xref: /openbmc/linux/arch/x86/entry/entry_64.S (revision f1575595)
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
145	.pushsection .entry_trampoline, "ax"
146
147/*
148 * The code in here gets remapped into cpu_entry_area's trampoline.  This means
149 * that the assembler and linker have the wrong idea as to where this code
150 * lives (and, in fact, it's mapped more than once, so it's not even at a
151 * fixed address).  So we can't reference any symbols outside the entry
152 * trampoline and expect it to work.
153 *
154 * Instead, we carefully abuse %rip-relative addressing.
155 * _entry_trampoline(%rip) refers to the start of the remapped) entry
156 * trampoline.  We can thus find cpu_entry_area with this macro:
157 */
158
159#define CPU_ENTRY_AREA \
160	_entry_trampoline - CPU_ENTRY_AREA_entry_trampoline(%rip)
161
162/* The top word of the SYSENTER stack is hot and is usable as scratch space. */
163#define RSP_SCRATCH	CPU_ENTRY_AREA_entry_stack + \
164			SIZEOF_entry_stack - 8 + CPU_ENTRY_AREA
165
166ENTRY(entry_SYSCALL_64_trampoline)
167	UNWIND_HINT_EMPTY
168	swapgs
169
170	/* Stash the user RSP. */
171	movq	%rsp, RSP_SCRATCH
172
173	/* Note: using %rsp as a scratch reg. */
174	SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
175
176	/* Load the top of the task stack into RSP */
177	movq	CPU_ENTRY_AREA_tss + TSS_sp1 + CPU_ENTRY_AREA, %rsp
178
179	/* Start building the simulated IRET frame. */
180	pushq	$__USER_DS			/* pt_regs->ss */
181	pushq	RSP_SCRATCH			/* pt_regs->sp */
182	pushq	%r11				/* pt_regs->flags */
183	pushq	$__USER_CS			/* pt_regs->cs */
184	pushq	%rcx				/* pt_regs->ip */
185
186	/*
187	 * x86 lacks a near absolute jump, and we can't jump to the real
188	 * entry text with a relative jump.  We could push the target
189	 * address and then use retq, but this destroys the pipeline on
190	 * many CPUs (wasting over 20 cycles on Sandy Bridge).  Instead,
191	 * spill RDI and restore it in a second-stage trampoline.
192	 */
193	pushq	%rdi
194	movq	$entry_SYSCALL_64_stage2, %rdi
195	JMP_NOSPEC %rdi
196END(entry_SYSCALL_64_trampoline)
197
198	.popsection
199
200ENTRY(entry_SYSCALL_64_stage2)
201	UNWIND_HINT_EMPTY
202	popq	%rdi
203	jmp	entry_SYSCALL_64_after_hwframe
204END(entry_SYSCALL_64_stage2)
205
206ENTRY(entry_SYSCALL_64)
207	UNWIND_HINT_EMPTY
208	/*
209	 * Interrupts are off on entry.
210	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
211	 * it is too small to ever cause noticeable irq latency.
212	 */
213
214	swapgs
215	/*
216	 * This path is only taken when PAGE_TABLE_ISOLATION is disabled so it
217	 * is not required to switch CR3.
218	 */
219	movq	%rsp, PER_CPU_VAR(rsp_scratch)
220	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
221
222	/* Construct struct pt_regs on stack */
223	pushq	$__USER_DS			/* pt_regs->ss */
224	pushq	PER_CPU_VAR(rsp_scratch)	/* pt_regs->sp */
225	pushq	%r11				/* pt_regs->flags */
226	pushq	$__USER_CS			/* pt_regs->cs */
227	pushq	%rcx				/* pt_regs->ip */
228GLOBAL(entry_SYSCALL_64_after_hwframe)
229	pushq	%rax				/* pt_regs->orig_ax */
230
231	PUSH_AND_CLEAR_REGS rax=$-ENOSYS
232
233	TRACE_IRQS_OFF
234
235	/* IRQs are off. */
236	movq	%rax, %rdi
237	movq	%rsp, %rsi
238	call	do_syscall_64		/* returns with IRQs disabled */
239
240	TRACE_IRQS_IRETQ		/* we're about to change IF */
241
242	/*
243	 * Try to use SYSRET instead of IRET if we're returning to
244	 * a completely clean 64-bit userspace context.  If we're not,
245	 * go to the slow exit path.
246	 */
247	movq	RCX(%rsp), %rcx
248	movq	RIP(%rsp), %r11
249
250	cmpq	%rcx, %r11	/* SYSRET requires RCX == RIP */
251	jne	swapgs_restore_regs_and_return_to_usermode
252
253	/*
254	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
255	 * in kernel space.  This essentially lets the user take over
256	 * the kernel, since userspace controls RSP.
257	 *
258	 * If width of "canonical tail" ever becomes variable, this will need
259	 * to be updated to remain correct on both old and new CPUs.
260	 *
261	 * Change top bits to match most significant bit (47th or 56th bit
262	 * depending on paging mode) in the address.
263	 */
264#ifdef CONFIG_X86_5LEVEL
265	ALTERNATIVE "shl $(64 - 48), %rcx; sar $(64 - 48), %rcx", \
266		"shl $(64 - 57), %rcx; sar $(64 - 57), %rcx", X86_FEATURE_LA57
267#else
268	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
269	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
270#endif
271
272	/* If this changed %rcx, it was not canonical */
273	cmpq	%rcx, %r11
274	jne	swapgs_restore_regs_and_return_to_usermode
275
276	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
277	jne	swapgs_restore_regs_and_return_to_usermode
278
279	movq	R11(%rsp), %r11
280	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
281	jne	swapgs_restore_regs_and_return_to_usermode
282
283	/*
284	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
285	 * restore RF properly. If the slowpath sets it for whatever reason, we
286	 * need to restore it correctly.
287	 *
288	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
289	 * trap from userspace immediately after SYSRET.  This would cause an
290	 * infinite loop whenever #DB happens with register state that satisfies
291	 * the opportunistic SYSRET conditions.  For example, single-stepping
292	 * this user code:
293	 *
294	 *           movq	$stuck_here, %rcx
295	 *           pushfq
296	 *           popq %r11
297	 *   stuck_here:
298	 *
299	 * would never get past 'stuck_here'.
300	 */
301	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
302	jnz	swapgs_restore_regs_and_return_to_usermode
303
304	/* nothing to check for RSP */
305
306	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
307	jne	swapgs_restore_regs_and_return_to_usermode
308
309	/*
310	 * We win! This label is here just for ease of understanding
311	 * perf profiles. Nothing jumps here.
312	 */
313syscall_return_via_sysret:
314	/* rcx and r11 are already restored (see code above) */
315	UNWIND_HINT_EMPTY
316	POP_REGS pop_rdi=0 skip_r11rcx=1
317
318	/*
319	 * Now all regs are restored except RSP and RDI.
320	 * Save old stack pointer and switch to trampoline stack.
321	 */
322	movq	%rsp, %rdi
323	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
324
325	pushq	RSP-RDI(%rdi)	/* RSP */
326	pushq	(%rdi)		/* RDI */
327
328	/*
329	 * We are on the trampoline stack.  All regs except RDI are live.
330	 * We can do future final exit work right here.
331	 */
332	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
333
334	popq	%rdi
335	popq	%rsp
336	USERGS_SYSRET64
337END(entry_SYSCALL_64)
338
339/*
340 * %rdi: prev task
341 * %rsi: next task
342 */
343ENTRY(__switch_to_asm)
344	UNWIND_HINT_FUNC
345	/*
346	 * Save callee-saved registers
347	 * This must match the order in inactive_task_frame
348	 */
349	pushq	%rbp
350	pushq	%rbx
351	pushq	%r12
352	pushq	%r13
353	pushq	%r14
354	pushq	%r15
355
356	/* switch stack */
357	movq	%rsp, TASK_threadsp(%rdi)
358	movq	TASK_threadsp(%rsi), %rsp
359
360#ifdef CONFIG_STACKPROTECTOR
361	movq	TASK_stack_canary(%rsi), %rbx
362	movq	%rbx, PER_CPU_VAR(irq_stack_union)+stack_canary_offset
363#endif
364
365#ifdef CONFIG_RETPOLINE
366	/*
367	 * When switching from a shallower to a deeper call stack
368	 * the RSB may either underflow or use entries populated
369	 * with userspace addresses. On CPUs where those concerns
370	 * exist, overwrite the RSB with entries which capture
371	 * speculative execution to prevent attack.
372	 */
373	FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW
374#endif
375
376	/* restore callee-saved registers */
377	popq	%r15
378	popq	%r14
379	popq	%r13
380	popq	%r12
381	popq	%rbx
382	popq	%rbp
383
384	jmp	__switch_to
385END(__switch_to_asm)
386
387/*
388 * A newly forked process directly context switches into this address.
389 *
390 * rax: prev task we switched from
391 * rbx: kernel thread func (NULL for user thread)
392 * r12: kernel thread arg
393 */
394ENTRY(ret_from_fork)
395	UNWIND_HINT_EMPTY
396	movq	%rax, %rdi
397	call	schedule_tail			/* rdi: 'prev' task parameter */
398
399	testq	%rbx, %rbx			/* from kernel_thread? */
400	jnz	1f				/* kernel threads are uncommon */
401
4022:
403	UNWIND_HINT_REGS
404	movq	%rsp, %rdi
405	call	syscall_return_slowpath	/* returns with IRQs disabled */
406	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
407	jmp	swapgs_restore_regs_and_return_to_usermode
408
4091:
410	/* kernel thread */
411	UNWIND_HINT_EMPTY
412	movq	%r12, %rdi
413	CALL_NOSPEC %rbx
414	/*
415	 * A kernel thread is allowed to return here after successfully
416	 * calling do_execve().  Exit to userspace to complete the execve()
417	 * syscall.
418	 */
419	movq	$0, RAX(%rsp)
420	jmp	2b
421END(ret_from_fork)
422
423/*
424 * Build the entry stubs with some assembler magic.
425 * We pack 1 stub into every 8-byte block.
426 */
427	.align 8
428ENTRY(irq_entries_start)
429    vector=FIRST_EXTERNAL_VECTOR
430    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
431	UNWIND_HINT_IRET_REGS
432	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
433	jmp	common_interrupt
434	.align	8
435	vector=vector+1
436    .endr
437END(irq_entries_start)
438
439.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
440#ifdef CONFIG_DEBUG_ENTRY
441	pushq %rax
442	SAVE_FLAGS(CLBR_RAX)
443	testl $X86_EFLAGS_IF, %eax
444	jz .Lokay_\@
445	ud2
446.Lokay_\@:
447	popq %rax
448#endif
449.endm
450
451/*
452 * Enters the IRQ stack if we're not already using it.  NMI-safe.  Clobbers
453 * flags and puts old RSP into old_rsp, and leaves all other GPRs alone.
454 * Requires kernel GSBASE.
455 *
456 * The invariant is that, if irq_count != -1, then the IRQ stack is in use.
457 */
458.macro ENTER_IRQ_STACK regs=1 old_rsp save_ret=0
459	DEBUG_ENTRY_ASSERT_IRQS_OFF
460
461	.if \save_ret
462	/*
463	 * If save_ret is set, the original stack contains one additional
464	 * entry -- the return address. Therefore, move the address one
465	 * entry below %rsp to \old_rsp.
466	 */
467	leaq	8(%rsp), \old_rsp
468	.else
469	movq	%rsp, \old_rsp
470	.endif
471
472	.if \regs
473	UNWIND_HINT_REGS base=\old_rsp
474	.endif
475
476	incl	PER_CPU_VAR(irq_count)
477	jnz	.Lirq_stack_push_old_rsp_\@
478
479	/*
480	 * Right now, if we just incremented irq_count to zero, we've
481	 * claimed the IRQ stack but we haven't switched to it yet.
482	 *
483	 * If anything is added that can interrupt us here without using IST,
484	 * it must be *extremely* careful to limit its stack usage.  This
485	 * could include kprobes and a hypothetical future IST-less #DB
486	 * handler.
487	 *
488	 * The OOPS unwinder relies on the word at the top of the IRQ
489	 * stack linking back to the previous RSP for the entire time we're
490	 * on the IRQ stack.  For this to work reliably, we need to write
491	 * it before we actually move ourselves to the IRQ stack.
492	 */
493
494	movq	\old_rsp, PER_CPU_VAR(irq_stack_union + IRQ_STACK_SIZE - 8)
495	movq	PER_CPU_VAR(irq_stack_ptr), %rsp
496
497#ifdef CONFIG_DEBUG_ENTRY
498	/*
499	 * If the first movq above becomes wrong due to IRQ stack layout
500	 * changes, the only way we'll notice is if we try to unwind right
501	 * here.  Assert that we set up the stack right to catch this type
502	 * of bug quickly.
503	 */
504	cmpq	-8(%rsp), \old_rsp
505	je	.Lirq_stack_okay\@
506	ud2
507	.Lirq_stack_okay\@:
508#endif
509
510.Lirq_stack_push_old_rsp_\@:
511	pushq	\old_rsp
512
513	.if \regs
514	UNWIND_HINT_REGS indirect=1
515	.endif
516
517	.if \save_ret
518	/*
519	 * Push the return address to the stack. This return address can
520	 * be found at the "real" original RSP, which was offset by 8 at
521	 * the beginning of this macro.
522	 */
523	pushq	-8(\old_rsp)
524	.endif
525.endm
526
527/*
528 * Undoes ENTER_IRQ_STACK.
529 */
530.macro LEAVE_IRQ_STACK regs=1
531	DEBUG_ENTRY_ASSERT_IRQS_OFF
532	/* We need to be off the IRQ stack before decrementing irq_count. */
533	popq	%rsp
534
535	.if \regs
536	UNWIND_HINT_REGS
537	.endif
538
539	/*
540	 * As in ENTER_IRQ_STACK, irq_count == 0, we are still claiming
541	 * the irq stack but we're not on it.
542	 */
543
544	decl	PER_CPU_VAR(irq_count)
545.endm
546
547/*
548 * Interrupt entry helper function.
549 *
550 * Entry runs with interrupts off. Stack layout at entry:
551 * +----------------------------------------------------+
552 * | regs->ss						|
553 * | regs->rsp						|
554 * | regs->eflags					|
555 * | regs->cs						|
556 * | regs->ip						|
557 * +----------------------------------------------------+
558 * | regs->orig_ax = ~(interrupt number)		|
559 * +----------------------------------------------------+
560 * | return address					|
561 * +----------------------------------------------------+
562 */
563ENTRY(interrupt_entry)
564	UNWIND_HINT_FUNC
565	ASM_CLAC
566	cld
567
568	testb	$3, CS-ORIG_RAX+8(%rsp)
569	jz	1f
570	SWAPGS
571
572	/*
573	 * Switch to the thread stack. The IRET frame and orig_ax are
574	 * on the stack, as well as the return address. RDI..R12 are
575	 * not (yet) on the stack and space has not (yet) been
576	 * allocated for them.
577	 */
578	pushq	%rdi
579
580	/* Need to switch before accessing the thread stack. */
581	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi
582	movq	%rsp, %rdi
583	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
584
585	 /*
586	  * We have RDI, return address, and orig_ax on the stack on
587	  * top of the IRET frame. That means offset=24
588	  */
589	UNWIND_HINT_IRET_REGS base=%rdi offset=24
590
591	pushq	7*8(%rdi)		/* regs->ss */
592	pushq	6*8(%rdi)		/* regs->rsp */
593	pushq	5*8(%rdi)		/* regs->eflags */
594	pushq	4*8(%rdi)		/* regs->cs */
595	pushq	3*8(%rdi)		/* regs->ip */
596	pushq	2*8(%rdi)		/* regs->orig_ax */
597	pushq	8(%rdi)			/* return address */
598	UNWIND_HINT_FUNC
599
600	movq	(%rdi), %rdi
6011:
602
603	PUSH_AND_CLEAR_REGS save_ret=1
604	ENCODE_FRAME_POINTER 8
605
606	testb	$3, CS+8(%rsp)
607	jz	1f
608
609	/*
610	 * IRQ from user mode.
611	 *
612	 * We need to tell lockdep that IRQs are off.  We can't do this until
613	 * we fix gsbase, and we should do it before enter_from_user_mode
614	 * (which can take locks).  Since TRACE_IRQS_OFF is idempotent,
615	 * the simplest way to handle it is to just call it twice if
616	 * we enter from user mode.  There's no reason to optimize this since
617	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
618	 */
619	TRACE_IRQS_OFF
620
621	CALL_enter_from_user_mode
622
6231:
624	ENTER_IRQ_STACK old_rsp=%rdi save_ret=1
625	/* We entered an interrupt context - irqs are off: */
626	TRACE_IRQS_OFF
627
628	ret
629END(interrupt_entry)
630
631
632/* Interrupt entry/exit. */
633
634	/*
635	 * The interrupt stubs push (~vector+0x80) onto the stack and
636	 * then jump to common_interrupt.
637	 */
638	.p2align CONFIG_X86_L1_CACHE_SHIFT
639common_interrupt:
640	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
641	call	interrupt_entry
642	UNWIND_HINT_REGS indirect=1
643	call	do_IRQ	/* rdi points to pt_regs */
644	/* 0(%rsp): old RSP */
645ret_from_intr:
646	DISABLE_INTERRUPTS(CLBR_ANY)
647	TRACE_IRQS_OFF
648
649	LEAVE_IRQ_STACK
650
651	testb	$3, CS(%rsp)
652	jz	retint_kernel
653
654	/* Interrupt came from user space */
655GLOBAL(retint_user)
656	mov	%rsp,%rdi
657	call	prepare_exit_to_usermode
658	TRACE_IRQS_IRETQ
659
660GLOBAL(swapgs_restore_regs_and_return_to_usermode)
661#ifdef CONFIG_DEBUG_ENTRY
662	/* Assert that pt_regs indicates user mode. */
663	testb	$3, CS(%rsp)
664	jnz	1f
665	ud2
6661:
667#endif
668	POP_REGS pop_rdi=0
669
670	/*
671	 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS.
672	 * Save old stack pointer and switch to trampoline stack.
673	 */
674	movq	%rsp, %rdi
675	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
676
677	/* Copy the IRET frame to the trampoline stack. */
678	pushq	6*8(%rdi)	/* SS */
679	pushq	5*8(%rdi)	/* RSP */
680	pushq	4*8(%rdi)	/* EFLAGS */
681	pushq	3*8(%rdi)	/* CS */
682	pushq	2*8(%rdi)	/* RIP */
683
684	/* Push user RDI on the trampoline stack. */
685	pushq	(%rdi)
686
687	/*
688	 * We are on the trampoline stack.  All regs except RDI are live.
689	 * We can do future final exit work right here.
690	 */
691
692	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
693
694	/* Restore RDI. */
695	popq	%rdi
696	SWAPGS
697	INTERRUPT_RETURN
698
699
700/* Returning to kernel space */
701retint_kernel:
702#ifdef CONFIG_PREEMPT
703	/* Interrupts are off */
704	/* Check if we need preemption */
705	btl	$9, EFLAGS(%rsp)		/* were interrupts off? */
706	jnc	1f
7070:	cmpl	$0, PER_CPU_VAR(__preempt_count)
708	jnz	1f
709	call	preempt_schedule_irq
710	jmp	0b
7111:
712#endif
713	/*
714	 * The iretq could re-enable interrupts:
715	 */
716	TRACE_IRQS_IRETQ
717
718GLOBAL(restore_regs_and_return_to_kernel)
719#ifdef CONFIG_DEBUG_ENTRY
720	/* Assert that pt_regs indicates kernel mode. */
721	testb	$3, CS(%rsp)
722	jz	1f
723	ud2
7241:
725#endif
726	POP_REGS
727	addq	$8, %rsp	/* skip regs->orig_ax */
728	/*
729	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
730	 * when returning from IPI handler.
731	 */
732	INTERRUPT_RETURN
733
734ENTRY(native_iret)
735	UNWIND_HINT_IRET_REGS
736	/*
737	 * Are we returning to a stack segment from the LDT?  Note: in
738	 * 64-bit mode SS:RSP on the exception stack is always valid.
739	 */
740#ifdef CONFIG_X86_ESPFIX64
741	testb	$4, (SS-RIP)(%rsp)
742	jnz	native_irq_return_ldt
743#endif
744
745.global native_irq_return_iret
746native_irq_return_iret:
747	/*
748	 * This may fault.  Non-paranoid faults on return to userspace are
749	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
750	 * Double-faults due to espfix64 are handled in do_double_fault.
751	 * Other faults here are fatal.
752	 */
753	iretq
754
755#ifdef CONFIG_X86_ESPFIX64
756native_irq_return_ldt:
757	/*
758	 * We are running with user GSBASE.  All GPRs contain their user
759	 * values.  We have a percpu ESPFIX stack that is eight slots
760	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
761	 * of the ESPFIX stack.
762	 *
763	 * We clobber RAX and RDI in this code.  We stash RDI on the
764	 * normal stack and RAX on the ESPFIX stack.
765	 *
766	 * The ESPFIX stack layout we set up looks like this:
767	 *
768	 * --- top of ESPFIX stack ---
769	 * SS
770	 * RSP
771	 * RFLAGS
772	 * CS
773	 * RIP  <-- RSP points here when we're done
774	 * RAX  <-- espfix_waddr points here
775	 * --- bottom of ESPFIX stack ---
776	 */
777
778	pushq	%rdi				/* Stash user RDI */
779	SWAPGS					/* to kernel GS */
780	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi	/* to kernel CR3 */
781
782	movq	PER_CPU_VAR(espfix_waddr), %rdi
783	movq	%rax, (0*8)(%rdi)		/* user RAX */
784	movq	(1*8)(%rsp), %rax		/* user RIP */
785	movq	%rax, (1*8)(%rdi)
786	movq	(2*8)(%rsp), %rax		/* user CS */
787	movq	%rax, (2*8)(%rdi)
788	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
789	movq	%rax, (3*8)(%rdi)
790	movq	(5*8)(%rsp), %rax		/* user SS */
791	movq	%rax, (5*8)(%rdi)
792	movq	(4*8)(%rsp), %rax		/* user RSP */
793	movq	%rax, (4*8)(%rdi)
794	/* Now RAX == RSP. */
795
796	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
797
798	/*
799	 * espfix_stack[31:16] == 0.  The page tables are set up such that
800	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
801	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
802	 * the same page.  Set up RSP so that RSP[31:16] contains the
803	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
804	 * still points to an RO alias of the ESPFIX stack.
805	 */
806	orq	PER_CPU_VAR(espfix_stack), %rax
807
808	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
809	SWAPGS					/* to user GS */
810	popq	%rdi				/* Restore user RDI */
811
812	movq	%rax, %rsp
813	UNWIND_HINT_IRET_REGS offset=8
814
815	/*
816	 * At this point, we cannot write to the stack any more, but we can
817	 * still read.
818	 */
819	popq	%rax				/* Restore user RAX */
820
821	/*
822	 * RSP now points to an ordinary IRET frame, except that the page
823	 * is read-only and RSP[31:16] are preloaded with the userspace
824	 * values.  We can now IRET back to userspace.
825	 */
826	jmp	native_irq_return_iret
827#endif
828END(common_interrupt)
829
830/*
831 * APIC interrupts.
832 */
833.macro apicinterrupt3 num sym do_sym
834ENTRY(\sym)
835	UNWIND_HINT_IRET_REGS
836	pushq	$~(\num)
837.Lcommon_\sym:
838	call	interrupt_entry
839	UNWIND_HINT_REGS indirect=1
840	call	\do_sym	/* rdi points to pt_regs */
841	jmp	ret_from_intr
842END(\sym)
843.endm
844
845/* Make sure APIC interrupt handlers end up in the irqentry section: */
846#define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
847#define POP_SECTION_IRQENTRY	.popsection
848
849.macro apicinterrupt num sym do_sym
850PUSH_SECTION_IRQENTRY
851apicinterrupt3 \num \sym \do_sym
852POP_SECTION_IRQENTRY
853.endm
854
855#ifdef CONFIG_SMP
856apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
857apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
858#endif
859
860#ifdef CONFIG_X86_UV
861apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
862#endif
863
864apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
865apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
866
867#ifdef CONFIG_HAVE_KVM
868apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
869apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
870apicinterrupt3 POSTED_INTR_NESTED_VECTOR	kvm_posted_intr_nested_ipi	smp_kvm_posted_intr_nested_ipi
871#endif
872
873#ifdef CONFIG_X86_MCE_THRESHOLD
874apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
875#endif
876
877#ifdef CONFIG_X86_MCE_AMD
878apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
879#endif
880
881#ifdef CONFIG_X86_THERMAL_VECTOR
882apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
883#endif
884
885#ifdef CONFIG_SMP
886apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
887apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
888apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
889#endif
890
891apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
892apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
893
894#ifdef CONFIG_IRQ_WORK
895apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
896#endif
897
898/*
899 * Exception entry points.
900 */
901#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss_rw) + (TSS_ist + ((x) - 1) * 8)
902
903.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
904ENTRY(\sym)
905	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
906
907	/* Sanity check */
908	.if \shift_ist != -1 && \paranoid == 0
909	.error "using shift_ist requires paranoid=1"
910	.endif
911
912	ASM_CLAC
913
914	.if \has_error_code == 0
915	pushq	$-1				/* ORIG_RAX: no syscall to restart */
916	.endif
917
918	.if \paranoid == 1
919	testb	$3, CS-ORIG_RAX(%rsp)		/* If coming from userspace, switch stacks */
920	jnz	.Lfrom_usermode_switch_stack_\@
921	.endif
922
923	.if \paranoid
924	call	paranoid_entry
925	.else
926	call	error_entry
927	.endif
928	UNWIND_HINT_REGS
929	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
930
931	.if \paranoid
932	.if \shift_ist != -1
933	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
934	.else
935	TRACE_IRQS_OFF
936	.endif
937	.endif
938
939	movq	%rsp, %rdi			/* pt_regs pointer */
940
941	.if \has_error_code
942	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
943	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
944	.else
945	xorl	%esi, %esi			/* no error code */
946	.endif
947
948	.if \shift_ist != -1
949	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
950	.endif
951
952	call	\do_sym
953
954	.if \shift_ist != -1
955	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
956	.endif
957
958	/* these procedures expect "no swapgs" flag in ebx */
959	.if \paranoid
960	jmp	paranoid_exit
961	.else
962	jmp	error_exit
963	.endif
964
965	.if \paranoid == 1
966	/*
967	 * Entry from userspace.  Switch stacks and treat it
968	 * as a normal entry.  This means that paranoid handlers
969	 * run in real process context if user_mode(regs).
970	 */
971.Lfrom_usermode_switch_stack_\@:
972	call	error_entry
973
974	movq	%rsp, %rdi			/* pt_regs pointer */
975
976	.if \has_error_code
977	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
978	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
979	.else
980	xorl	%esi, %esi			/* no error code */
981	.endif
982
983	call	\do_sym
984
985	jmp	error_exit
986	.endif
987END(\sym)
988.endm
989
990idtentry divide_error			do_divide_error			has_error_code=0
991idtentry overflow			do_overflow			has_error_code=0
992idtentry bounds				do_bounds			has_error_code=0
993idtentry invalid_op			do_invalid_op			has_error_code=0
994idtentry device_not_available		do_device_not_available		has_error_code=0
995idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
996idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
997idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
998idtentry segment_not_present		do_segment_not_present		has_error_code=1
999idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
1000idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
1001idtentry alignment_check		do_alignment_check		has_error_code=1
1002idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
1003
1004
1005	/*
1006	 * Reload gs selector with exception handling
1007	 * edi:  new selector
1008	 */
1009ENTRY(native_load_gs_index)
1010	FRAME_BEGIN
1011	pushfq
1012	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
1013	TRACE_IRQS_OFF
1014	SWAPGS
1015.Lgs_change:
1016	movl	%edi, %gs
10172:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
1018	SWAPGS
1019	TRACE_IRQS_FLAGS (%rsp)
1020	popfq
1021	FRAME_END
1022	ret
1023ENDPROC(native_load_gs_index)
1024EXPORT_SYMBOL(native_load_gs_index)
1025
1026	_ASM_EXTABLE(.Lgs_change, bad_gs)
1027	.section .fixup, "ax"
1028	/* running with kernelgs */
1029bad_gs:
1030	SWAPGS					/* switch back to user gs */
1031.macro ZAP_GS
1032	/* This can't be a string because the preprocessor needs to see it. */
1033	movl $__USER_DS, %eax
1034	movl %eax, %gs
1035.endm
1036	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
1037	xorl	%eax, %eax
1038	movl	%eax, %gs
1039	jmp	2b
1040	.previous
1041
1042/* Call softirq on interrupt stack. Interrupts are off. */
1043ENTRY(do_softirq_own_stack)
1044	pushq	%rbp
1045	mov	%rsp, %rbp
1046	ENTER_IRQ_STACK regs=0 old_rsp=%r11
1047	call	__do_softirq
1048	LEAVE_IRQ_STACK regs=0
1049	leaveq
1050	ret
1051ENDPROC(do_softirq_own_stack)
1052
1053#ifdef CONFIG_XEN
1054idtentry hypervisor_callback xen_do_hypervisor_callback has_error_code=0
1055
1056/*
1057 * A note on the "critical region" in our callback handler.
1058 * We want to avoid stacking callback handlers due to events occurring
1059 * during handling of the last event. To do this, we keep events disabled
1060 * until we've done all processing. HOWEVER, we must enable events before
1061 * popping the stack frame (can't be done atomically) and so it would still
1062 * be possible to get enough handler activations to overflow the stack.
1063 * Although unlikely, bugs of that kind are hard to track down, so we'd
1064 * like to avoid the possibility.
1065 * So, on entry to the handler we detect whether we interrupted an
1066 * existing activation in its critical region -- if so, we pop the current
1067 * activation and restart the handler using the previous one.
1068 */
1069ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
1070
1071/*
1072 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
1073 * see the correct pointer to the pt_regs
1074 */
1075	UNWIND_HINT_FUNC
1076	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
1077	UNWIND_HINT_REGS
1078
1079	ENTER_IRQ_STACK old_rsp=%r10
1080	call	xen_evtchn_do_upcall
1081	LEAVE_IRQ_STACK
1082
1083#ifndef CONFIG_PREEMPT
1084	call	xen_maybe_preempt_hcall
1085#endif
1086	jmp	error_exit
1087END(xen_do_hypervisor_callback)
1088
1089/*
1090 * Hypervisor uses this for application faults while it executes.
1091 * We get here for two reasons:
1092 *  1. Fault while reloading DS, ES, FS or GS
1093 *  2. Fault while executing IRET
1094 * Category 1 we do not need to fix up as Xen has already reloaded all segment
1095 * registers that could be reloaded and zeroed the others.
1096 * Category 2 we fix up by killing the current process. We cannot use the
1097 * normal Linux return path in this case because if we use the IRET hypercall
1098 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
1099 * We distinguish between categories by comparing each saved segment register
1100 * with its current contents: any discrepancy means we in category 1.
1101 */
1102ENTRY(xen_failsafe_callback)
1103	UNWIND_HINT_EMPTY
1104	movl	%ds, %ecx
1105	cmpw	%cx, 0x10(%rsp)
1106	jne	1f
1107	movl	%es, %ecx
1108	cmpw	%cx, 0x18(%rsp)
1109	jne	1f
1110	movl	%fs, %ecx
1111	cmpw	%cx, 0x20(%rsp)
1112	jne	1f
1113	movl	%gs, %ecx
1114	cmpw	%cx, 0x28(%rsp)
1115	jne	1f
1116	/* All segments match their saved values => Category 2 (Bad IRET). */
1117	movq	(%rsp), %rcx
1118	movq	8(%rsp), %r11
1119	addq	$0x30, %rsp
1120	pushq	$0				/* RIP */
1121	UNWIND_HINT_IRET_REGS offset=8
1122	jmp	general_protection
11231:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
1124	movq	(%rsp), %rcx
1125	movq	8(%rsp), %r11
1126	addq	$0x30, %rsp
1127	UNWIND_HINT_IRET_REGS
1128	pushq	$-1 /* orig_ax = -1 => not a system call */
1129	PUSH_AND_CLEAR_REGS
1130	ENCODE_FRAME_POINTER
1131	jmp	error_exit
1132END(xen_failsafe_callback)
1133
1134apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1135	xen_hvm_callback_vector xen_evtchn_do_upcall
1136
1137#endif /* CONFIG_XEN */
1138
1139#if IS_ENABLED(CONFIG_HYPERV)
1140apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
1141	hyperv_callback_vector hyperv_vector_handler
1142
1143apicinterrupt3 HYPERV_REENLIGHTENMENT_VECTOR \
1144	hyperv_reenlightenment_vector hyperv_reenlightenment_intr
1145
1146apicinterrupt3 HYPERV_STIMER0_VECTOR \
1147	hv_stimer0_callback_vector hv_stimer0_vector_handler
1148#endif /* CONFIG_HYPERV */
1149
1150idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1151idtentry int3			do_int3			has_error_code=0
1152idtentry stack_segment		do_stack_segment	has_error_code=1
1153
1154#ifdef CONFIG_XEN
1155idtentry xennmi			do_nmi			has_error_code=0
1156idtentry xendebug		do_debug		has_error_code=0
1157idtentry xenint3		do_int3			has_error_code=0
1158#endif
1159
1160idtentry general_protection	do_general_protection	has_error_code=1
1161idtentry page_fault		do_page_fault		has_error_code=1
1162
1163#ifdef CONFIG_KVM_GUEST
1164idtentry async_page_fault	do_async_page_fault	has_error_code=1
1165#endif
1166
1167#ifdef CONFIG_X86_MCE
1168idtentry machine_check		do_mce			has_error_code=0	paranoid=1
1169#endif
1170
1171/*
1172 * Save all registers in pt_regs, and switch gs if needed.
1173 * Use slow, but surefire "are we in kernel?" check.
1174 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
1175 */
1176ENTRY(paranoid_entry)
1177	UNWIND_HINT_FUNC
1178	cld
1179	PUSH_AND_CLEAR_REGS save_ret=1
1180	ENCODE_FRAME_POINTER 8
1181	movl	$1, %ebx
1182	movl	$MSR_GS_BASE, %ecx
1183	rdmsr
1184	testl	%edx, %edx
1185	js	1f				/* negative -> in kernel */
1186	SWAPGS
1187	xorl	%ebx, %ebx
1188
11891:
1190	SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
1191
1192	ret
1193END(paranoid_entry)
1194
1195/*
1196 * "Paranoid" exit path from exception stack.  This is invoked
1197 * only on return from non-NMI IST interrupts that came
1198 * from kernel space.
1199 *
1200 * We may be returning to very strange contexts (e.g. very early
1201 * in syscall entry), so checking for preemption here would
1202 * be complicated.  Fortunately, we there's no good reason
1203 * to try to handle preemption here.
1204 *
1205 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1206 */
1207ENTRY(paranoid_exit)
1208	UNWIND_HINT_REGS
1209	DISABLE_INTERRUPTS(CLBR_ANY)
1210	TRACE_IRQS_OFF_DEBUG
1211	testl	%ebx, %ebx			/* swapgs needed? */
1212	jnz	.Lparanoid_exit_no_swapgs
1213	TRACE_IRQS_IRETQ
1214	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1215	SWAPGS_UNSAFE_STACK
1216	jmp	.Lparanoid_exit_restore
1217.Lparanoid_exit_no_swapgs:
1218	TRACE_IRQS_IRETQ_DEBUG
1219	RESTORE_CR3	scratch_reg=%rbx save_reg=%r14
1220.Lparanoid_exit_restore:
1221	jmp restore_regs_and_return_to_kernel
1222END(paranoid_exit)
1223
1224/*
1225 * Save all registers in pt_regs, and switch GS if needed.
1226 */
1227ENTRY(error_entry)
1228	UNWIND_HINT_FUNC
1229	cld
1230	PUSH_AND_CLEAR_REGS save_ret=1
1231	ENCODE_FRAME_POINTER 8
1232	testb	$3, CS+8(%rsp)
1233	jz	.Lerror_kernelspace
1234
1235	/*
1236	 * We entered from user mode or we're pretending to have entered
1237	 * from user mode due to an IRET fault.
1238	 */
1239	SWAPGS
1240	/* We have user CR3.  Change to kernel CR3. */
1241	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1242
1243.Lerror_entry_from_usermode_after_swapgs:
1244	/* Put us onto the real thread stack. */
1245	popq	%r12				/* save return addr in %12 */
1246	movq	%rsp, %rdi			/* arg0 = pt_regs pointer */
1247	call	sync_regs
1248	movq	%rax, %rsp			/* switch stack */
1249	ENCODE_FRAME_POINTER
1250	pushq	%r12
1251
1252	/*
1253	 * We need to tell lockdep that IRQs are off.  We can't do this until
1254	 * we fix gsbase, and we should do it before enter_from_user_mode
1255	 * (which can take locks).
1256	 */
1257	TRACE_IRQS_OFF
1258	CALL_enter_from_user_mode
1259	ret
1260
1261.Lerror_entry_done:
1262	TRACE_IRQS_OFF
1263	ret
1264
1265	/*
1266	 * There are two places in the kernel that can potentially fault with
1267	 * usergs. Handle them here.  B stepping K8s sometimes report a
1268	 * truncated RIP for IRET exceptions returning to compat mode. Check
1269	 * for these here too.
1270	 */
1271.Lerror_kernelspace:
1272	leaq	native_irq_return_iret(%rip), %rcx
1273	cmpq	%rcx, RIP+8(%rsp)
1274	je	.Lerror_bad_iret
1275	movl	%ecx, %eax			/* zero extend */
1276	cmpq	%rax, RIP+8(%rsp)
1277	je	.Lbstep_iret
1278	cmpq	$.Lgs_change, RIP+8(%rsp)
1279	jne	.Lerror_entry_done
1280
1281	/*
1282	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1283	 * gsbase and proceed.  We'll fix up the exception and land in
1284	 * .Lgs_change's error handler with kernel gsbase.
1285	 */
1286	SWAPGS
1287	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1288	jmp .Lerror_entry_done
1289
1290.Lbstep_iret:
1291	/* Fix truncated RIP */
1292	movq	%rcx, RIP+8(%rsp)
1293	/* fall through */
1294
1295.Lerror_bad_iret:
1296	/*
1297	 * We came from an IRET to user mode, so we have user
1298	 * gsbase and CR3.  Switch to kernel gsbase and CR3:
1299	 */
1300	SWAPGS
1301	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1302
1303	/*
1304	 * Pretend that the exception came from user mode: set up pt_regs
1305	 * as if we faulted immediately after IRET.
1306	 */
1307	mov	%rsp, %rdi
1308	call	fixup_bad_iret
1309	mov	%rax, %rsp
1310	jmp	.Lerror_entry_from_usermode_after_swapgs
1311END(error_entry)
1312
1313ENTRY(error_exit)
1314	UNWIND_HINT_REGS
1315	DISABLE_INTERRUPTS(CLBR_ANY)
1316	TRACE_IRQS_OFF
1317	testb	$3, CS(%rsp)
1318	jz	retint_kernel
1319	jmp	retint_user
1320END(error_exit)
1321
1322/*
1323 * Runs on exception stack.  Xen PV does not go through this path at all,
1324 * so we can use real assembly here.
1325 *
1326 * Registers:
1327 *	%r14: Used to save/restore the CR3 of the interrupted context
1328 *	      when PAGE_TABLE_ISOLATION is in use.  Do not clobber.
1329 */
1330ENTRY(nmi)
1331	UNWIND_HINT_IRET_REGS
1332
1333	/*
1334	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1335	 * the iretq it performs will take us out of NMI context.
1336	 * This means that we can have nested NMIs where the next
1337	 * NMI is using the top of the stack of the previous NMI. We
1338	 * can't let it execute because the nested NMI will corrupt the
1339	 * stack of the previous NMI. NMI handlers are not re-entrant
1340	 * anyway.
1341	 *
1342	 * To handle this case we do the following:
1343	 *  Check the a special location on the stack that contains
1344	 *  a variable that is set when NMIs are executing.
1345	 *  The interrupted task's stack is also checked to see if it
1346	 *  is an NMI stack.
1347	 *  If the variable is not set and the stack is not the NMI
1348	 *  stack then:
1349	 *    o Set the special variable on the stack
1350	 *    o Copy the interrupt frame into an "outermost" location on the
1351	 *      stack
1352	 *    o Copy the interrupt frame into an "iret" location on the stack
1353	 *    o Continue processing the NMI
1354	 *  If the variable is set or the previous stack is the NMI stack:
1355	 *    o Modify the "iret" location to jump to the repeat_nmi
1356	 *    o return back to the first NMI
1357	 *
1358	 * Now on exit of the first NMI, we first clear the stack variable
1359	 * The NMI stack will tell any nested NMIs at that point that it is
1360	 * nested. Then we pop the stack normally with iret, and if there was
1361	 * a nested NMI that updated the copy interrupt stack frame, a
1362	 * jump will be made to the repeat_nmi code that will handle the second
1363	 * NMI.
1364	 *
1365	 * However, espfix prevents us from directly returning to userspace
1366	 * with a single IRET instruction.  Similarly, IRET to user mode
1367	 * can fault.  We therefore handle NMIs from user space like
1368	 * other IST entries.
1369	 */
1370
1371	ASM_CLAC
1372
1373	/* Use %rdx as our temp variable throughout */
1374	pushq	%rdx
1375
1376	testb	$3, CS-RIP+8(%rsp)
1377	jz	.Lnmi_from_kernel
1378
1379	/*
1380	 * NMI from user mode.  We need to run on the thread stack, but we
1381	 * can't go through the normal entry paths: NMIs are masked, and
1382	 * we don't want to enable interrupts, because then we'll end
1383	 * up in an awkward situation in which IRQs are on but NMIs
1384	 * are off.
1385	 *
1386	 * We also must not push anything to the stack before switching
1387	 * stacks lest we corrupt the "NMI executing" variable.
1388	 */
1389
1390	swapgs
1391	cld
1392	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx
1393	movq	%rsp, %rdx
1394	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1395	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1396	pushq	5*8(%rdx)	/* pt_regs->ss */
1397	pushq	4*8(%rdx)	/* pt_regs->rsp */
1398	pushq	3*8(%rdx)	/* pt_regs->flags */
1399	pushq	2*8(%rdx)	/* pt_regs->cs */
1400	pushq	1*8(%rdx)	/* pt_regs->rip */
1401	UNWIND_HINT_IRET_REGS
1402	pushq   $-1		/* pt_regs->orig_ax */
1403	PUSH_AND_CLEAR_REGS rdx=(%rdx)
1404	ENCODE_FRAME_POINTER
1405
1406	/*
1407	 * At this point we no longer need to worry about stack damage
1408	 * due to nesting -- we're on the normal thread stack and we're
1409	 * done with the NMI stack.
1410	 */
1411
1412	movq	%rsp, %rdi
1413	movq	$-1, %rsi
1414	call	do_nmi
1415
1416	/*
1417	 * Return back to user mode.  We must *not* do the normal exit
1418	 * work, because we don't want to enable interrupts.
1419	 */
1420	jmp	swapgs_restore_regs_and_return_to_usermode
1421
1422.Lnmi_from_kernel:
1423	/*
1424	 * Here's what our stack frame will look like:
1425	 * +---------------------------------------------------------+
1426	 * | original SS                                             |
1427	 * | original Return RSP                                     |
1428	 * | original RFLAGS                                         |
1429	 * | original CS                                             |
1430	 * | original RIP                                            |
1431	 * +---------------------------------------------------------+
1432	 * | temp storage for rdx                                    |
1433	 * +---------------------------------------------------------+
1434	 * | "NMI executing" variable                                |
1435	 * +---------------------------------------------------------+
1436	 * | iret SS          } Copied from "outermost" frame        |
1437	 * | iret Return RSP  } on each loop iteration; overwritten  |
1438	 * | iret RFLAGS      } by a nested NMI to force another     |
1439	 * | iret CS          } iteration if needed.                 |
1440	 * | iret RIP         }                                      |
1441	 * +---------------------------------------------------------+
1442	 * | outermost SS          } initialized in first_nmi;       |
1443	 * | outermost Return RSP  } will not be changed before      |
1444	 * | outermost RFLAGS      } NMI processing is done.         |
1445	 * | outermost CS          } Copied to "iret" frame on each  |
1446	 * | outermost RIP         } iteration.                      |
1447	 * +---------------------------------------------------------+
1448	 * | pt_regs                                                 |
1449	 * +---------------------------------------------------------+
1450	 *
1451	 * The "original" frame is used by hardware.  Before re-enabling
1452	 * NMIs, we need to be done with it, and we need to leave enough
1453	 * space for the asm code here.
1454	 *
1455	 * We return by executing IRET while RSP points to the "iret" frame.
1456	 * That will either return for real or it will loop back into NMI
1457	 * processing.
1458	 *
1459	 * The "outermost" frame is copied to the "iret" frame on each
1460	 * iteration of the loop, so each iteration starts with the "iret"
1461	 * frame pointing to the final return target.
1462	 */
1463
1464	/*
1465	 * Determine whether we're a nested NMI.
1466	 *
1467	 * If we interrupted kernel code between repeat_nmi and
1468	 * end_repeat_nmi, then we are a nested NMI.  We must not
1469	 * modify the "iret" frame because it's being written by
1470	 * the outer NMI.  That's okay; the outer NMI handler is
1471	 * about to about to call do_nmi anyway, so we can just
1472	 * resume the outer NMI.
1473	 */
1474
1475	movq	$repeat_nmi, %rdx
1476	cmpq	8(%rsp), %rdx
1477	ja	1f
1478	movq	$end_repeat_nmi, %rdx
1479	cmpq	8(%rsp), %rdx
1480	ja	nested_nmi_out
14811:
1482
1483	/*
1484	 * Now check "NMI executing".  If it's set, then we're nested.
1485	 * This will not detect if we interrupted an outer NMI just
1486	 * before IRET.
1487	 */
1488	cmpl	$1, -8(%rsp)
1489	je	nested_nmi
1490
1491	/*
1492	 * Now test if the previous stack was an NMI stack.  This covers
1493	 * the case where we interrupt an outer NMI after it clears
1494	 * "NMI executing" but before IRET.  We need to be careful, though:
1495	 * there is one case in which RSP could point to the NMI stack
1496	 * despite there being no NMI active: naughty userspace controls
1497	 * RSP at the very beginning of the SYSCALL targets.  We can
1498	 * pull a fast one on naughty userspace, though: we program
1499	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1500	 * if it controls the kernel's RSP.  We set DF before we clear
1501	 * "NMI executing".
1502	 */
1503	lea	6*8(%rsp), %rdx
1504	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1505	cmpq	%rdx, 4*8(%rsp)
1506	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1507	ja	first_nmi
1508
1509	subq	$EXCEPTION_STKSZ, %rdx
1510	cmpq	%rdx, 4*8(%rsp)
1511	/* If it is below the NMI stack, it is a normal NMI */
1512	jb	first_nmi
1513
1514	/* Ah, it is within the NMI stack. */
1515
1516	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1517	jz	first_nmi	/* RSP was user controlled. */
1518
1519	/* This is a nested NMI. */
1520
1521nested_nmi:
1522	/*
1523	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1524	 * iteration of NMI handling.
1525	 */
1526	subq	$8, %rsp
1527	leaq	-10*8(%rsp), %rdx
1528	pushq	$__KERNEL_DS
1529	pushq	%rdx
1530	pushfq
1531	pushq	$__KERNEL_CS
1532	pushq	$repeat_nmi
1533
1534	/* Put stack back */
1535	addq	$(6*8), %rsp
1536
1537nested_nmi_out:
1538	popq	%rdx
1539
1540	/* We are returning to kernel mode, so this cannot result in a fault. */
1541	iretq
1542
1543first_nmi:
1544	/* Restore rdx. */
1545	movq	(%rsp), %rdx
1546
1547	/* Make room for "NMI executing". */
1548	pushq	$0
1549
1550	/* Leave room for the "iret" frame */
1551	subq	$(5*8), %rsp
1552
1553	/* Copy the "original" frame to the "outermost" frame */
1554	.rept 5
1555	pushq	11*8(%rsp)
1556	.endr
1557	UNWIND_HINT_IRET_REGS
1558
1559	/* Everything up to here is safe from nested NMIs */
1560
1561#ifdef CONFIG_DEBUG_ENTRY
1562	/*
1563	 * For ease of testing, unmask NMIs right away.  Disabled by
1564	 * default because IRET is very expensive.
1565	 */
1566	pushq	$0		/* SS */
1567	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1568	addq	$8, (%rsp)	/* Fix up RSP */
1569	pushfq			/* RFLAGS */
1570	pushq	$__KERNEL_CS	/* CS */
1571	pushq	$1f		/* RIP */
1572	iretq			/* continues at repeat_nmi below */
1573	UNWIND_HINT_IRET_REGS
15741:
1575#endif
1576
1577repeat_nmi:
1578	/*
1579	 * If there was a nested NMI, the first NMI's iret will return
1580	 * here. But NMIs are still enabled and we can take another
1581	 * nested NMI. The nested NMI checks the interrupted RIP to see
1582	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1583	 * it will just return, as we are about to repeat an NMI anyway.
1584	 * This makes it safe to copy to the stack frame that a nested
1585	 * NMI will update.
1586	 *
1587	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1588	 * we're repeating an NMI, gsbase has the same value that it had on
1589	 * the first iteration.  paranoid_entry will load the kernel
1590	 * gsbase if needed before we call do_nmi.  "NMI executing"
1591	 * is zero.
1592	 */
1593	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1594
1595	/*
1596	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1597	 * here must not modify the "iret" frame while we're writing to
1598	 * it or it will end up containing garbage.
1599	 */
1600	addq	$(10*8), %rsp
1601	.rept 5
1602	pushq	-6*8(%rsp)
1603	.endr
1604	subq	$(5*8), %rsp
1605end_repeat_nmi:
1606
1607	/*
1608	 * Everything below this point can be preempted by a nested NMI.
1609	 * If this happens, then the inner NMI will change the "iret"
1610	 * frame to point back to repeat_nmi.
1611	 */
1612	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1613
1614	/*
1615	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1616	 * as we should not be calling schedule in NMI context.
1617	 * Even with normal interrupts enabled. An NMI should not be
1618	 * setting NEED_RESCHED or anything that normal interrupts and
1619	 * exceptions might do.
1620	 */
1621	call	paranoid_entry
1622	UNWIND_HINT_REGS
1623
1624	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1625	movq	%rsp, %rdi
1626	movq	$-1, %rsi
1627	call	do_nmi
1628
1629	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
1630
1631	testl	%ebx, %ebx			/* swapgs needed? */
1632	jnz	nmi_restore
1633nmi_swapgs:
1634	SWAPGS_UNSAFE_STACK
1635nmi_restore:
1636	POP_REGS
1637
1638	/*
1639	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1640	 * at the "iret" frame.
1641	 */
1642	addq	$6*8, %rsp
1643
1644	/*
1645	 * Clear "NMI executing".  Set DF first so that we can easily
1646	 * distinguish the remaining code between here and IRET from
1647	 * the SYSCALL entry and exit paths.
1648	 *
1649	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1650	 * this code when I had the misapprehension that Xen PV supported
1651	 * NMIs, and Xen PV would break that approach.
1652	 */
1653	std
1654	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1655
1656	/*
1657	 * iretq reads the "iret" frame and exits the NMI stack in a
1658	 * single instruction.  We are returning to kernel mode, so this
1659	 * cannot result in a fault.  Similarly, we don't need to worry
1660	 * about espfix64 on the way back to kernel mode.
1661	 */
1662	iretq
1663END(nmi)
1664
1665ENTRY(ignore_sysret)
1666	UNWIND_HINT_EMPTY
1667	mov	$-ENOSYS, %eax
1668	sysret
1669END(ignore_sysret)
1670
1671ENTRY(rewind_stack_do_exit)
1672	UNWIND_HINT_FUNC
1673	/* Prevent any naive code from trying to unwind to our caller. */
1674	xorl	%ebp, %ebp
1675
1676	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1677	leaq	-PTREGS_SIZE(%rax), %rsp
1678	UNWIND_HINT_FUNC sp_offset=PTREGS_SIZE
1679
1680	call	do_exit
1681END(rewind_stack_do_exit)
1682