kvm.c (28586B)
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 <asm/timer.h> 33#include <asm/cpu.h> 34#include <asm/traps.h> 35#include <asm/desc.h> 36#include <asm/tlbflush.h> 37#include <asm/apic.h> 38#include <asm/apicdef.h> 39#include <asm/hypervisor.h> 40#include <asm/tlb.h> 41#include <asm/cpuidle_haltpoll.h> 42#include <asm/ptrace.h> 43#include <asm/reboot.h> 44#include <asm/svm.h> 45#include <asm/e820/api.h> 46 47DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled); 48 49static int kvmapf = 1; 50 51static int __init parse_no_kvmapf(char *arg) 52{ 53 kvmapf = 0; 54 return 0; 55} 56 57early_param("no-kvmapf", parse_no_kvmapf); 58 59static int steal_acc = 1; 60static int __init parse_no_stealacc(char *arg) 61{ 62 steal_acc = 0; 63 return 0; 64} 65 66early_param("no-steal-acc", parse_no_stealacc); 67 68static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64); 69DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible; 70static int has_steal_clock = 0; 71 72static int has_guest_poll = 0; 73/* 74 * No need for any "IO delay" on KVM 75 */ 76static void kvm_io_delay(void) 77{ 78} 79 80#define KVM_TASK_SLEEP_HASHBITS 8 81#define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS) 82 83struct kvm_task_sleep_node { 84 struct hlist_node link; 85 struct swait_queue_head wq; 86 u32 token; 87 int cpu; 88}; 89 90static struct kvm_task_sleep_head { 91 raw_spinlock_t lock; 92 struct hlist_head list; 93} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE]; 94 95static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b, 96 u32 token) 97{ 98 struct hlist_node *p; 99 100 hlist_for_each(p, &b->list) { 101 struct kvm_task_sleep_node *n = 102 hlist_entry(p, typeof(*n), link); 103 if (n->token == token) 104 return n; 105 } 106 107 return NULL; 108} 109 110static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n) 111{ 112 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS); 113 struct kvm_task_sleep_head *b = &async_pf_sleepers[key]; 114 struct kvm_task_sleep_node *e; 115 116 raw_spin_lock(&b->lock); 117 e = _find_apf_task(b, token); 118 if (e) { 119 /* dummy entry exist -> wake up was delivered ahead of PF */ 120 hlist_del(&e->link); 121 raw_spin_unlock(&b->lock); 122 kfree(e); 123 return false; 124 } 125 126 n->token = token; 127 n->cpu = smp_processor_id(); 128 init_swait_queue_head(&n->wq); 129 hlist_add_head(&n->link, &b->list); 130 raw_spin_unlock(&b->lock); 131 return true; 132} 133 134/* 135 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled 136 * @token: Token to identify the sleep node entry 137 * 138 * Invoked from the async pagefault handling code or from the VM exit page 139 * fault handler. In both cases RCU is watching. 140 */ 141void kvm_async_pf_task_wait_schedule(u32 token) 142{ 143 struct kvm_task_sleep_node n; 144 DECLARE_SWAITQUEUE(wait); 145 146 lockdep_assert_irqs_disabled(); 147 148 if (!kvm_async_pf_queue_task(token, &n)) 149 return; 150 151 for (;;) { 152 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE); 153 if (hlist_unhashed(&n.link)) 154 break; 155 156 local_irq_enable(); 157 schedule(); 158 local_irq_disable(); 159 } 160 finish_swait(&n.wq, &wait); 161} 162EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule); 163 164static void apf_task_wake_one(struct kvm_task_sleep_node *n) 165{ 166 hlist_del_init(&n->link); 167 if (swq_has_sleeper(&n->wq)) 168 swake_up_one(&n->wq); 169} 170 171static void apf_task_wake_all(void) 172{ 173 int i; 174 175 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) { 176 struct kvm_task_sleep_head *b = &async_pf_sleepers[i]; 177 struct kvm_task_sleep_node *n; 178 struct hlist_node *p, *next; 179 180 raw_spin_lock(&b->lock); 181 hlist_for_each_safe(p, next, &b->list) { 182 n = hlist_entry(p, typeof(*n), link); 183 if (n->cpu == smp_processor_id()) 184 apf_task_wake_one(n); 185 } 186 raw_spin_unlock(&b->lock); 187 } 188} 189 190void kvm_async_pf_task_wake(u32 token) 191{ 192 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS); 193 struct kvm_task_sleep_head *b = &async_pf_sleepers[key]; 194 struct kvm_task_sleep_node *n, *dummy = NULL; 195 196 if (token == ~0) { 197 apf_task_wake_all(); 198 return; 199 } 200 201again: 202 raw_spin_lock(&b->lock); 203 n = _find_apf_task(b, token); 204 if (!n) { 205 /* 206 * Async #PF not yet handled, add a dummy entry for the token. 207 * Allocating the token must be down outside of the raw lock 208 * as the allocator is preemptible on PREEMPT_RT kernels. 209 */ 210 if (!dummy) { 211 raw_spin_unlock(&b->lock); 212 dummy = kzalloc(sizeof(*dummy), GFP_ATOMIC); 213 214 /* 215 * Continue looping on allocation failure, eventually 216 * the async #PF will be handled and allocating a new 217 * node will be unnecessary. 218 */ 219 if (!dummy) 220 cpu_relax(); 221 222 /* 223 * Recheck for async #PF completion before enqueueing 224 * the dummy token to avoid duplicate list entries. 225 */ 226 goto again; 227 } 228 dummy->token = token; 229 dummy->cpu = smp_processor_id(); 230 init_swait_queue_head(&dummy->wq); 231 hlist_add_head(&dummy->link, &b->list); 232 dummy = NULL; 233 } else { 234 apf_task_wake_one(n); 235 } 236 raw_spin_unlock(&b->lock); 237 238 /* A dummy token might be allocated and ultimately not used. */ 239 if (dummy) 240 kfree(dummy); 241} 242EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake); 243 244noinstr u32 kvm_read_and_reset_apf_flags(void) 245{ 246 u32 flags = 0; 247 248 if (__this_cpu_read(apf_reason.enabled)) { 249 flags = __this_cpu_read(apf_reason.flags); 250 __this_cpu_write(apf_reason.flags, 0); 251 } 252 253 return flags; 254} 255EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags); 256 257noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token) 258{ 259 u32 flags = kvm_read_and_reset_apf_flags(); 260 irqentry_state_t state; 261 262 if (!flags) 263 return false; 264 265 state = irqentry_enter(regs); 266 instrumentation_begin(); 267 268 /* 269 * If the host managed to inject an async #PF into an interrupt 270 * disabled region, then die hard as this is not going to end well 271 * and the host side is seriously broken. 272 */ 273 if (unlikely(!(regs->flags & X86_EFLAGS_IF))) 274 panic("Host injected async #PF in interrupt disabled region\n"); 275 276 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) { 277 if (unlikely(!(user_mode(regs)))) 278 panic("Host injected async #PF in kernel mode\n"); 279 /* Page is swapped out by the host. */ 280 kvm_async_pf_task_wait_schedule(token); 281 } else { 282 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags); 283 } 284 285 instrumentation_end(); 286 irqentry_exit(regs, state); 287 return true; 288} 289 290DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt) 291{ 292 struct pt_regs *old_regs = set_irq_regs(regs); 293 u32 token; 294 295 ack_APIC_irq(); 296 297 inc_irq_stat(irq_hv_callback_count); 298 299 if (__this_cpu_read(apf_reason.enabled)) { 300 token = __this_cpu_read(apf_reason.token); 301 kvm_async_pf_task_wake(token); 302 __this_cpu_write(apf_reason.token, 0); 303 wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1); 304 } 305 306 set_irq_regs(old_regs); 307} 308 309static void __init paravirt_ops_setup(void) 310{ 311 pv_info.name = "KVM"; 312 313 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY)) 314 pv_ops.cpu.io_delay = kvm_io_delay; 315 316#ifdef CONFIG_X86_IO_APIC 317 no_timer_check = 1; 318#endif 319} 320 321static void kvm_register_steal_time(void) 322{ 323 int cpu = smp_processor_id(); 324 struct kvm_steal_time *st = &per_cpu(steal_time, cpu); 325 326 if (!has_steal_clock) 327 return; 328 329 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED)); 330 pr_debug("stealtime: cpu %d, msr %llx\n", cpu, 331 (unsigned long long) slow_virt_to_phys(st)); 332} 333 334static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED; 335 336static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val) 337{ 338 /** 339 * This relies on __test_and_clear_bit to modify the memory 340 * in a way that is atomic with respect to the local CPU. 341 * The hypervisor only accesses this memory from the local CPU so 342 * there's no need for lock or memory barriers. 343 * An optimization barrier is implied in apic write. 344 */ 345 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi))) 346 return; 347 apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK); 348} 349 350static void kvm_guest_cpu_init(void) 351{ 352 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) { 353 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason)); 354 355 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled)); 356 357 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason)); 358 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT; 359 360 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT)) 361 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT; 362 363 wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR); 364 365 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa); 366 __this_cpu_write(apf_reason.enabled, 1); 367 pr_debug("setup async PF for cpu %d\n", smp_processor_id()); 368 } 369 370 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) { 371 unsigned long pa; 372 373 /* Size alignment is implied but just to make it explicit. */ 374 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4); 375 __this_cpu_write(kvm_apic_eoi, 0); 376 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi)) 377 | KVM_MSR_ENABLED; 378 wrmsrl(MSR_KVM_PV_EOI_EN, pa); 379 } 380 381 if (has_steal_clock) 382 kvm_register_steal_time(); 383} 384 385static void kvm_pv_disable_apf(void) 386{ 387 if (!__this_cpu_read(apf_reason.enabled)) 388 return; 389 390 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0); 391 __this_cpu_write(apf_reason.enabled, 0); 392 393 pr_debug("disable async PF for cpu %d\n", smp_processor_id()); 394} 395 396static void kvm_disable_steal_time(void) 397{ 398 if (!has_steal_clock) 399 return; 400 401 wrmsr(MSR_KVM_STEAL_TIME, 0, 0); 402} 403 404static u64 kvm_steal_clock(int cpu) 405{ 406 u64 steal; 407 struct kvm_steal_time *src; 408 int version; 409 410 src = &per_cpu(steal_time, cpu); 411 do { 412 version = src->version; 413 virt_rmb(); 414 steal = src->steal; 415 virt_rmb(); 416 } while ((version & 1) || (version != src->version)); 417 418 return steal; 419} 420 421static inline void __set_percpu_decrypted(void *ptr, unsigned long size) 422{ 423 early_set_memory_decrypted((unsigned long) ptr, size); 424} 425 426/* 427 * Iterate through all possible CPUs and map the memory region pointed 428 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once. 429 * 430 * Note: we iterate through all possible CPUs to ensure that CPUs 431 * hotplugged will have their per-cpu variable already mapped as 432 * decrypted. 433 */ 434static void __init sev_map_percpu_data(void) 435{ 436 int cpu; 437 438 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) 439 return; 440 441 for_each_possible_cpu(cpu) { 442 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason)); 443 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time)); 444 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi)); 445 } 446} 447 448static void kvm_guest_cpu_offline(bool shutdown) 449{ 450 kvm_disable_steal_time(); 451 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) 452 wrmsrl(MSR_KVM_PV_EOI_EN, 0); 453 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) 454 wrmsrl(MSR_KVM_MIGRATION_CONTROL, 0); 455 kvm_pv_disable_apf(); 456 if (!shutdown) 457 apf_task_wake_all(); 458 kvmclock_disable(); 459} 460 461static int kvm_cpu_online(unsigned int cpu) 462{ 463 unsigned long flags; 464 465 local_irq_save(flags); 466 kvm_guest_cpu_init(); 467 local_irq_restore(flags); 468 return 0; 469} 470 471#ifdef CONFIG_SMP 472 473static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask); 474 475static bool pv_tlb_flush_supported(void) 476{ 477 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) && 478 !kvm_para_has_hint(KVM_HINTS_REALTIME) && 479 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) && 480 !boot_cpu_has(X86_FEATURE_MWAIT) && 481 (num_possible_cpus() != 1)); 482} 483 484static bool pv_ipi_supported(void) 485{ 486 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) && 487 (num_possible_cpus() != 1)); 488} 489 490static bool pv_sched_yield_supported(void) 491{ 492 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) && 493 !kvm_para_has_hint(KVM_HINTS_REALTIME) && 494 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) && 495 !boot_cpu_has(X86_FEATURE_MWAIT) && 496 (num_possible_cpus() != 1)); 497} 498 499#define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG) 500 501static void __send_ipi_mask(const struct cpumask *mask, int vector) 502{ 503 unsigned long flags; 504 int cpu, apic_id, icr; 505 int min = 0, max = 0; 506#ifdef CONFIG_X86_64 507 __uint128_t ipi_bitmap = 0; 508#else 509 u64 ipi_bitmap = 0; 510#endif 511 long ret; 512 513 if (cpumask_empty(mask)) 514 return; 515 516 local_irq_save(flags); 517 518 switch (vector) { 519 default: 520 icr = APIC_DM_FIXED | vector; 521 break; 522 case NMI_VECTOR: 523 icr = APIC_DM_NMI; 524 break; 525 } 526 527 for_each_cpu(cpu, mask) { 528 apic_id = per_cpu(x86_cpu_to_apicid, cpu); 529 if (!ipi_bitmap) { 530 min = max = apic_id; 531 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) { 532 ipi_bitmap <<= min - apic_id; 533 min = apic_id; 534 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) { 535 max = apic_id < max ? max : apic_id; 536 } else { 537 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap, 538 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr); 539 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld", 540 ret); 541 min = max = apic_id; 542 ipi_bitmap = 0; 543 } 544 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap); 545 } 546 547 if (ipi_bitmap) { 548 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap, 549 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr); 550 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld", 551 ret); 552 } 553 554 local_irq_restore(flags); 555} 556 557static void kvm_send_ipi_mask(const struct cpumask *mask, int vector) 558{ 559 __send_ipi_mask(mask, vector); 560} 561 562static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector) 563{ 564 unsigned int this_cpu = smp_processor_id(); 565 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask); 566 const struct cpumask *local_mask; 567 568 cpumask_copy(new_mask, mask); 569 cpumask_clear_cpu(this_cpu, new_mask); 570 local_mask = new_mask; 571 __send_ipi_mask(local_mask, vector); 572} 573 574static int __init setup_efi_kvm_sev_migration(void) 575{ 576 efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled"; 577 efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID; 578 efi_status_t status; 579 unsigned long size; 580 bool enabled; 581 582 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) || 583 !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) 584 return 0; 585 586 if (!efi_enabled(EFI_BOOT)) 587 return 0; 588 589 if (!efi_enabled(EFI_RUNTIME_SERVICES)) { 590 pr_info("%s : EFI runtime services are not enabled\n", __func__); 591 return 0; 592 } 593 594 size = sizeof(enabled); 595 596 /* Get variable contents into buffer */ 597 status = efi.get_variable(efi_sev_live_migration_enabled, 598 &efi_variable_guid, NULL, &size, &enabled); 599 600 if (status == EFI_NOT_FOUND) { 601 pr_info("%s : EFI live migration variable not found\n", __func__); 602 return 0; 603 } 604 605 if (status != EFI_SUCCESS) { 606 pr_info("%s : EFI variable retrieval failed\n", __func__); 607 return 0; 608 } 609 610 if (enabled == 0) { 611 pr_info("%s: live migration disabled in EFI\n", __func__); 612 return 0; 613 } 614 615 pr_info("%s : live migration enabled in EFI\n", __func__); 616 wrmsrl(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY); 617 618 return 1; 619} 620 621late_initcall(setup_efi_kvm_sev_migration); 622 623/* 624 * Set the IPI entry points 625 */ 626static void kvm_setup_pv_ipi(void) 627{ 628 apic->send_IPI_mask = kvm_send_ipi_mask; 629 apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself; 630 pr_info("setup PV IPIs\n"); 631} 632 633static void kvm_smp_send_call_func_ipi(const struct cpumask *mask) 634{ 635 int cpu; 636 637 native_send_call_func_ipi(mask); 638 639 /* Make sure other vCPUs get a chance to run if they need to. */ 640 for_each_cpu(cpu, mask) { 641 if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) { 642 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu)); 643 break; 644 } 645 } 646} 647 648static void kvm_flush_tlb_multi(const struct cpumask *cpumask, 649 const struct flush_tlb_info *info) 650{ 651 u8 state; 652 int cpu; 653 struct kvm_steal_time *src; 654 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask); 655 656 cpumask_copy(flushmask, cpumask); 657 /* 658 * We have to call flush only on online vCPUs. And 659 * queue flush_on_enter for pre-empted vCPUs 660 */ 661 for_each_cpu(cpu, flushmask) { 662 /* 663 * The local vCPU is never preempted, so we do not explicitly 664 * skip check for local vCPU - it will never be cleared from 665 * flushmask. 666 */ 667 src = &per_cpu(steal_time, cpu); 668 state = READ_ONCE(src->preempted); 669 if ((state & KVM_VCPU_PREEMPTED)) { 670 if (try_cmpxchg(&src->preempted, &state, 671 state | KVM_VCPU_FLUSH_TLB)) 672 __cpumask_clear_cpu(cpu, flushmask); 673 } 674 } 675 676 native_flush_tlb_multi(flushmask, info); 677} 678 679static __init int kvm_alloc_cpumask(void) 680{ 681 int cpu; 682 683 if (!kvm_para_available() || nopv) 684 return 0; 685 686 if (pv_tlb_flush_supported() || pv_ipi_supported()) 687 for_each_possible_cpu(cpu) { 688 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu), 689 GFP_KERNEL, cpu_to_node(cpu)); 690 } 691 692 return 0; 693} 694arch_initcall(kvm_alloc_cpumask); 695 696static void __init kvm_smp_prepare_boot_cpu(void) 697{ 698 /* 699 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init() 700 * shares the guest physical address with the hypervisor. 701 */ 702 sev_map_percpu_data(); 703 704 kvm_guest_cpu_init(); 705 native_smp_prepare_boot_cpu(); 706 kvm_spinlock_init(); 707} 708 709static int kvm_cpu_down_prepare(unsigned int cpu) 710{ 711 unsigned long flags; 712 713 local_irq_save(flags); 714 kvm_guest_cpu_offline(false); 715 local_irq_restore(flags); 716 return 0; 717} 718 719#endif 720 721static int kvm_suspend(void) 722{ 723 u64 val = 0; 724 725 kvm_guest_cpu_offline(false); 726 727#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 728 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) 729 rdmsrl(MSR_KVM_POLL_CONTROL, val); 730 has_guest_poll = !(val & 1); 731#endif 732 return 0; 733} 734 735static void kvm_resume(void) 736{ 737 kvm_cpu_online(raw_smp_processor_id()); 738 739#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 740 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll) 741 wrmsrl(MSR_KVM_POLL_CONTROL, 0); 742#endif 743} 744 745static struct syscore_ops kvm_syscore_ops = { 746 .suspend = kvm_suspend, 747 .resume = kvm_resume, 748}; 749 750static void kvm_pv_guest_cpu_reboot(void *unused) 751{ 752 kvm_guest_cpu_offline(true); 753} 754 755static int kvm_pv_reboot_notify(struct notifier_block *nb, 756 unsigned long code, void *unused) 757{ 758 if (code == SYS_RESTART) 759 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1); 760 return NOTIFY_DONE; 761} 762 763static struct notifier_block kvm_pv_reboot_nb = { 764 .notifier_call = kvm_pv_reboot_notify, 765}; 766 767/* 768 * After a PV feature is registered, the host will keep writing to the 769 * registered memory location. If the guest happens to shutdown, this memory 770 * won't be valid. In cases like kexec, in which you install a new kernel, this 771 * means a random memory location will be kept being written. 772 */ 773#ifdef CONFIG_KEXEC_CORE 774static void kvm_crash_shutdown(struct pt_regs *regs) 775{ 776 kvm_guest_cpu_offline(true); 777 native_machine_crash_shutdown(regs); 778} 779#endif 780 781#if defined(CONFIG_X86_32) || !defined(CONFIG_SMP) 782bool __kvm_vcpu_is_preempted(long cpu); 783 784__visible bool __kvm_vcpu_is_preempted(long cpu) 785{ 786 struct kvm_steal_time *src = &per_cpu(steal_time, cpu); 787 788 return !!(src->preempted & KVM_VCPU_PREEMPTED); 789} 790PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted); 791 792#else 793 794#include <asm/asm-offsets.h> 795 796extern bool __raw_callee_save___kvm_vcpu_is_preempted(long); 797 798/* 799 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and 800 * restoring to/from the stack. 801 */ 802asm( 803".pushsection .text;" 804".global __raw_callee_save___kvm_vcpu_is_preempted;" 805".type __raw_callee_save___kvm_vcpu_is_preempted, @function;" 806"__raw_callee_save___kvm_vcpu_is_preempted:" 807ASM_ENDBR 808"movq __per_cpu_offset(,%rdi,8), %rax;" 809"cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);" 810"setne %al;" 811ASM_RET 812".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;" 813".popsection"); 814 815#endif 816 817static void __init kvm_guest_init(void) 818{ 819 int i; 820 821 paravirt_ops_setup(); 822 register_reboot_notifier(&kvm_pv_reboot_nb); 823 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) 824 raw_spin_lock_init(&async_pf_sleepers[i].lock); 825 826 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) { 827 has_steal_clock = 1; 828 static_call_update(pv_steal_clock, kvm_steal_clock); 829 830 pv_ops.lock.vcpu_is_preempted = 831 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted); 832 } 833 834 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) 835 apic_set_eoi_write(kvm_guest_apic_eoi_write); 836 837 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) { 838 static_branch_enable(&kvm_async_pf_enabled); 839 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt); 840 } 841 842#ifdef CONFIG_SMP 843 if (pv_tlb_flush_supported()) { 844 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi; 845 pv_ops.mmu.tlb_remove_table = tlb_remove_table; 846 pr_info("KVM setup pv remote TLB flush\n"); 847 } 848 849 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu; 850 if (pv_sched_yield_supported()) { 851 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi; 852 pr_info("setup PV sched yield\n"); 853 } 854 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online", 855 kvm_cpu_online, kvm_cpu_down_prepare) < 0) 856 pr_err("failed to install cpu hotplug callbacks\n"); 857#else 858 sev_map_percpu_data(); 859 kvm_guest_cpu_init(); 860#endif 861 862#ifdef CONFIG_KEXEC_CORE 863 machine_ops.crash_shutdown = kvm_crash_shutdown; 864#endif 865 866 register_syscore_ops(&kvm_syscore_ops); 867 868 /* 869 * Hard lockup detection is enabled by default. Disable it, as guests 870 * can get false positives too easily, for example if the host is 871 * overcommitted. 872 */ 873 hardlockup_detector_disable(); 874} 875 876static noinline uint32_t __kvm_cpuid_base(void) 877{ 878 if (boot_cpu_data.cpuid_level < 0) 879 return 0; /* So we don't blow up on old processors */ 880 881 if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) 882 return hypervisor_cpuid_base(KVM_SIGNATURE, 0); 883 884 return 0; 885} 886 887static inline uint32_t kvm_cpuid_base(void) 888{ 889 static int kvm_cpuid_base = -1; 890 891 if (kvm_cpuid_base == -1) 892 kvm_cpuid_base = __kvm_cpuid_base(); 893 894 return kvm_cpuid_base; 895} 896 897bool kvm_para_available(void) 898{ 899 return kvm_cpuid_base() != 0; 900} 901EXPORT_SYMBOL_GPL(kvm_para_available); 902 903unsigned int kvm_arch_para_features(void) 904{ 905 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES); 906} 907 908unsigned int kvm_arch_para_hints(void) 909{ 910 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES); 911} 912EXPORT_SYMBOL_GPL(kvm_arch_para_hints); 913 914static uint32_t __init kvm_detect(void) 915{ 916 return kvm_cpuid_base(); 917} 918 919static void __init kvm_apic_init(void) 920{ 921#ifdef CONFIG_SMP 922 if (pv_ipi_supported()) 923 kvm_setup_pv_ipi(); 924#endif 925} 926 927static bool __init kvm_msi_ext_dest_id(void) 928{ 929 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID); 930} 931 932static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc) 933{ 934 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages, 935 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K); 936} 937 938static void __init kvm_init_platform(void) 939{ 940 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) && 941 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) { 942 unsigned long nr_pages; 943 int i; 944 945 pv_ops.mmu.notify_page_enc_status_changed = 946 kvm_sev_hc_page_enc_status; 947 948 /* 949 * Reset the host's shared pages list related to kernel 950 * specific page encryption status settings before we load a 951 * new kernel by kexec. Reset the page encryption status 952 * during early boot intead of just before kexec to avoid SMP 953 * races during kvm_pv_guest_cpu_reboot(). 954 * NOTE: We cannot reset the complete shared pages list 955 * here as we need to retain the UEFI/OVMF firmware 956 * specific settings. 957 */ 958 959 for (i = 0; i < e820_table->nr_entries; i++) { 960 struct e820_entry *entry = &e820_table->entries[i]; 961 962 if (entry->type != E820_TYPE_RAM) 963 continue; 964 965 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE); 966 967 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr, 968 nr_pages, 969 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K); 970 } 971 972 /* 973 * Ensure that _bss_decrypted section is marked as decrypted in the 974 * shared pages list. 975 */ 976 nr_pages = DIV_ROUND_UP(__end_bss_decrypted - __start_bss_decrypted, 977 PAGE_SIZE); 978 early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted, 979 nr_pages, 0); 980 981 /* 982 * If not booted using EFI, enable Live migration support. 983 */ 984 if (!efi_enabled(EFI_BOOT)) 985 wrmsrl(MSR_KVM_MIGRATION_CONTROL, 986 KVM_MIGRATION_READY); 987 } 988 kvmclock_init(); 989 x86_platform.apic_post_init = kvm_apic_init; 990} 991 992#if defined(CONFIG_AMD_MEM_ENCRYPT) 993static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs) 994{ 995 /* RAX and CPL are already in the GHCB */ 996 ghcb_set_rbx(ghcb, regs->bx); 997 ghcb_set_rcx(ghcb, regs->cx); 998 ghcb_set_rdx(ghcb, regs->dx); 999 ghcb_set_rsi(ghcb, regs->si); 1000} 1001 1002static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs) 1003{ 1004 /* No checking of the return state needed */ 1005 return true; 1006} 1007#endif 1008 1009const __initconst struct hypervisor_x86 x86_hyper_kvm = { 1010 .name = "KVM", 1011 .detect = kvm_detect, 1012 .type = X86_HYPER_KVM, 1013 .init.guest_late_init = kvm_guest_init, 1014 .init.x2apic_available = kvm_para_available, 1015 .init.msi_ext_dest_id = kvm_msi_ext_dest_id, 1016 .init.init_platform = kvm_init_platform, 1017#if defined(CONFIG_AMD_MEM_ENCRYPT) 1018 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare, 1019 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish, 1020#endif 1021}; 1022 1023static __init int activate_jump_labels(void) 1024{ 1025 if (has_steal_clock) { 1026 static_key_slow_inc(¶virt_steal_enabled); 1027 if (steal_acc) 1028 static_key_slow_inc(¶virt_steal_rq_enabled); 1029 } 1030 1031 return 0; 1032} 1033arch_initcall(activate_jump_labels); 1034 1035#ifdef CONFIG_PARAVIRT_SPINLOCKS 1036 1037/* Kick a cpu by its apicid. Used to wake up a halted vcpu */ 1038static void kvm_kick_cpu(int cpu) 1039{ 1040 int apicid; 1041 unsigned long flags = 0; 1042 1043 apicid = per_cpu(x86_cpu_to_apicid, cpu); 1044 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid); 1045} 1046 1047#include <asm/qspinlock.h> 1048 1049static void kvm_wait(u8 *ptr, u8 val) 1050{ 1051 if (in_nmi()) 1052 return; 1053 1054 /* 1055 * halt until it's our turn and kicked. Note that we do safe halt 1056 * for irq enabled case to avoid hang when lock info is overwritten 1057 * in irq spinlock slowpath and no spurious interrupt occur to save us. 1058 */ 1059 if (irqs_disabled()) { 1060 if (READ_ONCE(*ptr) == val) 1061 halt(); 1062 } else { 1063 local_irq_disable(); 1064 1065 /* safe_halt() will enable IRQ */ 1066 if (READ_ONCE(*ptr) == val) 1067 safe_halt(); 1068 else 1069 local_irq_enable(); 1070 } 1071} 1072 1073/* 1074 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present. 1075 */ 1076void __init kvm_spinlock_init(void) 1077{ 1078 /* 1079 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an 1080 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is 1081 * preferred over native qspinlock when vCPU is preempted. 1082 */ 1083 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) { 1084 pr_info("PV spinlocks disabled, no host support\n"); 1085 return; 1086 } 1087 1088 /* 1089 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs 1090 * are available. 1091 */ 1092 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) { 1093 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n"); 1094 goto out; 1095 } 1096 1097 if (num_possible_cpus() == 1) { 1098 pr_info("PV spinlocks disabled, single CPU\n"); 1099 goto out; 1100 } 1101 1102 if (nopvspin) { 1103 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n"); 1104 goto out; 1105 } 1106 1107 pr_info("PV spinlocks enabled\n"); 1108 1109 __pv_init_lock_hash(); 1110 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath; 1111 pv_ops.lock.queued_spin_unlock = 1112 PV_CALLEE_SAVE(__pv_queued_spin_unlock); 1113 pv_ops.lock.wait = kvm_wait; 1114 pv_ops.lock.kick = kvm_kick_cpu; 1115 1116 /* 1117 * When PV spinlock is enabled which is preferred over 1118 * virt_spin_lock(), virt_spin_lock_key's value is meaningless. 1119 * Just disable it anyway. 1120 */ 1121out: 1122 static_branch_disable(&virt_spin_lock_key); 1123} 1124 1125#endif /* CONFIG_PARAVIRT_SPINLOCKS */ 1126 1127#ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 1128 1129static void kvm_disable_host_haltpoll(void *i) 1130{ 1131 wrmsrl(MSR_KVM_POLL_CONTROL, 0); 1132} 1133 1134static void kvm_enable_host_haltpoll(void *i) 1135{ 1136 wrmsrl(MSR_KVM_POLL_CONTROL, 1); 1137} 1138 1139void arch_haltpoll_enable(unsigned int cpu) 1140{ 1141 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) { 1142 pr_err_once("host does not support poll control\n"); 1143 pr_err_once("host upgrade recommended\n"); 1144 return; 1145 } 1146 1147 /* Enable guest halt poll disables host halt poll */ 1148 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1); 1149} 1150EXPORT_SYMBOL_GPL(arch_haltpoll_enable); 1151 1152void arch_haltpoll_disable(unsigned int cpu) 1153{ 1154 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) 1155 return; 1156 1157 /* Disable guest halt poll enables host halt poll */ 1158 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1); 1159} 1160EXPORT_SYMBOL_GPL(arch_haltpoll_disable); 1161#endif