• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 1994 Linus Torvalds
3  *
4  * Pentium III FXSR, SSE support
5  * General FPU state handling cleanups
6  *	Gareth Hughes <gareth@valinux.com>, May 2000
7  * x86-64 work by Andi Kleen 2002
8  */
9 
10 #ifndef _ASM_X86_FPU_INTERNAL_H
11 #define _ASM_X86_FPU_INTERNAL_H
12 
13 #include <linux/compat.h>
14 #include <linux/sched.h>
15 #include <linux/slab.h>
16 
17 #include <asm/user.h>
18 #include <asm/fpu/api.h>
19 #include <asm/fpu/xstate.h>
20 #include <asm/cpufeature.h>
21 
22 /*
23  * High level FPU state handling functions:
24  */
25 extern void fpu__activate_curr(struct fpu *fpu);
26 extern void fpu__activate_fpstate_read(struct fpu *fpu);
27 extern void fpu__activate_fpstate_write(struct fpu *fpu);
28 extern void fpu__save(struct fpu *fpu);
29 extern void fpu__restore(struct fpu *fpu);
30 extern int  fpu__restore_sig(void __user *buf, int ia32_frame);
31 extern void fpu__drop(struct fpu *fpu);
32 extern int  fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu);
33 extern void fpu__clear(struct fpu *fpu);
34 extern int  fpu__exception_code(struct fpu *fpu, int trap_nr);
35 extern int  dump_fpu(struct pt_regs *ptregs, struct user_i387_struct *fpstate);
36 
37 /*
38  * Boot time FPU initialization functions:
39  */
40 extern void fpu__init_cpu(void);
41 extern void fpu__init_system_xstate(void);
42 extern void fpu__init_cpu_xstate(void);
43 extern void fpu__init_system(struct cpuinfo_x86 *c);
44 extern void fpu__init_check_bugs(void);
45 extern void fpu__resume_cpu(void);
46 extern u64 fpu__get_supported_xfeatures_mask(void);
47 
48 /*
49  * Debugging facility:
50  */
51 #ifdef CONFIG_X86_DEBUG_FPU
52 # define WARN_ON_FPU(x) WARN_ON_ONCE(x)
53 #else
54 # define WARN_ON_FPU(x) ({ (void)(x); 0; })
55 #endif
56 
57 /*
58  * FPU related CPU feature flag helper routines:
59  */
use_xsaveopt(void)60 static __always_inline __pure bool use_xsaveopt(void)
61 {
62 	return static_cpu_has(X86_FEATURE_XSAVEOPT);
63 }
64 
use_xsave(void)65 static __always_inline __pure bool use_xsave(void)
66 {
67 	return static_cpu_has(X86_FEATURE_XSAVE);
68 }
69 
use_fxsr(void)70 static __always_inline __pure bool use_fxsr(void)
71 {
72 	return static_cpu_has(X86_FEATURE_FXSR);
73 }
74 
75 /*
76  * fpstate handling functions:
77  */
78 
79 extern union fpregs_state init_fpstate;
80 
81 extern void fpstate_init(union fpregs_state *state);
82 #ifdef CONFIG_MATH_EMULATION
83 extern void fpstate_init_soft(struct swregs_state *soft);
84 #else
fpstate_init_soft(struct swregs_state * soft)85 static inline void fpstate_init_soft(struct swregs_state *soft) {}
86 #endif
fpstate_init_fxstate(struct fxregs_state * fx)87 static inline void fpstate_init_fxstate(struct fxregs_state *fx)
88 {
89 	fx->cwd = 0x37f;
90 	fx->mxcsr = MXCSR_DEFAULT;
91 }
92 extern void fpstate_sanitize_xstate(struct fpu *fpu);
93 
94 #define user_insn(insn, output, input...)				\
95 ({									\
96 	int err;							\
97 									\
98 	might_fault();							\
99 									\
100 	asm volatile(ASM_STAC "\n"					\
101 		     "1:" #insn "\n\t"					\
102 		     "2: " ASM_CLAC "\n"				\
103 		     ".section .fixup,\"ax\"\n"				\
104 		     "3:  movl $-1,%[err]\n"				\
105 		     "    jmp  2b\n"					\
106 		     ".previous\n"					\
107 		     _ASM_EXTABLE(1b, 3b)				\
108 		     : [err] "=r" (err), output				\
109 		     : "0"(0), input);					\
110 	err;								\
111 })
112 
113 #define check_insn(insn, output, input...)				\
114 ({									\
115 	int err;							\
116 	asm volatile("1:" #insn "\n\t"					\
117 		     "2:\n"						\
118 		     ".section .fixup,\"ax\"\n"				\
119 		     "3:  movl $-1,%[err]\n"				\
120 		     "    jmp  2b\n"					\
121 		     ".previous\n"					\
122 		     _ASM_EXTABLE(1b, 3b)				\
123 		     : [err] "=r" (err), output				\
124 		     : "0"(0), input);					\
125 	err;								\
126 })
127 
copy_fregs_to_user(struct fregs_state __user * fx)128 static inline int copy_fregs_to_user(struct fregs_state __user *fx)
129 {
130 	return user_insn(fnsave %[fx]; fwait,  [fx] "=m" (*fx), "m" (*fx));
131 }
132 
copy_fxregs_to_user(struct fxregs_state __user * fx)133 static inline int copy_fxregs_to_user(struct fxregs_state __user *fx)
134 {
135 	if (config_enabled(CONFIG_X86_32))
136 		return user_insn(fxsave %[fx], [fx] "=m" (*fx), "m" (*fx));
137 	else if (config_enabled(CONFIG_AS_FXSAVEQ))
138 		return user_insn(fxsaveq %[fx], [fx] "=m" (*fx), "m" (*fx));
139 
140 	/* See comment in copy_fxregs_to_kernel() below. */
141 	return user_insn(rex64/fxsave (%[fx]), "=m" (*fx), [fx] "R" (fx));
142 }
143 
copy_kernel_to_fxregs(struct fxregs_state * fx)144 static inline void copy_kernel_to_fxregs(struct fxregs_state *fx)
145 {
146 	int err;
147 
148 	if (config_enabled(CONFIG_X86_32)) {
149 		err = check_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
150 	} else {
151 		if (config_enabled(CONFIG_AS_FXSAVEQ)) {
152 			err = check_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
153 		} else {
154 			/* See comment in copy_fxregs_to_kernel() below. */
155 			err = check_insn(rex64/fxrstor (%[fx]), "=m" (*fx), [fx] "R" (fx), "m" (*fx));
156 		}
157 	}
158 	/* Copying from a kernel buffer to FPU registers should never fail: */
159 	WARN_ON_FPU(err);
160 }
161 
copy_user_to_fxregs(struct fxregs_state __user * fx)162 static inline int copy_user_to_fxregs(struct fxregs_state __user *fx)
163 {
164 	if (config_enabled(CONFIG_X86_32))
165 		return user_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
166 	else if (config_enabled(CONFIG_AS_FXSAVEQ))
167 		return user_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
168 
169 	/* See comment in copy_fxregs_to_kernel() below. */
170 	return user_insn(rex64/fxrstor (%[fx]), "=m" (*fx), [fx] "R" (fx),
171 			  "m" (*fx));
172 }
173 
copy_kernel_to_fregs(struct fregs_state * fx)174 static inline void copy_kernel_to_fregs(struct fregs_state *fx)
175 {
176 	int err = check_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
177 
178 	WARN_ON_FPU(err);
179 }
180 
copy_user_to_fregs(struct fregs_state __user * fx)181 static inline int copy_user_to_fregs(struct fregs_state __user *fx)
182 {
183 	return user_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
184 }
185 
copy_fxregs_to_kernel(struct fpu * fpu)186 static inline void copy_fxregs_to_kernel(struct fpu *fpu)
187 {
188 	if (config_enabled(CONFIG_X86_32))
189 		asm volatile( "fxsave %[fx]" : [fx] "=m" (fpu->state.fxsave));
190 	else if (config_enabled(CONFIG_AS_FXSAVEQ))
191 		asm volatile("fxsaveq %[fx]" : [fx] "=m" (fpu->state.fxsave));
192 	else {
193 		/* Using "rex64; fxsave %0" is broken because, if the memory
194 		 * operand uses any extended registers for addressing, a second
195 		 * REX prefix will be generated (to the assembler, rex64
196 		 * followed by semicolon is a separate instruction), and hence
197 		 * the 64-bitness is lost.
198 		 *
199 		 * Using "fxsaveq %0" would be the ideal choice, but is only
200 		 * supported starting with gas 2.16.
201 		 *
202 		 * Using, as a workaround, the properly prefixed form below
203 		 * isn't accepted by any binutils version so far released,
204 		 * complaining that the same type of prefix is used twice if
205 		 * an extended register is needed for addressing (fix submitted
206 		 * to mainline 2005-11-21).
207 		 *
208 		 *  asm volatile("rex64/fxsave %0" : "=m" (fpu->state.fxsave));
209 		 *
210 		 * This, however, we can work around by forcing the compiler to
211 		 * select an addressing mode that doesn't require extended
212 		 * registers.
213 		 */
214 		asm volatile( "rex64/fxsave (%[fx])"
215 			     : "=m" (fpu->state.fxsave)
216 			     : [fx] "R" (&fpu->state.fxsave));
217 	}
218 }
219 
fxsave(struct fxregs_state * fx)220 static inline void fxsave(struct fxregs_state *fx)
221 {
222 	if (IS_ENABLED(CONFIG_X86_32))
223 		asm volatile( "fxsave %[fx]" : [fx] "=m" (*fx));
224 	else
225 		asm volatile("fxsaveq %[fx]" : [fx] "=m" (*fx));
226 }
227 
228 /* These macros all use (%edi)/(%rdi) as the single memory argument. */
229 #define XSAVE		".byte " REX_PREFIX "0x0f,0xae,0x27"
230 #define XSAVEOPT	".byte " REX_PREFIX "0x0f,0xae,0x37"
231 #define XSAVES		".byte " REX_PREFIX "0x0f,0xc7,0x2f"
232 #define XRSTOR		".byte " REX_PREFIX "0x0f,0xae,0x2f"
233 #define XRSTORS		".byte " REX_PREFIX "0x0f,0xc7,0x1f"
234 
235 #define XSTATE_OP(op, st, lmask, hmask, err)				\
236 	asm volatile("1:" op "\n\t"					\
237 		     "xor %[err], %[err]\n"				\
238 		     "2:\n\t"						\
239 		     ".pushsection .fixup,\"ax\"\n\t"			\
240 		     "3: movl $-2,%[err]\n\t"				\
241 		     "jmp 2b\n\t"					\
242 		     ".popsection\n\t"					\
243 		     _ASM_EXTABLE(1b, 3b)				\
244 		     : [err] "=r" (err)					\
245 		     : "D" (st), "m" (*st), "a" (lmask), "d" (hmask)	\
246 		     : "memory")
247 
248 /*
249  * If XSAVES is enabled, it replaces XSAVEOPT because it supports a compact
250  * format and supervisor states in addition to modified optimization in
251  * XSAVEOPT.
252  *
253  * Otherwise, if XSAVEOPT is enabled, XSAVEOPT replaces XSAVE because XSAVEOPT
254  * supports modified optimization which is not supported by XSAVE.
255  *
256  * We use XSAVE as a fallback.
257  *
258  * The 661 label is defined in the ALTERNATIVE* macros as the address of the
259  * original instruction which gets replaced. We need to use it here as the
260  * address of the instruction where we might get an exception at.
261  */
262 #define XSTATE_XSAVE(st, lmask, hmask, err)				\
263 	asm volatile(ALTERNATIVE_2(XSAVE,				\
264 				   XSAVEOPT, X86_FEATURE_XSAVEOPT,	\
265 				   XSAVES,   X86_FEATURE_XSAVES)	\
266 		     "\n"						\
267 		     "xor %[err], %[err]\n"				\
268 		     "3:\n"						\
269 		     ".pushsection .fixup,\"ax\"\n"			\
270 		     "4: movl $-2, %[err]\n"				\
271 		     "jmp 3b\n"						\
272 		     ".popsection\n"					\
273 		     _ASM_EXTABLE(661b, 4b)				\
274 		     : [err] "=r" (err)					\
275 		     : "D" (st), "m" (*st), "a" (lmask), "d" (hmask)	\
276 		     : "memory")
277 
278 /*
279  * Use XRSTORS to restore context if it is enabled. XRSTORS supports compact
280  * XSAVE area format.
281  */
282 #define XSTATE_XRESTORE(st, lmask, hmask, err)				\
283 	asm volatile(ALTERNATIVE(XRSTOR,				\
284 				 XRSTORS, X86_FEATURE_XSAVES)		\
285 		     "\n"						\
286 		     "xor %[err], %[err]\n"				\
287 		     "3:\n"						\
288 		     ".pushsection .fixup,\"ax\"\n"			\
289 		     "4: movl $-2, %[err]\n"				\
290 		     "jmp 3b\n"						\
291 		     ".popsection\n"					\
292 		     _ASM_EXTABLE(661b, 4b)				\
293 		     : [err] "=r" (err)					\
294 		     : "D" (st), "m" (*st), "a" (lmask), "d" (hmask)	\
295 		     : "memory")
296 
297 /*
298  * This function is called only during boot time when x86 caps are not set
299  * up and alternative can not be used yet.
300  */
copy_kernel_to_xregs_booting(struct xregs_state * xstate)301 static inline void copy_kernel_to_xregs_booting(struct xregs_state *xstate)
302 {
303 	u64 mask = -1;
304 	u32 lmask = mask;
305 	u32 hmask = mask >> 32;
306 	int err;
307 
308 	WARN_ON(system_state != SYSTEM_BOOTING);
309 
310 	if (static_cpu_has(X86_FEATURE_XSAVES))
311 		XSTATE_OP(XRSTORS, xstate, lmask, hmask, err);
312 	else
313 		XSTATE_OP(XRSTOR, xstate, lmask, hmask, err);
314 
315 	/* We should never fault when copying from a kernel buffer: */
316 	WARN_ON_FPU(err);
317 }
318 
319 /*
320  * Save processor xstate to xsave area.
321  */
copy_xregs_to_kernel(struct xregs_state * xstate)322 static inline void copy_xregs_to_kernel(struct xregs_state *xstate)
323 {
324 	u64 mask = -1;
325 	u32 lmask = mask;
326 	u32 hmask = mask >> 32;
327 	int err;
328 
329 	WARN_ON(!alternatives_patched);
330 
331 	XSTATE_XSAVE(xstate, lmask, hmask, err);
332 
333 	/* We should never fault when copying to a kernel buffer: */
334 	WARN_ON_FPU(err);
335 }
336 
337 /*
338  * Restore processor xstate from xsave area.
339  */
copy_kernel_to_xregs(struct xregs_state * xstate,u64 mask)340 static inline void copy_kernel_to_xregs(struct xregs_state *xstate, u64 mask)
341 {
342 	u32 lmask = mask;
343 	u32 hmask = mask >> 32;
344 	int err;
345 
346 	XSTATE_XRESTORE(xstate, lmask, hmask, err);
347 
348 	/* We should never fault when copying from a kernel buffer: */
349 	WARN_ON_FPU(err);
350 }
351 
352 /*
353  * Save xstate to user space xsave area.
354  *
355  * We don't use modified optimization because xrstor/xrstors might track
356  * a different application.
357  *
358  * We don't use compacted format xsave area for
359  * backward compatibility for old applications which don't understand
360  * compacted format of xsave area.
361  */
copy_xregs_to_user(struct xregs_state __user * buf)362 static inline int copy_xregs_to_user(struct xregs_state __user *buf)
363 {
364 	int err;
365 
366 	/*
367 	 * Clear the xsave header first, so that reserved fields are
368 	 * initialized to zero.
369 	 */
370 	err = __clear_user(&buf->header, sizeof(buf->header));
371 	if (unlikely(err))
372 		return -EFAULT;
373 
374 	stac();
375 	XSTATE_OP(XSAVE, buf, -1, -1, err);
376 	clac();
377 
378 	return err;
379 }
380 
381 /*
382  * Restore xstate from user space xsave area.
383  */
copy_user_to_xregs(struct xregs_state __user * buf,u64 mask)384 static inline int copy_user_to_xregs(struct xregs_state __user *buf, u64 mask)
385 {
386 	struct xregs_state *xstate = ((__force struct xregs_state *)buf);
387 	u32 lmask = mask;
388 	u32 hmask = mask >> 32;
389 	int err;
390 
391 	stac();
392 	XSTATE_OP(XRSTOR, xstate, lmask, hmask, err);
393 	clac();
394 
395 	return err;
396 }
397 
398 /*
399  * These must be called with preempt disabled. Returns
400  * 'true' if the FPU state is still intact and we can
401  * keep registers active.
402  *
403  * The legacy FNSAVE instruction cleared all FPU state
404  * unconditionally, so registers are essentially destroyed.
405  * Modern FPU state can be kept in registers, if there are
406  * no pending FP exceptions.
407  */
copy_fpregs_to_fpstate(struct fpu * fpu)408 static inline int copy_fpregs_to_fpstate(struct fpu *fpu)
409 {
410 	if (likely(use_xsave())) {
411 		copy_xregs_to_kernel(&fpu->state.xsave);
412 		return 1;
413 	}
414 
415 	if (likely(use_fxsr())) {
416 		copy_fxregs_to_kernel(fpu);
417 		return 1;
418 	}
419 
420 	/*
421 	 * Legacy FPU register saving, FNSAVE always clears FPU registers,
422 	 * so we have to mark them inactive:
423 	 */
424 	asm volatile("fnsave %[fp]; fwait" : [fp] "=m" (fpu->state.fsave));
425 
426 	return 0;
427 }
428 
__copy_kernel_to_fpregs(union fpregs_state * fpstate)429 static inline void __copy_kernel_to_fpregs(union fpregs_state *fpstate)
430 {
431 	if (use_xsave()) {
432 		copy_kernel_to_xregs(&fpstate->xsave, -1);
433 	} else {
434 		if (use_fxsr())
435 			copy_kernel_to_fxregs(&fpstate->fxsave);
436 		else
437 			copy_kernel_to_fregs(&fpstate->fsave);
438 	}
439 }
440 
copy_kernel_to_fpregs(union fpregs_state * fpstate)441 static inline void copy_kernel_to_fpregs(union fpregs_state *fpstate)
442 {
443 	/*
444 	 * AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception is
445 	 * pending. Clear the x87 state here by setting it to fixed values.
446 	 * "m" is a random variable that should be in L1.
447 	 */
448 	if (unlikely(static_cpu_has_bug(X86_BUG_FXSAVE_LEAK))) {
449 		asm volatile(
450 			"fnclex\n\t"
451 			"emms\n\t"
452 			"fildl %P[addr]"	/* set F?P to defined value */
453 			: : [addr] "m" (fpstate));
454 	}
455 
456 	__copy_kernel_to_fpregs(fpstate);
457 }
458 
459 extern int copy_fpstate_to_sigframe(void __user *buf, void __user *fp, int size);
460 
461 /*
462  * FPU context switch related helper methods:
463  */
464 
465 DECLARE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
466 
467 /*
468  * Must be run with preemption disabled: this clears the fpu_fpregs_owner_ctx,
469  * on this CPU.
470  *
471  * This will disable any lazy FPU state restore of the current FPU state,
472  * but if the current thread owns the FPU, it will still be saved by.
473  */
__cpu_disable_lazy_restore(unsigned int cpu)474 static inline void __cpu_disable_lazy_restore(unsigned int cpu)
475 {
476 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
477 }
478 
fpu_want_lazy_restore(struct fpu * fpu,unsigned int cpu)479 static inline int fpu_want_lazy_restore(struct fpu *fpu, unsigned int cpu)
480 {
481 	return fpu == this_cpu_read_stable(fpu_fpregs_owner_ctx) && cpu == fpu->last_cpu;
482 }
483 
484 
__fpregs_deactivate(struct fpu * fpu)485 static inline void __fpregs_deactivate(struct fpu *fpu)
486 {
487 	WARN_ON_FPU(!fpu->fpregs_active);
488 
489 	fpu->fpregs_active = 0;
490 	this_cpu_write(fpu_fpregs_owner_ctx, NULL);
491 }
492 
__fpregs_activate(struct fpu * fpu)493 static inline void __fpregs_activate(struct fpu *fpu)
494 {
495 	WARN_ON_FPU(fpu->fpregs_active);
496 
497 	fpu->fpregs_active = 1;
498 	this_cpu_write(fpu_fpregs_owner_ctx, fpu);
499 }
500 
501 /*
502  * The question "does this thread have fpu access?"
503  * is slightly racy, since preemption could come in
504  * and revoke it immediately after the test.
505  *
506  * However, even in that very unlikely scenario,
507  * we can just assume we have FPU access - typically
508  * to save the FP state - we'll just take a #NM
509  * fault and get the FPU access back.
510  */
fpregs_active(void)511 static inline int fpregs_active(void)
512 {
513 	return current->thread.fpu.fpregs_active;
514 }
515 
516 /*
517  * These generally need preemption protection to work,
518  * do try to avoid using these on their own.
519  */
fpregs_activate(struct fpu * fpu)520 static inline void fpregs_activate(struct fpu *fpu)
521 {
522 	__fpregs_activate(fpu);
523 }
524 
fpregs_deactivate(struct fpu * fpu)525 static inline void fpregs_deactivate(struct fpu *fpu)
526 {
527 	__fpregs_deactivate(fpu);
528 }
529 
530 /*
531  * FPU state switching for scheduling.
532  *
533  * This is a two-stage process:
534  *
535  *  - switch_fpu_prepare() saves the old state and
536  *    sets the new state of the CR0.TS bit. This is
537  *    done within the context of the old process.
538  *
539  *  - switch_fpu_finish() restores the new state as
540  *    necessary.
541  */
542 typedef struct { int preload; } fpu_switch_t;
543 
544 static inline fpu_switch_t
switch_fpu_prepare(struct fpu * old_fpu,struct fpu * new_fpu,int cpu)545 switch_fpu_prepare(struct fpu *old_fpu, struct fpu *new_fpu, int cpu)
546 {
547 	fpu_switch_t fpu;
548 
549 	/*
550 	 * If the task has used the math, pre-load the FPU on xsave processors
551 	 * or if the past 5 consecutive context-switches used math.
552 	 */
553 	fpu.preload = static_cpu_has(X86_FEATURE_FPU) &&
554 		      new_fpu->fpstate_active;
555 
556 	if (old_fpu->fpregs_active) {
557 		if (!copy_fpregs_to_fpstate(old_fpu))
558 			old_fpu->last_cpu = -1;
559 		else
560 			old_fpu->last_cpu = cpu;
561 
562 		/* But leave fpu_fpregs_owner_ctx! */
563 		old_fpu->fpregs_active = 0;
564 
565 		/* Don't change CR0.TS if we just switch! */
566 		if (fpu.preload) {
567 			__fpregs_activate(new_fpu);
568 			prefetch(&new_fpu->state);
569 		}
570 	} else {
571 		old_fpu->last_cpu = -1;
572 		if (fpu.preload) {
573 			if (fpu_want_lazy_restore(new_fpu, cpu))
574 				fpu.preload = 0;
575 			else
576 				prefetch(&new_fpu->state);
577 			fpregs_activate(new_fpu);
578 		}
579 	}
580 	return fpu;
581 }
582 
583 /*
584  * Misc helper functions:
585  */
586 
587 /*
588  * By the time this gets called, we've already cleared CR0.TS and
589  * given the process the FPU if we are going to preload the FPU
590  * state - all we need to do is to conditionally restore the register
591  * state itself.
592  */
switch_fpu_finish(struct fpu * new_fpu,fpu_switch_t fpu_switch)593 static inline void switch_fpu_finish(struct fpu *new_fpu, fpu_switch_t fpu_switch)
594 {
595 	if (fpu_switch.preload)
596 		copy_kernel_to_fpregs(&new_fpu->state);
597 }
598 
599 /*
600  * Needs to be preemption-safe.
601  *
602  * NOTE! user_fpu_begin() must be used only immediately before restoring
603  * the save state. It does not do any saving/restoring on its own. In
604  * lazy FPU mode, it is just an optimization to avoid a #NM exception,
605  * the task can lose the FPU right after preempt_enable().
606  */
user_fpu_begin(void)607 static inline void user_fpu_begin(void)
608 {
609 	struct fpu *fpu = &current->thread.fpu;
610 
611 	preempt_disable();
612 	if (!fpregs_active())
613 		fpregs_activate(fpu);
614 	preempt_enable();
615 }
616 
617 /*
618  * MXCSR and XCR definitions:
619  */
620 
621 extern unsigned int mxcsr_feature_mask;
622 
623 #define XCR_XFEATURE_ENABLED_MASK	0x00000000
624 
xgetbv(u32 index)625 static inline u64 xgetbv(u32 index)
626 {
627 	u32 eax, edx;
628 
629 	asm volatile(".byte 0x0f,0x01,0xd0" /* xgetbv */
630 		     : "=a" (eax), "=d" (edx)
631 		     : "c" (index));
632 	return eax + ((u64)edx << 32);
633 }
634 
xsetbv(u32 index,u64 value)635 static inline void xsetbv(u32 index, u64 value)
636 {
637 	u32 eax = value;
638 	u32 edx = value >> 32;
639 
640 	asm volatile(".byte 0x0f,0x01,0xd1" /* xsetbv */
641 		     : : "a" (eax), "d" (edx), "c" (index));
642 }
643 
644 #endif /* _ASM_X86_FPU_INTERNAL_H */
645