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>
9 #include <linux/hardirq.h>
12 * Track whether the kernel is using the FPU state
17 * - by IRQ context code to potentially use the FPU
20 * - to debug kernel_fpu_begin()/end() correctness
22 static DEFINE_PER_CPU(bool, in_kernel_fpu);
25 * Track which context is using the FPU on the CPU:
27 DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
29 static void kernel_fpu_disable(void)
31 WARN_ON(this_cpu_read(in_kernel_fpu));
32 this_cpu_write(in_kernel_fpu, true);
35 static void kernel_fpu_enable(void)
37 WARN_ON_ONCE(!this_cpu_read(in_kernel_fpu));
38 this_cpu_write(in_kernel_fpu, false);
41 static bool kernel_fpu_disabled(void)
43 return this_cpu_read(in_kernel_fpu);
47 * Were we in an interrupt that interrupted kernel mode?
49 * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
50 * pair does nothing at all: the thread must not have fpu (so
51 * that we don't try to save the FPU state), and TS must
52 * be set (so that the clts/stts pair does nothing that is
53 * visible in the interrupted kernel thread).
55 * Except for the eagerfpu case when we return true; in the likely case
56 * the thread has FPU but we are not going to set/clear TS.
58 static bool interrupted_kernel_fpu_idle(void)
60 if (kernel_fpu_disabled())
66 return !current->thread.fpu.fpregs_active && (read_cr0() & X86_CR0_TS);
70 * Were we in user mode (or vm86 mode) when we were
73 * Doing kernel_fpu_begin/end() is ok if we are running
74 * in an interrupt context from user mode - we'll just
75 * save the FPU state as required.
77 static bool interrupted_user_mode(void)
79 struct pt_regs *regs = get_irq_regs();
80 return regs && user_mode(regs);
84 * Can we use the FPU in kernel mode with the
85 * whole "kernel_fpu_begin/end()" sequence?
87 * It's always ok in process context (ie "not interrupt")
88 * but it is sometimes ok even from an irq.
90 bool irq_fpu_usable(void)
92 return !in_interrupt() ||
93 interrupted_user_mode() ||
94 interrupted_kernel_fpu_idle();
96 EXPORT_SYMBOL(irq_fpu_usable);
98 void __kernel_fpu_begin(void)
100 struct fpu *fpu = ¤t->thread.fpu;
102 kernel_fpu_disable();
104 if (fpu->fpregs_active) {
107 this_cpu_write(fpu_fpregs_owner_ctx, NULL);
108 if (!use_eager_fpu())
112 EXPORT_SYMBOL(__kernel_fpu_begin);
114 void __kernel_fpu_end(void)
116 struct fpu *fpu = ¤t->thread.fpu;
118 if (fpu->fpregs_active) {
119 if (WARN_ON(restore_fpu_checking(fpu)))
120 fpu_reset_state(fpu);
121 } else if (!use_eager_fpu()) {
127 EXPORT_SYMBOL(__kernel_fpu_end);
129 void kernel_fpu_begin(void)
132 WARN_ON_ONCE(!irq_fpu_usable());
133 __kernel_fpu_begin();
135 EXPORT_SYMBOL_GPL(kernel_fpu_begin);
137 void kernel_fpu_end(void)
142 EXPORT_SYMBOL_GPL(kernel_fpu_end);
145 * CR0::TS save/restore functions:
147 int irq_ts_save(void)
150 * If in process context and not atomic, we can take a spurious DNA fault.
151 * Otherwise, doing clts() in process context requires disabling preemption
152 * or some heavy lifting like kernel_fpu_begin()
157 if (read_cr0() & X86_CR0_TS) {
164 EXPORT_SYMBOL_GPL(irq_ts_save);
166 void irq_ts_restore(int TS_state)
171 EXPORT_SYMBOL_GPL(irq_ts_restore);
173 static void __save_fpu(struct fpu *fpu)
176 if (unlikely(system_state == SYSTEM_BOOTING))
177 xsave_state_booting(&fpu->state->xsave);
179 xsave_state(&fpu->state->xsave);
186 * Save the FPU state (initialize it if necessary):
188 * This only ever gets called for the current task.
190 void fpu__save(struct fpu *fpu)
192 WARN_ON(fpu != ¤t->thread.fpu);
195 if (fpu->fpregs_active) {
196 if (use_eager_fpu()) {
200 fpregs_deactivate(fpu);
205 EXPORT_SYMBOL_GPL(fpu__save);
207 void fpstate_init(struct fpu *fpu)
210 finit_soft_fpu(&fpu->state->soft);
214 memset(fpu->state, 0, xstate_size);
217 fx_finit(&fpu->state->fxsave);
219 struct i387_fsave_struct *fp = &fpu->state->fsave;
220 fp->cwd = 0xffff037fu;
221 fp->swd = 0xffff0000u;
222 fp->twd = 0xffffffffu;
223 fp->fos = 0xffff0000u;
226 EXPORT_SYMBOL_GPL(fpstate_init);
229 * FPU state allocation:
231 static struct kmem_cache *task_xstate_cachep;
233 void fpstate_cache_init(void)
236 kmem_cache_create("task_xstate", xstate_size,
237 __alignof__(union thread_xstate),
238 SLAB_PANIC | SLAB_NOTRACK, NULL);
242 int fpstate_alloc(struct fpu *fpu)
247 fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
251 /* The CPU requires the FPU state to be aligned to 16 byte boundaries: */
252 WARN_ON((unsigned long)fpu->state & 15);
256 EXPORT_SYMBOL_GPL(fpstate_alloc);
258 void fpstate_free(struct fpu *fpu)
261 kmem_cache_free(task_xstate_cachep, fpu->state);
265 EXPORT_SYMBOL_GPL(fpstate_free);
268 * Copy the current task's FPU state to a new task's FPU context.
270 * In the 'eager' case we just save to the destination context.
272 * In the 'lazy' case we save to the source context, mark the FPU lazy
273 * via stts() and copy the source context into the destination context.
275 static void fpu_copy(struct fpu *dst_fpu, struct fpu *src_fpu)
277 WARN_ON(src_fpu != ¤t->thread.fpu);
279 if (use_eager_fpu()) {
280 memset(&dst_fpu->state->xsave, 0, xstate_size);
284 memcpy(dst_fpu->state, src_fpu->state, xstate_size);
288 int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
290 dst_fpu->counter = 0;
291 dst_fpu->fpregs_active = 0;
292 dst_fpu->state = NULL;
293 dst_fpu->last_cpu = -1;
295 if (src_fpu->fpstate_active) {
296 int err = fpstate_alloc(dst_fpu);
300 fpu_copy(dst_fpu, src_fpu);
306 * Allocate the backing store for the current task's FPU registers
307 * and initialize the registers themselves as well.
311 int fpstate_alloc_init(struct fpu *fpu)
315 if (WARN_ON_ONCE(fpu != ¤t->thread.fpu))
317 if (WARN_ON_ONCE(fpu->fpstate_active))
321 * Memory allocation at the first usage of the FPU and other state.
323 ret = fpstate_alloc(fpu);
329 /* Safe to do for the current task: */
330 fpu->fpstate_active = 1;
334 EXPORT_SYMBOL_GPL(fpstate_alloc_init);
337 * This function is called before we modify a stopped child's
340 * If the child has not used the FPU before then initialize its
343 * If the child has used the FPU before then unlazy it.
345 * [ After this function call, after the context is modified and
346 * the child task is woken up, the child task will restore
347 * the modified FPU state from the modified context. If we
348 * didn't clear its lazy status here then the lazy in-registers
349 * state pending on its former CPU could be restored, losing
350 * the modifications. ]
352 * This function is also called before we read a stopped child's
353 * FPU state - to make sure it's modified.
355 * TODO: A future optimization would be to skip the unlazying in
356 * the read-only case, it's not strictly necessary for
357 * read-only access to the context.
359 static int fpu__unlazy_stopped(struct fpu *child_fpu)
363 if (WARN_ON_ONCE(child_fpu == ¤t->thread.fpu))
366 if (child_fpu->fpstate_active) {
367 child_fpu->last_cpu = -1;
372 * Memory allocation at the first usage of the FPU and other state.
374 ret = fpstate_alloc(child_fpu);
378 fpstate_init(child_fpu);
380 /* Safe to do for stopped child tasks: */
381 child_fpu->fpstate_active = 1;
387 * 'fpu__restore()' saves the current math information in the
388 * old math state array, and gets the new ones from the current task
390 * Careful.. There are problems with IBM-designed IRQ13 behaviour.
391 * Don't touch unless you *really* know how it works.
393 * Must be called with kernel preemption disabled (eg with local
394 * local interrupts as in the case of do_device_not_available).
396 void fpu__restore(void)
398 struct task_struct *tsk = current;
399 struct fpu *fpu = &tsk->thread.fpu;
401 if (!fpu->fpstate_active) {
404 * does a slab alloc which can sleep
406 if (fpstate_alloc_init(fpu)) {
410 do_group_exit(SIGKILL);
416 /* Avoid __kernel_fpu_begin() right after fpregs_activate() */
417 kernel_fpu_disable();
418 fpregs_activate(fpu);
419 if (unlikely(restore_fpu_checking(fpu))) {
420 fpu_reset_state(fpu);
421 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
423 tsk->thread.fpu.counter++;
427 EXPORT_SYMBOL_GPL(fpu__restore);
429 void fpu__clear(struct task_struct *tsk)
431 struct fpu *fpu = &tsk->thread.fpu;
433 WARN_ON_ONCE(tsk != current); /* Almost certainly an anomaly */
435 if (!use_eager_fpu()) {
436 /* FPU state will be reallocated lazily at the first use. */
440 if (!fpu->fpstate_active) {
441 /* kthread execs. TODO: cleanup this horror. */
442 if (WARN_ON(fpstate_alloc_init(fpu)))
443 force_sig(SIGKILL, tsk);
446 restore_init_xstate();
451 * The xstateregs_active() routine is the same as the regset_fpregs_active() routine,
452 * as the "regset->n" for the xstate regset will be updated based on the feature
453 * capabilites supported by the xsave.
455 int regset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
457 struct fpu *target_fpu = &target->thread.fpu;
459 return target_fpu->fpstate_active ? regset->n : 0;
462 int regset_xregset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
464 struct fpu *target_fpu = &target->thread.fpu;
466 return (cpu_has_fxsr && target_fpu->fpstate_active) ? regset->n : 0;
469 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
470 unsigned int pos, unsigned int count,
471 void *kbuf, void __user *ubuf)
473 struct fpu *fpu = &target->thread.fpu;
479 ret = fpu__unlazy_stopped(fpu);
483 sanitize_i387_state(target);
485 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
486 &fpu->state->fxsave, 0, -1);
489 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
490 unsigned int pos, unsigned int count,
491 const void *kbuf, const void __user *ubuf)
493 struct fpu *fpu = &target->thread.fpu;
499 ret = fpu__unlazy_stopped(fpu);
503 sanitize_i387_state(target);
505 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
506 &fpu->state->fxsave, 0, -1);
509 * mxcsr reserved bits must be masked to zero for security reasons.
511 fpu->state->fxsave.mxcsr &= mxcsr_feature_mask;
514 * update the header bits in the xsave header, indicating the
515 * presence of FP and SSE state.
518 fpu->state->xsave.header.xfeatures |= XSTATE_FPSSE;
523 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
524 unsigned int pos, unsigned int count,
525 void *kbuf, void __user *ubuf)
527 struct fpu *fpu = &target->thread.fpu;
528 struct xsave_struct *xsave;
534 ret = fpu__unlazy_stopped(fpu);
538 xsave = &fpu->state->xsave;
541 * Copy the 48bytes defined by the software first into the xstate
542 * memory layout in the thread struct, so that we can copy the entire
543 * xstateregs to the user using one user_regset_copyout().
545 memcpy(&xsave->i387.sw_reserved,
546 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
548 * Copy the xstate memory layout.
550 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
554 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
555 unsigned int pos, unsigned int count,
556 const void *kbuf, const void __user *ubuf)
558 struct fpu *fpu = &target->thread.fpu;
559 struct xsave_struct *xsave;
565 ret = fpu__unlazy_stopped(fpu);
569 xsave = &fpu->state->xsave;
571 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
573 * mxcsr reserved bits must be masked to zero for security reasons.
575 xsave->i387.mxcsr &= mxcsr_feature_mask;
576 xsave->header.xfeatures &= xfeatures_mask;
578 * These bits must be zero.
580 memset(&xsave->header.reserved, 0, 48);
585 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
588 * FPU tag word conversions.
591 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
593 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
595 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
597 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
598 /* and move the valid bits to the lower byte. */
599 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
600 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
601 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
606 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
607 #define FP_EXP_TAG_VALID 0
608 #define FP_EXP_TAG_ZERO 1
609 #define FP_EXP_TAG_SPECIAL 2
610 #define FP_EXP_TAG_EMPTY 3
612 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
615 u32 tos = (fxsave->swd >> 11) & 7;
616 u32 twd = (unsigned long) fxsave->twd;
618 u32 ret = 0xffff0000u;
621 for (i = 0; i < 8; i++, twd >>= 1) {
623 st = FPREG_ADDR(fxsave, (i - tos) & 7);
625 switch (st->exponent & 0x7fff) {
627 tag = FP_EXP_TAG_SPECIAL;
630 if (!st->significand[0] &&
631 !st->significand[1] &&
632 !st->significand[2] &&
634 tag = FP_EXP_TAG_ZERO;
636 tag = FP_EXP_TAG_SPECIAL;
639 if (st->significand[3] & 0x8000)
640 tag = FP_EXP_TAG_VALID;
642 tag = FP_EXP_TAG_SPECIAL;
646 tag = FP_EXP_TAG_EMPTY;
648 ret |= tag << (2 * i);
654 * FXSR floating point environment conversions.
658 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
660 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
661 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
662 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
665 env->cwd = fxsave->cwd | 0xffff0000u;
666 env->swd = fxsave->swd | 0xffff0000u;
667 env->twd = twd_fxsr_to_i387(fxsave);
670 env->fip = fxsave->rip;
671 env->foo = fxsave->rdp;
673 * should be actually ds/cs at fpu exception time, but
674 * that information is not available in 64bit mode.
676 env->fcs = task_pt_regs(tsk)->cs;
677 if (tsk == current) {
678 savesegment(ds, env->fos);
680 env->fos = tsk->thread.ds;
682 env->fos |= 0xffff0000;
684 env->fip = fxsave->fip;
685 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
686 env->foo = fxsave->foo;
687 env->fos = fxsave->fos;
690 for (i = 0; i < 8; ++i)
691 memcpy(&to[i], &from[i], sizeof(to[0]));
694 void convert_to_fxsr(struct task_struct *tsk,
695 const struct user_i387_ia32_struct *env)
698 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
699 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
700 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
703 fxsave->cwd = env->cwd;
704 fxsave->swd = env->swd;
705 fxsave->twd = twd_i387_to_fxsr(env->twd);
706 fxsave->fop = (u16) ((u32) env->fcs >> 16);
708 fxsave->rip = env->fip;
709 fxsave->rdp = env->foo;
710 /* cs and ds ignored */
712 fxsave->fip = env->fip;
713 fxsave->fcs = (env->fcs & 0xffff);
714 fxsave->foo = env->foo;
715 fxsave->fos = env->fos;
718 for (i = 0; i < 8; ++i)
719 memcpy(&to[i], &from[i], sizeof(from[0]));
722 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
723 unsigned int pos, unsigned int count,
724 void *kbuf, void __user *ubuf)
726 struct fpu *fpu = &target->thread.fpu;
727 struct user_i387_ia32_struct env;
730 ret = fpu__unlazy_stopped(fpu);
734 if (!static_cpu_has(X86_FEATURE_FPU))
735 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
738 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
739 &fpu->state->fsave, 0,
742 sanitize_i387_state(target);
744 if (kbuf && pos == 0 && count == sizeof(env)) {
745 convert_from_fxsr(kbuf, target);
749 convert_from_fxsr(&env, target);
751 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
754 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
755 unsigned int pos, unsigned int count,
756 const void *kbuf, const void __user *ubuf)
758 struct fpu *fpu = &target->thread.fpu;
759 struct user_i387_ia32_struct env;
762 ret = fpu__unlazy_stopped(fpu);
766 sanitize_i387_state(target);
768 if (!static_cpu_has(X86_FEATURE_FPU))
769 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
772 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
773 &fpu->state->fsave, 0,
776 if (pos > 0 || count < sizeof(env))
777 convert_from_fxsr(&env, target);
779 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
781 convert_to_fxsr(target, &env);
784 * update the header bit in the xsave header, indicating the
788 fpu->state->xsave.header.xfeatures |= XSTATE_FP;
793 * FPU state for core dumps.
794 * This is only used for a.out dumps now.
795 * It is declared generically using elf_fpregset_t (which is
796 * struct user_i387_struct) but is in fact only used for 32-bit
797 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
799 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *ufpu)
801 struct task_struct *tsk = current;
802 struct fpu *fpu = &tsk->thread.fpu;
805 fpvalid = fpu->fpstate_active;
807 fpvalid = !fpregs_get(tsk, NULL,
808 0, sizeof(struct user_i387_ia32_struct),
813 EXPORT_SYMBOL(dump_fpu);
815 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */