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