Merge tag 'v6.8-rc4' into x86/percpu, to resolve conflicts and refresh the branch
[linux-2.6-block.git] / arch / x86 / include / asm / nospec-branch.h
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  * 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_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 RESET_CALL_DEPTH                                        \
84         xor     %eax, %eax;                                     \
85         bts     $63, %rax;                                      \
86         movq    %rax, PER_CPU_VAR(pcpu_hot + 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(pcpu_hot + X86_call_depth);   \
92         CALL_THUNKS_DEBUG_INC_CALLS
93
94 #define INCREMENT_CALL_DEPTH                                    \
95         sarq    $5, PER_CPU_VAR(pcpu_hot + 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 #ifdef __ASSEMBLY__
181
182 /*
183  * This should be used immediately before an indirect jump/call. It tells
184  * objtool the subsequent indirect jump/call is vouched safe for retpoline
185  * builds.
186  */
187 .macro ANNOTATE_RETPOLINE_SAFE
188 .Lhere_\@:
189         .pushsection .discard.retpoline_safe
190         .long .Lhere_\@
191         .popsection
192 .endm
193
194 /*
195  * (ab)use RETPOLINE_SAFE on RET to annotate away 'bare' RET instructions
196  * vs RETBleed validation.
197  */
198 #define ANNOTATE_UNRET_SAFE ANNOTATE_RETPOLINE_SAFE
199
200 /*
201  * Abuse ANNOTATE_RETPOLINE_SAFE on a NOP to indicate UNRET_END, should
202  * eventually turn into its own annotation.
203  */
204 .macro VALIDATE_UNRET_END
205 #if defined(CONFIG_NOINSTR_VALIDATION) && \
206         (defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_SRSO))
207         ANNOTATE_RETPOLINE_SAFE
208         nop
209 #endif
210 .endm
211
212 /*
213  * Equivalent to -mindirect-branch-cs-prefix; emit the 5 byte jmp/call
214  * to the retpoline thunk with a CS prefix when the register requires
215  * a RAX prefix byte to encode. Also see apply_retpolines().
216  */
217 .macro __CS_PREFIX reg:req
218         .irp rs,r8,r9,r10,r11,r12,r13,r14,r15
219         .ifc \reg,\rs
220         .byte 0x2e
221         .endif
222         .endr
223 .endm
224
225 /*
226  * JMP_NOSPEC and CALL_NOSPEC macros can be used instead of a simple
227  * indirect jmp/call which may be susceptible to the Spectre variant 2
228  * attack.
229  *
230  * NOTE: these do not take kCFI into account and are thus not comparable to C
231  * indirect calls, take care when using. The target of these should be an ENDBR
232  * instruction irrespective of kCFI.
233  */
234 .macro JMP_NOSPEC reg:req
235 #ifdef CONFIG_RETPOLINE
236         __CS_PREFIX \reg
237         jmp     __x86_indirect_thunk_\reg
238 #else
239         jmp     *%\reg
240         int3
241 #endif
242 .endm
243
244 .macro CALL_NOSPEC reg:req
245 #ifdef CONFIG_RETPOLINE
246         __CS_PREFIX \reg
247         call    __x86_indirect_thunk_\reg
248 #else
249         call    *%\reg
250 #endif
251 .endm
252
253  /*
254   * A simpler FILL_RETURN_BUFFER macro. Don't make people use the CPP
255   * monstrosity above, manually.
256   */
257 .macro FILL_RETURN_BUFFER reg:req nr:req ftr:req ftr2=ALT_NOT(X86_FEATURE_ALWAYS)
258         ALTERNATIVE_2 "jmp .Lskip_rsb_\@", \
259                 __stringify(__FILL_RETURN_BUFFER(\reg,\nr)), \ftr, \
260                 __stringify(nop;nop;__FILL_ONE_RETURN), \ftr2
261
262 .Lskip_rsb_\@:
263 .endm
264
265 #if defined(CONFIG_CPU_UNRET_ENTRY) || defined(CONFIG_CPU_SRSO)
266 #define CALL_UNTRAIN_RET        "call entry_untrain_ret"
267 #else
268 #define CALL_UNTRAIN_RET        ""
269 #endif
270
271 /*
272  * Mitigate RETBleed for AMD/Hygon Zen uarch. Requires KERNEL CR3 because the
273  * return thunk isn't mapped into the userspace tables (then again, AMD
274  * typically has NO_MELTDOWN).
