Loading...
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 <linux/module.h>
9#include <linux/regset.h>
10#include <linux/sched.h>
11#include <linux/slab.h>
12
13#include <asm/sigcontext.h>
14#include <asm/processor.h>
15#include <asm/math_emu.h>
16#include <asm/uaccess.h>
17#include <asm/ptrace.h>
18#include <asm/i387.h>
19#include <asm/fpu-internal.h>
20#include <asm/user.h>
21
22#ifdef CONFIG_X86_64
23# include <asm/sigcontext32.h>
24# include <asm/user32.h>
25#else
26# define save_i387_xstate_ia32 save_i387_xstate
27# define restore_i387_xstate_ia32 restore_i387_xstate
28# define _fpstate_ia32 _fpstate
29# define _xstate_ia32 _xstate
30# define sig_xstate_ia32_size sig_xstate_size
31# define fx_sw_reserved_ia32 fx_sw_reserved
32# define user_i387_ia32_struct user_i387_struct
33# define user32_fxsr_struct user_fxsr_struct
34#endif
35
36/*
37 * Were we in an interrupt that interrupted kernel mode?
38 *
39 * We can do a kernel_fpu_begin/end() pair *ONLY* if that
40 * pair does nothing at all: the thread must not have fpu (so
41 * that we don't try to save the FPU state), and TS must
42 * be set (so that the clts/stts pair does nothing that is
43 * visible in the interrupted kernel thread).
44 */
45static inline bool interrupted_kernel_fpu_idle(void)
46{
47 return !__thread_has_fpu(current) &&
48 (read_cr0() & X86_CR0_TS);
49}
50
51/*
52 * Were we in user mode (or vm86 mode) when we were
53 * interrupted?
54 *
55 * Doing kernel_fpu_begin/end() is ok if we are running
56 * in an interrupt context from user mode - we'll just
57 * save the FPU state as required.
58 */
59static inline bool interrupted_user_mode(void)
60{
61 struct pt_regs *regs = get_irq_regs();
62 return regs && user_mode_vm(regs);
63}
64
65/*
66 * Can we use the FPU in kernel mode with the
67 * whole "kernel_fpu_begin/end()" sequence?
68 *
69 * It's always ok in process context (ie "not interrupt")
70 * but it is sometimes ok even from an irq.
71 */
72bool irq_fpu_usable(void)
73{
74 return !in_interrupt() ||
75 interrupted_user_mode() ||
76 interrupted_kernel_fpu_idle();
77}
78EXPORT_SYMBOL(irq_fpu_usable);
79
80void kernel_fpu_begin(void)
81{
82 struct task_struct *me = current;
83
84 WARN_ON_ONCE(!irq_fpu_usable());
85 preempt_disable();
86 if (__thread_has_fpu(me)) {
87 __save_init_fpu(me);
88 __thread_clear_has_fpu(me);
89 /* We do 'stts()' in kernel_fpu_end() */
90 } else {
91 this_cpu_write(fpu_owner_task, NULL);
92 clts();
93 }
94}
95EXPORT_SYMBOL(kernel_fpu_begin);
96
97void kernel_fpu_end(void)
98{
99 stts();
100 preempt_enable();
101}
102EXPORT_SYMBOL(kernel_fpu_end);
103
104void unlazy_fpu(struct task_struct *tsk)
105{
106 preempt_disable();
107 if (__thread_has_fpu(tsk)) {
108 __save_init_fpu(tsk);
109 __thread_fpu_end(tsk);
110 } else
111 tsk->fpu_counter = 0;
112 preempt_enable();
113}
114EXPORT_SYMBOL(unlazy_fpu);
115
116#ifdef CONFIG_MATH_EMULATION
117# define HAVE_HWFP (boot_cpu_data.hard_math)
118#else
119# define HAVE_HWFP 1
120#endif
121
122static unsigned int mxcsr_feature_mask __read_mostly = 0xffffffffu;
123unsigned int xstate_size;
124EXPORT_SYMBOL_GPL(xstate_size);
125unsigned int sig_xstate_ia32_size = sizeof(struct _fpstate_ia32);
126static struct i387_fxsave_struct fx_scratch __cpuinitdata;
127
128static void __cpuinit mxcsr_feature_mask_init(void)
129{
130 unsigned long mask = 0;
131
132 clts();
133 if (cpu_has_fxsr) {
134 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
135 asm volatile("fxsave %0" : : "m" (fx_scratch));
136 mask = fx_scratch.mxcsr_mask;
137 if (mask == 0)
138 mask = 0x0000ffbf;
139 }
140 mxcsr_feature_mask &= mask;
141 stts();
142}
143
144static void __cpuinit init_thread_xstate(void)
145{
146 /*
147 * Note that xstate_size might be overwriten later during
148 * xsave_init().
149 */
150
151 if (!HAVE_HWFP) {
152 /*
153 * Disable xsave as we do not support it if i387
154 * emulation is enabled.
155 */
156 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
157 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
158 xstate_size = sizeof(struct i387_soft_struct);
159 return;
160 }
161
162 if (cpu_has_fxsr)
163 xstate_size = sizeof(struct i387_fxsave_struct);
164 else
165 xstate_size = sizeof(struct i387_fsave_struct);
166}
167
168/*
169 * Called at bootup to set up the initial FPU state that is later cloned
170 * into all processes.
171 */
172
173void __cpuinit fpu_init(void)
174{
175 unsigned long cr0;
176 unsigned long cr4_mask = 0;
177
178 if (cpu_has_fxsr)
179 cr4_mask |= X86_CR4_OSFXSR;
180 if (cpu_has_xmm)
181 cr4_mask |= X86_CR4_OSXMMEXCPT;
182 if (cr4_mask)
183 set_in_cr4(cr4_mask);
184
185 cr0 = read_cr0();
186 cr0 &= ~(X86_CR0_TS|X86_CR0_EM); /* clear TS and EM */
187 if (!HAVE_HWFP)
188 cr0 |= X86_CR0_EM;
189 write_cr0(cr0);
190
191 if (!smp_processor_id())
192 init_thread_xstate();
193
194 mxcsr_feature_mask_init();
195 /* clean state in init */
196 current_thread_info()->status = 0;
197 clear_used_math();
198}
199
200void fpu_finit(struct fpu *fpu)
201{
202 if (!HAVE_HWFP) {
203 finit_soft_fpu(&fpu->state->soft);
204 return;
205 }
206
207 if (cpu_has_fxsr) {
208 struct i387_fxsave_struct *fx = &fpu->state->fxsave;
209
210 memset(fx, 0, xstate_size);
211 fx->cwd = 0x37f;
212 if (cpu_has_xmm)
213 fx->mxcsr = MXCSR_DEFAULT;
214 } else {
215 struct i387_fsave_struct *fp = &fpu->state->fsave;
216 memset(fp, 0, xstate_size);
217 fp->cwd = 0xffff037fu;
218 fp->swd = 0xffff0000u;
219 fp->twd = 0xffffffffu;
220 fp->fos = 0xffff0000u;
221 }
222}
223EXPORT_SYMBOL_GPL(fpu_finit);
224
225/*
226 * The _current_ task is using the FPU for the first time
227 * so initialize it and set the mxcsr to its default
228 * value at reset if we support XMM instructions and then
229 * remember the current task has used the FPU.
230 */
231int init_fpu(struct task_struct *tsk)
232{
233 int ret;
234
235 if (tsk_used_math(tsk)) {
236 if (HAVE_HWFP && tsk == current)
237 unlazy_fpu(tsk);
238 tsk->thread.fpu.last_cpu = ~0;
239 return 0;
240 }
241
242 /*
243 * Memory allocation at the first usage of the FPU and other state.
244 */
245 ret = fpu_alloc(&tsk->thread.fpu);
246 if (ret)
247 return ret;
248
249 fpu_finit(&tsk->thread.fpu);
250
251 set_stopped_child_used_math(tsk);
252 return 0;
253}
254EXPORT_SYMBOL_GPL(init_fpu);
255
256/*
257 * The xstateregs_active() routine is the same as the fpregs_active() routine,
258 * as the "regset->n" for the xstate regset will be updated based on the feature
259 * capabilites supported by the xsave.
260 */
261int fpregs_active(struct task_struct *target, const struct user_regset *regset)
262{
263 return tsk_used_math(target) ? regset->n : 0;
264}
265
266int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
267{
268 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
269}
270
271int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
272 unsigned int pos, unsigned int count,
273 void *kbuf, void __user *ubuf)
274{
275 int ret;
276
277 if (!cpu_has_fxsr)
278 return -ENODEV;
279
280 ret = init_fpu(target);
281 if (ret)
282 return ret;
283
284 sanitize_i387_state(target);
285
286 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
287 &target->thread.fpu.state->fxsave, 0, -1);
288}
289
290int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
291 unsigned int pos, unsigned int count,
292 const void *kbuf, const void __user *ubuf)
293{
294 int ret;
295
296 if (!cpu_has_fxsr)
297 return -ENODEV;
298
299 ret = init_fpu(target);
300 if (ret)
301 return ret;
302
303 sanitize_i387_state(target);
304
305 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
306 &target->thread.fpu.state->fxsave, 0, -1);
307
308 /*
309 * mxcsr reserved bits must be masked to zero for security reasons.
310 */
311 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
312
313 /*
314 * update the header bits in the xsave header, indicating the
315 * presence of FP and SSE state.
316 */
317 if (cpu_has_xsave)
318 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
319
320 return ret;
321}
322
323int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
324 unsigned int pos, unsigned int count,
325 void *kbuf, void __user *ubuf)
326{
327 int ret;
328
329 if (!cpu_has_xsave)
330 return -ENODEV;
331
332 ret = init_fpu(target);
333 if (ret)
334 return ret;
335
336 /*
337 * Copy the 48bytes defined by the software first into the xstate
338 * memory layout in the thread struct, so that we can copy the entire
339 * xstateregs to the user using one user_regset_copyout().
340 */
341 memcpy(&target->thread.fpu.state->fxsave.sw_reserved,
342 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
343
344 /*
345 * Copy the xstate memory layout.
346 */
347 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf,
348 &target->thread.fpu.state->xsave, 0, -1);
349 return ret;
350}
351
352int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
353 unsigned int pos, unsigned int count,
354 const void *kbuf, const void __user *ubuf)
355{
356 int ret;
357 struct xsave_hdr_struct *xsave_hdr;
358
359 if (!cpu_has_xsave)
360 return -ENODEV;
361
362 ret = init_fpu(target);
363 if (ret)
364 return ret;
365
366 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
367 &target->thread.fpu.state->xsave, 0, -1);
368
369 /*
370 * mxcsr reserved bits must be masked to zero for security reasons.
371 */
372 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
373
374 xsave_hdr = &target->thread.fpu.state->xsave.xsave_hdr;
375
376 xsave_hdr->xstate_bv &= pcntxt_mask;
377 /*
378 * These bits must be zero.
379 */
380 xsave_hdr->reserved1[0] = xsave_hdr->reserved1[1] = 0;
381
382 return ret;
383}
384
385#if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
386
387/*
388 * FPU tag word conversions.
389 */
390
391static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
392{
393 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
394
395 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
396 tmp = ~twd;
397 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
398 /* and move the valid bits to the lower byte. */
399 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
400 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
401 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
402
403 return tmp;
404}
405
406#define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
407#define FP_EXP_TAG_VALID 0
408#define FP_EXP_TAG_ZERO 1
409#define FP_EXP_TAG_SPECIAL 2
410#define FP_EXP_TAG_EMPTY 3
411
412static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
413{
414 struct _fpxreg *st;
415 u32 tos = (fxsave->swd >> 11) & 7;
416 u32 twd = (unsigned long) fxsave->twd;
417 u32 tag;
418 u32 ret = 0xffff0000u;
419 int i;
420
421 for (i = 0; i < 8; i++, twd >>= 1) {
422 if (twd & 0x1) {
423 st = FPREG_ADDR(fxsave, (i - tos) & 7);
424
425 switch (st->exponent & 0x7fff) {
426 case 0x7fff:
427 tag = FP_EXP_TAG_SPECIAL;
428 break;
429 case 0x0000:
430 if (!st->significand[0] &&
431 !st->significand[1] &&
432 !st->significand[2] &&
433 !st->significand[3])
434 tag = FP_EXP_TAG_ZERO;
435 else
436 tag = FP_EXP_TAG_SPECIAL;
437 break;
438 default:
439 if (st->significand[3] & 0x8000)
440 tag = FP_EXP_TAG_VALID;
441 else
442 tag = FP_EXP_TAG_SPECIAL;
443 break;
444 }
445 } else {
446 tag = FP_EXP_TAG_EMPTY;
447 }
448 ret |= tag << (2 * i);
449 }
450 return ret;
451}
452
453/*
454 * FXSR floating point environment conversions.
455 */
456
457static void
458convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
459{
460 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
461 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
462 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
463 int i;
464
465 env->cwd = fxsave->cwd | 0xffff0000u;
466 env->swd = fxsave->swd | 0xffff0000u;
467 env->twd = twd_fxsr_to_i387(fxsave);
468
469#ifdef CONFIG_X86_64
470 env->fip = fxsave->rip;
471 env->foo = fxsave->rdp;
472 /*
473 * should be actually ds/cs at fpu exception time, but
474 * that information is not available in 64bit mode.
475 */
476 env->fcs = task_pt_regs(tsk)->cs;
477 if (tsk == current) {
478 savesegment(ds, env->fos);
479 } else {
480 env->fos = tsk->thread.ds;
481 }
482 env->fos |= 0xffff0000;
483#else
484 env->fip = fxsave->fip;
485 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
486 env->foo = fxsave->foo;
487 env->fos = fxsave->fos;
488#endif
489
490 for (i = 0; i < 8; ++i)
491 memcpy(&to[i], &from[i], sizeof(to[0]));
492}
493
494static void convert_to_fxsr(struct task_struct *tsk,
495 const struct user_i387_ia32_struct *env)
496
497{
498 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
499 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
500 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
501 int i;
502
503 fxsave->cwd = env->cwd;
504 fxsave->swd = env->swd;
505 fxsave->twd = twd_i387_to_fxsr(env->twd);
506 fxsave->fop = (u16) ((u32) env->fcs >> 16);
507#ifdef CONFIG_X86_64
508 fxsave->rip = env->fip;
509 fxsave->rdp = env->foo;
510 /* cs and ds ignored */
511#else
512 fxsave->fip = env->fip;
513 fxsave->fcs = (env->fcs & 0xffff);
514 fxsave->foo = env->foo;
515 fxsave->fos = env->fos;
516#endif
517
518 for (i = 0; i < 8; ++i)
519 memcpy(&to[i], &from[i], sizeof(from[0]));
520}
521
522int fpregs_get(struct task_struct *target, const struct user_regset *regset,
523 unsigned int pos, unsigned int count,
524 void *kbuf, void __user *ubuf)
525{
526 struct user_i387_ia32_struct env;
527 int ret;
528
529 ret = init_fpu(target);
530 if (ret)
531 return ret;
532
533 if (!HAVE_HWFP)
534 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
535
536 if (!cpu_has_fxsr) {
537 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
538 &target->thread.fpu.state->fsave, 0,
539 -1);
540 }
541
542 sanitize_i387_state(target);
543
544 if (kbuf && pos == 0 && count == sizeof(env)) {
545 convert_from_fxsr(kbuf, target);
546 return 0;
547 }
548
549 convert_from_fxsr(&env, target);
550
551 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
552}
553
554int fpregs_set(struct task_struct *target, const struct user_regset *regset,
555 unsigned int pos, unsigned int count,
556 const void *kbuf, const void __user *ubuf)
557{
558 struct user_i387_ia32_struct env;
559 int ret;
560
561 ret = init_fpu(target);
562 if (ret)
563 return ret;
564
565 sanitize_i387_state(target);
566
567 if (!HAVE_HWFP)
568 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
569
570 if (!cpu_has_fxsr) {
571 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
572 &target->thread.fpu.state->fsave, 0, -1);
573 }
574
575 if (pos > 0 || count < sizeof(env))
576 convert_from_fxsr(&env, target);
577
578 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
579 if (!ret)
580 convert_to_fxsr(target, &env);
581
582 /*
583 * update the header bit in the xsave header, indicating the
584 * presence of FP.
585 */
586 if (cpu_has_xsave)
587 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
588 return ret;
589}
590
591/*
592 * Signal frame handlers.
593 */
594
595static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
596{
597 struct task_struct *tsk = current;
598 struct i387_fsave_struct *fp = &tsk->thread.fpu.state->fsave;
599
600 fp->status = fp->swd;
601 if (__copy_to_user(buf, fp, sizeof(struct i387_fsave_struct)))
602 return -1;
603 return 1;
604}
605
606static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
607{
608 struct task_struct *tsk = current;
609 struct i387_fxsave_struct *fx = &tsk->thread.fpu.state->fxsave;
610 struct user_i387_ia32_struct env;
611 int err = 0;
612
613 convert_from_fxsr(&env, tsk);
614 if (__copy_to_user(buf, &env, sizeof(env)))
615 return -1;
616
617 err |= __put_user(fx->swd, &buf->status);
618 err |= __put_user(X86_FXSR_MAGIC, &buf->magic);
619 if (err)
620 return -1;
621
622 if (__copy_to_user(&buf->_fxsr_env[0], fx, xstate_size))
623 return -1;
624 return 1;
625}
626
627static int save_i387_xsave(void __user *buf)
628{
629 struct task_struct *tsk = current;
630 struct _fpstate_ia32 __user *fx = buf;
631 int err = 0;
632
633
634 sanitize_i387_state(tsk);
635
636 /*
637 * For legacy compatible, we always set FP/SSE bits in the bit
638 * vector while saving the state to the user context.
639 * This will enable us capturing any changes(during sigreturn) to
640 * the FP/SSE bits by the legacy applications which don't touch
641 * xstate_bv in the xsave header.
642 *
643 * xsave aware applications can change the xstate_bv in the xsave
644 * header as well as change any contents in the memory layout.
645 * xrestore as part of sigreturn will capture all the changes.
646 */
647 tsk->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
648
649 if (save_i387_fxsave(fx) < 0)
650 return -1;
651
652 err = __copy_to_user(&fx->sw_reserved, &fx_sw_reserved_ia32,
653 sizeof(struct _fpx_sw_bytes));
654 err |= __put_user(FP_XSTATE_MAGIC2,
655 (__u32 __user *) (buf + sig_xstate_ia32_size
656 - FP_XSTATE_MAGIC2_SIZE));
657 if (err)
658 return -1;
659
660 return 1;
661}
662
663int save_i387_xstate_ia32(void __user *buf)
664{
665 struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
666 struct task_struct *tsk = current;
667
668 if (!used_math())
669 return 0;
670
671 if (!access_ok(VERIFY_WRITE, buf, sig_xstate_ia32_size))
672 return -EACCES;
673 /*
674 * This will cause a "finit" to be triggered by the next
675 * attempted FPU operation by the 'current' process.
676 */
677 clear_used_math();
678
679 if (!HAVE_HWFP) {
680 return fpregs_soft_get(current, NULL,
681 0, sizeof(struct user_i387_ia32_struct),
682 NULL, fp) ? -1 : 1;
683 }
684
685 unlazy_fpu(tsk);
686
687 if (cpu_has_xsave)
688 return save_i387_xsave(fp);
689 if (cpu_has_fxsr)
690 return save_i387_fxsave(fp);
691 else
692 return save_i387_fsave(fp);
693}
694
695static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
696{
697 struct task_struct *tsk = current;
698
699 return __copy_from_user(&tsk->thread.fpu.state->fsave, buf,
700 sizeof(struct i387_fsave_struct));
701}
702
703static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf,
704 unsigned int size)
705{
706 struct task_struct *tsk = current;
707 struct user_i387_ia32_struct env;
708 int err;
709
710 err = __copy_from_user(&tsk->thread.fpu.state->fxsave, &buf->_fxsr_env[0],
711 size);
712 /* mxcsr reserved bits must be masked to zero for security reasons */
713 tsk->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
714 if (err || __copy_from_user(&env, buf, sizeof(env)))
715 return 1;
716 convert_to_fxsr(tsk, &env);
717
718 return 0;
719}
720
721static int restore_i387_xsave(void __user *buf)
722{
723 struct _fpx_sw_bytes fx_sw_user;
724 struct _fpstate_ia32 __user *fx_user =
725 ((struct _fpstate_ia32 __user *) buf);
726 struct i387_fxsave_struct __user *fx =
727 (struct i387_fxsave_struct __user *) &fx_user->_fxsr_env[0];
728 struct xsave_hdr_struct *xsave_hdr =
729 ¤t->thread.fpu.state->xsave.xsave_hdr;
730 u64 mask;
731 int err;
732
733 if (check_for_xstate(fx, buf, &fx_sw_user))
734 goto fx_only;
735
736 mask = fx_sw_user.xstate_bv;
737
738 err = restore_i387_fxsave(buf, fx_sw_user.xstate_size);
739
740 xsave_hdr->xstate_bv &= pcntxt_mask;
741 /*
742 * These bits must be zero.
743 */
744 xsave_hdr->reserved1[0] = xsave_hdr->reserved1[1] = 0;
745
746 /*
747 * Init the state that is not present in the memory layout
748 * and enabled by the OS.
749 */
750 mask = ~(pcntxt_mask & ~mask);
751 xsave_hdr->xstate_bv &= mask;
752
753 return err;
754fx_only:
755 /*
756 * Couldn't find the extended state information in the memory
757 * layout. Restore the FP/SSE and init the other extended state
758 * enabled by the OS.
759 */
760 xsave_hdr->xstate_bv = XSTATE_FPSSE;
761 return restore_i387_fxsave(buf, sizeof(struct i387_fxsave_struct));
762}
763
764int restore_i387_xstate_ia32(void __user *buf)
765{
766 int err;
767 struct task_struct *tsk = current;
768 struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
769
770 if (HAVE_HWFP)
771 clear_fpu(tsk);
772
773 if (!buf) {
774 if (used_math()) {
775 clear_fpu(tsk);
776 clear_used_math();
777 }
778
779 return 0;
780 } else
781 if (!access_ok(VERIFY_READ, buf, sig_xstate_ia32_size))
782 return -EACCES;
783
784 if (!used_math()) {
785 err = init_fpu(tsk);
786 if (err)
787 return err;
788 }
789
790 if (HAVE_HWFP) {
791 if (cpu_has_xsave)
792 err = restore_i387_xsave(buf);
793 else if (cpu_has_fxsr)
794 err = restore_i387_fxsave(fp, sizeof(struct
795 i387_fxsave_struct));
796 else
797 err = restore_i387_fsave(fp);
798 } else {
799 err = fpregs_soft_set(current, NULL,
800 0, sizeof(struct user_i387_ia32_struct),
801 NULL, fp) != 0;
802 }
803 set_used_math();
804
805 return err;
806}
807
808/*
809 * FPU state for core dumps.
810 * This is only used for a.out dumps now.
811 * It is declared generically using elf_fpregset_t (which is
812 * struct user_i387_struct) but is in fact only used for 32-bit
813 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
814 */
815int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
816{
817 struct task_struct *tsk = current;
818 int fpvalid;
819
820 fpvalid = !!used_math();
821 if (fpvalid)
822 fpvalid = !fpregs_get(tsk, NULL,
823 0, sizeof(struct user_i387_ia32_struct),
824 fpu, NULL);
825
826 return fpvalid;
827}
828EXPORT_SYMBOL(dump_fpu);
829
830#endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */
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 <linux/module.h>
9#include <linux/regset.h>
10#include <linux/sched.h>
11#include <linux/slab.h>
12
13#include <asm/sigcontext.h>
14#include <asm/processor.h>
15#include <asm/math_emu.h>
16#include <asm/uaccess.h>
17#include <asm/ptrace.h>
18#include <asm/i387.h>
19#include <asm/fpu-internal.h>
20#include <asm/user.h>
21
22/*
23 * Were we in an interrupt that interrupted kernel mode?
24 *
25 * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
26 * pair does nothing at all: the thread must not have fpu (so
27 * that we don't try to save the FPU state), and TS must
28 * be set (so that the clts/stts pair does nothing that is
29 * visible in the interrupted kernel thread).
30 *
31 * Except for the eagerfpu case when we return 1 unless we've already
32 * been eager and saved the state in kernel_fpu_begin().
33 */
34static inline bool interrupted_kernel_fpu_idle(void)
35{
36 if (use_eager_fpu())
37 return __thread_has_fpu(current);
38
39 return !__thread_has_fpu(current) &&
40 (read_cr0() & X86_CR0_TS);
41}
42
43/*
44 * Were we in user mode (or vm86 mode) when we were
45 * interrupted?
46 *
47 * Doing kernel_fpu_begin/end() is ok if we are running
48 * in an interrupt context from user mode - we'll just
49 * save the FPU state as required.
50 */
51static inline bool interrupted_user_mode(void)
52{
53 struct pt_regs *regs = get_irq_regs();
54 return regs && user_mode_vm(regs);
55}
56
57/*
58 * Can we use the FPU in kernel mode with the
59 * whole "kernel_fpu_begin/end()" sequence?
60 *
61 * It's always ok in process context (ie "not interrupt")
62 * but it is sometimes ok even from an irq.
63 */
64bool irq_fpu_usable(void)
65{
66 return !in_interrupt() ||
67 interrupted_user_mode() ||
68 interrupted_kernel_fpu_idle();
69}
70EXPORT_SYMBOL(irq_fpu_usable);
71
72void __kernel_fpu_begin(void)
73{
74 struct task_struct *me = current;
75
76 if (__thread_has_fpu(me)) {
77 __thread_clear_has_fpu(me);
78 __save_init_fpu(me);
79 /* We do 'stts()' in __kernel_fpu_end() */
80 } else if (!use_eager_fpu()) {
81 this_cpu_write(fpu_owner_task, NULL);
82 clts();
83 }
84}
85EXPORT_SYMBOL(__kernel_fpu_begin);
86
87void __kernel_fpu_end(void)
88{
89 if (use_eager_fpu()) {
90 /*
91 * For eager fpu, most the time, tsk_used_math() is true.
92 * Restore the user math as we are done with the kernel usage.
93 * At few instances during thread exit, signal handling etc,
94 * tsk_used_math() is false. Those few places will take proper
95 * actions, so we don't need to restore the math here.
96 */
97 if (likely(tsk_used_math(current)))
98 math_state_restore();
99 } else {
100 stts();
101 }
102}
103EXPORT_SYMBOL(__kernel_fpu_end);
104
105void unlazy_fpu(struct task_struct *tsk)
106{
107 preempt_disable();
108 if (__thread_has_fpu(tsk)) {
109 __save_init_fpu(tsk);
110 __thread_fpu_end(tsk);
111 } else
112 tsk->thread.fpu_counter = 0;
113 preempt_enable();
114}
115EXPORT_SYMBOL(unlazy_fpu);
116
117unsigned int mxcsr_feature_mask __read_mostly = 0xffffffffu;
118unsigned int xstate_size;
119EXPORT_SYMBOL_GPL(xstate_size);
120static struct i387_fxsave_struct fx_scratch;
121
122static void mxcsr_feature_mask_init(void)
123{
124 unsigned long mask = 0;
125
126 if (cpu_has_fxsr) {
127 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
128 asm volatile("fxsave %0" : "+m" (fx_scratch));
129 mask = fx_scratch.mxcsr_mask;
130 if (mask == 0)
131 mask = 0x0000ffbf;
132 }
133 mxcsr_feature_mask &= mask;
134}
135
136static void init_thread_xstate(void)
137{
138 /*
139 * Note that xstate_size might be overwriten later during
140 * xsave_init().
141 */
142
143 if (!cpu_has_fpu) {
144 /*
145 * Disable xsave as we do not support it if i387
146 * emulation is enabled.
147 */
148 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
149 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
150 xstate_size = sizeof(struct i387_soft_struct);
151 return;
152 }
153
154 if (cpu_has_fxsr)
155 xstate_size = sizeof(struct i387_fxsave_struct);
156 else
157 xstate_size = sizeof(struct i387_fsave_struct);
158}
159
160/*
161 * Called at bootup to set up the initial FPU state that is later cloned
162 * into all processes.
163 */
164
165void fpu_init(void)
166{
167 unsigned long cr0;
168 unsigned long cr4_mask = 0;
169
170#ifndef CONFIG_MATH_EMULATION
171 if (!cpu_has_fpu) {
172 pr_emerg("No FPU found and no math emulation present\n");
173 pr_emerg("Giving up\n");
174 for (;;)
175 asm volatile("hlt");
176 }
177#endif
178 if (cpu_has_fxsr)
179 cr4_mask |= X86_CR4_OSFXSR;
180 if (cpu_has_xmm)
181 cr4_mask |= X86_CR4_OSXMMEXCPT;
182 if (cr4_mask)
183 set_in_cr4(cr4_mask);
184
185 cr0 = read_cr0();
186 cr0 &= ~(X86_CR0_TS|X86_CR0_EM); /* clear TS and EM */
187 if (!cpu_has_fpu)
188 cr0 |= X86_CR0_EM;
189 write_cr0(cr0);
190
191 /*
192 * init_thread_xstate is only called once to avoid overriding
193 * xstate_size during boot time or during CPU hotplug.
194 */
195 if (xstate_size == 0)
196 init_thread_xstate();
197
198 mxcsr_feature_mask_init();
199 xsave_init();
200 eager_fpu_init();
201}
202
203void fpu_finit(struct fpu *fpu)
204{
205 if (!cpu_has_fpu) {
206 finit_soft_fpu(&fpu->state->soft);
207 return;
208 }
209
210 if (cpu_has_fxsr) {
211 fx_finit(&fpu->state->fxsave);
212 } else {
213 struct i387_fsave_struct *fp = &fpu->state->fsave;
214 memset(fp, 0, xstate_size);
215 fp->cwd = 0xffff037fu;
216 fp->swd = 0xffff0000u;
217 fp->twd = 0xffffffffu;
218 fp->fos = 0xffff0000u;
219 }
220}
221EXPORT_SYMBOL_GPL(fpu_finit);
222
223/*
224 * The _current_ task is using the FPU for the first time
225 * so initialize it and set the mxcsr to its default
226 * value at reset if we support XMM instructions and then
227 * remember the current task has used the FPU.
228 */
229int init_fpu(struct task_struct *tsk)
230{
231 int ret;
232
233 if (tsk_used_math(tsk)) {
234 if (cpu_has_fpu && tsk == current)
235 unlazy_fpu(tsk);
236 tsk->thread.fpu.last_cpu = ~0;
237 return 0;
238 }
239
240 /*
241 * Memory allocation at the first usage of the FPU and other state.
242 */
243 ret = fpu_alloc(&tsk->thread.fpu);
244 if (ret)
245 return ret;
246
247 fpu_finit(&tsk->thread.fpu);
248
249 set_stopped_child_used_math(tsk);
250 return 0;
251}
252EXPORT_SYMBOL_GPL(init_fpu);
253
254/*
255 * The xstateregs_active() routine is the same as the fpregs_active() routine,
256 * as the "regset->n" for the xstate regset will be updated based on the feature
257 * capabilites supported by the xsave.
258 */
259int fpregs_active(struct task_struct *target, const struct user_regset *regset)
260{
261 return tsk_used_math(target) ? regset->n : 0;
262}
263
264int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
265{
266 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
267}
268
269int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
270 unsigned int pos, unsigned int count,
271 void *kbuf, void __user *ubuf)
272{
273 int ret;
274
275 if (!cpu_has_fxsr)
276 return -ENODEV;
277
278 ret = init_fpu(target);
279 if (ret)
280 return ret;
281
282 sanitize_i387_state(target);
283
284 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
285 &target->thread.fpu.state->fxsave, 0, -1);
286}
287
288int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
289 unsigned int pos, unsigned int count,
290 const void *kbuf, const void __user *ubuf)
291{
292 int ret;
293
294 if (!cpu_has_fxsr)
295 return -ENODEV;
296
297 ret = init_fpu(target);
298 if (ret)
299 return ret;
300
301 sanitize_i387_state(target);
302
303 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
304 &target->thread.fpu.state->fxsave, 0, -1);
305
306 /*
307 * mxcsr reserved bits must be masked to zero for security reasons.
308 */
309 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
310
311 /*
312 * update the header bits in the xsave header, indicating the
313 * presence of FP and SSE state.
314 */
315 if (cpu_has_xsave)
316 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
317
318 return ret;
319}
320
321int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
322 unsigned int pos, unsigned int count,
323 void *kbuf, void __user *ubuf)
324{
325 int ret;
326
327 if (!cpu_has_xsave)
328 return -ENODEV;
329
330 ret = init_fpu(target);
331 if (ret)
332 return ret;
333
334 /*
335 * Copy the 48bytes defined by the software first into the xstate
336 * memory layout in the thread struct, so that we can copy the entire
337 * xstateregs to the user using one user_regset_copyout().
338 */
339 memcpy(&target->thread.fpu.state->fxsave.sw_reserved,
340 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
341
342 /*
343 * Copy the xstate memory layout.
344 */
345 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf,
346 &target->thread.fpu.state->xsave, 0, -1);
347 return ret;
348}
349
350int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
351 unsigned int pos, unsigned int count,
352 const void *kbuf, const void __user *ubuf)
353{
354 int ret;
355 struct xsave_hdr_struct *xsave_hdr;
356
357 if (!cpu_has_xsave)
358 return -ENODEV;
359
360 ret = init_fpu(target);
361 if (ret)
362 return ret;
363
364 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
365 &target->thread.fpu.state->xsave, 0, -1);
366
367 /*
368 * mxcsr reserved bits must be masked to zero for security reasons.
369 */
370 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
371
372 xsave_hdr = &target->thread.fpu.state->xsave.xsave_hdr;
373
374 xsave_hdr->xstate_bv &= pcntxt_mask;
375 /*
376 * These bits must be zero.
377 */
378 xsave_hdr->reserved1[0] = xsave_hdr->reserved1[1] = 0;
379
380 return ret;
381}
382
383#if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
384
385/*
386 * FPU tag word conversions.
387 */
388
389static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
390{
391 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
392
393 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
394 tmp = ~twd;
395 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
396 /* and move the valid bits to the lower byte. */
397 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
398 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
399 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
400
401 return tmp;
402}
403
404#define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
405#define FP_EXP_TAG_VALID 0
406#define FP_EXP_TAG_ZERO 1
407#define FP_EXP_TAG_SPECIAL 2
408#define FP_EXP_TAG_EMPTY 3
409
410static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
411{
412 struct _fpxreg *st;
413 u32 tos = (fxsave->swd >> 11) & 7;
414 u32 twd = (unsigned long) fxsave->twd;
415 u32 tag;
416 u32 ret = 0xffff0000u;
417 int i;
418
419 for (i = 0; i < 8; i++, twd >>= 1) {
420 if (twd & 0x1) {
421 st = FPREG_ADDR(fxsave, (i - tos) & 7);
422
423 switch (st->exponent & 0x7fff) {
424 case 0x7fff:
425 tag = FP_EXP_TAG_SPECIAL;
426 break;
427 case 0x0000:
428 if (!st->significand[0] &&
429 !st->significand[1] &&
430 !st->significand[2] &&
431 !st->significand[3])
432 tag = FP_EXP_TAG_ZERO;
433 else
434 tag = FP_EXP_TAG_SPECIAL;
435 break;
436 default:
437 if (st->significand[3] & 0x8000)
438 tag = FP_EXP_TAG_VALID;
439 else
440 tag = FP_EXP_TAG_SPECIAL;
441 break;
442 }
443 } else {
444 tag = FP_EXP_TAG_EMPTY;
445 }
446 ret |= tag << (2 * i);
447 }
448 return ret;
449}
450
451/*
452 * FXSR floating point environment conversions.
453 */
454
455void
456convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
457{
458 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
459 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
460 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
461 int i;
462
463 env->cwd = fxsave->cwd | 0xffff0000u;
464 env->swd = fxsave->swd | 0xffff0000u;
465 env->twd = twd_fxsr_to_i387(fxsave);
466
467#ifdef CONFIG_X86_64
468 env->fip = fxsave->rip;
469 env->foo = fxsave->rdp;
470 /*
471 * should be actually ds/cs at fpu exception time, but
472 * that information is not available in 64bit mode.
473 */
474 env->fcs = task_pt_regs(tsk)->cs;
475 if (tsk == current) {
476 savesegment(ds, env->fos);
477 } else {
478 env->fos = tsk->thread.ds;
479 }
480 env->fos |= 0xffff0000;
481#else
482 env->fip = fxsave->fip;
483 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
484 env->foo = fxsave->foo;
485 env->fos = fxsave->fos;
486#endif
487
488 for (i = 0; i < 8; ++i)
489 memcpy(&to[i], &from[i], sizeof(to[0]));
490}
491
492void convert_to_fxsr(struct task_struct *tsk,
493 const struct user_i387_ia32_struct *env)
494
495{
496 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
497 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
498 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
499 int i;
500
501 fxsave->cwd = env->cwd;
502 fxsave->swd = env->swd;
503 fxsave->twd = twd_i387_to_fxsr(env->twd);
504 fxsave->fop = (u16) ((u32) env->fcs >> 16);
505#ifdef CONFIG_X86_64
506 fxsave->rip = env->fip;
507 fxsave->rdp = env->foo;
508 /* cs and ds ignored */
509#else
510 fxsave->fip = env->fip;
511 fxsave->fcs = (env->fcs & 0xffff);
512 fxsave->foo = env->foo;
513 fxsave->fos = env->fos;
514#endif
515
516 for (i = 0; i < 8; ++i)
517 memcpy(&to[i], &from[i], sizeof(from[0]));
518}
519
520int fpregs_get(struct task_struct *target, const struct user_regset *regset,
521 unsigned int pos, unsigned int count,
522 void *kbuf, void __user *ubuf)
523{
524 struct user_i387_ia32_struct env;
525 int ret;
526
527 ret = init_fpu(target);
528 if (ret)
529 return ret;
530
531 if (!static_cpu_has(X86_FEATURE_FPU))
532 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
533
534 if (!cpu_has_fxsr)
535 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
536 &target->thread.fpu.state->fsave, 0,
537 -1);
538
539 sanitize_i387_state(target);
540
541 if (kbuf && pos == 0 && count == sizeof(env)) {
542 convert_from_fxsr(kbuf, target);
543 return 0;
544 }
545
546 convert_from_fxsr(&env, target);
547
548 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
549}
550
551int fpregs_set(struct task_struct *target, const struct user_regset *regset,
552 unsigned int pos, unsigned int count,
553 const void *kbuf, const void __user *ubuf)
554{
555 struct user_i387_ia32_struct env;
556 int ret;
557
558 ret = init_fpu(target);
559 if (ret)
560 return ret;
561
562 sanitize_i387_state(target);
563
564 if (!static_cpu_has(X86_FEATURE_FPU))
565 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
566
567 if (!cpu_has_fxsr)
568 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
569 &target->thread.fpu.state->fsave, 0,
570 -1);
571
572 if (pos > 0 || count < sizeof(env))
573 convert_from_fxsr(&env, target);
574
575 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
576 if (!ret)
577 convert_to_fxsr(target, &env);
578
579 /*
580 * update the header bit in the xsave header, indicating the
581 * presence of FP.
582 */
583 if (cpu_has_xsave)
584 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
585 return ret;
586}
587
588/*
589 * FPU state for core dumps.
590 * This is only used for a.out dumps now.
591 * It is declared generically using elf_fpregset_t (which is
592 * struct user_i387_struct) but is in fact only used for 32-bit
593 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
594 */
595int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
596{
597 struct task_struct *tsk = current;
598 int fpvalid;
599
600 fpvalid = !!used_math();
601 if (fpvalid)
602 fpvalid = !fpregs_get(tsk, NULL,
603 0, sizeof(struct user_i387_ia32_struct),
604 fpu, NULL);
605
606 return fpvalid;
607}
608EXPORT_SYMBOL(dump_fpu);
609
610#endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */
611
612static int __init no_387(char *s)
613{
614 setup_clear_cpu_cap(X86_FEATURE_FPU);
615 return 1;
616}
617
618__setup("no387", no_387);
619
620void fpu_detect(struct cpuinfo_x86 *c)
621{
622 unsigned long cr0;
623 u16 fsw, fcw;
624
625 fsw = fcw = 0xffff;
626
627 cr0 = read_cr0();
628 cr0 &= ~(X86_CR0_TS | X86_CR0_EM);
629 write_cr0(cr0);
630
631 asm volatile("fninit ; fnstsw %0 ; fnstcw %1"
632 : "+m" (fsw), "+m" (fcw));
633
634 if (fsw == 0 && (fcw & 0x103f) == 0x003f)
635 set_cpu_cap(c, X86_FEATURE_FPU);
636 else
637 clear_cpu_cap(c, X86_FEATURE_FPU);
638
639 /* The final cr0 value is set in fpu_init() */
640}