arch/x86/kernel/kvm.c

   1 /*
   2  * KVM paravirt_ops implementation
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  17  *
  18  * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
  19  * Copyright IBM Corporation, 2007
  20  *   Authors: Anthony Liguori <aliguori@us.ibm.com>
  21  */
  22
  23 #include <linux/context_tracking.h>
  24 #include <linux/init.h>
  25 #include <linux/kernel.h>
  26 #include <linux/kvm_para.h>
  27 #include <linux/cpu.h>
  28 #include <linux/mm.h>
  29 #include <linux/highmem.h>
  30 #include <linux/hardirq.h>
  31 #include <linux/notifier.h>
  32 #include <linux/reboot.h>
  33 #include <linux/hash.h>
  34 #include <linux/sched.h>
  35 #include <linux/slab.h>
  36 #include <linux/kprobes.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/nmi.h>
  39 #include <linux/swait.h>
  40 #include <asm/timer.h>
  41 #include <asm/cpu.h>
  42 #include <asm/traps.h>
  43 #include <asm/desc.h>
  44 #include <asm/tlbflush.h>
  45 #include <asm/apic.h>
  46 #include <asm/apicdef.h>
  47 #include <asm/hypervisor.h>
  48 #include <asm/tlb.h>
  49
  50 static int kvmapf = 1;
  51
  52 static int __init parse_no_kvmapf(char *arg)
  53 {
  54         kvmapf = 0;
  55         return 0;
  56 }
  57
  58 early_param("no-kvmapf", parse_no_kvmapf);
  59
  60 static int steal_acc = 1;
  61 static int __init parse_no_stealacc(char *arg)
  62 {
  63         steal_acc = 0;
  64         return 0;
  65 }
  66
  67 early_param("no-steal-acc", parse_no_stealacc);
  68
  69 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  70 static DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64);
  71 static int has_steal_clock = 0;
  72
  73 /*
  74  * No need for any "IO delay" on KVM
  75  */
  76 static void kvm_io_delay(void)
  77 {
  78 }
  79
  80 #define KVM_TASK_SLEEP_HASHBITS 8
  81 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  82
  83 struct kvm_task_sleep_node {
  84         struct hlist_node link;
  85         struct swait_queue_head wq;
  86         u32 token;
  87         int cpu;
  88         bool halted;
  89 };
  90
  91 static struct kvm_task_sleep_head {
  92         raw_spinlock_t lock;
  93         struct hlist_head list;
  94 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
  95
  96 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
  97                                                   u32 token)
  98 {
  99         struct hlist_node *p;
 100
 101         hlist_for_each(p, &b->list) {
 102                 struct kvm_task_sleep_node *n =
 103                         hlist_entry(p, typeof(*n), link);
 104                 if (n->token == token)
 105                         return n;
 106         }
 107
 108         return NULL;
 109 }
 110
 111 /*
 112  * @interrupt_kernel: Is this called from a routine which interrupts the kernel
 113  *                    (other than user space)?
 114  */
 115 void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
 116 {
 117         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 118         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 119         struct kvm_task_sleep_node n, *e;
 120         DECLARE_SWAITQUEUE(wait);
 121
 122         rcu_irq_enter();
 123
 124         raw_spin_lock(&b->lock);
 125         e = _find_apf_task(b, token);
 126         if (e) {
 127                 /* dummy entry exist -> wake up was delivered ahead of PF */
 128                 hlist_del(&e->link);
 129                 kfree(e);
 130                 raw_spin_unlock(&b->lock);
 131
 132                 rcu_irq_exit();
 133                 return;
 134         }
 135
 136         n.token = token;
 137         n.cpu = smp_processor_id();
 138         n.halted = is_idle_task(current) ||
 139                    (IS_ENABLED(CONFIG_PREEMPT_COUNT)
 140                     ? preempt_count() > 1 || rcu_preempt_depth()
 141                     : interrupt_kernel);
 142         init_swait_queue_head(&n.wq);
 143         hlist_add_head(&n.link, &b->list);
 144         raw_spin_unlock(&b->lock);
 145
 146         for (;;) {
 147                 if (!n.halted)
 148                         prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 149                 if (hlist_unhashed(&n.link))
 150                         break;
 151
 152                 rcu_irq_exit();
 153
 154                 if (!n.halted) {
 155                         local_irq_enable();
 156                         schedule();
 157                         local_irq_disable();
 158                 } else {
 159                         /*
 160                          * We cannot reschedule. So halt.
 161                          */
 162                         native_safe_halt();
 163                         local_irq_disable();
 164                 }
 165
 166                 rcu_irq_enter();
 167         }
 168         if (!n.halted)
 169                 finish_swait(&n.wq, &wait);
 170
 171         rcu_irq_exit();
 172         return;
 173 }
 174 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 175
 176 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 177 {
 178         hlist_del_init(&n->link);
 179         if (n->halted)
 180                 smp_send_reschedule(n->cpu);
 181         else if (swq_has_sleeper(&n->wq))
 182                 swake_up_one(&n->wq);
 183 }
 184
 185 static void apf_task_wake_all(void)
 186 {
 187         int i;
 188
 189         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 190                 struct hlist_node *p, *next;
 191                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 192                 raw_spin_lock(&b->lock);
 193                 hlist_for_each_safe(p, next, &b->list) {
 194                         struct kvm_task_sleep_node *n =
 195                                 hlist_entry(p, typeof(*n), link);
 196                         if (n->cpu == smp_processor_id())
 197                                 apf_task_wake_one(n);
 198                 }
 199                 raw_spin_unlock(&b->lock);
 200         }
 201 }
 202
 203 void kvm_async_pf_task_wake(u32 token)
 204 {
 205         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 206         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 207         struct kvm_task_sleep_node *n;
 208
 209         if (token == ~0) {
 210                 apf_task_wake_all();
 211                 return;
 212         }
 213
 214 again:
 215         raw_spin_lock(&b->lock);
 216         n = _find_apf_task(b, token);
 217         if (!n) {
 218                 /*
 219                  * async PF was not yet handled.
 220                  * Add dummy entry for the token.
 221                  */
 222                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 223                 if (!n) {
 224                         /*
 225                          * Allocation failed! Busy wait while other cpu
 226                          * handles async PF.
 227                          */
 228                         raw_spin_unlock(&b->lock);
 229                         cpu_relax();
 230                         goto again;
 231                 }
 232                 n->token = token;
 233                 n->cpu = smp_processor_id();
 234                 init_swait_queue_head(&n->wq);
 235                 hlist_add_head(&n->link, &b->list);
 236         } else
 237                 apf_task_wake_one(n);
 238         raw_spin_unlock(&b->lock);
 239         return;
 240 }
 241 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 242
 243 u32 kvm_read_and_reset_pf_reason(void)
 244 {
 245         u32 reason = 0;
 246
 247         if (__this_cpu_read(apf_reason.enabled)) {
 248                 reason = __this_cpu_read(apf_reason.reason);
 249                 __this_cpu_write(apf_reason.reason, 0);
 250         }
 251
 252         return reason;
 253 }
 254 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 255 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 256
 257 dotraplinkage void
 258 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 259 {
 260         enum ctx_state prev_state;
 261
 262         switch (kvm_read_and_reset_pf_reason()) {
 263         default:
 264                 do_page_fault(regs, error_code);
 265                 break;
 266         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 267                 /* page is swapped out by the host. */
 268                 prev_state = exception_enter();
 269                 kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
 270                 exception_exit(prev_state);
 271                 break;
 272         case KVM_PV_REASON_PAGE_READY:
 273                 rcu_irq_enter();
 274                 kvm_async_pf_task_wake((u32)read_cr2());
 275                 rcu_irq_exit();
 276                 break;
 277         }
 278 }
 279 NOKPROBE_SYMBOL(do_async_page_fault);
 280
 281 static void __init paravirt_ops_setup(void)
 282 {
 283         pv_info.name = "KVM";
 284
 285         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 286                 pv_cpu_ops.io_delay = kvm_io_delay;
 287
 288 #ifdef CONFIG_X86_IO_APIC
 289         no_timer_check = 1;
 290 #endif
 291 }
 292
 293 static void kvm_register_steal_time(void)
 294 {
 295         int cpu = smp_processor_id();
 296         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 297
 298         if (!has_steal_clock)
 299                 return;
 300
 301         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 302         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
 303                 cpu, (unsigned long long) slow_virt_to_phys(st));
 304 }
 305
 306 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 307
 308 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 309 {
 310         /**
 311          * This relies on __test_and_clear_bit to modify the memory
 312          * in a way that is atomic with respect to the local CPU.
 313          * The hypervisor only accesses this memory from the local CPU so
 314          * there's no need for lock or memory barriers.
 315          * An optimization barrier is implied in apic write.
 316          */
 317         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 318                 return;
 319         apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
 320 }
 321
 322 static void kvm_guest_cpu_init(void)
 323 {
 324         if (!kvm_para_available())
 325                 return;
 326
 327         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 328                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 329
 330 #ifdef CONFIG_PREEMPT
 331                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 332 #endif
 333                 pa |= KVM_ASYNC_PF_ENABLED;
 334
 335                 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
 336                         pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
 337
 338                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 339                 __this_cpu_write(apf_reason.enabled, 1);
 340                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 341                        smp_processor_id());
 342         }
 343
 344         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 345                 unsigned long pa;
 346                 /* Size alignment is implied but just to make it explicit. */
 347                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 348                 __this_cpu_write(kvm_apic_eoi, 0);
 349                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 350                         | KVM_MSR_ENABLED;
 351                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 352         }
 353
 354         if (has_steal_clock)
 355                 kvm_register_steal_time();
 356 }
 357
 358 static void kvm_pv_disable_apf(void)
 359 {
 360         if (!__this_cpu_read(apf_reason.enabled))
 361                 return;
 362
 363         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 364         __this_cpu_write(apf_reason.enabled, 0);
 365
 366         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 367                smp_processor_id());
 368 }
 369
 370 static void kvm_pv_guest_cpu_reboot(void *unused)
 371 {
 372         /*
 373          * We disable PV EOI before we load a new kernel by kexec,
 374          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 375          * New kernel can re-enable when it boots.
 376          */
 377         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 378                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 379         kvm_pv_disable_apf();
 380         kvm_disable_steal_time();
 381 }
 382
 383 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 384                                 unsigned long code, void *unused)
 385 {
 386         if (code == SYS_RESTART)
 387                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 388         return NOTIFY_DONE;
 389 }
 390
 391 static struct notifier_block kvm_pv_reboot_nb = {
 392         .notifier_call = kvm_pv_reboot_notify,
 393 };
 394
 395 static u64 kvm_steal_clock(int cpu)
 396 {
 397         u64 steal;
 398         struct kvm_steal_time *src;
 399         int version;
 400
 401         src = &per_cpu(steal_time, cpu);
 402         do {
 403                 version = src->version;
 404                 virt_rmb();
 405                 steal = src->steal;
 406                 virt_rmb();
 407         } while ((version & 1) || (version != src->version));
 408
 409         return steal;
 410 }
 411
 412 void kvm_disable_steal_time(void)
 413 {
 414         if (!has_steal_clock)
 415                 return;
 416
 417         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 418 }
 419
 420 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
 421 {
 422         early_set_memory_decrypted((unsigned long) ptr, size);
 423 }
 424
 425 /*
 426  * Iterate through all possible CPUs and map the memory region pointed
 427  * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
 428  *
 429  * Note: we iterate through all possible CPUs to ensure that CPUs
 430  * hotplugged will have their per-cpu variable already mapped as
 431  * decrypted.
 432  */
 433 static void __init sev_map_percpu_data(void)
 434 {
 435         int cpu;
 436
 437         if (!sev_active())
 438                 return;
 439
 440         for_each_possible_cpu(cpu) {
 441                 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
 442                 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
 443                 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
 444         }
 445 }
 446
 447 #ifdef CONFIG_SMP
 448 #define KVM_IPI_CLUSTER_SIZE    (2 * BITS_PER_LONG)
 449
 450 static void __send_ipi_mask(const struct cpumask *mask, int vector)
 451 {
 452         unsigned long flags;
 453         int cpu, apic_id, icr;
 454         int min = 0, max = 0;
 455 #ifdef CONFIG_X86_64
 456         __uint128_t ipi_bitmap = 0;
 457 #else
 458         u64 ipi_bitmap = 0;
 459 #endif
 460         long ret;
 461
 462         if (cpumask_empty(mask))
 463                 return;
 464
 465         local_irq_save(flags);
 466
 467         switch (vector) {
 468         default:
 469                 icr = APIC_DM_FIXED | vector;
 470                 break;
 471         case NMI_VECTOR:
 472                 icr = APIC_DM_NMI;
 473                 break;
 474         }
 475
 476         for_each_cpu(cpu, mask) {
 477                 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
 478                 if (!ipi_bitmap) {
 479                         min = max = apic_id;
 480                 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
 481                         ipi_bitmap <<= min - apic_id;
 482                         min = apic_id;
 483                 } else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
 484                         max = apic_id < max ? max : apic_id;
 485                 } else {
 486                         ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 487                                 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 488                         WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
 489                         min = max = apic_id;
 490                         ipi_bitmap = 0;
 491                 }
 492                 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
 493         }
 494
 495         if (ipi_bitmap) {
 496                 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
 497                         (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
 498                 WARN_ONCE(ret < 0, "KVM: failed to send PV IPI: %ld", ret);
 499         }
 500
 501         local_irq_restore(flags);
 502 }
 503
 504 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
 505 {
 506         __send_ipi_mask(mask, vector);
 507 }
 508
 509 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
 510 {
 511         unsigned int this_cpu = smp_processor_id();
 512         struct cpumask new_mask;
 513         const struct cpumask *local_mask;
 514
 515         cpumask_copy(&new_mask, mask);
 516         cpumask_clear_cpu(this_cpu, &new_mask);
 517         local_mask = &new_mask;
 518         __send_ipi_mask(local_mask, vector);
 519 }
 520
 521 static void kvm_send_ipi_allbutself(int vector)
 522 {
 523         kvm_send_ipi_mask_allbutself(cpu_online_mask, vector);
 524 }
 525
 526 static void kvm_send_ipi_all(int vector)
 527 {
 528         __send_ipi_mask(cpu_online_mask, vector);
 529 }
 530
 531 /*
 532  * Set the IPI entry points
 533  */
 534 static void kvm_setup_pv_ipi(void)
 535 {
 536         apic->send_IPI_mask = kvm_send_ipi_mask;
 537         apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
 538         apic->send_IPI_allbutself = kvm_send_ipi_allbutself;
 539         apic->send_IPI_all = kvm_send_ipi_all;
 540         pr_info("KVM setup pv IPIs\n");
 541 }
 542
 543 static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
 544 {
 545         native_smp_prepare_cpus(max_cpus);
 546         if (kvm_para_has_hint(KVM_HINTS_REALTIME))
 547                 static_branch_disable(&virt_spin_lock_key);
 548 }
 549
 550 static void __init kvm_smp_prepare_boot_cpu(void)
 551 {
 552         /*
 553          * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
 554          * shares the guest physical address with the hypervisor.
 555          */
 556         sev_map_percpu_data();
 557
 558         kvm_guest_cpu_init();
 559         native_smp_prepare_boot_cpu();
 560         kvm_spinlock_init();
 561 }
 562
 563 static void kvm_guest_cpu_offline(void)
 564 {
 565         kvm_disable_steal_time();
 566         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 567                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 568         kvm_pv_disable_apf();
 569         apf_task_wake_all();
 570 }
 571
 572 static int kvm_cpu_online(unsigned int cpu)
 573 {
 574         local_irq_disable();
 575         kvm_guest_cpu_init();
 576         local_irq_enable();
 577         return 0;
 578 }
 579
 580 static int kvm_cpu_down_prepare(unsigned int cpu)
 581 {
 582         local_irq_disable();
 583         kvm_guest_cpu_offline();
 584         local_irq_enable();
 585         return 0;
 586 }
 587 #endif
 588
 589 static void __init kvm_apf_trap_init(void)
 590 {
 591         update_intr_gate(X86_TRAP_PF, async_page_fault);
 592 }
 593
 594 static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
 595
 596 static void kvm_flush_tlb_others(const struct cpumask *cpumask,
 597                         const struct flush_tlb_info *info)
 598 {
 599         u8 state;
 600         int cpu;
 601         struct kvm_steal_time *src;
 602         struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
 603
 604         cpumask_copy(flushmask, cpumask);
 605         /*
 606          * We have to call flush only on online vCPUs. And
 607          * queue flush_on_enter for pre-empted vCPUs
 608          */
 609         for_each_cpu(cpu, flushmask) {
 610                 src = &per_cpu(steal_time, cpu);
 611                 state = READ_ONCE(src->preempted);
 612                 if ((state & KVM_VCPU_PREEMPTED)) {
 613                         if (try_cmpxchg(&src->preempted, &state,
 614                                         state | KVM_VCPU_FLUSH_TLB))
 615                                 __cpumask_clear_cpu(cpu, flushmask);
 616                 }
 617         }
 618
 619         native_flush_tlb_others(flushmask, info);
 620 }
 621
 622 static void __init kvm_guest_init(void)
 623 {
 624         int i;
 625
 626         if (!kvm_para_available())
 627                 return;
 628
 629         paravirt_ops_setup();
 630         register_reboot_notifier(&kvm_pv_reboot_nb);
 631         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 632                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 633         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 634                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 635
 636         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 637                 has_steal_clock = 1;
 638                 pv_time_ops.steal_clock = kvm_steal_clock;
 639         }
 640
 641         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 642             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 643             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 644                 pv_mmu_ops.flush_tlb_others = kvm_flush_tlb_others;
 645                 pv_mmu_ops.tlb_remove_table = tlb_remove_table;
 646         }
 647
 648         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 649                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 650
 651 #ifdef CONFIG_SMP
 652         smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
 653         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 654         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 655                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 656                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 657 #else
 658         sev_map_percpu_data();
 659         kvm_guest_cpu_init();
 660 #endif
 661
 662         /*
 663          * Hard lockup detection is enabled by default. Disable it, as guests
 664          * can get false positives too easily, for example if the host is
 665          * overcommitted.
 666          */
 667         hardlockup_detector_disable();
 668 }
 669
 670 static noinline uint32_t __kvm_cpuid_base(void)
 671 {
 672         if (boot_cpu_data.cpuid_level < 0)
 673                 return 0;       /* So we don't blow up on old processors */
 674
 675         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 676                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 677
 678         return 0;
 679 }
 680
 681 static inline uint32_t kvm_cpuid_base(void)
 682 {
 683         static int kvm_cpuid_base = -1;
 684
 685         if (kvm_cpuid_base == -1)
 686                 kvm_cpuid_base = __kvm_cpuid_base();
 687
 688         return kvm_cpuid_base;
 689 }
 690
 691 bool kvm_para_available(void)
 692 {
 693         return kvm_cpuid_base() != 0;
 694 }
 695 EXPORT_SYMBOL_GPL(kvm_para_available);
 696
 697 unsigned int kvm_arch_para_features(void)
 698 {
 699         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 700 }
 701
 702 unsigned int kvm_arch_para_hints(void)
 703 {
 704         return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 705 }
 706
 707 static uint32_t __init kvm_detect(void)
 708 {
 709         return kvm_cpuid_base();
 710 }
 711
 712 static void __init kvm_apic_init(void)
 713 {
 714 #if defined(CONFIG_SMP)
 715         if (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI))
 716                 kvm_setup_pv_ipi();
 717 #endif
 718 }
 719
 720 static void __init kvm_init_platform(void)
 721 {
 722         kvmclock_init();
 723         x86_platform.apic_post_init = kvm_apic_init;
 724 }
 725
 726 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
 727         .name                   = "KVM",
 728         .detect                 = kvm_detect,
 729         .type                   = X86_HYPER_KVM,
 730         .init.guest_late_init   = kvm_guest_init,
 731         .init.x2apic_available  = kvm_para_available,
 732         .init.init_platform     = kvm_init_platform,
 733 };
 734
 735 static __init int activate_jump_labels(void)
 736 {
 737         if (has_steal_clock) {
 738                 static_key_slow_inc(&paravirt_steal_enabled);
 739                 if (steal_acc)
 740                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 741         }
 742
 743         return 0;
 744 }
 745 arch_initcall(activate_jump_labels);
 746
 747 static __init int kvm_setup_pv_tlb_flush(void)
 748 {
 749         int cpu;
 750
 751         if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 752             !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 753             kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 754                 for_each_possible_cpu(cpu) {
 755                         zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
 756                                 GFP_KERNEL, cpu_to_node(cpu));
 757                 }
 758                 pr_info("KVM setup pv remote TLB flush\n");
 759         }
 760
 761         return 0;
 762 }
 763 arch_initcall(kvm_setup_pv_tlb_flush);
 764
 765 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 766
 767 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 768 static void kvm_kick_cpu(int cpu)
 769 {
 770         int apicid;
 771         unsigned long flags = 0;
 772
 773         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 774         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 775 }
 776
 777 #include <asm/qspinlock.h>
 778
 779 static void kvm_wait(u8 *ptr, u8 val)
 780 {
 781         unsigned long flags;
 782
 783         if (in_nmi())
 784                 return;
 785
 786         local_irq_save(flags);
 787
 788         if (READ_ONCE(*ptr) != val)
 789                 goto out;
 790
 791         /*
 792          * halt until it's our turn and kicked. Note that we do safe halt
 793          * for irq enabled case to avoid hang when lock info is overwritten
 794          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 795          */
 796         if (arch_irqs_disabled_flags(flags))
 797                 halt();
 798         else
 799                 safe_halt();
 800
 801 out:
 802         local_irq_restore(flags);
 803 }
 804
 805 #ifdef CONFIG_X86_32
 806 __visible bool __kvm_vcpu_is_preempted(long cpu)
 807 {
 808         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 809
 810         return !!(src->preempted & KVM_VCPU_PREEMPTED);
 811 }
 812 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 813
 814 #else
 815
 816 #include <asm/asm-offsets.h>
 817
 818 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 819
 820 /*
 821  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 822  * restoring to/from the stack.
 823  */
 824 asm(
 825 ".pushsection .text;"
 826 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 827 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 828 "__raw_callee_save___kvm_vcpu_is_preempted:"
 829 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 830 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 831 "setne  %al;"
 832 "ret;"
 833 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
 834 ".popsection");
 835
 836 #endif
 837
 838 /*
 839  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 840  */
 841 void __init kvm_spinlock_init(void)
 842 {
 843         if (!kvm_para_available())
 844                 return;
 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_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 858         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 859         pv_lock_ops.wait = kvm_wait;
 860         pv_lock_ops.kick = kvm_kick_cpu;
 861
 862         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 863                 pv_lock_ops.vcpu_is_preempted =
 864                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 865         }
 866 }
 867
 868 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */