arch/x86/kernel/fpu/core.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 <asm/fpu-internal.h>
   9
  10 static DEFINE_PER_CPU(bool, in_kernel_fpu);
  11
  12 void kernel_fpu_disable(void)
  13 {
  14         WARN_ON(this_cpu_read(in_kernel_fpu));
  15         this_cpu_write(in_kernel_fpu, true);
  16 }
  17
  18 void kernel_fpu_enable(void)
  19 {
  20         this_cpu_write(in_kernel_fpu, false);
  21 }
  22
  23 /*
  24  * Were we in an interrupt that interrupted kernel mode?
  25  *
  26  * On others, we can do a kernel_fpu_begin/end() pair *ONLY* if that
  27  * pair does nothing at all: the thread must not have fpu (so
  28  * that we don't try to save the FPU state), and TS must
  29  * be set (so that the clts/stts pair does nothing that is
  30  * visible in the interrupted kernel thread).
  31  *
  32  * Except for the eagerfpu case when we return true; in the likely case
  33  * the thread has FPU but we are not going to set/clear TS.
  34  */
  35 static inline bool interrupted_kernel_fpu_idle(void)
  36 {
  37         if (this_cpu_read(in_kernel_fpu))
  38                 return false;
  39
  40         if (use_eager_fpu())
  41                 return true;
  42
  43         return !__thread_has_fpu(current) &&
  44                 (read_cr0() & X86_CR0_TS);
  45 }
  46
  47 /*
  48  * Were we in user mode (or vm86 mode) when we were
  49  * interrupted?
  50  *
  51  * Doing kernel_fpu_begin/end() is ok if we are running
  52  * in an interrupt context from user mode - we'll just
  53  * save the FPU state as required.
  54  */
  55 static inline bool interrupted_user_mode(void)
  56 {
  57         struct pt_regs *regs = get_irq_regs();
  58         return regs && user_mode(regs);
  59 }
  60
  61 /*
  62  * Can we use the FPU in kernel mode with the
  63  * whole "kernel_fpu_begin/end()" sequence?
  64  *
  65  * It's always ok in process context (ie "not interrupt")
  66  * but it is sometimes ok even from an irq.
  67  */
  68 bool irq_fpu_usable(void)
  69 {
  70         return !in_interrupt() ||
  71                 interrupted_user_mode() ||
  72                 interrupted_kernel_fpu_idle();
  73 }
  74 EXPORT_SYMBOL(irq_fpu_usable);
  75
  76 void __kernel_fpu_begin(void)
  77 {
  78         struct task_struct *me = current;
  79
  80         this_cpu_write(in_kernel_fpu, true);
  81
  82         if (__thread_has_fpu(me)) {
  83                 __save_init_fpu(me);
  84         } else {
  85                 this_cpu_write(fpu_owner_task, NULL);
  86                 if (!use_eager_fpu())
  87                         clts();
  88         }
  89 }
  90 EXPORT_SYMBOL(__kernel_fpu_begin);
  91
  92 void __kernel_fpu_end(void)
  93 {
  94         struct task_struct *me = current;
  95
  96         if (__thread_has_fpu(me)) {
  97                 if (WARN_ON(restore_fpu_checking(me)))
  98                         fpu_reset_state(me);
  99         } else if (!use_eager_fpu()) {
 100                 stts();
 101         }
 102
 103         this_cpu_write(in_kernel_fpu, false);
 104 }
 105 EXPORT_SYMBOL(__kernel_fpu_end);
 106
 107 /*
 108  * Save the FPU state (initialize it if necessary):
 109  *
 110  * This only ever gets called for the current task.
 111  */
 112 void fpu__save(struct task_struct *tsk)
 113 {
 114         WARN_ON(tsk != current);
 115
 116         preempt_disable();
 117         if (__thread_has_fpu(tsk)) {
 118                 if (use_eager_fpu()) {
 119                         __save_fpu(tsk);
 120                 } else {
 121                         __save_init_fpu(tsk);
 122                         __thread_fpu_end(tsk);
 123                 }
 124         }
 125         preempt_enable();
 126 }
 127 EXPORT_SYMBOL_GPL(fpu__save);
 128
 129 void fpstate_init(struct fpu *fpu)
 130 {
 131         if (!cpu_has_fpu) {
 132                 finit_soft_fpu(&fpu->state->soft);
 133                 return;
 134         }
 135
 136         memset(fpu->state, 0, xstate_size);
 137
 138         if (cpu_has_fxsr) {
 139                 fx_finit(&fpu->state->fxsave);
 140         } else {
 141                 struct i387_fsave_struct *fp = &fpu->state->fsave;
 142                 fp->cwd = 0xffff037fu;
 143                 fp->swd = 0xffff0000u;
 144                 fp->twd = 0xffffffffu;
 145                 fp->fos = 0xffff0000u;
 146         }
 147 }
 148 EXPORT_SYMBOL_GPL(fpstate_init);
 149
 150 /*
 151  * FPU state allocation:
 152  */
 153 struct kmem_cache *task_xstate_cachep;
 154 EXPORT_SYMBOL_GPL(task_xstate_cachep);
 155
 156 void fpstate_cache_init(void)
 157 {
 158         task_xstate_cachep =
 159                 kmem_cache_create("task_xstate", xstate_size,
 160                                   __alignof__(union thread_xstate),
 161                                   SLAB_PANIC | SLAB_NOTRACK, NULL);
 162         setup_xstate_comp();
 163 }
 164
 165 int fpstate_alloc(struct fpu *fpu)
 166 {
 167         if (fpu->state)
 168                 return 0;
 169
 170         fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
 171         if (!fpu->state)
 172                 return -ENOMEM;
 173
 174         /* The CPU requires the FPU state to be aligned to 16 byte boundaries: */
 175         WARN_ON((unsigned long)fpu->state & 15);
 176
 177         return 0;
 178 }
 179 EXPORT_SYMBOL_GPL(fpstate_alloc);
 180
 181 void fpstate_free(struct fpu *fpu)
 182 {
 183         if (fpu->state) {
 184                 kmem_cache_free(task_xstate_cachep, fpu->state);
 185                 fpu->state = NULL;
 186         }
 187 }
 188 EXPORT_SYMBOL_GPL(fpstate_free);
 189
 190 int fpu__copy(struct task_struct *dst, struct task_struct *src)
 191 {
 192         dst->thread.fpu.counter = 0;
 193         dst->thread.fpu.has_fpu = 0;
 194         dst->thread.fpu.state = NULL;
 195
 196         task_disable_lazy_fpu_restore(dst);
 197
 198         if (tsk_used_math(src)) {
 199                 int err = fpstate_alloc(&dst->thread.fpu);
 200
 201                 if (err)
 202                         return err;
 203                 fpu_copy(dst, src);
 204         }
 205         return 0;
 206 }
 207
 208 /*
 209  * Allocate the backing store for the current task's FPU registers
 210  * and initialize the registers themselves as well.
 211  *
 212  * Can fail.
 213  */
 214 int fpstate_alloc_init(struct task_struct *curr)
 215 {
 216         int ret;
 217
 218         if (WARN_ON_ONCE(curr != current))
 219                 return -EINVAL;
 220         if (WARN_ON_ONCE(curr->flags & PF_USED_MATH))
 221                 return -EINVAL;
 222
 223         /*
 224          * Memory allocation at the first usage of the FPU and other state.
 225          */
 226         ret = fpstate_alloc(&curr->thread.fpu);
 227         if (ret)
 228                 return ret;
 229
 230         fpstate_init(&curr->thread.fpu);
 231
 232         /* Safe to do for the current task: */
 233         curr->flags |= PF_USED_MATH;
 234
 235         return 0;
 236 }
 237 EXPORT_SYMBOL_GPL(fpstate_alloc_init);
 238
 239 /*
 240  * The _current_ task is using the FPU for the first time
 241  * so initialize it and set the mxcsr to its default
 242  * value at reset if we support XMM instructions and then
 243  * remember the current task has used the FPU.
 244  */
 245 static int fpu__unlazy_stopped(struct task_struct *child)
 246 {
 247         int ret;
 248
 249         if (WARN_ON_ONCE(child == current))
 250                 return -EINVAL;
 251
 252         if (child->flags & PF_USED_MATH) {
 253                 task_disable_lazy_fpu_restore(child);
 254                 return 0;
 255         }
 256
 257         /*
 258          * Memory allocation at the first usage of the FPU and other state.
 259          */
 260         ret = fpstate_alloc(&child->thread.fpu);
 261         if (ret)
 262                 return ret;
 263
 264         fpstate_init(&child->thread.fpu);
 265
 266         /* Safe to do for stopped child tasks: */
 267         child->flags |= PF_USED_MATH;
 268
 269         return 0;
 270 }
 271
 272 /*
 273  * 'fpu__restore()' saves the current math information in the
 274  * old math state array, and gets the new ones from the current task
 275  *
 276  * Careful.. There are problems with IBM-designed IRQ13 behaviour.
 277  * Don't touch unless you *really* know how it works.
 278  *
 279  * Must be called with kernel preemption disabled (eg with local
 280  * local interrupts as in the case of do_device_not_available).
 281  */
 282 void fpu__restore(void)
 283 {
 284         struct task_struct *tsk = current;
 285
 286         if (!tsk_used_math(tsk)) {
 287                 local_irq_enable();
 288                 /*
 289                  * does a slab alloc which can sleep
 290                  */
 291                 if (fpstate_alloc_init(tsk)) {
 292                         /*
 293                          * ran out of memory!
 294                          */
 295                         do_group_exit(SIGKILL);
 296                         return;
 297                 }
 298                 local_irq_disable();
 299         }
 300
 301         /* Avoid __kernel_fpu_begin() right after __thread_fpu_begin() */
 302         kernel_fpu_disable();
 303         __thread_fpu_begin(tsk);
 304         if (unlikely(restore_fpu_checking(tsk))) {
 305                 fpu_reset_state(tsk);
 306                 force_sig_info(SIGSEGV, SEND_SIG_PRIV, tsk);
 307         } else {
 308                 tsk->thread.fpu.counter++;
 309         }
 310         kernel_fpu_enable();
 311 }
 312 EXPORT_SYMBOL_GPL(fpu__restore);
 313
 314 void fpu__flush_thread(struct task_struct *tsk)
 315 {
 316         if (!use_eager_fpu()) {
 317                 /* FPU state will be reallocated lazily at the first use. */
 318                 drop_fpu(tsk);
 319                 fpstate_free(&tsk->thread.fpu);
 320         } else {
 321                 if (!tsk_used_math(tsk)) {
 322                         /* kthread execs. TODO: cleanup this horror. */
 323                 if (WARN_ON(fpstate_alloc_init(tsk)))
 324                                 force_sig(SIGKILL, tsk);
 325                         user_fpu_begin();
 326                 }
 327                 restore_init_xstate();
 328         }
 329 }
 330
 331 /*
 332  * The xstateregs_active() routine is the same as the fpregs_active() routine,
 333  * as the "regset->n" for the xstate regset will be updated based on the feature
 334  * capabilites supported by the xsave.
 335  */
 336 int fpregs_active(struct task_struct *target, const struct user_regset *regset)
 337 {
 338         return tsk_used_math(target) ? regset->n : 0;
 339 }
 340
 341 int xfpregs_active(struct task_struct *target, const struct user_regset *regset)
 342 {
 343         return (cpu_has_fxsr && tsk_used_math(target)) ? regset->n : 0;
 344 }
 345
 346 int xfpregs_get(struct task_struct *target, const struct user_regset *regset,
 347                 unsigned int pos, unsigned int count,
 348                 void *kbuf, void __user *ubuf)
 349 {
 350         int ret;
 351
 352         if (!cpu_has_fxsr)
 353                 return -ENODEV;
 354
 355         ret = fpu__unlazy_stopped(target);
 356         if (ret)
 357                 return ret;
 358
 359         sanitize_i387_state(target);
 360
 361         return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 362                                    &target->thread.fpu.state->fxsave, 0, -1);
 363 }
 364
 365 int xfpregs_set(struct task_struct *target, const struct user_regset *regset,
 366                 unsigned int pos, unsigned int count,
 367                 const void *kbuf, const void __user *ubuf)
 368 {
 369         int ret;
 370
 371         if (!cpu_has_fxsr)
 372                 return -ENODEV;
 373
 374         ret = fpu__unlazy_stopped(target);
 375         if (ret)
 376                 return ret;
 377
 378         sanitize_i387_state(target);
 379
 380         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 381                                  &target->thread.fpu.state->fxsave, 0, -1);
 382
 383         /*
 384          * mxcsr reserved bits must be masked to zero for security reasons.
 385          */
 386         target->thread.fpu.state->fxsave.mxcsr &= mxcsr_feature_mask;
 387
 388         /*
 389          * update the header bits in the xsave header, indicating the
 390          * presence of FP and SSE state.
 391          */
 392         if (cpu_has_xsave)
 393                 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FPSSE;
 394
 395         return ret;
 396 }
 397
 398 int xstateregs_get(struct task_struct *target, const struct user_regset *regset,
 399                 unsigned int pos, unsigned int count,
 400                 void *kbuf, void __user *ubuf)
 401 {
 402         struct xsave_struct *xsave;
 403         int ret;
 404
 405         if (!cpu_has_xsave)
 406                 return -ENODEV;
 407
 408         ret = fpu__unlazy_stopped(target);
 409         if (ret)
 410                 return ret;
 411
 412         xsave = &target->thread.fpu.state->xsave;
 413
 414         /*
 415          * Copy the 48bytes defined by the software first into the xstate
 416          * memory layout in the thread struct, so that we can copy the entire
 417          * xstateregs to the user using one user_regset_copyout().
 418          */
 419         memcpy(&xsave->i387.sw_reserved,
 420                 xstate_fx_sw_bytes, sizeof(xstate_fx_sw_bytes));
 421         /*
 422          * Copy the xstate memory layout.
 423          */
 424         ret = user_regset_copyout(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
 425         return ret;
 426 }
 427
 428 int xstateregs_set(struct task_struct *target, const struct user_regset *regset,
 429                   unsigned int pos, unsigned int count,
 430                   const void *kbuf, const void __user *ubuf)
 431 {
 432         struct xsave_struct *xsave;
 433         int ret;
 434
 435         if (!cpu_has_xsave)
 436                 return -ENODEV;
 437
 438         ret = fpu__unlazy_stopped(target);
 439         if (ret)
 440                 return ret;
 441
 442         xsave = &target->thread.fpu.state->xsave;
 443
 444         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, xsave, 0, -1);
 445         /*
 446          * mxcsr reserved bits must be masked to zero for security reasons.
 447          */
 448         xsave->i387.mxcsr &= mxcsr_feature_mask;
 449         xsave->xsave_hdr.xstate_bv &= pcntxt_mask;
 450         /*
 451          * These bits must be zero.
 452          */
 453         memset(&xsave->xsave_hdr.reserved, 0, 48);
 454         return ret;
 455 }
 456
 457 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
 458
 459 /*
 460  * FPU tag word conversions.
