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);
150 int fpstate_alloc(struct fpu *fpu)
155 fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
159 /* The CPU requires the FPU state to be aligned to 16 byte boundaries: */
160 WARN_ON((unsigned long)fpu->state & 15);
164 EXPORT_SYMBOL_GPL(fpstate_alloc);
167 * Allocate the backing store for the current task's FPU registers
168 * and initialize the registers themselves as well.
172 int fpstate_alloc_init(struct task_struct *curr)
176 if (WARN_ON_ONCE(curr != current))
178 if (WARN_ON_ONCE(curr->flags & PF_USED_MATH))
182 * Memory allocation at the first usage of the FPU and other state.
184 ret = fpstate_alloc(&curr->thread.fpu);
188 fpstate_init(&curr->thread.fpu);
190 /* Safe to do for the current task: */
191 curr->flags |= PF_USED_MATH;
195 EXPORT_SYMBOL_GPL(fpstate_alloc_init);
198 * The _current_ task is using the FPU for the first time
199 * so initialize it and set the mxcsr to its default
200 * value at reset if we support XMM instructions and then
201 * remember the current task has used the FPU.
203 static int fpu__unlazy_stopped(struct task_struct *child)
207 if (WARN_ON_ONCE(child == current))
210 if (child->flags & PF_USED_MATH) {
211 task_disable_lazy_fpu_restore(child);
216 * Memory allocation at the first usage of the FPU and other state.
218 ret = fpstate_alloc(&child->thread.fpu);
222 fpstate_init(&child->thread.fpu);
224 /* Safe to do for stopped child tasks: */
225 child->flags |= PF_USED_MATH;
231 * The xstateregs_active() routine is the same as the fpregs_active() routine,
232 * as the "regset->n" for the xstate regset will be updated based on the feature
233 * capabilites supported by the xsave.
235 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
237 return tsk_used_math(target) ? regset->n : 0;
240 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
242 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
245 int xfpregs_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 = fpu__unlazy_stopped(target);
258 sanitize_i387_state(target);
260 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
261 &target->thread.fpu.state->fxsave, 0, -1);
264 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
265 unsigned int pos, unsigned int count,
266 const void *kbuf, const void __user *ubuf)
273 ret = fpu__unlazy_stopped(target);
277 sanitize_i387_state(target);
279 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
280 &target->thread.fpu.state->fxsave, 0, -1);
283 * mxcsr reserved bits must be masked to zero for security reasons.
285 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
288 * update the header bits in the xsave header, indicating the
289 * presence of FP and SSE state.
292 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
297 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
298 unsigned int pos, unsigned int count,
299 void *kbuf, void __user *ubuf)
301 struct xsave_struct *xsave;
307 ret = fpu__unlazy_stopped(target);
311 xsave = &target->thread.fpu.state->xsave;
314 * Copy the 48bytes defined by the software first into the xstate
315 * memory layout in the thread struct, so that we can copy the entire
316 * xstateregs to the user using one user_regset_copyout().
318 memcpy(&xsave->i387.sw_reserved,
319 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
321 * Copy the xstate memory layout.
323 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
327 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
328 unsigned int pos, unsigned int count,
329 const void *kbuf, const void __user *ubuf)
331 struct xsave_struct *xsave;
337 ret = fpu__unlazy_stopped(target);
341 xsave = &target->thread.fpu.state->xsave;
343 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
345 * mxcsr reserved bits must be masked to zero for security reasons.
347 xsave->i387.mxcsr &= mxcsr_feature_mask;
348 xsave->xsave_hdr.xstate_bv &= pcntxt_mask;
350 * These bits must be zero.
352 memset(&xsave->xsave_hdr.reserved, 0, 48);
356 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
359 * FPU tag word conversions.
362 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
364 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
366 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
368 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
369 /* and move the valid bits to the lower byte. */
370 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
371 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
372 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
377 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
378 #define FP_EXP_TAG_VALID 0
379 #define FP_EXP_TAG_ZERO 1
380 #define FP_EXP_TAG_SPECIAL 2
381 #define FP_EXP_TAG_EMPTY 3
383 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
386 u32 tos = (fxsave->swd >> 11) & 7;
387 u32 twd = (unsigned long) fxsave->twd;
389 u32 ret = 0xffff0000u;
392 for (i = 0; i < 8; i++, twd >>= 1) {
394 st = FPREG_ADDR(fxsave, (i - tos) & 7);
396 switch (st->exponent & 0x7fff) {
398 tag = FP_EXP_TAG_SPECIAL;
401 if (!st->significand[0] &&
402 !st->significand[1] &&
403 !st->significand[2] &&
405 tag = FP_EXP_TAG_ZERO;
407 tag = FP_EXP_TAG_SPECIAL;
410 if (st->significand[3] & 0x8000)
411 tag = FP_EXP_TAG_VALID;
413 tag = FP_EXP_TAG_SPECIAL;
417 tag = FP_EXP_TAG_EMPTY;
419 ret |= tag << (2 * i);
425 * FXSR floating point environment conversions.
429 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
431 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
432 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
433 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
436 env->cwd = fxsave->cwd | 0xffff0000u;
437 env->swd = fxsave->swd | 0xffff0000u;
438 env->twd = twd_fxsr_to_i387(fxsave);
441 env->fip = fxsave->rip;
442 env->foo = fxsave->rdp;
444 * should be actually ds/cs at fpu exception time, but
445 * that information is not available in 64bit mode.
447 env->fcs = task_pt_regs(tsk)->cs;
448 if (tsk == current) {
449 savesegment(ds, env->fos);
451 env->fos = tsk->thread.ds;
453 env->fos |= 0xffff0000;
455 env->fip = fxsave->fip;
456 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
457 env->foo = fxsave->foo;
458 env->fos = fxsave->fos;
461 for (i = 0; i < 8; ++i)
462 memcpy(&to[i], &from[i], sizeof(to[0]));
465 void convert_to_fxsr(struct task_struct *tsk,
466 const struct user_i387_ia32_struct *env)
469 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
470 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
471 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
474 fxsave->cwd = env->cwd;
475 fxsave->swd = env->swd;
476 fxsave->twd = twd_i387_to_fxsr(env->twd);
477 fxsave->fop = (u16) ((u32) env->fcs >> 16);
479 fxsave->rip = env->fip;
480 fxsave->rdp = env->foo;
481 /* cs and ds ignored */
483 fxsave->fip = env->fip;
484 fxsave->fcs = (env->fcs & 0xffff);
485 fxsave->foo = env->foo;
486 fxsave->fos = env->fos;
489 for (i = 0; i < 8; ++i)
490 memcpy(&to[i], &from[i], sizeof(from[0]));
493 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
494 unsigned int pos, unsigned int count,
495 void *kbuf, void __user *ubuf)
497 struct user_i387_ia32_struct env;
500 ret = fpu__unlazy_stopped(target);
504 if (!static_cpu_has(X86_FEATURE_FPU))
505 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
508 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
509 &target->thread.fpu.state->fsave, 0,
512 sanitize_i387_state(target);
514 if (kbuf && pos == 0 && count == sizeof(env)) {
515 convert_from_fxsr(kbuf, target);
519 convert_from_fxsr(&env, target);
521 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
524 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
525 unsigned int pos, unsigned int count,
526 const void *kbuf, const void __user *ubuf)
528 struct user_i387_ia32_struct env;
531 ret = fpu__unlazy_stopped(target);
535 sanitize_i387_state(target);
537 if (!static_cpu_has(X86_FEATURE_FPU))
538 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
541 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
542 &target->thread.fpu.state->fsave, 0,
545 if (pos > 0 || count < sizeof(env))
546 convert_from_fxsr(&env, target);
548 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
550 convert_to_fxsr(target, &env);
553 * update the header bit in the xsave header, indicating the
557 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
562 * FPU state for core dumps.
563 * This is only used for a.out dumps now.
564 * It is declared generically using elf_fpregset_t (which is
565 * struct user_i387_struct) but is in fact only used for 32-bit
566 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
568 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
570 struct task_struct *tsk = current;
573 fpvalid = !!used_math();
575 fpvalid = !fpregs_get(tsk, NULL,
576 0, sizeof(struct user_i387_ia32_struct),
581 EXPORT_SYMBOL(dump_fpu);
583 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */