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_xstate_ia32          save_i387_xstate
  25 # define restore_i387_xstate_ia32       restore_i387_xstate
  26 # define _fpstate_ia32          _fpstate
  27 # define _xstate_ia32           _xstate
  28 # define sig_xstate_ia32_size   sig_xstate_size
  29 # define user_i387_ia32_struct  user_i387_struct
  30 # define user32_fxsr_struct     user_fxsr_struct
  31 #endif
  32
  33 #ifdef CONFIG_MATH_EMULATION
  34 # define HAVE_HWFP              (boot_cpu_data.hard_math)
  35 #else
  36 # define HAVE_HWFP              1
  37 #endif
  38
  39 static unsigned int             mxcsr_feature_mask __read_mostly = 0xffffffffu;
  40 unsigned int xstate_size;
  41 unsigned int sig_xstate_ia32_size = sizeof(struct _fpstate_ia32);
  42 static struct i387_fxsave_struct fx_scratch __cpuinitdata;
  43
  44 void __cpuinit mxcsr_feature_mask_init(void)
  45 {
  46         unsigned long mask = 0;
  47
  48         clts();
  49         if (cpu_has_fxsr) {
  50                 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
  51                 asm volatile("fxsave %0" : : "m" (fx_scratch));
  52                 mask = fx_scratch.mxcsr_mask;
  53                 if (mask == 0)
  54                         mask = 0x0000ffbf;
  55         }
  56         mxcsr_feature_mask &= mask;
  57         stts();
  58 }
  59
  60 void __init init_thread_xstate(void)
  61 {
  62         if (!HAVE_HWFP) {
  63                 xstate_size = sizeof(struct i387_soft_struct);
  64                 return;
  65         }
  66
  67         if (cpu_has_xsave) {
  68                 xsave_cntxt_init();
  69                 return;
  70         }
  71
  72         if (cpu_has_fxsr)
  73                 xstate_size = sizeof(struct i387_fxsave_struct);
  74 #ifdef CONFIG_X86_32
  75         else
  76                 xstate_size = sizeof(struct i387_fsave_struct);
  77 #endif
  78 }
  79
  80 #ifdef CONFIG_X86_64
  81 /*
  82  * Called at bootup to set up the initial FPU state that is later cloned
  83  * into all processes.
  84  */
  85 void __cpuinit fpu_init(void)
  86 {
  87         unsigned long oldcr0 = read_cr0();
  88
  89         set_in_cr4(X86_CR4_OSFXSR);
  90         set_in_cr4(X86_CR4_OSXMMEXCPT);
  91
  92         write_cr0(oldcr0 & ~(X86_CR0_TS|X86_CR0_EM)); /* clear TS and EM */
  93
  94         /*
  95          * Boot processor to setup the FP and extended state context info.
  96          */
  97         if (!smp_processor_id())
  98                 init_thread_xstate();
  99         xsave_init();
 100
 101         mxcsr_feature_mask_init();
 102         /* clean state in init */
 103         if (cpu_has_xsave)
 104                 current_thread_info()->status = TS_XSAVE;
 105         else
 106                 current_thread_info()->status = 0;
 107         clear_used_math();
 108 }
 109 #endif  /* CONFIG_X86_64 */
 110
 111 /*
 112  * The _current_ task is using the FPU for the first time
 113  * so initialize it and set the mxcsr to its default
 114  * value at reset if we support XMM instructions and then
 115  * remeber the current task has used the FPU.
 116  */
 117 int init_fpu(struct task_struct *tsk)
 118 {
 119         if (tsk_used_math(tsk)) {
 120                 if (HAVE_HWFP && tsk == current)
 121                         unlazy_fpu(tsk);
 122                 return 0;
 123         }
 124
 125         /*
 126          * Memory allocation at the first usage of the FPU and other state.
 127          */
 128         if (!tsk->thread.xstate) {
 129                 tsk->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
 130                                                       GFP_KERNEL);
 131                 if (!tsk->thread.xstate)
 132                         return -ENOMEM;
 133         }
 134
 135 #ifdef CONFIG_X86_32
 136         if (!HAVE_HWFP) {
 137                 memset(tsk->thread.xstate, 0, xstate_size);
 138                 finit();
 139                 set_stopped_child_used_math(tsk);
 140                 return 0;
 141         }
 142 #endif
 143
 144         if (cpu_has_fxsr) {
 145                 struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 146
 147                 memset(fx, 0, xstate_size);
 148                 fx->cwd = 0x37f;
 149                 if (cpu_has_xmm)
 150                         fx->mxcsr = MXCSR_DEFAULT;
 151         } else {
 152                 struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 153                 memset(fp, 0, xstate_size);
 154                 fp->cwd = 0xffff037fu;
 155                 fp->swd = 0xffff0000u;
 156                 fp->twd = 0xffffffffu;
 157                 fp->fos = 0xffff0000u;
 158         }
 159         /*
 160          * Only the device not available exception or ptrace can call init_fpu.
 161          */
 162         set_stopped_child_used_math(tsk);
 163         return 0;
 164 }
 165
 166 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
 167 {
 168         return tsk_used_math(target) ? regset->n : 0;
 169 }
 170
 171 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
 172 {
 173         return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
 174 }
 175
 176 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
 177                 unsigned int pos, unsigned int count,
 178                 void *kbuf, void __user *ubuf)
 179 {
 180         int ret;
 181
 182         if (!cpu_has_fxsr)
 183                 return -ENODEV;
 184
 185         ret = init_fpu(target);
 186         if (ret)
 187                 return ret;
 188
 189         return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 190                                    &target->thread.xstate->fxsave, 0, -1);
 191 }
 192
 193 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
 194                 unsigned int pos, unsigned int count,
 195                 const void *kbuf, const void __user *ubuf)
 196 {
 197         int ret;
 198
 199         if (!cpu_has_fxsr)
 200                 return -ENODEV;
 201
 202         ret = init_fpu(target);
 203         if (ret)
 204                 return ret;
 205
 206         set_stopped_child_used_math(target);
 207
 208         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 209                                  &target->thread.xstate->fxsave, 0, -1);
 210
 211         /*
 212          * mxcsr reserved bits must be masked to zero for security reasons.
 213          */
 214         target->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 215
 216         return ret;
 217 }
 218
 219 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
 220
 221 /*
 222  * FPU tag word conversions.
 223  */
 224
 225 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
 226 {
 227         unsigned int tmp; /* to avoid 16 bit prefixes in the code */
 228
 229         /* Transform each pair of bits into 01 (valid) or 00 (empty) */
 230         tmp = ~twd;
 231         tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
 232         /* and move the valid bits to the lower byte. */
 233         tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
 234         tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
 235         tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
 236
 237         return tmp;
 238 }
 239
 240 #define FPREG_ADDR(f, n)        ((void *)&(f)->st_space + (n) * 16);
 241 #define FP_EXP_TAG_VALID        0
 242 #define FP_EXP_TAG_ZERO         1
 243 #define FP_EXP_TAG_SPECIAL      2
 244 #define FP_EXP_TAG_EMPTY        3
 245
 246 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
 247 {
 248         struct _fpxreg *st;
 249         u32 tos = (fxsave->swd >> 11) & 7;
 250         u32 twd = (unsigned long) fxsave->twd;
 251         u32 tag;
 252         u32 ret = 0xffff0000u;
 253         int i;
 254
 255         for (i = 0; i < 8; i++, twd >>= 1) {
 256                 if (twd & 0x1) {
 257                         st = FPREG_ADDR(fxsave, (i - tos) & 7);
 258
 259                         switch (st->exponent & 0x7fff) {
 260                         case 0x7fff:
 261                                 tag = FP_EXP_TAG_SPECIAL;
 262                                 break;
 263                         case 0x0000:
 264                                 if (!st->significand[0] &&
 265                                     !st->significand[1] &&
 266                                     !st->significand[2] &&
 267                                     !st->significand[3])
 268                                         tag = FP_EXP_TAG_ZERO;
 269                                 else
 270                                         tag = FP_EXP_TAG_SPECIAL;
 271                                 break;
 272                         default:
 273                                 if (st->significand[3] & 0x8000)
 274                                         tag = FP_EXP_TAG_VALID;
 275                                 else
 276                                         tag = FP_EXP_TAG_SPECIAL;
 277                                 break;
 278                         }
 279                 } else {
 280                         tag = FP_EXP_TAG_EMPTY;
 281                 }
 282                 ret |= tag << (2 * i);
 283         }
 284         return ret;
 285 }
 286
 287 /*
 288  * FXSR floating point environment conversions.
 289  */
 290
 291 static void
 292 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
 293 {
 294         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 295         struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
 296         struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
 297         int i;
 298
 299         env->cwd = fxsave->cwd | 0xffff0000u;
 300         env->swd = fxsave->swd | 0xffff0000u;
 301         env->twd = twd_fxsr_to_i387(fxsave);
 302
 303 #ifdef CONFIG_X86_64
 304         env->fip = fxsave->rip;
 305         env->foo = fxsave->rdp;
 306         if (tsk == current) {
 307                 /*
 308                  * should be actually ds/cs at fpu exception time, but
 309                  * that information is not available in 64bit mode.
 310                  */
 311                 asm("mov %%ds, %[fos]" : [fos] "=r" (env->fos));
 312                 asm("mov %%cs, %[fcs]" : [fcs] "=r" (env->fcs));
 313         } else {
 314                 struct pt_regs *regs = task_pt_regs(tsk);
 315
 316                 env->fos = 0xffff0000 | tsk->thread.ds;
 317                 env->fcs = regs->cs;
 318         }
 319 #else
 320         env->fip = fxsave->fip;
 321         env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
 322         env->foo = fxsave->foo;
 323         env->fos = fxsave->fos;
 324 #endif
 325
 326         for (i = 0; i < 8; ++i)
 327                 memcpy(&to[i], &from[i], sizeof(to[0]));
 328 }
 329
 330 static void convert_to_fxsr(struct task_struct *tsk,
 331                             const struct user_i387_ia32_struct *env)
 332
 333 {
 334         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 335         struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
 336         struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
 337         int i;
 338
 339         fxsave->cwd = env->cwd;
 340         fxsave->swd = env->swd;
 341         fxsave->twd = twd_i387_to_fxsr(env->twd);
 342         fxsave->fop = (u16) ((u32) env->fcs >> 16);
 343 #ifdef CONFIG_X86_64
 344         fxsave->rip = env->fip;
 345         fxsave->rdp = env->foo;
 346         /* cs and ds ignored */
 347 #else
 348         fxsave->fip = env->fip;
 349         fxsave->fcs = (env->fcs & 0xffff);
 350         fxsave->foo = env->foo;
 351         fxsave->fos = env->fos;
 352 #endif
 353
 354         for (i = 0; i < 8; ++i)
 355                 memcpy(&to[i], &from[i], sizeof(from[0]));
 356 }
 357
 358 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
 359                unsigned int pos, unsigned int count,
 360                void *kbuf, void __user *ubuf)
 361 {
 362         struct user_i387_ia32_struct env;
 363         int ret;
 364
 365         ret = init_fpu(target);
 366         if (ret)
 367                 return ret;
 368
 369         if (!HAVE_HWFP)
 370                 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
 371
 372         if (!cpu_has_fxsr) {
 373                 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 374                                            &target->thread.xstate->fsave, 0,
 375                                            -1);
 376         }
 377
 378         if (kbuf && pos == 0 && count == sizeof(env)) {
 379                 convert_from_fxsr(kbuf, target);
 380                 return 0;
 381         }
 382
 383         convert_from_fxsr(&env, target);
 384
 385         return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 386 }
 387
 388 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
 389                unsigned int pos, unsigned int count,
 390                const void *kbuf, const void __user *ubuf)
 391 {
 392         struct user_i387_ia32_struct env;
 393         int ret;
 394
 395         ret = init_fpu(target);
 396         if (ret)
 397                 return ret;
 398
 399         set_stopped_child_used_math(target);
 400
 401         if (!HAVE_HWFP)
 402                 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
 403
 404         if (!cpu_has_fxsr) {
 405                 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 406                                           &target->thread.xstate->fsave, 0, -1);
 407         }
 408
 409         if (pos > 0 || count < sizeof(env))
 410                 convert_from_fxsr(&env, target);
 411
 412         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 413         if (!ret)
 414                 convert_to_fxsr(target, &env);
 415
 416         return ret;
 417 }
 418
 419 /*
 420  * Signal frame handlers.
 421  */
 422
 423 static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
 424 {
 425         struct task_struct *tsk = current;
 426         struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 427
 428         fp->status = fp->swd;
 429         if (__copy_to_user(buf, fp, sizeof(struct i387_fsave_struct)))
 430                 return -1;
 431         return 1;
 432 }
 433
 434 static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
 435 {
 436         struct task_struct *tsk = current;
 437         struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 438         struct user_i387_ia32_struct env;
 439         int err = 0;
 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_xstate_ia32(void __user *buf)
 457 {
 458         struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
 459         struct task_struct *tsk = current;
 460
 461         if (!used_math())
 462                 return 0;
 463
 464         if (!access_ok(VERIFY_WRITE, buf, sig_xstate_ia32_size))
 465                 return -EACCES;
 466         /*
 467          * This will cause a "finit" to be triggered by the next
 468          * attempted FPU operation by the 'current' process.
 469          */
 470         clear_used_math();
 471
 472         if (!HAVE_HWFP) {
 473                 return fpregs_soft_get(current, NULL,
 474                                        0, sizeof(struct user_i387_ia32_struct),
 475                                        NULL, fp) ? -1 : 1;
 476         }
 477
 478         unlazy_fpu(tsk);
 479
 480         if (cpu_has_fxsr)
 481                 return save_i387_fxsave(fp);
 482         else
 483                 return save_i387_fsave(fp);
 484 }
 485
 486 static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
 487 {
 488         struct task_struct *tsk = current;
 489
 490         return __copy_from_user(&tsk->thread.xstate->fsave, buf,
 491                                 sizeof(struct i387_fsave_struct));
 492 }
 493
 494 static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf)
 495 {
 496         struct task_struct *tsk = current;
 497         struct user_i387_ia32_struct env;
 498         int err;
 499
 500         err = __copy_from_user(&tsk->thread.xstate->fxsave, &buf->_fxsr_env[0],
 501                                sizeof(struct i387_fxsave_struct));
 502         /* mxcsr reserved bits must be masked to zero for security reasons */
 503         tsk->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 504         if (err || __copy_from_user(&env, buf, sizeof(env)))
 505                 return 1;
 506         convert_to_fxsr(tsk, &env);
 507
 508         return 0;
 509 }
 510
 511 int restore_i387_xstate_ia32(void __user *buf)
 512 {
 513         int err;
 514         struct task_struct *tsk = current;
 515         struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
 516
 517         if (HAVE_HWFP)
 518                 clear_fpu(tsk);
 519
 520         if (!buf) {
 521                 if (used_math()) {
 522                         clear_fpu(tsk);
 523                         clear_used_math();
 524                 }
 525
 526                 return 0;
 527         } else
 528                 if (!access_ok(VERIFY_READ, buf, sig_xstate_ia32_size))
 529                         return -EACCES;
 530
 531         if (!used_math()) {
 532                 err = init_fpu(tsk);
 533                 if (err)
 534                         return err;
 535         }
 536
 537         if (HAVE_HWFP) {
 538                 if (cpu_has_fxsr)
 539                         err = restore_i387_fxsave(fp);
 540                 else
 541                         err = restore_i387_fsave(fp);
 542         } else {
 543                 err = fpregs_soft_set(current, NULL,
 544                                       0, sizeof(struct user_i387_ia32_struct),
 545                                       NULL, fp) != 0;
 546         }
 547         set_used_math();
 548
 549         return err;
 550 }
 551
 552 /*
 553  * FPU state for core dumps.
 554  * This is only used for a.out dumps now.
 555  * It is declared generically using elf_fpregset_t (which is
 556  * struct user_i387_struct) but is in fact only used for 32-bit
 557  * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
 558  */
 559 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
 560 {
 561         struct task_struct *tsk = current;
 562         int fpvalid;
 563
 564         fpvalid = !!used_math();
 565         if (fpvalid)
 566                 fpvalid = !fpregs_get(tsk, NULL,
 567                                       0, sizeof(struct user_i387_ia32_struct),
 568                                       fpu, NULL);
 569
 570         return fpvalid;
 571 }
 572 EXPORT_SYMBOL(dump_fpu);
 573
 574 #endif  /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */