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 <linux/nmi.h> 39#include <asm/timer.h> 40#include <asm/cpu.h> 41#include <asm/traps.h> 42#include <asm/desc.h> 43#include <asm/tlbflush.h> 44#include <asm/idle.h> 45#include <asm/apic.h> 46#include <asm/apicdef.h> 47#include <asm/hypervisor.h> 48#include <asm/kvm_guest.h> 49 50static int kvmapf = 1; 51 52static int parse_no_kvmapf(char *arg) 53{ 54 kvmapf = 0; 55 return 0; 56} 57 58early_param("no-kvmapf", parse_no_kvmapf); 59 60static int steal_acc = 1; 61static int parse_no_stealacc(char *arg) 62{ 63 steal_acc = 0; 64 return 0; 65} 66 67early_param("no-steal-acc", parse_no_stealacc); 68 69static int kvmclock_vsyscall = 1; 70static int parse_no_kvmclock_vsyscall(char *arg) 71{ 72 kvmclock_vsyscall = 0; 73 return 0; 74} 75 76early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall); 77 78static DEFINE_PER_CPU(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64); 79static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64); 80static int has_steal_clock = 0; 81 82/* 83 * No need for any "IO delay" on KVM 84 */ 85static 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 92struct kvm_task_sleep_node { 93 struct hlist_node link; 94 wait_queue_head_t wq; 95 u32 token; 96 int cpu; 97 bool halted; 98}; 99 100static struct kvm_task_sleep_head { 101 spinlock_t lock; 102 struct hlist_head list; 103} async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE]; 104 105static 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 120void 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 DEFINE_WAIT(wait); 126 127 rcu_irq_enter(); 128 129 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 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_waitqueue_head(&n.wq); 145 hlist_add_head(&n.link, &b->list); 146 spin_unlock(&b->lock); 147 148 for (;;) { 149 if (!n.halted) 150 prepare_to_wait(&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 rcu_irq_enter(); 165 local_irq_disable(); 166 } 167 } 168 if (!n.halted) 169 finish_wait(&n.wq, &wait); 170 171 rcu_irq_exit(); 172 return; 173} 174EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait); 175 176static 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 (waitqueue_active(&n->wq)) 182 wake_up(&n->wq); 183} 184 185static 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 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 spin_unlock(&b->lock); 200 } 201} 202 203void 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 214again: 215 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 spin_unlock(&b->lock); 229 cpu_relax(); 230 goto again; 231 } 232 n->token = token; 233 n->cpu = smp_processor_id(); 234 init_waitqueue_head(&n->wq); 235 hlist_add_head(&n->link, &b->list); 236 } else 237 apf_task_wake_one(n); 238 spin_unlock(&b->lock); 239 return; 240} 241EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake); 242 243u32 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} 254EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason); 255NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason); 256 257dotraplinkage void 258do_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 exit_idle(); 270 kvm_async_pf_task_wait((u32)read_cr2()); 271 exception_exit(prev_state); 272 break; 273 case KVM_PV_REASON_PAGE_READY: 274 rcu_irq_enter(); 275 exit_idle(); 276 kvm_async_pf_task_wake((u32)read_cr2()); 277 rcu_irq_exit(); 278 break; 279 } 280} 281NOKPROBE_SYMBOL(do_async_page_fault); 282 283static void __init paravirt_ops_setup(void) 284{ 285 pv_info.name = "KVM"; 286 pv_info.paravirt_enabled = 1; 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 296static 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 memset(st, 0, sizeof(*st)); 305 306 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED)); 307 pr_info("kvm-stealtime: cpu %d, msr %llx\n", 308 cpu, (unsigned long long) slow_virt_to_phys(st)); 309} 310 311static DEFINE_PER_CPU(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED; 312 313static void kvm_guest_apic_eoi_write(u32 reg, u32 val) 314{ 315 /** 316 * This relies on __test_and_clear_bit to modify the memory 317 * in a way that is atomic with respect to the local CPU. 318 * The hypervisor only accesses this memory from the local CPU so 319 * there's no need for lock or memory barriers. 320 * An optimization barrier is implied in apic write. 321 */ 322 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi))) 323 return; 324 apic_write(APIC_EOI, APIC_EOI_ACK); 325} 326 327void kvm_guest_cpu_init(void) 328{ 329 if (!kvm_para_available()) 330 return; 331 332 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) { 333 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason)); 334 335#ifdef CONFIG_PREEMPT 336 pa |= KVM_ASYNC_PF_SEND_ALWAYS; 337#endif 338 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa | KVM_ASYNC_PF_ENABLED); 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 358static 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 370static 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 383static 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 391static struct notifier_block kvm_pv_reboot_nb = { 392 .notifier_call = kvm_pv_reboot_notify, 393}; 394 395static 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 rmb(); 405 steal = src->steal; 406 rmb(); 407 } while ((version & 1) || (version != src->version)); 408 409 return steal; 410} 411 412void 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 421static 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 428static void kvm_guest_cpu_online(void *dummy) 429{ 430 kvm_guest_cpu_init(); 431} 432 433static void kvm_guest_cpu_offline(void *dummy) 434{ 435 kvm_disable_steal_time(); 436 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) 437 wrmsrl(MSR_KVM_PV_EOI_EN, 0); 438 kvm_pv_disable_apf(); 439 apf_task_wake_all(); 440} 441 442static int kvm_cpu_notify(struct notifier_block *self, unsigned long action, 443 void *hcpu) 444{ 445 int cpu = (unsigned long)hcpu; 446 switch (action) { 447 case CPU_ONLINE: 448 case CPU_DOWN_FAILED: 449 case CPU_ONLINE_FROZEN: 450 smp_call_function_single(cpu, kvm_guest_cpu_online, NULL, 0); 451 break; 452 case CPU_DOWN_PREPARE: 453 case CPU_DOWN_PREPARE_FROZEN: 454 smp_call_function_single(cpu, kvm_guest_cpu_offline, NULL, 1); 455 break; 456 default: 457 break; 458 } 459 return NOTIFY_OK; 460} 461 462static struct notifier_block kvm_cpu_notifier = { 463 .notifier_call = kvm_cpu_notify, 464}; 465#endif 466 467static void __init kvm_apf_trap_init(void) 468{ 469 set_intr_gate(14, async_page_fault); 470} 471 472void __init kvm_guest_init(void) 473{ 474 int i; 475 476 if (!kvm_para_available()) 477 return; 478 479 paravirt_ops_setup(); 480 register_reboot_notifier(&kvm_pv_reboot_nb); 481 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) 482 spin_lock_init(&async_pf_sleepers[i].lock); 483 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF)) 484 x86_init.irqs.trap_init = kvm_apf_trap_init; 485 486 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) { 487 has_steal_clock = 1; 488 pv_time_ops.steal_clock = kvm_steal_clock; 489 } 490 491 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) 492 apic_set_eoi_write(kvm_guest_apic_eoi_write); 493 494 if (kvmclock_vsyscall) 495 kvm_setup_vsyscall_timeinfo(); 496 497#ifdef CONFIG_SMP 498 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu; 499 register_cpu_notifier(&kvm_cpu_notifier); 500#else 501 kvm_guest_cpu_init(); 502#endif 503 504 /* 505 * Hard lockup detection is enabled by default. Disable it, as guests 506 * can get false positives too easily, for example if the host is 507 * overcommitted. 508 */ 509 watchdog_enable_hardlockup_detector(false); 510} 511 512static noinline uint32_t __kvm_cpuid_base(void) 513{ 514 if (boot_cpu_data.cpuid_level < 0) 515 return 0; /* So we don't blow up on old processors */ 516 517 if (cpu_has_hypervisor) 518 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0); 519 520 return 0; 521} 522 523static inline uint32_t kvm_cpuid_base(void) 524{ 525 static int kvm_cpuid_base = -1; 526 527 if (kvm_cpuid_base == -1) 528 kvm_cpuid_base = __kvm_cpuid_base(); 529 530 return kvm_cpuid_base; 531} 532 533bool kvm_para_available(void) 534{ 535 return kvm_cpuid_base() != 0; 536} 537EXPORT_SYMBOL_GPL(kvm_para_available); 538 539unsigned int kvm_arch_para_features(void) 540{ 541 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES); 542} 543 544static uint32_t __init kvm_detect(void) 545{ 546 return kvm_cpuid_base(); 547} 548 549const struct hypervisor_x86 x86_hyper_kvm __refconst = { 550 .name = "KVM", 551 .detect = kvm_detect, 552 .x2apic_available = kvm_para_available, 553}; 554EXPORT_SYMBOL_GPL(x86_hyper_kvm); 555 556static __init int activate_jump_labels(void) 557{ 558 if (has_steal_clock) { 559 static_key_slow_inc(¶virt_steal_enabled); 560 if (steal_acc) 561 static_key_slow_inc(¶virt_steal_rq_enabled); 562 } 563 564 return 0; 565} 566arch_initcall(activate_jump_labels); 567 568#ifdef CONFIG_PARAVIRT_SPINLOCKS 569 570/* Kick a cpu by its apicid. Used to wake up a halted vcpu */ 571static void kvm_kick_cpu(int cpu) 572{ 573 int apicid; 574 unsigned long flags = 0; 575 576 apicid = per_cpu(x86_cpu_to_apicid, cpu); 577 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid); 578} 579 580enum kvm_contention_stat { 581 TAKEN_SLOW, 582 TAKEN_SLOW_PICKUP, 583 RELEASED_SLOW, 584 RELEASED_SLOW_KICKED, 585 NR_CONTENTION_STATS 586}; 587 588#ifdef CONFIG_KVM_DEBUG_FS 589#define HISTO_BUCKETS 30 590 591static struct kvm_spinlock_stats 592{ 593 u32 contention_stats[NR_CONTENTION_STATS]; 594 u32 histo_spin_blocked[HISTO_BUCKETS+1]; 595 u64 time_blocked; 596} spinlock_stats; 597 598static u8 zero_stats; 599 600static inline void check_zero(void) 601{ 602 u8 ret; 603 u8 old; 604 605 old = ACCESS_ONCE(zero_stats); 606 if (unlikely(old)) { 607 ret = cmpxchg(&zero_stats, old, 0); 608 /* This ensures only one fellow resets the stat */ 609 if (ret == old) 610 memset(&spinlock_stats, 0, sizeof(spinlock_stats)); 611 } 612} 613 614static inline void add_stats(enum kvm_contention_stat var, u32 val) 615{ 616 check_zero(); 617 spinlock_stats.contention_stats[var] += val; 618} 619 620 621static inline u64 spin_time_start(void) 622{ 623 return sched_clock(); 624} 625 626static void __spin_time_accum(u64 delta, u32 *array) 627{ 628 unsigned index; 629 630 index = ilog2(delta); 631 check_zero(); 632 633 if (index < HISTO_BUCKETS) 634 array[index]++; 635 else 636 array[HISTO_BUCKETS]++; 637} 638 639static inline void spin_time_accum_blocked(u64 start) 640{ 641 u32 delta; 642 643 delta = sched_clock() - start; 644 __spin_time_accum(delta, spinlock_stats.histo_spin_blocked); 645 spinlock_stats.time_blocked += delta; 646} 647 648static struct dentry *d_spin_debug; 649static struct dentry *d_kvm_debug; 650 651struct dentry *kvm_init_debugfs(void) 652{ 653 d_kvm_debug = debugfs_create_dir("kvm-guest", NULL); 654 if (!d_kvm_debug) 655 printk(KERN_WARNING "Could not create 'kvm' debugfs directory\n"); 656 657 return d_kvm_debug; 658} 659 660static int __init kvm_spinlock_debugfs(void) 661{ 662 struct dentry *d_kvm; 663 664 d_kvm = kvm_init_debugfs(); 665 if (d_kvm == NULL) 666 return -ENOMEM; 667 668 d_spin_debug = debugfs_create_dir("spinlocks", d_kvm); 669 670 debugfs_create_u8("zero_stats", 0644, d_spin_debug, &zero_stats); 671 672 debugfs_create_u32("taken_slow", 0444, d_spin_debug, 673 &spinlock_stats.contention_stats[TAKEN_SLOW]); 674 debugfs_create_u32("taken_slow_pickup", 0444, d_spin_debug, 675 &spinlock_stats.contention_stats[TAKEN_SLOW_PICKUP]); 676 677 debugfs_create_u32("released_slow", 0444, d_spin_debug, 678 &spinlock_stats.contention_stats[RELEASED_SLOW]); 679 debugfs_create_u32("released_slow_kicked", 0444, d_spin_debug, 680 &spinlock_stats.contention_stats[RELEASED_SLOW_KICKED]); 681 682 debugfs_create_u64("time_blocked", 0444, d_spin_debug, 683 &spinlock_stats.time_blocked); 684 685 debugfs_create_u32_array("histo_blocked", 0444, d_spin_debug, 686 spinlock_stats.histo_spin_blocked, HISTO_BUCKETS + 1); 687 688 return 0; 689} 690fs_initcall(kvm_spinlock_debugfs); 691#else /* !CONFIG_KVM_DEBUG_FS */ 692static inline void add_stats(enum kvm_contention_stat var, u32 val) 693{ 694} 695 696static inline u64 spin_time_start(void) 697{ 698 return 0; 699} 700 701static inline void spin_time_accum_blocked(u64 start) 702{ 703} 704#endif /* CONFIG_KVM_DEBUG_FS */ 705 706struct kvm_lock_waiting { 707 struct arch_spinlock *lock; 708 __ticket_t want; 709}; 710 711/* cpus 'waiting' on a spinlock to become available */ 712static cpumask_t waiting_cpus; 713 714/* Track spinlock on which a cpu is waiting */ 715static DEFINE_PER_CPU(struct kvm_lock_waiting, klock_waiting); 716 717__visible void kvm_lock_spinning(struct arch_spinlock *lock, __ticket_t want) 718{ 719 struct kvm_lock_waiting *w; 720 int cpu; 721 u64 start; 722 unsigned long flags; 723 724 if (in_nmi()) 725 return; 726 727 w = this_cpu_ptr(&klock_waiting); 728 cpu = smp_processor_id(); 729 start = spin_time_start(); 730 731 /* 732 * Make sure an interrupt handler can't upset things in a 733 * partially setup state. 734 */ 735 local_irq_save(flags); 736 737 /* 738 * The ordering protocol on this is that the "lock" pointer 739 * may only be set non-NULL if the "want" ticket is correct. 740 * If we're updating "want", we must first clear "lock". 741 */ 742 w->lock = NULL; 743 smp_wmb(); 744 w->want = want; 745 smp_wmb(); 746 w->lock = lock; 747 748 add_stats(TAKEN_SLOW, 1); 749 750 /* 751 * This uses set_bit, which is atomic but we should not rely on its 752 * reordering gurantees. So barrier is needed after this call. 753 */ 754 cpumask_set_cpu(cpu, &waiting_cpus); 755 756 barrier(); 757 758 /* 759 * Mark entry to slowpath before doing the pickup test to make 760 * sure we don't deadlock with an unlocker. 761 */ 762 __ticket_enter_slowpath(lock); 763 764 /* 765 * check again make sure it didn't become free while 766 * we weren't looking. 767 */ 768 if (ACCESS_ONCE(lock->tickets.head) == want) { 769 add_stats(TAKEN_SLOW_PICKUP, 1); 770 goto out; 771 } 772 773 /* 774 * halt until it's our turn and kicked. Note that we do safe halt 775 * for irq enabled case to avoid hang when lock info is overwritten 776 * in irq spinlock slowpath and no spurious interrupt occur to save us. 777 */ 778 if (arch_irqs_disabled_flags(flags)) 779 halt(); 780 else 781 safe_halt(); 782 783out: 784 cpumask_clear_cpu(cpu, &waiting_cpus); 785 w->lock = NULL; 786 local_irq_restore(flags); 787 spin_time_accum_blocked(start); 788} 789PV_CALLEE_SAVE_REGS_THUNK(kvm_lock_spinning); 790 791/* Kick vcpu waiting on @lock->head to reach value @ticket */ 792static void kvm_unlock_kick(struct arch_spinlock *lock, __ticket_t ticket) 793{ 794 int cpu; 795 796 add_stats(RELEASED_SLOW, 1); 797 for_each_cpu(cpu, &waiting_cpus) { 798 const struct kvm_lock_waiting *w = &per_cpu(klock_waiting, cpu); 799 if (ACCESS_ONCE(w->lock) == lock && 800 ACCESS_ONCE(w->want) == ticket) { 801 add_stats(RELEASED_SLOW_KICKED, 1); 802 kvm_kick_cpu(cpu); 803 break; 804 } 805 } 806} 807 808/* 809 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present. 810 */ 811void __init kvm_spinlock_init(void) 812{ 813 if (!kvm_para_available()) 814 return; 815 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */ 816 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) 817 return; 818 819 pv_lock_ops.lock_spinning = PV_CALLEE_SAVE(kvm_lock_spinning); 820 pv_lock_ops.unlock_kick = kvm_unlock_kick; 821} 822 823static __init int kvm_spinlock_init_jump(void) 824{ 825 if (!kvm_para_available()) 826 return 0; 827 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) 828 return 0; 829 830 static_key_slow_inc(¶virt_ticketlocks_enabled); 831 printk(KERN_INFO "KVM setup paravirtual spinlock\n"); 832 833 return 0; 834} 835early_initcall(kvm_spinlock_init_jump); 836 837#endif /* CONFIG_PARAVIRT_SPINLOCKS */ 838