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                 if (!n.halted) {
 155                         local_irq_enable();
 156                         schedule();
 157                         local_irq_disable();
 158                 } else {
 159                         /*
 160                          * We cannot reschedule. So halt.
 161                          */
 162                         rcu_irq_exit();
 163                         native_safe_halt();
 164                         local_irq_disable();
 165                         rcu_irq_enter();
 166                 }
 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 (swait_active(&n->wq))
 182                 swake_up(&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                 trace_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());
 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(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                 /* Async page fault support for L1 hypervisor is optional */
 336                 if (wrmsr_safe(MSR_KVM_ASYNC_PF_EN,
 337                         (pa | KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT) & 0xffffffff, pa >> 32) < 0)
 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 #ifdef CONFIG_SMP
 421 static void __init kvm_smp_prepare_boot_cpu(void)
 422 {
 423         kvm_guest_cpu_init();
 424         native_smp_prepare_boot_cpu();
 425         kvm_spinlock_init();
 426 }
 427
 428 static void kvm_guest_cpu_offline(void)
 429 {
 430         kvm_disable_steal_time();
 431         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 432                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 433         kvm_pv_disable_apf();
 434         apf_task_wake_all();
 435 }
 436
 437 static int kvm_cpu_online(unsigned int cpu)
 438 {
 439         local_irq_disable();
 440         kvm_guest_cpu_init();
 441         local_irq_enable();
 442         return 0;
 443 }
 444
 445 static int kvm_cpu_down_prepare(unsigned int cpu)
 446 {
 447         local_irq_disable();
 448         kvm_guest_cpu_offline();
 449         local_irq_enable();
 450         return 0;
 451 }
 452 #endif
 453
 454 static void __init kvm_apf_trap_init(void)
 455 {
 456         set_intr_gate(14, async_page_fault);
 457 }
 458
 459 void __init kvm_guest_init(void)
 460 {
 461         int i;
 462
 463         if (!kvm_para_available())
 464                 return;
 465
 466         paravirt_ops_setup();
 467         register_reboot_notifier(&kvm_pv_reboot_nb);
 468         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 469                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 470         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 471                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 472
 473         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 474                 has_steal_clock = 1;
 475                 pv_time_ops.steal_clock = kvm_steal_clock;
 476         }
 477
 478         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 479                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 480
 481         if (kvmclock_vsyscall)
 482                 kvm_setup_vsyscall_timeinfo();
 483
 484 #ifdef CONFIG_SMP
 485         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 486         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 487                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 488                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 489 #else
 490         kvm_guest_cpu_init();
 491 #endif
 492
 493         /*
 494          * Hard lockup detection is enabled by default. Disable it, as guests
 495          * can get false positives too easily, for example if the host is
 496          * overcommitted.
 497          */
 498         hardlockup_detector_disable();
 499 }
 500
 501 static noinline uint32_t __kvm_cpuid_base(void)
 502 {
 503         if (boot_cpu_data.cpuid_level < 0)
 504                 return 0;       /* So we don't blow up on old processors */
 505
 506         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 507                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 508
 509         return 0;
 510 }
 511
 512 static inline uint32_t kvm_cpuid_base(void)
 513 {
 514         static int kvm_cpuid_base = -1;
 515
 516         if (kvm_cpuid_base == -1)
 517                 kvm_cpuid_base = __kvm_cpuid_base();
 518
 519         return kvm_cpuid_base;
 520 }
 521
 522 bool kvm_para_available(void)
 523 {
 524         return kvm_cpuid_base() != 0;
 525 }
 526 EXPORT_SYMBOL_GPL(kvm_para_available);
 527
 528 unsigned int kvm_arch_para_features(void)
 529 {
 530         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 531 }
 532
 533 static uint32_t __init kvm_detect(void)
 534 {
 535         return kvm_cpuid_base();
 536 }
 537
 538 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
 539         .name                   = "KVM",
 540         .detect                 = kvm_detect,
 541         .x2apic_available       = kvm_para_available,
 542 };
 543 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
 544
 545 static __init int activate_jump_labels(void)
 546 {
 547         if (has_steal_clock) {
 548                 static_key_slow_inc(&paravirt_steal_enabled);
 549                 if (steal_acc)
 550                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 551         }
 552
 553         return 0;
 554 }
 555 arch_initcall(activate_jump_labels);
 556
 557 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 558
 559 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 560 static void kvm_kick_cpu(int cpu)
 561 {
 562         int apicid;
 563         unsigned long flags = 0;
 564
 565         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 566         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 567 }
 568
 569 #include <asm/qspinlock.h>
 570
 571 static void kvm_wait(u8 *ptr, u8 val)
 572 {
 573         unsigned long flags;
 574
 575         if (in_nmi())
 576                 return;
 577
 578         local_irq_save(flags);
 579
 580         if (READ_ONCE(*ptr) != val)
 581                 goto out;
 582
 583         /*
 584          * halt until it's our turn and kicked. Note that we do safe halt
 585          * for irq enabled case to avoid hang when lock info is overwritten
 586          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 587          */
 588         if (arch_irqs_disabled_flags(flags))
 589                 halt();
 590         else
 591                 safe_halt();
 592
 593 out:
 594         local_irq_restore(flags);
 595 }
 596
 597 #ifdef CONFIG_X86_32
 598 __visible bool __kvm_vcpu_is_preempted(long cpu)
 599 {
 600         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 601
 602         return !!src->preempted;
 603 }
 604 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 605
 606 #else
 607
 608 #include <asm/asm-offsets.h>
 609
 610 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 611
 612 /*
 613  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 614  * restoring to/from the stack.
 615  */
 616 asm(
 617 ".pushsection .text;"
 618 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 619 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 620 "__raw_callee_save___kvm_vcpu_is_preempted:"
 621 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 622 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 623 "setne  %al;"
 624 "ret;"
 625 ".popsection");
 626
 627 #endif
 628
 629 /*
 630  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 631  */
 632 void __init kvm_spinlock_init(void)
 633 {
 634         if (!kvm_para_available())
 635                 return;
 636         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 637         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 638                 return;
 639
 640         __pv_init_lock_hash();
 641         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 642         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 643         pv_lock_ops.wait = kvm_wait;
 644         pv_lock_ops.kick = kvm_kick_cpu;
 645
 646         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 647                 pv_lock_ops.vcpu_is_preempted =
 648                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 649         }
 650 }
 651
 652 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */