and land in
1347	 * .Lgs_change's error handler with kernel gsbase.
1348	 */
1349	SWAPGS
1350	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1351	jmp .Lerror_entry_done
1352
1353.Lbstep_iret:
1354	/* Fix truncated RIP */
1355	movq	%rcx, RIP+8(%rsp)
1356	/* fall through */
1357
1358.Lerror_bad_iret:
1359	/*
1360	 * We came from an IRET to user mode, so we have user
1361	 * gsbase and CR3.  Switch to kernel gsbase and CR3:
1362	 */
1363	SWAPGS
1364	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1365
1366	/*
1367	 * Pretend that the exception came from user mode: set up pt_regs
1368	 * as if we faulted immediately after IRET.
1369	 */
1370	mov	%rsp, %rdi
1371	call	fixup_bad_iret
1372	mov	%rax, %rsp
1373	jmp	.Lerror_entry_from_usermode_after_swapgs
1374END(error_entry)
1375
1376ENTRY(error_exit)
1377	UNWIND_HINT_REGS
1378	DISABLE_INTERRUPTS(CLBR_ANY)
1379	TRACE_IRQS_OFF
1380	testb	$3, CS(%rsp)
1381	jz	retint_kernel
1382	jmp	retint_user
1383END(error_exit)
1384
1385/*
1386 * Runs on exception stack.  Xen PV does not go through this path at all,
1387 * so we can use real assembly here.
1388 *
1389 * Registers:
1390 *	%r14: Used to save/restore the CR3 of the interrupted context
1391 *	      when PAGE_TABLE_ISOLATION is in use.  Do not clobber.
1392 */
1393ENTRY(nmi)
1394	UNWIND_HINT_IRET_REGS
1395
1396	/*
1397	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1398	 * the iretq it performs will take us out of NMI context.
1399	 * This means that we can have nested NMIs where the next
1400	 * NMI is using the top of the stack of the previous NMI. We
1401	 * can't let it execute because the nested NMI will corrupt the
1402	 * stack of the previous NMI. NMI handlers are not re-entrant
1403	 * anyway.
1404	 *
1405	 * To handle this case we do the following:
1406	 *  Check the a special location on the stack that contains
1407	 *  a variable that is set when NMIs are executing.
1408	 *  The interrupted task's stack is also checked to see if it
1409	 *  is an NMI stack.
1410	 *  If the variable is not set and the stack is not the NMI
1411	 *  stack then:
1412	 *    o Set the special variable on the stack
1413	 *    o Copy the interrupt frame into an "outermost" location on the
1414	 *      stack
1415	 *    o Copy the interrupt frame into an "iret" location on the stack
1416	 *    o Continue processing the NMI
1417	 *  If the variable is set or the previous stack is the NMI stack:
1418	 *    o Modify the "iret" location to jump to the repeat_nmi
1419	 *    o return back to the first NMI
1420	 *
1421	 * Now on exit of the first NMI, we first clear the stack variable
1422	 * The NMI stack will tell any nested NMIs at that point that it is
1423	 * nested. Then we pop the stack normally with iret, and if there was
1424	 * a nested NMI that updated the copy interrupt stack frame, a
1425	 * jump will be made to the repeat_nmi code that will handle the second
1426	 * NMI.
1427	 *
1428	 * However, espfix prevents us from directly returning to userspace
1429	 * with a single IRET instruction.  Similarly, IRET to user mode
1430	 * can fault.  We therefore handle NMIs from user space like
1431	 * other IST entries.
1432	 */
1433
1434	ASM_CLAC
1435
1436	/* Use %rdx as our temp variable throughout */
1437	pushq	%rdx
1438
1439	testb	$3, CS-RIP+8(%rsp)
1440	jz	.Lnmi_from_kernel
1441
1442	/*
1443	 * NMI from user mode.  We need to run on the thread stack, but we
1444	 * can't go through the normal entry paths: NMIs are masked, and
1445	 * we don't want to enable interrupts, because then we'll end
1446	 * up in an awkward situation in which IRQs are on but NMIs
1447	 * are off.
1448	 *
1449	 * We also must not push anything to the stack before switching
1450	 * stacks lest we corrupt the "NMI executing" variable.
1451	 */
1452
1453	swapgs
1454	cld
1455	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx
1456	movq	%rsp, %rdx
1457	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1458	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1459	pushq	5*8(%rdx)	/* pt_regs->ss */
1460	pushq	4*8(%rdx)	/* pt_regs->rsp */
1461	pushq	3*8(%rdx)	/* pt_regs->flags */
1462	pushq	2*8(%rdx)	/* pt_regs->cs */
1463	pushq	1*8(%rdx)	/* pt_regs->rip */
1464	UNWIND_HINT_IRET_REGS
1465	pushq   $-1		/* pt_regs->orig_ax */
1466	PUSH_AND_CLEAR_REGS rdx=(%rdx)
1467	ENCODE_FRAME_POINTER
1468
1469	/*
1470	 * At this point we no longer need to worry about stack damage
1471	 * due to nesting -- we're on the normal thread stack and we're
1472	 * done with the NMI stack.
1473	 */
1474
1475	movq	%rsp, %rdi
1476	movq	$-1, %rsi
1477	call	do_nmi
1478
1479	/*
1480	 * Return back to user mode.  We must *not* do the normal exit
1481	 * work, because we don't want to enable interrupts.
1482	 */
1483	jmp	swapgs_restore_regs_and_return_to_usermode
1484
1485.Lnmi_from_kernel:
1486	/*
1487	 * Here's what our stack frame will look like:
1488	 * +---------------------------------------------------------+
1489	 * | original SS                                             |
1490	 * | original Return RSP                                     |
1491	 * | original RFLAGS                                         |
1492	 * | original CS                                             |
1493	 * | original RIP                                            |
1494	 * +---------------------------------------------------------+
1495	 * | temp storage for rdx                                    |
1496	 * +---------------------------------------------------------+
1497	 * | "NMI executing" variable                                |
1498	 * +---------------------------------------------------------+
1499	 * | iret SS          } Copied from "outermost" frame        |
1500	 * | iret Return RSP  } on each loop iteration; overwritten  |
1501	 * | iret RFLAGS      } by a nested NMI to force another     |
1502	 * | iret CS          } iteration if needed.                 |
1503	 * | iret RIP         }                                      |
1504	 * +---------------------------------------------------------+
1505	 * | outermost SS          } initialized in first_nmi;       |
1506	 * | outermost Return RSP  } will not be changed before      |
1507	 * | outermost RFLAGS      } NMI processing is done.         |
1508	 * | outermost CS          } Copied to "iret" frame on each  |
1509	 * | outermost RIP         } iteration.                      |
1510	 * +---------------------------------------------------------+
1511	 * | pt_regs                                                 |
1512	 * +---------------------------------------------------------+
1513	 *
1514	 * The "original" frame is used by hardware.  Before re-enabling
1515	 * NMIs, we need to be done with it, and we need to leave enough
1516	 * space for the asm code here.
1517	 *
1518	 * We return by executing IRET while RSP points to the "iret" frame.
1519	 * That will either return for real or it will loop back into NMI
1520	 * processing.
1521	 *
1522	 * The "outermost" frame is copied to the "iret" frame on each
1523	 * iteration of the loop, so each iteration starts with the "iret"
1524	 * frame pointing to the final return target.
1525	 */
1526
1527	/*
1528	 * Determine whether we're a nested NMI.
1529	 *
1530	 * If we interrupted kernel code between repeat_nmi and
1531	 * end_repeat_nmi, then we are a nested NMI.  We must not
1532	 * modify the "iret" frame because it's being written by
1533	 * the outer NMI.  That's okay; the outer NMI handler is
1534	 * about to about to call do_nmi anyway, so we can just
1535	 * resume the outer NMI.
1536	 */
1537
1538	movq	$repeat_nmi, %rdx
1539	cmpq	8(%rsp), %rdx
1540	ja	1f
1541	movq	$end_repeat_nmi, %rdx
1542	cmpq	8(%rsp), %rdx
1543	ja	nested_nmi_out
15441:
1545
1546	/*
1547	 * Now check "NMI executing".  If it's set, then we're nested.
1548	 * This will not detect if we interrupted an outer NMI just
1549	 * before IRET.
1550	 */
1551	cmpl	$1, -8(%rsp)
1552	je	nested_nmi
1553
1554	/*
1555	 * Now test if the previous stack was an NMI stack.  This covers
1556	 * the case where we interrupt an outer NMI after it clears
1557	 * "NMI executing" but before IRET.  We need to be careful, though:
1558	 * there is one case in which RSP could point to the NMI stack
1559	 * despite there being no NMI active: naughty userspace controls
1560	 * RSP at the very beginning of the SYSCALL targets.  We can
1561	 * pull a fast one on naughty userspace, though: we program
1562	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1563	 * if it controls the kernel's RSP.  We set DF before we clear
1564	 * "NMI executing".
1565	 */
1566	lea	6*8(%rsp), %rdx
1567	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1568	cmpq	%rdx, 4*8(%rsp)
1569	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1570	ja	first_nmi
1571
1572	subq	$EXCEPTION_STKSZ, %rdx
1573	cmpq	%rdx, 4*8(%rsp)
1574	/* If it is below the NMI stack, it is a normal NMI */
1575	jb	first_nmi
1576
1577	/* Ah, it is within the NMI stack. */
1578
1579	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1580	jz	first_nmi	/* RSP was user controlled. */
1581
1582	/* This is a nested NMI. */
1583
1584nested_nmi:
1585	/*
1586	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1587	 * iteration of NMI handling.
1588	 */
1589	subq	$8, %rsp
1590	leaq	-10*8(%rsp), %rdx
1591	pushq	$__KERNEL_DS
1592	pushq	%rdx
1593	pushfq
1594	pushq	$__KERNEL_CS
1595	pushq	$repeat_nmi
1596
1597	/* Put stack back */
1598	addq	$(6*8), %rsp
1599
1600nested_nmi_out:
1601	popq	%rdx
1602
1603	/* We are returning to kernel mode, so this cannot result in a fault. */
1604	iretq
1605
1606first_nmi:
1607	/* Restore rdx. */
1608	movq	(%rsp), %rdx
1609
1610	/* Make room for "NMI executing". */
1611	pushq	$0
1612
1613	/* Leave room for the "iret" frame */
1614	subq	$(5*8), %rsp
1615
1616	/* Copy the "original" frame to the "outermost" frame */
1617	.rept 5
1618	pushq	11*8(%rsp)
1619	.endr
1620	UNWIND_HINT_IRET_REGS
1621
1622	/* Everything up to here is safe from nested NMIs */
1623
1624#ifdef CONFIG_DEBUG_ENTRY
1625	/*
1626	 * For ease of testing, unmask NMIs right away.  Disabled by
1627	 * default because IRET is very expensive.
1628	 */
1629	pushq	$0		/* SS */
1630	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1631	addq	$8, (%rsp)	/* Fix up RSP */
1632	pushfq			/* RFLAGS */
1633	pushq	$__KERNEL_CS	/* CS */
1634	pushq	$1f		/* RIP */
1635	iretq			/* continues at repeat_nmi below */
1636	UNWIND_HINT_IRET_REGS
16371:
1638#endif
1639
1640repeat_nmi:
1641	/*
1642	 * If there was a nested NMI, the first NMI's iret will return
1643	 * here. But NMIs are still enabled and we can take another
1644	 * nested NMI. The nested NMI checks the interrupted RIP to see
1645	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1646	 * it will just return, as we are about to repeat an NMI anyway.
1647	 * This makes it safe to copy to the stack frame that a nested
1648	 * NMI will update.
1649	 *
1650	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1651	 * we're repeating an NMI, gsbase has the same value that it had on
1652	 * the first iteration.  paranoid_entry will load the kernel
1653	 * gsbase if needed before we call do_nmi.  "NMI executing"
1654	 * is zero.
1655	 */
1656	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1657
1658	/*
1659	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1660	 * here must not modify the "iret" frame while we're writing to
1661	 * it or it will end up containing garbage.
1662	 */
1663	addq	$(10*8), %rsp
1664	.rept 5
1665	pushq	-6*8(%rsp)
1666	.endr
1667	subq	$(5*8), %rsp
1668end_repeat_nmi:
1669
1670	/*
1671	 * Everything below this point can be preempted by a nested NMI.
1672	 * If this happens, then the inner NMI will change the "iret"
1673	 * frame to point back to repeat_nmi.
1674	 */
1675	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1676
1677	/*
1678	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1679	 * as we should not be calling schedule in NMI context.
1680	 * Even with normal interrupts enabled. An NMI should not be
1681	 * setting NEED_RESCHED or anything that normal interrupts and
1682	 * exceptions might do.
1683	 */
1684	call	paranoid_entry
1685	UNWIND_HINT_REGS
1686
1687	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1688	movq	%rsp, %rdi
1689	movq	$-1, %rsi
1690	call	do_nmi
1691
1692	/* Always restore stashed CR3 value (see paranoid_entry) */
1693	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
1694
1695	/*
1696	 * The above invocation of paranoid_entry stored the GSBASE
1697	 * related information in R/EBX depending on the availability
1698	 * of FSGSBASE.
1699	 *
1700	 * If FSGSBASE is enabled, restore the saved GSBASE value
1701	 * unconditionally, otherwise take the conditional SWAPGS path.
1702	 */
1703	ALTERNATIVE "jmp nmi_no_fsgsbase", "", X86_FEATURE_FSGSBASE
1704
1705	wrgsbase	%rbx
1706	jmp	nmi_restore
1707
1708nmi_no_fsgsbase:
1709	/* EBX == 0 -> invoke SWAPGS */
1710	testl	%ebx, %ebx
1711	jnz	nmi_restore
1712
1713nmi_swapgs:
1714	SWAPGS_UNSAFE_STACK
1715
1716nmi_restore:
1717	POP_REGS
1718
1719	/*
1720	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1721	 * at the "iret" frame.
1722	 */
1723	addq	$6*8, %rsp
1724
1725	/*
1726	 * Clear "NMI executing".  Set DF first so that we can easily
1727	 * distinguish the remaining code between here and IRET from
1728	 * the SYSCALL entry and exit paths.
1729	 *
1730	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1731	 * this code when I had the misapprehension that Xen PV supported
1732	 * NMIs, and Xen PV would break that approach.
1733	 */
1734	std
1735	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1736
1737	/*
1738	 * iretq reads the "iret" frame and exits the NMI stack in a
1739	 * single instruction.  We are returning to kernel mode, so this
1740	 * cannot result in a fault.  Similarly, we don't need to worry
1741	 * about espfix64 on the way back to kernel mode.
1742	 */
1743	iretq
1744END(nmi)
1745
1746ENTRY(ignore_sysret)
1747	UNWIND_HINT_EMPTY
1748	mov	$-ENOSYS, %eax
1749	sysret
1750END(ignore_sysret)
1751
1752ENTRY(rewind_stack_do_exit)
1753	UNWIND_HINT_FUNC
1754	/* Prevent any naive code from trying to unwind to our caller. */
1755	xorl	%ebp, %ebp
1756
1757	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1758	leaq	-PTREGS_SIZE(%rax), %rsp
1759	UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE
1760
1761	call	do_exit
1762END(rewind_stack_do_exit)
1763