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