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                    rcu_preempt_depth();
 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                 rcu_irq_exit();
 156
 157                 if (!n.halted) {
 158                         local_irq_enable();
 159                         schedule();
 160                         local_irq_disable();
 161                 } else {
 162                         /*
 163                          * We cannot reschedule. So halt.
 164                          */
 165                         native_safe_halt();
 166                         local_irq_disable();
 167                 }
 168
 169                 rcu_irq_enter();
 170         }
 171         if (!n.halted)
 172                 finish_swait(&n.wq, &wait);
 173
 174         rcu_irq_exit();
 175         return;
 176 }
 177 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 178
 179 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 180 {
 181         hlist_del_init(&n->link);
 182         if (n->halted)
 183                 smp_send_reschedule(n->cpu);
 184         else if (swq_has_sleeper(&n->wq))
 185                 swake_up(&n->wq);
 186 }
 187
 188 static void apf_task_wake_all(void)
 189 {
 190         int i;
 191
 192         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 193                 struct hlist_node *p, *next;
 194                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 195                 raw_spin_lock(&b->lock);
 196                 hlist_for_each_safe(p, next, &b->list) {
 197                         struct kvm_task_sleep_node *n =
 198                                 hlist_entry(p, typeof(*n), link);
 199                         if (n->cpu == smp_processor_id())
 200                                 apf_task_wake_one(n);
 201                 }
 202                 raw_spin_unlock(&b->lock);
 203         }
 204 }
 205
 206 void kvm_async_pf_task_wake(u32 token)
 207 {
 208         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 209         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 210         struct kvm_task_sleep_node *n;
 211
 212         if (token == ~0) {
 213                 apf_task_wake_all();
 214                 return;
 215         }
 216
 217 again:
 218         raw_spin_lock(&b->lock);
 219         n = _find_apf_task(b, token);
 220         if (!n) {
 221                 /*
 222                  * async PF was not yet handled.
 223                  * Add dummy entry for the token.
 224                  */
 225                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 226                 if (!n) {
 227                         /*
 228                          * Allocation failed! Busy wait while other cpu
 229                          * handles async PF.
 230                          */
 231                         raw_spin_unlock(&b->lock);
 232                         cpu_relax();
 233                         goto again;
 234                 }
 235                 n->token = token;
 236                 n->cpu = smp_processor_id();
 237                 init_swait_queue_head(&n->wq);
 238                 hlist_add_head(&n->link, &b->list);
 239         } else
 240                 apf_task_wake_one(n);
 241         raw_spin_unlock(&b->lock);
 242         return;
 243 }
 244 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 245
 246 u32 kvm_read_and_reset_pf_reason(void)
 247 {
 248         u32 reason = 0;
 249
 250         if (__this_cpu_read(apf_reason.enabled)) {
 251                 reason = __this_cpu_read(apf_reason.reason);
 252                 __this_cpu_write(apf_reason.reason, 0);
 253         }
 254
 255         return reason;
 256 }
 257 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 258 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
 259
 260 dotraplinkage void
 261 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 262 {
 263         enum ctx_state prev_state;
 264
 265         switch (kvm_read_and_reset_pf_reason()) {
 266         default:
 267                 do_page_fault(regs, error_code);
 268                 break;
 269         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 270                 /* page is swapped out by the host. */
 271                 prev_state = exception_enter();
 272                 kvm_async_pf_task_wait((u32)read_cr2());
 273                 exception_exit(prev_state);
 274                 break;
 275         case KVM_PV_REASON_PAGE_READY:
 276                 rcu_irq_enter();
 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 notrace 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->native_eoi_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                 pa |= KVM_ASYNC_PF_ENABLED;
 337
 338                 /* Async page fault support for L1 hypervisor is optional */
 339                 if (wrmsr_safe(MSR_KVM_ASYNC_PF_EN,
 340                         (pa | KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT) & 0xffffffff, pa >> 32) < 0)
 341                         wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
 342                 __this_cpu_write(apf_reason.enabled, 1);
 343                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 344                        smp_processor_id());
 345         }
 346
 347         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 348                 unsigned long pa;
 349                 /* Size alignment is implied but just to make it explicit. */
 350                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 351                 __this_cpu_write(kvm_apic_eoi, 0);
 352                 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
 353                         | KVM_MSR_ENABLED;
 354                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 355         }
 356
 357         if (has_steal_clock)
 358                 kvm_register_steal_time();
 359 }
 360
 361 static void kvm_pv_disable_apf(void)
 362 {
 363         if (!__this_cpu_read(apf_reason.enabled))
 364                 return;
 365
 366         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 367         __this_cpu_write(apf_reason.enabled, 0);
 368
 369         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 370                smp_processor_id());
 371 }
 372
 373 static void kvm_pv_guest_cpu_reboot(void *unused)
 374 {
 375         /*
 376          * We disable PV EOI before we load a new kernel by kexec,
 377          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 378          * New kernel can re-enable when it boots.
 379          */
 380         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 381                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 382         kvm_pv_disable_apf();
 383         kvm_disable_steal_time();
 384 }
 385
 386 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 387                                 unsigned long code, void *unused)
 388 {
 389         if (code == SYS_RESTART)
 390                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 391         return NOTIFY_DONE;
 392 }
 393
 394 static struct notifier_block kvm_pv_reboot_nb = {
 395         .notifier_call = kvm_pv_reboot_notify,
 396 };
 397
 398 static u64 kvm_steal_clock(int cpu)
 399 {
 400         u64 steal;
 401         struct kvm_steal_time *src;
 402         int version;
 403
 404         src = &per_cpu(steal_time, cpu);
 405         do {
 406                 version = src->version;
 407                 virt_rmb();
 408                 steal = src->steal;
 409                 virt_rmb();
 410         } while ((version & 1) || (version != src->version));
 411
 412         return steal;
 413 }
 414
 415 void kvm_disable_steal_time(void)
 416 {
 417         if (!has_steal_clock)
 418                 return;
 419
 420         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 421 }
 422
 423 #ifdef CONFIG_SMP
 424 static void __init kvm_smp_prepare_boot_cpu(void)
 425 {
 426         kvm_guest_cpu_init();
 427         native_smp_prepare_boot_cpu();
 428         kvm_spinlock_init();
 429 }
 430
 431 static void kvm_guest_cpu_offline(void)
 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_online(unsigned int cpu)
 441 {
 442         local_irq_disable();
 443         kvm_guest_cpu_init();
 444         local_irq_enable();
 445         return 0;
 446 }
 447
 448 static int kvm_cpu_down_prepare(unsigned int cpu)
 449 {
 450         local_irq_disable();
 451         kvm_guest_cpu_offline();
 452         local_irq_enable();
 453         return 0;
 454 }
 455 #endif
 456
 457 static void __init kvm_apf_trap_init(void)
 458 {
 459         update_intr_gate(X86_TRAP_PF, async_page_fault);
 460 }
 461
 462 void __init kvm_guest_init(void)
 463 {
 464         int i;
 465
 466         if (!kvm_para_available())
 467                 return;
 468
 469         paravirt_ops_setup();
 470         register_reboot_notifier(&kvm_pv_reboot_nb);
 471         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
 472                 raw_spin_lock_init(&async_pf_sleepers[i].lock);
 473         if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
 474                 x86_init.irqs.trap_init = kvm_apf_trap_init;
 475
 476         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 477                 has_steal_clock = 1;
 478                 pv_time_ops.steal_clock = kvm_steal_clock;
 479         }
 480
 481         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 482                 apic_set_eoi_write(kvm_guest_apic_eoi_write);
 483
 484         if (kvmclock_vsyscall)
 485                 kvm_setup_vsyscall_timeinfo();
 486
 487 #ifdef CONFIG_SMP
 488         smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
 489         if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
 490                                       kvm_cpu_online, kvm_cpu_down_prepare) < 0)
 491                 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
 492 #else
 493         kvm_guest_cpu_init();
 494 #endif
 495
 496         /*
 497          * Hard lockup detection is enabled by default. Disable it, as guests
 498          * can get false positives too easily, for example if the host is
 499          * overcommitted.
 500          */
 501         hardlockup_detector_disable();
 502 }
 503
 504 static noinline uint32_t __kvm_cpuid_base(void)
 505 {
 506         if (boot_cpu_data.cpuid_level < 0)
 507                 return 0;       /* So we don't blow up on old processors */
 508
 509         if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
 510                 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
 511
 512         return 0;
 513 }
 514
 515 static inline uint32_t kvm_cpuid_base(void)
 516 {
 517         static int kvm_cpuid_base = -1;
 518
 519         if (kvm_cpuid_base == -1)
 520                 kvm_cpuid_base = __kvm_cpuid_base();
 521
 522         return kvm_cpuid_base;
 523 }
 524
 525 bool kvm_para_available(void)
 526 {
 527         return kvm_cpuid_base() != 0;
 528 }
 529 EXPORT_SYMBOL_GPL(kvm_para_available);
 530
 531 unsigned int kvm_arch_para_features(void)
 532 {
 533         return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
 534 }
 535
 536 static uint32_t __init kvm_detect(void)
 537 {
 538         return kvm_cpuid_base();
 539 }
 540
 541 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
 542         .name                   = "KVM",
 543         .detect                 = kvm_detect,
 544         .x2apic_available       = kvm_para_available,
 545 };
 546 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
 547
 548 static __init int activate_jump_labels(void)
 549 {
 550         if (has_steal_clock) {
 551                 static_key_slow_inc(&paravirt_steal_enabled);
 552                 if (steal_acc)
 553                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 554         }
 555
 556         return 0;
 557 }
 558 arch_initcall(activate_jump_labels);
 559
 560 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 561
 562 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 563 static void kvm_kick_cpu(int cpu)
 564 {
 565         int apicid;
 566         unsigned long flags = 0;
 567
 568         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 569         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 570 }
 571
 572 #include <asm/qspinlock.h>
 573
 574 static void kvm_wait(u8 *ptr, u8 val)
 575 {
 576         unsigned long flags;
 577
 578         if (in_nmi())
 579                 return;
 580
 581         local_irq_save(flags);
 582
 583         if (READ_ONCE(*ptr) != val)
 584                 goto out;
 585
 586         /*
 587          * halt until it's our turn and kicked. Note that we do safe halt
 588          * for irq enabled case to avoid hang when lock info is overwritten
 589          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 590          */
 591         if (arch_irqs_disabled_flags(flags))
 592                 halt();
 593         else
 594                 safe_halt();
 595
 596 out:
 597         local_irq_restore(flags);
 598 }
 599
 600 #ifdef CONFIG_X86_32
 601 __visible bool __kvm_vcpu_is_preempted(long cpu)
 602 {
 603         struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
 604
 605         return !!src->preempted;
 606 }
 607 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
 608
 609 #else
 610
 611 #include <asm/asm-offsets.h>
 612
 613 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
 614
 615 /*
 616  * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
 617  * restoring to/from the stack.
 618  */
 619 asm(
 620 ".pushsection .text;"
 621 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
 622 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
 623 "__raw_callee_save___kvm_vcpu_is_preempted:"
 624 "movq   __per_cpu_offset(,%rdi,8), %rax;"
 625 "cmpb   $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
 626 "setne  %al;"
 627 "ret;"
 628 ".popsection");
 629
 630 #endif
 631
 632 /*
 633  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 634  */
 635 void __init kvm_spinlock_init(void)
 636 {
 637         if (!kvm_para_available())
 638                 return;
 639         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 640         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 641                 return;
 642
 643         __pv_init_lock_hash();
 644         pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
 645         pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
 646         pv_lock_ops.wait = kvm_wait;
 647         pv_lock_ops.kick = kvm_kick_cpu;
 648
 649         if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
 650                 pv_lock_ops.vcpu_is_preempted =
 651                         PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
 652         }
 653 }
 654
 655 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */