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 <linux/cc_platform.h>
31 #include <linux/efi.h>
32 #include <asm/timer.h>
34 #include <asm/traps.h>
36 #include <asm/tlbflush.h>
38 #include <asm/apicdef.h>
39 #include <asm/hypervisor.h>
41 #include <asm/cpuidle_haltpoll.h>
42 #include <asm/ptrace.h>
43 #include <asm/reboot.h>
45 #include <asm/e820/api.h>
47 DEFINE_STATIC_KEY_FALSE(kvm_async_pf_enabled);
49 static int kvmapf = 1;
51 static int __init parse_no_kvmapf(char *arg)
57 early_param("no-kvmapf", parse_no_kvmapf);
59 static int steal_acc = 1;
60 static int __init parse_no_stealacc(char *arg)
66 early_param("no-steal-acc", parse_no_stealacc);
68 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
69 DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64) __visible;
70 static int has_steal_clock = 0;
72 static int has_guest_poll = 0;
74 * No need for any "IO delay" on KVM
76 static void kvm_io_delay(void)
80 #define KVM_TASK_SLEEP_HASHBITS 8
81 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
83 struct kvm_task_sleep_node {
84 struct hlist_node link;
85 struct swait_queue_head wq;
90 static struct kvm_task_sleep_head {
92 struct hlist_head list;
93 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
95 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
100 hlist_for_each(p, &b->list) {
101 struct kvm_task_sleep_node *n =
102 hlist_entry(p, typeof(*n), link);
103 if (n->token == token)
110 static bool kvm_async_pf_queue_task(u32 token, struct kvm_task_sleep_node *n)
112 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
113 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
114 struct kvm_task_sleep_node *e;
116 raw_spin_lock(&b->lock);
117 e = _find_apf_task(b, token);
119 /* dummy entry exist -> wake up was delivered ahead of PF */
121 raw_spin_unlock(&b->lock);
127 n->cpu = smp_processor_id();
128 init_swait_queue_head(&n->wq);
129 hlist_add_head(&n->link, &b->list);
130 raw_spin_unlock(&b->lock);
135 * kvm_async_pf_task_wait_schedule - Wait for pagefault to be handled
136 * @token: Token to identify the sleep node entry
138 * Invoked from the async pagefault handling code or from the VM exit page
139 * fault handler. In both cases RCU is watching.
141 void kvm_async_pf_task_wait_schedule(u32 token)
143 struct kvm_task_sleep_node n;
144 DECLARE_SWAITQUEUE(wait);
146 lockdep_assert_irqs_disabled();
148 if (!kvm_async_pf_queue_task(token, &n))
152 prepare_to_swait_exclusive(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
153 if (hlist_unhashed(&n.link))
160 finish_swait(&n.wq, &wait);
162 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait_schedule);
164 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
166 hlist_del_init(&n->link);
167 if (swq_has_sleeper(&n->wq))
168 swake_up_one(&n->wq);
171 static void apf_task_wake_all(void)
175 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
176 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
177 struct kvm_task_sleep_node *n;
178 struct hlist_node *p, *next;
180 raw_spin_lock(&b->lock);
181 hlist_for_each_safe(p, next, &b->list) {
182 n = hlist_entry(p, typeof(*n), link);
183 if (n->cpu == smp_processor_id())
184 apf_task_wake_one(n);
186 raw_spin_unlock(&b->lock);
190 void kvm_async_pf_task_wake(u32 token)
192 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
193 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
194 struct kvm_task_sleep_node *n, *dummy = NULL;
202 raw_spin_lock(&b->lock);
203 n = _find_apf_task(b, token);
206 * Async #PF not yet handled, add a dummy entry for the token.
207 * Allocating the token must be down outside of the raw lock
208 * as the allocator is preemptible on PREEMPT_RT kernels.
211 raw_spin_unlock(&b->lock);
212 dummy = kzalloc(sizeof(*dummy), GFP_ATOMIC);
215 * Continue looping on allocation failure, eventually
216 * the async #PF will be handled and allocating a new
217 * node will be unnecessary.
223 * Recheck for async #PF completion before enqueueing
224 * the dummy token to avoid duplicate list entries.
228 dummy->token = token;
229 dummy->cpu = smp_processor_id();
230 init_swait_queue_head(&dummy->wq);
231 hlist_add_head(&dummy->link, &b->list);
234 apf_task_wake_one(n);
236 raw_spin_unlock(&b->lock);
238 /* A dummy token might be allocated and ultimately not used. */
241 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
243 noinstr u32 kvm_read_and_reset_apf_flags(void)
247 if (__this_cpu_read(apf_reason.enabled)) {
248 flags = __this_cpu_read(apf_reason.flags);
249 __this_cpu_write(apf_reason.flags, 0);
254 EXPORT_SYMBOL_GPL(kvm_read_and_reset_apf_flags);
256 noinstr bool __kvm_handle_async_pf(struct pt_regs *regs, u32 token)
258 u32 flags = kvm_read_and_reset_apf_flags();
259 irqentry_state_t state;
264 state = irqentry_enter(regs);
265 instrumentation_begin();
268 * If the host managed to inject an async #PF into an interrupt
269 * disabled region, then die hard as this is not going to end well
270 * and the host side is seriously broken.
272 if (unlikely(!(regs->flags & X86_EFLAGS_IF)))
273 panic("Host injected async #PF in interrupt disabled region\n");
275 if (flags & KVM_PV_REASON_PAGE_NOT_PRESENT) {
276 if (unlikely(!(user_mode(regs))))
277 panic("Host injected async #PF in kernel mode\n");
278 /* Page is swapped out by the host. */
279 kvm_async_pf_task_wait_schedule(token);
281 WARN_ONCE(1, "Unexpected async PF flags: %x\n", flags);
284 instrumentation_end();
285 irqentry_exit(regs, state);
289 DEFINE_IDTENTRY_SYSVEC(sysvec_kvm_asyncpf_interrupt)
291 struct pt_regs *old_regs = set_irq_regs(regs);
296 inc_irq_stat(irq_hv_callback_count);
298 if (__this_cpu_read(apf_reason.enabled)) {
299 token = __this_cpu_read(apf_reason.token);
300 kvm_async_pf_task_wake(token);
301 __this_cpu_write(apf_reason.token, 0);
302 wrmsrl(MSR_KVM_ASYNC_PF_ACK, 1);
305 set_irq_regs(old_regs);
308 static void __init paravirt_ops_setup(void)
310 pv_info.name = "KVM";
312 if (kvm_para_has_feature(KVM_FEATURE_NOP_IO_DELAY))
313 pv_ops.cpu.io_delay = kvm_io_delay;
315 #ifdef CONFIG_X86_IO_APIC
320 static void kvm_register_steal_time(void)
322 int cpu = smp_processor_id();
323 struct kvm_steal_time *st = &per_cpu(steal_time, cpu);
325 if (!has_steal_clock)
328 wrmsrl(MSR_KVM_STEAL_TIME, (slow_virt_to_phys(st) | KVM_MSR_ENABLED));
329 pr_debug("stealtime: cpu %d, msr %llx\n", cpu,
330 (unsigned long long) slow_virt_to_phys(st));
333 static DEFINE_PER_CPU_DECRYPTED(unsigned long, kvm_apic_eoi) = KVM_PV_EOI_DISABLED;
335 static notrace __maybe_unused void kvm_guest_apic_eoi_write(void)
338 * This relies on __test_and_clear_bit to modify the memory
339 * in a way that is atomic with respect to the local CPU.
340 * The hypervisor only accesses this memory from the local CPU so
341 * there's no need for lock or memory barriers.
342 * An optimization barrier is implied in apic write.
344 if (__test_and_clear_bit(KVM_PV_EOI_BIT, this_cpu_ptr(&kvm_apic_eoi)))
349 static void kvm_guest_cpu_init(void)
351 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
354 WARN_ON_ONCE(!static_branch_likely(&kvm_async_pf_enabled));
356 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
357 pa |= KVM_ASYNC_PF_ENABLED | KVM_ASYNC_PF_DELIVERY_AS_INT;
359 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
360 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
362 wrmsrl(MSR_KVM_ASYNC_PF_INT, HYPERVISOR_CALLBACK_VECTOR);
364 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
365 __this_cpu_write(apf_reason.enabled, 1);
366 pr_debug("setup async PF for cpu %d\n", smp_processor_id());
369 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
372 /* Size alignment is implied but just to make it explicit. */
373 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
374 __this_cpu_write(kvm_apic_eoi, 0);
375 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
377 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
381 kvm_register_steal_time();
384 static void kvm_pv_disable_apf(void)
386 if (!__this_cpu_read(apf_reason.enabled))
389 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
390 __this_cpu_write(apf_reason.enabled, 0);
392 pr_debug("disable async PF for cpu %d\n", smp_processor_id());
395 static void kvm_disable_steal_time(void)
397 if (!has_steal_clock)
400 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
403 static u64 kvm_steal_clock(int cpu)
406 struct kvm_steal_time *src;
409 src = &per_cpu(steal_time, cpu);
411 version = src->version;
415 } while ((version & 1) || (version != src->version));
420 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
422 early_set_memory_decrypted((unsigned long) ptr, size);
426 * Iterate through all possible CPUs and map the memory region pointed
427 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
429 * Note: we iterate through all possible CPUs to ensure that CPUs
430 * hotplugged will have their per-cpu variable already mapped as
433 static void __init sev_map_percpu_data(void)
437 if (cc_vendor != CC_VENDOR_AMD ||
438 !cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT))
441 for_each_possible_cpu(cpu) {
442 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
443 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
444 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
448 static void kvm_guest_cpu_offline(bool shutdown)
450 kvm_disable_steal_time();
451 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
452 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
453 if (kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
454 wrmsrl(MSR_KVM_MIGRATION_CONTROL, 0);
455 kvm_pv_disable_apf();
461 static int kvm_cpu_online(unsigned int cpu)
465 local_irq_save(flags);
466 kvm_guest_cpu_init();
467 local_irq_restore(flags);
473 static DEFINE_PER_CPU(cpumask_var_t, __pv_cpu_mask);
475 static bool pv_tlb_flush_supported(void)
477 return (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
478 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
479 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
480 !boot_cpu_has(X86_FEATURE_MWAIT) &&
481 (num_possible_cpus() != 1));
484 static bool pv_ipi_supported(void)
486 return (kvm_para_has_feature(KVM_FEATURE_PV_SEND_IPI) &&
487 (num_possible_cpus() != 1));
490 static bool pv_sched_yield_supported(void)
492 return (kvm_para_has_feature(KVM_FEATURE_PV_SCHED_YIELD) &&
493 !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
494 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME) &&
495 !boot_cpu_has(X86_FEATURE_MWAIT) &&
496 (num_possible_cpus() != 1));
499 #define KVM_IPI_CLUSTER_SIZE (2 * BITS_PER_LONG)
501 static void __send_ipi_mask(const struct cpumask *mask, int vector)
504 int cpu, min = 0, max = 0;
506 __uint128_t ipi_bitmap = 0;
513 if (cpumask_empty(mask))
516 local_irq_save(flags);
520 icr = APIC_DM_FIXED | vector;
527 for_each_cpu(cpu, mask) {
528 apic_id = per_cpu(x86_cpu_to_apicid, cpu);
531 } else if (apic_id < min && max - apic_id < KVM_IPI_CLUSTER_SIZE) {
532 ipi_bitmap <<= min - apic_id;
534 } else if (apic_id > min && apic_id < min + KVM_IPI_CLUSTER_SIZE) {
535 max = apic_id < max ? max : apic_id;
537 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
538 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
539 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
544 __set_bit(apic_id - min, (unsigned long *)&ipi_bitmap);
548 ret = kvm_hypercall4(KVM_HC_SEND_IPI, (unsigned long)ipi_bitmap,
549 (unsigned long)(ipi_bitmap >> BITS_PER_LONG), min, icr);
550 WARN_ONCE(ret < 0, "kvm-guest: failed to send PV IPI: %ld",
554 local_irq_restore(flags);
557 static void kvm_send_ipi_mask(const struct cpumask *mask, int vector)
559 __send_ipi_mask(mask, vector);
562 static void kvm_send_ipi_mask_allbutself(const struct cpumask *mask, int vector)
564 unsigned int this_cpu = smp_processor_id();
565 struct cpumask *new_mask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
566 const struct cpumask *local_mask;
568 cpumask_copy(new_mask, mask);
569 cpumask_clear_cpu(this_cpu, new_mask);
570 local_mask = new_mask;
571 __send_ipi_mask(local_mask, vector);
574 static int __init setup_efi_kvm_sev_migration(void)
576 efi_char16_t efi_sev_live_migration_enabled[] = L"SevLiveMigrationEnabled";
577 efi_guid_t efi_variable_guid = AMD_SEV_MEM_ENCRYPT_GUID;
582 if (!cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) ||
583 !kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL))
586 if (!efi_enabled(EFI_BOOT))
589 if (!efi_enabled(EFI_RUNTIME_SERVICES)) {
590 pr_info("%s : EFI runtime services are not enabled\n", __func__);
594 size = sizeof(enabled);
596 /* Get variable contents into buffer */
597 status = efi.get_variable(efi_sev_live_migration_enabled,
598 &efi_variable_guid, NULL, &size, &enabled);
600 if (status == EFI_NOT_FOUND) {
601 pr_info("%s : EFI live migration variable not found\n", __func__);
605 if (status != EFI_SUCCESS) {
606 pr_info("%s : EFI variable retrieval failed\n", __func__);
611 pr_info("%s: live migration disabled in EFI\n", __func__);
615 pr_info("%s : live migration enabled in EFI\n", __func__);
616 wrmsrl(MSR_KVM_MIGRATION_CONTROL, KVM_MIGRATION_READY);
621 late_initcall(setup_efi_kvm_sev_migration);
624 * Set the IPI entry points
626 static __init void kvm_setup_pv_ipi(void)
628 apic_update_callback(send_IPI_mask, kvm_send_ipi_mask);
629 apic_update_callback(send_IPI_mask_allbutself, kvm_send_ipi_mask_allbutself);
630 pr_info("setup PV IPIs\n");
633 static void kvm_smp_send_call_func_ipi(const struct cpumask *mask)
637 native_send_call_func_ipi(mask);
639 /* Make sure other vCPUs get a chance to run if they need to. */
640 for_each_cpu(cpu, mask) {
641 if (!idle_cpu(cpu) && vcpu_is_preempted(cpu)) {
642 kvm_hypercall1(KVM_HC_SCHED_YIELD, per_cpu(x86_cpu_to_apicid, cpu));
648 static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
649 const struct flush_tlb_info *info)
653 struct kvm_steal_time *src;
654 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_cpu_mask);
656 cpumask_copy(flushmask, cpumask);
658 * We have to call flush only on online vCPUs. And
659 * queue flush_on_enter for pre-empted vCPUs
661 for_each_cpu(cpu, flushmask) {
663 * The local vCPU is never preempted, so we do not explicitly
664 * skip check for local vCPU - it will never be cleared from
667 src = &per_cpu(steal_time, cpu);
668 state = READ_ONCE(src->preempted);
669 if ((state & KVM_VCPU_PREEMPTED)) {
670 if (try_cmpxchg(&src->preempted, &state,
671 state | KVM_VCPU_FLUSH_TLB))
672 __cpumask_clear_cpu(cpu, flushmask);
676 native_flush_tlb_multi(flushmask, info);
679 static __init int kvm_alloc_cpumask(void)
683 if (!kvm_para_available() || nopv)
686 if (pv_tlb_flush_supported() || pv_ipi_supported())
687 for_each_possible_cpu(cpu) {
688 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_cpu_mask, cpu),
689 GFP_KERNEL, cpu_to_node(cpu));
694 arch_initcall(kvm_alloc_cpumask);
696 static void __init kvm_smp_prepare_boot_cpu(void)
699 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
700 * shares the guest physical address with the hypervisor.
702 sev_map_percpu_data();
704 kvm_guest_cpu_init();
705 native_smp_prepare_boot_cpu();
709 static int kvm_cpu_down_prepare(unsigned int cpu)
713 local_irq_save(flags);
714 kvm_guest_cpu_offline(false);
715 local_irq_restore(flags);
721 static int kvm_suspend(void)
725 kvm_guest_cpu_offline(false);
727 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
728 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
729 rdmsrl(MSR_KVM_POLL_CONTROL, val);
730 has_guest_poll = !(val & 1);
735 static void kvm_resume(void)
737 kvm_cpu_online(raw_smp_processor_id());
739 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
740 if (kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL) && has_guest_poll)
741 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
745 static struct syscore_ops kvm_syscore_ops = {
746 .suspend = kvm_suspend,
747 .resume = kvm_resume,
750 static void kvm_pv_guest_cpu_reboot(void *unused)
752 kvm_guest_cpu_offline(true);
755 static int kvm_pv_reboot_notify(struct notifier_block *nb,
756 unsigned long code, void *unused)
758 if (code == SYS_RESTART)
759 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
763 static struct notifier_block kvm_pv_reboot_nb = {
764 .notifier_call = kvm_pv_reboot_notify,
768 * After a PV feature is registered, the host will keep writing to the
769 * registered memory location. If the guest happens to shutdown, this memory
770 * won't be valid. In cases like kexec, in which you install a new kernel, this
771 * means a random memory location will be kept being written.
773 #ifdef CONFIG_CRASH_DUMP
774 static void kvm_crash_shutdown(struct pt_regs *regs)
776 kvm_guest_cpu_offline(true);
777 native_machine_crash_shutdown(regs);
781 #if defined(CONFIG_X86_32) || !defined(CONFIG_SMP)
782 bool __kvm_vcpu_is_preempted(long cpu);
784 __visible bool __kvm_vcpu_is_preempted(long cpu)
786 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
788 return !!(src->preempted & KVM_VCPU_PREEMPTED);
790 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
794 #include <asm/asm-offsets.h>
796 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
799 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
800 * restoring to/from the stack.
