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 * remember 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
)
157 tsk
->thread
.fpu
.last_cpu
= ~0;
162 * Memory allocation at the first usage of the FPU and other state.
164 ret
= fpu_alloc(&tsk
->thread
.fpu
);
168 fpu_finit(&tsk
->thread
.fpu
);
170 set_stopped_child_used_math(tsk
);
173 EXPORT_SYMBOL_GPL(init_fpu
);
176 * The xstateregs_active() routine is the same as the fpregs_active() routine,
177 * as the "regset->n" for the xstate regset will be updated based on the feature
178 * capabilites supported by the xsave.
180 int fpregs_active(struct task_struct
*target
, const struct user_regset
*regset
)
182 return tsk_used_math(target
) ? regset
->n
: 0;
185 int xfpregs_active(struct task_struct
*target
, const struct user_regset
*regset
)
187 return (cpu_has_fxsr
&& tsk_used_math(target
)) ? regset
->n
: 0;
190 int xfpregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
191 unsigned int pos
, unsigned int count
,
192 void *kbuf
, void __user
*ubuf
)
199 ret
= init_fpu(target
);
203 sanitize_i387_state(target
);
205 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
206 &target
->thread
.fpu
.state
->fxsave
, 0, -1);
209 int xfpregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
210 unsigned int pos
, unsigned int count
,
211 const void *kbuf
, const void __user
*ubuf
)
218 ret
= init_fpu(target
);
222 sanitize_i387_state(target
);
224 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
225 &target
->thread
.fpu
.state
->fxsave
, 0, -1);
228 * mxcsr reserved bits must be masked to zero for security reasons.
230 target
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
233 * update the header bits in the xsave header, indicating the
234 * presence of FP and SSE state.
237 target
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FPSSE
;
242 int xstateregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
243 unsigned int pos
, unsigned int count
,
244 void *kbuf
, void __user
*ubuf
)
251 ret
= init_fpu(target
);
256 * Copy the 48bytes defined by the software first into the xstate
257 * memory layout in the thread struct, so that we can copy the entire
258 * xstateregs to the user using one user_regset_copyout().
260 memcpy(&target
->thread
.fpu
.state
->fxsave
.sw_reserved
,
261 xstate_fx_sw_bytes
, sizeof(xstate_fx_sw_bytes
));
264 * Copy the xstate memory layout.
266 ret
= user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
267 &target
->thread
.fpu
.state
->xsave
, 0, -1);
271 int xstateregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
272 unsigned int pos
, unsigned int count
,
273 const void *kbuf
, const void __user
*ubuf
)
276 struct xsave_hdr_struct
*xsave_hdr
;
281 ret
= init_fpu(target
);
285 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
286 &target
->thread
.fpu
.state
->xsave
, 0, -1);
289 * mxcsr reserved bits must be masked to zero for security reasons.
291 target
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
293 xsave_hdr
= &target
->thread
.fpu
.state
->xsave
.xsave_hdr
;
295 xsave_hdr
->xstate_bv
&= pcntxt_mask
;
297 * These bits must be zero.
299 xsave_hdr
->reserved1
[0] = xsave_hdr
->reserved1
[1] = 0;
304 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
307 * FPU tag word conversions.
310 static inline unsigned short twd_i387_to_fxsr(unsigned short twd
)
312 unsigned int tmp
; /* to avoid 16 bit prefixes in the code */
314 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
316 tmp
= (tmp
| (tmp
>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
317 /* and move the valid bits to the lower byte. */
318 tmp
= (tmp
| (tmp
>> 1)) & 0x3333; /* 00VV00VV00VV00VV */
319 tmp
= (tmp
| (tmp
>> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
320 tmp
= (tmp
| (tmp
>> 4)) & 0x00ff; /* 00000000VVVVVVVV */
325 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
326 #define FP_EXP_TAG_VALID 0
327 #define FP_EXP_TAG_ZERO 1
328 #define FP_EXP_TAG_SPECIAL 2
329 #define FP_EXP_TAG_EMPTY 3
331 static inline u32
twd_fxsr_to_i387(struct i387_fxsave_struct
*fxsave
)
334 u32 tos
= (fxsave
->swd
>> 11) & 7;
335 u32 twd
= (unsigned long) fxsave
->twd
;
337 u32 ret
= 0xffff0000u
;
340 for (i
= 0; i
< 8; i
++, twd
>>= 1) {
342 st
= FPREG_ADDR(fxsave
, (i
- tos
) & 7);
344 switch (st
->exponent
& 0x7fff) {
346 tag
= FP_EXP_TAG_SPECIAL
;
349 if (!st
->significand
[0] &&
350 !st
->significand
[1] &&
351 !st
->significand
[2] &&
353 tag
= FP_EXP_TAG_ZERO
;
355 tag
= FP_EXP_TAG_SPECIAL
;
358 if (st
->significand
[3] & 0x8000)
359 tag
= FP_EXP_TAG_VALID
;
361 tag
= FP_EXP_TAG_SPECIAL
;
365 tag
= FP_EXP_TAG_EMPTY
;
367 ret
|= tag
<< (2 * i
);
373 * FXSR floating point environment conversions.
377 convert_from_fxsr(struct user_i387_ia32_struct
*env
, struct task_struct
*tsk
)
379 struct i387_fxsave_struct
*fxsave
= &tsk
->thread
.fpu
.state
->fxsave
;
380 struct _fpreg
*to
= (struct _fpreg
*) &env
->st_space
[0];
381 struct _fpxreg
*from
= (struct _fpxreg
*) &fxsave
->st_space
[0];
384 env
->cwd
= fxsave
->cwd
| 0xffff0000u
;
385 env
->swd
= fxsave
->swd
| 0xffff0000u
;
386 env
->twd
= twd_fxsr_to_i387(fxsave
);
389 env
->fip
= fxsave
->rip
;
390 env
->foo
= fxsave
->rdp
;
392 * should be actually ds/cs at fpu exception time, but
393 * that information is not available in 64bit mode.
395 env
->fcs
= task_pt_regs(tsk
)->cs
;
396 if (tsk
== current
) {
397 savesegment(ds
, env
->fos
);
399 env
->fos
= tsk
->thread
.ds
;
401 env
->fos
|= 0xffff0000;
403 env
->fip
= fxsave
->fip
;
404 env
->fcs
= (u16
) fxsave
->fcs
| ((u32
) fxsave
->fop
<< 16);
405 env
->foo
= fxsave
->foo
;
406 env
->fos
= fxsave
->fos
;
409 for (i
= 0; i
< 8; ++i
)
410 memcpy(&to
[i
], &from
[i
], sizeof(to
[0]));
413 static void convert_to_fxsr(struct task_struct
*tsk
,
414 const struct user_i387_ia32_struct
*env
)
417 struct i387_fxsave_struct
*fxsave
= &tsk
->thread
.fpu
.state
->fxsave
;
418 struct _fpreg
*from
= (struct _fpreg
*) &env
->st_space
[0];
419 struct _fpxreg
*to
= (struct _fpxreg
*) &fxsave
->st_space
[0];
422 fxsave
->cwd
= env
->cwd
;
423 fxsave
->swd
= env
->swd
;
424 fxsave
->twd
= twd_i387_to_fxsr(env
->twd
);
425 fxsave
->fop
= (u16
) ((u32
) env
->fcs
>> 16);
427 fxsave
->rip
= env
->fip
;
428 fxsave
->rdp
= env
->foo
;
429 /* cs and ds ignored */
431 fxsave
->fip
= env
->fip
;
432 fxsave
->fcs
= (env
->fcs
& 0xffff);
433 fxsave
->foo
= env
->foo
;
434 fxsave
->fos
= env
->fos
;
437 for (i
= 0; i
< 8; ++i
)
438 memcpy(&to
[i
], &from
[i
], sizeof(from
[0]));
441 int fpregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
442 unsigned int pos
, unsigned int count
,
443 void *kbuf
, void __user
*ubuf
)
445 struct user_i387_ia32_struct env
;
448 ret
= init_fpu(target
);
453 return fpregs_soft_get(target
, regset
, pos
, count
, kbuf
, ubuf
);
456 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
457 &target
->thread
.fpu
.state
->fsave
, 0,
461 sanitize_i387_state(target
);
463 if (kbuf
&& pos
== 0 && count
== sizeof(env
)) {
464 convert_from_fxsr(kbuf
, target
);
468 convert_from_fxsr(&env
, target
);
470 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
, &env
, 0, -1);
473 int fpregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
474 unsigned int pos
, unsigned int count
,
475 const void *kbuf
, const void __user
*ubuf
)
477 struct user_i387_ia32_struct env
;
480 ret
= init_fpu(target
);
484 sanitize_i387_state(target
);
487 return fpregs_soft_set(target
, regset
, pos
, count
, kbuf
, ubuf
);
490 return user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
491 &target
->thread
.fpu
.state
->fsave
, 0, -1);
494 if (pos
> 0 || count
< sizeof(env
))
495 convert_from_fxsr(&env
, target
);
497 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
, &env
, 0, -1);
499 convert_to_fxsr(target
, &env
);
502 * update the header bit in the xsave header, indicating the
506 target
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FP
;
511 * Signal frame handlers.
514 static inline int save_i387_fsave(struct _fpstate_ia32 __user
*buf
)
516 struct task_struct
*tsk
= current
;
517 struct i387_fsave_struct
*fp
= &tsk
->thread
.fpu
.state
->fsave
;
519 fp
->status
= fp
->swd
;
520 if (__copy_to_user(buf
, fp
, sizeof(struct i387_fsave_struct
)))
525 static int save_i387_fxsave(struct _fpstate_ia32 __user
*buf
)
527 struct task_struct
*tsk
= current
;
528 struct i387_fxsave_struct
*fx
= &tsk
->thread
.fpu
.state
->fxsave
;
529 struct user_i387_ia32_struct env
;
532 convert_from_fxsr(&env
, tsk
);
533 if (__copy_to_user(buf
, &env
, sizeof(env
)))
536 err
|= __put_user(fx
->swd
, &buf
->status
);
537 err
|= __put_user(X86_FXSR_MAGIC
, &buf
->magic
);
541 if (__copy_to_user(&buf
->_fxsr_env
[0], fx
, xstate_size
))
546 static int save_i387_xsave(void __user
*buf
)
548 struct task_struct
*tsk
= current
;
549 struct _fpstate_ia32 __user
*fx
= buf
;
553 sanitize_i387_state(tsk
);
556 * For legacy compatible, we always set FP/SSE bits in the bit
557 * vector while saving the state to the user context.
558 * This will enable us capturing any changes(during sigreturn) to
559 * the FP/SSE bits by the legacy applications which don't touch
560 * xstate_bv in the xsave header.
562 * xsave aware applications can change the xstate_bv in the xsave
563 * header as well as change any contents in the memory layout.
564 * xrestore as part of sigreturn will capture all the changes.
566 tsk
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FPSSE
;
568 if (save_i387_fxsave(fx
) < 0)
571 err
= __copy_to_user(&fx
->sw_reserved
, &fx_sw_reserved_ia32
,
572 sizeof(struct _fpx_sw_bytes
));
573 err
|= __put_user(FP_XSTATE_MAGIC2
,
574 (__u32 __user
*) (buf
+ sig_xstate_ia32_size
575 - FP_XSTATE_MAGIC2_SIZE
));
582 int save_i387_xstate_ia32(void __user
*buf
)
584 struct _fpstate_ia32 __user
*fp
= (struct _fpstate_ia32 __user
*) buf
;
585 struct task_struct
*tsk
= current
;
590 if (!access_ok(VERIFY_WRITE
, buf
, sig_xstate_ia32_size
))
593 * This will cause a "finit" to be triggered by the next
594 * attempted FPU operation by the 'current' process.
599 return fpregs_soft_get(current
, NULL
,
600 0, sizeof(struct user_i387_ia32_struct
),
607 return save_i387_xsave(fp
);
609 return save_i387_fxsave(fp
);
611 return save_i387_fsave(fp
);
614 static inline int restore_i387_fsave(struct _fpstate_ia32 __user
*buf
)
616 struct task_struct
*tsk
= current
;
618 return __copy_from_user(&tsk
->thread
.fpu
.state
->fsave
, buf
,
619 sizeof(struct i387_fsave_struct
));
622 static int restore_i387_fxsave(struct _fpstate_ia32 __user
*buf
,
625 struct task_struct
*tsk
= current
;
626 struct user_i387_ia32_struct env
;
629 err
= __copy_from_user(&tsk
->thread
.fpu
.state
->fxsave
, &buf
->_fxsr_env
[0],
631 /* mxcsr reserved bits must be masked to zero for security reasons */
632 tsk
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
633 if (err
|| __copy_from_user(&env
, buf
, sizeof(env
)))
635 convert_to_fxsr(tsk
, &env
);
640 static int restore_i387_xsave(void __user
*buf
)
642 struct _fpx_sw_bytes fx_sw_user
;
643 struct _fpstate_ia32 __user
*fx_user
=
644 ((struct _fpstate_ia32 __user
*) buf
);
645 struct i387_fxsave_struct __user
*fx
=
646 (struct i387_fxsave_struct __user
*) &fx_user
->_fxsr_env
[0];
647 struct xsave_hdr_struct
*xsave_hdr
=
648 ¤t
->thread
.fpu
.state
->xsave
.xsave_hdr
;
652 if (check_for_xstate(fx
, buf
, &fx_sw_user
))
655 mask
= fx_sw_user
.xstate_bv
;
657 err
= restore_i387_fxsave(buf
, fx_sw_user
.xstate_size
);
659 xsave_hdr
->xstate_bv
&= pcntxt_mask
;
661 * These bits must be zero.
663 xsave_hdr
->reserved1
[0] = xsave_hdr
->reserved1
[1] = 0;
666 * Init the state that is not present in the memory layout
667 * and enabled by the OS.
669 mask
= ~(pcntxt_mask
& ~mask
);
670 xsave_hdr
->xstate_bv
&= mask
;
675 * Couldn't find the extended state information in the memory
676 * layout. Restore the FP/SSE and init the other extended state
679 xsave_hdr
->xstate_bv
= XSTATE_FPSSE
;
680 return restore_i387_fxsave(buf
, sizeof(struct i387_fxsave_struct
));
683 int restore_i387_xstate_ia32(void __user
*buf
)
686 struct task_struct
*tsk
= current
;
687 struct _fpstate_ia32 __user
*fp
= (struct _fpstate_ia32 __user
*) buf
;
700 if (!access_ok(VERIFY_READ
, buf
, sig_xstate_ia32_size
))
711 err
= restore_i387_xsave(buf
);
712 else if (cpu_has_fxsr
)
713 err
= restore_i387_fxsave(fp
, sizeof(struct
714 i387_fxsave_struct
));
716 err
= restore_i387_fsave(fp
);
718 err
= fpregs_soft_set(current
, NULL
,
719 0, sizeof(struct user_i387_ia32_struct
),
728 * FPU state for core dumps.
729 * This is only used for a.out dumps now.
730 * It is declared generically using elf_fpregset_t (which is
731 * struct user_i387_struct) but is in fact only used for 32-bit
732 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
734 int dump_fpu(struct pt_regs
*regs
, struct user_i387_struct
*fpu
)
736 struct task_struct
*tsk
= current
;
739 fpvalid
= !!used_math();
741 fpvalid
= !fpregs_get(tsk
, NULL
,
742 0, sizeof(struct user_i387_ia32_struct
),
747 EXPORT_SYMBOL(dump_fpu
);
749 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */