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
  12 #include <asm/sigcontext.h>
  13 #include <asm/processor.h>
  14 #include <asm/math_emu.h>
  15 #include <asm/uaccess.h>
  16 #include <asm/ptrace.h>
  17 #include <asm/i387.h>
  18 #include <asm/user.h>
  19
  20 #ifdef CONFIG_X86_64
  21 # include <asm/sigcontext32.h>
  22 # include <asm/user32.h>
  23 #else
  24 # define save_i387_ia32         save_i387
  25 # define restore_i387_ia32      restore_i387
  26 # define _fpstate_ia32          _fpstate
  27 # define user_i387_ia32_struct  user_i387_struct
  28 # define user32_fxsr_struct     user_fxsr_struct
  29 #endif
  30
  31 #ifdef CONFIG_MATH_EMULATION
  32 # define HAVE_HWFP              (boot_cpu_data.hard_math)
  33 #else
  34 # define HAVE_HWFP              1
  35 #endif
  36
  37 static unsigned int             mxcsr_feature_mask __read_mostly = 0xffffffffu;
  38 unsigned int xstate_size;
  39 static struct i387_fxsave_struct fx_scratch __cpuinitdata;
  40
  41 void __cpuinit mxcsr_feature_mask_init(void)
  42 {
  43         unsigned long mask = 0;
  44
  45         clts();
  46         if (cpu_has_fxsr) {
  47                 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
  48                 asm volatile("fxsave %0" : : "m" (fx_scratch));
  49                 mask = fx_scratch.mxcsr_mask;
  50                 if (mask == 0)
  51                         mask = 0x0000ffbf;
  52         }
  53         mxcsr_feature_mask &= mask;
  54         stts();
  55 }
  56
  57 void __init init_thread_xstate(void)
  58 {
  59         if (!HAVE_HWFP) {
  60                 xstate_size = sizeof(struct i387_soft_struct);
  61                 return;
  62         }
  63
  64         if (cpu_has_xsave) {
  65                 xsave_cntxt_init();
  66                 return;
  67         }
  68
  69         if (cpu_has_fxsr)
  70                 xstate_size = sizeof(struct i387_fxsave_struct);
  71 #ifdef CONFIG_X86_32
  72         else
  73                 xstate_size = sizeof(struct i387_fsave_struct);
  74 #endif
  75 }
  76
  77 #ifdef CONFIG_X86_64
  78 /*
  79  * Called at bootup to set up the initial FPU state that is later cloned
  80  * into all processes.
  81  */
  82 void __cpuinit fpu_init(void)
  83 {
  84         unsigned long oldcr0 = read_cr0();
  85
  86         set_in_cr4(X86_CR4_OSFXSR);
  87         set_in_cr4(X86_CR4_OSXMMEXCPT);
  88
  89         write_cr0(oldcr0 & ~(X86_CR0_TS|X86_CR0_EM)); /* clear TS and EM */
  90
  91         /*
  92          * Boot processor to setup the FP and extended state context info.
  93          */
  94         if (!smp_processor_id())
  95                 init_thread_xstate();
  96         xsave_init();
  97
  98         mxcsr_feature_mask_init();
  99         /* clean state in init */
 100         current_thread_info()->status = 0;
 101         clear_used_math();
 102 }
 103 #endif  /* CONFIG_X86_64 */
 104
 105 /*
 106  * The _current_ task is using the FPU for the first time
 107  * so initialize it and set the mxcsr to its default
 108  * value at reset if we support XMM instructions and then
 109  * remeber the current task has used the FPU.
 110  */
 111 int init_fpu(struct task_struct *tsk)
 112 {
 113         if (tsk_used_math(tsk)) {
 114                 if (HAVE_HWFP && tsk == current)
 115                         unlazy_fpu(tsk);
 116                 return 0;
 117         }
 118
 119         /*
 120          * Memory allocation at the first usage of the FPU and other state.
 121          */
 122         if (!tsk->thread.xstate) {
 123                 tsk->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
 124                                                       GFP_KERNEL);
 125                 if (!tsk->thread.xstate)
 126                         return -ENOMEM;
 127         }
 128
 129 #ifdef CONFIG_X86_32
 130         if (!HAVE_HWFP) {
 131                 memset(tsk->thread.xstate, 0, xstate_size);
 132                 finit();
 133                 set_stopped_child_used_math(tsk);
 134                 return 0;
 135         }
 136 #endif
 137
 138         if (cpu_has_fxsr) {
 139                 struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 140
 141                 memset(fx, 0, xstate_size);
 142                 fx->cwd = 0x37f;
 143                 if (cpu_has_xmm)
 144                         fx->mxcsr = MXCSR_DEFAULT;
 145         } else {
 146                 struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 147                 memset(fp, 0, xstate_size);
 148                 fp->cwd = 0xffff037fu;
 149                 fp->swd = 0xffff0000u;
 150                 fp->twd = 0xffffffffu;
 151                 fp->fos = 0xffff0000u;
 152         }
 153         /*
 154          * Only the device not available exception or ptrace can call init_fpu.
 155          */
 156         set_stopped_child_used_math(tsk);
 157         return 0;
 158 }
 159
 160 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
 161 {
 162         return tsk_used_math(target) ? regset->n : 0;
 163 }
 164
 165 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
 166 {
 167         return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
 168 }
 169
 170 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
 171                 unsigned int pos, unsigned int count,
 172                 void *kbuf, void __user *ubuf)
 173 {
 174         int ret;
 175
 176         if (!cpu_has_fxsr)
 177                 return -ENODEV;
 178
 179         ret = init_fpu(target);
 180         if (ret)
 181                 return ret;
 182
 183         return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 184                                    &target->thread.xstate->fxsave, 0, -1);
 185 }
 186
 187 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
 188                 unsigned int pos, unsigned int count,
 189                 const void *kbuf, const void __user *ubuf)
 190 {
 191         int ret;
 192
 193         if (!cpu_has_fxsr)
 194                 return -ENODEV;
 195
 196         ret = init_fpu(target);
 197         if (ret)
 198                 return ret;
 199
 200         set_stopped_child_used_math(target);
 201
 202         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 203                                  &target->thread.xstate->fxsave, 0, -1);
 204
 205         /*
 206          * mxcsr reserved bits must be masked to zero for security reasons.
 207          */
 208         target->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 209
 210         return ret;
 211 }
 212
 213 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
 214
 215 /*
 216  * FPU tag word conversions.
 217  */
 218
 219 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
 220 {
 221         unsigned int tmp; /* to avoid 16 bit prefixes in the code */
 222
 223         /* Transform each pair of bits into 01 (valid) or 00 (empty) */
 224         tmp = ~twd;
 225         tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
 226         /* and move the valid bits to the lower byte. */
 227         tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
 228         tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
 229         tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
 230
 231         return tmp;
 232 }
 233
 234 #define FPREG_ADDR(f, n)        ((void *)&(f)->st_space + (n) * 16);
 235 #define FP_EXP_TAG_VALID        0
 236 #define FP_EXP_TAG_ZERO         1
 237 #define FP_EXP_TAG_SPECIAL      2
 238 #define FP_EXP_TAG_EMPTY        3
 239
 240 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
 241 {
 242         struct _fpxreg *st;
 243         u32 tos = (fxsave->swd >> 11) & 7;
 244         u32 twd = (unsigned long) fxsave->twd;
 245         u32 tag;
 246         u32 ret = 0xffff0000u;
 247         int i;
 248
 249         for (i = 0; i < 8; i++, twd >>= 1) {
 250                 if (twd & 0x1) {
 251                         st = FPREG_ADDR(fxsave, (i - tos) & 7);
 252
 253                         switch (st->exponent & 0x7fff) {
 254                         case 0x7fff:
 255                                 tag = FP_EXP_TAG_SPECIAL;
 256                                 break;
 257                         case 0x0000:
 258                                 if (!st->significand[0] &&
 259                                     !st->significand[1] &&
 260                                     !st->significand[2] &&
 261                                     !st->significand[3])
 262                                         tag = FP_EXP_TAG_ZERO;
 263                                 else
 264                                         tag = FP_EXP_TAG_SPECIAL;
 265                                 break;
 266                         default:
 267                                 if (st->significand[3] & 0x8000)
 268                                         tag = FP_EXP_TAG_VALID;
 269                                 else
 270                                         tag = FP_EXP_TAG_SPECIAL;
 271                                 break;
 272                         }
 273                 } else {
 274                         tag = FP_EXP_TAG_EMPTY;
 275                 }
 276                 ret |= tag << (2 * i);
 277         }
 278         return ret;
 279 }
 280
 281 /*
 282  * FXSR floating point environment conversions.
 283  */
 284
 285 static void
 286 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
 287 {
 288         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 289         struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
 290         struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
 291         int i;
 292
 293         env->cwd = fxsave->cwd | 0xffff0000u;
 294         env->swd = fxsave->swd | 0xffff0000u;
 295         env->twd = twd_fxsr_to_i387(fxsave);
 296
 297 #ifdef CONFIG_X86_64
 298         env->fip = fxsave->rip;
 299         env->foo = fxsave->rdp;
 300         if (tsk == current) {
 301                 /*
 302                  * should be actually ds/cs at fpu exception time, but
 303                  * that information is not available in 64bit mode.
 304                  */
 305                 asm("mov %%ds, %[fos]" : [fos] "=r" (env->fos));
 306                 asm("mov %%cs, %[fcs]" : [fcs] "=r" (env->fcs));
 307         } else {
 308                 struct pt_regs *regs = task_pt_regs(tsk);
 309
 310                 env->fos = 0xffff0000 | tsk->thread.ds;
 311                 env->fcs = regs->cs;
 312         }
 313 #else
 314         env->fip = fxsave->fip;
 315         env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
 316         env->foo = fxsave->foo;
 317         env->fos = fxsave->fos;
 318 #endif
 319
 320         for (i = 0; i < 8; ++i)
 321                 memcpy(&to[i], &from[i], sizeof(to[0]));
 322 }
 323
 324 static void convert_to_fxsr(struct task_struct *tsk,
 325                             const struct user_i387_ia32_struct *env)
 326
 327 {
 328         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 329         struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
 330         struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
 331         int i;
 332
 333         fxsave->cwd = env->cwd;
 334         fxsave->swd = env->swd;
 335         fxsave->twd = twd_i387_to_fxsr(env->twd);
 336         fxsave->fop = (u16) ((u32) env->fcs >> 16);
 337 #ifdef CONFIG_X86_64
 338         fxsave->rip = env->fip;
 339         fxsave->rdp = env->foo;
 340         /* cs and ds ignored */
 341 #else
 342         fxsave->fip = env->fip;
 343         fxsave->fcs = (env->fcs & 0xffff);
 344         fxsave->foo = env->foo;
 345         fxsave->fos = env->fos;
 346 #endif
 347
 348         for (i = 0; i < 8; ++i)
 349                 memcpy(&to[i], &from[i], sizeof(from[0]));
 350 }
 351
 352 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
 353                unsigned int pos, unsigned int count,
 354                void *kbuf, void __user *ubuf)
 355 {
 356         struct user_i387_ia32_struct env;
 357         int ret;
 358
 359         ret = init_fpu(target);
 360         if (ret)
 361                 return ret;
 362
 363         if (!HAVE_HWFP)
 364                 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
 365
 366         if (!cpu_has_fxsr) {
 367                 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 368                                            &target->thread.xstate->fsave, 0,
 369                                            -1);
 370         }
 371
 372         if (kbuf && pos == 0 && count == sizeof(env)) {
 373                 convert_from_fxsr(kbuf, target);
 374                 return 0;
 375         }
 376
 377         convert_from_fxsr(&env, target);
 378
 379         return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 380 }
 381
 382 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
 383                unsigned int pos, unsigned int count,
 384                const void *kbuf, const void __user *ubuf)
 385 {
 386         struct user_i387_ia32_struct env;
 387         int ret;
 388
 389         ret = init_fpu(target);
 390         if (ret)
 391                 return ret;
 392
 393         set_stopped_child_used_math(target);
 394
 395         if (!HAVE_HWFP)
 396                 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
 397
 398         if (!cpu_has_fxsr) {
 399                 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 400                                           &target->thread.xstate->fsave, 0, -1);
 401         }
 402
 403         if (pos > 0 || count < sizeof(env))
 404                 convert_from_fxsr(&env, target);
 405
 406         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 407         if (!ret)
 408                 convert_to_fxsr(target, &env);
 409
 410         return ret;
 411 }
 412
 413 /*
 414  * Signal frame handlers.
 415  */
 416
 417 static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
 418 {
 419         struct task_struct *tsk = current;
 420         struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 421
 422         unlazy_fpu(tsk);
 423         fp->status = fp->swd;
 424         if (__copy_to_user(buf, fp, sizeof(struct i387_fsave_struct)))
 425                 return -1;
 426         return 1;
 427 }
 428
 429 static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
 430 {
 431         struct task_struct *tsk = current;
 432         struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 433         struct user_i387_ia32_struct env;
 434         int err = 0;
 435
 436         unlazy_fpu(tsk);
 437
 438         convert_from_fxsr(&env, tsk);
 439         if (__copy_to_user(buf, &env, sizeof(env)))
 440                 return -1;
 441
 442         err |= __put_user(fx->swd, &buf->status);
 443         err |= __put_user(X86_FXSR_MAGIC, &buf->magic);
 444         if (err)
 445                 return -1;
 446
 447         if (__copy_to_user(&buf->_fxsr_env[0], fx,
 448                            sizeof(struct i387_fxsave_struct)))
 449                 return -1;
 450         return 1;
 451 }
 452
 453 int save_i387_ia32(struct _fpstate_ia32 __user *buf)
 454 {
 455         if (!used_math())
 456                 return 0;
 457         /*
 458          * This will cause a "finit" to be triggered by the next
 459          * attempted FPU operation by the 'current' process.
 460          */
 461         clear_used_math();
 462
 463         if (!HAVE_HWFP) {
 464                 return fpregs_soft_get(current, NULL,
 465                                        0, sizeof(struct user_i387_ia32_struct),
 466                                        NULL, buf) ? -1 : 1;
 467         }
 468
 469         if (cpu_has_fxsr)
 470                 return save_i387_fxsave(buf);
 471         else
 472                 return save_i387_fsave(buf);
 473 }
 474
 475 static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
 476 {
 477         struct task_struct *tsk = current;
 478
 479         return __copy_from_user(&tsk->thread.xstate->fsave, buf,
 480                                 sizeof(struct i387_fsave_struct));
 481 }
 482
 483 static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf)
 484 {
 485         struct task_struct *tsk = current;
 486         struct user_i387_ia32_struct env;
 487         int err;
 488
 489         err = __copy_from_user(&tsk->thread.xstate->fxsave, &buf->_fxsr_env[0],
 490                                sizeof(struct i387_fxsave_struct));
 491         /* mxcsr reserved bits must be masked to zero for security reasons */
 492         tsk->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 493         if (err || __copy_from_user(&env, buf, sizeof(env)))
 494                 return 1;
 495         convert_to_fxsr(tsk, &env);
 496
 497         return 0;
 498 }
 499
 500 int restore_i387_ia32(struct _fpstate_ia32 __user *buf)
 501 {
 502         int err;
 503         struct task_struct *tsk = current;
 504
 505         if (HAVE_HWFP)
 506                 clear_fpu(tsk);
 507
 508         if (!used_math()) {
 509                 err = init_fpu(tsk);
 510                 if (err)
 511                         return err;
 512         }
 513
 514         if (HAVE_HWFP) {
 515                 if (cpu_has_fxsr)
 516                         err = restore_i387_fxsave(buf);
 517                 else
 518                         err = restore_i387_fsave(buf);
 519         } else {
 520                 err = fpregs_soft_set(current, NULL,
 521                                       0, sizeof(struct user_i387_ia32_struct),
 522                                       NULL, buf) != 0;
 523         }
 524         set_used_math();
 525
 526         return err;
 527 }
 528
 529 /*
 530  * FPU state for core dumps.
 531  * This is only used for a.out dumps now.
 532  * It is declared generically using elf_fpregset_t (which is
 533  * struct user_i387_struct) but is in fact only used for 32-bit
 534  * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
 535  */
 536 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
 537 {
 538         struct task_struct *tsk = current;
 539         int fpvalid;
 540
 541         fpvalid = !!used_math();
 542         if (fpvalid)
 543                 fpvalid = !fpregs_get(tsk, NULL,
 544                                       0, sizeof(struct user_i387_ia32_struct),
 545                                       fpu, NULL);
 546
 547         return fpvalid;
 548 }
 549 EXPORT_SYMBOL(dump_fpu);
 550
 551 #endif  /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */