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>
11 * Track whether the kernel is using the FPU state
16 * - by IRQ context code to potentially use the FPU
19 * - to debug kernel_fpu_begin()/end() correctness
21 static DEFINE_PER_CPU(bool, in_kernel_fpu);
23 static void kernel_fpu_disable(void)
25 WARN_ON(this_cpu_read(in_kernel_fpu));
26 this_cpu_write(in_kernel_fpu, true);
29 static void kernel_fpu_enable(void)
31 WARN_ON_ONCE(!this_cpu_read(in_kernel_fpu));
32 this_cpu_write(in_kernel_fpu, false);
35 static bool kernel_fpu_disabled(void)
37 return this_cpu_read(in_kernel_fpu);
41 * Were we in an interrupt that interrupted kernel mode?
43 * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
44 * pair does nothing at all: the thread must not have fpu (so
45 * that we don't try to save the FPU state), and TS must
46 * be set (so that the clts/stts pair does nothing that is
47 * visible in the interrupted kernel thread).
49 * Except for the eagerfpu case when we return true; in the likely case
50 * the thread has FPU but we are not going to set/clear TS.
52 static bool interrupted_kernel_fpu_idle(void)
54 if (kernel_fpu_disabled())
60 return !current->thread.fpu.has_fpu && (read_cr0() & X86_CR0_TS);
64 * Were we in user mode (or vm86 mode) when we were
67 * Doing kernel_fpu_begin/end() is ok if we are running
68 * in an interrupt context from user mode - we'll just
69 * save the FPU state as required.
71 static bool interrupted_user_mode(void)
73 struct pt_regs *regs = get_irq_regs();
74 return regs && user_mode(regs);
78 * Can we use the FPU in kernel mode with the
79 * whole "kernel_fpu_begin/end()" sequence?
81 * It's always ok in process context (ie "not interrupt")
82 * but it is sometimes ok even from an irq.
84 bool irq_fpu_usable(void)
86 return !in_interrupt() ||
87 interrupted_user_mode() ||
88 interrupted_kernel_fpu_idle();
90 EXPORT_SYMBOL(irq_fpu_usable);
92 void __kernel_fpu_begin(void)
94 struct task_struct *me = current;
95 struct fpu *fpu = &me->thread.fpu;
102 this_cpu_write(fpu_owner_task, NULL);
103 if (!use_eager_fpu())
107 EXPORT_SYMBOL(__kernel_fpu_begin);
109 void __kernel_fpu_end(void)
111 struct task_struct *me = current;
112 struct fpu *fpu = &me->thread.fpu;
115 if (WARN_ON(restore_fpu_checking(me)))
117 } else if (!use_eager_fpu()) {
123 EXPORT_SYMBOL(__kernel_fpu_end);
126 * Save the FPU state (initialize it if necessary):
128 * This only ever gets called for the current task.
130 void fpu__save(struct task_struct *tsk)
132 struct fpu *fpu = &tsk->thread.fpu;
134 WARN_ON(tsk != current);
138 if (use_eager_fpu()) {
142 __thread_fpu_end(tsk);
147 EXPORT_SYMBOL_GPL(fpu__save);
149 void fpstate_init(struct fpu *fpu)
152 finit_soft_fpu(&fpu->state->soft);
156 memset(fpu->state, 0, xstate_size);
159 fx_finit(&fpu->state->fxsave);
161 struct i387_fsave_struct *fp = &fpu->state->fsave;
162 fp->cwd = 0xffff037fu;
163 fp->swd = 0xffff0000u;
164 fp->twd = 0xffffffffu;
165 fp->fos = 0xffff0000u;
168 EXPORT_SYMBOL_GPL(fpstate_init);
171 * FPU state allocation:
173 static struct kmem_cache *task_xstate_cachep;
175 void fpstate_cache_init(void)
178 kmem_cache_create("task_xstate", xstate_size,
179 __alignof__(union thread_xstate),
180 SLAB_PANIC | SLAB_NOTRACK, NULL);
184 int fpstate_alloc(struct fpu *fpu)
189 fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
193 /* The CPU requires the FPU state to be aligned to 16 byte boundaries: */
194 WARN_ON((unsigned long)fpu->state & 15);
198 EXPORT_SYMBOL_GPL(fpstate_alloc);
200 void fpstate_free(struct fpu *fpu)
203 kmem_cache_free(task_xstate_cachep, fpu->state);
207 EXPORT_SYMBOL_GPL(fpstate_free);
210 * Copy the current task's FPU state to a new task's FPU context.
212 * In the 'eager' case we just save to the destination context.
214 * In the 'lazy' case we save to the source context, mark the FPU lazy
215 * via stts() and copy the source context into the destination context.
217 static void fpu_copy(struct task_struct *dst, struct task_struct *src)
219 WARN_ON(src != current);
221 if (use_eager_fpu()) {
222 memset(&dst->thread.fpu.state->xsave, 0, xstate_size);
225 struct fpu *dfpu = &dst->thread.fpu;
226 struct fpu *sfpu = &src->thread.fpu;
229 memcpy(dfpu->state, sfpu->state, xstate_size);
233 int fpu__copy(struct task_struct *dst, struct task_struct *src)
235 dst->thread.fpu.counter = 0;
236 dst->thread.fpu.has_fpu = 0;
237 dst->thread.fpu.state = NULL;
239 task_disable_lazy_fpu_restore(dst);
241 if (tsk_used_math(src)) {
242 int err = fpstate_alloc(&dst->thread.fpu);
252 * Allocate the backing store for the current task's FPU registers
253 * and initialize the registers themselves as well.
257 int fpstate_alloc_init(struct task_struct *curr)
261 if (WARN_ON_ONCE(curr != current))
263 if (WARN_ON_ONCE(curr->flags & PF_USED_MATH))
267 * Memory allocation at the first usage of the FPU and other state.
269 ret = fpstate_alloc(&curr->thread.fpu);
273 fpstate_init(&curr->thread.fpu);
275 /* Safe to do for the current task: */
276 curr->flags |= PF_USED_MATH;
280 EXPORT_SYMBOL_GPL(fpstate_alloc_init);
283 * The _current_ task is using the FPU for the first time
284 * so initialize it and set the mxcsr to its default
285 * value at reset if we support XMM instructions and then
286 * remember the current task has used the FPU.
288 static int fpu__unlazy_stopped(struct task_struct *child)
292 if (WARN_ON_ONCE(child == current))
295 if (child->flags & PF_USED_MATH) {
296 task_disable_lazy_fpu_restore(child);
301 * Memory allocation at the first usage of the FPU and other state.
303 ret = fpstate_alloc(&child->thread.fpu);
307 fpstate_init(&child->thread.fpu);
309 /* Safe to do for stopped child tasks: */
310 child->flags |= PF_USED_MATH;
316 * 'fpu__restore()' saves the current math information in the
317 * old math state array, and gets the new ones from the current task
319 * Careful.. There are problems with IBM-designed IRQ13 behaviour.
320 * Don't touch unless you *really* know how it works.
322 * Must be called with kernel preemption disabled (eg with local
323 * local interrupts as in the case of do_device_not_available).
325 void fpu__restore(void)
327 struct task_struct *tsk = current;
329 if (!tsk_used_math(tsk)) {
332 * does a slab alloc which can sleep
334 if (fpstate_alloc_init(tsk)) {
338 do_group_exit(SIGKILL);
344 /* Avoid __kernel_fpu_begin() right after __thread_fpu_begin() */
345 kernel_fpu_disable();
346 __thread_fpu_begin(tsk);
347 if (unlikely(restore_fpu_checking(tsk))) {
348 fpu_reset_state(tsk);
349 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
351 tsk->thread.fpu.counter++;
355 EXPORT_SYMBOL_GPL(fpu__restore);
357 void fpu__flush_thread(struct task_struct *tsk)
359 if (!use_eager_fpu()) {
360 /* FPU state will be reallocated lazily at the first use. */
362 fpstate_free(&tsk->thread.fpu);
364 if (!tsk_used_math(tsk)) {
365 /* kthread execs. TODO: cleanup this horror. */
366 if (WARN_ON(fpstate_alloc_init(tsk)))
367 force_sig(SIGKILL, tsk);
370 restore_init_xstate();
375 * The xstateregs_active() routine is the same as the fpregs_active() routine,
376 * as the "regset->n" for the xstate regset will be updated based on the feature
377 * capabilites supported by the xsave.
379 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
381 return tsk_used_math(target) ? regset->n : 0;
384 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
386 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
389 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
390 unsigned int pos, unsigned int count,
391 void *kbuf, void __user *ubuf)
398 ret = fpu__unlazy_stopped(target);
402 sanitize_i387_state(target);
404 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
405 &target->thread.fpu.state->fxsave, 0, -1);
408 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
409 unsigned int pos, unsigned int count,
410 const void *kbuf, const void __user *ubuf)
417 ret = fpu__unlazy_stopped(target);
421 sanitize_i387_state(target);
423 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
424 &target->thread.fpu.state->fxsave, 0, -1);
427 * mxcsr reserved bits must be masked to zero for security reasons.
429 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
432 * update the header bits in the xsave header, indicating the
433 * presence of FP and SSE state.
436 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
441 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
442 unsigned int pos, unsigned int count,
443 void *kbuf, void __user *ubuf)
445 struct xsave_struct *xsave;
451 ret = fpu__unlazy_stopped(target);
455 xsave = &target->thread.fpu.state->xsave;
458 * Copy the 48bytes defined by the software first into the xstate
459 * memory layout in the thread struct, so that we can copy the entire
460 * xstateregs to the user using one user_regset_copyout().
462 memcpy(&xsave->i387.sw_reserved,
463 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
465 * Copy the xstate memory layout.
467 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
471 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
472 unsigned int pos, unsigned int count,
473 const void *kbuf, const void __user *ubuf)
475 struct xsave_struct *xsave;
481 ret = fpu__unlazy_stopped(target);
485 xsave = &target->thread.fpu.state->xsave;
487 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
489 * mxcsr reserved bits must be masked to zero for security reasons.
491 xsave->i387.mxcsr &= mxcsr_feature_mask;
492 xsave->xsave_hdr.xstate_bv &= pcntxt_mask;
494 * These bits must be zero.
496 memset(&xsave->xsave_hdr.reserved, 0, 48);
500 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
503 * FPU tag word conversions.
506 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
508 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
510 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
512 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
513 /* and move the valid bits to the lower byte. */
514 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
515 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
516 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
521 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
522 #define FP_EXP_TAG_VALID 0
523 #define FP_EXP_TAG_ZERO 1
524 #define FP_EXP_TAG_SPECIAL 2
525 #define FP_EXP_TAG_EMPTY 3
527 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
530 u32 tos = (fxsave->swd >> 11) & 7;
531 u32 twd = (unsigned long) fxsave->twd;
533 u32 ret = 0xffff0000u;
536 for (i = 0; i < 8; i++, twd >>= 1) {
538 st = FPREG_ADDR(fxsave, (i - tos) & 7);
540 switch (st->exponent & 0x7fff) {
542 tag = FP_EXP_TAG_SPECIAL;
545 if (!st->significand[0] &&
546 !st->significand[1] &&
547 !st->significand[2] &&
549 tag = FP_EXP_TAG_ZERO;
551 tag = FP_EXP_TAG_SPECIAL;
554 if (st->significand[3] & 0x8000)
555 tag = FP_EXP_TAG_VALID;
557 tag = FP_EXP_TAG_SPECIAL;
561 tag = FP_EXP_TAG_EMPTY;
563 ret |= tag << (2 * i);
569 * FXSR floating point environment conversions.
573 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
575 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
576 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
577 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
580 env->cwd = fxsave->cwd | 0xffff0000u;
581 env->swd = fxsave->swd | 0xffff0000u;
582 env->twd = twd_fxsr_to_i387(fxsave);
585 env->fip = fxsave->rip;
586 env->foo = fxsave->rdp;
588 * should be actually ds/cs at fpu exception time, but
589 * that information is not available in 64bit mode.
591 env->fcs = task_pt_regs(tsk)->cs;
592 if (tsk == current) {
593 savesegment(ds, env->fos);
595 env->fos = tsk->thread.ds;
597 env->fos |= 0xffff0000;
599 env->fip = fxsave->fip;
600 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
601 env->foo = fxsave->foo;
602 env->fos = fxsave->fos;
605 for (i = 0; i < 8; ++i)
606 memcpy(&to[i], &from[i], sizeof(to[0]));
609 void convert_to_fxsr(struct task_struct *tsk,
610 const struct user_i387_ia32_struct *env)
613 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
614 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
615 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
618 fxsave->cwd = env->cwd;
619 fxsave->swd = env->swd;
620 fxsave->twd = twd_i387_to_fxsr(env->twd);
621 fxsave->fop = (u16) ((u32) env->fcs >> 16);
623 fxsave->rip = env->fip;
624 fxsave->rdp = env->foo;
625 /* cs and ds ignored */
627 fxsave->fip = env->fip;
628 fxsave->fcs = (env->fcs & 0xffff);
629 fxsave->foo = env->foo;
630 fxsave->fos = env->fos;
633 for (i = 0; i < 8; ++i)
634 memcpy(&to[i], &from[i], sizeof(from[0]));
637 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
638 unsigned int pos, unsigned int count,
639 void *kbuf, void __user *ubuf)
641 struct user_i387_ia32_struct env;
644 ret = fpu__unlazy_stopped(target);
648 if (!static_cpu_has(X86_FEATURE_FPU))
649 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
652 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
653 &target->thread.fpu.state->fsave, 0,
656 sanitize_i387_state(target);
658 if (kbuf && pos == 0 && count == sizeof(env)) {
659 convert_from_fxsr(kbuf, target);
663 convert_from_fxsr(&env, target);
665 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
668 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
669 unsigned int pos, unsigned int count,
670 const void *kbuf, const void __user *ubuf)
672 struct user_i387_ia32_struct env;
675 ret = fpu__unlazy_stopped(target);
679 sanitize_i387_state(target);
681 if (!static_cpu_has(X86_FEATURE_FPU))
682 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
685 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
686 &target->thread.fpu.state->fsave, 0,
689 if (pos > 0 || count < sizeof(env))
690 convert_from_fxsr(&env, target);
692 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
694 convert_to_fxsr(target, &env);
697 * update the header bit in the xsave header, indicating the
701 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
706 * FPU state for core dumps.
707 * This is only used for a.out dumps now.
708 * It is declared generically using elf_fpregset_t (which is
709 * struct user_i387_struct) but is in fact only used for 32-bit
710 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
712 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
714 struct task_struct *tsk = current;
717 fpvalid = !!used_math();
719 fpvalid = !fpregs_get(tsk, NULL,
720 0, sizeof(struct user_i387_ia32_struct),
725 EXPORT_SYMBOL(dump_fpu);
727 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */