xref: /linux/arch/x86/include/asm/nospec-branch.h (revision 21ef2d065ad3f0cfbf2ae51260bf962a9fa2c643)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 
3 #ifndef _ASM_X86_NOSPEC_BRANCH_H_
4 #define _ASM_X86_NOSPEC_BRANCH_H_
5 
6 #include <linux/static_key.h>
7 #include <linux/objtool.h>
8 #include <linux/linkage.h>
9 
10 #include <asm/alternative.h>
11 #include <asm/cpufeatures.h>
12 #include <asm/msr-index.h>
13 #include <asm/unwind_hints.h>
14 #include <asm/percpu.h>
15 #include <asm/ptrace-abi.h>
16 
17 /*
18  * Call depth tracking for Intel SKL CPUs to address the RSB underflow
19  * issue in software.
20  *
21  * The tracking does not use a counter. It uses uses arithmetic shift
22  * right on call entry and logical shift left on return.
23  *
24  * The depth tracking variable is initialized to 0x8000.... when the call
25  * depth is zero. The arithmetic shift right sign extends the MSB and
26  * saturates after the 12th call. The shift count is 5 for both directions
27  * so the tracking covers 12 nested calls.
28  *
29  *  Call
30  *  0: 0x8000000000000000	0x0000000000000000
31  *  1: 0xfc00000000000000	0xf000000000000000
32  * ...
33  * 11: 0xfffffffffffffff8	0xfffffffffffffc00
34  * 12: 0xffffffffffffffff	0xffffffffffffffe0
35  *
36  * After a return buffer fill the depth is credited 12 calls before the
37  * next stuffing has to take place.
38  *
39  * There is a inaccuracy for situations like this:
40  *
41  *  10 calls
42  *   5 returns
43  *   3 calls
44  *   4 returns
45  *   3 calls
46  *   ....
47  *
48  * The shift count might cause this to be off by one in either direction,
49  * but there is still a cushion vs. the RSB depth. The algorithm does not
50  * claim to be perfect and it can be speculated around by the CPU, but it
51  * is considered that it obfuscates the problem enough to make exploitation
52  * extremely difficult.
53  */
54 #define RET_DEPTH_SHIFT			5
55 #define RSB_RET_STUFF_LOOPS		16
56 #define RET_DEPTH_INIT			0x8000000000000000ULL
57 #define RET_DEPTH_INIT_FROM_CALL	0xfc00000000000000ULL
58 #define RET_DEPTH_CREDIT		0xffffffffffffffffULL
59 
60 #ifdef CONFIG_CALL_THUNKS_DEBUG
61 # define CALL_THUNKS_DEBUG_INC_CALLS				\
62 	incq	PER_CPU_VAR(__x86_call_count);
63 # define CALL_THUNKS_DEBUG_INC_RETS				\
64 	incq	PER_CPU_VAR(__x86_ret_count);
65 # define CALL_THUNKS_DEBUG_INC_STUFFS				\
66 	incq	PER_CPU_VAR(__x86_stuffs_count);
67 # define CALL_THUNKS_DEBUG_INC_CTXSW				\
68 	incq	PER_CPU_VAR(__x86_ctxsw_count);
69 #else
70 # define CALL_THUNKS_DEBUG_INC_CALLS
71 # define CALL_THUNKS_DEBUG_INC_RETS
72 # define CALL_THUNKS_DEBUG_INC_STUFFS
73 # define CALL_THUNKS_DEBUG_INC_CTXSW
74 #endif
75 
76 #if defined(CONFIG_MITIGATION_CALL_DEPTH_TRACKING) && !defined(COMPILE_OFFSETS)
77 
78 #include <asm/asm-offsets.h>
79 
80 #define CREDIT_CALL_DEPTH					\
81 	movq	$-1, PER_CPU_VAR(__x86_call_depth);
82 
83 #define RESET_CALL_DEPTH					\
84 	xor	%eax, %eax;					\
85 	bts	$63, %rax;					\
86 	movq	%rax, PER_CPU_VAR(__x86_call_depth);
87 
88 #define RESET_CALL_DEPTH_FROM_CALL				\
89 	movb	$0xfc, %al;					\
90 	shl	$56, %rax;					\
91 	movq	%rax, PER_CPU_VAR(__x86_call_depth);		\
92 	CALL_THUNKS_DEBUG_INC_CALLS
93 
94 #define INCREMENT_CALL_DEPTH					\
95 	sarq	$5, PER_CPU_VAR(__x86_call_depth);		\
96 	CALL_THUNKS_DEBUG_INC_CALLS
97 
98 #else
99 #define CREDIT_CALL_DEPTH
100 #define RESET_CALL_DEPTH
101 #define RESET_CALL_DEPTH_FROM_CALL
102 #define INCREMENT_CALL_DEPTH
103 #endif
104 
105 /*
106  * Fill the CPU return stack buffer.
107  *
108  * Each entry in the RSB, if used for a speculative 'ret', contains an
109  * infinite 'pause; lfence; jmp' loop to capture speculative execution.
110  *
111  * This is required in various cases for retpoline and IBRS-based
112  * mitigations for the Spectre variant 2 vulnerability. Sometimes to
113  * eliminate potentially bogus entries from the RSB, and sometimes
114  * purely to ensure that it doesn't get empty, which on some CPUs would
115  * allow predictions from other (unwanted!) sources to be used.
116  *
117  * We define a CPP macro such that it can be used from both .S files and
118  * inline assembly. It's possible to do a .macro and then include that
119  * from C via asm(".include <asm/nospec-branch.h>") but let's not go there.
120  */
121 
122 #define RETPOLINE_THUNK_SIZE	32
123 #define RSB_CLEAR_LOOPS		32	/* To forcibly overwrite all entries */
124 
125 /*
126  * Common helper for __FILL_RETURN_BUFFER and __FILL_ONE_RETURN.
127  */
128 #define __FILL_RETURN_SLOT			\
129 	ANNOTATE_INTRA_FUNCTION_CALL;		\
130 	call	772f;				\
131 	int3;					\
132 772:
133 
134 /*
135  * Stuff the entire RSB.
136  *
137  * Google experimented with loop-unrolling and this turned out to be
138  * the optimal version - two calls, each with their own speculation
139  * trap should their return address end up getting used, in a loop.
140  */
141 #ifdef CONFIG_X86_64
142 #define __FILL_RETURN_BUFFER(reg, nr)			\
143 	mov	$(nr/2), reg;				\
144 771:							\
145 	__FILL_RETURN_SLOT				\
146 	__FILL_RETURN_SLOT				\
147 	add	$(BITS_PER_LONG/8) * 2, %_ASM_SP;	\
148 	dec	reg;					\
149 	jnz	771b;					\
150 	/* barrier for jnz misprediction */		\
151 	lfence;						\
152 	CREDIT_CALL_DEPTH				\
153 	CALL_THUNKS_DEBUG_INC_CTXSW
154 #else
155 /*
156  * i386 doesn't unconditionally have LFENCE, as such it can't
157  * do a loop.
158  */
159 #define __FILL_RETURN_BUFFER(reg, nr)			\
160 	.rept nr;					\
161 	__FILL_RETURN_SLOT;				\
162 	.endr;						\
163 	add	$(BITS_PER_LONG/8) * nr, %_ASM_SP;
164 #endif
165 
166 /*
167  * Stuff a single RSB slot.
168  *
169  * To mitigate Post-Barrier RSB speculation, one CALL instruction must be
170  * forced to retire before letting a RET instruction execute.
171  *
172  * On PBRSB-vulnerable CPUs, it is not safe for a RET to be executed
173  * before this point.
174  */
175 #define __FILL_ONE_RETURN				\
176 	__FILL_RETURN_SLOT				\
177 	add	$(BITS_PER_LONG/8), %_ASM_SP;		\
178 	lfence;
179 
180 /*
181  * Helper for detecting if an interrupt occurred at an unsafe location within
182  * Safe-RET.  If Safe-RET is interrupted after the CALL or LEA the RSB may get
183  * poisoned by the interrupt handler.
184  *
185  * The Safe-RET sequence is:
186  *
187  * CALL
188  * LEA 8(%RSP), %RSP
189  * RET
190  *
191  * The two CMPs below check whether RIP points to after the CALL or after the
192  * LEA.
193  *
194  * The LFENCE below is to address this particular speculation case:
195  *
196  * 1. Userspace runs and poisons the BTB around the safe-RET routine
197  *
198  * 2. Userspace triggers some kind of exception
199  *
200  * 3. Kernel executes error_entry() and mis-speculates the branch into thinking
201  *    it actually came from kernel space
202  *
203  * 4. The kernel then further mis-speculates that the exception occurred due
204  *    to an interrupted safe-RET
205  *
206  * 5. The handle_interrupted_saferet() routine speculatively executes and
207  *    speculatively does a safe-RET. But this is unsafe since it was never
208  *    untrained.
209  *
210  * The LFENCE fixes this by ensuring step 5 is never reached speculatively.
211  * Note that this LFENCE only occurs if safe-RET was actually interrupted (so
212  * it's outside of the normal path).
213  */
214 #define __HANDLE_INTR_SAFERET(name, pt_regs)		\
215 	cmpq	$(name), RIP+pt_regs;			\
216 	jb	1f;					\
217 	cmpq	$(name)+5, RIP+pt_regs;			\
218 	ja	1f;					\
219 	lfence;						\
220 	leaq	pt_regs, %rdi;				\
221 	call	handle_interrupted_saferet;		\
222 	1:
223 
224 #ifdef __ASSEMBLER__
225 
226 /*
227  * (ab)use RETPOLINE_SAFE on RET to annotate away 'bare' RET instructions
228  * vs RETBleed validation.
229  */
230 #define ANNOTATE_UNRET_SAFE ANNOTATE_RETPOLINE_SAFE
231 
232 /*
233  * Abuse ANNOTATE_RETPOLINE_SAFE on a NOP to indicate UNRET_END, should
234  * eventually turn into its own annotation.
235  */
236 .macro VALIDATE_UNRET_END
237 #if defined(CONFIG_NOINSTR_VALIDATION) && \
238 	(defined(CONFIG_MITIGATION_UNRET_ENTRY) || defined(CONFIG_MITIGATION_SRSO))
239 	ANNOTATE_RETPOLINE_SAFE
240 	nop
241 #endif
242 .endm
243 
244 /*
245  * Emits a conditional CS prefix that is compatible with
246  * -mindirect-branch-cs-prefix.
247  */
248 .macro __CS_PREFIX reg:req
249 	.irp rs,r8,r9,r10,r11,r12,r13,r14,r15
250 	.ifc \reg,\rs
251 	.byte 0x2e
252 	.endif
253 	.endr
254 .endm
255 
256 /*
257  * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
258  * indirect jmp/call which may be susceptible to the Spectre variant 2
259  * attack.
260  *
261  * NOTE: these do not take kCFI into account and are thus not comparable to C
262  * indirect calls, take care when using. The target of these should be an ENDBR
263  * instruction irrespective of kCFI.
264  */
265 .macro JMP_NOSPEC reg:req
266 #ifdef CONFIG_MITIGATION_RETPOLINE
267 	__CS_PREFIX \reg
268 	jmp	__x86_indirect_thunk_\reg
269 #else
270 	jmp	*%\reg
271 	int3
272 #endif
273 .endm
274 
275 .macro CALL_NOSPEC reg:req
276 #ifdef CONFIG_MITIGATION_RETPOLINE
277 	__CS_PREFIX \reg
278 	call	__x86_indirect_thunk_\reg
279 #else
280 	call	*%\reg
281 #endif
282 .endm
283 
284  /*
285   * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
286   * monstrosity above, manually.
287   */
288 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req ftr2=ALT_NOT(X86_FEATURE_ALWAYS)
289 	ALTERNATIVE_2 "jmp .Lskip_rsb_\@", \
290 		__stringify(__FILL_RETURN_BUFFER(\reg,\nr)), \ftr, \
291 		__stringify(nop;nop;__FILL_ONE_RETURN), \ftr2
292 
293 .Lskip_rsb_\@:
294 .endm
295 
296 /*
297  * The CALL to srso_alias_untrain_ret() must be patched in directly at
298  * the spot where untraining must be done, ie., srso_alias_untrain_ret()
299  * must be the target of a CALL instruction instead of indirectly
300  * jumping to a wrapper which then calls it. Therefore, this macro is
301  * called outside of __UNTRAIN_RET below, for the time being, before the
302  * kernel can support nested alternatives with arbitrary nesting.
303  */
304 .macro CALL_UNTRAIN_RET
305 #if defined(CONFIG_MITIGATION_UNRET_ENTRY) || defined(CONFIG_MITIGATION_SRSO)
306 	ALTERNATIVE_2 "", "call entry_untrain_ret", X86_FEATURE_UNRET, \
307 		          "call srso_alias_untrain_ret", X86_FEATURE_SRSO_ALIAS
308 #endif
309 .endm
310 
311 /*
312  * Mitigate RETBleed for AMD/Hygon Zen uarch. Requires KERNEL CR3 because the
313  * return thunk isn't mapped into the userspace tables (then again, AMD
314  * typically has NO_MELTDOWN).
315  *
316  * While retbleed_untrain_ret() doesn't clobber anything but requires stack,
317  * write_ibpb() will clobber AX, CX, DX.
318  *
319  * As such, this must be placed after every *SWITCH_TO_KERNEL_CR3 at a point
320  * where we have a stack but before any RET instruction.
321  */
322 .macro __UNTRAIN_RET ibpb_feature, call_depth_insns
323 #if defined(CONFIG_MITIGATION_RETHUNK) || defined(CONFIG_MITIGATION_IBPB_ENTRY)
324 	VALIDATE_UNRET_END
325 	CALL_UNTRAIN_RET
326 	ALTERNATIVE_2 "",						\
327 		      "call write_ibpb", \ibpb_feature,			\
328 		     __stringify(\call_depth_insns), X86_FEATURE_CALL_DEPTH
329 #endif
330 .endm
331 
332 #define UNTRAIN_RET \
333 	__UNTRAIN_RET X86_FEATURE_ENTRY_IBPB, __stringify(RESET_CALL_DEPTH)
334 
335 #define UNTRAIN_RET_VM \
336 	__UNTRAIN_RET X86_FEATURE_IBPB_ON_VMEXIT, __stringify(RESET_CALL_DEPTH)
337 
338 #define UNTRAIN_RET_FROM_CALL \
339 	__UNTRAIN_RET X86_FEATURE_ENTRY_IBPB, __stringify(RESET_CALL_DEPTH_FROM_CALL)
340 
341 .macro HANDLE_INTR_SAFERET pt_regs
342 #ifdef CONFIG_MITIGATION_SRSO
343 	ALTERNATIVE_2 "", \
344 	__stringify(__HANDLE_INTR_SAFERET(srso_safe_ret, \pt_regs)), X86_FEATURE_SRSO, \
345 	__stringify(__HANDLE_INTR_SAFERET(srso_alias_safe_ret, \pt_regs)), X86_FEATURE_SRSO_ALIAS
346 
347 #endif
348 .endm
349 
350 .macro CALL_DEPTH_ACCOUNT
351 #ifdef CONFIG_MITIGATION_CALL_DEPTH_TRACKING
352 	ALTERNATIVE "",							\
353 		    __stringify(INCREMENT_CALL_DEPTH), X86_FEATURE_CALL_DEPTH
354 #endif
355 .endm
356 
357 /*
358  * Macro to execute VERW insns that mitigate transient data sampling
359  * attacks such as MDS or TSA. On affected systems a microcode update
360  * overloaded VERW insns to also clear the CPU buffers. VERW clobbers
361  * CFLAGS.ZF.
362  * Note: Only the memory operand variant of VERW clears the CPU buffers.
363  */
364 #ifdef CONFIG_X86_64
365 #define VERW	verw x86_verw_sel(%rip)
366 #else
367 /*
368  * In 32bit mode, the memory operand must be a %cs reference. The data segments
369  * may not be usable (vm86 mode), and the stack segment may not be flat (ESPFIX32).
370  */
371 #define VERW	verw %cs:x86_verw_sel
372 #endif
373 
374 /*
375  * Provide a stringified VERW macro for simple usage, and a non-stringified
376  * VERW macro for use in more elaborate sequences, e.g. to encode a conditional
377  * VERW within an ALTERNATIVE.
378  */
379 #define __CLEAR_CPU_BUFFERS	__stringify(VERW)
380 
381 /* If necessary, emit VERW on exit-to-userspace to clear CPU buffers. */
382 #define CLEAR_CPU_BUFFERS \
383 	ALTERNATIVE "", __CLEAR_CPU_BUFFERS, X86_FEATURE_CLEAR_CPU_BUF
384 
385 #ifdef CONFIG_X86_64
386 .macro CLEAR_BRANCH_HISTORY
387 	ALTERNATIVE "", "call clear_bhb_loop", X86_FEATURE_CLEAR_BHB_LOOP
388 .endm
389 
390 .macro CLEAR_BRANCH_HISTORY_VMEXIT
391 	ALTERNATIVE "", "call clear_bhb_loop", X86_FEATURE_CLEAR_BHB_VMEXIT
392 .endm
393 #else
394 #define CLEAR_BRANCH_HISTORY
395 #define CLEAR_BRANCH_HISTORY_VMEXIT
396 #endif
397 
398 #else /* __ASSEMBLER__ */
399 
400 #define ITS_THUNK_SIZE	64
401 
402 typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE];
403 typedef u8 its_thunk_t[ITS_THUNK_SIZE];
404 extern retpoline_thunk_t __x86_indirect_thunk_array[];
405 extern retpoline_thunk_t __x86_indirect_call_thunk_array[];
406 extern retpoline_thunk_t __x86_indirect_jump_thunk_array[];
407 extern its_thunk_t	 __x86_indirect_its_thunk_array[];
408 
409 #ifdef CONFIG_MITIGATION_RETHUNK
410 extern void __x86_return_thunk(void);
411 #else
412 static inline void __x86_return_thunk(void) {}
413 #endif
414 
415 #ifdef CONFIG_MITIGATION_UNRET_ENTRY
416 extern void retbleed_return_thunk(void);
417 #else
418 static inline void retbleed_return_thunk(void) {}
419 #endif
420 
421 extern void srso_alias_untrain_ret(void);
422 
423 #ifdef CONFIG_MITIGATION_SRSO
424 extern void srso_return_thunk(void);
425 extern void srso_alias_return_thunk(void);
426 #else
427 static inline void srso_return_thunk(void) {}
428 static inline void srso_alias_return_thunk(void) {}
429 #endif
430 
431 #ifdef CONFIG_MITIGATION_ITS
432 extern void its_return_thunk(void);
433 #else
434 static inline void its_return_thunk(void) {}
435 #endif
436 
437 extern void retbleed_return_thunk(void);
438 extern void srso_return_thunk(void);
439 extern void srso_alias_return_thunk(void);
440 
441 extern void entry_untrain_ret(void);
442 extern void write_ibpb(void);
443 
444 #ifdef CONFIG_BPF_JIT
445 extern void bpf_arch_ibpb(void);
446 #endif
447 
448 #ifdef CONFIG_X86_64
449 extern void clear_bhb_loop(void);
450 #endif
451 
452 extern void (*x86_return_thunk)(void);
453 
454 extern void __warn_thunk(void);
455 
456 #ifdef CONFIG_MITIGATION_CALL_DEPTH_TRACKING
457 extern void call_depth_return_thunk(void);
458 
459 #define CALL_DEPTH_ACCOUNT					\
460 	ALTERNATIVE("",						\
461 		    __stringify(INCREMENT_CALL_DEPTH),		\
462 		    X86_FEATURE_CALL_DEPTH)
463 
464 DECLARE_PER_CPU_CACHE_HOT(u64, __x86_call_depth);
465 
466 #ifdef CONFIG_CALL_THUNKS_DEBUG
467 DECLARE_PER_CPU(u64, __x86_call_count);
468 DECLARE_PER_CPU(u64, __x86_ret_count);
469 DECLARE_PER_CPU(u64, __x86_stuffs_count);
470 DECLARE_PER_CPU(u64, __x86_ctxsw_count);
471 #endif
472 #else /* !CONFIG_MITIGATION_CALL_DEPTH_TRACKING */
473 
474 static inline void call_depth_return_thunk(void) {}
475 #define CALL_DEPTH_ACCOUNT ""
476 
477 #endif /* CONFIG_MITIGATION_CALL_DEPTH_TRACKING */
478 
479 #ifdef CONFIG_MITIGATION_RETPOLINE
480 
481 #define GEN(reg) \
482 	extern retpoline_thunk_t __x86_indirect_thunk_ ## reg;
483 #include <asm/GEN-for-each-reg.h>
484 #undef GEN
485 
486 #define GEN(reg)						\
487 	extern retpoline_thunk_t __x86_indirect_call_thunk_ ## reg;
488 #include <asm/GEN-for-each-reg.h>
489 #undef GEN
490 
491 #define GEN(reg)						\
492 	extern retpoline_thunk_t __x86_indirect_jump_thunk_ ## reg;
493 #include <asm/GEN-for-each-reg.h>
494 #undef GEN
495 
496 #ifdef CONFIG_X86_64
497 
498 /*
499  * Emits a conditional CS prefix that is compatible with
500  * -mindirect-branch-cs-prefix.
501  */
502 #define __CS_PREFIX(reg)				\
503 	".irp rs,r8,r9,r10,r11,r12,r13,r14,r15\n"	\
504 	".ifc \\rs," reg "\n"				\
505 	".byte 0x2e\n"					\
506 	".endif\n"					\
507 	".endr\n"
508 
509 /*
510  * Inline asm uses the %V modifier which is only in newer GCC
511  * which is ensured when CONFIG_MITIGATION_RETPOLINE is defined.
512  */
513 #define CALL_NOSPEC	__CS_PREFIX("%V[thunk_target]")	\
514 			"call __x86_indirect_thunk_%V[thunk_target]\n"
515 
516 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
517 
518 #else /* CONFIG_X86_32 */
519 /*
520  * For i386 we use the original ret-equivalent retpoline, because
521  * otherwise we'll run out of registers. We don't care about CET
522  * here, anyway.
523  */
524 # define CALL_NOSPEC						\
525 	ALTERNATIVE_2(						\
526 	ANNOTATE_RETPOLINE_SAFE "\n"				\
527 	"call *%[thunk_target]\n",				\
528 	"       jmp    904f;\n"					\
529 	"       .align 16\n"					\
530 	"901:	call   903f;\n"					\
531 	"902:	pause;\n"					\
532 	"    	lfence;\n"					\
533 	"       jmp    902b;\n"					\
534 	"       .align 16\n"					\
535 	"903:	lea    4(%%esp), %%esp;\n"			\
536 	"       pushl  %[thunk_target];\n"			\
537 	"       ret;\n"						\
538 	"       .align 16\n"					\
539 	"904:	call   901b;\n",				\
540 	X86_FEATURE_RETPOLINE,					\
541 	"lfence;\n"						\
542 	ANNOTATE_RETPOLINE_SAFE "\n"				\
543 	"call *%[thunk_target]\n",				\
544 	X86_FEATURE_RETPOLINE_LFENCE)
545 
546 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
547 #endif
548 #else /* No retpoline for C / inline asm */
549 # define CALL_NOSPEC "call *%[thunk_target]\n"
550 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
551 #endif
552 
553 /* The Spectre V2 mitigation variants */
554 enum spectre_v2_mitigation {
555 	SPECTRE_V2_NONE,
556 	SPECTRE_V2_RETPOLINE,
557 	SPECTRE_V2_LFENCE,
558 	SPECTRE_V2_EIBRS,
559 	SPECTRE_V2_EIBRS_RETPOLINE,
560 	SPECTRE_V2_EIBRS_LFENCE,
561 	SPECTRE_V2_IBRS,
562 };
563 
564 /* The indirect branch speculation control variants */
565 enum spectre_v2_user_mitigation {
566 	SPECTRE_V2_USER_NONE,
567 	SPECTRE_V2_USER_STRICT,
568 	SPECTRE_V2_USER_STRICT_PREFERRED,
569 	SPECTRE_V2_USER_PRCTL,
570 	SPECTRE_V2_USER_SECCOMP,
571 };
572 
573 /* The Speculative Store Bypass disable variants */
574 enum ssb_mitigation {
575 	SPEC_STORE_BYPASS_NONE,
576 	SPEC_STORE_BYPASS_AUTO,
577 	SPEC_STORE_BYPASS_DISABLE,
578 	SPEC_STORE_BYPASS_PRCTL,
579 	SPEC_STORE_BYPASS_SECCOMP,
580 };
581 
582 static __always_inline
583 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature)
584 {
585 	asm volatile(ALTERNATIVE("", "wrmsr", %c[feature])
586 		: : "c" (msr),
587 		    "a" ((u32)val),
588 		    "d" ((u32)(val >> 32)),
589 		    [feature] "i" (feature)
590 		: "memory");
591 }
592 
593 DECLARE_PER_CPU(bool, x86_ibpb_exit_to_user);
594 
595 static inline void indirect_branch_prediction_barrier(void)
596 {
597 	asm_inline volatile(ALTERNATIVE("", "call write_ibpb", X86_FEATURE_IBPB)
598 			    : ASM_CALL_CONSTRAINT
599 			    :: "rax", "rcx", "rdx", "memory");
600 }
601 
602 /* The Intel SPEC CTRL MSR base value cache */
603 extern u64 x86_spec_ctrl_base;
604 DECLARE_PER_CPU(u64, x86_spec_ctrl_current);
605 extern void update_spec_ctrl_cond(u64 val);
606 extern u64 spec_ctrl_current(void);
607 
608 /*
609  * With retpoline, we must use IBRS to restrict branch prediction
610  * before calling into firmware.
611  *
612  * (Implemented as CPP macros due to header hell.)
613  */
614 #define firmware_restrict_branch_speculation_start()			\
615 do {									\
616 	preempt_disable();						\
617 	alternative_msr_write(MSR_IA32_SPEC_CTRL,			\
618 			      spec_ctrl_current() | SPEC_CTRL_IBRS,	\
619 			      X86_FEATURE_USE_IBRS_FW);			\
620 	alternative_msr_write(MSR_IA32_PRED_CMD, PRED_CMD_IBPB,		\
621 			      X86_FEATURE_USE_IBPB_FW);			\
622 } while (0)
623 
624 #define firmware_restrict_branch_speculation_end()			\
625 do {									\
626 	alternative_msr_write(MSR_IA32_SPEC_CTRL,			\
627 			      spec_ctrl_current(),			\
628 			      X86_FEATURE_USE_IBRS_FW);			\
629 	preempt_enable();						\
630 } while (0)
631 
632 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp);
633 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb);
634 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb);
635 
636 DECLARE_STATIC_KEY_FALSE(switch_vcpu_ibpb);
637 
638 DECLARE_STATIC_KEY_FALSE(cpu_buf_idle_clear);
639 
640 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_l1d_flush);
641 
642 extern u16 x86_verw_sel;
643 
644 #include <asm/segment.h>
645 
646 /**
647  * x86_clear_cpu_buffers - Buffer clearing support for different x86 CPU vulns
648  *
649  * This uses the otherwise unused and obsolete VERW instruction in
650  * combination with microcode which triggers a CPU buffer flush when the
651  * instruction is executed.
652  */
653 static __always_inline void x86_clear_cpu_buffers(void)
654 {
655 	static const u16 ds = __KERNEL_DS;
656 
657 	/*
658 	 * Has to be the memory-operand variant because only that
659 	 * guarantees the CPU buffer flush functionality according to
660 	 * documentation. The register-operand variant does not.
661 	 * Works with any segment selector, but a valid writable
662 	 * data segment is the fastest variant.
663 	 *
664 	 * "cc" clobber is required because VERW modifies ZF.
665 	 */
666 	asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc");
667 }
668 
669 /**
670  * x86_idle_clear_cpu_buffers - Buffer clearing support in idle for the MDS
671  * and TSA vulnerabilities.
672  *
673  * Clear CPU buffers if the corresponding static key is enabled
674  */
675 static __always_inline void x86_idle_clear_cpu_buffers(void)
676 {
677 	if (static_branch_likely(&cpu_buf_idle_clear))
678 		x86_clear_cpu_buffers();
679 }
680 
681 void srso_safe_ret(void);
682 void srso_alias_safe_ret(void);
683 void handle_interrupted_saferet(struct pt_regs *regs);
684 
685 #endif /* __ASSEMBLER__ */
686 
687 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */
688