]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - arch/x86/include/asm/i387.h
Merge branch 'upstream' into for-linus
[mirror_ubuntu-artful-kernel.git] / arch / x86 / include / asm / i387.h
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
10 #ifndef _ASM_X86_I387_H
11 #define _ASM_X86_I387_H
12
13 #ifndef __ASSEMBLY__
14
15 #include <linux/sched.h>
16 #include <linux/kernel_stat.h>
17 #include <linux/regset.h>
18 #include <linux/hardirq.h>
19 #include <linux/slab.h>
20 #include <asm/asm.h>
21 #include <asm/cpufeature.h>
22 #include <asm/processor.h>
23 #include <asm/sigcontext.h>
24 #include <asm/user.h>
25 #include <asm/uaccess.h>
26 #include <asm/xsave.h>
27
28 extern unsigned int sig_xstate_size;
29 extern void fpu_init(void);
30 extern void mxcsr_feature_mask_init(void);
31 extern int init_fpu(struct task_struct *child);
32 extern void math_state_restore(void);
33 extern int dump_fpu(struct pt_regs *, struct user_i387_struct *);
34
35 DECLARE_PER_CPU(struct task_struct *, fpu_owner_task);
36
37 extern user_regset_active_fn fpregs_active, xfpregs_active;
38 extern user_regset_get_fn fpregs_get, xfpregs_get, fpregs_soft_get,
39 xstateregs_get;
40 extern user_regset_set_fn fpregs_set, xfpregs_set, fpregs_soft_set,
41 xstateregs_set;
42
43 /*
44 * xstateregs_active == fpregs_active. Please refer to the comment
45 * at the definition of fpregs_active.
46 */
47 #define xstateregs_active fpregs_active
48
49 extern struct _fpx_sw_bytes fx_sw_reserved;
50 #ifdef CONFIG_IA32_EMULATION
51 extern unsigned int sig_xstate_ia32_size;
52 extern struct _fpx_sw_bytes fx_sw_reserved_ia32;
53 struct _fpstate_ia32;
54 struct _xstate_ia32;
55 extern int save_i387_xstate_ia32(void __user *buf);
56 extern int restore_i387_xstate_ia32(void __user *buf);
57 #endif
58
59 #ifdef CONFIG_MATH_EMULATION
60 extern void finit_soft_fpu(struct i387_soft_struct *soft);
61 #else
62 static inline void finit_soft_fpu(struct i387_soft_struct *soft) {}
63 #endif
64
65 #define X87_FSW_ES (1 << 7) /* Exception Summary */
66
67 static __always_inline __pure bool use_xsaveopt(void)
68 {
69 return static_cpu_has(X86_FEATURE_XSAVEOPT);
70 }
71
72 static __always_inline __pure bool use_xsave(void)
73 {
74 return static_cpu_has(X86_FEATURE_XSAVE);
75 }
76
77 static __always_inline __pure bool use_fxsr(void)
78 {
79 return static_cpu_has(X86_FEATURE_FXSR);
80 }
81
82 extern void __sanitize_i387_state(struct task_struct *);
83
84 static inline void sanitize_i387_state(struct task_struct *tsk)
85 {
86 if (!use_xsaveopt())
87 return;
88 __sanitize_i387_state(tsk);
89 }
90
91 #ifdef CONFIG_X86_64
92 static inline int fxrstor_checking(struct i387_fxsave_struct *fx)
93 {
94 int err;
95
96 /* See comment in fxsave() below. */
97 #ifdef CONFIG_AS_FXSAVEQ
98 asm volatile("1: fxrstorq %[fx]\n\t"
99 "2:\n"
100 ".section .fixup,\"ax\"\n"
101 "3: movl $-1,%[err]\n"
102 " jmp 2b\n"
103 ".previous\n"
104 _ASM_EXTABLE(1b, 3b)
105 : [err] "=r" (err)
106 : [fx] "m" (*fx), "0" (0));
107 #else
108 asm volatile("1: rex64/fxrstor (%[fx])\n\t"
109 "2:\n"
110 ".section .fixup,\"ax\"\n"
111 "3: movl $-1,%[err]\n"
112 " jmp 2b\n"
113 ".previous\n"
114 _ASM_EXTABLE(1b, 3b)
115 : [err] "=r" (err)
116 : [fx] "R" (fx), "m" (*fx), "0" (0));
117 #endif
118 return err;
119 }
120
121 static inline int fxsave_user(struct i387_fxsave_struct __user *fx)
122 {
123 int err;
124
125 /*
126 * Clear the bytes not touched by the fxsave and reserved
127 * for the SW usage.
128 */
129 err = __clear_user(&fx->sw_reserved,
130 sizeof(struct _fpx_sw_bytes));
131 if (unlikely(err))
132 return -EFAULT;
133
134 /* See comment in fxsave() below. */
135 #ifdef CONFIG_AS_FXSAVEQ
136 asm volatile("1: fxsaveq %[fx]\n\t"
137 "2:\n"
138 ".section .fixup,\"ax\"\n"
139 "3: movl $-1,%[err]\n"
140 " jmp 2b\n"
141 ".previous\n"
142 _ASM_EXTABLE(1b, 3b)
143 : [err] "=r" (err), [fx] "=m" (*fx)
144 : "0" (0));
145 #else
146 asm volatile("1: rex64/fxsave (%[fx])\n\t"
147 "2:\n"
148 ".section .fixup,\"ax\"\n"
149 "3: movl $-1,%[err]\n"
150 " jmp 2b\n"
151 ".previous\n"
152 _ASM_EXTABLE(1b, 3b)
153 : [err] "=r" (err), "=m" (*fx)
154 : [fx] "R" (fx), "0" (0));
155 #endif
156 if (unlikely(err) &&
157 __clear_user(fx, sizeof(struct i387_fxsave_struct)))
158 err = -EFAULT;
159 /* No need to clear here because the caller clears USED_MATH */
160 return err;
161 }
162
163 static inline void fpu_fxsave(struct fpu *fpu)
164 {
165 /* Using "rex64; fxsave %0" is broken because, if the memory operand
166 uses any extended registers for addressing, a second REX prefix
167 will be generated (to the assembler, rex64 followed by semicolon
168 is a separate instruction), and hence the 64-bitness is lost. */
169
170 #ifdef CONFIG_AS_FXSAVEQ
171 /* Using "fxsaveq %0" would be the ideal choice, but is only supported
172 starting with gas 2.16. */
173 __asm__ __volatile__("fxsaveq %0"
174 : "=m" (fpu->state->fxsave));
175 #else
176 /* Using, as a workaround, the properly prefixed form below isn't
177 accepted by any binutils version so far released, complaining that
178 the same type of prefix is used twice if an extended register is
179 needed for addressing (fix submitted to mainline 2005-11-21).
180 asm volatile("rex64/fxsave %0"
181 : "=m" (fpu->state->fxsave));
182 This, however, we can work around by forcing the compiler to select
183 an addressing mode that doesn't require extended registers. */
184 asm volatile("rex64/fxsave (%[fx])"
185 : "=m" (fpu->state->fxsave)
186 : [fx] "R" (&fpu->state->fxsave));
187 #endif
188 }
189
190 #else /* CONFIG_X86_32 */
191
192 /* perform fxrstor iff the processor has extended states, otherwise frstor */
193 static inline int fxrstor_checking(struct i387_fxsave_struct *fx)
194 {
195 /*
196 * The "nop" is needed to make the instructions the same
197 * length.
198 */
199 alternative_input(
200 "nop ; frstor %1",
201 "fxrstor %1",
202 X86_FEATURE_FXSR,
203 "m" (*fx));
204
205 return 0;
206 }
207
208 static inline void fpu_fxsave(struct fpu *fpu)
209 {
210 asm volatile("fxsave %[fx]"
211 : [fx] "=m" (fpu->state->fxsave));
212 }
213
214 #endif /* CONFIG_X86_64 */
215
216 /*
217 * These must be called with preempt disabled. Returns
218 * 'true' if the FPU state is still intact.
219 */
220 static inline int fpu_save_init(struct fpu *fpu)
221 {
222 if (use_xsave()) {
223 fpu_xsave(fpu);
224
225 /*
226 * xsave header may indicate the init state of the FP.
227 */
228 if (!(fpu->state->xsave.xsave_hdr.xstate_bv & XSTATE_FP))
229 return 1;
230 } else if (use_fxsr()) {
231 fpu_fxsave(fpu);
232 } else {
233 asm volatile("fnsave %[fx]; fwait"
234 : [fx] "=m" (fpu->state->fsave));
235 return 0;
236 }
237
238 /*
239 * If exceptions are pending, we need to clear them so
240 * that we don't randomly get exceptions later.
241 *
242 * FIXME! Is this perhaps only true for the old-style
243 * irq13 case? Maybe we could leave the x87 state
244 * intact otherwise?
245 */
246 if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) {
247 asm volatile("fnclex");
248 return 0;
249 }
250 return 1;
251 }
252
253 static inline int __save_init_fpu(struct task_struct *tsk)
254 {
255 return fpu_save_init(&tsk->thread.fpu);
256 }
257
258 static inline int fpu_fxrstor_checking(struct fpu *fpu)
259 {
260 return fxrstor_checking(&fpu->state->fxsave);
261 }
262
263 static inline int fpu_restore_checking(struct fpu *fpu)
264 {
265 if (use_xsave())
266 return fpu_xrstor_checking(fpu);
267 else
268 return fpu_fxrstor_checking(fpu);
269 }
270
271 static inline int restore_fpu_checking(struct task_struct *tsk)
272 {
273 /* AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception
274 is pending. Clear the x87 state here by setting it to fixed
275 values. "m" is a random variable that should be in L1 */
276 alternative_input(
277 ASM_NOP8 ASM_NOP2,
278 "emms\n\t" /* clear stack tags */
279 "fildl %P[addr]", /* set F?P to defined value */
280 X86_FEATURE_FXSAVE_LEAK,
281 [addr] "m" (tsk->thread.fpu.has_fpu));
282
283 return fpu_restore_checking(&tsk->thread.fpu);
284 }
285
286 /*
287 * Software FPU state helpers. Careful: these need to
288 * be preemption protection *and* they need to be
289 * properly paired with the CR0.TS changes!
290 */
291 static inline int __thread_has_fpu(struct task_struct *tsk)
292 {
293 return tsk->thread.fpu.has_fpu;
294 }
295
296 /* Must be paired with an 'stts' after! */
297 static inline void __thread_clear_has_fpu(struct task_struct *tsk)
298 {
299 tsk->thread.fpu.has_fpu = 0;
300 percpu_write(fpu_owner_task, NULL);
301 }
302
303 /* Must be paired with a 'clts' before! */
304 static inline void __thread_set_has_fpu(struct task_struct *tsk)
305 {
306 tsk->thread.fpu.has_fpu = 1;
307 percpu_write(fpu_owner_task, tsk);
308 }
309
310 /*
311 * Encapsulate the CR0.TS handling together with the
312 * software flag.
313 *
314 * These generally need preemption protection to work,
315 * do try to avoid using these on their own.
316 */
317 static inline void __thread_fpu_end(struct task_struct *tsk)
318 {
319 __thread_clear_has_fpu(tsk);
320 stts();
321 }
322
323 static inline void __thread_fpu_begin(struct task_struct *tsk)
324 {
325 clts();
326 __thread_set_has_fpu(tsk);
327 }
328
329 /*
330 * FPU state switching for scheduling.
331 *
332 * This is a two-stage process:
333 *
334 * - switch_fpu_prepare() saves the old state and
335 * sets the new state of the CR0.TS bit. This is
336 * done within the context of the old process.
337 *
338 * - switch_fpu_finish() restores the new state as
339 * necessary.
340 */
341 typedef struct { int preload; } fpu_switch_t;
342
343 /*
344 * FIXME! We could do a totally lazy restore, but we need to
345 * add a per-cpu "this was the task that last touched the FPU
346 * on this CPU" variable, and the task needs to have a "I last
347 * touched the FPU on this CPU" and check them.
348 *
349 * We don't do that yet, so "fpu_lazy_restore()" always returns
350 * false, but some day..
351 */
352 static inline int fpu_lazy_restore(struct task_struct *new, unsigned int cpu)
353 {
354 return new == percpu_read_stable(fpu_owner_task) &&
355 cpu == new->thread.fpu.last_cpu;
356 }
357
358 static inline fpu_switch_t switch_fpu_prepare(struct task_struct *old, struct task_struct *new, int cpu)
359 {
360 fpu_switch_t fpu;
361
362 fpu.preload = tsk_used_math(new) && new->fpu_counter > 5;
363 if (__thread_has_fpu(old)) {
364 if (!__save_init_fpu(old))
365 cpu = ~0;
366 old->thread.fpu.last_cpu = cpu;
367 old->thread.fpu.has_fpu = 0; /* But leave fpu_owner_task! */
368
369 /* Don't change CR0.TS if we just switch! */
370 if (fpu.preload) {
371 new->fpu_counter++;
372 __thread_set_has_fpu(new);
373 prefetch(new->thread.fpu.state);
374 } else
375 stts();
376 } else {
377 old->fpu_counter = 0;
378 old->thread.fpu.last_cpu = ~0;
379 if (fpu.preload) {
380 new->fpu_counter++;
381 if (fpu_lazy_restore(new, cpu))
382 fpu.preload = 0;
383 else
384 prefetch(new->thread.fpu.state);
385 __thread_fpu_begin(new);
386 }
387 }
388 return fpu;
389 }
390
391 /*
392 * By the time this gets called, we've already cleared CR0.TS and
393 * given the process the FPU if we are going to preload the FPU
394 * state - all we need to do is to conditionally restore the register
395 * state itself.
396 */
397 static inline void switch_fpu_finish(struct task_struct *new, fpu_switch_t fpu)
398 {
399 if (fpu.preload) {
400 if (unlikely(restore_fpu_checking(new)))
401 __thread_fpu_end(new);
402 }
403 }
404
405 /*
406 * Signal frame handlers...
407 */
408 extern int save_i387_xstate(void __user *buf);
409 extern int restore_i387_xstate(void __user *buf);
410
411 static inline void __clear_fpu(struct task_struct *tsk)
412 {
413 if (__thread_has_fpu(tsk)) {
414 /* Ignore delayed exceptions from user space */
415 asm volatile("1: fwait\n"
416 "2:\n"
417 _ASM_EXTABLE(1b, 2b));
418 __thread_fpu_end(tsk);
419 }
420 }
421
422 /*
423 * Were we in an interrupt that interrupted kernel mode?
424 *
425 * We can do a kernel_fpu_begin/end() pair *ONLY* if that
426 * pair does nothing at all: the thread must not have fpu (so
427 * that we don't try to save the FPU state), and TS must
428 * be set (so that the clts/stts pair does nothing that is
429 * visible in the interrupted kernel thread).
430 */
431 static inline bool interrupted_kernel_fpu_idle(void)
432 {
433 return !__thread_has_fpu(current) &&
434 (read_cr0() & X86_CR0_TS);
435 }
436
437 /*
438 * Were we in user mode (or vm86 mode) when we were
439 * interrupted?
440 *
441 * Doing kernel_fpu_begin/end() is ok if we are running
442 * in an interrupt context from user mode - we'll just
443 * save the FPU state as required.
444 */
445 static inline bool interrupted_user_mode(void)
446 {
447 struct pt_regs *regs = get_irq_regs();
448 return regs && user_mode_vm(regs);
449 }
450
451 /*
452 * Can we use the FPU in kernel mode with the
453 * whole "kernel_fpu_begin/end()" sequence?
454 *
455 * It's always ok in process context (ie "not interrupt")
456 * but it is sometimes ok even from an irq.
457 */
458 static inline bool irq_fpu_usable(void)
459 {
460 return !in_interrupt() ||
461 interrupted_user_mode() ||
462 interrupted_kernel_fpu_idle();
463 }
464
465 static inline void kernel_fpu_begin(void)
466 {
467 struct task_struct *me = current;
468
469 WARN_ON_ONCE(!irq_fpu_usable());
470 preempt_disable();
471 if (__thread_has_fpu(me)) {
472 __save_init_fpu(me);
473 __thread_clear_has_fpu(me);
474 /* We do 'stts()' in kernel_fpu_end() */
475 } else {
476 percpu_write(fpu_owner_task, NULL);
477 clts();
478 }
479 }
480
481 static inline void kernel_fpu_end(void)
482 {
483 stts();
484 preempt_enable();
485 }
486
487 /*
488 * Some instructions like VIA's padlock instructions generate a spurious
489 * DNA fault but don't modify SSE registers. And these instructions
490 * get used from interrupt context as well. To prevent these kernel instructions
491 * in interrupt context interacting wrongly with other user/kernel fpu usage, we
492 * should use them only in the context of irq_ts_save/restore()
493 */
494 static inline int irq_ts_save(void)
495 {
496 /*
497 * If in process context and not atomic, we can take a spurious DNA fault.
498 * Otherwise, doing clts() in process context requires disabling preemption
499 * or some heavy lifting like kernel_fpu_begin()
500 */
501 if (!in_atomic())
502 return 0;
503
504 if (read_cr0() & X86_CR0_TS) {
505 clts();
506 return 1;
507 }
508
509 return 0;
510 }
511
512 static inline void irq_ts_restore(int TS_state)
513 {
514 if (TS_state)
515 stts();
516 }
517
518 /*
519 * The question "does this thread have fpu access?"
520 * is slightly racy, since preemption could come in
521 * and revoke it immediately after the test.
522 *
523 * However, even in that very unlikely scenario,
524 * we can just assume we have FPU access - typically
525 * to save the FP state - we'll just take a #NM
526 * fault and get the FPU access back.
527 *
528 * The actual user_fpu_begin/end() functions
529 * need to be preemption-safe, though.
530 *
531 * NOTE! user_fpu_end() must be used only after you
532 * have saved the FP state, and user_fpu_begin() must
533 * be used only immediately before restoring it.
534 * These functions do not do any save/restore on
535 * their own.
536 */
537 static inline int user_has_fpu(void)
538 {
539 return __thread_has_fpu(current);
540 }
541
542 static inline void user_fpu_end(void)
543 {
544 preempt_disable();
545 __thread_fpu_end(current);
546 preempt_enable();
547 }
548
549 static inline void user_fpu_begin(void)
550 {
551 preempt_disable();
552 if (!user_has_fpu())
553 __thread_fpu_begin(current);
554 preempt_enable();
555 }
556
557 /*
558 * These disable preemption on their own and are safe
559 */
560 static inline void save_init_fpu(struct task_struct *tsk)
561 {
562 WARN_ON_ONCE(!__thread_has_fpu(tsk));
563 preempt_disable();
564 __save_init_fpu(tsk);
565 __thread_fpu_end(tsk);
566 preempt_enable();
567 }
568
569 static inline void unlazy_fpu(struct task_struct *tsk)
570 {
571 preempt_disable();
572 if (__thread_has_fpu(tsk)) {
573 __save_init_fpu(tsk);
574 __thread_fpu_end(tsk);
575 } else
576 tsk->fpu_counter = 0;
577 preempt_enable();
578 }
579
580 static inline void clear_fpu(struct task_struct *tsk)
581 {
582 preempt_disable();
583 __clear_fpu(tsk);
584 preempt_enable();
585 }
586
587 /*
588 * i387 state interaction
589 */
590 static inline unsigned short get_fpu_cwd(struct task_struct *tsk)
591 {
592 if (cpu_has_fxsr) {
593 return tsk->thread.fpu.state->fxsave.cwd;
594 } else {
595 return (unsigned short)tsk->thread.fpu.state->fsave.cwd;
596 }
597 }
598
599 static inline unsigned short get_fpu_swd(struct task_struct *tsk)
600 {
601 if (cpu_has_fxsr) {
602 return tsk->thread.fpu.state->fxsave.swd;
603 } else {
604 return (unsigned short)tsk->thread.fpu.state->fsave.swd;
605 }
606 }
607
608 static inline unsigned short get_fpu_mxcsr(struct task_struct *tsk)
609 {
610 if (cpu_has_xmm) {
611 return tsk->thread.fpu.state->fxsave.mxcsr;
612 } else {
613 return MXCSR_DEFAULT;
614 }
615 }
616
617 static bool fpu_allocated(struct fpu *fpu)
618 {
619 return fpu->state != NULL;
620 }
621
622 static inline int fpu_alloc(struct fpu *fpu)
623 {
624 if (fpu_allocated(fpu))
625 return 0;
626 fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
627 if (!fpu->state)
628 return -ENOMEM;
629 WARN_ON((unsigned long)fpu->state & 15);
630 return 0;
631 }
632
633 static inline void fpu_free(struct fpu *fpu)
634 {
635 if (fpu->state) {
636 kmem_cache_free(task_xstate_cachep, fpu->state);
637 fpu->state = NULL;
638 }
639 }
640
641 static inline void fpu_copy(struct fpu *dst, struct fpu *src)
642 {
643 memcpy(dst->state, src->state, xstate_size);
644 }
645
646 extern void fpu_finit(struct fpu *fpu);
647
648 #endif /* __ASSEMBLY__ */
649
650 #endif /* _ASM_X86_I387_H */