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