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) {
105 copy_fpregs_to_fpstate(fpu);
107 this_cpu_write(fpu_fpregs_owner_ctx, NULL);
108 __fpregs_activate_hw();
111 EXPORT_SYMBOL(__kernel_fpu_begin);
113 void __kernel_fpu_end(void)
115 struct fpu *fpu = ¤t->thread.fpu;
117 if (fpu->fpregs_active) {
118 if (WARN_ON(copy_fpstate_to_fpregs(fpu)))
121 __fpregs_deactivate_hw();
126 EXPORT_SYMBOL(__kernel_fpu_end);
128 void kernel_fpu_begin(void)
131 WARN_ON_ONCE(!irq_fpu_usable());
132 __kernel_fpu_begin();
134 EXPORT_SYMBOL_GPL(kernel_fpu_begin);
136 void kernel_fpu_end(void)
141 EXPORT_SYMBOL_GPL(kernel_fpu_end);
144 * CR0::TS save/restore functions:
146 int irq_ts_save(void)
149 * If in process context and not atomic, we can take a spurious DNA fault.
150 * Otherwise, doing clts() in process context requires disabling preemption
151 * or some heavy lifting like kernel_fpu_begin()
156 if (read_cr0() & X86_CR0_TS) {
163 EXPORT_SYMBOL_GPL(irq_ts_save);
165 void irq_ts_restore(int TS_state)
170 EXPORT_SYMBOL_GPL(irq_ts_restore);
173 * Save the FPU state (mark it for reload if necessary):
175 * This only ever gets called for the current task.
177 void fpu__save(struct fpu *fpu)
179 WARN_ON(fpu != ¤t->thread.fpu);
182 if (fpu->fpregs_active) {
183 if (!copy_fpregs_to_fpstate(fpu))
184 fpregs_deactivate(fpu);
188 EXPORT_SYMBOL_GPL(fpu__save);
190 void fpstate_init(struct fpu *fpu)
193 finit_soft_fpu(&fpu->state.soft);
197 memset(&fpu->state, 0, xstate_size);
200 fx_finit(&fpu->state.fxsave);
202 struct i387_fsave_struct *fp = &fpu->state.fsave;
203 fp->cwd = 0xffff037fu;
204 fp->swd = 0xffff0000u;
205 fp->twd = 0xffffffffu;
206 fp->fos = 0xffff0000u;
209 EXPORT_SYMBOL_GPL(fpstate_init);
212 * Copy the current task's FPU state to a new task's FPU context.
214 * In the 'eager' case we just save to the destination context.
216 * In the 'lazy' case we save to the source context, mark the FPU lazy
217 * via stts() and copy the source context into the destination context.
219 static void fpu_copy(struct fpu *dst_fpu, struct fpu *src_fpu)
221 WARN_ON(src_fpu != ¤t->thread.fpu);
224 * Don't let 'init optimized' areas of the XSAVE area
225 * leak into the child task:
228 memset(&dst_fpu->state.xsave, 0, xstate_size);
231 * Save current FPU registers directly into the child
232 * FPU context, without any memory-to-memory copying.
234 * If the FPU context got destroyed in the process (FNSAVE
235 * done on old CPUs) then copy it back into the source
236 * context and mark the current task for lazy restore.
238 * We have to do all this with preemption disabled,
239 * mostly because of the FNSAVE case, because in that
240 * case we must not allow preemption in the window
241 * between the FNSAVE and us marking the context lazy.
243 * It shouldn't be an issue as even FNSAVE is plenty
244 * fast in terms of critical section length.
247 if (!copy_fpregs_to_fpstate(dst_fpu)) {
248 memcpy(&src_fpu->state, &dst_fpu->state, xstate_size);
249 fpregs_deactivate(src_fpu);
254 int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
256 dst_fpu->counter = 0;
257 dst_fpu->fpregs_active = 0;
258 dst_fpu->last_cpu = -1;
260 if (src_fpu->fpstate_active)
261 fpu_copy(dst_fpu, src_fpu);
267 * Activate the current task's in-memory FPU context,
268 * if it has not been used before:
270 void fpu__activate_curr(struct fpu *fpu)
272 WARN_ON_ONCE(fpu != ¤t->thread.fpu);
274 if (!fpu->fpstate_active) {
277 /* Safe to do for the current task: */
278 fpu->fpstate_active = 1;
281 EXPORT_SYMBOL_GPL(fpu__activate_curr);
284 * This function must be called before we modify a stopped child's
287 * If the child has not used the FPU before then initialize its
290 * If the child has used the FPU before then unlazy it.
292 * [ After this function call, after registers in the fpstate are
293 * modified and the child task has woken up, the child task will
294 * restore the modified FPU state from the modified context. If we
295 * didn't clear its lazy status here then the lazy in-registers
296 * state pending on its former CPU could be restored, corrupting
297 * the modifications. ]
299 * This function is also called before we read a stopped child's
300 * FPU state - to make sure it's initialized if the child has
301 * no active FPU state.
303 * TODO: A future optimization would be to skip the unlazying in
304 * the read-only case, it's not strictly necessary for
305 * read-only access to the context.
307 static void fpu__activate_stopped(struct fpu *child_fpu)
309 WARN_ON_ONCE(child_fpu == ¤t->thread.fpu);
311 if (child_fpu->fpstate_active) {
312 child_fpu->last_cpu = -1;
314 fpstate_init(child_fpu);
316 /* Safe to do for stopped child tasks: */
317 child_fpu->fpstate_active = 1;
322 * 'fpu__restore()' is called to copy FPU registers from
323 * the FPU fpstate to the live hw registers and to activate
324 * access to the hardware registers, so that FPU instructions
325 * can be used afterwards.
327 * Must be called with kernel preemption disabled (for example
328 * with local interrupts disabled, as it is in the case of
329 * do_device_not_available()).
331 void fpu__restore(void)
333 struct task_struct *tsk = current;
334 struct fpu *fpu = &tsk->thread.fpu;
336 fpu__activate_curr(fpu);
338 /* Avoid __kernel_fpu_begin() right after fpregs_activate() */
339 kernel_fpu_disable();
340 fpregs_activate(fpu);
341 if (unlikely(copy_fpstate_to_fpregs(fpu))) {
343 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
345 tsk->thread.fpu.counter++;
349 EXPORT_SYMBOL_GPL(fpu__restore);
352 * Drops current FPU state: deactivates the fpregs and
353 * the fpstate. NOTE: it still leaves previous contents
354 * in the fpregs in the eager-FPU case.
356 * This function can be used in cases where we know that
357 * a state-restore is coming: either an explicit one,
360 void fpu__drop(struct fpu *fpu)
365 if (fpu->fpregs_active) {
366 /* Ignore delayed exceptions from user space */
367 asm volatile("1: fwait\n"
369 _ASM_EXTABLE(1b, 2b));
370 fpregs_deactivate(fpu);
373 fpu->fpstate_active = 0;
379 * Reset the FPU state back to init state:
381 void fpu__reset(struct fpu *fpu)
383 if (!use_eager_fpu())
386 restore_init_xstate();
390 * Called by sys_execve() to clear the FPU fpregs, so that FPU state
391 * of the previous binary does not leak over into the exec()ed binary:
393 void fpu__clear(struct fpu *fpu)
395 WARN_ON_ONCE(fpu != ¤t->thread.fpu); /* Almost certainly an anomaly */
397 if (!use_eager_fpu()) {
398 /* FPU state will be reallocated lazily at the first use. */
401 if (!fpu->fpstate_active) {
402 fpu__activate_curr(fpu);
405 restore_init_xstate();
410 * The xstateregs_active() routine is the same as the regset_fpregs_active() routine,
411 * as the "regset->n" for the xstate regset will be updated based on the feature
412 * capabilites supported by the xsave.
414 int regset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
416 struct fpu *target_fpu = &target->thread.fpu;
418 return target_fpu->fpstate_active ? regset->n : 0;
421 int regset_xregset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
423 struct fpu *target_fpu = &target->thread.fpu;
425 return (cpu_has_fxsr && target_fpu->fpstate_active) ? regset->n : 0;
428 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
429 unsigned int pos, unsigned int count,
430 void *kbuf, void __user *ubuf)
432 struct fpu *fpu = &target->thread.fpu;
437 fpu__activate_stopped(fpu);
438 fpstate_sanitize_xstate(fpu);
440 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
441 &fpu->state.fxsave, 0, -1);
444 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
445 unsigned int pos, unsigned int count,
446 const void *kbuf, const void __user *ubuf)
448 struct fpu *fpu = &target->thread.fpu;
454 fpu__activate_stopped(fpu);
455 fpstate_sanitize_xstate(fpu);
457 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
458 &fpu->state.fxsave, 0, -1);
461 * mxcsr reserved bits must be masked to zero for security reasons.
463 fpu->state.fxsave.mxcsr &= mxcsr_feature_mask;
466 * update the header bits in the xsave header, indicating the
467 * presence of FP and SSE state.
470 fpu->state.xsave.header.xfeatures |= XSTATE_FPSSE;
475 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
476 unsigned int pos, unsigned int count,
477 void *kbuf, void __user *ubuf)
479 struct fpu *fpu = &target->thread.fpu;
480 struct xsave_struct *xsave;
486 fpu__activate_stopped(fpu);
488 xsave = &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 fpu *fpu = &target->thread.fpu;
509 struct xsave_struct *xsave;
515 fpu__activate_stopped(fpu);
517 xsave = &fpu->state.xsave;
519 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
521 * mxcsr reserved bits must be masked to zero for security reasons.
523 xsave->i387.mxcsr &= mxcsr_feature_mask;
524 xsave->header.xfeatures &= xfeatures_mask;
526 * These bits must be zero.
528 memset(&xsave->header.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 fpu *fpu = &target->thread.fpu;
675 struct user_i387_ia32_struct env;
677 fpu__activate_stopped(fpu);
679 if (!static_cpu_has(X86_FEATURE_FPU))
680 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
683 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
684 &fpu->state.fsave, 0,
687 fpstate_sanitize_xstate(fpu);
689 if (kbuf && pos == 0 && count == sizeof(env)) {
690 convert_from_fxsr(kbuf, target);
694 convert_from_fxsr(&env, target);
696 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
699 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
700 unsigned int pos, unsigned int count,
701 const void *kbuf, const void __user *ubuf)
703 struct fpu *fpu = &target->thread.fpu;
704 struct user_i387_ia32_struct env;
707 fpu__activate_stopped(fpu);
708 fpstate_sanitize_xstate(fpu);
710 if (!static_cpu_has(X86_FEATURE_FPU))
711 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
714 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
715 &fpu->state.fsave, 0,
718 if (pos > 0 || count < sizeof(env))
719 convert_from_fxsr(&env, target);
721 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
723 convert_to_fxsr(target, &env);
726 * update the header bit in the xsave header, indicating the
730 fpu->state.xsave.header.xfeatures |= XSTATE_FP;
735 * FPU state for core dumps.
736 * This is only used for a.out dumps now.
737 * It is declared generically using elf_fpregset_t (which is
738 * struct user_i387_struct) but is in fact only used for 32-bit
739 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
741 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *ufpu)
743 struct task_struct *tsk = current;
744 struct fpu *fpu = &tsk->thread.fpu;
747 fpvalid = fpu->fpstate_active;
749 fpvalid = !fpregs_get(tsk, NULL,
750 0, sizeof(struct user_i387_ia32_struct),
755 EXPORT_SYMBOL(dump_fpu);
757 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */