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 <asm/timer.h>
  28 #include <asm/cpu.h>
  29 #include <asm/traps.h>
  30 #include <asm/desc.h>
  31 #include <asm/tlbflush.h>
  32 #include <asm/apic.h>
  33 #include <asm/apicdef.h>
  34 #include <asm/hypervisor.h>
  35 #include <asm/tlb.h>
  36
  37 static int kvmapf = 1;
  38
  39 static int __init parse_no_kvmapf(char *arg)
  40 {
  41         kvmapf = 0;
  42         return 0;
  43 }
  44
  45 early_param("no-kvmapf", parse_no_kvmapf);
  46
  47 static int steal_acc = 1;
  48 static int __init parse_no_stealacc(char *arg)
  49 {
  50         steal_acc = 0;
  51         return 0;
  52 }
  53
  54 early_param("no-steal-acc", parse_no_stealacc);
  55
  56 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  57 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
  58 static int has_steal_clock = 0;
  59
  60 /*
  61  * No need for any "IO delay" on KVM
  62  */
  63 static void kvm_io_delay(void)
  64 {
  65 }
  66
  67 #define KVM_TASK_SLEEP_HASHBITS 8
  68 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  69
  70 struct kvm_task_sleep_node {
  71         struct hlist_node link;
  72         struct swait_queue_head wq;
  73         u32 token;
  74         int cpu;
  75         bool halted;
  76 };
  77
  78 static struct kvm_task_sleep_head {
  79         raw_spinlock_t lock;
  80         struct hlist_head list;
  81 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
  82
  83 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
  84                                                   u32 token)
  85 {
  86         struct hlist_node *p;
  87
  88         hlist_for_each(p, &b->list) {
  89                 struct kvm_task_sleep_node *n =
  90                         hlist_entry(p, typeof(*n), link);
  91                 if (n->token == token)
  92                         return n;
  93         }
  94
  95         return NULL;
  96 }
  97
  98 /*
  99  * @interrupt_kernel: Is this called from a routine which interrupts the kernel
 100  *                    (other than user space)?
 101  */
 102 void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
 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 n, *e;
 107         DECLARE_SWAITQUEUE(wait);
 108
 109         rcu_irq_enter();
 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                 kfree(e);
 117                 raw_spin_unlock(&b->lock);
 118
 119                 rcu_irq_exit();
 120                 return;
 121         }
 122
 123         n.token = token;
 124         n.cpu = smp_processor_id();
 125         n.halted = is_idle_task(current) ||
 126                    (IS_ENABLED(CONFIG_PREEMPT_COUNT)
 127                     ? preempt_count() > 1 || rcu_preempt_depth()
 128                     : interrupt_kernel);
 129         init_swait_queue_head(&n.wq);
 130         hlist_add_head(&n.link, &b->list);
 131         raw_spin_unlock(&b->lock);
 132
 133         for (;;) {
 134                 if (!n.halted)
 135                         prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 136                 if (hlist_unhashed(&n.link))
 137                         break;
 138
 139                 rcu_irq_exit();
 140
 141                 if (!n.halted) {
 142                         local_irq_enable();
 143                         schedule();
 144                         local_irq_disable();
 145                 } else {
 146                         /*
 147                          * We cannot reschedule. So halt.
 148                          */
 149                         native_safe_halt();
 150                         local_irq_disable();
 151                 }
 152
 153                 rcu_irq_enter();
 154         }
 155         if (!n.halted)
 156                 finish_swait(&n.wq, &wait);
 157
 158         rcu_irq_exit();
 159         return;
 160 }
 161 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 162
 163 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 164 {
 165         hlist_del_init(&n->link);
 166         if (n->halted)
 167                 smp_send_reschedule(n->cpu);
 168         else if (swq_has_sleeper(&n->wq))
 169                 swake_up_one(&n->wq);
 170 }
 171
 172 static void apf_task_wake_all(void)
 173 {
 174         int i;
 175
 176         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 177                 struct hlist_node *p, *next;
 178                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 179                 raw_spin_lock(&b->lock);
 180                 hlist_for_each_safe(p, next, &b->list) {
 181                         struct kvm_task_sleep_node *n =
 182                                 hlist_entry(p, typeof(*n), link);
 183                         if (n->cpu == smp_processor_id())
 184                                 apf_task_wake_one(n);
 185                 }
 186                 raw_spin_unlock(&b->lock);
 187         }
 188 }
 189
 190 void kvm_async_pf_task_wake(u32 token)
 191 {
 192         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 193         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 194         struct kvm_task_sleep_node *n;
 195
 196         if (token == ~0) {
 197                 apf_task_wake_all();
 198                 return;
 199         }
 200
 201 again:
 202         raw_spin_lock(&b->lock);
 203         n = _find_apf_task(b, token);
 204         if (!n) {
 205                 /*
 206                  * async PF was not yet handled.
 207                  * Add dummy entry for the token.
 208                  */
 209                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 210                 if (!n) {
 211                         /*
 212                          * Allocation failed! Busy wait while other cpu
 213                          * handles async PF.
 214                          */
 215                         raw_spin_unlock(&b->lock);
 216                         cpu_relax();
 217                         goto again;
 218                 }
 219                 n->token = token;
 220                 n->cpu = smp_processor_id();
 221                 init_swait_queue_head(&n->wq);
 222                 hlist_add_head(&n->link, &b->list);
 223         } else
 224                 apf_task_wake_one(n);
 225         raw_spin_unlock(&b->lock);
 226         return;
 227 }
 228 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 229
 230 u32 kvm_read_and_reset_pf_reason(void)
 231 {
 232         u32 reason = 0;
 233
 234         if (__this_cpu_read(apf_reason.enabled)) {
 235                 reason = __this_cpu_read(apf_reason.reason);
 236                 __this_cpu_write(apf_reason.reason, 0);
 237         }
 238
 239         return reason;
 240 }
 241 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 242 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 243
 244 dotraplinkage void
 245 do_async_page_fault(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 246 {
 247         enum ctx_state prev_state;
 248
 249         switch (kvm_read_and_reset_pf_reason()) {
 250         default:
 251                 do_page_fault(regs, error_code, address);
 252                 break;
 253         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 254                 /* page is swapped out by the host. */
 255                 prev_state = exception_enter();
 256                 kvm_async_pf_task_wait((u32)address, !user_mode(regs));
 257                 exception_exit(prev_state);
 258                 break;
 259         case KVM_PV_REASON_PAGE_READY:
 260                 rcu_irq_enter();
 261                 kvm_async_pf_task_wake((u32)address);
 262                 rcu_irq_exit();
 263                 break;
 264         }
 265 }
 266 NOKPROBE_SYMBOL(do_async_page_fault);
 267
 268 static void __init paravirt_ops_setup(void)
 269 {
 270         pv_info.name = "KVM";
 271
 272         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 273                 pv_ops.cpu.io_delay = kvm_io_delay;
 274
 275 #ifdef CONFIG_X86_IO_APIC
 276         no_timer_check = 1;
 277 #endif
 278 }
 279
 280 static void kvm_register_steal_time(void)
 281 {
 282         int cpu = smp_processor_id();
 283         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 284
 285         if (!has_steal_clock)
 286                 return;
 287
 288         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 289         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
 290                 cpu, (unsigned long long) slow_virt_to_phys(st));
 291 }
 292
 293 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 294
 295 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 296 {
 297         /**
 298          * This relies on __test_and_clear_bit to modify the memory
 299          * in a way that is atomic with respect to the local CPU.
 300          * The hypervisor only accesses this memory from the local CPU so
 301          * there's no need for lock or memory barriers.
 302          * An optimization barrier is implied in apic write.
 303          */
 304         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 305                 return;
 306         apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
 307 }
 308
 309 static void kvm_guest_cpu_init(void)
 310 {
 311         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 312                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 313
 314 #ifdef CONFIG_PREEMPT
 315                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 316 #endif
 317                 pa |= KVM_ASYNC_PF_ENABLED;
 318
 319                 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
 320                         pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
 321
 322                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 323                 __this_cpu_write(apf_reason.enabled, 1);
 324                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 325                        smp_processor_id());
 326         }
 327
 328         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 329                 unsigned long pa;
 330                 /* Size alignment is implied but just to make it explicit. */
 331                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 332                 __this_cpu_write(kvm_apic_eoi, 0);
 333                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 334                         | KVM_MSR_ENABLED;
 335                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 336         }
 337
 338         if (has_steal_clock)
 339                 kvm_register_steal_time();
 340 }
 341
 342 static void kvm_pv_disable_apf(void)
 343 {
 344         if (!__this_cpu_read(apf_reason.enabled))
 345                 return;
 346
 347         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 348         __this_cpu_write(apf_reason.enabled, 0);
 349
 350         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 351                smp_processor_id());
 352 }
 353
 354 static void kvm_pv_guest_cpu_reboot(void *unused)
 355 {
 356         /*
 357          * We disable PV EOI before we load a new kernel by kexec,
 358          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 359          * New kernel can re-enable when it boots.
 360          */
 361         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 362                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 363         kvm_pv_disable_apf();
 364         kvm_disable_steal_time();
 365 }
 366
 367 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 368                                 unsigned long code, void *unused)
 369 {
 370         if (code == SYS_RESTART)
 371                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 372         return NOTIFY_DONE;
 373 }
 374
 375 static struct notifier_block kvm_pv_reboot_nb = {
 376         .notifier_call = kvm_pv_reboot_notify,
 377 };
 378
 379 static u64 kvm_steal_clock(int cpu)
 380 {
 381         u64 steal;
 382         struct kvm_steal_time *src;
 383         int version;
 384
 385         src = &per_cpu(steal_time, cpu);
 386         do {
 387                 version = src->version;
 388                 virt_rmb();
 389                 steal = src->steal;
 390                 virt_rmb();
 391         } while ((version & 1) || (version != src->version));
 392
 393         return steal;
 394 }
 395
 396 void kvm_disable_steal_time(void)
 397 {
 398         if (!has_steal_clock)
 399                 return;
 400
 401         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 402 }
 403
 404 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
 405 {
 406         early_set_memory_decrypted((unsigned long) ptr, size);
 407 }
 408
 409 /*
 410  * Iterate through all possible CPUs and map the memory region pointed
 411  * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
 412  *
 413  * Note: we iterate through all possible CPUs to ensure that CPUs
 414  * hotplugged will have their per-cpu variable already mapped as
 415  * decrypted.
 416  */
 417 static void __init sev_map_percpu_data(void)
 418 {
 419         int cpu;
 420
 421         if (!sev_active())
 422                 return;
 423
 424         for_each_possible_cpu(cpu) {
 425                 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
 426                 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
 427                 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
 428         }
 429 }
 430
 431 #ifdef CONFIG_SMP
 432 #define KVM_IPI_CLUSTER_SIZE    (2 * BITS_PER_LONG)
 433
 434 static void __send_ipi_mask(const struct cpumask *mask, int vector)
 435 {
 436         unsigned long flags;
 437         int cpu, apic_id, icr;
 438         int min = 0, max = 0;
 439 #ifdef CONFIG_X86_64
 440         __uint128_t ipi_bitmap = 0;
 441 #else
 442         u64 ipi_bitmap = 0;
 443 #endif
 444         long ret;
 445
 446         if (cpumask_empty(mask))
 447                 return;
 448
 449         local_irq_save(flags);
 450
 451         switch (vector) {
 452         default:
 453                 icr = APIC_DM_FIXED | vector;
 454                 break;
 455         case NMI_VECTOR:
 456                 icr = APIC_DM_NMI;
 457                 break;
 458         }
 459
 460         for_each_cpu(cpu, mask) {
 461                 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
 462                 if (!ipi_bitmap) {
 463                         min = max = apic_id;
 464                 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
 465                         ipi_bitmap <<= min - apic_id;
 466                         min = apic_id;
 467                 } else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
 468                         max = apic_id < max ? max : apic_id;
 469                 } else {
 470                         ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 471                                 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 472                         WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
 473                         min = max = apic_id;
 474                         ipi_bitmap = 0;
 475                 }
 476                 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
 477         }
 478
 479         if (ipi_bitmap) {
 480                 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 481                         (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 482                 WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
 483         }
 484
 485         local_irq_restore(flags);
 486 }
 487
 488 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
 489 {
 490         __send_ipi_mask(mask, vector);
 491 }
 492
 493 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
 494 {
 495         unsigned int this_cpu = smp_processor_id();
 496         struct cpumask new_mask;
 497         const struct cpumask *local_mask;
 498
 499         cpumask_copy(&new_mask, mask);
 500         cpumask_clear_cpu(this_cpu, &new_mask);
 501         local_mask = &new_mask;
 502         __send_ipi_mask(local_mask, vector);
 503 }
 504
 505 static void kvm_send_ipi_allbutself(int vector)
 506 {
 507         kvm_send_ipi_mask_allbutself(cpu_online_mask, vector);
 508 }
 509
 510 static void kvm_send_ipi_all(int vector)
 511 {
 512         __send_ipi_mask(cpu_online_mask, vector);
 513 }
 514
 515 /*
 516  * Set the IPI entry points
 517  */
 518 static void kvm_setup_pv_ipi(void)
 519 {
 520         apic->send_IPI_mask = kvm_send_ipi_mask;
 521         apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
 522         apic->send_IPI_allbutself = kvm_send_ipi_allbutself;
 523         apic->send_IPI_all = kvm_send_ipi_all;
 524         pr_info("KVM setup pv IPIs\n");
 525 }
 526
 527 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
 528 {
 529         int cpu;
 530
 531         native_send_call_func_ipi(mask);
 532
 533         /* Make sure other vCPUs get a chance to run if they need to. */
 534         for_each_cpu(cpu, mask) {
 535                 if (vcpu_is_preempted(cpu)) {
 536                         kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
 537                         break;
 538                 }
 539         }
 540 }
 541
 542 static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
 543 {
 544         native_smp_prepare_cpus(max_cpus);
 545         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
 546                 static_branch_disable(&virt_spin_lock_key);
 547 }
 548
 549 static void __init kvm_smp_prepare_boot_cpu(void)
 550 {
 551         /*
 552          * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
 553          * shares the guest physical address with the hypervisor.
 554          */
 555         sev_map_percpu_data();
 556
 557         kvm_guest_cpu_init();
 558         native_smp_prepare_boot_cpu();
 559         kvm_spinlock_init();
 560 }
 561
 562 static void kvm_guest_cpu_offline(void)
 563 {
 564         kvm_disable_steal_time();
 565         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 566                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 567         kvm_pv_disable_apf();
 568         apf_task_wake_all();
 569 }
 570
 571 static int kvm_cpu_online(unsigned int cpu)
 572 {
 573         local_irq_disable();
 574         kvm_guest_cpu_init();
 575         local_irq_enable();
 576         return 0;
 577 }
 578
 579 static int kvm_cpu_down_prepare(unsigned int cpu)
 580 {
 581         local_irq_disable();
 582         kvm_guest_cpu_offline();
 583         local_irq_enable();
 584         return 0;
 585 }
 586 #endif
 587
 588 static void __init kvm_apf_trap_init(void)
 589 {
 590         update_intr_gate(X86_TRAP_PF, async_page_fault);
 591 }
 592
 593 static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
 594
 595 static void kvm_flush_tlb_others(const struct cpumask *cpumask,
 596                         const struct flush_tlb_info *info)
 597 {
 598         u8 state;
 599         int cpu;
 600         struct kvm_steal_time *src;
 601         struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
 602
 603         cpumask_copy(flushmask, cpumask);
 604         /*
 605          * We have to call flush only on online vCPUs. And
 606          * queue flush_on_enter for pre-empted vCPUs
 607          */
 608         for_each_cpu(cpu, flushmask) {
 609                 src = &per_cpu(steal_time, cpu);
 610                 state = READ_ONCE(src->preempted);
 611                 if ((state & KVM_VCPU_PREEMPTED)) {
 612                         if (try_cmpxchg(&src->preempted, &state,
 613                                         state | KVM_VCPU_FLUSH_TLB))
 614                                 __cpumask_clear_cpu(cpu, flushmask);
 615                 }
 616         }
 617
 618         native_flush_tlb_others(flushmask, info);
 619 }
 620
 621 static void __init kvm_guest_init(void)
 622 {
 623         int i;
 624
 625         paravirt_ops_setup();
 626         register_reboot_notifier(&kvm_pv_reboot_nb);
 627         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 628                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 629         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 630                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 631
 632         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 633                 has_steal_clock = 1;
 634                 pv_ops.time.steal_clock = kvm_steal_clock;
 635         }
 636
 637         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 638             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 639             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 640                 pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
 641                 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
 642         }
 643
 644         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 645                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 646
 647 #ifdef CONFIG_SMP
 648         smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
 649         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 650         if (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
 651             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 652             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 653                 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
 654                 pr_info("KVM setup pv sched yield\n");
 655         }
 656         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 657                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 658                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 659 #else
 660         sev_map_percpu_data();
 661         kvm_guest_cpu_init();
 662 #endif
 663
 664         /*
 665          * Hard lockup detection is enabled by default. Disable it, as guests
 666          * can get false positives too easily, for example if the host is
 667          * overcommitted.
 668          */
 669         hardlockup_detector_disable();
 670 }
 671
 672 static noinline uint32_t __kvm_cpuid_base(void)
 673 {
 674         if (boot_cpu_data.cpuid_level < 0)
 675                 return 0;       /* So we don't blow up on old processors */
 676
 677         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 678                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 679
 680         return 0;
 681 }
 682
 683 static inline uint32_t kvm_cpuid_base(void)
 684 {
 685         static int kvm_cpuid_base = -1;
 686
 687         if (kvm_cpuid_base == -1)
 688                 kvm_cpuid_base = __kvm_cpuid_base();
 689
 690         return kvm_cpuid_base;
 691 }
 692
 693 bool kvm_para_available(void)
 694 {
 695         return kvm_cpuid_base() != 0;
 696 }
 697 EXPORT_SYMBOL_GPL(kvm_para_available);
 698
 699 unsigned int kvm_arch_para_features(void)
 700 {
 701         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 702 }
 703
 704 unsigned int kvm_arch_para_hints(void)
 705 {
 706         return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 707 }
 708
 709 static uint32_t __init kvm_detect(void)
 710 {
 711         return kvm_cpuid_base();
 712 }
 713
 714 static void __init kvm_apic_init(void)
 715 {
 716 #if defined(CONFIG_SMP)
 717         if (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI))
 718                 kvm_setup_pv_ipi();
 719 #endif
 720 }
 721
 722 static void __init kvm_init_platform(void)
 723 {
 724         kvmclock_init();
 725         x86_platform.apic_post_init = kvm_apic_init;
 726 }
 727
 728 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
 729         .name                   = "KVM",
 730         .detect                 = kvm_detect,
 731         .type                   = X86_HYPER_KVM,
 732         .init.guest_late_init   = kvm_guest_init,
 733         .init.x2apic_available  = kvm_para_available,
 734         .init.init_platform     = kvm_init_platform,
 735 };
 736
 737 static __init int activate_jump_labels(void)
 738 {
 739         if (has_steal_clock) {
 740                 static_key_slow_inc(&paravirt_steal_enabled);
 741                 if (steal_acc)
 742                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 743         }
 744
 745         return 0;
 746 }
 747 arch_initcall(activate_jump_labels);
 748
 749 static __init int kvm_setup_pv_tlb_flush(void)
 750 {
 751         int cpu;
 752
 753         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 754             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 755             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 756                 for_each_possible_cpu(cpu) {
 757                         zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
 758                                 GFP_KERNEL, cpu_to_node(cpu));
 759                 }
 760                 pr_info("KVM setup pv remote TLB flush\n");
 761         }
 762
 763         return 0;
 764 }
 765 arch_initcall(kvm_setup_pv_tlb_flush);
 766
 767 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 768
 769 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 770 static void kvm_kick_cpu(int cpu)
 771 {
 772         int apicid;
 773         unsigned long flags = 0;
 774
 775         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 776         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 777 }
 778
 779 #include <asm/qspinlock.h>
 780
 781 static void kvm_wait(u8 *ptr, u8 val)
 782 {
 783         unsigned long flags;
 784
 785         if (in_nmi())
 786                 return;
 787
 788         local_irq_save(flags);
 789
 790         if (READ_ONCE(*ptr) != val)
 791                 goto out;
 792
 793         /*
 794          * halt until it's our turn and kicked. Note that we do safe halt
 795          * for irq enabled case to avoid hang when lock info is overwritten
 796          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 797          */
 798         if (arch_irqs_disabled_flags(flags))
 799                 halt();
 800         else
 801                 safe_halt();
 802
 803 out:
 804         local_irq_restore(flags);
 805 }
 806
 807 #ifdef CONFIG_X86_32
 808 __visible bool __kvm_vcpu_is_preempted(long cpu)
 809 {
 810         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 811
 812         return !!(src->preempted & KVM_VCPU_PREEMPTED);
 813 }
 814 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 815
 816 #else
 817
 818 #include <asm/asm-offsets.h>
 819
 820 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 821
 822 /*
 823  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 824  * restoring to/from the stack.
 825  */
 826 asm(
 827 ".pushsection .text;"
 828 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 829 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 830 "__raw_callee_save___kvm_vcpu_is_preempted:"
 831 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 832 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 833 "setne  %al;"
 834 "ret;"
 835 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
 836 ".popsection");
 837
 838 #endif
 839
 840 /*
 841  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 842  */
 843 void __init kvm_spinlock_init(void)
 844 {
 845         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 846         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 847                 return;
 848
 849         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
 850                 return;
 851
 852         /* Don't use the pvqspinlock code if there is only 1 vCPU. */
 853         if (num_possible_cpus() == 1)
 854                 return;
 855
 856         __pv_init_lock_hash();
 857         pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 858         pv_ops.lock.queued_spin_unlock =
 859                 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 860         pv_ops.lock.wait = kvm_wait;
 861         pv_ops.lock.kick = kvm_kick_cpu;
 862
 863         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 864                 pv_ops.lock.vcpu_is_preempted =
 865                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 866         }
 867 }
 868
 869 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */