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