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/kvm_guest.h>
  49
  50 static int kvmapf = 1;
  51
  52 static int 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 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 int kvmclock_vsyscall = 1;
  70 static int parse_no_kvmclock_vsyscall(char *arg)
  71 {
  72         kvmclock_vsyscall = 0;
  73         return 0;
  74 }
  75
  76 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
  77
  78 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  79 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
  80 static int has_steal_clock = 0;
  81
  82 /*
  83  * No need for any "IO delay" on KVM
  84  */
  85 static void kvm_io_delay(void)
  86 {
  87 }
  88
  89 #define KVM_TASK_SLEEP_HASHBITS 8
  90 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  91
  92 struct kvm_task_sleep_node {
  93         struct hlist_node link;
  94         struct swait_queue_head wq;
  95         u32 token;
  96         int cpu;
  97         bool halted;
  98 };
  99
 100 static struct kvm_task_sleep_head {
 101         raw_spinlock_t lock;
 102         struct hlist_head list;
 103 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
 104
 105 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
 106                                                   u32 token)
 107 {
 108         struct hlist_node *p;
 109
 110         hlist_for_each(p, &b->list) {
 111                 struct kvm_task_sleep_node *n =
 112                         hlist_entry(p, typeof(*n), link);
 113                 if (n->token == token)
 114                         return n;
 115         }
 116
 117         return NULL;
 118 }
 119
 120 /*
 121  * @interrupt_kernel: Is this called from a routine which interrupts the kernel
 122  *                    (other than user space)?
 123  */
 124 void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
 125 {
 126         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 127         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 128         struct kvm_task_sleep_node n, *e;
 129         DECLARE_SWAITQUEUE(wait);
 130
 131         rcu_irq_enter();
 132
 133         raw_spin_lock(&b->lock);
 134         e = _find_apf_task(b, token);
 135         if (e) {
 136                 /* dummy entry exist -> wake up was delivered ahead of PF */
 137                 hlist_del(&e->link);
 138                 kfree(e);
 139                 raw_spin_unlock(&b->lock);
 140
 141                 rcu_irq_exit();
 142                 return;
 143         }
 144
 145         n.token = token;
 146         n.cpu = smp_processor_id();
 147         n.halted = is_idle_task(current) ||
 148                    (IS_ENABLED(CONFIG_PREEMPT_COUNT)
 149                     ? preempt_count() > 1 || rcu_preempt_depth()
 150                     : interrupt_kernel);
 151         init_swait_queue_head(&n.wq);
 152         hlist_add_head(&n.link, &b->list);
 153         raw_spin_unlock(&b->lock);
 154
 155         for (;;) {
 156                 if (!n.halted)
 157                         prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 158                 if (hlist_unhashed(&n.link))
 159                         break;
 160
 161                 rcu_irq_exit();
 162
 163                 if (!n.halted) {
 164                         local_irq_enable();
 165                         schedule();
 166                         local_irq_disable();
 167                 } else {
 168                         /*
 169                          * We cannot reschedule. So halt.
 170                          */
 171                         native_safe_halt();
 172                         local_irq_disable();
 173                 }
 174
 175                 rcu_irq_enter();
 176         }
 177         if (!n.halted)
 178                 finish_swait(&n.wq, &wait);
 179
 180         rcu_irq_exit();
 181         return;
 182 }
 183 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 184
 185 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 186 {
 187         hlist_del_init(&n->link);
 188         if (n->halted)
 189                 smp_send_reschedule(n->cpu);
 190         else if (swq_has_sleeper(&n->wq))
 191                 swake_up(&n->wq);
 192 }
 193
 194 static void apf_task_wake_all(void)
 195 {
 196         int i;
 197
 198         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 199                 struct hlist_node *p, *next;
 200                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 201                 raw_spin_lock(&b->lock);
 202                 hlist_for_each_safe(p, next, &b->list) {
 203                         struct kvm_task_sleep_node *n =
 204                                 hlist_entry(p, typeof(*n), link);
 205                         if (n->cpu == smp_processor_id())
 206                                 apf_task_wake_one(n);
 207                 }
 208                 raw_spin_unlock(&b->lock);
 209         }
 210 }
 211
 212 void kvm_async_pf_task_wake(u32 token)
 213 {
 214         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 215         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 216         struct kvm_task_sleep_node *n;
 217
 218         if (token == ~0) {
 219                 apf_task_wake_all();
 220                 return;
 221         }
 222
 223 again:
 224         raw_spin_lock(&b->lock);
 225         n = _find_apf_task(b, token);
 226         if (!n) {
 227                 /*
 228                  * async PF was not yet handled.
 229                  * Add dummy entry for the token.
 230                  */
 231                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 232                 if (!n) {
 233                         /*
 234                          * Allocation failed! Busy wait while other cpu
 235                          * handles async PF.
 236                          */
 237                         raw_spin_unlock(&b->lock);
 238                         cpu_relax();
 239                         goto again;
 240                 }
 241                 n->token = token;
 242                 n->cpu = smp_processor_id();
 243                 init_swait_queue_head(&n->wq);
 244                 hlist_add_head(&n->link, &b->list);
 245         } else
 246                 apf_task_wake_one(n);
 247         raw_spin_unlock(&b->lock);
 248         return;
 249 }
 250 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 251
 252 u32 kvm_read_and_reset_pf_reason(void)
 253 {
 254         u32 reason = 0;
 255
 256         if (__this_cpu_read(apf_reason.enabled)) {
 257                 reason = __this_cpu_read(apf_reason.reason);
 258                 __this_cpu_write(apf_reason.reason, 0);
 259         }
 260
 261         return reason;
 262 }
 263 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 264 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 265
 266 dotraplinkage void
 267 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 268 {
 269         enum ctx_state prev_state;
 270
 271         switch (kvm_read_and_reset_pf_reason()) {
 272         default:
 273                 do_page_fault(regs, error_code);
 274                 break;
 275         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 276                 /* page is swapped out by the host. */
 277                 prev_state = exception_enter();
 278                 kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
 279                 exception_exit(prev_state);
 280                 break;
 281         case KVM_PV_REASON_PAGE_READY:
 282                 rcu_irq_enter();
 283                 kvm_async_pf_task_wake((u32)read_cr2());
 284                 rcu_irq_exit();
 285                 break;
 286         }
 287 }
 288 NOKPROBE_SYMBOL(do_async_page_fault);
 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_cpu_ops.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("kvm-stealtime: cpu %d, msr %llx\n",
 312                 cpu, (unsigned long long) slow_virt_to_phys(st));
 313 }
 314
 315 static DEFINE_PER_CPU(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_available())
 334                 return;
 335
 336         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 337                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 338
 339 #ifdef CONFIG_PREEMPT
 340                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 341 #endif
 342                 pa |= KVM_ASYNC_PF_ENABLED;
 343
 344                 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
 345                         pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
 346
 347                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 348                 __this_cpu_write(apf_reason.enabled, 1);
 349                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 350                        smp_processor_id());
 351         }
 352
 353         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 354                 unsigned long pa;
 355                 /* Size alignment is implied but just to make it explicit. */
 356                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 357                 __this_cpu_write(kvm_apic_eoi, 0);
 358                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 359                         | KVM_MSR_ENABLED;
 360                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 361         }
 362
 363         if (has_steal_clock)
 364                 kvm_register_steal_time();
 365 }
 366
 367 static void kvm_pv_disable_apf(void)
 368 {
 369         if (!__this_cpu_read(apf_reason.enabled))
 370                 return;
 371
 372         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 373         __this_cpu_write(apf_reason.enabled, 0);
 374
 375         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 376                smp_processor_id());
 377 }
 378
 379 static void kvm_pv_guest_cpu_reboot(void *unused)
 380 {
 381         /*
 382          * We disable PV EOI before we load a new kernel by kexec,
 383          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 384          * New kernel can re-enable when it boots.
 385          */
 386         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 387                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 388         kvm_pv_disable_apf();
 389         kvm_disable_steal_time();
 390 }
 391
 392 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 393                                 unsigned long code, void *unused)
 394 {
 395         if (code == SYS_RESTART)
 396                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 397         return NOTIFY_DONE;
 398 }
 399
 400 static struct notifier_block kvm_pv_reboot_nb = {
 401         .notifier_call = kvm_pv_reboot_notify,
 402 };
 403
 404 static u64 kvm_steal_clock(int cpu)
 405 {
 406         u64 steal;
 407         struct kvm_steal_time *src;
 408         int version;
 409
 410         src = &per_cpu(steal_time, cpu);
 411         do {
 412                 version = src->version;
 413                 virt_rmb();
 414                 steal = src->steal;
 415                 virt_rmb();
 416         } while ((version & 1) || (version != src->version));
 417
 418         return steal;
 419 }
 420
 421 void kvm_disable_steal_time(void)
 422 {
 423         if (!has_steal_clock)
 424                 return;
 425
 426         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 427 }
 428
 429 #ifdef CONFIG_SMP
 430 static void __init kvm_smp_prepare_boot_cpu(void)
 431 {
 432         kvm_guest_cpu_init();
 433         native_smp_prepare_boot_cpu();
 434         kvm_spinlock_init();
 435 }
 436
 437 static void kvm_guest_cpu_offline(void)
 438 {
 439         kvm_disable_steal_time();
 440         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 441                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 442         kvm_pv_disable_apf();
 443         apf_task_wake_all();
 444 }
 445
 446 static int kvm_cpu_online(unsigned int cpu)
 447 {
 448         local_irq_disable();
 449         kvm_guest_cpu_init();
 450         local_irq_enable();
 451         return 0;
 452 }
 453
 454 static int kvm_cpu_down_prepare(unsigned int cpu)
 455 {
 456         local_irq_disable();
 457         kvm_guest_cpu_offline();
 458         local_irq_enable();
 459         return 0;
 460 }
 461 #endif
 462
 463 static void __init kvm_apf_trap_init(void)
 464 {
 465         update_intr_gate(X86_TRAP_PF, async_page_fault);
 466 }
 467
 468 void __init kvm_guest_init(void)
 469 {
 470         int i;
 471
 472         if (!kvm_para_available())
 473                 return;
 474
 475         paravirt_ops_setup();
 476         register_reboot_notifier(&kvm_pv_reboot_nb);
 477         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 478                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 479         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 480                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 481
 482         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 483                 has_steal_clock = 1;
 484                 pv_time_ops.steal_clock = kvm_steal_clock;
 485         }
 486
 487         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 488                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 489
 490         if (kvmclock_vsyscall)
 491                 kvm_setup_vsyscall_timeinfo();
 492
 493 #ifdef CONFIG_SMP
 494         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 495         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 496                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 497                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 498 #else
 499         kvm_guest_cpu_init();
 500 #endif
 501
 502         /*
 503          * Hard lockup detection is enabled by default. Disable it, as guests
 504          * can get false positives too easily, for example if the host is
 505          * overcommitted.
 506          */
 507         hardlockup_detector_disable();
 508 }
 509
 510 static noinline uint32_t __kvm_cpuid_base(void)
 511 {
 512         if (boot_cpu_data.cpuid_level < 0)
 513                 return 0;       /* So we don't blow up on old processors */
 514
 515         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 516                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 517
 518         return 0;
 519 }
 520
 521 static inline uint32_t kvm_cpuid_base(void)
 522 {
 523         static int kvm_cpuid_base = -1;
 524
 525         if (kvm_cpuid_base == -1)
 526                 kvm_cpuid_base = __kvm_cpuid_base();
 527
 528         return kvm_cpuid_base;
 529 }
 530
 531 bool kvm_para_available(void)
 532 {
 533         return kvm_cpuid_base() != 0;
 534 }
 535 EXPORT_SYMBOL_GPL(kvm_para_available);
 536
 537 unsigned int kvm_arch_para_features(void)
 538 {
 539         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 540 }
 541
 542 static uint32_t __init kvm_detect(void)
 543 {
 544         return kvm_cpuid_base();
 545 }
 546
 547 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
 548         .name                   = "KVM",
 549         .detect                 = kvm_detect,
 550         .type                   = X86_HYPER_KVM,
 551         .init.x2apic_available  = kvm_para_available,
 552 };
 553
 554 static __init int activate_jump_labels(void)
 555 {
 556         if (has_steal_clock) {
 557                 static_key_slow_inc(&paravirt_steal_enabled);
 558                 if (steal_acc)
 559                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 560         }
 561
 562         return 0;
 563 }
 564 arch_initcall(activate_jump_labels);
 565
 566 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 567
 568 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 569 static void kvm_kick_cpu(int cpu)
 570 {
 571         int apicid;
 572         unsigned long flags = 0;
 573
 574         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 575         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 576 }
 577
 578 #include <asm/qspinlock.h>
 579
 580 static void kvm_wait(u8 *ptr, u8 val)
 581 {
 582         unsigned long flags;
 583
 584         if (in_nmi())
 585                 return;
 586
 587         local_irq_save(flags);
 588
 589         if (READ_ONCE(*ptr) != val)
 590                 goto out;
 591
 592         /*
 593          * halt until it's our turn and kicked. Note that we do safe halt
 594          * for irq enabled case to avoid hang when lock info is overwritten
 595          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 596          */
 597         if (arch_irqs_disabled_flags(flags))
 598                 halt();
 599         else
 600                 safe_halt();
 601
 602 out:
 603         local_irq_restore(flags);
 604 }
 605
 606 #ifdef CONFIG_X86_32
 607 __visible bool __kvm_vcpu_is_preempted(long cpu)
 608 {
 609         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 610
 611         return !!src->preempted;
 612 }
 613 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 614
 615 #else
 616
 617 #include <asm/asm-offsets.h>
 618
 619 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 620
 621 /*
 622  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 623  * restoring to/from the stack.
 624  */
 625 asm(
 626 ".pushsection .text;"
 627 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 628 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 629 "__raw_callee_save___kvm_vcpu_is_preempted:"
 630 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 631 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 632 "setne  %al;"
 633 "ret;"
 634 ".popsection");
 635
 636 #endif
 637
 638 /*
 639  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 640  */
 641 void __init kvm_spinlock_init(void)
 642 {
 643         if (!kvm_para_available())
 644                 return;
 645         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 646         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 647                 return;
 648
 649         __pv_init_lock_hash();
 650         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 651         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 652         pv_lock_ops.wait = kvm_wait;
 653         pv_lock_ops.kick = kvm_kick_cpu;
 654
 655         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 656                 pv_lock_ops.vcpu_is_preempted =
 657                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 658         }
 659 }
 660
 661 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */