xref: /linux/arch/x86/entry/entry_64.S (revision c063a217bc0726c2560138229de5673dbb253a02)
1/* SPDX-License-Identifier: GPL-2.0 */
2/*
3 *  linux/arch/x86_64/entry.S
4 *
5 *  Copyright (C) 1991, 1992  Linus Torvalds
6 *  Copyright (C) 2000, 2001, 2002  Andi Kleen SuSE Labs
7 *  Copyright (C) 2000  Pavel Machek <pavel@suse.cz>
8 *
9 * entry.S contains the system-call and fault low-level handling routines.
10 *
11 * Some of this is documented in Documentation/x86/entry_64.rst
12 *
13 * A note on terminology:
14 * - iret frame:	Architecture defined interrupt frame from SS to RIP
15 *			at the top of the kernel process stack.
16 *
17 * Some macro usage:
18 * - SYM_FUNC_START/END:Define functions in the symbol table.
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 <asm/asm-offsets.h>
26#include <asm/msr.h>
27#include <asm/unistd.h>
28#include <asm/thread_info.h>
29#include <asm/hw_irq.h>
30#include <asm/page_types.h>
31#include <asm/irqflags.h>
32#include <asm/paravirt.h>
33#include <asm/percpu.h>
34#include <asm/asm.h>
35#include <asm/smap.h>
36#include <asm/pgtable_types.h>
37#include <asm/export.h>
38#include <asm/frame.h>
39#include <asm/trapnr.h>
40#include <asm/nospec-branch.h>
41#include <asm/fsgsbase.h>
42#include <linux/err.h>
43
44#include "calling.h"
45
46.code64
47.section .entry.text, "ax"
48
49/*
50 * 64-bit SYSCALL instruction entry. Up to 6 arguments in registers.
51 *
52 * This is the only entry point used for 64-bit system calls.  The
53 * hardware interface is reasonably well designed and the register to
54 * argument mapping Linux uses fits well with the registers that are
55 * available when SYSCALL is used.
56 *
57 * SYSCALL instructions can be found inlined in libc implementations as
58 * well as some other programs and libraries.  There are also a handful
59 * of SYSCALL instructions in the vDSO used, for example, as a
60 * clock_gettimeofday fallback.
61 *
62 * 64-bit SYSCALL saves rip to rcx, clears rflags.RF, then saves rflags to r11,
63 * then loads new ss, cs, and rip from previously programmed MSRs.
64 * rflags gets masked by a value from another MSR (so CLD and CLAC
65 * are not needed). SYSCALL does not save anything on the stack
66 * and does not change rsp.
67 *
68 * Registers on entry:
69 * rax  system call number
70 * rcx  return address
71 * r11  saved rflags (note: r11 is callee-clobbered register in C ABI)
72 * rdi  arg0
73 * rsi  arg1
74 * rdx  arg2
75 * r10  arg3 (needs to be moved to rcx to conform to C ABI)
76 * r8   arg4
77 * r9   arg5
78 * (note: r12-r15, rbp, rbx are callee-preserved in C ABI)
79 *
80 * Only called from user space.
81 *
82 * When user can change pt_regs->foo always force IRET. That is because
83 * it deals with uncanonical addresses better. SYSRET has trouble
84 * with them due to bugs in both AMD and Intel CPUs.
85 */
86
87SYM_CODE_START(entry_SYSCALL_64)
88	UNWIND_HINT_ENTRY
89	ENDBR
90
91	swapgs
92	/* tss.sp2 is scratch space. */
93	movq	%rsp, PER_CPU_VAR(cpu_tss_rw + TSS_sp2)
94	SWITCH_TO_KERNEL_CR3 scratch_reg=%rsp
95	movq	PER_CPU_VAR(pcpu_hot + X86_top_of_stack), %rsp
96
97SYM_INNER_LABEL(entry_SYSCALL_64_safe_stack, SYM_L_GLOBAL)
98	ANNOTATE_NOENDBR
99
100	/* Construct struct pt_regs on stack */
101	pushq	$__USER_DS				/* pt_regs->ss */
102	pushq	PER_CPU_VAR(cpu_tss_rw + TSS_sp2)	/* pt_regs->sp */
103	pushq	%r11					/* pt_regs->flags */
104	pushq	$__USER_CS				/* pt_regs->cs */
105	pushq	%rcx					/* pt_regs->ip */
106SYM_INNER_LABEL(entry_SYSCALL_64_after_hwframe, SYM_L_GLOBAL)
107	pushq	%rax					/* pt_regs->orig_ax */
108
109	PUSH_AND_CLEAR_REGS rax=$-ENOSYS
110
111	/* IRQs are off. */
112	movq	%rsp, %rdi
113	/* Sign extend the lower 32bit as syscall numbers are treated as int */
114	movslq	%eax, %rsi
115
116	/* clobbers %rax, make sure it is after saving the syscall nr */
117	IBRS_ENTER
118	UNTRAIN_RET
119
120	call	do_syscall_64		/* returns with IRQs disabled */
121
122	/*
123	 * Try to use SYSRET instead of IRET if we're returning to
124	 * a completely clean 64-bit userspace context.  If we're not,
125	 * go to the slow exit path.
126	 * In the Xen PV case we must use iret anyway.
127	 */
128
129	ALTERNATIVE "", "jmp	swapgs_restore_regs_and_return_to_usermode", \
130		X86_FEATURE_XENPV
131
132	movq	RCX(%rsp), %rcx
133	movq	RIP(%rsp), %r11
134
135	cmpq	%rcx, %r11	/* SYSRET requires RCX == RIP */
136	jne	swapgs_restore_regs_and_return_to_usermode
137
138	/*
139	 * On Intel CPUs, SYSRET with non-canonical RCX/RIP will #GP
140	 * in kernel space.  This essentially lets the user take over
141	 * the kernel, since userspace controls RSP.
142	 *
143	 * If width of "canonical tail" ever becomes variable, this will need
144	 * to be updated to remain correct on both old and new CPUs.
145	 *
146	 * Change top bits to match most significant bit (47th or 56th bit
147	 * depending on paging mode) in the address.
148	 */
149#ifdef CONFIG_X86_5LEVEL
150	ALTERNATIVE "shl $(64 - 48), %rcx; sar $(64 - 48), %rcx", \
151		"shl $(64 - 57), %rcx; sar $(64 - 57), %rcx", X86_FEATURE_LA57
152#else
153	shl	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
154	sar	$(64 - (__VIRTUAL_MASK_SHIFT+1)), %rcx
155#endif
156
157	/* If this changed %rcx, it was not canonical */
158	cmpq	%rcx, %r11
159	jne	swapgs_restore_regs_and_return_to_usermode
160
161	cmpq	$__USER_CS, CS(%rsp)		/* CS must match SYSRET */
162	jne	swapgs_restore_regs_and_return_to_usermode
163
164	movq	R11(%rsp), %r11
165	cmpq	%r11, EFLAGS(%rsp)		/* R11 == RFLAGS */
166	jne	swapgs_restore_regs_and_return_to_usermode
167
168	/*
169	 * SYSCALL clears RF when it saves RFLAGS in R11 and SYSRET cannot
170	 * restore RF properly. If the slowpath sets it for whatever reason, we
171	 * need to restore it correctly.
172	 *
173	 * SYSRET can restore TF, but unlike IRET, restoring TF results in a
174	 * trap from userspace immediately after SYSRET.  This would cause an
175	 * infinite loop whenever #DB happens with register state that satisfies
176	 * the opportunistic SYSRET conditions.  For example, single-stepping
177	 * this user code:
178	 *
179	 *           movq	$stuck_here, %rcx
180	 *           pushfq
181	 *           popq %r11
182	 *   stuck_here:
183	 *
184	 * would never get past 'stuck_here'.
185	 */
186	testq	$(X86_EFLAGS_RF|X86_EFLAGS_TF), %r11
187	jnz	swapgs_restore_regs_and_return_to_usermode
188
189	/* nothing to check for RSP */
190
191	cmpq	$__USER_DS, SS(%rsp)		/* SS must match SYSRET */
192	jne	swapgs_restore_regs_and_return_to_usermode
193
194	/*
195	 * We win! This label is here just for ease of understanding
196	 * perf profiles. Nothing jumps here.
197	 */
198syscall_return_via_sysret:
199	IBRS_EXIT
200	POP_REGS pop_rdi=0
201
202	/*
203	 * Now all regs are restored except RSP and RDI.
204	 * Save old stack pointer and switch to trampoline stack.
205	 */
206	movq	%rsp, %rdi
207	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
208	UNWIND_HINT_EMPTY
209
210	pushq	RSP-RDI(%rdi)	/* RSP */
211	pushq	(%rdi)		/* RDI */
212
213	/*
214	 * We are on the trampoline stack.  All regs except RDI are live.
215	 * We can do future final exit work right here.
216	 */
217	STACKLEAK_ERASE_NOCLOBBER
218
219	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
220
221	popq	%rdi
222	popq	%rsp
223SYM_INNER_LABEL(entry_SYSRETQ_unsafe_stack, SYM_L_GLOBAL)
224	ANNOTATE_NOENDBR
225	swapgs
226	sysretq
227SYM_INNER_LABEL(entry_SYSRETQ_end, SYM_L_GLOBAL)
228	ANNOTATE_NOENDBR
229	int3
230SYM_CODE_END(entry_SYSCALL_64)
231
232/*
233 * %rdi: prev task
234 * %rsi: next task
235 */
236.pushsection .text, "ax"
237SYM_FUNC_START(__switch_to_asm)
238	/*
239	 * Save callee-saved registers
240	 * This must match the order in inactive_task_frame
241	 */
242	pushq	%rbp
243	pushq	%rbx
244	pushq	%r12
245	pushq	%r13
246	pushq	%r14
247	pushq	%r15
248
249	/* switch stack */
250	movq	%rsp, TASK_threadsp(%rdi)
251	movq	TASK_threadsp(%rsi), %rsp
252
253#ifdef CONFIG_STACKPROTECTOR
254	movq	TASK_stack_canary(%rsi), %rbx
255	movq	%rbx, PER_CPU_VAR(fixed_percpu_data) + stack_canary_offset
256#endif
257
258	/*
259	 * When switching from a shallower to a deeper call stack
260	 * the RSB may either underflow or use entries populated
261	 * with userspace addresses. On CPUs where those concerns
262	 * exist, overwrite the RSB with entries which capture
263	 * speculative execution to prevent attack.
264	 */
265	FILL_RETURN_BUFFER %r12, RSB_CLEAR_LOOPS, X86_FEATURE_RSB_CTXSW
266
267	/* restore callee-saved registers */
268	popq	%r15
269	popq	%r14
270	popq	%r13
271	popq	%r12
272	popq	%rbx
273	popq	%rbp
274
275	jmp	__switch_to
276SYM_FUNC_END(__switch_to_asm)
277.popsection
278
279/*
280 * A newly forked process directly context switches into this address.
281 *
282 * rax: prev task we switched from
283 * rbx: kernel thread func (NULL for user thread)
284 * r12: kernel thread arg
285 */
286.pushsection .text, "ax"
287	__FUNC_ALIGN
288SYM_CODE_START_NOALIGN(ret_from_fork)
289	UNWIND_HINT_EMPTY
290	ANNOTATE_NOENDBR // copy_thread
291	movq	%rax, %rdi
292	call	schedule_tail			/* rdi: 'prev' task parameter */
293
294	testq	%rbx, %rbx			/* from kernel_thread? */
295	jnz	1f				/* kernel threads are uncommon */
296
2972:
298	UNWIND_HINT_REGS
299	movq	%rsp, %rdi
300	call	syscall_exit_to_user_mode	/* returns with IRQs disabled */
301	jmp	swapgs_restore_regs_and_return_to_usermode
302
3031:
304	/* kernel thread */
305	UNWIND_HINT_EMPTY
306	movq	%r12, %rdi
307	CALL_NOSPEC rbx
308	/*
309	 * A kernel thread is allowed to return here after successfully
310	 * calling kernel_execve().  Exit to userspace to complete the execve()
311	 * syscall.
312	 */
313	movq	$0, RAX(%rsp)
314	jmp	2b
315SYM_CODE_END(ret_from_fork)
316.popsection
317
318.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
319#ifdef CONFIG_DEBUG_ENTRY
320	pushq %rax
321	SAVE_FLAGS
322	testl $X86_EFLAGS_IF, %eax
323	jz .Lokay_\@
324	ud2
325.Lokay_\@:
326	popq %rax
327#endif
328.endm
329
330SYM_CODE_START_LOCAL(xen_error_entry)
331	UNWIND_HINT_FUNC
332	PUSH_AND_CLEAR_REGS save_ret=1
333	ENCODE_FRAME_POINTER 8
334	UNTRAIN_RET
335	RET
336SYM_CODE_END(xen_error_entry)
337
338/**
339 * idtentry_body - Macro to emit code calling the C function
340 * @cfunc:		C function to be called
341 * @has_error_code:	Hardware pushed error code on stack
342 */
343.macro idtentry_body cfunc has_error_code:req
344
345	/*
346	 * Call error_entry() and switch to the task stack if from userspace.
347	 *
348	 * When in XENPV, it is already in the task stack, and it can't fault
349	 * for native_iret() nor native_load_gs_index() since XENPV uses its
350	 * own pvops for IRET and load_gs_index().  And it doesn't need to
351	 * switch the CR3.  So it can skip invoking error_entry().
352	 */
353	ALTERNATIVE "call error_entry; movq %rax, %rsp", \
354		    "call xen_error_entry", X86_FEATURE_XENPV
355
356	ENCODE_FRAME_POINTER
357	UNWIND_HINT_REGS
358
359	movq	%rsp, %rdi			/* pt_regs pointer into 1st argument*/
360
361	.if \has_error_code == 1
362		movq	ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
363		movq	$-1, ORIG_RAX(%rsp)	/* no syscall to restart */
364	.endif
365
366	call	\cfunc
367
368	/* For some configurations \cfunc ends up being a noreturn. */
369	REACHABLE
370
371	jmp	error_return
372.endm
373
374/**
375 * idtentry - Macro to generate entry stubs for simple IDT entries
376 * @vector:		Vector number
377 * @asmsym:		ASM symbol for the entry point
378 * @cfunc:		C function to be called
379 * @has_error_code:	Hardware pushed error code on stack
380 *
381 * The macro emits code to set up the kernel context for straight forward
382 * and simple IDT entries. No IST stack, no paranoid entry checks.
