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);
128 static void __save_fpu(struct fpu *fpu)
131 if (unlikely(system_state == SYSTEM_BOOTING))
132 xsave_state_booting(&fpu->state->xsave);
134 xsave_state(&fpu->state->xsave);
141 * Save the FPU state (initialize it if necessary):
143 * This only ever gets called for the current task.
145 void fpu__save(struct task_struct *tsk)
147 struct fpu *fpu = &tsk->thread.fpu;
149 WARN_ON(tsk != current);
153 if (use_eager_fpu()) {
157 __thread_fpu_end(fpu);
162 EXPORT_SYMBOL_GPL(fpu__save);
164 void fpstate_init(struct fpu *fpu)
167 finit_soft_fpu(&fpu->state->soft);
171 memset(fpu->state, 0, xstate_size);
174 fx_finit(&fpu->state->fxsave);
176 struct i387_fsave_struct *fp = &fpu->state->fsave;
177 fp->cwd = 0xffff037fu;
178 fp->swd = 0xffff0000u;
179 fp->twd = 0xffffffffu;
180 fp->fos = 0xffff0000u;
183 EXPORT_SYMBOL_GPL(fpstate_init);
186 * FPU state allocation:
188 static struct kmem_cache *task_xstate_cachep;
190 void fpstate_cache_init(void)
193 kmem_cache_create("task_xstate", xstate_size,
194 __alignof__(union thread_xstate),
195 SLAB_PANIC | SLAB_NOTRACK, NULL);
199 int fpstate_alloc(struct fpu *fpu)
204 fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
208 /* The CPU requires the FPU state to be aligned to 16 byte boundaries: */
209 WARN_ON((unsigned long)fpu->state & 15);
213 EXPORT_SYMBOL_GPL(fpstate_alloc);
215 void fpstate_free(struct fpu *fpu)
218 kmem_cache_free(task_xstate_cachep, fpu->state);
222 EXPORT_SYMBOL_GPL(fpstate_free);
225 * Copy the current task's FPU state to a new task's FPU context.
227 * In the 'eager' case we just save to the destination context.
229 * In the 'lazy' case we save to the source context, mark the FPU lazy
230 * via stts() and copy the source context into the destination context.
232 static void fpu_copy(struct task_struct *dst, struct task_struct *src)
234 struct fpu *dst_fpu = &dst->thread.fpu;
235 struct fpu *src_fpu = &src->thread.fpu;
237 WARN_ON(src != current);
239 if (use_eager_fpu()) {
240 memset(&dst->thread.fpu.state->xsave, 0, xstate_size);
244 memcpy(dst_fpu->state, src_fpu->state, xstate_size);
248 int fpu__copy(struct task_struct *dst, struct task_struct *src)
250 struct fpu *dst_fpu = &dst->thread.fpu;
251 struct fpu *src_fpu = &src->thread.fpu;
253 dst->thread.fpu.counter = 0;
254 dst->thread.fpu.has_fpu = 0;
255 dst->thread.fpu.state = NULL;
256 dst->thread.fpu.last_cpu = -1;
258 if (src_fpu->fpstate_active) {
259 int err = fpstate_alloc(dst_fpu);
269 * Allocate the backing store for the current task's FPU registers
270 * and initialize the registers themselves as well.
274 int fpstate_alloc_init(struct task_struct *curr)
276 struct fpu *fpu = &curr->thread.fpu;
279 if (WARN_ON_ONCE(curr != current))
281 if (WARN_ON_ONCE(fpu->fpstate_active))
285 * Memory allocation at the first usage of the FPU and other state.
287 ret = fpstate_alloc(&curr->thread.fpu);
291 fpstate_init(&curr->thread.fpu);
293 /* Safe to do for the current task: */
294 fpu->fpstate_active = 1;
298 EXPORT_SYMBOL_GPL(fpstate_alloc_init);
301 * This function is called before we modify a stopped child's
304 * If the child has not used the FPU before then initialize its
307 * If the child has used the FPU before then unlazy it.
309 * [ After this function call, after the context is modified and
310 * the child task is woken up, the child task will restore
311 * the modified FPU state from the modified context. If we
312 * didn't clear its lazy status here then the lazy in-registers
313 * state pending on its former CPU could be restored, losing
314 * the modifications. ]
316 * This function is also called before we read a stopped child's
317 * FPU state - to make sure it's modified.
319 * TODO: A future optimization would be to skip the unlazying in
320 * the read-only case, it's not strictly necessary for
321 * read-only access to the context.
323 static int fpu__unlazy_stopped(struct task_struct *child)
325 struct fpu *child_fpu = &child->thread.fpu;
328 if (WARN_ON_ONCE(child == current))
331 if (child_fpu->fpstate_active) {
332 child->thread.fpu.last_cpu = -1;
337 * Memory allocation at the first usage of the FPU and other state.
339 ret = fpstate_alloc(&child->thread.fpu);
343 fpstate_init(&child->thread.fpu);
345 /* Safe to do for stopped child tasks: */
346 child_fpu->fpstate_active = 1;
352 * 'fpu__restore()' saves the current math information in the
353 * old math state array, and gets the new ones from the current task
355 * Careful.. There are problems with IBM-designed IRQ13 behaviour.
356 * Don't touch unless you *really* know how it works.
358 * Must be called with kernel preemption disabled (eg with local
359 * local interrupts as in the case of do_device_not_available).
361 void fpu__restore(void)
363 struct task_struct *tsk = current;
364 struct fpu *fpu = &tsk->thread.fpu;
366 if (!fpu->fpstate_active) {
369 * does a slab alloc which can sleep
371 if (fpstate_alloc_init(tsk)) {
375 do_group_exit(SIGKILL);
381 /* Avoid __kernel_fpu_begin() right after __thread_fpu_begin() */
382 kernel_fpu_disable();
383 __thread_fpu_begin(fpu);
384 if (unlikely(restore_fpu_checking(fpu))) {
385 fpu_reset_state(fpu);
386 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
388 tsk->thread.fpu.counter++;
392 EXPORT_SYMBOL_GPL(fpu__restore);
394 void fpu__flush_thread(struct task_struct *tsk)
396 struct fpu *fpu = &tsk->thread.fpu;
398 WARN_ON(tsk != current);
400 if (!use_eager_fpu()) {
401 /* FPU state will be reallocated lazily at the first use. */
403 fpstate_free(&tsk->thread.fpu);
405 if (!fpu->fpstate_active) {
406 /* kthread execs. TODO: cleanup this horror. */
407 if (WARN_ON(fpstate_alloc_init(tsk)))
408 force_sig(SIGKILL, tsk);
411 restore_init_xstate();
416 * The xstateregs_active() routine is the same as the fpregs_active() routine,
417 * as the "regset->n" for the xstate regset will be updated based on the feature
418 * capabilites supported by the xsave.
420 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
422 struct fpu *target_fpu = &target->thread.fpu;
424 return target_fpu->fpstate_active ? regset->n : 0;
427 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
429 struct fpu *target_fpu = &target->thread.fpu;
431 return (cpu_has_fxsr && target_fpu->fpstate_active) ? regset->n : 0;
434 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
435 unsigned int pos, unsigned int count,
436 void *kbuf, void __user *ubuf)
443 ret = fpu__unlazy_stopped(target);
447 sanitize_i387_state(target);
449 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
450 &target->thread.fpu.state->fxsave, 0, -1);
453 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
454 unsigned int pos, unsigned int count,
455 const void *kbuf, const void __user *ubuf)
462 ret = fpu__unlazy_stopped(target);
466 sanitize_i387_state(target);
468 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
469 &target->thread.fpu.state->fxsave, 0, -1);
472 * mxcsr reserved bits must be masked to zero for security reasons.
474 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
477 * update the header bits in the xsave header, indicating the
478 * presence of FP and SSE state.
481 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
486 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
487 unsigned int pos, unsigned int count,
488 void *kbuf, void __user *ubuf)
490 struct xsave_struct *xsave;
496 ret = fpu__unlazy_stopped(target);
500 xsave = &target->thread.fpu.state->xsave;
503 * Copy the 48bytes defined by the software first into the xstate
504 * memory layout in the thread struct, so that we can copy the entire
505 * xstateregs to the user using one user_regset_copyout().
507 memcpy(&xsave->i387.sw_reserved,
508 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
510 * Copy the xstate memory layout.
512 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
516 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
517 unsigned int pos, unsigned int count,
518 const void *kbuf, const void __user *ubuf)
520 struct xsave_struct *xsave;
526 ret = fpu__unlazy_stopped(target);
530 xsave = &target->thread.fpu.state->xsave;
532 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
534 * mxcsr reserved bits must be masked to zero for security reasons.
536 xsave->i387.mxcsr &= mxcsr_feature_mask;
537 xsave->xsave_hdr.xstate_bv &= pcntxt_mask;
539 * These bits must be zero.
541 memset(&xsave->xsave_hdr.reserved, 0, 48);
545 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
548 * FPU tag word conversions.
551 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
553 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
555 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
557 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
558 /* and move the valid bits to the lower byte. */
559 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
560 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
561 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
566 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
567 #define FP_EXP_TAG_VALID 0
568 #define FP_EXP_TAG_ZERO 1
569 #define FP_EXP_TAG_SPECIAL 2
570 #define FP_EXP_TAG_EMPTY 3
572 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
575 u32 tos = (fxsave->swd >> 11) & 7;
576 u32 twd = (unsigned long) fxsave->twd;
578 u32 ret = 0xffff0000u;
581 for (i = 0; i < 8; i++, twd >>= 1) {
583 st = FPREG_ADDR(fxsave, (i - tos) & 7);
585 switch (st->exponent & 0x7fff) {
587 tag = FP_EXP_TAG_SPECIAL;
590 if (!st->significand[0] &&
591 !st->significand[1] &&
592 !st->significand[2] &&
594 tag = FP_EXP_TAG_ZERO;
596 tag = FP_EXP_TAG_SPECIAL;
599 if (st->significand[3] & 0x8000)
600 tag = FP_EXP_TAG_VALID;
602 tag = FP_EXP_TAG_SPECIAL;
606 tag = FP_EXP_TAG_EMPTY;
608 ret |= tag << (2 * i);
614 * FXSR floating point environment conversions.
618 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
620 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
621 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
622 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
625 env->cwd = fxsave->cwd | 0xffff0000u;
626 env->swd = fxsave->swd | 0xffff0000u;
627 env->twd = twd_fxsr_to_i387(fxsave);
630 env->fip = fxsave->rip;
631 env->foo = fxsave->rdp;
633 * should be actually ds/cs at fpu exception time, but
634 * that information is not available in 64bit mode.
636 env->fcs = task_pt_regs(tsk)->cs;
637 if (tsk == current) {
638 savesegment(ds, env->fos);
640 env->fos = tsk->thread.ds;
642 env->fos |= 0xffff0000;
644 env->fip = fxsave->fip;
645 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
646 env->foo = fxsave->foo;
647 env->fos = fxsave->fos;
650 for (i = 0; i < 8; ++i)
651 memcpy(&to[i], &from[i], sizeof(to[0]));
654 void convert_to_fxsr(struct task_struct *tsk,
655 const struct user_i387_ia32_struct *env)
658 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
659 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
660 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
663 fxsave->cwd = env->cwd;
664 fxsave->swd = env->swd;
665 fxsave->twd = twd_i387_to_fxsr(env->twd);
666 fxsave->fop = (u16) ((u32) env->fcs >> 16);
668 fxsave->rip = env->fip;
669 fxsave->rdp = env->foo;
670 /* cs and ds ignored */
672 fxsave->fip = env->fip;
673 fxsave->fcs = (env->fcs & 0xffff);
674 fxsave->foo = env->foo;
675 fxsave->fos = env->fos;
678 for (i = 0; i < 8; ++i)
679 memcpy(&to[i], &from[i], sizeof(from[0]));
682 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
683 unsigned int pos, unsigned int count,
684 void *kbuf, void __user *ubuf)
686 struct user_i387_ia32_struct env;
689 ret = fpu__unlazy_stopped(target);
693 if (!static_cpu_has(X86_FEATURE_FPU))
694 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
697 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
698 &target->thread.fpu.state->fsave, 0,
701 sanitize_i387_state(target);
703 if (kbuf && pos == 0 && count == sizeof(env)) {
704 convert_from_fxsr(kbuf, target);
708 convert_from_fxsr(&env, target);
710 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
713 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
714 unsigned int pos, unsigned int count,
715 const void *kbuf, const void __user *ubuf)
717 struct user_i387_ia32_struct env;
720 ret = fpu__unlazy_stopped(target);
724 sanitize_i387_state(target);
726 if (!static_cpu_has(X86_FEATURE_FPU))
727 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
730 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
731 &target->thread.fpu.state->fsave, 0,
734 if (pos > 0 || count < sizeof(env))
735 convert_from_fxsr(&env, target);
737 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
739 convert_to_fxsr(target, &env);
742 * update the header bit in the xsave header, indicating the
746 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
751 * FPU state for core dumps.
752 * This is only used for a.out dumps now.
753 * It is declared generically using elf_fpregset_t (which is
754 * struct user_i387_struct) but is in fact only used for 32-bit
755 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
757 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *ufpu)
759 struct task_struct *tsk = current;
760 struct fpu *fpu = &tsk->thread.fpu;
763 fpvalid = fpu->fpstate_active;
765 fpvalid = !fpregs_get(tsk, NULL,
766 0, sizeof(struct user_i387_ia32_struct),
771 EXPORT_SYMBOL(dump_fpu);
773 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */