2 * Copyright (C) 1994 Linus Torvalds
4 * Pentium III FXSR, SSE support
5 * General FPU state handling cleanups
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 static void __cpuinit init_thread_xstate(void)
65 * Note that xstate_size might be overwriten later during
71 * Disable xsave as we do not support it if i387
72 * emulation is enabled.
74 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
75 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
76 xstate_size = sizeof(struct i387_soft_struct);
81 xstate_size = sizeof(struct i387_fxsave_struct);
84 xstate_size = sizeof(struct i387_fsave_struct);
90 * Called at bootup to set up the initial FPU state that is later cloned
94 void __cpuinit fpu_init(void)
96 unsigned long oldcr0 = read_cr0();
98 set_in_cr4(X86_CR4_OSFXSR);
99 set_in_cr4(X86_CR4_OSXMMEXCPT);
101 write_cr0(oldcr0 & ~(X86_CR0_TS|X86_CR0_EM)); /* clear TS and EM */
103 if (!smp_processor_id())
104 init_thread_xstate();
106 mxcsr_feature_mask_init();
107 /* clean state in init */
108 current_thread_info()->status = 0;
112 #else /* CONFIG_X86_64 */
114 void __cpuinit fpu_init(void)
116 if (!smp_processor_id())
117 init_thread_xstate();
120 #endif /* CONFIG_X86_32 */
122 void fpu_finit(struct fpu *fpu)
126 finit_soft_fpu(&fpu->state->soft);
132 struct i387_fxsave_struct *fx = &fpu->state->fxsave;
134 memset(fx, 0, xstate_size);
137 fx->mxcsr = MXCSR_DEFAULT;
139 struct i387_fsave_struct *fp = &fpu->state->fsave;
140 memset(fp, 0, xstate_size);
141 fp->cwd = 0xffff037fu;
142 fp->swd = 0xffff0000u;
143 fp->twd = 0xffffffffu;
144 fp->fos = 0xffff0000u;
147 EXPORT_SYMBOL_GPL(fpu_finit);
150 * The _current_ task is using the FPU for the first time
151 * so initialize it and set the mxcsr to its default
152 * value at reset if we support XMM instructions and then
153 * remeber the current task has used the FPU.
155 int init_fpu(struct task_struct *tsk)
159 if (tsk_used_math(tsk)) {
160 if (HAVE_HWFP && tsk == current)
166 * Memory allocation at the first usage of the FPU and other state.
168 ret = fpu_alloc(&tsk->thread.fpu);
172 fpu_finit(&tsk->thread.fpu);
174 set_stopped_child_used_math(tsk);
179 * The xstateregs_active() routine is the same as the fpregs_active() routine,
180 * as the "regset->n" for the xstate regset will be updated based on the feature
181 * capabilites supported by the xsave.
183 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
185 return tsk_used_math(target) ? regset->n : 0;
188 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
190 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
193 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
194 unsigned int pos, unsigned int count,
195 void *kbuf, void __user *ubuf)
202 ret = init_fpu(target);
206 sanitize_i387_state(target);
208 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
209 &target->thread.fpu.state->fxsave, 0, -1);
212 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
213 unsigned int pos, unsigned int count,
214 const void *kbuf, const void __user *ubuf)
221 ret = init_fpu(target);
225 sanitize_i387_state(target);
227 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
228 &target->thread.fpu.state->fxsave, 0, -1);
231 * mxcsr reserved bits must be masked to zero for security reasons.
233 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
236 * update the header bits in the xsave header, indicating the
237 * presence of FP and SSE state.
240 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
245 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
246 unsigned int pos, unsigned int count,
247 void *kbuf, void __user *ubuf)
254 ret = init_fpu(target);
259 * Copy the 48bytes defined by the software first into the xstate
260 * memory layout in the thread struct, so that we can copy the entire
261 * xstateregs to the user using one user_regset_copyout().
263 memcpy(&target->thread.fpu.state->fxsave.sw_reserved,
264 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
267 * Copy the xstate memory layout.
269 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf,
270 &target->thread.fpu.state->xsave, 0, -1);
274 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
275 unsigned int pos, unsigned int count,
276 const void *kbuf, const void __user *ubuf)
279 struct xsave_hdr_struct *xsave_hdr;
284 ret = init_fpu(target);
288 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
289 &target->thread.fpu.state->xsave, 0, -1);
292 * mxcsr reserved bits must be masked to zero for security reasons.
294 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
296 xsave_hdr = &target->thread.fpu.state->xsave.xsave_hdr;
298 xsave_hdr->xstate_bv &= pcntxt_mask;
300 * These bits must be zero.
302 xsave_hdr->reserved1[0] = xsave_hdr->reserved1[1] = 0;
307 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
310 * FPU tag word conversions.
313 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
315 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
317 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
319 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
320 /* and move the valid bits to the lower byte. */
321 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
322 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
323 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
328 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16);
329 #define FP_EXP_TAG_VALID 0
330 #define FP_EXP_TAG_ZERO 1
331 #define FP_EXP_TAG_SPECIAL 2
332 #define FP_EXP_TAG_EMPTY 3
334 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
337 u32 tos = (fxsave->swd >> 11) & 7;
338 u32 twd = (unsigned long) fxsave->twd;
340 u32 ret = 0xffff0000u;
343 for (i = 0; i < 8; i++, twd >>= 1) {
345 st = FPREG_ADDR(fxsave, (i - tos) & 7);
347 switch (st->exponent & 0x7fff) {
349 tag = FP_EXP_TAG_SPECIAL;
352 if (!st->significand[0] &&
353 !st->significand[1] &&
354 !st->significand[2] &&
356 tag = FP_EXP_TAG_ZERO;
358 tag = FP_EXP_TAG_SPECIAL;
361 if (st->significand[3] & 0x8000)
362 tag = FP_EXP_TAG_VALID;
364 tag = FP_EXP_TAG_SPECIAL;
368 tag = FP_EXP_TAG_EMPTY;
370 ret |= tag << (2 * i);
376 * FXSR floating point environment conversions.
380 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
382 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
383 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
384 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
387 env->cwd = fxsave->cwd | 0xffff0000u;
388 env->swd = fxsave->swd | 0xffff0000u;
389 env->twd = twd_fxsr_to_i387(fxsave);
392 env->fip = fxsave->rip;
393 env->foo = fxsave->rdp;
394 if (tsk == current) {
396 * should be actually ds/cs at fpu exception time, but
397 * that information is not available in 64bit mode.
399 asm("mov %%ds, %[fos]" : [fos] "=r" (env->fos));
400 asm("mov %%cs, %[fcs]" : [fcs] "=r" (env->fcs));
402 struct pt_regs *regs = task_pt_regs(tsk);
404 env->fos = 0xffff0000 | tsk->thread.ds;
408 env->fip = fxsave->fip;
409 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
410 env->foo = fxsave->foo;
411 env->fos = fxsave->fos;
414 for (i = 0; i < 8; ++i)
415 memcpy(&to[i], &from[i], sizeof(to[0]));
418 static void convert_to_fxsr(struct task_struct *tsk,
419 const struct user_i387_ia32_struct *env)
422 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
423 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
424 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
427 fxsave->cwd = env->cwd;
428 fxsave->swd = env->swd;
429 fxsave->twd = twd_i387_to_fxsr(env->twd);
430 fxsave->fop = (u16) ((u32) env->fcs >> 16);
432 fxsave->rip = env->fip;
433 fxsave->rdp = env->foo;
434 /* cs and ds ignored */
436 fxsave->fip = env->fip;
437 fxsave->fcs = (env->fcs & 0xffff);
438 fxsave->foo = env->foo;
439 fxsave->fos = env->fos;
442 for (i = 0; i < 8; ++i)
443 memcpy(&to[i], &from[i], sizeof(from[0]));
446 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
447 unsigned int pos, unsigned int count,
448 void *kbuf, void __user *ubuf)
450 struct user_i387_ia32_struct env;
453 ret = init_fpu(target);
458 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
461 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
462 &target->thread.fpu.state->fsave, 0,
466 sanitize_i387_state(target);
468 if (kbuf && pos == 0 && count == sizeof(env)) {
469 convert_from_fxsr(kbuf, target);
473 convert_from_fxsr(&env, target);
475 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
478 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
479 unsigned int pos, unsigned int count,
480 const void *kbuf, const void __user *ubuf)
482 struct user_i387_ia32_struct env;
485 ret = init_fpu(target);
489 sanitize_i387_state(target);
492 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
495 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
496 &target->thread.fpu.state->fsave, 0, -1);
499 if (pos > 0 || count < sizeof(env))
500 convert_from_fxsr(&env, target);
502 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
504 convert_to_fxsr(target, &env);
507 * update the header bit in the xsave header, indicating the
511 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
516 * Signal frame handlers.
519 static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
521 struct task_struct *tsk = current;
522 struct i387_fsave_struct *fp = &tsk->thread.fpu.state->fsave;
524 fp->status = fp->swd;
525 if (__copy_to_user(buf, fp, sizeof(struct i387_fsave_struct)))
530 static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
532 struct task_struct *tsk = current;
533 struct i387_fxsave_struct *fx = &tsk->thread.fpu.state->fxsave;
534 struct user_i387_ia32_struct env;
537 convert_from_fxsr(&env, tsk);
538 if (__copy_to_user(buf, &env, sizeof(env)))
541 err |= __put_user(fx->swd, &buf->status);
542 err |= __put_user(X86_FXSR_MAGIC, &buf->magic);
546 if (__copy_to_user(&buf->_fxsr_env[0], fx, xstate_size))
551 static int save_i387_xsave(void __user *buf)
553 struct task_struct *tsk = current;
554 struct _fpstate_ia32 __user *fx = buf;
558 sanitize_i387_state(tsk);
561 * For legacy compatible, we always set FP/SSE bits in the bit
562 * vector while saving the state to the user context.
563 * This will enable us capturing any changes(during sigreturn) to
564 * the FP/SSE bits by the legacy applications which don't touch
565 * xstate_bv in the xsave header.
567 * xsave aware applications can change the xstate_bv in the xsave
568 * header as well as change any contents in the memory layout.
569 * xrestore as part of sigreturn will capture all the changes.
571 tsk->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
573 if (save_i387_fxsave(fx) < 0)
576 err = __copy_to_user(&fx->sw_reserved, &fx_sw_reserved_ia32,
577 sizeof(struct _fpx_sw_bytes));
578 err |= __put_user(FP_XSTATE_MAGIC2,
579 (__u32 __user *) (buf + sig_xstate_ia32_size
580 - FP_XSTATE_MAGIC2_SIZE));
587 int save_i387_xstate_ia32(void __user *buf)
589 struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
590 struct task_struct *tsk = current;
595 if (!access_ok(VERIFY_WRITE, buf, sig_xstate_ia32_size))
598 * This will cause a "finit" to be triggered by the next
599 * attempted FPU operation by the 'current' process.
604 return fpregs_soft_get(current, NULL,
605 0, sizeof(struct user_i387_ia32_struct),
612 return save_i387_xsave(fp);
614 return save_i387_fxsave(fp);
616 return save_i387_fsave(fp);
619 static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
621 struct task_struct *tsk = current;
623 return __copy_from_user(&tsk->thread.fpu.state->fsave, buf,
624 sizeof(struct i387_fsave_struct));
627 static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf,
630 struct task_struct *tsk = current;
631 struct user_i387_ia32_struct env;
634 err = __copy_from_user(&tsk->thread.fpu.state->fxsave, &buf->_fxsr_env[0],
636 /* mxcsr reserved bits must be masked to zero for security reasons */
637 tsk->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
638 if (err || __copy_from_user(&env, buf, sizeof(env)))
640 convert_to_fxsr(tsk, &env);
645 static int restore_i387_xsave(void __user *buf)
647 struct _fpx_sw_bytes fx_sw_user;
648 struct _fpstate_ia32 __user *fx_user =
649 ((struct _fpstate_ia32 __user *) buf);
650 struct i387_fxsave_struct __user *fx =
651 (struct i387_fxsave_struct __user *) &fx_user->_fxsr_env[0];
652 struct xsave_hdr_struct *xsave_hdr =
653 ¤t->thread.fpu.state->xsave.xsave_hdr;
657 if (check_for_xstate(fx, buf, &fx_sw_user))
660 mask = fx_sw_user.xstate_bv;
662 err = restore_i387_fxsave(buf, fx_sw_user.xstate_size);
664 xsave_hdr->xstate_bv &= pcntxt_mask;
666 * These bits must be zero.
668 xsave_hdr->reserved1[0] = xsave_hdr->reserved1[1] = 0;
671 * Init the state that is not present in the memory layout
672 * and enabled by the OS.
674 mask = ~(pcntxt_mask & ~mask);
675 xsave_hdr->xstate_bv &= mask;
680 * Couldn't find the extended state information in the memory
681 * layout. Restore the FP/SSE and init the other extended state
684 xsave_hdr->xstate_bv = XSTATE_FPSSE;
685 return restore_i387_fxsave(buf, sizeof(struct i387_fxsave_struct));
688 int restore_i387_xstate_ia32(void __user *buf)
691 struct task_struct *tsk = current;
692 struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
705 if (!access_ok(VERIFY_READ, buf, sig_xstate_ia32_size))
716 err = restore_i387_xsave(buf);
717 else if (cpu_has_fxsr)
718 err = restore_i387_fxsave(fp, sizeof(struct
719 i387_fxsave_struct));
721 err = restore_i387_fsave(fp);
723 err = fpregs_soft_set(current, NULL,
724 0, sizeof(struct user_i387_ia32_struct),
733 * FPU state for core dumps.
734 * This is only used for a.out dumps now.
735 * It is declared generically using elf_fpregset_t (which is
736 * struct user_i387_struct) but is in fact only used for 32-bit
737 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
739 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
741 struct task_struct *tsk = current;
744 fpvalid = !!used_math();
746 fpvalid = !fpregs_get(tsk, NULL,
747 0, sizeof(struct user_i387_ia32_struct),
752 EXPORT_SYMBOL(dump_fpu);
754 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */