2 * Copyright (C) 1994 Linus Torvalds
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
10 #ifndef _FPU_INTERNAL_H
11 #define _FPU_INTERNAL_H
13 #include <linux/kernel_stat.h>
14 #include <linux/regset.h>
15 #include <linux/compat.h>
16 #include <linux/slab.h>
18 #include <asm/cpufeature.h>
19 #include <asm/processor.h>
20 #include <asm/sigcontext.h>
22 #include <asm/uaccess.h>
23 #include <asm/xsave.h>
27 # include <asm/sigcontext32.h>
28 # include <asm/user32.h>
30 int ia32_setup_rt_frame(int sig
, struct ksignal
*ksig
,
31 compat_sigset_t
*set
, struct pt_regs
*regs
);
32 int ia32_setup_frame(int sig
, struct ksignal
*ksig
,
33 compat_sigset_t
*set
, struct pt_regs
*regs
);
35 # define user_i387_ia32_struct user_i387_struct
36 # define user32_fxsr_struct user_fxsr_struct
37 # define ia32_setup_frame __setup_frame
38 # define ia32_setup_rt_frame __setup_rt_frame
41 extern unsigned int mxcsr_feature_mask
;
42 extern void fpu_init(void);
43 extern void eager_fpu_init(void);
45 DECLARE_PER_CPU(struct task_struct
*, fpu_owner_task
);
47 extern void convert_from_fxsr(struct user_i387_ia32_struct
*env
,
48 struct task_struct
*tsk
);
49 extern void convert_to_fxsr(struct task_struct
*tsk
,
50 const struct user_i387_ia32_struct
*env
);
52 extern user_regset_active_fn fpregs_active
, xfpregs_active
;
53 extern user_regset_get_fn fpregs_get
, xfpregs_get
, fpregs_soft_get
,
55 extern user_regset_set_fn fpregs_set
, xfpregs_set
, fpregs_soft_set
,
59 * xstateregs_active == fpregs_active. Please refer to the comment
60 * at the definition of fpregs_active.
62 #define xstateregs_active fpregs_active
64 #ifdef CONFIG_MATH_EMULATION
65 extern void finit_soft_fpu(struct i387_soft_struct
*soft
);
67 static inline void finit_soft_fpu(struct i387_soft_struct
*soft
) {}
71 * Must be run with preemption disabled: this clears the fpu_owner_task,
74 * This will disable any lazy FPU state restore of the current FPU state,
75 * but if the current thread owns the FPU, it will still be saved by.
77 static inline void __cpu_disable_lazy_restore(unsigned int cpu
)
79 per_cpu(fpu_owner_task
, cpu
) = NULL
;
83 * Used to indicate that the FPU state in memory is newer than the FPU
84 * state in registers, and the FPU state should be reloaded next time the
85 * task is run. Only safe on the current task, or non-running tasks.
87 static inline void task_disable_lazy_fpu_restore(struct task_struct
*tsk
)
89 tsk
->thread
.fpu
.last_cpu
= ~0;
92 static inline int fpu_lazy_restore(struct task_struct
*new, unsigned int cpu
)
94 return new == this_cpu_read_stable(fpu_owner_task
) &&
95 cpu
== new->thread
.fpu
.last_cpu
;
98 static inline int is_ia32_compat_frame(void)
100 return config_enabled(CONFIG_IA32_EMULATION
) &&
101 test_thread_flag(TIF_IA32
);
104 static inline int is_ia32_frame(void)
106 return config_enabled(CONFIG_X86_32
) || is_ia32_compat_frame();
109 static inline int is_x32_frame(void)
111 return config_enabled(CONFIG_X86_X32_ABI
) && test_thread_flag(TIF_X32
);
114 #define X87_FSW_ES (1 << 7) /* Exception Summary */
116 static __always_inline __pure
bool use_eager_fpu(void)
118 return static_cpu_has_safe(X86_FEATURE_EAGER_FPU
);
121 static __always_inline __pure
bool use_xsaveopt(void)
123 return static_cpu_has_safe(X86_FEATURE_XSAVEOPT
);
126 static __always_inline __pure
bool use_xsave(void)
128 return static_cpu_has_safe(X86_FEATURE_XSAVE
);
131 static __always_inline __pure
bool use_fxsr(void)
133 return static_cpu_has_safe(X86_FEATURE_FXSR
);
136 static inline void fx_finit(struct i387_fxsave_struct
*fx
)
139 fx
->mxcsr
= MXCSR_DEFAULT
;
142 extern void __sanitize_i387_state(struct task_struct
*);
144 static inline void sanitize_i387_state(struct task_struct
*tsk
)
148 __sanitize_i387_state(tsk
);
151 #define user_insn(insn, output, input...) \
154 asm volatile(ASM_STAC "\n" \
156 "2: " ASM_CLAC "\n" \
157 ".section .fixup,\"ax\"\n" \
158 "3: movl $-1,%[err]\n" \
161 _ASM_EXTABLE(1b, 3b) \
162 : [err] "=r" (err), output \
167 #define check_insn(insn, output, input...) \
170 asm volatile("1:" #insn "\n\t" \
172 ".section .fixup,\"ax\"\n" \
173 "3: movl $-1,%[err]\n" \
176 _ASM_EXTABLE(1b, 3b) \
177 : [err] "=r" (err), output \
182 static inline int fsave_user(struct i387_fsave_struct __user
*fx
)
184 return user_insn(fnsave
%[fx
]; fwait
, [fx
] "=m" (*fx
), "m" (*fx
));
187 static inline int fxsave_user(struct i387_fxsave_struct __user
*fx
)
189 if (config_enabled(CONFIG_X86_32
))
190 return user_insn(fxsave
%[fx
], [fx
] "=m" (*fx
), "m" (*fx
));
191 else if (config_enabled(CONFIG_AS_FXSAVEQ
))
192 return user_insn(fxsaveq
%[fx
], [fx
] "=m" (*fx
), "m" (*fx
));
194 /* See comment in fpu_fxsave() below. */
195 return user_insn(rex64
/fxsave (%[fx
]), "=m" (*fx
), [fx
] "R" (fx
));
198 static inline int fxrstor_checking(struct i387_fxsave_struct
*fx
)
200 if (config_enabled(CONFIG_X86_32
))
201 return check_insn(fxrstor
%[fx
], "=m" (*fx
), [fx
] "m" (*fx
));
202 else if (config_enabled(CONFIG_AS_FXSAVEQ
))
203 return check_insn(fxrstorq
%[fx
], "=m" (*fx
), [fx
] "m" (*fx
));
205 /* See comment in fpu_fxsave() below. */
206 return check_insn(rex64
/fxrstor (%[fx
]), "=m" (*fx
), [fx
] "R" (fx
),
210 static inline int fxrstor_user(struct i387_fxsave_struct __user
*fx
)
212 if (config_enabled(CONFIG_X86_32
))
213 return user_insn(fxrstor
%[fx
], "=m" (*fx
), [fx
] "m" (*fx
));
214 else if (config_enabled(CONFIG_AS_FXSAVEQ
))
215 return user_insn(fxrstorq
%[fx
], "=m" (*fx
), [fx
] "m" (*fx
));
217 /* See comment in fpu_fxsave() below. */
218 return user_insn(rex64
/fxrstor (%[fx
]), "=m" (*fx
), [fx
] "R" (fx
),
222 static inline int frstor_checking(struct i387_fsave_struct
*fx
)
224 return check_insn(frstor
%[fx
], "=m" (*fx
), [fx
] "m" (*fx
));
227 static inline int frstor_user(struct i387_fsave_struct __user
*fx
)
229 return user_insn(frstor
%[fx
], "=m" (*fx
), [fx
] "m" (*fx
));
232 static inline void fpu_fxsave(struct fpu
*fpu
)
234 if (config_enabled(CONFIG_X86_32
))
235 asm volatile( "fxsave %[fx]" : [fx
] "=m" (fpu
->state
->fxsave
));
236 else if (config_enabled(CONFIG_AS_FXSAVEQ
))
237 asm volatile("fxsaveq %[fx]" : [fx
] "=m" (fpu
->state
->fxsave
));
239 /* Using "rex64; fxsave %0" is broken because, if the memory
240 * operand uses any extended registers for addressing, a second
241 * REX prefix will be generated (to the assembler, rex64
242 * followed by semicolon is a separate instruction), and hence
243 * the 64-bitness is lost.
245 * Using "fxsaveq %0" would be the ideal choice, but is only
246 * supported starting with gas 2.16.
248 * Using, as a workaround, the properly prefixed form below
249 * isn't accepted by any binutils version so far released,
250 * complaining that the same type of prefix is used twice if
251 * an extended register is needed for addressing (fix submitted
252 * to mainline 2005-11-21).
254 * asm volatile("rex64/fxsave %0" : "=m" (fpu->state->fxsave));
256 * This, however, we can work around by forcing the compiler to
257 * select an addressing mode that doesn't require extended
260 asm volatile( "rex64/fxsave (%[fx])"
261 : "=m" (fpu
->state
->fxsave
)
262 : [fx
] "R" (&fpu
->state
->fxsave
));
267 * These must be called with preempt disabled. Returns
268 * 'true' if the FPU state is still intact.
270 static inline int fpu_save_init(struct fpu
*fpu
)
276 * xsave header may indicate the init state of the FP.
278 if (!(fpu
->state
->xsave
.xsave_hdr
.xstate_bv
& XSTATE_FP
))
280 } else if (use_fxsr()) {
283 asm volatile("fnsave %[fx]; fwait"
284 : [fx
] "=m" (fpu
->state
->fsave
));
289 * If exceptions are pending, we need to clear them so
290 * that we don't randomly get exceptions later.
292 * FIXME! Is this perhaps only true for the old-style
293 * irq13 case? Maybe we could leave the x87 state
296 if (unlikely(fpu
->state
->fxsave
.swd
& X87_FSW_ES
)) {
297 asm volatile("fnclex");
303 static inline int __save_init_fpu(struct task_struct
*tsk
)
305 return fpu_save_init(&tsk
->thread
.fpu
);
308 static inline int fpu_restore_checking(struct fpu
*fpu
)
311 return fpu_xrstor_checking(&fpu
->state
->xsave
);
313 return fxrstor_checking(&fpu
->state
->fxsave
);
315 return frstor_checking(&fpu
->state
->fsave
);
318 static inline int restore_fpu_checking(struct task_struct
*tsk
)
321 * AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception is
322 * pending. Clear the x87 state here by setting it to fixed values.
323 * "m" is a random variable that should be in L1.
325 if (unlikely(static_cpu_has_bug_safe(X86_BUG_FXSAVE_LEAK
))) {
329 "fildl %P[addr]" /* set F?P to defined value */
330 : : [addr
] "m" (tsk
->thread
.fpu
.has_fpu
));
333 return fpu_restore_checking(&tsk
->thread
.fpu
);
337 * Software FPU state helpers. Careful: these need to
338 * be preemption protection *and* they need to be
339 * properly paired with the CR0.TS changes!
341 static inline int __thread_has_fpu(struct task_struct
*tsk
)
343 return tsk
->thread
.fpu
.has_fpu
;
346 /* Must be paired with an 'stts' after! */
347 static inline void __thread_clear_has_fpu(struct task_struct
*tsk
)
349 tsk
->thread
.fpu
.has_fpu
= 0;
350 this_cpu_write(fpu_owner_task
, NULL
);
353 /* Must be paired with a 'clts' before! */
354 static inline void __thread_set_has_fpu(struct task_struct
*tsk
)
356 tsk
->thread
.fpu
.has_fpu
= 1;
357 this_cpu_write(fpu_owner_task
, tsk
);
361 * Encapsulate the CR0.TS handling together with the
364 * These generally need preemption protection to work,
365 * do try to avoid using these on their own.
367 static inline void __thread_fpu_end(struct task_struct
*tsk
)
369 __thread_clear_has_fpu(tsk
);
370 if (!use_eager_fpu())
374 static inline void __thread_fpu_begin(struct task_struct
*tsk
)
376 if (!use_eager_fpu())
378 __thread_set_has_fpu(tsk
);
381 static inline void __drop_fpu(struct task_struct
*tsk
)
383 if (__thread_has_fpu(tsk
)) {
384 /* Ignore delayed exceptions from user space */
385 asm volatile("1: fwait\n"
387 _ASM_EXTABLE(1b
, 2b
));
388 __thread_fpu_end(tsk
);
392 static inline void drop_fpu(struct task_struct
*tsk
)
395 * Forget coprocessor state..
398 tsk
->thread
.fpu_counter
= 0;
400 clear_stopped_child_used_math(tsk
);
404 static inline void drop_init_fpu(struct task_struct
*tsk
)
406 if (!use_eager_fpu())
410 xrstor_state(init_xstate_buf
, -1);
412 fxrstor_checking(&init_xstate_buf
->i387
);
417 * FPU state switching for scheduling.
419 * This is a two-stage process:
421 * - switch_fpu_prepare() saves the old state and
422 * sets the new state of the CR0.TS bit. This is
423 * done within the context of the old process.
425 * - switch_fpu_finish() restores the new state as
428 typedef struct { int preload
; } fpu_switch_t
;
430 static inline fpu_switch_t
switch_fpu_prepare(struct task_struct
*old
, struct task_struct
*new, int cpu
)
435 * If the task has used the math, pre-load the FPU on xsave processors
436 * or if the past 5 consecutive context-switches used math.
438 fpu
.preload
= tsk_used_math(new) &&
439 (use_eager_fpu() || new->thread
.fpu_counter
> 5);
441 if (__thread_has_fpu(old
)) {
442 if (!__save_init_fpu(old
))
443 task_disable_lazy_fpu_restore(old
);
445 old
->thread
.fpu
.last_cpu
= cpu
;
447 /* But leave fpu_owner_task! */
448 old
->thread
.fpu
.has_fpu
= 0;
450 /* Don't change CR0.TS if we just switch! */
452 new->thread
.fpu_counter
++;
453 __thread_set_has_fpu(new);
454 prefetch(new->thread
.fpu
.state
);
455 } else if (!use_eager_fpu())
458 old
->thread
.fpu_counter
= 0;
459 task_disable_lazy_fpu_restore(old
);
461 new->thread
.fpu_counter
++;
462 if (fpu_lazy_restore(new, cpu
))
465 prefetch(new->thread
.fpu
.state
);
466 __thread_fpu_begin(new);
473 * By the time this gets called, we've already cleared CR0.TS and
474 * given the process the FPU if we are going to preload the FPU
475 * state - all we need to do is to conditionally restore the register
478 static inline void switch_fpu_finish(struct task_struct
*new, fpu_switch_t fpu
)
481 if (unlikely(restore_fpu_checking(new)))
487 * Signal frame handlers...
489 extern int save_xstate_sig(void __user
*buf
, void __user
*fx
, int size
);
490 extern int __restore_xstate_sig(void __user
*buf
, void __user
*fx
, int size
);
492 static inline int xstate_sigframe_size(void)
494 return use_xsave() ? xstate_size
+ FP_XSTATE_MAGIC2_SIZE
: xstate_size
;
497 static inline int restore_xstate_sig(void __user
*buf
, int ia32_frame
)
499 void __user
*buf_fx
= buf
;
500 int size
= xstate_sigframe_size();
502 if (ia32_frame
&& use_fxsr()) {
503 buf_fx
= buf
+ sizeof(struct i387_fsave_struct
);
504 size
+= sizeof(struct i387_fsave_struct
);
507 return __restore_xstate_sig(buf
, buf_fx
, size
);
511 * Need to be preemption-safe.
513 * NOTE! user_fpu_begin() must be used only immediately before restoring
514 * it. This function does not do any save/restore on their own.
516 static inline void user_fpu_begin(void)
520 __thread_fpu_begin(current
);
524 static inline void __save_fpu(struct task_struct
*tsk
)
527 if (unlikely(system_state
== SYSTEM_BOOTING
))
528 xsave_state_booting(&tsk
->thread
.fpu
.state
->xsave
, -1);
530 xsave_state(&tsk
->thread
.fpu
.state
->xsave
, -1);
532 fpu_fxsave(&tsk
->thread
.fpu
);
536 * i387 state interaction
538 static inline unsigned short get_fpu_cwd(struct task_struct
*tsk
)
541 return tsk
->thread
.fpu
.state
->fxsave
.cwd
;
543 return (unsigned short)tsk
->thread
.fpu
.state
->fsave
.cwd
;
547 static inline unsigned short get_fpu_swd(struct task_struct
*tsk
)
550 return tsk
->thread
.fpu
.state
->fxsave
.swd
;
552 return (unsigned short)tsk
->thread
.fpu
.state
->fsave
.swd
;
556 static inline unsigned short get_fpu_mxcsr(struct task_struct
*tsk
)
559 return tsk
->thread
.fpu
.state
->fxsave
.mxcsr
;
561 return MXCSR_DEFAULT
;
565 static bool fpu_allocated(struct fpu
*fpu
)
567 return fpu
->state
!= NULL
;
570 static inline int fpu_alloc(struct fpu
*fpu
)
572 if (fpu_allocated(fpu
))
574 fpu
->state
= kmem_cache_alloc(task_xstate_cachep
, GFP_KERNEL
);
577 WARN_ON((unsigned long)fpu
->state
& 15);
581 static inline void fpu_free(struct fpu
*fpu
)
584 kmem_cache_free(task_xstate_cachep
, fpu
->state
);
589 static inline void fpu_copy(struct task_struct
*dst
, struct task_struct
*src
)
591 if (use_eager_fpu()) {
592 memset(&dst
->thread
.fpu
.state
->xsave
, 0, xstate_size
);
595 struct fpu
*dfpu
= &dst
->thread
.fpu
;
596 struct fpu
*sfpu
= &src
->thread
.fpu
;
599 memcpy(dfpu
->state
, sfpu
->state
, xstate_size
);
603 static inline unsigned long
604 alloc_mathframe(unsigned long sp
, int ia32_frame
, unsigned long *buf_fx
,
607 unsigned long frame_size
= xstate_sigframe_size();
609 *buf_fx
= sp
= round_down(sp
- frame_size
, 64);
610 if (ia32_frame
&& use_fxsr()) {
611 frame_size
+= sizeof(struct i387_fsave_struct
);
612 sp
-= sizeof(struct i387_fsave_struct
);