arch/x86/kernel/kvm.c

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