275  *
276  * While retbleed_untrain_ret() doesn't clobber anything but requires stack,
277  * entry_ibpb() will clobber AX, CX, DX.
278  *
279  * As such, this must be placed after every *SWITCH_TO_KERNEL_CR3 at a point
280  * where we have a stack but before any RET instruction.
281  */
282 .macro __UNTRAIN_RET ibpb_feature, call_depth_insns
283 #if defined(CONFIG_RETHUNK) || defined(CONFIG_CPU_IBPB_ENTRY)
284         VALIDATE_UNRET_END
285         ALTERNATIVE_3 "",                                               \
286                       CALL_UNTRAIN_RET, X86_FEATURE_UNRET,              \
287                       "call entry_ibpb", \ibpb_feature,                 \
288                      __stringify(\call_depth_insns), X86_FEATURE_CALL_DEPTH
289 #endif
290 .endm
291
292 #define UNTRAIN_RET \
293         __UNTRAIN_RET X86_FEATURE_ENTRY_IBPB, __stringify(RESET_CALL_DEPTH)
294
295 #define UNTRAIN_RET_VM \
296         __UNTRAIN_RET X86_FEATURE_IBPB_ON_VMEXIT, __stringify(RESET_CALL_DEPTH)
297
298 #define UNTRAIN_RET_FROM_CALL \
299         __UNTRAIN_RET X86_FEATURE_ENTRY_IBPB, __stringify(RESET_CALL_DEPTH_FROM_CALL)
300
301
302 .macro CALL_DEPTH_ACCOUNT
303 #ifdef CONFIG_CALL_DEPTH_TRACKING
304         ALTERNATIVE "",                                                 \
305                     __stringify(INCREMENT_CALL_DEPTH), X86_FEATURE_CALL_DEPTH
306 #endif
307 .endm
308
309 #else /* __ASSEMBLY__ */
310
311 #define ANNOTATE_RETPOLINE_SAFE                                 \
312         "999:\n\t"                                              \
313         ".pushsection .discard.retpoline_safe\n\t"              \
314         ".long 999b\n\t"                                        \
315         ".popsection\n\t"
316
317 typedef u8 retpoline_thunk_t[RETPOLINE_THUNK_SIZE];
318 extern retpoline_thunk_t __x86_indirect_thunk_array[];
319 extern retpoline_thunk_t __x86_indirect_call_thunk_array[];
320 extern retpoline_thunk_t __x86_indirect_jump_thunk_array[];
321
322 #ifdef CONFIG_RETHUNK
323 extern void __x86_return_thunk(void);
324 #else
325 static inline void __x86_return_thunk(void) {}
326 #endif
327
328 #ifdef CONFIG_CPU_UNRET_ENTRY
329 extern void retbleed_return_thunk(void);
330 #else
331 static inline void retbleed_return_thunk(void) {}
332 #endif
333
334 #ifdef CONFIG_CPU_SRSO
335 extern void srso_return_thunk(void);
336 extern void srso_alias_return_thunk(void);
337 #else
338 static inline void srso_return_thunk(void) {}
339 static inline void srso_alias_return_thunk(void) {}
340 #endif
341
342 extern void retbleed_return_thunk(void);
343 extern void srso_return_thunk(void);
344 extern void srso_alias_return_thunk(void);
345
346 extern void entry_untrain_ret(void);
347 extern void entry_ibpb(void);
348
349 extern void (*x86_return_thunk)(void);
350
351 #ifdef CONFIG_CALL_DEPTH_TRACKING
352 extern void call_depth_return_thunk(void);
353
354 #define CALL_DEPTH_ACCOUNT                                      \
355         ALTERNATIVE("",                                         \
356                     __stringify(INCREMENT_CALL_DEPTH),          \
357                     X86_FEATURE_CALL_DEPTH)
358
359 #ifdef CONFIG_CALL_THUNKS_DEBUG
360 DECLARE_PER_CPU(u64, __x86_call_count);
361 DECLARE_PER_CPU(u64, __x86_ret_count);
362 DECLARE_PER_CPU(u64, __x86_stuffs_count);
363 DECLARE_PER_CPU(u64, __x86_ctxsw_count);
364 #endif
365 #else /* !CONFIG_CALL_DEPTH_TRACKING */
366
367 static inline void call_depth_return_thunk(void) {}
368 #define CALL_DEPTH_ACCOUNT ""
369
370 #endif /* CONFIG_CALL_DEPTH_TRACKING */
371
372 #ifdef CONFIG_RETPOLINE
373
374 #define GEN(reg) \
375         extern retpoline_thunk_t __x86_indirect_thunk_ ## reg;
376 #include <asm/GEN-for-each-reg.h>
377 #undef GEN
378
379 #define GEN(reg)                                                \
380         extern retpoline_thunk_t __x86_indirect_call_thunk_ ## reg;
381 #include <asm/GEN-for-each-reg.h>
382 #undef GEN
383
384 #define GEN(reg)                                                \
385         extern retpoline_thunk_t __x86_indirect_jump_thunk_ ## reg;
386 #include <asm/GEN-for-each-reg.h>
387 #undef GEN
388
389 #ifdef CONFIG_X86_64
390
391 /*
392  * Inline asm uses the %V modifier which is only in newer GCC
393  * which is ensured when CONFIG_RETPOLINE is defined.
394  */
395 # define CALL_NOSPEC                                            \
396         ALTERNATIVE_2(                                          \
397         ANNOTATE_RETPOLINE_SAFE                                 \
398         "call *%[thunk_target]\n",                              \
399         "call __x86_indirect_thunk_%V[thunk_target]\n",         \
400         X86_FEATURE_RETPOLINE,                                  \
401         "lfence;\n"                                             \
402         ANNOTATE_RETPOLINE_SAFE                                 \
403         "call *%[thunk_target]\n",                              \
404         X86_FEATURE_RETPOLINE_LFENCE)
405
406 # define THUNK_TARGET(addr) [thunk_target] "r" (addr)
407
408 #else /* CONFIG_X86_32 */
409 /*
410  * For i386 we use the original ret-equivalent retpoline, because
411  * otherwise we'll run out of registers. We don't care about CET
412  * here, anyway.
413  */
414 # define CALL_NOSPEC                                            \
415         ALTERNATIVE_2(                                          \
416         ANNOTATE_RETPOLINE_SAFE                                 \
417         "call *%[thunk_target]\n",                              \
418         "       jmp    904f;\n"                                 \
419         "       .align 16\n"                                    \
420         "901:   call   903f;\n"                                 \
421         "902:   pause;\n"                                       \
422         "       lfence;\n"                                      \
423         "       jmp    902b;\n"                                 \
424         "       .align 16\n"                                    \
425         "903:   lea    4(%%esp), %%esp;\n"                      \
426         "       pushl  %[thunk_target];\n"                      \
427         "       ret;\n"                                         \
428         "       .align 16\n"                                    \
429         "904:   call   901b;\n",                                \
430         X86_FEATURE_RETPOLINE,                                  \
431         "lfence;\n"                                             \
432         ANNOTATE_RETPOLINE_SAFE                                 \
433         "call *%[thunk_target]\n",                              \
434         X86_FEATURE_RETPOLINE_LFENCE)
435
436 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
437 #endif
438 #else /* No retpoline for C / inline asm */
439 # define CALL_NOSPEC "call *%[thunk_target]\n"
440 # define THUNK_TARGET(addr) [thunk_target] "rm" (addr)
441 #endif
442
443 /* The Spectre V2 mitigation variants */
444 enum spectre_v2_mitigation {
445         SPECTRE_V2_NONE,
446         SPECTRE_V2_RETPOLINE,
447         SPECTRE_V2_LFENCE,
448         SPECTRE_V2_EIBRS,
449         SPECTRE_V2_EIBRS_RETPOLINE,
450         SPECTRE_V2_EIBRS_LFENCE,
451         SPECTRE_V2_IBRS,
452 };
453
454 /* The indirect branch speculation control variants */
455 enum spectre_v2_user_mitigation {
456         SPECTRE_V2_USER_NONE,
457         SPECTRE_V2_USER_STRICT,
458         SPECTRE_V2_USER_STRICT_PREFERRED,
459         SPECTRE_V2_USER_PRCTL,
460         SPECTRE_V2_USER_SECCOMP,
461 };
462
463 /* The Speculative Store Bypass disable variants */
464 enum ssb_mitigation {
465         SPEC_STORE_BYPASS_NONE,
466         SPEC_STORE_BYPASS_DISABLE,
467         SPEC_STORE_BYPASS_PRCTL,
468         SPEC_STORE_BYPASS_SECCOMP,
469 };
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 extern u64 x86_pred_cmd;
483
484 static inline void indirect_branch_prediction_barrier(void)
485 {
486         alternative_msr_write(MSR_IA32_PRED_CMD, x86_pred_cmd, 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_ */