arm64: vdso32: Fix '--prefix=' value for newer versions of clang
[linux-2.6-microblaze.git] / arch / x86 / include / asm / fpu / internal.h
1 /* SPDX-License-Identifier: GPL-2.0 */
2 /*
3  * Copyright (C) 1994 Linus Torvalds
4  *
5  * Pentium III FXSR, SSE support
6  * General FPU state handling cleanups
7  *      Gareth Hughes <gareth@valinux.com>, May 2000
8  * x86-64 work by Andi Kleen 2002
9  */
10
11 #ifndef _ASM_X86_FPU_INTERNAL_H
12 #define _ASM_X86_FPU_INTERNAL_H
13
14 #include <linux/compat.h>
15 #include <linux/sched.h>
16 #include <linux/slab.h>
17 #include <linux/mm.h>
18
19 #include <asm/user.h>
20 #include <asm/fpu/api.h>
21 #include <asm/fpu/xstate.h>
22 #include <asm/cpufeature.h>
23 #include <asm/trace/fpu.h>
24
25 /*
26  * High level FPU state handling functions:
27  */
28 extern void fpu__prepare_read(struct fpu *fpu);
29 extern void fpu__prepare_write(struct fpu *fpu);
30 extern void fpu__save(struct fpu *fpu);
31 extern int  fpu__restore_sig(void __user *buf, int ia32_frame);
32 extern void fpu__drop(struct fpu *fpu);
33 extern int  fpu__copy(struct task_struct *dst, struct task_struct *src);
34 extern void fpu__clear_user_states(struct fpu *fpu);
35 extern void fpu__clear_all(struct fpu *fpu);
36 extern int  fpu__exception_code(struct fpu *fpu, int trap_nr);
37 extern int  dump_fpu(struct pt_regs *ptregs, struct user_i387_struct *fpstate);
38
39 /*
40  * Boot time FPU initialization functions:
41  */
42 extern void fpu__init_cpu(void);
43 extern void fpu__init_system_xstate(void);
44 extern void fpu__init_cpu_xstate(void);
45 extern void fpu__init_system(struct cpuinfo_x86 *c);
46 extern void fpu__init_check_bugs(void);
47 extern void fpu__resume_cpu(void);
48 extern u64 fpu__get_supported_xfeatures_mask(void);
49
50 /*
51  * Debugging facility:
52  */
53 #ifdef CONFIG_X86_DEBUG_FPU
54 # define WARN_ON_FPU(x) WARN_ON_ONCE(x)
55 #else
56 # define WARN_ON_FPU(x) ({ (void)(x); 0; })
57 #endif
58
59 /*
60  * FPU related CPU feature flag helper routines:
61  */
62 static __always_inline __pure bool use_xsaveopt(void)
63 {
64         return static_cpu_has(X86_FEATURE_XSAVEOPT);
65 }
66
67 static __always_inline __pure bool use_xsave(void)
68 {
69         return static_cpu_has(X86_FEATURE_XSAVE);
70 }
71
72 static __always_inline __pure bool use_fxsr(void)
73 {
74         return static_cpu_has(X86_FEATURE_FXSR);
75 }
76
77 /*
78  * fpstate handling functions:
79  */
80
81 extern union fpregs_state init_fpstate;
82
83 extern void fpstate_init(union fpregs_state *state);
84 #ifdef CONFIG_MATH_EMULATION
85 extern void fpstate_init_soft(struct swregs_state *soft);
86 #else
87 static inline void fpstate_init_soft(struct swregs_state *soft) {}
88 #endif
89
90 static inline void fpstate_init_xstate(struct xregs_state *xsave)
91 {
92         /*
93          * XRSTORS requires these bits set in xcomp_bv, or it will
94          * trigger #GP:
95          */
96         xsave->header.xcomp_bv = XCOMP_BV_COMPACTED_FORMAT | xfeatures_mask_all;
97 }
98
99 static inline void fpstate_init_fxstate(struct fxregs_state *fx)
100 {
101         fx->cwd = 0x37f;
102         fx->mxcsr = MXCSR_DEFAULT;
103 }
104 extern void fpstate_sanitize_xstate(struct fpu *fpu);
105
106 #define user_insn(insn, output, input...)                               \
107 ({                                                                      \
108         int err;                                                        \
109                                                                         \
110         might_fault();                                                  \
111                                                                         \
112         asm volatile(ASM_STAC "\n"                                      \
113                      "1:" #insn "\n\t"                                  \
114                      "2: " ASM_CLAC "\n"                                \
115                      ".section .fixup,\"ax\"\n"                         \
116                      "3:  movl $-1,%[err]\n"                            \
117                      "    jmp  2b\n"                                    \
118                      ".previous\n"                                      \
119                      _ASM_EXTABLE(1b, 3b)                               \
120                      : [err] "=r" (err), output                         \
121                      : "0"(0), input);                                  \
122         err;                                                            \
123 })
124
125 #define kernel_insn_err(insn, output, input...)                         \
126 ({                                                                      \
127         int err;                                                        \
128         asm volatile("1:" #insn "\n\t"                                  \
129                      "2:\n"                                             \
130                      ".section .fixup,\"ax\"\n"                         \
131                      "3:  movl $-1,%[err]\n"                            \
132                      "    jmp  2b\n"                                    \
133                      ".previous\n"                                      \
134                      _ASM_EXTABLE(1b, 3b)                               \
135                      : [err] "=r" (err), output                         \
136                      : "0"(0), input);                                  \
137         err;                                                            \
138 })
139
140 #define kernel_insn(insn, output, input...)                             \
141         asm volatile("1:" #insn "\n\t"                                  \
142                      "2:\n"                                             \
143                      _ASM_EXTABLE_HANDLE(1b, 2b, ex_handler_fprestore)  \
144                      : output : input)
145
146 static inline int copy_fregs_to_user(struct fregs_state __user *fx)
147 {
148         return user_insn(fnsave %[fx]; fwait,  [fx] "=m" (*fx), "m" (*fx));
149 }
150
151 static inline int copy_fxregs_to_user(struct fxregs_state __user *fx)
152 {
153         if (IS_ENABLED(CONFIG_X86_32))
154                 return user_insn(fxsave %[fx], [fx] "=m" (*fx), "m" (*fx));
155         else
156                 return user_insn(fxsaveq %[fx], [fx] "=m" (*fx), "m" (*fx));
157
158 }
159
160 static inline void copy_kernel_to_fxregs(struct fxregs_state *fx)
161 {
162         if (IS_ENABLED(CONFIG_X86_32))
163                 kernel_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
164         else
165                 kernel_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
166 }
167
168 static inline int copy_kernel_to_fxregs_err(struct fxregs_state *fx)
169 {
170         if (IS_ENABLED(CONFIG_X86_32))
171                 return kernel_insn_err(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
172         else
173                 return kernel_insn_err(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
174 }
175
176 static inline int copy_user_to_fxregs(struct fxregs_state __user *fx)
177 {
178         if (IS_ENABLED(CONFIG_X86_32))
179                 return user_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
180         else
181                 return user_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
182 }
183
184 static inline void copy_kernel_to_fregs(struct fregs_state *fx)
185 {
186         kernel_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
187 }
188
189 static inline int copy_kernel_to_fregs_err(struct fregs_state *fx)
190 {
191         return kernel_insn_err(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
192 }
193
194 static inline int copy_user_to_fregs(struct fregs_state __user *fx)
195 {
196         return user_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
197 }
198
199 static inline void copy_fxregs_to_kernel(struct fpu *fpu)
200 {
201         if (IS_ENABLED(CONFIG_X86_32))
202                 asm volatile( "fxsave %[fx]" : [fx] "=m" (fpu->state.fxsave));
203         else
204                 asm volatile("fxsaveq %[fx]" : [fx] "=m" (fpu->state.fxsave));
205 }
206
207 /* These macros all use (%edi)/(%rdi) as the single memory argument. */
208 #define XSAVE           ".byte " REX_PREFIX "0x0f,0xae,0x27"
209 #define XSAVEOPT        ".byte " REX_PREFIX "0x0f,0xae,0x37"
210 #define XSAVES          ".byte " REX_PREFIX "0x0f,0xc7,0x2f"
211 #define XRSTOR          ".byte " REX_PREFIX "0x0f,0xae,0x2f"
212 #define XRSTORS         ".byte " REX_PREFIX "0x0f,0xc7,0x1f"
213
214 #define XSTATE_OP(op, st, lmask, hmask, err)                            \
215         asm volatile("1:" op "\n\t"                                     \
216                      "xor %[err], %[err]\n"                             \
217                      "2:\n\t"                                           \
218                      ".pushsection .fixup,\"ax\"\n\t"                   \
219                      "3: movl $-2,%[err]\n\t"                           \
220                      "jmp 2b\n\t"                                       \
221                      ".popsection\n\t"                                  \
222                      _ASM_EXTABLE(1b, 3b)                               \
223                      : [err] "=r" (err)                                 \
224                      : "D" (st), "m" (*st), "a" (lmask), "d" (hmask)    \
225                      : "memory")
226
227 /*
228  * If XSAVES is enabled, it replaces XSAVEOPT because it supports a compact
229  * format and supervisor states in addition to modified optimization in
230  * XSAVEOPT.
231  *
232  * Otherwise, if XSAVEOPT is enabled, XSAVEOPT replaces XSAVE because XSAVEOPT
233  * supports modified optimization which is not supported by XSAVE.
234  *
235  * We use XSAVE as a fallback.
236  *
237  * The 661 label is defined in the ALTERNATIVE* macros as the address of the
238  * original instruction which gets replaced. We need to use it here as the
239  * address of the instruction where we might get an exception at.
240  */
241 #define XSTATE_XSAVE(st, lmask, hmask, err)                             \
242         asm volatile(ALTERNATIVE_2(XSAVE,                               \
243                                    XSAVEOPT, X86_FEATURE_XSAVEOPT,      \
244                                    XSAVES,   X86_FEATURE_XSAVES)        \
245                      "\n"                                               \
246                      "xor %[err], %[err]\n"                             \
247                      "3:\n"                                             \
248                      ".pushsection .fixup,\"ax\"\n"                     \
249                      "4: movl $-2, %[err]\n"                            \
250                      "jmp 3b\n"                                         \
251                      ".popsection\n"                                    \
252                      _ASM_EXTABLE(661b, 4b)                             \
253                      : [err] "=r" (err)                                 \
254                      : "D" (st), "m" (*st), "a" (lmask), "d" (hmask)    \
255                      : "memory")
256
257 /*
258  * Use XRSTORS to restore context if it is enabled. XRSTORS supports compact
259  * XSAVE area format.
260  */
261 #define XSTATE_XRESTORE(st, lmask, hmask)                               \
262         asm volatile(ALTERNATIVE(XRSTOR,                                \
263                                  XRSTORS, X86_FEATURE_XSAVES)           \
264                      "\n"                                               \
265                      "3:\n"                                             \
266                      _ASM_EXTABLE_HANDLE(661b, 3b, ex_handler_fprestore)\
267                      :                                                  \
268                      : "D" (st), "m" (*st), "a" (lmask), "d" (hmask)    \
269                      : "memory")
270
271 /*
272  * This function is called only during boot time when x86 caps are not set
273  * up and alternative can not be used yet.
274  */
275 static inline void copy_xregs_to_kernel_booting(struct xregs_state *xstate)
276 {
277         u64 mask = -1;
278         u32 lmask = mask;
279         u32 hmask = mask >> 32;
280         int err;
281
282         WARN_ON(system_state != SYSTEM_BOOTING);
283
284         if (boot_cpu_has(X86_FEATURE_XSAVES))
285                 XSTATE_OP(XSAVES, xstate, lmask, hmask, err);
286         else
287                 XSTATE_OP(XSAVE, xstate, lmask, hmask, err);
288
289         /* We should never fault when copying to a kernel buffer: */
290         WARN_ON_FPU(err);
291 }
292
293 /*
294  * This function is called only during boot time when x86 caps are not set
295  * up and alternative can not be used yet.
296  */
297 static inline void copy_kernel_to_xregs_booting(struct xregs_state *xstate)
298 {
299         u64 mask = -1;
300         u32 lmask = mask;
301         u32 hmask = mask >> 32;
302         int err;
303
304         WARN_ON(system_state != SYSTEM_BOOTING);
305
306         if (boot_cpu_has(X86_FEATURE_XSAVES))
307                 XSTATE_OP(XRSTORS, xstate, lmask, hmask, err);
308         else
309                 XSTATE_OP(XRSTOR, xstate, lmask, hmask, err);
310
311         /*
312          * We should never fault when copying from a kernel buffer, and the FPU
313          * state we set at boot time should be valid.
314          */
315         WARN_ON_FPU(err);
316 }
317
318 /*
319  * Save processor xstate to xsave area.
320  */
321 static inline void copy_xregs_to_kernel(struct xregs_state *xstate)
322 {
323         u64 mask = -1;
324         u32 lmask = mask;
325         u32 hmask = mask >> 32;
326         int err;
327
328         WARN_ON_FPU(!alternatives_patched);
329
330         XSTATE_XSAVE(xstate, lmask, hmask, err);
331
332         /* We should never fault when copying to a kernel buffer: */
333         WARN_ON_FPU(err);
334 }
335
336 /*
337  * Restore processor xstate from xsave area.
338  */
339 static inline void copy_kernel_to_xregs(struct xregs_state *xstate, u64 mask)
340 {
341         u32 lmask = mask;
342         u32 hmask = mask >> 32;
343
344         XSTATE_XRESTORE(xstate, lmask, hmask);
345 }
346
347 /*
348  * Save xstate to user space xsave area.
349  *
350  * We don't use modified optimization because xrstor/xrstors might track
351  * a different application.
352  *
353  * We don't use compacted format xsave area for
354  * backward compatibility for old applications which don't understand
355  * compacted format of xsave area.
356  */
357 static inline int copy_xregs_to_user(struct xregs_state __user *buf)
358 {
359         int err;
360
361         /*
362          * Clear the xsave header first, so that reserved fields are
363          * initialized to zero.
364          */
365         err = __clear_user(&buf->header, sizeof(buf->header));
366         if (unlikely(err))
367                 return -EFAULT;
368
369         stac();
370         XSTATE_OP(XSAVE, buf, -1, -1, err);
371         clac();
372
373         return err;
374 }
375
376 /*
377  * Restore xstate from user space xsave area.
378  */
379 static inline int copy_user_to_xregs(struct xregs_state __user *buf, u64 mask)
380 {
381         struct xregs_state *xstate = ((__force struct xregs_state *)buf);
382         u32 lmask = mask;
383         u32 hmask = mask >> 32;
384         int err;
385
386         stac();
387         XSTATE_OP(XRSTOR, xstate, lmask, hmask, err);
388         clac();
389
390         return err;
391 }
392
393 /*
394  * Restore xstate from kernel space xsave area, return an error code instead of
395  * an exception.
396  */
397 static inline int copy_kernel_to_xregs_err(struct xregs_state *xstate, u64 mask)
398 {
399         u32 lmask = mask;
400         u32 hmask = mask >> 32;
401         int err;
402
403         if (static_cpu_has(X86_FEATURE_XSAVES))
404                 XSTATE_OP(XRSTORS, xstate, lmask, hmask, err);
405         else
406                 XSTATE_OP(XRSTOR, xstate, lmask, hmask, err);
407
408         return err;
409 }
410
411 /*
412  * These must be called with preempt disabled. Returns
413  * 'true' if the FPU state is still intact and we can
414  * keep registers active.
415  *
416  * The legacy FNSAVE instruction cleared all FPU state
417  * unconditionally, so registers are essentially destroyed.
418  * Modern FPU state can be kept in registers, if there are
419  * no pending FP exceptions.
420  */
421 static inline int copy_fpregs_to_fpstate(struct fpu *fpu)
422 {
423         if (likely(use_xsave())) {
424                 copy_xregs_to_kernel(&fpu->state.xsave);
425
426                 /*
427                  * AVX512 state is tracked here because its use is
428                  * known to slow the max clock speed of the core.
429                  */
430                 if (fpu->state.xsave.header.xfeatures & XFEATURE_MASK_AVX512)
431                         fpu->avx512_timestamp = jiffies;
432                 return 1;
433         }
434
435         if (likely(use_fxsr())) {
436                 copy_fxregs_to_kernel(fpu);
437                 return 1;
438         }
439
440         /*
441          * Legacy FPU register saving, FNSAVE always clears FPU registers,
442          * so we have to mark them inactive:
443          */
444         asm volatile("fnsave %[fp]; fwait" : [fp] "=m" (fpu->state.fsave));
445
446         return 0;
447 }
448
449 static inline void __copy_kernel_to_fpregs(union fpregs_state *fpstate, u64 mask)
450 {
451         if (use_xsave()) {
452                 copy_kernel_to_xregs(&fpstate->xsave, mask);
453         } else {
454                 if (use_fxsr())
455                         copy_kernel_to_fxregs(&fpstate->fxsave);
456                 else
457                         copy_kernel_to_fregs(&fpstate->fsave);
458         }
459 }
460
461 static inline void copy_kernel_to_fpregs(union fpregs_state *fpstate)
462 {
463         /*
464          * AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception is
465          * pending. Clear the x87 state here by setting it to fixed values.
466          * "m" is a random variable that should be in L1.
467          */
468         if (unlikely(static_cpu_has_bug(X86_BUG_FXSAVE_LEAK))) {
469                 asm volatile(
470                         "fnclex\n\t"
471                         "emms\n\t"
472                         "fildl %P[addr]"        /* set F?P to defined value */
473                         : : [addr] "m" (fpstate));
474         }
475
476         __copy_kernel_to_fpregs(fpstate, -1);
477 }
478
479 extern int copy_fpstate_to_sigframe(void __user *buf, void __user *fp, int size);
480
481 /*
482  * FPU context switch related helper methods:
483  */
484
485 DECLARE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
486
487 /*
488  * The in-register FPU state for an FPU context on a CPU is assumed to be
489  * valid if the fpu->last_cpu matches the CPU, and the fpu_fpregs_owner_ctx
490  * matches the FPU.
491  *
492  * If the FPU register state is valid, the kernel can skip restoring the
493  * FPU state from memory.
494  *
495  * Any code that clobbers the FPU registers or updates the in-memory
496  * FPU state for a task MUST let the rest of the kernel know that the
497  * FPU registers are no longer valid for this task.
498  *
499  * Either one of these invalidation functions is enough. Invalidate
500  * a resource you control: CPU if using the CPU for something else
501  * (with preemption disabled), FPU for the current task, or a task that
502  * is prevented from running by the current task.
503  */
504 static inline void __cpu_invalidate_fpregs_state(void)
505 {
506         __this_cpu_write(fpu_fpregs_owner_ctx, NULL);
507 }
508
509 static inline void __fpu_invalidate_fpregs_state(struct fpu *fpu)
510 {
511         fpu->last_cpu = -1;
512 }
513
514 static inline int fpregs_state_valid(struct fpu *fpu, unsigned int cpu)
515 {
516         return fpu == this_cpu_read(fpu_fpregs_owner_ctx) && cpu == fpu->last_cpu;
517 }
518
519 /*
520  * These generally need preemption protection to work,
521  * do try to avoid using these on their own:
522  */
523 static inline void fpregs_deactivate(struct fpu *fpu)
524 {
525         this_cpu_write(fpu_fpregs_owner_ctx, NULL);
526         trace_x86_fpu_regs_deactivated(fpu);
527 }
528
529 static inline void fpregs_activate(struct fpu *fpu)
530 {
531         this_cpu_write(fpu_fpregs_owner_ctx, fpu);
532         trace_x86_fpu_regs_activated(fpu);
533 }
534
535 /*
536  * Internal helper, do not use directly. Use switch_fpu_return() instead.
537  */
538 static inline void __fpregs_load_activate(void)
539 {
540         struct fpu *fpu = &current->thread.fpu;
541         int cpu = smp_processor_id();
542
543         if (WARN_ON_ONCE(current->flags & PF_KTHREAD))
544                 return;
545
546         if (!fpregs_state_valid(fpu, cpu)) {
547                 copy_kernel_to_fpregs(&fpu->state);
548                 fpregs_activate(fpu);
549                 fpu->last_cpu = cpu;
550         }
551         clear_thread_flag(TIF_NEED_FPU_LOAD);
552 }
553
554 /*
555  * FPU state switching for scheduling.
556  *
557  * This is a two-stage process:
558  *
559  *  - switch_fpu_prepare() saves the old state.
560  *    This is done within the context of the old process.
561  *
562  *  - switch_fpu_finish() sets TIF_NEED_FPU_LOAD; the floating point state
563  *    will get loaded on return to userspace, or when the kernel needs it.
564  *
565  * If TIF_NEED_FPU_LOAD is cleared then the CPU's FPU registers
566  * are saved in the current thread's FPU register state.
567  *
568  * If TIF_NEED_FPU_LOAD is set then CPU's FPU registers may not
569  * hold current()'s FPU registers. It is required to load the
570  * registers before returning to userland or using the content
571  * otherwise.
572  *
573  * The FPU context is only stored/restored for a user task and
574  * PF_KTHREAD is used to distinguish between kernel and user threads.
575  */
576 static inline void switch_fpu_prepare(struct fpu *old_fpu, int cpu)
577 {
578         if (static_cpu_has(X86_FEATURE_FPU) && !(current->flags & PF_KTHREAD)) {
579                 if (!copy_fpregs_to_fpstate(old_fpu))
580                         old_fpu->last_cpu = -1;
581                 else
582                         old_fpu->last_cpu = cpu;
583
584                 /* But leave fpu_fpregs_owner_ctx! */
585                 trace_x86_fpu_regs_deactivated(old_fpu);
586         }
587 }
588
589 /*
590  * Misc helper functions:
591  */
592
593 /*
594  * Load PKRU from the FPU context if available. Delay loading of the
595  * complete FPU state until the return to userland.
596  */
597 static inline void switch_fpu_finish(struct fpu *new_fpu)
598 {
599         u32 pkru_val = init_pkru_value;
600         struct pkru_state *pk;
601
602         if (!static_cpu_has(X86_FEATURE_FPU))
603                 return;
604
605         set_thread_flag(TIF_NEED_FPU_LOAD);
606
607         if (!cpu_feature_enabled(X86_FEATURE_OSPKE))
608                 return;
609
610         /*
611          * PKRU state is switched eagerly because it needs to be valid before we
612          * return to userland e.g. for a copy_to_user() operation.
613          */
614         if (current->mm) {
615                 pk = get_xsave_addr(&new_fpu->state.xsave, XFEATURE_PKRU);
616                 if (pk)
617                         pkru_val = pk->pkru;
618         }
619         __write_pkru(pkru_val);
620 }
621
622 /*
623  * MXCSR and XCR definitions:
624  */
625
626 extern unsigned int mxcsr_feature_mask;
627
628 #define XCR_XFEATURE_ENABLED_MASK       0x00000000
629
630 static inline u64 xgetbv(u32 index)
631 {
632         u32 eax, edx;
633
634         asm volatile(".byte 0x0f,0x01,0xd0" /* xgetbv */
635                      : "=a" (eax), "=d" (edx)
636                      : "c" (index));
637         return eax + ((u64)edx << 32);
638 }
639
640 static inline void xsetbv(u32 index, u64 value)
641 {
642         u32 eax = value;
643         u32 edx = value >> 32;
644
645         asm volatile(".byte 0x0f,0x01,0xd1" /* xsetbv */
646                      : : "a" (eax), "d" (edx), "c" (index));
647 }
648
649 #endif /* _ASM_X86_FPU_INTERNAL_H */