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 sig_xstate_ia32_size   sig_xstate_size
  28 # define user_i387_ia32_struct  user_i387_struct
  29 # define user32_fxsr_struct     user_fxsr_struct
  30 #endif
  31
  32 #ifdef CONFIG_MATH_EMULATION
  33 # define HAVE_HWFP              (boot_cpu_data.hard_math)
  34 #else
  35 # define HAVE_HWFP              1
  36 #endif
  37
  38 static unsigned int             mxcsr_feature_mask __read_mostly = 0xffffffffu;
  39 unsigned int xstate_size;
  40 unsigned int sig_xstate_ia32_size = sizeof(struct _fpstate_ia32);
  41 static struct i387_fxsave_struct fx_scratch __cpuinitdata;
  42
  43 void __cpuinit mxcsr_feature_mask_init(void)
  44 {
  45         unsigned long mask = 0;
  46
  47         clts();
  48         if (cpu_has_fxsr) {
  49                 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
  50                 asm volatile("fxsave %0" : : "m" (fx_scratch));
  51                 mask = fx_scratch.mxcsr_mask;
  52                 if (mask == 0)
  53                         mask = 0x0000ffbf;
  54         }
  55         mxcsr_feature_mask &= mask;
  56         stts();
  57 }
  58
  59 void __init init_thread_xstate(void)
  60 {
  61         if (!HAVE_HWFP) {
  62                 xstate_size = sizeof(struct i387_soft_struct);
  63                 return;
  64         }
  65
  66         if (cpu_has_xsave) {
  67                 xsave_cntxt_init();
  68                 return;
  69         }
  70
  71         if (cpu_has_fxsr)
  72                 xstate_size = sizeof(struct i387_fxsave_struct);
  73 #ifdef CONFIG_X86_32
  74         else
  75                 xstate_size = sizeof(struct i387_fsave_struct);
  76 #endif
  77 }
  78
  79 #ifdef CONFIG_X86_64
  80 /*
  81  * Called at bootup to set up the initial FPU state that is later cloned
  82  * into all processes.
  83  */
  84 void __cpuinit fpu_init(void)
  85 {
  86         unsigned long oldcr0 = read_cr0();
  87
  88         set_in_cr4(X86_CR4_OSFXSR);
  89         set_in_cr4(X86_CR4_OSXMMEXCPT);
  90
  91         write_cr0(oldcr0 & ~(X86_CR0_TS|X86_CR0_EM)); /* clear TS and EM */
  92
  93         /*
  94          * Boot processor to setup the FP and extended state context info.
  95          */
  96         if (!smp_processor_id())
  97                 init_thread_xstate();
  98         xsave_init();
  99
 100         mxcsr_feature_mask_init();
 101         /* clean state in init */
 102         if (cpu_has_xsave)
 103                 current_thread_info()->status = TS_XSAVE;
 104         else
 105                 current_thread_info()->status = 0;
 106         clear_used_math();
 107 }
 108 #endif  /* CONFIG_X86_64 */
 109
 110 /*
 111  * The _current_ task is using the FPU for the first time
 112  * so initialize it and set the mxcsr to its default
 113  * value at reset if we support XMM instructions and then
 114  * remeber the current task has used the FPU.
 115  */
 116 int init_fpu(struct task_struct *tsk)
 117 {
 118         if (tsk_used_math(tsk)) {
 119                 if (HAVE_HWFP && tsk == current)
 120                         unlazy_fpu(tsk);
 121                 return 0;
 122         }
 123
 124         /*
 125          * Memory allocation at the first usage of the FPU and other state.
 126          */
 127         if (!tsk->thread.xstate) {
 128                 tsk->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
 129                                                       GFP_KERNEL);
 130                 if (!tsk->thread.xstate)
 131                         return -ENOMEM;
 132         }
 133
 134 #ifdef CONFIG_X86_32
 135         if (!HAVE_HWFP) {
 136                 memset(tsk->thread.xstate, 0, xstate_size);
 137                 finit();
 138                 set_stopped_child_used_math(tsk);
 139                 return 0;
 140         }
 141 #endif
 142
 143         if (cpu_has_fxsr) {
 144                 struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 145
 146                 memset(fx, 0, xstate_size);
 147                 fx->cwd = 0x37f;
 148                 if (cpu_has_xmm)
 149                         fx->mxcsr = MXCSR_DEFAULT;
 150         } else {
 151                 struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 152                 memset(fp, 0, xstate_size);
 153                 fp->cwd = 0xffff037fu;
 154                 fp->swd = 0xffff0000u;
 155                 fp->twd = 0xffffffffu;
 156                 fp->fos = 0xffff0000u;
 157         }
 158         /*
 159          * Only the device not available exception or ptrace can call init_fpu.
 160          */
 161         set_stopped_child_used_math(tsk);
 162         return 0;
 163 }
 164
 165 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
 166 {
 167         return tsk_used_math(target) ? regset->n : 0;
 168 }
 169
 170 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
 171 {
 172         return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
 173 }
 174
 175 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
 176                 unsigned int pos, unsigned int count,
 177                 void *kbuf, void __user *ubuf)
 178 {
 179         int ret;
 180
 181         if (!cpu_has_fxsr)
 182                 return -ENODEV;
 183
 184         ret = init_fpu(target);
 185         if (ret)
 186                 return ret;
 187
 188         return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 189                                    &target->thread.xstate->fxsave, 0, -1);
 190 }
 191
 192 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
 193                 unsigned int pos, unsigned int count,
 194                 const void *kbuf, const void __user *ubuf)
 195 {
 196         int ret;
 197
 198         if (!cpu_has_fxsr)
 199                 return -ENODEV;
 200
 201         ret = init_fpu(target);
 202         if (ret)
 203                 return ret;
 204
 205         set_stopped_child_used_math(target);
 206
 207         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 208                                  &target->thread.xstate->fxsave, 0, -1);
 209
 210         /*
 211          * mxcsr reserved bits must be masked to zero for security reasons.
 212          */
 213         target->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 214
 215         return ret;
 216 }
 217
 218 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
 219
 220 /*
 221  * FPU tag word conversions.
 222  */
 223
 224 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
 225 {
 226         unsigned int tmp; /* to avoid 16 bit prefixes in the code */
 227
 228         /* Transform each pair of bits into 01 (valid) or 00 (empty) */
 229         tmp = ~twd;
 230         tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
 231         /* and move the valid bits to the lower byte. */
 232         tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
 233         tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
 234         tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
 235
 236         return tmp;
 237 }
 238
 239 #define FPREG_ADDR(f, n)        ((void *)&(f)->st_space + (n) * 16);
 240 #define FP_EXP_TAG_VALID        0
 241 #define FP_EXP_TAG_ZERO         1
 242 #define FP_EXP_TAG_SPECIAL      2
 243 #define FP_EXP_TAG_EMPTY        3
 244
 245 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
 246 {
 247         struct _fpxreg *st;
 248         u32 tos = (fxsave->swd >> 11) & 7;
 249         u32 twd = (unsigned long) fxsave->twd;
 250         u32 tag;
 251         u32 ret = 0xffff0000u;
 252         int i;
 253
 254         for (i = 0; i < 8; i++, twd >>= 1) {
 255                 if (twd & 0x1) {
 256                         st = FPREG_ADDR(fxsave, (i - tos) & 7);
 257
 258                         switch (st->exponent & 0x7fff) {
 259                         case 0x7fff:
 260                                 tag = FP_EXP_TAG_SPECIAL;
 261                                 break;
 262                         case 0x0000:
 263                                 if (!st->significand[0] &&
 264                                     !st->significand[1] &&
 265                                     !st->significand[2] &&
 266                                     !st->significand[3])
 267                                         tag = FP_EXP_TAG_ZERO;
 268                                 else
 269                                         tag = FP_EXP_TAG_SPECIAL;
 270                                 break;
 271                         default:
 272                                 if (st->significand[3] & 0x8000)
 273                                         tag = FP_EXP_TAG_VALID;
 274                                 else
 275                                         tag = FP_EXP_TAG_SPECIAL;
 276                                 break;
 277                         }
 278                 } else {
 279                         tag = FP_EXP_TAG_EMPTY;
 280                 }
 281                 ret |= tag << (2 * i);
 282         }
 283         return ret;
 284 }
 285
 286 /*
 287  * FXSR floating point environment conversions.
 288  */
 289
 290 static void
 291 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
 292 {
 293         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 294         struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
 295         struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
 296         int i;
 297
 298         env->cwd = fxsave->cwd | 0xffff0000u;
 299         env->swd = fxsave->swd | 0xffff0000u;
 300         env->twd = twd_fxsr_to_i387(fxsave);
 301
 302 #ifdef CONFIG_X86_64
 303         env->fip = fxsave->rip;
 304         env->foo = fxsave->rdp;
 305         if (tsk == current) {
 306                 /*
 307                  * should be actually ds/cs at fpu exception time, but
 308                  * that information is not available in 64bit mode.
 309                  */
 310                 asm("mov %%ds, %[fos]" : [fos] "=r" (env->fos));
 311                 asm("mov %%cs, %[fcs]" : [fcs] "=r" (env->fcs));
 312         } else {
 313                 struct pt_regs *regs = task_pt_regs(tsk);
 314
 315                 env->fos = 0xffff0000 | tsk->thread.ds;
 316                 env->fcs = regs->cs;
 317         }
 318 #else
 319         env->fip = fxsave->fip;
 320         env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
 321         env->foo = fxsave->foo;
 322         env->fos = fxsave->fos;
 323 #endif
 324
 325         for (i = 0; i < 8; ++i)
 326                 memcpy(&to[i], &from[i], sizeof(to[0]));
 327 }
 328
 329 static void convert_to_fxsr(struct task_struct *tsk,
 330                             const struct user_i387_ia32_struct *env)
 331
 332 {
 333         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 334         struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
 335         struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
 336         int i;
 337
 338         fxsave->cwd = env->cwd;
 339         fxsave->swd = env->swd;
 340         fxsave->twd = twd_i387_to_fxsr(env->twd);
 341         fxsave->fop = (u16) ((u32) env->fcs >> 16);
 342 #ifdef CONFIG_X86_64
 343         fxsave->rip = env->fip;
 344         fxsave->rdp = env->foo;
 345         /* cs and ds ignored */
 346 #else
 347         fxsave->fip = env->fip;
 348         fxsave->fcs = (env->fcs & 0xffff);
 349         fxsave->foo = env->foo;
 350         fxsave->fos = env->fos;
 351 #endif
 352
 353         for (i = 0; i < 8; ++i)
 354                 memcpy(&to[i], &from[i], sizeof(from[0]));
 355 }
 356
 357 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
 358                unsigned int pos, unsigned int count,
 359                void *kbuf, void __user *ubuf)
 360 {
 361         struct user_i387_ia32_struct env;
 362         int ret;
 363
 364         ret = init_fpu(target);
 365         if (ret)
 366                 return ret;
 367
 368         if (!HAVE_HWFP)
 369                 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
 370
 371         if (!cpu_has_fxsr) {
 372                 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 373                                            &target->thread.xstate->fsave, 0,
 374                                            -1);
 375         }
 376
 377         if (kbuf && pos == 0 && count == sizeof(env)) {
 378                 convert_from_fxsr(kbuf, target);
 379                 return 0;
 380         }
 381
 382         convert_from_fxsr(&env, target);
 383
 384         return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 385 }
 386
 387 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
 388                unsigned int pos, unsigned int count,
 389                const void *kbuf, const void __user *ubuf)
 390 {
 391         struct user_i387_ia32_struct env;
 392         int ret;
 393
 394         ret = init_fpu(target);
 395         if (ret)
 396                 return ret;
 397
 398         set_stopped_child_used_math(target);
 399
 400         if (!HAVE_HWFP)
 401                 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
 402
 403         if (!cpu_has_fxsr) {
 404                 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 405                                           &target->thread.xstate->fsave, 0, -1);
 406         }
 407
 408         if (pos > 0 || count < sizeof(env))
 409                 convert_from_fxsr(&env, target);
 410
 411         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 412         if (!ret)
 413                 convert_to_fxsr(target, &env);
 414
 415         return ret;
 416 }
 417
 418 /*
 419  * Signal frame handlers.
 420  */
 421
 422 static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
 423 {
 424         struct task_struct *tsk = current;
 425         struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 426
 427         unlazy_fpu(tsk);
 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         unlazy_fpu(tsk);
 442
 443         convert_from_fxsr(&env, tsk);
 444         if (__copy_to_user(buf, &env, sizeof(env)))
 445                 return -1;
 446
 447         err |= __put_user(fx->swd, &buf->status);
 448         err |= __put_user(X86_FXSR_MAGIC, &buf->magic);
 449         if (err)
 450                 return -1;
 451
 452         if (__copy_to_user(&buf->_fxsr_env[0], fx,
 453                            sizeof(struct i387_fxsave_struct)))
 454                 return -1;
 455         return 1;
 456 }
 457
 458 int save_i387_ia32(struct _fpstate_ia32 __user *buf)
 459 {
 460         if (!used_math())
 461                 return 0;
 462         /*
 463          * This will cause a "finit" to be triggered by the next
 464          * attempted FPU operation by the 'current' process.
 465          */
 466         clear_used_math();
 467
 468         if (!HAVE_HWFP) {
 469                 return fpregs_soft_get(current, NULL,
 470                                        0, sizeof(struct user_i387_ia32_struct),
 471                                        NULL, buf) ? -1 : 1;
 472         }
 473
 474         if (cpu_has_fxsr)
 475                 return save_i387_fxsave(buf);
 476         else
 477                 return save_i387_fsave(buf);
 478 }
 479
 480 static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
 481 {
 482         struct task_struct *tsk = current;
 483
 484         return __copy_from_user(&tsk->thread.xstate->fsave, buf,
 485                                 sizeof(struct i387_fsave_struct));
 486 }
 487
 488 static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf)
 489 {
 490         struct task_struct *tsk = current;
 491         struct user_i387_ia32_struct env;
 492         int err;
 493
 494         err = __copy_from_user(&tsk->thread.xstate->fxsave, &buf->_fxsr_env[0],
 495                                sizeof(struct i387_fxsave_struct));
 496         /* mxcsr reserved bits must be masked to zero for security reasons */
 497         tsk->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 498         if (err || __copy_from_user(&env, buf, sizeof(env)))
 499                 return 1;
 500         convert_to_fxsr(tsk, &env);
 501
 502         return 0;
 503 }
 504
 505 int restore_i387_ia32(struct _fpstate_ia32 __user *buf)
 506 {
 507         int err;
 508         struct task_struct *tsk = current;
 509
 510         if (HAVE_HWFP)
 511                 clear_fpu(tsk);
 512
 513         if (!used_math()) {
 514                 err = init_fpu(tsk);
 515                 if (err)
 516                         return err;
 517         }
 518
 519         if (HAVE_HWFP) {
 520                 if (cpu_has_fxsr)
 521                         err = restore_i387_fxsave(buf);
 522                 else
 523                         err = restore_i387_fsave(buf);
 524         } else {
 525                 err = fpregs_soft_set(current, NULL,
 526                                       0, sizeof(struct user_i387_ia32_struct),
 527                                       NULL, buf) != 0;
 528         }
 529         set_used_math();
 530
 531         return err;
 532 }
 533
 534 /*
 535  * FPU state for core dumps.
 536  * This is only used for a.out dumps now.
 537  * It is declared generically using elf_fpregset_t (which is
 538  * struct user_i387_struct) but is in fact only used for 32-bit
 539  * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
 540  */
 541 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
 542 {
 543         struct task_struct *tsk = current;
 544         int fpvalid;
 545
 546         fpvalid = !!used_math();
 547         if (fpvalid)
 548                 fpvalid = !fpregs_get(tsk, NULL,
 549                                       0, sizeof(struct user_i387_ia32_struct),
 550                                       fpu, NULL);
 551
 552         return fpvalid;
 553 }
 554 EXPORT_SYMBOL(dump_fpu);
 555
 556 #endif  /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */