1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * KVM paravirt_ops implementation
5 * Copyright (C) 2007, Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
6 * Copyright IBM Corporation, 2007
7 * Authors: Anthony Liguori <aliguori@us.ibm.com>
10 #define pr_fmt(fmt) "kvm-guest: " fmt
12 #include <linux/context_tracking.h>
13 #include <linux/init.h>
14 #include <linux/irq.h>
15 #include <linux/kernel.h>
16 #include <linux/kvm_para.h>
17 #include <linux/cpu.h>
19 #include <linux/highmem.h>
20 #include <linux/hardirq.h>
21 #include <linux/notifier.h>
22 #include <linux/reboot.h>
23 #include <linux/hash.h>
24 #include <linux/sched.h>
25 #include <linux/slab.h>
26 #include <linux/kprobes.h>
27 #include <linux/nmi.h>
28 #include <linux/swait.h>
29 #include <linux/syscore_ops.h>
30 #include <asm/timer.h>
32 #include <asm/traps.h>
34 #include <asm/tlbflush.h>
36 #include <asm/apicdef.h>
37 #include <asm/hypervisor.h>
39 #include <asm/cpuidle_haltpoll.h>
40 #include <asm/ptrace.h>
41 #include <asm/reboot.h>
44 DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
46 static int kvmapf = 1;
48 static int __init parse_no_kvmapf(char *arg)
54 early_param("no-kvmapf", parse_no_kvmapf);
56 static int steal_acc = 1;
57 static int __init parse_no_stealacc(char *arg)
63 early_param("no-steal-acc", parse_no_stealacc);
65 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
66 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
67 static int has_steal_clock = 0;
70 * No need for any "IO delay" on KVM
72 static void kvm_io_delay(void)
76 #define KVM_TASK_SLEEP_HASHBITS 8
77 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
79 struct kvm_task_sleep_node {
80 struct hlist_node link;
81 struct swait_queue_head wq;
86 static struct kvm_task_sleep_head {
88 struct hlist_head list;
89 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
91 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
96 hlist_for_each(p, &b->list) {
97 struct kvm_task_sleep_node *n =
98 hlist_entry(p, typeof(*n), link);
99 if (n->token == token)
106 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
108 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
109 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
110 struct kvm_task_sleep_node *e;
112 raw_spin_lock(&b->lock);
113 e = _find_apf_task(b, token);
115 /* dummy entry exist -> wake up was delivered ahead of PF */
117 raw_spin_unlock(&b->lock);
123 n->cpu = smp_processor_id();
124 init_swait_queue_head(&n->wq);
125 hlist_add_head(&n->link, &b->list);
126 raw_spin_unlock(&b->lock);
131 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
132 * @token: Token to identify the sleep node entry
134 * Invoked from the async pagefault handling code or from the VM exit page
135 * fault handler. In both cases RCU is watching.
137 void kvm_async_pf_task_wait_schedule(u32 token)
139 struct kvm_task_sleep_node n;
140 DECLARE_SWAITQUEUE(wait);
142 lockdep_assert_irqs_disabled();
144 if (!kvm_async_pf_queue_task(token, &n))
148 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
149 if (hlist_unhashed(&n.link))
156 finish_swait(&n.wq, &wait);
158 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
160 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
162 hlist_del_init(&n->link);
163 if (swq_has_sleeper(&n->wq))
164 swake_up_one(&n->wq);
167 static void apf_task_wake_all(void)
171 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
172 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
173 struct kvm_task_sleep_node *n;
174 struct hlist_node *p, *next;
176 raw_spin_lock(&b->lock);
177 hlist_for_each_safe(p, next, &b->list) {
178 n = hlist_entry(p, typeof(*n), link);
179 if (n->cpu == smp_processor_id())
180 apf_task_wake_one(n);
182 raw_spin_unlock(&b->lock);
186 void kvm_async_pf_task_wake(u32 token)
188 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
189 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
190 struct kvm_task_sleep_node *n;
198 raw_spin_lock(&b->lock);
199 n = _find_apf_task(b, token);
202 * async PF was not yet handled.
203 * Add dummy entry for the token.
205 n = kzalloc(sizeof(*n), GFP_ATOMIC);
208 * Allocation failed! Busy wait while other cpu
211 raw_spin_unlock(&b->lock);
216 n->cpu = smp_processor_id();
217 init_swait_queue_head(&n->wq);
218 hlist_add_head(&n->link, &b->list);
220 apf_task_wake_one(n);
222 raw_spin_unlock(&b->lock);
225 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
227 noinstr u32 kvm_read_and_reset_apf_flags(void)
231 if (__this_cpu_read(apf_reason.enabled)) {
232 flags = __this_cpu_read(apf_reason.flags);
233 __this_cpu_write(apf_reason.flags, 0);
238 EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
240 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
242 u32 flags = kvm_read_and_reset_apf_flags();
243 irqentry_state_t state;
248 state = irqentry_enter(regs);
249 instrumentation_begin();
252 * If the host managed to inject an async #PF into an interrupt
253 * disabled region, then die hard as this is not going to end well
254 * and the host side is seriously broken.
256 if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
257 panic("Host injected async #PF in interrupt disabled region\n");
259 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
260 if (unlikely(!(user_mode(regs))))
261 panic("Host injected async #PF in kernel mode\n");
262 /* Page is swapped out by the host. */
263 kvm_async_pf_task_wait_schedule(token);
265 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
268 instrumentation_end();
269 irqentry_exit(regs, state);
273 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
275 struct pt_regs *old_regs = set_irq_regs(regs);
280 inc_irq_stat(irq_hv_callback_count);
282 if (__this_cpu_read(apf_reason.enabled)) {
283 token = __this_cpu_read(apf_reason.token);
284 kvm_async_pf_task_wake(token);
285 __this_cpu_write(apf_reason.token, 0);
286 wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
289 set_irq_regs(old_regs);
292 static void __init paravirt_ops_setup(void)
294 pv_info.name = "KVM";
296 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
297 pv_ops.cpu.io_delay = kvm_io_delay;
299 #ifdef CONFIG_X86_IO_APIC
304 static void kvm_register_steal_time(void)
306 int cpu = smp_processor_id();
307 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
309 if (!has_steal_clock)
312 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
313 pr_info("stealtime: cpu %d, msr %llx\n", cpu,
314 (unsigned long long) slow_virt_to_phys(st));
317 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
319 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
322 * This relies on __test_and_clear_bit to modify the memory
323 * in a way that is atomic with respect to the local CPU.
324 * The hypervisor only accesses this memory from the local CPU so
325 * there's no need for lock or memory barriers.
326 * An optimization barrier is implied in apic write.
328 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
330 apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
333 static void kvm_guest_cpu_init(void)
335 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
336 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
338 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
340 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
341 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
343 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
344 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
346 wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
348 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
349 __this_cpu_write(apf_reason.enabled, 1);
350 pr_info("setup async PF for cpu %d\n", smp_processor_id());
353 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
356 /* Size alignment is implied but just to make it explicit. */
357 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
358 __this_cpu_write(kvm_apic_eoi, 0);
359 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
361 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
365 kvm_register_steal_time();
368 static void kvm_pv_disable_apf(void)
370 if (!__this_cpu_read(apf_reason.enabled))
373 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
374 __this_cpu_write(apf_reason.enabled, 0);
376 pr_info("disable async PF for cpu %d\n", smp_processor_id());
379 static void kvm_disable_steal_time(void)
381 if (!has_steal_clock)
384 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
387 static u64 kvm_steal_clock(int cpu)
390 struct kvm_steal_time *src;
393 src = &per_cpu(steal_time, cpu);
395 version = src->version;
399 } while ((version & 1) || (version != src->version));
404 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
406 early_set_memory_decrypted((unsigned long) ptr, size);
410 * Iterate through all possible CPUs and map the memory region pointed
411 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
413 * Note: we iterate through all possible CPUs to ensure that CPUs
414 * hotplugged will have their per-cpu variable already mapped as
417 static void __init sev_map_percpu_data(void)
424 for_each_possible_cpu(cpu) {
425 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
426 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
427 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
431 static void kvm_guest_cpu_offline(bool shutdown)
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();
442 static int kvm_cpu_online(unsigned int cpu)
446 local_irq_save(flags);
447 kvm_guest_cpu_init();
448 local_irq_restore(flags);
454 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
456 static bool pv_tlb_flush_supported(void)
458 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
459 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
460 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME));
463 static bool pv_ipi_supported(void)
465 return kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI);
468 static bool pv_sched_yield_supported(void)
470 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
471 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
472 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME));
475 #define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG)
477 static void __send_ipi_mask(const struct cpumask *mask, int vector)
480 int cpu, apic_id, icr;
481 int min = 0, max = 0;
483 __uint128_t ipi_bitmap = 0;
489 if (cpumask_empty(mask))
492 local_irq_save(flags);
496 icr = APIC_DM_FIXED | vector;
503 for_each_cpu(cpu, mask) {
504 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
507 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
508 ipi_bitmap <<= min - apic_id;
510 } else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
511 max = apic_id < max ? max : apic_id;
513 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
514 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
515 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
520 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
524 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
525 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
526 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
530 local_irq_restore(flags);
533 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
535 __send_ipi_mask(mask, vector);
538 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
540 unsigned int this_cpu = smp_processor_id();
541 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
542 const struct cpumask *local_mask;
544 cpumask_copy(new_mask, mask);
545 cpumask_clear_cpu(this_cpu, new_mask);
546 local_mask = new_mask;
547 __send_ipi_mask(local_mask, vector);
551 * Set the IPI entry points
553 static void kvm_setup_pv_ipi(void)
555 apic->send_IPI_mask = kvm_send_ipi_mask;
556 apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
557 pr_info("setup PV IPIs\n");
560 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
564 native_send_call_func_ipi(mask);
566 /* Make sure other vCPUs get a chance to run if they need to. */
567 for_each_cpu(cpu, mask) {
568 if (vcpu_is_preempted(cpu)) {
569 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
575 static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
576 const struct flush_tlb_info *info)
580 struct kvm_steal_time *src;
581 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
583 cpumask_copy(flushmask, cpumask);
585 * We have to call flush only on online vCPUs. And
586 * queue flush_on_enter for pre-empted vCPUs
588 for_each_cpu(cpu, flushmask) {
590 * The local vCPU is never preempted, so we do not explicitly
591 * skip check for local vCPU - it will never be cleared from
594 src = &per_cpu(steal_time, cpu);
595 state = READ_ONCE(src->preempted);
596 if ((state & KVM_VCPU_PREEMPTED)) {
597 if (try_cmpxchg(&src->preempted, &state,
598 state | KVM_VCPU_FLUSH_TLB))
599 __cpumask_clear_cpu(cpu, flushmask);
603 native_flush_tlb_multi(flushmask, info);
606 static __init int kvm_alloc_cpumask(void)
610 if (!kvm_para_available() || nopv)
613 if (pv_tlb_flush_supported() || pv_ipi_supported())
614 for_each_possible_cpu(cpu) {
615 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
616 GFP_KERNEL, cpu_to_node(cpu));
621 arch_initcall(kvm_alloc_cpumask);
623 static void __init kvm_smp_prepare_boot_cpu(void)
626 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
627 * shares the guest physical address with the hypervisor.
629 sev_map_percpu_data();
631 kvm_guest_cpu_init();
632 native_smp_prepare_boot_cpu();
636 static int kvm_cpu_down_prepare(unsigned int cpu)
640 local_irq_save(flags);
641 kvm_guest_cpu_offline(false);
642 local_irq_restore(flags);
648 static int kvm_suspend(void)
650 kvm_guest_cpu_offline(false);
655 static void kvm_resume(void)
657 kvm_cpu_online(raw_smp_processor_id());
660 static struct syscore_ops kvm_syscore_ops = {
661 .suspend = kvm_suspend,
662 .resume = kvm_resume,
665 static void kvm_pv_guest_cpu_reboot(void *unused)
667 kvm_guest_cpu_offline(true);
670 static int kvm_pv_reboot_notify(struct notifier_block *nb,
671 unsigned long code, void *unused)
673 if (code == SYS_RESTART)
674 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
678 static struct notifier_block kvm_pv_reboot_nb = {
679 .notifier_call = kvm_pv_reboot_notify,
683 * After a PV feature is registered, the host will keep writing to the
684 * registered memory location. If the guest happens to shutdown, this memory
685 * won't be valid. In cases like kexec, in which you install a new kernel, this
686 * means a random memory location will be kept being written.
688 #ifdef CONFIG_KEXEC_CORE
689 static void kvm_crash_shutdown(struct pt_regs *regs)
691 kvm_guest_cpu_offline(true);
692 native_machine_crash_shutdown(regs);
696 static void __init kvm_guest_init(void)
700 paravirt_ops_setup();
701 register_reboot_notifier(&kvm_pv_reboot_nb);
702 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
703 raw_spin_lock_init(&async_pf_sleepers[i].lock);
705 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
707 static_call_update(pv_steal_clock, kvm_steal_clock);
710 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
711 apic_set_eoi_write(kvm_guest_apic_eoi_write);
713 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
714 static_branch_enable(&kvm_async_pf_enabled);
715 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
719 if (pv_tlb_flush_supported()) {
720 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
721 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
722 pr_info("KVM setup pv remote TLB flush\n");
725 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
726 if (pv_sched_yield_supported()) {
727 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
728 pr_info("setup PV sched yield\n");
730 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
731 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
732 pr_err("failed to install cpu hotplug callbacks\n");
734 sev_map_percpu_data();
735 kvm_guest_cpu_init();
738 #ifdef CONFIG_KEXEC_CORE
739 machine_ops.crash_shutdown = kvm_crash_shutdown;
742 register_syscore_ops(&kvm_syscore_ops);
745 * Hard lockup detection is enabled by default. Disable it, as guests
746 * can get false positives too easily, for example if the host is
749 hardlockup_detector_disable();
752 static noinline uint32_t __kvm_cpuid_base(void)
754 if (boot_cpu_data.cpuid_level < 0)
755 return 0; /* So we don't blow up on old processors */
757 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
758 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
763 static inline uint32_t kvm_cpuid_base(void)
765 static int kvm_cpuid_base = -1;
767 if (kvm_cpuid_base == -1)
768 kvm_cpuid_base = __kvm_cpuid_base();
770 return kvm_cpuid_base;
773 bool kvm_para_available(void)
775 return kvm_cpuid_base() != 0;
777 EXPORT_SYMBOL_GPL(kvm_para_available);
779 unsigned int kvm_arch_para_features(void)
781 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
784 unsigned int kvm_arch_para_hints(void)
786 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
788 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
790 static uint32_t __init kvm_detect(void)
792 return kvm_cpuid_base();
795 static void __init kvm_apic_init(void)
798 if (pv_ipi_supported())
803 static bool __init kvm_msi_ext_dest_id(void)
805 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
808 static void __init kvm_init_platform(void)
811 x86_platform.apic_post_init = kvm_apic_init;
814 #if defined(CONFIG_AMD_MEM_ENCRYPT)
815 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
817 /* RAX and CPL are already in the GHCB */
818 ghcb_set_rbx(ghcb, regs->bx);
819 ghcb_set_rcx(ghcb, regs->cx);
820 ghcb_set_rdx(ghcb, regs->dx);
821 ghcb_set_rsi(ghcb, regs->si);
824 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
826 /* No checking of the return state needed */
831 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
833 .detect = kvm_detect,
834 .type = X86_HYPER_KVM,
835 .init.guest_late_init = kvm_guest_init,
836 .init.x2apic_available = kvm_para_available,
837 .init.msi_ext_dest_id = kvm_msi_ext_dest_id,
838 .init.init_platform = kvm_init_platform,
839 #if defined(CONFIG_AMD_MEM_ENCRYPT)
840 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare,
841 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish,
845 static __init int activate_jump_labels(void)
847 if (has_steal_clock) {
848 static_key_slow_inc(¶virt_steal_enabled);
850 static_key_slow_inc(¶virt_steal_rq_enabled);
855 arch_initcall(activate_jump_labels);
857 #ifdef CONFIG_PARAVIRT_SPINLOCKS
859 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
860 static void kvm_kick_cpu(int cpu)
863 unsigned long flags = 0;
865 apicid = per_cpu(x86_cpu_to_apicid, cpu);
866 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
869 #include <asm/qspinlock.h>
871 static void kvm_wait(u8 *ptr, u8 val)
877 * halt until it's our turn and kicked. Note that we do safe halt
878 * for irq enabled case to avoid hang when lock info is overwritten
879 * in irq spinlock slowpath and no spurious interrupt occur to save us.
881 if (irqs_disabled()) {
882 if (READ_ONCE(*ptr) == val)
887 /* safe_halt() will enable IRQ */
888 if (READ_ONCE(*ptr) == val)
896 __visible bool __kvm_vcpu_is_preempted(long cpu)
898 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
900 return !!(src->preempted & KVM_VCPU_PREEMPTED);
902 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
906 #include <asm/asm-offsets.h>
908 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
911 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
912 * restoring to/from the stack.
915 ".pushsection .text;"
916 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
917 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
918 "__raw_callee_save___kvm_vcpu_is_preempted:"
919 "movq __per_cpu_offset(,%rdi,8), %rax;"
920 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
923 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
929 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
931 void __init kvm_spinlock_init(void)
934 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
935 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
936 * preferred over native qspinlock when vCPU is preempted.
938 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
939 pr_info("PV spinlocks disabled, no host support\n");
944 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
947 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
948 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
952 if (num_possible_cpus() == 1) {
953 pr_info("PV spinlocks disabled, single CPU\n");
958 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
962 pr_info("PV spinlocks enabled\n");
964 __pv_init_lock_hash();
965 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
966 pv_ops.lock.queued_spin_unlock =
967 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
968 pv_ops.lock.wait = kvm_wait;
969 pv_ops.lock.kick = kvm_kick_cpu;
971 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
972 pv_ops.lock.vcpu_is_preempted =
973 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
976 * When PV spinlock is enabled which is preferred over
977 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
978 * Just disable it anyway.
981 static_branch_disable(&virt_spin_lock_key);
984 #endif /* CONFIG_PARAVIRT_SPINLOCKS */
986 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
988 static void kvm_disable_host_haltpoll(void *i)
990 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
993 static void kvm_enable_host_haltpoll(void *i)
995 wrmsrl(MSR_KVM_POLL_CONTROL, 1);
998 void arch_haltpoll_enable(unsigned int cpu)
1000 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1001 pr_err_once("host does not support poll control\n");
1002 pr_err_once("host upgrade recommended\n");
1006 /* Enable guest halt poll disables host halt poll */
1007 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1009 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1011 void arch_haltpoll_disable(unsigned int cpu)
1013 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1016 /* Disable guest halt poll enables host halt poll */
1017 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1019 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);