383 */
384.macro idtentry vector asmsym cfunc has_error_code:req
385SYM_CODE_START(\asmsym)
386	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
387	ENDBR
388	ASM_CLAC
389	cld
390
391	.if \has_error_code == 0
392		pushq	$-1			/* ORIG_RAX: no syscall to restart */
393	.endif
394
395	.if \vector == X86_TRAP_BP
396		/*
397		 * If coming from kernel space, create a 6-word gap to allow the
398		 * int3 handler to emulate a call instruction.
399		 */
400		testb	$3, CS-ORIG_RAX(%rsp)
401		jnz	.Lfrom_usermode_no_gap_\@
402		.rept	6
403		pushq	5*8(%rsp)
404		.endr
405		UNWIND_HINT_IRET_REGS offset=8
406.Lfrom_usermode_no_gap_\@:
407	.endif
408
409	idtentry_body \cfunc \has_error_code
410
411_ASM_NOKPROBE(\asmsym)
412SYM_CODE_END(\asmsym)
413.endm
414
415/*
416 * Interrupt entry/exit.
417 *
418 + The interrupt stubs push (vector) onto the stack, which is the error_code
419 * position of idtentry exceptions, and jump to one of the two idtentry points
420 * (common/spurious).
421 *
422 * common_interrupt is a hotpath, align it to a cache line
423 */
424.macro idtentry_irq vector cfunc
425	.p2align CONFIG_X86_L1_CACHE_SHIFT
426	idtentry \vector asm_\cfunc \cfunc has_error_code=1
427.endm
428
429/*
430 * System vectors which invoke their handlers directly and are not
431 * going through the regular common device interrupt handling code.
432 */
433.macro idtentry_sysvec vector cfunc
434	idtentry \vector asm_\cfunc \cfunc has_error_code=0
435.endm
436
437/**
438 * idtentry_mce_db - Macro to generate entry stubs for #MC and #DB
439 * @vector:		Vector number
440 * @asmsym:		ASM symbol for the entry point
441 * @cfunc:		C function to be called
442 *
443 * The macro emits code to set up the kernel context for #MC and #DB
444 *
445 * If the entry comes from user space it uses the normal entry path
446 * including the return to user space work and preemption checks on
447 * exit.
448 *
449 * If hits in kernel mode then it needs to go through the paranoid
450 * entry as the exception can hit any random state. No preemption
451 * check on exit to keep the paranoid path simple.
452 */
453.macro idtentry_mce_db vector asmsym cfunc
454SYM_CODE_START(\asmsym)
455	UNWIND_HINT_IRET_REGS
456	ENDBR
457	ASM_CLAC
458	cld
459
460	pushq	$-1			/* ORIG_RAX: no syscall to restart */
461
462	/*
463	 * If the entry is from userspace, switch stacks and treat it as
464	 * a normal entry.
465	 */
466	testb	$3, CS-ORIG_RAX(%rsp)
467	jnz	.Lfrom_usermode_switch_stack_\@
468
469	/* paranoid_entry returns GS information for paranoid_exit in EBX. */
470	call	paranoid_entry
471
472	UNWIND_HINT_REGS
473
474	movq	%rsp, %rdi		/* pt_regs pointer */
475
476	call	\cfunc
477
478	jmp	paranoid_exit
479
480	/* Switch to the regular task stack and use the noist entry point */
481.Lfrom_usermode_switch_stack_\@:
482	idtentry_body noist_\cfunc, has_error_code=0
483
484_ASM_NOKPROBE(\asmsym)
485SYM_CODE_END(\asmsym)
486.endm
487
488#ifdef CONFIG_AMD_MEM_ENCRYPT
489/**
490 * idtentry_vc - Macro to generate entry stub for #VC
491 * @vector:		Vector number
492 * @asmsym:		ASM symbol for the entry point
493 * @cfunc:		C function to be called
494 *
495 * The macro emits code to set up the kernel context for #VC. The #VC handler
496 * runs on an IST stack and needs to be able to cause nested #VC exceptions.
497 *
498 * To make this work the #VC entry code tries its best to pretend it doesn't use
499 * an IST stack by switching to the task stack if coming from user-space (which
500 * includes early SYSCALL entry path) or back to the stack in the IRET frame if
501 * entered from kernel-mode.
502 *
503 * If entered from kernel-mode the return stack is validated first, and if it is
504 * not safe to use (e.g. because it points to the entry stack) the #VC handler
505 * will switch to a fall-back stack (VC2) and call a special handler function.
506 *
507 * The macro is only used for one vector, but it is planned to be extended in
508 * the future for the #HV exception.
509 */
510.macro idtentry_vc vector asmsym cfunc
511SYM_CODE_START(\asmsym)
512	UNWIND_HINT_IRET_REGS
513	ENDBR
514	ASM_CLAC
515	cld
516
517	/*
518	 * If the entry is from userspace, switch stacks and treat it as
519	 * a normal entry.
520	 */
521	testb	$3, CS-ORIG_RAX(%rsp)
522	jnz	.Lfrom_usermode_switch_stack_\@
523
524	/*
525	 * paranoid_entry returns SWAPGS flag for paranoid_exit in EBX.
526	 * EBX == 0 -> SWAPGS, EBX == 1 -> no SWAPGS
527	 */
528	call	paranoid_entry
529
530	UNWIND_HINT_REGS
531
532	/*
533	 * Switch off the IST stack to make it free for nested exceptions. The
534	 * vc_switch_off_ist() function will switch back to the interrupted
535	 * stack if it is safe to do so. If not it switches to the VC fall-back
536	 * stack.
537	 */
538	movq	%rsp, %rdi		/* pt_regs pointer */
539	call	vc_switch_off_ist
540	movq	%rax, %rsp		/* Switch to new stack */
541
542	ENCODE_FRAME_POINTER
543	UNWIND_HINT_REGS
544
545	/* Update pt_regs */
546	movq	ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
547	movq	$-1, ORIG_RAX(%rsp)	/* no syscall to restart */
548
549	movq	%rsp, %rdi		/* pt_regs pointer */
550
551	call	kernel_\cfunc
552
553	/*
554	 * No need to switch back to the IST stack. The current stack is either
555	 * identical to the stack in the IRET frame or the VC fall-back stack,
556	 * so it is definitely mapped even with PTI enabled.
557	 */
558	jmp	paranoid_exit
559
560	/* Switch to the regular task stack */
561.Lfrom_usermode_switch_stack_\@:
562	idtentry_body user_\cfunc, has_error_code=1
563
564_ASM_NOKPROBE(\asmsym)
565SYM_CODE_END(\asmsym)
566.endm
567#endif
568
569/*
570 * Double fault entry. Straight paranoid. No checks from which context
571 * this comes because for the espfix induced #DF this would do the wrong
572 * thing.
573 */
574.macro idtentry_df vector asmsym cfunc
575SYM_CODE_START(\asmsym)
576	UNWIND_HINT_IRET_REGS offset=8
577	ENDBR
578	ASM_CLAC
579	cld
580
581	/* paranoid_entry returns GS information for paranoid_exit in EBX. */
582	call	paranoid_entry
583	UNWIND_HINT_REGS
584
585	movq	%rsp, %rdi		/* pt_regs pointer into first argument */
586	movq	ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
587	movq	$-1, ORIG_RAX(%rsp)	/* no syscall to restart */
588	call	\cfunc
589
590	/* For some configurations \cfunc ends up being a noreturn. */
591	REACHABLE
592
593	jmp	paranoid_exit
594
595_ASM_NOKPROBE(\asmsym)
596SYM_CODE_END(\asmsym)
597.endm
598
599/*
600 * Include the defines which emit the idt entries which are shared
601 * shared between 32 and 64 bit and emit the __irqentry_text_* markers
602 * so the stacktrace boundary checks work.
603 */
604	__ALIGN
605	.globl __irqentry_text_start
606__irqentry_text_start:
607
608#include <asm/idtentry.h>
609
610	__ALIGN
611	.globl __irqentry_text_end
612__irqentry_text_end:
613	ANNOTATE_NOENDBR
614
615SYM_CODE_START_LOCAL(common_interrupt_return)
616SYM_INNER_LABEL(swapgs_restore_regs_and_return_to_usermode, SYM_L_GLOBAL)
617	IBRS_EXIT
618#ifdef CONFIG_DEBUG_ENTRY
619	/* Assert that pt_regs indicates user mode. */
620	testb	$3, CS(%rsp)
621	jnz	1f
622	ud2
6231:
624#endif
625#ifdef CONFIG_XEN_PV
626	ALTERNATIVE "", "jmp xenpv_restore_regs_and_return_to_usermode", X86_FEATURE_XENPV
627#endif
628
629	POP_REGS pop_rdi=0
630
631	/*
632	 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS.
633	 * Save old stack pointer and switch to trampoline stack.
634	 */
635	movq	%rsp, %rdi
636	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
637	UNWIND_HINT_EMPTY
638
639	/* Copy the IRET frame to the trampoline stack. */
640	pushq	6*8(%rdi)	/* SS */
641	pushq	5*8(%rdi)	/* RSP */
642	pushq	4*8(%rdi)	/* EFLAGS */
643	pushq	3*8(%rdi)	/* CS */
644	pushq	2*8(%rdi)	/* RIP */
645
646	/* Push user RDI on the trampoline stack. */
647	pushq	(%rdi)
648
649	/*
650	 * We are on the trampoline stack.  All regs except RDI are live.
651	 * We can do future final exit work right here.
652	 */
653	STACKLEAK_ERASE_NOCLOBBER
654
655	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
656
657	/* Restore RDI. */
658	popq	%rdi
659	swapgs
660	jmp	.Lnative_iret
661
662
663SYM_INNER_LABEL(restore_regs_and_return_to_kernel, SYM_L_GLOBAL)
664#ifdef CONFIG_DEBUG_ENTRY
665	/* Assert that pt_regs indicates kernel mode. */
666	testb	$3, CS(%rsp)
667	jz	1f
668	ud2
6691:
670#endif
671	POP_REGS
672	addq	$8, %rsp	/* skip regs->orig_ax */
673	/*
674	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
675	 * when returning from IPI handler.
676	 */
677#ifdef CONFIG_XEN_PV
678SYM_INNER_LABEL(early_xen_iret_patch, SYM_L_GLOBAL)
679	ANNOTATE_NOENDBR
680	.byte 0xe9
681	.long .Lnative_iret - (. + 4)
682#endif
683
684.Lnative_iret:
685	UNWIND_HINT_IRET_REGS
686	/*
687	 * Are we returning to a stack segment from the LDT?  Note: in
688	 * 64-bit mode SS:RSP on the exception stack is always valid.
689	 */
690#ifdef CONFIG_X86_ESPFIX64
691	testb	$4, (SS-RIP)(%rsp)
692	jnz	native_irq_return_ldt
693#endif
694
695SYM_INNER_LABEL(native_irq_return_iret, SYM_L_GLOBAL)
696	ANNOTATE_NOENDBR // exc_double_fault
697	/*
698	 * This may fault.  Non-paranoid faults on return to userspace are
699	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
700	 * Double-faults due to espfix64 are handled in exc_double_fault.
701	 * Other faults here are fatal.
702	 */
703	iretq
704
705#ifdef CONFIG_X86_ESPFIX64
706native_irq_return_ldt:
707	/*
708	 * We are running with user GSBASE.  All GPRs contain their user
709	 * values.  We have a percpu ESPFIX stack that is eight slots
710	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
711	 * of the ESPFIX stack.
712	 *
713	 * We clobber RAX and RDI in this code.  We stash RDI on the
714	 * normal stack and RAX on the ESPFIX stack.
715	 *
716	 * The ESPFIX stack layout we set up looks like this:
717	 *
718	 * --- top of ESPFIX stack ---
719	 * SS
720	 * RSP
721	 * RFLAGS
722	 * CS
723	 * RIP  <-- RSP points here when we're done
724	 * RAX  <-- espfix_waddr points here
725	 * --- bottom of ESPFIX stack ---
726	 */
727
728	pushq	%rdi				/* Stash user RDI */
729	swapgs					/* to kernel GS */
730	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi	/* to kernel CR3 */
731
732	movq	PER_CPU_VAR(espfix_waddr), %rdi
733	movq	%rax, (0*8)(%rdi)		/* user RAX */
734	movq	(1*8)(%rsp), %rax		/* user RIP */
735	movq	%rax, (1*8)(%rdi)
736	movq	(2*8)(%rsp), %rax		/* user CS */
737	movq	%rax, (2*8)(%rdi)
738	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
739	movq	%rax, (3*8)(%rdi)
740	movq	(5*8)(%rsp), %rax		/* user SS */
741	movq	%rax, (5*8)(%rdi)
742	movq	(4*8)(%rsp), %rax		/* user RSP */
743	movq	%rax, (4*8)(%rdi)
744	/* Now RAX == RSP. */
745
746	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
747
748	/*
749	 * espfix_stack[31:16] == 0.  The page tables are set up such that
750	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
751	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
752	 * the same page.  Set up RSP so that RSP[31:16] contains the
753	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
754	 * still points to an RO alias of the ESPFIX stack.
755	 */
756	orq	PER_CPU_VAR(espfix_stack), %rax
757
758	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
759	swapgs					/* to user GS */
760	popq	%rdi				/* Restore user RDI */
761
762	movq	%rax, %rsp
763	UNWIND_HINT_IRET_REGS offset=8
764
765	/*
766	 * At this point, we cannot write to the stack any more, but we can
767	 * still read.
768	 */
769	popq	%rax				/* Restore user RAX */
770
771	/*
772	 * RSP now points to an ordinary IRET frame, except that the page
773	 * is read-only and RSP[31:16] are preloaded with the userspace
774	 * values.  We can now IRET back to userspace.
775	 */
776	jmp	native_irq_return_iret
777#endif
778SYM_CODE_END(common_interrupt_return)
779_ASM_NOKPROBE(common_interrupt_return)
780
781/*
782 * Reload gs selector with exception handling
783 * edi:  new selector
784 *
785 * Is in entry.text as it shouldn't be instrumented.
786 */
787SYM_FUNC_START(asm_load_gs_index)
788	FRAME_BEGIN
789	swapgs
790.Lgs_change:
791	ANNOTATE_NOENDBR // error_entry
792	movl	%edi, %gs
7932:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
794	swapgs
795	FRAME_END
796	RET
797
798	/* running with kernelgs */
799.Lbad_gs:
800	swapgs					/* switch back to user gs */
801.macro ZAP_GS
802	/* This can't be a string because the preprocessor needs to see it. */
803	movl $__USER_DS, %eax
804	movl %eax, %gs
805.endm
806	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
807	xorl	%eax, %eax
808	movl	%eax, %gs
809	jmp	2b
810
811	_ASM_EXTABLE(.Lgs_change, .Lbad_gs)
812
813SYM_FUNC_END(asm_load_gs_index)
814EXPORT_SYMBOL(asm_load_gs_index)
815
816#ifdef CONFIG_XEN_PV
817/*
818 * A note on the "critical region" in our callback handler.
819 * We want to avoid stacking callback handlers due to events occurring
820 * during handling of the last event. To do this, we keep events disabled
821 * until we've done all processing. HOWEVER, we must enable events before
822 * popping the stack frame (can't be done atomically) and so it would still
823 * be possible to get enough handler activations to overflow the stack.
824 * Although unlikely, bugs of that kind are hard to track down, so we'd
825 * like to avoid the possibility.
826 * So, on entry to the handler we detect whether we interrupted an
827 * existing activation in its critical region -- if so, we pop the current
828 * activation and restart the handler using the previous one.
829 *
830 * C calling convention: exc_xen_hypervisor_callback(struct *pt_regs)
831 */
832	__FUNC_ALIGN
833SYM_CODE_START_LOCAL_NOALIGN(exc_xen_hypervisor_callback)
834
835/*
836 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
837 * see the correct pointer to the pt_regs
838 */
839	UNWIND_HINT_FUNC
840	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
841	UNWIND_HINT_REGS
842
843	call	xen_pv_evtchn_do_upcall
844
845	jmp	error_return
846SYM_CODE_END(exc_xen_hypervisor_callback)
847
848/*
849 * Hypervisor uses this for application faults while it executes.
850 * We get here for two reasons:
851 *  1. Fault while reloading DS, ES, FS or GS
852 *  2. Fault while executing IRET
853 * Category 1 we do not need to fix up as Xen has already reloaded all segment
854 * registers that could be reloaded and zeroed the others.
855 * Category 2 we fix up by killing the current process. We cannot use the
856 * normal Linux return path in this case because if we use the IRET hypercall
857 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
858 * We distinguish between categories by comparing each saved segment register
859 * with its current contents: any discrepancy means we in category 1.
860 */
861	__FUNC_ALIGN
862SYM_CODE_START_NOALIGN(xen_failsafe_callback)
863	UNWIND_HINT_EMPTY
864	ENDBR
865	movl	%ds, %ecx
866	cmpw	%cx, 0x10(%rsp)
867	jne	1f
868	movl	%es, %ecx
869	cmpw	%cx, 0x18(%rsp)
870	jne	1f
871	movl	%fs, %ecx
872	cmpw	%cx, 0x20(%rsp)
873	jne	1f
874	movl	%gs, %ecx
875	cmpw	%cx, 0x28(%rsp)
876	jne	1f
877	/* All segments match their saved values => Category 2 (Bad IRET). */
878	movq	(%rsp), %rcx
879	movq	8(%rsp), %r11
880	addq	$0x30, %rsp
881	pushq	$0				/* RIP */
882	UNWIND_HINT_IRET_REGS offset=8
883	jmp	asm_exc_general_protection
8841:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
885	movq	(%rsp), %rcx
886	movq	8(%rsp), %r11
887	addq	$0x30, %rsp
888	UNWIND_HINT_IRET_REGS
889	pushq	$-1 /* orig_ax = -1 => not a system call */
890	PUSH_AND_CLEAR_REGS
891	ENCODE_FRAME_POINTER
892	jmp	error_return
893SYM_CODE_END(xen_failsafe_callback)
894#endif /* CONFIG_XEN_PV */
895
896/*
897 * Save all registers in pt_regs. Return GSBASE related information
898 * in EBX depending on the availability of the FSGSBASE instructions:
899 *
900 * FSGSBASE	R/EBX
901 *     N        0 -> SWAPGS on exit
902 *              1 -> no SWAPGS on exit
903 *
904 *     Y        GSBASE value at entry, must be restored in paranoid_exit
905 *
906 * R14 - old CR3
907 * R15 - old SPEC_CTRL
908 */
909SYM_CODE_START_LOCAL(paranoid_entry)
910	UNWIND_HINT_FUNC
911	PUSH_AND_CLEAR_REGS save_ret=1
912	ENCODE_FRAME_POINTER 8
913
914	/*
915	 * Always stash CR3 in %r14.  This value will be restored,
916	 * verbatim, at exit.  Needed if paranoid_entry interrupted
917	 * another entry that already switched to the user CR3 value
918	 * but has not yet returned to userspace.
919	 *
920	 * This is also why CS (stashed in the "iret frame" by the
921	 * hardware at entry) can not be used: this may be a return
922	 * to kernel code, but with a user CR3 value.
923	 *
924	 * Switching CR3 does not depend on kernel GSBASE so it can
925	 * be done before switching to the kernel GSBASE. This is
926	 * required for FSGSBASE because the kernel GSBASE has to
927	 * be retrieved from a kernel internal table.
928	 */
929	SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
930
931	/*
932	 * Handling GSBASE depends on the availability of FSGSBASE.
933	 *
934	 * Without FSGSBASE the kernel enforces that negative GSBASE
935	 * values indicate kernel GSBASE. With FSGSBASE no assumptions
936	 * can be made about the GSBASE value when entering from user
937	 * space.
938	 */
939	ALTERNATIVE "jmp .Lparanoid_entry_checkgs", "", X86_FEATURE_FSGSBASE
940
941	/*
942	 * Read the current GSBASE and store it in %rbx unconditionally,
943	 * retrieve and set the current CPUs kernel GSBASE. The stored value
944	 * has to be restored in paranoid_exit unconditionally.
945	 *
946	 * The unconditional write to GS base below ensures that no subsequent
947	 * loads based on a mispredicted GS base can happen, therefore no LFENCE
948	 * is needed here.
949	 */
950	SAVE_AND_SET_GSBASE scratch_reg=%rax save_reg=%rbx
951	jmp .Lparanoid_gsbase_done
952
953.Lparanoid_entry_checkgs:
954	/* EBX = 1 -> kernel GSBASE active, no restore required */
955	movl	$1, %ebx
956
957	/*
958	 * The kernel-enforced convention is a negative GSBASE indicates
959	 * a kernel value. No SWAPGS needed on entry and exit.
960	 */
961	movl	$MSR_GS_BASE, %ecx
962	rdmsr
963	testl	%edx, %edx
964	js	.Lparanoid_kernel_gsbase
965
966	/* EBX = 0 -> SWAPGS required on exit */
967	xorl	%ebx, %ebx
968	swapgs
969.Lparanoid_kernel_gsbase:
970	FENCE_SWAPGS_KERNEL_ENTRY
971.Lparanoid_gsbase_done:
972
973	/*
974	 * Once we have CR3 and %GS setup save and set SPEC_CTRL. Just like
975	 * CR3 above, keep the old value in a callee saved register.
976	 */
977	IBRS_ENTER save_reg=%r15
978	UNTRAIN_RET
979
980	RET
981SYM_CODE_END(paranoid_entry)
982
983/*
984 * "Paranoid" exit path from exception stack.  This is invoked
985 * only on return from non-NMI IST interrupts that came
986 * from kernel space.
987 *
988 * We may be returning to very strange contexts (e.g. very early
989 * in syscall entry), so checking for preemption here would
990 * be complicated.  Fortunately, there's no good reason to try
991 * to handle preemption here.
992 *
993 * R/EBX contains the GSBASE related information depending on the
994 * availability of the FSGSBASE instructions:
995 *
996 * FSGSBASE	R/EBX
997 *     N        0 -> SWAPGS on exit
998 *              1 -> no SWAPGS on exit
999 *
1000 *     Y        User space GSBASE, must be restored unconditionally
1001 *
1002 * R14 - old CR3
1003 * R15 - old SPEC_CTRL
1004 */
1005SYM_CODE_START_LOCAL(paranoid_exit)
1006	UNWIND_HINT_REGS
1007
1008	/*
1009	 * Must restore IBRS state before both CR3 and %GS since we need access
1010	 * to the per-CPU x86_spec_ctrl_shadow variable.
1011	 */
1012	IBRS_EXIT save_reg=%r15
1013
1014	/*
1015	 * The order of operations is important. RESTORE_CR3 requires
1016	 * kernel GSBASE.
1017	 *
1018	 * NB to anyone to try to optimize this code: this code does
1019	 * not execute at all for exceptions from user mode. Those
1020	 * exceptions go through error_exit instead.
1021	 */
1022	RESTORE_CR3	scratch_reg=%rax save_reg=%r14
1023
1024	/* Handle the three GSBASE cases */
1025	ALTERNATIVE "jmp .Lparanoid_exit_checkgs", "", X86_FEATURE_FSGSBASE
1026
1027	/* With FSGSBASE enabled, unconditionally restore GSBASE */
1028	wrgsbase	%rbx
1029	jmp		restore_regs_and_return_to_kernel
1030
1031.Lparanoid_exit_checkgs:
1032	/* On non-FSGSBASE systems, conditionally do SWAPGS */
1033	testl		%ebx, %ebx
1034	jnz		restore_regs_and_return_to_kernel
1035
1036	/* We are returning to a context with user GSBASE */
1037	swapgs
1038	jmp		restore_regs_and_return_to_kernel
1039SYM_CODE_END(paranoid_exit)
1040
1041/*
1042 * Switch GS and CR3 if needed.
1043 */
1044SYM_CODE_START_LOCAL(error_entry)
1045	UNWIND_HINT_FUNC
1046
1047	PUSH_AND_CLEAR_REGS save_ret=1
1048	ENCODE_FRAME_POINTER 8
1049
1050	testb	$3, CS+8(%rsp)
1051	jz	.Lerror_kernelspace
1052
1053	/*
1054	 * We entered from user mode or we're pretending to have entered
1055	 * from user mode due to an IRET fault.
1056	 */
1057	swapgs
1058	FENCE_SWAPGS_USER_ENTRY
1059	/* We have user CR3.  Change to kernel CR3. */
1060	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1061	IBRS_ENTER
1062	UNTRAIN_RET
1063
1064	leaq	8(%rsp), %rdi			/* arg0 = pt_regs pointer */
1065.Lerror_entry_from_usermode_after_swapgs:
1066
1067	/* Put us onto the real thread stack. */
1068	call	sync_regs
1069	RET
1070
1071	/*
1072	 * There are two places in the kernel that can potentially fault with
1073	 * usergs. Handle them here.  B stepping K8s sometimes report a
1074	 * truncated RIP for IRET exceptions returning to compat mode. Check
1075	 * for these here too.
1076	 */
1077.Lerror_kernelspace:
1078	leaq	native_irq_return_iret(%rip), %rcx
1079	cmpq	%rcx, RIP+8(%rsp)
1080	je	.Lerror_bad_iret
1081	movl	%ecx, %eax			/* zero extend */
1082	cmpq	%rax, RIP+8(%rsp)
1083	je	.Lbstep_iret
1084	cmpq	$.Lgs_change, RIP+8(%rsp)
1085	jne	.Lerror_entry_done_lfence
1086
1087	/*
1088	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1089	 * gsbase and proceed.  We'll fix up the exception and land in
1090	 * .Lgs_change's error handler with kernel gsbase.
1091	 */
1092	swapgs
1093
1094	/*
1095	 * Issue an LFENCE to prevent GS speculation, regardless of whether it is a
1096	 * kernel or user gsbase.
1097	 */
1098.Lerror_entry_done_lfence:
1099	FENCE_SWAPGS_KERNEL_ENTRY
1100	leaq	8(%rsp), %rax			/* return pt_regs pointer */
1101	ANNOTATE_UNRET_END
1102	RET
1103
1104.Lbstep_iret:
1105	/* Fix truncated RIP */
1106	movq	%rcx, RIP+8(%rsp)
1107	/* fall through */
1108
1109.Lerror_bad_iret:
1110	/*
1111	 * We came from an IRET to user mode, so we have user
1112	 * gsbase and CR3.  Switch to kernel gsbase and CR3:
1113	 */
1114	swapgs
1115	FENCE_SWAPGS_USER_ENTRY
1116	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1117	IBRS_ENTER
1118	UNTRAIN_RET
1119
1120	/*
1121	 * Pretend that the exception came from user mode: set up pt_regs
1122	 * as if we faulted immediately after IRET.
1123	 */
1124	leaq	8(%rsp), %rdi			/* arg0 = pt_regs pointer */
1125	call	fixup_bad_iret
1126	mov	%rax, %rdi
1127	jmp	.Lerror_entry_from_usermode_after_swapgs
1128SYM_CODE_END(error_entry)
1129
1130SYM_CODE_START_LOCAL(error_return)
1131	UNWIND_HINT_REGS
1132	DEBUG_ENTRY_ASSERT_IRQS_OFF
1133	testb	$3, CS(%rsp)
1134	jz	restore_regs_and_return_to_kernel
1135	jmp	swapgs_restore_regs_and_return_to_usermode
1136SYM_CODE_END(error_return)
1137
1138/*
1139 * Runs on exception stack.  Xen PV does not go through this path at all,
1140 * so we can use real assembly here.
1141 *
1142 * Registers:
1143 *	%r14: Used to save/restore the CR3 of the interrupted context
1144 *	      when PAGE_TABLE_ISOLATION is in use.  Do not clobber.
1145 */
1146SYM_CODE_START(asm_exc_nmi)
1147	UNWIND_HINT_IRET_REGS
1148	ENDBR
1149
1150	/*
1151	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1152	 * the iretq it performs will take us out of NMI context.
1153	 * This means that we can have nested NMIs where the next
1154	 * NMI is using the top of the stack of the previous NMI. We
1155	 * can't let it execute because the nested NMI will corrupt the
1156	 * stack of the previous NMI. NMI handlers are not re-entrant
1157	 * anyway.
1158	 *
1159	 * To handle this case we do the following:
1160	 *  Check the a special location on the stack that contains
1161	 *  a variable that is set when NMIs are executing.
1162	 *  The interrupted task's stack is also checked to see if it
1163	 *  is an NMI stack.
1164	 *  If the variable is not set and the stack is not the NMI
1165	 *  stack then:
1166	 *    o Set the special variable on the stack
1167	 *    o Copy the interrupt frame into an "outermost" location on the
1168	 *      stack
1169	 *    o Copy the interrupt frame into an "iret" location on the stack
1170	 *    o Continue processing the NMI
1171	 *  If the variable is set or the previous stack is the NMI stack:
1172	 *    o Modify the "iret" location to jump to the repeat_nmi
1173	 *    o return back to the first NMI
1174	 *
1175	 * Now on exit of the first NMI, we first clear the stack variable
1176	 * The NMI stack will tell any nested NMIs at that point that it is
1177	 * nested. Then we pop the stack normally with iret, and if there was
1178	 * a nested NMI that updated the copy interrupt stack frame, a
1179	 * jump will be made to the repeat_nmi code that will handle the second
1180	 * NMI.
1181	 *
1182	 * However, espfix prevents us from directly returning to userspace
1183	 * with a single IRET instruction.  Similarly, IRET to user mode
1184	 * can fault.  We therefore handle NMIs from user space like
1185	 * other IST entries.
1186	 */
1187
1188	ASM_CLAC
1189	cld
1190
1191	/* Use %rdx as our temp variable throughout */
1192	pushq	%rdx
1193
1194	testb	$3, CS-RIP+8(%rsp)
1195	jz	.Lnmi_from_kernel
1196
1197	/*
1198	 * NMI from user mode.  We need to run on the thread stack, but we
1199	 * can't go through the normal entry paths: NMIs are masked, and
1200	 * we don't want to enable interrupts, because then we'll end
1201	 * up in an awkward situation in which IRQs are on but NMIs
1202	 * are off.
1203	 *
1204	 * We also must not push anything to the stack before switching
1205	 * stacks lest we corrupt the "NMI executing" variable.
1206	 */
1207
1208	swapgs
1209	FENCE_SWAPGS_USER_ENTRY
1210	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx
1211	movq	%rsp, %rdx
1212	movq	PER_CPU_VAR(pcpu_hot + X86_top_of_stack), %rsp
1213	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1214	pushq	5*8(%rdx)	/* pt_regs->ss */
1215	pushq	4*8(%rdx)	/* pt_regs->rsp */
1216	pushq	3*8(%rdx)	/* pt_regs->flags */
1217	pushq	2*8(%rdx)	/* pt_regs->cs */
1218	pushq	1*8(%rdx)	/* pt_regs->rip */
1219	UNWIND_HINT_IRET_REGS
1220	pushq   $-1		/* pt_regs->orig_ax */
1221	PUSH_AND_CLEAR_REGS rdx=(%rdx)
1222	ENCODE_FRAME_POINTER
1223
1224	IBRS_ENTER
1225	UNTRAIN_RET
1226
1227	/*
1228	 * At this point we no longer need to worry about stack damage
1229	 * due to nesting -- we're on the normal thread stack and we're
1230	 * done with the NMI stack.
1231	 */
1232
1233	movq	%rsp, %rdi
1234	movq	$-1, %rsi
1235	call	exc_nmi
1236
1237	/*
1238	 * Return back to user mode.  We must *not* do the normal exit
1239	 * work, because we don't want to enable interrupts.
1240	 */
1241	jmp	swapgs_restore_regs_and_return_to_usermode
1242
1243.Lnmi_from_kernel:
1244	/*
1245	 * Here's what our stack frame will look like:
1246	 * +---------------------------------------------------------+
1247	 * | original SS                                             |
1248	 * | original Return RSP                                     |
1249	 * | original RFLAGS                                         |
1250	 * | original CS                                             |
1251	 * | original RIP                                            |
1252	 * +---------------------------------------------------------+
1253	 * | temp storage for rdx                                    |
1254	 * +---------------------------------------------------------+
1255	 * | "NMI executing" variable                                |
1256	 * +---------------------------------------------------------+
1257	 * | iret SS          } Copied from "outermost" frame        |
1258	 * | iret Return RSP  } on each loop iteration; overwritten  |
1259	 * | iret RFLAGS      } by a nested NMI to force another     |
1260	 * | iret CS          } iteration if needed.                 |
1261	 * | iret RIP         }                                      |
1262	 * +---------------------------------------------------------+
1263	 * | outermost SS          } initialized in first_nmi;       |
1264	 * | outermost Return RSP  } will not be changed before      |
1265	 * | outermost RFLAGS      } NMI processing is done.         |
1266	 * | outermost CS          } Copied to "iret" frame on each  |
1267	 * | outermost RIP         } iteration.                      |
1268	 * +---------------------------------------------------------+
1269	 * | pt_regs                                                 |
1270	 * +---------------------------------------------------------+
1271	 *
1272	 * The "original" frame is used by hardware.  Before re-enabling
1273	 * NMIs, we need to be done with it, and we need to leave enough
1274	 * space for the asm code here.
1275	 *
1276	 * We return by executing IRET while RSP points to the "iret" frame.
1277	 * That will either return for real or it will loop back into NMI
1278	 * processing.
1279	 *
1280	 * The "outermost" frame is copied to the "iret" frame on each
1281	 * iteration of the loop, so each iteration starts with the "iret"
1282	 * frame pointing to the final return target.
1283	 */
1284
1285	/*
1286	 * Determine whether we're a nested NMI.
1287	 *
1288	 * If we interrupted kernel code between repeat_nmi and
1289	 * end_repeat_nmi, then we are a nested NMI.  We must not
1290	 * modify the "iret" frame because it's being written by
1291	 * the outer NMI.  That's okay; the outer NMI handler is
1292	 * about to about to call exc_nmi() anyway, so we can just
1293	 * resume the outer NMI.
1294	 */
1295
1296	movq	$repeat_nmi, %rdx
1297	cmpq	8(%rsp), %rdx
1298	ja	1f
1299	movq	$end_repeat_nmi, %rdx
1300	cmpq	8(%rsp), %rdx
1301	ja	nested_nmi_out
13021:
1303
1304	/*
1305	 * Now check "NMI executing".  If it's set, then we're nested.
1306	 * This will not detect if we interrupted an outer NMI just
1307	 * before IRET.
1308	 */
1309	cmpl	$1, -8(%rsp)
1310	je	nested_nmi
1311
1312	/*
1313	 * Now test if the previous stack was an NMI stack.  This covers
1314	 * the case where we interrupt an outer NMI after it clears
1315	 * "NMI executing" but before IRET.  We need to be careful, though:
1316	 * there is one case in which RSP could point to the NMI stack
1317	 * despite there being no NMI active: naughty userspace controls
1318	 * RSP at the very beginning of the SYSCALL targets.  We can
1319	 * pull a fast one on naughty userspace, though: we program
1320	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1321	 * if it controls the kernel's RSP.  We set DF before we clear
1322	 * "NMI executing".
1323	 */
1324	lea	6*8(%rsp), %rdx
1325	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1326	cmpq	%rdx, 4*8(%rsp)
1327	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1328	ja	first_nmi
1329
1330	subq	$EXCEPTION_STKSZ, %rdx
1331	cmpq	%rdx, 4*8(%rsp)
1332	/* If it is below the NMI stack, it is a normal NMI */
1333	jb	first_nmi
1334
1335	/* Ah, it is within the NMI stack. */
1336
1337	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1338	jz	first_nmi	/* RSP was user controlled. */
1339
1340	/* This is a nested NMI. */
1341
1342nested_nmi:
1343	/*
1344	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1345	 * iteration of NMI handling.
1346	 */
1347	subq	$8, %rsp
1348	leaq	-10*8(%rsp), %rdx
1349	pushq	$__KERNEL_DS
1350	pushq	%rdx
1351	pushfq
1352	pushq	$__KERNEL_CS
1353	pushq	$repeat_nmi
1354
1355	/* Put stack back */
1356	addq	$(6*8), %rsp
1357
1358nested_nmi_out:
1359	popq	%rdx
1360
1361	/* We are returning to kernel mode, so this cannot result in a fault. */
1362	iretq
1363
1364first_nmi:
1365	/* Restore rdx. */
1366	movq	(%rsp), %rdx
1367
1368	/* Make room for "NMI executing". */
1369	pushq	$0
1370
1371	/* Leave room for the "iret" frame */
1372	subq	$(5*8), %rsp
1373
1374	/* Copy the "original" frame to the "outermost" frame */
1375	.rept 5
1376	pushq	11*8(%rsp)
1377	.endr
1378	UNWIND_HINT_IRET_REGS
1379
1380	/* Everything up to here is safe from nested NMIs */
1381
1382#ifdef CONFIG_DEBUG_ENTRY
1383	/*
1384	 * For ease of testing, unmask NMIs right away.  Disabled by
1385	 * default because IRET is very expensive.
1386	 */
1387	pushq	$0		/* SS */
1388	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1389	addq	$8, (%rsp)	/* Fix up RSP */
1390	pushfq			/* RFLAGS */
1391	pushq	$__KERNEL_CS	/* CS */
1392	pushq	$1f		/* RIP */
1393	iretq			/* continues at repeat_nmi below */
1394	UNWIND_HINT_IRET_REGS
13951:
1396#endif
1397
1398repeat_nmi:
1399	ANNOTATE_NOENDBR // this code
1400	/*
1401	 * If there was a nested NMI, the first NMI's iret will return
1402	 * here. But NMIs are still enabled and we can take another
1403	 * nested NMI. The nested NMI checks the interrupted RIP to see
1404	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1405	 * it will just return, as we are about to repeat an NMI anyway.
1406	 * This makes it safe to copy to the stack frame that a nested
1407	 * NMI will update.
1408	 *
1409	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1410	 * we're repeating an NMI, gsbase has the same value that it had on
1411	 * the first iteration.  paranoid_entry will load the kernel
1412	 * gsbase if needed before we call exc_nmi().  "NMI executing"
1413	 * is zero.
1414	 */
1415	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1416
1417	/*
1418	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1419	 * here must not modify the "iret" frame while we're writing to
1420	 * it or it will end up containing garbage.
1421	 */
1422	addq	$(10*8), %rsp
1423	.rept 5
1424	pushq	-6*8(%rsp)
1425	.endr
1426	subq	$(5*8), %rsp
1427end_repeat_nmi:
1428	ANNOTATE_NOENDBR // this code
1429
1430	/*
1431	 * Everything below this point can be preempted by a nested NMI.
1432	 * If this happens, then the inner NMI will change the "iret"
1433	 * frame to point back to repeat_nmi.
1434	 */
1435	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1436
1437	/*
1438	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1439	 * as we should not be calling schedule in NMI context.
1440	 * Even with normal interrupts enabled. An NMI should not be
1441	 * setting NEED_RESCHED or anything that normal interrupts and
1442	 * exceptions might do.
1443	 */
1444	call	paranoid_entry
1445	UNWIND_HINT_REGS
1446
1447	movq	%rsp, %rdi
1448	movq	$-1, %rsi
1449	call	exc_nmi
1450
1451	/* Always restore stashed SPEC_CTRL value (see paranoid_entry) */
1452	IBRS_EXIT save_reg=%r15
1453
1454	/* Always restore stashed CR3 value (see paranoid_entry) */
1455	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
1456
1457	/*
1458	 * The above invocation of paranoid_entry stored the GSBASE
1459	 * related information in R/EBX depending on the availability
1460	 * of FSGSBASE.
1461	 *
1462	 * If FSGSBASE is enabled, restore the saved GSBASE value
1463	 * unconditionally, otherwise take the conditional SWAPGS path.
1464	 */
1465	ALTERNATIVE "jmp nmi_no_fsgsbase", "", X86_FEATURE_FSGSBASE
1466
1467	wrgsbase	%rbx
1468	jmp	nmi_restore
1469
1470nmi_no_fsgsbase:
1471	/* EBX == 0 -> invoke SWAPGS */
1472	testl	%ebx, %ebx
1473	jnz	nmi_restore
1474
1475nmi_swapgs:
1476	swapgs
1477
1478nmi_restore:
1479	POP_REGS
1480
1481	/*
1482	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1483	 * at the "iret" frame.
1484	 */
1485	addq	$6*8, %rsp
1486
1487	/*
1488	 * Clear "NMI executing".  Set DF first so that we can easily
1489	 * distinguish the remaining code between here and IRET from
1490	 * the SYSCALL entry and exit paths.
1491	 *
1492	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1493	 * this code when I had the misapprehension that Xen PV supported
1494	 * NMIs, and Xen PV would break that approach.
1495	 */
1496	std
1497	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1498
1499	/*
1500	 * iretq reads the "iret" frame and exits the NMI stack in a
1501	 * single instruction.  We are returning to kernel mode, so this
1502	 * cannot result in a fault.  Similarly, we don't need to worry
1503	 * about espfix64 on the way back to kernel mode.
1504	 */
1505	iretq
1506SYM_CODE_END(asm_exc_nmi)
1507
1508#ifndef CONFIG_IA32_EMULATION
1509/*
1510 * This handles SYSCALL from 32-bit code.  There is no way to program
1511 * MSRs to fully disable 32-bit SYSCALL.
1512 */
1513SYM_CODE_START(ignore_sysret)
1514	UNWIND_HINT_EMPTY
1515	ENDBR
1516	mov	$-ENOSYS, %eax
1517	sysretl
1518SYM_CODE_END(ignore_sysret)
1519#endif
1520
1521.pushsection .text, "ax"
1522	__FUNC_ALIGN
1523SYM_CODE_START_NOALIGN(rewind_stack_and_make_dead)
1524	UNWIND_HINT_FUNC
1525	/* Prevent any naive code from trying to unwind to our caller. */
1526	xorl	%ebp, %ebp
1527
1528	movq	PER_CPU_VAR(pcpu_hot + X86_top_of_stack), %rax
1529	leaq	-PTREGS_SIZE(%rax), %rsp
1530	UNWIND_HINT_REGS
1531
1532	call	make_task_dead
1533SYM_CODE_END(rewind_stack_and_make_dead)
1534.popsection
1535