arch/x86/kernel/fpu/core.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Copyright (C) 1994 Linus Torvalds
   4  *
   5  *  Pentium III FXSR, SSE support
   6  *  General FPU state handling cleanups
   7  *      Gareth Hughes <gareth@valinux.com>, May 2000
   8  */
   9 #include <asm/fpu/api.h>
  10 #include <asm/fpu/regset.h>
  11 #include <asm/fpu/sched.h>
  12 #include <asm/fpu/signal.h>
  13 #include <asm/fpu/types.h>
  14 #include <asm/traps.h>
  15 #include <asm/irq_regs.h>
  16
  17 #include <uapi/asm/kvm.h>
  18
  19 #include <linux/hardirq.h>
  20 #include <linux/pkeys.h>
  21 #include <linux/vmalloc.h>
  22
  23 #include "context.h"
  24 #include "internal.h"
  25 #include "legacy.h"
  26 #include "xstate.h"
  27
  28 #define CREATE_TRACE_POINTS
  29 #include <asm/trace/fpu.h>
  30
  31 #ifdef CONFIG_X86_64
  32 DEFINE_STATIC_KEY_FALSE(__fpu_state_size_dynamic);
  33 DEFINE_PER_CPU(u64, xfd_state);
  34 #endif
  35
  36 /* The FPU state configuration data for kernel and user space */
  37 struct fpu_state_config fpu_kernel_cfg __ro_after_init;
  38 struct fpu_state_config fpu_user_cfg __ro_after_init;
  39
  40 /*
  41  * Represents the initial FPU state. It's mostly (but not completely) zeroes,
  42  * depending on the FPU hardware format:
  43  */
  44 struct fpstate init_fpstate __ro_after_init;
  45
  46 /* Track in-kernel FPU usage */
  47 static DEFINE_PER_CPU(bool, in_kernel_fpu);
  48
  49 /*
  50  * Track which context is using the FPU on the CPU:
  51  */
  52 DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx);
  53
  54 /*
  55  * Can we use the FPU in kernel mode with the
  56  * whole "kernel_fpu_begin/end()" sequence?
  57  */
  58 bool irq_fpu_usable(void)
  59 {
  60         if (WARN_ON_ONCE(in_nmi()))
  61                 return false;
  62
  63         /* In kernel FPU usage already active? */
  64         if (this_cpu_read(in_kernel_fpu))
  65                 return false;
  66
  67         /*
  68          * When not in NMI or hard interrupt context, FPU can be used in:
  69          *
  70          * - Task context except from within fpregs_lock()'ed critical
  71          *   regions.
  72          *
  73          * - Soft interrupt processing context which cannot happen
  74          *   while in a fpregs_lock()'ed critical region.
  75          */
  76         if (!in_hardirq())
  77                 return true;
  78
  79         /*
  80          * In hard interrupt context it's safe when soft interrupts
  81          * are enabled, which means the interrupt did not hit in
  82          * a fpregs_lock()'ed critical region.
  83          */
  84         return !softirq_count();
  85 }
  86 EXPORT_SYMBOL(irq_fpu_usable);
  87
  88 /*
  89  * Track AVX512 state use because it is known to slow the max clock
  90  * speed of the core.
  91  */
  92 static void update_avx_timestamp(struct fpu *fpu)
  93 {
  94
  95 #define AVX512_TRACKING_MASK    (XFEATURE_MASK_ZMM_Hi256 | XFEATURE_MASK_Hi16_ZMM)
  96
  97         if (fpu->fpstate->regs.xsave.header.xfeatures & AVX512_TRACKING_MASK)
  98                 fpu->avx512_timestamp = jiffies;
  99 }
 100
 101 /*
 102  * Save the FPU register state in fpu->fpstate->regs. The register state is
 103  * preserved.
 104  *
 105  * Must be called with fpregs_lock() held.
 106  *
 107  * The legacy FNSAVE instruction clears all FPU state unconditionally, so
 108  * register state has to be reloaded. That might be a pointless exercise
 109  * when the FPU is going to be used by another task right after that. But
 110  * this only affects 20+ years old 32bit systems and avoids conditionals all
 111  * over the place.
 112  *
 113  * FXSAVE and all XSAVE variants preserve the FPU register state.
 114  */
 115 void save_fpregs_to_fpstate(struct fpu *fpu)
 116 {
 117         if (likely(use_xsave())) {
 118                 os_xsave(fpu->fpstate);
 119                 update_avx_timestamp(fpu);
 120                 return;
 121         }
 122
 123         if (likely(use_fxsr())) {
 124                 fxsave(&fpu->fpstate->regs.fxsave);
 125                 return;
 126         }
 127
 128         /*
 129          * Legacy FPU register saving, FNSAVE always clears FPU registers,
 130          * so we have to reload them from the memory state.
 131          */
 132         asm volatile("fnsave %[fp]; fwait" : [fp] "=m" (fpu->fpstate->regs.fsave));
 133         frstor(&fpu->fpstate->regs.fsave);
 134 }
 135
 136 void restore_fpregs_from_fpstate(struct fpstate *fpstate, u64 mask)
 137 {
 138         /*
 139          * AMD K7/K8 and later CPUs up to Zen don't save/restore
 140          * FDP/FIP/FOP unless an exception is pending. Clear the x87 state
 141          * here by setting it to fixed values.  "m" is a random variable
 142          * that should be in L1.
 143          */
 144         if (unlikely(static_cpu_has_bug(X86_BUG_FXSAVE_LEAK))) {
 145                 asm volatile(
 146                         "fnclex\n\t"
 147                         "emms\n\t"
 148                         "fildl %P[addr]"        /* set F?P to defined value */
 149                         : : [addr] "m" (fpstate));
 150         }
 151
 152         if (use_xsave()) {
 153                 /*
 154                  * Dynamically enabled features are enabled in XCR0, but
 155                  * usage requires also that the corresponding bits in XFD
 156                  * are cleared.  If the bits are set then using a related
 157                  * instruction will raise #NM. This allows to do the
 158                  * allocation of the larger FPU buffer lazy from #NM or if
 159                  * the task has no permission to kill it which would happen
 160                  * via #UD if the feature is disabled in XCR0.
 161                  *
 162                  * XFD state is following the same life time rules as
 163                  * XSTATE and to restore state correctly XFD has to be
 164                  * updated before XRSTORS otherwise the component would
 165                  * stay in or go into init state even if the bits are set
 166                  * in fpstate::regs::xsave::xfeatures.
 167                  */
 168                 xfd_update_state(fpstate);
 169
 170                 /*
 171                  * Restoring state always needs to modify all features
 172                  * which are in @mask even if the current task cannot use
 173                  * extended features.
 174                  *
 175                  * So fpstate->xfeatures cannot be used here, because then
 176                  * a feature for which the task has no permission but was
 177                  * used by the previous task would not go into init state.
 178                  */
 179                 mask = fpu_kernel_cfg.max_features & mask;
 180
 181                 os_xrstor(fpstate, mask);
 182         } else {
 183                 if (use_fxsr())
 184                         fxrstor(&fpstate->regs.fxsave);
 185                 else
 186                         frstor(&fpstate->regs.fsave);
 187         }
 188 }
 189
 190 void fpu_reset_from_exception_fixup(void)
 191 {
 192         restore_fpregs_from_fpstate(&init_fpstate, XFEATURE_MASK_FPSTATE);
 193 }
 194
 195 #if IS_ENABLED(CONFIG_KVM)
 196 static void __fpstate_reset(struct fpstate *fpstate, u64 xfd);
 197
 198 static void fpu_init_guest_permissions(struct fpu_guest *gfpu)
 199 {
 200         struct fpu_state_perm *fpuperm;
 201         u64 perm;
 202
 203         if (!IS_ENABLED(CONFIG_X86_64))
 204                 return;
 205
 206         spin_lock_irq(&current->sighand->siglock);
 207         fpuperm = &current->group_leader->thread.fpu.guest_perm;
 208         perm = fpuperm->__state_perm;
 209
 210         /* First fpstate allocation locks down permissions. */
 211         WRITE_ONCE(fpuperm->__state_perm, perm | FPU_GUEST_PERM_LOCKED);
 212
 213         spin_unlock_irq(&current->sighand->siglock);
 214
 215         gfpu->perm = perm & ~FPU_GUEST_PERM_LOCKED;
 216 }
 217
 218 bool fpu_alloc_guest_fpstate(struct fpu_guest *gfpu)
 219 {
 220         struct fpstate *fpstate;
 221         unsigned int size;
 222
 223         size = fpu_user_cfg.default_size + ALIGN(offsetof(struct fpstate, regs), 64);
 224         fpstate = vzalloc(size);
 225         if (!fpstate)
 226                 return false;
 227
 228         /* Leave xfd to 0 (the reset value defined by spec) */
 229         __fpstate_reset(fpstate, 0);
 230         fpstate_init_user(fpstate);
 231         fpstate->is_valloc      = true;
 232         fpstate->is_guest       = true;
 233
 234         gfpu->fpstate           = fpstate;
 235         gfpu->xfeatures         = fpu_user_cfg.default_features;
 236         gfpu->perm              = fpu_user_cfg.default_features;
 237
 238         /*
 239          * KVM sets the FP+SSE bits in the XSAVE header when copying FPU state
 240          * to userspace, even when XSAVE is unsupported, so that restoring FPU
 241          * state on a different CPU that does support XSAVE can cleanly load
 242          * the incoming state using its natural XSAVE.  In other words, KVM's
 243          * uABI size may be larger than this host's default size.  Conversely,
 244          * the default size should never be larger than KVM's base uABI size;
 245          * all features that can expand the uABI size must be opt-in.
 246          */
 247         gfpu->uabi_size         = sizeof(struct kvm_xsave);
 248         if (WARN_ON_ONCE(fpu_user_cfg.default_size > gfpu->uabi_size))
 249                 gfpu->uabi_size = fpu_user_cfg.default_size;
 250
 251         fpu_init_guest_permissions(gfpu);
 252
 253         return true;
 254 }
 255 EXPORT_SYMBOL_GPL(fpu_alloc_guest_fpstate);
 256
 257 void fpu_free_guest_fpstate(struct fpu_guest *gfpu)
 258 {
 259         struct fpstate *fps = gfpu->fpstate;
 260
 261         if (!fps)
 262                 return;
 263
 264         if (WARN_ON_ONCE(!fps->is_valloc || !fps->is_guest || fps->in_use))
 265                 return;
 266
 267         gfpu->fpstate = NULL;
 268         vfree(fps);
 269 }
 270 EXPORT_SYMBOL_GPL(fpu_free_guest_fpstate);
 271
 272 /*
 273   * fpu_enable_guest_xfd_features - Check xfeatures against guest perm and enable
 274   * @guest_fpu:         Pointer to the guest FPU container
 275   * @xfeatures:         Features requested by guest CPUID
 276   *
 277   * Enable all dynamic xfeatures according to guest perm and requested CPUID.
 278   *
 279   * Return: 0 on success, error code otherwise
 280   */
 281 int fpu_enable_guest_xfd_features(struct fpu_guest *guest_fpu, u64 xfeatures)
 282 {
 283         lockdep_assert_preemption_enabled();
 284
 285         /* Nothing to do if all requested features are already enabled. */
 286         xfeatures &= ~guest_fpu->xfeatures;
 287         if (!xfeatures)
 288                 return 0;
 289
 290         return __xfd_enable_feature(xfeatures, guest_fpu);
 291 }
 292 EXPORT_SYMBOL_GPL(fpu_enable_guest_xfd_features);
 293
 294 #ifdef CONFIG_X86_64
 295 void fpu_update_guest_xfd(struct fpu_guest *guest_fpu, u64 xfd)
 296 {
 297         fpregs_lock();
 298         guest_fpu->fpstate->xfd = xfd;
 299         if (guest_fpu->fpstate->in_use)
 300                 xfd_update_state(guest_fpu->fpstate);
 301         fpregs_unlock();
 302 }
 303 EXPORT_SYMBOL_GPL(fpu_update_guest_xfd);
 304
 305 /**
 306  * fpu_sync_guest_vmexit_xfd_state - Synchronize XFD MSR and software state
 307  *
 308  * Must be invoked from KVM after a VMEXIT before enabling interrupts when
 309  * XFD write emulation is disabled. This is required because the guest can
 310  * freely modify XFD and the state at VMEXIT is not guaranteed to be the
 311  * same as the state on VMENTER. So software state has to be udpated before
 312  * any operation which depends on it can take place.
 313  *
 314  * Note: It can be invoked unconditionally even when write emulation is
 315  * enabled for the price of a then pointless MSR read.
 316  */
 317 void fpu_sync_guest_vmexit_xfd_state(void)
 318 {
 319         struct fpstate *fps = current->thread.fpu.fpstate;
 320
 321         lockdep_assert_irqs_disabled();
 322         if (fpu_state_size_dynamic()) {
 323                 rdmsrl(MSR_IA32_XFD, fps->xfd);
 324                 __this_cpu_write(xfd_state, fps->xfd);
 325         }
 326 }
 327 EXPORT_SYMBOL_GPL(fpu_sync_guest_vmexit_xfd_state);
 328 #endif /* CONFIG_X86_64 */
 329
 330 int fpu_swap_kvm_fpstate(struct fpu_guest *guest_fpu, bool enter_guest)
 331 {
 332         struct fpstate *guest_fps = guest_fpu->fpstate;
 333         struct fpu *fpu = &current->thread.fpu;
 334         struct fpstate *cur_fps = fpu->fpstate;
 335
 336         fpregs_lock();
 337         if (!cur_fps->is_confidential && !test_thread_flag(TIF_NEED_FPU_LOAD))
 338                 save_fpregs_to_fpstate(fpu);
 339
 340         /* Swap fpstate */
 341         if (enter_guest) {
 342                 fpu->__task_fpstate = cur_fps;
 343                 fpu->fpstate = guest_fps;
 344                 guest_fps->in_use = true;
 345         } else {
 346                 guest_fps->in_use = false;
 347                 fpu->fpstate = fpu->__task_fpstate;
 348                 fpu->__task_fpstate = NULL;
 349         }
 350
 351         cur_fps = fpu->fpstate;
 352
 353         if (!cur_fps->is_confidential) {
 354                 /* Includes XFD update */
 355                 restore_fpregs_from_fpstate(cur_fps, XFEATURE_MASK_FPSTATE);
 356         } else {
 357                 /*
 358                  * XSTATE is restored by firmware from encrypted
 359                  * memory. Make sure XFD state is correct while
 360                  * running with guest fpstate
 361                  */
 362                 xfd_update_state(cur_fps);
 363         }
 364
 365         fpregs_mark_activate();
 366         fpregs_unlock();
 367         return 0;
 368 }
 369 EXPORT_SYMBOL_GPL(fpu_swap_kvm_fpstate);
 370
 371 void fpu_copy_guest_fpstate_to_uabi(struct fpu_guest *gfpu, void *buf,
 372                                     unsigned int size, u32 pkru)
 373 {
 374         struct fpstate *kstate = gfpu->fpstate;
 375         union fpregs_state *ustate = buf;
 376         struct membuf mb = { .p = buf, .left = size };
 377
 378         if (cpu_feature_enabled(X86_FEATURE_XSAVE)) {
 379                 __copy_xstate_to_uabi_buf(mb, kstate, pkru, XSTATE_COPY_XSAVE);
 380         } else {
 381                 memcpy(&ustate->fxsave, &kstate->regs.fxsave,
 382                        sizeof(ustate->fxsave));
 383                 /* Make it restorable on a XSAVE enabled host */
 384                 ustate->xsave.header.xfeatures = XFEATURE_MASK_FPSSE;
 385         }
 386 }
 387 EXPORT_SYMBOL_GPL(fpu_copy_guest_fpstate_to_uabi);
 388
 389 int fpu_copy_uabi_to_guest_fpstate(struct fpu_guest *gfpu, const void *buf,
 390                                    u64 xcr0, u32 *vpkru)
 391 {
 392         struct fpstate *kstate = gfpu->fpstate;
 393         const union fpregs_state *ustate = buf;
 394
 395         if (!cpu_feature_enabled(X86_FEATURE_XSAVE)) {
 396                 if (ustate->xsave.header.xfeatures & ~XFEATURE_MASK_FPSSE)
 397                         return -EINVAL;
 398                 if (ustate->fxsave.mxcsr & ~mxcsr_feature_mask)
 399                         return -EINVAL;
 400                 memcpy(&kstate->regs.fxsave, &ustate->fxsave, sizeof(ustate->fxsave));
 401                 return 0;
 402         }
 403
 404         if (ustate->xsave.header.xfeatures & ~xcr0)
 405                 return -EINVAL;
 406
 407         /*
 408          * Nullify @vpkru to preserve its current value if PKRU's bit isn't set
 409          * in the header.  KVM's odd ABI is to leave PKRU untouched in this
 410          * case (all other components are eventually re-initialized).
 411          */
 412         if (!(ustate->xsave.header.xfeatures & XFEATURE_MASK_PKRU))
 413                 vpkru = NULL;
 414
 415         return copy_uabi_from_kernel_to_xstate(kstate, ustate, vpkru);
 416 }
 417 EXPORT_SYMBOL_GPL(fpu_copy_uabi_to_guest_fpstate);
 418 #endif /* CONFIG_KVM */
 419
 420 void kernel_fpu_begin_mask(unsigned int kfpu_mask)
 421 {
 422         preempt_disable();
 423
 424         WARN_ON_FPU(!irq_fpu_usable());
 425         WARN_ON_FPU(this_cpu_read(in_kernel_fpu));
 426
 427         this_cpu_write(in_kernel_fpu, true);
 428
 429         if (!(current->flags & (PF_KTHREAD | PF_USER_WORKER)) &&
 430             !test_thread_flag(TIF_NEED_FPU_LOAD)) {
 431                 set_thread_flag(TIF_NEED_FPU_LOAD);
 432                 save_fpregs_to_fpstate(&current->thread.fpu);
 433         }
 434         __cpu_invalidate_fpregs_state();
 435
 436         /* Put sane initial values into the control registers. */
 437         if (likely(kfpu_mask & KFPU_MXCSR) && boot_cpu_has(X86_FEATURE_XMM))
 438                 ldmxcsr(MXCSR_DEFAULT);
 439
 440         if (unlikely(kfpu_mask & KFPU_387) && boot_cpu_has(X86_FEATURE_FPU))
 441                 asm volatile ("fninit");
 442 }
 443 EXPORT_SYMBOL_GPL(kernel_fpu_begin_mask);
 444
 445 void kernel_fpu_end(void)
 446 {
 447         WARN_ON_FPU(!this_cpu_read(in_kernel_fpu));
 448
 449         this_cpu_write(in_kernel_fpu, false);
 450         preempt_enable();
 451 }
 452 EXPORT_SYMBOL_GPL(kernel_fpu_end);
 453
 454 /*
 455  * Sync the FPU register state to current's memory register state when the
 456  * current task owns the FPU. The hardware register state is preserved.
 457  */
 458 void fpu_sync_fpstate(struct fpu *fpu)
 459 {
 460         WARN_ON_FPU(fpu != &current->thread.fpu);
 461
 462         fpregs_lock();
 463         trace_x86_fpu_before_save(fpu);
 464
 465         if (!test_thread_flag(TIF_NEED_FPU_LOAD))
 466                 save_fpregs_to_fpstate(fpu);
 467
 468         trace_x86_fpu_after_save(fpu);
 469         fpregs_unlock();
 470 }
 471
 472 static inline unsigned int init_fpstate_copy_size(void)
 473 {
 474         if (!use_xsave())
 475                 return fpu_kernel_cfg.default_size;
 476
 477         /* XSAVE(S) just needs the legacy and the xstate header part */
 478         return sizeof(init_fpstate.regs.xsave);
 479 }
 480
 481 static inline void fpstate_init_fxstate(struct fpstate *fpstate)
 482 {
 483         fpstate->regs.fxsave.cwd = 0x37f;
 484         fpstate->regs.fxsave.mxcsr = MXCSR_DEFAULT;
 485 }
 486
 487 /*
 488  * Legacy x87 fpstate state init:
 489  */
 490 static inline void fpstate_init_fstate(struct fpstate *fpstate)
 491 {
 492         fpstate->regs.fsave.cwd = 0xffff037fu;
 493         fpstate->regs.fsave.swd = 0xffff0000u;
 494         fpstate->regs.fsave.twd = 0xffffffffu;
 495         fpstate->regs.fsave.fos = 0xffff0000u;
 496 }
 497
 498 /*
 499  * Used in two places:
 500  * 1) Early boot to setup init_fpstate for non XSAVE systems
 501  * 2) fpu_init_fpstate_user() which is invoked from KVM
 502  */
 503 void fpstate_init_user(struct fpstate *fpstate)
 504 {
 505         if (!cpu_feature_enabled(X86_FEATURE_FPU)) {
 506                 fpstate_init_soft(&fpstate->regs.soft);
 507                 return;
 508         }
 509
 510         xstate_init_xcomp_bv(&fpstate->regs.xsave, fpstate->xfeatures);
 511
 512         if (cpu_feature_enabled(X86_FEATURE_FXSR))
 513                 fpstate_init_fxstate(fpstate);
 514         else
 515                 fpstate_init_fstate(fpstate);
 516 }
 517
 518 static void __fpstate_reset(struct fpstate *fpstate, u64 xfd)
 519 {
 520         /* Initialize sizes and feature masks */
 521         fpstate->size           = fpu_kernel_cfg.default_size;
 522         fpstate->user_size      = fpu_user_cfg.default_size;
 523         fpstate->xfeatures      = fpu_kernel_cfg.default_features;
 524         fpstate->user_xfeatures = fpu_user_cfg.default_features;
 525         fpstate->xfd            = xfd;
 526 }
 527
 528 void fpstate_reset(struct fpu *fpu)
 529 {
 530         /* Set the fpstate pointer to the default fpstate */
 531         fpu->fpstate = &fpu->__fpstate;
 532         __fpstate_reset(fpu->fpstate, init_fpstate.xfd);
 533
 534         /* Initialize the permission related info in fpu */
 535         fpu->perm.__state_perm          = fpu_kernel_cfg.default_features;
 536         fpu->perm.__state_size          = fpu_kernel_cfg.default_size;
 537         fpu->perm.__user_state_size     = fpu_user_cfg.default_size;
 538         /* Same defaults for guests */
 539         fpu->guest_perm = fpu->perm;
 540 }
 541
 542 static inline void fpu_inherit_perms(struct fpu *dst_fpu)
 543 {
 544         if (fpu_state_size_dynamic()) {
 545                 struct fpu *src_fpu = &current->group_leader->thread.fpu;
 546
 547                 spin_lock_irq(&current->sighand->siglock);
 548                 /* Fork also inherits the permissions of the parent */
 549                 dst_fpu->perm = src_fpu->perm;
 550                 dst_fpu->guest_perm = src_fpu->guest_perm;
 551                 spin_unlock_irq(&current->sighand->siglock);
 552         }
 553 }
 554
 555 /* Clone current's FPU state on fork */
 556 int fpu_clone(struct task_struct *dst, unsigned long clone_flags, bool minimal)
 557 {
 558         struct fpu *src_fpu = &current->thread.fpu;
 559         struct fpu *dst_fpu = &dst->thread.fpu;
 560
 561         /* The new task's FPU state cannot be valid in the hardware. */
 562         dst_fpu->last_cpu = -1;
 563
 564         fpstate_reset(dst_fpu);
 565
 566         if (!cpu_feature_enabled(X86_FEATURE_FPU))
 567                 return 0;
 568
 569         /*
 570          * Enforce reload for user space tasks and prevent kernel threads
 571          * from trying to save the FPU registers on context switch.
 572          */
 573         set_tsk_thread_flag(dst, TIF_NEED_FPU_LOAD);
 574
 575         /*
 576          * No FPU state inheritance for kernel threads and IO
 577          * worker threads.
 578          */
 579         if (minimal) {
 580                 /* Clear out the minimal state */
 581                 memcpy(&dst_fpu->fpstate->regs, &init_fpstate.regs,
 582                        init_fpstate_copy_size());
 583                 return 0;
 584         }
 585
 586         /*
 587          * If a new feature is added, ensure all dynamic features are
 588          * caller-saved from here!
 589          */
 590         BUILD_BUG_ON(XFEATURE_MASK_USER_DYNAMIC != XFEATURE_MASK_XTILE_DATA);
 591
 592         /*
 593          * Save the default portion of the current FPU state into the
 594          * clone. Assume all dynamic features to be defined as caller-
 595          * saved, which enables skipping both the expansion of fpstate
 596          * and the copying of any dynamic state.
 597          *
 598          * Do not use memcpy() when TIF_NEED_FPU_LOAD is set because
 599          * copying is not valid when current uses non-default states.
 600          */
 601         fpregs_lock();
 602         if (test_thread_flag(TIF_NEED_FPU_LOAD))
 603                 fpregs_restore_userregs();
 604         save_fpregs_to_fpstate(dst_fpu);
 605         fpregs_unlock();
 606         if (!(clone_flags & CLONE_THREAD))
 607                 fpu_inherit_perms(dst_fpu);
 608
 609         /*
 610          * Children never inherit PASID state.
 611          * Force it to have its init value:
 612          */
 613         if (use_xsave())
 614                 dst_fpu->fpstate->regs.xsave.header.xfeatures &= ~XFEATURE_MASK_PASID;
 615
 616         trace_x86_fpu_copy_src(src_fpu);
 617         trace_x86_fpu_copy_dst(dst_fpu);
 618
 619         return 0;
 620 }
 621
 622 /*
 623  * Whitelist the FPU register state embedded into task_struct for hardened
 624  * usercopy.
 625  */
 626 void fpu_thread_struct_whitelist(unsigned long *offset, unsigned long *size)
 627 {
 628         *offset = offsetof(struct thread_struct, fpu.__fpstate.regs);
 629         *size = fpu_kernel_cfg.default_size;
 630 }
 631
 632 /*
 633  * Drops current FPU state: deactivates the fpregs and
 634  * the fpstate. NOTE: it still leaves previous contents
 635  * in the fpregs in the eager-FPU case.
 636  *
 637  * This function can be used in cases where we know that
 638  * a state-restore is coming: either an explicit one,
 639  * or a reschedule.
 640  */
 641 void fpu__drop(struct fpu *fpu)
 642 {
 643         preempt_disable();
 644
 645         if (fpu == &current->thread.fpu) {
 646                 /* Ignore delayed exceptions from user space */
 647                 asm volatile("1: fwait\n"
 648                              "2:\n"
 649                              _ASM_EXTABLE(1b, 2b));
 650                 fpregs_deactivate(fpu);
 651         }
 652
 653         trace_x86_fpu_dropped(fpu);
 654
 655         preempt_enable();
 656 }
 657
 658 /*
 659  * Clear FPU registers by setting them up from the init fpstate.
 660  * Caller must do fpregs_[un]lock() around it.
 661  */
 662 static inline void restore_fpregs_from_init_fpstate(u64 features_mask)
 663 {
 664         if (use_xsave())
 665                 os_xrstor(&init_fpstate, features_mask);
 666         else if (use_fxsr())
 667                 fxrstor(&init_fpstate.regs.fxsave);
 668         else
 669                 frstor(&init_fpstate.regs.fsave);
 670
 671         pkru_write_default();
 672 }
 673
 674 /*
 675  * Reset current->fpu memory state to the init values.
 676  */
 677 static void fpu_reset_fpregs(void)
 678 {
 679         struct fpu *fpu = &current->thread.fpu;
 680
 681         fpregs_lock();
 682         fpu__drop(fpu);
 683         /*
 684          * This does not change the actual hardware registers. It just
 685          * resets the memory image and sets TIF_NEED_FPU_LOAD so a
 686          * subsequent return to usermode will reload the registers from the
 687          * task's memory image.
 688          *
 689          * Do not use fpstate_init() here. Just copy init_fpstate which has
 690          * the correct content already except for PKRU.
 691          *
 692          * PKRU handling does not rely on the xstate when restoring for
 693          * user space as PKRU is eagerly written in switch_to() and
 694          * flush_thread().
 695          */
 696         memcpy(&fpu->fpstate->regs, &init_fpstate.regs, init_fpstate_copy_size());
 697         set_thread_flag(TIF_NEED_FPU_LOAD);
 698         fpregs_unlock();
 699 }
 700
 701 /*
 702  * Reset current's user FPU states to the init states.  current's
 703  * supervisor states, if any, are not modified by this function.  The
 704  * caller guarantees that the XSTATE header in memory is intact.
 705  */
 706 void fpu__clear_user_states(struct fpu *fpu)
 707 {
 708         WARN_ON_FPU(fpu != &current->thread.fpu);
 709
 710         fpregs_lock();
 711         if (!cpu_feature_enabled(X86_FEATURE_FPU)) {
 712                 fpu_reset_fpregs();
 713                 fpregs_unlock();
 714                 return;
 715         }
 716
 717         /*
 718          * Ensure that current's supervisor states are loaded into their
 719          * corresponding registers.
 720          */
 721         if (xfeatures_mask_supervisor() &&
 722             !fpregs_state_valid(fpu, smp_processor_id()))
 723                 os_xrstor_supervisor(fpu->fpstate);
 724
 725         /* Reset user states in registers. */
 726         restore_fpregs_from_init_fpstate(XFEATURE_MASK_USER_RESTORE);
 727
 728         /*
 729          * Now all FPU registers have their desired values.  Inform the FPU
 730          * state machine that current's FPU registers are in the hardware
 731          * registers. The memory image does not need to be updated because
 732          * any operation relying on it has to save the registers first when
 733          * current's FPU is marked active.
 734          */
 735         fpregs_mark_activate();
 736         fpregs_unlock();
 737 }
 738
 739 void fpu_flush_thread(void)
 740 {
 741         fpstate_reset(&current->thread.fpu);
 742         fpu_reset_fpregs();
 743 }
 744 /*
 745  * Load FPU context before returning to userspace.
 746  */
 747 void switch_fpu_return(void)
 748 {
 749         if (!static_cpu_has(X86_FEATURE_FPU))
 750                 return;
 751
 752         fpregs_restore_userregs();
 753 }
 754 EXPORT_SYMBOL_GPL(switch_fpu_return);
 755
 756 #ifdef CONFIG_X86_DEBUG_FPU
 757 /*
 758  * If current FPU state according to its tracking (loaded FPU context on this
 759  * CPU) is not valid then we must have TIF_NEED_FPU_LOAD set so the context is
 760  * loaded on return to userland.
 761  */
 762 void fpregs_assert_state_consistent(void)
 763 {
 764         struct fpu *fpu = &current->thread.fpu;
 765
 766         if (test_thread_flag(TIF_NEED_FPU_LOAD))
 767                 return;
 768
 769         WARN_ON_FPU(!fpregs_state_valid(fpu, smp_processor_id()));
 770 }
 771 EXPORT_SYMBOL_GPL(fpregs_assert_state_consistent);
 772 #endif
 773
 774 void fpregs_mark_activate(void)
 775 {
 776         struct fpu *fpu = &current->thread.fpu;
 777
 778         fpregs_activate(fpu);
 779         fpu->last_cpu = smp_processor_id();
 780         clear_thread_flag(TIF_NEED_FPU_LOAD);
 781 }
 782
 783 /*
 784  * x87 math exception handling:
 785  */
 786
 787 int fpu__exception_code(struct fpu *fpu, int trap_nr)
 788 {
 789         int err;
 790
 791         if (trap_nr == X86_TRAP_MF) {
 792                 unsigned short cwd, swd;
 793                 /*
 794                  * (~cwd & swd) will mask out exceptions that are not set to unmasked
 795                  * status.  0x3f is the exception bits in these regs, 0x200 is the
 796                  * C1 reg you need in case of a stack fault, 0x040 is the stack
 797                  * fault bit.  We should only be taking one exception at a time,
 798                  * so if this combination doesn't produce any single exception,
 799                  * then we have a bad program that isn't synchronizing its FPU usage
 800                  * and it will suffer the consequences since we won't be able to
 801                  * fully reproduce the context of the exception.
 802                  */
 803                 if (boot_cpu_has(X86_FEATURE_FXSR)) {
 804                         cwd = fpu->fpstate->regs.fxsave.cwd;
 805                         swd = fpu->fpstate->regs.fxsave.swd;
 806                 } else {
 807                         cwd = (unsigned short)fpu->fpstate->regs.fsave.cwd;
 808                         swd = (unsigned short)fpu->fpstate->regs.fsave.swd;
 809                 }
 810
 811                 err = swd & ~cwd;
 812         } else {
 813                 /*
 814                  * The SIMD FPU exceptions are handled a little differently, as there
 815                  * is only a single status/control register.  Thus, to determine which
 816                  * unmasked exception was caught we must mask the exception mask bits
 817                  * at 0x1f80, and then use these to mask the exception bits at 0x3f.
 818                  */
 819                 unsigned short mxcsr = MXCSR_DEFAULT;
 820
 821                 if (boot_cpu_has(X86_FEATURE_XMM))
 822                         mxcsr = fpu->fpstate->regs.fxsave.mxcsr;
 823
 824                 err = ~(mxcsr >> 7) & mxcsr;
 825         }
 826
 827         if (err & 0x001) {      /* Invalid op */
 828                 /*
 829                  * swd & 0x240 == 0x040: Stack Underflow
 830                  * swd & 0x240 == 0x240: Stack Overflow
 831                  * User must clear the SF bit (0x40) if set
 832                  */
 833                 return FPE_FLTINV;
 834         } else if (err & 0x004) { /* Divide by Zero */
 835                 return FPE_FLTDIV;
 836         } else if (err & 0x008) { /* Overflow */
 837                 return FPE_FLTOVF;
 838         } else if (err & 0x012) { /* Denormal, Underflow */
 839                 return FPE_FLTUND;
 840         } else if (err & 0x020) { /* Precision */
 841                 return FPE_FLTRES;
 842         }
 843
 844         /*
 845          * If we're using IRQ 13, or supposedly even some trap
 846          * X86_TRAP_MF implementations, it's possible
 847          * we get a spurious trap, which is not an error.
 848          */
 849         return 0;
 850 }
 851
 852 /*
 853  * Initialize register state that may prevent from entering low-power idle.
 854  * This function will be invoked from the cpuidle driver only when needed.
 855  */
 856 noinstr void fpu_idle_fpregs(void)
 857 {
 858         /* Note: AMX_TILE being enabled implies XGETBV1 support */
 859         if (cpu_feature_enabled(X86_FEATURE_AMX_TILE) &&
 860             (xfeatures_in_use() & XFEATURE_MASK_XTILE)) {
 861                 tile_release();
 862                 __this_cpu_write(fpu_fpregs_owner_ctx, NULL);
 863         }
 864 }