xref: /linux/arch/x86/entry/entry_64.S (revision 2209fda323e2fd2a2d0885595fd5097717f8d2aa)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 *  linux/arch/x86_64/entry.S
4 *
5 *  Copyright (C) 1991, 1992  Linus Torvalds
6 *  Copyright (C) 2000, 2001, 2002  Andi Kleen SuSE Labs
7 *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
8 *
9 * entry.S contains the system-call and fault low-level handling routines.
10 *
11 * Some of this is documented in Documentation/x86/entry_64.txt
12 *
13 * A note on terminology:
14 * - iret frame:	Architecture defined interrupt frame from SS to RIP
15 *			at the top of the kernel process stack.
16 *
17 * Some macro usage:
18 * - ENTRY/END:		Define functions in the symbol table.
19 * - TRACE_IRQ_*:	Trace hardirq state for lock debugging.
20 * - idtentry:		Define exception entry points.
21 */
22#include <linux/linkage.h>
23#include <asm/segment.h>
24#include <asm/cache.h>
25#include <asm/errno.h>
26#include <asm/asm-offsets.h>
27#include <asm/msr.h>
28#include <asm/unistd.h>
29#include <asm/thread_info.h>
30#include <asm/hw_irq.h>
31#include <asm/page_types.h>
32#include <asm/irqflags.h>
33#include <asm/paravirt.h>
34#include <asm/percpu.h>
35#include <asm/asm.h>
36#include <asm/smap.h>
37#include <asm/pgtable_types.h>
38#include <asm/export.h>
39#include <asm/frame.h>
40#include <asm/nospec-branch.h>
41#include <linux/err.h>
42
43#include "calling.h"
44
45.code64
46.section .entry.text, "ax"
47
48#ifdef CONFIG_PARAVIRT
49ENTRY(native_usergs_sysret64)
50	UNWIND_HINT_EMPTY
51	swapgs
52	sysretq
53END(native_usergs_sysret64)
54#endif /* CONFIG_PARAVIRT */
55
56.macro TRACE_IRQS_FLAGS flags:req
57#ifdef CONFIG_TRACE_IRQFLAGS
58	btl	$9, \flags		/* interrupts off? */
59	jnc	1f
60	TRACE_IRQS_ON
611:
62#endif
63.endm
64
65.macro TRACE_IRQS_IRETQ
66	TRACE_IRQS_FLAGS EFLAGS(%rsp)
67.endm
68
69/*
70 * When dynamic function tracer is enabled it will add a breakpoint
71 * to all locations that it is about to modify, sync CPUs, update
72 * all the code, sync CPUs, then remove the breakpoints. In this time
73 * if lockdep is enabled, it might jump back into the debug handler
74 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF).
75 *
76 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to
77 * make sure the stack pointer does not get reset back to the top
78 * of the debug stack, and instead just reuses the current stack.
79 */
80#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS)
81
82.macro TRACE_IRQS_OFF_DEBUG
83	call	debug_stack_set_zero
84	TRACE_IRQS_OFF
85	call	debug_stack_reset
86.endm
87
88.macro TRACE_IRQS_ON_DEBUG
89	call	debug_stack_set_zero
90	TRACE_IRQS_ON
91	call	debug_stack_reset
92.endm
93
94.macro TRACE_IRQS_IRETQ_DEBUG
95	btl	$9, EFLAGS(%rsp)		/* interrupts off? */
96	jnc	1f
97	TRACE_IRQS_ON_DEBUG
981:
99.endm
100
101#else
102# define TRACE_IRQS_OFF_DEBUG			TRACE_IRQS_OFF
103# define TRACE_IRQS_ON_DEBUG			TRACE_IRQS_ON
104# define TRACE_IRQS_IRETQ_DEBUG			TRACE_IRQS_IRETQ
105#endif
106
107/*
108 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
109 *
110 * This is the only entry point used for 64-bit system calls.  The
111 * hardware interface is reasonably well designed and the register to
112 * argument mapping Linux uses fits well with the registers that are
113 * available when SYSCALL is used.
114 *
115 * SYSCALL instructions can be found inlined in libc implementations as
116 * well as some other programs and libraries.  There are also a handful
117 * of SYSCALL instructions in the vDSO used, for example, as a
118 * clock_gettimeofday fallback.
119 *
120 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
121 * then loads new ss, cs, and rip from previously programmed MSRs.
122 * rflags gets masked by a value from another MSR (so CLD and CLAC
123 * are not needed). SYSCALL does not save anything on the stack
124 * and does not change rsp.
125 *
126 * Registers on entry:
127 * rax  system call number
128 * rcx  return address
129 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
130 * rdi  arg0
131 * rsi  arg1
132 * rdx  arg2
133 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
134 * r8   arg4
135 * r9   arg5
136 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
137 *
138 * Only called from user space.
139 *
140 * When user can change pt_regs->foo always force IRET. That is because
141 * it deals with uncanonical addresses better. SYSRET has trouble
142 * with them due to bugs in both AMD and Intel CPUs.
143 */
144
145ENTRY(entry_SYSCALL_64)
146	UNWIND_HINT_EMPTY
147	/*
148	 * Interrupts are off on entry.
149	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
150	 * it is too small to ever cause noticeable irq latency.
151	 */
152
153	swapgs
154	/* tss.sp2 is scratch space. */
155	movq	%rsp, PER_CPU_VAR(cpu_tss_rw + TSS_sp2)
156	SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
157	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
158
159	/* Construct struct pt_regs on stack */
160	pushq	$__USER_DS				/* pt_regs->ss */
161	pushq	PER_CPU_VAR(cpu_tss_rw + TSS_sp2)	/* pt_regs->sp */
162	pushq	%r11					/* pt_regs->flags */
163	pushq	$__USER_CS				/* pt_regs->cs */
164	pushq	%rcx					/* pt_regs->ip */
165GLOBAL(entry_SYSCALL_64_after_hwframe)
166	pushq	%rax					/* pt_regs->orig_ax */
167
168	PUSH_AND_CLEAR_REGS rax=$-ENOSYS
169
170	TRACE_IRQS_OFF
171
172	/* IRQs are off. */
173	movq	%rax, %rdi
174	movq	%rsp, %rsi
175	call	do_syscall_64		/* returns with IRQs disabled */
176
177	TRACE_IRQS_IRETQ		/* we're about to change IF */
178
179	/*
180	 * Try to use SYSRET instead of IRET if we're returning to
181	 * a completely clean 64-bit userspace context.  If we're not,
182	 * go to the slow exit path.
183	 */
184	movq	RCX(%rsp), %rcx
185	movq	RIP(%rsp), %r11
186
187	cmpq	%rcx, %r11	/* SYSRET requires RCX == RIP */
188	jne	swapgs_restore_regs_and_return_to_usermode
189
190	/*
191	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
192	 * in kernel space.  This essentially lets the user take over
193	 * the kernel, since userspace controls RSP.
194	 *
195	 * If width of "canonical tail" ever becomes variable, this will need
196	 * to be updated to remain correct on both old and new CPUs.
197	 *
198	 * Change top bits to match most significant bit (47th or 56th bit
199	 * depending on paging mode) in the address.
200	 */
201#ifdef CONFIG_X86_5LEVEL
202	ALTERNATIVE "shl $(64 - 48), %rcx; sar $(64 - 48), %rcx", \
203		"shl $(64 - 57), %rcx; sar $(64 - 57), %rcx", X86_FEATURE_LA57
204#else
205	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
206	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
207#endif
208
209	/* If this changed %rcx, it was not canonical */
210	cmpq	%rcx, %r11
211	jne	swapgs_restore_regs_and_return_to_usermode
212
213	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
214	jne	swapgs_restore_regs_and_return_to_usermode
215
216	movq	R11(%rsp), %r11
217	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
218	jne	swapgs_restore_regs_and_return_to_usermode
219
220	/*
221	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
222	 * restore RF properly. If the slowpath sets it for whatever reason, we
223	 * need to restore it correctly.
224	 *
225	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
226	 * trap from userspace immediately after SYSRET.  This would cause an
227	 * infinite loop whenever #DB happens with register state that satisfies
228	 * the opportunistic SYSRET conditions.  For example, single-stepping
229	 * this user code:
230	 *
231	 *           movq	$stuck_here, %rcx
232	 *           pushfq
233	 *           popq %r11
234	 *   stuck_here:
235	 *
236	 * would never get past 'stuck_here'.
237	 */
238	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
239	jnz	swapgs_restore_regs_and_return_to_usermode
240
241	/* nothing to check for RSP */
242
243	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
244	jne	swapgs_restore_regs_and_return_to_usermode
245
246	/*
247	 * We win! This label is here just for ease of understanding
248	 * perf profiles. Nothing jumps here.
249	 */
250syscall_return_via_sysret:
251	/* rcx and r11 are already restored (see code above) */
252	UNWIND_HINT_EMPTY
253	POP_REGS pop_rdi=0 skip_r11rcx=1
254
255	/*
256	 * Now all regs are restored except RSP and RDI.
257	 * Save old stack pointer and switch to trampoline stack.
258	 */
259	movq	%rsp, %rdi
260	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
261
262	pushq	RSP-RDI(%rdi)	/* RSP */
263	pushq	(%rdi)		/* RDI */
264
265	/*
266	 * We are on the trampoline stack.  All regs except RDI are live.
267	 * We can do future final exit work right here.
268	 */
269	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
270
271	popq	%rdi
272	popq	%rsp
273	USERGS_SYSRET64
274END(entry_SYSCALL_64)
275
276/*
277 * %rdi: prev task
278 * %rsi: next task
279 */
280ENTRY(__switch_to_asm)
281	UNWIND_HINT_FUNC
282	/*
283	 * Save callee-saved registers
284	 * This must match the order in inactive_task_frame
285	 */
286	pushq	%rbp
287	pushq	%rbx
288	pushq	%r12
289	pushq	%r13
290	pushq	%r14
291	pushq	%r15
292
293	/* switch stack */
294	movq	%rsp, TASK_threadsp(%rdi)
295	movq	TASK_threadsp(%rsi), %rsp
296
297#ifdef CONFIG_STACKPROTECTOR
298	movq	TASK_stack_canary(%rsi), %rbx
299	movq	%rbx, PER_CPU_VAR(irq_stack_union)+stack_canary_offset
300#endif
301
302#ifdef CONFIG_RETPOLINE
303	/*
304	 * When switching from a shallower to a deeper call stack
305	 * the RSB may either underflow or use entries populated
306	 * with userspace addresses. On CPUs where those concerns
307	 * exist, overwrite the RSB with entries which capture
308	 * speculative execution to prevent attack.
309	 */
310	FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW
311#endif
312
313	/* restore callee-saved registers */
314	popq	%r15
315	popq	%r14
316	popq	%r13
317	popq	%r12
318	popq	%rbx
319	popq	%rbp
320
321	jmp	__switch_to
322END(__switch_to_asm)
323
324/*
325 * A newly forked process directly context switches into this address.
326 *
327 * rax: prev task we switched from
328 * rbx: kernel thread func (NULL for user thread)
329 * r12: kernel thread arg
330 */
331ENTRY(ret_from_fork)
332	UNWIND_HINT_EMPTY
333	movq	%rax, %rdi
334	call	schedule_tail			/* rdi: 'prev' task parameter */
335
336	testq	%rbx, %rbx			/* from kernel_thread? */
337	jnz	1f				/* kernel threads are uncommon */
338
3392:
340	UNWIND_HINT_REGS
341	movq	%rsp, %rdi
342	call	syscall_return_slowpath	/* returns with IRQs disabled */
343	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
344	jmp	swapgs_restore_regs_and_return_to_usermode
345
3461:
347	/* kernel thread */
348	UNWIND_HINT_EMPTY
349	movq	%r12, %rdi
350	CALL_NOSPEC %rbx
351	/*
352	 * A kernel thread is allowed to return here after successfully
353	 * calling do_execve().  Exit to userspace to complete the execve()
354	 * syscall.
355	 */
356	movq	$0, RAX(%rsp)
357	jmp	2b
358END(ret_from_fork)
359
360/*
361 * Build the entry stubs with some assembler magic.
362 * We pack 1 stub into every 8-byte block.
363 */
364	.align 8
365ENTRY(irq_entries_start)
366    vector=FIRST_EXTERNAL_VECTOR
367    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
368	UNWIND_HINT_IRET_REGS
369	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
370	jmp	common_interrupt
371	.align	8
372	vector=vector+1
373    .endr
374END(irq_entries_start)
375
376.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
377#ifdef CONFIG_DEBUG_ENTRY
378	pushq %rax
379	SAVE_FLAGS(CLBR_RAX)
380	testl $X86_EFLAGS_IF, %eax
381	jz .Lokay_\@
382	ud2
383.Lokay_\@:
384	popq %rax
385#endif
386.endm
387
388/*
389 * Enters the IRQ stack if we're not already using it.  NMI-safe.  Clobbers
390 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone.
391 * Requires kernel GSBASE.
392 *
393 * The invariant is that, if irq_count != -1, then the IRQ stack is in use.
394 */
395.macro ENTER_IRQ_STACK regs=1 old_rsp save_ret=0
396	DEBUG_ENTRY_ASSERT_IRQS_OFF
397
398	.if \save_ret
399	/*
400	 * If save_ret is set, the original stack contains one additional
401	 * entry -- the return address. Therefore, move the address one
402	 * entry below %rsp to \old_rsp.
403	 */
404	leaq	8(%rsp), \old_rsp
405	.else
406	movq	%rsp, \old_rsp
407	.endif
408
409	.if \regs
410	UNWIND_HINT_REGS base=\old_rsp
411	.endif
412
413	incl	PER_CPU_VAR(irq_count)
414	jnz	.Lirq_stack_push_old_rsp_\@
415
416	/*
417	 * Right now, if we just incremented irq_count to zero, we've
418	 * claimed the IRQ stack but we haven't switched to it yet.
419	 *
420	 * If anything is added that can interrupt us here without using IST,
421	 * it must be *extremely* careful to limit its stack usage.  This
422	 * could include kprobes and a hypothetical future IST-less #DB
423	 * handler.
424	 *
425	 * The OOPS unwinder relies on the word at the top of the IRQ
426	 * stack linking back to the previous RSP for the entire time we're
427	 * on the IRQ stack.  For this to work reliably, we need to write
428	 * it before we actually move ourselves to the IRQ stack.
429	 */
430
431	movq	\old_rsp, PER_CPU_VAR(irq_stack_union + IRQ_STACK_SIZE - 8)
432	movq	PER_CPU_VAR(irq_stack_ptr), %rsp
433
434#ifdef CONFIG_DEBUG_ENTRY
435	/*
436	 * If the first movq above becomes wrong due to IRQ stack layout
437	 * changes, the only way we'll notice is if we try to unwind right
438	 * here.  Assert that we set up the stack right to catch this type
439	 * of bug quickly.
440	 */
441	cmpq	-8(%rsp), \old_rsp
442	je	.Lirq_stack_okay\@
443	ud2
444	.Lirq_stack_okay\@:
445#endif
446
447.Lirq_stack_push_old_rsp_\@:
448	pushq	\old_rsp
449
450	.if \regs
451	UNWIND_HINT_REGS indirect=1
452	.endif
453
454	.if \save_ret
455	/*
456	 * Push the return address to the stack. This return address can
457	 * be found at the "real" original RSP, which was offset by 8 at
458	 * the beginning of this macro.
459	 */
460	pushq	-8(\old_rsp)
461	.endif
462.endm
463
464/*
465 * Undoes ENTER_IRQ_STACK.
466 */
467.macro LEAVE_IRQ_STACK regs=1
468	DEBUG_ENTRY_ASSERT_IRQS_OFF
469	/* We need to be off the IRQ stack before decrementing irq_count. */
470	popq	%rsp
471
472	.if \regs
473	UNWIND_HINT_REGS
474	.endif
475
476	/*
477	 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming
478	 * the irq stack but we're not on it.
479	 */
480
481	decl	PER_CPU_VAR(irq_count)
482.endm
483
484/*
485 * Interrupt entry helper function.
486 *
487 * Entry runs with interrupts off. Stack layout at entry:
488 * +----------------------------------------------------+
489 * | regs->ss						|
490 * | regs->rsp						|
491 * | regs->eflags					|
492 * | regs->cs						|
493 * | regs->ip						|
494 * +----------------------------------------------------+
495 * | regs->orig_ax = ~(interrupt number)		|
496 * +----------------------------------------------------+
497 * | return address					|
498 * +----------------------------------------------------+
499 */
500ENTRY(interrupt_entry)
501	UNWIND_HINT_FUNC
502	ASM_CLAC
503	cld
504
505	testb	$3, CS-ORIG_RAX+8(%rsp)
506	jz	1f
507	SWAPGS
508
509	/*
510	 * Switch to the thread stack. The IRET frame and orig_ax are
511	 * on the stack, as well as the return address. RDI..R12 are
512	 * not (yet) on the stack and space has not (yet) been
513	 * allocated for them.
514	 */
515	pushq	%rdi
516
517	/* Need to switch before accessing the thread stack. */
518	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi
519	movq	%rsp, %rdi
520	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
521
522	 /*
523	  * We have RDI, return address, and orig_ax on the stack on
524	  * top of the IRET frame. That means offset=24
525	  */
526	UNWIND_HINT_IRET_REGS base=%rdi offset=24
527
528	pushq	7*8(%rdi)		/* regs->ss */
529	pushq	6*8(%rdi)		/* regs->rsp */
530	pushq	5*8(%rdi)		/* regs->eflags */
531	pushq	4*8(%rdi)		/* regs->cs */
532	pushq	3*8(%rdi)		/* regs->ip */
533	pushq	2*8(%rdi)		/* regs->orig_ax */
534	pushq	8(%rdi)			/* return address */
535	UNWIND_HINT_FUNC
536
537	movq	(%rdi), %rdi
5381:
539
540	PUSH_AND_CLEAR_REGS save_ret=1
541	ENCODE_FRAME_POINTER 8
542
543	testb	$3, CS+8(%rsp)
544	jz	1f
545
546	/*
547	 * IRQ from user mode.
548	 *
549	 * We need to tell lockdep that IRQs are off.  We can't do this until
550	 * we fix gsbase, and we should do it before enter_from_user_mode
551	 * (which can take locks).  Since TRACE_IRQS_OFF is idempotent,
552	 * the simplest way to handle it is to just call it twice if
553	 * we enter from user mode.  There's no reason to optimize this since
554	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
555	 */
556	TRACE_IRQS_OFF
557
558	CALL_enter_from_user_mode
559
5601:
561	ENTER_IRQ_STACK old_rsp=%rdi save_ret=1
562	/* We entered an interrupt context - irqs are off: */
563	TRACE_IRQS_OFF
564
565	ret
566END(interrupt_entry)
567
568
569/* Interrupt entry/exit. */
570
571	/*
572	 * The interrupt stubs push (~vector+0x80) onto the stack and
573	 * then jump to common_interrupt.
574	 */
575	.p2align CONFIG_X86_L1_CACHE_SHIFT
576common_interrupt:
577	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
578	call	interrupt_entry
579	UNWIND_HINT_REGS indirect=1
580	call	do_IRQ	/* rdi points to pt_regs */
581	/* 0(%rsp): old RSP */
582ret_from_intr:
583	DISABLE_INTERRUPTS(CLBR_ANY)
584	TRACE_IRQS_OFF
585
586	LEAVE_IRQ_STACK
587
588	testb	$3, CS(%rsp)
589	jz	retint_kernel
590
591	/* Interrupt came from user space */
592GLOBAL(retint_user)
593	mov	%rsp,%rdi
594	call	prepare_exit_to_usermode
595	TRACE_IRQS_IRETQ
596
597GLOBAL(swapgs_restore_regs_and_return_to_usermode)
598#ifdef CONFIG_DEBUG_ENTRY
599	/* Assert that pt_regs indicates user mode. */
600	testb	$3, CS(%rsp)
601	jnz	1f
602	ud2
6031:
604#endif
605	POP_REGS pop_rdi=0
606
607	/*
608	 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS.
609	 * Save old stack pointer and switch to trampoline stack.
610	 */
611	movq	%rsp, %rdi
612	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
613
614	/* Copy the IRET frame to the trampoline stack. */
615	pushq	6*8(%rdi)	/* SS */
616	pushq	5*8(%rdi)	/* RSP */
617	pushq	4*8(%rdi)	/* EFLAGS */
618	pushq	3*8(%rdi)	/* CS */
619	pushq	2*8(%rdi)	/* RIP */
620
621	/* Push user RDI on the trampoline stack. */
622	pushq	(%rdi)
623
624	/*
625	 * We are on the trampoline stack.  All regs except RDI are live.
626	 * We can do future final exit work right here.
627	 */
628
629	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
630
631	/* Restore RDI. */
632	popq	%rdi
633	SWAPGS
634	INTERRUPT_RETURN
635
636
637/* Returning to kernel space */
638retint_kernel:
639#ifdef CONFIG_PREEMPT
640	/* Interrupts are off */
641	/* Check if we need preemption */
642	btl	$9, EFLAGS(%rsp)		/* were interrupts off? */
643	jnc	1f
6440:	cmpl	$0, PER_CPU_VAR(__preempt_count)
645	jnz	1f
646	call	preempt_schedule_irq
647	jmp	0b
6481:
649#endif
650	/*
651	 * The iretq could re-enable interrupts:
652	 */
653	TRACE_IRQS_IRETQ
654
655GLOBAL(restore_regs_and_return_to_kernel)
656#ifdef CONFIG_DEBUG_ENTRY
657	/* Assert that pt_regs indicates kernel mode. */
658	testb	$3, CS(%rsp)
659	jz	1f
660	ud2
6611:
662#endif
663	POP_REGS
664	addq	$8, %rsp	/* skip regs->orig_ax */
665	/*
666	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
667	 * when returning from IPI handler.
668	 */
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					/* to kernel GS */
717	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi	/* to kernel CR3 */
718
719	movq	PER_CPU_VAR(espfix_waddr), %rdi
720	movq	%rax, (0*8)(%rdi)		/* user RAX */
721	movq	(1*8)(%rsp), %rax		/* user RIP */
722	movq	%rax, (1*8)(%rdi)
723	movq	(2*8)(%rsp), %rax		/* user CS */
724	movq	%rax, (2*8)(%rdi)
725	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
726	movq	%rax, (3*8)(%rdi)
727	movq	(5*8)(%rsp), %rax		/* user SS */
728	movq	%rax, (5*8)(%rdi)
729	movq	(4*8)(%rsp), %rax		/* user RSP */
730	movq	%rax, (4*8)(%rdi)
731	/* Now RAX == RSP. */
732
733	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
734
735	/*
736	 * espfix_stack[31:16] == 0.  The page tables are set up such that
737	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
738	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
739	 * the same page.  Set up RSP so that RSP[31:16] contains the
740	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
741	 * still points to an RO alias of the ESPFIX stack.
742	 */
743	orq	PER_CPU_VAR(espfix_stack), %rax
744
745	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
746	SWAPGS					/* to user GS */
747	popq	%rdi				/* Restore user RDI */
748
749	movq	%rax, %rsp
750	UNWIND_HINT_IRET_REGS offset=8
751
752	/*
753	 * At this point, we cannot write to the stack any more, but we can
754	 * still read.
755	 */
756	popq	%rax				/* Restore user RAX */
757
758	/*
759	 * RSP now points to an ordinary IRET frame, except that the page
760	 * is read-only and RSP[31:16] are preloaded with the userspace
761	 * values.  We can now IRET back to userspace.
762	 */
763	jmp	native_irq_return_iret
764#endif
765END(common_interrupt)
766
767/*
768 * APIC interrupts.
769 */
770.macro apicinterrupt3 num sym do_sym
771ENTRY(\sym)
772	UNWIND_HINT_IRET_REGS
773	pushq	$~(\num)
774.Lcommon_\sym:
775	call	interrupt_entry
776	UNWIND_HINT_REGS indirect=1
777	call	\do_sym	/* rdi points to pt_regs */
778	jmp	ret_from_intr
779END(\sym)
780.endm
781
782/* Make sure APIC interrupt handlers end up in the irqentry section: */
783#define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
784#define POP_SECTION_IRQENTRY	.popsection
785
786.macro apicinterrupt num sym do_sym
787PUSH_SECTION_IRQENTRY
788apicinterrupt3 \num \sym \do_sym
789POP_SECTION_IRQENTRY
790.endm
791
792#ifdef CONFIG_SMP
793apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
794apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
795#endif
796
797#ifdef CONFIG_X86_UV
798apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
799#endif
800
801apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
802apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
803
804#ifdef CONFIG_HAVE_KVM
805apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
806apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
807apicinterrupt3 POSTED_INTR_NESTED_VECTOR	kvm_posted_intr_nested_ipi	smp_kvm_posted_intr_nested_ipi
808#endif
809
810#ifdef CONFIG_X86_MCE_THRESHOLD
811apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
812#endif
813
814#ifdef CONFIG_X86_MCE_AMD
815apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
816#endif
817
818#ifdef CONFIG_X86_THERMAL_VECTOR
819apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
820#endif
821
822#ifdef CONFIG_SMP
823apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
824apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
825apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
826#endif
827
828apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
829apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
830
831#ifdef CONFIG_IRQ_WORK
832apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
833#endif
834
835/*
836 * Exception entry points.
837 */
838#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss_rw) + (TSS_ist + ((x) - 1) * 8)
839
840/**
841 * idtentry - Generate an IDT entry stub
842 * @sym:		Name of the generated entry point
843 * @do_sym: 		C function to be called
844 * @has_error_code: 	True if this IDT vector has an error code on the stack
845 * @paranoid: 		non-zero means that this vector may be invoked from
846 *			kernel mode with user GSBASE and/or user CR3.
847 *			2 is special -- see below.
848 * @shift_ist:		Set to an IST index if entries from kernel mode should
849 *             		decrement the IST stack so that nested entries get a
850 *			fresh stack.  (This is for #DB, which has a nasty habit
851 *             		of recursing.)
852 *
853 * idtentry generates an IDT stub that sets up a usable kernel context,
854 * creates struct pt_regs, and calls @do_sym.  The stub has the following
855 * special behaviors:
856 *
857 * On an entry from user mode, the stub switches from the trampoline or
858 * IST stack to the normal thread stack.  On an exit to user mode, the
859 * normal exit-to-usermode path is invoked.
860 *
861 * On an exit to kernel mode, if @paranoid == 0, we check for preemption,
862 * whereas we omit the preemption check if @paranoid != 0.  This is purely
863 * because the implementation is simpler this way.  The kernel only needs
864 * to check for asynchronous kernel preemption when IRQ handlers return.
865 *
866 * If @paranoid == 0, then the stub will handle IRET faults by pretending
867 * that the fault came from user mode.  It will handle gs_change faults by
868 * pretending that the fault happened with kernel GSBASE.  Since this handling
869 * is omitted for @paranoid != 0, the #GP, #SS, and #NP stubs must have
870 * @paranoid == 0.  This special handling will do the wrong thing for
871 * espfix-induced #DF on IRET, so #DF must not use @paranoid == 0.
872 *
873 * @paranoid == 2 is special: the stub will never switch stacks.  This is for
874 * #DF: if the thread stack is somehow unusable, we'll still get a useful OOPS.
875 */
876.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
877ENTRY(\sym)
878	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
879
880	/* Sanity check */
881	.if \shift_ist != -1 && \paranoid == 0
882	.error "using shift_ist requires paranoid=1"
883	.endif
884
885	ASM_CLAC
886
887	.if \has_error_code == 0
888	pushq	$-1				/* ORIG_RAX: no syscall to restart */
889	.endif
890
891	.if \paranoid == 1
892	testb	$3, CS-ORIG_RAX(%rsp)		/* If coming from userspace, switch stacks */
893	jnz	.Lfrom_usermode_switch_stack_\@
894	.endif
895
896	.if \paranoid
897	call	paranoid_entry
898	.else
899	call	error_entry
900	.endif
901	UNWIND_HINT_REGS
902	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
903
904	.if \paranoid
905	.if \shift_ist != -1
906	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
907	.else
908	TRACE_IRQS_OFF
909	.endif
910	.endif
911
912	movq	%rsp, %rdi			/* pt_regs pointer */
913
914	.if \has_error_code
915	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
916	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
917	.else
918	xorl	%esi, %esi			/* no error code */
919	.endif
920
921	.if \shift_ist != -1
922	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
923	.endif
924
925	call	\do_sym
926
927	.if \shift_ist != -1
928	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
929	.endif
930
931	/* these procedures expect "no swapgs" flag in ebx */
932	.if \paranoid
933	jmp	paranoid_exit
934	.else
935	jmp	error_exit
936	.endif
937
938	.if \paranoid == 1
939	/*
940	 * Entry from userspace.  Switch stacks and treat it
941	 * as a normal entry.  This means that paranoid handlers
942	 * run in real process context if user_mode(regs).
943	 */
944.Lfrom_usermode_switch_stack_\@:
945	call	error_entry
946
947	movq	%rsp, %rdi			/* pt_regs pointer */
948
949	.if \has_error_code
950	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
951	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
952	.else
953	xorl	%esi, %esi			/* no error code */
954	.endif
955
956	call	\do_sym
957
958	jmp	error_exit
959	.endif
960END(\sym)
961.endm
962
963idtentry divide_error			do_divide_error			has_error_code=0
964idtentry overflow			do_overflow			has_error_code=0
965idtentry bounds				do_bounds			has_error_code=0
966idtentry invalid_op			do_invalid_op			has_error_code=0
967idtentry device_not_available		do_device_not_available		has_error_code=0
968idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
969idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
970idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
971idtentry segment_not_present		do_segment_not_present		has_error_code=1
972idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
973idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
974idtentry alignment_check		do_alignment_check		has_error_code=1
975idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
976
977
978	/*
979	 * Reload gs selector with exception handling
980	 * edi:  new selector
981	 */
982ENTRY(native_load_gs_index)
983	FRAME_BEGIN
984	pushfq
985	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
986	TRACE_IRQS_OFF
987	SWAPGS
988.Lgs_change:
989	movl	%edi, %gs
9902:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
991	SWAPGS
992	TRACE_IRQS_FLAGS (%rsp)
993	popfq
994	FRAME_END
995	ret
996ENDPROC(native_load_gs_index)
997EXPORT_SYMBOL(native_load_gs_index)
998
999	_ASM_EXTABLE(.Lgs_change, bad_gs)
1000	.section .fixup, "ax"
1001	/* running with kernelgs */
1002bad_gs:
1003	SWAPGS					/* switch back to user gs */
1004.macro ZAP_GS
1005	/* This can't be a string because the preprocessor needs to see it. */
1006	movl $__USER_DS, %eax
1007	movl %eax, %gs
1008.endm
1009	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
1010	xorl	%eax, %eax
1011	movl	%eax, %gs
1012	jmp	2b
1013	.previous
1014
1015/* Call softirq on interrupt stack. Interrupts are off. */
1016ENTRY(do_softirq_own_stack)
1017	pushq	%rbp
1018	mov	%rsp, %rbp
1019	ENTER_IRQ_STACK regs=0 old_rsp=%r11
1020	call	__do_softirq
1021	LEAVE_IRQ_STACK regs=0
1022	leaveq
1023	ret
1024ENDPROC(do_softirq_own_stack)
1025
1026#ifdef CONFIG_XEN_PV
1027idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0
1028
1029/*
1030 * A note on the "critical region" in our callback handler.
1031 * We want to avoid stacking callback handlers due to events occurring
1032 * during handling of the last event. To do this, we keep events disabled
1033 * until we've done all processing. HOWEVER, we must enable events before
1034 * popping the stack frame (can't be done atomically) and so it would still
1035 * be possible to get enough handler activations to overflow the stack.
1036 * Although unlikely, bugs of that kind are hard to track down, so we'd
1037 * like to avoid the possibility.
1038 * So, on entry to the handler we detect whether we interrupted an
1039 * existing activation in its critical region -- if so, we pop the current
1040 * activation and restart the handler using the previous one.
1041 */
1042ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
1043
1044/*
1045 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
1046 * see the correct pointer to the pt_regs
1047 */
1048	UNWIND_HINT_FUNC
1049	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
1050	UNWIND_HINT_REGS
1051
1052	ENTER_IRQ_STACK old_rsp=%r10
1053	call	xen_evtchn_do_upcall
1054	LEAVE_IRQ_STACK
1055
1056#ifndef CONFIG_PREEMPT
1057	call	xen_maybe_preempt_hcall
1058#endif
1059	jmp	error_exit
1060END(xen_do_hypervisor_callback)
1061
1062/*
1063 * Hypervisor uses this for application faults while it executes.
1064 * We get here for two reasons:
1065 *  1. Fault while reloading DS, ES, FS or GS
1066 *  2. Fault while executing IRET
1067 * Category 1 we do not need to fix up as Xen has already reloaded all segment
1068 * registers that could be reloaded and zeroed the others.
1069 * Category 2 we fix up by killing the current process. We cannot use the
1070 * normal Linux return path in this case because if we use the IRET hypercall
1071 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
1072 * We distinguish between categories by comparing each saved segment register
1073 * with its current contents: any discrepancy means we in category 1.
1074 */
1075ENTRY(xen_failsafe_callback)
1076	UNWIND_HINT_EMPTY
1077	movl	%ds, %ecx
1078	cmpw	%cx, 0x10(%rsp)
1079	jne	1f
1080	movl	%es, %ecx
1081	cmpw	%cx, 0x18(%rsp)
1082	jne	1f
1083	movl	%fs, %ecx
1084	cmpw	%cx, 0x20(%rsp)
1085	jne	1f
1086	movl	%gs, %ecx
1087	cmpw	%cx, 0x28(%rsp)
1088	jne	1f
1089	/* All segments match their saved values => Category 2 (Bad IRET). */
1090	movq	(%rsp), %rcx
1091	movq	8(%rsp), %r11
1092	addq	$0x30, %rsp
1093	pushq	$0				/* RIP */
1094	UNWIND_HINT_IRET_REGS offset=8
1095	jmp	general_protection
10961:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
1097	movq	(%rsp), %rcx
1098	movq	8(%rsp), %r11
1099	addq	$0x30, %rsp
1100	UNWIND_HINT_IRET_REGS
1101	pushq	$-1 /* orig_ax = -1 => not a system call */
1102	PUSH_AND_CLEAR_REGS
1103	ENCODE_FRAME_POINTER
1104	jmp	error_exit
1105END(xen_failsafe_callback)
1106#endif /* CONFIG_XEN_PV */
1107
1108#ifdef CONFIG_XEN_PVHVM
1109apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1110	xen_hvm_callback_vector xen_evtchn_do_upcall
1111#endif
1112
1113
1114#if IS_ENABLED(CONFIG_HYPERV)
1115apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1116	hyperv_callback_vector hyperv_vector_handler
1117
1118apicinterrupt3 HYPERV_REENLIGHTENMENT_VECTOR \
1119	hyperv_reenlightenment_vector hyperv_reenlightenment_intr
1120
1121apicinterrupt3 HYPERV_STIMER0_VECTOR \
1122	hv_stimer0_callback_vector hv_stimer0_vector_handler
1123#endif /* CONFIG_HYPERV */
1124
1125idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1126idtentry int3			do_int3			has_error_code=0
1127idtentry stack_segment		do_stack_segment	has_error_code=1
1128
1129#ifdef CONFIG_XEN_PV
1130idtentry xennmi			do_nmi			has_error_code=0
1131idtentry xendebug		do_debug		has_error_code=0
1132idtentry xenint3		do_int3			has_error_code=0
1133#endif
1134
1135idtentry general_protection	do_general_protection	has_error_code=1
1136idtentry page_fault		do_page_fault		has_error_code=1
1137
1138#ifdef CONFIG_KVM_GUEST
1139idtentry async_page_fault	do_async_page_fault	has_error_code=1
1140#endif
1141
1142#ifdef CONFIG_X86_MCE
1143idtentry machine_check		do_mce			has_error_code=0	paranoid=1
1144#endif
1145
1146/*
1147 * Save all registers in pt_regs, and switch gs if needed.
1148 * Use slow, but surefire "are we in kernel?" check.
1149 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
1150 */
1151ENTRY(paranoid_entry)
1152	UNWIND_HINT_FUNC
1153	cld
1154	PUSH_AND_CLEAR_REGS save_ret=1
1155	ENCODE_FRAME_POINTER 8
1156	movl	$1, %ebx
1157	movl	$MSR_GS_BASE, %ecx
1158	rdmsr
1159	testl	%edx, %edx
1160	js	1f				/* negative -> in kernel */
1161	SWAPGS
1162	xorl	%ebx, %ebx
1163
11641:
1165	/*
1166	 * Always stash CR3 in %r14.  This value will be restored,
1167	 * verbatim, at exit.  Needed if paranoid_entry interrupted
1168	 * another entry that already switched to the user CR3 value
1169	 * but has not yet returned to userspace.
1170	 *
1171	 * This is also why CS (stashed in the "iret frame" by the
1172	 * hardware at entry) can not be used: this may be a return
1173	 * to kernel code, but with a user CR3 value.
1174	 */
1175	SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
1176
1177	ret
1178END(paranoid_entry)
1179
1180/*
1181 * "Paranoid" exit path from exception stack.  This is invoked
1182 * only on return from non-NMI IST interrupts that came
1183 * from kernel space.
1184 *
1185 * We may be returning to very strange contexts (e.g. very early
1186 * in syscall entry), so checking for preemption here would
1187 * be complicated.  Fortunately, we there's no good reason
1188 * to try to handle preemption here.
1189 *
1190 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1191 */
1192ENTRY(paranoid_exit)
1193	UNWIND_HINT_REGS
1194	DISABLE_INTERRUPTS(CLBR_ANY)
1195	TRACE_IRQS_OFF_DEBUG
1196	testl	%ebx, %ebx			/* swapgs needed? */
1197	jnz	.Lparanoid_exit_no_swapgs
1198	TRACE_IRQS_IRETQ
1199	/* Always restore stashed CR3 value (see paranoid_entry) */
1200	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1201	SWAPGS_UNSAFE_STACK
1202	jmp	.Lparanoid_exit_restore
1203.Lparanoid_exit_no_swapgs:
1204	TRACE_IRQS_IRETQ_DEBUG
1205	/* Always restore stashed CR3 value (see paranoid_entry) */
1206	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1207.Lparanoid_exit_restore:
1208	jmp restore_regs_and_return_to_kernel
1209END(paranoid_exit)
1210
1211/*
1212 * Save all registers in pt_regs, and switch GS if needed.
1213 */
1214ENTRY(error_entry)
1215	UNWIND_HINT_FUNC
1216	cld
1217	PUSH_AND_CLEAR_REGS save_ret=1
1218	ENCODE_FRAME_POINTER 8
1219	testb	$3, CS+8(%rsp)
1220	jz	.Lerror_kernelspace
1221
1222	/*
1223	 * We entered from user mode or we're pretending to have entered
1224	 * from user mode due to an IRET fault.
1225	 */
1226	SWAPGS
1227	/* We have user CR3.  Change to kernel CR3. */
1228	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1229
1230.Lerror_entry_from_usermode_after_swapgs:
1231	/* Put us onto the real thread stack. */
1232	popq	%r12				/* save return addr in %12 */
1233	movq	%rsp, %rdi			/* arg0 = pt_regs pointer */
1234	call	sync_regs
1235	movq	%rax, %rsp			/* switch stack */
1236	ENCODE_FRAME_POINTER
1237	pushq	%r12
1238
1239	/*
1240	 * We need to tell lockdep that IRQs are off.  We can't do this until
1241	 * we fix gsbase, and we should do it before enter_from_user_mode
1242	 * (which can take locks).
1243	 */
1244	TRACE_IRQS_OFF
1245	CALL_enter_from_user_mode
1246	ret
1247
1248.Lerror_entry_done:
1249	TRACE_IRQS_OFF
1250	ret
1251
1252	/*
1253	 * There are two places in the kernel that can potentially fault with
1254	 * usergs. Handle them here.  B stepping K8s sometimes report a
1255	 * truncated RIP for IRET exceptions returning to compat mode. Check
1256	 * for these here too.
1257	 */
1258.Lerror_kernelspace:
1259	leaq	native_irq_return_iret(%rip), %rcx
1260	cmpq	%rcx, RIP+8(%rsp)
1261	je	.Lerror_bad_iret
1262	movl	%ecx, %eax			/* zero extend */
1263	cmpq	%rax, RIP+8(%rsp)
1264	je	.Lbstep_iret
1265	cmpq	$.Lgs_change, RIP+8(%rsp)
1266	jne	.Lerror_entry_done
1267
1268	/*
1269	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1270	 * gsbase and proceed.  We'll fix up the exception and land in
1271	 * .Lgs_change's error handler with kernel gsbase.
1272	 */
1273	SWAPGS
1274	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1275	jmp .Lerror_entry_done
1276
1277.Lbstep_iret:
1278	/* Fix truncated RIP */
1279	movq	%rcx, RIP+8(%rsp)
1280	/* fall through */
1281
1282.Lerror_bad_iret:
1283	/*
1284	 * We came from an IRET to user mode, so we have user
1285	 * gsbase and CR3.  Switch to kernel gsbase and CR3:
1286	 */
1287	SWAPGS
1288	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1289
1290	/*
1291	 * Pretend that the exception came from user mode: set up pt_regs
1292	 * as if we faulted immediately after IRET.
1293	 */
1294	mov	%rsp, %rdi
1295	call	fixup_bad_iret
1296	mov	%rax, %rsp
1297	jmp	.Lerror_entry_from_usermode_after_swapgs
1298END(error_entry)
1299
1300ENTRY(error_exit)
1301	UNWIND_HINT_REGS
1302	DISABLE_INTERRUPTS(CLBR_ANY)
1303	TRACE_IRQS_OFF
1304	testb	$3, CS(%rsp)
1305	jz	retint_kernel
1306	jmp	retint_user
1307END(error_exit)
1308
1309/*
1310 * Runs on exception stack.  Xen PV does not go through this path at all,
1311 * so we can use real assembly here.
1312 *
1313 * Registers:
1314 *	%r14: Used to save/restore the CR3 of the interrupted context
1315 *	      when PAGE_TABLE_ISOLATION is in use.  Do not clobber.
1316 */
1317ENTRY(nmi)
1318	UNWIND_HINT_IRET_REGS
1319
1320	/*
1321	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1322	 * the iretq it performs will take us out of NMI context.
1323	 * This means that we can have nested NMIs where the next
1324	 * NMI is using the top of the stack of the previous NMI. We
1325	 * can't let it execute because the nested NMI will corrupt the
1326	 * stack of the previous NMI. NMI handlers are not re-entrant
1327	 * anyway.
1328	 *
1329	 * To handle this case we do the following:
1330	 *  Check the a special location on the stack that contains
1331	 *  a variable that is set when NMIs are executing.
1332	 *  The interrupted task's stack is also checked to see if it
1333	 *  is an NMI stack.
1334	 *  If the variable is not set and the stack is not the NMI
1335	 *  stack then:
1336	 *    o Set the special variable on the stack
1337	 *    o Copy the interrupt frame into an "outermost" location on the
1338	 *      stack
1339	 *    o Copy the interrupt frame into an "iret" location on the stack
1340	 *    o Continue processing the NMI
1341	 *  If the variable is set or the previous stack is the NMI stack:
1342	 *    o Modify the "iret" location to jump to the repeat_nmi
1343	 *    o return back to the first NMI
1344	 *
1345	 * Now on exit of the first NMI, we first clear the stack variable
1346	 * The NMI stack will tell any nested NMIs at that point that it is
1347	 * nested. Then we pop the stack normally with iret, and if there was
1348	 * a nested NMI that updated the copy interrupt stack frame, a
1349	 * jump will be made to the repeat_nmi code that will handle the second
1350	 * NMI.
1351	 *
1352	 * However, espfix prevents us from directly returning to userspace
1353	 * with a single IRET instruction.  Similarly, IRET to user mode
1354	 * can fault.  We therefore handle NMIs from user space like
1355	 * other IST entries.
1356	 */
1357
1358	ASM_CLAC
1359
1360	/* Use %rdx as our temp variable throughout */
1361	pushq	%rdx
1362
1363	testb	$3, CS-RIP+8(%rsp)
1364	jz	.Lnmi_from_kernel
1365
1366	/*
1367	 * NMI from user mode.  We need to run on the thread stack, but we
1368	 * can't go through the normal entry paths: NMIs are masked, and
1369	 * we don't want to enable interrupts, because then we'll end
1370	 * up in an awkward situation in which IRQs are on but NMIs
1371	 * are off.
1372	 *
1373	 * We also must not push anything to the stack before switching
1374	 * stacks lest we corrupt the "NMI executing" variable.
1375	 */
1376
1377	swapgs
1378	cld
1379	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx
1380	movq	%rsp, %rdx
1381	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1382	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1383	pushq	5*8(%rdx)	/* pt_regs->ss */
1384	pushq	4*8(%rdx)	/* pt_regs->rsp */
1385	pushq	3*8(%rdx)	/* pt_regs->flags */
1386	pushq	2*8(%rdx)	/* pt_regs->cs */
1387	pushq	1*8(%rdx)	/* pt_regs->rip */
1388	UNWIND_HINT_IRET_REGS
1389	pushq   $-1		/* pt_regs->orig_ax */
1390	PUSH_AND_CLEAR_REGS rdx=(%rdx)
1391	ENCODE_FRAME_POINTER
1392
1393	/*
1394	 * At this point we no longer need to worry about stack damage
1395	 * due to nesting -- we're on the normal thread stack and we're
1396	 * done with the NMI stack.
1397	 */
1398
1399	movq	%rsp, %rdi
1400	movq	$-1, %rsi
1401	call	do_nmi
1402
1403	/*
1404	 * Return back to user mode.  We must *not* do the normal exit
1405	 * work, because we don't want to enable interrupts.
1406	 */
1407	jmp	swapgs_restore_regs_and_return_to_usermode
1408
1409.Lnmi_from_kernel:
1410	/*
1411	 * Here's what our stack frame will look like:
1412	 * +---------------------------------------------------------+
1413	 * | original SS                                             |
1414	 * | original Return RSP                                     |
1415	 * | original RFLAGS                                         |
1416	 * | original CS                                             |
1417	 * | original RIP                                            |
1418	 * +---------------------------------------------------------+
1419	 * | temp storage for rdx                                    |
1420	 * +---------------------------------------------------------+
1421	 * | "NMI executing" variable                                |
1422	 * +---------------------------------------------------------+
1423	 * | iret SS          } Copied from "outermost" frame        |
1424	 * | iret Return RSP  } on each loop iteration; overwritten  |
1425	 * | iret RFLAGS      } by a nested NMI to force another     |
1426	 * | iret CS          } iteration if needed.                 |
1427	 * | iret RIP         }                                      |
1428	 * +---------------------------------------------------------+
1429	 * | outermost SS          } initialized in first_nmi;       |
1430	 * | outermost Return RSP  } will not be changed before      |
1431	 * | outermost RFLAGS      } NMI processing is done.         |
1432	 * | outermost CS          } Copied to "iret" frame on each  |
1433	 * | outermost RIP         } iteration.                      |
1434	 * +---------------------------------------------------------+
1435	 * | pt_regs                                                 |
1436	 * +---------------------------------------------------------+
1437	 *
1438	 * The "original" frame is used by hardware.  Before re-enabling
1439	 * NMIs, we need to be done with it, and we need to leave enough
1440	 * space for the asm code here.
1441	 *
1442	 * We return by executing IRET while RSP points to the "iret" frame.
1443	 * That will either return for real or it will loop back into NMI
1444	 * processing.
1445	 *
1446	 * The "outermost" frame is copied to the "iret" frame on each
1447	 * iteration of the loop, so each iteration starts with the "iret"
1448	 * frame pointing to the final return target.
1449	 */
1450
1451	/*
1452	 * Determine whether we're a nested NMI.
1453	 *
1454	 * If we interrupted kernel code between repeat_nmi and
1455	 * end_repeat_nmi, then we are a nested NMI.  We must not
1456	 * modify the "iret" frame because it's being written by
1457	 * the outer NMI.  That's okay; the outer NMI handler is
1458	 * about to about to call do_nmi anyway, so we can just
1459	 * resume the outer NMI.
