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 fx_sw_reserved_ia32    fx_sw_reserved
  30 # define user_i387_ia32_struct  user_i387_struct
  31 # define user32_fxsr_struct     user_fxsr_struct
  32 #endif
  33
  34 #ifdef CONFIG_MATH_EMULATION
  35 # define HAVE_HWFP              (boot_cpu_data.hard_math)
  36 #else
  37 # define HAVE_HWFP              1
  38 #endif
  39
  40 static unsigned int             mxcsr_feature_mask __read_mostly = 0xffffffffu;
  41 unsigned int xstate_size;
  42 unsigned int sig_xstate_ia32_size = sizeof(struct _fpstate_ia32);
  43 static struct i387_fxsave_struct fx_scratch __cpuinitdata;
  44
  45 void __cpuinit mxcsr_feature_mask_init(void)
  46 {
  47         unsigned long mask = 0;
  48
  49         clts();
  50         if (cpu_has_fxsr) {
  51                 memset(&fx_scratch, 0, sizeof(struct i387_fxsave_struct));
  52                 asm volatile("fxsave %0" : : "m" (fx_scratch));
  53                 mask = fx_scratch.mxcsr_mask;
  54                 if (mask == 0)
  55                         mask = 0x0000ffbf;
  56         }
  57         mxcsr_feature_mask &= mask;
  58         stts();
  59 }
  60
  61 void __init init_thread_xstate(void)
  62 {
  63         if (!HAVE_HWFP) {
  64                 xstate_size = sizeof(struct i387_soft_struct);
  65                 return;
  66         }
  67
  68         if (cpu_has_xsave) {
  69                 xsave_cntxt_init();
  70                 return;
  71         }
  72
  73         if (cpu_has_fxsr)
  74                 xstate_size = sizeof(struct i387_fxsave_struct);
  75 #ifdef CONFIG_X86_32
  76         else
  77                 xstate_size = sizeof(struct i387_fsave_struct);
  78 #endif
  79 }
  80
  81 #ifdef CONFIG_X86_64
  82 /*
  83  * Called at bootup to set up the initial FPU state that is later cloned
  84  * into all processes.
  85  */
  86 void __cpuinit fpu_init(void)
  87 {
  88         unsigned long oldcr0 = read_cr0();
  89
  90         set_in_cr4(X86_CR4_OSFXSR);
  91         set_in_cr4(X86_CR4_OSXMMEXCPT);
  92
  93         write_cr0(oldcr0 & ~(X86_CR0_TS|X86_CR0_EM)); /* clear TS and EM */
  94
  95         /*
  96          * Boot processor to setup the FP and extended state context info.
  97          */
  98         if (!smp_processor_id())
  99                 init_thread_xstate();
 100         xsave_init();
 101
 102         mxcsr_feature_mask_init();
 103         /* clean state in init */
 104         if (cpu_has_xsave)
 105                 current_thread_info()->status = TS_XSAVE;
 106         else
 107                 current_thread_info()->status = 0;
 108         clear_used_math();
 109 }
 110 #endif  /* CONFIG_X86_64 */
 111
 112 /*
 113  * The _current_ task is using the FPU for the first time
 114  * so initialize it and set the mxcsr to its default
 115  * value at reset if we support XMM instructions and then
 116  * remeber the current task has used the FPU.
 117  */
 118 int init_fpu(struct task_struct *tsk)
 119 {
 120         if (tsk_used_math(tsk)) {
 121                 if (HAVE_HWFP && tsk == current)
 122                         unlazy_fpu(tsk);
 123                 return 0;
 124         }
 125
 126         /*
 127          * Memory allocation at the first usage of the FPU and other state.
 128          */
 129         if (!tsk->thread.xstate) {
 130                 tsk->thread.xstate = kmem_cache_alloc(task_xstate_cachep,
 131                                                       GFP_KERNEL);
 132                 if (!tsk->thread.xstate)
 133                         return -ENOMEM;
 134         }
 135
 136 #ifdef CONFIG_X86_32
 137         if (!HAVE_HWFP) {
 138                 memset(tsk->thread.xstate, 0, xstate_size);
 139                 finit();
 140                 set_stopped_child_used_math(tsk);
 141                 return 0;
 142         }
 143 #endif
 144
 145         if (cpu_has_fxsr) {
 146                 struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 147
 148                 memset(fx, 0, xstate_size);
 149                 fx->cwd = 0x37f;
 150                 if (cpu_has_xmm)
 151                         fx->mxcsr = MXCSR_DEFAULT;
 152         } else {
 153                 struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 154                 memset(fp, 0, xstate_size);
 155                 fp->cwd = 0xffff037fu;
 156                 fp->swd = 0xffff0000u;
 157                 fp->twd = 0xffffffffu;
 158                 fp->fos = 0xffff0000u;
 159         }
 160         /*
 161          * Only the device not available exception or ptrace can call init_fpu.
 162          */
 163         set_stopped_child_used_math(tsk);
 164         return 0;
 165 }
 166
 167 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
 168 {
 169         return tsk_used_math(target) ? regset->n : 0;
 170 }
 171
 172 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
 173 {
 174         return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
 175 }
 176
 177 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
 178                 unsigned int pos, unsigned int count,
 179                 void *kbuf, void __user *ubuf)
 180 {
 181         int ret;
 182
 183         if (!cpu_has_fxsr)
 184                 return -ENODEV;
 185
 186         ret = init_fpu(target);
 187         if (ret)
 188                 return ret;
 189
 190         return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 191                                    &target->thread.xstate->fxsave, 0, -1);
 192 }
 193
 194 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
 195                 unsigned int pos, unsigned int count,
 196                 const void *kbuf, const void __user *ubuf)
 197 {
 198         int ret;
 199
 200         if (!cpu_has_fxsr)
 201                 return -ENODEV;
 202
 203         ret = init_fpu(target);
 204         if (ret)
 205                 return ret;
 206
 207         set_stopped_child_used_math(target);
 208
 209         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 210                                  &target->thread.xstate->fxsave, 0, -1);
 211
 212         /*
 213          * mxcsr reserved bits must be masked to zero for security reasons.
 214          */
 215         target->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 216
 217         /*
 218          * update the header bits in the xsave header, indicating the
 219          * presence of FP and SSE state.
 220          */
 221         if (cpu_has_xsave)
 222                 target->thread.xstate->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
 223
 224         return ret;
 225 }
 226
 227 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
 228
 229 /*
 230  * FPU tag word conversions.
 231  */
 232
 233 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
 234 {
 235         unsigned int tmp; /* to avoid 16 bit prefixes in the code */
 236
 237         /* Transform each pair of bits into 01 (valid) or 00 (empty) */
 238         tmp = ~twd;
 239         tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
 240         /* and move the valid bits to the lower byte. */
 241         tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
 242         tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
 243         tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
 244
 245         return tmp;
 246 }
 247
 248 #define FPREG_ADDR(f, n)        ((void *)&(f)->st_space + (n) * 16);
 249 #define FP_EXP_TAG_VALID        0
 250 #define FP_EXP_TAG_ZERO         1
 251 #define FP_EXP_TAG_SPECIAL      2
 252 #define FP_EXP_TAG_EMPTY        3
 253
 254 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
 255 {
 256         struct _fpxreg *st;
 257         u32 tos = (fxsave->swd >> 11) & 7;
 258         u32 twd = (unsigned long) fxsave->twd;
 259         u32 tag;
 260         u32 ret = 0xffff0000u;
 261         int i;
 262
 263         for (i = 0; i < 8; i++, twd >>= 1) {
 264                 if (twd & 0x1) {
 265                         st = FPREG_ADDR(fxsave, (i - tos) & 7);
 266
 267                         switch (st->exponent & 0x7fff) {
 268                         case 0x7fff:
 269                                 tag = FP_EXP_TAG_SPECIAL;
 270                                 break;
 271                         case 0x0000:
 272                                 if (!st->significand[0] &&
 273                                     !st->significand[1] &&
 274                                     !st->significand[2] &&
 275                                     !st->significand[3])
 276                                         tag = FP_EXP_TAG_ZERO;
 277                                 else
 278                                         tag = FP_EXP_TAG_SPECIAL;
 279                                 break;
 280                         default:
 281                                 if (st->significand[3] & 0x8000)
 282                                         tag = FP_EXP_TAG_VALID;
 283                                 else
 284                                         tag = FP_EXP_TAG_SPECIAL;
 285                                 break;
 286                         }
 287                 } else {
 288                         tag = FP_EXP_TAG_EMPTY;
 289                 }
 290                 ret |= tag << (2 * i);
 291         }
 292         return ret;
 293 }
 294
 295 /*
 296  * FXSR floating point environment conversions.
 297  */
 298
 299 static void
 300 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
 301 {
 302         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 303         struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
 304         struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
 305         int i;
 306
 307         env->cwd = fxsave->cwd | 0xffff0000u;
 308         env->swd = fxsave->swd | 0xffff0000u;
 309         env->twd = twd_fxsr_to_i387(fxsave);
 310
 311 #ifdef CONFIG_X86_64
 312         env->fip = fxsave->rip;
 313         env->foo = fxsave->rdp;
 314         if (tsk == current) {
 315                 /*
 316                  * should be actually ds/cs at fpu exception time, but
 317                  * that information is not available in 64bit mode.
 318                  */
 319                 asm("mov %%ds, %[fos]" : [fos] "=r" (env->fos));
 320                 asm("mov %%cs, %[fcs]" : [fcs] "=r" (env->fcs));
 321         } else {
 322                 struct pt_regs *regs = task_pt_regs(tsk);
 323
 324                 env->fos = 0xffff0000 | tsk->thread.ds;
 325                 env->fcs = regs->cs;
 326         }
 327 #else
 328         env->fip = fxsave->fip;
 329         env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
 330         env->foo = fxsave->foo;
 331         env->fos = fxsave->fos;
 332 #endif
 333
 334         for (i = 0; i < 8; ++i)
 335                 memcpy(&to[i], &from[i], sizeof(to[0]));
 336 }
 337
 338 static void convert_to_fxsr(struct task_struct *tsk,
 339                             const struct user_i387_ia32_struct *env)
 340
 341 {
 342         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 343         struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
 344         struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
 345         int i;
 346
 347         fxsave->cwd = env->cwd;
 348         fxsave->swd = env->swd;
 349         fxsave->twd = twd_i387_to_fxsr(env->twd);
 350         fxsave->fop = (u16) ((u32) env->fcs >> 16);
 351 #ifdef CONFIG_X86_64
 352         fxsave->rip = env->fip;
 353         fxsave->rdp = env->foo;
 354         /* cs and ds ignored */
 355 #else
 356         fxsave->fip = env->fip;
 357         fxsave->fcs = (env->fcs & 0xffff);
 358         fxsave->foo = env->foo;
 359         fxsave->fos = env->fos;
 360 #endif
 361
 362         for (i = 0; i < 8; ++i)
 363                 memcpy(&to[i], &from[i], sizeof(from[0]));
 364 }
 365
 366 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
 367                unsigned int pos, unsigned int count,
 368                void *kbuf, void __user *ubuf)
 369 {
 370         struct user_i387_ia32_struct env;
 371         int ret;
 372
 373         ret = init_fpu(target);
 374         if (ret)
 375                 return ret;
 376
 377         if (!HAVE_HWFP)
 378                 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
 379
 380         if (!cpu_has_fxsr) {
 381                 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 382                                            &target->thread.xstate->fsave, 0,
 383                                            -1);
 384         }
 385
 386         if (kbuf && pos == 0 && count == sizeof(env)) {
 387                 convert_from_fxsr(kbuf, target);
 388                 return 0;
 389         }
 390
 391         convert_from_fxsr(&env, target);
 392
 393         return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 394 }
 395
 396 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
 397                unsigned int pos, unsigned int count,
 398                const void *kbuf, const void __user *ubuf)
 399 {
 400         struct user_i387_ia32_struct env;
 401         int ret;
 402
 403         ret = init_fpu(target);
 404         if (ret)
 405                 return ret;
 406
 407         set_stopped_child_used_math(target);
 408
 409         if (!HAVE_HWFP)
 410                 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
 411
 412         if (!cpu_has_fxsr) {
 413                 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 414                                           &target->thread.xstate->fsave, 0, -1);
 415         }
 416
 417         if (pos > 0 || count < sizeof(env))
 418                 convert_from_fxsr(&env, target);
 419
 420         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 421         if (!ret)
 422                 convert_to_fxsr(target, &env);
 423
 424         /*
 425          * update the header bit in the xsave header, indicating the
 426          * presence of FP.
 427          */
 428         if (cpu_has_xsave)
 429                 target->thread.xstate->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
 430         return ret;
 431 }
 432
 433 /*
 434  * Signal frame handlers.
 435  */
 436
 437 static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
 438 {
 439         struct task_struct *tsk = current;
 440         struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 441
 442         fp->status = fp->swd;
 443         if (__copy_to_user(buf, fp, sizeof(struct i387_fsave_struct)))
 444                 return -1;
 445         return 1;
 446 }
 447
 448 static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
 449 {
 450         struct task_struct *tsk = current;
 451         struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 452         struct user_i387_ia32_struct env;
 453         int err = 0;
 454
 455         convert_from_fxsr(&env, tsk);
 456         if (__copy_to_user(buf, &env, sizeof(env)))
 457                 return -1;
 458
 459         err |= __put_user(fx->swd, &buf->status);
 460         err |= __put_user(X86_FXSR_MAGIC, &buf->magic);
 461         if (err)
 462                 return -1;
 463
 464         if (__copy_to_user(&buf->_fxsr_env[0], fx, xstate_size))
 465                 return -1;
 466         return 1;
 467 }
 468
 469 static int save_i387_xsave(void __user *buf)
 470 {
 471         struct _fpstate_ia32 __user *fx = buf;
 472         int err = 0;
 473
 474         if (save_i387_fxsave(fx) < 0)
 475                 return -1;
 476
 477         err = __copy_to_user(&fx->sw_reserved, &fx_sw_reserved_ia32,
 478                              sizeof(struct _fpx_sw_bytes));
 479         err |= __put_user(FP_XSTATE_MAGIC2,
 480                           (__u32 __user *) (buf + sig_xstate_ia32_size
 481                                             - FP_XSTATE_MAGIC2_SIZE));
 482         if (err)
 483                 return -1;
 484
 485         return 1;
 486 }
 487
 488 int save_i387_xstate_ia32(void __user *buf)
 489 {
 490         struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
 491         struct task_struct *tsk = current;
 492
 493         if (!used_math())
 494                 return 0;
 495
 496         if (!access_ok(VERIFY_WRITE, buf, sig_xstate_ia32_size))
 497                 return -EACCES;
 498         /*
 499          * This will cause a "finit" to be triggered by the next
 500          * attempted FPU operation by the 'current' process.
 501          */
 502         clear_used_math();
 503
 504         if (!HAVE_HWFP) {
 505                 return fpregs_soft_get(current, NULL,
 506                                        0, sizeof(struct user_i387_ia32_struct),
 507                                        NULL, fp) ? -1 : 1;
 508         }
 509
 510         unlazy_fpu(tsk);
 511
 512         if (cpu_has_xsave)
 513                 return save_i387_xsave(fp);
 514         if (cpu_has_fxsr)
 515                 return save_i387_fxsave(fp);
 516         else
 517                 return save_i387_fsave(fp);
 518 }
 519
 520 static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
 521 {
 522         struct task_struct *tsk = current;
 523
 524         return __copy_from_user(&tsk->thread.xstate->fsave, buf,
 525                                 sizeof(struct i387_fsave_struct));
 526 }
 527
 528 static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf,
 529                                unsigned int size)
 530 {
 531         struct task_struct *tsk = current;
 532         struct user_i387_ia32_struct env;
 533         int err;
 534
 535         err = __copy_from_user(&tsk->thread.xstate->fxsave, &buf->_fxsr_env[0],
 536                                size);
 537         /* mxcsr reserved bits must be masked to zero for security reasons */
 538         tsk->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 539         if (err || __copy_from_user(&env, buf, sizeof(env)))
 540                 return 1;
 541         convert_to_fxsr(tsk, &env);
 542
 543         return 0;
 544 }
 545
 546 static int restore_i387_xsave(void __user *buf)
 547 {
 548         struct _fpx_sw_bytes fx_sw_user;
 549         struct _fpstate_ia32 __user *fx_user =
 550                         ((struct _fpstate_ia32 __user *) buf);
 551         struct i387_fxsave_struct __user *fx =
 552                 (struct i387_fxsave_struct __user *) &fx_user->_fxsr_env[0];
 553         struct xsave_hdr_struct *xsave_hdr =
 554                                 &current->thread.xstate->xsave.xsave_hdr;
 555         u64 mask;
 556         int err;
 557
 558         if (check_for_xstate(fx, buf, &fx_sw_user))
 559                 goto fx_only;
 560
 561         mask = fx_sw_user.xstate_bv;
 562
 563         err = restore_i387_fxsave(buf, fx_sw_user.xstate_size);
 564
 565         xsave_hdr->xstate_bv &= pcntxt_mask;
 566         /*
 567          * These bits must be zero.
 568          */
 569         xsave_hdr->reserved1[0] = xsave_hdr->reserved1[1] = 0;
 570
 571         /*
 572          * Init the state that is not present in the memory layout
 573          * and enabled by the OS.
 574          */
 575         mask = ~(pcntxt_mask & ~mask);
 576         xsave_hdr->xstate_bv &= mask;
 577
 578         return err;
 579 fx_only:
 580         /*
 581          * Couldn't find the extended state information in the memory
 582          * layout. Restore the FP/SSE and init the other extended state
 583          * enabled by the OS.
 584          */
 585         xsave_hdr->xstate_bv = XSTATE_FPSSE;
 586         return restore_i387_fxsave(buf, sizeof(struct i387_fxsave_struct));
 587 }
 588
 589 int restore_i387_xstate_ia32(void __user *buf)
 590 {
 591         int err;
 592         struct task_struct *tsk = current;
 593         struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
 594
 595         if (HAVE_HWFP)
 596                 clear_fpu(tsk);
 597
 598         if (!buf) {
 599                 if (used_math()) {
 600                         clear_fpu(tsk);
 601                         clear_used_math();
 602                 }
 603
 604                 return 0;
 605         } else
 606                 if (!access_ok(VERIFY_READ, buf, sig_xstate_ia32_size))
 607                         return -EACCES;
 608
 609         if (!used_math()) {
 610                 err = init_fpu(tsk);
 611                 if (err)
 612                         return err;
 613         }
 614
 615         if (HAVE_HWFP) {
 616                 if (cpu_has_xsave)
 617                         err = restore_i387_xsave(buf);
 618                 else if (cpu_has_fxsr)
 619                         err = restore_i387_fxsave(fp, sizeof(struct
 620                                                            i387_fxsave_struct));
 621                 else
 622                         err = restore_i387_fsave(fp);
 623         } else {
 624                 err = fpregs_soft_set(current, NULL,
 625                                       0, sizeof(struct user_i387_ia32_struct),
 626                                       NULL, fp) != 0;
 627         }
 628         set_used_math();
 629
 630         return err;
 631 }
 632
 633 /*
 634  * FPU state for core dumps.
 635  * This is only used for a.out dumps now.
 636  * It is declared generically using elf_fpregset_t (which is
 637  * struct user_i387_struct) but is in fact only used for 32-bit
 638  * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
 639  */
 640 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
 641 {
 642         struct task_struct *tsk = current;
 643         int fpvalid;
 644
 645         fpvalid = !!used_math();
 646         if (fpvalid)
 647                 fpvalid = !fpregs_get(tsk, NULL,
 648                                       0, sizeof(struct user_i387_ia32_struct),
 649                                       fpu, NULL);
 650
 651         return fpvalid;
 652 }
 653 EXPORT_SYMBOL(dump_fpu);
 654
 655 #endif  /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */