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 unsigned int sig_xstate_ia32_size
= sizeof(struct _fpstate_ia32
);
44 static struct i387_fxsave_struct fx_scratch __cpuinitdata
;
46 void __cpuinit
mxcsr_feature_mask_init(void)
48 unsigned long mask
= 0;
52 memset(&fx_scratch
, 0, sizeof(struct i387_fxsave_struct
));
53 asm volatile("fxsave %0" : : "m" (fx_scratch
));
54 mask
= fx_scratch
.mxcsr_mask
;
58 mxcsr_feature_mask
&= mask
;
62 void __cpuinit
init_thread_xstate(void)
65 xstate_size
= sizeof(struct i387_soft_struct
);
75 xstate_size
= sizeof(struct i387_fxsave_struct
);
78 xstate_size
= sizeof(struct i387_fsave_struct
);
84 * Called at bootup to set up the initial FPU state that is later cloned
87 void __cpuinit
fpu_init(void)
89 unsigned long oldcr0
= read_cr0();
91 set_in_cr4(X86_CR4_OSFXSR
);
92 set_in_cr4(X86_CR4_OSXMMEXCPT
);
94 write_cr0(oldcr0
& ~(X86_CR0_TS
|X86_CR0_EM
)); /* clear TS and EM */
97 * Boot processor to setup the FP and extended state context info.
99 if (!smp_processor_id())
100 init_thread_xstate();
103 mxcsr_feature_mask_init();
104 /* clean state in init */
106 current_thread_info()->status
= TS_XSAVE
;
108 current_thread_info()->status
= 0;
111 #endif /* CONFIG_X86_64 */
114 * The _current_ task is using the FPU for the first time
115 * so initialize it and set the mxcsr to its default
116 * value at reset if we support XMM instructions and then
117 * remeber the current task has used the FPU.
119 int init_fpu(struct task_struct
*tsk
)
121 if (tsk_used_math(tsk
)) {
122 if (HAVE_HWFP
&& tsk
== current
)
128 * Memory allocation at the first usage of the FPU and other state.
130 if (!tsk
->thread
.xstate
) {
131 tsk
->thread
.xstate
= kmem_cache_alloc(task_xstate_cachep
,
133 if (!tsk
->thread
.xstate
)
139 memset(tsk
->thread
.xstate
, 0, xstate_size
);
141 set_stopped_child_used_math(tsk
);
147 struct i387_fxsave_struct
*fx
= &tsk
->thread
.xstate
->fxsave
;
149 memset(fx
, 0, xstate_size
);
152 fx
->mxcsr
= MXCSR_DEFAULT
;
154 struct i387_fsave_struct
*fp
= &tsk
->thread
.xstate
->fsave
;
155 memset(fp
, 0, xstate_size
);
156 fp
->cwd
= 0xffff037fu
;
157 fp
->swd
= 0xffff0000u
;
158 fp
->twd
= 0xffffffffu
;
159 fp
->fos
= 0xffff0000u
;
162 * Only the device not available exception or ptrace can call init_fpu.
164 set_stopped_child_used_math(tsk
);
169 * The xstateregs_active() routine is the same as the fpregs_active() routine,
170 * as the "regset->n" for the xstate regset will be updated based on the feature
171 * capabilites supported by the xsave.
173 int fpregs_active(struct task_struct
*target
, const struct user_regset
*regset
)
175 return tsk_used_math(target
) ? regset
->n
: 0;
178 int xfpregs_active(struct task_struct
*target
, const struct user_regset
*regset
)
180 return (cpu_has_fxsr
&& tsk_used_math(target
)) ? regset
->n
: 0;
183 int xfpregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
184 unsigned int pos
, unsigned int count
,
185 void *kbuf
, void __user
*ubuf
)
192 ret
= init_fpu(target
);
196 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
197 &target
->thread
.xstate
->fxsave
, 0, -1);
200 int xfpregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
201 unsigned int pos
, unsigned int count
,
202 const void *kbuf
, const void __user
*ubuf
)
209 ret
= init_fpu(target
);
213 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
214 &target
->thread
.xstate
->fxsave
, 0, -1);
217 * mxcsr reserved bits must be masked to zero for security reasons.
219 target
->thread
.xstate
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
222 * update the header bits in the xsave header, indicating the
223 * presence of FP and SSE state.
226 target
->thread
.xstate
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FPSSE
;
231 int xstateregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
232 unsigned int pos
, unsigned int count
,
233 void *kbuf
, void __user
*ubuf
)
240 ret
= init_fpu(target
);
245 * Copy the 48bytes defined by the software first into the xstate
246 * memory layout in the thread struct, so that we can copy the entire
247 * xstateregs to the user using one user_regset_copyout().
249 memcpy(&target
->thread
.xstate
->fxsave
.sw_reserved
,
250 xstate_fx_sw_bytes
, sizeof(xstate_fx_sw_bytes
));
253 * Copy the xstate memory layout.
255 ret
= user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
256 &target
->thread
.xstate
->xsave
, 0, -1);
260 int xstateregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
261 unsigned int pos
, unsigned int count
,
262 const void *kbuf
, const void __user
*ubuf
)
265 struct xsave_hdr_struct
*xsave_hdr
;
270 ret
= init_fpu(target
);
274 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
275 &target
->thread
.xstate
->xsave
, 0, -1);
278 * mxcsr reserved bits must be masked to zero for security reasons.
280 target
->thread
.xstate
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
282 xsave_hdr
= &target
->thread
.xstate
->xsave
.xsave_hdr
;
284 xsave_hdr
->xstate_bv
&= pcntxt_mask
;
286 * These bits must be zero.
288 xsave_hdr
->reserved1
[0] = xsave_hdr
->reserved1
[1] = 0;
293 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
296 * FPU tag word conversions.
299 static inline unsigned short twd_i387_to_fxsr(unsigned short twd
)
301 unsigned int tmp
; /* to avoid 16 bit prefixes in the code */
303 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
305 tmp
= (tmp
| (tmp
>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
306 /* and move the valid bits to the lower byte. */
307 tmp
= (tmp
| (tmp
>> 1)) & 0x3333; /* 00VV00VV00VV00VV */
308 tmp
= (tmp
| (tmp
>> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
309 tmp
= (tmp
| (tmp
>> 4)) & 0x00ff; /* 00000000VVVVVVVV */
314 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16);
315 #define FP_EXP_TAG_VALID 0
316 #define FP_EXP_TAG_ZERO 1
317 #define FP_EXP_TAG_SPECIAL 2
318 #define FP_EXP_TAG_EMPTY 3
320 static inline u32
twd_fxsr_to_i387(struct i387_fxsave_struct
*fxsave
)
323 u32 tos
= (fxsave
->swd
>> 11) & 7;
324 u32 twd
= (unsigned long) fxsave
->twd
;
326 u32 ret
= 0xffff0000u
;
329 for (i
= 0; i
< 8; i
++, twd
>>= 1) {
331 st
= FPREG_ADDR(fxsave
, (i
- tos
) & 7);
333 switch (st
->exponent
& 0x7fff) {
335 tag
= FP_EXP_TAG_SPECIAL
;
338 if (!st
->significand
[0] &&
339 !st
->significand
[1] &&
340 !st
->significand
[2] &&
342 tag
= FP_EXP_TAG_ZERO
;
344 tag
= FP_EXP_TAG_SPECIAL
;
347 if (st
->significand
[3] & 0x8000)
348 tag
= FP_EXP_TAG_VALID
;
350 tag
= FP_EXP_TAG_SPECIAL
;
354 tag
= FP_EXP_TAG_EMPTY
;
356 ret
|= tag
<< (2 * i
);
362 * FXSR floating point environment conversions.
366 convert_from_fxsr(struct user_i387_ia32_struct
*env
, struct task_struct
*tsk
)
368 struct i387_fxsave_struct
*fxsave
= &tsk
->thread
.xstate
->fxsave
;
369 struct _fpreg
*to
= (struct _fpreg
*) &env
->st_space
[0];
370 struct _fpxreg
*from
= (struct _fpxreg
*) &fxsave
->st_space
[0];
373 env
->cwd
= fxsave
->cwd
| 0xffff0000u
;
374 env
->swd
= fxsave
->swd
| 0xffff0000u
;
375 env
->twd
= twd_fxsr_to_i387(fxsave
);
378 env
->fip
= fxsave
->rip
;
379 env
->foo
= fxsave
->rdp
;
380 if (tsk
== current
) {
382 * should be actually ds/cs at fpu exception time, but
383 * that information is not available in 64bit mode.
385 asm("mov %%ds, %[fos]" : [fos
] "=r" (env
->fos
));
386 asm("mov %%cs, %[fcs]" : [fcs
] "=r" (env
->fcs
));
388 struct pt_regs
*regs
= task_pt_regs(tsk
);
390 env
->fos
= 0xffff0000 | tsk
->thread
.ds
;
394 env
->fip
= fxsave
->fip
;
395 env
->fcs
= (u16
) fxsave
->fcs
| ((u32
) fxsave
->fop
<< 16);
396 env
->foo
= fxsave
->foo
;
397 env
->fos
= fxsave
->fos
;
400 for (i
= 0; i
< 8; ++i
)
401 memcpy(&to
[i
], &from
[i
], sizeof(to
[0]));
404 static void convert_to_fxsr(struct task_struct
*tsk
,
405 const struct user_i387_ia32_struct
*env
)
408 struct i387_fxsave_struct
*fxsave
= &tsk
->thread
.xstate
->fxsave
;
409 struct _fpreg
*from
= (struct _fpreg
*) &env
->st_space
[0];
410 struct _fpxreg
*to
= (struct _fpxreg
*) &fxsave
->st_space
[0];
413 fxsave
->cwd
= env
->cwd
;
414 fxsave
->swd
= env
->swd
;
415 fxsave
->twd
= twd_i387_to_fxsr(env
->twd
);
416 fxsave
->fop
= (u16
) ((u32
) env
->fcs
>> 16);
418 fxsave
->rip
= env
->fip
;
419 fxsave
->rdp
= env
->foo
;
420 /* cs and ds ignored */
422 fxsave
->fip
= env
->fip
;
423 fxsave
->fcs
= (env
->fcs
& 0xffff);
424 fxsave
->foo
= env
->foo
;
425 fxsave
->fos
= env
->fos
;
428 for (i
= 0; i
< 8; ++i
)
429 memcpy(&to
[i
], &from
[i
], sizeof(from
[0]));
432 int fpregs_get(struct task_struct
*target
, const struct user_regset
*regset
,
433 unsigned int pos
, unsigned int count
,
434 void *kbuf
, void __user
*ubuf
)
436 struct user_i387_ia32_struct env
;
439 ret
= init_fpu(target
);
444 return fpregs_soft_get(target
, regset
, pos
, count
, kbuf
, ubuf
);
447 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
,
448 &target
->thread
.xstate
->fsave
, 0,
452 if (kbuf
&& pos
== 0 && count
== sizeof(env
)) {
453 convert_from_fxsr(kbuf
, target
);
457 convert_from_fxsr(&env
, target
);
459 return user_regset_copyout(&pos
, &count
, &kbuf
, &ubuf
, &env
, 0, -1);
462 int fpregs_set(struct task_struct
*target
, const struct user_regset
*regset
,
463 unsigned int pos
, unsigned int count
,
464 const void *kbuf
, const void __user
*ubuf
)
466 struct user_i387_ia32_struct env
;
469 ret
= init_fpu(target
);
474 return fpregs_soft_set(target
, regset
, pos
, count
, kbuf
, ubuf
);
477 return user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
,
478 &target
->thread
.xstate
->fsave
, 0, -1);
481 if (pos
> 0 || count
< sizeof(env
))
482 convert_from_fxsr(&env
, target
);
484 ret
= user_regset_copyin(&pos
, &count
, &kbuf
, &ubuf
, &env
, 0, -1);
486 convert_to_fxsr(target
, &env
);
489 * update the header bit in the xsave header, indicating the
493 target
->thread
.xstate
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FP
;
498 * Signal frame handlers.
501 static inline int save_i387_fsave(struct _fpstate_ia32 __user
*buf
)
503 struct task_struct
*tsk
= current
;
504 struct i387_fsave_struct
*fp
= &tsk
->thread
.xstate
->fsave
;
506 fp
->status
= fp
->swd
;
507 if (__copy_to_user(buf
, fp
, sizeof(struct i387_fsave_struct
)))
512 static int save_i387_fxsave(struct _fpstate_ia32 __user
*buf
)
514 struct task_struct
*tsk
= current
;
515 struct i387_fxsave_struct
*fx
= &tsk
->thread
.xstate
->fxsave
;
516 struct user_i387_ia32_struct env
;
519 convert_from_fxsr(&env
, tsk
);
520 if (__copy_to_user(buf
, &env
, sizeof(env
)))
523 err
|= __put_user(fx
->swd
, &buf
->status
);
524 err
|= __put_user(X86_FXSR_MAGIC
, &buf
->magic
);
528 if (__copy_to_user(&buf
->_fxsr_env
[0], fx
, xstate_size
))
533 static int save_i387_xsave(void __user
*buf
)
535 struct task_struct
*tsk
= current
;
536 struct _fpstate_ia32 __user
*fx
= buf
;
540 * For legacy compatible, we always set FP/SSE bits in the bit
541 * vector while saving the state to the user context.
542 * This will enable us capturing any changes(during sigreturn) to
543 * the FP/SSE bits by the legacy applications which don't touch
544 * xstate_bv in the xsave header.
546 * xsave aware applications can change the xstate_bv in the xsave
547 * header as well as change any contents in the memory layout.
548 * xrestore as part of sigreturn will capture all the changes.
550 tsk
->thread
.xstate
->xsave
.xsave_hdr
.xstate_bv
|= XSTATE_FPSSE
;
552 if (save_i387_fxsave(fx
) < 0)
555 err
= __copy_to_user(&fx
->sw_reserved
, &fx_sw_reserved_ia32
,
556 sizeof(struct _fpx_sw_bytes
));
557 err
|= __put_user(FP_XSTATE_MAGIC2
,
558 (__u32 __user
*) (buf
+ sig_xstate_ia32_size
559 - FP_XSTATE_MAGIC2_SIZE
));
566 int save_i387_xstate_ia32(void __user
*buf
)
568 struct _fpstate_ia32 __user
*fp
= (struct _fpstate_ia32 __user
*) buf
;
569 struct task_struct
*tsk
= current
;
574 if (!access_ok(VERIFY_WRITE
, buf
, sig_xstate_ia32_size
))
577 * This will cause a "finit" to be triggered by the next
578 * attempted FPU operation by the 'current' process.
583 return fpregs_soft_get(current
, NULL
,
584 0, sizeof(struct user_i387_ia32_struct
),
591 return save_i387_xsave(fp
);
593 return save_i387_fxsave(fp
);
595 return save_i387_fsave(fp
);
598 static inline int restore_i387_fsave(struct _fpstate_ia32 __user
*buf
)
600 struct task_struct
*tsk
= current
;
602 return __copy_from_user(&tsk
->thread
.xstate
->fsave
, buf
,
603 sizeof(struct i387_fsave_struct
));
606 static int restore_i387_fxsave(struct _fpstate_ia32 __user
*buf
,
609 struct task_struct
*tsk
= current
;
610 struct user_i387_ia32_struct env
;
613 err
= __copy_from_user(&tsk
->thread
.xstate
->fxsave
, &buf
->_fxsr_env
[0],
615 /* mxcsr reserved bits must be masked to zero for security reasons */
616 tsk
->thread
.xstate
->fxsave
.mxcsr
&= mxcsr_feature_mask
;
617 if (err
|| __copy_from_user(&env
, buf
, sizeof(env
)))
619 convert_to_fxsr(tsk
, &env
);
624 static int restore_i387_xsave(void __user
*buf
)
626 struct _fpx_sw_bytes fx_sw_user
;
627 struct _fpstate_ia32 __user
*fx_user
=
628 ((struct _fpstate_ia32 __user
*) buf
);
629 struct i387_fxsave_struct __user
*fx
=
630 (struct i387_fxsave_struct __user
*) &fx_user
->_fxsr_env
[0];
631 struct xsave_hdr_struct
*xsave_hdr
=
632 ¤t
->thread
.xstate
->xsave
.xsave_hdr
;
636 if (check_for_xstate(fx
, buf
, &fx_sw_user
))
639 mask
= fx_sw_user
.xstate_bv
;
641 err
= restore_i387_fxsave(buf
, fx_sw_user
.xstate_size
);
643 xsave_hdr
->xstate_bv
&= pcntxt_mask
;
645 * These bits must be zero.
647 xsave_hdr
->reserved1
[0] = xsave_hdr
->reserved1
[1] = 0;
650 * Init the state that is not present in the memory layout
651 * and enabled by the OS.
653 mask
= ~(pcntxt_mask
& ~mask
);
654 xsave_hdr
->xstate_bv
&= mask
;
659 * Couldn't find the extended state information in the memory
660 * layout. Restore the FP/SSE and init the other extended state
663 xsave_hdr
->xstate_bv
= XSTATE_FPSSE
;
664 return restore_i387_fxsave(buf
, sizeof(struct i387_fxsave_struct
));
667 int restore_i387_xstate_ia32(void __user
*buf
)
670 struct task_struct
*tsk
= current
;
671 struct _fpstate_ia32 __user
*fp
= (struct _fpstate_ia32 __user
*) buf
;
684 if (!access_ok(VERIFY_READ
, buf
, sig_xstate_ia32_size
))
695 err
= restore_i387_xsave(buf
);
696 else if (cpu_has_fxsr
)
697 err
= restore_i387_fxsave(fp
, sizeof(struct
698 i387_fxsave_struct
));
700 err
= restore_i387_fsave(fp
);
702 err
= fpregs_soft_set(current
, NULL
,
703 0, sizeof(struct user_i387_ia32_struct
),
712 * FPU state for core dumps.
713 * This is only used for a.out dumps now.
714 * It is declared generically using elf_fpregset_t (which is
715 * struct user_i387_struct) but is in fact only used for 32-bit
716 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
718 int dump_fpu(struct pt_regs
*regs
, struct user_i387_struct
*fpu
)
720 struct task_struct
*tsk
= current
;
723 fpvalid
= !!used_math();
725 fpvalid
= !fpregs_get(tsk
, NULL
,
726 0, sizeof(struct user_i387_ia32_struct
),
731 EXPORT_SYMBOL(dump_fpu
);
733 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */