arch/x86/kernel/kvm.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * KVM paravirt_ops implementation
   4  *
   5  * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   6  * Copyright IBM Corporation, 2007
   7  *   Authors: Anthony Liguori <aliguori@us.ibm.com>
   8  */
   9
  10 #define pr_fmt(fmt) "kvm-guest: " fmt
  11
  12 #include <linux/context_tracking.h>
  13 #include <linux/init.h>
  14 #include <linux/irq.h>
  15 #include <linux/kernel.h>
  16 #include <linux/kvm_para.h>
  17 #include <linux/cpu.h>
  18 #include <linux/mm.h>
  19 #include <linux/highmem.h>
  20 #include <linux/hardirq.h>
  21 #include <linux/notifier.h>
  22 #include <linux/reboot.h>
  23 #include <linux/hash.h>
  24 #include <linux/sched.h>
  25 #include <linux/slab.h>
  26 #include <linux/kprobes.h>
  27 #include <linux/nmi.h>
  28 #include <linux/swait.h>
  29 #include <linux/syscore_ops.h>
  30 #include <asm/timer.h>
  31 #include <asm/cpu.h>
  32 #include <asm/traps.h>
  33 #include <asm/desc.h>
  34 #include <asm/tlbflush.h>
  35 #include <asm/apic.h>
  36 #include <asm/apicdef.h>
  37 #include <asm/hypervisor.h>
  38 #include <asm/tlb.h>
  39 #include <asm/cpuidle_haltpoll.h>
  40 #include <asm/ptrace.h>
  41 #include <asm/reboot.h>
  42 #include <asm/svm.h>
  43
  44 DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
  45
  46 static int kvmapf = 1;
  47
  48 static int __init parse_no_kvmapf(char *arg)
  49 {
  50         kvmapf = 0;
  51         return 0;
  52 }
  53
  54 early_param("no-kvmapf", parse_no_kvmapf);
  55
  56 static int steal_acc = 1;
  57 static int __init parse_no_stealacc(char *arg)
  58 {
  59         steal_acc = 0;
  60         return 0;
  61 }
  62
  63 early_param("no-steal-acc", parse_no_stealacc);
  64
  65 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  66 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
  67 static int has_steal_clock = 0;
  68
  69 /*
  70  * No need for any "IO delay" on KVM
  71  */
  72 static void kvm_io_delay(void)
  73 {
  74 }
  75
  76 #define KVM_TASK_SLEEP_HASHBITS 8
  77 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  78
  79 struct kvm_task_sleep_node {
  80         struct hlist_node link;
  81         struct swait_queue_head wq;
  82         u32 token;
  83         int cpu;
  84 };
  85
  86 static struct kvm_task_sleep_head {
  87         raw_spinlock_t lock;
  88         struct hlist_head list;
  89 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
  90
  91 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
  92                                                   u32 token)
  93 {
  94         struct hlist_node *p;
  95
  96         hlist_for_each(p, &b->list) {
  97                 struct kvm_task_sleep_node *n =
  98                         hlist_entry(p, typeof(*n), link);
  99                 if (n->token == token)
 100                         return n;
 101         }
 102
 103         return NULL;
 104 }
 105
 106 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
 107 {
 108         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 109         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 110         struct kvm_task_sleep_node *e;
 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                 raw_spin_unlock(&b->lock);
 118                 kfree(e);
 119                 return false;
 120         }
 121
 122         n->token = token;
 123         n->cpu = smp_processor_id();
 124         init_swait_queue_head(&n->wq);
 125         hlist_add_head(&n->link, &b->list);
 126         raw_spin_unlock(&b->lock);
 127         return true;
 128 }
 129
 130 /*
 131  * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
 132  * @token:      Token to identify the sleep node entry
 133  *
 134  * Invoked from the async pagefault handling code or from the VM exit page
 135  * fault handler. In both cases RCU is watching.
 136  */
 137 void kvm_async_pf_task_wait_schedule(u32 token)
 138 {
 139         struct kvm_task_sleep_node n;
 140         DECLARE_SWAITQUEUE(wait);
 141
 142         lockdep_assert_irqs_disabled();
 143
 144         if (!kvm_async_pf_queue_task(token, &n))
 145                 return;
 146
 147         for (;;) {
 148                 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 149                 if (hlist_unhashed(&n.link))
 150                         break;
 151
 152                 local_irq_enable();
 153                 schedule();
 154                 local_irq_disable();
 155         }
 156         finish_swait(&n.wq, &wait);
 157 }
 158 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
 159
 160 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 161 {
 162         hlist_del_init(&n->link);
 163         if (swq_has_sleeper(&n->wq))
 164                 swake_up_one(&n->wq);
 165 }
 166
 167 static void apf_task_wake_all(void)
 168 {
 169         int i;
 170
 171         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 172                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 173                 struct kvm_task_sleep_node *n;
 174                 struct hlist_node *p, *next;
 175
 176                 raw_spin_lock(&b->lock);
 177                 hlist_for_each_safe(p, next, &b->list) {
 178                         n = hlist_entry(p, typeof(*n), link);
 179                         if (n->cpu == smp_processor_id())
 180                                 apf_task_wake_one(n);
 181                 }
 182                 raw_spin_unlock(&b->lock);
 183         }
 184 }
 185
 186 void kvm_async_pf_task_wake(u32 token)
 187 {
 188         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 189         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 190         struct kvm_task_sleep_node *n;
 191
 192         if (token == ~0) {
 193                 apf_task_wake_all();
 194                 return;
 195         }
 196
 197 again:
 198         raw_spin_lock(&b->lock);
 199         n = _find_apf_task(b, token);
 200         if (!n) {
 201                 /*
 202                  * async PF was not yet handled.
 203                  * Add dummy entry for the token.
 204                  */
 205                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 206                 if (!n) {
 207                         /*
 208                          * Allocation failed! Busy wait while other cpu
 209                          * handles async PF.
 210                          */
 211                         raw_spin_unlock(&b->lock);
 212                         cpu_relax();
 213                         goto again;
 214                 }
 215                 n->token = token;
 216                 n->cpu = smp_processor_id();
 217                 init_swait_queue_head(&n->wq);
 218                 hlist_add_head(&n->link, &b->list);
 219         } else {
 220                 apf_task_wake_one(n);
 221         }
 222         raw_spin_unlock(&b->lock);
 223         return;
 224 }
 225 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 226
 227 noinstr u32 kvm_read_and_reset_apf_flags(void)
 228 {
 229         u32 flags = 0;
 230
 231         if (__this_cpu_read(apf_reason.enabled)) {
 232                 flags = __this_cpu_read(apf_reason.flags);
 233                 __this_cpu_write(apf_reason.flags, 0);
 234         }
 235
 236         return flags;
 237 }
 238 EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
 239
 240 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
 241 {
 242         u32 flags = kvm_read_and_reset_apf_flags();
 243         irqentry_state_t state;
 244
 245         if (!flags)
 246                 return false;
 247
 248         state = irqentry_enter(regs);
 249         instrumentation_begin();
 250
 251         /*
 252          * If the host managed to inject an async #PF into an interrupt
 253          * disabled region, then die hard as this is not going to end well
 254          * and the host side is seriously broken.
 255          */
 256         if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
 257                 panic("Host injected async #PF in interrupt disabled region\n");
 258
 259         if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
 260                 if (unlikely(!(user_mode(regs))))
 261                         panic("Host injected async #PF in kernel mode\n");
 262                 /* Page is swapped out by the host. */
 263                 kvm_async_pf_task_wait_schedule(token);
 264         } else {
 265                 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
 266         }
 267
 268         instrumentation_end();
 269         irqentry_exit(regs, state);
 270         return true;
 271 }
 272
 273 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
 274 {
 275         struct pt_regs *old_regs = set_irq_regs(regs);
 276         u32 token;
 277
 278         ack_APIC_irq();
 279
 280         inc_irq_stat(irq_hv_callback_count);
 281
 282         if (__this_cpu_read(apf_reason.enabled)) {
 283                 token = __this_cpu_read(apf_reason.token);
 284                 kvm_async_pf_task_wake(token);
 285                 __this_cpu_write(apf_reason.token, 0);
 286                 wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
 287         }
 288
 289         set_irq_regs(old_regs);
 290 }
 291
 292 static void __init paravirt_ops_setup(void)
 293 {
 294         pv_info.name = "KVM";
 295
 296         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 297                 pv_ops.cpu.io_delay = kvm_io_delay;
 298
 299 #ifdef CONFIG_X86_IO_APIC
 300         no_timer_check = 1;
 301 #endif
 302 }
 303
 304 static void kvm_register_steal_time(void)
 305 {
 306         int cpu = smp_processor_id();
 307         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 308
 309         if (!has_steal_clock)
 310                 return;
 311
 312         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 313         pr_info("stealtime: cpu %d, msr %llx\n", cpu,
 314                 (unsigned long long) slow_virt_to_phys(st));
 315 }
 316
 317 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 318
 319 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 320 {
 321         /**
 322          * This relies on __test_and_clear_bit to modify the memory
 323          * in a way that is atomic with respect to the local CPU.
 324          * The hypervisor only accesses this memory from the local CPU so
 325          * there's no need for lock or memory barriers.
 326          * An optimization barrier is implied in apic write.
 327          */
 328         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 329                 return;
 330         apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
 331 }
 332
 333 static void kvm_guest_cpu_init(void)
 334 {
 335         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
 336                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 337
 338                 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
 339
 340                 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 341                 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
 342
 343                 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
 344                         pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
 345
 346                 wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
 347
 348                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 349                 __this_cpu_write(apf_reason.enabled, 1);
 350                 pr_info("setup async PF for cpu %d\n", smp_processor_id());
 351         }
 352
 353         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 354                 unsigned long pa;
 355
 356                 /* Size alignment is implied but just to make it explicit. */
 357                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 358                 __this_cpu_write(kvm_apic_eoi, 0);
 359                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 360                         | KVM_MSR_ENABLED;
 361                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 362         }
 363
 364         if (has_steal_clock)
 365                 kvm_register_steal_time();
 366 }
 367
 368 static void kvm_pv_disable_apf(void)
 369 {
 370         if (!__this_cpu_read(apf_reason.enabled))
 371                 return;
 372
 373         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 374         __this_cpu_write(apf_reason.enabled, 0);
 375
 376         pr_info("disable async PF for cpu %d\n", smp_processor_id());
 377 }
 378
 379 static void kvm_disable_steal_time(void)
 380 {
 381         if (!has_steal_clock)
 382                 return;
 383
 384         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 385 }
 386
 387 static u64 kvm_steal_clock(int cpu)
 388 {
 389         u64 steal;
 390         struct kvm_steal_time *src;
 391         int version;
 392
 393         src = &per_cpu(steal_time, cpu);
 394         do {
 395                 version = src->version;
 396                 virt_rmb();
 397                 steal = src->steal;
 398                 virt_rmb();
 399         } while ((version & 1) || (version != src->version));
 400
 401         return steal;
 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 void kvm_guest_cpu_offline(bool shutdown)
 432 {
 433         kvm_disable_steal_time();
 434         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 435                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 436         kvm_pv_disable_apf();
 437         if (!shutdown)
 438                 apf_task_wake_all();
 439         kvmclock_disable();
 440 }
 441
 442 static int kvm_cpu_online(unsigned int cpu)
 443 {
 444         unsigned long flags;
 445
 446         local_irq_save(flags);
 447         kvm_guest_cpu_init();
 448         local_irq_restore(flags);
 449         return 0;
 450 }
 451
 452 #ifdef CONFIG_SMP
 453
 454 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
 455
 456 static bool pv_tlb_flush_supported(void)
 457 {
 458         return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 459                 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 460                 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME));
 461 }
 462
 463 static bool pv_ipi_supported(void)
 464 {
 465         return kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI);
 466 }
 467
 468 static bool pv_sched_yield_supported(void)
 469 {
 470         return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
 471                 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 472             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME));
 473 }
 474
 475 #define KVM_IPI_CLUSTER_SIZE    (2 * BITS_PER_LONG)
 476
 477 static void __send_ipi_mask(const struct cpumask *mask, int vector)
 478 {
 479         unsigned long flags;
 480         int cpu, apic_id, icr;
 481         int min = 0, max = 0;
 482 #ifdef CONFIG_X86_64
 483         __uint128_t ipi_bitmap = 0;
 484 #else
 485         u64 ipi_bitmap = 0;
 486 #endif
 487         long ret;
 488
 489         if (cpumask_empty(mask))
 490                 return;
 491
 492         local_irq_save(flags);
 493
 494         switch (vector) {
 495         default:
 496                 icr = APIC_DM_FIXED | vector;
 497                 break;
 498         case NMI_VECTOR:
 499                 icr = APIC_DM_NMI;
 500                 break;
 501         }
 502
 503         for_each_cpu(cpu, mask) {
 504                 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
 505                 if (!ipi_bitmap) {
 506                         min = max = apic_id;
 507                 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
 508                         ipi_bitmap <<= min - apic_id;
 509                         min = apic_id;
 510                 } else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
 511                         max = apic_id < max ? max : apic_id;
 512                 } else {
 513                         ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 514                                 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 515                         WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
 516                                   ret);
 517                         min = max = apic_id;
 518                         ipi_bitmap = 0;
 519                 }
 520                 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
 521         }
 522
 523         if (ipi_bitmap) {
 524                 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 525                         (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 526                 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
 527                           ret);
 528         }
 529
 530         local_irq_restore(flags);
 531 }
 532
 533 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
 534 {
 535         __send_ipi_mask(mask, vector);
 536 }
 537
 538 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
 539 {
 540         unsigned int this_cpu = smp_processor_id();
 541         struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
 542         const struct cpumask *local_mask;
 543
 544         cpumask_copy(new_mask, mask);
 545         cpumask_clear_cpu(this_cpu, new_mask);
 546         local_mask = new_mask;
 547         __send_ipi_mask(local_mask, vector);
 548 }
 549
 550 /*
 551  * Set the IPI entry points
 552  */
 553 static void kvm_setup_pv_ipi(void)
 554 {
 555         apic->send_IPI_mask = kvm_send_ipi_mask;
 556         apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
 557         pr_info("setup PV IPIs\n");
 558 }
 559
 560 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
 561 {
 562         int cpu;
 563
 564         native_send_call_func_ipi(mask);
 565
 566         /* Make sure other vCPUs get a chance to run if they need to. */
 567         for_each_cpu(cpu, mask) {
 568                 if (vcpu_is_preempted(cpu)) {
 569                         kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
 570                         break;
 571                 }
 572         }
 573 }
 574
 575 static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
 576                         const struct flush_tlb_info *info)
 577 {
 578         u8 state;
 579         int cpu;
 580         struct kvm_steal_time *src;
 581         struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
 582
 583         cpumask_copy(flushmask, cpumask);
 584         /*
 585          * We have to call flush only on online vCPUs. And
 586          * queue flush_on_enter for pre-empted vCPUs
 587          */
 588         for_each_cpu(cpu, flushmask) {
 589                 /*
 590                  * The local vCPU is never preempted, so we do not explicitly
 591                  * skip check for local vCPU - it will never be cleared from
 592                  * flushmask.
 593                  */
 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_multi(flushmask, info);
 604 }
 605
 606 static __init int kvm_alloc_cpumask(void)
 607 {
 608         int cpu;
 609
 610         if (!kvm_para_available() || nopv)
 611                 return 0;
 612
 613         if (pv_tlb_flush_supported() || pv_ipi_supported())
 614                 for_each_possible_cpu(cpu) {
 615                         zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
 616                                 GFP_KERNEL, cpu_to_node(cpu));
 617                 }
 618
 619         return 0;
 620 }
 621 arch_initcall(kvm_alloc_cpumask);
 622
 623 static void __init kvm_smp_prepare_boot_cpu(void)
 624 {
 625         /*
 626          * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
 627          * shares the guest physical address with the hypervisor.
 628          */
 629         sev_map_percpu_data();
 630
 631         kvm_guest_cpu_init();
 632         native_smp_prepare_boot_cpu();
 633         kvm_spinlock_init();
 634 }
 635
 636 static int kvm_cpu_down_prepare(unsigned int cpu)
 637 {
 638         unsigned long flags;
 639
 640         local_irq_save(flags);
 641         kvm_guest_cpu_offline(false);
 642         local_irq_restore(flags);
 643         return 0;
 644 }
 645
 646 #endif
 647
 648 static int kvm_suspend(void)
 649 {
 650         kvm_guest_cpu_offline(false);
 651
 652         return 0;
 653 }
 654
 655 static void kvm_resume(void)
 656 {
 657         kvm_cpu_online(raw_smp_processor_id());
 658 }
 659
 660 static struct syscore_ops kvm_syscore_ops = {
 661         .suspend        = kvm_suspend,
 662         .resume         = kvm_resume,
 663 };
 664
 665 static void kvm_pv_guest_cpu_reboot(void *unused)
 666 {
 667         kvm_guest_cpu_offline(true);
 668 }
 669
 670 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 671                                 unsigned long code, void *unused)
 672 {
 673         if (code == SYS_RESTART)
 674                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 675         return NOTIFY_DONE;
 676 }
 677
 678 static struct notifier_block kvm_pv_reboot_nb = {
 679         .notifier_call = kvm_pv_reboot_notify,
 680 };
 681
 682 /*
 683  * After a PV feature is registered, the host will keep writing to the
 684  * registered memory location. If the guest happens to shutdown, this memory
 685  * won't be valid. In cases like kexec, in which you install a new kernel, this
 686  * means a random memory location will be kept being written.
 687  */
 688 #ifdef CONFIG_KEXEC_CORE
 689 static void kvm_crash_shutdown(struct pt_regs *regs)
 690 {
 691         kvm_guest_cpu_offline(true);
 692         native_machine_crash_shutdown(regs);
 693 }
 694 #endif
 695
 696 static void __init kvm_guest_init(void)
 697 {
 698         int i;
 699
 700         paravirt_ops_setup();
 701         register_reboot_notifier(&kvm_pv_reboot_nb);
 702         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 703                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 704
 705         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 706                 has_steal_clock = 1;
 707                 static_call_update(pv_steal_clock, kvm_steal_clock);
 708         }
 709
 710         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 711                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 712
 713         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
 714                 static_branch_enable(&kvm_async_pf_enabled);
 715                 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
 716         }
 717
 718 #ifdef CONFIG_SMP
 719         if (pv_tlb_flush_supported()) {
 720                 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
 721                 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
 722                 pr_info("KVM setup pv remote TLB flush\n");
 723         }
 724
 725         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 726         if (pv_sched_yield_supported()) {
 727                 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
 728                 pr_info("setup PV sched yield\n");
 729         }
 730         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 731                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 732                 pr_err("failed to install cpu hotplug callbacks\n");
 733 #else
 734         sev_map_percpu_data();
 735         kvm_guest_cpu_init();
 736 #endif
 737
 738 #ifdef CONFIG_KEXEC_CORE
 739         machine_ops.crash_shutdown = kvm_crash_shutdown;
 740 #endif
 741
 742         register_syscore_ops(&kvm_syscore_ops);
 743
 744         /*
 745          * Hard lockup detection is enabled by default. Disable it, as guests
 746          * can get false positives too easily, for example if the host is
 747          * overcommitted.
 748          */
 749         hardlockup_detector_disable();
 750 }
 751
 752 static noinline uint32_t __kvm_cpuid_base(void)
 753 {
 754         if (boot_cpu_data.cpuid_level < 0)
 755                 return 0;       /* So we don't blow up on old processors */
 756
 757         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 758                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 759
 760         return 0;
 761 }
 762
 763 static inline uint32_t kvm_cpuid_base(void)
 764 {
 765         static int kvm_cpuid_base = -1;
 766
 767         if (kvm_cpuid_base == -1)
 768                 kvm_cpuid_base = __kvm_cpuid_base();
 769
 770         return kvm_cpuid_base;
 771 }
 772
 773 bool kvm_para_available(void)
 774 {
 775         return kvm_cpuid_base() != 0;
 776 }
 777 EXPORT_SYMBOL_GPL(kvm_para_available);
 778
 779 unsigned int kvm_arch_para_features(void)
 780 {
 781         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 782 }
 783
 784 unsigned int kvm_arch_para_hints(void)
 785 {
 786         return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 787 }
 788 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
 789
 790 static uint32_t __init kvm_detect(void)
 791 {
 792         return kvm_cpuid_base();
 793 }
 794
 795 static void __init kvm_apic_init(void)
 796 {
 797 #ifdef CONFIG_SMP
 798         if (pv_ipi_supported())
 799                 kvm_setup_pv_ipi();
 800 #endif
 801 }
 802
 803 static bool __init kvm_msi_ext_dest_id(void)
 804 {
 805         return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
 806 }
 807
 808 static void __init kvm_init_platform(void)
 809 {
 810         kvmclock_init();
 811         x86_platform.apic_post_init = kvm_apic_init;
 812 }
 813
 814 #if defined(CONFIG_AMD_MEM_ENCRYPT)
 815 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
 816 {
 817         /* RAX and CPL are already in the GHCB */
 818         ghcb_set_rbx(ghcb, regs->bx);
 819         ghcb_set_rcx(ghcb, regs->cx);
 820         ghcb_set_rdx(ghcb, regs->dx);
 821         ghcb_set_rsi(ghcb, regs->si);
 822 }
 823
 824 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
 825 {
 826         /* No checking of the return state needed */
 827         return true;
 828 }
 829 #endif
 830
 831 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
 832         .name                           = "KVM",
 833         .detect                         = kvm_detect,
 834         .type                           = X86_HYPER_KVM,
 835         .init.guest_late_init           = kvm_guest_init,
 836         .init.x2apic_available          = kvm_para_available,
 837         .init.msi_ext_dest_id           = kvm_msi_ext_dest_id,
 838         .init.init_platform             = kvm_init_platform,
 839 #if defined(CONFIG_AMD_MEM_ENCRYPT)
 840         .runtime.sev_es_hcall_prepare   = kvm_sev_es_hcall_prepare,
 841         .runtime.sev_es_hcall_finish    = kvm_sev_es_hcall_finish,
 842 #endif
 843 };
 844
 845 static __init int activate_jump_labels(void)
 846 {
 847         if (has_steal_clock) {
 848                 static_key_slow_inc(&paravirt_steal_enabled);
 849                 if (steal_acc)
 850                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 851         }
 852
 853         return 0;
 854 }
 855 arch_initcall(activate_jump_labels);
 856
 857 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 858
 859 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 860 static void kvm_kick_cpu(int cpu)
 861 {
 862         int apicid;
 863         unsigned long flags = 0;
 864
 865         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 866         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 867 }
 868
 869 #include <asm/qspinlock.h>
 870
 871 static void kvm_wait(u8 *ptr, u8 val)
 872 {
 873         if (in_nmi())
 874                 return;
 875
 876         /*
 877          * halt until it's our turn and kicked. Note that we do safe halt
 878          * for irq enabled case to avoid hang when lock info is overwritten
 879          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 880          */
 881         if (irqs_disabled()) {
 882                 if (READ_ONCE(*ptr) == val)
 883                         halt();
 884         } else {
 885                 local_irq_disable();
 886
 887                 /* safe_halt() will enable IRQ */
 888                 if (READ_ONCE(*ptr) == val)
 889                         safe_halt();
 890                 else
 891                         local_irq_enable();
 892         }
 893 }
 894
 895 #ifdef CONFIG_X86_32
 896 __visible bool __kvm_vcpu_is_preempted(long cpu)
 897 {
 898         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 899
 900         return !!(src->preempted & KVM_VCPU_PREEMPTED);
 901 }
 902 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 903
 904 #else
 905
 906 #include <asm/asm-offsets.h>
 907
 908 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 909
 910 /*
 911  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 912  * restoring to/from the stack.
 913  */
 914 asm(
 915 ".pushsection .text;"
 916 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 917 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 918 "__raw_callee_save___kvm_vcpu_is_preempted:"
 919 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 920 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 921 "setne  %al;"
 922 "ret;"
 923 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
 924 ".popsection");
 925
 926 #endif
 927
 928 /*
 929  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 930  */
 931 void __init kvm_spinlock_init(void)
 932 {
 933         /*
 934          * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
 935          * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
 936          * preferred over native qspinlock when vCPU is preempted.
 937          */
 938         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
 939                 pr_info("PV spinlocks disabled, no host support\n");
 940                 return;
 941         }
 942
 943         /*
 944          * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
 945          * are available.
 946          */
 947         if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
 948                 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
 949                 goto out;
 950         }
 951
 952         if (num_possible_cpus() == 1) {
 953                 pr_info("PV spinlocks disabled, single CPU\n");
 954                 goto out;
 955         }
 956
 957         if (nopvspin) {
 958                 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
 959                 goto out;
 960         }
 961
 962         pr_info("PV spinlocks enabled\n");
 963
 964         __pv_init_lock_hash();
 965         pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 966         pv_ops.lock.queued_spin_unlock =
 967                 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 968         pv_ops.lock.wait = kvm_wait;
 969         pv_ops.lock.kick = kvm_kick_cpu;
 970
 971         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 972                 pv_ops.lock.vcpu_is_preempted =
 973                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 974         }
 975         /*
 976          * When PV spinlock is enabled which is preferred over
 977          * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
 978          * Just disable it anyway.
 979          */
 980 out:
 981         static_branch_disable(&virt_spin_lock_key);
 982 }
 983
 984 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */
 985
 986 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
 987
 988 static void kvm_disable_host_haltpoll(void *i)
 989 {
 990         wrmsrl(MSR_KVM_POLL_CONTROL, 0);
 991 }
 992
 993 static void kvm_enable_host_haltpoll(void *i)
 994 {
 995         wrmsrl(MSR_KVM_POLL_CONTROL, 1);
 996 }
 997
 998 void arch_haltpoll_enable(unsigned int cpu)
 999 {
1000         if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1001                 pr_err_once("host does not support poll control\n");
1002                 pr_err_once("host upgrade recommended\n");
1003                 return;
1004         }
1005
1006         /* Enable guest halt poll disables host halt poll */
1007         smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1008 }
1009 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1010
1011 void arch_haltpoll_disable(unsigned int cpu)
1012 {
1013         if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1014                 return;
1015
1016         /* Disable guest halt poll enables host halt poll */
1017         smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1018 }
1019 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
1020 #endif