• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/kaiser.h>
40#include <asm/nospec-branch.h>
41#include <linux/err.h>
42
43/* Avoid __ASSEMBLER__'ifying <linux/audit.h> just for this.  */
44#include <linux/elf-em.h>
45#define AUDIT_ARCH_X86_64			(EM_X86_64|__AUDIT_ARCH_64BIT|__AUDIT_ARCH_LE)
46#define __AUDIT_ARCH_64BIT			0x80000000
47#define __AUDIT_ARCH_LE				0x40000000
48
49.code64
50.section .entry.text, "ax"
51
52#ifdef CONFIG_PARAVIRT
53ENTRY(native_usergs_sysret64)
54	swapgs
55	sysretq
56ENDPROC(native_usergs_sysret64)
57#endif /* CONFIG_PARAVIRT */
58
59.macro TRACE_IRQS_IRETQ
60#ifdef CONFIG_TRACE_IRQFLAGS
61	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
62	jnc	1f
63	TRACE_IRQS_ON
641:
65#endif
66.endm
67
68/*
69 * When dynamic function tracer is enabled it will add a breakpoint
70 * to all locations that it is about to modify, sync CPUs, update
71 * all the code, sync CPUs, then remove the breakpoints. In this time
72 * if lockdep is enabled, it might jump back into the debug handler
73 * outside the updating of the IST protection. (TRACE_IRQS_ON/OFF).
74 *
75 * We need to change the IDT table before calling TRACE_IRQS_ON/OFF to
76 * make sure the stack pointer does not get reset back to the top
77 * of the debug stack, and instead just reuses the current stack.
78 */
79#if defined(CONFIG_DYNAMIC_FTRACE) && defined(CONFIG_TRACE_IRQFLAGS)
80
81.macro TRACE_IRQS_OFF_DEBUG
82	call	debug_stack_set_zero
83	TRACE_IRQS_OFF
84	call	debug_stack_reset
85.endm
86
87.macro TRACE_IRQS_ON_DEBUG
88	call	debug_stack_set_zero
89	TRACE_IRQS_ON
90	call	debug_stack_reset
91.endm
92
93.macro TRACE_IRQS_IRETQ_DEBUG
94	bt	$9, EFLAGS(%rsp)		/* interrupts off? */
95	jnc	1f
96	TRACE_IRQS_ON_DEBUG
971:
98.endm
99
100#else
101# define TRACE_IRQS_OFF_DEBUG			TRACE_IRQS_OFF
102# define TRACE_IRQS_ON_DEBUG			TRACE_IRQS_ON
103# define TRACE_IRQS_IRETQ_DEBUG			TRACE_IRQS_IRETQ
104#endif
105
106/*
107 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
108 *
109 * This is the only entry point used for 64-bit system calls.  The
110 * hardware interface is reasonably well designed and the register to
111 * argument mapping Linux uses fits well with the registers that are
112 * available when SYSCALL is used.
113 *
114 * SYSCALL instructions can be found inlined in libc implementations as
115 * well as some other programs and libraries.  There are also a handful
116 * of SYSCALL instructions in the vDSO used, for example, as a
117 * clock_gettimeofday fallback.
118 *
119 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
120 * then loads new ss, cs, and rip from previously programmed MSRs.
121 * rflags gets masked by a value from another MSR (so CLD and CLAC
122 * are not needed). SYSCALL does not save anything on the stack
123 * and does not change rsp.
124 *
125 * Registers on entry:
126 * rax  system call number
127 * rcx  return address
128 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
129 * rdi  arg0
130 * rsi  arg1
131 * rdx  arg2
132 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
133 * r8   arg4
134 * r9   arg5
135 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
136 *
137 * Only called from user space.
138 *
139 * When user can change pt_regs->foo always force IRET. That is because
140 * it deals with uncanonical addresses better. SYSRET has trouble
141 * with them due to bugs in both AMD and Intel CPUs.
142 */
143
144ENTRY(entry_SYSCALL_64)
145	/*
146	 * Interrupts are off on entry.
147	 * We do not frame this tiny irq-off block with TRACE_IRQS_OFF/ON,
148	 * it is too small to ever cause noticeable irq latency.
149	 */
150	SWAPGS_UNSAFE_STACK
151	SWITCH_KERNEL_CR3_NO_STACK
152	/*
153	 * A hypervisor implementation might want to use a label
154	 * after the swapgs, so that it can do the swapgs
155	 * for the guest and jump here on syscall.
156	 */
157GLOBAL(entry_SYSCALL_64_after_swapgs)
158
159	movq	%rsp, PER_CPU_VAR(rsp_scratch)
160	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
161
162	TRACE_IRQS_OFF
163
164	/* Construct struct pt_regs on stack */
165	pushq	$__USER_DS			/* pt_regs->ss */
166	pushq	PER_CPU_VAR(rsp_scratch)	/* pt_regs->sp */
167	pushq	%r11				/* pt_regs->flags */
168	pushq	$__USER_CS			/* pt_regs->cs */
169	pushq	%rcx				/* pt_regs->ip */
170	pushq	%rax				/* pt_regs->orig_ax */
171	pushq	%rdi				/* pt_regs->di */
172	pushq	%rsi				/* pt_regs->si */
173	pushq	%rdx				/* pt_regs->dx */
174	pushq	%rcx				/* pt_regs->cx */
175	pushq	$-ENOSYS			/* pt_regs->ax */
176	pushq	%r8				/* pt_regs->r8 */
177	pushq	%r9				/* pt_regs->r9 */
178	pushq	%r10				/* pt_regs->r10 */
179	/*
180	 * Clear extra registers that a speculation attack might
181	 * otherwise want to exploit. Interleave XOR with PUSH
182	 * for better uop scheduling:
183	 */
184	xorq	%r10, %r10			/* nospec   r10 */
185	pushq	%r11				/* pt_regs->r11 */
186	xorq	%r11, %r11			/* nospec   r11 */
187	pushq	%rbx				/* pt_regs->rbx */
188	xorl	%ebx, %ebx			/* nospec   rbx */
189	pushq	%rbp				/* pt_regs->rbp */
190	xorl	%ebp, %ebp			/* nospec   rbp */
191	pushq	%r12				/* pt_regs->r12 */
192	xorq	%r12, %r12			/* nospec   r12 */
193	pushq	%r13				/* pt_regs->r13 */
194	xorq	%r13, %r13			/* nospec   r13 */
195	pushq	%r14				/* pt_regs->r14 */
196	xorq	%r14, %r14			/* nospec   r14 */
197	pushq	%r15				/* pt_regs->r15 */
198	xorq	%r15, %r15			/* nospec   r15 */
199
200	/* IRQs are off. */
201	movq	%rsp, %rdi
202	call	do_syscall_64		/* returns with IRQs disabled */
203
204	RESTORE_EXTRA_REGS
205	TRACE_IRQS_IRETQ		/* we're about to change IF */
206
207	/*
208	 * Try to use SYSRET instead of IRET if we're returning to
209	 * a completely clean 64-bit userspace context.
210	 */
211	movq	RCX(%rsp), %rcx
212	movq	RIP(%rsp), %r11
213	cmpq	%rcx, %r11			/* RCX == RIP */
214	jne	opportunistic_sysret_failed
215
216	/*
217	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
218	 * in kernel space.  This essentially lets the user take over
219	 * the kernel, since userspace controls RSP.
220	 *
221	 * If width of "canonical tail" ever becomes variable, this will need
222	 * to be updated to remain correct on both old and new CPUs.
223	 */
224	.ifne __VIRTUAL_MASK_SHIFT - 47
225	.error "virtual address width changed -- SYSRET checks need update"
226	.endif
227
228	/* Change top 16 bits to be the sign-extension of 47th bit */
229	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
230	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
231
232	/* If this changed %rcx, it was not canonical */
233	cmpq	%rcx, %r11
234	jne	opportunistic_sysret_failed
235
236	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
237	jne	opportunistic_sysret_failed
238
239	movq	R11(%rsp), %r11
240	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
241	jne	opportunistic_sysret_failed
242
243	/*
244	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
245	 * restore RF properly. If the slowpath sets it for whatever reason, we
246	 * need to restore it correctly.
247	 *
248	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
249	 * trap from userspace immediately after SYSRET.  This would cause an
250	 * infinite loop whenever #DB happens with register state that satisfies
251	 * the opportunistic SYSRET conditions.  For example, single-stepping
252	 * this user code:
253	 *
254	 *           movq	$stuck_here, %rcx
255	 *           pushfq
256	 *           popq %r11
257	 *   stuck_here:
258	 *
259	 * would never get past 'stuck_here'.
260	 */
261	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
262	jnz	opportunistic_sysret_failed
263
264	/* nothing to check for RSP */
265
266	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
267	jne	opportunistic_sysret_failed
268
269	/*
270	 * We win! This label is here just for ease of understanding
271	 * perf profiles. Nothing jumps here.
272	 */
273syscall_return_via_sysret:
274	/* rcx and r11 are already restored (see code above) */
275	RESTORE_C_REGS_EXCEPT_RCX_R11
276
277	/*
278	 * This opens a window where we have a user CR3, but are
279	 * running in the kernel.  This makes using the CS
280	 * register useless for telling whether or not we need to
281	 * switch CR3 in NMIs.  Normal interrupts are OK because
282	 * they are off here.
283	 */
284	SWITCH_USER_CR3
285	movq	RSP(%rsp), %rsp
286	USERGS_SYSRET64
287
288opportunistic_sysret_failed:
289	/*
290	 * This opens a window where we have a user CR3, but are
291	 * running in the kernel.  This makes using the CS
292	 * register useless for telling whether or not we need to
293	 * switch CR3 in NMIs.  Normal interrupts are OK because
294	 * they are off here.
295	 */
296	SWITCH_USER_CR3
297	SWAPGS
298	jmp	restore_c_regs_and_iret
299END(entry_SYSCALL_64)
300
301/*
302 * %rdi: prev task
303 * %rsi: next task
304 */
305ENTRY(__switch_to_asm)
306	/*
307	 * Save callee-saved registers
308	 * This must match the order in inactive_task_frame
309	 */
310	pushq	%rbp
311	pushq	%rbx
312	pushq	%r12
313	pushq	%r13
314	pushq	%r14
315	pushq	%r15
316
317	/* switch stack */
318	movq	%rsp, TASK_threadsp(%rdi)
319	movq	TASK_threadsp(%rsi), %rsp
320
321#ifdef CONFIG_CC_STACKPROTECTOR
322	movq	TASK_stack_canary(%rsi), %rbx
323	movq	%rbx, PER_CPU_VAR(irq_stack_union)+stack_canary_offset
324#endif
325
326#ifdef CONFIG_RETPOLINE
327	/*
328	 * When switching from a shallower to a deeper call stack
329	 * the RSB may either underflow or use entries populated
330	 * with userspace addresses. On CPUs where those concerns
331	 * exist, overwrite the RSB with entries which capture
332	 * speculative execution to prevent attack.
333	 */
334	FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW
335#endif
336
337	/* restore callee-saved registers */
338	popq	%r15
339	popq	%r14
340	popq	%r13
341	popq	%r12
342	popq	%rbx
343	popq	%rbp
344
345	jmp	__switch_to
346END(__switch_to_asm)
347
348/*
349 * A newly forked process directly context switches into this address.
350 *
351 * rax: prev task we switched from
352 * rbx: kernel thread func (NULL for user thread)
353 * r12: kernel thread arg
354 */
355ENTRY(ret_from_fork)
356	movq	%rax, %rdi
357	call	schedule_tail			/* rdi: 'prev' task parameter */
358
359	testq	%rbx, %rbx			/* from kernel_thread? */
360	jnz	1f				/* kernel threads are uncommon */
361
3622:
363	movq	%rsp, %rdi
364	call	syscall_return_slowpath	/* returns with IRQs disabled */
365	TRACE_IRQS_ON			/* user mode is traced as IRQS on */
366	SWITCH_USER_CR3
367	SWAPGS
368	jmp	restore_regs_and_iret
369
3701:
371	/* kernel thread */
372	movq	%r12, %rdi
373	CALL_NOSPEC %rbx
374	/*
375	 * A kernel thread is allowed to return here after successfully
376	 * calling do_execve().  Exit to userspace to complete the execve()
377	 * syscall.
378	 */
379	movq	$0, RAX(%rsp)
380	jmp	2b
381END(ret_from_fork)
382
383/*
384 * Build the entry stubs with some assembler magic.
385 * We pack 1 stub into every 8-byte block.
386 */
387	.align 8
388ENTRY(irq_entries_start)
389    vector=FIRST_EXTERNAL_VECTOR
390    .rept (FIRST_SYSTEM_VECTOR - FIRST_EXTERNAL_VECTOR)
391	pushq	$(~vector+0x80)			/* Note: always in signed byte range */
392    vector=vector+1
393	jmp	common_interrupt
394	.align	8
395    .endr
396END(irq_entries_start)
397
398/*
399 * Interrupt entry/exit.
400 *
401 * Interrupt entry points save only callee clobbered registers in fast path.
402 *
403 * Entry runs with interrupts off.
404 */
405
406/* 0(%rsp): ~(interrupt number) */
407	.macro interrupt func
408	cld
409	ALLOC_PT_GPREGS_ON_STACK
410	SAVE_C_REGS
411	SAVE_EXTRA_REGS
412
413	testb	$3, CS(%rsp)
414	jz	1f
415
416	/*
417	 * IRQ from user mode.  Switch to kernel gsbase and inform context
418	 * tracking that we're in kernel mode.
419	 */
420	SWAPGS
421	SWITCH_KERNEL_CR3
422
423	/*
424	 * We need to tell lockdep that IRQs are off.  We can't do this until
425	 * we fix gsbase, and we should do it before enter_from_user_mode
426	 * (which can take locks).  Since TRACE_IRQS_OFF idempotent,
427	 * the simplest way to handle it is to just call it twice if
428	 * we enter from user mode.  There's no reason to optimize this since
429	 * TRACE_IRQS_OFF is a no-op if lockdep is off.
430	 */
431	TRACE_IRQS_OFF
432
433	CALL_enter_from_user_mode
434
4351:
436	/*
437	 * Save previous stack pointer, optionally switch to interrupt stack.
438	 * irq_count is used to check if a CPU is already on an interrupt stack
439	 * or not. While this is essentially redundant with preempt_count it is
440	 * a little cheaper to use a separate counter in the PDA (short of
441	 * moving irq_enter into assembly, which would be too much work)
442	 */
443	movq	%rsp, %rdi
444	incl	PER_CPU_VAR(irq_count)
445	cmovzq	PER_CPU_VAR(irq_stack_ptr), %rsp
446	pushq	%rdi
447	/* We entered an interrupt context - irqs are off: */
448	TRACE_IRQS_OFF
449
450	call	\func	/* rdi points to pt_regs */
451	.endm
452
453	/*
454	 * The interrupt stubs push (~vector+0x80) onto the stack and
455	 * then jump to common_interrupt.
456	 */
457	.p2align CONFIG_X86_L1_CACHE_SHIFT
458common_interrupt:
459	ASM_CLAC
460	addq	$-0x80, (%rsp)			/* Adjust vector to [-256, -1] range */
461	interrupt do_IRQ
462	/* 0(%rsp): old RSP */
463ret_from_intr:
464	DISABLE_INTERRUPTS(CLBR_NONE)
465	TRACE_IRQS_OFF
466	decl	PER_CPU_VAR(irq_count)
467
468	/* Restore saved previous stack */
469	popq	%rsp
470
471	testb	$3, CS(%rsp)
472	jz	retint_kernel
473
474	/* Interrupt came from user space */
475GLOBAL(retint_user)
476	mov	%rsp,%rdi
477	call	prepare_exit_to_usermode
478	TRACE_IRQS_IRETQ
479	SWITCH_USER_CR3
480	SWAPGS
481	jmp	restore_regs_and_iret
482
483/* Returning to kernel space */
484retint_kernel:
485#ifdef CONFIG_PREEMPT
486	/* Interrupts are off */
487	/* Check if we need preemption */
488	bt	$9, EFLAGS(%rsp)		/* were interrupts off? */
489	jnc	1f
4900:	cmpl	$0, PER_CPU_VAR(__preempt_count)
491	jnz	1f
492	call	preempt_schedule_irq
493	jmp	0b
4941:
495#endif
496	/*
497	 * The iretq could re-enable interrupts:
498	 */
499	TRACE_IRQS_IRETQ
500
501/*
502 * At this label, code paths which return to kernel and to user,
503 * which come from interrupts/exception and from syscalls, merge.
504 */
505GLOBAL(restore_regs_and_iret)
506	RESTORE_EXTRA_REGS
507restore_c_regs_and_iret:
508	RESTORE_C_REGS
509	REMOVE_PT_GPREGS_FROM_STACK 8
510	INTERRUPT_RETURN
511
512ENTRY(native_iret)
513	/*
514	 * Are we returning to a stack segment from the LDT?  Note: in
515	 * 64-bit mode SS:RSP on the exception stack is always valid.
516	 */
517#ifdef CONFIG_X86_ESPFIX64
518	testb	$4, (SS-RIP)(%rsp)
519	jnz	native_irq_return_ldt
520#endif
521
522.global native_irq_return_iret
523native_irq_return_iret:
524	/*
525	 * This may fault.  Non-paranoid faults on return to userspace are
526	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
527	 * Double-faults due to espfix64 are handled in do_double_fault.
528	 * Other faults here are fatal.
529	 */
530	iretq
531
532#ifdef CONFIG_X86_ESPFIX64
533native_irq_return_ldt:
534	/*
535	 * We are running with user GSBASE.  All GPRs contain their user
536	 * values.  We have a percpu ESPFIX stack that is eight slots
537	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
538	 * of the ESPFIX stack.
539	 *
540	 * We clobber RAX and RDI in this code.  We stash RDI on the
541	 * normal stack and RAX on the ESPFIX stack.
542	 *
543	 * The ESPFIX stack layout we set up looks like this:
544	 *
545	 * --- top of ESPFIX stack ---
546	 * SS
547	 * RSP
548	 * RFLAGS
549	 * CS
550	 * RIP  <-- RSP points here when we're done
551	 * RAX  <-- espfix_waddr points here
552	 * --- bottom of ESPFIX stack ---
553	 */
554
555	pushq	%rdi				/* Stash user RDI */
556	SWAPGS
557	SWITCH_KERNEL_CR3
558	movq	PER_CPU_VAR(espfix_waddr), %rdi
559	movq	%rax, (0*8)(%rdi)		/* user RAX */
560	movq	(1*8)(%rsp), %rax		/* user RIP */
561	movq	%rax, (1*8)(%rdi)
562	movq	(2*8)(%rsp), %rax		/* user CS */
563	movq	%rax, (2*8)(%rdi)
564	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
565	movq	%rax, (3*8)(%rdi)
566	movq	(5*8)(%rsp), %rax		/* user SS */
567	movq	%rax, (5*8)(%rdi)
568	movq	(4*8)(%rsp), %rax		/* user RSP */
569	movq	%rax, (4*8)(%rdi)
570	/* Now RAX == RSP. */
571
572	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
573	popq	%rdi				/* Restore user RDI */
574
575	/*
576	 * espfix_stack[31:16] == 0.  The page tables are set up such that
577	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
578	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
579	 * the same page.  Set up RSP so that RSP[31:16] contains the
580	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
581	 * still points to an RO alias of the ESPFIX stack.
582	 */
583	orq	PER_CPU_VAR(espfix_stack), %rax
584	SWITCH_USER_CR3
585	SWAPGS
586	movq	%rax, %rsp
587
588	/*
589	 * At this point, we cannot write to the stack any more, but we can
590	 * still read.
591	 */
592	popq	%rax				/* Restore user RAX */
593
594	/*
595	 * RSP now points to an ordinary IRET frame, except that the page
596	 * is read-only and RSP[31:16] are preloaded with the userspace
597	 * values.  We can now IRET back to userspace.
598	 */
599	jmp	native_irq_return_iret
600#endif
601END(common_interrupt)
602
603/*
604 * APIC interrupts.
605 */
606.macro apicinterrupt3 num sym do_sym
607ENTRY(\sym)
608	ASM_CLAC
609	pushq	$~(\num)
610.Lcommon_\sym:
611	interrupt \do_sym
612	jmp	ret_from_intr
613END(\sym)
614.endm
615
616#ifdef CONFIG_TRACING
617#define trace(sym) trace_##sym
618#define smp_trace(sym) smp_trace_##sym
619
620.macro trace_apicinterrupt num sym
621apicinterrupt3 \num trace(\sym) smp_trace(\sym)
622.endm
623#else
624.macro trace_apicinterrupt num sym do_sym
625.endm
626#endif
627
628/* Make sure APIC interrupt handlers end up in the irqentry section: */
629#define PUSH_SECTION_IRQENTRY	.pushsection .irqentry.text, "ax"
630#define POP_SECTION_IRQENTRY	.popsection
631
632.macro apicinterrupt num sym do_sym
633PUSH_SECTION_IRQENTRY
634apicinterrupt3 \num \sym \do_sym
635trace_apicinterrupt \num \sym
636POP_SECTION_IRQENTRY
637.endm
638
639#ifdef CONFIG_SMP
640apicinterrupt3 IRQ_MOVE_CLEANUP_VECTOR		irq_move_cleanup_interrupt	smp_irq_move_cleanup_interrupt
641apicinterrupt3 REBOOT_VECTOR			reboot_interrupt		smp_reboot_interrupt
642#endif
643
644#ifdef CONFIG_X86_UV
645apicinterrupt3 UV_BAU_MESSAGE			uv_bau_message_intr1		uv_bau_message_interrupt
646#endif
647
648apicinterrupt LOCAL_TIMER_VECTOR		apic_timer_interrupt		smp_apic_timer_interrupt
649apicinterrupt X86_PLATFORM_IPI_VECTOR		x86_platform_ipi		smp_x86_platform_ipi
650
651#ifdef CONFIG_HAVE_KVM
652apicinterrupt3 POSTED_INTR_VECTOR		kvm_posted_intr_ipi		smp_kvm_posted_intr_ipi
653apicinterrupt3 POSTED_INTR_WAKEUP_VECTOR	kvm_posted_intr_wakeup_ipi	smp_kvm_posted_intr_wakeup_ipi
654#endif
655
656#ifdef CONFIG_X86_MCE_THRESHOLD
657apicinterrupt THRESHOLD_APIC_VECTOR		threshold_interrupt		smp_threshold_interrupt
658#endif
659
660#ifdef CONFIG_X86_MCE_AMD
661apicinterrupt DEFERRED_ERROR_VECTOR		deferred_error_interrupt	smp_deferred_error_interrupt
662#endif
663
664#ifdef CONFIG_X86_THERMAL_VECTOR
665apicinterrupt THERMAL_APIC_VECTOR		thermal_interrupt		smp_thermal_interrupt
666#endif
667
668#ifdef CONFIG_SMP
669apicinterrupt CALL_FUNCTION_SINGLE_VECTOR	call_function_single_interrupt	smp_call_function_single_interrupt
670apicinterrupt CALL_FUNCTION_VECTOR		call_function_interrupt		smp_call_function_interrupt
671apicinterrupt RESCHEDULE_VECTOR			reschedule_interrupt		smp_reschedule_interrupt
672#endif
673
674apicinterrupt ERROR_APIC_VECTOR			error_interrupt			smp_error_interrupt
675apicinterrupt SPURIOUS_APIC_VECTOR		spurious_interrupt		smp_spurious_interrupt
676
677#ifdef CONFIG_IRQ_WORK
678apicinterrupt IRQ_WORK_VECTOR			irq_work_interrupt		smp_irq_work_interrupt
679#endif
680
681/*
682 * Exception entry points.
683 */
684#define CPU_TSS_IST(x) PER_CPU_VAR(cpu_tss) + (TSS_ist + ((x) - 1) * 8)
685
686.macro idtentry sym do_sym has_error_code:req paranoid=0 shift_ist=-1
687ENTRY(\sym)
688	/* Sanity check */
689	.if \shift_ist != -1 && \paranoid == 0
690	.error "using shift_ist requires paranoid=1"
691	.endif
692
693	ASM_CLAC
694	PARAVIRT_ADJUST_EXCEPTION_FRAME
695
696	.ifeq \has_error_code
697	pushq	$-1				/* ORIG_RAX: no syscall to restart */
698	.endif
699
700	ALLOC_PT_GPREGS_ON_STACK
701
702	.if \paranoid
703	.if \paranoid == 1
704	testb	$3, CS(%rsp)			/* If coming from userspace, switch stacks */
705	jnz	1f
706	.endif
707	call	paranoid_entry
708	.else
709	call	error_entry
710	.endif
711	/* returned flag: ebx=0: need swapgs on exit, ebx=1: don't need it */
712
713	.if \paranoid
714	.if \shift_ist != -1
715	TRACE_IRQS_OFF_DEBUG			/* reload IDT in case of recursion */
716	.else
717	TRACE_IRQS_OFF
718	.endif
719	.endif
720
721	movq	%rsp, %rdi			/* pt_regs pointer */
722
723	.if \has_error_code
724	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
725	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
726	.else
727	xorl	%esi, %esi			/* no error code */
728	.endif
729
730	.if \shift_ist != -1
731	subq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
732	.endif
733
734	call	\do_sym
735
736	.if \shift_ist != -1
737	addq	$EXCEPTION_STKSZ, CPU_TSS_IST(\shift_ist)
738	.endif
739
740	/* these procedures expect "no swapgs" flag in ebx */
741	.if \paranoid
742	jmp	paranoid_exit
743	.else
744	jmp	error_exit
745	.endif
746
747	.if \paranoid == 1
748	/*
749	 * Paranoid entry from userspace.  Switch stacks and treat it
750	 * as a normal entry.  This means that paranoid handlers
751	 * run in real process context if user_mode(regs).
752	 */
7531:
754	call	error_entry
755
756
757	movq	%rsp, %rdi			/* pt_regs pointer */
758	call	sync_regs
759	movq	%rax, %rsp			/* switch stack */
760
761	movq	%rsp, %rdi			/* pt_regs pointer */
762
763	.if \has_error_code
764	movq	ORIG_RAX(%rsp), %rsi		/* get error code */
765	movq	$-1, ORIG_RAX(%rsp)		/* no syscall to restart */
766	.else
767	xorl	%esi, %esi			/* no error code */
768	.endif
769
770	call	\do_sym
771
772	jmp	error_exit			/* %ebx: no swapgs flag */
773	.endif
774END(\sym)
775.endm
776
777#ifdef CONFIG_TRACING
778.macro trace_idtentry sym do_sym has_error_code:req
779idtentry trace(\sym) trace(\do_sym) has_error_code=\has_error_code
780idtentry \sym \do_sym has_error_code=\has_error_code
781.endm
782#else
783.macro trace_idtentry sym do_sym has_error_code:req
784idtentry \sym \do_sym has_error_code=\has_error_code
785.endm
786#endif
787
788idtentry divide_error			do_divide_error			has_error_code=0
789idtentry overflow			do_overflow			has_error_code=0
790idtentry bounds				do_bounds			has_error_code=0
791idtentry invalid_op			do_invalid_op			has_error_code=0
792idtentry device_not_available		do_device_not_available		has_error_code=0
793idtentry double_fault			do_double_fault			has_error_code=1 paranoid=2
794idtentry coprocessor_segment_overrun	do_coprocessor_segment_overrun	has_error_code=0
795idtentry invalid_TSS			do_invalid_TSS			has_error_code=1
796idtentry segment_not_present		do_segment_not_present		has_error_code=1
797idtentry spurious_interrupt_bug		do_spurious_interrupt_bug	has_error_code=0
798idtentry coprocessor_error		do_coprocessor_error		has_error_code=0
799idtentry alignment_check		do_alignment_check		has_error_code=1
800idtentry simd_coprocessor_error		do_simd_coprocessor_error	has_error_code=0
801
802
803	/*
804	 * Reload gs selector with exception handling
805	 * edi:  new selector
806	 */
807ENTRY(native_load_gs_index)
808	pushfq
809	DISABLE_INTERRUPTS(CLBR_ANY & ~CLBR_RDI)
810	SWAPGS
811.Lgs_change:
812	movl	%edi, %gs
8132:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
814	SWAPGS
815	popfq
816	ret
817END(native_load_gs_index)
818EXPORT_SYMBOL(native_load_gs_index)
819
820	_ASM_EXTABLE(.Lgs_change, bad_gs)
821	.section .fixup, "ax"
822	/* running with kernelgs */
823bad_gs:
824	SWAPGS					/* switch back to user gs */
825.macro ZAP_GS
826	/* This can't be a string because the preprocessor needs to see it. */
827	movl $__USER_DS, %eax
828	movl %eax, %gs
829.endm
830	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
831	xorl	%eax, %eax
832	movl	%eax, %gs
833	jmp	2b
834	.previous
835
836/* Call softirq on interrupt stack. Interrupts are off. */
837ENTRY(do_softirq_own_stack)
838	pushq	%rbp
839	mov	%rsp, %rbp
840	incl	PER_CPU_VAR(irq_count)
841	cmove	PER_CPU_VAR(irq_stack_ptr), %rsp
842	push	%rbp				/* frame pointer backlink */
843	call	__do_softirq
844	leaveq
845	decl	PER_CPU_VAR(irq_count)
846	ret
847END(do_softirq_own_stack)
848
849#ifdef CONFIG_XEN
850idtentry xen_hypervisor_callback xen_do_hypervisor_callback has_error_code=0
851
852/*
853 * A note on the "critical region" in our callback handler.
854 * We want to avoid stacking callback handlers due to events occurring
855 * during handling of the last event. To do this, we keep events disabled
856 * until we've done all processing. HOWEVER, we must enable events before
857 * popping the stack frame (can't be done atomically) and so it would still
858 * be possible to get enough handler activations to overflow the stack.
859 * Although unlikely, bugs of that kind are hard to track down, so we'd
860 * like to avoid the possibility.
861 * So, on entry to the handler we detect whether we interrupted an
862 * existing activation in its critical region -- if so, we pop the current
863 * activation and restart the handler using the previous one.
864 */
865ENTRY(xen_do_hypervisor_callback)		/* do_hypervisor_callback(struct *pt_regs) */
866
867/*
868 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
869 * see the correct pointer to the pt_regs
870 */
871	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
87211:	incl	PER_CPU_VAR(irq_count)
873	movq	%rsp, %rbp
874	cmovzq	PER_CPU_VAR(irq_stack_ptr), %rsp
875	pushq	%rbp				/* frame pointer backlink */
876	call	xen_evtchn_do_upcall
877	popq	%rsp
878	decl	PER_CPU_VAR(irq_count)
879#ifndef CONFIG_PREEMPT
880	call	xen_maybe_preempt_hcall
881#endif
882	jmp	error_exit
883END(xen_do_hypervisor_callback)
884
885/*
886 * Hypervisor uses this for application faults while it executes.
887 * We get here for two reasons:
888 *  1. Fault while reloading DS, ES, FS or GS
889 *  2. Fault while executing IRET
890 * Category 1 we do not need to fix up as Xen has already reloaded all segment
891 * registers that could be reloaded and zeroed the others.
892 * Category 2 we fix up by killing the current process. We cannot use the
893 * normal Linux return path in this case because if we use the IRET hypercall
894 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
895 * We distinguish between categories by comparing each saved segment register
896 * with its current contents: any discrepancy means we in category 1.
897 */
898ENTRY(xen_failsafe_callback)
899	movl	%ds, %ecx
900	cmpw	%cx, 0x10(%rsp)
901	jne	1f
902	movl	%es, %ecx
903	cmpw	%cx, 0x18(%rsp)
904	jne	1f
905	movl	%fs, %ecx
906	cmpw	%cx, 0x20(%rsp)
907	jne	1f
908	movl	%gs, %ecx
909	cmpw	%cx, 0x28(%rsp)
910	jne	1f
911	/* All segments match their saved values => Category 2 (Bad IRET). */
912	movq	(%rsp), %rcx
913	movq	8(%rsp), %r11
914	addq	$0x30, %rsp
915	pushq	$0				/* RIP */
916	pushq	%r11
917	pushq	%rcx
918	jmp	general_protection
9191:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
920	movq	(%rsp), %rcx
921	movq	8(%rsp), %r11
922	addq	$0x30, %rsp
923	pushq	$-1 /* orig_ax = -1 => not a system call */
924	ALLOC_PT_GPREGS_ON_STACK
925	SAVE_C_REGS
926	SAVE_EXTRA_REGS
927	jmp	error_exit
928END(xen_failsafe_callback)
929
930apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
931	xen_hvm_callback_vector xen_evtchn_do_upcall
932
933#endif /* CONFIG_XEN */
934
935#if IS_ENABLED(CONFIG_HYPERV)
936apicinterrupt3 HYPERVISOR_CALLBACK_VECTOR \
937	hyperv_callback_vector hyperv_vector_handler
938#endif /* CONFIG_HYPERV */
939
940idtentry debug			do_debug		has_error_code=0	paranoid=1 shift_ist=DEBUG_STACK
941idtentry int3			do_int3			has_error_code=0
942idtentry stack_segment		do_stack_segment	has_error_code=1
943
944#ifdef CONFIG_XEN
945idtentry xen_debug		do_debug		has_error_code=0
946idtentry xen_int3		do_int3			has_error_code=0
947idtentry xen_stack_segment	do_stack_segment	has_error_code=1
948#endif
949
950idtentry general_protection	do_general_protection	has_error_code=1
951trace_idtentry page_fault	do_page_fault		has_error_code=1
952
953#ifdef CONFIG_KVM_GUEST
954idtentry async_page_fault	do_async_page_fault	has_error_code=1
955#endif
956
957#ifdef CONFIG_X86_MCE
958idtentry machine_check		do_mce			has_error_code=0	paranoid=1
959#endif
960
961/*
962 * Save all registers in pt_regs, and switch gs if needed.
963 * Use slow, but surefire "are we in kernel?" check.
964 *
965 * Return: ebx=0: needs swapgs but not SWITCH_USER_CR3 in paranoid_exit
966 *         ebx=1: needs neither swapgs nor SWITCH_USER_CR3 in paranoid_exit
967 *         ebx=2: needs both swapgs and SWITCH_USER_CR3 in paranoid_exit
968 *         ebx=3: needs SWITCH_USER_CR3 but not swapgs in paranoid_exit
969 */
970ENTRY(paranoid_entry)
971	cld
972	SAVE_C_REGS 8
973	SAVE_EXTRA_REGS 8
974	movl	$1, %ebx
975	movl	$MSR_GS_BASE, %ecx
976	rdmsr
977	testl	%edx, %edx
978	js	1f				/* negative -> in kernel */
979	SWAPGS
980	xorl	%ebx, %ebx
9811:
982#ifdef CONFIG_PAGE_TABLE_ISOLATION
983	/*
984	 * We might have come in between a swapgs and a SWITCH_KERNEL_CR3
985	 * on entry, or between a SWITCH_USER_CR3 and a swapgs on exit.
986	 * Do a conditional SWITCH_KERNEL_CR3: this could safely be done
987	 * unconditionally, but we need to find out whether the reverse
988	 * should be done on return (conveyed to paranoid_exit in %ebx).
989	 */
990	ALTERNATIVE "jmp 2f", "movq %cr3, %rax", X86_FEATURE_KAISER
991	testl	$KAISER_SHADOW_PGD_OFFSET, %eax
992	jz	2f
993	orl	$2, %ebx
994	andq	$(~(X86_CR3_PCID_ASID_MASK | KAISER_SHADOW_PGD_OFFSET)), %rax
995	/* If PCID enabled, set X86_CR3_PCID_NOFLUSH_BIT */
996	ALTERNATIVE "", "bts $63, %rax", X86_FEATURE_PCID
997	movq	%rax, %cr3
9982:
999#endif
1000	ret
1001END(paranoid_entry)
1002
1003/*
1004 * "Paranoid" exit path from exception stack.  This is invoked
1005 * only on return from non-NMI IST interrupts that came
1006 * from kernel space.
1007 *
1008 * We may be returning to very strange contexts (e.g. very early
1009 * in syscall entry), so checking for preemption here would
1010 * be complicated.  Fortunately, we there's no good reason
1011 * to try to handle preemption here.
1012 *
1013 * On entry: ebx=0: needs swapgs but not SWITCH_USER_CR3
1014 *           ebx=1: needs neither swapgs nor SWITCH_USER_CR3
1015 *           ebx=2: needs both swapgs and SWITCH_USER_CR3
1016 *           ebx=3: needs SWITCH_USER_CR3 but not swapgs
1017 */
1018ENTRY(paranoid_exit)
1019	DISABLE_INTERRUPTS(CLBR_NONE)
1020	TRACE_IRQS_OFF_DEBUG
1021	TRACE_IRQS_IRETQ_DEBUG
1022#ifdef CONFIG_PAGE_TABLE_ISOLATION
1023	/* No ALTERNATIVE for X86_FEATURE_KAISER: paranoid_entry sets %ebx */
1024	testl	$2, %ebx			/* SWITCH_USER_CR3 needed? */
1025	jz	paranoid_exit_no_switch
1026	SWITCH_USER_CR3
1027paranoid_exit_no_switch:
1028#endif
1029	testl	$1, %ebx			/* swapgs needed? */
1030	jnz	paranoid_exit_no_swapgs
1031	SWAPGS_UNSAFE_STACK
1032paranoid_exit_no_swapgs:
1033	RESTORE_EXTRA_REGS
1034	RESTORE_C_REGS
1035	REMOVE_PT_GPREGS_FROM_STACK 8
1036	INTERRUPT_RETURN
1037END(paranoid_exit)
1038
1039/*
1040 * Save all registers in pt_regs, and switch gs if needed.
1041 * Return: EBX=0: came from user mode; EBX=1: otherwise
1042 */
1043ENTRY(error_entry)
1044	cld
1045	SAVE_C_REGS 8
1046	SAVE_EXTRA_REGS 8
1047	/*
1048	 * error_entry() always returns with a kernel gsbase and
1049	 * CR3.  We must also have a kernel CR3/gsbase before
1050	 * calling TRACE_IRQS_*.  Just unconditionally switch to
1051	 * the kernel CR3 here.
1052	 */
1053	SWITCH_KERNEL_CR3
1054	xorl	%ebx, %ebx
1055	testb	$3, CS+8(%rsp)
1056	jz	.Lerror_kernelspace
1057
1058	/*
1059	 * We entered from user mode or we're pretending to have entered
1060	 * from user mode due to an IRET fault.
1061	 */
1062	SWAPGS
1063
1064.Lerror_entry_from_usermode_after_swapgs:
1065	/*
1066	 * We need to tell lockdep that IRQs are off.  We can't do this until
1067	 * we fix gsbase, and we should do it before enter_from_user_mode
1068	 * (which can take locks).
1069	 */
1070	TRACE_IRQS_OFF
1071	CALL_enter_from_user_mode
1072	ret
1073
1074.Lerror_entry_done:
1075	TRACE_IRQS_OFF
1076	ret
1077
1078	/*
1079	 * There are two places in the kernel that can potentially fault with
1080	 * usergs. Handle them here.  B stepping K8s sometimes report a
1081	 * truncated RIP for IRET exceptions returning to compat mode. Check
1082	 * for these here too.
1083	 */
1084.Lerror_kernelspace:
1085	incl	%ebx
1086	leaq	native_irq_return_iret(%rip), %rcx
1087	cmpq	%rcx, RIP+8(%rsp)
1088	je	.Lerror_bad_iret
1089	movl	%ecx, %eax			/* zero extend */
1090	cmpq	%rax, RIP+8(%rsp)
1091	je	.Lbstep_iret
1092	cmpq	$.Lgs_change, RIP+8(%rsp)
1093	jne	.Lerror_entry_done
1094
1095	/*
1096	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1097	 * gsbase and proceed.  We'll fix up the exception and land in
1098	 * .Lgs_change's error handler with kernel gsbase.
1099	 */
1100	SWAPGS
1101	jmp .Lerror_entry_done
1102
1103.Lbstep_iret:
1104	/* Fix truncated RIP */
1105	movq	%rcx, RIP+8(%rsp)
1106	/* fall through */
1107
1108.Lerror_bad_iret:
1109	/*
1110	 * We came from an IRET to user mode, so we have user gsbase.
1111	 * Switch to kernel gsbase:
1112	 */
1113	SWAPGS
1114
1115	/*
1116	 * Pretend that the exception came from user mode: set up pt_regs
1117	 * as if we faulted immediately after IRET and clear EBX so that
1118	 * error_exit knows that we will be returning to user mode.
1119	 */
1120	mov	%rsp, %rdi
1121	call	fixup_bad_iret
1122	mov	%rax, %rsp
1123	decl	%ebx
1124	jmp	.Lerror_entry_from_usermode_after_swapgs
1125END(error_entry)
1126
1127
1128/*
1129 * On entry, EBX is a "return to kernel mode" flag:
1130 *   1: already in kernel mode, don't need SWAPGS
1131 *   0: user gsbase is loaded, we need SWAPGS and standard preparation for return to usermode
1132 */
1133ENTRY(error_exit)
1134	movl	%ebx, %eax
1135	DISABLE_INTERRUPTS(CLBR_NONE)
1136	TRACE_IRQS_OFF
1137	testl	%eax, %eax
1138	jnz	retint_kernel
1139	jmp	retint_user
1140END(error_exit)
1141
1142/* Runs on exception stack */
1143ENTRY(nmi)
1144	/*
1145	 * Fix up the exception frame if we're on Xen.
1146	 * PARAVIRT_ADJUST_EXCEPTION_FRAME is guaranteed to push at most
1147	 * one value to the stack on native, so it may clobber the rdx
1148	 * scratch slot, but it won't clobber any of the important
1149	 * slots past it.
1150	 *
1151	 * Xen is a different story, because the Xen frame itself overlaps
1152	 * the "NMI executing" variable.
1153	 */
1154	PARAVIRT_ADJUST_EXCEPTION_FRAME
1155
1156	/*
1157	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1158	 * the iretq it performs will take us out of NMI context.
1159	 * This means that we can have nested NMIs where the next
1160	 * NMI is using the top of the stack of the previous NMI. We
1161	 * can't let it execute because the nested NMI will corrupt the
1162	 * stack of the previous NMI. NMI handlers are not re-entrant
1163	 * anyway.
1164	 *
1165	 * To handle this case we do the following:
1166	 *  Check the a special location on the stack that contains
1167	 *  a variable that is set when NMIs are executing.
1168	 *  The interrupted task's stack is also checked to see if it
1169	 *  is an NMI stack.
1170	 *  If the variable is not set and the stack is not the NMI
1171	 *  stack then:
1172	 *    o Set the special variable on the stack
1173	 *    o Copy the interrupt frame into an "outermost" location on the
1174	 *      stack
1175	 *    o Copy the interrupt frame into an "iret" location on the stack
1176	 *    o Continue processing the NMI
1177	 *  If the variable is set or the previous stack is the NMI stack:
1178	 *    o Modify the "iret" location to jump to the repeat_nmi
1179	 *    o return back to the first NMI
1180	 *
1181	 * Now on exit of the first NMI, we first clear the stack variable
1182	 * The NMI stack will tell any nested NMIs at that point that it is
1183	 * nested. Then we pop the stack normally with iret, and if there was
1184	 * a nested NMI that updated the copy interrupt stack frame, a
1185	 * jump will be made to the repeat_nmi code that will handle the second
1186	 * NMI.
1187	 *
1188	 * However, espfix prevents us from directly returning to userspace
1189	 * with a single IRET instruction.  Similarly, IRET to user mode
1190	 * can fault.  We therefore handle NMIs from user space like
1191	 * other IST entries.
1192	 */
1193
1194	ASM_CLAC
1195
1196	/* Use %rdx as our temp variable throughout */
1197	pushq	%rdx
1198
1199	testb	$3, CS-RIP+8(%rsp)
1200	jz	.Lnmi_from_kernel
1201
1202	/*
1203	 * NMI from user mode.  We need to run on the thread stack, but we
1204	 * can't go through the normal entry paths: NMIs are masked, and
1205	 * we don't want to enable interrupts, because then we'll end
1206	 * up in an awkward situation in which IRQs are on but NMIs
1207	 * are off.
1208	 *
1209	 * We also must not push anything to the stack before switching
1210	 * stacks lest we corrupt the "NMI executing" variable.
1211	 */
1212
1213	SWAPGS_UNSAFE_STACK
1214	/*
1215	 * percpu variables are mapped with user CR3, so no need
1216	 * to switch CR3 here.
1217	 */
1218	cld
1219	movq	%rsp, %rdx
1220	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1221	pushq	5*8(%rdx)	/* pt_regs->ss */
1222	pushq	4*8(%rdx)	/* pt_regs->rsp */
1223	pushq	3*8(%rdx)	/* pt_regs->flags */
1224	pushq	2*8(%rdx)	/* pt_regs->cs */
1225	pushq	1*8(%rdx)	/* pt_regs->rip */
1226	pushq   $-1		/* pt_regs->orig_ax */
1227	pushq   %rdi		/* pt_regs->di */
1228	pushq   %rsi		/* pt_regs->si */
1229	pushq   (%rdx)		/* pt_regs->dx */
1230	pushq   %rcx		/* pt_regs->cx */
1231	pushq   %rax		/* pt_regs->ax */
1232	pushq   %r8		/* pt_regs->r8 */
1233	pushq   %r9		/* pt_regs->r9 */
1234	pushq   %r10		/* pt_regs->r10 */
1235	pushq   %r11		/* pt_regs->r11 */
1236	pushq	%rbx		/* pt_regs->rbx */
1237	pushq	%rbp		/* pt_regs->rbp */
1238	pushq	%r12		/* pt_regs->r12 */
1239	pushq	%r13		/* pt_regs->r13 */
1240	pushq	%r14		/* pt_regs->r14 */
1241	pushq	%r15		/* pt_regs->r15 */
1242
1243	/*
1244	 * At this point we no longer need to worry about stack damage
1245	 * due to nesting -- we're on the normal thread stack and we're
1246	 * done with the NMI stack.
1247	 */
1248
1249	movq	%rsp, %rdi
1250	movq	$-1, %rsi
1251#ifdef CONFIG_PAGE_TABLE_ISOLATION
1252	/* Unconditionally use kernel CR3 for do_nmi() */
1253	/* %rax is saved above, so OK to clobber here */
1254	ALTERNATIVE "jmp 2f", "movq %cr3, %rax", X86_FEATURE_KAISER
1255	/* If PCID enabled, NOFLUSH now and NOFLUSH on return */
1256	ALTERNATIVE "", "bts $63, %rax", X86_FEATURE_PCID
1257	pushq	%rax
1258	/* mask off "user" bit of pgd address and 12 PCID bits: */
1259	andq	$(~(X86_CR3_PCID_ASID_MASK | KAISER_SHADOW_PGD_OFFSET)), %rax
1260	movq	%rax, %cr3
12612:
1262#endif
1263	call	do_nmi
1264
1265#ifdef CONFIG_PAGE_TABLE_ISOLATION
1266	/*
1267	 * Unconditionally restore CR3.  I know we return to
1268	 * kernel code that needs user CR3, but do we ever return
1269	 * to "user mode" where we need the kernel CR3?
1270	 */
1271	ALTERNATIVE "", "popq %rax; movq %rax, %cr3", X86_FEATURE_KAISER
1272#endif
1273
1274	/*
1275	 * Return back to user mode.  We must *not* do the normal exit
1276	 * work, because we don't want to enable interrupts.  Do not
1277	 * switch to user CR3: we might be going back to kernel code
1278	 * that had a user CR3 set.
1279	 */
1280	SWAPGS
1281	jmp	restore_c_regs_and_iret
1282
1283.Lnmi_from_kernel:
1284	/*
1285	 * Here's what our stack frame will look like:
1286	 * +---------------------------------------------------------+
1287	 * | original SS                                             |
1288	 * | original Return RSP                                     |
1289	 * | original RFLAGS                                         |
1290	 * | original CS                                             |
1291	 * | original RIP                                            |
1292	 * +---------------------------------------------------------+
1293	 * | temp storage for rdx                                    |
1294	 * +---------------------------------------------------------+
1295	 * | "NMI executing" variable                                |
1296	 * +---------------------------------------------------------+
1297	 * | iret SS          } Copied from "outermost" frame        |
1298	 * | iret Return RSP  } on each loop iteration; overwritten  |
1299	 * | iret RFLAGS      } by a nested NMI to force another     |
1300	 * | iret CS          } iteration if needed.                 |
1301	 * | iret RIP         }                                      |
1302	 * +---------------------------------------------------------+
1303	 * | outermost SS          } initialized in first_nmi;       |
1304	 * | outermost Return RSP  } will not be changed before      |
1305	 * | outermost RFLAGS      } NMI processing is done.         |
1306	 * | outermost CS          } Copied to "iret" frame on each  |
1307	 * | outermost RIP         } iteration.                      |
1308	 * +---------------------------------------------------------+
1309	 * | pt_regs                                                 |
1310	 * +---------------------------------------------------------+
1311	 *
1312	 * The "original" frame is used by hardware.  Before re-enabling
1313	 * NMIs, we need to be done with it, and we need to leave enough
1314	 * space for the asm code here.
1315	 *
1316	 * We return by executing IRET while RSP points to the "iret" frame.
1317	 * That will either return for real or it will loop back into NMI
1318	 * processing.
1319	 *
1320	 * The "outermost" frame is copied to the "iret" frame on each
1321	 * iteration of the loop, so each iteration starts with the "iret"
1322	 * frame pointing to the final return target.
1323	 */
1324
1325	/*
1326	 * Determine whether we're a nested NMI.
1327	 *
1328	 * If we interrupted kernel code between repeat_nmi and
1329	 * end_repeat_nmi, then we are a nested NMI.  We must not
1330	 * modify the "iret" frame because it's being written by
1331	 * the outer NMI.  That's okay; the outer NMI handler is
1332	 * about to about to call do_nmi anyway, so we can just
1333	 * resume the outer NMI.
1334	 */
1335
1336	movq	$repeat_nmi, %rdx
1337	cmpq	8(%rsp), %rdx
1338	ja	1f
1339	movq	$end_repeat_nmi, %rdx
1340	cmpq	8(%rsp), %rdx
1341	ja	nested_nmi_out
13421:
1343
1344	/*
1345	 * Now check "NMI executing".  If it's set, then we're nested.
1346	 * This will not detect if we interrupted an outer NMI just
1347	 * before IRET.
1348	 */
1349	cmpl	$1, -8(%rsp)
1350	je	nested_nmi
1351
1352	/*
1353	 * Now test if the previous stack was an NMI stack.  This covers
1354	 * the case where we interrupt an outer NMI after it clears
1355	 * "NMI executing" but before IRET.  We need to be careful, though:
1356	 * there is one case in which RSP could point to the NMI stack
1357	 * despite there being no NMI active: naughty userspace controls
1358	 * RSP at the very beginning of the SYSCALL targets.  We can
1359	 * pull a fast one on naughty userspace, though: we program
1360	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1361	 * if it controls the kernel's RSP.  We set DF before we clear
1362	 * "NMI executing".
1363	 */
1364	lea	6*8(%rsp), %rdx
1365	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1366	cmpq	%rdx, 4*8(%rsp)
1367	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1368	ja	first_nmi
1369
1370	subq	$EXCEPTION_STKSZ, %rdx
1371	cmpq	%rdx, 4*8(%rsp)
1372	/* If it is below the NMI stack, it is a normal NMI */
1373	jb	first_nmi
1374
1375	/* Ah, it is within the NMI stack. */
1376
1377	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1378	jz	first_nmi	/* RSP was user controlled. */
1379
1380	/* This is a nested NMI. */
1381
1382nested_nmi:
1383	/*
1384	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1385	 * iteration of NMI handling.
1386	 */
1387	subq	$8, %rsp
1388	leaq	-10*8(%rsp), %rdx
1389	pushq	$__KERNEL_DS
1390	pushq	%rdx
1391	pushfq
1392	pushq	$__KERNEL_CS
1393	pushq	$repeat_nmi
1394
1395	/* Put stack back */
1396	addq	$(6*8), %rsp
1397
1398nested_nmi_out:
1399	popq	%rdx
1400
1401	/* We are returning to kernel mode, so this cannot result in a fault. */
1402	INTERRUPT_RETURN
1403
1404first_nmi:
1405	/* Restore rdx. */
1406	movq	(%rsp), %rdx
1407
1408	/* Make room for "NMI executing". */
1409	pushq	$0
1410
1411	/* Leave room for the "iret" frame */
1412	subq	$(5*8), %rsp
1413
1414	/* Copy the "original" frame to the "outermost" frame */
1415	.rept 5
1416	pushq	11*8(%rsp)
1417	.endr
1418
1419	/* Everything up to here is safe from nested NMIs */
1420
1421#ifdef CONFIG_DEBUG_ENTRY
1422	/*
1423	 * For ease of testing, unmask NMIs right away.  Disabled by
1424	 * default because IRET is very expensive.
1425	 */
1426	pushq	$0		/* SS */
1427	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1428	addq	$8, (%rsp)	/* Fix up RSP */
1429	pushfq			/* RFLAGS */
1430	pushq	$__KERNEL_CS	/* CS */
1431	pushq	$1f		/* RIP */
1432	INTERRUPT_RETURN	/* continues at repeat_nmi below */
14331:
1434#endif
1435
1436repeat_nmi:
1437	/*
1438	 * If there was a nested NMI, the first NMI's iret will return
1439	 * here. But NMIs are still enabled and we can take another
1440	 * nested NMI. The nested NMI checks the interrupted RIP to see
1441	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1442	 * it will just return, as we are about to repeat an NMI anyway.
1443	 * This makes it safe to copy to the stack frame that a nested
1444	 * NMI will update.
1445	 *
1446	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1447	 * we're repeating an NMI, gsbase has the same value that it had on
1448	 * the first iteration.  paranoid_entry will load the kernel
1449	 * gsbase if needed before we call do_nmi.  "NMI executing"
1450	 * is zero.
1451	 */
1452	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1453
1454	/*
1455	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1456	 * here must not modify the "iret" frame while we're writing to
1457	 * it or it will end up containing garbage.
1458	 */
1459	addq	$(10*8), %rsp
1460	.rept 5
1461	pushq	-6*8(%rsp)
1462	.endr
1463	subq	$(5*8), %rsp
1464end_repeat_nmi:
1465
1466	/*
1467	 * Everything below this point can be preempted by a nested NMI.
1468	 * If this happens, then the inner NMI will change the "iret"
1469	 * frame to point back to repeat_nmi.
1470	 */
1471	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1472	ALLOC_PT_GPREGS_ON_STACK
1473
1474	/*
1475	 * Use the same approach as paranoid_entry to handle SWAPGS, but
1476	 * without CR3 handling since we do that differently in NMIs.  No
1477	 * need to use paranoid_exit as we should not be calling schedule
1478	 * in NMI context.  Even with normal interrupts enabled. An NMI
1479	 * should not be setting NEED_RESCHED or anything that normal
1480	 * interrupts and exceptions might do.
1481	 */
1482	cld
1483	SAVE_C_REGS
1484	SAVE_EXTRA_REGS
1485	movl	$1, %ebx
1486	movl	$MSR_GS_BASE, %ecx
1487	rdmsr
1488	testl	%edx, %edx
1489	js	1f				/* negative -> in kernel */
1490	SWAPGS
1491	xorl	%ebx, %ebx
14921:
1493	movq	%rsp, %rdi
1494	movq	$-1, %rsi
1495#ifdef CONFIG_PAGE_TABLE_ISOLATION
1496	/* Unconditionally use kernel CR3 for do_nmi() */
1497	/* %rax is saved above, so OK to clobber here */
1498	ALTERNATIVE "jmp 2f", "movq %cr3, %rax", X86_FEATURE_KAISER
1499	/* If PCID enabled, NOFLUSH now and NOFLUSH on return */
1500	ALTERNATIVE "", "bts $63, %rax", X86_FEATURE_PCID
1501	pushq	%rax
1502	/* mask off "user" bit of pgd address and 12 PCID bits: */
1503	andq	$(~(X86_CR3_PCID_ASID_MASK | KAISER_SHADOW_PGD_OFFSET)), %rax
1504	movq	%rax, %cr3
15052:
1506#endif
1507
1508	/* paranoidentry do_nmi, 0; without TRACE_IRQS_OFF */
1509	call	do_nmi
1510
1511#ifdef CONFIG_PAGE_TABLE_ISOLATION
1512	/*
1513	 * Unconditionally restore CR3.  We might be returning to
1514	 * kernel code that needs user CR3, like just just before
1515	 * a sysret.
1516	 */
1517	ALTERNATIVE "", "popq %rax; movq %rax, %cr3", X86_FEATURE_KAISER
1518#endif
1519
1520	testl	%ebx, %ebx			/* swapgs needed? */
1521	jnz	nmi_restore
1522nmi_swapgs:
1523	/* We fixed up CR3 above, so no need to switch it here */
1524	SWAPGS_UNSAFE_STACK
1525nmi_restore:
1526	RESTORE_EXTRA_REGS
1527	RESTORE_C_REGS
1528
1529	/* Point RSP at the "iret" frame. */
1530	REMOVE_PT_GPREGS_FROM_STACK 6*8
1531
1532	/*
1533	 * Clear "NMI executing".  Set DF first so that we can easily
1534	 * distinguish the remaining code between here and IRET from
1535	 * the SYSCALL entry and exit paths.  On a native kernel, we
1536	 * could just inspect RIP, but, on paravirt kernels,
1537	 * INTERRUPT_RETURN can translate into a jump into a
1538	 * hypercall page.
1539	 */
1540	std
1541	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1542
1543	/*
1544	 * INTERRUPT_RETURN reads the "iret" frame and exits the NMI
1545	 * stack in a single instruction.  We are returning to kernel
1546	 * mode, so this cannot result in a fault.
1547	 */
1548	INTERRUPT_RETURN
1549END(nmi)
1550
1551ENTRY(ignore_sysret)
1552	mov	$-ENOSYS, %eax
1553	sysret
1554END(ignore_sysret)
1555
1556ENTRY(rewind_stack_do_exit)
1557	/* Prevent any naive code from trying to unwind to our caller. */
1558	xorl	%ebp, %ebp
1559
1560	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1561	leaq	-TOP_OF_KERNEL_STACK_PADDING-PTREGS_SIZE(%rax), %rsp
1562
1563	call	do_exit
15641:	jmp 1b
1565END(rewind_stack_do_exit)
1566