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/idle.h>
  46 #include <asm/apic.h>
  47 #include <asm/apicdef.h>
  48 #include <asm/hypervisor.h>
  49 #include <asm/kvm_guest.h>
  50
  51 static int kvmapf = 1;
  52
  53 static int parse_no_kvmapf(char *arg)
  54 {
  55         kvmapf = 0;
  56         return 0;
  57 }
  58
  59 early_param("no-kvmapf", parse_no_kvmapf);
  60
  61 static int steal_acc = 1;
  62 static int parse_no_stealacc(char *arg)
  63 {
  64         steal_acc = 0;
  65         return 0;
  66 }
  67
  68 early_param("no-steal-acc", parse_no_stealacc);
  69
  70 static int kvmclock_vsyscall = 1;
  71 static int parse_no_kvmclock_vsyscall(char *arg)
  72 {
  73         kvmclock_vsyscall = 0;
  74         return 0;
  75 }
  76
  77 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
  78
  79 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  80 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
  81 static int has_steal_clock = 0;
  82
  83 /*
  84  * No need for any "IO delay" on KVM
  85  */
  86 static void kvm_io_delay(void)
  87 {
  88 }
  89
  90 #define KVM_TASK_SLEEP_HASHBITS 8
  91 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  92
  93 struct kvm_task_sleep_node {
  94         struct hlist_node link;
  95         struct swait_queue_head wq;
  96         u32 token;
  97         int cpu;
  98         bool halted;
  99 };
 100
 101 static struct kvm_task_sleep_head {
 102         raw_spinlock_t lock;
 103         struct hlist_head list;
 104 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
 105
 106 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
 107                                                   u32 token)
 108 {
 109         struct hlist_node *p;
 110
 111         hlist_for_each(p, &b->list) {
 112                 struct kvm_task_sleep_node *n =
 113                         hlist_entry(p, typeof(*n), link);
 114                 if (n->token == token)
 115                         return n;
 116         }
 117
 118         return NULL;
 119 }
 120
 121 void kvm_async_pf_task_wait(u32 token)
 122 {
 123         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 124         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 125         struct kvm_task_sleep_node n, *e;
 126         DECLARE_SWAITQUEUE(wait);
 127
 128         rcu_irq_enter();
 129
 130         raw_spin_lock(&b->lock);
 131         e = _find_apf_task(b, token);
 132         if (e) {
 133                 /* dummy entry exist -> wake up was delivered ahead of PF */
 134                 hlist_del(&e->link);
 135                 kfree(e);
 136                 raw_spin_unlock(&b->lock);
 137
 138                 rcu_irq_exit();
 139                 return;
 140         }
 141
 142         n.token = token;
 143         n.cpu = smp_processor_id();
 144         n.halted = is_idle_task(current) || preempt_count() > 1;
 145         init_swait_queue_head(&n.wq);
 146         hlist_add_head(&n.link, &b->list);
 147         raw_spin_unlock(&b->lock);
 148
 149         for (;;) {
 150                 if (!n.halted)
 151                         prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 152                 if (hlist_unhashed(&n.link))
 153                         break;
 154
 155                 if (!n.halted) {
 156                         local_irq_enable();
 157                         schedule();
 158                         local_irq_disable();
 159                 } else {
 160                         /*
 161                          * We cannot reschedule. So halt.
 162                          */
 163                         rcu_irq_exit();
 164                         native_safe_halt();
 165                         rcu_irq_enter();
 166                         local_irq_disable();
 167                 }
 168         }
 169         if (!n.halted)
 170                 finish_swait(&n.wq, &wait);
 171
 172         rcu_irq_exit();
 173         return;
 174 }
 175 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 176
 177 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 178 {
 179         hlist_del_init(&n->link);
 180         if (n->halted)
 181                 smp_send_reschedule(n->cpu);
 182         else if (swait_active(&n->wq))
 183                 swake_up(&n->wq);
 184 }
 185
 186 static void apf_task_wake_all(void)
 187 {
 188         int i;
 189
 190         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 191                 struct hlist_node *p, *next;
 192                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 193                 raw_spin_lock(&b->lock);
 194                 hlist_for_each_safe(p, next, &b->list) {
 195                         struct kvm_task_sleep_node *n =
 196                                 hlist_entry(p, typeof(*n), link);
 197                         if (n->cpu == smp_processor_id())
 198                                 apf_task_wake_one(n);
 199                 }
 200                 raw_spin_unlock(&b->lock);
 201         }
 202 }
 203
 204 void kvm_async_pf_task_wake(u32 token)
 205 {
 206         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 207         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 208         struct kvm_task_sleep_node *n;
 209
 210         if (token == ~0) {
 211                 apf_task_wake_all();
 212                 return;
 213         }
 214
 215 again:
 216         raw_spin_lock(&b->lock);
 217         n = _find_apf_task(b, token);
 218         if (!n) {
 219                 /*
 220                  * async PF was not yet handled.
 221                  * Add dummy entry for the token.
 222                  */
 223                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 224                 if (!n) {
 225                         /*
 226                          * Allocation failed! Busy wait while other cpu
 227                          * handles async PF.
 228                          */
 229                         raw_spin_unlock(&b->lock);
 230                         cpu_relax();
 231                         goto again;
 232                 }
 233                 n->token = token;
 234                 n->cpu = smp_processor_id();
 235                 init_swait_queue_head(&n->wq);
 236                 hlist_add_head(&n->link, &b->list);
 237         } else
 238                 apf_task_wake_one(n);
 239         raw_spin_unlock(&b->lock);
 240         return;
 241 }
 242 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 243
 244 u32 kvm_read_and_reset_pf_reason(void)
 245 {
 246         u32 reason = 0;
 247
 248         if (__this_cpu_read(apf_reason.enabled)) {
 249                 reason = __this_cpu_read(apf_reason.reason);
 250                 __this_cpu_write(apf_reason.reason, 0);
 251         }
 252
 253         return reason;
 254 }
 255 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 256 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 257
 258 dotraplinkage void
 259 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 260 {
 261         enum ctx_state prev_state;
 262
 263         switch (kvm_read_and_reset_pf_reason()) {
 264         default:
 265                 trace_do_page_fault(regs, error_code);
 266                 break;
 267         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 268                 /* page is swapped out by the host. */
 269                 prev_state = exception_enter();
 270                 exit_idle();
 271                 kvm_async_pf_task_wait((u32)read_cr2());
 272                 exception_exit(prev_state);
 273                 break;
 274         case KVM_PV_REASON_PAGE_READY:
 275                 rcu_irq_enter();
 276                 exit_idle();
 277                 kvm_async_pf_task_wake((u32)read_cr2());
 278                 rcu_irq_exit();
 279                 break;
 280         }
 281 }
 282 NOKPROBE_SYMBOL(do_async_page_fault);
 283
 284 static void __init paravirt_ops_setup(void)
 285 {
 286         pv_info.name = "KVM";
 287
 288         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 289                 pv_cpu_ops.io_delay = kvm_io_delay;
 290
 291 #ifdef CONFIG_X86_IO_APIC
 292         no_timer_check = 1;
 293 #endif
 294 }
 295
 296 static void kvm_register_steal_time(void)
 297 {
 298         int cpu = smp_processor_id();
 299         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 300
 301         if (!has_steal_clock)
 302                 return;
 303
 304         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 305         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
 306                 cpu, (unsigned long long) slow_virt_to_phys(st));
 307 }
 308
 309 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 310
 311 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 312 {
 313         /**
 314          * This relies on __test_and_clear_bit to modify the memory
 315          * in a way that is atomic with respect to the local CPU.
 316          * The hypervisor only accesses this memory from the local CPU so
 317          * there's no need for lock or memory barriers.
 318          * An optimization barrier is implied in apic write.
 319          */
 320         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 321                 return;
 322         apic_write(APIC_EOI, APIC_EOI_ACK);
 323 }
 324
 325 static void kvm_guest_cpu_init(void)
 326 {
 327         if (!kvm_para_available())
 328                 return;
 329
 330         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 331                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 332
 333 #ifdef CONFIG_PREEMPT
 334                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 335 #endif
 336                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
 337                 __this_cpu_write(apf_reason.enabled, 1);
 338                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 339                        smp_processor_id());
 340         }
 341
 342         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 343                 unsigned long pa;
 344                 /* Size alignment is implied but just to make it explicit. */
 345                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 346                 __this_cpu_write(kvm_apic_eoi, 0);
 347                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 348                         | KVM_MSR_ENABLED;
 349                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 350         }
 351
 352         if (has_steal_clock)
 353                 kvm_register_steal_time();
 354 }
 355
 356 static void kvm_pv_disable_apf(void)
 357 {
 358         if (!__this_cpu_read(apf_reason.enabled))
 359                 return;
 360
 361         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 362         __this_cpu_write(apf_reason.enabled, 0);
 363
 364         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 365                smp_processor_id());
 366 }
 367
 368 static void kvm_pv_guest_cpu_reboot(void *unused)
 369 {
 370         /*
 371          * We disable PV EOI before we load a new kernel by kexec,
 372          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 373          * New kernel can re-enable when it boots.
 374          */
 375         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 376                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 377         kvm_pv_disable_apf();
 378         kvm_disable_steal_time();
 379 }
 380
 381 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 382                                 unsigned long code, void *unused)
 383 {
 384         if (code == SYS_RESTART)
 385                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 386         return NOTIFY_DONE;
 387 }
 388
 389 static struct notifier_block kvm_pv_reboot_nb = {
 390         .notifier_call = kvm_pv_reboot_notify,
 391 };
 392
 393 static u64 kvm_steal_clock(int cpu)
 394 {
 395         u64 steal;
 396         struct kvm_steal_time *src;
 397         int version;
 398
 399         src = &per_cpu(steal_time, cpu);
 400         do {
 401                 version = src->version;
 402                 rmb();
 403                 steal = src->steal;
 404                 rmb();
 405         } while ((version & 1) || (version != src->version));
 406
 407         return steal;
 408 }
 409
 410 void kvm_disable_steal_time(void)
 411 {
 412         if (!has_steal_clock)
 413                 return;
 414
 415         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 416 }
 417
 418 #ifdef CONFIG_SMP
 419 static void __init kvm_smp_prepare_boot_cpu(void)
 420 {
 421         kvm_guest_cpu_init();
 422         native_smp_prepare_boot_cpu();
 423         kvm_spinlock_init();
 424 }
 425
 426 static void kvm_guest_cpu_online(void *dummy)
 427 {
 428         kvm_guest_cpu_init();
 429 }
 430
 431 static void kvm_guest_cpu_offline(void *dummy)
 432 {
 433         kvm_disable_steal_time();
 434         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 435                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 436         kvm_pv_disable_apf();
 437         apf_task_wake_all();
 438 }
 439
 440 static int kvm_cpu_notify(struct notifier_block *self, unsigned long action,
 441                           void *hcpu)
 442 {
 443         int cpu = (unsigned long)hcpu;
 444         switch (action) {
 445         case CPU_ONLINE:
 446         case CPU_DOWN_FAILED:
 447         case CPU_ONLINE_FROZEN:
 448                 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0);
 449                 break;
 450         case CPU_DOWN_PREPARE:
 451         case CPU_DOWN_PREPARE_FROZEN:
 452                 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1);
 453                 break;
 454         default:
 455                 break;
 456         }
 457         return NOTIFY_OK;
 458 }
 459
 460 static struct notifier_block kvm_cpu_notifier = {
 461         .notifier_call  = kvm_cpu_notify,
 462 };
 463 #endif
 464
 465 static void __init kvm_apf_trap_init(void)
 466 {
 467         set_intr_gate(14, async_page_fault);
 468 }
 469
 470 void __init kvm_guest_init(void)
 471 {
 472         int i;
 473
 474         if (!kvm_para_available())
 475                 return;
 476
 477         paravirt_ops_setup();
 478         register_reboot_notifier(&kvm_pv_reboot_nb);
 479         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 480                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 481         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 482                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 483
 484         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 485                 has_steal_clock = 1;
 486                 pv_time_ops.steal_clock = kvm_steal_clock;
 487         }
 488
 489         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 490                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 491
 492         if (kvmclock_vsyscall)
 493                 kvm_setup_vsyscall_timeinfo();
 494
 495 #ifdef CONFIG_SMP
 496         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 497         register_cpu_notifier(&kvm_cpu_notifier);
 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 struct hypervisor_x86 x86_hyper_kvm __refconst = {
 548         .name                   = "KVM",
 549         .detect                 = kvm_detect,
 550         .x2apic_available       = kvm_para_available,
 551 };
 552 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
 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 /*
 607  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 608  */
 609 void __init kvm_spinlock_init(void)
 610 {
 611         if (!kvm_para_available())
 612                 return;
 613         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 614         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 615                 return;
 616
 617         __pv_init_lock_hash();
 618         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 619         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 620         pv_lock_ops.wait = kvm_wait;
 621         pv_lock_ops.kick = kvm_kick_cpu;
 622 }
 623
 624 static __init int kvm_spinlock_init_jump(void)
 625 {
 626         if (!kvm_para_available())
 627                 return 0;
 628         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 629                 return 0;
 630
 631         static_key_slow_inc(&paravirt_ticketlocks_enabled);
 632         printk(KERN_INFO "KVM setup paravirtual spinlock\n");
 633
 634         return 0;
 635 }
 636 early_initcall(kvm_spinlock_init_jump);
 637
 638 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */