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/module.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 <asm/timer.h>
  39 #include <asm/cpu.h>
  40 #include <asm/traps.h>
  41 #include <asm/desc.h>
  42 #include <asm/tlbflush.h>
  43 #include <asm/idle.h>
  44 #include <asm/apic.h>
  45 #include <asm/apicdef.h>
  46 #include <asm/hypervisor.h>
  47 #include <asm/kvm_guest.h>
  48
  49 static int kvmapf = 1;
  50
  51 static int parse_no_kvmapf(char *arg)
  52 {
  53         kvmapf = 0;
  54         return 0;
  55 }
  56
  57 early_param("no-kvmapf", parse_no_kvmapf);
  58
  59 static int steal_acc = 1;
  60 static int parse_no_stealacc(char *arg)
  61 {
  62         steal_acc = 0;
  63         return 0;
  64 }
  65
  66 early_param("no-steal-acc", parse_no_stealacc);
  67
  68 static int kvmclock_vsyscall = 1;
  69 static int parse_no_kvmclock_vsyscall(char *arg)
  70 {
  71         kvmclock_vsyscall = 0;
  72         return 0;
  73 }
  74
  75 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
  76
  77 static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
  78 static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
  79 static int has_steal_clock = 0;
  80
  81 /*
  82  * No need for any "IO delay" on KVM
  83  */
  84 static void kvm_io_delay(void)
  85 {
  86 }
  87
  88 #define KVM_TASK_SLEEP_HASHBITS 8
  89 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
  90
  91 struct kvm_task_sleep_node {
  92         struct hlist_node link;
  93         wait_queue_head_t wq;
  94         u32 token;
  95         int cpu;
  96         bool halted;
  97 };
  98
  99 static struct kvm_task_sleep_head {
 100         spinlock_t lock;
 101         struct hlist_head list;
 102 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
 103
 104 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
 105                                                   u32 token)
 106 {
 107         struct hlist_node *p;
 108
 109         hlist_for_each(p, &b->list) {
 110                 struct kvm_task_sleep_node *n =
 111                         hlist_entry(p, typeof(*n), link);
 112                 if (n->token == token)
 113                         return n;
 114         }
 115
 116         return NULL;
 117 }
 118
 119 void kvm_async_pf_task_wait(u32 token)
 120 {
 121         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 122         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 123         struct kvm_task_sleep_node n, *e;
 124         DEFINE_WAIT(wait);
 125
 126         rcu_irq_enter();
 127
 128         spin_lock(&b->lock);
 129         e = _find_apf_task(b, token);
 130         if (e) {
 131                 /* dummy entry exist -> wake up was delivered ahead of PF */
 132                 hlist_del(&e->link);
 133                 kfree(e);
 134                 spin_unlock(&b->lock);
 135
 136                 rcu_irq_exit();
 137                 return;
 138         }
 139
 140         n.token = token;
 141         n.cpu = smp_processor_id();
 142         n.halted = is_idle_task(current) || preempt_count() > 1;
 143         init_waitqueue_head(&n.wq);
 144         hlist_add_head(&n.link, &b->list);
 145         spin_unlock(&b->lock);
 146
 147         for (;;) {
 148                 if (!n.halted)
 149                         prepare_to_wait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
 150                 if (hlist_unhashed(&n.link))
 151                         break;
 152
 153                 if (!n.halted) {
 154                         local_irq_enable();
 155                         schedule();
 156                         local_irq_disable();
 157                 } else {
 158                         /*
 159                          * We cannot reschedule. So halt.
 160                          */
 161                         rcu_irq_exit();
 162                         native_safe_halt();
 163                         rcu_irq_enter();
 164                         local_irq_disable();
 165                 }
 166         }
 167         if (!n.halted)
 168                 finish_wait(&n.wq, &wait);
 169
 170         rcu_irq_exit();
 171         return;
 172 }
 173 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
 174
 175 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
 176 {
 177         hlist_del_init(&n->link);
 178         if (n->halted)
 179                 smp_send_reschedule(n->cpu);
 180         else if (waitqueue_active(&n->wq))
 181                 wake_up(&n->wq);
 182 }
 183
 184 static void apf_task_wake_all(void)
 185 {
 186         int i;
 187
 188         for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
 189                 struct hlist_node *p, *next;
 190                 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
 191                 spin_lock(&b->lock);
 192                 hlist_for_each_safe(p, next, &b->list) {
 193                         struct kvm_task_sleep_node *n =
 194                                 hlist_entry(p, typeof(*n), link);
 195                         if (n->cpu == smp_processor_id())
 196                                 apf_task_wake_one(n);
 197                 }
 198                 spin_unlock(&b->lock);
 199         }
 200 }
 201
 202 void kvm_async_pf_task_wake(u32 token)
 203 {
 204         u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
 205         struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
 206         struct kvm_task_sleep_node *n;
 207
 208         if (token == ~0) {
 209                 apf_task_wake_all();
 210                 return;
 211         }
 212
 213 again:
 214         spin_lock(&b->lock);
 215         n = _find_apf_task(b, token);
 216         if (!n) {
 217                 /*
 218                  * async PF was not yet handled.
 219                  * Add dummy entry for the token.
 220                  */
 221                 n = kzalloc(sizeof(*n), GFP_ATOMIC);
 222                 if (!n) {
 223                         /*
 224                          * Allocation failed! Busy wait while other cpu
 225                          * handles async PF.
 226                          */
 227                         spin_unlock(&b->lock);
 228                         cpu_relax();
 229                         goto again;
 230                 }
 231                 n->token = token;
 232                 n->cpu = smp_processor_id();
 233                 init_waitqueue_head(&n->wq);
 234                 hlist_add_head(&n->link, &b->list);
 235         } else
 236                 apf_task_wake_one(n);
 237         spin_unlock(&b->lock);
 238         return;
 239 }
 240 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
 241
 242 u32 kvm_read_and_reset_pf_reason(void)
 243 {
 244         u32 reason = 0;
 245
 246         if (__get_cpu_var(apf_reason).enabled) {
 247                 reason = __get_cpu_var(apf_reason).reason;
 248                 __get_cpu_var(apf_reason).reason = 0;
 249         }
 250
 251         return reason;
 252 }
 253 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
 254
 255 dotraplinkage void __kprobes
 256 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
 257 {
 258         enum ctx_state prev_state;
 259
 260         switch (kvm_read_and_reset_pf_reason()) {
 261         default:
 262                 do_page_fault(regs, error_code);
 263                 break;
 264         case KVM_PV_REASON_PAGE_NOT_PRESENT:
 265                 /* page is swapped out by the host. */
 266                 prev_state = exception_enter();
 267                 exit_idle();
 268                 kvm_async_pf_task_wait((u32)read_cr2());
 269                 exception_exit(prev_state);
 270                 break;
 271         case KVM_PV_REASON_PAGE_READY:
 272                 rcu_irq_enter();
 273                 exit_idle();
 274                 kvm_async_pf_task_wake((u32)read_cr2());
 275                 rcu_irq_exit();
 276                 break;
 277         }
 278 }
 279
 280 static void __init paravirt_ops_setup(void)
 281 {
 282         pv_info.name = "KVM";
 283         pv_info.paravirt_enabled = 1;
 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         memset(st, 0, sizeof(*st));
 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 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, &__get_cpu_var(kvm_apic_eoi)))
 320                 return;
 321         apic_write(APIC_EOI, APIC_EOI_ACK);
 322 }
 323
 324 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(&__get_cpu_var(apf_reason));
 331
 332 #ifdef CONFIG_PREEMPT
 333                 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
 334 #endif
 335                 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED);
 336                 __get_cpu_var(apf_reason).enabled = 1;
 337                 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
 338                        smp_processor_id());
 339         }
 340
 341         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
 342                 unsigned long pa;
 343                 /* Size alignment is implied but just to make it explicit. */
 344                 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
 345                 __get_cpu_var(kvm_apic_eoi) = 0;
 346                 pa = slow_virt_to_phys(&__get_cpu_var(kvm_apic_eoi))
 347                         | KVM_MSR_ENABLED;
 348                 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
 349         }
 350
 351         if (has_steal_clock)
 352                 kvm_register_steal_time();
 353 }
 354
 355 static void kvm_pv_disable_apf(void)
 356 {
 357         if (!__get_cpu_var(apf_reason).enabled)
 358                 return;
 359
 360         wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
 361         __get_cpu_var(apf_reason).enabled = 0;
 362
 363         printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
 364                smp_processor_id());
 365 }
 366
 367 static void kvm_pv_guest_cpu_reboot(void *unused)
 368 {
 369         /*
 370          * We disable PV EOI before we load a new kernel by kexec,
 371          * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
 372          * New kernel can re-enable when it boots.
 373          */
 374         if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
 375                 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
 376         kvm_pv_disable_apf();
 377         kvm_disable_steal_time();
 378 }
 379
 380 static int kvm_pv_reboot_notify(struct notifier_block *nb,
 381                                 unsigned long code, void *unused)
 382 {
 383         if (code == SYS_RESTART)
 384                 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
 385         return NOTIFY_DONE;
 386 }
 387
 388 static struct notifier_block kvm_pv_reboot_nb = {
 389         .notifier_call = kvm_pv_reboot_notify,
 390 };
 391
 392 static u64 kvm_steal_clock(int cpu)
 393 {
 394         u64 steal;
 395         struct kvm_steal_time *src;
 396         int version;
 397
 398         src = &per_cpu(steal_time, cpu);
 399         do {
 400                 version = src->version;
 401                 rmb();
 402                 steal = src->steal;
 403                 rmb();
 404         } while ((version & 1) || (version != src->version));
 405
 406         return steal;
 407 }
 408
 409 void kvm_disable_steal_time(void)
 410 {
 411         if (!has_steal_clock)
 412                 return;
 413
 414         wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
 415 }
 416
 417 #ifdef CONFIG_SMP
 418 static void __init kvm_smp_prepare_boot_cpu(void)
 419 {
 420         WARN_ON(kvm_register_clock("primary cpu clock"));
 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                 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 static uint32_t __init kvm_detect(void)
 504 {
 505         return kvm_cpuid_base();
 506 }
 507
 508 const struct hypervisor_x86 x86_hyper_kvm __refconst = {
 509         .name                   = "KVM",
 510         .detect                 = kvm_detect,
 511         .x2apic_available       = kvm_para_available,
 512 };
 513 EXPORT_SYMBOL_GPL(x86_hyper_kvm);
 514
 515 static __init int activate_jump_labels(void)
 516 {
 517         if (has_steal_clock) {
 518                 static_key_slow_inc(&paravirt_steal_enabled);
 519                 if (steal_acc)
 520                         static_key_slow_inc(&paravirt_steal_rq_enabled);
 521         }
 522
 523         return 0;
 524 }
 525 arch_initcall(activate_jump_labels);
 526
 527 #ifdef CONFIG_PARAVIRT_SPINLOCKS
 528
 529 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
 530 static void kvm_kick_cpu(int cpu)
 531 {
 532         int apicid;
 533         unsigned long flags = 0;
 534
 535         apicid = per_cpu(x86_cpu_to_apicid, cpu);
 536         kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
 537 }
 538
 539 enum kvm_contention_stat {
 540         TAKEN_SLOW,
 541         TAKEN_SLOW_PICKUP,
 542         RELEASED_SLOW,
 543         RELEASED_SLOW_KICKED,
 544         NR_CONTENTION_STATS
 545 };
 546
 547 #ifdef CONFIG_KVM_DEBUG_FS
 548 #define HISTO_BUCKETS   30
 549
 550 static struct kvm_spinlock_stats
 551 {
 552         u32 contention_stats[NR_CONTENTION_STATS];
 553         u32 histo_spin_blocked[HISTO_BUCKETS+1];
 554         u64 time_blocked;
 555 } spinlock_stats;
 556
 557 static u8 zero_stats;
 558
 559 static inline void check_zero(void)
 560 {
 561         u8 ret;
 562         u8 old;
 563
 564         old = ACCESS_ONCE(zero_stats);
 565         if (unlikely(old)) {
 566                 ret = cmpxchg(&zero_stats, old, 0);
 567                 /* This ensures only one fellow resets the stat */
 568                 if (ret == old)
 569                         memset(&spinlock_stats, 0, sizeof(spinlock_stats));
 570         }
 571 }
 572
 573 static inline void add_stats(enum kvm_contention_stat var, u32 val)
 574 {
 575         check_zero();
 576         spinlock_stats.contention_stats[var] += val;
 577 }
 578
 579
 580 static inline u64 spin_time_start(void)
 581 {
 582         return sched_clock();
 583 }
 584
 585 static void __spin_time_accum(u64 delta, u32 *array)
 586 {
 587         unsigned index;
 588
 589         index = ilog2(delta);
 590         check_zero();
 591
 592         if (index < HISTO_BUCKETS)
 593                 array[index]++;
 594         else
 595                 array[HISTO_BUCKETS]++;
 596 }
 597
 598 static inline void spin_time_accum_blocked(u64 start)
 599 {
 600         u32 delta;
 601
 602         delta = sched_clock() - start;
 603         __spin_time_accum(delta, spinlock_stats.histo_spin_blocked);
 604         spinlock_stats.time_blocked += delta;
 605 }
 606
 607 static struct dentry *d_spin_debug;
 608 static struct dentry *d_kvm_debug;
 609
 610 struct dentry *kvm_init_debugfs(void)
 611 {
 612         d_kvm_debug = debugfs_create_dir("kvm-guest", NULL);
 613         if (!d_kvm_debug)
 614                 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n");
 615
 616         return d_kvm_debug;
 617 }
 618
 619 static int __init kvm_spinlock_debugfs(void)
 620 {
 621         struct dentry *d_kvm;
 622
 623         d_kvm = kvm_init_debugfs();
 624         if (d_kvm == NULL)
 625                 return -ENOMEM;
 626
 627         d_spin_debug = debugfs_create_dir("spinlocks", d_kvm);
 628
 629         debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats);
 630
 631         debugfs_create_u32("taken_slow", 0444, d_spin_debug,
 632                    &spinlock_stats.contention_stats[TAKEN_SLOW]);
 633         debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug,
 634                    &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]);
 635
 636         debugfs_create_u32("released_slow", 0444, d_spin_debug,
 637                    &spinlock_stats.contention_stats[RELEASED_SLOW]);
 638         debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug,
 639                    &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]);
 640
 641         debugfs_create_u64("time_blocked", 0444, d_spin_debug,
 642                            &spinlock_stats.time_blocked);
 643
 644         debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug,
 645                      spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1);
 646
 647         return 0;
 648 }
 649 fs_initcall(kvm_spinlock_debugfs);
 650 #else  /* !CONFIG_KVM_DEBUG_FS */
 651 static inline void add_stats(enum kvm_contention_stat var, u32 val)
 652 {
 653 }
 654
 655 static inline u64 spin_time_start(void)
 656 {
 657         return 0;
 658 }
 659
 660 static inline void spin_time_accum_blocked(u64 start)
 661 {
 662 }
 663 #endif  /* CONFIG_KVM_DEBUG_FS */
 664
 665 struct kvm_lock_waiting {
 666         struct arch_spinlock *lock;
 667         __ticket_t want;
 668 };
 669
 670 /* cpus 'waiting' on a spinlock to become available */
 671 static cpumask_t waiting_cpus;
 672
 673 /* Track spinlock on which a cpu is waiting */
 674 static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting);
 675
 676 __visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want)
 677 {
 678         struct kvm_lock_waiting *w;
 679         int cpu;
 680         u64 start;
 681         unsigned long flags;
 682
 683         if (in_nmi())
 684                 return;
 685
 686         w = &__get_cpu_var(klock_waiting);
 687         cpu = smp_processor_id();
 688         start = spin_time_start();
 689
 690         /*
 691          * Make sure an interrupt handler can't upset things in a
 692          * partially setup state.
 693          */
 694         local_irq_save(flags);
 695
 696         /*
 697          * The ordering protocol on this is that the "lock" pointer
 698          * may only be set non-NULL if the "want" ticket is correct.
 699          * If we're updating "want", we must first clear "lock".
 700          */
 701         w->lock = NULL;
 702         smp_wmb();
 703         w->want = want;
 704         smp_wmb();
 705         w->lock = lock;
 706
 707         add_stats(TAKEN_SLOW, 1);
 708
 709         /*
 710          * This uses set_bit, which is atomic but we should not rely on its
 711          * reordering gurantees. So barrier is needed after this call.
 712          */
 713         cpumask_set_cpu(cpu, &waiting_cpus);
 714
 715         barrier();
 716
 717         /*
 718          * Mark entry to slowpath before doing the pickup test to make
 719          * sure we don't deadlock with an unlocker.
 720          */
 721         __ticket_enter_slowpath(lock);
 722
 723         /*
 724          * check again make sure it didn't become free while
 725          * we weren't looking.
 726          */
 727         if (ACCESS_ONCE(lock->tickets.head) == want) {
 728                 add_stats(TAKEN_SLOW_PICKUP, 1);
 729                 goto out;
 730         }
 731
 732         /*
 733          * halt until it's our turn and kicked. Note that we do safe halt
 734          * for irq enabled case to avoid hang when lock info is overwritten
 735          * in irq spinlock slowpath and no spurious interrupt occur to save us.
 736          */
 737         if (arch_irqs_disabled_flags(flags))
 738                 halt();
 739         else
 740                 safe_halt();
 741
 742 out:
 743         cpumask_clear_cpu(cpu, &waiting_cpus);
 744         w->lock = NULL;
 745         local_irq_restore(flags);
 746         spin_time_accum_blocked(start);
 747 }
 748 PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning);
 749
 750 /* Kick vcpu waiting on @lock->head to reach value @ticket */
 751 static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket)
 752 {
 753         int cpu;
 754
 755         add_stats(RELEASED_SLOW, 1);
 756         for_each_cpu(cpu, &waiting_cpus) {
 757                 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu);
 758                 if (ACCESS_ONCE(w->lock) == lock &&
 759                     ACCESS_ONCE(w->want) == ticket) {
 760                         add_stats(RELEASED_SLOW_KICKED, 1);
 761                         kvm_kick_cpu(cpu);
 762                         break;
 763                 }
 764         }
 765 }
 766
 767 /*
 768  * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
 769  */
 770 void __init kvm_spinlock_init(void)
 771 {
 772         if (!kvm_para_available())
 773                 return;
 774         /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
 775         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 776                 return;
 777
 778         pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning);
 779         pv_lock_ops.unlock_kick = kvm_unlock_kick;
 780 }
 781
 782 static __init int kvm_spinlock_init_jump(void)
 783 {
 784         if (!kvm_para_available())
 785                 return 0;
 786         if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
 787                 return 0;
 788
 789         static_key_slow_inc(&paravirt_ticketlocks_enabled);
 790         printk(KERN_INFO "KVM setup paravirtual spinlock\n");
 791
 792         return 0;
 793 }
 794 early_initcall(kvm_spinlock_init_jump);
 795
 796 #endif  /* CONFIG_PARAVIRT_SPINLOCKS */