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