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(restore_fpu_checking(fpu)))
119 fpu_reset_state(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);
172 static void __save_fpu(struct fpu *fpu)
175 if (unlikely(system_state == SYSTEM_BOOTING))
176 xsave_state_booting(&fpu->state.xsave);
178 xsave_state(&fpu->state.xsave);
185 * Save the FPU state (initialize it if necessary):
187 * This only ever gets called for the current task.
189 void fpu__save(struct fpu *fpu)
191 WARN_ON(fpu != ¤t->thread.fpu);
194 if (fpu->fpregs_active) {
195 if (use_eager_fpu()) {
198 copy_fpregs_to_fpstate(fpu);
199 fpregs_deactivate(fpu);
204 EXPORT_SYMBOL_GPL(fpu__save);
206 void fpstate_init(struct fpu *fpu)
209 finit_soft_fpu(&fpu->state.soft);
213 memset(&fpu->state, 0, xstate_size);
216 fx_finit(&fpu->state.fxsave);
218 struct i387_fsave_struct *fp = &fpu->state.fsave;
219 fp->cwd = 0xffff037fu;
220 fp->swd = 0xffff0000u;
221 fp->twd = 0xffffffffu;
222 fp->fos = 0xffff0000u;
225 EXPORT_SYMBOL_GPL(fpstate_init);
228 * Copy the current task's FPU state to a new task's FPU context.
230 * In the 'eager' case we just save to the destination context.
232 * In the 'lazy' case we save to the source context, mark the FPU lazy
233 * via stts() and copy the source context into the destination context.
235 static void fpu_copy(struct fpu *dst_fpu, struct fpu *src_fpu)
237 WARN_ON(src_fpu != ¤t->thread.fpu);
239 if (use_eager_fpu()) {
240 memset(&dst_fpu->state.xsave, 0, xstate_size);
244 memcpy(&dst_fpu->state, &src_fpu->state, xstate_size);
248 int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
250 dst_fpu->counter = 0;
251 dst_fpu->fpregs_active = 0;
252 dst_fpu->last_cpu = -1;
254 if (src_fpu->fpstate_active)
255 fpu_copy(dst_fpu, src_fpu);
261 * Activate the current task's in-memory FPU context,
262 * if it has not been used before:
264 void fpu__activate_curr(struct fpu *fpu)
266 WARN_ON_ONCE(fpu != ¤t->thread.fpu);
268 if (!fpu->fpstate_active) {
271 /* Safe to do for the current task: */
272 fpu->fpstate_active = 1;
275 EXPORT_SYMBOL_GPL(fpu__activate_curr);
278 * This function must be called before we modify a stopped child's
281 * If the child has not used the FPU before then initialize its
284 * If the child has used the FPU before then unlazy it.
286 * [ After this function call, after registers in the fpstate are
287 * modified and the child task has woken up, the child task will
288 * restore the modified FPU state from the modified context. If we
289 * didn't clear its lazy status here then the lazy in-registers
290 * state pending on its former CPU could be restored, corrupting
291 * the modifications. ]
293 * This function is also called before we read a stopped child's
294 * FPU state - to make sure it's initialized if the child has
295 * no active FPU state.
297 * TODO: A future optimization would be to skip the unlazying in
298 * the read-only case, it's not strictly necessary for
299 * read-only access to the context.
301 static void fpu__activate_stopped(struct fpu *child_fpu)
303 WARN_ON_ONCE(child_fpu == ¤t->thread.fpu);
305 if (child_fpu->fpstate_active) {
306 child_fpu->last_cpu = -1;
308 fpstate_init(child_fpu);
310 /* Safe to do for stopped child tasks: */
311 child_fpu->fpstate_active = 1;
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;
328 struct fpu *fpu = &tsk->thread.fpu;
330 fpu__activate_curr(fpu);
332 /* Avoid __kernel_fpu_begin() right after fpregs_activate() */
333 kernel_fpu_disable();
334 fpregs_activate(fpu);
335 if (unlikely(restore_fpu_checking(fpu))) {
336 fpu_reset_state(fpu);
337 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
339 tsk->thread.fpu.counter++;
343 EXPORT_SYMBOL_GPL(fpu__restore);
345 void fpu__clear(struct task_struct *tsk)
347 struct fpu *fpu = &tsk->thread.fpu;
349 WARN_ON_ONCE(tsk != current); /* Almost certainly an anomaly */
351 if (!use_eager_fpu()) {
352 /* FPU state will be reallocated lazily at the first use. */
355 if (!fpu->fpstate_active) {
356 fpu__activate_curr(fpu);
359 restore_init_xstate();
364 * The xstateregs_active() routine is the same as the regset_fpregs_active() routine,
365 * as the "regset->n" for the xstate regset will be updated based on the feature
366 * capabilites supported by the xsave.
368 int regset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
370 struct fpu *target_fpu = &target->thread.fpu;
372 return target_fpu->fpstate_active ? regset->n : 0;
375 int regset_xregset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
377 struct fpu *target_fpu = &target->thread.fpu;
379 return (cpu_has_fxsr && target_fpu->fpstate_active) ? regset->n : 0;
382 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
383 unsigned int pos, unsigned int count,
384 void *kbuf, void __user *ubuf)
386 struct fpu *fpu = &target->thread.fpu;
391 fpu__activate_stopped(fpu);
392 sanitize_i387_state(target);
394 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
395 &fpu->state.fxsave, 0, -1);
398 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
399 unsigned int pos, unsigned int count,
400 const void *kbuf, const void __user *ubuf)
402 struct fpu *fpu = &target->thread.fpu;
408 fpu__activate_stopped(fpu);
409 sanitize_i387_state(target);
411 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
412 &fpu->state.fxsave, 0, -1);
415 * mxcsr reserved bits must be masked to zero for security reasons.
417 fpu->state.fxsave.mxcsr &= mxcsr_feature_mask;
420 * update the header bits in the xsave header, indicating the
421 * presence of FP and SSE state.
424 fpu->state.xsave.header.xfeatures |= XSTATE_FPSSE;
429 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
430 unsigned int pos, unsigned int count,
431 void *kbuf, void __user *ubuf)
433 struct fpu *fpu = &target->thread.fpu;
434 struct xsave_struct *xsave;
440 fpu__activate_stopped(fpu);
442 xsave = &fpu->state.xsave;
445 * Copy the 48bytes defined by the software first into the xstate
446 * memory layout in the thread struct, so that we can copy the entire
447 * xstateregs to the user using one user_regset_copyout().
449 memcpy(&xsave->i387.sw_reserved,
450 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
452 * Copy the xstate memory layout.
454 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
458 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
459 unsigned int pos, unsigned int count,
460 const void *kbuf, const void __user *ubuf)
462 struct fpu *fpu = &target->thread.fpu;
463 struct xsave_struct *xsave;
469 fpu__activate_stopped(fpu);
471 xsave = &fpu->state.xsave;
473 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
475 * mxcsr reserved bits must be masked to zero for security reasons.
477 xsave->i387.mxcsr &= mxcsr_feature_mask;
478 xsave->header.xfeatures &= xfeatures_mask;
480 * These bits must be zero.
482 memset(&xsave->header.reserved, 0, 48);
487 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
490 * FPU tag word conversions.
493 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
495 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
497 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
499 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
500 /* and move the valid bits to the lower byte. */
501 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
502 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
503 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
508 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
509 #define FP_EXP_TAG_VALID 0
510 #define FP_EXP_TAG_ZERO 1
511 #define FP_EXP_TAG_SPECIAL 2
512 #define FP_EXP_TAG_EMPTY 3
514 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
517 u32 tos = (fxsave->swd >> 11) & 7;
518 u32 twd = (unsigned long) fxsave->twd;
520 u32 ret = 0xffff0000u;
523 for (i = 0; i < 8; i++, twd >>= 1) {
525 st = FPREG_ADDR(fxsave, (i - tos) & 7);
527 switch (st->exponent & 0x7fff) {
529 tag = FP_EXP_TAG_SPECIAL;
532 if (!st->significand[0] &&
533 !st->significand[1] &&
534 !st->significand[2] &&
536 tag = FP_EXP_TAG_ZERO;
538 tag = FP_EXP_TAG_SPECIAL;
541 if (st->significand[3] & 0x8000)
542 tag = FP_EXP_TAG_VALID;
544 tag = FP_EXP_TAG_SPECIAL;
548 tag = FP_EXP_TAG_EMPTY;
550 ret |= tag << (2 * i);
556 * FXSR floating point environment conversions.
560 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
562 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state.fxsave;
563 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
564 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
567 env->cwd = fxsave->cwd | 0xffff0000u;
568 env->swd = fxsave->swd | 0xffff0000u;
569 env->twd = twd_fxsr_to_i387(fxsave);
572 env->fip = fxsave->rip;
573 env->foo = fxsave->rdp;
575 * should be actually ds/cs at fpu exception time, but
576 * that information is not available in 64bit mode.
578 env->fcs = task_pt_regs(tsk)->cs;
579 if (tsk == current) {
580 savesegment(ds, env->fos);
582 env->fos = tsk->thread.ds;
584 env->fos |= 0xffff0000;
586 env->fip = fxsave->fip;
587 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
588 env->foo = fxsave->foo;
589 env->fos = fxsave->fos;
592 for (i = 0; i < 8; ++i)
593 memcpy(&to[i], &from[i], sizeof(to[0]));
596 void convert_to_fxsr(struct task_struct *tsk,
597 const struct user_i387_ia32_struct *env)
600 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state.fxsave;
601 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
602 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
605 fxsave->cwd = env->cwd;
606 fxsave->swd = env->swd;
607 fxsave->twd = twd_i387_to_fxsr(env->twd);
608 fxsave->fop = (u16) ((u32) env->fcs >> 16);
610 fxsave->rip = env->fip;
611 fxsave->rdp = env->foo;
612 /* cs and ds ignored */
614 fxsave->fip = env->fip;
615 fxsave->fcs = (env->fcs & 0xffff);
616 fxsave->foo = env->foo;
617 fxsave->fos = env->fos;
620 for (i = 0; i < 8; ++i)
621 memcpy(&to[i], &from[i], sizeof(from[0]));
624 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
625 unsigned int pos, unsigned int count,
626 void *kbuf, void __user *ubuf)
628 struct fpu *fpu = &target->thread.fpu;
629 struct user_i387_ia32_struct env;
631 fpu__activate_stopped(fpu);
633 if (!static_cpu_has(X86_FEATURE_FPU))
634 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
637 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
638 &fpu->state.fsave, 0,
641 sanitize_i387_state(target);
643 if (kbuf && pos == 0 && count == sizeof(env)) {
644 convert_from_fxsr(kbuf, target);
648 convert_from_fxsr(&env, target);
650 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
653 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
654 unsigned int pos, unsigned int count,
655 const void *kbuf, const void __user *ubuf)
657 struct fpu *fpu = &target->thread.fpu;
658 struct user_i387_ia32_struct env;
661 fpu__activate_stopped(fpu);
663 sanitize_i387_state(target);
665 if (!static_cpu_has(X86_FEATURE_FPU))
666 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
669 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
670 &fpu->state.fsave, 0,
673 if (pos > 0 || count < sizeof(env))
674 convert_from_fxsr(&env, target);
676 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
678 convert_to_fxsr(target, &env);
681 * update the header bit in the xsave header, indicating the
685 fpu->state.xsave.header.xfeatures |= XSTATE_FP;
690 * FPU state for core dumps.
691 * This is only used for a.out dumps now.
692 * It is declared generically using elf_fpregset_t (which is
693 * struct user_i387_struct) but is in fact only used for 32-bit
694 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
696 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *ufpu)
698 struct task_struct *tsk = current;
699 struct fpu *fpu = &tsk->thread.fpu;
702 fpvalid = fpu->fpstate_active;
704 fpvalid = !fpregs_get(tsk, NULL,
705 0, sizeof(struct user_i387_ia32_struct),
710 EXPORT_SYMBOL(dump_fpu);
712 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */