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