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         if (cpu_has_xsave)
 101                 current_thread_info()->status = TS_XSAVE;
 102         else
 103                 current_thread_info()->status = 0;
 104         clear_used_math();
 105 }
 106 #endif  /* CONFIG_X86_64 */
 107
 108 /*
 109  * The _current_ task is using the FPU for the first time
 110  * so initialize it and set the mxcsr to its default
 111  * value at reset if we support XMM instructions and then
 112  * remeber the current task has used the FPU.
 113  */
 114 int init_fpu(struct task_struct *tsk)
 115 {
 116         if (tsk_used_math(tsk)) {
 117                 if (HAVE_HWFP && tsk == current)
 118                         unlazy_fpu(tsk);
 119                 return 0;
 120         }
 121
 122         /*
 123          * Memory allocation at the first usage of the FPU and other state.
 124          */
 125         if (!tsk->thread.xstate) {
 126                 tsk->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
 127                                                       GFP_KERNEL);
 128                 if (!tsk->thread.xstate)
 129                         return -ENOMEM;
 130         }
 131
 132 #ifdef CONFIG_X86_32
 133         if (!HAVE_HWFP) {
 134                 memset(tsk->thread.xstate, 0, xstate_size);
 135                 finit();
 136                 set_stopped_child_used_math(tsk);
 137                 return 0;
 138         }
 139 #endif
 140
 141         if (cpu_has_fxsr) {
 142                 struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 143
 144                 memset(fx, 0, xstate_size);
 145                 fx->cwd = 0x37f;
 146                 if (cpu_has_xmm)
 147                         fx->mxcsr = MXCSR_DEFAULT;
 148         } else {
 149                 struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 150                 memset(fp, 0, xstate_size);
 151                 fp->cwd = 0xffff037fu;
 152                 fp->swd = 0xffff0000u;
 153                 fp->twd = 0xffffffffu;
 154                 fp->fos = 0xffff0000u;
 155         }
 156         /*
 157          * Only the device not available exception or ptrace can call init_fpu.
 158          */
 159         set_stopped_child_used_math(tsk);
 160         return 0;
 161 }
 162
 163 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
 164 {
 165         return tsk_used_math(target) ? regset->n : 0;
 166 }
 167
 168 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
 169 {
 170         return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
 171 }
 172
 173 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
 174                 unsigned int pos, unsigned int count,
 175                 void *kbuf, void __user *ubuf)
 176 {
 177         int ret;
 178
 179         if (!cpu_has_fxsr)
 180                 return -ENODEV;
 181
 182         ret = init_fpu(target);
 183         if (ret)
 184                 return ret;
 185
 186         return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 187                                    &target->thread.xstate->fxsave, 0, -1);
 188 }
 189
 190 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
 191                 unsigned int pos, unsigned int count,
 192                 const void *kbuf, const void __user *ubuf)
 193 {
 194         int ret;
 195
 196         if (!cpu_has_fxsr)
 197                 return -ENODEV;
 198
 199         ret = init_fpu(target);
 200         if (ret)
 201                 return ret;
 202
 203         set_stopped_child_used_math(target);
 204
 205         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 206                                  &target->thread.xstate->fxsave, 0, -1);
 207
 208         /*
 209          * mxcsr reserved bits must be masked to zero for security reasons.
 210          */
 211         target->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 212
 213         return ret;
 214 }
 215
 216 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
 217
 218 /*
 219  * FPU tag word conversions.
 220  */
 221
 222 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
 223 {
 224         unsigned int tmp; /* to avoid 16 bit prefixes in the code */
 225
 226         /* Transform each pair of bits into 01 (valid) or 00 (empty) */
 227         tmp = ~twd;
 228         tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
 229         /* and move the valid bits to the lower byte. */
 230         tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
 231         tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
 232         tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
 233
 234         return tmp;
 235 }
 236
 237 #define FPREG_ADDR(f, n)        ((void *)&(f)->st_space + (n) * 16);
 238 #define FP_EXP_TAG_VALID        0
 239 #define FP_EXP_TAG_ZERO         1
 240 #define FP_EXP_TAG_SPECIAL      2
 241 #define FP_EXP_TAG_EMPTY        3
 242
 243 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
 244 {
 245         struct _fpxreg *st;
 246         u32 tos = (fxsave->swd >> 11) & 7;
 247         u32 twd = (unsigned long) fxsave->twd;
 248         u32 tag;
 249         u32 ret = 0xffff0000u;
 250         int i;
 251
 252         for (i = 0; i < 8; i++, twd >>= 1) {
 253                 if (twd & 0x1) {
 254                         st = FPREG_ADDR(fxsave, (i - tos) & 7);
 255
 256                         switch (st->exponent & 0x7fff) {
 257                         case 0x7fff:
 258                                 tag = FP_EXP_TAG_SPECIAL;
 259                                 break;
 260                         case 0x0000:
 261                                 if (!st->significand[0] &&
 262                                     !st->significand[1] &&
 263                                     !st->significand[2] &&
 264                                     !st->significand[3])
 265                                         tag = FP_EXP_TAG_ZERO;
 266                                 else
 267                                         tag = FP_EXP_TAG_SPECIAL;
 268                                 break;
 269                         default:
 270                                 if (st->significand[3] & 0x8000)
 271                                         tag = FP_EXP_TAG_VALID;
 272                                 else
 273                                         tag = FP_EXP_TAG_SPECIAL;
 274                                 break;
 275                         }
 276                 } else {
 277                         tag = FP_EXP_TAG_EMPTY;
 278                 }
 279                 ret |= tag << (2 * i);
 280         }
 281         return ret;
 282 }
 283
 284 /*
 285  * FXSR floating point environment conversions.
 286  */
 287
 288 static void
 289 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
 290 {
 291         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 292         struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
 293         struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
 294         int i;
 295
 296         env->cwd = fxsave->cwd | 0xffff0000u;
 297         env->swd = fxsave->swd | 0xffff0000u;
 298         env->twd = twd_fxsr_to_i387(fxsave);
 299
 300 #ifdef CONFIG_X86_64
 301         env->fip = fxsave->rip;
 302         env->foo = fxsave->rdp;
 303         if (tsk == current) {
 304                 /*
 305                  * should be actually ds/cs at fpu exception time, but
 306                  * that information is not available in 64bit mode.
 307                  */
 308                 asm("mov %%ds, %[fos]" : [fos] "=r" (env->fos));
 309                 asm("mov %%cs, %[fcs]" : [fcs] "=r" (env->fcs));
 310         } else {
 311                 struct pt_regs *regs = task_pt_regs(tsk);
 312
 313                 env->fos = 0xffff0000 | tsk->thread.ds;
 314                 env->fcs = regs->cs;
 315         }
 316 #else
 317         env->fip = fxsave->fip;
 318         env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
 319         env->foo = fxsave->foo;
 320         env->fos = fxsave->fos;
 321 #endif
 322
 323         for (i = 0; i < 8; ++i)
 324                 memcpy(&to[i], &from[i], sizeof(to[0]));
 325 }
 326
 327 static void convert_to_fxsr(struct task_struct *tsk,
 328                             const struct user_i387_ia32_struct *env)
 329
 330 {
 331         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 332         struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
 333         struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
 334         int i;
 335
 336         fxsave->cwd = env->cwd;
 337         fxsave->swd = env->swd;
 338         fxsave->twd = twd_i387_to_fxsr(env->twd);
 339         fxsave->fop = (u16) ((u32) env->fcs >> 16);
 340 #ifdef CONFIG_X86_64
 341         fxsave->rip = env->fip;
 342         fxsave->rdp = env->foo;
 343         /* cs and ds ignored */
 344 #else
 345         fxsave->fip = env->fip;
 346         fxsave->fcs = (env->fcs & 0xffff);
 347         fxsave->foo = env->foo;
 348         fxsave->fos = env->fos;
 349 #endif
 350
 351         for (i = 0; i < 8; ++i)
 352                 memcpy(&to[i], &from[i], sizeof(from[0]));
 353 }
 354
 355 int fpregs_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 user_i387_ia32_struct env;
 360         int ret;
 361
 362         ret = init_fpu(target);
 363         if (ret)
 364                 return ret;
 365
 366         if (!HAVE_HWFP)
 367                 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
 368
 369         if (!cpu_has_fxsr) {
 370                 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 371                                            &target->thread.xstate->fsave, 0,
 372                                            -1);
 373         }
 374
 375         if (kbuf && pos == 0 && count == sizeof(env)) {
 376                 convert_from_fxsr(kbuf, target);
 377                 return 0;
 378         }
 379
 380         convert_from_fxsr(&env, target);
 381
 382         return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 383 }
 384
 385 int fpregs_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 user_i387_ia32_struct env;
 390         int ret;
 391
 392         ret = init_fpu(target);
 393         if (ret)
 394                 return ret;
 395
 396         set_stopped_child_used_math(target);
 397
 398         if (!HAVE_HWFP)
 399                 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
 400
 401         if (!cpu_has_fxsr) {
 402                 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 403                                           &target->thread.xstate->fsave, 0, -1);
 404         }
 405
 406         if (pos > 0 || count < sizeof(env))
 407                 convert_from_fxsr(&env, target);
 408
 409         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 410         if (!ret)
 411                 convert_to_fxsr(target, &env);
 412
 413         return ret;
 414 }
 415
 416 /*
 417  * Signal frame handlers.
 418  */
 419
 420 static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
 421 {
 422         struct task_struct *tsk = current;
 423         struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 424
 425         unlazy_fpu(tsk);
 426         fp->status = fp->swd;
 427         if (__copy_to_user(buf, fp, sizeof(struct i387_fsave_struct)))
 428                 return -1;
 429         return 1;
 430 }
 431
 432 static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
 433 {
 434         struct task_struct *tsk = current;
 435         struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 436         struct user_i387_ia32_struct env;
 437         int err = 0;
 438
 439         unlazy_fpu(tsk);
 440
 441         convert_from_fxsr(&env, tsk);
 442         if (__copy_to_user(buf, &env, sizeof(env)))
 443                 return -1;
 444
 445         err |= __put_user(fx->swd, &buf->status);
 446         err |= __put_user(X86_FXSR_MAGIC, &buf->magic);
 447         if (err)
 448                 return -1;
 449
 450         if (__copy_to_user(&buf->_fxsr_env[0], fx,
 451                            sizeof(struct i387_fxsave_struct)))
 452                 return -1;
 453         return 1;
 454 }
 455
 456 int save_i387_ia32(struct _fpstate_ia32 __user *buf)
 457 {
 458         if (!used_math())
 459                 return 0;
 460         /*
 461          * This will cause a "finit" to be triggered by the next
 462          * attempted FPU operation by the 'current' process.
 463          */
 464         clear_used_math();
 465
 466         if (!HAVE_HWFP) {
 467                 return fpregs_soft_get(current, NULL,
 468                                        0, sizeof(struct user_i387_ia32_struct),
 469                                        NULL, buf) ? -1 : 1;
 470         }
 471
 472         if (cpu_has_fxsr)
 473                 return save_i387_fxsave(buf);
 474         else
 475                 return save_i387_fsave(buf);
 476 }
 477
 478 static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
 479 {
 480         struct task_struct *tsk = current;
 481
 482         return __copy_from_user(&tsk->thread.xstate->fsave, buf,
 483                                 sizeof(struct i387_fsave_struct));
 484 }
 485
 486 static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf)
 487 {
 488         struct task_struct *tsk = current;
 489         struct user_i387_ia32_struct env;
 490         int err;
 491
 492         err = __copy_from_user(&tsk->thread.xstate->fxsave, &buf->_fxsr_env[0],
 493                                sizeof(struct i387_fxsave_struct));
 494         /* mxcsr reserved bits must be masked to zero for security reasons */
 495         tsk->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 496         if (err || __copy_from_user(&env, buf, sizeof(env)))
 497                 return 1;
 498         convert_to_fxsr(tsk, &env);
 499
 500         return 0;
 501 }
 502
 503 int restore_i387_ia32(struct _fpstate_ia32 __user *buf)
 504 {
 505         int err;
 506         struct task_struct *tsk = current;
 507
 508         if (HAVE_HWFP)
 509                 clear_fpu(tsk);
 510
 511         if (!used_math()) {
 512                 err = init_fpu(tsk);
 513                 if (err)
 514                         return err;
 515         }
 516
 517         if (HAVE_HWFP) {
 518                 if (cpu_has_fxsr)
 519                         err = restore_i387_fxsave(buf);
 520                 else
 521                         err = restore_i387_fsave(buf);
 522         } else {
 523                 err = fpregs_soft_set(current, NULL,
 524                                       0, sizeof(struct user_i387_ia32_struct),
 525                                       NULL, buf) != 0;
 526         }
 527         set_used_math();
 528
 529         return err;
 530 }
 531
 532 /*
 533  * FPU state for core dumps.
 534  * This is only used for a.out dumps now.
 535  * It is declared generically using elf_fpregset_t (which is
 536  * struct user_i387_struct) but is in fact only used for 32-bit
 537  * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
 538  */
 539 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
 540 {
 541         struct task_struct *tsk = current;
 542         int fpvalid;
 543
 544         fpvalid = !!used_math();
 545         if (fpvalid)
 546                 fpvalid = !fpregs_get(tsk, NULL,
 547                                       0, sizeof(struct user_i387_ia32_struct),
 548                                       fpu, NULL);
 549
 550         return fpvalid;
 551 }
 552 EXPORT_SYMBOL(dump_fpu);
 553
 554 #endif  /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */