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/current.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  * extremly 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	%gs:__x86_call_count;
63 # define CALL_THUNKS_DEBUG_INC_RETS				\
64 	incq	%gs:__x86_ret_count;
65 # define CALL_THUNKS_DEBUG_INC_STUFFS				\
66 	incq	%gs:__x86_stuffs_count;
67 # define CALL_THUNKS_DEBUG_INC_CTXSW				\
68 	incq	%gs:__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_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(pcpu_hot + X86_call_depth);
82 
83 #define ASM_CREDIT_CALL_DEPTH					\
84 	movq	$-1, PER_CPU_VAR(pcpu_hot + X86_call_depth);
85 
86 #define RESET_CALL_DEPTH					\
87 	xor	%eax, %eax;					\
88 	bts	$63, %rax;					\
89 	movq	%rax, PER_CPU_VAR(pcpu_hot + X86_call_depth);
90 
91 #define RESET_CALL_DEPTH_FROM_CALL				\
92 	movb	$0xfc, %al;					\
93 	shl	$56, %rax;					\
94 	movq	%rax, PER_CPU_VAR(pcpu_hot + X86_call_depth);	\
95 	CALL_THUNKS_DEBUG_INC_CALLS
96 
97 #define INCREMENT_CALL_DEPTH					\
98 	sarq	$5, %gs:pcpu_hot + X86_call_depth;		\
99 	CALL_THUNKS_DEBUG_INC_CALLS
100 
101 #define ASM_INCREMENT_CALL_DEPTH				\
102 	sarq	$5, PER_CPU_VAR(pcpu_hot + X86_call_depth);	\
103 	CALL_THUNKS_DEBUG_INC_CALLS
104 
105 #else
106 #define CREDIT_CALL_DEPTH
107 #define ASM_CREDIT_CALL_DEPTH
108 #define RESET_CALL_DEPTH
109 #define INCREMENT_CALL_DEPTH
110 #define ASM_INCREMENT_CALL_DEPTH
111 #define RESET_CALL_DEPTH_FROM_CALL
112 #endif
113 
114 /*
115  * Fill the CPU return stack buffer.
116  *
117  * Each entry in the RSB, if used for a speculative 'ret', contains an
118  * infinite 'pause; lfence; jmp' loop to capture speculative execution.
119  *
120  * This is required in various cases for retpoline and IBRS-based
121  * mitigations for the Spectre variant 2 vulnerability. Sometimes to
122  * eliminate potentially bogus entries from the RSB, and sometimes
123  * purely to ensure that it doesn't get empty, which on some CPUs would
124  * allow predictions from other (unwanted!) sources to be used.
125  *
126  * We define a CPP macro such that it can be used from both .S files and
127  * inline assembly. It's possible to do a .macro and then include that
128  * from C via asm(".include <asm/nospec-branch.h>") but let's not go there.
129  */
130 
131 #define RETPOLINE_THUNK_SIZE	32
132 #define RSB_CLEAR_LOOPS		32	/* To forcibly overwrite all entries */
133 
134 /*
135  * Common helper for __FILL_RETURN_BUFFER and __FILL_ONE_RETURN.
136  */
137 #define __FILL_RETURN_SLOT			\
138 	ANNOTATE_INTRA_FUNCTION_CALL;		\
139 	call	772f;				\
140 	int3;					\
141 772:
142 
143 /*
144  * Stuff the entire RSB.
145  *
146  * Google experimented with loop-unrolling and this turned out to be
147  * the optimal version - two calls, each with their own speculation
148  * trap should their return address end up getting used, in a loop.
149  */
150 #ifdef CONFIG_X86_64
151 #define __FILL_RETURN_BUFFER(reg, nr)			\
152 	mov	$(nr/2), reg;				\
153 771:							\
154 	__FILL_RETURN_SLOT				\
155 	__FILL_RETURN_SLOT				\
156 	add	$(BITS_PER_LONG/8) * 2, %_ASM_SP;	\
157 	dec	reg;					\
158 	jnz	771b;					\
159 	/* barrier for jnz misprediction */		\
160 	lfence;						\
161 	ASM_CREDIT_CALL_DEPTH				\
162 	CALL_THUNKS_DEBUG_INC_CTXSW
163 #else
164 /*
165  * i386 doesn't unconditionally have LFENCE, as such it can't
166  * do a loop.
