1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * KVM paravirt_ops implementation 4 * 5 * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com> 6 * Copyright IBM Corporation, 2007 7 * Authors: Anthony Liguori <aliguori@us.ibm.com> 8 */ 9 10 #define pr_fmt(fmt) "kvm-guest: " fmt 11 12 #include <linux/context_tracking.h> 13 #include <linux/init.h> 14 #include <linux/irq.h> 15 #include <linux/kernel.h> 16 #include <linux/kvm_para.h> 17 #include <linux/cpu.h> 18 #include <linux/mm.h> 19 #include <linux/highmem.h> 20 #include <linux/hardirq.h> 21 #include <linux/notifier.h> 22 #include <linux/reboot.h> 23 #include <linux/hash.h> 24 #include <linux/sched.h> 25 #include <linux/slab.h> 26 #include <linux/kprobes.h> 27 #include <linux/nmi.h> 28 #include <linux/swait.h> 29 #include <linux/syscore_ops.h> 30 #include <linux/cc_platform.h> 31 #include <linux/efi.h> 32 #include <linux/kvm_types.h> 33 #include <linux/sched/cputime.h> 34 #include <asm/timer.h> 35 #include <asm/cpu.h> 36 #include <asm/traps.h> 37 #include <asm/desc.h> 38 #include <asm/tlbflush.h> 39 #include <asm/apic.h> 40 #include <asm/apicdef.h> 41 #include <asm/hypervisor.h> 42 #include <asm/mtrr.h> 43 #include <asm/tlb.h> 44 #include <asm/cpuidle_haltpoll.h> 45 #include <asm/msr.h> 46 #include <asm/ptrace.h> 47 #include <asm/reboot.h> 48 #include <asm/svm.h> 49 #include <asm/e820/api.h> 50 51 DEFINE_STATIC_KEY_FALSE_RO(kvm_async_pf_enabled); 52 53 static int kvmapf = 1; 54 55 static int __init parse_no_kvmapf(char *arg) 56 { 57 kvmapf = 0; 58 return 0; 59 } 60 61 early_param("no-kvmapf", parse_no_kvmapf); 62 63 static int steal_acc = 1; 64 static int __init parse_no_stealacc(char *arg) 65 { 66 steal_acc = 0; 67 return 0; 68 } 69 70 early_param("no-steal-acc", parse_no_stealacc); 71 72 static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled); 73 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64); 74 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible; 75 static int has_steal_clock = 0; 76 77 static int has_guest_poll = 0; 78 /* 79 * No need for any "IO delay" on KVM 80 */ 81 static void kvm_io_delay(void) 82 { 83 } 84 85 #define KVM_TASK_SLEEP_HASHBITS 8 86 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS) 87 88 struct kvm_task_sleep_node { 89 struct hlist_node link; 90 struct swait_queue_head wq; 91 u32 token; 92 int cpu; 93 }; 94 95 static struct kvm_task_sleep_head { 96 raw_spinlock_t lock; 97 struct hlist_head list; 98 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE]; 99 100 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b, 101 u32 token) 102 { 103 struct hlist_node *p; 104 105 hlist_for_each(p, &b->list) { 106 struct kvm_task_sleep_node *n = 107 hlist_entry(p, typeof(*n), link); 108 if (n->token == token) 109 return n; 110 } 111 112 return NULL; 113 } 114 115 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n) 116 { 117 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS); 118 struct kvm_task_sleep_head *b = &async_pf_sleepers[key]; 119 struct kvm_task_sleep_node *e; 120 121 raw_spin_lock(&b->lock); 122 e = _find_apf_task(b, token); 123 if (e) { 124 /* dummy entry exist -> wake up was delivered ahead of PF */ 125 hlist_del(&e->link); 126 raw_spin_unlock(&b->lock); 127 kfree(e); 128 return false; 129 } 130 131 n->token = token; 132 n->cpu = smp_processor_id(); 133 init_swait_queue_head(&n->wq); 134 hlist_add_head(&n->link, &b->list); 135 raw_spin_unlock(&b->lock); 136 return true; 137 } 138 139 /* 140 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled 141 * @token: Token to identify the sleep node entry 142 * 143 * Invoked from the async pagefault handling code or from the VM exit page 144 * fault handler. In both cases RCU is watching. 145 */ 146 void kvm_async_pf_task_wait_schedule(u32 token) 147 { 148 struct kvm_task_sleep_node n; 149 DECLARE_SWAITQUEUE(wait); 150 151 lockdep_assert_irqs_disabled(); 152 153 if (!kvm_async_pf_queue_task(token, &n)) 154 return; 155 156 for (;;) { 157 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE); 158 if (hlist_unhashed(&n.link)) 159 break; 160 161 local_irq_enable(); 162 schedule(); 163 local_irq_disable(); 164 } 165 finish_swait(&n.wq, &wait); 166 } 167 EXPORT_SYMBOL_FOR_KVM(kvm_async_pf_task_wait_schedule); 168 169 static void apf_task_wake_one(struct kvm_task_sleep_node *n) 170 { 171 hlist_del_init(&n->link); 172 if (swq_has_sleeper(&n->wq)) 173 swake_up_one(&n->wq); 174 } 175 176 static void apf_task_wake_all(void) 177 { 178 int i; 179 180 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) { 181 struct kvm_task_sleep_head *b = &async_pf_sleepers[i]; 182 struct kvm_task_sleep_node *n; 183 struct hlist_node *p, *next; 184 185 raw_spin_lock(&b->lock); 186 hlist_for_each_safe(p, next, &b->list) { 187 n = hlist_entry(p, typeof(*n), link); 188 if (n->cpu == smp_processor_id()) 189 apf_task_wake_one(n); 190 } 191 raw_spin_unlock(&b->lock); 192 } 193 } 194 195 static void kvm_async_pf_task_wake(u32 token) 196 { 197 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS); 198 struct kvm_task_sleep_head *b = &async_pf_sleepers[key]; 199 struct kvm_task_sleep_node *n, *dummy = NULL; 200 201 if (token == ~0) { 202 apf_task_wake_all(); 203 return; 204 } 205 206 again: 207 raw_spin_lock(&b->lock); 208 n = _find_apf_task(b, token); 209 if (!n) { 210 /* 211 * Async #PF not yet handled, add a dummy entry for the token. 212 * Allocating the token must be down outside of the raw lock 213 * as the allocator is preemptible on PREEMPT_RT kernels. 214 */ 215 if (!dummy) { 216 raw_spin_unlock(&b->lock); 217 dummy = kzalloc(sizeof(*dummy), GFP_ATOMIC); 218 219 /* 220 * Continue looping on allocation failure, eventually 221 * the async #PF will be handled and allocating a new 222 * node will be unnecessary. 223 */ 224 if (!dummy) 225 cpu_relax(); 226 227 /* 228 * Recheck for async #PF completion before enqueueing 229 * the dummy token to avoid duplicate list entries. 230 */ 231 goto again; 232 } 233 dummy->token = token; 234 dummy->cpu = smp_processor_id(); 235 init_swait_queue_head(&dummy->wq); 236 hlist_add_head(&dummy->link, &b->list); 237 dummy = NULL; 238 } else { 239 apf_task_wake_one(n); 240 } 241 raw_spin_unlock(&b->lock); 242 243 /* A dummy token might be allocated and ultimately not used. */ 244 kfree(dummy); 245 } 246 247 noinstr u32 kvm_read_and_reset_apf_flags(void) 248 { 249 u32 flags = 0; 250 251 if (__this_cpu_read(async_pf_enabled)) { 252 flags = __this_cpu_read(apf_reason.flags); 253 __this_cpu_write(apf_reason.flags, 0); 254 } 255 256 return flags; 257 } 258 EXPORT_SYMBOL_FOR_KVM(kvm_read_and_reset_apf_flags); 259 260 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token) 261 { 262 u32 flags = kvm_read_and_reset_apf_flags(); 263 irqentry_state_t state; 264 265 if (!flags) 266 return false; 267 268 state = irqentry_enter(regs); 269 instrumentation_begin(); 270 271 /* 272 * If the host managed to inject an async #PF into an interrupt 273 * disabled region, then die hard as this is not going to end well 274 * and the host side is seriously broken. 275 */ 276 if (unlikely(!(regs->flags & X86_EFLAGS_IF))) 277 panic("Host injected async #PF in interrupt disabled region\n"); 278 279 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) { 280 if (unlikely(!(user_mode(regs)))) 281 panic("Host injected async #PF in kernel mode\n"); 282 /* Page is swapped out by the host. */ 283 kvm_async_pf_task_wait_schedule(token); 284 } else { 285 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags); 286 } 287 288 instrumentation_end(); 289 irqentry_exit(regs, state); 290 return true; 291 } 292 293 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt) 294 { 295 struct pt_regs *old_regs = set_irq_regs(regs); 296 u32 token; 297 298 apic_eoi(); 299 300 inc_irq_stat(irq_hv_callback_count); 301 302 if (__this_cpu_read(async_pf_enabled)) { 303 token = __this_cpu_read(apf_reason.token); 304 kvm_async_pf_task_wake(token); 305 __this_cpu_write(apf_reason.token, 0); 306 wrmsrq(MSR_KVM_ASYNC_PF_ACK, 1); 307 } 308 309 set_irq_regs(old_regs); 310 } 311 312 static void __init paravirt_ops_setup(void) 313 { 314 pv_info.name = "KVM"; 315 316 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY)) 317 pv_ops.cpu.io_delay = kvm_io_delay; 318 319 #ifdef CONFIG_X86_IO_APIC 320 no_timer_check = 1; 321 #endif 322 } 323 324 static void kvm_register_steal_time(void) 325 { 326 int cpu = smp_processor_id(); 327 struct kvm_steal_time *st = &per_cpu(steal_time, cpu); 328 329 if (!has_steal_clock) 330 return; 331 332 wrmsrq(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED)); 333 pr_debug("stealtime: cpu %d, msr %llx\n", cpu, 334 (unsigned long long) slow_virt_to_phys(st)); 335 } 336 337 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED; 338 339 static notrace __maybe_unused void kvm_guest_apic_eoi_write(void) 340 { 341 /** 342 * This relies on __test_and_clear_bit to modify the memory 343 * in a way that is atomic with respect to the local CPU. 344 * The hypervisor only accesses this memory from the local CPU so 345 * there's no need for lock or memory barriers. 346 * An optimization barrier is implied in apic write. 347 */ 348 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi))) 349 return; 350 apic_native_eoi(); 351 } 352 353 static void kvm_guest_cpu_init(void) 354 { 355 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) { 356 u64 pa; 357 358 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled)); 359 360 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason)); 361 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT; 362 363 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT)) 364 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT; 365 366 wrmsrq(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR); 367 368 wrmsrq(MSR_KVM_ASYNC_PF_EN, pa); 369 __this_cpu_write(async_pf_enabled, true); 370 pr_debug("setup async PF for cpu %d\n", smp_processor_id()); 371 } 372 373 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) { 374 unsigned long pa; 375 376 /* Size alignment is implied but just to make it explicit. */ 377 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4); 378 __this_cpu_write(kvm_apic_eoi, 0); 379 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi)) 380 | KVM_MSR_ENABLED; 381 wrmsrq(MSR_KVM_PV_EOI_EN, pa); 382 } 383 384 if (has_steal_clock) 385 kvm_register_steal_time(); 386 } 387 388 static void kvm_pv_disable_apf(void) 389 { 390 if (!__this_cpu_read(async_pf_enabled)) 391 return; 392 393 wrmsrq(MSR_KVM_ASYNC_PF_EN, 0); 394 __this_cpu_write(async_pf_enabled, false); 395 396 pr_debug("disable async PF for cpu %d\n", smp_processor_id()); 397 } 398 399 static void kvm_disable_steal_time(void) 400 { 401 if (!has_steal_clock) 402 return; 403 404 wrmsrq(MSR_KVM_STEAL_TIME, 0); 405 } 406 407 static u64 kvm_steal_clock(int cpu) 408 { 409 u64 steal; 410 struct kvm_steal_time *src; 411 int version; 412 413 src = &per_cpu(steal_time, cpu); 414 do { 415 version = src->version; 416 virt_rmb(); 417 steal = src->steal; 418 virt_rmb(); 419 } while ((version & 1) || (version != src->version)); 420 421 return steal; 422 } 423 424 static inline __init void __set_percpu_decrypted(void *ptr, unsigned long size) 425 { 426 early_set_memory_decrypted((unsigned long) ptr, size); 427 } 428 429 /* 430 * Iterate through all possible CPUs and map the memory region pointed 431 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once. 432 * 433 * Note: we iterate through all possible CPUs to ensure that CPUs 434 * hotplugged will have their per-cpu variable already mapped as 435 * decrypted. 436 */ 437 static void __init sev_map_percpu_data(void) 438 { 439 int cpu; 440 441 if (cc_vendor != CC_VENDOR_AMD || 442 !cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) 443 return; 444 445 for_each_possible_cpu(cpu) { 446 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason)); 447 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time)); 448 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi)); 449 } 450 } 451 452 static void kvm_guest_cpu_offline(bool shutdown) 453 { 454 kvm_disable_steal_time(); 455 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) 456 wrmsrq(MSR_KVM_PV_EOI_EN, 0); 457 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) 458 wrmsrq(MSR_KVM_MIGRATION_CONTROL, 0); 459 kvm_pv_disable_apf(); 460 if (!shutdown) 461 apf_task_wake_all(); 462 kvmclock_disable(); 463 } 464 465 static int kvm_cpu_online(unsigned int cpu) 466 { 467 unsigned long flags; 468 469 local_irq_save(flags); 470 kvm_guest_cpu_init(); 471 local_irq_restore(flags); 472 return 0; 473 } 474 475 #ifdef CONFIG_SMP 476 477 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask); 478 479 static bool pv_tlb_flush_supported(void) 480 { 481 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) && 482 !kvm_para_has_hint(KVM_HINTS_REALTIME) && 483 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) && 484 !boot_cpu_has(X86_FEATURE_MWAIT) && 485 (num_possible_cpus() != 1)); 486 } 487 488 static bool pv_ipi_supported(void) 489 { 490 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) && 491 (num_possible_cpus() != 1)); 492 } 493 494 static bool pv_sched_yield_supported(void) 495 { 496 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) && 497 !kvm_para_has_hint(KVM_HINTS_REALTIME) && 498 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) && 499 !boot_cpu_has(X86_FEATURE_MWAIT) && 500 (num_possible_cpus() != 1)); 501 } 502 503 #define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG) 504 505 static void __send_ipi_mask(const struct cpumask *mask, int vector) 506 { 507 unsigned long flags; 508 int cpu, min = 0, max = 0; 509 #ifdef CONFIG_X86_64 510 __uint128_t ipi_bitmap = 0; 511 #else 512 u64 ipi_bitmap = 0; 513 #endif 514 u32 apic_id, icr; 515 long ret; 516 517 if (cpumask_empty(mask)) 518 return; 519 520 local_irq_save(flags); 521 522 switch (vector) { 523 default: 524 icr = APIC_DM_FIXED | vector; 525 break; 526 case NMI_VECTOR: 527 icr = APIC_DM_NMI; 528 break; 529 } 530 531 for_each_cpu(cpu, mask) { 532 apic_id = per_cpu(x86_cpu_to_apicid, cpu); 533 if (!ipi_bitmap) { 534 min = max = apic_id; 535 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) { 536 ipi_bitmap <<= min - apic_id; 537 min = apic_id; 538 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) { 539 max = apic_id < max ? max : apic_id; 540 } else { 541 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap, 542 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr); 543 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld", 544 ret); 545 min = max = apic_id; 546 ipi_bitmap = 0; 547 } 548 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap); 549 } 550 551 if (ipi_bitmap) { 552 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap, 553 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr); 554 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld", 555 ret); 556 } 557 558 local_irq_restore(flags); 559 } 560 561 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector) 562 { 563 __send_ipi_mask(mask, vector); 564 } 565 566 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector) 567 { 568 unsigned int this_cpu = smp_processor_id(); 569 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask); 570 const struct cpumask *local_mask; 571 572 cpumask_copy(new_mask, mask); 573 cpumask_clear_cpu(this_cpu, new_mask); 574 local_mask = new_mask; 575 __send_ipi_mask(local_mask, vector); 576 } 577 578 static int __init setup_efi_kvm_sev_migration(void) 579 { 580 efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled"; 581 efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID; 582 efi_status_t status; 583 unsigned long size; 584 bool enabled; 585 586 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) || 587 !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) 588 return 0; 589 590 if (!efi_enabled(EFI_BOOT)) 591 return 0; 592 593 if (!efi_enabled(EFI_RUNTIME_SERVICES)) { 594 pr_info("%s : EFI runtime services are not enabled\n", __func__); 595 return 0; 596 } 597 598 size = sizeof(enabled); 599 600 /* Get variable contents into buffer */ 601 status = efi.get_variable(efi_sev_live_migration_enabled, 602 &efi_variable_guid, NULL, &size, &enabled); 603 604 if (status == EFI_NOT_FOUND) { 605 pr_info("%s : EFI live migration variable not found\n", __func__); 606 return 0; 607 } 608 609 if (status != EFI_SUCCESS) { 610 pr_info("%s : EFI variable retrieval failed\n", __func__); 611 return 0; 612 } 613 614 if (enabled == 0) { 615 pr_info("%s: live migration disabled in EFI\n", __func__); 616 return 0; 617 } 618 619 pr_info("%s : live migration enabled in EFI\n", __func__); 620 wrmsrq(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY); 621 622 return 1; 623 } 624 625 late_initcall(setup_efi_kvm_sev_migration); 626 627 /* 628 * Set the IPI entry points 629 */ 630 static __init void kvm_setup_pv_ipi(void) 631 { 632 apic_update_callback(send_IPI_mask, kvm_send_ipi_mask); 633 apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself); 634 pr_info("setup PV IPIs\n"); 635 } 636 637 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask) 638 { 639 int cpu; 640 641 native_send_call_func_ipi(mask); 642 643 /* Make sure other vCPUs get a chance to run if they need to. */ 644 for_each_cpu(cpu, mask) { 645 if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) { 646 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu)); 647 break; 648 } 649 } 650 } 651 652 static void kvm_flush_tlb_multi(const struct cpumask *cpumask, 653 const struct flush_tlb_info *info) 654 { 655 u8 state; 656 int cpu; 657 struct kvm_steal_time *src; 658 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask); 659 660 cpumask_copy(flushmask, cpumask); 661 /* 662 * We have to call flush only on online vCPUs. And 663 * queue flush_on_enter for pre-empted vCPUs 664 */ 665 for_each_cpu(cpu, flushmask) { 666 /* 667 * The local vCPU is never preempted, so we do not explicitly 668 * skip check for local vCPU - it will never be cleared from 669 * flushmask. 670 */ 671 src = &per_cpu(steal_time, cpu); 672 state = READ_ONCE(src->preempted); 673 if ((state & KVM_VCPU_PREEMPTED)) { 674 if (try_cmpxchg(&src->preempted, &state, 675 state | KVM_VCPU_FLUSH_TLB)) 676 __cpumask_clear_cpu(cpu, flushmask); 677 } 678 } 679 680 native_flush_tlb_multi(flushmask, info); 681 } 682 683 static __init int kvm_alloc_cpumask(void) 684 { 685 int cpu; 686 687 if (!kvm_para_available() || nopv) 688 return 0; 689 690 if (pv_tlb_flush_supported() || pv_ipi_supported()) 691 for_each_possible_cpu(cpu) { 692 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu), 693 GFP_KERNEL, cpu_to_node(cpu)); 694 } 695 696 return 0; 697 } 698 arch_initcall(kvm_alloc_cpumask); 699 700 static void __init kvm_smp_prepare_boot_cpu(void) 701 { 702 /* 703 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init() 704 * shares the guest physical address with the hypervisor. 705 */ 706 sev_map_percpu_data(); 707 708 kvm_guest_cpu_init(); 709 native_smp_prepare_boot_cpu(); 710 kvm_spinlock_init(); 711 } 712 713 static int kvm_cpu_down_prepare(unsigned int cpu) 714 { 715 unsigned long flags; 716 717 local_irq_save(flags); 718 kvm_guest_cpu_offline(false); 719 local_irq_restore(flags); 720 return 0; 721 } 722 723 #endif 724 725 static int kvm_suspend(void *data) 726 { 727 u64 val = 0; 728 729 kvm_guest_cpu_offline(false); 730 731 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 732 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) 733 rdmsrq(MSR_KVM_POLL_CONTROL, val); 734 has_guest_poll = !(val & 1); 735 #endif 736 return 0; 737 } 738 739 static void kvm_resume(void *data) 740 { 741 kvm_cpu_online(raw_smp_processor_id()); 742 743 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 744 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll) 745 wrmsrq(MSR_KVM_POLL_CONTROL, 0); 746 #endif 747 } 748 749 static const struct syscore_ops kvm_syscore_ops = { 750 .suspend = kvm_suspend, 751 .resume = kvm_resume, 752 }; 753 754 static struct syscore kvm_syscore = { 755 .ops = &kvm_syscore_ops, 756 }; 757 758 static void kvm_pv_guest_cpu_reboot(void *unused) 759 { 760 kvm_guest_cpu_offline(true); 761 } 762 763 static int kvm_pv_reboot_notify(struct notifier_block *nb, 764 unsigned long code, void *unused) 765 { 766 if (code == SYS_RESTART) 767 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1); 768 return NOTIFY_DONE; 769 } 770 771 static struct notifier_block kvm_pv_reboot_nb = { 772 .notifier_call = kvm_pv_reboot_notify, 773 }; 774 775 /* 776 * After a PV feature is registered, the host will keep writing to the 777 * registered memory location. If the guest happens to shutdown, this memory 778 * won't be valid. In cases like kexec, in which you install a new kernel, this 779 * means a random memory location will be kept being written. 780 */ 781 #ifdef CONFIG_CRASH_DUMP 782 static void kvm_crash_shutdown(struct pt_regs *regs) 783 { 784 kvm_guest_cpu_offline(true); 785 native_machine_crash_shutdown(regs); 786 } 787 #endif 788 789 #if defined(CONFIG_X86_32) || !defined(CONFIG_SMP) 790 bool __kvm_vcpu_is_preempted(long cpu); 791 792 __visible bool __kvm_vcpu_is_preempted(long cpu) 793 { 794 struct kvm_steal_time *src = &per_cpu(steal_time, cpu); 795 796 return !!(src->preempted & KVM_VCPU_PREEMPTED); 797 } 798 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted); 799 800 #else 801 802 #include <asm/asm-offsets.h> 803 804 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long); 805 806 /* 807 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and 808 * restoring to/from the stack. 809 */ 810 #define PV_VCPU_PREEMPTED_ASM \ 811 "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \ 812 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \ 813 "setne %al\n\t" 814 815 DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted, 816 PV_VCPU_PREEMPTED_ASM, .text); 817 #endif 818 819 static void __init kvm_guest_init(void) 820 { 821 int i; 822 823 paravirt_ops_setup(); 824 register_reboot_notifier(&kvm_pv_reboot_nb); 825 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) 826 raw_spin_lock_init(&async_pf_sleepers[i].lock); 827 828 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) { 829 has_steal_clock = 1; 830 static_call_update(pv_steal_clock, kvm_steal_clock); 831 832 pv_ops.lock.vcpu_is_preempted = 833 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted); 834 } 835 836 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) 837 apic_update_callback(eoi, kvm_guest_apic_eoi_write); 838 839 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) { 840 static_branch_enable(&kvm_async_pf_enabled); 841 sysvec_install(HYPERVISOR_CALLBACK_VECTOR, sysvec_kvm_asyncpf_interrupt); 842 } 843 844 #ifdef CONFIG_SMP 845 if (pv_tlb_flush_supported()) { 846 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi; 847 pr_info("KVM setup pv remote TLB flush\n"); 848 } 849 850 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu; 851 if (pv_sched_yield_supported()) { 852 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi; 853 pr_info("setup PV sched yield\n"); 854 } 855 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online", 856 kvm_cpu_online, kvm_cpu_down_prepare) < 0) 857 pr_err("failed to install cpu hotplug callbacks\n"); 858 #else 859 sev_map_percpu_data(); 860 kvm_guest_cpu_init(); 861 #endif 862 863 #ifdef CONFIG_CRASH_DUMP 864 machine_ops.crash_shutdown = kvm_crash_shutdown; 865 #endif 866 867 register_syscore(&kvm_syscore); 868 869 /* 870 * Hard lockup detection is enabled by default. Disable it, as guests 871 * can get false positives too easily, for example if the host is 872 * overcommitted. 873 */ 874 hardlockup_detector_disable(); 875 } 876 877 static noinline uint32_t __kvm_cpuid_base(void) 878 { 879 if (boot_cpu_data.cpuid_level < 0) 880 return 0; /* So we don't blow up on old processors */ 881 882 if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) 883 return cpuid_base_hypervisor(KVM_SIGNATURE, 0); 884 885 return 0; 886 } 887 888 static inline uint32_t kvm_cpuid_base(void) 889 { 890 static int kvm_cpuid_base = -1; 891 892 if (kvm_cpuid_base == -1) 893 kvm_cpuid_base = __kvm_cpuid_base(); 894 895 return kvm_cpuid_base; 896 } 897 898 bool kvm_para_available(void) 899 { 900 return kvm_cpuid_base() != 0; 901 } 902 EXPORT_SYMBOL_GPL(kvm_para_available); 903 904 unsigned int kvm_arch_para_features(void) 905 { 906 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES); 907 } 908 909 unsigned int kvm_arch_para_hints(void) 910 { 911 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES); 912 } 913 EXPORT_SYMBOL_GPL(kvm_arch_para_hints); 914 915 static uint32_t __init kvm_detect(void) 916 { 917 return kvm_cpuid_base(); 918 } 919 920 static void __init kvm_apic_init(void) 921 { 922 #ifdef CONFIG_SMP 923 if (pv_ipi_supported()) 924 kvm_setup_pv_ipi(); 925 #endif 926 } 927 928 static bool __init kvm_msi_ext_dest_id(void) 929 { 930 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID); 931 } 932 933 static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc) 934 { 935 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages, 936 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K); 937 } 938 939 static void __init kvm_init_platform(void) 940 { 941 u64 tolud = PFN_PHYS(e820__end_of_low_ram_pfn()); 942 /* 943 * Note, hardware requires variable MTRR ranges to be power-of-2 sized 944 * and naturally aligned. But when forcing guest MTRR state, Linux 945 * doesn't program the forced ranges into hardware. Don't bother doing 946 * the math to generate a technically-legal range. 947 */ 948 struct mtrr_var_range pci_hole = { 949 .base_lo = tolud | X86_MEMTYPE_UC, 950 .mask_lo = (u32)(~(SZ_4G - tolud - 1)) | MTRR_PHYSMASK_V, 951 .mask_hi = (BIT_ULL(boot_cpu_data.x86_phys_bits) - 1) >> 32, 952 }; 953 954 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) && 955 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) { 956 unsigned long nr_pages; 957 int i; 958 959 pv_ops.mmu.notify_page_enc_status_changed = 960 kvm_sev_hc_page_enc_status; 961 962 /* 963 * Reset the host's shared pages list related to kernel 964 * specific page encryption status settings before we load a 965 * new kernel by kexec. Reset the page encryption status 966 * during early boot instead of just before kexec to avoid SMP 967 * races during kvm_pv_guest_cpu_reboot(). 968 * NOTE: We cannot reset the complete shared pages list 969 * here as we need to retain the UEFI/OVMF firmware 970 * specific settings. 971 */ 972 973 for (i = 0; i < e820_table->nr_entries; i++) { 974 struct e820_entry *entry = &e820_table->entries[i]; 975 976 if (entry->type != E820_TYPE_RAM) 977 continue; 978 979 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE); 980 981 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr, 982 nr_pages, 983 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K); 984 } 985 986 /* 987 * Ensure that _bss_decrypted section is marked as decrypted in the 988 * shared pages list. 989 */ 990 early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted, 991 __end_bss_decrypted - __start_bss_decrypted, 0); 992 993 /* 994 * If not booted using EFI, enable Live migration support. 995 */ 996 if (!efi_enabled(EFI_BOOT)) 997 wrmsrq(MSR_KVM_MIGRATION_CONTROL, 998 KVM_MIGRATION_READY); 999 } 1000 kvmclock_init(); 1001 x86_platform.apic_post_init = kvm_apic_init; 1002 1003 /* 1004 * Set WB as the default cache mode for SEV-SNP and TDX, with a single 1005 * UC range for the legacy PCI hole, e.g. so that devices that expect 1006 * to get UC/WC mappings don't get surprised with WB. 1007 */ 1008 guest_force_mtrr_state(&pci_hole, 1, MTRR_TYPE_WRBACK); 1009 } 1010 1011 #if defined(CONFIG_AMD_MEM_ENCRYPT) 1012 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs) 1013 { 1014 /* RAX and CPL are already in the GHCB */ 1015 ghcb_set_rbx(ghcb, regs->bx); 1016 ghcb_set_rcx(ghcb, regs->cx); 1017 ghcb_set_rdx(ghcb, regs->dx); 1018 ghcb_set_rsi(ghcb, regs->si); 1019 } 1020 1021 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs) 1022 { 1023 /* No checking of the return state needed */ 1024 return true; 1025 } 1026 #endif 1027 1028 const __initconst struct hypervisor_x86 x86_hyper_kvm = { 1029 .name = "KVM", 1030 .detect = kvm_detect, 1031 .type = X86_HYPER_KVM, 1032 .init.guest_late_init = kvm_guest_init, 1033 .init.x2apic_available = kvm_para_available, 1034 .init.msi_ext_dest_id = kvm_msi_ext_dest_id, 1035 .init.init_platform = kvm_init_platform, 1036 #if defined(CONFIG_AMD_MEM_ENCRYPT) 1037 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare, 1038 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish, 1039 #endif 1040 }; 1041 1042 static __init int activate_jump_labels(void) 1043 { 1044 if (has_steal_clock) { 1045 static_key_slow_inc(¶virt_steal_enabled); 1046 if (steal_acc) 1047 static_key_slow_inc(¶virt_steal_rq_enabled); 1048 } 1049 1050 return 0; 1051 } 1052 arch_initcall(activate_jump_labels); 1053 1054 #ifdef CONFIG_PARAVIRT_SPINLOCKS 1055 1056 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */ 1057 static void kvm_kick_cpu(int cpu) 1058 { 1059 unsigned long flags = 0; 1060 u32 apicid; 1061 1062 apicid = per_cpu(x86_cpu_to_apicid, cpu); 1063 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid); 1064 } 1065 1066 #include <asm/qspinlock.h> 1067 1068 static void kvm_wait(u8 *ptr, u8 val) 1069 { 1070 if (in_nmi()) 1071 return; 1072 1073 /* 1074 * halt until it's our turn and kicked. Note that we do safe halt 1075 * for irq enabled case to avoid hang when lock info is overwritten 1076 * in irq spinlock slowpath and no spurious interrupt occur to save us. 1077 */ 1078 if (irqs_disabled()) { 1079 if (READ_ONCE(*ptr) == val) 1080 halt(); 1081 } else { 1082 local_irq_disable(); 1083 1084 /* safe_halt() will enable IRQ */ 1085 if (READ_ONCE(*ptr) == val) 1086 safe_halt(); 1087 else 1088 local_irq_enable(); 1089 } 1090 } 1091 1092 /* 1093 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present. 1094 */ 1095 void __init kvm_spinlock_init(void) 1096 { 1097 /* 1098 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs 1099 * are available. 1100 */ 1101 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) { 1102 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n"); 1103 goto out; 1104 } 1105 1106 if (num_possible_cpus() == 1) { 1107 pr_info("PV spinlocks disabled, single CPU\n"); 1108 goto out; 1109 } 1110 1111 if (nopvspin) { 1112 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n"); 1113 goto out; 1114 } 1115 1116 /* 1117 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an 1118 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is 1119 * preferred over native qspinlock when vCPU is preempted. 1120 */ 1121 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) { 1122 pr_info("PV spinlocks disabled, no host support\n"); 1123 return; 1124 } 1125 1126 pr_info("PV spinlocks enabled\n"); 1127 1128 __pv_init_lock_hash(); 1129 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath; 1130 pv_ops.lock.queued_spin_unlock = 1131 PV_CALLEE_SAVE(__pv_queued_spin_unlock); 1132 pv_ops.lock.wait = kvm_wait; 1133 pv_ops.lock.kick = kvm_kick_cpu; 1134 1135 /* 1136 * When PV spinlock is enabled which is preferred over 1137 * virt_spin_lock(), virt_spin_lock_key's value is meaningless. 1138 * Just disable it anyway. 1139 */ 1140 out: 1141 static_branch_disable(&virt_spin_lock_key); 1142 } 1143 1144 #endif /* CONFIG_PARAVIRT_SPINLOCKS */ 1145 1146 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 1147 1148 static void kvm_disable_host_haltpoll(void *i) 1149 { 1150 wrmsrq(MSR_KVM_POLL_CONTROL, 0); 1151 } 1152 1153 static void kvm_enable_host_haltpoll(void *i) 1154 { 1155 wrmsrq(MSR_KVM_POLL_CONTROL, 1); 1156 } 1157 1158 void arch_haltpoll_enable(unsigned int cpu) 1159 { 1160 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) { 1161 pr_err_once("host does not support poll control\n"); 1162 pr_err_once("host upgrade recommended\n"); 1163 return; 1164 } 1165 1166 /* Enable guest halt poll disables host halt poll */ 1167 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1); 1168 } 1169 EXPORT_SYMBOL_GPL(arch_haltpoll_enable); 1170 1171 void arch_haltpoll_disable(unsigned int cpu) 1172 { 1173 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) 1174 return; 1175 1176 /* Disable guest halt poll enables host halt poll */ 1177 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1); 1178 } 1179 EXPORT_SYMBOL_GPL(arch_haltpoll_disable); 1180 #endif 1181