arch/i386/kernel/i387.c

   1 /*
   2  *  linux/arch/i386/kernel/i387.c
   3  *
   4  *  Copyright (C) 1994 Linus Torvalds
   5  *
   6  *  Pentium III FXSR, SSE support
   7  *  General FPU state handling cleanups
   8  *      Gareth Hughes <gareth@valinux.com>, May 2000
   9  */
  10
  11 #include <linux/sched.h>
  12 #include <linux/module.h>
  13 #include <asm/processor.h>
  14 #include <asm/i387.h>
  15 #include <asm/math_emu.h>
  16 #include <asm/sigcontext.h>
  17 #include <asm/user.h>
  18 #include <asm/ptrace.h>
  19 #include <asm/uaccess.h>
  20
  21 #ifdef CONFIG_MATH_EMULATION
  22 #define HAVE_HWFP (boot_cpu_data.hard_math)
  23 #else
  24 #define HAVE_HWFP 1
  25 #endif
  26
  27 static unsigned long mxcsr_feature_mask __read_mostly = 0xffffffff;
  28
  29 void mxcsr_feature_mask_init(void)
  30 {
  31         unsigned long mask = 0;
  32         clts();
  33         if (cpu_has_fxsr) {
  34                 memset(&current->thread.i387.fxsave, 0, sizeof(struct i387_fxsave_struct));
  35                 asm volatile("fxsave %0" : : "m" (current->thread.i387.fxsave));
  36                 mask = current->thread.i387.fxsave.mxcsr_mask;
  37                 if (mask == 0) mask = 0x0000ffbf;
  38         }
  39         mxcsr_feature_mask &= mask;
  40         stts();
  41 }
  42
  43 /*
  44  * The _current_ task is using the FPU for the first time
  45  * so initialize it and set the mxcsr to its default
  46  * value at reset if we support XMM instructions and then
  47  * remeber the current task has used the FPU.
  48  */
  49 void init_fpu(struct task_struct *tsk)
  50 {
  51         if (cpu_has_fxsr) {
  52                 memset(&tsk->thread.i387.fxsave, 0, sizeof(struct i387_fxsave_struct));
  53                 tsk->thread.i387.fxsave.cwd = 0x37f;
  54                 if (cpu_has_xmm)
  55                         tsk->thread.i387.fxsave.mxcsr = 0x1f80;
  56         } else {
  57                 memset(&tsk->thread.i387.fsave, 0, sizeof(struct i387_fsave_struct));
  58                 tsk->thread.i387.fsave.cwd = 0xffff037fu;
  59                 tsk->thread.i387.fsave.swd = 0xffff0000u;
  60                 tsk->thread.i387.fsave.twd = 0xffffffffu;
  61                 tsk->thread.i387.fsave.fos = 0xffff0000u;
  62         }
  63         /* only the device not available exception or ptrace can call init_fpu */
  64         set_stopped_child_used_math(tsk);
  65 }
  66
  67 /*
  68  * FPU lazy state save handling.
  69  */
  70
  71 void kernel_fpu_begin(void)
  72 {
  73         struct thread_info *thread = current_thread_info();
  74
  75         preempt_disable();
  76         if (thread->status & TS_USEDFPU) {
  77                 __save_init_fpu(thread->task);
  78                 return;
  79         }
  80         clts();
  81 }
  82 EXPORT_SYMBOL_GPL(kernel_fpu_begin);
  83
  84 /*
  85  * FPU tag word conversions.
  86  */
  87
  88 static inline unsigned short twd_i387_to_fxsr( unsigned short twd )
  89 {
  90         unsigned int tmp; /* to avoid 16 bit prefixes in the code */
  91
  92         /* Transform each pair of bits into 01 (valid) or 00 (empty) */
  93         tmp = ~twd;
  94         tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
  95         /* and move the valid bits to the lower byte. */
  96         tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
  97         tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
  98         tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
  99         return tmp;
 100 }
 101
 102 static inline unsigned long twd_fxsr_to_i387( struct i387_fxsave_struct *fxsave )
 103 {
 104         struct _fpxreg *st = NULL;
 105         unsigned long tos = (fxsave->swd >> 11) & 7;
 106         unsigned long twd = (unsigned long) fxsave->twd;
 107         unsigned long tag;
 108         unsigned long ret = 0xffff0000u;
 109         int i;
 110
 111 #define FPREG_ADDR(f, n)        ((void *)&(f)->st_space + (n) * 16);
 112
 113         for ( i = 0 ; i < 8 ; i++ ) {
 114                 if ( twd & 0x1 ) {
 115                         st = FPREG_ADDR( fxsave, (i - tos) & 7 );
 116
 117                         switch ( st->exponent & 0x7fff ) {
 118                         case 0x7fff:
 119                                 tag = 2;                /* Special */
 120                                 break;
 121                         case 0x0000:
 122                                 if ( !st->significand[0] &&
 123                                      !st->significand[1] &&
 124                                      !st->significand[2] &&
 125                                      !st->significand[3] ) {
 126                                         tag = 1;        /* Zero */
 127                                 } else {
 128                                         tag = 2;        /* Special */
 129                                 }
 130                                 break;
 131                         default:
 132                                 if ( st->significand[3] & 0x8000 ) {
 133                                         tag = 0;        /* Valid */
 134                                 } else {
 135                                         tag = 2;        /* Special */
 136                                 }
 137                                 break;
 138                         }
 139                 } else {
 140                         tag = 3;                        /* Empty */
 141                 }
 142                 ret |= (tag << (2 * i));
 143                 twd = twd >> 1;
 144         }
 145         return ret;
 146 }
 147
 148 /*
 149  * FPU state interaction.
 150  */
 151
 152 unsigned short get_fpu_cwd( struct task_struct *tsk )
 153 {
 154         if ( cpu_has_fxsr ) {
 155                 return tsk->thread.i387.fxsave.cwd;
 156         } else {
 157                 return (unsigned short)tsk->thread.i387.fsave.cwd;
 158         }
 159 }
 160
 161 unsigned short get_fpu_swd( struct task_struct *tsk )
 162 {
 163         if ( cpu_has_fxsr ) {
 164                 return tsk->thread.i387.fxsave.swd;
 165         } else {
 166                 return (unsigned short)tsk->thread.i387.fsave.swd;
 167         }
 168 }
 169
 170 #if 0
 171 unsigned short get_fpu_twd( struct task_struct *tsk )
 172 {
 173         if ( cpu_has_fxsr ) {
 174                 return tsk->thread.i387.fxsave.twd;
 175         } else {
 176                 return (unsigned short)tsk->thread.i387.fsave.twd;
 177         }
 178 }
 179 #endif  /*  0  */
 180
 181 unsigned short get_fpu_mxcsr( struct task_struct *tsk )
 182 {
 183         if ( cpu_has_xmm ) {
 184                 return tsk->thread.i387.fxsave.mxcsr;
 185         } else {
 186                 return 0x1f80;
 187         }
 188 }
 189
 190 #if 0
 191
 192 void set_fpu_cwd( struct task_struct *tsk, unsigned short cwd )
 193 {
 194         if ( cpu_has_fxsr ) {
 195                 tsk->thread.i387.fxsave.cwd = cwd;
 196         } else {
 197                 tsk->thread.i387.fsave.cwd = ((long)cwd | 0xffff0000u);
 198         }
 199 }
 200
 201 void set_fpu_swd( struct task_struct *tsk, unsigned short swd )
 202 {
 203         if ( cpu_has_fxsr ) {
 204                 tsk->thread.i387.fxsave.swd = swd;
 205         } else {
 206                 tsk->thread.i387.fsave.swd = ((long)swd | 0xffff0000u);
 207         }
 208 }
 209
 210 void set_fpu_twd( struct task_struct *tsk, unsigned short twd )
 211 {
 212         if ( cpu_has_fxsr ) {
 213                 tsk->thread.i387.fxsave.twd = twd_i387_to_fxsr(twd);
 214         } else {
 215                 tsk->thread.i387.fsave.twd = ((long)twd | 0xffff0000u);
 216         }
 217 }
 218
 219 #endif  /*  0  */
 220
 221 /*
 222  * FXSR floating point environment conversions.
 223  */
 224
 225 static inline void
 226 convert_fxsr_env_to_i387(unsigned long env[7],
 227                          struct i387_fxsave_struct *fxsave)
 228 {
 229         env[0] = (unsigned long)fxsave->cwd | 0xffff0000ul;
 230         env[1] = (unsigned long)fxsave->swd | 0xffff0000ul;
 231         env[2] = twd_fxsr_to_i387(fxsave);
 232         env[3] = fxsave->fip;
 233         env[4] = fxsave->fcs | ((unsigned long)fxsave->fop << 16);
 234         env[5] = fxsave->foo;
 235         env[6] = fxsave->fos;
 236 }
 237
 238 static int convert_fxsr_to_user(struct _fpstate __user *buf,
 239                                 struct i387_fxsave_struct *fxsave)
 240 {
 241         unsigned long env[7];
 242         struct _fpreg __user *to;
 243         struct _fpxreg *from;
 244         int i;
 245
 246         convert_fxsr_env_to_i387(env, fxsave);
 247         if ( __copy_to_user( buf, env, 7 * sizeof(unsigned long) ) )
 248                 return 1;
 249
 250         to = &buf->_st[0];
 251         from = (struct _fpxreg *) &fxsave->st_space[0];
 252         for ( i = 0 ; i < 8 ; i++, to++, from++ ) {
 253                 unsigned long __user *t = (unsigned long __user *)to;
 254                 unsigned long *f = (unsigned long *)from;
 255
 256                 if (__put_user(*f, t) ||
 257                                 __put_user(*(f + 1), t + 1) ||
 258                                 __put_user(from->exponent, &to->exponent))
 259                         return 1;
 260         }
 261         return 0;
 262 }
 263
 264 static inline void
 265 convert_fxsr_env_from_i387(struct i387_fxsave_struct *fxsave,
 266                            const unsigned long env[7])
 267 {
 268         fxsave->cwd = (unsigned short)(env[0] & 0xffff);
 269         fxsave->swd = (unsigned short)(env[1] & 0xffff);
 270         fxsave->twd = twd_i387_to_fxsr((unsigned short)(env[2] & 0xffff));
 271         fxsave->fip = env[3];
 272         fxsave->fop = (unsigned short)((env[4] & 0xffff0000ul) >> 16);
 273         fxsave->fcs = (env[4] & 0xffff);
 274         fxsave->foo = env[5];
 275         fxsave->fos = env[6];
 276 }
 277
 278 static int convert_fxsr_from_user( struct i387_fxsave_struct *fxsave,
 279                                           struct _fpstate __user *buf )
 280 {
 281         unsigned long env[7];
 282         struct _fpxreg *to;
 283         struct _fpreg __user *from;
 284         int i;
 285
 286         if ( __copy_from_user( env, buf, 7 * sizeof(long) ) )
 287                 return 1;
 288
 289         convert_fxsr_env_from_i387(fxsave, env);
 290
 291         to = (struct _fpxreg *) &fxsave->st_space[0];
 292         from = &buf->_st[0];
 293         for ( i = 0 ; i < 8 ; i++, to++, from++ ) {
 294                 unsigned long *t = (unsigned long *)to;
 295                 unsigned long __user *f = (unsigned long __user *)from;
 296
 297                 if (__get_user(*t, f) ||
 298                                 __get_user(*(t + 1), f + 1) ||
 299                                 __get_user(to->exponent, &from->exponent))
 300                         return 1;
 301         }
 302         return 0;
 303 }
 304
 305 /*
 306  * Signal frame handlers.
 307  */
 308
 309 static inline int save_i387_fsave( struct _fpstate __user *buf )
 310 {
 311         struct task_struct *tsk = current;
 312
 313         unlazy_fpu( tsk );
 314         tsk->thread.i387.fsave.status = tsk->thread.i387.fsave.swd;
 315         if ( __copy_to_user( buf, &tsk->thread.i387.fsave,
 316                              sizeof(struct i387_fsave_struct) ) )
 317                 return -1;
 318         return 1;
 319 }
 320
 321 static int save_i387_fxsave( struct _fpstate __user *buf )
 322 {
 323         struct task_struct *tsk = current;
 324         int err = 0;
 325
 326         unlazy_fpu( tsk );
 327
 328         if ( convert_fxsr_to_user( buf, &tsk->thread.i387.fxsave ) )
 329                 return -1;
 330
 331         err |= __put_user( tsk->thread.i387.fxsave.swd, &buf->status );
 332         err |= __put_user( X86_FXSR_MAGIC, &buf->magic );
 333         if ( err )
 334                 return -1;
 335
 336         if ( __copy_to_user( &buf->_fxsr_env[0], &tsk->thread.i387.fxsave,
 337                              sizeof(struct i387_fxsave_struct) ) )
 338                 return -1;
 339         return 1;
 340 }
 341
 342 int save_i387( struct _fpstate __user *buf )
 343 {
 344         if ( !used_math() )
 345                 return 0;
 346
 347         /* This will cause a "finit" to be triggered by the next
 348          * attempted FPU operation by the 'current' process.
 349          */
 350         clear_used_math();
 351
 352         if ( HAVE_HWFP ) {
 353                 if ( cpu_has_fxsr ) {
 354                         return save_i387_fxsave( buf );
 355                 } else {
 356                         return save_i387_fsave( buf );
 357                 }
 358         } else {
 359                 return save_i387_soft( &current->thread.i387.soft, buf );
 360         }
 361 }
 362
 363 static inline int restore_i387_fsave( struct _fpstate __user *buf )
 364 {
 365         struct task_struct *tsk = current;
 366         clear_fpu( tsk );
 367         return __copy_from_user( &tsk->thread.i387.fsave, buf,
 368                                  sizeof(struct i387_fsave_struct) );
 369 }
 370
 371 static int restore_i387_fxsave( struct _fpstate __user *buf )
 372 {
 373         int err;
 374         struct task_struct *tsk = current;
 375         clear_fpu( tsk );
 376         err = __copy_from_user( &tsk->thread.i387.fxsave, &buf->_fxsr_env[0],
 377                                 sizeof(struct i387_fxsave_struct) );
 378         /* mxcsr reserved bits must be masked to zero for security reasons */
 379         tsk->thread.i387.fxsave.mxcsr &= mxcsr_feature_mask;
 380         return err ? 1 : convert_fxsr_from_user( &tsk->thread.i387.fxsave, buf );
 381 }
 382
 383 int restore_i387( struct _fpstate __user *buf )
 384 {
 385         int err;
 386
 387         if ( HAVE_HWFP ) {
 388                 if ( cpu_has_fxsr ) {
 389                         err = restore_i387_fxsave( buf );
 390                 } else {
 391                         err = restore_i387_fsave( buf );
 392                 }
 393         } else {
 394                 err = restore_i387_soft( &current->thread.i387.soft, buf );
 395         }
 396         set_used_math();
 397         return err;
 398 }
 399
 400 /*
 401  * ptrace request handlers.
 402  */
 403
 404 static inline void get_fpregs_fsave(struct user_i387_struct *buf,
 405                                     struct task_struct *tsk)
 406 {
 407         memcpy(buf, &tsk->thread.i387.fsave, sizeof(struct user_i387_struct));
 408 }
 409
 410 static inline void get_fpregs_fxsave(struct user_i387_struct *buf,
 411                                      struct task_struct *tsk)
 412 {
 413         struct _fpreg *to;
 414         const struct _fpxreg *from;
 415         unsigned int i;
 416
 417         convert_fxsr_env_to_i387((unsigned long *) buf,
 418                                  &tsk->thread.i387.fxsave);
 419
 420         to = (struct _fpreg *) buf->st_space;
 421         from = (const struct _fpxreg *) &tsk->thread.i387.fxsave.st_space[0];
 422         for (i = 0; i < 8; i++, to++, from++)
 423                 *to = *(const struct _fpreg *) from;
 424 }
 425
 426 int get_fpregs(struct user_i387_struct *buf, struct task_struct *tsk)
 427 {
 428         if ( HAVE_HWFP ) {
 429                 if (cpu_has_fxsr)
 430                         get_fpregs_fxsave(buf, tsk);
 431                 else
 432                         get_fpregs_fsave(buf, tsk);
 433                 return 0;
 434         } else {
 435                 return save_i387_soft( &tsk->thread.i387.soft,
 436                                        (struct _fpstate __user *)buf );
 437         }
 438 }
 439
 440 static inline void set_fpregs_fsave(struct task_struct *tsk,
 441                                     const struct user_i387_struct *buf)
 442 {
 443         memcpy(&tsk->thread.i387.fsave, buf, sizeof(struct user_i387_struct));
 444 }
 445
 446 static inline void set_fpregs_fxsave(struct task_struct *tsk,
 447                                      const struct user_i387_struct *buf)
 448 {
 449         struct _fpxreg *to;
 450         const struct _fpreg *from;
 451         unsigned int i;
 452
 453         convert_fxsr_env_from_i387(&tsk->thread.i387.fxsave,
 454                                    (unsigned long *) buf);
 455
 456         to = (struct _fpxreg *) &tsk->thread.i387.fxsave.st_space[0];
 457         from = (const struct _fpreg *) buf->st_space;
 458         for (i = 0; i < 8; i++, to++, from++)
 459                 *(struct _fpreg *) to = *from;
 460 }
 461
 462 int set_fpregs(struct task_struct *tsk, const struct user_i387_struct *buf)
 463 {
 464         if ( HAVE_HWFP ) {
 465                 if (cpu_has_fxsr)
 466                         set_fpregs_fxsave(tsk, buf);
 467                 else
 468                         set_fpregs_fsave(tsk, buf);
 469                 return 0;
 470         } else {
 471                 return restore_i387_soft( &tsk->thread.i387.soft,
 472                                           (struct _fpstate __user *)buf );
 473         }
 474 }
 475
 476 void updated_fpxregs(struct task_struct *tsk)
 477 {
 478         /* mxcsr reserved bits must be masked to zero for security reasons */
 479         tsk->thread.i387.fxsave.mxcsr &= mxcsr_feature_mask;
 480 }
 481
 482 /*
 483  * FPU state for core dumps.
 484  */
 485
 486 static inline void copy_fpu_fsave( struct task_struct *tsk,
 487                                    struct user_i387_struct *fpu )
 488 {
 489         memcpy( fpu, &tsk->thread.i387.fsave,
 490                 sizeof(struct user_i387_struct) );
 491 }
 492
 493 static inline void copy_fpu_fxsave( struct task_struct *tsk,
 494                                    struct user_i387_struct *fpu )
 495 {
 496         unsigned short *to;
 497         unsigned short *from;
 498         int i;
 499
 500         memcpy( fpu, &tsk->thread.i387.fxsave, 7 * sizeof(long) );
 501
 502         to = (unsigned short *)&fpu->st_space[0];
 503         from = (unsigned short *)&tsk->thread.i387.fxsave.st_space[0];
 504         for ( i = 0 ; i < 8 ; i++, to += 5, from += 8 ) {
 505                 memcpy( to, from, 5 * sizeof(unsigned short) );
 506         }
 507 }
 508
 509 int dump_fpu( struct pt_regs *regs, struct user_i387_struct *fpu )
 510 {
 511         int fpvalid;
 512         struct task_struct *tsk = current;
 513
 514         fpvalid = !!used_math();
 515         if ( fpvalid ) {
 516                 unlazy_fpu( tsk );
 517                 if ( cpu_has_fxsr ) {
 518                         copy_fpu_fxsave( tsk, fpu );
 519                 } else {
 520                         copy_fpu_fsave( tsk, fpu );
 521                 }
 522         }
 523
 524         return fpvalid;
 525 }
 526 EXPORT_SYMBOL(dump_fpu);
 527
 528 int dump_task_fpu(struct task_struct *tsk, struct user_i387_struct *fpu)
 529 {
 530         int fpvalid = !!tsk_used_math(tsk);
 531
 532         if (fpvalid) {
 533                 if (tsk == current)
 534                         unlazy_fpu(tsk);
 535                 if (cpu_has_fxsr)
 536                         copy_fpu_fxsave(tsk, fpu);
 537                 else
 538                         copy_fpu_fsave(tsk, fpu);
 539         }
 540         return fpvalid;
 541 }
 542
 543 int dump_task_extended_fpu(struct task_struct *tsk, struct user_fxsr_struct *fpu)
 544 {
 545         int fpvalid = tsk_used_math(tsk) && cpu_has_fxsr;
 546
 547         if (fpvalid) {
 548                 if (tsk == current)
 549                        unlazy_fpu(tsk);
 550                 memcpy(fpu, &tsk->thread.i387.fxsave, sizeof(*fpu));
 551         }
 552         return fpvalid;
 553 }