x86/fpu: Disable XSAVES* support for now
[deliverable/linux.git] / arch / x86 / kernel / i387.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 <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/tlbflush.h>
17 #include <asm/uaccess.h>
18 #include <asm/ptrace.h>
19 #include <asm/i387.h>
20 #include <asm/fpu-internal.h>
21 #include <asm/user.h>
22
23 static DEFINE_PER_CPU(bool, in_kernel_fpu);
24
25 void kernel_fpu_disable(void)
26 {
27 WARN_ON(this_cpu_read(in_kernel_fpu));
28 this_cpu_write(in_kernel_fpu, true);
29 }
30
31 void kernel_fpu_enable(void)
32 {
33 this_cpu_write(in_kernel_fpu, false);
34 }
35
36 /*
37 * Were we in an interrupt that interrupted kernel mode?
38 *
39 * On others, 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 *
45 * Except for the eagerfpu case when we return true; in the likely case
46 * the thread has FPU but we are not going to set/clear TS.
47 */
48 static inline bool interrupted_kernel_fpu_idle(void)
49 {
50 if (this_cpu_read(in_kernel_fpu))
51 return false;
52
53 if (use_eager_fpu())
54 return true;
55
56 return !__thread_has_fpu(current) &&
57 (read_cr0() & X86_CR0_TS);
58 }
59
60 /*
61 * Were we in user mode (or vm86 mode) when we were
62 * interrupted?
63 *
64 * Doing kernel_fpu_begin/end() is ok if we are running
65 * in an interrupt context from user mode - we'll just
66 * save the FPU state as required.
67 */
68 static inline bool interrupted_user_mode(void)
69 {
70 struct pt_regs *regs = get_irq_regs();
71 return regs && user_mode(regs);
72 }
73
74 /*
75 * Can we use the FPU in kernel mode with the
76 * whole "kernel_fpu_begin/end()" sequence?
77 *
78 * It's always ok in process context (ie "not interrupt")
79 * but it is sometimes ok even from an irq.
80 */
81 bool irq_fpu_usable(void)
82 {
83 return !in_interrupt() ||
84 interrupted_user_mode() ||
85 interrupted_kernel_fpu_idle();
86 }
87 EXPORT_SYMBOL(irq_fpu_usable);
88
89 void __kernel_fpu_begin(void)
90 {
91 struct task_struct *me = current;
92
93 this_cpu_write(in_kernel_fpu, true);
94
95 if (__thread_has_fpu(me)) {
96 __save_init_fpu(me);
97 } else {
98 this_cpu_write(fpu_owner_task, NULL);
99 if (!use_eager_fpu())
100 clts();
101 }
102 }
103 EXPORT_SYMBOL(__kernel_fpu_begin);
104
105 void __kernel_fpu_end(void)
106 {
107 struct task_struct *me = current;
108
109 if (__thread_has_fpu(me)) {
110 if (WARN_ON(restore_fpu_checking(me)))
111 fpu_reset_state(me);
112 } else if (!use_eager_fpu()) {
113 stts();
114 }
115
116 this_cpu_write(in_kernel_fpu, false);
117 }
118 EXPORT_SYMBOL(__kernel_fpu_end);
119
120 void unlazy_fpu(struct task_struct *tsk)
121 {
122 preempt_disable();
123 if (__thread_has_fpu(tsk)) {
124 if (use_eager_fpu()) {
125 __save_fpu(tsk);
126 } else {
127 __save_init_fpu(tsk);
128 __thread_fpu_end(tsk);
129 }
130 }
131 preempt_enable();
132 }
133 EXPORT_SYMBOL(unlazy_fpu);
134
135 unsigned int mxcsr_feature_mask __read_mostly = 0xffffffffu;
136 unsigned int xstate_size;
137 EXPORT_SYMBOL_GPL(xstate_size);
138 static struct i387_fxsave_struct fx_scratch;
139
140 static void mxcsr_feature_mask_init(void)
141 {
142 unsigned long mask = 0;
143
144 if (cpu_has_fxsr) {
145 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
146 asm volatile("fxsave %0" : "+m" (fx_scratch));
147 mask = fx_scratch.mxcsr_mask;
148 if (mask == 0)
149 mask = 0x0000ffbf;
150 }
151 mxcsr_feature_mask &= mask;
152 }
153
154 static void init_thread_xstate(void)
155 {
156 /*
157 * Note that xstate_size might be overwriten later during
158 * xsave_init().
159 */
160
161 if (!cpu_has_fpu) {
162 /*
163 * Disable xsave as we do not support it if i387
164 * emulation is enabled.
165 */
166 setup_clear_cpu_cap(X86_FEATURE_XSAVE);
167 setup_clear_cpu_cap(X86_FEATURE_XSAVEOPT);
168 xstate_size = sizeof(struct i387_soft_struct);
169 return;
170 }
171
172 if (cpu_has_fxsr)
173 xstate_size = sizeof(struct i387_fxsave_struct);
174 else
175 xstate_size = sizeof(struct i387_fsave_struct);
176
177 /*
178 * Quirk: we don't yet handle the XSAVES* instructions
179 * correctly, as we don't correctly convert between
180 * standard and compacted format when interfacing
181 * with user-space - so disable it for now.
182 *
183 * The difference is small: with recent CPUs the
184 * compacted format is only marginally smaller than
185 * the standard FPU state format.
186 *
187 * ( This is easy to backport while we are fixing
188 * XSAVES* support. )
189 */
190 setup_clear_cpu_cap(X86_FEATURE_XSAVES);
191 }
192
193 /*
194 * Called at bootup to set up the initial FPU state that is later cloned
195 * into all processes.
196 */
197
198 void fpu_init(void)
199 {
200 unsigned long cr0;
201 unsigned long cr4_mask = 0;
202
203 #ifndef CONFIG_MATH_EMULATION
204 if (!cpu_has_fpu) {
205 pr_emerg("No FPU found and no math emulation present\n");
206 pr_emerg("Giving up\n");
207 for (;;)
208 asm volatile("hlt");
209 }
210 #endif
211 if (cpu_has_fxsr)
212 cr4_mask |= X86_CR4_OSFXSR;
213 if (cpu_has_xmm)
214 cr4_mask |= X86_CR4_OSXMMEXCPT;
215 if (cr4_mask)
216 cr4_set_bits(cr4_mask);
217
218 cr0 = read_cr0();
219 cr0 &= ~(X86_CR0_TS|X86_CR0_EM); /* clear TS and EM */
220 if (!cpu_has_fpu)
221 cr0 |= X86_CR0_EM;
222 write_cr0(cr0);
223
224 /*
225 * init_thread_xstate is only called once to avoid overriding
226 * xstate_size during boot time or during CPU hotplug.
227 */
228 if (xstate_size == 0)
229 init_thread_xstate();
230
231 mxcsr_feature_mask_init();
232 xsave_init();
233 eager_fpu_init();
234 }
235
236 void fpu_finit(struct fpu *fpu)
237 {
238 if (!cpu_has_fpu) {
239 finit_soft_fpu(&fpu->state->soft);
240 return;
241 }
242
243 memset(fpu->state, 0, xstate_size);
244
245 if (cpu_has_fxsr) {
246 fx_finit(&fpu->state->fxsave);
247 } else {
248 struct i387_fsave_struct *fp = &fpu->state->fsave;
249 fp->cwd = 0xffff037fu;
250 fp->swd = 0xffff0000u;
251 fp->twd = 0xffffffffu;
252 fp->fos = 0xffff0000u;
253 }
254 }
255 EXPORT_SYMBOL_GPL(fpu_finit);
256
257 /*
258 * The _current_ task is using the FPU for the first time
259 * so initialize it and set the mxcsr to its default
260 * value at reset if we support XMM instructions and then
261 * remember the current task has used the FPU.
262 */
263 int init_fpu(struct task_struct *tsk)
264 {
265 int ret;
266
267 if (tsk_used_math(tsk)) {
268 if (cpu_has_fpu && tsk == current)
269 unlazy_fpu(tsk);
270 task_disable_lazy_fpu_restore(tsk);
271 return 0;
272 }
273
274 /*
275 * Memory allocation at the first usage of the FPU and other state.
276 */
277 ret = fpu_alloc(&tsk->thread.fpu);
278 if (ret)
279 return ret;
280
281 fpu_finit(&tsk->thread.fpu);
282
283 set_stopped_child_used_math(tsk);
284 return 0;
285 }
286 EXPORT_SYMBOL_GPL(init_fpu);
287
288 /*
289 * The xstateregs_active() routine is the same as the fpregs_active() routine,
290 * as the "regset->n" for the xstate regset will be updated based on the feature
291 * capabilites supported by the xsave.
292 */
293 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
294 {
295 return tsk_used_math(target) ? regset->n : 0;
296 }
297
298 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
299 {
300 return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
301 }
302
303 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
304 unsigned int pos, unsigned int count,
305 void *kbuf, void __user *ubuf)
306 {
307 int ret;
308
309 if (!cpu_has_fxsr)
310 return -ENODEV;
311
312 ret = init_fpu(target);
313 if (ret)
314 return ret;
315
316 sanitize_i387_state(target);
317
318 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
319 &target->thread.fpu.state->fxsave, 0, -1);
320 }
321
322 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
323 unsigned int pos, unsigned int count,
324 const void *kbuf, const void __user *ubuf)
325 {
326 int ret;
327
328 if (!cpu_has_fxsr)
329 return -ENODEV;
330
331 ret = init_fpu(target);
332 if (ret)
333 return ret;
334
335 sanitize_i387_state(target);
336
337 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
338 &target->thread.fpu.state->fxsave, 0, -1);
339
340 /*
341 * mxcsr reserved bits must be masked to zero for security reasons.
342 */
343 target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
344
345 /*
346 * update the header bits in the xsave header, indicating the
347 * presence of FP and SSE state.
348 */
349 if (cpu_has_xsave)
350 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
351
352 return ret;
353 }
354
355 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
356 unsigned int pos, unsigned int count,
357 void *kbuf, void __user *ubuf)
358 {
359 struct xsave_struct *xsave;
360 int ret;
361
362 if (!cpu_has_xsave)
363 return -ENODEV;
364
365 ret = init_fpu(target);
366 if (ret)
367 return ret;
368
369 xsave = &target->thread.fpu.state->xsave;
370
371 /*
372 * Copy the 48bytes defined by the software first into the xstate
373 * memory layout in the thread struct, so that we can copy the entire
374 * xstateregs to the user using one user_regset_copyout().
375 */
376 memcpy(&xsave->i387.sw_reserved,
377 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
378 /*
379 * Copy the xstate memory layout.
380 */
381 ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
382 return ret;
383 }
384
385 int xstateregs_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 xsave_struct *xsave;
390 int ret;
391
392 if (!cpu_has_xsave)
393 return -ENODEV;
394
395 ret = init_fpu(target);
396 if (ret)
397 return ret;
398
399 xsave = &target->thread.fpu.state->xsave;
400
401 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
402 /*
403 * mxcsr reserved bits must be masked to zero for security reasons.
404 */
405 xsave->i387.mxcsr &= mxcsr_feature_mask;
406 xsave->xsave_hdr.xstate_bv &= pcntxt_mask;
407 /*
408 * These bits must be zero.
409 */
410 memset(&xsave->xsave_hdr.reserved, 0, 48);
411 return ret;
412 }
413
414 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
415
416 /*
417 * FPU tag word conversions.
418 */
419
420 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
421 {
422 unsigned int tmp; /* to avoid 16 bit prefixes in the code */
423
424 /* Transform each pair of bits into 01 (valid) or 00 (empty) */
425 tmp = ~twd;
426 tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
427 /* and move the valid bits to the lower byte. */
428 tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
429 tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
430 tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
431
432 return tmp;
433 }
434
435 #define FPREG_ADDR(f, n) ((void *)&(f)->st_space + (n) * 16)
436 #define FP_EXP_TAG_VALID 0
437 #define FP_EXP_TAG_ZERO 1
438 #define FP_EXP_TAG_SPECIAL 2
439 #define FP_EXP_TAG_EMPTY 3
440
441 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
442 {
443 struct _fpxreg *st;
444 u32 tos = (fxsave->swd >> 11) & 7;
445 u32 twd = (unsigned long) fxsave->twd;
446 u32 tag;
447 u32 ret = 0xffff0000u;
448 int i;
449
450 for (i = 0; i < 8; i++, twd >>= 1) {
451 if (twd & 0x1) {
452 st = FPREG_ADDR(fxsave, (i - tos) & 7);
453
454 switch (st->exponent & 0x7fff) {
455 case 0x7fff:
456 tag = FP_EXP_TAG_SPECIAL;
457 break;
458 case 0x0000:
459 if (!st->significand[0] &&
460 !st->significand[1] &&
461 !st->significand[2] &&
462 !st->significand[3])
463 tag = FP_EXP_TAG_ZERO;
464 else
465 tag = FP_EXP_TAG_SPECIAL;
466 break;
467 default:
468 if (st->significand[3] & 0x8000)
469 tag = FP_EXP_TAG_VALID;
470 else
471 tag = FP_EXP_TAG_SPECIAL;
472 break;
473 }
474 } else {
475 tag = FP_EXP_TAG_EMPTY;
476 }
477 ret |= tag << (2 * i);
478 }
479 return ret;
480 }
481
482 /*
483 * FXSR floating point environment conversions.
484 */
485
486 void
487 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
488 {
489 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
490 struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
491 struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
492 int i;
493
494 env->cwd = fxsave->cwd | 0xffff0000u;
495 env->swd = fxsave->swd | 0xffff0000u;
496 env->twd = twd_fxsr_to_i387(fxsave);
497
498 #ifdef CONFIG_X86_64
499 env->fip = fxsave->rip;
500 env->foo = fxsave->rdp;
501 /*
502 * should be actually ds/cs at fpu exception time, but
503 * that information is not available in 64bit mode.
504 */
505 env->fcs = task_pt_regs(tsk)->cs;
506 if (tsk == current) {
507 savesegment(ds, env->fos);
508 } else {
509 env->fos = tsk->thread.ds;
510 }
511 env->fos |= 0xffff0000;
512 #else
513 env->fip = fxsave->fip;
514 env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
515 env->foo = fxsave->foo;
516 env->fos = fxsave->fos;
517 #endif
518
519 for (i = 0; i < 8; ++i)
520 memcpy(&to[i], &from[i], sizeof(to[0]));
521 }
522
523 void convert_to_fxsr(struct task_struct *tsk,
524 const struct user_i387_ia32_struct *env)
525
526 {
527 struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
528 struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
529 struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
530 int i;
531
532 fxsave->cwd = env->cwd;
533 fxsave->swd = env->swd;
534 fxsave->twd = twd_i387_to_fxsr(env->twd);
535 fxsave->fop = (u16) ((u32) env->fcs >> 16);
536 #ifdef CONFIG_X86_64
537 fxsave->rip = env->fip;
538 fxsave->rdp = env->foo;
539 /* cs and ds ignored */
540 #else
541 fxsave->fip = env->fip;
542 fxsave->fcs = (env->fcs & 0xffff);
543 fxsave->foo = env->foo;
544 fxsave->fos = env->fos;
545 #endif
546
547 for (i = 0; i < 8; ++i)
548 memcpy(&to[i], &from[i], sizeof(from[0]));
549 }
550
551 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
552 unsigned int pos, unsigned int count,
553 void *kbuf, 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 if (!static_cpu_has(X86_FEATURE_FPU))
563 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
564
565 if (!cpu_has_fxsr)
566 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
567 &target->thread.fpu.state->fsave, 0,
568 -1);
569
570 sanitize_i387_state(target);
571
572 if (kbuf && pos == 0 && count == sizeof(env)) {
573 convert_from_fxsr(kbuf, target);
574 return 0;
575 }
576
577 convert_from_fxsr(&env, target);
578
579 return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
580 }
581
582 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
583 unsigned int pos, unsigned int count,
584 const void *kbuf, const void __user *ubuf)
585 {
586 struct user_i387_ia32_struct env;
587 int ret;
588
589 ret = init_fpu(target);
590 if (ret)
591 return ret;
592
593 sanitize_i387_state(target);
594
595 if (!static_cpu_has(X86_FEATURE_FPU))
596 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
597
598 if (!cpu_has_fxsr)
599 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
600 &target->thread.fpu.state->fsave, 0,
601 -1);
602
603 if (pos > 0 || count < sizeof(env))
604 convert_from_fxsr(&env, target);
605
606 ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
607 if (!ret)
608 convert_to_fxsr(target, &env);
609
610 /*
611 * update the header bit in the xsave header, indicating the
612 * presence of FP.
613 */
614 if (cpu_has_xsave)
615 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
616 return ret;
617 }
618
619 /*
620 * FPU state for core dumps.
621 * This is only used for a.out dumps now.
622 * It is declared generically using elf_fpregset_t (which is
623 * struct user_i387_struct) but is in fact only used for 32-bit
624 * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
625 */
626 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
627 {
628 struct task_struct *tsk = current;
629 int fpvalid;
630
631 fpvalid = !!used_math();
632 if (fpvalid)
633 fpvalid = !fpregs_get(tsk, NULL,
634 0, sizeof(struct user_i387_ia32_struct),
635 fpu, NULL);
636
637 return fpvalid;
638 }
639 EXPORT_SYMBOL(dump_fpu);
640
641 #endif /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */
642
643 static int __init no_387(char *s)
644 {
645 setup_clear_cpu_cap(X86_FEATURE_FPU);
646 return 1;
647 }
648
649 __setup("no387", no_387);
650
651 void fpu_detect(struct cpuinfo_x86 *c)
652 {
653 unsigned long cr0;
654 u16 fsw, fcw;
655
656 fsw = fcw = 0xffff;
657
658 cr0 = read_cr0();
659 cr0 &= ~(X86_CR0_TS | X86_CR0_EM);
660 write_cr0(cr0);
661
662 asm volatile("fninit ; fnstsw %0 ; fnstcw %1"
663 : "+m" (fsw), "+m" (fcw));
664
665 if (fsw == 0 && (fcw & 0x103f) == 0x003f)
666 set_cpu_cap(c, X86_FEATURE_FPU);
667 else
668 clear_cpu_cap(c, X86_FEATURE_FPU);
669
670 /* The final cr0 value is set in fpu_init() */
671 }
This page took 0.051632 seconds and 5 git commands to generate.