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         return ret;
 218 }
 219
 220 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
 221
 222 /*
 223  * FPU tag word conversions.
 224  */
 225
 226 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
 227 {
 228         unsigned int tmp; /* to avoid 16 bit prefixes in the code */
 229
 230         /* Transform each pair of bits into 01 (valid) or 00 (empty) */
 231         tmp = ~twd;
 232         tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
 233         /* and move the valid bits to the lower byte. */
 234         tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
 235         tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
 236         tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
 237
 238         return tmp;
 239 }
 240
 241 #define FPREG_ADDR(f, n)        ((void *)&(f)->st_space + (n) * 16);
 242 #define FP_EXP_TAG_VALID        0
 243 #define FP_EXP_TAG_ZERO         1
 244 #define FP_EXP_TAG_SPECIAL      2
 245 #define FP_EXP_TAG_EMPTY        3
 246
 247 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
 248 {
 249         struct _fpxreg *st;
 250         u32 tos = (fxsave->swd >> 11) & 7;
 251         u32 twd = (unsigned long) fxsave->twd;
 252         u32 tag;
 253         u32 ret = 0xffff0000u;
 254         int i;
 255
 256         for (i = 0; i < 8; i++, twd >>= 1) {
 257                 if (twd & 0x1) {
 258                         st = FPREG_ADDR(fxsave, (i - tos) & 7);
 259
 260                         switch (st->exponent & 0x7fff) {
 261                         case 0x7fff:
 262                                 tag = FP_EXP_TAG_SPECIAL;
 263                                 break;
 264                         case 0x0000:
 265                                 if (!st->significand[0] &&
 266                                     !st->significand[1] &&
 267                                     !st->significand[2] &&
 268                                     !st->significand[3])
 269                                         tag = FP_EXP_TAG_ZERO;
 270                                 else
 271                                         tag = FP_EXP_TAG_SPECIAL;
 272                                 break;
 273                         default:
 274                                 if (st->significand[3] & 0x8000)
 275                                         tag = FP_EXP_TAG_VALID;
 276                                 else
 277                                         tag = FP_EXP_TAG_SPECIAL;
 278                                 break;
 279                         }
 280                 } else {
 281                         tag = FP_EXP_TAG_EMPTY;
 282                 }
 283                 ret |= tag << (2 * i);
 284         }
 285         return ret;
 286 }
 287
 288 /*
 289  * FXSR floating point environment conversions.
 290  */
 291
 292 static void
 293 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
 294 {
 295         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 296         struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
 297         struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
 298         int i;
 299
 300         env->cwd = fxsave->cwd | 0xffff0000u;
 301         env->swd = fxsave->swd | 0xffff0000u;
 302         env->twd = twd_fxsr_to_i387(fxsave);
 303
 304 #ifdef CONFIG_X86_64
 305         env->fip = fxsave->rip;
 306         env->foo = fxsave->rdp;
 307         if (tsk == current) {
 308                 /*
 309                  * should be actually ds/cs at fpu exception time, but
 310                  * that information is not available in 64bit mode.
 311                  */
 312                 asm("mov %%ds, %[fos]" : [fos] "=r" (env->fos));
 313                 asm("mov %%cs, %[fcs]" : [fcs] "=r" (env->fcs));
 314         } else {
 315                 struct pt_regs *regs = task_pt_regs(tsk);
 316
 317                 env->fos = 0xffff0000 | tsk->thread.ds;
 318                 env->fcs = regs->cs;
 319         }
 320 #else
 321         env->fip = fxsave->fip;
 322         env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
 323         env->foo = fxsave->foo;
 324         env->fos = fxsave->fos;
 325 #endif
 326
 327         for (i = 0; i < 8; ++i)
 328                 memcpy(&to[i], &from[i], sizeof(to[0]));
 329 }
 330
 331 static void convert_to_fxsr(struct task_struct *tsk,
 332                             const struct user_i387_ia32_struct *env)
 333
 334 {
 335         struct i387_fxsave_struct *fxsave = &tsk->thread.xstate->fxsave;
 336         struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
 337         struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
 338         int i;
 339
 340         fxsave->cwd = env->cwd;
 341         fxsave->swd = env->swd;
 342         fxsave->twd = twd_i387_to_fxsr(env->twd);
 343         fxsave->fop = (u16) ((u32) env->fcs >> 16);
 344 #ifdef CONFIG_X86_64
 345         fxsave->rip = env->fip;
 346         fxsave->rdp = env->foo;
 347         /* cs and ds ignored */
 348 #else
 349         fxsave->fip = env->fip;
 350         fxsave->fcs = (env->fcs & 0xffff);
 351         fxsave->foo = env->foo;
 352         fxsave->fos = env->fos;
 353 #endif
 354
 355         for (i = 0; i < 8; ++i)
 356                 memcpy(&to[i], &from[i], sizeof(from[0]));
 357 }
 358
 359 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
 360                unsigned int pos, unsigned int count,
 361                void *kbuf, void __user *ubuf)
 362 {
 363         struct user_i387_ia32_struct env;
 364         int ret;
 365
 366         ret = init_fpu(target);
 367         if (ret)
 368                 return ret;
 369
 370         if (!HAVE_HWFP)
 371                 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
 372
 373         if (!cpu_has_fxsr) {
 374                 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 375                                            &target->thread.xstate->fsave, 0,
 376                                            -1);
 377         }
 378
 379         if (kbuf && pos == 0 && count == sizeof(env)) {
 380                 convert_from_fxsr(kbuf, target);
 381                 return 0;
 382         }
 383
 384         convert_from_fxsr(&env, target);
 385
 386         return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 387 }
 388
 389 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
 390                unsigned int pos, unsigned int count,
 391                const void *kbuf, const void __user *ubuf)
 392 {
 393         struct user_i387_ia32_struct env;
 394         int ret;
 395
 396         ret = init_fpu(target);
 397         if (ret)
 398                 return ret;
 399
 400         set_stopped_child_used_math(target);
 401
 402         if (!HAVE_HWFP)
 403                 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
 404
 405         if (!cpu_has_fxsr) {
 406                 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 407                                           &target->thread.xstate->fsave, 0, -1);
 408         }
 409
 410         if (pos > 0 || count < sizeof(env))
 411                 convert_from_fxsr(&env, target);
 412
 413         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 414         if (!ret)
 415                 convert_to_fxsr(target, &env);
 416
 417         return ret;
 418 }
 419
 420 /*
 421  * Signal frame handlers.
 422  */
 423
 424 static inline int save_i387_fsave(struct _fpstate_ia32 __user *buf)
 425 {
 426         struct task_struct *tsk = current;
 427         struct i387_fsave_struct *fp = &tsk->thread.xstate->fsave;
 428
 429         fp->status = fp->swd;
 430         if (__copy_to_user(buf, fp, sizeof(struct i387_fsave_struct)))
 431                 return -1;
 432         return 1;
 433 }
 434
 435 static int save_i387_fxsave(struct _fpstate_ia32 __user *buf)
 436 {
 437         struct task_struct *tsk = current;
 438         struct i387_fxsave_struct *fx = &tsk->thread.xstate->fxsave;
 439         struct user_i387_ia32_struct env;
 440         int err = 0;
 441
 442         convert_from_fxsr(&env, tsk);
 443         if (__copy_to_user(buf, &env, sizeof(env)))
 444                 return -1;
 445
 446         err |= __put_user(fx->swd, &buf->status);
 447         err |= __put_user(X86_FXSR_MAGIC, &buf->magic);
 448         if (err)
 449                 return -1;
 450
 451         if (__copy_to_user(&buf->_fxsr_env[0], fx, xstate_size))
 452                 return -1;
 453         return 1;
 454 }
 455
 456 static int save_i387_xsave(void __user *buf)
 457 {
 458         struct _fpstate_ia32 __user *fx = buf;
 459         int err = 0;
 460
 461         if (save_i387_fxsave(fx) < 0)
 462                 return -1;
 463
 464         err = __copy_to_user(&fx->sw_reserved, &fx_sw_reserved_ia32,
 465                              sizeof(struct _fpx_sw_bytes));
 466         err |= __put_user(FP_XSTATE_MAGIC2,
 467                           (__u32 __user *) (buf + sig_xstate_ia32_size
 468                                             - FP_XSTATE_MAGIC2_SIZE));
 469         if (err)
 470                 return -1;
 471
 472         return 1;
 473 }
 474
 475 int save_i387_xstate_ia32(void __user *buf)
 476 {
 477         struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
 478         struct task_struct *tsk = current;
 479
 480         if (!used_math())
 481                 return 0;
 482
 483         if (!access_ok(VERIFY_WRITE, buf, sig_xstate_ia32_size))
 484                 return -EACCES;
 485         /*
 486          * This will cause a "finit" to be triggered by the next
 487          * attempted FPU operation by the 'current' process.
 488          */
 489         clear_used_math();
 490
 491         if (!HAVE_HWFP) {
 492                 return fpregs_soft_get(current, NULL,
 493                                        0, sizeof(struct user_i387_ia32_struct),
 494                                        NULL, fp) ? -1 : 1;
 495         }
 496
 497         unlazy_fpu(tsk);
 498
 499         if (cpu_has_xsave)
 500                 return save_i387_xsave(fp);
 501         if (cpu_has_fxsr)
 502                 return save_i387_fxsave(fp);
 503         else
 504                 return save_i387_fsave(fp);
 505 }
 506
 507 static inline int restore_i387_fsave(struct _fpstate_ia32 __user *buf)
 508 {
 509         struct task_struct *tsk = current;
 510
 511         return __copy_from_user(&tsk->thread.xstate->fsave, buf,
 512                                 sizeof(struct i387_fsave_struct));
 513 }
 514
 515 static int restore_i387_fxsave(struct _fpstate_ia32 __user *buf,
 516                                unsigned int size)
 517 {
 518         struct task_struct *tsk = current;
 519         struct user_i387_ia32_struct env;
 520         int err;
 521
 522         err = __copy_from_user(&tsk->thread.xstate->fxsave, &buf->_fxsr_env[0],
 523                                size);
 524         /* mxcsr reserved bits must be masked to zero for security reasons */
 525         tsk->thread.xstate->fxsave.mxcsr &= mxcsr_feature_mask;
 526         if (err || __copy_from_user(&env, buf, sizeof(env)))
 527                 return 1;
 528         convert_to_fxsr(tsk, &env);
 529
 530         return 0;
 531 }
 532
 533 static int restore_i387_xsave(void __user *buf)
 534 {
 535         struct _fpx_sw_bytes fx_sw_user;
 536         struct _fpstate_ia32 __user *fx_user =
 537                         ((struct _fpstate_ia32 __user *) buf);
 538         struct i387_fxsave_struct __user *fx =
 539                 (struct i387_fxsave_struct __user *) &fx_user->_fxsr_env[0];
 540         struct xsave_hdr_struct *xsave_hdr =
 541                                 &current->thread.xstate->xsave.xsave_hdr;
 542         unsigned int lmask, hmask;
 543         int err;
 544
 545         if (check_for_xstate(fx, buf, &fx_sw_user))
 546                 goto fx_only;
 547
 548         lmask = fx_sw_user.xstate_bv;
 549         hmask = fx_sw_user.xstate_bv >> 32;
 550
 551         err = restore_i387_fxsave(buf, fx_sw_user.xstate_size);
 552
 553         xsave_hdr->xstate_bv &=  (pcntxt_lmask | (((u64) pcntxt_hmask) << 32));
 554         /*
 555          * These bits must be zero.
 556          */
 557         xsave_hdr->reserved1[0] = xsave_hdr->reserved1[1] = 0;
 558
 559         /*
 560          * Init the state that is not present in the memory layout
 561          * and enabled by the OS.
 562          */
 563         lmask = ~(pcntxt_lmask & ~lmask);
 564         hmask = ~(pcntxt_hmask & ~hmask);
 565         xsave_hdr->xstate_bv &=  (lmask | (((u64) hmask) << 32));
 566
 567         return err;
 568 fx_only:
 569         /*
 570          * Couldn't find the extended state information in the memory
 571          * layout. Restore the FP/SSE and init the other extended state
 572          * enabled by the OS.
 573          */
 574         xsave_hdr->xstate_bv = XSTATE_FPSSE;
 575         return restore_i387_fxsave(buf, sizeof(struct i387_fxsave_struct));
 576 }
 577
 578 int restore_i387_xstate_ia32(void __user *buf)
 579 {
 580         int err;
 581         struct task_struct *tsk = current;
 582         struct _fpstate_ia32 __user *fp = (struct _fpstate_ia32 __user *) buf;
 583
 584         if (HAVE_HWFP)
 585                 clear_fpu(tsk);
 586
 587         if (!buf) {
 588                 if (used_math()) {
 589                         clear_fpu(tsk);
 590                         clear_used_math();
 591                 }
 592
 593                 return 0;
 594         } else
 595                 if (!access_ok(VERIFY_READ, buf, sig_xstate_ia32_size))
 596                         return -EACCES;
 597
 598         if (!used_math()) {
 599                 err = init_fpu(tsk);
 600                 if (err)
 601                         return err;
 602         }
 603
 604         if (HAVE_HWFP) {
 605                 if (cpu_has_xsave)
 606                         err = restore_i387_xsave(buf);
 607                 else if (cpu_has_fxsr)
 608                         err = restore_i387_fxsave(fp, sizeof(struct
 609                                                            i387_fxsave_struct));
 610                 else
 611                         err = restore_i387_fsave(fp);
 612         } else {
 613                 err = fpregs_soft_set(current, NULL,
 614                                       0, sizeof(struct user_i387_ia32_struct),
 615                                       NULL, fp) != 0;
 616         }
 617         set_used_math();
 618
 619         return err;
 620 }
 621
 622 /*
 623  * FPU state for core dumps.
 624  * This is only used for a.out dumps now.
 625  * It is declared generically using elf_fpregset_t (which is
 626  * struct user_i387_struct) but is in fact only used for 32-bit
 627  * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
 628  */
 629 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
 630 {
 631         struct task_struct *tsk = current;
 632         int fpvalid;
 633
 634         fpvalid = !!used_math();
 635         if (fpvalid)
 636                 fpvalid = !fpregs_get(tsk, NULL,
 637                                       0, sizeof(struct user_i387_ia32_struct),
 638                                       fpu, NULL);
 639
 640         return fpvalid;
 641 }
 642 EXPORT_SYMBOL(dump_fpu);
 643
 644 #endif  /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */