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/cpuid/api.h> 45 #include <asm/cpuidle_haltpoll.h> 46 #include <asm/msr.h> 47 #include <asm/ptrace.h> 48 #include <asm/reboot.h> 49 #include <asm/svm.h> 50 #include <asm/e820/api.h> 51 52 DEFINE_STATIC_KEY_FALSE_RO(kvm_async_pf_enabled); 53 54 static int kvmapf = 1; 55 56 static int __init parse_no_kvmapf(char *arg) 57 { 58 kvmapf = 0; 59 return 0; 60 } 61 62 early_param("no-kvmapf", parse_no_kvmapf); 63 64 static int steal_acc = 1; 65 static int __init parse_no_stealacc(char *arg) 66 { 67 steal_acc = 0; 68 return 0; 69 } 70 71 early_param("no-steal-acc", parse_no_stealacc); 72 73 static DEFINE_PER_CPU_READ_MOSTLY(bool, async_pf_enabled); 74 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64); 75 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible; 76 static int has_steal_clock = 0; 77 78 static int has_guest_poll = 0; 79 80 #define KVM_TASK_SLEEP_HASHBITS 8 81 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS) 82 83 struct kvm_task_sleep_node { 84 struct hlist_node link; 85 struct swait_queue_head wq; 86 u32 token; 87 int cpu; 88 bool dummy; 89 }; 90 91 static struct kvm_task_sleep_head { 92 raw_spinlock_t lock; 93 struct hlist_head list; 94 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE]; 95 96 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b, 97 u32 token) 98 { 99 struct hlist_node *p; 100 101 hlist_for_each(p, &b->list) { 102 struct kvm_task_sleep_node *n = 103 hlist_entry(p, typeof(*n), link); 104 if (n->token == token) 105 return n; 106 } 107 108 return NULL; 109 } 110 111 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n) 112 { 113 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS); 114 struct kvm_task_sleep_head *b = &async_pf_sleepers[key]; 115 struct kvm_task_sleep_node *e; 116 117 raw_spin_lock(&b->lock); 118 e = _find_apf_task(b, token); 119 if (e) { 120 struct kvm_task_sleep_node *dummy = NULL; 121 122 /* 123 * The entry can either be a 'dummy' entry (which is put on the 124 * list when wake-up happens ahead of APF handling completion) 125 * or a token from another task which should not be touched. 126 */ 127 if (e->dummy) { 128 hlist_del(&e->link); 129 dummy = e; 130 } 131 132 raw_spin_unlock(&b->lock); 133 kfree(dummy); 134 return false; 135 } 136 137 n->token = token; 138 n->cpu = smp_processor_id(); 139 n->dummy = false; 140 init_swait_queue_head(&n->wq); 141 hlist_add_head(&n->link, &b->list); 142 raw_spin_unlock(&b->lock); 143 return true; 144 } 145 146 /* 147 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled 148 * @token: Token to identify the sleep node entry 149 * 150 * Invoked from the async pagefault handling code or from the VM exit page 151 * fault handler. In both cases RCU is watching. 152 */ 153 void kvm_async_pf_task_wait_schedule(u32 token) 154 { 155 struct kvm_task_sleep_node n; 156 DECLARE_SWAITQUEUE(wait); 157 158 lockdep_assert_irqs_disabled(); 159 160 if (!kvm_async_pf_queue_task(token, &n)) 161 return; 162 163 for (;;) { 164 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE); 165 if (hlist_unhashed(&n.link)) 166 break; 167 168 local_irq_enable(); 169 schedule(); 170 local_irq_disable(); 171 } 172 finish_swait(&n.wq, &wait); 173 } 174 EXPORT_SYMBOL_FOR_KVM(kvm_async_pf_task_wait_schedule); 175 176 static void apf_task_wake_one(struct kvm_task_sleep_node *n) 177 { 178 hlist_del_init(&n->link); 179 if (swq_has_sleeper(&n->wq)) 180 swake_up_one(&n->wq); 181 } 182 183 static void apf_task_wake_all(void) 184 { 185 int i; 186 187 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) { 188 struct kvm_task_sleep_head *b = &async_pf_sleepers[i]; 189 struct kvm_task_sleep_node *n; 190 struct hlist_node *p, *next; 191 192 raw_spin_lock(&b->lock); 193 hlist_for_each_safe(p, next, &b->list) { 194 n = hlist_entry(p, typeof(*n), link); 195 if (n->cpu == smp_processor_id()) 196 apf_task_wake_one(n); 197 } 198 raw_spin_unlock(&b->lock); 199 } 200 } 201 202 static void kvm_async_pf_task_wake(u32 token) 203 { 204 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS); 205 struct kvm_task_sleep_head *b = &async_pf_sleepers[key]; 206 struct kvm_task_sleep_node *n, *dummy = NULL; 207 208 if (token == ~0) { 209 apf_task_wake_all(); 210 return; 211 } 212 213 again: 214 raw_spin_lock(&b->lock); 215 n = _find_apf_task(b, token); 216 if (!n) { 217 /* 218 * Async #PF not yet handled, add a dummy entry for the token. 219 * Allocating the token must be down outside of the raw lock 220 * as the allocator is preemptible on PREEMPT_RT kernels. 221 */ 222 if (!dummy) { 223 raw_spin_unlock(&b->lock); 224 dummy = kzalloc_obj(*dummy, GFP_ATOMIC); 225 226 /* 227 * Continue looping on allocation failure, eventually 228 * the async #PF will be handled and allocating a new 229 * node will be unnecessary. 230 */ 231 if (!dummy) 232 cpu_relax(); 233 234 /* 235 * Recheck for async #PF completion before enqueueing 236 * the dummy token to avoid duplicate list entries. 237 */ 238 goto again; 239 } 240 dummy->token = token; 241 dummy->cpu = smp_processor_id(); 242 dummy->dummy = true; 243 init_swait_queue_head(&dummy->wq); 244 hlist_add_head(&dummy->link, &b->list); 245 dummy = NULL; 246 } else { 247 apf_task_wake_one(n); 248 } 249 raw_spin_unlock(&b->lock); 250 251 /* A dummy token might be allocated and ultimately not used. */ 252 kfree(dummy); 253 } 254 255 noinstr u32 kvm_read_and_reset_apf_flags(void) 256 { 257 u32 flags = 0; 258 259 if (__this_cpu_read(async_pf_enabled)) { 260 flags = __this_cpu_read(apf_reason.flags); 261 __this_cpu_write(apf_reason.flags, 0); 262 } 263 264 return flags; 265 } 266 EXPORT_SYMBOL_FOR_KVM(kvm_read_and_reset_apf_flags); 267 268 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token) 269 { 270 u32 flags = kvm_read_and_reset_apf_flags(); 271 irqentry_state_t state; 272 273 if (!flags) 274 return false; 275 276 state = irqentry_enter(regs); 277 instrumentation_begin(); 278 279 /* 280 * If the host managed to inject an async #PF into an interrupt 281 * disabled region, then die hard as this is not going to end well 282 * and the host side is seriously broken. 283 */ 284 if (unlikely(!(regs->flags & X86_EFLAGS_IF))) 285 panic("Host injected async #PF in interrupt disabled region\n"); 286 287 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) { 288 if (unlikely(!(user_mode(regs)))) 289 panic("Host injected async #PF in kernel mode\n"); 290 /* Page is swapped out by the host. */ 291 kvm_async_pf_task_wait_schedule(token); 292 } else { 293 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags); 294 } 295 296 instrumentation_end(); 297 irqentry_exit(regs, state); 298 return true; 299 } 300 301 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt) 302 { 303 struct pt_regs *old_regs = set_irq_regs(regs); 304 u32 token; 305 306 apic_eoi(); 307 308 inc_irq_stat(HYPERVISOR_CALLBACK); 309 310 if (__this_cpu_read(async_pf_enabled)) { 311 token = __this_cpu_read(apf_reason.token); 312 kvm_async_pf_task_wake(token); 313 __this_cpu_write(apf_reason.token, 0); 314 wrmsrq(MSR_KVM_ASYNC_PF_ACK, 1); 315 } 316 317 set_irq_regs(old_regs); 318 } 319 320 static void __init paravirt_ops_setup(void) 321 { 322 pv_info.name = "KVM"; 323 324 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY)) 325 pv_info.io_delay = false; 326 327 #ifdef CONFIG_X86_IO_APIC 328 no_timer_check = 1; 329 #endif 330 } 331 332 static void kvm_register_steal_time(void) 333 { 334 int cpu = smp_processor_id(); 335 struct kvm_steal_time *st = &per_cpu(steal_time, cpu); 336 337 if (!has_steal_clock) 338 return; 339 340 wrmsrq(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED)); 341 pr_debug("stealtime: cpu %d, msr %llx\n", cpu, 342 (unsigned long long) slow_virt_to_phys(st)); 343 } 344 345 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED; 346 347 static notrace __maybe_unused void kvm_guest_apic_eoi_write(void) 348 { 349 /** 350 * This relies on __test_and_clear_bit to modify the memory 351 * in a way that is atomic with respect to the local CPU. 352 * The hypervisor only accesses this memory from the local CPU so 353 * there's no need for lock or memory barriers. 354 * An optimization barrier is implied in apic write. 355 */ 356 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi))) 357 return; 358 apic_native_eoi(); 359 } 360 361 static void kvm_guest_cpu_init(void) 362 { 363 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) { 364 u64 pa; 365 366 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled)); 367 368 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason)); 369 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT; 370 371 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT)) 372 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT; 373 374 wrmsrq(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR); 375 376 wrmsrq(MSR_KVM_ASYNC_PF_EN, pa); 377 __this_cpu_write(async_pf_enabled, true); 378 pr_debug("setup async PF for cpu %d\n", smp_processor_id()); 379 } 380 381 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) { 382 unsigned long pa; 383 384 /* Size alignment is implied but just to make it explicit. */ 385 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4); 386 __this_cpu_write(kvm_apic_eoi, 0); 387 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi)) 388 | KVM_MSR_ENABLED; 389 wrmsrq(MSR_KVM_PV_EOI_EN, pa); 390 } 391 392 if (has_steal_clock) 393 kvm_register_steal_time(); 394 } 395 396 static void kvm_pv_disable_apf(void) 397 { 398 if (!__this_cpu_read(async_pf_enabled)) 399 return; 400 401 wrmsrq(MSR_KVM_ASYNC_PF_EN, 0); 402 __this_cpu_write(async_pf_enabled, false); 403 404 pr_debug("disable async PF for cpu %d\n", smp_processor_id()); 405 } 406 407 static void kvm_disable_steal_time(void) 408 { 409 if (!has_steal_clock) 410 return; 411 412 wrmsrq(MSR_KVM_STEAL_TIME, 0); 413 } 414 415 static u64 kvm_steal_clock(int cpu) 416 { 417 u64 steal; 418 struct kvm_steal_time *src; 419 int version; 420 421 src = &per_cpu(steal_time, cpu); 422 do { 423 version = src->version; 424 virt_rmb(); 425 steal = src->steal; 426 virt_rmb(); 427 } while ((version & 1) || (version != src->version)); 428 429 return steal; 430 } 431 432 static inline __init void __set_percpu_decrypted(void *ptr, unsigned long size) 433 { 434 early_set_memory_decrypted((unsigned long) ptr, size); 435 } 436 437 /* 438 * Iterate through all possible CPUs and map the memory region pointed 439 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once. 440 * 441 * Note: we iterate through all possible CPUs to ensure that CPUs 442 * hotplugged will have their per-cpu variable already mapped as 443 * decrypted. 444 */ 445 static void __init sev_map_percpu_data(void) 446 { 447 int cpu; 448 449 if (cc_vendor != CC_VENDOR_AMD || 450 !cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT)) 451 return; 452 453 for_each_possible_cpu(cpu) { 454 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason)); 455 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time)); 456 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi)); 457 } 458 } 459 460 static void kvm_guest_cpu_offline(bool shutdown) 461 { 462 kvm_disable_steal_time(); 463 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) 464 wrmsrq(MSR_KVM_PV_EOI_EN, 0); 465 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) 466 wrmsrq(MSR_KVM_MIGRATION_CONTROL, 0); 467 kvm_pv_disable_apf(); 468 if (!shutdown) 469 apf_task_wake_all(); 470 kvmclock_disable(); 471 } 472 473 static int kvm_cpu_online(unsigned int cpu) 474 { 475 unsigned long flags; 476 477 local_irq_save(flags); 478 kvm_guest_cpu_init(); 479 local_irq_restore(flags); 480 return 0; 481 } 482 483 #ifdef CONFIG_SMP 484 485 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask); 486 487 static bool pv_tlb_flush_supported(void) 488 { 489 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) && 490 !kvm_para_has_hint(KVM_HINTS_REALTIME) && 491 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) && 492 !boot_cpu_has(X86_FEATURE_MWAIT) && 493 (num_possible_cpus() != 1)); 494 } 495 496 static bool pv_ipi_supported(void) 497 { 498 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) && 499 (num_possible_cpus() != 1)); 500 } 501 502 static bool pv_sched_yield_supported(void) 503 { 504 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) && 505 !kvm_para_has_hint(KVM_HINTS_REALTIME) && 506 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) && 507 !boot_cpu_has(X86_FEATURE_MWAIT) && 508 (num_possible_cpus() != 1)); 509 } 510 511 #define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG) 512 513 static void __send_ipi_mask(const struct cpumask *mask, int vector) 514 { 515 unsigned long flags; 516 int cpu, min = 0, max = 0; 517 #ifdef CONFIG_X86_64 518 __uint128_t ipi_bitmap = 0; 519 #else 520 u64 ipi_bitmap = 0; 521 #endif 522 u32 apic_id, icr; 523 long ret; 524 525 if (cpumask_empty(mask)) 526 return; 527 528 local_irq_save(flags); 529 530 switch (vector) { 531 default: 532 icr = APIC_DM_FIXED | vector; 533 break; 534 case NMI_VECTOR: 535 icr = APIC_DM_NMI; 536 break; 537 } 538 539 for_each_cpu(cpu, mask) { 540 apic_id = per_cpu(x86_cpu_to_apicid, cpu); 541 if (!ipi_bitmap) { 542 min = max = apic_id; 543 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) { 544 ipi_bitmap <<= min - apic_id; 545 min = apic_id; 546 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) { 547 max = apic_id < max ? max : apic_id; 548 } else { 549 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap, 550 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr); 551 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld", 552 ret); 553 min = max = apic_id; 554 ipi_bitmap = 0; 555 } 556 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap); 557 } 558 559 if (ipi_bitmap) { 560 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap, 561 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr); 562 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld", 563 ret); 564 } 565 566 local_irq_restore(flags); 567 } 568 569 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector) 570 { 571 __send_ipi_mask(mask, vector); 572 } 573 574 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector) 575 { 576 unsigned int this_cpu = smp_processor_id(); 577 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask); 578 const struct cpumask *local_mask; 579 580 cpumask_copy(new_mask, mask); 581 cpumask_clear_cpu(this_cpu, new_mask); 582 local_mask = new_mask; 583 __send_ipi_mask(local_mask, vector); 584 } 585 586 static int __init setup_efi_kvm_sev_migration(void) 587 { 588 efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled"; 589 efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID; 590 efi_status_t status; 591 unsigned long size; 592 bool enabled; 593 594 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) || 595 !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) 596 return 0; 597 598 if (!efi_enabled(EFI_BOOT)) 599 return 0; 600 601 if (!efi_enabled(EFI_RUNTIME_SERVICES)) { 602 pr_info("%s : EFI runtime services are not enabled\n", __func__); 603 return 0; 604 } 605 606 size = sizeof(enabled); 607 608 /* Get variable contents into buffer */ 609 status = efi.get_variable(efi_sev_live_migration_enabled, 610 &efi_variable_guid, NULL, &size, &enabled); 611 612 if (status == EFI_NOT_FOUND) { 613 pr_info("%s : EFI live migration variable not found\n", __func__); 614 return 0; 615 } 616 617 if (status != EFI_SUCCESS) { 618 pr_info("%s : EFI variable retrieval failed\n", __func__); 619 return 0; 620 } 621 622 if (enabled == 0) { 623 pr_info("%s: live migration disabled in EFI\n", __func__); 624 return 0; 625 } 626 627 pr_info("%s : live migration enabled in EFI\n", __func__); 628 wrmsrq(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY); 629 630 return 1; 631 } 632 633 late_initcall(setup_efi_kvm_sev_migration); 634 635 /* 636 * Set the IPI entry points 637 */ 638 static __init void kvm_setup_pv_ipi(void) 639 { 640 apic_update_callback(send_IPI_mask, kvm_send_ipi_mask); 641 apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself); 642 pr_info("setup PV IPIs\n"); 643 } 644 645 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask) 646 { 647 int cpu; 648 649 native_send_call_func_ipi(mask); 650 651 /* Make sure other vCPUs get a chance to run if they need to. */ 652 for_each_cpu(cpu, mask) { 653 if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) { 654 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu)); 655 break; 656 } 657 } 658 } 659 660 static void kvm_flush_tlb_multi(const struct cpumask *cpumask, 661 const struct flush_tlb_info *info) 662 { 663 u8 state; 664 int cpu; 665 struct kvm_steal_time *src; 666 struct cpumask *flushmask; 667 668 guard(preempt)(); 669 flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask); 670 cpumask_copy(flushmask, cpumask); 671 /* 672 * We have to call flush only on online vCPUs. And 673 * queue flush_on_enter for pre-empted vCPUs 674 */ 675 for_each_cpu(cpu, flushmask) { 676 /* 677 * The local vCPU is never preempted, so we do not explicitly 678 * skip check for local vCPU - it will never be cleared from 679 * flushmask. 680 */ 681 src = &per_cpu(steal_time, cpu); 682 state = READ_ONCE(src->preempted); 683 if ((state & KVM_VCPU_PREEMPTED)) { 684 if (try_cmpxchg(&src->preempted, &state, 685 state | KVM_VCPU_FLUSH_TLB)) 686 __cpumask_clear_cpu(cpu, flushmask); 687 } 688 } 689 690 native_flush_tlb_multi(flushmask, info); 691 } 692 693 static __init int kvm_alloc_cpumask(void) 694 { 695 int cpu; 696 697 if (!kvm_para_available() || nopv) 698 return 0; 699 700 if (pv_tlb_flush_supported() || pv_ipi_supported()) 701 for_each_possible_cpu(cpu) { 702 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu), 703 GFP_KERNEL, cpu_to_node(cpu)); 704 } 705 706 return 0; 707 } 708 arch_initcall(kvm_alloc_cpumask); 709 710 static void __init kvm_smp_prepare_boot_cpu(void) 711 { 712 /* 713 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init() 714 * shares the guest physical address with the hypervisor. 715 */ 716 sev_map_percpu_data(); 717 718 kvm_guest_cpu_init(); 719 native_smp_prepare_boot_cpu(); 720 kvm_spinlock_init(); 721 } 722 723 static int kvm_cpu_down_prepare(unsigned int cpu) 724 { 725 unsigned long flags; 726 727 local_irq_save(flags); 728 kvm_guest_cpu_offline(false); 729 local_irq_restore(flags); 730 return 0; 731 } 732 733 #endif 734 735 static int kvm_suspend(void *data) 736 { 737 u64 val = 0; 738 739 kvm_guest_cpu_offline(false); 740 741 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 742 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) 743 rdmsrq(MSR_KVM_POLL_CONTROL, val); 744 has_guest_poll = !(val & 1); 745 #endif 746 return 0; 747 } 748 749 static void kvm_resume(void *data) 750 { 751 kvm_cpu_online(raw_smp_processor_id()); 752 753 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 754 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll) 755 wrmsrq(MSR_KVM_POLL_CONTROL, 0); 756 #endif 757 } 758 759 static const struct syscore_ops kvm_syscore_ops = { 760 .suspend = kvm_suspend, 761 .resume = kvm_resume, 762 }; 763 764 static struct syscore kvm_syscore = { 765 .ops = &kvm_syscore_ops, 766 }; 767 768 static void kvm_pv_guest_cpu_reboot(void *unused) 769 { 770 kvm_guest_cpu_offline(true); 771 } 772 773 static int kvm_pv_reboot_notify(struct notifier_block *nb, 774 unsigned long code, void *unused) 775 { 776 if (code == SYS_RESTART) 777 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1); 778 return NOTIFY_DONE; 779 } 780 781 static struct notifier_block kvm_pv_reboot_nb = { 782 .notifier_call = kvm_pv_reboot_notify, 783 }; 784 785 /* 786 * After a PV feature is registered, the host will keep writing to the 787 * registered memory location. If the guest happens to shutdown, this memory 788 * won't be valid. In cases like kexec, in which you install a new kernel, this 789 * means a random memory location will be kept being written. 790 */ 791 #ifdef CONFIG_CRASH_DUMP 792 static void kvm_crash_shutdown(struct pt_regs *regs) 793 { 794 kvm_guest_cpu_offline(true); 795 native_machine_crash_shutdown(regs); 796 } 797 #endif 798 799 #if defined(CONFIG_X86_32) || !defined(CONFIG_SMP) 800 bool __kvm_vcpu_is_preempted(long cpu); 801 802 __visible bool __kvm_vcpu_is_preempted(long cpu) 803 { 804 struct kvm_steal_time *src = &per_cpu(steal_time, cpu); 805 806 return !!(src->preempted & KVM_VCPU_PREEMPTED); 807 } 808 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted); 809 810 #else 811 812 #include <asm/asm-offsets.h> 813 814 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long); 815 816 /* 817 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and 818 * restoring to/from the stack. 819 */ 820 #define PV_VCPU_PREEMPTED_ASM \ 821 "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \ 822 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \ 823 "setne %al\n\t" 824 825 DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted, 826 PV_VCPU_PREEMPTED_ASM, .text); 827 #endif 828 829 static void __init kvm_guest_init(void) 830 { 831 int i; 832 833 paravirt_ops_setup(); 834 register_reboot_notifier(&kvm_pv_reboot_nb); 835 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) 836 raw_spin_lock_init(&async_pf_sleepers[i].lock); 837 838 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) { 839 has_steal_clock = 1; 840 static_call_update(pv_steal_clock, kvm_steal_clock); 841 842 #ifdef CONFIG_PARAVIRT_SPINLOCKS 843 pv_ops_lock.vcpu_is_preempted = 844 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted); 845 #endif 846 } 847 848 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) 849 apic_update_callback(eoi, kvm_guest_apic_eoi_write); 850 851 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) { 852 static_branch_enable(&kvm_async_pf_enabled); 853 sysvec_install(HYPERVISOR_CALLBACK_VECTOR, sysvec_kvm_asyncpf_interrupt); 854 } 855 856 #ifdef CONFIG_SMP 857 if (pv_tlb_flush_supported()) { 858 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi; 859 pr_info("KVM setup pv remote TLB flush\n"); 860 } 861 862 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu; 863 if (pv_sched_yield_supported()) { 864 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi; 865 pr_info("setup PV sched yield\n"); 866 } 867 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online", 868 kvm_cpu_online, kvm_cpu_down_prepare) < 0) 869 pr_err("failed to install cpu hotplug callbacks\n"); 870 #else 871 sev_map_percpu_data(); 872 kvm_guest_cpu_init(); 873 #endif 874 875 #ifdef CONFIG_CRASH_DUMP 876 machine_ops.crash_shutdown = kvm_crash_shutdown; 877 #endif 878 879 register_syscore(&kvm_syscore); 880 881 /* 882 * Hard lockup detection is enabled by default. Disable it, as guests 883 * can get false positives too easily, for example if the host is 884 * overcommitted. 885 */ 886 hardlockup_detector_disable(); 887 } 888 889 static noinline uint32_t __kvm_cpuid_base(void) 890 { 891 if (boot_cpu_data.cpuid_level < 0) 892 return 0; /* So we don't blow up on old processors */ 893 894 if (boot_cpu_has(X86_FEATURE_HYPERVISOR)) 895 return cpuid_base_hypervisor(KVM_SIGNATURE, 0); 896 897 return 0; 898 } 899 900 static inline uint32_t kvm_cpuid_base(void) 901 { 902 static int kvm_cpuid_base = -1; 903 904 if (kvm_cpuid_base == -1) 905 kvm_cpuid_base = __kvm_cpuid_base(); 906 907 return kvm_cpuid_base; 908 } 909 910 bool kvm_para_available(void) 911 { 912 return kvm_cpuid_base() != 0; 913 } 914 EXPORT_SYMBOL_GPL(kvm_para_available); 915 916 unsigned int kvm_arch_para_features(void) 917 { 918 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES); 919 } 920 921 unsigned int kvm_arch_para_hints(void) 922 { 923 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES); 924 } 925 EXPORT_SYMBOL_GPL(kvm_arch_para_hints); 926 927 static uint32_t __init kvm_detect(void) 928 { 929 return kvm_cpuid_base(); 930 } 931 932 static void __init kvm_apic_init(void) 933 { 934 #ifdef CONFIG_SMP 935 if (pv_ipi_supported()) 936 kvm_setup_pv_ipi(); 937 #endif 938 } 939 940 static bool __init kvm_msi_ext_dest_id(void) 941 { 942 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID); 943 } 944 945 static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc) 946 { 947 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages, 948 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K); 949 } 950 951 static void __init kvm_init_platform(void) 952 { 953 u64 tolud = PFN_PHYS(e820__end_of_low_ram_pfn()); 954 /* 955 * Note, hardware requires variable MTRR ranges to be power-of-2 sized 956 * and naturally aligned. But when forcing guest MTRR state, Linux 957 * doesn't program the forced ranges into hardware. Don't bother doing 958 * the math to generate a technically-legal range. 959 */ 960 struct mtrr_var_range pci_hole = { 961 .base_lo = tolud | X86_MEMTYPE_UC, 962 .mask_lo = (u32)(~(SZ_4G - tolud - 1)) | MTRR_PHYSMASK_V, 963 .mask_hi = (BIT_ULL(boot_cpu_data.x86_phys_bits) - 1) >> 32, 964 }; 965 966 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) && 967 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) { 968 unsigned long nr_pages; 969 int i; 970 971 pv_ops.mmu.notify_page_enc_status_changed = 972 kvm_sev_hc_page_enc_status; 973 974 /* 975 * Reset the host's shared pages list related to kernel 976 * specific page encryption status settings before we load a 977 * new kernel by kexec. Reset the page encryption status 978 * during early boot instead of just before kexec to avoid SMP 979 * races during kvm_pv_guest_cpu_reboot(). 980 * NOTE: We cannot reset the complete shared pages list 981 * here as we need to retain the UEFI/OVMF firmware 982 * specific settings. 983 */ 984 985 for (i = 0; i < e820_table->nr_entries; i++) { 986 struct e820_entry *entry = &e820_table->entries[i]; 987 988 if (entry->type != E820_TYPE_RAM) 989 continue; 990 991 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE); 992 993 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr, 994 nr_pages, 995 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K); 996 } 997 998 /* 999 * Ensure that _bss_decrypted section is marked as decrypted in the 1000 * shared pages list. 1001 */ 1002 early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted, 1003 __end_bss_decrypted - __start_bss_decrypted, 0); 1004 1005 /* 1006 * If not booted using EFI, enable Live migration support. 1007 */ 1008 if (!efi_enabled(EFI_BOOT)) 1009 wrmsrq(MSR_KVM_MIGRATION_CONTROL, 1010 KVM_MIGRATION_READY); 1011 } 1012 kvmclock_init(); 1013 x86_platform.apic_post_init = kvm_apic_init; 1014 1015 /* 1016 * Set WB as the default cache mode for SEV-SNP and TDX, with a single 1017 * UC range for the legacy PCI hole, e.g. so that devices that expect 1018 * to get UC/WC mappings don't get surprised with WB. 1019 */ 1020 guest_force_mtrr_state(&pci_hole, 1, MTRR_TYPE_WRBACK); 1021 } 1022 1023 #if defined(CONFIG_AMD_MEM_ENCRYPT) 1024 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs) 1025 { 1026 /* RAX and CPL are already in the GHCB */ 1027 ghcb_set_rbx(ghcb, regs->bx); 1028 ghcb_set_rcx(ghcb, regs->cx); 1029 ghcb_set_rdx(ghcb, regs->dx); 1030 ghcb_set_rsi(ghcb, regs->si); 1031 } 1032 1033 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs) 1034 { 1035 /* No checking of the return state needed */ 1036 return true; 1037 } 1038 #endif 1039 1040 const __initconst struct hypervisor_x86 x86_hyper_kvm = { 1041 .name = "KVM", 1042 .detect = kvm_detect, 1043 .type = X86_HYPER_KVM, 1044 .init.guest_late_init = kvm_guest_init, 1045 .init.x2apic_available = kvm_para_available, 1046 .init.msi_ext_dest_id = kvm_msi_ext_dest_id, 1047 .init.init_platform = kvm_init_platform, 1048 #if defined(CONFIG_AMD_MEM_ENCRYPT) 1049 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare, 1050 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish, 1051 #endif 1052 }; 1053 1054 static __init int activate_jump_labels(void) 1055 { 1056 if (has_steal_clock) { 1057 static_key_slow_inc(¶virt_steal_enabled); 1058 if (steal_acc) 1059 static_key_slow_inc(¶virt_steal_rq_enabled); 1060 } 1061 1062 return 0; 1063 } 1064 arch_initcall(activate_jump_labels); 1065 1066 #ifdef CONFIG_PARAVIRT_SPINLOCKS 1067 1068 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */ 1069 static void kvm_kick_cpu(int cpu) 1070 { 1071 unsigned long flags = 0; 1072 u32 apicid; 1073 1074 apicid = per_cpu(x86_cpu_to_apicid, cpu); 1075 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid); 1076 } 1077 1078 #include <asm/qspinlock.h> 1079 1080 static void kvm_wait(u8 *ptr, u8 val) 1081 { 1082 if (in_nmi()) 1083 return; 1084 1085 /* 1086 * halt until it's our turn and kicked. Note that we do safe halt 1087 * for irq enabled case to avoid hang when lock info is overwritten 1088 * in irq spinlock slowpath and no spurious interrupt occur to save us. 1089 */ 1090 if (irqs_disabled()) { 1091 if (READ_ONCE(*ptr) == val) 1092 halt(); 1093 } else { 1094 local_irq_disable(); 1095 1096 /* safe_halt() will enable IRQ */ 1097 if (READ_ONCE(*ptr) == val) 1098 safe_halt(); 1099 else 1100 local_irq_enable(); 1101 } 1102 } 1103 1104 /* 1105 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present. 1106 */ 1107 void __init kvm_spinlock_init(void) 1108 { 1109 /* 1110 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs 1111 * are available. 1112 */ 1113 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) { 1114 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n"); 1115 goto out; 1116 } 1117 1118 if (num_possible_cpus() == 1) { 1119 pr_info("PV spinlocks disabled, single CPU\n"); 1120 goto out; 1121 } 1122 1123 if (nopvspin) { 1124 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n"); 1125 goto out; 1126 } 1127 1128 /* 1129 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an 1130 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is 1131 * preferred over native qspinlock when vCPU is preempted. 1132 */ 1133 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) { 1134 pr_info("PV spinlocks disabled, no host support\n"); 1135 return; 1136 } 1137 1138 pr_info("PV spinlocks enabled\n"); 1139 1140 __pv_init_lock_hash(); 1141 static_call_update(queued_spin_lock_slowpath, __pv_queued_spin_lock_slowpath); 1142 static_call_update(queued_spin_unlock, __raw_callee_save___pv_queued_spin_unlock); 1143 pv_ops_lock.wait = kvm_wait; 1144 pv_ops_lock.kick = kvm_kick_cpu; 1145 1146 /* 1147 * When PV spinlock is enabled which is preferred over 1148 * virt_spin_lock(), virt_spin_lock_key's value is meaningless. 1149 * Just disable it anyway. 1150 */ 1151 out: 1152 static_branch_disable(&virt_spin_lock_key); 1153 } 1154 1155 #endif /* CONFIG_PARAVIRT_SPINLOCKS */ 1156 1157 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL 1158 1159 static void kvm_disable_host_haltpoll(void *i) 1160 { 1161 wrmsrq(MSR_KVM_POLL_CONTROL, 0); 1162 } 1163 1164 static void kvm_enable_host_haltpoll(void *i) 1165 { 1166 wrmsrq(MSR_KVM_POLL_CONTROL, 1); 1167 } 1168 1169 void arch_haltpoll_enable(unsigned int cpu) 1170 { 1171 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) { 1172 pr_err_once("host does not support poll control\n"); 1173 pr_err_once("host upgrade recommended\n"); 1174 return; 1175 } 1176 1177 /* Enable guest halt poll disables host halt poll */ 1178 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1); 1179 } 1180 EXPORT_SYMBOL_GPL(arch_haltpoll_enable); 1181 1182 void arch_haltpoll_disable(unsigned int cpu) 1183 { 1184 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) 1185 return; 1186 1187 /* Disable guest halt poll enables host halt poll */ 1188 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1); 1189 } 1190 EXPORT_SYMBOL_GPL(arch_haltpoll_disable); 1191 #endif 1192