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 #include <linux/context_tracking.h>
  11 #include <linux/init.h>
  12 #include <linux/kernel.h>
  13 #include <linux/kvm_para.h>
  14 #include <linux/cpu.h>
  15 #include <linux/mm.h>
  16 #include <linux/highmem.h>
  17 #include <linux/hardirq.h>
  18 #include <linux/notifier.h>
  19 #include <linux/reboot.h>
  20 #include <linux/hash.h>
  21 #include <linux/sched.h>
  22 #include <linux/slab.h>
  23 #include <linux/kprobes.h>
  24 #include <linux/debugfs.h>
  25 #include <linux/nmi.h>
  26 #include <linux/swait.h>
  27 #include <linux/syscore_ops.h>
  28 #include <asm/timer.h>
  29 #include <asm/cpu.h>
  30 #include <asm/traps.h>
  31 #include <asm/desc.h>
  32 #include <asm/tlbflush.h>
  33 #include <asm/apic.h>
  34 #include <asm/apicdef.h>
  35 #include <asm/hypervisor.h>
  36 #include <asm/tlb.h>
  37 #include <asm/reboot.h>
  38
  39 static int kvmapf = 1;
  40
  41 static int __init parse_no_kvmapf(char *arg)
  42 {
  43         kvmapf = 0;
  44         return 0;
  45 }
  46
  47 early_param("no-kvmapf", parse_no_kvmapf);
  48
  49 static int steal_acc = 1;
  50 static int __init parse_no_stealacc(char *arg)
  51 {
  52         steal_acc = 0;
  53         return 0;
  54 }
  55
  56 early_param("no-steal-acc", parse_no_stealacc);
  57
  58 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  59 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
  60 static int has_steal_clock = 0;
  61
  62 static int has_guest_poll = 0;
  63 /*
  64  * No need for any "IO delay" on KVM
  65  */
  66 static void kvm_io_delay(void)
  67 {
  68 }
  69
  70 #define KVM_TASK_SLEEP_HASHBITS 8
  71 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  72
  73 struct kvm_task_sleep_node {
  74         struct hlist_node link;
  75         struct swait_queue_head wq;
  76         u32 token;
  77         int cpu;
  78         bool halted;
  79 };
  80
  81 static struct kvm_task_sleep_head {
  82         raw_spinlock_t lock;
  83         struct hlist_head list;
  84 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
  85
  86 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
  87                                                   u32 token)
  88 {
  89         struct hlist_node *p;
  90
  91         hlist_for_each(p, &b->list) {
  92                 struct kvm_task_sleep_node *n =
  93                         hlist_entry(p, typeof(*n), link);
  94                 if (n->token == token)
  95                         return n;
  96         }
  97
  98         return NULL;
  99 }
 100
 101 /*
 102  * @interrupt_kernel: Is this called from a routine which interrupts the kernel
 103  *                    (other than user space)?
 104  */
 105 void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
 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 n, *e;
 110         DECLARE_SWAITQUEUE(wait);
 111
 112         rcu_irq_enter();
 113
 114         raw_spin_lock(&b->lock);
 115         e = _find_apf_task(b, token);
 116         if (e) {
 117                 /* dummy entry exist -> wake up was delivered ahead of PF */
 118                 hlist_del(&e->link);
 119                 kfree(e);
 120                 raw_spin_unlock(&b->lock);
 121
 122                 rcu_irq_exit();
 123                 return;
 124         }
 125
 126         n.token = token;
 127         n.cpu = smp_processor_id();
 128         n.halted = is_idle_task(current) ||
 129                    (IS_ENABLED(CONFIG_PREEMPT_COUNT)
 130                     ? preempt_count() > 1 || rcu_preempt_depth()
 131                     : interrupt_kernel);
 132         init_swait_queue_head(&n.wq);
 133         hlist_add_head(&n.link, &b->list);
 134         raw_spin_unlock(&b->lock);
 135
 136         for (;;) {
 137                 if (!n.halted)
 138                         prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 139                 if (hlist_unhashed(&n.link))
 140                         break;
 141
 142                 rcu_irq_exit();
 143
 144                 if (!n.halted) {
 145                         local_irq_enable();
 146                         schedule();
 147                         local_irq_disable();
 148                 } else {
 149                         /*
 150                          * We cannot reschedule. So halt.
 151                          */
 152                         native_safe_halt();
 153                         local_irq_disable();
 154                 }
 155
 156                 rcu_irq_enter();
 157         }
 158         if (!n.halted)
 159                 finish_swait(&n.wq, &wait);
 160
 161         rcu_irq_exit();
 162         return;
 163 }
 164 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 165
 166 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 167 {
 168         hlist_del_init(&n->link);
 169         if (n->halted)
 170                 smp_send_reschedule(n->cpu);
 171         else if (swq_has_sleeper(&n->wq))
 172                 swake_up_one(&n->wq);
 173 }
 174
 175 static void apf_task_wake_all(void)
 176 {
 177         int i;
 178
 179         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 180                 struct hlist_node *p, *next;
 181                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 182                 raw_spin_lock(&b->lock);
 183                 hlist_for_each_safe(p, next, &b->list) {
 184                         struct kvm_task_sleep_node *n =
 185                                 hlist_entry(p, typeof(*n), link);
 186                         if (n->cpu == smp_processor_id())
 187                                 apf_task_wake_one(n);
 188                 }
 189                 raw_spin_unlock(&b->lock);
 190         }
 191 }
 192
 193 void kvm_async_pf_task_wake(u32 token)
 194 {
 195         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 196         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 197         struct kvm_task_sleep_node *n;
 198
 199         if (token == ~0) {
 200                 apf_task_wake_all();
 201                 return;
 202         }
 203
 204 again:
 205         raw_spin_lock(&b->lock);
 206         n = _find_apf_task(b, token);
 207         if (!n) {
 208                 /*
 209                  * async PF was not yet handled.
 210                  * Add dummy entry for the token.
 211                  */
 212                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 213                 if (!n) {
 214                         /*
 215                          * Allocation failed! Busy wait while other cpu
 216                          * handles async PF.
 217                          */
 218                         raw_spin_unlock(&b->lock);
 219                         cpu_relax();
 220                         goto again;
 221                 }
 222                 n->token = token;
 223                 n->cpu = smp_processor_id();
 224                 init_swait_queue_head(&n->wq);
 225                 hlist_add_head(&n->link, &b->list);
 226         } else
 227                 apf_task_wake_one(n);
 228         raw_spin_unlock(&b->lock);
 229         return;
 230 }
 231 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 232
 233 u32 kvm_read_and_reset_pf_reason(void)
 234 {
 235         u32 reason = 0;
 236
 237         if (__this_cpu_read(apf_reason.enabled)) {
 238                 reason = __this_cpu_read(apf_reason.reason);
 239                 __this_cpu_write(apf_reason.reason, 0);
 240         }
 241
 242         return reason;
 243 }
 244 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 245 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 246
 247 dotraplinkage void
 248 do_async_page_fault(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 249 {
 250         enum ctx_state prev_state;
 251
 252         switch (kvm_read_and_reset_pf_reason()) {
 253         default:
 254                 do_page_fault(regs, error_code, address);
 255                 break;
 256         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 257                 /* page is swapped out by the host. */
 258                 prev_state = exception_enter();
 259                 kvm_async_pf_task_wait((u32)address, !user_mode(regs));
 260                 exception_exit(prev_state);
 261                 break;
 262         case KVM_PV_REASON_PAGE_READY:
 263                 rcu_irq_enter();
 264                 kvm_async_pf_task_wake((u32)address);
 265                 rcu_irq_exit();
 266                 break;
 267         }
 268 }
 269 NOKPROBE_SYMBOL(do_async_page_fault);
 270
 271 static void __init paravirt_ops_setup(void)
 272 {
 273         pv_info.name = "KVM";
 274
 275         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 276                 pv_ops.cpu.io_delay = kvm_io_delay;
 277
 278 #ifdef CONFIG_X86_IO_APIC
 279         no_timer_check = 1;
 280 #endif
 281 }
 282
 283 static void kvm_register_steal_time(void)
 284 {
 285         int cpu = smp_processor_id();
 286         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 287
 288         if (!has_steal_clock)
 289                 return;
 290
 291         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 292         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
 293                 cpu, (unsigned long long) slow_virt_to_phys(st));
 294 }
 295
 296 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 297
 298 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 299 {
 300         /**
 301          * This relies on __test_and_clear_bit to modify the memory
 302          * in a way that is atomic with respect to the local CPU.
 303          * The hypervisor only accesses this memory from the local CPU so
 304          * there's no need for lock or memory barriers.
 305          * An optimization barrier is implied in apic write.
 306          */
 307         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 308                 return;
 309         apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
 310 }
 311
 312 static void kvm_guest_cpu_init(void)
 313 {
 314         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 315                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 316
 317 #ifdef CONFIG_PREEMPTION
 318                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 319 #endif
 320                 pa |= KVM_ASYNC_PF_ENABLED;
 321
 322                 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
 323                         pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
 324
 325                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 326                 __this_cpu_write(apf_reason.enabled, 1);
 327                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 328                        smp_processor_id());
 329         }
 330
 331         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 332                 unsigned long pa;
 333                 /* Size alignment is implied but just to make it explicit. */
 334                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 335                 __this_cpu_write(kvm_apic_eoi, 0);
 336                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 337                         | KVM_MSR_ENABLED;
 338                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 339         }
 340
 341         if (has_steal_clock)
 342                 kvm_register_steal_time();
 343 }
 344
 345 static void kvm_pv_disable_apf(void)
 346 {
 347         if (!__this_cpu_read(apf_reason.enabled))
 348                 return;
 349
 350         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 351         __this_cpu_write(apf_reason.enabled, 0);
 352
 353         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 354                smp_processor_id());
 355 }
 356
 357 static void kvm_disable_steal_time(void)
 358 {
 359         if (!has_steal_clock)
 360                 return;
 361
 362         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 363 }
 364
 365 static void kvm_pv_guest_cpu_reboot(void *unused)
 366 {
 367         /*
 368          * We disable PV EOI before we load a new kernel by kexec,
 369          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 370          * New kernel can re-enable when it boots.
 371          */
 372         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 373                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 374         kvm_pv_disable_apf();
 375         kvm_disable_steal_time();
 376 }
 377
 378 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 379                                 unsigned long code, void *unused)
 380 {
 381         if (code == SYS_RESTART)
 382                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 383         return NOTIFY_DONE;
 384 }
 385
 386 static struct notifier_block kvm_pv_reboot_nb = {
 387         .notifier_call = kvm_pv_reboot_notify,
 388 };
 389
 390 static u64 kvm_steal_clock(int cpu)
 391 {
 392         u64 steal;
 393         struct kvm_steal_time *src;
 394         int version;
 395
 396         src = &per_cpu(steal_time, cpu);
 397         do {
 398                 version = src->version;
 399                 virt_rmb();
 400                 steal = src->steal;
 401                 virt_rmb();
 402         } while ((version & 1) || (version != src->version));
 403
 404         return steal;
 405 }
 406
 407 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
 408 {
 409         early_set_memory_decrypted((unsigned long) ptr, size);
 410 }
 411
 412 /*
 413  * Iterate through all possible CPUs and map the memory region pointed
 414  * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
 415  *
 416  * Note: we iterate through all possible CPUs to ensure that CPUs
 417  * hotplugged will have their per-cpu variable already mapped as
 418  * decrypted.
 419  */
 420 static void __init sev_map_percpu_data(void)
 421 {
 422         int cpu;
 423
 424         if (!sev_active())
 425                 return;
 426
 427         for_each_possible_cpu(cpu) {
 428                 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
 429                 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
 430                 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
 431         }
 432 }
 433
 434 static void kvm_guest_cpu_offline(bool shutdown)
 435 {
 436         kvm_disable_steal_time();
 437         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 438                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 439         kvm_pv_disable_apf();
 440         if (!shutdown)
 441                 apf_task_wake_all();
 442         kvmclock_disable();
 443 }
 444
 445 static int kvm_cpu_online(unsigned int cpu)
 446 {
 447         unsigned long flags;
 448
 449         local_irq_save(flags);
 450         kvm_guest_cpu_init();
 451         local_irq_restore(flags);
 452         return 0;
 453 }
 454
 455 #ifdef CONFIG_SMP
 456 #define KVM_IPI_CLUSTER_SIZE    (2 * BITS_PER_LONG)
 457
 458 static void __send_ipi_mask(const struct cpumask *mask, int vector)
 459 {
 460         unsigned long flags;
 461         int cpu, apic_id, icr;
 462         int min = 0, max = 0;
 463 #ifdef CONFIG_X86_64
 464         __uint128_t ipi_bitmap = 0;
 465 #else
 466         u64 ipi_bitmap = 0;
 467 #endif
 468         long ret;
 469
 470         if (cpumask_empty(mask))
 471                 return;
 472
 473         local_irq_save(flags);
 474
 475         switch (vector) {
 476         default:
 477                 icr = APIC_DM_FIXED | vector;
 478                 break;
 479         case NMI_VECTOR:
 480                 icr = APIC_DM_NMI;
 481                 break;
 482         }
 483
 484         for_each_cpu(cpu, mask) {
 485                 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
 486                 if (!ipi_bitmap) {
 487                         min = max = apic_id;
 488                 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
 489                         ipi_bitmap <<= min - apic_id;
 490                         min = apic_id;
 491                 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
 492                         max = apic_id < max ? max : apic_id;
 493                 } else {
 494                         ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 495                                 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 496                         WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
 497                         min = max = apic_id;
 498                         ipi_bitmap = 0;
 499                 }
 500                 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
 501         }
 502
 503         if (ipi_bitmap) {
 504                 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 505                         (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 506                 WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
 507         }
 508
 509         local_irq_restore(flags);
 510 }
 511
 512 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
 513 {
 514         __send_ipi_mask(mask, vector);
 515 }
 516
 517 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
 518 {
 519         unsigned int this_cpu = smp_processor_id();
 520         struct cpumask new_mask;
 521         const struct cpumask *local_mask;
 522
 523         cpumask_copy(&new_mask, mask);
 524         cpumask_clear_cpu(this_cpu, &new_mask);
 525         local_mask = &new_mask;
 526         __send_ipi_mask(local_mask, vector);
 527 }
 528
 529 /*
 530  * Set the IPI entry points
 531  */
 532 static void kvm_setup_pv_ipi(void)
 533 {
 534         apic->send_IPI_mask = kvm_send_ipi_mask;
 535         apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
 536         pr_info("KVM setup pv IPIs\n");
 537 }
 538
 539 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
 540 {
 541         int cpu;
 542
 543         native_send_call_func_ipi(mask);
 544
 545         /* Make sure other vCPUs get a chance to run if they need to. */
 546         for_each_cpu(cpu, mask) {
 547                 if (vcpu_is_preempted(cpu)) {
 548                         kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
 549                         break;
 550                 }
 551         }
 552 }
 553
 554 static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
 555 {
 556         native_smp_prepare_cpus(max_cpus);
 557         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
 558                 static_branch_disable(&virt_spin_lock_key);
 559 }
 560
 561 static void __init kvm_smp_prepare_boot_cpu(void)
 562 {
 563         /*
 564          * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
 565          * shares the guest physical address with the hypervisor.
 566          */
 567         sev_map_percpu_data();
 568
 569         kvm_guest_cpu_init();
 570         native_smp_prepare_boot_cpu();
 571         kvm_spinlock_init();
 572 }
 573
 574 static int kvm_cpu_down_prepare(unsigned int cpu)
 575 {
 576         unsigned long flags;
 577
 578         local_irq_save(flags);
 579         kvm_guest_cpu_offline(false);
 580         local_irq_restore(flags);
 581         return 0;
 582 }
 583
 584 #endif
 585
 586 static int kvm_suspend(void)
 587 {
 588         u64 val = 0;
 589
 590         kvm_guest_cpu_offline(false);
 591
 592 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
 593         if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
 594                 rdmsrl(MSR_KVM_POLL_CONTROL, val);
 595         has_guest_poll = !(val & 1);
 596 #endif
 597         return 0;
 598 }
 599
 600 static void kvm_resume(void)
 601 {
 602         kvm_cpu_online(raw_smp_processor_id());
 603
 604 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
 605         if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
 606                 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
 607 #endif
 608 }
 609
 610 static struct syscore_ops kvm_syscore_ops = {
 611         .suspend        = kvm_suspend,
 612         .resume         = kvm_resume,
 613 };
 614
 615 /*
 616  * After a PV feature is registered, the host will keep writing to the
 617  * registered memory location. If the guest happens to shutdown, this memory
 618  * won't be valid. In cases like kexec, in which you install a new kernel, this
 619  * means a random memory location will be kept being written.
 620  */
 621 #ifdef CONFIG_KEXEC_CORE
 622 static void kvm_crash_shutdown(struct pt_regs *regs)
 623 {
 624         kvm_guest_cpu_offline(true);
 625         native_machine_crash_shutdown(regs);
 626 }
 627 #endif
 628
 629 static void __init kvm_apf_trap_init(void)
 630 {
 631         update_intr_gate(X86_TRAP_PF, async_page_fault);
 632 }
 633
 634 static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
 635
 636 static void kvm_flush_tlb_others(const struct cpumask *cpumask,
 637                         const struct flush_tlb_info *info)
 638 {
 639         u8 state;
 640         int cpu;
 641         struct kvm_steal_time *src;
 642         struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
 643
 644         cpumask_copy(flushmask, cpumask);
 645         /*
 646          * We have to call flush only on online vCPUs. And
 647          * queue flush_on_enter for pre-empted vCPUs
 648          */
 649         for_each_cpu(cpu, flushmask) {
 650                 src = &per_cpu(steal_time, cpu);
 651                 state = READ_ONCE(src->preempted);
 652                 if ((state & KVM_VCPU_PREEMPTED)) {
 653                         if (try_cmpxchg(&src->preempted, &state,
 654                                         state | KVM_VCPU_FLUSH_TLB))
 655                                 __cpumask_clear_cpu(cpu, flushmask);
 656                 }
 657         }
 658
 659         native_flush_tlb_others(flushmask, info);
 660 }
 661
 662 static void __init kvm_guest_init(void)
 663 {
 664         int i;
 665
 666         paravirt_ops_setup();
 667         register_reboot_notifier(&kvm_pv_reboot_nb);
 668         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 669                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 670         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 671                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 672
 673         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 674                 has_steal_clock = 1;
 675                 pv_ops.time.steal_clock = kvm_steal_clock;
 676         }
 677
 678         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 679             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 680             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 681                 pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
 682                 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
 683         }
 684
 685         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 686                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 687
 688 #ifdef CONFIG_SMP
 689         smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
 690         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 691         if (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
 692             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 693             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 694                 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
 695                 pr_info("KVM setup pv sched yield\n");
 696         }
 697         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 698                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 699                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 700 #else
 701         sev_map_percpu_data();
 702         kvm_guest_cpu_init();
 703 #endif
 704
 705 #ifdef CONFIG_KEXEC_CORE
 706         machine_ops.crash_shutdown = kvm_crash_shutdown;
 707 #endif
 708
 709         register_syscore_ops(&kvm_syscore_ops);
 710
 711         /*
 712          * Hard lockup detection is enabled by default. Disable it, as guests
 713          * can get false positives too easily, for example if the host is
 714          * overcommitted.
 715          */
 716         hardlockup_detector_disable();
 717 }
 718
 719 static noinline uint32_t __kvm_cpuid_base(void)
 720 {
 721         if (boot_cpu_data.cpuid_level < 0)
 722                 return 0;       /* So we don't blow up on old processors */
 723
 724         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 725                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 726
 727         return 0;
 728 }
 729
 730 static inline uint32_t kvm_cpuid_base(void)
 731 {
 732         static int kvm_cpuid_base = -1;
 733
 734         if (kvm_cpuid_base == -1)
 735                 kvm_cpuid_base = __kvm_cpuid_base();
 736
 737         return kvm_cpuid_base;
 738 }
 739
 740 bool kvm_para_available(void)
 741 {
 742         return kvm_cpuid_base() != 0;
 743 }
 744 EXPORT_SYMBOL_GPL(kvm_para_available);
 745
 746 unsigned int kvm_arch_para_features(void)
 747 {
 748         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 749 }
 750
 751 unsigned int kvm_arch_para_hints(void)
 752 {
 753         return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 754 }
 755 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
 756
 757 static uint32_t __init kvm_detect(void)
 758 {
 759         return kvm_cpuid_base();
 760 }
 761
 762 static void __init kvm_apic_init(void)
 763 {
 764 #if defined(CONFIG_SMP)
 765         if (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI))
 766                 kvm_setup_pv_ipi();
 767 #endif
 768 }
 769
 770 static void __init kvm_init_platform(void)
 771 {
 772         kvmclock_init();
 773         x86_platform.apic_post_init = kvm_apic_init;
 774 }
 775
 776 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
 777         .name                   = "KVM",
 778         .detect                 = kvm_detect,
 779         .type                   = X86_HYPER_KVM,
 780         .init.guest_late_init   = kvm_guest_init,
 781         .init.x2apic_available  = kvm_para_available,
 782         .init.init_platform     = kvm_init_platform,
 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_setup_pv_tlb_flush(void)
 798 {
 799         int cpu;
 800
 801         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 802             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 803             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 804                 for_each_possible_cpu(cpu) {
 805                         zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
 806                                 GFP_KERNEL, cpu_to_node(cpu));
 807                 }
 808                 pr_info("KVM setup pv remote TLB flush\n");
 809         }
 810
 811         return 0;
 812 }
 813 arch_initcall(kvm_setup_pv_tlb_flush);
 814
 815 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 816
 817 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 818 static void kvm_kick_cpu(int cpu)
 819 {
 820         int apicid;
 821         unsigned long flags = 0;
 822
 823         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 824         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 825 }
 826
 827 #include <asm/qspinlock.h>
 828
 829 static void kvm_wait(u8 *ptr, u8 val)
 830 {
 831         unsigned long flags;
 832
 833         if (in_nmi())
 834                 return;
 835
 836         local_irq_save(flags);
 837
 838         if (READ_ONCE(*ptr) != val)
 839                 goto out;
 840
 841         /*
 842          * halt until it's our turn and kicked. Note that we do safe halt
 843          * for irq enabled case to avoid hang when lock info is overwritten
 844          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 845          */
 846         if (arch_irqs_disabled_flags(flags))
 847                 halt();
 848         else
 849                 safe_halt();
 850
 851 out:
 852         local_irq_restore(flags);
 853 }
 854
 855 #ifdef CONFIG_X86_32
 856 __visible bool __kvm_vcpu_is_preempted(long cpu)
 857 {
 858         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 859
 860         return !!(src->preempted & KVM_VCPU_PREEMPTED);
 861 }
 862 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 863
 864 #else
 865
 866 #include <asm/asm-offsets.h>
 867
 868 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 869
 870 /*
 871  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 872  * restoring to/from the stack.
 873  */
 874 asm(
 875 ".pushsection .text;"
 876 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 877 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 878 "__raw_callee_save___kvm_vcpu_is_preempted:"
 879 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 880 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 881 "setne  %al;"
 882 "ret;"
 883 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
 884 ".popsection");
 885
 886 #endif
 887
 888 /*
 889  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 890  */
 891 void __init kvm_spinlock_init(void)
 892 {
 893         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 894         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 895                 return;
 896
 897         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
 898                 return;
 899
 900         /* Don't use the pvqspinlock code if there is only 1 vCPU. */
 901         if (num_possible_cpus() == 1)
 902                 return;
 903
 904         __pv_init_lock_hash();
 905         pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 906         pv_ops.lock.queued_spin_unlock =
 907                 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 908         pv_ops.lock.wait = kvm_wait;
 909         pv_ops.lock.kick = kvm_kick_cpu;
 910
 911         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 912                 pv_ops.lock.vcpu_is_preempted =
 913                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 914         }
 915 }
 916
 917 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */
 918
 919 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
 920
 921 static void kvm_disable_host_haltpoll(void *i)
 922 {
 923         wrmsrl(MSR_KVM_POLL_CONTROL, 0);
 924 }
 925
 926 static void kvm_enable_host_haltpoll(void *i)
 927 {
 928         wrmsrl(MSR_KVM_POLL_CONTROL, 1);
 929 }
 930
 931 void arch_haltpoll_enable(unsigned int cpu)
 932 {
 933         if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
 934                 pr_err_once("kvm: host does not support poll control\n");
 935                 pr_err_once("kvm: host upgrade recommended\n");
 936                 return;
 937         }
 938
 939         /* Enable guest halt poll disables host halt poll */
 940         smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
 941 }
 942 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
 943
 944 void arch_haltpoll_disable(unsigned int cpu)
 945 {
 946         if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
 947                 return;
 948
 949         /* Enable guest halt poll disables host halt poll */
 950         smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
 951 }
 952 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);
 953 #endif