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
)
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
);
172 EXPORT_SYMBOL_GPL(init_fpu
);
175 * The xstateregs_active() routine is the same as the fpregs_active() routine,
176 * as the "regset->n" for the xstate regset will be updated based on the feature
177 * capabilites supported by the xsave.
179 int fpregs_active(struct task_struct
*target
, const struct user_regset
*regset
)
181 return tsk_used_math(target
) ? regset
->n
: 0;
184 int xfpregs_active(struct task_struct
*target
, const struct user_regset
*regset
)
186 return (cpu_has_fxsr
&& tsk_used_math(target
)) ? regset
->n
: 0;
189 int xfpregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
190 unsigned int pos
, unsigned int count
,
191 void *kbuf
, void __user
*ubuf
)
198 ret
= init_fpu(target
);
202 sanitize_i387_state(target
);
204 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
205 &target
->thread
.fpu
.state
->fxsave
, 0, -1);
208 int xfpregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
209 unsigned int pos
, unsigned int count
,
210 const void *kbuf
, const void __user
*ubuf
)
217 ret
= init_fpu(target
);
221 sanitize_i387_state(target
);
223 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
224 &target
->thread
.fpu
.state
->fxsave
, 0, -1);
227 * mxcsr reserved bits must be masked to zero for security reasons.
229 target
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
232 * update the header bits in the xsave header, indicating the
233 * presence of FP and SSE state.
236 target
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FPSSE
;
241 int xstateregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
242 unsigned int pos
, unsigned int count
,
243 void *kbuf
, void __user
*ubuf
)
250 ret
= init_fpu(target
);
255 * Copy the 48bytes defined by the software first into the xstate
256 * memory layout in the thread struct, so that we can copy the entire
257 * xstateregs to the user using one user_regset_copyout().
259 memcpy(&target
->thread
.fpu
.state
->fxsave
.sw_reserved
,
260 xstate_fx_sw_bytes
, sizeof(xstate_fx_sw_bytes
));
263 * Copy the xstate memory layout.
265 ret
= user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
266 &target
->thread
.fpu
.state
->xsave
, 0, -1);
270 int xstateregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
271 unsigned int pos
, unsigned int count
,
272 const void *kbuf
, const void __user
*ubuf
)
275 struct xsave_hdr_struct
*xsave_hdr
;
280 ret
= init_fpu(target
);
284 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
285 &target
->thread
.fpu
.state
->xsave
, 0, -1);
288 * mxcsr reserved bits must be masked to zero for security reasons.
290 target
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
292 xsave_hdr
= &target
->thread
.fpu
.state
->xsave
.xsave_hdr
;
294 xsave_hdr
->xstate_bv
&= pcntxt_mask
;
296 * These bits must be zero.
298 xsave_hdr
->reserved1
[0] = xsave_hdr
->reserved1
[1] = 0;
303 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
306 * FPU tag word conversions.
309 static inline unsigned short twd_i387_to_fxsr(unsigned short twd
)
311 unsigned int tmp
; /* to avoid 16 bit prefixes in the code */
313 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
315 tmp
= (tmp
| (tmp
>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
316 /* and move the valid bits to the lower byte. */
317 tmp
= (tmp
| (tmp
>> 1)) & 0x3333; /* 00VV00VV00VV00VV */
318 tmp
= (tmp
| (tmp
>> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
319 tmp
= (tmp
| (tmp
>> 4)) & 0x00ff; /* 00000000VVVVVVVV */
324 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
325 #define FP_EXP_TAG_VALID 0
326 #define FP_EXP_TAG_ZERO 1
327 #define FP_EXP_TAG_SPECIAL 2
328 #define FP_EXP_TAG_EMPTY 3
330 static inline u32
twd_fxsr_to_i387(struct i387_fxsave_struct
*fxsave
)
333 u32 tos
= (fxsave
->swd
>> 11) & 7;
334 u32 twd
= (unsigned long) fxsave
->twd
;
336 u32 ret
= 0xffff0000u
;
339 for (i
= 0; i
< 8; i
++, twd
>>= 1) {
341 st
= FPREG_ADDR(fxsave
, (i
- tos
) & 7);
343 switch (st
->exponent
& 0x7fff) {
345 tag
= FP_EXP_TAG_SPECIAL
;
348 if (!st
->significand
[0] &&
349 !st
->significand
[1] &&
350 !st
->significand
[2] &&
352 tag
= FP_EXP_TAG_ZERO
;
354 tag
= FP_EXP_TAG_SPECIAL
;
357 if (st
->significand
[3] & 0x8000)
358 tag
= FP_EXP_TAG_VALID
;
360 tag
= FP_EXP_TAG_SPECIAL
;
364 tag
= FP_EXP_TAG_EMPTY
;
366 ret
|= tag
<< (2 * i
);
372 * FXSR floating point environment conversions.
376 convert_from_fxsr(struct user_i387_ia32_struct
*env
, struct task_struct
*tsk
)
378 struct i387_fxsave_struct
*fxsave
= &tsk
->thread
.fpu
.state
->fxsave
;
379 struct _fpreg
*to
= (struct _fpreg
*) &env
->st_space
[0];
380 struct _fpxreg
*from
= (struct _fpxreg
*) &fxsave
->st_space
[0];
383 env
->cwd
= fxsave
->cwd
| 0xffff0000u
;
384 env
->swd
= fxsave
->swd
| 0xffff0000u
;
385 env
->twd
= twd_fxsr_to_i387(fxsave
);
388 env
->fip
= fxsave
->rip
;
389 env
->foo
= fxsave
->rdp
;
391 * should be actually ds/cs at fpu exception time, but
392 * that information is not available in 64bit mode.
394 env
->fcs
= task_pt_regs(tsk
)->cs
;
395 if (tsk
== current
) {
396 savesegment(ds
, env
->fos
);
398 env
->fos
= tsk
->thread
.ds
;
400 env
->fos
|= 0xffff0000;
402 env
->fip
= fxsave
->fip
;
403 env
->fcs
= (u16
) fxsave
->fcs
| ((u32
) fxsave
->fop
<< 16);
404 env
->foo
= fxsave
->foo
;
405 env
->fos
= fxsave
->fos
;
408 for (i
= 0; i
< 8; ++i
)
409 memcpy(&to
[i
], &from
[i
], sizeof(to
[0]));
412 static void convert_to_fxsr(struct task_struct
*tsk
,
413 const struct user_i387_ia32_struct
*env
)
416 struct i387_fxsave_struct
*fxsave
= &tsk
->thread
.fpu
.state
->fxsave
;
417 struct _fpreg
*from
= (struct _fpreg
*) &env
->st_space
[0];
418 struct _fpxreg
*to
= (struct _fpxreg
*) &fxsave
->st_space
[0];
421 fxsave
->cwd
= env
->cwd
;
422 fxsave
->swd
= env
->swd
;
423 fxsave
->twd
= twd_i387_to_fxsr(env
->twd
);
424 fxsave
->fop
= (u16
) ((u32
) env
->fcs
>> 16);
426 fxsave
->rip
= env
->fip
;
427 fxsave
->rdp
= env
->foo
;
428 /* cs and ds ignored */
430 fxsave
->fip
= env
->fip
;
431 fxsave
->fcs
= (env
->fcs
& 0xffff);
432 fxsave
->foo
= env
->foo
;
433 fxsave
->fos
= env
->fos
;
436 for (i
= 0; i
< 8; ++i
)
437 memcpy(&to
[i
], &from
[i
], sizeof(from
[0]));
440 int fpregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
441 unsigned int pos
, unsigned int count
,
442 void *kbuf
, void __user
*ubuf
)
444 struct user_i387_ia32_struct env
;
447 ret
= init_fpu(target
);
452 return fpregs_soft_get(target
, regset
, pos
, count
, kbuf
, ubuf
);
455 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
456 &target
->thread
.fpu
.state
->fsave
, 0,
460 sanitize_i387_state(target
);
462 if (kbuf
&& pos
== 0 && count
== sizeof(env
)) {
463 convert_from_fxsr(kbuf
, target
);
467 convert_from_fxsr(&env
, target
);
469 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
, &env
, 0, -1);
472 int fpregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
473 unsigned int pos
, unsigned int count
,
474 const void *kbuf
, const void __user
*ubuf
)
476 struct user_i387_ia32_struct env
;
479 ret
= init_fpu(target
);
483 sanitize_i387_state(target
);
486 return fpregs_soft_set(target
, regset
, pos
, count
, kbuf
, ubuf
);
489 return user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
490 &target
->thread
.fpu
.state
->fsave
, 0, -1);
493 if (pos
> 0 || count
< sizeof(env
))
494 convert_from_fxsr(&env
, target
);
496 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
, &env
, 0, -1);
498 convert_to_fxsr(target
, &env
);
501 * update the header bit in the xsave header, indicating the
505 target
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FP
;
510 * Signal frame handlers.
513 static inline int save_i387_fsave(struct _fpstate_ia32 __user
*buf
)
515 struct task_struct
*tsk
= current
;
516 struct i387_fsave_struct
*fp
= &tsk
->thread
.fpu
.state
->fsave
;
518 fp
->status
= fp
->swd
;
519 if (__copy_to_user(buf
, fp
, sizeof(struct i387_fsave_struct
)))
524 static int save_i387_fxsave(struct _fpstate_ia32 __user
*buf
)
526 struct task_struct
*tsk
= current
;
527 struct i387_fxsave_struct
*fx
= &tsk
->thread
.fpu
.state
->fxsave
;
528 struct user_i387_ia32_struct env
;
531 convert_from_fxsr(&env
, tsk
);
532 if (__copy_to_user(buf
, &env
, sizeof(env
)))
535 err
|= __put_user(fx
->swd
, &buf
->status
);
536 err
|= __put_user(X86_FXSR_MAGIC
, &buf
->magic
);
540 if (__copy_to_user(&buf
->_fxsr_env
[0], fx
, xstate_size
))
545 static int save_i387_xsave(void __user
*buf
)
547 struct task_struct
*tsk
= current
;
548 struct _fpstate_ia32 __user
*fx
= buf
;
552 sanitize_i387_state(tsk
);
555 * For legacy compatible, we always set FP/SSE bits in the bit
556 * vector while saving the state to the user context.
557 * This will enable us capturing any changes(during sigreturn) to
558 * the FP/SSE bits by the legacy applications which don't touch
559 * xstate_bv in the xsave header.
561 * xsave aware applications can change the xstate_bv in the xsave
562 * header as well as change any contents in the memory layout.
563 * xrestore as part of sigreturn will capture all the changes.
565 tsk
->thread
.fpu
.state
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FPSSE
;
567 if (save_i387_fxsave(fx
) < 0)
570 err
= __copy_to_user(&fx
->sw_reserved
, &fx_sw_reserved_ia32
,
571 sizeof(struct _fpx_sw_bytes
));
572 err
|= __put_user(FP_XSTATE_MAGIC2
,
573 (__u32 __user
*) (buf
+ sig_xstate_ia32_size
574 - FP_XSTATE_MAGIC2_SIZE
));
581 int save_i387_xstate_ia32(void __user
*buf
)
583 struct _fpstate_ia32 __user
*fp
= (struct _fpstate_ia32 __user
*) buf
;
584 struct task_struct
*tsk
= current
;
589 if (!access_ok(VERIFY_WRITE
, buf
, sig_xstate_ia32_size
))
592 * This will cause a "finit" to be triggered by the next
593 * attempted FPU operation by the 'current' process.
598 return fpregs_soft_get(current
, NULL
,
599 0, sizeof(struct user_i387_ia32_struct
),
606 return save_i387_xsave(fp
);
608 return save_i387_fxsave(fp
);
610 return save_i387_fsave(fp
);
613 static inline int restore_i387_fsave(struct _fpstate_ia32 __user
*buf
)
615 struct task_struct
*tsk
= current
;
617 return __copy_from_user(&tsk
->thread
.fpu
.state
->fsave
, buf
,
618 sizeof(struct i387_fsave_struct
));
621 static int restore_i387_fxsave(struct _fpstate_ia32 __user
*buf
,
624 struct task_struct
*tsk
= current
;
625 struct user_i387_ia32_struct env
;
628 err
= __copy_from_user(&tsk
->thread
.fpu
.state
->fxsave
, &buf
->_fxsr_env
[0],
630 /* mxcsr reserved bits must be masked to zero for security reasons */
631 tsk
->thread
.fpu
.state
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
632 if (err
|| __copy_from_user(&env
, buf
, sizeof(env
)))
634 convert_to_fxsr(tsk
, &env
);
639 static int restore_i387_xsave(void __user
*buf
)
641 struct _fpx_sw_bytes fx_sw_user
;
642 struct _fpstate_ia32 __user
*fx_user
=
643 ((struct _fpstate_ia32 __user
*) buf
);
644 struct i387_fxsave_struct __user
*fx
=
645 (struct i387_fxsave_struct __user
*) &fx_user
->_fxsr_env
[0];
646 struct xsave_hdr_struct
*xsave_hdr
=
647 ¤t
->thread
.fpu
.state
->xsave
.xsave_hdr
;
651 if (check_for_xstate(fx
, buf
, &fx_sw_user
))
654 mask
= fx_sw_user
.xstate_bv
;
656 err
= restore_i387_fxsave(buf
, fx_sw_user
.xstate_size
);
658 xsave_hdr
->xstate_bv
&= pcntxt_mask
;
660 * These bits must be zero.
662 xsave_hdr
->reserved1
[0] = xsave_hdr
->reserved1
[1] = 0;
665 * Init the state that is not present in the memory layout
666 * and enabled by the OS.
668 mask
= ~(pcntxt_mask
& ~mask
);
669 xsave_hdr
->xstate_bv
&= mask
;
674 * Couldn't find the extended state information in the memory
675 * layout. Restore the FP/SSE and init the other extended state
678 xsave_hdr
->xstate_bv
= XSTATE_FPSSE
;
679 return restore_i387_fxsave(buf
, sizeof(struct i387_fxsave_struct
));
682 int restore_i387_xstate_ia32(void __user
*buf
)
685 struct task_struct
*tsk
= current
;
686 struct _fpstate_ia32 __user
*fp
= (struct _fpstate_ia32 __user
*) buf
;
699 if (!access_ok(VERIFY_READ
, buf
, sig_xstate_ia32_size
))
710 err
= restore_i387_xsave(buf
);
711 else if (cpu_has_fxsr
)
712 err
= restore_i387_fxsave(fp
, sizeof(struct
713 i387_fxsave_struct
));
715 err
= restore_i387_fsave(fp
);
717 err
= fpregs_soft_set(current
, NULL
,
718 0, sizeof(struct user_i387_ia32_struct
),
727 * FPU state for core dumps.
728 * This is only used for a.out dumps now.
729 * It is declared generically using elf_fpregset_t (which is
730 * struct user_i387_struct) but is in fact only used for 32-bit
731 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
733 int dump_fpu(struct pt_regs
*regs
, struct user_i387_struct
*fpu
)
735 struct task_struct
*tsk
= current
;
738 fpvalid
= !!used_math();
740 fpvalid
= !fpregs_get(tsk
, NULL
,
741 0, sizeof(struct user_i387_ia32_struct
),
746 EXPORT_SYMBOL(dump_fpu
);
748 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */