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 void kvm_async_pf_task_wait(u32 token)
 121 {
 122         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 123         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 124         struct kvm_task_sleep_node n, *e;
 125         DECLARE_SWAITQUEUE(wait);
 126
 127         rcu_irq_enter();
 128
 129         raw_spin_lock(&b->lock);
 130         e = _find_apf_task(b, token);
 131         if (e) {
 132                 /* dummy entry exist -> wake up was delivered ahead of PF */
 133                 hlist_del(&e->link);
 134                 kfree(e);
 135                 raw_spin_unlock(&b->lock);
 136
 137                 rcu_irq_exit();
 138                 return;
 139         }
 140
 141         n.token = token;
 142         n.cpu = smp_processor_id();
 143         n.halted = is_idle_task(current) || preempt_count() > 1;
 144         init_swait_queue_head(&n.wq);
 145         hlist_add_head(&n.link, &b->list);
 146         raw_spin_unlock(&b->lock);
 147
 148         for (;;) {
 149                 if (!n.halted)
 150                         prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 151                 if (hlist_unhashed(&n.link))
 152                         break;
 153
 154                 rcu_irq_exit();
 155
 156                 if (!n.halted) {
 157                         local_irq_enable();
 158                         schedule();
 159                         local_irq_disable();
 160                 } else {
 161                         /*
 162                          * We cannot reschedule. So halt.
 163                          */
 164                         native_safe_halt();
 165                         local_irq_disable();
 166                 }
 167
 168                 rcu_irq_enter();
 169         }
 170         if (!n.halted)
 171                 finish_swait(&n.wq, &wait);
 172
 173         rcu_irq_exit();
 174         return;
 175 }
 176 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 177
 178 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 179 {
 180         hlist_del_init(&n->link);
 181         if (n->halted)
 182                 smp_send_reschedule(n->cpu);
 183         else if (swait_active(&n->wq))
 184                 swake_up(&n->wq);
 185 }
 186
 187 static void apf_task_wake_all(void)
 188 {
 189         int i;
 190
 191         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 192                 struct hlist_node *p, *next;
 193                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 194                 raw_spin_lock(&b->lock);
 195                 hlist_for_each_safe(p, next, &b->list) {
 196                         struct kvm_task_sleep_node *n =
 197                                 hlist_entry(p, typeof(*n), link);
 198                         if (n->cpu == smp_processor_id())
 199                                 apf_task_wake_one(n);
 200                 }
 201                 raw_spin_unlock(&b->lock);
 202         }
 203 }
 204
 205 void kvm_async_pf_task_wake(u32 token)
 206 {
 207         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 208         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 209         struct kvm_task_sleep_node *n;
 210
 211         if (token == ~0) {
 212                 apf_task_wake_all();
 213                 return;
 214         }
 215
 216 again:
 217         raw_spin_lock(&b->lock);
 218         n = _find_apf_task(b, token);
 219         if (!n) {
 220                 /*
 221                  * async PF was not yet handled.
 222                  * Add dummy entry for the token.
 223                  */
 224                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 225                 if (!n) {
 226                         /*
 227                          * Allocation failed! Busy wait while other cpu
 228                          * handles async PF.
 229                          */
 230                         raw_spin_unlock(&b->lock);
 231                         cpu_relax();
 232                         goto again;
 233                 }
 234                 n->token = token;
 235                 n->cpu = smp_processor_id();
 236                 init_swait_queue_head(&n->wq);
 237                 hlist_add_head(&n->link, &b->list);
 238         } else
 239                 apf_task_wake_one(n);
 240         raw_spin_unlock(&b->lock);
 241         return;
 242 }
 243 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 244
 245 u32 kvm_read_and_reset_pf_reason(void)
 246 {
 247         u32 reason = 0;
 248
 249         if (__this_cpu_read(apf_reason.enabled)) {
 250                 reason = __this_cpu_read(apf_reason.reason);
 251                 __this_cpu_write(apf_reason.reason, 0);
 252         }
 253
 254         return reason;
 255 }
 256 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 257 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 258
 259 dotraplinkage void
 260 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 261 {
 262         enum ctx_state prev_state;
 263
 264         switch (kvm_read_and_reset_pf_reason()) {
 265         default:
 266                 trace_do_page_fault(regs, error_code);
 267                 break;
 268         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 269                 /* page is swapped out by the host. */
 270                 prev_state = exception_enter();
 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                 kvm_async_pf_task_wake((u32)read_cr2());
 277                 rcu_irq_exit();
 278                 break;
 279         }
 280 }
 281 NOKPROBE_SYMBOL(do_async_page_fault);
 282
 283 static void __init paravirt_ops_setup(void)
 284 {
 285         pv_info.name = "KVM";
 286
 287         if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
 288                 pv_cpu_ops.io_delay = kvm_io_delay;
 289
 290 #ifdef CONFIG_X86_IO_APIC
 291         no_timer_check = 1;
 292 #endif
 293 }
 294
 295 static void kvm_register_steal_time(void)
 296 {
 297         int cpu = smp_processor_id();
 298         struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
 299
 300         if (!has_steal_clock)
 301                 return;
 302
 303         wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
 304         pr_info("kvm-stealtime: cpu %d, msr %llx\n",
 305                 cpu, (unsigned long long) slow_virt_to_phys(st));
 306 }
 307
 308 static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
 309
 310 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
 311 {
 312         /**
 313          * This relies on __test_and_clear_bit to modify the memory
 314          * in a way that is atomic with respect to the local CPU.
 315          * The hypervisor only accesses this memory from the local CPU so
 316          * there's no need for lock or memory barriers.
 317          * An optimization barrier is implied in apic write.
 318          */
 319         if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
 320                 return;
 321         apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
 322 }
 323
 324 static void kvm_guest_cpu_init(void)
 325 {
 326         if (!kvm_para_available())
 327                 return;
 328
 329         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
 330                 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
 331
 332 #ifdef CONFIG_PREEMPT
 333                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 334 #endif
 335                 pa |= KVM_ASYNC_PF_ENABLED;
 336
 337                 /* Async page fault support for L1 hypervisor is optional */
 338                 if (wrmsr_safe(MSR_KVM_ASYNC_PF_EN,
 339                         (pa | KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT) & 0xffffffff, pa >> 32) < 0)
 340                         wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 341                 __this_cpu_write(apf_reason.enabled, 1);
 342                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 343                        smp_processor_id());
 344         }
 345
 346         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 347                 unsigned long pa;
 348                 /* Size alignment is implied but just to make it explicit. */
 349                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 350                 __this_cpu_write(kvm_apic_eoi, 0);
 351                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 352                         | KVM_MSR_ENABLED;
 353                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 354         }
 355
 356         if (has_steal_clock)
 357                 kvm_register_steal_time();
 358 }
 359
 360 static void kvm_pv_disable_apf(void)
 361 {
 362         if (!__this_cpu_read(apf_reason.enabled))
 363                 return;
 364
 365         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 366         __this_cpu_write(apf_reason.enabled, 0);
 367
 368         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 369                smp_processor_id());
 370 }
 371
 372 static void kvm_pv_guest_cpu_reboot(void *unused)
 373 {
 374         /*
 375          * We disable PV EOI before we load a new kernel by kexec,
 376          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 377          * New kernel can re-enable when it boots.
 378          */
 379         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 380                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 381         kvm_pv_disable_apf();
 382         kvm_disable_steal_time();
 383 }
 384
 385 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 386                                 unsigned long code, void *unused)
 387 {
 388         if (code == SYS_RESTART)
 389                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 390         return NOTIFY_DONE;
 391 }
 392
 393 static struct notifier_block kvm_pv_reboot_nb = {
 394         .notifier_call = kvm_pv_reboot_notify,
 395 };
 396
 397 static u64 kvm_steal_clock(int cpu)
 398 {
 399         u64 steal;
 400         struct kvm_steal_time *src;
 401         int version;
 402
 403         src = &per_cpu(steal_time, cpu);
 404         do {
 405                 version = src->version;
 406                 virt_rmb();
 407                 steal = src->steal;
 408                 virt_rmb();
 409         } while ((version & 1) || (version != src->version));
 410
 411         return steal;
 412 }
 413
 414 void kvm_disable_steal_time(void)
 415 {
 416         if (!has_steal_clock)
 417                 return;
 418
 419         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 420 }
 421
 422 #ifdef CONFIG_SMP
 423 static void __init kvm_smp_prepare_boot_cpu(void)
 424 {
 425         kvm_guest_cpu_init();
 426         native_smp_prepare_boot_cpu();
 427         kvm_spinlock_init();
 428 }
 429
 430 static void kvm_guest_cpu_offline(void)
 431 {
 432         kvm_disable_steal_time();
 433         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 434                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 435         kvm_pv_disable_apf();
 436         apf_task_wake_all();
 437 }
 438
 439 static int kvm_cpu_online(unsigned int cpu)
 440 {
 441         local_irq_disable();
 442         kvm_guest_cpu_init();
 443         local_irq_enable();
 444         return 0;
 445 }
 446
 447 static int kvm_cpu_down_prepare(unsigned int cpu)
 448 {
 449         local_irq_disable();
 450         kvm_guest_cpu_offline();
 451         local_irq_enable();
 452         return 0;
 453 }
 454 #endif
 455
 456 static void __init kvm_apf_trap_init(void)
 457 {
 458         set_intr_gate(14, async_page_fault);
 459 }
 460
 461 void __init kvm_guest_init(void)
 462 {
 463         int i;
 464
 465         if (!kvm_para_available())
 466                 return;
 467
 468         paravirt_ops_setup();
 469         register_reboot_notifier(&kvm_pv_reboot_nb);
 470         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 471                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 472         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 473                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 474
 475         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 476                 has_steal_clock = 1;
 477                 pv_time_ops.steal_clock = kvm_steal_clock;
 478         }
 479
 480         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 481                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 482
 483         if (kvmclock_vsyscall)
 484                 kvm_setup_vsyscall_timeinfo();
 485
 486 #ifdef CONFIG_SMP
 487         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 488         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 489                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 490                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 491 #else
 492         kvm_guest_cpu_init();
 493 #endif
 494
 495         /*
 496          * Hard lockup detection is enabled by default. Disable it, as guests
 497          * can get false positives too easily, for example if the host is
 498          * overcommitted.
 499          */
 500         hardlockup_detector_disable();
 501 }
 502
 503 static noinline uint32_t __kvm_cpuid_base(void)
 504 {
 505         if (boot_cpu_data.cpuid_level < 0)
 506                 return 0;       /* So we don't blow up on old processors */
 507
 508         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 509                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 510
 511         return 0;
 512 }
 513
 514 static inline uint32_t kvm_cpuid_base(void)
 515 {
 516         static int kvm_cpuid_base = -1;
 517
 518         if (kvm_cpuid_base == -1)
 519                 kvm_cpuid_base = __kvm_cpuid_base();
 520
 521         return kvm_cpuid_base;
 522 }
 523
 524 bool kvm_para_available(void)
 525 {
 526         return kvm_cpuid_base() != 0;
 527 }
 528 EXPORT_SYMBOL_GPL(kvm_para_available);
 529
 530 unsigned int kvm_arch_para_features(void)
 531 {
 532         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 533 }
 534
 535 static uint32_t __init kvm_detect(void)
 536 {
 537         return kvm_cpuid_base();
 538 }
 539
 540 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
 541         .name                   = "KVM",
 542         .detect                 = kvm_detect,
 543         .x2apic_available       = kvm_para_available,
 544 };
 545 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
 546
 547 static __init int activate_jump_labels(void)
 548 {
 549         if (has_steal_clock) {
 550                 static_key_slow_inc(&paravirt_steal_enabled);
 551                 if (steal_acc)
 552                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 553         }
 554
 555         return 0;
 556 }
 557 arch_initcall(activate_jump_labels);
 558
 559 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 560
 561 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 562 static void kvm_kick_cpu(int cpu)
 563 {
 564         int apicid;
 565         unsigned long flags = 0;
 566
 567         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 568         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 569 }
 570
 571 #include <asm/qspinlock.h>
 572
 573 static void kvm_wait(u8 *ptr, u8 val)
 574 {
 575         unsigned long flags;
 576
 577         if (in_nmi())
 578                 return;
 579
 580         local_irq_save(flags);
 581
 582         if (READ_ONCE(*ptr) != val)
 583                 goto out;
 584
 585         /*
 586          * halt until it's our turn and kicked. Note that we do safe halt
 587          * for irq enabled case to avoid hang when lock info is overwritten
 588          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 589          */
 590         if (arch_irqs_disabled_flags(flags))
 591                 halt();
 592         else
 593                 safe_halt();
 594
 595 out:
 596         local_irq_restore(flags);
 597 }
 598
 599 #ifdef CONFIG_X86_32
 600 __visible bool __kvm_vcpu_is_preempted(long cpu)
 601 {
 602         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 603
 604         return !!src->preempted;
 605 }
 606 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 607
 608 #else
 609
 610 #include <asm/asm-offsets.h>
 611
 612 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 613
 614 /*
 615  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 616  * restoring to/from the stack.
 617  */
 618 asm(
 619 ".pushsection .text;"
 620 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 621 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 622 "__raw_callee_save___kvm_vcpu_is_preempted:"
 623 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 624 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 625 "setne  %al;"
 626 "ret;"
 627 ".popsection");
 628
 629 #endif
 630
 631 /*
 632  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 633  */
 634 void __init kvm_spinlock_init(void)
 635 {
 636         if (!kvm_para_available())
 637                 return;
 638         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 639         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 640                 return;
 641
 642         __pv_init_lock_hash();
 643         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 644         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 645         pv_lock_ops.wait = kvm_wait;
 646         pv_lock_ops.kick = kvm_kick_cpu;
 647
 648         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 649                 pv_lock_ops.vcpu_is_preempted =
 650                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 651         }
 652 }
 653
 654 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */