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
7 * x86-64 work by Andi Kleen 2002
10 #ifndef _FPU_INTERNAL_H
11 #define _FPU_INTERNAL_H
13 #include <linux/kernel_stat.h>
14 #include <linux/regset.h>
15 #include <linux/compat.h>
16 #include <linux/slab.h>
18 #include <asm/cpufeature.h>
19 #include <asm/processor.h>
20 #include <asm/sigcontext.h>
22 #include <asm/uaccess.h>
23 #include <asm/xsave.h>
27 # include <asm/sigcontext32.h>
28 # include <asm/user32.h>
30 int ia32_setup_rt_frame(int sig, struct ksignal *ksig,
31 compat_sigset_t *set, struct pt_regs *regs);
32 int ia32_setup_frame(int sig, struct ksignal *ksig,
33 compat_sigset_t *set, struct pt_regs *regs);
35 # define user_i387_ia32_struct user_i387_struct
36 # define user32_fxsr_struct user_fxsr_struct
37 # define ia32_setup_frame __setup_frame
38 # define ia32_setup_rt_frame __setup_rt_frame
41 extern unsigned int mxcsr_feature_mask;
42 extern void fpu_init(void);
43 extern void eager_fpu_init(void);
45 DECLARE_PER_CPU(struct task_struct *, fpu_owner_task);
47 extern void convert_from_fxsr(struct user_i387_ia32_struct *env,
48 struct task_struct *tsk);
49 extern void convert_to_fxsr(struct task_struct *tsk,
50 const struct user_i387_ia32_struct *env);
52 extern user_regset_active_fn fpregs_active, xfpregs_active;
53 extern user_regset_get_fn fpregs_get, xfpregs_get, fpregs_soft_get,
55 extern user_regset_set_fn fpregs_set, xfpregs_set, fpregs_soft_set,
59 * xstateregs_active == fpregs_active. Please refer to the comment
60 * at the definition of fpregs_active.
62 #define xstateregs_active fpregs_active
64 #ifdef CONFIG_MATH_EMULATION
65 extern void finit_soft_fpu(struct i387_soft_struct *soft);
67 static inline void finit_soft_fpu(struct i387_soft_struct *soft) {}
71 * Must be run with preemption disabled: this clears the fpu_owner_task,
74 * This will disable any lazy FPU state restore of the current FPU state,
75 * but if the current thread owns the FPU, it will still be saved by.
77 static inline void __cpu_disable_lazy_restore(unsigned int cpu)
79 per_cpu(fpu_owner_task, cpu) = NULL;
82 static inline int fpu_lazy_restore(struct task_struct *new, unsigned int cpu)
84 return new == this_cpu_read_stable(fpu_owner_task) &&
85 cpu == new->thread.fpu.last_cpu;
88 static inline int is_ia32_compat_frame(void)
90 return config_enabled(CONFIG_IA32_EMULATION) &&
91 test_thread_flag(TIF_IA32);
94 static inline int is_ia32_frame(void)
96 return config_enabled(CONFIG_X86_32) || is_ia32_compat_frame();
99 static inline int is_x32_frame(void)
101 return config_enabled(CONFIG_X86_X32_ABI) && test_thread_flag(TIF_X32);
104 #define X87_FSW_ES (1 << 7) /* Exception Summary */
106 static __always_inline __pure bool use_eager_fpu(void)
108 return static_cpu_has_safe(X86_FEATURE_EAGER_FPU);
111 static __always_inline __pure bool use_xsaveopt(void)
113 return static_cpu_has_safe(X86_FEATURE_XSAVEOPT);
116 static __always_inline __pure bool use_xsave(void)
118 return static_cpu_has_safe(X86_FEATURE_XSAVE);
121 static __always_inline __pure bool use_fxsr(void)
123 return static_cpu_has_safe(X86_FEATURE_FXSR);
126 static inline void fx_finit(struct i387_fxsave_struct *fx)
128 memset(fx, 0, xstate_size);
130 fx->mxcsr = MXCSR_DEFAULT;
133 extern void __sanitize_i387_state(struct task_struct *);
135 static inline void sanitize_i387_state(struct task_struct *tsk)
139 __sanitize_i387_state(tsk);
142 #define user_insn(insn, output, input...) \
145 asm volatile(ASM_STAC "\n" \
147 "2: " ASM_CLAC "\n" \
148 ".section .fixup,\"ax\"\n" \
149 "3: movl $-1,%[err]\n" \
152 _ASM_EXTABLE(1b, 3b) \
153 : [err] "=r" (err), output \
158 #define check_insn(insn, output, input...) \
161 asm volatile("1:" #insn "\n\t" \
163 ".section .fixup,\"ax\"\n" \
164 "3: movl $-1,%[err]\n" \
167 _ASM_EXTABLE(1b, 3b) \
168 : [err] "=r" (err), output \
173 static inline int fsave_user(struct i387_fsave_struct __user *fx)
175 return user_insn(fnsave %[fx]; fwait, [fx] "=m" (*fx), "m" (*fx));
178 static inline int fxsave_user(struct i387_fxsave_struct __user *fx)
180 if (config_enabled(CONFIG_X86_32))
181 return user_insn(fxsave %[fx], [fx] "=m" (*fx), "m" (*fx));
182 else if (config_enabled(CONFIG_AS_FXSAVEQ))
183 return user_insn(fxsaveq %[fx], [fx] "=m" (*fx), "m" (*fx));
185 /* See comment in fpu_fxsave() below. */
186 return user_insn(rex64/fxsave (%[fx]), "=m" (*fx), [fx] "R" (fx));
189 static inline int fxrstor_checking(struct i387_fxsave_struct *fx)
191 if (config_enabled(CONFIG_X86_32))
192 return check_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
193 else if (config_enabled(CONFIG_AS_FXSAVEQ))
194 return check_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
196 /* See comment in fpu_fxsave() below. */
197 return check_insn(rex64/fxrstor (%[fx]), "=m" (*fx), [fx] "R" (fx),
201 static inline int fxrstor_user(struct i387_fxsave_struct __user *fx)
203 if (config_enabled(CONFIG_X86_32))
204 return user_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
205 else if (config_enabled(CONFIG_AS_FXSAVEQ))
206 return user_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
208 /* See comment in fpu_fxsave() below. */
209 return user_insn(rex64/fxrstor (%[fx]), "=m" (*fx), [fx] "R" (fx),
213 static inline int frstor_checking(struct i387_fsave_struct *fx)
215 return check_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
218 static inline int frstor_user(struct i387_fsave_struct __user *fx)
220 return user_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
223 static inline void fpu_fxsave(struct fpu *fpu)
225 if (config_enabled(CONFIG_X86_32))
226 asm volatile( "fxsave %[fx]" : [fx] "=m" (fpu->state->fxsave));
227 else if (config_enabled(CONFIG_AS_FXSAVEQ))
228 asm volatile("fxsaveq %0" : "=m" (fpu->state->fxsave));
230 /* Using "rex64; fxsave %0" is broken because, if the memory
231 * operand uses any extended registers for addressing, a second
232 * REX prefix will be generated (to the assembler, rex64
233 * followed by semicolon is a separate instruction), and hence
234 * the 64-bitness is lost.
236 * Using "fxsaveq %0" would be the ideal choice, but is only
237 * supported starting with gas 2.16.
239 * Using, as a workaround, the properly prefixed form below
240 * isn't accepted by any binutils version so far released,
241 * complaining that the same type of prefix is used twice if
242 * an extended register is needed for addressing (fix submitted
243 * to mainline 2005-11-21).
245 * asm volatile("rex64/fxsave %0" : "=m" (fpu->state->fxsave));
247 * This, however, we can work around by forcing the compiler to
248 * select an addressing mode that doesn't require extended
251 asm volatile( "rex64/fxsave (%[fx])"
252 : "=m" (fpu->state->fxsave)
253 : [fx] "R" (&fpu->state->fxsave));
258 * These must be called with preempt disabled. Returns
259 * 'true' if the FPU state is still intact.
261 static inline int fpu_save_init(struct fpu *fpu)
267 * xsave header may indicate the init state of the FP.
269 if (!(fpu->state->xsave.xsave_hdr.xstate_bv & XSTATE_FP))
271 } else if (use_fxsr()) {
274 asm volatile("fnsave %[fx]; fwait"
275 : [fx] "=m" (fpu->state->fsave));
280 * If exceptions are pending, we need to clear them so
281 * that we don't randomly get exceptions later.
283 * FIXME! Is this perhaps only true for the old-style
284 * irq13 case? Maybe we could leave the x87 state
287 if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) {
288 asm volatile("fnclex");
294 static inline int __save_init_fpu(struct task_struct *tsk)
296 return fpu_save_init(&tsk->thread.fpu);
299 static inline int fpu_restore_checking(struct fpu *fpu)
302 return fpu_xrstor_checking(&fpu->state->xsave);
304 return fxrstor_checking(&fpu->state->fxsave);
306 return frstor_checking(&fpu->state->fsave);
309 static inline int restore_fpu_checking(struct task_struct *tsk)
311 /* AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception
312 is pending. Clear the x87 state here by setting it to fixed
313 values. "m" is a random variable that should be in L1 */
314 if (unlikely(static_cpu_has_bug_safe(X86_BUG_FXSAVE_LEAK))) {
318 "fildl %P[addr]" /* set F?P to defined value */
319 : : [addr] "m" (tsk->thread.fpu.has_fpu));
322 return fpu_restore_checking(&tsk->thread.fpu);
326 * Software FPU state helpers. Careful: these need to
327 * be preemption protection *and* they need to be
328 * properly paired with the CR0.TS changes!
330 static inline int __thread_has_fpu(struct task_struct *tsk)
332 return tsk->thread.fpu.has_fpu;
335 /* Must be paired with an 'stts' after! */
336 static inline void __thread_clear_has_fpu(struct task_struct *tsk)
338 tsk->thread.fpu.has_fpu = 0;
339 this_cpu_write(fpu_owner_task, NULL);
342 /* Must be paired with a 'clts' before! */
343 static inline void __thread_set_has_fpu(struct task_struct *tsk)
345 tsk->thread.fpu.has_fpu = 1;
346 this_cpu_write(fpu_owner_task, tsk);
350 * Encapsulate the CR0.TS handling together with the
353 * These generally need preemption protection to work,
354 * do try to avoid using these on their own.
356 static inline void __thread_fpu_end(struct task_struct *tsk)
358 __thread_clear_has_fpu(tsk);
359 if (!use_eager_fpu())
363 static inline void __thread_fpu_begin(struct task_struct *tsk)
365 if (!use_eager_fpu())
367 __thread_set_has_fpu(tsk);
370 static inline void __drop_fpu(struct task_struct *tsk)
372 if (__thread_has_fpu(tsk)) {
373 /* Ignore delayed exceptions from user space */
374 asm volatile("1: fwait\n"
376 _ASM_EXTABLE(1b, 2b));
377 __thread_fpu_end(tsk);
381 static inline void drop_fpu(struct task_struct *tsk)
384 * Forget coprocessor state..
387 tsk->thread.fpu_counter = 0;
393 static inline void drop_init_fpu(struct task_struct *tsk)
395 if (!use_eager_fpu())
399 xrstor_state(init_xstate_buf, -1);
401 fxrstor_checking(&init_xstate_buf->i387);
406 * FPU state switching for scheduling.
408 * This is a two-stage process:
410 * - switch_fpu_prepare() saves the old state and
411 * sets the new state of the CR0.TS bit. This is
412 * done within the context of the old process.
414 * - switch_fpu_finish() restores the new state as
417 typedef struct { int preload; } fpu_switch_t;
419 static inline fpu_switch_t switch_fpu_prepare(struct task_struct *old, struct task_struct *new, int cpu)
424 * If the task has used the math, pre-load the FPU on xsave processors
425 * or if the past 5 consecutive context-switches used math.
427 fpu.preload = tsk_used_math(new) && (use_eager_fpu() ||
428 new->thread.fpu_counter > 5);
429 if (__thread_has_fpu(old)) {
430 if (!__save_init_fpu(old))
432 old->thread.fpu.last_cpu = cpu;
433 old->thread.fpu.has_fpu = 0; /* But leave fpu_owner_task! */
435 /* Don't change CR0.TS if we just switch! */
437 new->thread.fpu_counter++;
438 __thread_set_has_fpu(new);
439 prefetch(new->thread.fpu.state);
440 } else if (!use_eager_fpu())
443 old->thread.fpu_counter = 0;
444 old->thread.fpu.last_cpu = ~0;
446 new->thread.fpu_counter++;
447 if (!use_eager_fpu() && fpu_lazy_restore(new, cpu))
450 prefetch(new->thread.fpu.state);
451 __thread_fpu_begin(new);
458 * By the time this gets called, we've already cleared CR0.TS and
459 * given the process the FPU if we are going to preload the FPU
460 * state - all we need to do is to conditionally restore the register
463 static inline void switch_fpu_finish(struct task_struct *new, fpu_switch_t fpu)
466 if (unlikely(restore_fpu_checking(new)))
472 * Signal frame handlers...
474 extern int save_xstate_sig(void __user *buf, void __user *fx, int size);
475 extern int __restore_xstate_sig(void __user *buf, void __user *fx, int size);
477 static inline int xstate_sigframe_size(void)
479 return use_xsave() ? xstate_size + FP_XSTATE_MAGIC2_SIZE : xstate_size;
482 static inline int restore_xstate_sig(void __user *buf, int ia32_frame)
484 void __user *buf_fx = buf;
485 int size = xstate_sigframe_size();
487 if (ia32_frame && use_fxsr()) {
488 buf_fx = buf + sizeof(struct i387_fsave_struct);
489 size += sizeof(struct i387_fsave_struct);
492 return __restore_xstate_sig(buf, buf_fx, size);
496 * Need to be preemption-safe.
498 * NOTE! user_fpu_begin() must be used only immediately before restoring
499 * it. This function does not do any save/restore on their own.
501 static inline void user_fpu_begin(void)
505 __thread_fpu_begin(current);
509 static inline void __save_fpu(struct task_struct *tsk)
512 if (unlikely(system_state == SYSTEM_BOOTING))
513 xsave_state_booting(&tsk->thread.fpu.state->xsave, -1);
515 xsave_state(&tsk->thread.fpu.state->xsave, -1);
517 fpu_fxsave(&tsk->thread.fpu);
521 * i387 state interaction
523 static inline unsigned short get_fpu_cwd(struct task_struct *tsk)
526 return tsk->thread.fpu.state->fxsave.cwd;
528 return (unsigned short)tsk->thread.fpu.state->fsave.cwd;
532 static inline unsigned short get_fpu_swd(struct task_struct *tsk)
535 return tsk->thread.fpu.state->fxsave.swd;
537 return (unsigned short)tsk->thread.fpu.state->fsave.swd;
541 static inline unsigned short get_fpu_mxcsr(struct task_struct *tsk)
544 return tsk->thread.fpu.state->fxsave.mxcsr;
546 return MXCSR_DEFAULT;
550 static bool fpu_allocated(struct fpu *fpu)
552 return fpu->state != NULL;
555 static inline int fpu_alloc(struct fpu *fpu)
557 if (fpu_allocated(fpu))
559 fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
562 WARN_ON((unsigned long)fpu->state & 15);
566 static inline void fpu_free(struct fpu *fpu)
569 kmem_cache_free(task_xstate_cachep, fpu->state);
574 static inline void fpu_copy(struct task_struct *dst, struct task_struct *src)
576 if (use_eager_fpu()) {
577 memset(&dst->thread.fpu.state->xsave, 0, xstate_size);
580 struct fpu *dfpu = &dst->thread.fpu;
581 struct fpu *sfpu = &src->thread.fpu;
584 memcpy(dfpu->state, sfpu->state, xstate_size);
588 static inline unsigned long
589 alloc_mathframe(unsigned long sp, int ia32_frame, unsigned long *buf_fx,
592 unsigned long frame_size = xstate_sigframe_size();
594 *buf_fx = sp = round_down(sp - frame_size, 64);
595 if (ia32_frame && use_fxsr()) {
596 frame_size += sizeof(struct i387_fsave_struct);
597 sp -= sizeof(struct i387_fsave_struct);