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 <asm/fpu-internal.h>
10 static DEFINE_PER_CPU(bool, in_kernel_fpu);
12 void kernel_fpu_disable(void)
14 WARN_ON(this_cpu_read(in_kernel_fpu));
15 this_cpu_write(in_kernel_fpu, true);
18 void kernel_fpu_enable(void)
20 this_cpu_write(in_kernel_fpu, false);
24 * Were we in an interrupt that interrupted kernel mode?
26 * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
27 * pair does nothing at all: the thread must not have fpu (so
28 * that we don't try to save the FPU state), and TS must
29 * be set (so that the clts/stts pair does nothing that is
30 * visible in the interrupted kernel thread).
32 * Except for the eagerfpu case when we return true; in the likely case
33 * the thread has FPU but we are not going to set/clear TS.
35 static inline bool interrupted_kernel_fpu_idle(void)
37 if (this_cpu_read(in_kernel_fpu))
43 return !__thread_has_fpu(current) &&
44 (read_cr0() & X86_CR0_TS);
48 * Were we in user mode (or vm86 mode) when we were
51 * Doing kernel_fpu_begin/end() is ok if we are running
52 * in an interrupt context from user mode - we'll just
53 * save the FPU state as required.
55 static inline bool interrupted_user_mode(void)
57 struct pt_regs *regs = get_irq_regs();
58 return regs && user_mode(regs);
62 * Can we use the FPU in kernel mode with the
63 * whole "kernel_fpu_begin/end()" sequence?
65 * It's always ok in process context (ie "not interrupt")
66 * but it is sometimes ok even from an irq.
68 bool irq_fpu_usable(void)
70 return !in_interrupt() ||
71 interrupted_user_mode() ||
72 interrupted_kernel_fpu_idle();
74 EXPORT_SYMBOL(irq_fpu_usable);
76 void __kernel_fpu_begin(void)
78 struct task_struct *me = current;
80 this_cpu_write(in_kernel_fpu, true);
82 if (__thread_has_fpu(me)) {
85 this_cpu_write(fpu_owner_task, NULL);
90 EXPORT_SYMBOL(__kernel_fpu_begin);
92 void __kernel_fpu_end(void)
94 struct task_struct *me = current;
96 if (__thread_has_fpu(me)) {
97 if (WARN_ON(restore_fpu_checking(me)))
99 } else if (!use_eager_fpu()) {
103 this_cpu_write(in_kernel_fpu, false);
105 EXPORT_SYMBOL(__kernel_fpu_end);
108 * Save the FPU state (initialize it if necessary):
110 * This only ever gets called for the current task.
112 void fpu__save(struct task_struct *tsk)
114 WARN_ON(tsk != current);
117 if (__thread_has_fpu(tsk)) {
118 if (use_eager_fpu()) {
121 __save_init_fpu(tsk);
122 __thread_fpu_end(tsk);
127 EXPORT_SYMBOL_GPL(fpu__save);
129 void fpstate_init(struct fpu *fpu)
132 finit_soft_fpu(&fpu->state->soft);
136 memset(fpu->state, 0, xstate_size);
139 fx_finit(&fpu->state->fxsave);
141 struct i387_fsave_struct *fp = &fpu->state->fsave;
142 fp->cwd = 0xffff037fu;
143 fp->swd = 0xffff0000u;
144 fp->twd = 0xffffffffu;
145 fp->fos = 0xffff0000u;
148 EXPORT_SYMBOL_GPL(fpstate_init);
151 * FPU state allocation:
153 static struct kmem_cache *task_xstate_cachep;
155 void fpstate_cache_init(void)
158 kmem_cache_create("task_xstate", xstate_size,
159 __alignof__(union thread_xstate),
160 SLAB_PANIC | SLAB_NOTRACK, NULL);
164 int fpstate_alloc(struct fpu *fpu)
169 fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
173 /* The CPU requires the FPU state to be aligned to 16 byte boundaries: */
174 WARN_ON((unsigned long)fpu->state & 15);
178 EXPORT_SYMBOL_GPL(fpstate_alloc);
180 void fpstate_free(struct fpu *fpu)
183 kmem_cache_free(task_xstate_cachep, fpu->state);
187 EXPORT_SYMBOL_GPL(fpstate_free);
189 int fpu__copy(struct task_struct *dst, struct task_struct *src)
191 dst->thread.fpu.counter = 0;
192 dst->thread.fpu.has_fpu = 0;
193 dst->thread.fpu.state = NULL;
195 task_disable_lazy_fpu_restore(dst);
197 if (tsk_used_math(src)) {
198 int err = fpstate_alloc(&dst->thread.fpu);
208 * Allocate the backing store for the current task's FPU registers
209 * and initialize the registers themselves as well.
213 int fpstate_alloc_init(struct task_struct *curr)
217 if (WARN_ON_ONCE(curr != current))
219 if (WARN_ON_ONCE(curr->flags & PF_USED_MATH))
223 * Memory allocation at the first usage of the FPU and other state.
225 ret = fpstate_alloc(&curr->thread.fpu);
229 fpstate_init(&curr->thread.fpu);
231 /* Safe to do for the current task: */
232 curr->flags |= PF_USED_MATH;
236 EXPORT_SYMBOL_GPL(fpstate_alloc_init);
239 * The _current_ task is using the FPU for the first time
240 * so initialize it and set the mxcsr to its default
241 * value at reset if we support XMM instructions and then
242 * remember the current task has used the FPU.
244 static int fpu__unlazy_stopped(struct task_struct *child)
248 if (WARN_ON_ONCE(child == current))
251 if (child->flags & PF_USED_MATH) {
252 task_disable_lazy_fpu_restore(child);
257 * Memory allocation at the first usage of the FPU and other state.
259 ret = fpstate_alloc(&child->thread.fpu);
263 fpstate_init(&child->thread.fpu);
265 /* Safe to do for stopped child tasks: */
266 child->flags |= PF_USED_MATH;
272 * 'fpu__restore()' saves the current math information in the
273 * old math state array, and gets the new ones from the current task
275 * Careful.. There are problems with IBM-designed IRQ13 behaviour.
276 * Don't touch unless you *really* know how it works.
278 * Must be called with kernel preemption disabled (eg with local
279 * local interrupts as in the case of do_device_not_available).
281 void fpu__restore(void)
283 struct task_struct *tsk = current;
285 if (!tsk_used_math(tsk)) {
288 * does a slab alloc which can sleep
290 if (fpstate_alloc_init(tsk)) {
294 do_group_exit(SIGKILL);
300 /* Avoid __kernel_fpu_begin() right after __thread_fpu_begin() */
301 kernel_fpu_disable();
302 __thread_fpu_begin(tsk);
303 if (unlikely(restore_fpu_checking(tsk))) {
304 fpu_reset_state(tsk);
305 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
307 tsk->thread.fpu.counter++;
311 EXPORT_SYMBOL_GPL(fpu__restore);
313 void fpu__flush_thread(struct task_struct *tsk)
315 if (!use_eager_fpu()) {
316 /* FPU state will be reallocated lazily at the first use. */
318 fpstate_free(&tsk->thread.fpu);
320 if (!tsk_used_math(tsk)) {
321 /* kthread execs. TODO: cleanup this horror. */
322 if (WARN_ON(fpstate_alloc_init(tsk)))
323 force_sig(SIGKILL, tsk);
326 restore_init_xstate();
331 * The xstateregs_active() routine is the same as the fpregs_active() routine,
332 * as the "regset->n" for the xstate regset will be updated based on the feature
333 * capabilites supported by the xsave.
335 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
337 return tsk_used_math(target) ? regset->n : 0;
340 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
342 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
345 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
346 unsigned int pos, unsigned int count,
347 void *kbuf, void __user *ubuf)
354 ret = fpu__unlazy_stopped(target);
358 sanitize_i387_state(target);
360 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
361 &target->thread.fpu.state->fxsave, 0, -1);
364 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
365 unsigned int pos, unsigned int count,
366 const void *kbuf, const void __user *ubuf)
373 ret = fpu__unlazy_stopped(target);
377 sanitize_i387_state(target);
379 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
380 &target->thread.fpu.state->fxsave, 0, -1);
383 * mxcsr reserved bits must be masked to zero for security reasons.
385 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
388 * update the header bits in the xsave header, indicating the
389 * presence of FP and SSE state.
392 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
397 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
398 unsigned int pos, unsigned int count,
399 void *kbuf, void __user *ubuf)
401 struct xsave_struct *xsave;
407 ret = fpu__unlazy_stopped(target);
411 xsave = &target->thread.fpu.state->xsave;
414 * Copy the 48bytes defined by the software first into the xstate
415 * memory layout in the thread struct, so that we can copy the entire
416 * xstateregs to the user using one user_regset_copyout().
418 memcpy(&xsave->i387.sw_reserved,
419 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
421 * Copy the xstate memory layout.
423 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
427 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
428 unsigned int pos, unsigned int count,
429 const void *kbuf, const void __user *ubuf)
431 struct xsave_struct *xsave;
437 ret = fpu__unlazy_stopped(target);
441 xsave = &target->thread.fpu.state->xsave;
443 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
445 * mxcsr reserved bits must be masked to zero for security reasons.
447 xsave->i387.mxcsr &= mxcsr_feature_mask;
448 xsave->xsave_hdr.xstate_bv &= pcntxt_mask;
450 * These bits must be zero.
452 memset(&xsave->xsave_hdr.reserved, 0, 48);
456 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
459 * FPU tag word conversions.
462 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
464 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
466 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
468 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
469 /* and move the valid bits to the lower byte. */
470 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
471 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
472 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
477 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
478 #define FP_EXP_TAG_VALID 0
479 #define FP_EXP_TAG_ZERO 1
480 #define FP_EXP_TAG_SPECIAL 2
481 #define FP_EXP_TAG_EMPTY 3
483 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
486 u32 tos = (fxsave->swd >> 11) & 7;
487 u32 twd = (unsigned long) fxsave->twd;
489 u32 ret = 0xffff0000u;
492 for (i = 0; i < 8; i++, twd >>= 1) {
494 st = FPREG_ADDR(fxsave, (i - tos) & 7);
496 switch (st->exponent & 0x7fff) {
498 tag = FP_EXP_TAG_SPECIAL;
501 if (!st->significand[0] &&
502 !st->significand[1] &&
503 !st->significand[2] &&
505 tag = FP_EXP_TAG_ZERO;
507 tag = FP_EXP_TAG_SPECIAL;
510 if (st->significand[3] & 0x8000)
511 tag = FP_EXP_TAG_VALID;
513 tag = FP_EXP_TAG_SPECIAL;
517 tag = FP_EXP_TAG_EMPTY;
519 ret |= tag << (2 * i);
525 * FXSR floating point environment conversions.
529 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
531 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
532 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
533 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
536 env->cwd = fxsave->cwd | 0xffff0000u;
537 env->swd = fxsave->swd | 0xffff0000u;
538 env->twd = twd_fxsr_to_i387(fxsave);
541 env->fip = fxsave->rip;
542 env->foo = fxsave->rdp;
544 * should be actually ds/cs at fpu exception time, but
545 * that information is not available in 64bit mode.
547 env->fcs = task_pt_regs(tsk)->cs;
548 if (tsk == current) {
549 savesegment(ds, env->fos);
551 env->fos = tsk->thread.ds;
553 env->fos |= 0xffff0000;
555 env->fip = fxsave->fip;
556 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
557 env->foo = fxsave->foo;
558 env->fos = fxsave->fos;
561 for (i = 0; i < 8; ++i)
562 memcpy(&to[i], &from[i], sizeof(to[0]));
565 void convert_to_fxsr(struct task_struct *tsk,
566 const struct user_i387_ia32_struct *env)
569 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
570 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
571 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
574 fxsave->cwd = env->cwd;
575 fxsave->swd = env->swd;
576 fxsave->twd = twd_i387_to_fxsr(env->twd);
577 fxsave->fop = (u16) ((u32) env->fcs >> 16);
579 fxsave->rip = env->fip;
580 fxsave->rdp = env->foo;
581 /* cs and ds ignored */
583 fxsave->fip = env->fip;
584 fxsave->fcs = (env->fcs & 0xffff);
585 fxsave->foo = env->foo;
586 fxsave->fos = env->fos;
589 for (i = 0; i < 8; ++i)
590 memcpy(&to[i], &from[i], sizeof(from[0]));
593 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
594 unsigned int pos, unsigned int count,
595 void *kbuf, void __user *ubuf)
597 struct user_i387_ia32_struct env;
600 ret = fpu__unlazy_stopped(target);
604 if (!static_cpu_has(X86_FEATURE_FPU))
605 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
608 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
609 &target->thread.fpu.state->fsave, 0,
612 sanitize_i387_state(target);
614 if (kbuf && pos == 0 && count == sizeof(env)) {
615 convert_from_fxsr(kbuf, target);
619 convert_from_fxsr(&env, target);
621 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
624 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
625 unsigned int pos, unsigned int count,
626 const void *kbuf, const void __user *ubuf)
628 struct user_i387_ia32_struct env;
631 ret = fpu__unlazy_stopped(target);
635 sanitize_i387_state(target);
637 if (!static_cpu_has(X86_FEATURE_FPU))
638 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
641 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
642 &target->thread.fpu.state->fsave, 0,
645 if (pos > 0 || count < sizeof(env))
646 convert_from_fxsr(&env, target);
648 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
650 convert_to_fxsr(target, &env);
653 * update the header bit in the xsave header, indicating the
657 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
662 * FPU state for core dumps.
663 * This is only used for a.out dumps now.
664 * It is declared generically using elf_fpregset_t (which is
665 * struct user_i387_struct) but is in fact only used for 32-bit
666 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
668 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
670 struct task_struct *tsk = current;
673 fpvalid = !!used_math();
675 fpvalid = !fpregs_get(tsk, NULL,
676 0, sizeof(struct user_i387_ia32_struct),
681 EXPORT_SYMBOL(dump_fpu);
683 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */