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
8 #include <linux/module.h>
9 #include <linux/regset.h>
10 #include <linux/sched.h>
11 #include <linux/slab.h>
13 #include <asm/sigcontext.h>
14 #include <asm/processor.h>
15 #include <asm/math_emu.h>
16 #include <asm/uaccess.h>
17 #include <asm/ptrace.h>
22 # include <asm/sigcontext32.h>
23 # include <asm/user32.h>
25 # define save_i387_xstate_ia32 save_i387_xstate
26 # define restore_i387_xstate_ia32 restore_i387_xstate
27 # define _fpstate_ia32 _fpstate
28 # define _xstate_ia32 _xstate
29 # define sig_xstate_ia32_size sig_xstate_size
30 # define fx_sw_reserved_ia32 fx_sw_reserved
31 # define user_i387_ia32_struct user_i387_struct
32 # define user32_fxsr_struct user_fxsr_struct
35 #ifdef CONFIG_MATH_EMULATION
36 # define HAVE_HWFP (boot_cpu_data.hard_math)
41 static unsigned int mxcsr_feature_mask __read_mostly
= 0xffffffffu
;
42 unsigned int xstate_size
;
43 EXPORT_SYMBOL_GPL(xstate_size
);
44 unsigned int sig_xstate_ia32_size
= sizeof(struct _fpstate_ia32
);
45 static struct i387_fxsave_struct fx_scratch __cpuinitdata
;
47 void __cpuinit
mxcsr_feature_mask_init(void)
49 unsigned long mask
= 0;
53 memset(&fx_scratch
, 0, sizeof(struct i387_fxsave_struct
));
54 asm volatile("fxsave %0" : : "m" (fx_scratch
));
55 mask
= fx_scratch
.mxcsr_mask
;
59 mxcsr_feature_mask
&= mask
;
63 static void __cpuinit
init_thread_xstate(void)
66 * Note that xstate_size might be overwriten later during
72 * Disable xsave as we do not support it if i387
73 * emulation is enabled.
75 setup_clear_cpu_cap(X86_FEATURE_XSAVE
);
76 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT
);
77 xstate_size
= sizeof(struct i387_soft_struct
);
82 xstate_size
= sizeof(struct i387_fxsave_struct
);
84 xstate_size
= sizeof(struct i387_fsave_struct
);
88 * Called at bootup to set up the initial FPU state that is later cloned
92 void __cpuinit
fpu_init(void)
95 unsigned long cr4_mask
= 0;
98 cr4_mask
|= X86_CR4_OSFXSR
;
100 cr4_mask
|= X86_CR4_OSXMMEXCPT
;
102 set_in_cr4(cr4_mask
);
105 cr0
&= ~(X86_CR0_TS
|X86_CR0_EM
); /* clear TS and EM */
110 if (!smp_processor_id())
111 init_thread_xstate();
113 mxcsr_feature_mask_init();
114 /* clean state in init */
115 current_thread_info()->status
= 0;
119 void fpu_finit(struct fpu
*fpu
)
122 finit_soft_fpu(&fpu
->state
->soft
);
127 struct i387_fxsave_struct
*fx
= &fpu
->state
->fxsave
;
129 memset(fx
, 0, xstate_size
);
132 fx
->mxcsr
= MXCSR_DEFAULT
;
134 struct i387_fsave_struct
*fp
= &fpu
->state
->fsave
;
135 memset(fp
, 0, xstate_size
);
136 fp
->cwd
= 0xffff037fu
;
137 fp
->swd
= 0xffff0000u
;
138 fp
->twd
= 0xffffffffu
;
139 fp
->fos
= 0xffff0000u
;
142 EXPORT_SYMBOL_GPL(fpu_finit
);
145 * The _current_ task is using the FPU for the first time
146 * so initialize it and set the mxcsr to its default
147 * value at reset if we support XMM instructions and then
148 * remeber the current task has used the FPU.
150 int init_fpu(struct task_struct
*tsk
)
154 if (tsk_used_math(tsk
)) {
155 if (HAVE_HWFP
&& tsk
== current
)
161 * Memory allocation at the first usage of the FPU and other state.
163 ret
= fpu_alloc(&tsk
->thread
.fpu
);
167 fpu_finit(&tsk
->thread
.fpu
);
169 set_stopped_child_used_math(tsk
);
174 * The xstateregs_active() routine is the same as the fpregs_active() routine,
175 * as the "regset->n" for the xstate regset will be updated based on the feature
176 * capabilites supported by the xsave.
178 int fpregs_active(struct task_struct
*target
, const struct user_regset
*regset
)
180 return tsk_used_math(target
) ? regset
->n
: 0;
183 int xfpregs_active(struct task_struct
*target
, const struct user_regset
*regset
)
185 return (cpu_has_fxsr
&& tsk_used_math(target
)) ? regset
->n
: 0;
188 int xfpregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
189 unsigned int pos
, unsigned int count
,
190 void *kbuf
, void __user
*ubuf
)
197 ret
= init_fpu(target
);
201 sanitize_i387_state(target
);
203 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
204 &target
->thread
.fpu
.state
->fxsave
, 0, -1);
207 int xfpregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
208 unsigned int pos
, unsigned int count
,
209 const void *kbuf
, const void __user
*ubuf
)
216 ret
= init_fpu(target
);
220 sanitize_i387_state(target
);
222 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
223 &target
->thread
.fpu
.state
->fxsave
, 0, -1);
226 * mxcsr reserved bits must be masked to zero for security reasons.
228 target
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
231 * update the header bits in the xsave header, indicating the
232 * presence of FP and SSE state.
235 target
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FPSSE
;
240 int xstateregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
241 unsigned int pos
, unsigned int count
,
242 void *kbuf
, void __user
*ubuf
)
249 ret
= init_fpu(target
);
254 * Copy the 48bytes defined by the software first into the xstate
255 * memory layout in the thread struct, so that we can copy the entire
256 * xstateregs to the user using one user_regset_copyout().
258 memcpy(&target
->thread
.fpu
.state
->fxsave
.sw_reserved
,
259 xstate_fx_sw_bytes
, sizeof(xstate_fx_sw_bytes
));
262 * Copy the xstate memory layout.
264 ret
= user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
265 &target
->thread
.fpu
.state
->xsave
, 0, -1);
269 int xstateregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
270 unsigned int pos
, unsigned int count
,
271 const void *kbuf
, const void __user
*ubuf
)
274 struct xsave_hdr_struct
*xsave_hdr
;
279 ret
= init_fpu(target
);
283 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
284 &target
->thread
.fpu
.state
->xsave
, 0, -1);
287 * mxcsr reserved bits must be masked to zero for security reasons.
289 target
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
291 xsave_hdr
= &target
->thread
.fpu
.state
->xsave
.xsave_hdr
;
293 xsave_hdr
->xstate_bv
&= pcntxt_mask
;
295 * These bits must be zero.
297 xsave_hdr
->reserved1
[0] = xsave_hdr
->reserved1
[1] = 0;
302 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
305 * FPU tag word conversions.
308 static inline unsigned short twd_i387_to_fxsr(unsigned short twd
)
310 unsigned int tmp
; /* to avoid 16 bit prefixes in the code */
312 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
314 tmp
= (tmp
| (tmp
>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
315 /* and move the valid bits to the lower byte. */
316 tmp
= (tmp
| (tmp
>> 1)) & 0x3333; /* 00VV00VV00VV00VV */
317 tmp
= (tmp
| (tmp
>> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
318 tmp
= (tmp
| (tmp
>> 4)) & 0x00ff; /* 00000000VVVVVVVV */
323 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16);
324 #define FP_EXP_TAG_VALID 0
325 #define FP_EXP_TAG_ZERO 1
326 #define FP_EXP_TAG_SPECIAL 2
327 #define FP_EXP_TAG_EMPTY 3
329 static inline u32
twd_fxsr_to_i387(struct i387_fxsave_struct
*fxsave
)
332 u32 tos
= (fxsave
->swd
>> 11) & 7;
333 u32 twd
= (unsigned long) fxsave
->twd
;
335 u32 ret
= 0xffff0000u
;
338 for (i
= 0; i
< 8; i
++, twd
>>= 1) {
340 st
= FPREG_ADDR(fxsave
, (i
- tos
) & 7);
342 switch (st
->exponent
& 0x7fff) {
344 tag
= FP_EXP_TAG_SPECIAL
;
347 if (!st
->significand
[0] &&
348 !st
->significand
[1] &&
349 !st
->significand
[2] &&
351 tag
= FP_EXP_TAG_ZERO
;
353 tag
= FP_EXP_TAG_SPECIAL
;
356 if (st
->significand
[3] & 0x8000)
357 tag
= FP_EXP_TAG_VALID
;
359 tag
= FP_EXP_TAG_SPECIAL
;
363 tag
= FP_EXP_TAG_EMPTY
;
365 ret
|= tag
<< (2 * i
);
371 * FXSR floating point environment conversions.
375 convert_from_fxsr(struct user_i387_ia32_struct
*env
, struct task_struct
*tsk
)
377 struct i387_fxsave_struct
*fxsave
= &tsk
->thread
.fpu
.state
->fxsave
;
378 struct _fpreg
*to
= (struct _fpreg
*) &env
->st_space
[0];
379 struct _fpxreg
*from
= (struct _fpxreg
*) &fxsave
->st_space
[0];
382 env
->cwd
= fxsave
->cwd
| 0xffff0000u
;
383 env
->swd
= fxsave
->swd
| 0xffff0000u
;
384 env
->twd
= twd_fxsr_to_i387(fxsave
);
387 env
->fip
= fxsave
->rip
;
388 env
->foo
= fxsave
->rdp
;
390 * should be actually ds/cs at fpu exception time, but
391 * that information is not available in 64bit mode.
393 env
->fcs
= task_pt_regs(tsk
)->cs
;
394 if (tsk
== current
) {
395 savesegment(ds
, env
->fos
);
397 env
->fos
= tsk
->thread
.ds
;
399 env
->fos
|= 0xffff0000;
401 env
->fip
= fxsave
->fip
;
402 env
->fcs
= (u16
) fxsave
->fcs
| ((u32
) fxsave
->fop
<< 16);
403 env
->foo
= fxsave
->foo
;
404 env
->fos
= fxsave
->fos
;
407 for (i
= 0; i
< 8; ++i
)
408 memcpy(&to
[i
], &from
[i
], sizeof(to
[0]));
411 static void convert_to_fxsr(struct task_struct
*tsk
,
412 const struct user_i387_ia32_struct
*env
)
415 struct i387_fxsave_struct
*fxsave
= &tsk
->thread
.fpu
.state
->fxsave
;
416 struct _fpreg
*from
= (struct _fpreg
*) &env
->st_space
[0];
417 struct _fpxreg
*to
= (struct _fpxreg
*) &fxsave
->st_space
[0];
420 fxsave
->cwd
= env
->cwd
;
421 fxsave
->swd
= env
->swd
;
422 fxsave
->twd
= twd_i387_to_fxsr(env
->twd
);
423 fxsave
->fop
= (u16
) ((u32
) env
->fcs
>> 16);
425 fxsave
->rip
= env
->fip
;
426 fxsave
->rdp
= env
->foo
;
427 /* cs and ds ignored */
429 fxsave
->fip
= env
->fip
;
430 fxsave
->fcs
= (env
->fcs
& 0xffff);
431 fxsave
->foo
= env
->foo
;
432 fxsave
->fos
= env
->fos
;
435 for (i
= 0; i
< 8; ++i
)
436 memcpy(&to
[i
], &from
[i
], sizeof(from
[0]));
439 int fpregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
440 unsigned int pos
, unsigned int count
,
441 void *kbuf
, void __user
*ubuf
)
443 struct user_i387_ia32_struct env
;
446 ret
= init_fpu(target
);
451 return fpregs_soft_get(target
, regset
, pos
, count
, kbuf
, ubuf
);
454 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
455 &target
->thread
.fpu
.state
->fsave
, 0,
459 sanitize_i387_state(target
);
461 if (kbuf
&& pos
== 0 && count
== sizeof(env
)) {
462 convert_from_fxsr(kbuf
, target
);
466 convert_from_fxsr(&env
, target
);
468 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
, &env
, 0, -1);
471 int fpregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
472 unsigned int pos
, unsigned int count
,
473 const void *kbuf
, const void __user
*ubuf
)
475 struct user_i387_ia32_struct env
;
478 ret
= init_fpu(target
);
482 sanitize_i387_state(target
);
485 return fpregs_soft_set(target
, regset
, pos
, count
, kbuf
, ubuf
);
488 return user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
489 &target
->thread
.fpu
.state
->fsave
, 0, -1);
492 if (pos
> 0 || count
< sizeof(env
))
493 convert_from_fxsr(&env
, target
);
495 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
, &env
, 0, -1);
497 convert_to_fxsr(target
, &env
);
500 * update the header bit in the xsave header, indicating the
504 target
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FP
;
509 * Signal frame handlers.
512 static inline int save_i387_fsave(struct _fpstate_ia32 __user
*buf
)
514 struct task_struct
*tsk
= current
;
515 struct i387_fsave_struct
*fp
= &tsk
->thread
.fpu
.state
->fsave
;
517 fp
->status
= fp
->swd
;
518 if (__copy_to_user(buf
, fp
, sizeof(struct i387_fsave_struct
)))
523 static int save_i387_fxsave(struct _fpstate_ia32 __user
*buf
)
525 struct task_struct
*tsk
= current
;
526 struct i387_fxsave_struct
*fx
= &tsk
->thread
.fpu
.state
->fxsave
;
527 struct user_i387_ia32_struct env
;
530 convert_from_fxsr(&env
, tsk
);
531 if (__copy_to_user(buf
, &env
, sizeof(env
)))
534 err
|= __put_user(fx
->swd
, &buf
->status
);
535 err
|= __put_user(X86_FXSR_MAGIC
, &buf
->magic
);
539 if (__copy_to_user(&buf
->_fxsr_env
[0], fx
, xstate_size
))
544 static int save_i387_xsave(void __user
*buf
)
546 struct task_struct
*tsk
= current
;
547 struct _fpstate_ia32 __user
*fx
= buf
;
551 sanitize_i387_state(tsk
);
554 * For legacy compatible, we always set FP/SSE bits in the bit
555 * vector while saving the state to the user context.
556 * This will enable us capturing any changes(during sigreturn) to
557 * the FP/SSE bits by the legacy applications which don't touch
558 * xstate_bv in the xsave header.
560 * xsave aware applications can change the xstate_bv in the xsave
561 * header as well as change any contents in the memory layout.
562 * xrestore as part of sigreturn will capture all the changes.
564 tsk
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FPSSE
;
566 if (save_i387_fxsave(fx
) < 0)
569 err
= __copy_to_user(&fx
->sw_reserved
, &fx_sw_reserved_ia32
,
570 sizeof(struct _fpx_sw_bytes
));
571 err
|= __put_user(FP_XSTATE_MAGIC2
,
572 (__u32 __user
*) (buf
+ sig_xstate_ia32_size
573 - FP_XSTATE_MAGIC2_SIZE
));
580 int save_i387_xstate_ia32(void __user
*buf
)
582 struct _fpstate_ia32 __user
*fp
= (struct _fpstate_ia32 __user
*) buf
;
583 struct task_struct
*tsk
= current
;
588 if (!access_ok(VERIFY_WRITE
, buf
, sig_xstate_ia32_size
))
591 * This will cause a "finit" to be triggered by the next
592 * attempted FPU operation by the 'current' process.
597 return fpregs_soft_get(current
, NULL
,
598 0, sizeof(struct user_i387_ia32_struct
),
605 return save_i387_xsave(fp
);
607 return save_i387_fxsave(fp
);
609 return save_i387_fsave(fp
);
612 static inline int restore_i387_fsave(struct _fpstate_ia32 __user
*buf
)
614 struct task_struct
*tsk
= current
;
616 return __copy_from_user(&tsk
->thread
.fpu
.state
->fsave
, buf
,
617 sizeof(struct i387_fsave_struct
));
620 static int restore_i387_fxsave(struct _fpstate_ia32 __user
*buf
,
623 struct task_struct
*tsk
= current
;
624 struct user_i387_ia32_struct env
;
627 err
= __copy_from_user(&tsk
->thread
.fpu
.state
->fxsave
, &buf
->_fxsr_env
[0],
629 /* mxcsr reserved bits must be masked to zero for security reasons */
630 tsk
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
631 if (err
|| __copy_from_user(&env
, buf
, sizeof(env
)))
633 convert_to_fxsr(tsk
, &env
);
638 static int restore_i387_xsave(void __user
*buf
)
640 struct _fpx_sw_bytes fx_sw_user
;
641 struct _fpstate_ia32 __user
*fx_user
=
642 ((struct _fpstate_ia32 __user
*) buf
);
643 struct i387_fxsave_struct __user
*fx
=
644 (struct i387_fxsave_struct __user
*) &fx_user
->_fxsr_env
[0];
645 struct xsave_hdr_struct
*xsave_hdr
=
646 ¤t
->thread
.fpu
.state
->xsave
.xsave_hdr
;
650 if (check_for_xstate(fx
, buf
, &fx_sw_user
))
653 mask
= fx_sw_user
.xstate_bv
;
655 err
= restore_i387_fxsave(buf
, fx_sw_user
.xstate_size
);
657 xsave_hdr
->xstate_bv
&= pcntxt_mask
;
659 * These bits must be zero.
661 xsave_hdr
->reserved1
[0] = xsave_hdr
->reserved1
[1] = 0;
664 * Init the state that is not present in the memory layout
665 * and enabled by the OS.
667 mask
= ~(pcntxt_mask
& ~mask
);
668 xsave_hdr
->xstate_bv
&= mask
;
673 * Couldn't find the extended state information in the memory
674 * layout. Restore the FP/SSE and init the other extended state
677 xsave_hdr
->xstate_bv
= XSTATE_FPSSE
;
678 return restore_i387_fxsave(buf
, sizeof(struct i387_fxsave_struct
));
681 int restore_i387_xstate_ia32(void __user
*buf
)
684 struct task_struct
*tsk
= current
;
685 struct _fpstate_ia32 __user
*fp
= (struct _fpstate_ia32 __user
*) buf
;
698 if (!access_ok(VERIFY_READ
, buf
, sig_xstate_ia32_size
))
709 err
= restore_i387_xsave(buf
);
710 else if (cpu_has_fxsr
)
711 err
= restore_i387_fxsave(fp
, sizeof(struct
712 i387_fxsave_struct
));
714 err
= restore_i387_fsave(fp
);
716 err
= fpregs_soft_set(current
, NULL
,
717 0, sizeof(struct user_i387_ia32_struct
),
726 * FPU state for core dumps.
727 * This is only used for a.out dumps now.
728 * It is declared generically using elf_fpregset_t (which is
729 * struct user_i387_struct) but is in fact only used for 32-bit
730 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
732 int dump_fpu(struct pt_regs
*regs
, struct user_i387_struct
*fpu
)
734 struct task_struct
*tsk
= current
;
737 fpvalid
= !!used_math();
739 fpvalid
= !fpregs_get(tsk
, NULL
,
740 0, sizeof(struct user_i387_ia32_struct
),
745 EXPORT_SYMBOL(dump_fpu
);
747 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */