1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2019 Western Digital Corporation or its affiliates. 4 * 5 * Authors: 6 * Anup Patel <anup.patel@wdc.com> 7 */ 8 9 #include <linux/bitops.h> 10 #include <linux/errno.h> 11 #include <linux/err.h> 12 #include <linux/kdebug.h> 13 #include <linux/module.h> 14 #include <linux/percpu.h> 15 #include <linux/vmalloc.h> 16 #include <linux/sched/signal.h> 17 #include <linux/fs.h> 18 #include <linux/kvm_host.h> 19 #include <asm/cacheflush.h> 20 #include <asm/kvm_mmu.h> 21 #include <asm/kvm_nacl.h> 22 #include <asm/kvm_vcpu_vector.h> 23 24 #define CREATE_TRACE_POINTS 25 #include "trace.h" 26 27 static DEFINE_PER_CPU(struct kvm_vcpu *, kvm_former_vcpu); 28 29 void kvm_riscv_clear_former_vcpu(void) 30 { 31 /* 32 * Clear the per-CPU former VCPU pointer because hypervisor CSR state 33 * will be lost. This ensures that the next VCPU entry will properly 34 * restore all CSRs instead of incorrectly skipping CSR restoration 35 * via the fast-path optimization. 36 */ 37 __this_cpu_write(kvm_former_vcpu, NULL); 38 } 39 40 const struct kvm_stats_desc kvm_vcpu_stats_desc[] = { 41 KVM_GENERIC_VCPU_STATS(), 42 STATS_DESC_COUNTER(VCPU, ecall_exit_stat), 43 STATS_DESC_COUNTER(VCPU, wfi_exit_stat), 44 STATS_DESC_COUNTER(VCPU, wrs_exit_stat), 45 STATS_DESC_COUNTER(VCPU, mmio_exit_user), 46 STATS_DESC_COUNTER(VCPU, mmio_exit_kernel), 47 STATS_DESC_COUNTER(VCPU, csr_exit_user), 48 STATS_DESC_COUNTER(VCPU, csr_exit_kernel), 49 STATS_DESC_COUNTER(VCPU, signal_exits), 50 STATS_DESC_COUNTER(VCPU, exits), 51 STATS_DESC_COUNTER(VCPU, instr_illegal_exits), 52 STATS_DESC_COUNTER(VCPU, load_misaligned_exits), 53 STATS_DESC_COUNTER(VCPU, store_misaligned_exits), 54 STATS_DESC_COUNTER(VCPU, load_access_exits), 55 STATS_DESC_COUNTER(VCPU, store_access_exits), 56 }; 57 58 const struct kvm_stats_header kvm_vcpu_stats_header = { 59 .name_size = KVM_STATS_NAME_SIZE, 60 .num_desc = ARRAY_SIZE(kvm_vcpu_stats_desc), 61 .id_offset = sizeof(struct kvm_stats_header), 62 .desc_offset = sizeof(struct kvm_stats_header) + KVM_STATS_NAME_SIZE, 63 .data_offset = sizeof(struct kvm_stats_header) + KVM_STATS_NAME_SIZE + 64 sizeof(kvm_vcpu_stats_desc), 65 }; 66 67 static void kvm_riscv_vcpu_context_reset(struct kvm_vcpu *vcpu, 68 bool kvm_sbi_reset) 69 { 70 struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr; 71 struct kvm_cpu_context *cntx = &vcpu->arch.guest_context; 72 void *vector_datap = cntx->vector.datap; 73 74 memset(cntx, 0, sizeof(*cntx)); 75 memset(csr, 0, sizeof(*csr)); 76 memset(&vcpu->arch.smstateen_csr, 0, sizeof(vcpu->arch.smstateen_csr)); 77 memset(&vcpu->arch.zicfiss_csr, 0, sizeof(vcpu->arch.zicfiss_csr)); 78 79 /* Restore datap as it's not a part of the guest context. */ 80 cntx->vector.datap = vector_datap; 81 82 if (kvm_sbi_reset) 83 kvm_riscv_vcpu_sbi_load_reset_state(vcpu); 84 85 /* Setup reset state of shadow SSTATUS and HSTATUS CSRs */ 86 cntx->sstatus = SR_SPP | SR_SPIE; 87 88 cntx->hstatus |= HSTATUS_VTW; 89 cntx->hstatus |= HSTATUS_SPVP; 90 cntx->hstatus |= HSTATUS_SPV; 91 } 92 93 static void kvm_riscv_reset_vcpu(struct kvm_vcpu *vcpu, bool kvm_sbi_reset) 94 { 95 unsigned long flags; 96 bool loaded; 97 98 /** 99 * The preemption should be disabled here because it races with 100 * kvm_sched_out/kvm_sched_in(called from preempt notifiers) which 101 * also calls vcpu_load/put. 102 */ 103 get_cpu(); 104 loaded = (vcpu->cpu != -1); 105 if (loaded) 106 kvm_arch_vcpu_put(vcpu); 107 108 vcpu->arch.last_exit_cpu = -1; 109 110 kvm_riscv_vcpu_context_reset(vcpu, kvm_sbi_reset); 111 112 kvm_riscv_vcpu_fp_reset(vcpu); 113 114 kvm_riscv_vcpu_vector_reset(vcpu); 115 116 kvm_riscv_vcpu_timer_reset(vcpu); 117 118 kvm_riscv_vcpu_aia_reset(vcpu); 119 120 raw_spin_lock_irqsave(&vcpu->arch.irqs_pending_lock, flags); 121 bitmap_zero(vcpu->arch.irqs_pending, KVM_RISCV_VCPU_NR_IRQS); 122 bitmap_zero(vcpu->arch.irqs_pending_mask, KVM_RISCV_VCPU_NR_IRQS); 123 raw_spin_unlock_irqrestore(&vcpu->arch.irqs_pending_lock, flags); 124 125 kvm_riscv_vcpu_pmu_reset(vcpu); 126 127 vcpu->arch.hfence_head = 0; 128 vcpu->arch.hfence_tail = 0; 129 memset(vcpu->arch.hfence_queue, 0, sizeof(vcpu->arch.hfence_queue)); 130 131 kvm_riscv_vcpu_sbi_reset(vcpu); 132 133 /* Reset the guest CSRs for hotplug usecase */ 134 if (loaded) 135 kvm_arch_vcpu_load(vcpu, smp_processor_id()); 136 put_cpu(); 137 } 138 139 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id) 140 { 141 return 0; 142 } 143 144 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu) 145 { 146 int rc; 147 148 spin_lock_init(&vcpu->arch.mp_state_lock); 149 150 /* Mark this VCPU never ran */ 151 vcpu->arch.ran_atleast_once = false; 152 153 vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO; 154 bitmap_zero(vcpu->arch.isa, RISCV_ISA_EXT_MAX); 155 156 /* Setup VCPU config */ 157 kvm_riscv_vcpu_config_init(vcpu); 158 159 /* Setup ISA features available to VCPU */ 160 kvm_riscv_vcpu_setup_isa(vcpu); 161 162 /* Setup vendor, arch, and implementation details */ 163 vcpu->arch.mvendorid = sbi_get_mvendorid(); 164 vcpu->arch.marchid = sbi_get_marchid(); 165 vcpu->arch.mimpid = sbi_get_mimpid(); 166 167 /* Setup VCPU hfence queue */ 168 spin_lock_init(&vcpu->arch.hfence_lock); 169 raw_spin_lock_init(&vcpu->arch.irqs_pending_lock); 170 171 spin_lock_init(&vcpu->arch.reset_state.lock); 172 173 rc = kvm_riscv_vcpu_alloc_vector_context(vcpu); 174 if (rc) 175 return rc; 176 177 /* Setup VCPU timer */ 178 kvm_riscv_vcpu_timer_init(vcpu); 179 180 /* setup performance monitoring */ 181 kvm_riscv_vcpu_pmu_init(vcpu); 182 183 /* Setup VCPU AIA */ 184 kvm_riscv_vcpu_aia_init(vcpu); 185 186 /* 187 * Setup SBI extensions 188 * NOTE: This must be the last thing to be initialized. 189 */ 190 kvm_riscv_vcpu_sbi_init(vcpu); 191 192 /* Reset VCPU */ 193 kvm_riscv_reset_vcpu(vcpu, false); 194 195 return 0; 196 } 197 198 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu) 199 { 200 /** 201 * vcpu with id 0 is the designated boot cpu. 202 * Keep all vcpus with non-zero id in power-off state so that 203 * they can be brought up using SBI HSM extension. 204 */ 205 if (vcpu->vcpu_idx != 0) 206 kvm_riscv_vcpu_power_off(vcpu); 207 } 208 209 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu) 210 { 211 kvm_riscv_vcpu_sbi_deinit(vcpu); 212 213 /* Cleanup VCPU AIA context */ 214 kvm_riscv_vcpu_aia_deinit(vcpu); 215 216 /* Cleanup VCPU timer */ 217 kvm_riscv_vcpu_timer_deinit(vcpu); 218 219 kvm_riscv_vcpu_pmu_deinit(vcpu); 220 221 /* Free unused pages pre-allocated for G-stage page table mappings */ 222 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache); 223 224 /* Free vector context space for host and guest kernel */ 225 kvm_riscv_vcpu_free_vector_context(vcpu); 226 } 227 228 int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu) 229 { 230 return kvm_riscv_vcpu_timer_pending(vcpu); 231 } 232 233 int kvm_arch_vcpu_runnable(struct kvm_vcpu *vcpu) 234 { 235 return (kvm_riscv_vcpu_has_interrupts(vcpu, -1ULL) && 236 !kvm_riscv_vcpu_stopped(vcpu) && !vcpu->arch.pause); 237 } 238 239 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu) 240 { 241 return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE; 242 } 243 244 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu) 245 { 246 return (vcpu->arch.guest_context.sstatus & SR_SPP) ? true : false; 247 } 248 249 #ifdef CONFIG_GUEST_PERF_EVENTS 250 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu) 251 { 252 return vcpu->arch.guest_context.sepc; 253 } 254 #endif 255 256 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf) 257 { 258 return VM_FAULT_SIGBUS; 259 } 260 261 long kvm_arch_vcpu_unlocked_ioctl(struct file *filp, unsigned int ioctl, 262 unsigned long arg) 263 { 264 struct kvm_vcpu *vcpu = filp->private_data; 265 void __user *argp = (void __user *)arg; 266 267 if (ioctl == KVM_INTERRUPT) { 268 struct kvm_interrupt irq; 269 270 if (copy_from_user(&irq, argp, sizeof(irq))) 271 return -EFAULT; 272 273 if (irq.irq == KVM_INTERRUPT_SET) 274 return kvm_riscv_vcpu_set_interrupt(vcpu, IRQ_VS_EXT); 275 else 276 return kvm_riscv_vcpu_unset_interrupt(vcpu, IRQ_VS_EXT); 277 } 278 279 return -ENOIOCTLCMD; 280 } 281 282 long kvm_arch_vcpu_ioctl(struct file *filp, 283 unsigned int ioctl, unsigned long arg) 284 { 285 struct kvm_vcpu *vcpu = filp->private_data; 286 void __user *argp = (void __user *)arg; 287 long r = -EINVAL; 288 289 switch (ioctl) { 290 case KVM_SET_ONE_REG: 291 case KVM_GET_ONE_REG: { 292 struct kvm_one_reg reg; 293 294 r = -EFAULT; 295 if (copy_from_user(®, argp, sizeof(reg))) 296 break; 297 298 if (ioctl == KVM_SET_ONE_REG) 299 r = kvm_riscv_vcpu_set_reg(vcpu, ®); 300 else 301 r = kvm_riscv_vcpu_get_reg(vcpu, ®); 302 break; 303 } 304 case KVM_GET_REG_LIST: { 305 struct kvm_reg_list __user *user_list = argp; 306 struct kvm_reg_list reg_list; 307 unsigned int n; 308 309 r = -EFAULT; 310 if (copy_from_user(®_list, user_list, sizeof(reg_list))) 311 break; 312 n = reg_list.n; 313 reg_list.n = kvm_riscv_vcpu_num_regs(vcpu); 314 if (copy_to_user(user_list, ®_list, sizeof(reg_list))) 315 break; 316 r = -E2BIG; 317 if (n < reg_list.n) 318 break; 319 r = kvm_riscv_vcpu_copy_reg_indices(vcpu, user_list->reg); 320 break; 321 } 322 default: 323 break; 324 } 325 326 return r; 327 } 328 329 int kvm_arch_vcpu_ioctl_get_sregs(struct kvm_vcpu *vcpu, 330 struct kvm_sregs *sregs) 331 { 332 return -EINVAL; 333 } 334 335 int kvm_arch_vcpu_ioctl_set_sregs(struct kvm_vcpu *vcpu, 336 struct kvm_sregs *sregs) 337 { 338 return -EINVAL; 339 } 340 341 int kvm_arch_vcpu_ioctl_get_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu) 342 { 343 return -EINVAL; 344 } 345 346 int kvm_arch_vcpu_ioctl_set_fpu(struct kvm_vcpu *vcpu, struct kvm_fpu *fpu) 347 { 348 return -EINVAL; 349 } 350 351 int kvm_arch_vcpu_ioctl_translate(struct kvm_vcpu *vcpu, 352 struct kvm_translation *tr) 353 { 354 return -EINVAL; 355 } 356 357 int kvm_arch_vcpu_ioctl_get_regs(struct kvm_vcpu *vcpu, struct kvm_regs *regs) 358 { 359 return -EINVAL; 360 } 361 362 int kvm_arch_vcpu_ioctl_set_regs(struct kvm_vcpu *vcpu, struct kvm_regs *regs) 363 { 364 return -EINVAL; 365 } 366 367 void kvm_riscv_vcpu_flush_interrupts(struct kvm_vcpu *vcpu) 368 { 369 struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr; 370 unsigned long mask, val; 371 unsigned long flags; 372 373 raw_spin_lock_irqsave(&vcpu->arch.irqs_pending_lock, flags); 374 375 mask = vcpu->arch.irqs_pending_mask[0]; 376 if (mask) { 377 vcpu->arch.irqs_pending_mask[0] = 0; 378 val = vcpu->arch.irqs_pending[0] & mask; 379 380 csr->hvip &= ~mask; 381 csr->hvip |= val; 382 } 383 384 /* Flush AIA high interrupts */ 385 kvm_riscv_vcpu_aia_flush_interrupts(vcpu); 386 387 raw_spin_unlock_irqrestore(&vcpu->arch.irqs_pending_lock, flags); 388 } 389 390 void kvm_riscv_vcpu_sync_interrupts(struct kvm_vcpu *vcpu) 391 { 392 unsigned long hvip; 393 unsigned long flags; 394 struct kvm_vcpu_arch *v = &vcpu->arch; 395 struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr; 396 397 /* Read current HVIP and VSIE CSRs */ 398 csr->vsie = ncsr_read(CSR_VSIE); 399 400 /* Sync-up HVIP.VSSIP bit changes does by Guest */ 401 hvip = ncsr_read(CSR_HVIP); 402 403 raw_spin_lock_irqsave(&v->irqs_pending_lock, flags); 404 405 if ((csr->hvip ^ hvip) & (1UL << IRQ_VS_SOFT)) { 406 if (hvip & (1UL << IRQ_VS_SOFT)) { 407 if (!__test_and_set_bit(IRQ_VS_SOFT, 408 v->irqs_pending_mask)) 409 __set_bit(IRQ_VS_SOFT, v->irqs_pending); 410 } else { 411 if (!__test_and_set_bit(IRQ_VS_SOFT, 412 v->irqs_pending_mask)) 413 __clear_bit(IRQ_VS_SOFT, v->irqs_pending); 414 } 415 } 416 417 /* Sync up the HVIP.LCOFIP bit changes (only clear) by the guest */ 418 if ((csr->hvip ^ hvip) & (1UL << IRQ_PMU_OVF)) { 419 if (!(hvip & (1UL << IRQ_PMU_OVF)) && 420 !__test_and_set_bit(IRQ_PMU_OVF, v->irqs_pending_mask)) 421 __clear_bit(IRQ_PMU_OVF, v->irqs_pending); 422 } 423 424 /* Sync-up AIA high interrupts */ 425 kvm_riscv_vcpu_aia_sync_interrupts(vcpu); 426 427 raw_spin_unlock_irqrestore(&v->irqs_pending_lock, flags); 428 429 /* Sync-up timer CSRs */ 430 kvm_riscv_vcpu_timer_sync(vcpu); 431 } 432 433 int kvm_riscv_vcpu_set_interrupt(struct kvm_vcpu *vcpu, unsigned int irq) 434 { 435 unsigned long flags; 436 437 /* 438 * We only allow VS-mode software, timer, and external 439 * interrupts when irq is one of the local interrupts 440 * defined by RISC-V privilege specification. 441 */ 442 if (irq < IRQ_LOCAL_MAX && 443 irq != IRQ_VS_SOFT && 444 irq != IRQ_VS_TIMER && 445 irq != IRQ_VS_EXT && 446 irq != IRQ_PMU_OVF) 447 return -EINVAL; 448 449 raw_spin_lock_irqsave(&vcpu->arch.irqs_pending_lock, flags); 450 __set_bit(irq, vcpu->arch.irqs_pending); 451 __set_bit(irq, vcpu->arch.irqs_pending_mask); 452 raw_spin_unlock_irqrestore(&vcpu->arch.irqs_pending_lock, flags); 453 454 trace_kvm_vcpu_irq(vcpu->vcpu_id, irq, 1); 455 456 kvm_vcpu_kick(vcpu); 457 458 return 0; 459 } 460 461 int kvm_riscv_vcpu_unset_interrupt(struct kvm_vcpu *vcpu, unsigned int irq) 462 { 463 unsigned long flags; 464 465 /* 466 * We only allow VS-mode software, timer, counter overflow and external 467 * interrupts when irq is one of the local interrupts 468 * defined by RISC-V privilege specification. 469 */ 470 if (irq < IRQ_LOCAL_MAX && 471 irq != IRQ_VS_SOFT && 472 irq != IRQ_VS_TIMER && 473 irq != IRQ_VS_EXT && 474 irq != IRQ_PMU_OVF) 475 return -EINVAL; 476 477 raw_spin_lock_irqsave(&vcpu->arch.irqs_pending_lock, flags); 478 __clear_bit(irq, vcpu->arch.irqs_pending); 479 __set_bit(irq, vcpu->arch.irqs_pending_mask); 480 raw_spin_unlock_irqrestore(&vcpu->arch.irqs_pending_lock, flags); 481 482 trace_kvm_vcpu_irq(vcpu->vcpu_id, irq, 0); 483 484 return 0; 485 } 486 487 bool kvm_riscv_vcpu_has_interrupts(struct kvm_vcpu *vcpu, u64 mask) 488 { 489 unsigned long flags; 490 unsigned long ie; 491 bool ret; 492 493 raw_spin_lock_irqsave(&vcpu->arch.irqs_pending_lock, flags); 494 ie = ((vcpu->arch.guest_csr.vsie & VSIP_VALID_MASK) 495 << VSIP_TO_HVIP_SHIFT) & (unsigned long)mask; 496 ie |= vcpu->arch.guest_csr.vsie & ~IRQ_LOCAL_MASK & 497 (unsigned long)mask; 498 ret = vcpu->arch.irqs_pending[0] & ie; 499 raw_spin_unlock_irqrestore(&vcpu->arch.irqs_pending_lock, flags); 500 501 /* Check AIA high interrupts */ 502 if (!ret) 503 ret = kvm_riscv_vcpu_aia_has_interrupts(vcpu, mask); 504 505 return ret; 506 } 507 508 void __kvm_riscv_vcpu_power_off(struct kvm_vcpu *vcpu) 509 { 510 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED); 511 kvm_make_request(KVM_REQ_SLEEP, vcpu); 512 kvm_vcpu_kick(vcpu); 513 } 514 515 void kvm_riscv_vcpu_power_off(struct kvm_vcpu *vcpu) 516 { 517 spin_lock(&vcpu->arch.mp_state_lock); 518 __kvm_riscv_vcpu_power_off(vcpu); 519 spin_unlock(&vcpu->arch.mp_state_lock); 520 } 521 522 void __kvm_riscv_vcpu_power_on(struct kvm_vcpu *vcpu) 523 { 524 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE); 525 kvm_vcpu_wake_up(vcpu); 526 } 527 528 void kvm_riscv_vcpu_power_on(struct kvm_vcpu *vcpu) 529 { 530 spin_lock(&vcpu->arch.mp_state_lock); 531 __kvm_riscv_vcpu_power_on(vcpu); 532 spin_unlock(&vcpu->arch.mp_state_lock); 533 } 534 535 bool kvm_riscv_vcpu_stopped(struct kvm_vcpu *vcpu) 536 { 537 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED; 538 } 539 540 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu, 541 struct kvm_mp_state *mp_state) 542 { 543 *mp_state = READ_ONCE(vcpu->arch.mp_state); 544 545 return 0; 546 } 547 548 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu, 549 struct kvm_mp_state *mp_state) 550 { 551 int ret = 0; 552 553 spin_lock(&vcpu->arch.mp_state_lock); 554 555 switch (mp_state->mp_state) { 556 case KVM_MP_STATE_RUNNABLE: 557 WRITE_ONCE(vcpu->arch.mp_state, *mp_state); 558 break; 559 case KVM_MP_STATE_STOPPED: 560 __kvm_riscv_vcpu_power_off(vcpu); 561 break; 562 case KVM_MP_STATE_INIT_RECEIVED: 563 if (vcpu->kvm->arch.mp_state_reset) 564 kvm_riscv_reset_vcpu(vcpu, false); 565 else 566 ret = -EINVAL; 567 break; 568 default: 569 ret = -EINVAL; 570 } 571 572 spin_unlock(&vcpu->arch.mp_state_lock); 573 574 return ret; 575 } 576 577 int kvm_arch_vcpu_ioctl_set_guest_debug(struct kvm_vcpu *vcpu, 578 struct kvm_guest_debug *dbg) 579 { 580 if (dbg->control & KVM_GUESTDBG_ENABLE) 581 vcpu->guest_debug = dbg->control; 582 else 583 vcpu->guest_debug = 0; 584 585 kvm_riscv_vcpu_config_guest_debug(vcpu); 586 return 0; 587 } 588 589 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu) 590 { 591 void *nsh; 592 struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr; 593 594 /* 595 * If VCPU is being reloaded on the same physical CPU and no 596 * other KVM VCPU has run on this CPU since it was last put, 597 * we can skip the expensive CSR and HGATP writes. 598 * 599 * Note: If a new CSR is added to this fast-path skip block, 600 * make sure that 'csr_dirty' is set to true in any 601 * ioctl (e.g., KVM_SET_ONE_REG) that modifies it. 602 */ 603 if (vcpu != __this_cpu_read(kvm_former_vcpu)) 604 __this_cpu_write(kvm_former_vcpu, vcpu); 605 else if (vcpu->arch.last_exit_cpu == cpu && !vcpu->arch.csr_dirty) 606 goto csr_restore_done; 607 608 vcpu->arch.csr_dirty = false; 609 610 /* 611 * Load VCPU config CSRs before other CSRs because 612 * the read/write behaviour of certain CSRs change 613 * based on VCPU config CSRs. 614 */ 615 kvm_riscv_vcpu_config_load(vcpu); 616 617 if (kvm_riscv_nacl_sync_csr_available()) { 618 nsh = nacl_shmem(); 619 nacl_csr_write(nsh, CSR_VSSTATUS, csr->vsstatus); 620 nacl_csr_write(nsh, CSR_VSIE, csr->vsie); 621 nacl_csr_write(nsh, CSR_VSTVEC, csr->vstvec); 622 nacl_csr_write(nsh, CSR_VSSCRATCH, csr->vsscratch); 623 nacl_csr_write(nsh, CSR_VSEPC, csr->vsepc); 624 nacl_csr_write(nsh, CSR_VSCAUSE, csr->vscause); 625 nacl_csr_write(nsh, CSR_VSTVAL, csr->vstval); 626 nacl_csr_write(nsh, CSR_HVIP, csr->hvip); 627 nacl_csr_write(nsh, CSR_VSATP, csr->vsatp); 628 } else { 629 csr_write(CSR_VSSTATUS, csr->vsstatus); 630 csr_write(CSR_VSIE, csr->vsie); 631 csr_write(CSR_VSTVEC, csr->vstvec); 632 csr_write(CSR_VSSCRATCH, csr->vsscratch); 633 csr_write(CSR_VSEPC, csr->vsepc); 634 csr_write(CSR_VSCAUSE, csr->vscause); 635 csr_write(CSR_VSTVAL, csr->vstval); 636 csr_write(CSR_HVIP, csr->hvip); 637 csr_write(CSR_VSATP, csr->vsatp); 638 } 639 640 kvm_riscv_mmu_update_hgatp(vcpu); 641 642 kvm_riscv_vcpu_aia_load(vcpu, cpu); 643 644 csr_restore_done: 645 kvm_riscv_vcpu_timer_restore(vcpu); 646 647 kvm_riscv_vcpu_host_fp_save(&vcpu->arch.host_context); 648 kvm_riscv_vcpu_guest_fp_restore(&vcpu->arch.guest_context, 649 vcpu->arch.isa); 650 get_cpu_vector_context(); 651 kvm_riscv_vcpu_host_vector_save(&vcpu->arch.host_context); 652 kvm_riscv_vcpu_guest_vector_restore(&vcpu->arch.guest_context, 653 vcpu->arch.isa); 654 put_cpu_vector_context(); 655 656 kvm_make_request(KVM_REQ_STEAL_UPDATE, vcpu); 657 658 vcpu->cpu = cpu; 659 } 660 661 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu) 662 { 663 void *nsh; 664 struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr; 665 666 vcpu->cpu = -1; 667 668 kvm_riscv_vcpu_aia_put(vcpu); 669 670 kvm_riscv_vcpu_guest_fp_save(&vcpu->arch.guest_context, 671 vcpu->arch.isa); 672 kvm_riscv_vcpu_host_fp_restore(&vcpu->arch.host_context); 673 674 kvm_riscv_vcpu_timer_save(vcpu); 675 get_cpu_vector_context(); 676 kvm_riscv_vcpu_guest_vector_save(&vcpu->arch.guest_context, 677 vcpu->arch.isa); 678 kvm_riscv_vcpu_host_vector_restore(&vcpu->arch.host_context); 679 put_cpu_vector_context(); 680 681 if (kvm_riscv_nacl_available()) { 682 nsh = nacl_shmem(); 683 csr->vsstatus = nacl_csr_read(nsh, CSR_VSSTATUS); 684 csr->vsie = nacl_csr_read(nsh, CSR_VSIE); 685 csr->vstvec = nacl_csr_read(nsh, CSR_VSTVEC); 686 csr->vsscratch = nacl_csr_read(nsh, CSR_VSSCRATCH); 687 csr->vsepc = nacl_csr_read(nsh, CSR_VSEPC); 688 csr->vscause = nacl_csr_read(nsh, CSR_VSCAUSE); 689 csr->vstval = nacl_csr_read(nsh, CSR_VSTVAL); 690 csr->hvip = nacl_csr_read(nsh, CSR_HVIP); 691 csr->vsatp = nacl_csr_read(nsh, CSR_VSATP); 692 } else { 693 csr->vsstatus = csr_read(CSR_VSSTATUS); 694 csr->vsie = csr_read(CSR_VSIE); 695 csr->vstvec = csr_read(CSR_VSTVEC); 696 csr->vsscratch = csr_read(CSR_VSSCRATCH); 697 csr->vsepc = csr_read(CSR_VSEPC); 698 csr->vscause = csr_read(CSR_VSCAUSE); 699 csr->vstval = csr_read(CSR_VSTVAL); 700 csr->hvip = csr_read(CSR_HVIP); 701 csr->vsatp = csr_read(CSR_VSATP); 702 } 703 } 704 705 /** 706 * kvm_riscv_check_vcpu_requests - check and handle pending vCPU requests 707 * @vcpu: the VCPU pointer 708 * 709 * Return: 1 if we should enter the guest 710 * 0 if we should exit to userspace 711 */ 712 static int kvm_riscv_check_vcpu_requests(struct kvm_vcpu *vcpu) 713 { 714 struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu); 715 716 if (kvm_request_pending(vcpu)) { 717 if (kvm_check_request(KVM_REQ_SLEEP, vcpu)) { 718 kvm_vcpu_srcu_read_unlock(vcpu); 719 rcuwait_wait_event(wait, 720 (!kvm_riscv_vcpu_stopped(vcpu)) && (!vcpu->arch.pause), 721 TASK_INTERRUPTIBLE); 722 kvm_vcpu_srcu_read_lock(vcpu); 723 724 if (kvm_riscv_vcpu_stopped(vcpu) || vcpu->arch.pause) { 725 /* 726 * Awaken to handle a signal, request to 727 * sleep again later. 728 */ 729 kvm_make_request(KVM_REQ_SLEEP, vcpu); 730 } 731 } 732 733 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 734 kvm_riscv_reset_vcpu(vcpu, true); 735 736 if (kvm_check_request(KVM_REQ_UPDATE_HGATP, vcpu)) 737 kvm_riscv_mmu_update_hgatp(vcpu); 738 739 if (kvm_check_request(KVM_REQ_FENCE_I, vcpu)) 740 kvm_riscv_fence_i_process(vcpu); 741 742 if (kvm_check_request(KVM_REQ_TLB_FLUSH, vcpu)) 743 kvm_riscv_tlb_flush_process(vcpu); 744 745 if (kvm_check_request(KVM_REQ_HFENCE_VVMA_ALL, vcpu)) 746 kvm_riscv_hfence_vvma_all_process(vcpu); 747 748 if (kvm_check_request(KVM_REQ_HFENCE, vcpu)) 749 kvm_riscv_hfence_process(vcpu); 750 751 if (kvm_check_request(KVM_REQ_STEAL_UPDATE, vcpu)) 752 kvm_riscv_vcpu_record_steal_time(vcpu); 753 754 if (kvm_dirty_ring_check_request(vcpu)) 755 return 0; 756 } 757 758 return 1; 759 } 760 761 static void kvm_riscv_update_hvip(struct kvm_vcpu *vcpu) 762 { 763 struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr; 764 765 ncsr_write(CSR_HVIP, csr->hvip); 766 kvm_riscv_vcpu_aia_update_hvip(vcpu); 767 } 768 769 static __always_inline void kvm_riscv_vcpu_swap_in_guest_state(struct kvm_vcpu *vcpu) 770 { 771 struct kvm_vcpu_zicfiss_csr *zicficsr = &vcpu->arch.zicfiss_csr; 772 struct kvm_vcpu_smstateen_csr *smcsr = &vcpu->arch.smstateen_csr; 773 struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr; 774 775 vcpu->arch.host_scounteren = csr_swap(CSR_SCOUNTEREN, csr->scounteren); 776 vcpu->arch.host_senvcfg = csr_swap(CSR_SENVCFG, csr->senvcfg); 777 if (riscv_has_extension_unlikely(RISCV_ISA_EXT_SMSTATEEN)) 778 vcpu->arch.host_sstateen0 = csr_swap(CSR_SSTATEEN0, smcsr->sstateen0); 779 if (riscv_has_extension_unlikely(RISCV_ISA_EXT_ZICFISS)) 780 csr_write(CSR_SSP, zicficsr->ssp); 781 } 782 783 static __always_inline void kvm_riscv_vcpu_swap_in_host_state(struct kvm_vcpu *vcpu) 784 { 785 struct kvm_vcpu_zicfiss_csr *zicficsr = &vcpu->arch.zicfiss_csr; 786 struct kvm_vcpu_smstateen_csr *smcsr = &vcpu->arch.smstateen_csr; 787 struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr; 788 789 csr->scounteren = csr_swap(CSR_SCOUNTEREN, vcpu->arch.host_scounteren); 790 csr->senvcfg = csr_swap(CSR_SENVCFG, vcpu->arch.host_senvcfg); 791 if (riscv_has_extension_unlikely(RISCV_ISA_EXT_SMSTATEEN)) 792 smcsr->sstateen0 = csr_swap(CSR_SSTATEEN0, vcpu->arch.host_sstateen0); 793 if (riscv_has_extension_unlikely(RISCV_ISA_EXT_ZICFISS)) 794 zicficsr->ssp = csr_swap(CSR_SSP, 0); 795 } 796 797 /* 798 * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while 799 * the vCPU is running. 800 * 801 * This must be noinstr as instrumentation may make use of RCU, and this is not 802 * safe during the EQS. 803 */ 804 static void noinstr kvm_riscv_vcpu_enter_exit(struct kvm_vcpu *vcpu, 805 struct kvm_cpu_trap *trap) 806 { 807 void *nsh; 808 struct kvm_cpu_context *gcntx = &vcpu->arch.guest_context; 809 struct kvm_cpu_context *hcntx = &vcpu->arch.host_context; 810 811 /* 812 * We save trap CSRs (such as SEPC, SCAUSE, STVAL, HTVAL, and 813 * HTINST) here because we do local_irq_enable() after this 814 * function in kvm_arch_vcpu_ioctl_run() which can result in 815 * an interrupt immediately after local_irq_enable() and can 816 * potentially change trap CSRs. 817 */ 818 819 kvm_riscv_vcpu_swap_in_guest_state(vcpu); 820 guest_state_enter_irqoff(); 821 822 /* sstatus.VS != SR_VS_OFF is guaranteed when NEED_RESTORE is set */ 823 if (current->thread.riscv_v_flags & RISCV_V_VCPU_NEED_RESTORE) { 824 current->thread.riscv_v_flags &= ~RISCV_V_VCPU_NEED_RESTORE; 825 current->thread.riscv_v_flags |= RISCV_V_VCPU_CTX; 826 __kvm_riscv_vector_restore(gcntx); 827 gcntx->sstatus = (gcntx->sstatus & ~SR_VS) | SR_VS_CLEAN; 828 } 829 830 if (kvm_riscv_nacl_sync_sret_available()) { 831 nsh = nacl_shmem(); 832 833 if (kvm_riscv_nacl_autoswap_csr_available()) { 834 hcntx->hstatus = 835 nacl_csr_read(nsh, CSR_HSTATUS); 836 nacl_scratch_write_long(nsh, 837 SBI_NACL_SHMEM_AUTOSWAP_OFFSET + 838 SBI_NACL_SHMEM_AUTOSWAP_HSTATUS, 839 gcntx->hstatus); 840 nacl_scratch_write_long(nsh, 841 SBI_NACL_SHMEM_AUTOSWAP_OFFSET, 842 SBI_NACL_SHMEM_AUTOSWAP_FLAG_HSTATUS); 843 } else if (kvm_riscv_nacl_sync_csr_available()) { 844 hcntx->hstatus = nacl_csr_swap(nsh, 845 CSR_HSTATUS, gcntx->hstatus); 846 } else { 847 hcntx->hstatus = csr_swap(CSR_HSTATUS, gcntx->hstatus); 848 } 849 850 nacl_scratch_write_longs(nsh, 851 SBI_NACL_SHMEM_SRET_OFFSET + 852 SBI_NACL_SHMEM_SRET_X(1), 853 &gcntx->ra, 854 SBI_NACL_SHMEM_SRET_X_LAST); 855 856 __kvm_riscv_nacl_switch_to(&vcpu->arch, SBI_EXT_NACL, 857 SBI_EXT_NACL_SYNC_SRET); 858 859 if (kvm_riscv_nacl_autoswap_csr_available()) { 860 nacl_scratch_write_long(nsh, 861 SBI_NACL_SHMEM_AUTOSWAP_OFFSET, 862 0); 863 gcntx->hstatus = nacl_scratch_read_long(nsh, 864 SBI_NACL_SHMEM_AUTOSWAP_OFFSET + 865 SBI_NACL_SHMEM_AUTOSWAP_HSTATUS); 866 } else { 867 gcntx->hstatus = csr_swap(CSR_HSTATUS, hcntx->hstatus); 868 } 869 870 trap->htval = nacl_csr_read(nsh, CSR_HTVAL); 871 trap->htinst = nacl_csr_read(nsh, CSR_HTINST); 872 } else { 873 hcntx->hstatus = csr_swap(CSR_HSTATUS, gcntx->hstatus); 874 875 __kvm_riscv_switch_to(&vcpu->arch); 876 877 gcntx->hstatus = csr_swap(CSR_HSTATUS, hcntx->hstatus); 878 879 trap->htval = csr_read(CSR_HTVAL); 880 trap->htinst = csr_read(CSR_HTINST); 881 } 882 883 trap->sepc = gcntx->sepc; 884 trap->scause = csr_read(CSR_SCAUSE); 885 trap->stval = csr_read(CSR_STVAL); 886 887 vcpu->arch.last_exit_cpu = vcpu->cpu; 888 guest_state_exit_irqoff(); 889 kvm_riscv_vcpu_swap_in_host_state(vcpu); 890 } 891 892 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) 893 { 894 int ret; 895 struct kvm_cpu_trap trap; 896 struct kvm_run *run = vcpu->run; 897 898 if (!vcpu->arch.ran_atleast_once) 899 kvm_riscv_vcpu_config_ran_once(vcpu); 900 901 /* Mark this VCPU ran at least once */ 902 vcpu->arch.ran_atleast_once = true; 903 904 kvm_vcpu_srcu_read_lock(vcpu); 905 906 switch (run->exit_reason) { 907 case KVM_EXIT_MMIO: 908 /* Process MMIO value returned from user-space */ 909 ret = kvm_riscv_vcpu_mmio_return(vcpu, vcpu->run); 910 break; 911 case KVM_EXIT_RISCV_SBI: 912 /* Process SBI value returned from user-space */ 913 ret = kvm_riscv_vcpu_sbi_return(vcpu, vcpu->run); 914 break; 915 case KVM_EXIT_RISCV_CSR: 916 /* Process CSR value returned from user-space */ 917 ret = kvm_riscv_vcpu_csr_return(vcpu, vcpu->run); 918 break; 919 default: 920 ret = 0; 921 break; 922 } 923 if (ret) { 924 kvm_vcpu_srcu_read_unlock(vcpu); 925 return ret; 926 } 927 928 if (!vcpu->wants_to_run) { 929 kvm_vcpu_srcu_read_unlock(vcpu); 930 return -EINTR; 931 } 932 933 vcpu_load(vcpu); 934 935 kvm_sigset_activate(vcpu); 936 937 ret = 1; 938 run->exit_reason = KVM_EXIT_UNKNOWN; 939 while (ret > 0) { 940 /* Check conditions before entering the guest */ 941 ret = kvm_xfer_to_guest_mode_handle_work(vcpu); 942 if (ret) 943 continue; 944 ret = 1; 945 946 kvm_riscv_gstage_vmid_update(vcpu); 947 948 ret = kvm_riscv_check_vcpu_requests(vcpu); 949 if (ret <= 0) 950 continue; 951 952 preempt_disable(); 953 954 /* Update AIA HW state before entering guest */ 955 ret = kvm_riscv_vcpu_aia_update(vcpu); 956 if (ret <= 0) { 957 preempt_enable(); 958 continue; 959 } 960 961 local_irq_disable(); 962 963 /* 964 * Ensure we set mode to IN_GUEST_MODE after we disable 965 * interrupts and before the final VCPU requests check. 966 * See the comment in kvm_vcpu_exiting_guest_mode() and 967 * Documentation/virt/kvm/vcpu-requests.rst 968 */ 969 vcpu->mode = IN_GUEST_MODE; 970 971 kvm_vcpu_srcu_read_unlock(vcpu); 972 smp_mb__after_srcu_read_unlock(); 973 974 /* 975 * We might have got VCPU interrupts updated asynchronously 976 * so update it in HW. 977 */ 978 kvm_riscv_vcpu_flush_interrupts(vcpu); 979 980 /* Update HVIP CSR for current CPU */ 981 kvm_riscv_update_hvip(vcpu); 982 983 if (kvm_riscv_gstage_vmid_ver_changed(&vcpu->kvm->arch.vmid) || 984 kvm_request_pending(vcpu) || 985 xfer_to_guest_mode_work_pending()) { 986 vcpu->mode = OUTSIDE_GUEST_MODE; 987 local_irq_enable(); 988 preempt_enable(); 989 kvm_vcpu_srcu_read_lock(vcpu); 990 continue; 991 } 992 993 /* 994 * Sanitize VMID mappings cached (TLB) on current CPU 995 * 996 * Note: This should be done after G-stage VMID has been 997 * updated using kvm_riscv_gstage_vmid_ver_changed() 998 */ 999 kvm_riscv_local_tlb_sanitize(vcpu); 1000 1001 trace_kvm_entry(vcpu); 1002 1003 guest_timing_enter_irqoff(); 1004 1005 kvm_riscv_vcpu_enter_exit(vcpu, &trap); 1006 1007 vcpu->mode = OUTSIDE_GUEST_MODE; 1008 vcpu->stat.exits++; 1009 1010 /* Syncup interrupts state with HW */ 1011 kvm_riscv_vcpu_sync_interrupts(vcpu); 1012 1013 /* 1014 * We must ensure that any pending interrupts are taken before 1015 * we exit guest timing so that timer ticks are accounted as 1016 * guest time. Transiently unmask interrupts so that any 1017 * pending interrupts are taken. 1018 * 1019 * There's no barrier which ensures that pending interrupts are 1020 * recognised, so we just hope that the CPU takes any pending 1021 * interrupts between the enable and disable. 1022 */ 1023 local_irq_enable(); 1024 local_irq_disable(); 1025 1026 guest_timing_exit_irqoff(); 1027 1028 local_irq_enable(); 1029 1030 trace_kvm_exit(&trap); 1031 1032 preempt_enable(); 1033 1034 kvm_vcpu_srcu_read_lock(vcpu); 1035 1036 ret = kvm_riscv_vcpu_exit(vcpu, run, &trap); 1037 } 1038 1039 kvm_sigset_deactivate(vcpu); 1040 1041 vcpu_put(vcpu); 1042 1043 kvm_vcpu_srcu_read_unlock(vcpu); 1044 1045 return ret; 1046 } 1047