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);
24 * Track which context is using the FPU on the CPU:
26 DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
28 static void kernel_fpu_disable(void)
30 WARN_ON(this_cpu_read(in_kernel_fpu));
31 this_cpu_write(in_kernel_fpu, true);
34 static void kernel_fpu_enable(void)
36 WARN_ON_ONCE(!this_cpu_read(in_kernel_fpu));
37 this_cpu_write(in_kernel_fpu, false);
40 static bool kernel_fpu_disabled(void)
42 return this_cpu_read(in_kernel_fpu);
46 * Were we in an interrupt that interrupted kernel mode?
48 * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
49 * pair does nothing at all: the thread must not have fpu (so
50 * that we don't try to save the FPU state), and TS must
51 * be set (so that the clts/stts pair does nothing that is
52 * visible in the interrupted kernel thread).
54 * Except for the eagerfpu case when we return true; in the likely case
55 * the thread has FPU but we are not going to set/clear TS.
57 static bool interrupted_kernel_fpu_idle(void)
59 if (kernel_fpu_disabled())
65 return !current->thread.fpu.has_fpu && (read_cr0() & X86_CR0_TS);
69 * Were we in user mode (or vm86 mode) when we were
72 * Doing kernel_fpu_begin/end() is ok if we are running
73 * in an interrupt context from user mode - we'll just
74 * save the FPU state as required.
76 static bool interrupted_user_mode(void)
78 struct pt_regs *regs = get_irq_regs();
79 return regs && user_mode(regs);
83 * Can we use the FPU in kernel mode with the
84 * whole "kernel_fpu_begin/end()" sequence?
86 * It's always ok in process context (ie "not interrupt")
87 * but it is sometimes ok even from an irq.
89 bool irq_fpu_usable(void)
91 return !in_interrupt() ||
92 interrupted_user_mode() ||
93 interrupted_kernel_fpu_idle();
95 EXPORT_SYMBOL(irq_fpu_usable);
97 void __kernel_fpu_begin(void)
99 struct fpu *fpu = ¤t->thread.fpu;
101 kernel_fpu_disable();
106 this_cpu_write(fpu_fpregs_owner_ctx, NULL);
107 if (!use_eager_fpu())
111 EXPORT_SYMBOL(__kernel_fpu_begin);
113 void __kernel_fpu_end(void)
115 struct fpu *fpu = ¤t->thread.fpu;
118 if (WARN_ON(restore_fpu_checking(fpu)))
119 fpu_reset_state(fpu);
120 } else if (!use_eager_fpu()) {
126 EXPORT_SYMBOL(__kernel_fpu_end);
129 * Save the FPU state (initialize it if necessary):
131 * This only ever gets called for the current task.
133 void fpu__save(struct task_struct *tsk)
135 struct fpu *fpu = &tsk->thread.fpu;
137 WARN_ON(tsk != current);
141 if (use_eager_fpu()) {
145 __thread_fpu_end(fpu);
150 EXPORT_SYMBOL_GPL(fpu__save);
152 void fpstate_init(struct fpu *fpu)
155 finit_soft_fpu(&fpu->state->soft);
159 memset(fpu->state, 0, xstate_size);
162 fx_finit(&fpu->state->fxsave);
164 struct i387_fsave_struct *fp = &fpu->state->fsave;
165 fp->cwd = 0xffff037fu;
166 fp->swd = 0xffff0000u;
167 fp->twd = 0xffffffffu;
168 fp->fos = 0xffff0000u;
171 EXPORT_SYMBOL_GPL(fpstate_init);
174 * FPU state allocation:
176 static struct kmem_cache *task_xstate_cachep;
178 void fpstate_cache_init(void)
181 kmem_cache_create("task_xstate", xstate_size,
182 __alignof__(union thread_xstate),
183 SLAB_PANIC | SLAB_NOTRACK, NULL);
187 int fpstate_alloc(struct fpu *fpu)
192 fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
196 /* The CPU requires the FPU state to be aligned to 16 byte boundaries: */
197 WARN_ON((unsigned long)fpu->state & 15);
201 EXPORT_SYMBOL_GPL(fpstate_alloc);
203 void fpstate_free(struct fpu *fpu)
206 kmem_cache_free(task_xstate_cachep, fpu->state);
210 EXPORT_SYMBOL_GPL(fpstate_free);
213 * Copy the current task's FPU state to a new task's FPU context.
215 * In the 'eager' case we just save to the destination context.
217 * In the 'lazy' case we save to the source context, mark the FPU lazy
218 * via stts() and copy the source context into the destination context.
220 static void fpu_copy(struct task_struct *dst, struct task_struct *src)
222 WARN_ON(src != current);
224 if (use_eager_fpu()) {
225 memset(&dst->thread.fpu.state->xsave, 0, xstate_size);
228 struct fpu *dfpu = &dst->thread.fpu;
229 struct fpu *sfpu = &src->thread.fpu;
232 memcpy(dfpu->state, sfpu->state, xstate_size);
236 int fpu__copy(struct task_struct *dst, struct task_struct *src)
238 struct fpu *dst_fpu = &dst->thread.fpu;
239 struct fpu *src_fpu = &src->thread.fpu;
241 dst->thread.fpu.counter = 0;
242 dst->thread.fpu.has_fpu = 0;
243 dst->thread.fpu.state = NULL;
244 dst->thread.fpu.last_cpu = -1;
246 if (src_fpu->fpstate_active) {
247 int err = fpstate_alloc(dst_fpu);
257 * Allocate the backing store for the current task's FPU registers
258 * and initialize the registers themselves as well.
262 int fpstate_alloc_init(struct task_struct *curr)
264 struct fpu *fpu = &curr->thread.fpu;
267 if (WARN_ON_ONCE(curr != current))
269 if (WARN_ON_ONCE(fpu->fpstate_active))
273 * Memory allocation at the first usage of the FPU and other state.
275 ret = fpstate_alloc(&curr->thread.fpu);
279 fpstate_init(&curr->thread.fpu);
281 /* Safe to do for the current task: */
282 fpu->fpstate_active = 1;
286 EXPORT_SYMBOL_GPL(fpstate_alloc_init);
289 * This function is called before we modify a stopped child's
292 * If the child has not used the FPU before then initialize its
295 * If the child has used the FPU before then unlazy it.
297 * [ After this function call, after the context is modified and
298 * the child task is woken up, the child task will restore
299 * the modified FPU state from the modified context. If we
300 * didn't clear its lazy status here then the lazy in-registers
301 * state pending on its former CPU could be restored, losing
302 * the modifications. ]
304 * This function is also called before we read a stopped child's
305 * FPU state - to make sure it's modified.
307 * TODO: A future optimization would be to skip the unlazying in
308 * the read-only case, it's not strictly necessary for
309 * read-only access to the context.
311 static int fpu__unlazy_stopped(struct task_struct *child)
313 struct fpu *child_fpu = &child->thread.fpu;
316 if (WARN_ON_ONCE(child == current))
319 if (child_fpu->fpstate_active) {
320 child->thread.fpu.last_cpu = -1;
325 * Memory allocation at the first usage of the FPU and other state.
327 ret = fpstate_alloc(&child->thread.fpu);
331 fpstate_init(&child->thread.fpu);
333 /* Safe to do for stopped child tasks: */
334 child_fpu->fpstate_active = 1;
340 * 'fpu__restore()' saves the current math information in the
341 * old math state array, and gets the new ones from the current task
343 * Careful.. There are problems with IBM-designed IRQ13 behaviour.
344 * Don't touch unless you *really* know how it works.
346 * Must be called with kernel preemption disabled (eg with local
347 * local interrupts as in the case of do_device_not_available).
349 void fpu__restore(void)
351 struct task_struct *tsk = current;
352 struct fpu *fpu = &tsk->thread.fpu;
354 if (!fpu->fpstate_active) {
357 * does a slab alloc which can sleep
359 if (fpstate_alloc_init(tsk)) {
363 do_group_exit(SIGKILL);
369 /* Avoid __kernel_fpu_begin() right after __thread_fpu_begin() */
370 kernel_fpu_disable();
371 __thread_fpu_begin(fpu);
372 if (unlikely(restore_fpu_checking(fpu))) {
373 fpu_reset_state(fpu);
374 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
376 tsk->thread.fpu.counter++;
380 EXPORT_SYMBOL_GPL(fpu__restore);
382 void fpu__flush_thread(struct task_struct *tsk)
384 struct fpu *fpu = &tsk->thread.fpu;
386 WARN_ON(tsk != current);
388 if (!use_eager_fpu()) {
389 /* FPU state will be reallocated lazily at the first use. */
391 fpstate_free(&tsk->thread.fpu);
393 if (!fpu->fpstate_active) {
394 /* kthread execs. TODO: cleanup this horror. */
395 if (WARN_ON(fpstate_alloc_init(tsk)))
396 force_sig(SIGKILL, tsk);
399 restore_init_xstate();
404 * The xstateregs_active() routine is the same as the fpregs_active() routine,
405 * as the "regset->n" for the xstate regset will be updated based on the feature
406 * capabilites supported by the xsave.
408 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
410 struct fpu *target_fpu = &target->thread.fpu;
412 return target_fpu->fpstate_active ? regset->n : 0;
415 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
417 struct fpu *target_fpu = &target->thread.fpu;
419 return (cpu_has_fxsr && target_fpu->fpstate_active) ? regset->n : 0;
422 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
423 unsigned int pos, unsigned int count,
424 void *kbuf, void __user *ubuf)
431 ret = fpu__unlazy_stopped(target);
435 sanitize_i387_state(target);
437 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
438 &target->thread.fpu.state->fxsave, 0, -1);
441 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
442 unsigned int pos, unsigned int count,
443 const void *kbuf, const void __user *ubuf)
450 ret = fpu__unlazy_stopped(target);
454 sanitize_i387_state(target);
456 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
457 &target->thread.fpu.state->fxsave, 0, -1);
460 * mxcsr reserved bits must be masked to zero for security reasons.
462 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
465 * update the header bits in the xsave header, indicating the
466 * presence of FP and SSE state.
469 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
474 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
475 unsigned int pos, unsigned int count,
476 void *kbuf, void __user *ubuf)
478 struct xsave_struct *xsave;
484 ret = fpu__unlazy_stopped(target);
488 xsave = &target->thread.fpu.state->xsave;
491 * Copy the 48bytes defined by the software first into the xstate
492 * memory layout in the thread struct, so that we can copy the entire
493 * xstateregs to the user using one user_regset_copyout().
495 memcpy(&xsave->i387.sw_reserved,
496 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
498 * Copy the xstate memory layout.
500 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
504 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
505 unsigned int pos, unsigned int count,
506 const void *kbuf, const void __user *ubuf)
508 struct xsave_struct *xsave;
514 ret = fpu__unlazy_stopped(target);
518 xsave = &target->thread.fpu.state->xsave;
520 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
522 * mxcsr reserved bits must be masked to zero for security reasons.
524 xsave->i387.mxcsr &= mxcsr_feature_mask;
525 xsave->xsave_hdr.xstate_bv &= pcntxt_mask;
527 * These bits must be zero.
529 memset(&xsave->xsave_hdr.reserved, 0, 48);
533 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
536 * FPU tag word conversions.
539 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
541 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
543 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
545 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
546 /* and move the valid bits to the lower byte. */
547 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
548 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
549 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
554 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
555 #define FP_EXP_TAG_VALID 0
556 #define FP_EXP_TAG_ZERO 1
557 #define FP_EXP_TAG_SPECIAL 2
558 #define FP_EXP_TAG_EMPTY 3
560 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
563 u32 tos = (fxsave->swd >> 11) & 7;
564 u32 twd = (unsigned long) fxsave->twd;
566 u32 ret = 0xffff0000u;
569 for (i = 0; i < 8; i++, twd >>= 1) {
571 st = FPREG_ADDR(fxsave, (i - tos) & 7);
573 switch (st->exponent & 0x7fff) {
575 tag = FP_EXP_TAG_SPECIAL;
578 if (!st->significand[0] &&
579 !st->significand[1] &&
580 !st->significand[2] &&
582 tag = FP_EXP_TAG_ZERO;
584 tag = FP_EXP_TAG_SPECIAL;
587 if (st->significand[3] & 0x8000)
588 tag = FP_EXP_TAG_VALID;
590 tag = FP_EXP_TAG_SPECIAL;
594 tag = FP_EXP_TAG_EMPTY;
596 ret |= tag << (2 * i);
602 * FXSR floating point environment conversions.
606 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
608 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
609 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
610 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
613 env->cwd = fxsave->cwd | 0xffff0000u;
614 env->swd = fxsave->swd | 0xffff0000u;
615 env->twd = twd_fxsr_to_i387(fxsave);
618 env->fip = fxsave->rip;
619 env->foo = fxsave->rdp;
621 * should be actually ds/cs at fpu exception time, but
622 * that information is not available in 64bit mode.
624 env->fcs = task_pt_regs(tsk)->cs;
625 if (tsk == current) {
626 savesegment(ds, env->fos);
628 env->fos = tsk->thread.ds;
630 env->fos |= 0xffff0000;
632 env->fip = fxsave->fip;
633 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
634 env->foo = fxsave->foo;
635 env->fos = fxsave->fos;
638 for (i = 0; i < 8; ++i)
639 memcpy(&to[i], &from[i], sizeof(to[0]));
642 void convert_to_fxsr(struct task_struct *tsk,
643 const struct user_i387_ia32_struct *env)
646 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
647 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
648 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
651 fxsave->cwd = env->cwd;
652 fxsave->swd = env->swd;
653 fxsave->twd = twd_i387_to_fxsr(env->twd);
654 fxsave->fop = (u16) ((u32) env->fcs >> 16);
656 fxsave->rip = env->fip;
657 fxsave->rdp = env->foo;
658 /* cs and ds ignored */
660 fxsave->fip = env->fip;
661 fxsave->fcs = (env->fcs & 0xffff);
662 fxsave->foo = env->foo;
663 fxsave->fos = env->fos;
666 for (i = 0; i < 8; ++i)
667 memcpy(&to[i], &from[i], sizeof(from[0]));
670 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
671 unsigned int pos, unsigned int count,
672 void *kbuf, void __user *ubuf)
674 struct user_i387_ia32_struct env;
677 ret = fpu__unlazy_stopped(target);
681 if (!static_cpu_has(X86_FEATURE_FPU))
682 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
685 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
686 &target->thread.fpu.state->fsave, 0,
689 sanitize_i387_state(target);
691 if (kbuf && pos == 0 && count == sizeof(env)) {
692 convert_from_fxsr(kbuf, target);
696 convert_from_fxsr(&env, target);
698 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
701 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
702 unsigned int pos, unsigned int count,
703 const void *kbuf, const void __user *ubuf)
705 struct user_i387_ia32_struct env;
708 ret = fpu__unlazy_stopped(target);
712 sanitize_i387_state(target);
714 if (!static_cpu_has(X86_FEATURE_FPU))
715 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
718 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
719 &target->thread.fpu.state->fsave, 0,
722 if (pos > 0 || count < sizeof(env))
723 convert_from_fxsr(&env, target);
725 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
727 convert_to_fxsr(target, &env);
730 * update the header bit in the xsave header, indicating the
734 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
739 * FPU state for core dumps.
740 * This is only used for a.out dumps now.
741 * It is declared generically using elf_fpregset_t (which is
742 * struct user_i387_struct) but is in fact only used for 32-bit
743 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
745 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *ufpu)
747 struct task_struct *tsk = current;
748 struct fpu *fpu = &tsk->thread.fpu;
751 fpvalid = fpu->fpstate_active;
753 fpvalid = !fpregs_get(tsk, NULL,
754 0, sizeof(struct user_i387_ia32_struct),
759 EXPORT_SYMBOL(dump_fpu);
761 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */