x86/fpu: Eliminate __save_fpu()
[deliverable/linux.git] / arch / x86 / kernel / fpu / core.c
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 */
8 #include <asm/fpu/internal.h>
9 #include <linux/hardirq.h>
10
11 /*
12 * Track whether the kernel is using the FPU state
13 * currently.
14 *
15 * This flag is used:
16 *
17 * - by IRQ context code to potentially use the FPU
18 * if it's unused.
19 *
20 * - to debug kernel_fpu_begin()/end() correctness
21 */
22 static DEFINE_PER_CPU(bool, in_kernel_fpu);
23
24 /*
25 * Track which context is using the FPU on the CPU:
26 */
27 DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
28
29 static void kernel_fpu_disable(void)
30 {
31 WARN_ON(this_cpu_read(in_kernel_fpu));
32 this_cpu_write(in_kernel_fpu, true);
33 }
34
35 static void kernel_fpu_enable(void)
36 {
37 WARN_ON_ONCE(!this_cpu_read(in_kernel_fpu));
38 this_cpu_write(in_kernel_fpu, false);
39 }
40
41 static bool kernel_fpu_disabled(void)
42 {
43 return this_cpu_read(in_kernel_fpu);
44 }
45
46 /*
47 * Were we in an interrupt that interrupted kernel mode?
48 *
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).
54 *
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.
57 */
58 static bool interrupted_kernel_fpu_idle(void)
59 {
60 if (kernel_fpu_disabled())
61 return false;
62
63 if (use_eager_fpu())
64 return true;
65
66 return !current->thread.fpu.fpregs_active && (read_cr0() & X86_CR0_TS);
67 }
68
69 /*
70 * Were we in user mode (or vm86 mode) when we were
71 * interrupted?
72 *
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.
76 */
77 static bool interrupted_user_mode(void)
78 {
79 struct pt_regs *regs = get_irq_regs();
80 return regs && user_mode(regs);
81 }
82
83 /*
84 * Can we use the FPU in kernel mode with the
85 * whole "kernel_fpu_begin/end()" sequence?
86 *
87 * It's always ok in process context (ie "not interrupt")
88 * but it is sometimes ok even from an irq.
89 */
90 bool irq_fpu_usable(void)
91 {
92 return !in_interrupt() ||
93 interrupted_user_mode() ||
94 interrupted_kernel_fpu_idle();
95 }
96 EXPORT_SYMBOL(irq_fpu_usable);
97
98 void __kernel_fpu_begin(void)
99 {
100 struct fpu *fpu = &current->thread.fpu;
101
102 kernel_fpu_disable();
103
104 if (fpu->fpregs_active) {
105 copy_fpregs_to_fpstate(fpu);
106 } else {
107 this_cpu_write(fpu_fpregs_owner_ctx, NULL);
108 __fpregs_activate_hw();
109 }
110 }
111 EXPORT_SYMBOL(__kernel_fpu_begin);
112
113 void __kernel_fpu_end(void)
114 {
115 struct fpu *fpu = &current->thread.fpu;
116
117 if (fpu->fpregs_active) {
118 if (WARN_ON(restore_fpu_checking(fpu)))
119 fpu_reset_state(fpu);
120 } else {
121 __fpregs_deactivate_hw();
122 }
123
124 kernel_fpu_enable();
125 }
126 EXPORT_SYMBOL(__kernel_fpu_end);
127
128 void kernel_fpu_begin(void)
129 {
130 preempt_disable();
131 WARN_ON_ONCE(!irq_fpu_usable());
132 __kernel_fpu_begin();
133 }
134 EXPORT_SYMBOL_GPL(kernel_fpu_begin);
135
136 void kernel_fpu_end(void)
137 {
138 __kernel_fpu_end();
139 preempt_enable();
140 }
141 EXPORT_SYMBOL_GPL(kernel_fpu_end);
142
143 /*
144 * CR0::TS save/restore functions:
145 */
146 int irq_ts_save(void)
147 {
148 /*
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()
152 */
153 if (!in_atomic())
154 return 0;
155
156 if (read_cr0() & X86_CR0_TS) {
157 clts();
158 return 1;
159 }
160
161 return 0;
162 }
163 EXPORT_SYMBOL_GPL(irq_ts_save);
164
165 void irq_ts_restore(int TS_state)
166 {
167 if (TS_state)
168 stts();
169 }
170 EXPORT_SYMBOL_GPL(irq_ts_restore);
171
172 /*
173 * Save the FPU state (initialize it if necessary):
174 *
175 * This only ever gets called for the current task.
176 */
177 void fpu__save(struct fpu *fpu)
178 {
179 WARN_ON(fpu != &current->thread.fpu);
180
181 preempt_disable();
182 if (fpu->fpregs_active) {
183 if (use_eager_fpu()) {
184 copy_fpregs_to_fpstate(fpu);
185 } else {
186 copy_fpregs_to_fpstate(fpu);
187 fpregs_deactivate(fpu);
188 }
189 }
190 preempt_enable();
191 }
192 EXPORT_SYMBOL_GPL(fpu__save);
193
194 void fpstate_init(struct fpu *fpu)
195 {
196 if (!cpu_has_fpu) {
197 finit_soft_fpu(&fpu->state.soft);
198 return;
199 }
200
201 memset(&fpu->state, 0, xstate_size);
202
203 if (cpu_has_fxsr) {
204 fx_finit(&fpu->state.fxsave);
205 } else {
206 struct i387_fsave_struct *fp = &fpu->state.fsave;
207 fp->cwd = 0xffff037fu;
208 fp->swd = 0xffff0000u;
209 fp->twd = 0xffffffffu;
210 fp->fos = 0xffff0000u;
211 }
212 }
213 EXPORT_SYMBOL_GPL(fpstate_init);
214
215 /*
216 * Copy the current task's FPU state to a new task's FPU context.
217 *
218 * In the 'eager' case we just save to the destination context.
219 *
220 * In the 'lazy' case we save to the source context, mark the FPU lazy
221 * via stts() and copy the source context into the destination context.
222 */
223 static void fpu_copy(struct fpu *dst_fpu, struct fpu *src_fpu)
224 {
225 WARN_ON(src_fpu != &current->thread.fpu);
226
227 if (use_eager_fpu()) {
228 memset(&dst_fpu->state.xsave, 0, xstate_size);
229 copy_fpregs_to_fpstate(dst_fpu);
230 } else {
231 fpu__save(src_fpu);
232 memcpy(&dst_fpu->state, &src_fpu->state, xstate_size);
233 }
234 }
235
236 int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
237 {
238 dst_fpu->counter = 0;
239 dst_fpu->fpregs_active = 0;
240 dst_fpu->last_cpu = -1;
241
242 if (src_fpu->fpstate_active)
243 fpu_copy(dst_fpu, src_fpu);
244
245 return 0;
246 }
247
248 /*
249 * Activate the current task's in-memory FPU context,
250 * if it has not been used before:
251 */
252 void fpu__activate_curr(struct fpu *fpu)
253 {
254 WARN_ON_ONCE(fpu != &current->thread.fpu);
255
256 if (!fpu->fpstate_active) {
257 fpstate_init(fpu);
258
259 /* Safe to do for the current task: */
260 fpu->fpstate_active = 1;
261 }
262 }
263 EXPORT_SYMBOL_GPL(fpu__activate_curr);
264
265 /*
266 * This function must be called before we modify a stopped child's
267 * fpstate.
268 *
269 * If the child has not used the FPU before then initialize its
270 * fpstate.
271 *
272 * If the child has used the FPU before then unlazy it.
273 *
274 * [ After this function call, after registers in the fpstate are
275 * modified and the child task has woken up, the child task will
276 * restore the modified FPU state from the modified context. If we
277 * didn't clear its lazy status here then the lazy in-registers
278 * state pending on its former CPU could be restored, corrupting
279 * the modifications. ]
280 *
281 * This function is also called before we read a stopped child's
282 * FPU state - to make sure it's initialized if the child has
283 * no active FPU state.
284 *
285 * TODO: A future optimization would be to skip the unlazying in
286 * the read-only case, it's not strictly necessary for
287 * read-only access to the context.
288 */
289 static void fpu__activate_stopped(struct fpu *child_fpu)
290 {
291 WARN_ON_ONCE(child_fpu == &current->thread.fpu);
292
293 if (child_fpu->fpstate_active) {
294 child_fpu->last_cpu = -1;
295 } else {
296 fpstate_init(child_fpu);
297
298 /* Safe to do for stopped child tasks: */
299 child_fpu->fpstate_active = 1;
300 }
301 }
302
303 /*
304 * 'fpu__restore()' saves the current math information in the
305 * old math state array, and gets the new ones from the current task
306 *
307 * Careful.. There are problems with IBM-designed IRQ13 behaviour.
308 * Don't touch unless you *really* know how it works.
309 *
310 * Must be called with kernel preemption disabled (eg with local
311 * local interrupts as in the case of do_device_not_available).
312 */
313 void fpu__restore(void)
314 {
315 struct task_struct *tsk = current;
316 struct fpu *fpu = &tsk->thread.fpu;
317
318 fpu__activate_curr(fpu);
319
320 /* Avoid __kernel_fpu_begin() right after fpregs_activate() */
321 kernel_fpu_disable();
322 fpregs_activate(fpu);
323 if (unlikely(restore_fpu_checking(fpu))) {
324 fpu_reset_state(fpu);
325 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
326 } else {
327 tsk->thread.fpu.counter++;
328 }
329 kernel_fpu_enable();
330 }
331 EXPORT_SYMBOL_GPL(fpu__restore);
332
333 void fpu__clear(struct task_struct *tsk)
334 {
335 struct fpu *fpu = &tsk->thread.fpu;
336
337 WARN_ON_ONCE(tsk != current); /* Almost certainly an anomaly */
338
339 if (!use_eager_fpu()) {
340 /* FPU state will be reallocated lazily at the first use. */
341 drop_fpu(fpu);
342 } else {
343 if (!fpu->fpstate_active) {
344 fpu__activate_curr(fpu);
345 user_fpu_begin();
346 }
347 restore_init_xstate();
348 }
349 }
350
351 /*
352 * The xstateregs_active() routine is the same as the regset_fpregs_active() routine,
353 * as the "regset->n" for the xstate regset will be updated based on the feature
354 * capabilites supported by the xsave.
355 */
356 int regset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
357 {
358 struct fpu *target_fpu = &target->thread.fpu;
359
360 return target_fpu->fpstate_active ? regset->n : 0;
361 }
362
363 int regset_xregset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
364 {
365 struct fpu *target_fpu = &target->thread.fpu;
366
367 return (cpu_has_fxsr && target_fpu->fpstate_active) ? regset->n : 0;
368 }
369
370 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
371 unsigned int pos, unsigned int count,
372 void *kbuf, void __user *ubuf)
373 {
374 struct fpu *fpu = &target->thread.fpu;
375
376 if (!cpu_has_fxsr)
377 return -ENODEV;
378
379 fpu__activate_stopped(fpu);
380 sanitize_i387_state(target);
381
382 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
383 &fpu->state.fxsave, 0, -1);
384 }
385
386 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
387 unsigned int pos, unsigned int count,
388 const void *kbuf, const void __user *ubuf)
389 {
390 struct fpu *fpu = &target->thread.fpu;
391 int ret;
392
393 if (!cpu_has_fxsr)
394 return -ENODEV;
395
396 fpu__activate_stopped(fpu);
397 sanitize_i387_state(target);
398
399 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
400 &fpu->state.fxsave, 0, -1);
401
402 /*
403 * mxcsr reserved bits must be masked to zero for security reasons.
404 */
405 fpu->state.fxsave.mxcsr &= mxcsr_feature_mask;
406
407 /*
408 * update the header bits in the xsave header, indicating the
409 * presence of FP and SSE state.
410 */
411 if (cpu_has_xsave)
412 fpu->state.xsave.header.xfeatures |= XSTATE_FPSSE;
413
414 return ret;
415 }
416
417 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
418 unsigned int pos, unsigned int count,
419 void *kbuf, void __user *ubuf)
420 {
421 struct fpu *fpu = &target->thread.fpu;
422 struct xsave_struct *xsave;
423 int ret;
424
425 if (!cpu_has_xsave)
426 return -ENODEV;
427
428 fpu__activate_stopped(fpu);
429
430 xsave = &fpu->state.xsave;
431
432 /*
433 * Copy the 48bytes defined by the software first into the xstate
434 * memory layout in the thread struct, so that we can copy the entire
435 * xstateregs to the user using one user_regset_copyout().
436 */
437 memcpy(&xsave->i387.sw_reserved,
438 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
439 /*
440 * Copy the xstate memory layout.
441 */
442 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
443 return ret;
444 }
445
446 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
447 unsigned int pos, unsigned int count,
448 const void *kbuf, const void __user *ubuf)
449 {
450 struct fpu *fpu = &target->thread.fpu;
451 struct xsave_struct *xsave;
452 int ret;
453
454 if (!cpu_has_xsave)
455 return -ENODEV;
456
457 fpu__activate_stopped(fpu);
458
459 xsave = &fpu->state.xsave;
460
461 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
462 /*
463 * mxcsr reserved bits must be masked to zero for security reasons.
464 */
465 xsave->i387.mxcsr &= mxcsr_feature_mask;
466 xsave->header.xfeatures &= xfeatures_mask;
467 /*
468 * These bits must be zero.
469 */
470 memset(&xsave->header.reserved, 0, 48);
471
472 return ret;
473 }
474
475 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
476
477 /*
478 * FPU tag word conversions.
479 */
480
481 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
482 {
483 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
484
485 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
486 tmp = ~twd;
487 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
488 /* and move the valid bits to the lower byte. */
489 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
490 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
491 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
492
493 return tmp;
494 }
495
496 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
497 #define FP_EXP_TAG_VALID 0
498 #define FP_EXP_TAG_ZERO 1
499 #define FP_EXP_TAG_SPECIAL 2
500 #define FP_EXP_TAG_EMPTY 3
501
502 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
503 {
504 struct _fpxreg *st;
505 u32 tos = (fxsave->swd >> 11) & 7;
506 u32 twd = (unsigned long) fxsave->twd;
507 u32 tag;
508 u32 ret = 0xffff0000u;
509 int i;
510
511 for (i = 0; i < 8; i++, twd >>= 1) {
512 if (twd & 0x1) {
513 st = FPREG_ADDR(fxsave, (i - tos) & 7);
514
515 switch (st->exponent & 0x7fff) {
516 case 0x7fff:
517 tag = FP_EXP_TAG_SPECIAL;
518 break;
519 case 0x0000:
520 if (!st->significand[0] &&
521 !st->significand[1] &&
522 !st->significand[2] &&
523 !st->significand[3])
524 tag = FP_EXP_TAG_ZERO;
525 else
526 tag = FP_EXP_TAG_SPECIAL;
527 break;
528 default:
529 if (st->significand[3] & 0x8000)
530 tag = FP_EXP_TAG_VALID;
531 else
532 tag = FP_EXP_TAG_SPECIAL;
533 break;
534 }
535 } else {
536 tag = FP_EXP_TAG_EMPTY;
537 }
538 ret |= tag << (2 * i);
539 }
540 return ret;
541 }
542
543 /*
544 * FXSR floating point environment conversions.
545 */
546
547 void
548 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
549 {
550 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state.fxsave;
551 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
552 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
553 int i;
554
555 env->cwd = fxsave->cwd | 0xffff0000u;
556 env->swd = fxsave->swd | 0xffff0000u;
557 env->twd = twd_fxsr_to_i387(fxsave);
558
559 #ifdef CONFIG_X86_64
560 env->fip = fxsave->rip;
561 env->foo = fxsave->rdp;
562 /*
563 * should be actually ds/cs at fpu exception time, but
564 * that information is not available in 64bit mode.
565 */
566 env->fcs = task_pt_regs(tsk)->cs;
567 if (tsk == current) {
568 savesegment(ds, env->fos);
569 } else {
570 env->fos = tsk->thread.ds;
571 }
572 env->fos |= 0xffff0000;
573 #else
574 env->fip = fxsave->fip;
575 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
576 env->foo = fxsave->foo;
577 env->fos = fxsave->fos;
578 #endif
579
580 for (i = 0; i < 8; ++i)
581 memcpy(&to[i], &from[i], sizeof(to[0]));
582 }
583
584 void convert_to_fxsr(struct task_struct *tsk,
585 const struct user_i387_ia32_struct *env)
586
587 {
588 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state.fxsave;
589 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
590 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
591 int i;
592
593 fxsave->cwd = env->cwd;
594 fxsave->swd = env->swd;
595 fxsave->twd = twd_i387_to_fxsr(env->twd);
596 fxsave->fop = (u16) ((u32) env->fcs >> 16);
597 #ifdef CONFIG_X86_64
598 fxsave->rip = env->fip;
599 fxsave->rdp = env->foo;
600 /* cs and ds ignored */
601 #else
602 fxsave->fip = env->fip;
603 fxsave->fcs = (env->fcs & 0xffff);
604 fxsave->foo = env->foo;
605 fxsave->fos = env->fos;
606 #endif
607
608 for (i = 0; i < 8; ++i)
609 memcpy(&to[i], &from[i], sizeof(from[0]));
610 }
611
612 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
613 unsigned int pos, unsigned int count,
614 void *kbuf, void __user *ubuf)
615 {
616 struct fpu *fpu = &target->thread.fpu;
617 struct user_i387_ia32_struct env;
618
619 fpu__activate_stopped(fpu);
620
621 if (!static_cpu_has(X86_FEATURE_FPU))
622 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
623
624 if (!cpu_has_fxsr)
625 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
626 &fpu->state.fsave, 0,
627 -1);
628
629 sanitize_i387_state(target);
630
631 if (kbuf && pos == 0 && count == sizeof(env)) {
632 convert_from_fxsr(kbuf, target);
633 return 0;
634 }
635
636 convert_from_fxsr(&env, target);
637
638 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
639 }
640
641 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
642 unsigned int pos, unsigned int count,
643 const void *kbuf, const void __user *ubuf)
644 {
645 struct fpu *fpu = &target->thread.fpu;
646 struct user_i387_ia32_struct env;
647 int ret;
648
649 fpu__activate_stopped(fpu);
650
651 sanitize_i387_state(target);
652
653 if (!static_cpu_has(X86_FEATURE_FPU))
654 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
655
656 if (!cpu_has_fxsr)
657 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
658 &fpu->state.fsave, 0,
659 -1);
660
661 if (pos > 0 || count < sizeof(env))
662 convert_from_fxsr(&env, target);
663
664 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
665 if (!ret)
666 convert_to_fxsr(target, &env);
667
668 /*
669 * update the header bit in the xsave header, indicating the
670 * presence of FP.
671 */
672 if (cpu_has_xsave)
673 fpu->state.xsave.header.xfeatures |= XSTATE_FP;
674 return ret;
675 }
676
677 /*
678 * FPU state for core dumps.
679 * This is only used for a.out dumps now.
680 * It is declared generically using elf_fpregset_t (which is
681 * struct user_i387_struct) but is in fact only used for 32-bit
682 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
683 */
684 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *ufpu)
685 {
686 struct task_struct *tsk = current;
687 struct fpu *fpu = &tsk->thread.fpu;
688 int fpvalid;
689
690 fpvalid = fpu->fpstate_active;
691 if (fpvalid)
692 fpvalid = !fpregs_get(tsk, NULL,
693 0, sizeof(struct user_i387_ia32_struct),
694 ufpu, NULL);
695
696 return fpvalid;
697 }
698 EXPORT_SYMBOL(dump_fpu);
699
700 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */
This page took 0.045422 seconds and 5 git commands to generate.