 461  */
 462
 463 static inline unsigned short twd_i387_to_fxsr(unsigned short twd)
 464 {
 465         unsigned int tmp; /* to avoid 16 bit prefixes in the code */
 466
 467         /* Transform each pair of bits into 01 (valid) or 00 (empty) */
 468         tmp = ~twd;
 469         tmp = (tmp | (tmp>>1)) & 0x5555; /* 0V0V0V0V0V0V0V0V */
 470         /* and move the valid bits to the lower byte. */
 471         tmp = (tmp | (tmp >> 1)) & 0x3333; /* 00VV00VV00VV00VV */
 472         tmp = (tmp | (tmp >> 2)) & 0x0f0f; /* 0000VVVV0000VVVV */
 473         tmp = (tmp | (tmp >> 4)) & 0x00ff; /* 00000000VVVVVVVV */
 474
 475         return tmp;
 476 }
 477
 478 #define FPREG_ADDR(f, n)        ((void *)&(f)->st_space + (n) * 16)
 479 #define FP_EXP_TAG_VALID        0
 480 #define FP_EXP_TAG_ZERO         1
 481 #define FP_EXP_TAG_SPECIAL      2
 482 #define FP_EXP_TAG_EMPTY        3
 483
 484 static inline u32 twd_fxsr_to_i387(struct i387_fxsave_struct *fxsave)
 485 {
 486         struct _fpxreg *st;
 487         u32 tos = (fxsave->swd >> 11) & 7;
 488         u32 twd = (unsigned long) fxsave->twd;
 489         u32 tag;
 490         u32 ret = 0xffff0000u;
 491         int i;
 492
 493         for (i = 0; i < 8; i++, twd >>= 1) {
 494                 if (twd & 0x1) {
 495                         st = FPREG_ADDR(fxsave, (i - tos) & 7);
 496
 497                         switch (st->exponent & 0x7fff) {
 498                         case 0x7fff:
 499                                 tag = FP_EXP_TAG_SPECIAL;
 500                                 break;
 501                         case 0x0000:
 502                                 if (!st->significand[0] &&
 503                                     !st->significand[1] &&
 504                                     !st->significand[2] &&
 505                                     !st->significand[3])
 506                                         tag = FP_EXP_TAG_ZERO;
 507                                 else
 508                                         tag = FP_EXP_TAG_SPECIAL;
 509                                 break;
 510                         default:
 511                                 if (st->significand[3] & 0x8000)
 512                                         tag = FP_EXP_TAG_VALID;
 513                                 else
 514                                         tag = FP_EXP_TAG_SPECIAL;
 515                                 break;
 516                         }
 517                 } else {
 518                         tag = FP_EXP_TAG_EMPTY;
 519                 }
 520                 ret |= tag << (2 * i);
 521         }
 522         return ret;
 523 }
 524
 525 /*
 526  * FXSR floating point environment conversions.
 527  */
 528
 529 void
 530 convert_from_fxsr(struct user_i387_ia32_struct *env, struct task_struct *tsk)
 531 {
 532         struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
 533         struct _fpreg *to = (struct _fpreg *) &env->st_space[0];
 534         struct _fpxreg *from = (struct _fpxreg *) &fxsave->st_space[0];
 535         int i;
 536
 537         env->cwd = fxsave->cwd | 0xffff0000u;
 538         env->swd = fxsave->swd | 0xffff0000u;
 539         env->twd = twd_fxsr_to_i387(fxsave);
 540
 541 #ifdef CONFIG_X86_64
 542         env->fip = fxsave->rip;
 543         env->foo = fxsave->rdp;
 544         /*
 545          * should be actually ds/cs at fpu exception time, but
 546          * that information is not available in 64bit mode.
 547          */
 548         env->fcs = task_pt_regs(tsk)->cs;
 549         if (tsk == current) {
 550                 savesegment(ds, env->fos);
 551         } else {
 552                 env->fos = tsk->thread.ds;
 553         }
 554         env->fos |= 0xffff0000;
 555 #else
 556         env->fip = fxsave->fip;
 557         env->fcs = (u16) fxsave->fcs | ((u32) fxsave->fop << 16);
 558         env->foo = fxsave->foo;
 559         env->fos = fxsave->fos;
 560 #endif
 561
 562         for (i = 0; i < 8; ++i)
 563                 memcpy(&to[i], &from[i], sizeof(to[0]));
 564 }
 565
 566 void convert_to_fxsr(struct task_struct *tsk,
 567                      const struct user_i387_ia32_struct *env)
 568
 569 {
 570         struct i387_fxsave_struct *fxsave = &tsk->thread.fpu.state->fxsave;
 571         struct _fpreg *from = (struct _fpreg *) &env->st_space[0];
 572         struct _fpxreg *to = (struct _fpxreg *) &fxsave->st_space[0];
 573         int i;
 574
 575         fxsave->cwd = env->cwd;
 576         fxsave->swd = env->swd;
 577         fxsave->twd = twd_i387_to_fxsr(env->twd);
 578         fxsave->fop = (u16) ((u32) env->fcs >> 16);
 579 #ifdef CONFIG_X86_64
 580         fxsave->rip = env->fip;
 581         fxsave->rdp = env->foo;
 582         /* cs and ds ignored */
 583 #else
 584         fxsave->fip = env->fip;
 585         fxsave->fcs = (env->fcs & 0xffff);
 586         fxsave->foo = env->foo;
 587         fxsave->fos = env->fos;
 588 #endif
 589
 590         for (i = 0; i < 8; ++i)
 591                 memcpy(&to[i], &from[i], sizeof(from[0]));
 592 }
 593
 594 int fpregs_get(struct task_struct *target, const struct user_regset *regset,
 595                unsigned int pos, unsigned int count,
 596                void *kbuf, void __user *ubuf)
 597 {
 598         struct user_i387_ia32_struct env;
 599         int ret;
 600
 601         ret = fpu__unlazy_stopped(target);
 602         if (ret)
 603                 return ret;
 604
 605         if (!static_cpu_has(X86_FEATURE_FPU))
 606                 return fpregs_soft_get(target, regset, pos, count, kbuf, ubuf);
 607
 608         if (!cpu_has_fxsr)
 609                 return user_regset_copyout(&pos, &count, &kbuf, &ubuf,
 610                                            &target->thread.fpu.state->fsave, 0,
 611                                            -1);
 612
 613         sanitize_i387_state(target);
 614
 615         if (kbuf && pos == 0 && count == sizeof(env)) {
 616                 convert_from_fxsr(kbuf, target);
 617                 return 0;
 618         }
 619
 620         convert_from_fxsr(&env, target);
 621
 622         return user_regset_copyout(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 623 }
 624
 625 int fpregs_set(struct task_struct *target, const struct user_regset *regset,
 626                unsigned int pos, unsigned int count,
 627                const void *kbuf, const void __user *ubuf)
 628 {
 629         struct user_i387_ia32_struct env;
 630         int ret;
 631
 632         ret = fpu__unlazy_stopped(target);
 633         if (ret)
 634                 return ret;
 635
 636         sanitize_i387_state(target);
 637
 638         if (!static_cpu_has(X86_FEATURE_FPU))
 639                 return fpregs_soft_set(target, regset, pos, count, kbuf, ubuf);
 640
 641         if (!cpu_has_fxsr)
 642                 return user_regset_copyin(&pos, &count, &kbuf, &ubuf,
 643                                           &target->thread.fpu.state->fsave, 0,
 644                                           -1);
 645
 646         if (pos > 0 || count < sizeof(env))
 647                 convert_from_fxsr(&env, target);
 648
 649         ret = user_regset_copyin(&pos, &count, &kbuf, &ubuf, &env, 0, -1);
 650         if (!ret)
 651                 convert_to_fxsr(target, &env);
 652
 653         /*
 654          * update the header bit in the xsave header, indicating the
 655          * presence of FP.
 656          */
 657         if (cpu_has_xsave)
 658                 target->thread.fpu.state->xsave.xsave_hdr.xstate_bv |= XSTATE_FP;
 659         return ret;
 660 }
 661
 662 /*
 663  * FPU state for core dumps.
 664  * This is only used for a.out dumps now.
 665  * It is declared generically using elf_fpregset_t (which is
 666  * struct user_i387_struct) but is in fact only used for 32-bit
 667  * dumps, so on 64-bit it is really struct user_i387_ia32_struct.
 668  */
 669 int dump_fpu(struct pt_regs *regs, struct user_i387_struct *fpu)
 670 {
 671         struct task_struct *tsk = current;
 672         int fpvalid;
 673
 674         fpvalid = !!used_math();
 675         if (fpvalid)
 676                 fpvalid = !fpregs_get(tsk, NULL,
 677                                       0, sizeof(struct user_i387_ia32_struct),
 678                                       fpu, NULL);
 679
 680         return fpvalid;
 681 }
 682 EXPORT_SYMBOL(dump_fpu);
 683
 684 #endif  /* CONFIG_X86_32 || CONFIG_IA32_EMULATION */