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