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(cpu_current_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"
287SYM_CODE_START(ret_from_fork)
288	UNWIND_HINT_EMPTY
289	ANNOTATE_NOENDBR // copy_thread
290	movq	%rax, %rdi
291	call	schedule_tail			/* rdi: 'prev' task parameter */
292
293	testq	%rbx, %rbx			/* from kernel_thread? */
294	jnz	1f				/* kernel threads are uncommon */
295
2962:
297	UNWIND_HINT_REGS
298	movq	%rsp, %rdi
299	call	syscall_exit_to_user_mode	/* returns with IRQs disabled */
300	jmp	swapgs_restore_regs_and_return_to_usermode
301
3021:
303	/* kernel thread */
304	UNWIND_HINT_EMPTY
305	movq	%r12, %rdi
306	CALL_NOSPEC rbx
307	/*
308	 * A kernel thread is allowed to return here after successfully
309	 * calling kernel_execve().  Exit to userspace to complete the execve()
310	 * syscall.
311	 */
312	movq	$0, RAX(%rsp)
313	jmp	2b
314SYM_CODE_END(ret_from_fork)
315.popsection
316
317.macro DEBUG_ENTRY_ASSERT_IRQS_OFF
318#ifdef CONFIG_DEBUG_ENTRY
319	pushq %rax
320	SAVE_FLAGS
321	testl $X86_EFLAGS_IF, %eax
322	jz .Lokay_\@
323	ud2
324.Lokay_\@:
325	popq %rax
326#endif
327.endm
328
329SYM_CODE_START_LOCAL(xen_error_entry)
330	UNWIND_HINT_FUNC
331	PUSH_AND_CLEAR_REGS save_ret=1
332	ENCODE_FRAME_POINTER 8
333	UNTRAIN_RET
334	RET
335SYM_CODE_END(xen_error_entry)
336
337/**
338 * idtentry_body - Macro to emit code calling the C function
339 * @cfunc:		C function to be called
340 * @has_error_code:	Hardware pushed error code on stack
341 */
342.macro idtentry_body cfunc has_error_code:req
343
344	/*
345	 * Call error_entry() and switch to the task stack if from userspace.
346	 *
347	 * When in XENPV, it is already in the task stack, and it can't fault
348	 * for native_iret() nor native_load_gs_index() since XENPV uses its
349	 * own pvops for IRET and load_gs_index().  And it doesn't need to
350	 * switch the CR3.  So it can skip invoking error_entry().
351	 */
352	ALTERNATIVE "call error_entry; movq %rax, %rsp", \
353		    "call xen_error_entry", X86_FEATURE_XENPV
354
355	ENCODE_FRAME_POINTER
356	UNWIND_HINT_REGS
357
358	movq	%rsp, %rdi			/* pt_regs pointer into 1st argument*/
359
360	.if \has_error_code == 1
361		movq	ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
362		movq	$-1, ORIG_RAX(%rsp)	/* no syscall to restart */
363	.endif
364
365	call	\cfunc
366
367	/* For some configurations \cfunc ends up being a noreturn. */
368	REACHABLE
369
370	jmp	error_return
371.endm
372
373/**
374 * idtentry - Macro to generate entry stubs for simple IDT entries
375 * @vector:		Vector number
376 * @asmsym:		ASM symbol for the entry point
377 * @cfunc:		C function to be called
378 * @has_error_code:	Hardware pushed error code on stack
379 *
380 * The macro emits code to set up the kernel context for straight forward
381 * and simple IDT entries. No IST stack, no paranoid entry checks.
382 */
383.macro idtentry vector asmsym cfunc has_error_code:req
384SYM_CODE_START(\asmsym)
385	UNWIND_HINT_IRET_REGS offset=\has_error_code*8
386	ENDBR
387	ASM_CLAC
388	cld
389
390	.if \has_error_code == 0
391		pushq	$-1			/* ORIG_RAX: no syscall to restart */
392	.endif
393
394	.if \vector == X86_TRAP_BP
395		/*
396		 * If coming from kernel space, create a 6-word gap to allow the
397		 * int3 handler to emulate a call instruction.
398		 */
399		testb	$3, CS-ORIG_RAX(%rsp)
400		jnz	.Lfrom_usermode_no_gap_\@
401		.rept	6
402		pushq	5*8(%rsp)
403		.endr
404		UNWIND_HINT_IRET_REGS offset=8
405.Lfrom_usermode_no_gap_\@:
406	.endif
407
408	idtentry_body \cfunc \has_error_code
409
410_ASM_NOKPROBE(\asmsym)
411SYM_CODE_END(\asmsym)
412.endm
413
414/*
415 * Interrupt entry/exit.
416 *
417 + The interrupt stubs push (vector) onto the stack, which is the error_code
418 * position of idtentry exceptions, and jump to one of the two idtentry points
419 * (common/spurious).
420 *
421 * common_interrupt is a hotpath, align it to a cache line
422 */
423.macro idtentry_irq vector cfunc
424	.p2align CONFIG_X86_L1_CACHE_SHIFT
425	idtentry \vector asm_\cfunc \cfunc has_error_code=1
426.endm
427
428/*
429 * System vectors which invoke their handlers directly and are not
430 * going through the regular common device interrupt handling code.
431 */
432.macro idtentry_sysvec vector cfunc
433	idtentry \vector asm_\cfunc \cfunc has_error_code=0
434.endm
435
436/**
437 * idtentry_mce_db - Macro to generate entry stubs for #MC and #DB
438 * @vector:		Vector number
439 * @asmsym:		ASM symbol for the entry point
440 * @cfunc:		C function to be called
441 *
442 * The macro emits code to set up the kernel context for #MC and #DB
443 *
444 * If the entry comes from user space it uses the normal entry path
445 * including the return to user space work and preemption checks on
446 * exit.
447 *
448 * If hits in kernel mode then it needs to go through the paranoid
449 * entry as the exception can hit any random state. No preemption
450 * check on exit to keep the paranoid path simple.
451 */
452.macro idtentry_mce_db vector asmsym cfunc
453SYM_CODE_START(\asmsym)
454	UNWIND_HINT_IRET_REGS
455	ENDBR
456	ASM_CLAC
457	cld
458
459	pushq	$-1			/* ORIG_RAX: no syscall to restart */
460
461	/*
462	 * If the entry is from userspace, switch stacks and treat it as
463	 * a normal entry.
464	 */
465	testb	$3, CS-ORIG_RAX(%rsp)
466	jnz	.Lfrom_usermode_switch_stack_\@
467
468	/* paranoid_entry returns GS information for paranoid_exit in EBX. */
469	call	paranoid_entry
470
471	UNWIND_HINT_REGS
472
473	movq	%rsp, %rdi		/* pt_regs pointer */
474
475	call	\cfunc
476
477	jmp	paranoid_exit
478
479	/* Switch to the regular task stack and use the noist entry point */
480.Lfrom_usermode_switch_stack_\@:
481	idtentry_body noist_\cfunc, has_error_code=0
482
483_ASM_NOKPROBE(\asmsym)
484SYM_CODE_END(\asmsym)
485.endm
486
487#ifdef CONFIG_AMD_MEM_ENCRYPT
488/**
489 * idtentry_vc - Macro to generate entry stub for #VC
490 * @vector:		Vector number
491 * @asmsym:		ASM symbol for the entry point
492 * @cfunc:		C function to be called
493 *
494 * The macro emits code to set up the kernel context for #VC. The #VC handler
495 * runs on an IST stack and needs to be able to cause nested #VC exceptions.
496 *
497 * To make this work the #VC entry code tries its best to pretend it doesn't use
498 * an IST stack by switching to the task stack if coming from user-space (which
499 * includes early SYSCALL entry path) or back to the stack in the IRET frame if
500 * entered from kernel-mode.
501 *
502 * If entered from kernel-mode the return stack is validated first, and if it is
503 * not safe to use (e.g. because it points to the entry stack) the #VC handler
504 * will switch to a fall-back stack (VC2) and call a special handler function.
505 *
506 * The macro is only used for one vector, but it is planned to be extended in
507 * the future for the #HV exception.
508 */
509.macro idtentry_vc vector asmsym cfunc
510SYM_CODE_START(\asmsym)
511	UNWIND_HINT_IRET_REGS
512	ENDBR
513	ASM_CLAC
514	cld
515
516	/*
517	 * If the entry is from userspace, switch stacks and treat it as
518	 * a normal entry.
519	 */
520	testb	$3, CS-ORIG_RAX(%rsp)
521	jnz	.Lfrom_usermode_switch_stack_\@
522
523	/*
524	 * paranoid_entry returns SWAPGS flag for paranoid_exit in EBX.
525	 * EBX == 0 -> SWAPGS, EBX == 1 -> no SWAPGS
526	 */
527	call	paranoid_entry
528
529	UNWIND_HINT_REGS
530
531	/*
532	 * Switch off the IST stack to make it free for nested exceptions. The
533	 * vc_switch_off_ist() function will switch back to the interrupted
534	 * stack if it is safe to do so. If not it switches to the VC fall-back
535	 * stack.
536	 */
537	movq	%rsp, %rdi		/* pt_regs pointer */
538	call	vc_switch_off_ist
539	movq	%rax, %rsp		/* Switch to new stack */
540
541	ENCODE_FRAME_POINTER
542	UNWIND_HINT_REGS
543
544	/* Update pt_regs */
545	movq	ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
546	movq	$-1, ORIG_RAX(%rsp)	/* no syscall to restart */
547
548	movq	%rsp, %rdi		/* pt_regs pointer */
549
550	call	kernel_\cfunc
551
552	/*
553	 * No need to switch back to the IST stack. The current stack is either
554	 * identical to the stack in the IRET frame or the VC fall-back stack,
555	 * so it is definitely mapped even with PTI enabled.
556	 */
557	jmp	paranoid_exit
558
559	/* Switch to the regular task stack */
560.Lfrom_usermode_switch_stack_\@:
561	idtentry_body user_\cfunc, has_error_code=1
562
563_ASM_NOKPROBE(\asmsym)
564SYM_CODE_END(\asmsym)
565.endm
566#endif
567
568/*
569 * Double fault entry. Straight paranoid. No checks from which context
570 * this comes because for the espfix induced #DF this would do the wrong
571 * thing.
572 */
573.macro idtentry_df vector asmsym cfunc
574SYM_CODE_START(\asmsym)
575	UNWIND_HINT_IRET_REGS offset=8
576	ENDBR
577	ASM_CLAC
578	cld
579
580	/* paranoid_entry returns GS information for paranoid_exit in EBX. */
581	call	paranoid_entry
582	UNWIND_HINT_REGS
583
584	movq	%rsp, %rdi		/* pt_regs pointer into first argument */
585	movq	ORIG_RAX(%rsp), %rsi	/* get error code into 2nd argument*/
586	movq	$-1, ORIG_RAX(%rsp)	/* no syscall to restart */
587	call	\cfunc
588
589	/* For some configurations \cfunc ends up being a noreturn. */
590	REACHABLE
591
592	jmp	paranoid_exit
593
594_ASM_NOKPROBE(\asmsym)
595SYM_CODE_END(\asmsym)
596.endm
597
598/*
599 * Include the defines which emit the idt entries which are shared
600 * shared between 32 and 64 bit and emit the __irqentry_text_* markers
601 * so the stacktrace boundary checks work.
602 */
603	.align 16
604	.globl __irqentry_text_start
605__irqentry_text_start:
606
607#include <asm/idtentry.h>
608
609	.align 16
610	.globl __irqentry_text_end
611__irqentry_text_end:
612	ANNOTATE_NOENDBR
613
614SYM_CODE_START_LOCAL(common_interrupt_return)
615SYM_INNER_LABEL(swapgs_restore_regs_and_return_to_usermode, SYM_L_GLOBAL)
616	IBRS_EXIT
617#ifdef CONFIG_DEBUG_ENTRY
618	/* Assert that pt_regs indicates user mode. */
619	testb	$3, CS(%rsp)
620	jnz	1f
621	ud2
6221:
623#endif
624#ifdef CONFIG_XEN_PV
625	ALTERNATIVE "", "jmp xenpv_restore_regs_and_return_to_usermode", X86_FEATURE_XENPV
626#endif
627
628	POP_REGS pop_rdi=0
629
630	/*
631	 * The stack is now user RDI, orig_ax, RIP, CS, EFLAGS, RSP, SS.
632	 * Save old stack pointer and switch to trampoline stack.
633	 */
634	movq	%rsp, %rdi
635	movq	PER_CPU_VAR(cpu_tss_rw + TSS_sp0), %rsp
636	UNWIND_HINT_EMPTY
637
638	/* Copy the IRET frame to the trampoline stack. */
639	pushq	6*8(%rdi)	/* SS */
640	pushq	5*8(%rdi)	/* RSP */
641	pushq	4*8(%rdi)	/* EFLAGS */
642	pushq	3*8(%rdi)	/* CS */
643	pushq	2*8(%rdi)	/* RIP */
644
645	/* Push user RDI on the trampoline stack. */
646	pushq	(%rdi)
647
648	/*
649	 * We are on the trampoline stack.  All regs except RDI are live.
650	 * We can do future final exit work right here.
651	 */
652	STACKLEAK_ERASE_NOCLOBBER
653
654	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
655
656	/* Restore RDI. */
657	popq	%rdi
658	swapgs
659	jmp	.Lnative_iret
660
661
662SYM_INNER_LABEL(restore_regs_and_return_to_kernel, SYM_L_GLOBAL)
663#ifdef CONFIG_DEBUG_ENTRY
664	/* Assert that pt_regs indicates kernel mode. */
665	testb	$3, CS(%rsp)
666	jz	1f
667	ud2
6681:
669#endif
670	POP_REGS
671	addq	$8, %rsp	/* skip regs->orig_ax */
672	/*
673	 * ARCH_HAS_MEMBARRIER_SYNC_CORE rely on IRET core serialization
674	 * when returning from IPI handler.
675	 */
676#ifdef CONFIG_XEN_PV
677SYM_INNER_LABEL(early_xen_iret_patch, SYM_L_GLOBAL)
678	ANNOTATE_NOENDBR
679	.byte 0xe9
680	.long .Lnative_iret - (. + 4)
681#endif
682
683.Lnative_iret:
684	UNWIND_HINT_IRET_REGS
685	/*
686	 * Are we returning to a stack segment from the LDT?  Note: in
687	 * 64-bit mode SS:RSP on the exception stack is always valid.
688	 */
689#ifdef CONFIG_X86_ESPFIX64
690	testb	$4, (SS-RIP)(%rsp)
691	jnz	native_irq_return_ldt
692#endif
693
694SYM_INNER_LABEL(native_irq_return_iret, SYM_L_GLOBAL)
695	ANNOTATE_NOENDBR // exc_double_fault
696	/*
697	 * This may fault.  Non-paranoid faults on return to userspace are
698	 * handled by fixup_bad_iret.  These include #SS, #GP, and #NP.
699	 * Double-faults due to espfix64 are handled in exc_double_fault.
700	 * Other faults here are fatal.
701	 */
702	iretq
703
704#ifdef CONFIG_X86_ESPFIX64
705native_irq_return_ldt:
706	/*
707	 * We are running with user GSBASE.  All GPRs contain their user
708	 * values.  We have a percpu ESPFIX stack that is eight slots
709	 * long (see ESPFIX_STACK_SIZE).  espfix_waddr points to the bottom
710	 * of the ESPFIX stack.
711	 *
712	 * We clobber RAX and RDI in this code.  We stash RDI on the
713	 * normal stack and RAX on the ESPFIX stack.
714	 *
715	 * The ESPFIX stack layout we set up looks like this:
716	 *
717	 * --- top of ESPFIX stack ---
718	 * SS
719	 * RSP
720	 * RFLAGS
721	 * CS
722	 * RIP  <-- RSP points here when we're done
723	 * RAX  <-- espfix_waddr points here
724	 * --- bottom of ESPFIX stack ---
725	 */
726
727	pushq	%rdi				/* Stash user RDI */
728	swapgs					/* to kernel GS */
729	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdi	/* to kernel CR3 */
730
731	movq	PER_CPU_VAR(espfix_waddr), %rdi
732	movq	%rax, (0*8)(%rdi)		/* user RAX */
733	movq	(1*8)(%rsp), %rax		/* user RIP */
734	movq	%rax, (1*8)(%rdi)
735	movq	(2*8)(%rsp), %rax		/* user CS */
736	movq	%rax, (2*8)(%rdi)
737	movq	(3*8)(%rsp), %rax		/* user RFLAGS */
738	movq	%rax, (3*8)(%rdi)
739	movq	(5*8)(%rsp), %rax		/* user SS */
740	movq	%rax, (5*8)(%rdi)
741	movq	(4*8)(%rsp), %rax		/* user RSP */
742	movq	%rax, (4*8)(%rdi)
743	/* Now RAX == RSP. */
744
745	andl	$0xffff0000, %eax		/* RAX = (RSP & 0xffff0000) */
746
747	/*
748	 * espfix_stack[31:16] == 0.  The page tables are set up such that
749	 * (espfix_stack | (X & 0xffff0000)) points to a read-only alias of
750	 * espfix_waddr for any X.  That is, there are 65536 RO aliases of
751	 * the same page.  Set up RSP so that RSP[31:16] contains the
752	 * respective 16 bits of the /userspace/ RSP and RSP nonetheless
753	 * still points to an RO alias of the ESPFIX stack.
754	 */
755	orq	PER_CPU_VAR(espfix_stack), %rax
756
757	SWITCH_TO_USER_CR3_STACK scratch_reg=%rdi
758	swapgs					/* to user GS */
759	popq	%rdi				/* Restore user RDI */
760
761	movq	%rax, %rsp
762	UNWIND_HINT_IRET_REGS offset=8
763
764	/*
765	 * At this point, we cannot write to the stack any more, but we can
766	 * still read.
767	 */
768	popq	%rax				/* Restore user RAX */
769
770	/*
771	 * RSP now points to an ordinary IRET frame, except that the page
772	 * is read-only and RSP[31:16] are preloaded with the userspace
773	 * values.  We can now IRET back to userspace.
774	 */
775	jmp	native_irq_return_iret
776#endif
777SYM_CODE_END(common_interrupt_return)
778_ASM_NOKPROBE(common_interrupt_return)
779
780/*
781 * Reload gs selector with exception handling
782 * edi:  new selector
783 *
784 * Is in entry.text as it shouldn't be instrumented.
785 */
786SYM_FUNC_START(asm_load_gs_index)
787	FRAME_BEGIN
788	swapgs
789.Lgs_change:
790	ANNOTATE_NOENDBR // error_entry
791	movl	%edi, %gs
7922:	ALTERNATIVE "", "mfence", X86_BUG_SWAPGS_FENCE
793	swapgs
794	FRAME_END
795	RET
796
797	/* running with kernelgs */
798.Lbad_gs:
799	swapgs					/* switch back to user gs */
800.macro ZAP_GS
801	/* This can't be a string because the preprocessor needs to see it. */
802	movl $__USER_DS, %eax
803	movl %eax, %gs
804.endm
805	ALTERNATIVE "", "ZAP_GS", X86_BUG_NULL_SEG
806	xorl	%eax, %eax
807	movl	%eax, %gs
808	jmp	2b
809
810	_ASM_EXTABLE(.Lgs_change, .Lbad_gs)
811
812SYM_FUNC_END(asm_load_gs_index)
813EXPORT_SYMBOL(asm_load_gs_index)
814
815#ifdef CONFIG_XEN_PV
816/*
817 * A note on the "critical region" in our callback handler.
818 * We want to avoid stacking callback handlers due to events occurring
819 * during handling of the last event. To do this, we keep events disabled
820 * until we've done all processing. HOWEVER, we must enable events before
821 * popping the stack frame (can't be done atomically) and so it would still
822 * be possible to get enough handler activations to overflow the stack.
823 * Although unlikely, bugs of that kind are hard to track down, so we'd
824 * like to avoid the possibility.
825 * So, on entry to the handler we detect whether we interrupted an
826 * existing activation in its critical region -- if so, we pop the current
827 * activation and restart the handler using the previous one.
828 *
829 * C calling convention: exc_xen_hypervisor_callback(struct *pt_regs)
830 */
831SYM_CODE_START_LOCAL(exc_xen_hypervisor_callback)
832
833/*
834 * Since we don't modify %rdi, evtchn_do_upall(struct *pt_regs) will
835 * see the correct pointer to the pt_regs
836 */
837	UNWIND_HINT_FUNC
838	movq	%rdi, %rsp			/* we don't return, adjust the stack frame */
839	UNWIND_HINT_REGS
840
841	call	xen_pv_evtchn_do_upcall
842
843	jmp	error_return
844SYM_CODE_END(exc_xen_hypervisor_callback)
845
846/*
847 * Hypervisor uses this for application faults while it executes.
848 * We get here for two reasons:
849 *  1. Fault while reloading DS, ES, FS or GS
850 *  2. Fault while executing IRET
851 * Category 1 we do not need to fix up as Xen has already reloaded all segment
852 * registers that could be reloaded and zeroed the others.
853 * Category 2 we fix up by killing the current process. We cannot use the
854 * normal Linux return path in this case because if we use the IRET hypercall
855 * to pop the stack frame we end up in an infinite loop of failsafe callbacks.
856 * We distinguish between categories by comparing each saved segment register
857 * with its current contents: any discrepancy means we in category 1.
858 */
859SYM_CODE_START(xen_failsafe_callback)
860	UNWIND_HINT_EMPTY
861	ENDBR
862	movl	%ds, %ecx
863	cmpw	%cx, 0x10(%rsp)
864	jne	1f
865	movl	%es, %ecx
866	cmpw	%cx, 0x18(%rsp)
867	jne	1f
868	movl	%fs, %ecx
869	cmpw	%cx, 0x20(%rsp)
870	jne	1f
871	movl	%gs, %ecx
872	cmpw	%cx, 0x28(%rsp)
873	jne	1f
874	/* All segments match their saved values => Category 2 (Bad IRET). */
875	movq	(%rsp), %rcx
876	movq	8(%rsp), %r11
877	addq	$0x30, %rsp
878	pushq	$0				/* RIP */
879	UNWIND_HINT_IRET_REGS offset=8
880	jmp	asm_exc_general_protection
8811:	/* Segment mismatch => Category 1 (Bad segment). Retry the IRET. */
882	movq	(%rsp), %rcx
883	movq	8(%rsp), %r11
884	addq	$0x30, %rsp
885	UNWIND_HINT_IRET_REGS
886	pushq	$-1 /* orig_ax = -1 => not a system call */
887	PUSH_AND_CLEAR_REGS
888	ENCODE_FRAME_POINTER
889	jmp	error_return
890SYM_CODE_END(xen_failsafe_callback)
891#endif /* CONFIG_XEN_PV */
892
893/*
894 * Save all registers in pt_regs. Return GSBASE related information
895 * in EBX depending on the availability of the FSGSBASE instructions:
896 *
897 * FSGSBASE	R/EBX
898 *     N        0 -> SWAPGS on exit
899 *              1 -> no SWAPGS on exit
900 *
901 *     Y        GSBASE value at entry, must be restored in paranoid_exit
902 *
903 * R14 - old CR3
904 * R15 - old SPEC_CTRL
905 */
906SYM_CODE_START_LOCAL(paranoid_entry)
907	UNWIND_HINT_FUNC
908	PUSH_AND_CLEAR_REGS save_ret=1
909	ENCODE_FRAME_POINTER 8
910
911	/*
912	 * Always stash CR3 in %r14.  This value will be restored,
913	 * verbatim, at exit.  Needed if paranoid_entry interrupted
914	 * another entry that already switched to the user CR3 value
915	 * but has not yet returned to userspace.
916	 *
917	 * This is also why CS (stashed in the "iret frame" by the
918	 * hardware at entry) can not be used: this may be a return
919	 * to kernel code, but with a user CR3 value.
920	 *
921	 * Switching CR3 does not depend on kernel GSBASE so it can
922	 * be done before switching to the kernel GSBASE. This is
923	 * required for FSGSBASE because the kernel GSBASE has to
924	 * be retrieved from a kernel internal table.
925	 */
926	SAVE_AND_SWITCH_TO_KERNEL_CR3 scratch_reg=%rax save_reg=%r14
927
928	/*
929	 * Handling GSBASE depends on the availability of FSGSBASE.
930	 *
931	 * Without FSGSBASE the kernel enforces that negative GSBASE
932	 * values indicate kernel GSBASE. With FSGSBASE no assumptions
933	 * can be made about the GSBASE value when entering from user
934	 * space.
935	 */
936	ALTERNATIVE "jmp .Lparanoid_entry_checkgs", "", X86_FEATURE_FSGSBASE
937
938	/*
939	 * Read the current GSBASE and store it in %rbx unconditionally,
940	 * retrieve and set the current CPUs kernel GSBASE. The stored value
941	 * has to be restored in paranoid_exit unconditionally.
942	 *
943	 * The unconditional write to GS base below ensures that no subsequent
944	 * loads based on a mispredicted GS base can happen, therefore no LFENCE
945	 * is needed here.
946	 */
947	SAVE_AND_SET_GSBASE scratch_reg=%rax save_reg=%rbx
948	jmp .Lparanoid_gsbase_done
949
950.Lparanoid_entry_checkgs:
951	/* EBX = 1 -> kernel GSBASE active, no restore required */
952	movl	$1, %ebx
953
954	/*
955	 * The kernel-enforced convention is a negative GSBASE indicates
956	 * a kernel value. No SWAPGS needed on entry and exit.
957	 */
958	movl	$MSR_GS_BASE, %ecx
959	rdmsr
960	testl	%edx, %edx
961	js	.Lparanoid_kernel_gsbase
962
963	/* EBX = 0 -> SWAPGS required on exit */
964	xorl	%ebx, %ebx
965	swapgs
966.Lparanoid_kernel_gsbase:
967	FENCE_SWAPGS_KERNEL_ENTRY
968.Lparanoid_gsbase_done:
969
970	/*
971	 * Once we have CR3 and %GS setup save and set SPEC_CTRL. Just like
972	 * CR3 above, keep the old value in a callee saved register.
973	 */
974	IBRS_ENTER save_reg=%r15
975	UNTRAIN_RET
976
977	RET
978SYM_CODE_END(paranoid_entry)
979
980/*
981 * "Paranoid" exit path from exception stack.  This is invoked
982 * only on return from non-NMI IST interrupts that came
983 * from kernel space.
984 *
985 * We may be returning to very strange contexts (e.g. very early
986 * in syscall entry), so checking for preemption here would
987 * be complicated.  Fortunately, there's no good reason to try
988 * to handle preemption here.
989 *
990 * R/EBX contains the GSBASE related information depending on the
991 * availability of the FSGSBASE instructions:
992 *
993 * FSGSBASE	R/EBX
994 *     N        0 -> SWAPGS on exit
995 *              1 -> no SWAPGS on exit
996 *
997 *     Y        User space GSBASE, must be restored unconditionally
998 *
999 * R14 - old CR3
1000 * R15 - old SPEC_CTRL
1001 */
1002SYM_CODE_START_LOCAL(paranoid_exit)
1003	UNWIND_HINT_REGS
1004
1005	/*
1006	 * Must restore IBRS state before both CR3 and %GS since we need access
1007	 * to the per-CPU x86_spec_ctrl_shadow variable.
1008	 */
1009	IBRS_EXIT save_reg=%r15
1010
1011	/*
1012	 * The order of operations is important. RESTORE_CR3 requires
1013	 * kernel GSBASE.
1014	 *
1015	 * NB to anyone to try to optimize this code: this code does
1016	 * not execute at all for exceptions from user mode. Those
1017	 * exceptions go through error_exit instead.
1018	 */
1019	RESTORE_CR3	scratch_reg=%rax save_reg=%r14
1020
1021	/* Handle the three GSBASE cases */
1022	ALTERNATIVE "jmp .Lparanoid_exit_checkgs", "", X86_FEATURE_FSGSBASE
1023
1024	/* With FSGSBASE enabled, unconditionally restore GSBASE */
1025	wrgsbase	%rbx
1026	jmp		restore_regs_and_return_to_kernel
1027
1028.Lparanoid_exit_checkgs:
1029	/* On non-FSGSBASE systems, conditionally do SWAPGS */
1030	testl		%ebx, %ebx
1031	jnz		restore_regs_and_return_to_kernel
1032
1033	/* We are returning to a context with user GSBASE */
1034	swapgs
1035	jmp		restore_regs_and_return_to_kernel
1036SYM_CODE_END(paranoid_exit)
1037
1038/*
1039 * Switch GS and CR3 if needed.
1040 */
1041SYM_CODE_START_LOCAL(error_entry)
1042	UNWIND_HINT_FUNC
1043
1044	PUSH_AND_CLEAR_REGS save_ret=1
1045	ENCODE_FRAME_POINTER 8
1046
1047	testb	$3, CS+8(%rsp)
1048	jz	.Lerror_kernelspace
1049
1050	/*
1051	 * We entered from user mode or we're pretending to have entered
1052	 * from user mode due to an IRET fault.
1053	 */
1054	swapgs
1055	FENCE_SWAPGS_USER_ENTRY
1056	/* We have user CR3.  Change to kernel CR3. */
1057	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1058	IBRS_ENTER
1059	UNTRAIN_RET
1060
1061	leaq	8(%rsp), %rdi			/* arg0 = pt_regs pointer */
1062.Lerror_entry_from_usermode_after_swapgs:
1063
1064	/* Put us onto the real thread stack. */
1065	call	sync_regs
1066	RET
1067
1068	/*
1069	 * There are two places in the kernel that can potentially fault with
1070	 * usergs. Handle them here.  B stepping K8s sometimes report a
1071	 * truncated RIP for IRET exceptions returning to compat mode. Check
1072	 * for these here too.
1073	 */
1074.Lerror_kernelspace:
1075	leaq	native_irq_return_iret(%rip), %rcx
1076	cmpq	%rcx, RIP+8(%rsp)
1077	je	.Lerror_bad_iret
1078	movl	%ecx, %eax			/* zero extend */
1079	cmpq	%rax, RIP+8(%rsp)
1080	je	.Lbstep_iret
1081	cmpq	$.Lgs_change, RIP+8(%rsp)
1082	jne	.Lerror_entry_done_lfence
1083
1084	/*
1085	 * hack: .Lgs_change can fail with user gsbase.  If this happens, fix up
1086	 * gsbase and proceed.  We'll fix up the exception and land in
1087	 * .Lgs_change's error handler with kernel gsbase.
1088	 */
1089	swapgs
1090
1091	/*
1092	 * Issue an LFENCE to prevent GS speculation, regardless of whether it is a
1093	 * kernel or user gsbase.
1094	 */
1095.Lerror_entry_done_lfence:
1096	FENCE_SWAPGS_KERNEL_ENTRY
1097	leaq	8(%rsp), %rax			/* return pt_regs pointer */
1098	ANNOTATE_UNRET_END
1099	RET
1100
1101.Lbstep_iret:
1102	/* Fix truncated RIP */
1103	movq	%rcx, RIP+8(%rsp)
1104	/* fall through */
1105
1106.Lerror_bad_iret:
1107	/*
1108	 * We came from an IRET to user mode, so we have user
1109	 * gsbase and CR3.  Switch to kernel gsbase and CR3:
1110	 */
1111	swapgs
1112	FENCE_SWAPGS_USER_ENTRY
1113	SWITCH_TO_KERNEL_CR3 scratch_reg=%rax
1114	IBRS_ENTER
1115	UNTRAIN_RET
1116
1117	/*
1118	 * Pretend that the exception came from user mode: set up pt_regs
1119	 * as if we faulted immediately after IRET.
1120	 */
1121	leaq	8(%rsp), %rdi			/* arg0 = pt_regs pointer */
1122	call	fixup_bad_iret
1123	mov	%rax, %rdi
1124	jmp	.Lerror_entry_from_usermode_after_swapgs
1125SYM_CODE_END(error_entry)
1126
1127SYM_CODE_START_LOCAL(error_return)
1128	UNWIND_HINT_REGS
1129	DEBUG_ENTRY_ASSERT_IRQS_OFF
1130	testb	$3, CS(%rsp)
1131	jz	restore_regs_and_return_to_kernel
1132	jmp	swapgs_restore_regs_and_return_to_usermode
1133SYM_CODE_END(error_return)
1134
1135/*
1136 * Runs on exception stack.  Xen PV does not go through this path at all,
1137 * so we can use real assembly here.
1138 *
1139 * Registers:
1140 *	%r14: Used to save/restore the CR3 of the interrupted context
1141 *	      when PAGE_TABLE_ISOLATION is in use.  Do not clobber.
1142 */
1143SYM_CODE_START(asm_exc_nmi)
1144	UNWIND_HINT_IRET_REGS
1145	ENDBR
1146
1147	/*
1148	 * We allow breakpoints in NMIs. If a breakpoint occurs, then
1149	 * the iretq it performs will take us out of NMI context.
1150	 * This means that we can have nested NMIs where the next
1151	 * NMI is using the top of the stack of the previous NMI. We
1152	 * can't let it execute because the nested NMI will corrupt the
1153	 * stack of the previous NMI. NMI handlers are not re-entrant
1154	 * anyway.
1155	 *
1156	 * To handle this case we do the following:
1157	 *  Check the a special location on the stack that contains
1158	 *  a variable that is set when NMIs are executing.
1159	 *  The interrupted task's stack is also checked to see if it
1160	 *  is an NMI stack.
1161	 *  If the variable is not set and the stack is not the NMI
1162	 *  stack then:
1163	 *    o Set the special variable on the stack
1164	 *    o Copy the interrupt frame into an "outermost" location on the
1165	 *      stack
1166	 *    o Copy the interrupt frame into an "iret" location on the stack
1167	 *    o Continue processing the NMI
1168	 *  If the variable is set or the previous stack is the NMI stack:
1169	 *    o Modify the "iret" location to jump to the repeat_nmi
1170	 *    o return back to the first NMI
1171	 *
1172	 * Now on exit of the first NMI, we first clear the stack variable
1173	 * The NMI stack will tell any nested NMIs at that point that it is
1174	 * nested. Then we pop the stack normally with iret, and if there was
1175	 * a nested NMI that updated the copy interrupt stack frame, a
1176	 * jump will be made to the repeat_nmi code that will handle the second
1177	 * NMI.
1178	 *
1179	 * However, espfix prevents us from directly returning to userspace
1180	 * with a single IRET instruction.  Similarly, IRET to user mode
1181	 * can fault.  We therefore handle NMIs from user space like
1182	 * other IST entries.
1183	 */
1184
1185	ASM_CLAC
1186	cld
1187
1188	/* Use %rdx as our temp variable throughout */
1189	pushq	%rdx
1190
1191	testb	$3, CS-RIP+8(%rsp)
1192	jz	.Lnmi_from_kernel
1193
1194	/*
1195	 * NMI from user mode.  We need to run on the thread stack, but we
1196	 * can't go through the normal entry paths: NMIs are masked, and
1197	 * we don't want to enable interrupts, because then we'll end
1198	 * up in an awkward situation in which IRQs are on but NMIs
1199	 * are off.
1200	 *
1201	 * We also must not push anything to the stack before switching
1202	 * stacks lest we corrupt the "NMI executing" variable.
1203	 */
1204
1205	swapgs
1206	FENCE_SWAPGS_USER_ENTRY
1207	SWITCH_TO_KERNEL_CR3 scratch_reg=%rdx
1208	movq	%rsp, %rdx
1209	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rsp
1210	UNWIND_HINT_IRET_REGS base=%rdx offset=8
1211	pushq	5*8(%rdx)	/* pt_regs->ss */
1212	pushq	4*8(%rdx)	/* pt_regs->rsp */
1213	pushq	3*8(%rdx)	/* pt_regs->flags */
1214	pushq	2*8(%rdx)	/* pt_regs->cs */
1215	pushq	1*8(%rdx)	/* pt_regs->rip */
1216	UNWIND_HINT_IRET_REGS
1217	pushq   $-1		/* pt_regs->orig_ax */
1218	PUSH_AND_CLEAR_REGS rdx=(%rdx)
1219	ENCODE_FRAME_POINTER
1220
1221	IBRS_ENTER
1222	UNTRAIN_RET
1223
1224	/*
1225	 * At this point we no longer need to worry about stack damage
1226	 * due to nesting -- we're on the normal thread stack and we're
1227	 * done with the NMI stack.
