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
  40 DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
  41
  42 static int kvmapf = 1;
  43
  44 static int __init parse_no_kvmapf(char *arg)
  45 {
  46         kvmapf = 0;
  47         return 0;
  48 }
  49
  50 early_param("no-kvmapf", parse_no_kvmapf);
  51
  52 static int steal_acc = 1;
  53 static int __init parse_no_stealacc(char *arg)
  54 {
  55         steal_acc = 0;
  56         return 0;
  57 }
  58
  59 early_param("no-steal-acc", parse_no_stealacc);
  60
  61 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  62 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
  63 static int has_steal_clock = 0;
  64
  65 /*
  66  * No need for any "IO delay" on KVM
  67  */
  68 static void kvm_io_delay(void)
  69 {
  70 }
  71
  72 #define KVM_TASK_SLEEP_HASHBITS 8
  73 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  74
  75 struct kvm_task_sleep_node {
  76         struct hlist_node link;
  77         struct swait_queue_head wq;
  78         u32 token;
  79         int cpu;
  80 };
  81
  82 static struct kvm_task_sleep_head {
  83         raw_spinlock_t lock;
  84         struct hlist_head list;
  85 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
  86
  87 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
  88                                                   u32 token)
  89 {
  90         struct hlist_node *p;
  91
  92         hlist_for_each(p, &b->list) {
  93                 struct kvm_task_sleep_node *n =
  94                         hlist_entry(p, typeof(*n), link);
  95                 if (n->token == token)
  96                         return n;
  97         }
  98
  99         return NULL;
 100 }
 101
 102 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
 103 {
 104         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 105         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 106         struct kvm_task_sleep_node *e;
 107
 108         raw_spin_lock(&b->lock);
 109         e = _find_apf_task(b, token);
 110         if (e) {
 111                 /* dummy entry exist -> wake up was delivered ahead of PF */
 112                 hlist_del(&e->link);
 113                 raw_spin_unlock(&b->lock);
 114                 kfree(e);
 115                 return false;
 116         }
 117
 118         n->token = token;
 119         n->cpu = smp_processor_id();
 120         init_swait_queue_head(&n->wq);
 121         hlist_add_head(&n->link, &b->list);
 122         raw_spin_unlock(&b->lock);
 123         return true;
 124 }
 125
 126 /*
 127  * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
 128  * @token:      Token to identify the sleep node entry
 129  *
 130  * Invoked from the async pagefault handling code or from the VM exit page
 131  * fault handler. In both cases RCU is watching.
 132  */
 133 void kvm_async_pf_task_wait_schedule(u32 token)
 134 {
 135         struct kvm_task_sleep_node n;
 136         DECLARE_SWAITQUEUE(wait);
 137
 138         lockdep_assert_irqs_disabled();
 139
 140         if (!kvm_async_pf_queue_task(token, &n))
 141                 return;
 142
 143         for (;;) {
 144                 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 145                 if (hlist_unhashed(&n.link))
 146                         break;
 147
 148                 local_irq_enable();
 149                 schedule();
 150                 local_irq_disable();
 151         }
 152         finish_swait(&n.wq, &wait);
 153 }
 154 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
 155
 156 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 157 {
 158         hlist_del_init(&n->link);
 159         if (swq_has_sleeper(&n->wq))
 160                 swake_up_one(&n->wq);
 161 }
 162
 163 static void apf_task_wake_all(void)
 164 {
 165         int i;
 166
 167         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 168                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 169                 struct kvm_task_sleep_node *n;
 170                 struct hlist_node *p, *next;
 171
 172                 raw_spin_lock(&b->lock);
 173                 hlist_for_each_safe(p, next, &b->list) {
 174                         n = hlist_entry(p, typeof(*n), link);
 175                         if (n->cpu == smp_processor_id())
 176                                 apf_task_wake_one(n);
 177                 }
 178                 raw_spin_unlock(&b->lock);
 179         }
 180 }
 181
 182 void kvm_async_pf_task_wake(u32 token)
 183 {
 184         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 185         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 186         struct kvm_task_sleep_node *n;
 187
 188         if (token == ~0) {
 189                 apf_task_wake_all();
 190                 return;
 191         }
 192
 193 again:
 194         raw_spin_lock(&b->lock);
 195         n = _find_apf_task(b, token);
 196         if (!n) {
 197                 /*
 198                  * async PF was not yet handled.
 199                  * Add dummy entry for the token.
 200                  */
 201                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 202                 if (!n) {
 203                         /*
 204                          * Allocation failed! Busy wait while other cpu
 205                          * handles async PF.
 206                          */
 207                         raw_spin_unlock(&b->lock);
 208                         cpu_relax();
 209                         goto again;
 210                 }
 211                 n->token = token;
 212                 n->cpu = smp_processor_id();
 213                 init_swait_queue_head(&n->wq);
 214                 hlist_add_head(&n->link, &b->list);
 215         } else {
 216                 apf_task_wake_one(n);
 217         }
 218         raw_spin_unlock(&b->lock);
 219         return;
 220 }
 221 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 222
 223 noinstr u32 kvm_read_and_reset_apf_flags(void)
 224 {
 225         u32 flags = 0;
 226
 227         if (__this_cpu_read(apf_reason.enabled)) {
 228                 flags = __this_cpu_read(apf_reason.flags);
 229                 __this_cpu_write(apf_reason.flags, 0);
 230         }
 231
 232         return flags;
 233 }
 234 EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
 235
 236 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
 237 {
 238         u32 flags = kvm_read_and_reset_apf_flags();
 239         irqentry_state_t state;
 240
 241         if (!flags)
 242                 return false;
 243
 244         state = irqentry_enter(regs);
 245         instrumentation_begin();
 246
 247         /*
 248          * If the host managed to inject an async #PF into an interrupt
 249          * disabled region, then die hard as this is not going to end well
 250          * and the host side is seriously broken.
 251          */
 252         if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
 253                 panic("Host injected async #PF in interrupt disabled region\n");
 254
 255         if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
 256                 if (unlikely(!(user_mode(regs))))
 257                         panic("Host injected async #PF in kernel mode\n");
 258                 /* Page is swapped out by the host. */
 259                 kvm_async_pf_task_wait_schedule(token);
 260         } else {
 261                 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
 262         }
 263
 264         instrumentation_end();
 265         irqentry_exit(regs, state);
 266         return true;
 267 }
 268
 269 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
 270 {
 271         struct pt_regs *old_regs = set_irq_regs(regs);
 272         u32 token;
 273         irqentry_state_t state;
 274
 275         state = irqentry_enter(regs);
 276
 277         inc_irq_stat(irq_hv_callback_count);
 278
 279         if (__this_cpu_read(apf_reason.enabled)) {
 280                 token = __this_cpu_read(apf_reason.token);
 281                 kvm_async_pf_task_wake(token);
 282                 __this_cpu_write(apf_reason.token, 0);
 283                 wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
 284         }
 285
 286         irqentry_exit(regs, state);
 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.tlb_remove_table = tlb_remove_table;
 658                 pr_info("KVM setup pv remote TLB flush\n");
 659         }
 660
 661         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 662                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 663
 664         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
 665                 static_branch_enable(&kvm_async_pf_enabled);
 666                 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
 667         }
 668
 669 #ifdef CONFIG_SMP
 670         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 671         if (pv_sched_yield_supported()) {
 672                 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
 673                 pr_info("setup PV sched yield\n");
 674         }
 675         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 676                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 677                 pr_err("failed to install cpu hotplug callbacks\n");
 678 #else
 679         sev_map_percpu_data();
 680         kvm_guest_cpu_init();
 681 #endif
 682
 683         /*
 684          * Hard lockup detection is enabled by default. Disable it, as guests
 685          * can get false positives too easily, for example if the host is
 686          * overcommitted.
 687          */
 688         hardlockup_detector_disable();
 689 }
 690
 691 static noinline uint32_t __kvm_cpuid_base(void)
 692 {
 693         if (boot_cpu_data.cpuid_level < 0)
 694                 return 0;       /* So we don't blow up on old processors */
 695
 696         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 697                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 698
 699         return 0;
 700 }
 701
 702 static inline uint32_t kvm_cpuid_base(void)
 703 {
 704         static int kvm_cpuid_base = -1;
 705
 706         if (kvm_cpuid_base == -1)
 707                 kvm_cpuid_base = __kvm_cpuid_base();
 708
 709         return kvm_cpuid_base;
 710 }
 711
 712 bool kvm_para_available(void)
 713 {
 714         return kvm_cpuid_base() != 0;
 715 }
 716 EXPORT_SYMBOL_GPL(kvm_para_available);
 717
 718 unsigned int kvm_arch_para_features(void)
 719 {
 720         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 721 }
 722
 723 unsigned int kvm_arch_para_hints(void)
 724 {
 725         return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 726 }
 727 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
 728
 729 static uint32_t __init kvm_detect(void)
 730 {
 731         return kvm_cpuid_base();
 732 }
 733
 734 static void __init kvm_apic_init(void)
 735 {
 736 #if defined(CONFIG_SMP)
 737         if (pv_ipi_supported())
 738                 kvm_setup_pv_ipi();
 739 #endif
 740 }
 741
 742 static void __init kvm_init_platform(void)
 743 {
 744         kvmclock_init();
 745         x86_platform.apic_post_init = kvm_apic_init;
 746 }
 747
 748 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
 749         .name                   = "KVM",
 750         .detect                 = kvm_detect,
 751         .type                   = X86_HYPER_KVM,
 752         .init.guest_late_init   = kvm_guest_init,
 753         .init.x2apic_available  = kvm_para_available,
 754         .init.init_platform     = kvm_init_platform,
 755 };
 756
 757 static __init int activate_jump_labels(void)
 758 {
 759         if (has_steal_clock) {
 760                 static_key_slow_inc(&paravirt_steal_enabled);
 761                 if (steal_acc)
 762                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 763         }
 764
 765         return 0;
 766 }
 767 arch_initcall(activate_jump_labels);
 768
 769 static void kvm_free_pv_cpu_mask(void)
 770 {
 771         unsigned int cpu;
 772
 773         for_each_possible_cpu(cpu)
 774                 free_cpumask_var(per_cpu(__pv_cpu_mask, cpu));
 775 }
 776
 777 static __init int kvm_alloc_cpumask(void)
 778 {
 779         int cpu;
 780         bool alloc = false;
 781
 782         if (!kvm_para_available() || nopv)
 783                 return 0;
 784
 785         if (pv_tlb_flush_supported())
 786                 alloc = true;
 787
 788 #if defined(CONFIG_SMP)
 789         if (pv_ipi_supported())
 790                 alloc = true;
 791 #endif
 792
 793         if (alloc)
 794                 for_each_possible_cpu(cpu) {
 795                         if (!zalloc_cpumask_var_node(
 796                                 per_cpu_ptr(&__pv_cpu_mask, cpu),
 797                                 GFP_KERNEL, cpu_to_node(cpu))) {
 798                                 goto zalloc_cpumask_fail;
 799                         }
 800                 }
 801
 802         apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
 803         pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
 804         return 0;
 805
 806 zalloc_cpumask_fail:
 807         kvm_free_pv_cpu_mask();
 808         return -ENOMEM;
 809 }
 810 arch_initcall(kvm_alloc_cpumask);
 811
 812 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 813
 814 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 815 static void kvm_kick_cpu(int cpu)
 816 {
 817         int apicid;
 818         unsigned long flags = 0;
 819
 820         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 821         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 822 }
 823
 824 #include <asm/qspinlock.h>
 825
 826 static void kvm_wait(u8 *ptr, u8 val)
 827 {
 828         unsigned long flags;
 829
 830         if (in_nmi())
 831                 return;
 832
 833         local_irq_save(flags);
 834
 835         if (READ_ONCE(*ptr) != val)
 836                 goto out;
 837
 838         /*
 839          * halt until it's our turn and kicked. Note that we do safe halt
 840          * for irq enabled case to avoid hang when lock info is overwritten
 841          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 842          */
 843         if (arch_irqs_disabled_flags(flags))
 844                 halt();
 845         else
 846                 safe_halt();
 847
 848 out:
 849         local_irq_restore(flags);
 850 }
 851
 852 #ifdef CONFIG_X86_32
 853 __visible bool __kvm_vcpu_is_preempted(long cpu)
 854 {
 855         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 856
 857         return !!(src->preempted & KVM_VCPU_PREEMPTED);
 858 }
 859 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 860
 861 #else
 862
 863 #include <asm/asm-offsets.h>
 864
 865 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 866
 867 /*
 868  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 869  * restoring to/from the stack.
 870  */
 871 asm(
 872 ".pushsection .text;"
 873 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 874 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 875 "__raw_callee_save___kvm_vcpu_is_preempted:"
 876 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 877 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 878 "setne  %al;"
 879 "ret;"
 880 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
 881 ".popsection");
 882
 883 #endif
 884
 885 /*
 886  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 887  */
 888 void __init kvm_spinlock_init(void)
 889 {
 890         /*
 891          * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
 892          * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
 893          * preferred over native qspinlock when vCPU is preempted.
 894          */
 895         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
 896                 pr_info("PV spinlocks disabled, no host support\n");
 897                 return;
 898         }
 899
 900         /*
 901          * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
 902          * are available.
 903          */
 904         if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
 905                 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
 906                 goto out;
 907         }
 908
 909         if (num_possible_cpus() == 1) {
 910                 pr_info("PV spinlocks disabled, single CPU\n");
 911                 goto out;
 912         }
 913
 914         if (nopvspin) {
 915                 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
 916                 goto out;
 917         }
 918
 919         pr_info("PV spinlocks enabled\n");
 920
 921         __pv_init_lock_hash();
 922         pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 923         pv_ops.lock.queued_spin_unlock =
 924                 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 925         pv_ops.lock.wait = kvm_wait;
 926         pv_ops.lock.kick = kvm_kick_cpu;
 927
 928         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 929                 pv_ops.lock.vcpu_is_preempted =
 930                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 931         }
 932         /*
 933          * When PV spinlock is enabled which is preferred over
 934          * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
 935          * Just disable it anyway.
 936          */
 937 out:
 938         static_branch_disable(&virt_spin_lock_key);
 939 }
 940
 941 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */
 942
 943 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
 944
 945 static void kvm_disable_host_haltpoll(void *i)
 946 {
 947         wrmsrl(MSR_KVM_POLL_CONTROL, 0);
 948 }
 949
 950 static void kvm_enable_host_haltpoll(void *i)
 951 {
 952         wrmsrl(MSR_KVM_POLL_CONTROL, 1);
 953 }
 954
 955 void arch_haltpoll_enable(unsigned int cpu)
 956 {
 957         if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
 958                 pr_err_once("host does not support poll control\n");
 959                 pr_err_once("host upgrade recommended\n");
 960                 return;
 961         }
 962
 963         /* Enable guest halt poll disables host halt poll */
 964         smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
 965 }
 966 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
 967
 968 void arch_haltpoll_disable(unsigned int cpu)
 969 {
 970         if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
 971                 return;
 972
 973         /* Enable guest halt poll disables host halt poll */
 974         smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
 975 }
 976 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
 977 #endif