| 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/msr.h> |
| 15 | #include <asm/traps.h> |
| 16 | #include <asm/irq_regs.h> |
| 17 | |
| 18 | #include <uapi/asm/kvm.h> |
| 19 | |
| 20 | #include <linux/hardirq.h> |
| 21 | #include <linux/kvm_types.h> |
| 22 | #include <linux/pkeys.h> |
| 23 | #include <linux/vmalloc.h> |
| 24 | |
| 25 | #include "context.h" |
| 26 | #include "internal.h" |
| 27 | #include "legacy.h" |
| 28 | #include "xstate.h" |
| 29 | |
| 30 | #define CREATE_TRACE_POINTS |
| 31 | #include <asm/trace/fpu.h> |
| 32 | |
| 33 | #ifdef CONFIG_X86_64 |
| 34 | DEFINE_STATIC_KEY_FALSE(__fpu_state_size_dynamic); |
| 35 | DEFINE_PER_CPU(u64, xfd_state); |
| 36 | #endif |
| 37 | |
| 38 | /* The FPU state configuration data for kernel and user space */ |
| 39 | struct fpu_state_config fpu_kernel_cfg __ro_after_init; |
| 40 | struct fpu_state_config fpu_user_cfg __ro_after_init; |
| 41 | struct vcpu_fpu_config guest_default_cfg __ro_after_init; |
| 42 | |
| 43 | /* |
| 44 | * Represents the initial FPU state. It's mostly (but not completely) zeroes, |
| 45 | * depending on the FPU hardware format: |
| 46 | */ |
| 47 | struct fpstate init_fpstate __ro_after_init; |
| 48 | |
| 49 | /* |
| 50 | * Track FPU initialization and kernel-mode usage. 'true' means the FPU is |
| 51 | * initialized and is not currently being used by the kernel: |
| 52 | */ |
| 53 | DEFINE_PER_CPU(bool, kernel_fpu_allowed); |
| 54 | |
| 55 | /* |
| 56 | * Track which context is using the FPU on the CPU: |
| 57 | */ |
| 58 | DEFINE_PER_CPU(struct fpu *, fpu_fpregs_owner_ctx); |
| 59 | |
| 60 | #ifdef CONFIG_X86_DEBUG_FPU |
| 61 | struct fpu *x86_task_fpu(struct task_struct *task) |
| 62 | { |
| 63 | if (WARN_ON_ONCE(task->flags & PF_KTHREAD)) |
| 64 | return NULL; |
| 65 | |
| 66 | return (void *)task + sizeof(*task); |
| 67 | } |
| 68 | #endif |
| 69 | |
| 70 | /* |
| 71 | * Can we use the FPU in kernel mode with the |
| 72 | * whole "kernel_fpu_begin/end()" sequence? |
| 73 | */ |
| 74 | bool irq_fpu_usable(void) |
| 75 | { |
| 76 | if (WARN_ON_ONCE(in_nmi())) |
| 77 | return false; |
| 78 | |
| 79 | /* |
| 80 | * Return false in the following cases: |
| 81 | * |
| 82 | * - FPU is not yet initialized. This can happen only when the call is |
| 83 | * coming from CPU onlining, for example for microcode checksumming. |
| 84 | * - The kernel is already using the FPU, either because of explicit |
| 85 | * nesting (which should never be done), or because of implicit |
| 86 | * nesting when a hardirq interrupted a kernel-mode FPU section. |
| 87 | * |
| 88 | * The single boolean check below handles both cases: |
| 89 | */ |
| 90 | if (!this_cpu_read(kernel_fpu_allowed)) |
| 91 | return false; |
| 92 | |
| 93 | /* |
| 94 | * When not in NMI or hard interrupt context, FPU can be used in: |
| 95 | * |
| 96 | * - Task context except from within fpregs_lock()'ed critical |
| 97 | * regions. |
| 98 | * |
| 99 | * - Soft interrupt processing context which cannot happen |
| 100 | * while in a fpregs_lock()'ed critical region. |
| 101 | */ |
| 102 | if (!in_hardirq()) |
| 103 | return true; |
| 104 | |
| 105 | /* |
| 106 | * In hard interrupt context it's safe when soft interrupts |
| 107 | * are enabled, which means the interrupt did not hit in |
| 108 | * a fpregs_lock()'ed critical region. |
| 109 | */ |
| 110 | return !softirq_count(); |
| 111 | } |
| 112 | EXPORT_SYMBOL(irq_fpu_usable); |
| 113 | |
| 114 | /* |
| 115 | * Track AVX512 state use because it is known to slow the max clock |
| 116 | * speed of the core. |
| 117 | */ |
| 118 | static void update_avx_timestamp(struct fpu *fpu) |
| 119 | { |
| 120 | |
| 121 | #define AVX512_TRACKING_MASK (XFEATURE_MASK_ZMM_Hi256 | XFEATURE_MASK_Hi16_ZMM) |
| 122 | |
| 123 | if (fpu->fpstate->regs.xsave.header.xfeatures & AVX512_TRACKING_MASK) |
| 124 | fpu->avx512_timestamp = jiffies; |
| 125 | } |
| 126 | |
| 127 | /* |
| 128 | * Save the FPU register state in fpu->fpstate->regs. The register state is |
| 129 | * preserved. |
| 130 | * |
| 131 | * Must be called with fpregs_lock() held. |
| 132 | * |
| 133 | * The legacy FNSAVE instruction clears all FPU state unconditionally, so |
| 134 | * register state has to be reloaded. That might be a pointless exercise |
| 135 | * when the FPU is going to be used by another task right after that. But |
| 136 | * this only affects 20+ years old 32bit systems and avoids conditionals all |
| 137 | * over the place. |
| 138 | * |
| 139 | * FXSAVE and all XSAVE variants preserve the FPU register state. |
| 140 | */ |
| 141 | void save_fpregs_to_fpstate(struct fpu *fpu) |
| 142 | { |
| 143 | if (likely(use_xsave())) { |
| 144 | os_xsave(fpstate: fpu->fpstate); |
| 145 | update_avx_timestamp(fpu); |
| 146 | return; |
| 147 | } |
| 148 | |
| 149 | if (likely(use_fxsr())) { |
| 150 | fxsave(fx: &fpu->fpstate->regs.fxsave); |
| 151 | return; |
| 152 | } |
| 153 | |
| 154 | /* |
| 155 | * Legacy FPU register saving, FNSAVE always clears FPU registers, |
| 156 | * so we have to reload them from the memory state. |
| 157 | */ |
| 158 | asm volatile("fnsave %[fp]; fwait" : [fp] "=m" (fpu->fpstate->regs.fsave)); |
| 159 | frstor(fx: &fpu->fpstate->regs.fsave); |
| 160 | } |
| 161 | |
| 162 | void restore_fpregs_from_fpstate(struct fpstate *fpstate, u64 mask) |
| 163 | { |
| 164 | /* |
| 165 | * AMD K7/K8 and later CPUs up to Zen don't save/restore |
| 166 | * FDP/FIP/FOP unless an exception is pending. Clear the x87 state |
| 167 | * here by setting it to fixed values. "m" is a random variable |
| 168 | * that should be in L1. |
| 169 | */ |
| 170 | if (unlikely(static_cpu_has_bug(X86_BUG_FXSAVE_LEAK))) { |
| 171 | asm volatile( |
| 172 | "fnclex\n\t" |
| 173 | "emms\n\t" |
| 174 | "fildl %[addr]" /* set F?P to defined value */ |
| 175 | : : [addr] "m" (*fpstate)); |
| 176 | } |
| 177 | |
| 178 | if (use_xsave()) { |
| 179 | /* |
| 180 | * Dynamically enabled features are enabled in XCR0, but |
| 181 | * usage requires also that the corresponding bits in XFD |
| 182 | * are cleared. If the bits are set then using a related |
| 183 | * instruction will raise #NM. This allows to do the |
| 184 | * allocation of the larger FPU buffer lazy from #NM or if |
| 185 | * the task has no permission to kill it which would happen |
| 186 | * via #UD if the feature is disabled in XCR0. |
| 187 | * |
| 188 | * XFD state is following the same life time rules as |
| 189 | * XSTATE and to restore state correctly XFD has to be |
| 190 | * updated before XRSTORS otherwise the component would |
| 191 | * stay in or go into init state even if the bits are set |
| 192 | * in fpstate::regs::xsave::xfeatures. |
| 193 | */ |
| 194 | xfd_update_state(fpstate); |
| 195 | |
| 196 | /* |
| 197 | * Restoring state always needs to modify all features |
| 198 | * which are in @mask even if the current task cannot use |
| 199 | * extended features. |
| 200 | * |
| 201 | * So fpstate->xfeatures cannot be used here, because then |
| 202 | * a feature for which the task has no permission but was |
| 203 | * used by the previous task would not go into init state. |
| 204 | */ |
| 205 | mask = fpu_kernel_cfg.max_features & mask; |
| 206 | |
| 207 | os_xrstor(fpstate, mask); |
| 208 | } else { |
| 209 | if (use_fxsr()) |
| 210 | fxrstor(fx: &fpstate->regs.fxsave); |
| 211 | else |
| 212 | frstor(fx: &fpstate->regs.fsave); |
| 213 | } |
| 214 | } |
| 215 | |
| 216 | void fpu_reset_from_exception_fixup(void) |
| 217 | { |
| 218 | restore_fpregs_from_fpstate(fpstate: &init_fpstate, XFEATURE_MASK_FPSTATE); |
| 219 | } |
| 220 | |
| 221 | #if IS_ENABLED(CONFIG_KVM) |
| 222 | static void __fpstate_reset(struct fpstate *fpstate); |
| 223 | |
| 224 | static void fpu_lock_guest_permissions(void) |
| 225 | { |
| 226 | struct fpu_state_perm *fpuperm; |
| 227 | u64 perm; |
| 228 | |
| 229 | if (!IS_ENABLED(CONFIG_X86_64)) |
| 230 | return; |
| 231 | |
| 232 | spin_lock_irq(lock: ¤t->sighand->siglock); |
| 233 | fpuperm = &x86_task_fpu(current->group_leader)->guest_perm; |
| 234 | perm = fpuperm->__state_perm; |
| 235 | |
| 236 | /* First fpstate allocation locks down permissions. */ |
| 237 | WRITE_ONCE(fpuperm->__state_perm, perm | FPU_GUEST_PERM_LOCKED); |
| 238 | |
| 239 | spin_unlock_irq(lock: ¤t->sighand->siglock); |
| 240 | } |
| 241 | |
| 242 | bool fpu_alloc_guest_fpstate(struct fpu_guest *gfpu) |
| 243 | { |
| 244 | struct fpstate *fpstate; |
| 245 | unsigned int size; |
| 246 | |
| 247 | size = guest_default_cfg.size + ALIGN(offsetof(struct fpstate, regs), 64); |
| 248 | |
| 249 | fpstate = vzalloc(size); |
| 250 | if (!fpstate) |
| 251 | return false; |
| 252 | |
| 253 | /* Initialize indicators to reflect properties of the fpstate */ |
| 254 | fpstate->is_valloc = true; |
| 255 | fpstate->is_guest = true; |
| 256 | |
| 257 | __fpstate_reset(fpstate); |
| 258 | fpstate_init_user(fpstate); |
| 259 | |
| 260 | gfpu->fpstate = fpstate; |
| 261 | gfpu->xfeatures = guest_default_cfg.features; |
| 262 | |
| 263 | /* |
| 264 | * KVM sets the FP+SSE bits in the XSAVE header when copying FPU state |
| 265 | * to userspace, even when XSAVE is unsupported, so that restoring FPU |
| 266 | * state on a different CPU that does support XSAVE can cleanly load |
| 267 | * the incoming state using its natural XSAVE. In other words, KVM's |
| 268 | * uABI size may be larger than this host's default size. Conversely, |
| 269 | * the default size should never be larger than KVM's base uABI size; |
| 270 | * all features that can expand the uABI size must be opt-in. |
| 271 | */ |
| 272 | gfpu->uabi_size = sizeof(struct kvm_xsave); |
| 273 | if (WARN_ON_ONCE(fpu_user_cfg.default_size > gfpu->uabi_size)) |
| 274 | gfpu->uabi_size = fpu_user_cfg.default_size; |
| 275 | |
| 276 | fpu_lock_guest_permissions(); |
| 277 | |
| 278 | return true; |
| 279 | } |
| 280 | EXPORT_SYMBOL_FOR_KVM(fpu_alloc_guest_fpstate); |
| 281 | |
| 282 | void fpu_free_guest_fpstate(struct fpu_guest *gfpu) |
| 283 | { |
| 284 | struct fpstate *fpstate = gfpu->fpstate; |
| 285 | |
| 286 | if (!fpstate) |
| 287 | return; |
| 288 | |
| 289 | if (WARN_ON_ONCE(!fpstate->is_valloc || !fpstate->is_guest || fpstate->in_use)) |
| 290 | return; |
| 291 | |
| 292 | gfpu->fpstate = NULL; |
| 293 | vfree(addr: fpstate); |
| 294 | } |
| 295 | EXPORT_SYMBOL_FOR_KVM(fpu_free_guest_fpstate); |
| 296 | |
| 297 | /* |
| 298 | * fpu_enable_guest_xfd_features - Check xfeatures against guest perm and enable |
| 299 | * @guest_fpu: Pointer to the guest FPU container |
| 300 | * @xfeatures: Features requested by guest CPUID |
| 301 | * |
| 302 | * Enable all dynamic xfeatures according to guest perm and requested CPUID. |
| 303 | * |
| 304 | * Return: 0 on success, error code otherwise |
| 305 | */ |
| 306 | int fpu_enable_guest_xfd_features(struct fpu_guest *guest_fpu, u64 xfeatures) |
| 307 | { |
| 308 | lockdep_assert_preemption_enabled(); |
| 309 | |
| 310 | /* Nothing to do if all requested features are already enabled. */ |
| 311 | xfeatures &= ~guest_fpu->xfeatures; |
| 312 | if (!xfeatures) |
| 313 | return 0; |
| 314 | |
| 315 | return __xfd_enable_feature(which: xfeatures, guest_fpu); |
| 316 | } |
| 317 | EXPORT_SYMBOL_FOR_KVM(fpu_enable_guest_xfd_features); |
| 318 | |
| 319 | #ifdef CONFIG_X86_64 |
| 320 | void fpu_update_guest_xfd(struct fpu_guest *guest_fpu, u64 xfd) |
| 321 | { |
| 322 | struct fpstate *fpstate = guest_fpu->fpstate; |
| 323 | |
| 324 | fpregs_lock(); |
| 325 | |
| 326 | /* |
| 327 | * KVM's guest ABI is that setting XFD[i]=1 *can* immediately revert the |
| 328 | * save state to its initial configuration. Likewise, KVM_GET_XSAVE does |
| 329 | * the same as XSAVE and returns XSTATE_BV[i]=0 whenever XFD[i]=1. |
| 330 | * |
| 331 | * If the guest's FPU state is in hardware, just update XFD: the XSAVE |
| 332 | * in fpu_swap_kvm_fpstate will clear XSTATE_BV[i] whenever XFD[i]=1. |
| 333 | * |
| 334 | * If however the guest's FPU state is NOT resident in hardware, clear |
| 335 | * disabled components in XSTATE_BV now, or a subsequent XRSTOR will |
| 336 | * attempt to load disabled components and generate #NM _in the host_. |
| 337 | */ |
| 338 | if (xfd && test_thread_flag(TIF_NEED_FPU_LOAD)) |
| 339 | fpstate->regs.xsave.header.xfeatures &= ~xfd; |
| 340 | |
| 341 | fpstate->xfd = xfd; |
| 342 | if (fpstate->in_use) |
| 343 | xfd_update_state(fpstate); |
| 344 | |
| 345 | fpregs_unlock(); |
| 346 | } |
| 347 | EXPORT_SYMBOL_FOR_KVM(fpu_update_guest_xfd); |
| 348 | |
| 349 | /** |
| 350 | * fpu_sync_guest_vmexit_xfd_state - Synchronize XFD MSR and software state |
| 351 | * |
| 352 | * Must be invoked from KVM after a VMEXIT before enabling interrupts when |
| 353 | * XFD write emulation is disabled. This is required because the guest can |
| 354 | * freely modify XFD and the state at VMEXIT is not guaranteed to be the |
| 355 | * same as the state on VMENTER. So software state has to be updated before |
| 356 | * any operation which depends on it can take place. |
| 357 | * |
| 358 | * Note: It can be invoked unconditionally even when write emulation is |
| 359 | * enabled for the price of a then pointless MSR read. |
| 360 | */ |
| 361 | void fpu_sync_guest_vmexit_xfd_state(void) |
| 362 | { |
| 363 | struct fpstate *fpstate = x86_task_fpu(current)->fpstate; |
| 364 | |
| 365 | lockdep_assert_irqs_disabled(); |
| 366 | if (fpu_state_size_dynamic()) { |
| 367 | rdmsrq(MSR_IA32_XFD, fpstate->xfd); |
| 368 | __this_cpu_write(xfd_state, fpstate->xfd); |
| 369 | } |
| 370 | } |
| 371 | EXPORT_SYMBOL_FOR_KVM(fpu_sync_guest_vmexit_xfd_state); |
| 372 | #endif /* CONFIG_X86_64 */ |
| 373 | |
| 374 | int fpu_swap_kvm_fpstate(struct fpu_guest *guest_fpu, bool enter_guest) |
| 375 | { |
| 376 | struct fpstate *guest_fps = guest_fpu->fpstate; |
| 377 | struct fpu *fpu = x86_task_fpu(current); |
| 378 | struct fpstate *cur_fps = fpu->fpstate; |
| 379 | |
| 380 | fpregs_lock(); |
| 381 | if (!cur_fps->is_confidential && !test_thread_flag(TIF_NEED_FPU_LOAD)) |
| 382 | save_fpregs_to_fpstate(fpu); |
| 383 | |
| 384 | /* Swap fpstate */ |
| 385 | if (enter_guest) { |
| 386 | fpu->__task_fpstate = cur_fps; |
| 387 | fpu->fpstate = guest_fps; |
| 388 | guest_fps->in_use = true; |
| 389 | } else { |
| 390 | guest_fps->in_use = false; |
| 391 | fpu->fpstate = fpu->__task_fpstate; |
| 392 | fpu->__task_fpstate = NULL; |
| 393 | } |
| 394 | |
| 395 | cur_fps = fpu->fpstate; |
| 396 | |
| 397 | if (!cur_fps->is_confidential) { |
| 398 | /* Includes XFD update */ |
| 399 | restore_fpregs_from_fpstate(fpstate: cur_fps, XFEATURE_MASK_FPSTATE); |
| 400 | } else { |
| 401 | /* |
| 402 | * XSTATE is restored by firmware from encrypted |
| 403 | * memory. Make sure XFD state is correct while |
| 404 | * running with guest fpstate |
| 405 | */ |
| 406 | xfd_update_state(fpstate: cur_fps); |
| 407 | } |
| 408 | |
| 409 | fpregs_mark_activate(); |
| 410 | fpregs_unlock(); |
| 411 | return 0; |
| 412 | } |
| 413 | EXPORT_SYMBOL_FOR_KVM(fpu_swap_kvm_fpstate); |
| 414 | |
| 415 | void fpu_copy_guest_fpstate_to_uabi(struct fpu_guest *gfpu, void *buf, |
| 416 | unsigned int size, u64 xfeatures, u32 pkru) |
| 417 | { |
| 418 | struct fpstate *kstate = gfpu->fpstate; |
| 419 | union fpregs_state *ustate = buf; |
| 420 | struct membuf mb = { .p = buf, .left = size }; |
| 421 | |
| 422 | if (cpu_feature_enabled(X86_FEATURE_XSAVE)) { |
| 423 | __copy_xstate_to_uabi_buf(to: mb, fpstate: kstate, xfeatures, pkru_val: pkru, |
| 424 | copy_mode: XSTATE_COPY_XSAVE); |
| 425 | } else { |
| 426 | memcpy(&ustate->fxsave, &kstate->regs.fxsave, |
| 427 | sizeof(ustate->fxsave)); |
| 428 | /* Make it restorable on a XSAVE enabled host */ |
| 429 | ustate->xsave.header.xfeatures = XFEATURE_MASK_FPSSE; |
| 430 | } |
| 431 | } |
| 432 | EXPORT_SYMBOL_FOR_KVM(fpu_copy_guest_fpstate_to_uabi); |
| 433 | |
| 434 | int fpu_copy_uabi_to_guest_fpstate(struct fpu_guest *gfpu, const void *buf, |
| 435 | u64 xcr0, u32 *vpkru) |
| 436 | { |
| 437 | struct fpstate *kstate = gfpu->fpstate; |
| 438 | const union fpregs_state *ustate = buf; |
| 439 | |
| 440 | if (!cpu_feature_enabled(X86_FEATURE_XSAVE)) { |
| 441 | if (ustate->xsave.header.xfeatures & ~XFEATURE_MASK_FPSSE) |
| 442 | return -EINVAL; |
| 443 | if (ustate->fxsave.mxcsr & ~mxcsr_feature_mask) |
| 444 | return -EINVAL; |
| 445 | memcpy(&kstate->regs.fxsave, &ustate->fxsave, sizeof(ustate->fxsave)); |
| 446 | return 0; |
| 447 | } |
| 448 | |
| 449 | if (ustate->xsave.header.xfeatures & ~xcr0) |
| 450 | return -EINVAL; |
| 451 | |
| 452 | /* |
| 453 | * Disabled features must be in their initial state, otherwise XRSTOR |
| 454 | * causes an exception. |
| 455 | */ |
| 456 | if (WARN_ON_ONCE(ustate->xsave.header.xfeatures & kstate->xfd)) |
| 457 | return -EINVAL; |
| 458 | |
| 459 | /* |
| 460 | * Nullify @vpkru to preserve its current value if PKRU's bit isn't set |
| 461 | * in the header. KVM's odd ABI is to leave PKRU untouched in this |
| 462 | * case (all other components are eventually re-initialized). |
| 463 | */ |
| 464 | if (!(ustate->xsave.header.xfeatures & XFEATURE_MASK_PKRU)) |
| 465 | vpkru = NULL; |
| 466 | |
| 467 | return copy_uabi_from_kernel_to_xstate(fpstate: kstate, kbuf: ustate, pkru: vpkru); |
| 468 | } |
| 469 | EXPORT_SYMBOL_FOR_KVM(fpu_copy_uabi_to_guest_fpstate); |
| 470 | #endif /* CONFIG_KVM */ |
| 471 | |
| 472 | void kernel_fpu_begin_mask(unsigned int kfpu_mask) |
| 473 | { |
| 474 | if (!irqs_disabled()) |
| 475 | fpregs_lock(); |
| 476 | |
| 477 | WARN_ON_FPU(!irq_fpu_usable()); |
| 478 | |
| 479 | /* Toggle kernel_fpu_allowed to false: */ |
| 480 | WARN_ON_FPU(!this_cpu_read(kernel_fpu_allowed)); |
| 481 | this_cpu_write(kernel_fpu_allowed, false); |
| 482 | |
| 483 | if (!(current->flags & (PF_KTHREAD | PF_USER_WORKER)) && |
| 484 | !test_thread_flag(TIF_NEED_FPU_LOAD)) { |
| 485 | set_thread_flag(TIF_NEED_FPU_LOAD); |
| 486 | save_fpregs_to_fpstate(fpu: x86_task_fpu(current)); |
| 487 | } |
| 488 | __cpu_invalidate_fpregs_state(); |
| 489 | |
| 490 | /* Put sane initial values into the control registers. */ |
| 491 | if (likely(kfpu_mask & KFPU_MXCSR) && boot_cpu_has(X86_FEATURE_XMM)) |
| 492 | ldmxcsr(MXCSR_DEFAULT); |
| 493 | |
| 494 | if (unlikely(kfpu_mask & KFPU_387) && boot_cpu_has(X86_FEATURE_FPU)) |
| 495 | asm volatile ("fninit" ); |
| 496 | } |
| 497 | EXPORT_SYMBOL_GPL(kernel_fpu_begin_mask); |
| 498 | |
| 499 | void kernel_fpu_end(void) |
| 500 | { |
| 501 | /* Toggle kernel_fpu_allowed back to true: */ |
| 502 | WARN_ON_FPU(this_cpu_read(kernel_fpu_allowed)); |
| 503 | this_cpu_write(kernel_fpu_allowed, true); |
| 504 | |
| 505 | if (!irqs_disabled()) |
| 506 | fpregs_unlock(); |
| 507 | } |
| 508 | EXPORT_SYMBOL_GPL(kernel_fpu_end); |
| 509 | |
| 510 | /* |
| 511 | * Sync the FPU register state to current's memory register state when the |
| 512 | * current task owns the FPU. The hardware register state is preserved. |
| 513 | */ |
| 514 | void fpu_sync_fpstate(struct fpu *fpu) |
| 515 | { |
| 516 | WARN_ON_FPU(fpu != x86_task_fpu(current)); |
| 517 | |
| 518 | fpregs_lock(); |
| 519 | trace_x86_fpu_before_save(fpu); |
| 520 | |
| 521 | if (!test_thread_flag(TIF_NEED_FPU_LOAD)) |
| 522 | save_fpregs_to_fpstate(fpu); |
| 523 | |
| 524 | trace_x86_fpu_after_save(fpu); |
| 525 | fpregs_unlock(); |
| 526 | } |
| 527 | |
| 528 | static inline unsigned int init_fpstate_copy_size(void) |
| 529 | { |
| 530 | if (!use_xsave()) |
| 531 | return fpu_kernel_cfg.default_size; |
| 532 | |
| 533 | /* XSAVE(S) just needs the legacy and the xstate header part */ |
| 534 | return sizeof(init_fpstate.regs.xsave); |
| 535 | } |
| 536 | |
| 537 | static inline void fpstate_init_fxstate(struct fpstate *fpstate) |
| 538 | { |
| 539 | fpstate->regs.fxsave.cwd = 0x37f; |
| 540 | fpstate->regs.fxsave.mxcsr = MXCSR_DEFAULT; |
| 541 | } |
| 542 | |
| 543 | /* |
| 544 | * Legacy x87 fpstate state init: |
| 545 | */ |
| 546 | static inline void fpstate_init_fstate(struct fpstate *fpstate) |
| 547 | { |
| 548 | fpstate->regs.fsave.cwd = 0xffff037fu; |
| 549 | fpstate->regs.fsave.swd = 0xffff0000u; |
| 550 | fpstate->regs.fsave.twd = 0xffffffffu; |
| 551 | fpstate->regs.fsave.fos = 0xffff0000u; |
| 552 | } |
| 553 | |
| 554 | /* |
| 555 | * Used in two places: |
| 556 | * 1) Early boot to setup init_fpstate for non XSAVE systems |
| 557 | * 2) fpu_alloc_guest_fpstate() which is invoked from KVM |
| 558 | */ |
| 559 | void fpstate_init_user(struct fpstate *fpstate) |
| 560 | { |
| 561 | if (!cpu_feature_enabled(X86_FEATURE_FPU)) { |
| 562 | fpstate_init_soft(soft: &fpstate->regs.soft); |
| 563 | return; |
| 564 | } |
| 565 | |
| 566 | xstate_init_xcomp_bv(xsave: &fpstate->regs.xsave, mask: fpstate->xfeatures); |
| 567 | |
| 568 | if (cpu_feature_enabled(X86_FEATURE_FXSR)) |
| 569 | fpstate_init_fxstate(fpstate); |
| 570 | else |
| 571 | fpstate_init_fstate(fpstate); |
| 572 | } |
| 573 | |
| 574 | static void __fpstate_reset(struct fpstate *fpstate) |
| 575 | { |
| 576 | /* |
| 577 | * Supervisor features (and thus sizes) may diverge between guest |
| 578 | * FPUs and host FPUs, as some supervisor features are supported |
| 579 | * for guests despite not being utilized by the host. User |
| 580 | * features and sizes are always identical, which allows for |
| 581 | * common guest and userspace ABI. |
| 582 | * |
| 583 | * For the host, set XFD to the kernel's desired initialization |
| 584 | * value. For guests, set XFD to its architectural RESET value. |
| 585 | */ |
| 586 | if (fpstate->is_guest) { |
| 587 | fpstate->size = guest_default_cfg.size; |
| 588 | fpstate->xfeatures = guest_default_cfg.features; |
| 589 | fpstate->xfd = 0; |
| 590 | } else { |
| 591 | fpstate->size = fpu_kernel_cfg.default_size; |
| 592 | fpstate->xfeatures = fpu_kernel_cfg.default_features; |
| 593 | fpstate->xfd = init_fpstate.xfd; |
| 594 | } |
| 595 | |
| 596 | fpstate->user_size = fpu_user_cfg.default_size; |
| 597 | fpstate->user_xfeatures = fpu_user_cfg.default_features; |
| 598 | } |
| 599 | |
| 600 | void fpstate_reset(struct fpu *fpu) |
| 601 | { |
| 602 | /* Set the fpstate pointer to the default fpstate */ |
| 603 | fpu->fpstate = &fpu->__fpstate; |
| 604 | __fpstate_reset(fpstate: fpu->fpstate); |
| 605 | |
| 606 | /* Initialize the permission related info in fpu */ |
| 607 | fpu->perm.__state_perm = fpu_kernel_cfg.default_features; |
| 608 | fpu->perm.__state_size = fpu_kernel_cfg.default_size; |
| 609 | fpu->perm.__user_state_size = fpu_user_cfg.default_size; |
| 610 | |
| 611 | fpu->guest_perm.__state_perm = guest_default_cfg.features; |
| 612 | fpu->guest_perm.__state_size = guest_default_cfg.size; |
| 613 | /* |
| 614 | * User features and sizes are always identical between host and |
| 615 | * guest FPUs, which allows for common guest and userspace ABI. |
| 616 | */ |
| 617 | fpu->guest_perm.__user_state_size = fpu_user_cfg.default_size; |
| 618 | } |
| 619 | |
| 620 | static inline void fpu_inherit_perms(struct fpu *dst_fpu) |
| 621 | { |
| 622 | if (fpu_state_size_dynamic()) { |
| 623 | struct fpu *src_fpu = x86_task_fpu(current->group_leader); |
| 624 | |
| 625 | spin_lock_irq(lock: ¤t->sighand->siglock); |
| 626 | /* Fork also inherits the permissions of the parent */ |
| 627 | dst_fpu->perm = src_fpu->perm; |
| 628 | dst_fpu->guest_perm = src_fpu->guest_perm; |
| 629 | spin_unlock_irq(lock: ¤t->sighand->siglock); |
| 630 | } |
| 631 | } |
| 632 | |
| 633 | /* A passed ssp of zero will not cause any update */ |
| 634 | static int update_fpu_shstk(struct task_struct *dst, unsigned long ssp) |
| 635 | { |
| 636 | #ifdef CONFIG_X86_USER_SHADOW_STACK |
| 637 | struct cet_user_state *xstate; |
| 638 | |
| 639 | /* If ssp update is not needed. */ |
| 640 | if (!ssp) |
| 641 | return 0; |
| 642 | |
| 643 | xstate = get_xsave_addr(xsave: &x86_task_fpu(task: dst)->fpstate->regs.xsave, |
| 644 | xfeature_nr: XFEATURE_CET_USER); |
| 645 | |
| 646 | /* |
| 647 | * If there is a non-zero ssp, then 'dst' must be configured with a shadow |
| 648 | * stack and the fpu state should be up to date since it was just copied |
| 649 | * from the parent in fpu_clone(). So there must be a valid non-init CET |
| 650 | * state location in the buffer. |
| 651 | */ |
| 652 | if (WARN_ON_ONCE(!xstate)) |
| 653 | return 1; |
| 654 | |
| 655 | xstate->user_ssp = (u64)ssp; |
| 656 | #endif |
| 657 | return 0; |
| 658 | } |
| 659 | |
| 660 | /* Clone current's FPU state on fork */ |
| 661 | int fpu_clone(struct task_struct *dst, u64 clone_flags, bool minimal, |
| 662 | unsigned long ssp) |
| 663 | { |
| 664 | /* |
| 665 | * We allocate the new FPU structure right after the end of the task struct. |
| 666 | * task allocation size already took this into account. |
| 667 | * |
| 668 | * This is safe because task_struct size is a multiple of cacheline size, |
| 669 | * thus x86_task_fpu() will always be cacheline aligned as well. |
| 670 | */ |
| 671 | struct fpu *dst_fpu = (void *)dst + sizeof(*dst); |
| 672 | |
| 673 | BUILD_BUG_ON(sizeof(*dst) % SMP_CACHE_BYTES != 0); |
| 674 | |
| 675 | /* The new task's FPU state cannot be valid in the hardware. */ |
| 676 | dst_fpu->last_cpu = -1; |
| 677 | |
| 678 | fpstate_reset(fpu: dst_fpu); |
| 679 | |
| 680 | if (!cpu_feature_enabled(X86_FEATURE_FPU)) |
| 681 | return 0; |
| 682 | |
| 683 | /* |
| 684 | * Enforce reload for user space tasks and prevent kernel threads |
| 685 | * from trying to save the FPU registers on context switch. |
| 686 | */ |
| 687 | set_tsk_thread_flag(tsk: dst, TIF_NEED_FPU_LOAD); |
| 688 | |
| 689 | /* |
| 690 | * No FPU state inheritance for kernel threads and IO |
| 691 | * worker threads. |
| 692 | */ |
| 693 | if (minimal) { |
| 694 | /* Clear out the minimal state */ |
| 695 | memcpy(&dst_fpu->fpstate->regs, &init_fpstate.regs, |
| 696 | init_fpstate_copy_size()); |
| 697 | return 0; |
| 698 | } |
| 699 | |
| 700 | /* |
| 701 | * If a new feature is added, ensure all dynamic features are |
| 702 | * caller-saved from here! |
| 703 | */ |
| 704 | BUILD_BUG_ON(XFEATURE_MASK_USER_DYNAMIC != XFEATURE_MASK_XTILE_DATA); |
| 705 | |
| 706 | /* |
| 707 | * Save the default portion of the current FPU state into the |
| 708 | * clone. Assume all dynamic features to be defined as caller- |
| 709 | * saved, which enables skipping both the expansion of fpstate |
| 710 | * and the copying of any dynamic state. |
| 711 | * |
| 712 | * Do not use memcpy() when TIF_NEED_FPU_LOAD is set because |
| 713 | * copying is not valid when current uses non-default states. |
| 714 | */ |
| 715 | fpregs_lock(); |
| 716 | if (test_thread_flag(TIF_NEED_FPU_LOAD)) |
| 717 | fpregs_restore_userregs(); |
| 718 | save_fpregs_to_fpstate(fpu: dst_fpu); |
| 719 | fpregs_unlock(); |
| 720 | if (!(clone_flags & CLONE_THREAD)) |
| 721 | fpu_inherit_perms(dst_fpu); |
| 722 | |
| 723 | /* |
| 724 | * Children never inherit PASID state. |
| 725 | * Force it to have its init value: |
| 726 | */ |
| 727 | if (use_xsave()) |
| 728 | dst_fpu->fpstate->regs.xsave.header.xfeatures &= ~XFEATURE_MASK_PASID; |
| 729 | |
| 730 | /* |
| 731 | * Update shadow stack pointer, in case it changed during clone. |
| 732 | */ |
| 733 | if (update_fpu_shstk(dst, ssp)) |
| 734 | return 1; |
| 735 | |
| 736 | trace_x86_fpu_copy_dst(fpu: dst_fpu); |
| 737 | |
| 738 | return 0; |
| 739 | } |
| 740 | |
| 741 | /* |
| 742 | * While struct fpu is no longer part of struct thread_struct, it is still |
| 743 | * allocated after struct task_struct in the "task_struct" kmem cache. But |
| 744 | * since FPU is expected to be part of struct thread_struct, we have to |
| 745 | * adjust for it here. |
| 746 | */ |
| 747 | void fpu_thread_struct_whitelist(unsigned long *offset, unsigned long *size) |
| 748 | { |
| 749 | /* The allocation follows struct task_struct. */ |
| 750 | *offset = sizeof(struct task_struct) - offsetof(struct task_struct, thread); |
| 751 | *offset += offsetof(struct fpu, __fpstate.regs); |
| 752 | *size = fpu_kernel_cfg.default_size; |
| 753 | } |
| 754 | |
| 755 | /* |
| 756 | * Drops current FPU state: deactivates the fpregs and |
| 757 | * the fpstate. NOTE: it still leaves previous contents |
| 758 | * in the fpregs in the eager-FPU case. |
| 759 | * |
| 760 | * This function can be used in cases where we know that |
| 761 | * a state-restore is coming: either an explicit one, |
| 762 | * or a reschedule. |
| 763 | */ |
| 764 | void fpu__drop(struct task_struct *tsk) |
| 765 | { |
| 766 | struct fpu *fpu; |
| 767 | |
| 768 | if (test_tsk_thread_flag(tsk, TIF_NEED_FPU_LOAD)) |
| 769 | return; |
| 770 | |
| 771 | fpu = x86_task_fpu(task: tsk); |
| 772 | |
| 773 | preempt_disable(); |
| 774 | |
| 775 | if (fpu == x86_task_fpu(current)) { |
| 776 | /* Ignore delayed exceptions from user space */ |
| 777 | asm volatile("1: fwait\n" |
| 778 | "2:\n" |
| 779 | _ASM_EXTABLE(1b, 2b)); |
| 780 | fpregs_deactivate(fpu); |
| 781 | } |
| 782 | |
| 783 | trace_x86_fpu_dropped(fpu); |
| 784 | |
| 785 | preempt_enable(); |
| 786 | } |
| 787 | |
| 788 | /* |
| 789 | * Clear FPU registers by setting them up from the init fpstate. |
| 790 | * Caller must do fpregs_[un]lock() around it. |
| 791 | */ |
| 792 | static inline void restore_fpregs_from_init_fpstate(u64 features_mask) |
| 793 | { |
| 794 | if (use_xsave()) |
| 795 | os_xrstor(fpstate: &init_fpstate, mask: features_mask); |
| 796 | else if (use_fxsr()) |
| 797 | fxrstor(fx: &init_fpstate.regs.fxsave); |
| 798 | else |
| 799 | frstor(fx: &init_fpstate.regs.fsave); |
| 800 | |
| 801 | pkru_write_default(); |
| 802 | } |
| 803 | |
| 804 | /* |
| 805 | * Reset current->fpu memory state to the init values. |
| 806 | */ |
| 807 | static void fpu_reset_fpstate_regs(void) |
| 808 | { |
| 809 | struct fpu *fpu = x86_task_fpu(current); |
| 810 | |
| 811 | fpregs_lock(); |
| 812 | __fpu_invalidate_fpregs_state(fpu); |
| 813 | /* |
| 814 | * This does not change the actual hardware registers. It just |
| 815 | * resets the memory image and sets TIF_NEED_FPU_LOAD so a |
| 816 | * subsequent return to usermode will reload the registers from the |
| 817 | * task's memory image. |
| 818 | * |
| 819 | * Do not use fpstate_init() here. Just copy init_fpstate which has |
| 820 | * the correct content already except for PKRU. |
| 821 | * |
| 822 | * PKRU handling does not rely on the xstate when restoring for |
| 823 | * user space as PKRU is eagerly written in switch_to() and |
| 824 | * flush_thread(). |
| 825 | */ |
| 826 | memcpy(&fpu->fpstate->regs, &init_fpstate.regs, init_fpstate_copy_size()); |
| 827 | set_thread_flag(TIF_NEED_FPU_LOAD); |
| 828 | fpregs_unlock(); |
| 829 | } |
| 830 | |
| 831 | /* |
| 832 | * Reset current's user FPU states to the init states. current's |
| 833 | * supervisor states, if any, are not modified by this function. The |
| 834 | * caller guarantees that the XSTATE header in memory is intact. |
| 835 | */ |
| 836 | void fpu__clear_user_states(struct fpu *fpu) |
| 837 | { |
| 838 | WARN_ON_FPU(fpu != x86_task_fpu(current)); |
| 839 | |
| 840 | fpregs_lock(); |
| 841 | if (!cpu_feature_enabled(X86_FEATURE_FPU)) { |
| 842 | fpu_reset_fpstate_regs(); |
| 843 | fpregs_unlock(); |
| 844 | return; |
| 845 | } |
| 846 | |
| 847 | /* |
| 848 | * Ensure that current's supervisor states are loaded into their |
| 849 | * corresponding registers. |
| 850 | */ |
| 851 | if (xfeatures_mask_supervisor() && |
| 852 | !fpregs_state_valid(fpu, smp_processor_id())) |
| 853 | os_xrstor_supervisor(fpstate: fpu->fpstate); |
| 854 | |
| 855 | /* Ensure XFD state is in sync before reloading XSTATE */ |
| 856 | xfd_update_state(fpstate: fpu->fpstate); |
| 857 | |
| 858 | /* Reset user states in registers. */ |
| 859 | restore_fpregs_from_init_fpstate(XFEATURE_MASK_USER_RESTORE); |
| 860 | |
| 861 | /* |
| 862 | * Now all FPU registers have their desired values. Inform the FPU |
| 863 | * state machine that current's FPU registers are in the hardware |
| 864 | * registers. The memory image does not need to be updated because |
| 865 | * any operation relying on it has to save the registers first when |
| 866 | * current's FPU is marked active. |
| 867 | */ |
| 868 | fpregs_mark_activate(); |
| 869 | fpregs_unlock(); |
| 870 | } |
| 871 | |
| 872 | void fpu_flush_thread(void) |
| 873 | { |
| 874 | fpstate_reset(fpu: x86_task_fpu(current)); |
| 875 | fpu_reset_fpstate_regs(); |
| 876 | } |
| 877 | /* |
| 878 | * Load FPU context before returning to userspace. |
| 879 | */ |
| 880 | void switch_fpu_return(void) |
| 881 | { |
| 882 | if (!static_cpu_has(X86_FEATURE_FPU)) |
| 883 | return; |
| 884 | |
| 885 | fpregs_restore_userregs(); |
| 886 | } |
| 887 | EXPORT_SYMBOL_FOR_KVM(switch_fpu_return); |
| 888 | |
| 889 | void fpregs_lock_and_load(void) |
| 890 | { |
| 891 | /* |
| 892 | * fpregs_lock() only disables preemption (mostly). So modifying state |
| 893 | * in an interrupt could screw up some in progress fpregs operation. |
| 894 | * Warn about it. |
| 895 | */ |
| 896 | WARN_ON_ONCE(!irq_fpu_usable()); |
| 897 | WARN_ON_ONCE(current->flags & PF_KTHREAD); |
| 898 | |
| 899 | fpregs_lock(); |
| 900 | |
| 901 | fpregs_assert_state_consistent(); |
| 902 | |
| 903 | if (test_thread_flag(TIF_NEED_FPU_LOAD)) |
| 904 | fpregs_restore_userregs(); |
| 905 | } |
| 906 | |
| 907 | #ifdef CONFIG_X86_DEBUG_FPU |
| 908 | /* |
| 909 | * If current FPU state according to its tracking (loaded FPU context on this |
| 910 | * CPU) is not valid then we must have TIF_NEED_FPU_LOAD set so the context is |
| 911 | * loaded on return to userland. |
| 912 | */ |
| 913 | void fpregs_assert_state_consistent(void) |
| 914 | { |
| 915 | struct fpu *fpu = x86_task_fpu(current); |
| 916 | |
| 917 | if (test_thread_flag(TIF_NEED_FPU_LOAD)) |
| 918 | return; |
| 919 | |
| 920 | WARN_ON_FPU(!fpregs_state_valid(fpu, smp_processor_id())); |
| 921 | } |
| 922 | EXPORT_SYMBOL_FOR_KVM(fpregs_assert_state_consistent); |
| 923 | #endif |
| 924 | |
| 925 | void fpregs_mark_activate(void) |
| 926 | { |
| 927 | struct fpu *fpu = x86_task_fpu(current); |
| 928 | |
| 929 | fpregs_activate(fpu); |
| 930 | fpu->last_cpu = smp_processor_id(); |
| 931 | clear_thread_flag(TIF_NEED_FPU_LOAD); |
| 932 | } |
| 933 | |
| 934 | /* |
| 935 | * x87 math exception handling: |
| 936 | */ |
| 937 | |
| 938 | int fpu__exception_code(struct fpu *fpu, int trap_nr) |
| 939 | { |
| 940 | int err; |
| 941 | |
| 942 | if (trap_nr == X86_TRAP_MF) { |
| 943 | unsigned short cwd, swd; |
| 944 | /* |
| 945 | * (~cwd & swd) will mask out exceptions that are not set to unmasked |
| 946 | * status. 0x3f is the exception bits in these regs, 0x200 is the |
| 947 | * C1 reg you need in case of a stack fault, 0x040 is the stack |
| 948 | * fault bit. We should only be taking one exception at a time, |
| 949 | * so if this combination doesn't produce any single exception, |
| 950 | * then we have a bad program that isn't synchronizing its FPU usage |
| 951 | * and it will suffer the consequences since we won't be able to |
| 952 | * fully reproduce the context of the exception. |
| 953 | */ |
| 954 | if (boot_cpu_has(X86_FEATURE_FXSR)) { |
| 955 | cwd = fpu->fpstate->regs.fxsave.cwd; |
| 956 | swd = fpu->fpstate->regs.fxsave.swd; |
| 957 | } else { |
| 958 | cwd = (unsigned short)fpu->fpstate->regs.fsave.cwd; |
| 959 | swd = (unsigned short)fpu->fpstate->regs.fsave.swd; |
| 960 | } |
| 961 | |
| 962 | err = swd & ~cwd; |
| 963 | } else { |
| 964 | /* |
| 965 | * The SIMD FPU exceptions are handled a little differently, as there |
| 966 | * is only a single status/control register. Thus, to determine which |
| 967 | * unmasked exception was caught we must mask the exception mask bits |
| 968 | * at 0x1f80, and then use these to mask the exception bits at 0x3f. |
| 969 | */ |
| 970 | unsigned short mxcsr = MXCSR_DEFAULT; |
| 971 | |
| 972 | if (boot_cpu_has(X86_FEATURE_XMM)) |
| 973 | mxcsr = fpu->fpstate->regs.fxsave.mxcsr; |
| 974 | |
| 975 | err = ~(mxcsr >> 7) & mxcsr; |
| 976 | } |
| 977 | |
| 978 | if (err & 0x001) { /* Invalid op */ |
| 979 | /* |
| 980 | * swd & 0x240 == 0x040: Stack Underflow |
| 981 | * swd & 0x240 == 0x240: Stack Overflow |
| 982 | * User must clear the SF bit (0x40) if set |
| 983 | */ |
| 984 | return FPE_FLTINV; |
| 985 | } else if (err & 0x004) { /* Divide by Zero */ |
| 986 | return FPE_FLTDIV; |
| 987 | } else if (err & 0x008) { /* Overflow */ |
| 988 | return FPE_FLTOVF; |
| 989 | } else if (err & 0x012) { /* Denormal, Underflow */ |
| 990 | return FPE_FLTUND; |
| 991 | } else if (err & 0x020) { /* Precision */ |
| 992 | return FPE_FLTRES; |
| 993 | } |
| 994 | |
| 995 | /* |
| 996 | * If we're using IRQ 13, or supposedly even some trap |
| 997 | * X86_TRAP_MF implementations, it's possible |
| 998 | * we get a spurious trap, which is not an error. |
| 999 | */ |
| 1000 | return 0; |
| 1001 | } |
| 1002 | |
| 1003 | /* |
| 1004 | * Initialize register state that may prevent from entering low-power idle. |
| 1005 | * This function will be invoked from the cpuidle driver only when needed. |
| 1006 | */ |
| 1007 | noinstr void fpu_idle_fpregs(void) |
| 1008 | { |
| 1009 | /* Note: AMX_TILE being enabled implies XGETBV1 support */ |
| 1010 | if (cpu_feature_enabled(X86_FEATURE_AMX_TILE) && |
| 1011 | (xfeatures_in_use() & XFEATURE_MASK_XTILE)) { |
| 1012 | tile_release(); |
| 1013 | __this_cpu_write(fpu_fpregs_owner_ctx, NULL); |
| 1014 | } |
| 1015 | } |
| 1016 | |