1228	 */
1229
1230	movq	%rsp, %rdi
1231	movq	$-1, %rsi
1232	call	exc_nmi
1233
1234	/*
1235	 * Return back to user mode.  We must *not* do the normal exit
1236	 * work, because we don't want to enable interrupts.
1237	 */
1238	jmp	swapgs_restore_regs_and_return_to_usermode
1239
1240.Lnmi_from_kernel:
1241	/*
1242	 * Here's what our stack frame will look like:
1243	 * +---------------------------------------------------------+
1244	 * | original SS                                             |
1245	 * | original Return RSP                                     |
1246	 * | original RFLAGS                                         |
1247	 * | original CS                                             |
1248	 * | original RIP                                            |
1249	 * +---------------------------------------------------------+
1250	 * | temp storage for rdx                                    |
1251	 * +---------------------------------------------------------+
1252	 * | "NMI executing" variable                                |
1253	 * +---------------------------------------------------------+
1254	 * | iret SS          } Copied from "outermost" frame        |
1255	 * | iret Return RSP  } on each loop iteration; overwritten  |
1256	 * | iret RFLAGS      } by a nested NMI to force another     |
1257	 * | iret CS          } iteration if needed.                 |
1258	 * | iret RIP         }                                      |
1259	 * +---------------------------------------------------------+
1260	 * | outermost SS          } initialized in first_nmi;       |
1261	 * | outermost Return RSP  } will not be changed before      |
1262	 * | outermost RFLAGS      } NMI processing is done.         |
1263	 * | outermost CS          } Copied to "iret" frame on each  |
1264	 * | outermost RIP         } iteration.                      |
1265	 * +---------------------------------------------------------+
1266	 * | pt_regs                                                 |
1267	 * +---------------------------------------------------------+
1268	 *
1269	 * The "original" frame is used by hardware.  Before re-enabling
1270	 * NMIs, we need to be done with it, and we need to leave enough
1271	 * space for the asm code here.
1272	 *
1273	 * We return by executing IRET while RSP points to the "iret" frame.
1274	 * That will either return for real or it will loop back into NMI
1275	 * processing.
1276	 *
1277	 * The "outermost" frame is copied to the "iret" frame on each
1278	 * iteration of the loop, so each iteration starts with the "iret"
1279	 * frame pointing to the final return target.
1280	 */
1281
1282	/*
1283	 * Determine whether we're a nested NMI.
1284	 *
1285	 * If we interrupted kernel code between repeat_nmi and
1286	 * end_repeat_nmi, then we are a nested NMI.  We must not
1287	 * modify the "iret" frame because it's being written by
1288	 * the outer NMI.  That's okay; the outer NMI handler is
1289	 * about to about to call exc_nmi() anyway, so we can just
1290	 * resume the outer NMI.
1291	 */
1292
1293	movq	$repeat_nmi, %rdx
1294	cmpq	8(%rsp), %rdx
1295	ja	1f
1296	movq	$end_repeat_nmi, %rdx
1297	cmpq	8(%rsp), %rdx
1298	ja	nested_nmi_out
12991:
1300
1301	/*
1302	 * Now check "NMI executing".  If it's set, then we're nested.
1303	 * This will not detect if we interrupted an outer NMI just
1304	 * before IRET.
1305	 */
1306	cmpl	$1, -8(%rsp)
1307	je	nested_nmi
1308
1309	/*
1310	 * Now test if the previous stack was an NMI stack.  This covers
1311	 * the case where we interrupt an outer NMI after it clears
1312	 * "NMI executing" but before IRET.  We need to be careful, though:
1313	 * there is one case in which RSP could point to the NMI stack
1314	 * despite there being no NMI active: naughty userspace controls
1315	 * RSP at the very beginning of the SYSCALL targets.  We can
1316	 * pull a fast one on naughty userspace, though: we program
1317	 * SYSCALL to mask DF, so userspace cannot cause DF to be set
1318	 * if it controls the kernel's RSP.  We set DF before we clear
1319	 * "NMI executing".
1320	 */
1321	lea	6*8(%rsp), %rdx
1322	/* Compare the NMI stack (rdx) with the stack we came from (4*8(%rsp)) */
1323	cmpq	%rdx, 4*8(%rsp)
1324	/* If the stack pointer is above the NMI stack, this is a normal NMI */
1325	ja	first_nmi
1326
1327	subq	$EXCEPTION_STKSZ, %rdx
1328	cmpq	%rdx, 4*8(%rsp)
1329	/* If it is below the NMI stack, it is a normal NMI */
1330	jb	first_nmi
1331
1332	/* Ah, it is within the NMI stack. */
1333
1334	testb	$(X86_EFLAGS_DF >> 8), (3*8 + 1)(%rsp)
1335	jz	first_nmi	/* RSP was user controlled. */
1336
1337	/* This is a nested NMI. */
1338
1339nested_nmi:
1340	/*
1341	 * Modify the "iret" frame to point to repeat_nmi, forcing another
1342	 * iteration of NMI handling.
1343	 */
1344	subq	$8, %rsp
1345	leaq	-10*8(%rsp), %rdx
1346	pushq	$__KERNEL_DS
1347	pushq	%rdx
1348	pushfq
1349	pushq	$__KERNEL_CS
1350	pushq	$repeat_nmi
1351
1352	/* Put stack back */
1353	addq	$(6*8), %rsp
1354
1355nested_nmi_out:
1356	popq	%rdx
1357
1358	/* We are returning to kernel mode, so this cannot result in a fault. */
1359	iretq
1360
1361first_nmi:
1362	/* Restore rdx. */
1363	movq	(%rsp), %rdx
1364
1365	/* Make room for "NMI executing". */
1366	pushq	$0
1367
1368	/* Leave room for the "iret" frame */
1369	subq	$(5*8), %rsp
1370
1371	/* Copy the "original" frame to the "outermost" frame */
1372	.rept 5
1373	pushq	11*8(%rsp)
1374	.endr
1375	UNWIND_HINT_IRET_REGS
1376
1377	/* Everything up to here is safe from nested NMIs */
1378
1379#ifdef CONFIG_DEBUG_ENTRY
1380	/*
1381	 * For ease of testing, unmask NMIs right away.  Disabled by
1382	 * default because IRET is very expensive.
1383	 */
1384	pushq	$0		/* SS */
1385	pushq	%rsp		/* RSP (minus 8 because of the previous push) */
1386	addq	$8, (%rsp)	/* Fix up RSP */
1387	pushfq			/* RFLAGS */
1388	pushq	$__KERNEL_CS	/* CS */
1389	pushq	$1f		/* RIP */
1390	iretq			/* continues at repeat_nmi below */
1391	UNWIND_HINT_IRET_REGS
13921:
1393#endif
1394
1395repeat_nmi:
1396	ANNOTATE_NOENDBR // this code
1397	/*
1398	 * If there was a nested NMI, the first NMI's iret will return
1399	 * here. But NMIs are still enabled and we can take another
1400	 * nested NMI. The nested NMI checks the interrupted RIP to see
1401	 * if it is between repeat_nmi and end_repeat_nmi, and if so
1402	 * it will just return, as we are about to repeat an NMI anyway.
1403	 * This makes it safe to copy to the stack frame that a nested
1404	 * NMI will update.
1405	 *
1406	 * RSP is pointing to "outermost RIP".  gsbase is unknown, but, if
1407	 * we're repeating an NMI, gsbase has the same value that it had on
1408	 * the first iteration.  paranoid_entry will load the kernel
1409	 * gsbase if needed before we call exc_nmi().  "NMI executing"
1410	 * is zero.
1411	 */
1412	movq	$1, 10*8(%rsp)		/* Set "NMI executing". */
1413
1414	/*
1415	 * Copy the "outermost" frame to the "iret" frame.  NMIs that nest
1416	 * here must not modify the "iret" frame while we're writing to
1417	 * it or it will end up containing garbage.
1418	 */
1419	addq	$(10*8), %rsp
1420	.rept 5
1421	pushq	-6*8(%rsp)
1422	.endr
1423	subq	$(5*8), %rsp
1424end_repeat_nmi:
1425	ANNOTATE_NOENDBR // this code
1426
1427	/*
1428	 * Everything below this point can be preempted by a nested NMI.
1429	 * If this happens, then the inner NMI will change the "iret"
1430	 * frame to point back to repeat_nmi.
1431	 */
1432	pushq	$-1				/* ORIG_RAX: no syscall to restart */
1433
1434	/*
1435	 * Use paranoid_entry to handle SWAPGS, but no need to use paranoid_exit
1436	 * as we should not be calling schedule in NMI context.
1437	 * Even with normal interrupts enabled. An NMI should not be
1438	 * setting NEED_RESCHED or anything that normal interrupts and
1439	 * exceptions might do.
1440	 */
1441	call	paranoid_entry
1442	UNWIND_HINT_REGS
1443
1444	movq	%rsp, %rdi
1445	movq	$-1, %rsi
1446	call	exc_nmi
1447
1448	/* Always restore stashed SPEC_CTRL value (see paranoid_entry) */
1449	IBRS_EXIT save_reg=%r15
1450
1451	/* Always restore stashed CR3 value (see paranoid_entry) */
1452	RESTORE_CR3 scratch_reg=%r15 save_reg=%r14
1453
1454	/*
1455	 * The above invocation of paranoid_entry stored the GSBASE
1456	 * related information in R/EBX depending on the availability
1457	 * of FSGSBASE.
1458	 *
1459	 * If FSGSBASE is enabled, restore the saved GSBASE value
1460	 * unconditionally, otherwise take the conditional SWAPGS path.
1461	 */
1462	ALTERNATIVE "jmp nmi_no_fsgsbase", "", X86_FEATURE_FSGSBASE
1463
1464	wrgsbase	%rbx
1465	jmp	nmi_restore
1466
1467nmi_no_fsgsbase:
1468	/* EBX == 0 -> invoke SWAPGS */
1469	testl	%ebx, %ebx
1470	jnz	nmi_restore
1471
1472nmi_swapgs:
1473	swapgs
1474
1475nmi_restore:
1476	POP_REGS
1477
1478	/*
1479	 * Skip orig_ax and the "outermost" frame to point RSP at the "iret"
1480	 * at the "iret" frame.
1481	 */
1482	addq	$6*8, %rsp
1483
1484	/*
1485	 * Clear "NMI executing".  Set DF first so that we can easily
1486	 * distinguish the remaining code between here and IRET from
1487	 * the SYSCALL entry and exit paths.
1488	 *
1489	 * We arguably should just inspect RIP instead, but I (Andy) wrote
1490	 * this code when I had the misapprehension that Xen PV supported
1491	 * NMIs, and Xen PV would break that approach.
1492	 */
1493	std
1494	movq	$0, 5*8(%rsp)		/* clear "NMI executing" */
1495
1496	/*
1497	 * iretq reads the "iret" frame and exits the NMI stack in a
1498	 * single instruction.  We are returning to kernel mode, so this
1499	 * cannot result in a fault.  Similarly, we don't need to worry
1500	 * about espfix64 on the way back to kernel mode.
1501	 */
1502	iretq
1503SYM_CODE_END(asm_exc_nmi)
1504
1505#ifndef CONFIG_IA32_EMULATION
1506/*
1507 * This handles SYSCALL from 32-bit code.  There is no way to program
1508 * MSRs to fully disable 32-bit SYSCALL.
1509 */
1510SYM_CODE_START(ignore_sysret)
1511	UNWIND_HINT_EMPTY
1512	ENDBR
1513	mov	$-ENOSYS, %eax
1514	sysretl
1515SYM_CODE_END(ignore_sysret)
1516#endif
1517
1518.pushsection .text, "ax"
1519SYM_CODE_START(rewind_stack_and_make_dead)
1520	UNWIND_HINT_FUNC
1521	/* Prevent any naive code from trying to unwind to our caller. */
1522	xorl	%ebp, %ebp
1523
1524	movq	PER_CPU_VAR(cpu_current_top_of_stack), %rax
1525	leaq	-PTREGS_SIZE(%rax), %rsp
1526	UNWIND_HINT_REGS
1527
1528	call	make_task_dead
1529SYM_CODE_END(rewind_stack_and_make_dead)
1530.popsection
1531