167  */
168 #define __FILL_RETURN_BUFFER(reg, nr)			\
169 	.rept nr;					\
170 	__FILL_RETURN_SLOT;				\
171 	.endr;						\
172 	add	$(BITS_PER_LONG/8) * nr, %_ASM_SP;
173 #endif
174 
175 /*
176  * Stuff a single RSB slot.
177  *
178  * To mitigate Post-Barrier RSB speculation, one CALL instruction must be
179  * forced to retire before letting a RET instruction execute.
180  *
181  * On PBRSB-vulnerable CPUs, it is not safe for a RET to be executed
182  * before this point.
183  */
184 #define __FILL_ONE_RETURN				\
185 	__FILL_RETURN_SLOT				\
186 	add	$(BITS_PER_LONG/8), %_ASM_SP;		\
187 	lfence;
188 
189 #ifdef __ASSEMBLY__
190 
191 /*
192  * This should be used immediately before an indirect jump/call. It tells
193  * objtool the subsequent indirect jump/call is vouched safe for retpoline
194  * builds.
195  */
196 .macro ANNOTATE_RETPOLINE_SAFE
197 .Lhere_\@:
198 	.pushsection .discard.retpoline_safe
199 	.long .Lhere_\@ - .
200 	.popsection
201 .endm
202 
203 /*
204  * (ab)use RETPOLINE_SAFE on RET to annotate away 'bare' RET instructions
205  * vs RETBleed validation.
206  */
207 #define ANNOTATE_UNRET_SAFE ANNOTATE_RETPOLINE_SAFE
208 
209 /*
210  * Abuse ANNOTATE_RETPOLINE_SAFE on a NOP to indicate UNRET_END, should
211  * eventually turn into it's own annotation.
212  */
213 .macro VALIDATE_UNRET_END
214 #if defined(CONFIG_NOINSTR_VALIDATION) && defined(CONFIG_CPU_UNRET_ENTRY)
215 	ANNOTATE_RETPOLINE_SAFE
216 	nop
217 #endif
218 .endm
219 
220 /*
221  * Equivalent to -mindirect-branch-cs-prefix; emit the 5 byte jmp/call
222  * to the retpoline thunk with a CS prefix when the register requires
223  * a RAX prefix byte to encode. Also see apply_retpolines().
224  */
225 .macro __CS_PREFIX reg:req
226 	.irp rs,r8,r9,r10,r11,r12,r13,r14,r15
227 	.ifc \reg,\rs
228 	.byte 0x2e
229 	.endif
230 	.endr
231 .endm
232 
233 /*
234  * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
235  * indirect jmp/call which may be susceptible to the Spectre variant 2
236  * attack.
237  *
238  * NOTE: these do not take kCFI into account and are thus not comparable to C
239  * indirect calls, take care when using. The target of these should be an ENDBR
240  * instruction irrespective of kCFI.
241  */
242 .macro JMP_NOSPEC reg:req
243 #ifdef CONFIG_RETPOLINE
244 	__CS_PREFIX \reg
245 	jmp	__x86_indirect_thunk_\reg
246 #else
247 	jmp	*%\reg
248 	int3
249 #endif
250 .endm
251 
252 .macro CALL_NOSPEC reg:req
253 #ifdef CONFIG_RETPOLINE
254 	__CS_PREFIX \reg
255 	call	__x86_indirect_thunk_\reg
256 #else
257 	call	*%\reg
258 #endif
259 .endm
260 
261  /*
262   * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
263   * monstrosity above, manually.
264   */
265 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req ftr2=ALT_NOT(X86_FEATURE_ALWAYS)
266 	ALTERNATIVE_2 "jmp .Lskip_rsb_\@", \
267 		__stringify(__FILL_RETURN_BUFFER(\reg,\nr)), \ftr, \
268 		__stringify(nop;nop;__FILL_ONE_RETURN), \ftr2
269 
270 .Lskip_rsb_\@:
271 .endm
272 
273 #ifdef CONFIG_CPU_UNRET_ENTRY
274 #define CALL_ZEN_UNTRAIN_RET	"call zen_untrain_ret"
275 #else
276 #define CALL_ZEN_UNTRAIN_RET	""
277 #endif
278 
279 /*
280  * Mitigate RETBleed for AMD/Hygon Zen uarch. Requires KERNEL CR3 because the
281  * return thunk isn't mapped into the userspace tables (then again, AMD
282  * typically has NO_MELTDOWN).
