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 static int kvmapf = 1;
  39
  40 static int __init parse_no_kvmapf(char *arg)
  41 {
  42         kvmapf = 0;
  43         return 0;
  44 }
  45
  46 early_param("no-kvmapf", parse_no_kvmapf);
  47
  48 static int steal_acc = 1;
  49 static int __init parse_no_stealacc(char *arg)
  50 {
  51         steal_acc = 0;
  52         return 0;
  53 }
  54
  55 early_param("no-steal-acc", parse_no_stealacc);
  56
  57 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  58 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
  59 static int has_steal_clock = 0;
  60
  61 /*
  62  * No need for any "IO delay" on KVM
  63  */
  64 static void kvm_io_delay(void)
  65 {
  66 }
  67
  68 #define KVM_TASK_SLEEP_HASHBITS 8
  69 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  70
  71 struct kvm_task_sleep_node {
  72         struct hlist_node link;
  73         struct swait_queue_head wq;
  74         u32 token;
  75         int cpu;
  76         bool halted;
  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 /*
 100  * @interrupt_kernel: Is this called from a routine which interrupts the kernel
 101  *                    (other than user space)?
 102  */
 103 void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
 104 {
 105         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 106         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 107         struct kvm_task_sleep_node n, *e;
 108         DECLARE_SWAITQUEUE(wait);
 109
 110         rcu_irq_enter();
 111
 112         raw_spin_lock(&b->lock);
 113         e = _find_apf_task(b, token);
 114         if (e) {
 115                 /* dummy entry exist -> wake up was delivered ahead of PF */
 116                 hlist_del(&e->link);
 117                 kfree(e);
 118                 raw_spin_unlock(&b->lock);
 119
 120                 rcu_irq_exit();
 121                 return;
 122         }
 123
 124         n.token = token;
 125         n.cpu = smp_processor_id();
 126         n.halted = is_idle_task(current) ||
 127                    (IS_ENABLED(CONFIG_PREEMPT_COUNT)
 128                     ? preempt_count() > 1 || rcu_preempt_depth()
 129                     : interrupt_kernel);
 130         init_swait_queue_head(&n.wq);
 131         hlist_add_head(&n.link, &b->list);
 132         raw_spin_unlock(&b->lock);
 133
 134         for (;;) {
 135                 if (!n.halted)
 136                         prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 137                 if (hlist_unhashed(&n.link))
 138                         break;
 139
 140                 rcu_irq_exit();
 141
 142                 if (!n.halted) {
 143                         local_irq_enable();
 144                         schedule();
 145                         local_irq_disable();
 146                 } else {
 147                         /*
 148                          * We cannot reschedule. So halt.
 149                          */
 150                         native_safe_halt();
 151                         local_irq_disable();
 152                 }
 153
 154                 rcu_irq_enter();
 155         }
 156         if (!n.halted)
 157                 finish_swait(&n.wq, &wait);
 158
 159         rcu_irq_exit();
 160         return;
 161 }
 162 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 163
 164 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 165 {
 166         hlist_del_init(&n->link);
 167         if (n->halted)
 168                 smp_send_reschedule(n->cpu);
 169         else if (swq_has_sleeper(&n->wq))
 170                 swake_up_one(&n->wq);
 171 }
 172
 173 static void apf_task_wake_all(void)
 174 {
 175         int i;
 176
 177         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 178                 struct hlist_node *p, *next;
 179                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 180                 raw_spin_lock(&b->lock);
 181                 hlist_for_each_safe(p, next, &b->list) {
 182                         struct kvm_task_sleep_node *n =
 183                                 hlist_entry(p, typeof(*n), link);
 184                         if (n->cpu == smp_processor_id())
 185                                 apf_task_wake_one(n);
 186                 }
 187                 raw_spin_unlock(&b->lock);
 188         }
 189 }
 190
 191 void kvm_async_pf_task_wake(u32 token)
 192 {
 193         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 194         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 195         struct kvm_task_sleep_node *n;
 196
 197         if (token == ~0) {
 198                 apf_task_wake_all();
 199                 return;
 200         }
 201
 202 again:
 203         raw_spin_lock(&b->lock);
 204         n = _find_apf_task(b, token);
 205         if (!n) {
 206                 /*
 207                  * async PF was not yet handled.
 208                  * Add dummy entry for the token.
 209                  */
 210                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 211                 if (!n) {
 212                         /*
 213                          * Allocation failed! Busy wait while other cpu
 214                          * handles async PF.
 215                          */
 216                         raw_spin_unlock(&b->lock);
 217                         cpu_relax();
 218                         goto again;
 219                 }
 220                 n->token = token;
 221                 n->cpu = smp_processor_id();
 222                 init_swait_queue_head(&n->wq);
 223                 hlist_add_head(&n->link, &b->list);
 224         } else
 225                 apf_task_wake_one(n);
 226         raw_spin_unlock(&b->lock);
 227         return;
 228 }
 229 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 230
 231 u32 kvm_read_and_reset_pf_reason(void)
 232 {
 233         u32 reason = 0;
 234
 235         if (__this_cpu_read(apf_reason.enabled)) {
 236                 reason = __this_cpu_read(apf_reason.reason);
 237                 __this_cpu_write(apf_reason.reason, 0);
 238         }
 239
 240         return reason;
 241 }
 242 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 243 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 244
 245 dotraplinkage void
 246 do_async_page_fault(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 247 {
 248         switch (kvm_read_and_reset_pf_reason()) {
 249         default:
 250                 do_page_fault(regs, error_code, address);
 251                 break;
 252         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 253                 /* page is swapped out by the host. */
 254                 kvm_async_pf_task_wait((u32)address, !user_mode(regs));
 255                 break;
 256         case KVM_PV_REASON_PAGE_READY:
 257                 rcu_irq_enter();
 258                 kvm_async_pf_task_wake((u32)address);
 259                 rcu_irq_exit();
 260                 break;
 261         }
 262 }
 263 NOKPROBE_SYMBOL(do_async_page_fault);
 264
 265 static void __init paravirt_ops_setup(void)
 266 {
 267         pv_info.name = "KVM";
 268
 269         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 270                 pv_ops.cpu.io_delay = kvm_io_delay;
 271
 272 #ifdef CONFIG_X86_IO_APIC
 273         no_timer_check = 1;
 274 #endif
 275 }
 276
 277 static void kvm_register_steal_time(void)
 278 {
 279         int cpu = smp_processor_id();
 280         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 281
 282         if (!has_steal_clock)
 283                 return;
 284
 285         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 286         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
 287                 cpu, (unsigned long long) slow_virt_to_phys(st));
 288 }
 289
 290 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 291
 292 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 293 {
 294         /**
 295          * This relies on __test_and_clear_bit to modify the memory
 296          * in a way that is atomic with respect to the local CPU.
 297          * The hypervisor only accesses this memory from the local CPU so
 298          * there's no need for lock or memory barriers.
 299          * An optimization barrier is implied in apic write.
 300          */
 301         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 302                 return;
 303         apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
 304 }
 305
 306 static void kvm_guest_cpu_init(void)
 307 {
 308         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 309                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 310
 311 #ifdef CONFIG_PREEMPTION
 312                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 313 #endif
 314                 pa |= KVM_ASYNC_PF_ENABLED;
 315
 316                 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
 317                         pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
 318
 319                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 320                 __this_cpu_write(apf_reason.enabled, 1);
 321                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 322                        smp_processor_id());
 323         }
 324
 325         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 326                 unsigned long pa;
 327                 /* Size alignment is implied but just to make it explicit. */
 328                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 329                 __this_cpu_write(kvm_apic_eoi, 0);
 330                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 331                         | KVM_MSR_ENABLED;
 332                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 333         }
 334
 335         if (has_steal_clock)
 336                 kvm_register_steal_time();
 337 }
 338
 339 static void kvm_pv_disable_apf(void)
 340 {
 341         if (!__this_cpu_read(apf_reason.enabled))
 342                 return;
 343
 344         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 345         __this_cpu_write(apf_reason.enabled, 0);
 346
 347         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 348                smp_processor_id());
 349 }
 350
 351 static void kvm_pv_guest_cpu_reboot(void *unused)
 352 {
 353         /*
 354          * We disable PV EOI before we load a new kernel by kexec,
 355          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 356          * New kernel can re-enable when it boots.
 357          */
 358         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 359                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 360         kvm_pv_disable_apf();
 361         kvm_disable_steal_time();
 362 }
 363
 364 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 365                                 unsigned long code, void *unused)
 366 {
 367         if (code == SYS_RESTART)
 368                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 369         return NOTIFY_DONE;
 370 }
 371
 372 static struct notifier_block kvm_pv_reboot_nb = {
 373         .notifier_call = kvm_pv_reboot_notify,
 374 };
 375
 376 static u64 kvm_steal_clock(int cpu)
 377 {
 378         u64 steal;
 379         struct kvm_steal_time *src;
 380         int version;
 381
 382         src = &per_cpu(steal_time, cpu);
 383         do {
 384                 version = src->version;
 385                 virt_rmb();
 386                 steal = src->steal;
 387                 virt_rmb();
 388         } while ((version & 1) || (version != src->version));
 389
 390         return steal;
 391 }
 392
 393 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 inline void __set_percpu_decrypted(void *ptr, unsigned long size)
 402 {
 403         early_set_memory_decrypted((unsigned long) ptr, size);
 404 }
 405
 406 /*
 407  * Iterate through all possible CPUs and map the memory region pointed
 408  * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
 409  *
 410  * Note: we iterate through all possible CPUs to ensure that CPUs
 411  * hotplugged will have their per-cpu variable already mapped as
 412  * decrypted.
 413  */
 414 static void __init sev_map_percpu_data(void)
 415 {
 416         int cpu;
 417
 418         if (!sev_active())
 419                 return;
 420
 421         for_each_possible_cpu(cpu) {
 422                 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
 423                 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
 424                 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
 425         }
 426 }
 427
 428 #ifdef CONFIG_SMP
 429 #define KVM_IPI_CLUSTER_SIZE    (2 * BITS_PER_LONG)
 430
 431 static void __send_ipi_mask(const struct cpumask *mask, int vector)
 432 {
 433         unsigned long flags;
 434         int cpu, apic_id, icr;
 435         int min = 0, max = 0;
 436 #ifdef CONFIG_X86_64
 437         __uint128_t ipi_bitmap = 0;
 438 #else
 439         u64 ipi_bitmap = 0;
 440 #endif
 441         long ret;
 442
 443         if (cpumask_empty(mask))
 444                 return;
 445
 446         local_irq_save(flags);
 447
 448         switch (vector) {
 449         default:
 450                 icr = APIC_DM_FIXED | vector;
 451                 break;
 452         case NMI_VECTOR:
 453                 icr = APIC_DM_NMI;
 454                 break;
 455         }
 456
 457         for_each_cpu(cpu, mask) {
 458                 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
 459                 if (!ipi_bitmap) {
 460                         min = max = apic_id;
 461                 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
 462                         ipi_bitmap <<= min - apic_id;
 463                         min = apic_id;
 464                 } else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
 465                         max = apic_id < max ? max : apic_id;
 466                 } else {
 467                         ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 468                                 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 469                         WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
 470                         min = max = apic_id;
 471                         ipi_bitmap = 0;
 472                 }
 473                 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
 474         }
 475
 476         if (ipi_bitmap) {
 477                 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 478                         (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 479                 WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
 480         }
 481
 482         local_irq_restore(flags);
 483 }
 484
 485 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
 486 {
 487         __send_ipi_mask(mask, vector);
 488 }
 489
 490 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
 491 {
 492         unsigned int this_cpu = smp_processor_id();
 493         struct cpumask new_mask;
 494         const struct cpumask *local_mask;
 495
 496         cpumask_copy(&new_mask, mask);
 497         cpumask_clear_cpu(this_cpu, &new_mask);
 498         local_mask = &new_mask;
 499         __send_ipi_mask(local_mask, vector);
 500 }
 501
 502 /*
 503  * Set the IPI entry points
 504  */
 505 static void kvm_setup_pv_ipi(void)
 506 {
 507         apic->send_IPI_mask = kvm_send_ipi_mask;
 508         apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
 509         pr_info("KVM setup pv IPIs\n");
 510 }
 511
 512 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
 513 {
 514         int cpu;
 515
 516         native_send_call_func_ipi(mask);
 517
 518         /* Make sure other vCPUs get a chance to run if they need to. */
 519         for_each_cpu(cpu, mask) {
 520                 if (vcpu_is_preempted(cpu)) {
 521                         kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
 522                         break;
 523                 }
 524         }
 525 }
 526
 527 static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
 528 {
 529         native_smp_prepare_cpus(max_cpus);
 530         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
 531                 static_branch_disable(&virt_spin_lock_key);
 532 }
 533
 534 static void __init kvm_smp_prepare_boot_cpu(void)
 535 {
 536         /*
 537          * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
 538          * shares the guest physical address with the hypervisor.
 539          */
 540         sev_map_percpu_data();
 541
 542         kvm_guest_cpu_init();
 543         native_smp_prepare_boot_cpu();
 544         kvm_spinlock_init();
 545 }
 546
 547 static void kvm_guest_cpu_offline(void)
 548 {
 549         kvm_disable_steal_time();
 550         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 551                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 552         kvm_pv_disable_apf();
 553         apf_task_wake_all();
 554 }
 555
 556 static int kvm_cpu_online(unsigned int cpu)
 557 {
 558         local_irq_disable();
 559         kvm_guest_cpu_init();
 560         local_irq_enable();
 561         return 0;
 562 }
 563
 564 static int kvm_cpu_down_prepare(unsigned int cpu)
 565 {
 566         local_irq_disable();
 567         kvm_guest_cpu_offline();
 568         local_irq_enable();
 569         return 0;
 570 }
 571 #endif
 572
 573 static void __init kvm_apf_trap_init(void)
 574 {
 575         update_intr_gate(X86_TRAP_PF, async_page_fault);
 576 }
 577
 578 static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
 579
 580 static void kvm_flush_tlb_others(const struct cpumask *cpumask,
 581                         const struct flush_tlb_info *info)
 582 {
 583         u8 state;
 584         int cpu;
 585         struct kvm_steal_time *src;
 586         struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
 587
 588         cpumask_copy(flushmask, cpumask);
 589         /*
 590          * We have to call flush only on online vCPUs. And
 591          * queue flush_on_enter for pre-empted vCPUs
 592          */
 593         for_each_cpu(cpu, flushmask) {
 594                 src = &per_cpu(steal_time, cpu);
 595                 state = READ_ONCE(src->preempted);
 596                 if ((state & KVM_VCPU_PREEMPTED)) {
 597                         if (try_cmpxchg(&src->preempted, &state,
 598                                         state | KVM_VCPU_FLUSH_TLB))
 599                                 __cpumask_clear_cpu(cpu, flushmask);
 600                 }
 601         }
 602
 603         native_flush_tlb_others(flushmask, info);
 604 }
 605
 606 static void __init kvm_guest_init(void)
 607 {
 608         int i;
 609
 610         paravirt_ops_setup();
 611         register_reboot_notifier(&kvm_pv_reboot_nb);
 612         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 613                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 614         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 615                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 616
 617         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 618                 has_steal_clock = 1;
 619                 pv_ops.time.steal_clock = kvm_steal_clock;
 620         }
 621
 622         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 623             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 624             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 625                 pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
 626                 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
 627         }
 628
 629         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 630                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 631
 632 #ifdef CONFIG_SMP
 633         smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
 634         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 635         if (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
 636             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 637             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 638                 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
 639                 pr_info("KVM setup pv sched yield\n");
 640         }
 641         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 642                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 643                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 644 #else
 645         sev_map_percpu_data();
 646         kvm_guest_cpu_init();
 647 #endif
 648
 649         /*
 650          * Hard lockup detection is enabled by default. Disable it, as guests
 651          * can get false positives too easily, for example if the host is
 652          * overcommitted.
 653          */
 654         hardlockup_detector_disable();
 655 }
 656
 657 static noinline uint32_t __kvm_cpuid_base(void)
 658 {
 659         if (boot_cpu_data.cpuid_level < 0)
 660                 return 0;       /* So we don't blow up on old processors */
 661
 662         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 663                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 664
 665         return 0;
 666 }
 667
 668 static inline uint32_t kvm_cpuid_base(void)
 669 {
 670         static int kvm_cpuid_base = -1;
 671
 672         if (kvm_cpuid_base == -1)
 673                 kvm_cpuid_base = __kvm_cpuid_base();
 674
 675         return kvm_cpuid_base;
 676 }
 677
 678 bool kvm_para_available(void)
 679 {
 680         return kvm_cpuid_base() != 0;
 681 }
 682 EXPORT_SYMBOL_GPL(kvm_para_available);
 683
 684 unsigned int kvm_arch_para_features(void)
 685 {
 686         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 687 }
 688
 689 unsigned int kvm_arch_para_hints(void)
 690 {
 691         return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 692 }
 693 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
 694
 695 static uint32_t __init kvm_detect(void)
 696 {
 697         return kvm_cpuid_base();
 698 }
 699
 700 static void __init kvm_apic_init(void)
 701 {
 702 #if defined(CONFIG_SMP)
 703         if (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI))
 704                 kvm_setup_pv_ipi();
 705 #endif
 706 }
 707
 708 static void __init kvm_init_platform(void)
 709 {
 710         kvmclock_init();
 711         x86_platform.apic_post_init = kvm_apic_init;
 712 }
 713
 714 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
 715         .name                   = "KVM",
 716         .detect                 = kvm_detect,
 717         .type                   = X86_HYPER_KVM,
 718         .init.guest_late_init   = kvm_guest_init,
 719         .init.x2apic_available  = kvm_para_available,
 720         .init.init_platform     = kvm_init_platform,
 721 };
 722
 723 static __init int activate_jump_labels(void)
 724 {
 725         if (has_steal_clock) {
 726                 static_key_slow_inc(&paravirt_steal_enabled);
 727                 if (steal_acc)
 728                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 729         }
 730
 731         return 0;
 732 }
 733 arch_initcall(activate_jump_labels);
 734
 735 static __init int kvm_setup_pv_tlb_flush(void)
 736 {
 737         int cpu;
 738
 739         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 740             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 741             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 742                 for_each_possible_cpu(cpu) {
 743                         zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
 744                                 GFP_KERNEL, cpu_to_node(cpu));
 745                 }
 746                 pr_info("KVM setup pv remote TLB flush\n");
 747         }
 748
 749         return 0;
 750 }
 751 arch_initcall(kvm_setup_pv_tlb_flush);
 752
 753 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 754
 755 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 756 static void kvm_kick_cpu(int cpu)
 757 {
 758         int apicid;
 759         unsigned long flags = 0;
 760
 761         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 762         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 763 }
 764
 765 #include <asm/qspinlock.h>
 766
 767 static void kvm_wait(u8 *ptr, u8 val)
 768 {
 769         unsigned long flags;
 770
 771         if (in_nmi())
 772                 return;
 773
 774         local_irq_save(flags);
 775
 776         if (READ_ONCE(*ptr) != val)
 777                 goto out;
 778
 779         /*
 780          * halt until it's our turn and kicked. Note that we do safe halt
 781          * for irq enabled case to avoid hang when lock info is overwritten
 782          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 783          */
 784         if (arch_irqs_disabled_flags(flags))
 785                 halt();
 786         else
 787                 safe_halt();
 788
 789 out:
 790         local_irq_restore(flags);
 791 }
 792
 793 #ifdef CONFIG_X86_32
 794 __visible bool __kvm_vcpu_is_preempted(long cpu)
 795 {
 796         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 797
 798         return !!(src->preempted & KVM_VCPU_PREEMPTED);
 799 }
 800 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 801
 802 #else
 803
 804 #include <asm/asm-offsets.h>
 805
 806 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 807
 808 /*
 809  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 810  * restoring to/from the stack.
 811  */
 812 asm(
 813 ".pushsection .text;"
 814 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 815 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 816 "__raw_callee_save___kvm_vcpu_is_preempted:"
 817 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 818 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 819 "setne  %al;"
 820 "ret;"
 821 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
 822 ".popsection");
 823
 824 #endif
 825
 826 /*
 827  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 828  */
 829 void __init kvm_spinlock_init(void)
 830 {
 831         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 832         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 833                 return;
 834
 835         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
 836                 return;
 837
 838         /* Don't use the pvqspinlock code if there is only 1 vCPU. */
 839         if (num_possible_cpus() == 1)
 840                 return;
 841
 842         __pv_init_lock_hash();
 843         pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 844         pv_ops.lock.queued_spin_unlock =
 845                 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 846         pv_ops.lock.wait = kvm_wait;
 847         pv_ops.lock.kick = kvm_kick_cpu;
 848
 849         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 850                 pv_ops.lock.vcpu_is_preempted =
 851                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 852         }
 853 }
 854
 855 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */
 856
 857 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
 858
 859 static void kvm_disable_host_haltpoll(void *i)
 860 {
 861         wrmsrl(MSR_KVM_POLL_CONTROL, 0);
 862 }
 863
 864 static void kvm_enable_host_haltpoll(void *i)
 865 {
 866         wrmsrl(MSR_KVM_POLL_CONTROL, 1);
 867 }
 868
 869 void arch_haltpoll_enable(unsigned int cpu)
 870 {
 871         if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
 872                 pr_err_once("kvm: host does not support poll control\n");
 873                 pr_err_once("kvm: host upgrade recommended\n");
 874                 return;
 875         }
 876
 877         /* Enable guest halt poll disables host halt poll */
 878         smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
 879 }
 880 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
 881
 882 void arch_haltpoll_disable(unsigned int cpu)
 883 {
 884         if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
 885                 return;
 886
 887         /* Enable guest halt poll disables host halt poll */
 888         smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
 889 }
 890 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
 891 #endif