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                    rcu_preempt_depth();
 146         init_swait_queue_head(&n.wq);
 147         hlist_add_head(&n.link, &b->list);
 148         raw_spin_unlock(&b->lock);
 149
 150         for (;;) {
 151                 if (!n.halted)
 152                         prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 153                 if (hlist_unhashed(&n.link))
 154                         break;
 155
 156                 rcu_irq_exit();
 157
 158                 if (!n.halted) {
 159                         local_irq_enable();
 160                         schedule();
 161                         local_irq_disable();
 162                 } else {
 163                         /*
 164                          * We cannot reschedule. So halt.
 165                          */
 166                         native_safe_halt();
 167                         local_irq_disable();
 168                 }
 169
 170                 rcu_irq_enter();
 171         }
 172         if (!n.halted)
 173                 finish_swait(&n.wq, &wait);
 174
 175         rcu_irq_exit();
 176         return;
 177 }
 178 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 179
 180 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 181 {
 182         hlist_del_init(&n->link);
 183         if (n->halted)
 184                 smp_send_reschedule(n->cpu);
 185         else if (swait_active(&n->wq))
 186                 swake_up(&n->wq);
 187 }
 188
 189 static void apf_task_wake_all(void)
 190 {
 191         int i;
 192
 193         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 194                 struct hlist_node *p, *next;
 195                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 196                 raw_spin_lock(&b->lock);
 197                 hlist_for_each_safe(p, next, &b->list) {
 198                         struct kvm_task_sleep_node *n =
 199                                 hlist_entry(p, typeof(*n), link);
 200                         if (n->cpu == smp_processor_id())
 201                                 apf_task_wake_one(n);
 202                 }
 203                 raw_spin_unlock(&b->lock);
 204         }
 205 }
 206
 207 void kvm_async_pf_task_wake(u32 token)
 208 {
 209         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 210         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 211         struct kvm_task_sleep_node *n;
 212
 213         if (token == ~0) {
 214                 apf_task_wake_all();
 215                 return;
 216         }
 217
 218 again:
 219         raw_spin_lock(&b->lock);
 220         n = _find_apf_task(b, token);
 221         if (!n) {
 222                 /*
 223                  * async PF was not yet handled.
 224                  * Add dummy entry for the token.
 225                  */
 226                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 227                 if (!n) {
 228                         /*
 229                          * Allocation failed! Busy wait while other cpu
 230                          * handles async PF.
 231                          */
 232                         raw_spin_unlock(&b->lock);
 233                         cpu_relax();
 234                         goto again;
 235                 }
 236                 n->token = token;
 237                 n->cpu = smp_processor_id();
 238                 init_swait_queue_head(&n->wq);
 239                 hlist_add_head(&n->link, &b->list);
 240         } else
 241                 apf_task_wake_one(n);
 242         raw_spin_unlock(&b->lock);
 243         return;
 244 }
 245 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 246
 247 u32 kvm_read_and_reset_pf_reason(void)
 248 {
 249         u32 reason = 0;
 250
 251         if (__this_cpu_read(apf_reason.enabled)) {
 252                 reason = __this_cpu_read(apf_reason.reason);
 253                 __this_cpu_write(apf_reason.reason, 0);
 254         }
 255
 256         return reason;
 257 }
 258 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 259 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 260
 261 dotraplinkage void
 262 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 263 {
 264         enum ctx_state prev_state;
 265
 266         switch (kvm_read_and_reset_pf_reason()) {
 267         default:
 268                 trace_do_page_fault(regs, error_code);
 269                 break;
 270         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 271                 /* page is swapped out by the host. */
 272                 prev_state = exception_enter();
 273                 exit_idle();
 274                 kvm_async_pf_task_wait((u32)read_cr2());
 275                 exception_exit(prev_state);
 276                 break;
 277         case KVM_PV_REASON_PAGE_READY:
 278                 rcu_irq_enter();
 279                 exit_idle();
 280                 kvm_async_pf_task_wake((u32)read_cr2());
 281                 rcu_irq_exit();
 282                 break;
 283         }
 284 }
 285 NOKPROBE_SYMBOL(do_async_page_fault);
 286
 287 static void __init paravirt_ops_setup(void)
 288 {
 289         pv_info.name = "KVM";
 290
 291         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 292                 pv_cpu_ops.io_delay = kvm_io_delay;
 293
 294 #ifdef CONFIG_X86_IO_APIC
 295         no_timer_check = 1;
 296 #endif
 297 }
 298
 299 static void kvm_register_steal_time(void)
 300 {
 301         int cpu = smp_processor_id();
 302         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 303
 304         if (!has_steal_clock)
 305                 return;
 306
 307         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 308         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
 309                 cpu, (unsigned long long) slow_virt_to_phys(st));
 310 }
 311
 312 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 313
 314 static void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 315 {
 316         /**
 317          * This relies on __test_and_clear_bit to modify the memory
 318          * in a way that is atomic with respect to the local CPU.
 319          * The hypervisor only accesses this memory from the local CPU so
 320          * there's no need for lock or memory barriers.
 321          * An optimization barrier is implied in apic write.
 322          */
 323         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 324                 return;
 325         apic_write(APIC_EOI, APIC_EOI_ACK);
 326 }
 327
 328 static void kvm_guest_cpu_init(void)
 329 {
 330         if (!kvm_para_available())
 331                 return;
 332
 333         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 334                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 335
 336 #ifdef CONFIG_PREEMPT
 337                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 338 #endif
 339                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
 340                 __this_cpu_write(apf_reason.enabled, 1);
 341                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 342                        smp_processor_id());
 343         }
 344
 345         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 346                 unsigned long pa;
 347                 /* Size alignment is implied but just to make it explicit. */
 348                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 349                 __this_cpu_write(kvm_apic_eoi, 0);
 350                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 351                         | KVM_MSR_ENABLED;
 352                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 353         }
 354
 355         if (has_steal_clock)
 356                 kvm_register_steal_time();
 357 }
 358
 359 static void kvm_pv_disable_apf(void)
 360 {
 361         if (!__this_cpu_read(apf_reason.enabled))
 362                 return;
 363
 364         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 365         __this_cpu_write(apf_reason.enabled, 0);
 366
 367         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 368                smp_processor_id());
 369 }
 370
 371 static void kvm_pv_guest_cpu_reboot(void *unused)
 372 {
 373         /*
 374          * We disable PV EOI before we load a new kernel by kexec,
 375          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 376          * New kernel can re-enable when it boots.
 377          */
 378         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 379                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 380         kvm_pv_disable_apf();
 381         kvm_disable_steal_time();
 382 }
 383
 384 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 385                                 unsigned long code, void *unused)
 386 {
 387         if (code == SYS_RESTART)
 388                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 389         return NOTIFY_DONE;
 390 }
 391
 392 static struct notifier_block kvm_pv_reboot_nb = {
 393         .notifier_call = kvm_pv_reboot_notify,
 394 };
 395
 396 static u64 kvm_steal_clock(int cpu)
 397 {
 398         u64 steal;
 399         struct kvm_steal_time *src;
 400         int version;
 401
 402         src = &per_cpu(steal_time, cpu);
 403         do {
 404                 version = src->version;
 405                 rmb();
 406                 steal = src->steal;
 407                 rmb();
 408         } while ((version & 1) || (version != src->version));
 409
 410         return steal;
 411 }
 412
 413 void kvm_disable_steal_time(void)
 414 {
 415         if (!has_steal_clock)
 416                 return;
 417
 418         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 419 }
 420
 421 #ifdef CONFIG_SMP
 422 static void __init kvm_smp_prepare_boot_cpu(void)
 423 {
 424         kvm_guest_cpu_init();
 425         native_smp_prepare_boot_cpu();
 426         kvm_spinlock_init();
 427 }
 428
 429 static void kvm_guest_cpu_offline(void)
 430 {
 431         kvm_disable_steal_time();
 432         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 433                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 434         kvm_pv_disable_apf();
 435         apf_task_wake_all();
 436 }
 437
 438 static int kvm_cpu_online(unsigned int cpu)
 439 {
 440         local_irq_disable();
 441         kvm_guest_cpu_init();
 442         local_irq_enable();
 443         return 0;
 444 }
 445
 446 static int kvm_cpu_down_prepare(unsigned int cpu)
 447 {
 448         local_irq_disable();
 449         kvm_guest_cpu_offline();
 450         local_irq_enable();
 451         return 0;
 452 }
 453 #endif
 454
 455 static void __init kvm_apf_trap_init(void)
 456 {
 457         set_intr_gate(14, async_page_fault);
 458 }
 459
 460 void __init kvm_guest_init(void)
 461 {
 462         int i;
 463
 464         if (!kvm_para_available())
 465                 return;
 466
 467         paravirt_ops_setup();
 468         register_reboot_notifier(&kvm_pv_reboot_nb);
 469         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 470                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 471         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 472                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 473
 474         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 475                 has_steal_clock = 1;
 476                 pv_time_ops.steal_clock = kvm_steal_clock;
 477         }
 478
 479         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 480                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 481
 482         if (kvmclock_vsyscall)
 483                 kvm_setup_vsyscall_timeinfo();
 484
 485 #ifdef CONFIG_SMP
 486         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 487         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 488                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 489                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 490 #else
 491         kvm_guest_cpu_init();
 492 #endif
 493
 494         /*
 495          * Hard lockup detection is enabled by default. Disable it, as guests
 496          * can get false positives too easily, for example if the host is
 497          * overcommitted.
 498          */
 499         hardlockup_detector_disable();
 500 }
 501
 502 static noinline uint32_t __kvm_cpuid_base(void)
 503 {
 504         if (boot_cpu_data.cpuid_level < 0)
 505                 return 0;       /* So we don't blow up on old processors */
 506
 507         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 508                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 509
 510         return 0;
 511 }
 512
 513 static inline uint32_t kvm_cpuid_base(void)
 514 {
 515         static int kvm_cpuid_base = -1;
 516
 517         if (kvm_cpuid_base == -1)
 518                 kvm_cpuid_base = __kvm_cpuid_base();
 519
 520         return kvm_cpuid_base;
 521 }
 522
 523 bool kvm_para_available(void)
 524 {
 525         return kvm_cpuid_base() != 0;
 526 }
 527 EXPORT_SYMBOL_GPL(kvm_para_available);
 528
 529 unsigned int kvm_arch_para_features(void)
 530 {
 531         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 532 }
 533
 534 static uint32_t __init kvm_detect(void)
 535 {
 536         return kvm_cpuid_base();
 537 }
 538
 539 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
 540         .name                   = "KVM",
 541         .detect                 = kvm_detect,
 542         .x2apic_available       = kvm_para_available,
 543 };
 544 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
 545
 546 static __init int activate_jump_labels(void)
 547 {
 548         if (has_steal_clock) {
 549                 static_key_slow_inc(&paravirt_steal_enabled);
 550                 if (steal_acc)
 551                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 552         }
 553
 554         return 0;
 555 }
 556 arch_initcall(activate_jump_labels);
 557
 558 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 559
 560 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 561 static void kvm_kick_cpu(int cpu)
 562 {
 563         int apicid;
 564         unsigned long flags = 0;
 565
 566         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 567         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 568 }
 569
 570 #include <asm/qspinlock.h>
 571
 572 static void kvm_wait(u8 *ptr, u8 val)
 573 {
 574         unsigned long flags;
 575
 576         if (in_nmi())
 577                 return;
 578
 579         local_irq_save(flags);
 580
 581         if (READ_ONCE(*ptr) != val)
 582                 goto out;
 583
 584         /*
 585          * halt until it's our turn and kicked. Note that we do safe halt
 586          * for irq enabled case to avoid hang when lock info is overwritten
 587          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 588          */
 589         if (arch_irqs_disabled_flags(flags))
 590                 halt();
 591         else
 592                 safe_halt();
 593
 594 out:
 595         local_irq_restore(flags);
 596 }
 597
 598 /*
 599  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 600  */
 601 void __init kvm_spinlock_init(void)
 602 {
 603         if (!kvm_para_available())
 604                 return;
 605         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 606         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 607                 return;
 608
 609         __pv_init_lock_hash();
 610         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 611         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 612         pv_lock_ops.wait = kvm_wait;
 613         pv_lock_ops.kick = kvm_kick_cpu;
 614 }
 615
 616 static __init int kvm_spinlock_init_jump(void)
 617 {
 618         if (!kvm_para_available())
 619                 return 0;
 620         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 621                 return 0;
 622
 623         static_key_slow_inc(&paravirt_ticketlocks_enabled);
 624         printk(KERN_INFO "KVM setup paravirtual spinlock\n");
 625
 626         return 0;
 627 }
 628 early_initcall(kvm_spinlock_init_jump);
 629
 630 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */