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 <asm/timer.h>
31 #include <asm/traps.h>
33 #include <asm/tlbflush.h>
35 #include <asm/apicdef.h>
36 #include <asm/hypervisor.h>
38 #include <asm/cpuidle_haltpoll.h>
39 #include <asm/ptrace.h>
42 DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
44 static int kvmapf = 1;
46 static int __init parse_no_kvmapf(char *arg)
52 early_param("no-kvmapf", parse_no_kvmapf);
54 static int steal_acc = 1;
55 static int __init parse_no_stealacc(char *arg)
61 early_param("no-steal-acc", parse_no_stealacc);
63 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
64 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
65 static int has_steal_clock = 0;
68 * No need for any "IO delay" on KVM
70 static void kvm_io_delay(void)
74 #define KVM_TASK_SLEEP_HASHBITS 8
75 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
77 struct kvm_task_sleep_node {
78 struct hlist_node link;
79 struct swait_queue_head wq;
84 static struct kvm_task_sleep_head {
86 struct hlist_head list;
87 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
89 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
94 hlist_for_each(p, &b->list) {
95 struct kvm_task_sleep_node *n =
96 hlist_entry(p, typeof(*n), link);
97 if (n->token == token)
104 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
106 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
107 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
108 struct kvm_task_sleep_node *e;
110 raw_spin_lock(&b->lock);
111 e = _find_apf_task(b, token);
113 /* dummy entry exist -> wake up was delivered ahead of PF */
115 raw_spin_unlock(&b->lock);
121 n->cpu = smp_processor_id();
122 init_swait_queue_head(&n->wq);
123 hlist_add_head(&n->link, &b->list);
124 raw_spin_unlock(&b->lock);
129 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
130 * @token: Token to identify the sleep node entry
132 * Invoked from the async pagefault handling code or from the VM exit page
133 * fault handler. In both cases RCU is watching.
135 void kvm_async_pf_task_wait_schedule(u32 token)
137 struct kvm_task_sleep_node n;
138 DECLARE_SWAITQUEUE(wait);
140 lockdep_assert_irqs_disabled();
142 if (!kvm_async_pf_queue_task(token, &n))
146 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
147 if (hlist_unhashed(&n.link))
154 finish_swait(&n.wq, &wait);
156 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
158 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
160 hlist_del_init(&n->link);
161 if (swq_has_sleeper(&n->wq))
162 swake_up_one(&n->wq);
165 static void apf_task_wake_all(void)
169 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
170 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
171 struct kvm_task_sleep_node *n;
172 struct hlist_node *p, *next;
174 raw_spin_lock(&b->lock);
175 hlist_for_each_safe(p, next, &b->list) {
176 n = hlist_entry(p, typeof(*n), link);
177 if (n->cpu == smp_processor_id())
178 apf_task_wake_one(n);
180 raw_spin_unlock(&b->lock);
184 void kvm_async_pf_task_wake(u32 token)
186 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
187 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
188 struct kvm_task_sleep_node *n;
196 raw_spin_lock(&b->lock);
197 n = _find_apf_task(b, token);
200 * async PF was not yet handled.
201 * Add dummy entry for the token.
203 n = kzalloc(sizeof(*n), GFP_ATOMIC);
206 * Allocation failed! Busy wait while other cpu
209 raw_spin_unlock(&b->lock);
214 n->cpu = smp_processor_id();
215 init_swait_queue_head(&n->wq);
216 hlist_add_head(&n->link, &b->list);
218 apf_task_wake_one(n);
220 raw_spin_unlock(&b->lock);
223 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
225 noinstr u32 kvm_read_and_reset_apf_flags(void)
229 if (__this_cpu_read(apf_reason.enabled)) {
230 flags = __this_cpu_read(apf_reason.flags);
231 __this_cpu_write(apf_reason.flags, 0);
236 EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
238 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
240 u32 flags = kvm_read_and_reset_apf_flags();
241 irqentry_state_t state;
246 state = irqentry_enter(regs);
247 instrumentation_begin();
250 * If the host managed to inject an async #PF into an interrupt
251 * disabled region, then die hard as this is not going to end well
252 * and the host side is seriously broken.
254 if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
255 panic("Host injected async #PF in interrupt disabled region\n");
257 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
258 if (unlikely(!(user_mode(regs))))
259 panic("Host injected async #PF in kernel mode\n");
260 /* Page is swapped out by the host. */
261 kvm_async_pf_task_wait_schedule(token);
263 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
266 instrumentation_end();
267 irqentry_exit(regs, state);
271 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
273 struct pt_regs *old_regs = set_irq_regs(regs);
278 inc_irq_stat(irq_hv_callback_count);
280 if (__this_cpu_read(apf_reason.enabled)) {
281 token = __this_cpu_read(apf_reason.token);
282 kvm_async_pf_task_wake(token);
283 __this_cpu_write(apf_reason.token, 0);
284 wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
287 set_irq_regs(old_regs);
290 static void __init paravirt_ops_setup(void)
292 pv_info.name = "KVM";
294 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
295 pv_ops.cpu.io_delay = kvm_io_delay;
297 #ifdef CONFIG_X86_IO_APIC
302 static void kvm_register_steal_time(void)
304 int cpu = smp_processor_id();
305 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
307 if (!has_steal_clock)
310 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
311 pr_info("stealtime: cpu %d, msr %llx\n", cpu,
312 (unsigned long long) slow_virt_to_phys(st));
315 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
317 static notrace void kvm_guest_apic_eoi_write(u32 reg, u32 val)
320 * This relies on __test_and_clear_bit to modify the memory
321 * in a way that is atomic with respect to the local CPU.
322 * The hypervisor only accesses this memory from the local CPU so
323 * there's no need for lock or memory barriers.
324 * An optimization barrier is implied in apic write.
326 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
328 apic->native_eoi_write(APIC_EOI, APIC_EOI_ACK);
331 static void kvm_guest_cpu_init(void)
333 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
334 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
336 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
338 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
339 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
341 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
342 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
344 wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
346 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
347 __this_cpu_write(apf_reason.enabled, 1);
348 pr_info("KVM setup async PF for cpu %d\n", smp_processor_id());
351 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
354 /* Size alignment is implied but just to make it explicit. */
355 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
356 __this_cpu_write(kvm_apic_eoi, 0);
357 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
359 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
363 kvm_register_steal_time();
366 static void kvm_pv_disable_apf(void)
368 if (!__this_cpu_read(apf_reason.enabled))
371 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
372 __this_cpu_write(apf_reason.enabled, 0);
374 pr_info("Unregister pv shared memory for cpu %d\n", smp_processor_id());
377 static void kvm_pv_guest_cpu_reboot(void *unused)
380 * We disable PV EOI before we load a new kernel by kexec,
381 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
382 * New kernel can re-enable when it boots.
384 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
385 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
386 kvm_pv_disable_apf();
387 kvm_disable_steal_time();
390 static int kvm_pv_reboot_notify(struct notifier_block *nb,
391 unsigned long code, void *unused)
393 if (code == SYS_RESTART)
394 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
398 static struct notifier_block kvm_pv_reboot_nb = {
399 .notifier_call = kvm_pv_reboot_notify,
402 static u64 kvm_steal_clock(int cpu)
405 struct kvm_steal_time *src;
408 src = &per_cpu(steal_time, cpu);
410 version = src->version;
414 } while ((version & 1) || (version != src->version));
419 void kvm_disable_steal_time(void)
421 if (!has_steal_clock)
424 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
427 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
429 early_set_memory_decrypted((unsigned long) ptr, size);
433 * Iterate through all possible CPUs and map the memory region pointed
434 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
436 * Note: we iterate through all possible CPUs to ensure that CPUs
437 * hotplugged will have their per-cpu variable already mapped as
440 static void __init sev_map_percpu_data(void)
447 for_each_possible_cpu(cpu) {
448 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
449 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
450 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
454 static bool pv_tlb_flush_supported(void)
456 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
457 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
458 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME));
461 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
465 static bool pv_ipi_supported(void)
467 return kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI);
470 static bool pv_sched_yield_supported(void)
472 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
473 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
474 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME));
477 #define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG)
479 static void __send_ipi_mask(const struct cpumask *mask, int vector)
482 int cpu, apic_id, icr;
483 int min = 0, max = 0;
485 __uint128_t ipi_bitmap = 0;
491 if (cpumask_empty(mask))
494 local_irq_save(flags);
498 icr = APIC_DM_FIXED | vector;
505 for_each_cpu(cpu, mask) {
506 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
509 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
510 ipi_bitmap <<= min - apic_id;
512 } else if (apic_id < min + KVM_IPI_CLUSTER_SIZE) {
513 max = apic_id < max ? max : apic_id;
515 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
516 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
517 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
522 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
526 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
527 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
528 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
532 local_irq_restore(flags);
535 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
537 __send_ipi_mask(mask, vector);
540 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
542 unsigned int this_cpu = smp_processor_id();
543 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
544 const struct cpumask *local_mask;
546 cpumask_copy(new_mask, mask);
547 cpumask_clear_cpu(this_cpu, new_mask);
548 local_mask = new_mask;
549 __send_ipi_mask(local_mask, vector);
553 * Set the IPI entry points
555 static void kvm_setup_pv_ipi(void)
557 apic->send_IPI_mask = kvm_send_ipi_mask;
558 apic->send_IPI_mask_allbutself = kvm_send_ipi_mask_allbutself;
559 pr_info("setup PV IPIs\n");
562 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
566 native_send_call_func_ipi(mask);
568 /* Make sure other vCPUs get a chance to run if they need to. */
569 for_each_cpu(cpu, mask) {
570 if (vcpu_is_preempted(cpu)) {
571 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
577 static void __init kvm_smp_prepare_boot_cpu(void)
580 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
581 * shares the guest physical address with the hypervisor.
583 sev_map_percpu_data();
585 kvm_guest_cpu_init();
586 native_smp_prepare_boot_cpu();
590 static void kvm_guest_cpu_offline(void)
592 kvm_disable_steal_time();
593 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
594 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
595 kvm_pv_disable_apf();
599 static int kvm_cpu_online(unsigned int cpu)
602 kvm_guest_cpu_init();
607 static int kvm_cpu_down_prepare(unsigned int cpu)
610 kvm_guest_cpu_offline();
616 static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
617 const struct flush_tlb_info *info)
621 struct kvm_steal_time *src;
622 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
624 cpumask_copy(flushmask, cpumask);
626 * We have to call flush only on online vCPUs. And
627 * queue flush_on_enter for pre-empted vCPUs
629 for_each_cpu(cpu, flushmask) {
631 * The local vCPU is never preempted, so we do not explicitly
632 * skip check for local vCPU - it will never be cleared from
635 src = &per_cpu(steal_time, cpu);
636 state = READ_ONCE(src->preempted);
637 if ((state & KVM_VCPU_PREEMPTED)) {
638 if (try_cmpxchg(&src->preempted, &state,
639 state | KVM_VCPU_FLUSH_TLB))
640 __cpumask_clear_cpu(cpu, flushmask);
644 native_flush_tlb_multi(flushmask, info);
647 static void __init kvm_guest_init(void)
651 paravirt_ops_setup();
652 register_reboot_notifier(&kvm_pv_reboot_nb);
653 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
654 raw_spin_lock_init(&async_pf_sleepers[i].lock);
656 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
658 static_call_update(pv_steal_clock, kvm_steal_clock);
661 if (pv_tlb_flush_supported()) {
662 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
663 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
664 pr_info("KVM setup pv remote TLB flush\n");
667 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
668 apic_set_eoi_write(kvm_guest_apic_eoi_write);
670 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
671 static_branch_enable(&kvm_async_pf_enabled);
672 alloc_intr_gate(HYPERVISOR_CALLBACK_VECTOR, asm_sysvec_kvm_asyncpf_interrupt);
676 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
677 if (pv_sched_yield_supported()) {
678 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
679 pr_info("setup PV sched yield\n");
681 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
682 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
683 pr_err("failed to install cpu hotplug callbacks\n");
685 sev_map_percpu_data();
686 kvm_guest_cpu_init();
690 * Hard lockup detection is enabled by default. Disable it, as guests
691 * can get false positives too easily, for example if the host is
694 hardlockup_detector_disable();
697 static noinline uint32_t __kvm_cpuid_base(void)
699 if (boot_cpu_data.cpuid_level < 0)
700 return 0; /* So we don't blow up on old processors */
702 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
703 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
708 static inline uint32_t kvm_cpuid_base(void)
710 static int kvm_cpuid_base = -1;
712 if (kvm_cpuid_base == -1)
713 kvm_cpuid_base = __kvm_cpuid_base();
715 return kvm_cpuid_base;
718 bool kvm_para_available(void)
720 return kvm_cpuid_base() != 0;
722 EXPORT_SYMBOL_GPL(kvm_para_available);
724 unsigned int kvm_arch_para_features(void)
726 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
729 unsigned int kvm_arch_para_hints(void)
731 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
733 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
735 static uint32_t __init kvm_detect(void)
737 return kvm_cpuid_base();
740 static void __init kvm_apic_init(void)
742 #if defined(CONFIG_SMP)
743 if (pv_ipi_supported())
748 static bool __init kvm_msi_ext_dest_id(void)
750 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
753 static void __init kvm_init_platform(void)
756 x86_platform.apic_post_init = kvm_apic_init;
759 #if defined(CONFIG_AMD_MEM_ENCRYPT)
760 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
762 /* RAX and CPL are already in the GHCB */
763 ghcb_set_rbx(ghcb, regs->bx);
764 ghcb_set_rcx(ghcb, regs->cx);
765 ghcb_set_rdx(ghcb, regs->dx);
766 ghcb_set_rsi(ghcb, regs->si);
769 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
771 /* No checking of the return state needed */
776 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
778 .detect = kvm_detect,
779 .type = X86_HYPER_KVM,
780 .init.guest_late_init = kvm_guest_init,
781 .init.x2apic_available = kvm_para_available,
782 .init.msi_ext_dest_id = kvm_msi_ext_dest_id,
783 .init.init_platform = kvm_init_platform,
784 #if defined(CONFIG_AMD_MEM_ENCRYPT)
785 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare,
786 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish,
790 static __init int activate_jump_labels(void)
792 if (has_steal_clock) {
793 static_key_slow_inc(¶virt_steal_enabled);
795 static_key_slow_inc(¶virt_steal_rq_enabled);
800 arch_initcall(activate_jump_labels);
802 static __init int kvm_alloc_cpumask(void)
807 if (!kvm_para_available() || nopv)
810 if (pv_tlb_flush_supported())
813 #if defined(CONFIG_SMP)
814 if (pv_ipi_supported())
819 for_each_possible_cpu(cpu) {
820 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
821 GFP_KERNEL, cpu_to_node(cpu));
826 arch_initcall(kvm_alloc_cpumask);
828 #ifdef CONFIG_PARAVIRT_SPINLOCKS
830 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
831 static void kvm_kick_cpu(int cpu)
834 unsigned long flags = 0;
836 apicid = per_cpu(x86_cpu_to_apicid, cpu);
837 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
840 #include <asm/qspinlock.h>
842 static void kvm_wait(u8 *ptr, u8 val)
848 * halt until it's our turn and kicked. Note that we do safe halt
849 * for irq enabled case to avoid hang when lock info is overwritten
850 * in irq spinlock slowpath and no spurious interrupt occur to save us.
852 if (irqs_disabled()) {
853 if (READ_ONCE(*ptr) == val)
858 if (READ_ONCE(*ptr) == val)
866 __visible bool __kvm_vcpu_is_preempted(long cpu)
868 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
870 return !!(src->preempted & KVM_VCPU_PREEMPTED);
872 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
876 #include <asm/asm-offsets.h>
878 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
881 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
882 * restoring to/from the stack.
885 ".pushsection .text;"
886 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
887 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
888 "__raw_callee_save___kvm_vcpu_is_preempted:"
889 "movq __per_cpu_offset(,%rdi,8), %rax;"
890 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
893 ".size __raw_callee_save___kvm_vcpu_is_preempted, .-__raw_callee_save___kvm_vcpu_is_preempted;"
899 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
901 void __init kvm_spinlock_init(void)
904 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
905 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
906 * preferred over native qspinlock when vCPU is preempted.
908 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
909 pr_info("PV spinlocks disabled, no host support\n");
914 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
917 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
918 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
922 if (num_possible_cpus() == 1) {
923 pr_info("PV spinlocks disabled, single CPU\n");
928 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
932 pr_info("PV spinlocks enabled\n");
934 __pv_init_lock_hash();
935 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
936 pv_ops.lock.queued_spin_unlock =
937 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
938 pv_ops.lock.wait = kvm_wait;
939 pv_ops.lock.kick = kvm_kick_cpu;
941 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
942 pv_ops.lock.vcpu_is_preempted =
943 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
946 * When PV spinlock is enabled which is preferred over
947 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
948 * Just disable it anyway.
951 static_branch_disable(&virt_spin_lock_key);
954 #endif /* CONFIG_PARAVIRT_SPINLOCKS */
956 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
958 static void kvm_disable_host_haltpoll(void *i)
960 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
963 static void kvm_enable_host_haltpoll(void *i)
965 wrmsrl(MSR_KVM_POLL_CONTROL, 1);
968 void arch_haltpoll_enable(unsigned int cpu)
970 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
971 pr_err_once("host does not support poll control\n");
972 pr_err_once("host upgrade recommended\n");
976 /* Enable guest halt poll disables host halt poll */
977 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
979 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
981 void arch_haltpoll_disable(unsigned int cpu)
983 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
986 /* Disable guest halt poll enables host halt poll */
987 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
989 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);