Commit | Line | Data |
---|---|---|
1eeaed76 RM |
1 | /* |
2 | * Copyright (C) 1994 Linus Torvalds | |
3 | * | |
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 | |
8 | */ | |
9 | ||
77ef50a5 VN |
10 | #ifndef ASM_X86__I387_H |
11 | #define ASM_X86__I387_H | |
1eeaed76 RM |
12 | |
13 | #include <linux/sched.h> | |
14 | #include <linux/kernel_stat.h> | |
15 | #include <linux/regset.h> | |
92c37fa3 | 16 | #include <asm/asm.h> |
1eeaed76 RM |
17 | #include <asm/processor.h> |
18 | #include <asm/sigcontext.h> | |
19 | #include <asm/user.h> | |
20 | #include <asm/uaccess.h> | |
dc1e35c6 | 21 | #include <asm/xsave.h> |
1eeaed76 | 22 | |
3c1c7f10 | 23 | extern unsigned int sig_xstate_size; |
1eeaed76 | 24 | extern void fpu_init(void); |
1eeaed76 | 25 | extern void mxcsr_feature_mask_init(void); |
aa283f49 | 26 | extern int init_fpu(struct task_struct *child); |
1eeaed76 | 27 | extern asmlinkage void math_state_restore(void); |
61c4628b | 28 | extern void init_thread_xstate(void); |
1eeaed76 RM |
29 | |
30 | extern user_regset_active_fn fpregs_active, xfpregs_active; | |
31 | extern user_regset_get_fn fpregs_get, xfpregs_get, fpregs_soft_get; | |
32 | extern user_regset_set_fn fpregs_set, xfpregs_set, fpregs_soft_set; | |
33 | ||
34 | #ifdef CONFIG_IA32_EMULATION | |
3c1c7f10 | 35 | extern unsigned int sig_xstate_ia32_size; |
1eeaed76 | 36 | struct _fpstate_ia32; |
ab513701 SS |
37 | struct _xstate_ia32; |
38 | extern int save_i387_xstate_ia32(void __user *buf); | |
39 | extern int restore_i387_xstate_ia32(void __user *buf); | |
1eeaed76 RM |
40 | #endif |
41 | ||
b359e8a4 SS |
42 | #define X87_FSW_ES (1 << 7) /* Exception Summary */ |
43 | ||
1eeaed76 RM |
44 | #ifdef CONFIG_X86_64 |
45 | ||
46 | /* Ignore delayed exceptions from user space */ | |
47 | static inline void tolerant_fwait(void) | |
48 | { | |
49 | asm volatile("1: fwait\n" | |
50 | "2:\n" | |
affe6637 | 51 | _ASM_EXTABLE(1b, 2b)); |
1eeaed76 RM |
52 | } |
53 | ||
b359e8a4 | 54 | static inline int fxrstor_checking(struct i387_fxsave_struct *fx) |
1eeaed76 RM |
55 | { |
56 | int err; | |
57 | ||
58 | asm volatile("1: rex64/fxrstor (%[fx])\n\t" | |
59 | "2:\n" | |
60 | ".section .fixup,\"ax\"\n" | |
61 | "3: movl $-1,%[err]\n" | |
62 | " jmp 2b\n" | |
63 | ".previous\n" | |
affe6637 | 64 | _ASM_EXTABLE(1b, 3b) |
1eeaed76 RM |
65 | : [err] "=r" (err) |
66 | #if 0 /* See comment in __save_init_fpu() below. */ | |
67 | : [fx] "r" (fx), "m" (*fx), "0" (0)); | |
68 | #else | |
69 | : [fx] "cdaSDb" (fx), "m" (*fx), "0" (0)); | |
70 | #endif | |
1eeaed76 RM |
71 | return err; |
72 | } | |
73 | ||
b359e8a4 SS |
74 | static inline int restore_fpu_checking(struct task_struct *tsk) |
75 | { | |
76 | if (task_thread_info(tsk)->status & TS_XSAVE) | |
77 | return xrstor_checking(&tsk->thread.xstate->xsave); | |
78 | else | |
79 | return fxrstor_checking(&tsk->thread.xstate->fxsave); | |
80 | } | |
1eeaed76 RM |
81 | |
82 | /* AMD CPUs don't save/restore FDP/FIP/FOP unless an exception | |
83 | is pending. Clear the x87 state here by setting it to fixed | |
84 | values. The kernel data segment can be sometimes 0 and sometimes | |
85 | new user value. Both should be ok. | |
86 | Use the PDA as safe address because it should be already in L1. */ | |
b359e8a4 | 87 | static inline void clear_fpu_state(struct task_struct *tsk) |
1eeaed76 | 88 | { |
b359e8a4 SS |
89 | struct xsave_struct *xstate = &tsk->thread.xstate->xsave; |
90 | struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave; | |
91 | ||
92 | /* | |
93 | * xsave header may indicate the init state of the FP. | |
94 | */ | |
95 | if ((task_thread_info(tsk)->status & TS_XSAVE) && | |
96 | !(xstate->xsave_hdr.xstate_bv & XSTATE_FP)) | |
97 | return; | |
98 | ||
1eeaed76 | 99 | if (unlikely(fx->swd & X87_FSW_ES)) |
affe6637 | 100 | asm volatile("fnclex"); |
1eeaed76 | 101 | alternative_input(ASM_NOP8 ASM_NOP2, |
affe6637 JP |
102 | " emms\n" /* clear stack tags */ |
103 | " fildl %%gs:0", /* load to clear state */ | |
104 | X86_FEATURE_FXSAVE_LEAK); | |
1eeaed76 RM |
105 | } |
106 | ||
107 | static inline int save_i387_checking(struct i387_fxsave_struct __user *fx) | |
108 | { | |
109 | int err; | |
110 | ||
111 | asm volatile("1: rex64/fxsave (%[fx])\n\t" | |
112 | "2:\n" | |
113 | ".section .fixup,\"ax\"\n" | |
114 | "3: movl $-1,%[err]\n" | |
115 | " jmp 2b\n" | |
116 | ".previous\n" | |
affe6637 | 117 | _ASM_EXTABLE(1b, 3b) |
1eeaed76 RM |
118 | : [err] "=r" (err), "=m" (*fx) |
119 | #if 0 /* See comment in __fxsave_clear() below. */ | |
120 | : [fx] "r" (fx), "0" (0)); | |
121 | #else | |
122 | : [fx] "cdaSDb" (fx), "0" (0)); | |
123 | #endif | |
affe6637 JP |
124 | if (unlikely(err) && |
125 | __clear_user(fx, sizeof(struct i387_fxsave_struct))) | |
1eeaed76 RM |
126 | err = -EFAULT; |
127 | /* No need to clear here because the caller clears USED_MATH */ | |
128 | return err; | |
129 | } | |
130 | ||
b359e8a4 | 131 | static inline void fxsave(struct task_struct *tsk) |
1eeaed76 RM |
132 | { |
133 | /* Using "rex64; fxsave %0" is broken because, if the memory operand | |
134 | uses any extended registers for addressing, a second REX prefix | |
135 | will be generated (to the assembler, rex64 followed by semicolon | |
136 | is a separate instruction), and hence the 64-bitness is lost. */ | |
137 | #if 0 | |
138 | /* Using "fxsaveq %0" would be the ideal choice, but is only supported | |
139 | starting with gas 2.16. */ | |
140 | __asm__ __volatile__("fxsaveq %0" | |
61c4628b | 141 | : "=m" (tsk->thread.xstate->fxsave)); |
1eeaed76 RM |
142 | #elif 0 |
143 | /* Using, as a workaround, the properly prefixed form below isn't | |
144 | accepted by any binutils version so far released, complaining that | |
145 | the same type of prefix is used twice if an extended register is | |
146 | needed for addressing (fix submitted to mainline 2005-11-21). */ | |
147 | __asm__ __volatile__("rex64/fxsave %0" | |
61c4628b | 148 | : "=m" (tsk->thread.xstate->fxsave)); |
1eeaed76 RM |
149 | #else |
150 | /* This, however, we can work around by forcing the compiler to select | |
151 | an addressing mode that doesn't require extended registers. */ | |
61c4628b SS |
152 | __asm__ __volatile__("rex64/fxsave (%1)" |
153 | : "=m" (tsk->thread.xstate->fxsave) | |
154 | : "cdaSDb" (&tsk->thread.xstate->fxsave)); | |
1eeaed76 | 155 | #endif |
b359e8a4 SS |
156 | } |
157 | ||
158 | static inline void __save_init_fpu(struct task_struct *tsk) | |
159 | { | |
160 | if (task_thread_info(tsk)->status & TS_XSAVE) | |
161 | xsave(tsk); | |
162 | else | |
163 | fxsave(tsk); | |
164 | ||
165 | clear_fpu_state(tsk); | |
1eeaed76 RM |
166 | task_thread_info(tsk)->status &= ~TS_USEDFPU; |
167 | } | |
168 | ||
1eeaed76 RM |
169 | #else /* CONFIG_X86_32 */ |
170 | ||
e8a496ac SS |
171 | extern void finit(void); |
172 | ||
1eeaed76 RM |
173 | static inline void tolerant_fwait(void) |
174 | { | |
175 | asm volatile("fnclex ; fwait"); | |
176 | } | |
177 | ||
178 | static inline void restore_fpu(struct task_struct *tsk) | |
179 | { | |
b359e8a4 SS |
180 | if (task_thread_info(tsk)->status & TS_XSAVE) { |
181 | xrstor_checking(&tsk->thread.xstate->xsave); | |
182 | return; | |
183 | } | |
1eeaed76 RM |
184 | /* |
185 | * The "nop" is needed to make the instructions the same | |
186 | * length. | |
187 | */ | |
188 | alternative_input( | |
189 | "nop ; frstor %1", | |
190 | "fxrstor %1", | |
191 | X86_FEATURE_FXSR, | |
61c4628b | 192 | "m" (tsk->thread.xstate->fxsave)); |
1eeaed76 RM |
193 | } |
194 | ||
195 | /* We need a safe address that is cheap to find and that is already | |
196 | in L1 during context switch. The best choices are unfortunately | |
197 | different for UP and SMP */ | |
198 | #ifdef CONFIG_SMP | |
199 | #define safe_address (__per_cpu_offset[0]) | |
200 | #else | |
201 | #define safe_address (kstat_cpu(0).cpustat.user) | |
202 | #endif | |
203 | ||
204 | /* | |
205 | * These must be called with preempt disabled | |
206 | */ | |
207 | static inline void __save_init_fpu(struct task_struct *tsk) | |
208 | { | |
b359e8a4 SS |
209 | if (task_thread_info(tsk)->status & TS_XSAVE) { |
210 | struct xsave_struct *xstate = &tsk->thread.xstate->xsave; | |
211 | struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave; | |
212 | ||
213 | xsave(tsk); | |
214 | ||
215 | /* | |
216 | * xsave header may indicate the init state of the FP. | |
217 | */ | |
218 | if (!(xstate->xsave_hdr.xstate_bv & XSTATE_FP)) | |
219 | goto end; | |
220 | ||
221 | if (unlikely(fx->swd & X87_FSW_ES)) | |
222 | asm volatile("fnclex"); | |
223 | ||
224 | /* | |
225 | * we can do a simple return here or be paranoid :) | |
226 | */ | |
227 | goto clear_state; | |
228 | } | |
229 | ||
1eeaed76 RM |
230 | /* Use more nops than strictly needed in case the compiler |
231 | varies code */ | |
232 | alternative_input( | |
233 | "fnsave %[fx] ;fwait;" GENERIC_NOP8 GENERIC_NOP4, | |
234 | "fxsave %[fx]\n" | |
235 | "bt $7,%[fsw] ; jnc 1f ; fnclex\n1:", | |
236 | X86_FEATURE_FXSR, | |
61c4628b SS |
237 | [fx] "m" (tsk->thread.xstate->fxsave), |
238 | [fsw] "m" (tsk->thread.xstate->fxsave.swd) : "memory"); | |
b359e8a4 | 239 | clear_state: |
1eeaed76 RM |
240 | /* AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception |
241 | is pending. Clear the x87 state here by setting it to fixed | |
242 | values. safe_address is a random variable that should be in L1 */ | |
243 | alternative_input( | |
244 | GENERIC_NOP8 GENERIC_NOP2, | |
245 | "emms\n\t" /* clear stack tags */ | |
246 | "fildl %[addr]", /* set F?P to defined value */ | |
247 | X86_FEATURE_FXSAVE_LEAK, | |
248 | [addr] "m" (safe_address)); | |
b359e8a4 | 249 | end: |
1eeaed76 RM |
250 | task_thread_info(tsk)->status &= ~TS_USEDFPU; |
251 | } | |
252 | ||
ab513701 SS |
253 | #endif /* CONFIG_X86_64 */ |
254 | ||
1eeaed76 RM |
255 | /* |
256 | * Signal frame handlers... | |
257 | */ | |
ab513701 SS |
258 | extern int save_i387_xstate(void __user *buf); |
259 | extern int restore_i387_xstate(void __user *buf); | |
1eeaed76 RM |
260 | |
261 | static inline void __unlazy_fpu(struct task_struct *tsk) | |
262 | { | |
263 | if (task_thread_info(tsk)->status & TS_USEDFPU) { | |
264 | __save_init_fpu(tsk); | |
265 | stts(); | |
266 | } else | |
267 | tsk->fpu_counter = 0; | |
268 | } | |
269 | ||
270 | static inline void __clear_fpu(struct task_struct *tsk) | |
271 | { | |
272 | if (task_thread_info(tsk)->status & TS_USEDFPU) { | |
273 | tolerant_fwait(); | |
274 | task_thread_info(tsk)->status &= ~TS_USEDFPU; | |
275 | stts(); | |
276 | } | |
277 | } | |
278 | ||
279 | static inline void kernel_fpu_begin(void) | |
280 | { | |
281 | struct thread_info *me = current_thread_info(); | |
282 | preempt_disable(); | |
283 | if (me->status & TS_USEDFPU) | |
284 | __save_init_fpu(me->task); | |
285 | else | |
286 | clts(); | |
287 | } | |
288 | ||
289 | static inline void kernel_fpu_end(void) | |
290 | { | |
291 | stts(); | |
292 | preempt_enable(); | |
293 | } | |
294 | ||
295 | #ifdef CONFIG_X86_64 | |
296 | ||
297 | static inline void save_init_fpu(struct task_struct *tsk) | |
298 | { | |
299 | __save_init_fpu(tsk); | |
300 | stts(); | |
301 | } | |
302 | ||
303 | #define unlazy_fpu __unlazy_fpu | |
304 | #define clear_fpu __clear_fpu | |
305 | ||
306 | #else /* CONFIG_X86_32 */ | |
307 | ||
308 | /* | |
309 | * These disable preemption on their own and are safe | |
310 | */ | |
311 | static inline void save_init_fpu(struct task_struct *tsk) | |
312 | { | |
313 | preempt_disable(); | |
314 | __save_init_fpu(tsk); | |
315 | stts(); | |
316 | preempt_enable(); | |
317 | } | |
318 | ||
319 | static inline void unlazy_fpu(struct task_struct *tsk) | |
320 | { | |
321 | preempt_disable(); | |
322 | __unlazy_fpu(tsk); | |
323 | preempt_enable(); | |
324 | } | |
325 | ||
326 | static inline void clear_fpu(struct task_struct *tsk) | |
327 | { | |
328 | preempt_disable(); | |
329 | __clear_fpu(tsk); | |
330 | preempt_enable(); | |
331 | } | |
332 | ||
333 | #endif /* CONFIG_X86_64 */ | |
334 | ||
1eeaed76 RM |
335 | /* |
336 | * i387 state interaction | |
337 | */ | |
338 | static inline unsigned short get_fpu_cwd(struct task_struct *tsk) | |
339 | { | |
340 | if (cpu_has_fxsr) { | |
61c4628b | 341 | return tsk->thread.xstate->fxsave.cwd; |
1eeaed76 | 342 | } else { |
1679f271 | 343 | return (unsigned short)tsk->thread.xstate->fsave.cwd; |
1eeaed76 RM |
344 | } |
345 | } | |
346 | ||
347 | static inline unsigned short get_fpu_swd(struct task_struct *tsk) | |
348 | { | |
349 | if (cpu_has_fxsr) { | |
61c4628b | 350 | return tsk->thread.xstate->fxsave.swd; |
1eeaed76 | 351 | } else { |
1679f271 | 352 | return (unsigned short)tsk->thread.xstate->fsave.swd; |
1eeaed76 RM |
353 | } |
354 | } | |
355 | ||
356 | static inline unsigned short get_fpu_mxcsr(struct task_struct *tsk) | |
357 | { | |
358 | if (cpu_has_xmm) { | |
61c4628b | 359 | return tsk->thread.xstate->fxsave.mxcsr; |
1eeaed76 RM |
360 | } else { |
361 | return MXCSR_DEFAULT; | |
362 | } | |
363 | } | |
364 | ||
77ef50a5 | 365 | #endif /* ASM_X86__I387_H */ |