arch/x86/include/asm/fpu-internal.h

   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  * x86-64 work by Andi Kleen 2002
   8  */
   9
  10 #ifndef _FPU_INTERNAL_H
  11 #define _FPU_INTERNAL_H
  12
  13 #include <linux/kernel_stat.h>
  14 #include <linux/regset.h>
  15 #include <linux/compat.h>
  16 #include <linux/slab.h>
  17 #include <asm/asm.h>
  18 #include <asm/cpufeature.h>
  19 #include <asm/processor.h>
  20 #include <asm/sigcontext.h>
  21 #include <asm/user.h>
  22 #include <asm/uaccess.h>
  23 #include <asm/xsave.h>
  24 #include <asm/smap.h>
  25
  26 #ifdef CONFIG_X86_64
  27 # include <asm/sigcontext32.h>
  28 # include <asm/user32.h>
  29 struct ksignal;
  30 int ia32_setup_rt_frame(int sig, struct ksignal *ksig,
  31                         compat_sigset_t *set, struct pt_regs *regs);
  32 int ia32_setup_frame(int sig, struct ksignal *ksig,
  33                      compat_sigset_t *set, struct pt_regs *regs);
  34 #else
  35 # define user_i387_ia32_struct  user_i387_struct
  36 # define user32_fxsr_struct     user_fxsr_struct
  37 # define ia32_setup_frame       __setup_frame
  38 # define ia32_setup_rt_frame    __setup_rt_frame
  39 #endif
  40
  41 extern unsigned int mxcsr_feature_mask;
  42 extern void fpu_init(void);
  43 extern void eager_fpu_init(void);
  44
  45 DECLARE_PER_CPU(struct task_struct *, fpu_owner_task);
  46
  47 extern void convert_from_fxsr(struct user_i387_ia32_struct *env,
  48                               struct task_struct *tsk);
  49 extern void convert_to_fxsr(struct task_struct *tsk,
  50                             const struct user_i387_ia32_struct *env);
  51
  52 extern user_regset_active_fn fpregs_active, xfpregs_active;
  53 extern user_regset_get_fn fpregs_get, xfpregs_get, fpregs_soft_get,
  54                                 xstateregs_get;
  55 extern user_regset_set_fn fpregs_set, xfpregs_set, fpregs_soft_set,
  56                                  xstateregs_set;
  57
  58 /*
  59  * xstateregs_active == fpregs_active. Please refer to the comment
  60  * at the definition of fpregs_active.
  61  */
  62 #define xstateregs_active       fpregs_active
  63
  64 #ifdef CONFIG_MATH_EMULATION
  65 extern void finit_soft_fpu(struct i387_soft_struct *soft);
  66 #else
  67 static inline void finit_soft_fpu(struct i387_soft_struct *soft) {}
  68 #endif
  69
  70 static inline int is_ia32_compat_frame(void)
  71 {
  72         return config_enabled(CONFIG_IA32_EMULATION) &&
  73                test_thread_flag(TIF_IA32);
  74 }
  75
  76 static inline int is_ia32_frame(void)
  77 {
  78         return config_enabled(CONFIG_X86_32) || is_ia32_compat_frame();
  79 }
  80
  81 static inline int is_x32_frame(void)
  82 {
  83         return config_enabled(CONFIG_X86_X32_ABI) && test_thread_flag(TIF_X32);
  84 }
  85
  86 #define X87_FSW_ES (1 << 7)     /* Exception Summary */
  87
  88 static __always_inline __pure bool use_eager_fpu(void)
  89 {
  90         return static_cpu_has(X86_FEATURE_EAGER_FPU);
  91 }
  92
  93 static __always_inline __pure bool use_xsaveopt(void)
  94 {
  95         return static_cpu_has(X86_FEATURE_XSAVEOPT);
  96 }
  97
  98 static __always_inline __pure bool use_xsave(void)
  99 {
 100         return static_cpu_has(X86_FEATURE_XSAVE);
 101 }
 102
 103 static __always_inline __pure bool use_fxsr(void)
 104 {
 105         return static_cpu_has(X86_FEATURE_FXSR);
 106 }
 107
 108 static inline void fx_finit(struct i387_fxsave_struct *fx)
 109 {
 110         memset(fx, 0, xstate_size);
 111         fx->cwd = 0x37f;
 112         fx->mxcsr = MXCSR_DEFAULT;
 113 }
 114
 115 extern void __sanitize_i387_state(struct task_struct *);
 116
 117 static inline void sanitize_i387_state(struct task_struct *tsk)
 118 {
 119         if (!use_xsaveopt())
 120                 return;
 121         __sanitize_i387_state(tsk);
 122 }
 123
 124 #define user_insn(insn, output, input...)                               \
 125 ({                                                                      \
 126         int err;                                                        \
 127         asm volatile(ASM_STAC "\n"                                      \
 128                      "1:" #insn "\n\t"                                  \
 129                      "2: " ASM_CLAC "\n"                                \
 130                      ".section .fixup,\"ax\"\n"                         \
 131                      "3:  movl $-1,%[err]\n"                            \
 132                      "    jmp  2b\n"                                    \
 133                      ".previous\n"                                      \
 134                      _ASM_EXTABLE(1b, 3b)                               \
 135                      : [err] "=r" (err), output                         \
 136                      : "0"(0), input);                                  \
 137         err;                                                            \
 138 })
 139
 140 #define check_insn(insn, output, input...)                              \
 141 ({                                                                      \
 142         int err;                                                        \
 143         asm volatile("1:" #insn "\n\t"                                  \
 144                      "2:\n"                                             \
 145                      ".section .fixup,\"ax\"\n"                         \
 146                      "3:  movl $-1,%[err]\n"                            \
 147                      "    jmp  2b\n"                                    \
 148                      ".previous\n"                                      \
 149                      _ASM_EXTABLE(1b, 3b)                               \
 150                      : [err] "=r" (err), output                         \
 151                      : "0"(0), input);                                  \
 152         err;                                                            \
 153 })
 154
 155 static inline int fsave_user(struct i387_fsave_struct __user *fx)
 156 {
 157         return user_insn(fnsave %[fx]; fwait,  [fx] "=m" (*fx), "m" (*fx));
 158 }
 159
 160 static inline int fxsave_user(struct i387_fxsave_struct __user *fx)
 161 {
 162         if (config_enabled(CONFIG_X86_32))
 163                 return user_insn(fxsave %[fx], [fx] "=m" (*fx), "m" (*fx));
 164         else if (config_enabled(CONFIG_AS_FXSAVEQ))
 165                 return user_insn(fxsaveq %[fx], [fx] "=m" (*fx), "m" (*fx));
 166
 167         /* See comment in fpu_fxsave() below. */
 168         return user_insn(rex64/fxsave (%[fx]), "=m" (*fx), [fx] "R" (fx));
 169 }
 170
 171 static inline int fxrstor_checking(struct i387_fxsave_struct *fx)
 172 {
 173         if (config_enabled(CONFIG_X86_32))
 174                 return check_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
 175         else if (config_enabled(CONFIG_AS_FXSAVEQ))
 176                 return check_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
 177
 178         /* See comment in fpu_fxsave() below. */
 179         return check_insn(rex64/fxrstor (%[fx]), "=m" (*fx), [fx] "R" (fx),
 180                           "m" (*fx));
 181 }
 182
 183 static inline int fxrstor_user(struct i387_fxsave_struct __user *fx)
 184 {
 185         if (config_enabled(CONFIG_X86_32))
 186                 return user_insn(fxrstor %[fx], "=m" (*fx), [fx] "m" (*fx));
 187         else if (config_enabled(CONFIG_AS_FXSAVEQ))
 188                 return user_insn(fxrstorq %[fx], "=m" (*fx), [fx] "m" (*fx));
 189
 190         /* See comment in fpu_fxsave() below. */
 191         return user_insn(rex64/fxrstor (%[fx]), "=m" (*fx), [fx] "R" (fx),
 192                           "m" (*fx));
 193 }
 194
 195 static inline int frstor_checking(struct i387_fsave_struct *fx)
 196 {
 197         return check_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
 198 }
 199
 200 static inline int frstor_user(struct i387_fsave_struct __user *fx)
 201 {
 202         return user_insn(frstor %[fx], "=m" (*fx), [fx] "m" (*fx));
 203 }
 204
 205 static inline void fpu_fxsave(struct fpu *fpu)
 206 {
 207         if (config_enabled(CONFIG_X86_32))
 208                 asm volatile( "fxsave %[fx]" : [fx] "=m" (fpu->state->fxsave));
 209         else if (config_enabled(CONFIG_AS_FXSAVEQ))
 210                 asm volatile("fxsaveq %0" : "=m" (fpu->state->fxsave));
 211         else {
 212                 /* Using "rex64; fxsave %0" is broken because, if the memory
 213                  * operand uses any extended registers for addressing, a second
 214                  * REX prefix will be generated (to the assembler, rex64
 215                  * followed by semicolon is a separate instruction), and hence
 216                  * the 64-bitness is lost.
 217                  *
 218                  * Using "fxsaveq %0" would be the ideal choice, but is only
 219                  * supported starting with gas 2.16.
 220                  *
 221                  * Using, as a workaround, the properly prefixed form below
 222                  * isn't accepted by any binutils version so far released,
 223                  * complaining that the same type of prefix is used twice if
 224                  * an extended register is needed for addressing (fix submitted
 225                  * to mainline 2005-11-21).
 226                  *
 227                  *  asm volatile("rex64/fxsave %0" : "=m" (fpu->state->fxsave));
 228                  *
 229                  * This, however, we can work around by forcing the compiler to
 230                  * select an addressing mode that doesn't require extended
 231                  * registers.
 232                  */
 233                 asm volatile( "rex64/fxsave (%[fx])"
 234                              : "=m" (fpu->state->fxsave)
 235                              : [fx] "R" (&fpu->state->fxsave));
 236         }
 237 }
 238
 239 /*
 240  * These must be called with preempt disabled. Returns
 241  * 'true' if the FPU state is still intact.
 242  */
 243 static inline int fpu_save_init(struct fpu *fpu)
 244 {
 245         if (use_xsave()) {
 246                 fpu_xsave(fpu);
 247
 248                 /*
 249                  * xsave header may indicate the init state of the FP.
 250                  */
 251                 if (!(fpu->state->xsave.xsave_hdr.xstate_bv & XSTATE_FP))
 252                         return 1;
 253         } else if (use_fxsr()) {
 254                 fpu_fxsave(fpu);
 255         } else {
 256                 asm volatile("fnsave %[fx]; fwait"
 257                              : [fx] "=m" (fpu->state->fsave));
 258                 return 0;
 259         }
 260
 261         /*
 262          * If exceptions are pending, we need to clear them so
 263          * that we don't randomly get exceptions later.
 264          *
 265          * FIXME! Is this perhaps only true for the old-style
 266          * irq13 case? Maybe we could leave the x87 state
 267          * intact otherwise?
 268          */
 269         if (unlikely(fpu->state->fxsave.swd & X87_FSW_ES)) {
 270                 asm volatile("fnclex");
 271                 return 0;
 272         }
 273         return 1;
 274 }
 275
 276 static inline int __save_init_fpu(struct task_struct *tsk)
 277 {
 278         return fpu_save_init(&tsk->thread.fpu);
 279 }
 280
 281 static inline int fpu_restore_checking(struct fpu *fpu)
 282 {
 283         if (use_xsave())
 284                 return fpu_xrstor_checking(&fpu->state->xsave);
 285         else if (use_fxsr())
 286                 return fxrstor_checking(&fpu->state->fxsave);
 287         else
 288                 return frstor_checking(&fpu->state->fsave);
 289 }
 290
 291 static inline int restore_fpu_checking(struct task_struct *tsk)
 292 {
 293         /* AMD K7/K8 CPUs don't save/restore FDP/FIP/FOP unless an exception
 294            is pending.  Clear the x87 state here by setting it to fixed
 295            values. "m" is a random variable that should be in L1 */
 296         alternative_input(
 297                 ASM_NOP8 ASM_NOP2,
 298                 "emms\n\t"              /* clear stack tags */
 299                 "fildl %P[addr]",       /* set F?P to defined value */
 300                 X86_FEATURE_FXSAVE_LEAK,
 301                 [addr] "m" (tsk->thread.fpu.has_fpu));
 302
 303         return fpu_restore_checking(&tsk->thread.fpu);
 304 }
 305
 306 /*
 307  * Software FPU state helpers. Careful: these need to
 308  * be preemption protection *and* they need to be
 309  * properly paired with the CR0.TS changes!
 310  */
 311 static inline int __thread_has_fpu(struct task_struct *tsk)
 312 {
 313         return tsk->thread.fpu.has_fpu;
 314 }
 315
 316 /* Must be paired with an 'stts' after! */
 317 static inline void __thread_clear_has_fpu(struct task_struct *tsk)
 318 {
 319         tsk->thread.fpu.has_fpu = 0;
 320         this_cpu_write(fpu_owner_task, NULL);
 321 }
 322
 323 /* Must be paired with a 'clts' before! */
 324 static inline void __thread_set_has_fpu(struct task_struct *tsk)
 325 {
 326         tsk->thread.fpu.has_fpu = 1;
 327         this_cpu_write(fpu_owner_task, tsk);
 328 }
 329
 330 /*
 331  * Encapsulate the CR0.TS handling together with the
 332  * software flag.
 333  *
 334  * These generally need preemption protection to work,
 335  * do try to avoid using these on their own.
 336  */
 337 static inline void __thread_fpu_end(struct task_struct *tsk)
 338 {
 339         __thread_clear_has_fpu(tsk);
 340         if (!use_eager_fpu())
 341                 stts();
 342 }
 343
 344 static inline void __thread_fpu_begin(struct task_struct *tsk)
 345 {
 346         if (!static_cpu_has_safe(X86_FEATURE_EAGER_FPU))
 347                 clts();
 348         __thread_set_has_fpu(tsk);
 349 }
 350
 351 static inline void __drop_fpu(struct task_struct *tsk)
 352 {
 353         if (__thread_has_fpu(tsk)) {
 354                 /* Ignore delayed exceptions from user space */
 355                 asm volatile("1: fwait\n"
 356                              "2:\n"
 357                              _ASM_EXTABLE(1b, 2b));
 358                 __thread_fpu_end(tsk);
 359         }
 360 }
 361
 362 static inline void drop_fpu(struct task_struct *tsk)
 363 {
 364         /*
 365          * Forget coprocessor state..
 366          */
 367         preempt_disable();
 368         tsk->fpu_counter = 0;
 369         __drop_fpu(tsk);
 370         clear_used_math();
 371         preempt_enable();
 372 }
 373
 374 static inline void drop_init_fpu(struct task_struct *tsk)
 375 {
 376         if (!use_eager_fpu())
 377                 drop_fpu(tsk);
 378         else {
 379                 if (use_xsave())
 380                         xrstor_state(init_xstate_buf, -1);
 381                 else
 382                         fxrstor_checking(&init_xstate_buf->i387);
 383         }
 384 }
 385
 386 /*
 387  * FPU state switching for scheduling.
 388  *
 389  * This is a two-stage process:
 390  *
 391  *  - switch_fpu_prepare() saves the old state and
 392  *    sets the new state of the CR0.TS bit. This is
 393  *    done within the context of the old process.
 394  *
 395  *  - switch_fpu_finish() restores the new state as
 396  *    necessary.
 397  */
 398 typedef struct { int preload; } fpu_switch_t;
 399
 400 /*
 401  * Must be run with preemption disabled: this clears the fpu_owner_task,
 402  * on this CPU.
 403  *
 404  * This will disable any lazy FPU state restore of the current FPU state,
 405  * but if the current thread owns the FPU, it will still be saved by.
 406  */
 407 static inline void __cpu_disable_lazy_restore(unsigned int cpu)
 408 {
 409         per_cpu(fpu_owner_task, cpu) = NULL;
 410 }
 411
 412 static inline int fpu_lazy_restore(struct task_struct *new, unsigned int cpu)
 413 {
 414         return new == this_cpu_read_stable(fpu_owner_task) &&
 415                 cpu == new->thread.fpu.last_cpu;
 416 }
 417
 418 static inline fpu_switch_t switch_fpu_prepare(struct task_struct *old, struct task_struct *new, int cpu)
 419 {
 420         fpu_switch_t fpu;
 421
 422         /*
 423          * If the task has used the math, pre-load the FPU on xsave processors
 424          * or if the past 5 consecutive context-switches used math.
 425          */
 426         fpu.preload = tsk_used_math(new) && (use_eager_fpu() ||
 427                                              new->fpu_counter > 5);
 428         if (__thread_has_fpu(old)) {
 429                 if (!__save_init_fpu(old))
 430                         cpu = ~0;
 431                 old->thread.fpu.last_cpu = cpu;
 432                 old->thread.fpu.has_fpu = 0;    /* But leave fpu_owner_task! */
 433
 434                 /* Don't change CR0.TS if we just switch! */
 435                 if (fpu.preload) {
 436                         new->fpu_counter++;
 437                         __thread_set_has_fpu(new);
 438                         prefetch(new->thread.fpu.state);
 439                 } else if (!use_eager_fpu())
 440                         stts();
 441         } else {
 442                 old->fpu_counter = 0;
 443                 old->thread.fpu.last_cpu = ~0;
 444                 if (fpu.preload) {
 445                         new->fpu_counter++;
 446                         if (!use_eager_fpu() && fpu_lazy_restore(new, cpu))
 447                                 fpu.preload = 0;
 448                         else
 449                                 prefetch(new->thread.fpu.state);
 450                         __thread_fpu_begin(new);
 451                 }
 452         }
 453         return fpu;
 454 }
 455
 456 /*
 457  * By the time this gets called, we've already cleared CR0.TS and
 458  * given the process the FPU if we are going to preload the FPU
 459  * state - all we need to do is to conditionally restore the register
 460  * state itself.
 461  */
 462 static inline void switch_fpu_finish(struct task_struct *new, fpu_switch_t fpu)
 463 {
 464         if (fpu.preload) {
 465                 if (unlikely(restore_fpu_checking(new)))
 466                         drop_init_fpu(new);
 467         }
 468 }
 469
 470 /*
 471  * Signal frame handlers...
 472  */
 473 extern int save_xstate_sig(void __user *buf, void __user *fx, int size);
 474 extern int __restore_xstate_sig(void __user *buf, void __user *fx, int size);
 475
 476 static inline int xstate_sigframe_size(void)
 477 {
 478         return use_xsave() ? xstate_size + FP_XSTATE_MAGIC2_SIZE : xstate_size;
 479 }
 480
 481 static inline int restore_xstate_sig(void __user *buf, int ia32_frame)
 482 {
 483         void __user *buf_fx = buf;
 484         int size = xstate_sigframe_size();
 485
 486         if (ia32_frame && use_fxsr()) {
 487                 buf_fx = buf + sizeof(struct i387_fsave_struct);
 488                 size += sizeof(struct i387_fsave_struct);
 489         }
 490
 491         return __restore_xstate_sig(buf, buf_fx, size);
 492 }
 493
 494 /*
 495  * Need to be preemption-safe.
 496  *
 497  * NOTE! user_fpu_begin() must be used only immediately before restoring
 498  * it. This function does not do any save/restore on their own.
 499  */
 500 static inline void user_fpu_begin(void)
 501 {
 502         preempt_disable();
 503         if (!user_has_fpu())
 504                 __thread_fpu_begin(current);
 505         preempt_enable();
 506 }
 507
 508 static inline void __save_fpu(struct task_struct *tsk)
 509 {
 510         if (use_xsave())
 511                 xsave_state(&tsk->thread.fpu.state->xsave, -1);
 512         else
 513                 fpu_fxsave(&tsk->thread.fpu);
 514 }
 515
 516 /*
 517  * These disable preemption on their own and are safe
 518  */
 519 static inline void save_init_fpu(struct task_struct *tsk)
 520 {
 521         WARN_ON_ONCE(!__thread_has_fpu(tsk));
 522
 523         if (use_eager_fpu()) {
 524                 __save_fpu(tsk);
 525                 return;
 526         }
 527
 528         preempt_disable();
 529         __save_init_fpu(tsk);
 530         __thread_fpu_end(tsk);
 531         preempt_enable();
 532 }
 533
 534 /*
 535  * i387 state interaction
 536  */
 537 static inline unsigned short get_fpu_cwd(struct task_struct *tsk)
 538 {
 539         if (cpu_has_fxsr) {
 540                 return tsk->thread.fpu.state->fxsave.cwd;
 541         } else {
 542                 return (unsigned short)tsk->thread.fpu.state->fsave.cwd;
 543         }
 544 }
 545
 546 static inline unsigned short get_fpu_swd(struct task_struct *tsk)
 547 {
 548         if (cpu_has_fxsr) {
 549                 return tsk->thread.fpu.state->fxsave.swd;
 550         } else {
 551                 return (unsigned short)tsk->thread.fpu.state->fsave.swd;
 552         }
 553 }
 554
 555 static inline unsigned short get_fpu_mxcsr(struct task_struct *tsk)
 556 {
 557         if (cpu_has_xmm) {
 558                 return tsk->thread.fpu.state->fxsave.mxcsr;
 559         } else {
 560                 return MXCSR_DEFAULT;
 561         }
 562 }
 563
 564 static bool fpu_allocated(struct fpu *fpu)
 565 {
 566         return fpu->state != NULL;
 567 }
 568
 569 static inline int fpu_alloc(struct fpu *fpu)
 570 {
 571         if (fpu_allocated(fpu))
 572                 return 0;
 573         fpu->state = kmem_cache_alloc(task_xstate_cachep, GFP_KERNEL);
 574         if (!fpu->state)
 575                 return -ENOMEM;
 576         WARN_ON((unsigned long)fpu->state & 15);
 577         return 0;
 578 }
 579
 580 static inline void fpu_free(struct fpu *fpu)
 581 {
 582         if (fpu->state) {
 583                 kmem_cache_free(task_xstate_cachep, fpu->state);
 584                 fpu->state = NULL;
 585         }
 586 }
 587
 588 static inline void fpu_copy(struct task_struct *dst, struct task_struct *src)
 589 {
 590         if (use_eager_fpu()) {
 591                 memset(&dst->thread.fpu.state->xsave, 0, xstate_size);
 592                 __save_fpu(dst);
 593         } else {
 594                 struct fpu *dfpu = &dst->thread.fpu;
 595                 struct fpu *sfpu = &src->thread.fpu;
 596
 597                 unlazy_fpu(src);
 598                 memcpy(dfpu->state, sfpu->state, xstate_size);
 599         }
 600 }
 601
 602 static inline unsigned long
 603 alloc_mathframe(unsigned long sp, int ia32_frame, unsigned long *buf_fx,
 604                 unsigned long *size)
 605 {
 606         unsigned long frame_size = xstate_sigframe_size();
 607
 608         *buf_fx = sp = round_down(sp - frame_size, 64);
 609         if (ia32_frame && use_fxsr()) {
 610                 frame_size += sizeof(struct i387_fsave_struct);
 611                 sp -= sizeof(struct i387_fsave_struct);
 612         }
 613
 614         *size = frame_size;
 615         return sp;
 616 }
 617
 618 #endif