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