283  *
284  * While zen_untrain_ret() doesn't clobber anything but requires stack,
285  * entry_ibpb() will clobber AX, CX, DX.
286  *
287  * As such, this must be placed after every *SWITCH_TO_KERNEL_CR3 at a point
288  * where we have a stack but before any RET instruction.
289  */
290 .macro UNTRAIN_RET
291 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY) || \
292 	defined(CONFIG_CALL_DEPTH_TRACKING)
293 	VALIDATE_UNRET_END
294 	ALTERNATIVE_3 "",						\
295 		      CALL_ZEN_UNTRAIN_RET, X86_FEATURE_UNRET,		\
296 		      "call entry_ibpb", X86_FEATURE_ENTRY_IBPB,	\
297 		      __stringify(RESET_CALL_DEPTH), X86_FEATURE_CALL_DEPTH
298 #endif
299 .endm
300 
301 .macro UNTRAIN_RET_FROM_CALL
302 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_IBPB_ENTRY) || \
303 	defined(CONFIG_CALL_DEPTH_TRACKING)
304 	VALIDATE_UNRET_END
305 	ALTERNATIVE_3 "",						\
306 		      CALL_ZEN_UNTRAIN_RET, X86_FEATURE_UNRET,		\
307 		      "call entry_ibpb", X86_FEATURE_ENTRY_IBPB,	\
308 		      __stringify(RESET_CALL_DEPTH_FROM_CALL), X86_FEATURE_CALL_DEPTH
309 #endif
310 .endm
311 
312 
313 .macro CALL_DEPTH_ACCOUNT
314 #ifdef CONFIG_CALL_DEPTH_TRACKING
315 	ALTERNATIVE "",							\
316 		    __stringify(ASM_INCREMENT_CALL_DEPTH), X86_FEATURE_CALL_DEPTH
317 #endif
318 .endm
319 
320 #else /* __ASSEMBLY__ */
321 
322 #define ANNOTATE_RETPOLINE_SAFE					\
323 	"999:\n\t"						\
324 	".pushsection .discard.retpoline_safe\n\t"		\
325 	".long 999b - .\n\t"					\
326 	".popsection\n\t"
327 
328 typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE];
329 extern retpoline_thunk_t __x86_indirect_thunk_array[];
330 extern retpoline_thunk_t __x86_indirect_call_thunk_array[];
331 extern retpoline_thunk_t __x86_indirect_jump_thunk_array[];
332 
333 extern void __x86_return_thunk(void);
334 extern void zen_untrain_ret(void);
335 extern void entry_ibpb(void);
336 
337 #ifdef CONFIG_CALL_THUNKS
338 extern void (*x86_return_thunk)(void);
339 #else
340 #define x86_return_thunk	(&__x86_return_thunk)
341 #endif
342 
343 #ifdef CONFIG_CALL_DEPTH_TRACKING
344 extern void __x86_return_skl(void);
345 
346 static inline void x86_set_skl_return_thunk(void)
347 {
348 	x86_return_thunk = &__x86_return_skl;
349 }
350 
351 #define CALL_DEPTH_ACCOUNT					\
352 	ALTERNATIVE("",						\
353 		    __stringify(INCREMENT_CALL_DEPTH),		\
354 		    X86_FEATURE_CALL_DEPTH)
355 
356 #ifdef CONFIG_CALL_THUNKS_DEBUG
357 DECLARE_PER_CPU(u64, __x86_call_count);
358 DECLARE_PER_CPU(u64, __x86_ret_count);
359 DECLARE_PER_CPU(u64, __x86_stuffs_count);
360 DECLARE_PER_CPU(u64, __x86_ctxsw_count);
361 #endif
362 #else
363 static inline void x86_set_skl_return_thunk(void) {}
364 
365 #define CALL_DEPTH_ACCOUNT ""
366 
367 #endif
368 
369 #ifdef CONFIG_RETPOLINE
370 
371 #define GEN(reg) \
372 	extern retpoline_thunk_t __x86_indirect_thunk_ ## reg;
373 #include <asm/GEN-for-each-reg.h>
374 #undef GEN
375 
376 #define GEN(reg)						\
377 	extern retpoline_thunk_t __x86_indirect_call_thunk_ ## reg;
378 #include <asm/GEN-for-each-reg.h>
379 #undef GEN
380 
381 #define GEN(reg)						\
382 	extern retpoline_thunk_t __x86_indirect_jump_thunk_ ## reg;
383 #include <asm/GEN-for-each-reg.h>
384 #undef GEN
385 
386 #ifdef CONFIG_X86_64
387 
388 /*
389  * Inline asm uses the %V modifier which is only in newer GCC
390  * which is ensured when CONFIG_RETPOLINE is defined.
391  */
392 # define CALL_NOSPEC						\
393 	ALTERNATIVE_2(						\
394 	ANNOTATE_RETPOLINE_SAFE					\
395 	"call *%[thunk_target]\n",				\
396 	"call __x86_indirect_thunk_%V[thunk_target]\n",		\
397 	X86_FEATURE_RETPOLINE,					\
398 	"lfence;\n"						\
399 	ANNOTATE_RETPOLINE_SAFE					\
400 	"call *%[thunk_target]\n",				\
401 	X86_FEATURE_RETPOLINE_LFENCE)
402 
403 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
404 
405 #else /* CONFIG_X86_32 */
406 /*
407  * For i386 we use the original ret-equivalent retpoline, because
408  * otherwise we'll run out of registers. We don't care about CET
409  * here, anyway.
410  */
411 # define CALL_NOSPEC						\
412 	ALTERNATIVE_2(						\
413 	ANNOTATE_RETPOLINE_SAFE					\
414 	"call *%[thunk_target]\n",				\
415 	"       jmp    904f;\n"					\
416 	"       .align 16\n"					\
417 	"901:	call   903f;\n"					\
418 	"902:	pause;\n"					\
419 	"    	lfence;\n"					\
420 	"       jmp    902b;\n"					\
421 	"       .align 16\n"					\
422 	"903:	lea    4(%%esp), %%esp;\n"			\
423 	"       pushl  %[thunk_target];\n"			\
424 	"       ret;\n"						\
425 	"       .align 16\n"					\
426 	"904:	call   901b;\n",				\
427 	X86_FEATURE_RETPOLINE,					\
428 	"lfence;\n"						\
429 	ANNOTATE_RETPOLINE_SAFE					\
430 	"call *%[thunk_target]\n",				\
431 	X86_FEATURE_RETPOLINE_LFENCE)
432 
433 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
434 #endif
435 #else /* No retpoline for C / inline asm */
436 # define CALL_NOSPEC "call *%[thunk_target]\n"
437 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
438 #endif
439 
440 /* The Spectre V2 mitigation variants */
441 enum spectre_v2_mitigation {
442 	SPECTRE_V2_NONE,
443 	SPECTRE_V2_RETPOLINE,
444 	SPECTRE_V2_LFENCE,
445 	SPECTRE_V2_EIBRS,
446 	SPECTRE_V2_EIBRS_RETPOLINE,
447 	SPECTRE_V2_EIBRS_LFENCE,
448 	SPECTRE_V2_IBRS,
449 };
450 
451 /* The indirect branch speculation control variants */
452 enum spectre_v2_user_mitigation {
453 	SPECTRE_V2_USER_NONE,
454 	SPECTRE_V2_USER_STRICT,
455 	SPECTRE_V2_USER_STRICT_PREFERRED,
456 	SPECTRE_V2_USER_PRCTL,
457 	SPECTRE_V2_USER_SECCOMP,
458 };
459 
460 /* The Speculative Store Bypass disable variants */
461 enum ssb_mitigation {
462 	SPEC_STORE_BYPASS_NONE,
463 	SPEC_STORE_BYPASS_DISABLE,
464 	SPEC_STORE_BYPASS_PRCTL,
465 	SPEC_STORE_BYPASS_SECCOMP,
466 };
467 
468 extern char __indirect_thunk_start[];
469 extern char __indirect_thunk_end[];
470 
471 static __always_inline
472 void alternative_msr_write(unsigned int msr, u64 val, unsigned int feature)
473 {
474 	asm volatile(ALTERNATIVE("", "wrmsr", %c[feature])
475 		: : "c" (msr),
476 		    "a" ((u32)val),
477 		    "d" ((u32)(val >> 32)),
478 		    [feature] "i" (feature)
479 		: "memory");
480 }
481 
482 static inline void indirect_branch_prediction_barrier(void)
483 {
484 	u64 val = PRED_CMD_IBPB;
485 
486 	alternative_msr_write(MSR_IA32_PRED_CMD, val, X86_FEATURE_USE_IBPB);
487 }
488 
489 /* The Intel SPEC CTRL MSR base value cache */
490 extern u64 x86_spec_ctrl_base;
491 DECLARE_PER_CPU(u64, x86_spec_ctrl_current);
492 extern void update_spec_ctrl_cond(u64 val);
493 extern u64 spec_ctrl_current(void);
494 
495 /*
496  * With retpoline, we must use IBRS to restrict branch prediction
497  * before calling into firmware.
498  *
499  * (Implemented as CPP macros due to header hell.)
500  */
501 #define firmware_restrict_branch_speculation_start()			\
502 do {									\
503 	preempt_disable();						\
504 	alternative_msr_write(MSR_IA32_SPEC_CTRL,			\
505 			      spec_ctrl_current() | SPEC_CTRL_IBRS,	\
506 			      X86_FEATURE_USE_IBRS_FW);			\
507 	alternative_msr_write(MSR_IA32_PRED_CMD, PRED_CMD_IBPB,		\
508 			      X86_FEATURE_USE_IBPB_FW);			\
509 } while (0)
510 
511 #define firmware_restrict_branch_speculation_end()			\
512 do {									\
513 	alternative_msr_write(MSR_IA32_SPEC_CTRL,			\
514 			      spec_ctrl_current(),			\
515 			      X86_FEATURE_USE_IBRS_FW);			\
516 	preempt_enable();						\
517 } while (0)
518 
519 DECLARE_STATIC_KEY_FALSE(switch_to_cond_stibp);
520 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_ibpb);
521 DECLARE_STATIC_KEY_FALSE(switch_mm_always_ibpb);
522 
523 DECLARE_STATIC_KEY_FALSE(mds_user_clear);
524 DECLARE_STATIC_KEY_FALSE(mds_idle_clear);
525 
526 DECLARE_STATIC_KEY_FALSE(switch_mm_cond_l1d_flush);
527 
528 DECLARE_STATIC_KEY_FALSE(mmio_stale_data_clear);
529 
530 #include <asm/segment.h>
531 
532 /**
533  * mds_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
534  *
535  * This uses the otherwise unused and obsolete VERW instruction in
536  * combination with microcode which triggers a CPU buffer flush when the
537  * instruction is executed.
538  */
539 static __always_inline void mds_clear_cpu_buffers(void)
540 {
541 	static const u16 ds = __KERNEL_DS;
542 
543 	/*
544 	 * Has to be the memory-operand variant because only that
545 	 * guarantees the CPU buffer flush functionality according to
546 	 * documentation. The register-operand variant does not.
547 	 * Works with any segment selector, but a valid writable
548 	 * data segment is the fastest variant.
549 	 *
550 	 * "cc" clobber is required because VERW modifies ZF.
551 	 */
552 	asm volatile("verw %[ds]" : : [ds] "m" (ds) : "cc");
553 }
554 
555 /**
556  * mds_user_clear_cpu_buffers - Mitigation for MDS and TAA vulnerability
557  *
558  * Clear CPU buffers if the corresponding static key is enabled
559  */
560 static __always_inline void mds_user_clear_cpu_buffers(void)
561 {
562 	if (static_branch_likely(&mds_user_clear))
563 		mds_clear_cpu_buffers();
564 }
565 
566 /**
567  * mds_idle_clear_cpu_buffers - Mitigation for MDS vulnerability
568  *
569  * Clear CPU buffers if the corresponding static key is enabled
570  */
571 static __always_inline void mds_idle_clear_cpu_buffers(void)
572 {
573 	if (static_branch_likely(&mds_idle_clear))
574 		mds_clear_cpu_buffers();
575 }
576 
577 #endif /* __ASSEMBLY__ */
578 
579 #endif /* _ASM_X86_NOSPEC_BRANCH_H_ */
580