x86/fpu: Optimize fpu_copy()
[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 (mark it for reload 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 (!copy_fpregs_to_fpstate(fpu))
184 fpregs_deactivate(fpu);
185 }
186 preempt_enable();
187 }
188 EXPORT_SYMBOL_GPL(fpu__save);
189
190 void fpstate_init(struct fpu *fpu)
191 {
192 if (!cpu_has_fpu) {
193 finit_soft_fpu(&fpu->state.soft);
194 return;
195 }
196
197 memset(&fpu->state, 0, xstate_size);
198
199 if (cpu_has_fxsr) {
200 fx_finit(&fpu->state.fxsave);
201 } else {
202 struct i387_fsave_struct *fp = &fpu->state.fsave;
203 fp->cwd = 0xffff037fu;
204 fp->swd = 0xffff0000u;
205 fp->twd = 0xffffffffu;
206 fp->fos = 0xffff0000u;
207 }
208 }
209 EXPORT_SYMBOL_GPL(fpstate_init);
210
211 /*
212 * Copy the current task's FPU state to a new task's FPU context.
213 *
214 * In the 'eager' case we just save to the destination context.
215 *
216 * In the 'lazy' case we save to the source context, mark the FPU lazy
217 * via stts() and copy the source context into the destination context.
218 */
219 static void fpu_copy(struct fpu *dst_fpu, struct fpu *src_fpu)
220 {
221 WARN_ON(src_fpu != &current->thread.fpu);
222
223 if (use_eager_fpu()) {
224 memset(&dst_fpu->state.xsave, 0, xstate_size);
225 copy_fpregs_to_fpstate(dst_fpu);
226 } else {
227 preempt_disable();
228 if (!copy_fpregs_to_fpstate(src_fpu))
229 fpregs_deactivate(src_fpu);
230 preempt_enable();
231 memcpy(&dst_fpu->state, &src_fpu->state, xstate_size);
232 }
233 }
234
235 int fpu__copy(struct fpu *dst_fpu, struct fpu *src_fpu)
236 {
237 dst_fpu->counter = 0;
238 dst_fpu->fpregs_active = 0;
239 dst_fpu->last_cpu = -1;
240
241 if (src_fpu->fpstate_active)
242 fpu_copy(dst_fpu, src_fpu);
243
244 return 0;
245 }
246
247 /*
248 * Activate the current task's in-memory FPU context,
249 * if it has not been used before:
250 */
251 void fpu__activate_curr(struct fpu *fpu)
252 {
253 WARN_ON_ONCE(fpu != &current->thread.fpu);
254
255 if (!fpu->fpstate_active) {
256 fpstate_init(fpu);
257
258 /* Safe to do for the current task: */
259 fpu->fpstate_active = 1;
260 }
261 }
262 EXPORT_SYMBOL_GPL(fpu__activate_curr);
263
264 /*
265 * This function must be called before we modify a stopped child's
266 * fpstate.
267 *
268 * If the child has not used the FPU before then initialize its
269 * fpstate.
270 *
271 * If the child has used the FPU before then unlazy it.
272 *
273 * [ After this function call, after registers in the fpstate are
274 * modified and the child task has woken up, the child task will
275 * restore the modified FPU state from the modified context. If we
276 * didn't clear its lazy status here then the lazy in-registers
277 * state pending on its former CPU could be restored, corrupting
278 * the modifications. ]
279 *
280 * This function is also called before we read a stopped child's
281 * FPU state - to make sure it's initialized if the child has
282 * no active FPU state.
283 *
284 * TODO: A future optimization would be to skip the unlazying in
285 * the read-only case, it's not strictly necessary for
286 * read-only access to the context.
287 */
288 static void fpu__activate_stopped(struct fpu *child_fpu)
289 {
290 WARN_ON_ONCE(child_fpu == &current->thread.fpu);
291
292 if (child_fpu->fpstate_active) {
293 child_fpu->last_cpu = -1;
294 } else {
295 fpstate_init(child_fpu);
296
297 /* Safe to do for stopped child tasks: */
298 child_fpu->fpstate_active = 1;
299 }
300 }
301
302 /*
303 * 'fpu__restore()' saves the current math information in the
304 * old math state array, and gets the new ones from the current task
305 *
306 * Careful.. There are problems with IBM-designed IRQ13 behaviour.
307 * Don't touch unless you *really* know how it works.
308 *
309 * Must be called with kernel preemption disabled (eg with local
310 * local interrupts as in the case of do_device_not_available).
311 */
312 void fpu__restore(void)
313 {
314 struct task_struct *tsk = current;
315 struct fpu *fpu = &tsk->thread.fpu;
316
317 fpu__activate_curr(fpu);
318
319 /* Avoid __kernel_fpu_begin() right after fpregs_activate() */
320 kernel_fpu_disable();
321 fpregs_activate(fpu);
322 if (unlikely(restore_fpu_checking(fpu))) {
323 fpu_reset_state(fpu);
324 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
325 } else {
326 tsk->thread.fpu.counter++;
327 }
328 kernel_fpu_enable();
329 }
330 EXPORT_SYMBOL_GPL(fpu__restore);
331
332 void fpu__clear(struct task_struct *tsk)
333 {
334 struct fpu *fpu = &tsk->thread.fpu;
335
336 WARN_ON_ONCE(tsk != current); /* Almost certainly an anomaly */
337
338 if (!use_eager_fpu()) {
339 /* FPU state will be reallocated lazily at the first use. */
340 drop_fpu(fpu);
341 } else {
342 if (!fpu->fpstate_active) {
343 fpu__activate_curr(fpu);
344 user_fpu_begin();
345 }
346 restore_init_xstate();
347 }
348 }
349
350 /*
351 * The xstateregs_active() routine is the same as the regset_fpregs_active() routine,
352 * as the "regset->n" for the xstate regset will be updated based on the feature
353 * capabilites supported by the xsave.
354 */
355 int regset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
356 {
357 struct fpu *target_fpu = &target->thread.fpu;
358
359 return target_fpu->fpstate_active ? regset->n : 0;
360 }
361
362 int regset_xregset_fpregs_active(struct task_struct *target, const struct user_regset *regset)
363 {
364 struct fpu *target_fpu = &target->thread.fpu;
365
366 return (cpu_has_fxsr && target_fpu->fpstate_active) ? regset->n : 0;
367 }
368
369 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
370 unsigned int pos, unsigned int count,
371 void *kbuf, void __user *ubuf)
372 {
373 struct fpu *fpu = &target->thread.fpu;
374
375 if (!cpu_has_fxsr)
376 return -ENODEV;
377
378 fpu__activate_stopped(fpu);
379 sanitize_i387_state(target);
380
381 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
382 &fpu->state.fxsave, 0, -1);
383 }
384
385 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
386 unsigned int pos, unsigned int count,
387 const void *kbuf, const void __user *ubuf)
388 {
389 struct fpu *fpu = &target->thread.fpu;
390 int ret;
391
392 if (!cpu_has_fxsr)
393 return -ENODEV;
394
395 fpu__activate_stopped(fpu);
396 sanitize_i387_state(target);
397
398 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
399 &fpu->state.fxsave, 0, -1);
400
401 /*
402 * mxcsr reserved bits must be masked to zero for security reasons.
403 */
404 fpu->state.fxsave.mxcsr &= mxcsr_feature_mask;
405
406 /*
407 * update the header bits in the xsave header, indicating the
408 * presence of FP and SSE state.
409 */
410 if (cpu_has_xsave)
411 fpu->state.xsave.header.xfeatures |= XSTATE_FPSSE;
412
413 return ret;
414 }
415
416 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
417 unsigned int pos, unsigned int count,
418 void *kbuf, void __user *ubuf)
419 {
420 struct fpu *fpu = &target->thread.fpu;
421 struct xsave_struct *xsave;
422 int ret;
423
424 if (!cpu_has_xsave)
425 return -ENODEV;
426
427 fpu__activate_stopped(fpu);
428
429 xsave = &fpu->state.xsave;
430
431 /*
432 * Copy the 48bytes defined by the software first into the xstate
433 * memory layout in the thread struct, so that we can copy the entire
434 * xstateregs to the user using one user_regset_copyout().
435 */
436 memcpy(&xsave->i387.sw_reserved,
437 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
438 /*
439 * Copy the xstate memory layout.
440 */
441 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
442 return ret;
443 }
444
445 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
446 unsigned int pos, unsigned int count,
447 const void *kbuf, const void __user *ubuf)
448 {
449 struct fpu *fpu = &target->thread.fpu;
450 struct xsave_struct *xsave;
451 int ret;
452
453 if (!cpu_has_xsave)
454 return -ENODEV;
455
456 fpu__activate_stopped(fpu);
457
458 xsave = &fpu->state.xsave;
459
460 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
461 /*
462 * mxcsr reserved bits must be masked to zero for security reasons.
463 */
464 xsave->i387.mxcsr &= mxcsr_feature_mask;
465 xsave->header.xfeatures &= xfeatures_mask;
466 /*
467 * These bits must be zero.
468 */
469 memset(&xsave->header.reserved, 0, 48);
470
471 return ret;
472 }
473
474 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
475
476 /*
477 * FPU tag word conversions.
478 */
479
480 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
481 {
482 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
483
484 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
485 tmp = ~twd;
486 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
487 /* and move the valid bits to the lower byte. */
488 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
489 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
490 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
491
492 return tmp;
493 }
494
495 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
496 #define FP_EXP_TAG_VALID 0
497 #define FP_EXP_TAG_ZERO 1
498 #define FP_EXP_TAG_SPECIAL 2
499 #define FP_EXP_TAG_EMPTY 3
500
501 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
502 {
503 struct _fpxreg *st;
504 u32 tos = (fxsave->swd >> 11) & 7;
505 u32 twd = (unsigned long) fxsave->twd;
506 u32 tag;
507 u32 ret = 0xffff0000u;
508 int i;
509
510 for (i = 0; i < 8; i++, twd >>= 1) {
511 if (twd & 0x1) {
512 st = FPREG_ADDR(fxsave, (i - tos) & 7);
513
514 switch (st->exponent & 0x7fff) {
515 case 0x7fff:
516 tag = FP_EXP_TAG_SPECIAL;
517 break;
518 case 0x0000:
519 if (!st->significand[0] &&
520 !st->significand[1] &&
521 !st->significand[2] &&
522 !st->significand[3])
523 tag = FP_EXP_TAG_ZERO;
524 else
525 tag = FP_EXP_TAG_SPECIAL;
526 break;
527 default:
528 if (st->significand[3] & 0x8000)
529 tag = FP_EXP_TAG_VALID;
530 else
531 tag = FP_EXP_TAG_SPECIAL;
532 break;
533 }
534 } else {
535 tag = FP_EXP_TAG_EMPTY;
536 }
537 ret |= tag << (2 * i);
538 }
539 return ret;
540 }
541
542 /*
543 * FXSR floating point environment conversions.
544 */
545
546 void
547 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
548 {
549 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state.fxsave;
550 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
551 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
552 int i;
553
554 env->cwd = fxsave->cwd | 0xffff0000u;
555 env->swd = fxsave->swd | 0xffff0000u;
556 env->twd = twd_fxsr_to_i387(fxsave);
557
558 #ifdef CONFIG_X86_64
559 env->fip = fxsave->rip;
560 env->foo = fxsave->rdp;
561 /*
562 * should be actually ds/cs at fpu exception time, but
563 * that information is not available in 64bit mode.
564 */
565 env->fcs = task_pt_regs(tsk)->cs;
566 if (tsk == current) {
567 savesegment(ds, env->fos);
568 } else {
569 env->fos = tsk->thread.ds;
570 }
571 env->fos |= 0xffff0000;
572 #else
573 env->fip = fxsave->fip;
574 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
575 env->foo = fxsave->foo;
576 env->fos = fxsave->fos;
577 #endif
578
579 for (i = 0; i < 8; ++i)
580 memcpy(&to[i], &from[i], sizeof(to[0]));
581 }
582
583 void convert_to_fxsr(struct task_struct *tsk,
584 const struct user_i387_ia32_struct *env)
585
586 {
587 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state.fxsave;
588 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
589 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
590 int i;
591
592 fxsave->cwd = env->cwd;
593 fxsave->swd = env->swd;
594 fxsave->twd = twd_i387_to_fxsr(env->twd);
595 fxsave->fop = (u16) ((u32) env->fcs >> 16);
596 #ifdef CONFIG_X86_64
597 fxsave->rip = env->fip;
598 fxsave->rdp = env->foo;
599 /* cs and ds ignored */
600 #else
601 fxsave->fip = env->fip;
602 fxsave->fcs = (env->fcs & 0xffff);
603 fxsave->foo = env->foo;
604 fxsave->fos = env->fos;
605 #endif
606
607 for (i = 0; i < 8; ++i)
608 memcpy(&to[i], &from[i], sizeof(from[0]));
609 }
610
611 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
612 unsigned int pos, unsigned int count,
613 void *kbuf, void __user *ubuf)
614 {
615 struct fpu *fpu = &target->thread.fpu;
616 struct user_i387_ia32_struct env;
617
618 fpu__activate_stopped(fpu);
619
620 if (!static_cpu_has(X86_FEATURE_FPU))
621 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
622
623 if (!cpu_has_fxsr)
624 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
625 &fpu->state.fsave, 0,
626 -1);
627
628 sanitize_i387_state(target);
629
630 if (kbuf && pos == 0 && count == sizeof(env)) {
631 convert_from_fxsr(kbuf, target);
632 return 0;
633 }
634
635 convert_from_fxsr(&env, target);
636
637 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
638 }
639
640 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
641 unsigned int pos, unsigned int count,
642 const void *kbuf, const void __user *ubuf)
643 {
644 struct fpu *fpu = &target->thread.fpu;
645 struct user_i387_ia32_struct env;
646 int ret;
647
648 fpu__activate_stopped(fpu);
649
650 sanitize_i387_state(target);
651
652 if (!static_cpu_has(X86_FEATURE_FPU))
653 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
654
655 if (!cpu_has_fxsr)
656 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
657 &fpu->state.fsave, 0,
658 -1);
659
660 if (pos > 0 || count < sizeof(env))
661 convert_from_fxsr(&env, target);
662
663 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
664 if (!ret)
665 convert_to_fxsr(target, &env);
666
667 /*
668 * update the header bit in the xsave header, indicating the
669 * presence of FP.
670 */
671 if (cpu_has_xsave)
672 fpu->state.xsave.header.xfeatures |= XSTATE_FP;
673 return ret;
674 }
675
676 /*
677 * FPU state for core dumps.
678 * This is only used for a.out dumps now.
679 * It is declared generically using elf_fpregset_t (which is
680 * struct user_i387_struct) but is in fact only used for 32-bit
681 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
682 */
683 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *ufpu)
684 {
685 struct task_struct *tsk = current;
686 struct fpu *fpu = &tsk->thread.fpu;
687 int fpvalid;
688
689 fpvalid = fpu->fpstate_active;
690 if (fpvalid)
691 fpvalid = !fpregs_get(tsk, NULL,
692 0, sizeof(struct user_i387_ia32_struct),
693 ufpu, NULL);
694
695 return fpvalid;
696 }
697 EXPORT_SYMBOL(dump_fpu);
698
699 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */
This page took 0.046602 seconds and 5 git commands to generate.