xref: /openbmc/linux/arch/x86/entry/entry_64.S (revision 8730046c)
1/*
2 *  linux/arch/x86_64/entry.S
3 *
4 *  Copyright (C) 1991, 1992  Linus Torvalds
5 *  Copyright (C) 2000, 2001, 2002  Andi Kleen SuSE Labs
6 *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
7 *
8 * entry.S contains the system-call and fault low-level handling routines.
9 *
10 * Some of this is documented in Documentation/x86/entry_64.txt
11 *
12 * A note on terminology:
13 * - iret frame:	Architecture defined interrupt frame from SS to RIP
14 *			at the top of the kernel process stack.
15 *
16 * Some macro usage:
17 * - ENTRY/END:		Define functions in the symbol table.
18 * - TRACE_IRQ_*:	Trace hardirq state for lock debugging.
19 * - idtentry:		Define exception entry points.
20 */
21#include <linux/linkage.h>
22#include <asm/segment.h>
23#include <asm/cache.h>
24#include <asm/errno.h>
25#include "calling.h"
26#include <asm/asm-offsets.h>
27#include <asm/msr.h>
28#include <asm/unistd.h>
29#include <asm/thread_info.h>
30#include <asm/hw_irq.h>
31#include <asm/page_types.h>
32#include <asm/irqflags.h>
33#include <asm/paravirt.h>
34#include <asm/percpu.h>
35#include <asm/asm.h>
36#include <asm/smap.h>
37#include <asm/pgtable_types.h>
38#include <asm/export.h>
39#include <asm/frame.h>
40#include <linux/err.h>
41
42.code64
43.section .entry.text, "ax"
44
45#ifdef CONFIG_PARAVIRT
46ENTRY(native_usergs_sysret64)
47	swapgs
48	sysretq
49ENDPROC(native_usergs_sysret64)
50#endif /* CONFIG_PARAVIRT */
51
52.macro TRACE_IRQS_IRETQ
53#ifdef CONFIG_TRACE_IRQFLAGS
54	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
55	jnc	1f
56	TRACE_IRQS_ON
571:
58#endif
59.endm
60
61/*
62 * When dynamic function tracer is enabled it will add a breakpoint
63 * to all locations that it is about to modify, sync CPUs, update
64 * all the code, sync CPUs, then remove the breakpoints. In this time
65 * if lockdep is enabled, it might jump back into the debug handler
66 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF).
67 *
68 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to
69 * make sure the stack pointer does not get reset back to the top
70 * of the debug stack, and instead just reuses the current stack.
71 */
72#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS)
73
74.macro TRACE_IRQS_OFF_DEBUG
75	call	debug_stack_set_zero
76	TRACE_IRQS_OFF
77	call	debug_stack_reset
78.endm
79
80.macro TRACE_IRQS_ON_DEBUG
81	call	debug_stack_set_zero
82	TRACE_IRQS_ON
83	call	debug_stack_reset
84.endm
85
86.macro TRACE_IRQS_IRETQ_DEBUG
87	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
88	jnc	1f
89	TRACE_IRQS_ON_DEBUG
901:
91.endm
92
93#else
94# define TRACE_IRQS_OFF_DEBUG			TRACE_IRQS_OFF
95# define TRACE_IRQS_ON_DEBUG			TRACE_IRQS_ON
96# define TRACE_IRQS_IRETQ_DEBUG			TRACE_IRQS_IRETQ
97#endif
98
99/*
100 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
101 *
102 * This is the only entry point used for 64-bit system calls.  The
103 * hardware interface is reasonably well designed and the register to
104 * argument mapping Linux uses fits well with the registers that are
105 * available when SYSCALL is used.
106 *
107 * SYSCALL instructions can be found inlined in libc implementations as
108 * well as some other programs and libraries.  There are also a handful
109 * of SYSCALL instructions in the vDSO used, for example, as a
110 * clock_gettimeofday fallback.
111 *
112 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
113 * then loads new ss, cs, and rip from previously programmed MSRs.
114 * rflags gets masked by a value from another MSR (so CLD and CLAC
115 * are not needed). SYSCALL does not save anything on the stack
116 * and does not change rsp.
117 *
118 * Registers on entry:
119 * rax  system call number
120 * rcx  return address
121 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
122 * rdi  arg0
123 * rsi  arg1
124 * rdx  arg2
125 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
126 * r8   arg4
127 * r9   arg5
128 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
129 *
130 * Only called from user space.
131 *
132 * When user can change pt_regs->foo always force IRET. That is because
133 * it deals with uncanonical addresses better. SYSRET has trouble
134 * with them due to bugs in both AMD and Intel CPUs.
135 */
136
137ENTRY(entry_SYSCALL_64)
138	/*
139	 * Interrupts are off on entry.
140	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
141	 * it is too small to ever cause noticeable irq latency.
142	 */
143	SWAPGS_UNSAFE_STACK
144	/*
145	 * A hypervisor implementation might want to use a label
146	 * after the swapgs, so that it can do the swapgs
147	 * for the guest and jump here on syscall.
148	 */
149GLOBAL(entry_SYSCALL_64_after_swapgs)
150
151	movq	%rsp, PER_CPU_VAR(rsp_scratch)
152	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
153
154	TRACE_IRQS_OFF
155
156	/* Construct struct pt_regs on stack */
157	pushq	$__USER_DS			/* pt_regs->ss */
158	pushq	PER_CPU_VAR(rsp_scratch)	/* pt_regs->sp */
159	pushq	%r11				/* pt_regs->flags */
160	pushq	$__USER_CS			/* pt_regs->cs */
161	pushq	%rcx				/* pt_regs->ip */
162	pushq	%rax				/* pt_regs->orig_ax */
163	pushq	%rdi				/* pt_regs->di */
164	pushq	%rsi				/* pt_regs->si */
165	pushq	%rdx				/* pt_regs->dx */
166	pushq	%rcx				/* pt_regs->cx */
167	pushq	$-ENOSYS			/* pt_regs->ax */
168	pushq	%r8				/* pt_regs->r8 */
169	pushq	%r9				/* pt_regs->r9 */
170	pushq	%r10				/* pt_regs->r10 */
171	pushq	%r11				/* pt_regs->r11 */
172	sub	$(6*8), %rsp			/* pt_regs->bp, bx, r12-15 not saved */
173
174	/*
175	 * If we need to do entry work or if we guess we'll need to do
176	 * exit work, go straight to the slow path.
177	 */
178	movq	PER_CPU_VAR(current_task), %r11
179	testl	$_TIF_WORK_SYSCALL_ENTRY|_TIF_ALLWORK_MASK, TASK_TI_flags(%r11)
180	jnz	entry_SYSCALL64_slow_path
181
182entry_SYSCALL_64_fastpath:
183	/*
184	 * Easy case: enable interrupts and issue the syscall.  If the syscall
185	 * needs pt_regs, we'll call a stub that disables interrupts again
186	 * and jumps to the slow path.
187	 */
188	TRACE_IRQS_ON
189	ENABLE_INTERRUPTS(CLBR_NONE)
190#if __SYSCALL_MASK == ~0
191	cmpq	$__NR_syscall_max, %rax
192#else
193	andl	$__SYSCALL_MASK, %eax
194	cmpl	$__NR_syscall_max, %eax
195#endif
196	ja	1f				/* return -ENOSYS (already in pt_regs->ax) */
197	movq	%r10, %rcx
198
199	/*
200	 * This call instruction is handled specially in stub_ptregs_64.
201	 * It might end up jumping to the slow path.  If it jumps, RAX
202	 * and all argument registers are clobbered.
203	 */
204	call	*sys_call_table(, %rax, 8)
205.Lentry_SYSCALL_64_after_fastpath_call:
206
207	movq	%rax, RAX(%rsp)
2081:
209
210	/*
211	 * If we get here, then we know that pt_regs is clean for SYSRET64.
212	 * If we see that no exit work is required (which we are required
213	 * to check with IRQs off), then we can go straight to SYSRET64.
214	 */
215	DISABLE_INTERRUPTS(CLBR_NONE)
216	TRACE_IRQS_OFF
217	movq	PER_CPU_VAR(current_task), %r11
218	testl	$_TIF_ALLWORK_MASK, TASK_TI_flags(%r11)
219	jnz	1f
220
221	LOCKDEP_SYS_EXIT
222	TRACE_IRQS_ON		/* user mode is traced as IRQs on */
223	movq	RIP(%rsp), %rcx
224	movq	EFLAGS(%rsp), %r11
225	RESTORE_C_REGS_EXCEPT_RCX_R11
226	movq	RSP(%rsp), %rsp
227	USERGS_SYSRET64
228
2291:
230	/*
231	 * The fast path looked good when we started, but something changed
232	 * along the way and we need to switch to the slow path.  Calling
233	 * raise(3) will trigger this, for example.  IRQs are off.
234	 */
235	TRACE_IRQS_ON
236	ENABLE_INTERRUPTS(CLBR_NONE)
237	SAVE_EXTRA_REGS
238	movq	%rsp, %rdi
239	call	syscall_return_slowpath	/* returns with IRQs disabled */
240	jmp	return_from_SYSCALL_64
241
242entry_SYSCALL64_slow_path:
243	/* IRQs are off. */
244	SAVE_EXTRA_REGS
245	movq	%rsp, %rdi
246	call	do_syscall_64		/* returns with IRQs disabled */
247
248return_from_SYSCALL_64:
249	RESTORE_EXTRA_REGS
250	TRACE_IRQS_IRETQ		/* we're about to change IF */
251
252	/*
253	 * Try to use SYSRET instead of IRET if we're returning to
254	 * a completely clean 64-bit userspace context.
255	 */
256	movq	RCX(%rsp), %rcx
257	movq	RIP(%rsp), %r11
258	cmpq	%rcx, %r11			/* RCX == RIP */
259	jne	opportunistic_sysret_failed
260
261	/*
262	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
263	 * in kernel space.  This essentially lets the user take over
264	 * the kernel, since userspace controls RSP.
265	 *
266	 * If width of "canonical tail" ever becomes variable, this will need
267	 * to be updated to remain correct on both old and new CPUs.
268	 */
269	.ifne __VIRTUAL_MASK_SHIFT - 47
270	.error "virtual address width changed -- SYSRET checks need update"
271	.endif
272
273	/* Change top 16 bits to be the sign-extension of 47th bit */
274	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
275	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
276
277	/* If this changed %rcx, it was not canonical */
278	cmpq	%rcx, %r11
279	jne	opportunistic_sysret_failed
280
281	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
282	jne	opportunistic_sysret_failed
283
284	movq	R11(%rsp), %r11
285	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
286	jne	opportunistic_sysret_failed
287
288	/*
289	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
290	 * restore RF properly. If the slowpath sets it for whatever reason, we
291	 * need to restore it correctly.
292	 *
293	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
294	 * trap from userspace immediately after SYSRET.  This would cause an
295	 * infinite loop whenever #DB happens with register state that satisfies
296	 * the opportunistic SYSRET conditions.  For example, single-stepping
297	 * this user code:
298	 *
299	 *           movq	$stuck_here, %rcx
300	 *           pushfq
301	 *           popq %r11
302	 *   stuck_here:
303	 *
304	 * would never get past 'stuck_here'.
305	 */
306	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
307	jnz	opportunistic_sysret_failed
308
309	/* nothing to check for RSP */
310
311	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
312	jne	opportunistic_sysret_failed
313
314	/*
315	 * We win! This label is here just for ease of understanding
316	 * perf profiles. Nothing jumps here.
317	 */
318syscall_return_via_sysret:
319	/* rcx and r11 are already restored (see code above) */
320	RESTORE_C_REGS_EXCEPT_RCX_R11
321	movq	RSP(%rsp), %rsp
322	USERGS_SYSRET64
323
324opportunistic_sysret_failed:
325	SWAPGS
326	jmp	restore_c_regs_and_iret
327END(entry_SYSCALL_64)
328
329ENTRY(stub_ptregs_64)
330	/*
331	 * Syscalls marked as needing ptregs land here.
332	 * If we are on the fast path, we need to save the extra regs,
333	 * which we achieve by trying again on the slow path.  If we are on
334	 * the slow path, the extra regs are already saved.
335	 *
336	 * RAX stores a pointer to the C function implementing the syscall.
337	 * IRQs are on.
338	 */
339	cmpq	$.Lentry_SYSCALL_64_after_fastpath_call, (%rsp)
340	jne	1f
341
342	/*
343	 * Called from fast path -- disable IRQs again, pop return address
344	 * and jump to slow path
345	 */
346	DISABLE_INTERRUPTS(CLBR_NONE)
347	TRACE_IRQS_OFF
348	popq	%rax
349	jmp	entry_SYSCALL64_slow_path
350
3511:
352	jmp	*%rax				/* Called from C */
353END(stub_ptregs_64)
354
355.macro ptregs_stub func
356ENTRY(ptregs_\func)
357	leaq	\func(%rip), %rax
358	jmp	stub_ptregs_64
359END(ptregs_\func)
360.endm
361
362/* Instantiate ptregs_stub for each ptregs-using syscall */
363#define __SYSCALL_64_QUAL_(sym)
364#define __SYSCALL_64_QUAL_ptregs(sym) ptregs_stub sym
365#define __SYSCALL_64(nr, sym, qual) __SYSCALL_64_QUAL_##qual(sym)
366#include <asm/syscalls_64.h>
367
368/*
369 * %rdi: prev task
370 * %rsi: next task
371 */
372ENTRY(__switch_to_asm)
373	/*
374	 * Save callee-saved registers
375	 * This must match the order in inactive_task_frame
376	 */
377	pushq	%rbp
378	pushq	%rbx
379	pushq	%r12
380	pushq	%r13
381	pushq	%r14
382	pushq	%r15
383
384	/* switch stack */
385	movq	%rsp, TASK_threadsp(%rdi)
386	movq	TASK_threadsp(%rsi), %rsp
387
388#ifdef CONFIG_CC_STACKPROTECTOR
389	movq	TASK_stack_canary(%rsi), %rbx
390	movq	%rbx, PER_CPU_VAR(irq_stack_union)+stack_canary_offset
391#endif
392
393	/* restore callee-saved registers */
394	popq	%r15
395	popq	%r14
396	popq	%r13
397	popq	%r12
398	popq	%rbx
399	popq	%rbp
400
401	jmp	__switch_to
402END(__switch_to_asm)
403
404/*
405 * A newly forked process directly context switches into this address.
406 *
407 * rax: prev task we switched from
408 * rbx: kernel thread func (NULL for user thread)
409 * r12: kernel thread arg
410 */
411ENTRY(ret_from_fork)
412	FRAME_BEGIN			/* help unwinder find end of stack */
413	movq	%rax, %rdi
414	call	schedule_tail		/* rdi: 'prev' task parameter */
415
416	testq	%rbx, %rbx		/* from kernel_thread? */
417	jnz	1f			/* kernel threads are uncommon */
418
4192:
420	leaq	FRAME_OFFSET(%rsp),%rdi	/* pt_regs pointer */
421	call	syscall_return_slowpath	/* returns with IRQs disabled */
422	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
423	SWAPGS
424	FRAME_END
425	jmp	restore_regs_and_iret
426
4271:
428	/* kernel thread */
429	movq	%r12, %rdi
430	call	*%rbx
431	/*
432	 * A kernel thread is allowed to return here after successfully
433	 * calling do_execve().  Exit to userspace to complete the execve()
434	 * syscall.
435	 */
436	movq	$0, RAX(%rsp)
437	jmp	2b
438END(ret_from_fork)
439
440/*
441 * Build the entry stubs with some assembler magic.
442 * We pack 1 stub into every 8-byte block.
443 */
444	.align 8
445ENTRY(irq_entries_start)
446    vector=FIRST_EXTERNAL_VECTOR
447    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
448	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
449    vector=vector+1
450	jmp	common_interrupt
451	.align	8
452    .endr
453END(irq_entries_start)
454
455/*
456 * Interrupt entry/exit.
457 *
458 * Interrupt entry points save only callee clobbered registers in fast path.
459 *
460 * Entry runs with interrupts off.
461 */
462
463/* 0(%rsp): ~(interrupt number) */
464	.macro interrupt func
465	cld
466	ALLOC_PT_GPREGS_ON_STACK
467	SAVE_C_REGS
468	SAVE_EXTRA_REGS
469	ENCODE_FRAME_POINTER
470
471	testb	$3, CS(%rsp)
472	jz	1f
473
474	/*
475	 * IRQ from user mode.  Switch to kernel gsbase and inform context
476	 * tracking that we're in kernel mode.
477	 */
478	SWAPGS
479
480	/*
481	 * We need to tell lockdep that IRQs are off.  We can't do this until
482	 * we fix gsbase, and we should do it before enter_from_user_mode
483	 * (which can take locks).  Since TRACE_IRQS_OFF idempotent,
484	 * the simplest way to handle it is to just call it twice if
485	 * we enter from user mode.  There's no reason to optimize this since
486	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
487	 */
488	TRACE_IRQS_OFF
489
490	CALL_enter_from_user_mode
491
4921:
493	/*
494	 * Save previous stack pointer, optionally switch to interrupt stack.
495	 * irq_count is used to check if a CPU is already on an interrupt stack
496	 * or not. While this is essentially redundant with preempt_count it is
497	 * a little cheaper to use a separate counter in the PDA (short of
498	 * moving irq_enter into assembly, which would be too much work)
499	 */
500	movq	%rsp, %rdi
501	incl	PER_CPU_VAR(irq_count)
502	cmovzq	PER_CPU_VAR(irq_stack_ptr), %rsp
503	pushq	%rdi
504	/* We entered an interrupt context - irqs are off: */
505	TRACE_IRQS_OFF
506
507	call	\func	/* rdi points to pt_regs */
508	.endm
509
510	/*
511	 * The interrupt stubs push (~vector+0x80) onto the stack and
512	 * then jump to common_interrupt.
513	 */
514	.p2align CONFIG_X86_L1_CACHE_SHIFT
515common_interrupt:
516	ASM_CLAC
517	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
518	interrupt do_IRQ
519	/* 0(%rsp): old RSP */
520ret_from_intr:
521	DISABLE_INTERRUPTS(CLBR_NONE)
522	TRACE_IRQS_OFF
523	decl	PER_CPU_VAR(irq_count)
524
525	/* Restore saved previous stack */
526	popq	%rsp
527
528	testb	$3, CS(%rsp)
529	jz	retint_kernel
530
531	/* Interrupt came from user space */
532GLOBAL(retint_user)
533	mov	%rsp,%rdi
534	call	prepare_exit_to_usermode
535	TRACE_IRQS_IRETQ
536	SWAPGS
537	jmp	restore_regs_and_iret
538
539/* Returning to kernel space */
540retint_kernel:
541#ifdef CONFIG_PREEMPT
542	/* Interrupts are off */
543	/* Check if we need preemption */
544	bt	$9, EFLAGS(%rsp)		/* were interrupts off? */
545	jnc	1f
5460:	cmpl	$0, PER_CPU_VAR(__preempt_count)
547	jnz	1f
548	call	preempt_schedule_irq
549	jmp	0b
5501:
551#endif
552	/*
553	 * The iretq could re-enable interrupts:
554	 */
555	TRACE_IRQS_IRETQ
556
557/*
558 * At this label, code paths which return to kernel and to user,
559 * which come from interrupts/exception and from syscalls, merge.
560 */
561GLOBAL(restore_regs_and_iret)
562	RESTORE_EXTRA_REGS
563restore_c_regs_and_iret:
564	RESTORE_C_REGS
565	REMOVE_PT_GPREGS_FROM_STACK 8
566	INTERRUPT_RETURN
567
568ENTRY(native_iret)
569	/*
570	 * Are we returning to a stack segment from the LDT?  Note: in
571	 * 64-bit mode SS:RSP on the exception stack is always valid.
572	 */
573#ifdef CONFIG_X86_ESPFIX64
574	testb	$4, (SS-RIP)(%rsp)
575	jnz	native_irq_return_ldt
576#endif
577
578.global native_irq_return_iret
579native_irq_return_iret:
580	/*
581	 * This may fault.  Non-paranoid faults on return to userspace are
582	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
583	 * Double-faults due to espfix64 are handled in do_double_fault.
584	 * Other faults here are fatal.
585	 */
586	iretq
587
588#ifdef CONFIG_X86_ESPFIX64
589native_irq_return_ldt:
590	/*
591	 * We are running with user GSBASE.  All GPRs contain their user
592	 * values.  We have a percpu ESPFIX stack that is eight slots
593	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
594	 * of the ESPFIX stack.
595	 *
596	 * We clobber RAX and RDI in this code.  We stash RDI on the
597	 * normal stack and RAX on the ESPFIX stack.
598	 *
599	 * The ESPFIX stack layout we set up looks like this:
600	 *
601	 * --- top of ESPFIX stack ---
602	 * SS
603	 * RSP
604	 * RFLAGS
605	 * CS
606	 * RIP  <-- RSP points here when we're done
607	 * RAX  <-- espfix_waddr points here
608	 * --- bottom of ESPFIX stack ---
609	 */
610
611	pushq	%rdi				/* Stash user RDI */
612	SWAPGS
613	movq	PER_CPU_VAR(espfix_waddr), %rdi
614	movq	%rax, (0*8)(%rdi)		/* user RAX */
615	movq	(1*8)(%rsp), %rax		/* user RIP */
616	movq	%rax, (1*8)(%rdi)
617	movq	(2*8)(%rsp), %rax		/* user CS */
618	movq	%rax, (2*8)(%rdi)
619	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
620	movq	%rax, (3*8)(%rdi)
621	movq	(5*8)(%rsp), %rax		/* user SS */
622	movq	%rax, (5*8)(%rdi)
623	movq	(4*8)(%rsp), %rax		/* user RSP */
624	movq	%rax, (4*8)(%rdi)
625	/* Now RAX == RSP. */
626
627	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
628	popq	%rdi				/* Restore user RDI */
629
630	/*
631	 * espfix_stack[31:16] == 0.  The page tables are set up such that
632	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
633	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
634	 * the same page.  Set up RSP so that RSP[31:16] contains the
635	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
636	 * still points to an RO alias of the ESPFIX stack.
637	 */
638	orq	PER_CPU_VAR(espfix_stack), %rax
639	SWAPGS
640	movq	%rax, %rsp
641
642	/*
643	 * At this point, we cannot write to the stack any more, but we can
644	 * still read.
645	 */
646	popq	%rax				/* Restore user RAX */
647
648	/*
649	 * RSP now points to an ordinary IRET frame, except that the page
650	 * is read-only and RSP[31:16] are preloaded with the userspace
651	 * values.  We can now IRET back to userspace.
652	 */
653	jmp	native_irq_return_iret
654#endif
655END(common_interrupt)
656
657/*
658 * APIC interrupts.
659 */
660.macro apicinterrupt3 num sym do_sym
661ENTRY(\sym)
662	ASM_CLAC
663	pushq	$~(\num)
664.Lcommon_\sym:
665	interrupt \do_sym
666	jmp	ret_from_intr
667END(\sym)
668.endm
669
670#ifdef CONFIG_TRACING
671#define trace(sym) trace_##sym
672#define smp_trace(sym) smp_trace_##sym
673
674.macro trace_apicinterrupt num sym
675apicinterrupt3 \num trace(\sym) smp_trace(\sym)
676.endm
677#else
678.macro trace_apicinterrupt num sym do_sym
679.endm
680#endif
681
682/* Make sure APIC interrupt handlers end up in the irqentry section: */
683#if defined(CONFIG_FUNCTION_GRAPH_TRACER) || defined(CONFIG_KASAN)
684# define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
685# define POP_SECTION_IRQENTRY	.popsection
686#else
687# define PUSH_SECTION_IRQENTRY
688# define POP_SECTION_IRQENTRY
689#endif
690
691.macro apicinterrupt num sym do_sym
692PUSH_SECTION_IRQENTRY
693apicinterrupt3 \num \sym \do_sym
694trace_apicinterrupt \num \sym
695POP_SECTION_IRQENTRY
696.endm
697
698#ifdef CONFIG_SMP
699apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
700apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
701#endif
702
703#ifdef CONFIG_X86_UV
704apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
705#endif
706
707apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
708apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
709
710#ifdef CONFIG_HAVE_KVM
711apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
712apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
713#endif
714
715#ifdef CONFIG_X86_MCE_THRESHOLD
716apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
717#endif
718
719#ifdef CONFIG_X86_MCE_AMD
720apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
721#endif
722
723#ifdef CONFIG_X86_THERMAL_VECTOR
724apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
725#endif
726
727#ifdef CONFIG_SMP
728apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
729apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
730apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
731#endif
732
733apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
734apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
735
736#ifdef CONFIG_IRQ_WORK
737apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
738#endif
739
740/*
741 * Exception entry points.
742 */
743#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss) + (TSS_ist + ((x) - 1) * 8)
744
745.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
746ENTRY(\sym)
747	/* Sanity check */
748	.if \shift_ist != -1 && \paranoid == 0
749	.error "using shift_ist requires paranoid=1"
750	.endif
751
752	ASM_CLAC
753	PARAVIRT_ADJUST_EXCEPTION_FRAME
754
755	.ifeq \has_error_code
756	pushq	$-1				/* ORIG_RAX: no syscall to restart */
757	.endif
758
759	ALLOC_PT_GPREGS_ON_STACK
760
761	.if \paranoid
762	.if \paranoid == 1
763	testb	$3, CS(%rsp)			/* If coming from userspace, switch stacks */
764	jnz	1f
765	.endif
766	call	paranoid_entry
767	.else
768	call	error_entry
769	.endif
770	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
771
772	.if \paranoid
773	.if \shift_ist != -1
774	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
775	.else
776	TRACE_IRQS_OFF
777	.endif
778	.endif
779
780	movq	%rsp, %rdi			/* pt_regs pointer */
781
782	.if \has_error_code
783	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
784	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
785	.else
786	xorl	%esi, %esi			/* no error code */
787	.endif
788
789	.if \shift_ist != -1
790	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
791	.endif
792
793	call	\do_sym
794
795	.if \shift_ist != -1
796	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
797	.endif
798
799	/* these procedures expect "no swapgs" flag in ebx */
800	.if \paranoid
801	jmp	paranoid_exit
802	.else
803	jmp	error_exit
804	.endif
805
806	.if \paranoid == 1
807	/*
808	 * Paranoid entry from userspace.  Switch stacks and treat it
809	 * as a normal entry.  This means that paranoid handlers
810	 * run in real process context if user_mode(regs).
811	 */
8121:
813	call	error_entry
814
815
816	movq	%rsp, %rdi			/* pt_regs pointer */
817	call	sync_regs
818	movq	%rax, %rsp			/* switch stack */
819
820	movq	%rsp, %rdi			/* pt_regs pointer */
821
822	.if \has_error_code
823	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
824	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
825	.else
826	xorl	%esi, %esi			/* no error code */
827	.endif
828
829	call	\do_sym
830
831	jmp	error_exit			/* %ebx: no swapgs flag */
832	.endif
833END(\sym)
834.endm
835
836#ifdef CONFIG_TRACING
837.macro trace_idtentry sym do_sym has_error_code:req
838idtentry trace(\sym) trace(\do_sym) has_error_code=\has_error_code
839idtentry \sym \do_sym has_error_code=\has_error_code
840.endm
841#else
842.macro trace_idtentry sym do_sym has_error_code:req
843idtentry \sym \do_sym has_error_code=\has_error_code
844.endm
845#endif
846
847idtentry divide_error			do_divide_error			has_error_code=0
848idtentry overflow			do_overflow			has_error_code=0
849idtentry bounds				do_bounds			has_error_code=0
850idtentry invalid_op			do_invalid_op			has_error_code=0
851idtentry device_not_available		do_device_not_available		has_error_code=0
852idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
853idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
854idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
855idtentry segment_not_present		do_segment_not_present		has_error_code=1
856idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
857idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
858idtentry alignment_check		do_alignment_check		has_error_code=1
859idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
860
861
862	/*
863	 * Reload gs selector with exception handling
864	 * edi:  new selector
865	 */
866ENTRY(native_load_gs_index)
867	pushfq
868	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
869	SWAPGS
870.Lgs_change:
871	movl	%edi, %gs
8722:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
873	SWAPGS
874	popfq
875	ret
876END(native_load_gs_index)
877EXPORT_SYMBOL(native_load_gs_index)
878
879	_ASM_EXTABLE(.Lgs_change, bad_gs)
880	.section .fixup, "ax"
881	/* running with kernelgs */
882bad_gs:
883	SWAPGS					/* switch back to user gs */
884.macro ZAP_GS
885	/* This can't be a string because the preprocessor needs to see it. */
886	movl $__USER_DS, %eax
887	movl %eax, %gs
888.endm
889	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
890	xorl	%eax, %eax
891	movl	%eax, %gs
892	jmp	2b
893	.previous
894
895/* Call softirq on interrupt stack. Interrupts are off. */
896ENTRY(do_softirq_own_stack)
897	pushq	%rbp
898	mov	%rsp, %rbp
899	incl	PER_CPU_VAR(irq_count)
900	cmove	PER_CPU_VAR(irq_stack_ptr), %rsp
901	push	%rbp				/* frame pointer backlink */
902	call	__do_softirq
903	leaveq
904	decl	PER_CPU_VAR(irq_count)
905	ret
906END(do_softirq_own_stack)
907
908#ifdef CONFIG_XEN
909idtentry xen_hypervisor_callback xen_do_hypervisor_callback has_error_code=0
910
911/*
912 * A note on the "critical region" in our callback handler.
913 * We want to avoid stacking callback handlers due to events occurring
914 * during handling of the last event. To do this, we keep events disabled
915 * until we've done all processing. HOWEVER, we must enable events before
916 * popping the stack frame (can't be done atomically) and so it would still
917 * be possible to get enough handler activations to overflow the stack.
918 * Although unlikely, bugs of that kind are hard to track down, so we'd
919 * like to avoid the possibility.
920 * So, on entry to the handler we detect whether we interrupted an
921 * existing activation in its critical region -- if so, we pop the current
922 * activation and restart the handler using the previous one.
923 */
924ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
925
926/*
927 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
928 * see the correct pointer to the pt_regs
929 */
930	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
93111:	incl	PER_CPU_VAR(irq_count)
932	movq	%rsp, %rbp
933	cmovzq	PER_CPU_VAR(irq_stack_ptr), %rsp
934	pushq	%rbp				/* frame pointer backlink */
935	call	xen_evtchn_do_upcall
936	popq	%rsp
937	decl	PER_CPU_VAR(irq_count)
938#ifndef CONFIG_PREEMPT
939	call	xen_maybe_preempt_hcall
940#endif
941	jmp	error_exit
942END(xen_do_hypervisor_callback)
943
944/*
945 * Hypervisor uses this for application faults while it executes.
946 * We get here for two reasons:
947 *  1. Fault while reloading DS, ES, FS or GS
948 *  2. Fault while executing IRET
949 * Category 1 we do not need to fix up as Xen has already reloaded all segment
950 * registers that could be reloaded and zeroed the others.
951 * Category 2 we fix up by killing the current process. We cannot use the
952 * normal Linux return path in this case because if we use the IRET hypercall
953 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
954 * We distinguish between categories by comparing each saved segment register
955 * with its current contents: any discrepancy means we in category 1.
956 */
957ENTRY(xen_failsafe_callback)
958	movl	%ds, %ecx
959	cmpw	%cx, 0x10(%rsp)
960	jne	1f
961	movl	%es, %ecx
962	cmpw	%cx, 0x18(%rsp)
963	jne	1f
964	movl	%fs, %ecx
965	cmpw	%cx, 0x20(%rsp)
966	jne	1f
967	movl	%gs, %ecx
968	cmpw	%cx, 0x28(%rsp)
969	jne	1f
970	/* All segments match their saved values => Category 2 (Bad IRET). */
971	movq	(%rsp), %rcx
972	movq	8(%rsp), %r11
973	addq	$0x30, %rsp
974	pushq	$0				/* RIP */
975	pushq	%r11
976	pushq	%rcx
977	jmp	general_protection
9781:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
979	movq	(%rsp), %rcx
980	movq	8(%rsp), %r11
981	addq	$0x30, %rsp
982	pushq	$-1 /* orig_ax = -1 => not a system call */
983	ALLOC_PT_GPREGS_ON_STACK
984	SAVE_C_REGS
985	SAVE_EXTRA_REGS
986	ENCODE_FRAME_POINTER
987	jmp	error_exit
988END(xen_failsafe_callback)
989
990apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
991	xen_hvm_callback_vector xen_evtchn_do_upcall
992
993#endif /* CONFIG_XEN */
994
995#if IS_ENABLED(CONFIG_HYPERV)
996apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
997	hyperv_callback_vector hyperv_vector_handler
998#endif /* CONFIG_HYPERV */
999
1000idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1001idtentry int3			do_int3			has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
1002idtentry stack_segment		do_stack_segment	has_error_code=1
1003
1004#ifdef CONFIG_XEN
1005idtentry xen_debug		do_debug		has_error_code=0
1006idtentry xen_int3		do_int3			has_error_code=0
1007idtentry xen_stack_segment	do_stack_segment	has_error_code=1
1008#endif
1009
1010idtentry general_protection	do_general_protection	has_error_code=1
1011trace_idtentry page_fault	do_page_fault		has_error_code=1
1012
1013#ifdef CONFIG_KVM_GUEST
1014idtentry async_page_fault	do_async_page_fault	has_error_code=1
1015#endif
1016
1017#ifdef CONFIG_X86_MCE
1018idtentry machine_check					has_error_code=0	paranoid=1 do_sym=*machine_check_vector(%rip)
1019#endif
1020
1021/*
1022 * Save all registers in pt_regs, and switch gs if needed.
1023 * Use slow, but surefire "are we in kernel?" check.
1024 * Return: ebx=0: need swapgs on exit, ebx=1: otherwise
1025 */
1026ENTRY(paranoid_entry)
1027	cld
1028	SAVE_C_REGS 8
1029	SAVE_EXTRA_REGS 8
1030	ENCODE_FRAME_POINTER 8
1031	movl	$1, %ebx
1032	movl	$MSR_GS_BASE, %ecx
1033	rdmsr
1034	testl	%edx, %edx
1035	js	1f				/* negative -> in kernel */
1036	SWAPGS
1037	xorl	%ebx, %ebx
10381:	ret
1039END(paranoid_entry)
1040
1041/*
1042 * "Paranoid" exit path from exception stack.  This is invoked
1043 * only on return from non-NMI IST interrupts that came
1044 * from kernel space.
1045 *
1046 * We may be returning to very strange contexts (e.g. very early
1047 * in syscall entry), so checking for preemption here would
1048 * be complicated.  Fortunately, we there's no good reason
1049 * to try to handle preemption here.
1050 *
1051 * On entry, ebx is "no swapgs" flag (1: don't need swapgs, 0: need it)
1052 */
1053ENTRY(paranoid_exit)
1054	DISABLE_INTERRUPTS(CLBR_NONE)
1055	TRACE_IRQS_OFF_DEBUG
1056	testl	%ebx, %ebx			/* swapgs needed? */
1057	jnz	paranoid_exit_no_swapgs
1058	TRACE_IRQS_IRETQ
1059	SWAPGS_UNSAFE_STACK
1060	jmp	paranoid_exit_restore
1061paranoid_exit_no_swapgs:
1062	TRACE_IRQS_IRETQ_DEBUG
1063paranoid_exit_restore:
1064	RESTORE_EXTRA_REGS
1065	RESTORE_C_REGS
1066	REMOVE_PT_GPREGS_FROM_STACK 8
1067	INTERRUPT_RETURN
1068END(paranoid_exit)
1069
1070/*
1071 * Save all registers in pt_regs, and switch gs if needed.
1072 * Return: EBX=0: came from user mode; EBX=1: otherwise
1073 */
1074ENTRY(error_entry)
1075	cld
1076	SAVE_C_REGS 8
1077	SAVE_EXTRA_REGS 8
1078	ENCODE_FRAME_POINTER 8
1079	xorl	%ebx, %ebx
1080	testb	$3, CS+8(%rsp)
1081	jz	.Lerror_kernelspace
1082
1083	/*
1084	 * We entered from user mode or we're pretending to have entered
1085	 * from user mode due to an IRET fault.
1086	 */
1087	SWAPGS
1088
1089.Lerror_entry_from_usermode_after_swapgs:
1090	/*
1091	 * We need to tell lockdep that IRQs are off.  We can't do this until
1092	 * we fix gsbase, and we should do it before enter_from_user_mode
1093	 * (which can take locks).
1094	 */
1095	TRACE_IRQS_OFF
1096	CALL_enter_from_user_mode
1097	ret
1098
1099.Lerror_entry_done:
1100	TRACE_IRQS_OFF
1101	ret
1102
1103	/*
1104	 * There are two places in the kernel that can potentially fault with
1105	 * usergs. Handle them here.  B stepping K8s sometimes report a
1106	 * truncated RIP for IRET exceptions returning to compat mode. Check
1107	 * for these here too.
1108	 */
1109.Lerror_kernelspace:
1110	incl	%ebx
1111	leaq	native_irq_return_iret(%rip), %rcx
1112	cmpq	%rcx, RIP+8(%rsp)
1113	je	.Lerror_bad_iret
1114	movl	%ecx, %eax			/* zero extend */
1115	cmpq	%rax, RIP+8(%rsp)
1116	je	.Lbstep_iret
1117	cmpq	$.Lgs_change, RIP+8(%rsp)
1118	jne	.Lerror_entry_done
1119
1120	/*
1121	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1122	 * gsbase and proceed.  We'll fix up the exception and land in
1123	 * .Lgs_change's error handler with kernel gsbase.
1124	 */
1125	SWAPGS
1126	jmp .Lerror_entry_done
1127
1128.Lbstep_iret:
1129	/* Fix truncated RIP */
1130	movq	%rcx, RIP+8(%rsp)
1131	/* fall through */
1132
1133.Lerror_bad_iret:
1134	/*
1135	 * We came from an IRET to user mode, so we have user gsbase.
1136	 * Switch to kernel gsbase:
1137	 */
1138	SWAPGS
1139
1140	/*
1141	 * Pretend that the exception came from user mode: set up pt_regs
1142	 * as if we faulted immediately after IRET and clear EBX so that
1143	 * error_exit knows that we will be returning to user mode.
1144	 */
1145	mov	%rsp, %rdi
1146	call	fixup_bad_iret
1147	mov	%rax, %rsp
1148	decl	%ebx
1149	jmp	.Lerror_entry_from_usermode_after_swapgs
1150END(error_entry)
1151
1152
1153/*
1154 * On entry, EBX is a "return to kernel mode" flag:
1155 *   1: already in kernel mode, don't need SWAPGS
1156 *   0: user gsbase is loaded, we need SWAPGS and standard preparation for return to usermode
1157 */
1158ENTRY(error_exit)
1159	movl	%ebx, %eax
1160	DISABLE_INTERRUPTS(CLBR_NONE)
1161	TRACE_IRQS_OFF
1162	testl	%eax, %eax
1163	jnz	retint_kernel
1164	jmp	retint_user
1165END(error_exit)
1166
1167/* Runs on exception stack */
1168ENTRY(nmi)
1169	/*
1170	 * Fix up the exception frame if we're on Xen.
1171	 * PARAVIRT_ADJUST_EXCEPTION_FRAME is guaranteed to push at most
1172	 * one value to the stack on native, so it may clobber the rdx
1173	 * scratch slot, but it won't clobber any of the important
1174	 * slots past it.
1175	 *
1176	 * Xen is a different story, because the Xen frame itself overlaps
1177	 * the "NMI executing" variable.
1178	 */
1179	PARAVIRT_ADJUST_EXCEPTION_FRAME
1180
1181	/*
1182	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1183	 * the iretq it performs will take us out of NMI context.
1184	 * This means that we can have nested NMIs where the next
1185	 * NMI is using the top of the stack of the previous NMI. We
1186	 * can't let it execute because the nested NMI will corrupt the
1187	 * stack of the previous NMI. NMI handlers are not re-entrant
1188	 * anyway.
1189	 *
1190	 * To handle this case we do the following:
1191	 *  Check the a special location on the stack that contains
1192	 *  a variable that is set when NMIs are executing.
1193	 *  The interrupted task's stack is also checked to see if it
1194	 *  is an NMI stack.
1195	 *  If the variable is not set and the stack is not the NMI
1196	 *  stack then:
1197	 *    o Set the special variable on the stack
1198	 *    o Copy the interrupt frame into an "outermost" location on the
1199	 *      stack
1200	 *    o Copy the interrupt frame into an "iret" location on the stack
1201	 *    o Continue processing the NMI
1202	 *  If the variable is set or the previous stack is the NMI stack:
1203	 *    o Modify the "iret" location to jump to the repeat_nmi
1204	 *    o return back to the first NMI
1205	 *
1206	 * Now on exit of the first NMI, we first clear the stack variable
1207	 * The NMI stack will tell any nested NMIs at that point that it is
1208	 * nested. Then we pop the stack normally with iret, and if there was
1209	 * a nested NMI that updated the copy interrupt stack frame, a
1210	 * jump will be made to the repeat_nmi code that will handle the second
1211	 * NMI.
1212	 *
1213	 * However, espfix prevents us from directly returning to userspace
1214	 * with a single IRET instruction.  Similarly, IRET to user mode
1215	 * can fault.  We therefore handle NMIs from user space like
1216	 * other IST entries.
1217	 */
1218
1219	/* Use %rdx as our temp variable throughout */
1220	pushq	%rdx
1221
1222	testb	$3, CS-RIP+8(%rsp)
1223	jz	.Lnmi_from_kernel
1224
1225	/*
1226	 * NMI from user mode.  We need to run on the thread stack, but we
1227	 * can't go through the normal entry paths: NMIs are masked, and
1228	 * we don't want to enable interrupts, because then we'll end
1229	 * up in an awkward situation in which IRQs are on but NMIs
1230	 * are off.
1231	 *
1232	 * We also must not push anything to the stack before switching
1233	 * stacks lest we corrupt the "NMI executing" variable.
1234	 */
1235
1236	SWAPGS_UNSAFE_STACK
1237	cld
1238	movq	%rsp, %rdx
1239	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1240	pushq	5*8(%rdx)	/* pt_regs->ss */
1241	pushq	4*8(%rdx)	/* pt_regs->rsp */
1242	pushq	3*8(%rdx)	/* pt_regs->flags */
1243	pushq	2*8(%rdx)	/* pt_regs->cs */
1244	pushq	1*8(%rdx)	/* pt_regs->rip */
1245	pushq   $-1		/* pt_regs->orig_ax */
1246	pushq   %rdi		/* pt_regs->di */
1247	pushq   %rsi		/* pt_regs->si */
1248	pushq   (%rdx)		/* pt_regs->dx */
1249	pushq   %rcx		/* pt_regs->cx */
1250	pushq   %rax		/* pt_regs->ax */
1251	pushq   %r8		/* pt_regs->r8 */
1252	pushq   %r9		/* pt_regs->r9 */
1253	pushq   %r10		/* pt_regs->r10 */
1254	pushq   %r11		/* pt_regs->r11 */
1255	pushq	%rbx		/* pt_regs->rbx */
1256	pushq	%rbp		/* pt_regs->rbp */
1257	pushq	%r12		/* pt_regs->r12 */
1258	pushq	%r13		/* pt_regs->r13 */
1259	pushq	%r14		/* pt_regs->r14 */
1260	pushq	%r15		/* pt_regs->r15 */
1261	ENCODE_FRAME_POINTER
1262
1263	/*
1264	 * At this point we no longer need to worry about stack damage
1265	 * due to nesting -- we're on the normal thread stack and we're
1266	 * done with the NMI stack.
1267	 */
1268
1269	movq	%rsp, %rdi
1270	movq	$-1, %rsi
1271	call	do_nmi
1272
1273	/*
1274	 * Return back to user mode.  We must *not* do the normal exit
1275	 * work, because we don't want to enable interrupts.
1276	 */
1277	SWAPGS
1278	jmp	restore_regs_and_iret
1279
1280.Lnmi_from_kernel:
1281	/*
1282	 * Here's what our stack frame will look like:
1283	 * +---------------------------------------------------------+
1284	 * | original SS                                             |
1285	 * | original Return RSP                                     |
1286	 * | original RFLAGS                                         |
1287	 * | original CS                                             |
1288	 * | original RIP                                            |
1289	 * +---------------------------------------------------------+
1290	 * | temp storage for rdx                                    |
1291	 * +---------------------------------------------------------+
1292	 * | "NMI executing" variable                                |
1293	 * +---------------------------------------------------------+
1294	 * | iret SS          } Copied from "outermost" frame        |
1295	 * | iret Return RSP  } on each loop iteration; overwritten  |
1296	 * | iret RFLAGS      } by a nested NMI to force another     |
1297	 * | iret CS          } iteration if needed.                 |
1298	 * | iret RIP         }                                      |
1299	 * +---------------------------------------------------------+
1300	 * | outermost SS          } initialized in first_nmi;       |
1301	 * | outermost Return RSP  } will not be changed before      |
1302	 * | outermost RFLAGS      } NMI processing is done.         |
1303	 * | outermost CS          } Copied to "iret" frame on each  |
1304	 * | outermost RIP         } iteration.                      |
1305	 * +---------------------------------------------------------+
1306	 * | pt_regs                                                 |
1307	 * +---------------------------------------------------------+
1308	 *
1309	 * The "original" frame is used by hardware.  Before re-enabling
1310	 * NMIs, we need to be done with it, and we need to leave enough
1311	 * space for the asm code here.
1312	 *
1313	 * We return by executing IRET while RSP points to the "iret" frame.
1314	 * That will either return for real or it will loop back into NMI
1315	 * processing.
1316	 *
1317	 * The "outermost" frame is copied to the "iret" frame on each
1318	 * iteration of the loop, so each iteration starts with the "iret"
1319	 * frame pointing to the final return target.
1320	 */
1321
1322	/*
1323	 * Determine whether we're a nested NMI.
1324	 *
1325	 * If we interrupted kernel code between repeat_nmi and
1326	 * end_repeat_nmi, then we are a nested NMI.  We must not
1327	 * modify the "iret" frame because it's being written by
1328	 * the outer NMI.  That's okay; the outer NMI handler is
1329	 * about to about to call do_nmi anyway, so we can just
1330	 * resume the outer NMI.
1331	 */
1332
1333	movq	$repeat_nmi, %rdx
1334	cmpq	8(%rsp), %rdx
1335	ja	1f
1336	movq	$end_repeat_nmi, %rdx
1337	cmpq	8(%rsp), %rdx
1338	ja	nested_nmi_out
13391:
1340
1341	/*
1342	 * Now check "NMI executing".  If it's set, then we're nested.
1343	 * This will not detect if we interrupted an outer NMI just
1344	 * before IRET.
1345	 */
1346	cmpl	$1, -8(%rsp)
1347	je	nested_nmi
1348
1349	/*
1350	 * Now test if the previous stack was an NMI stack.  This covers
1351	 * the case where we interrupt an outer NMI after it clears
1352	 * "NMI executing" but before IRET.  We need to be careful, though:
1353	 * there is one case in which RSP could point to the NMI stack
1354	 * despite there being no NMI active: naughty userspace controls
1355	 * RSP at the very beginning of the SYSCALL targets.  We can
1356	 * pull a fast one on naughty userspace, though: we program
1357	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1358	 * if it controls the kernel's RSP.  We set DF before we clear
1359	 * "NMI executing".
1360	 */
1361	lea	6*8(%rsp), %rdx
1362	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1363	cmpq	%rdx, 4*8(%rsp)
1364	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1365	ja	first_nmi
1366
1367	subq	$EXCEPTION_STKSZ, %rdx
1368	cmpq	%rdx, 4*8(%rsp)
1369	/* If it is below the NMI stack, it is a normal NMI */
1370	jb	first_nmi
1371
1372	/* Ah, it is within the NMI stack. */
1373
1374	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1375	jz	first_nmi	/* RSP was user controlled. */
1376
1377	/* This is a nested NMI. */
1378
1379nested_nmi:
1380	/*
1381	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1382	 * iteration of NMI handling.
1383	 */
1384	subq	$8, %rsp
1385	leaq	-10*8(%rsp), %rdx
1386	pushq	$__KERNEL_DS
1387	pushq	%rdx
1388	pushfq
1389	pushq	$__KERNEL_CS
1390	pushq	$repeat_nmi
1391
1392	/* Put stack back */
1393	addq	$(6*8), %rsp
1394
1395nested_nmi_out:
1396	popq	%rdx
1397
1398	/* We are returning to kernel mode, so this cannot result in a fault. */
1399	INTERRUPT_RETURN
1400
1401first_nmi:
1402	/* Restore rdx. */
1403	movq	(%rsp), %rdx
1404
1405	/* Make room for "NMI executing". */
1406	pushq	$0
1407
1408	/* Leave room for the "iret" frame */
1409	subq	$(5*8), %rsp
1410
1411	/* Copy the "original" frame to the "outermost" frame */
1412	.rept 5
1413	pushq	11*8(%rsp)
1414	.endr
1415
1416	/* Everything up to here is safe from nested NMIs */
1417
1418#ifdef CONFIG_DEBUG_ENTRY
1419	/*
1420	 * For ease of testing, unmask NMIs right away.  Disabled by
1421	 * default because IRET is very expensive.
1422	 */
1423	pushq	$0		/* SS */
1424	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1425	addq	$8, (%rsp)	/* Fix up RSP */
1426	pushfq			/* RFLAGS */
1427	pushq	$__KERNEL_CS	/* CS */
1428	pushq	$1f		/* RIP */
1429	INTERRUPT_RETURN	/* continues at repeat_nmi below */
14301:
1431#endif
1432
1433repeat_nmi:
1434	/*
1435	 * If there was a nested NMI, the first NMI's iret will return
1436	 * here. But NMIs are still enabled and we can take another
1437	 * nested NMI. The nested NMI checks the interrupted RIP to see
1438	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1439	 * it will just return, as we are about to repeat an NMI anyway.
1440	 * This makes it safe to copy to the stack frame that a nested
1441	 * NMI will update.
1442	 *
1443	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1444	 * we're repeating an NMI, gsbase has the same value that it had on
1445	 * the first iteration.  paranoid_entry will load the kernel
1446	 * gsbase if needed before we call do_nmi.  "NMI executing"
1447	 * is zero.
1448	 */
1449	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1450
1451	/*
1452	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1453	 * here must not modify the "iret" frame while we're writing to
1454	 * it or it will end up containing garbage.
1455	 */
1456	addq	$(10*8), %rsp
1457	.rept 5
1458	pushq	-6*8(%rsp)
1459	.endr
1460	subq	$(5*8), %rsp
1461end_repeat_nmi:
1462
1463	/*
1464	 * Everything below this point can be preempted by a nested NMI.
1465	 * If this happens, then the inner NMI will change the "iret"
1466	 * frame to point back to repeat_nmi.
1467	 */
1468	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1469	ALLOC_PT_GPREGS_ON_STACK
1470
1471	/*
1472	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1473	 * as we should not be calling schedule in NMI context.
1474	 * Even with normal interrupts enabled. An NMI should not be
1475	 * setting NEED_RESCHED or anything that normal interrupts and
1476	 * exceptions might do.
1477	 */
1478	call	paranoid_entry
1479
1480	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1481	movq	%rsp, %rdi
1482	movq	$-1, %rsi
1483	call	do_nmi
1484
1485	testl	%ebx, %ebx			/* swapgs needed? */
1486	jnz	nmi_restore
1487nmi_swapgs:
1488	SWAPGS_UNSAFE_STACK
1489nmi_restore:
1490	RESTORE_EXTRA_REGS
1491	RESTORE_C_REGS
1492
1493	/* Point RSP at the "iret" frame. */
1494	REMOVE_PT_GPREGS_FROM_STACK 6*8
1495
1496	/*
1497	 * Clear "NMI executing".  Set DF first so that we can easily
1498	 * distinguish the remaining code between here and IRET from
1499	 * the SYSCALL entry and exit paths.  On a native kernel, we
1500	 * could just inspect RIP, but, on paravirt kernels,
1501	 * INTERRUPT_RETURN can translate into a jump into a
1502	 * hypercall page.
1503	 */
1504	std
1505	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1506
1507	/*
1508	 * INTERRUPT_RETURN reads the "iret" frame and exits the NMI
1509	 * stack in a single instruction.  We are returning to kernel
1510	 * mode, so this cannot result in a fault.
1511	 */
1512	INTERRUPT_RETURN
1513END(nmi)
1514
1515ENTRY(ignore_sysret)
1516	mov	$-ENOSYS, %eax
1517	sysret
1518END(ignore_sysret)
1519
1520ENTRY(rewind_stack_do_exit)
1521	/* Prevent any naive code from trying to unwind to our caller. */
1522	xorl	%ebp, %ebp
1523
1524	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1525	leaq	-TOP_OF_KERNEL_STACK_PADDING-PTREGS_SIZE(%rax), %rsp
1526
1527	call	do_exit
15281:	jmp 1b
1529END(rewind_stack_do_exit)
1530