1460	 */
1461
1462	movq	$repeat_nmi, %rdx
1463	cmpq	8(%rsp), %rdx
1464	ja	1f
1465	movq	$end_repeat_nmi, %rdx
1466	cmpq	8(%rsp), %rdx
1467	ja	nested_nmi_out
14681:
1469
1470	/*
1471	 * Now check "NMI executing".  If it's set, then we're nested.
1472	 * This will not detect if we interrupted an outer NMI just
1473	 * before IRET.
1474	 */
1475	cmpl	$1, -8(%rsp)
1476	je	nested_nmi
1477
1478	/*
1479	 * Now test if the previous stack was an NMI stack.  This covers
1480	 * the case where we interrupt an outer NMI after it clears
1481	 * "NMI executing" but before IRET.  We need to be careful, though:
1482	 * there is one case in which RSP could point to the NMI stack
1483	 * despite there being no NMI active: naughty userspace controls
1484	 * RSP at the very beginning of the SYSCALL targets.  We can
1485	 * pull a fast one on naughty userspace, though: we program
1486	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1487	 * if it controls the kernel's RSP.  We set DF before we clear
1488	 * "NMI executing".
1489	 */
1490	lea	6*8(%rsp), %rdx
1491	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1492	cmpq	%rdx, 4*8(%rsp)
1493	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1494	ja	first_nmi
1495
1496	subq	$EXCEPTION_STKSZ, %rdx
1497	cmpq	%rdx, 4*8(%rsp)
1498	/* If it is below the NMI stack, it is a normal NMI */
1499	jb	first_nmi
1500
1501	/* Ah, it is within the NMI stack. */
1502
1503	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1504	jz	first_nmi	/* RSP was user controlled. */
1505
1506	/* This is a nested NMI. */
1507
1508nested_nmi:
1509	/*
1510	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1511	 * iteration of NMI handling.
1512	 */
1513	subq	$8, %rsp
1514	leaq	-10*8(%rsp), %rdx
1515	pushq	$__KERNEL_DS
1516	pushq	%rdx
1517	pushfq
1518	pushq	$__KERNEL_CS
1519	pushq	$repeat_nmi
1520
1521	/* Put stack back */
1522	addq	$(6*8), %rsp
1523
1524nested_nmi_out:
1525	popq	%rdx
1526
1527	/* We are returning to kernel mode, so this cannot result in a fault. */
1528	iretq
1529
1530first_nmi:
1531	/* Restore rdx. */
1532	movq	(%rsp), %rdx
1533
1534	/* Make room for "NMI executing". */
1535	pushq	$0
1536
1537	/* Leave room for the "iret" frame */
1538	subq	$(5*8), %rsp
1539
1540	/* Copy the "original" frame to the "outermost" frame */
1541	.rept 5
1542	pushq	11*8(%rsp)
1543	.endr
1544	UNWIND_HINT_IRET_REGS
1545
1546	/* Everything up to here is safe from nested NMIs */
1547
1548#ifdef CONFIG_DEBUG_ENTRY
1549	/*
1550	 * For ease of testing, unmask NMIs right away.  Disabled by
1551	 * default because IRET is very expensive.
1552	 */
1553	pushq	$0		/* SS */
1554	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1555	addq	$8, (%rsp)	/* Fix up RSP */
1556	pushfq			/* RFLAGS */
1557	pushq	$__KERNEL_CS	/* CS */
1558	pushq	$1f		/* RIP */
1559	iretq			/* continues at repeat_nmi below */
1560	UNWIND_HINT_IRET_REGS
15611:
1562#endif
1563
1564repeat_nmi:
1565	/*
1566	 * If there was a nested NMI, the first NMI's iret will return
1567	 * here. But NMIs are still enabled and we can take another
1568	 * nested NMI. The nested NMI checks the interrupted RIP to see
1569	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1570	 * it will just return, as we are about to repeat an NMI anyway.
1571	 * This makes it safe to copy to the stack frame that a nested
1572	 * NMI will update.
1573	 *
1574	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1575	 * we're repeating an NMI, gsbase has the same value that it had on
1576	 * the first iteration.  paranoid_entry will load the kernel
1577	 * gsbase if needed before we call do_nmi.  "NMI executing"
1578	 * is zero.
1579	 */
1580	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1581
1582	/*
1583	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1584	 * here must not modify the "iret" frame while we're writing to
1585	 * it or it will end up containing garbage.
1586	 */
1587	addq	$(10*8), %rsp
1588	.rept 5
1589	pushq	-6*8(%rsp)
1590	.endr
1591	subq	$(5*8), %rsp
1592end_repeat_nmi:
1593
1594	/*
1595	 * Everything below this point can be preempted by a nested NMI.
1596	 * If this happens, then the inner NMI will change the "iret"
1597	 * frame to point back to repeat_nmi.
1598	 */
1599	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1600
1601	/*
1602	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1603	 * as we should not be calling schedule in NMI context.
1604	 * Even with normal interrupts enabled. An NMI should not be
1605	 * setting NEED_RESCHED or anything that normal interrupts and
1606	 * exceptions might do.
1607	 */
1608	call	paranoid_entry
1609	UNWIND_HINT_REGS
1610
1611	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1612	movq	%rsp, %rdi
1613	movq	$-1, %rsi
1614	call	do_nmi
1615
1616	/* Always restore stashed CR3 value (see paranoid_entry) */
1617	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
1618
1619	testl	%ebx, %ebx			/* swapgs needed? */
1620	jnz	nmi_restore
1621nmi_swapgs:
1622	SWAPGS_UNSAFE_STACK
1623nmi_restore:
1624	POP_REGS
1625
1626	/*
1627	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1628	 * at the "iret" frame.
1629	 */
1630	addq	$6*8, %rsp
1631
1632	/*
1633	 * Clear "NMI executing".  Set DF first so that we can easily
1634	 * distinguish the remaining code between here and IRET from
1635	 * the SYSCALL entry and exit paths.
1636	 *
1637	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1638	 * this code when I had the misapprehension that Xen PV supported
1639	 * NMIs, and Xen PV would break that approach.
1640	 */
1641	std
1642	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1643
1644	/*
1645	 * iretq reads the "iret" frame and exits the NMI stack in a
1646	 * single instruction.  We are returning to kernel mode, so this
1647	 * cannot result in a fault.  Similarly, we don't need to worry
1648	 * about espfix64 on the way back to kernel mode.
1649	 */
1650	iretq
1651END(nmi)
1652
1653ENTRY(ignore_sysret)
1654	UNWIND_HINT_EMPTY
1655	mov	$-ENOSYS, %eax
1656	sysret
1657END(ignore_sysret)
1658
1659ENTRY(rewind_stack_do_exit)
1660	UNWIND_HINT_FUNC
1661	/* Prevent any naive code from trying to unwind to our caller. */
1662	xorl	%ebp, %ebp
1663
1664	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1665	leaq	-PTREGS_SIZE(%rax), %rsp
1666	UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE
1667
1668	call	do_exit
1669END(rewind_stack_do_exit)
1670