802 #define PV_VCPU_PREEMPTED_ASM \
803 "movq __per_cpu_offset(,%rdi,8), %rax\n\t" \
804 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax)\n\t" \
807 DEFINE_ASM_FUNC(__raw_callee_save___kvm_vcpu_is_preempted,
808 PV_VCPU_PREEMPTED_ASM, .text);
811 static void __init kvm_guest_init(void)
815 paravirt_ops_setup();
816 register_reboot_notifier(&kvm_pv_reboot_nb);
817 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
818 raw_spin_lock_init(&async_pf_sleepers[i].lock);
820 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
822 static_call_update(pv_steal_clock, kvm_steal_clock);
824 pv_ops.lock.vcpu_is_preempted =
825 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
828 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
829 apic_update_callback(eoi, kvm_guest_apic_eoi_write);
831 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_INT) && kvmapf) {
832 static_branch_enable(&kvm_async_pf_enabled);
833 sysvec_install(HYPERVISOR_CALLBACK_VECTOR, sysvec_kvm_asyncpf_interrupt);
837 if (pv_tlb_flush_supported()) {
838 pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
839 pv_ops.mmu.tlb_remove_table = tlb_remove_table;
840 pr_info("KVM setup pv remote TLB flush\n");
843 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
844 if (pv_sched_yield_supported()) {
845 smp_ops.send_call_func_ipi = kvm_smp_send_call_func_ipi;
846 pr_info("setup PV sched yield\n");
848 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
849 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
850 pr_err("failed to install cpu hotplug callbacks\n");
852 sev_map_percpu_data();
853 kvm_guest_cpu_init();
856 #ifdef CONFIG_CRASH_DUMP
857 machine_ops.crash_shutdown = kvm_crash_shutdown;
860 register_syscore_ops(&kvm_syscore_ops);
863 * Hard lockup detection is enabled by default. Disable it, as guests
864 * can get false positives too easily, for example if the host is
867 hardlockup_detector_disable();
870 static noinline uint32_t __kvm_cpuid_base(void)
872 if (boot_cpu_data.cpuid_level < 0)
873 return 0; /* So we don't blow up on old processors */
875 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
876 return hypervisor_cpuid_base(KVM_SIGNATURE, 0);
881 static inline uint32_t kvm_cpuid_base(void)
883 static int kvm_cpuid_base = -1;
885 if (kvm_cpuid_base == -1)
886 kvm_cpuid_base = __kvm_cpuid_base();
888 return kvm_cpuid_base;
891 bool kvm_para_available(void)
893 return kvm_cpuid_base() != 0;
895 EXPORT_SYMBOL_GPL(kvm_para_available);
897 unsigned int kvm_arch_para_features(void)
899 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
902 unsigned int kvm_arch_para_hints(void)
904 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
906 EXPORT_SYMBOL_GPL(kvm_arch_para_hints);
908 static uint32_t __init kvm_detect(void)
910 return kvm_cpuid_base();
913 static void __init kvm_apic_init(void)
916 if (pv_ipi_supported())
921 static bool __init kvm_msi_ext_dest_id(void)
923 return kvm_para_has_feature(KVM_FEATURE_MSI_EXT_DEST_ID);
926 static void kvm_sev_hc_page_enc_status(unsigned long pfn, int npages, bool enc)
928 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, pfn << PAGE_SHIFT, npages,
929 KVM_MAP_GPA_RANGE_ENC_STAT(enc) | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
932 static void __init kvm_init_platform(void)
934 if (cc_platform_has(CC_ATTR_GUEST_MEM_ENCRYPT) &&
935 kvm_para_has_feature(KVM_FEATURE_MIGRATION_CONTROL)) {
936 unsigned long nr_pages;
939 pv_ops.mmu.notify_page_enc_status_changed =
940 kvm_sev_hc_page_enc_status;
943 * Reset the host's shared pages list related to kernel
944 * specific page encryption status settings before we load a
945 * new kernel by kexec. Reset the page encryption status
946 * during early boot instead of just before kexec to avoid SMP
947 * races during kvm_pv_guest_cpu_reboot().
948 * NOTE: We cannot reset the complete shared pages list
949 * here as we need to retain the UEFI/OVMF firmware
953 for (i = 0; i < e820_table->nr_entries; i++) {
954 struct e820_entry *entry = &e820_table->entries[i];
956 if (entry->type != E820_TYPE_RAM)
959 nr_pages = DIV_ROUND_UP(entry->size, PAGE_SIZE);
961 kvm_sev_hypercall3(KVM_HC_MAP_GPA_RANGE, entry->addr,
963 KVM_MAP_GPA_RANGE_ENCRYPTED | KVM_MAP_GPA_RANGE_PAGE_SZ_4K);
967 * Ensure that _bss_decrypted section is marked as decrypted in the
970 early_set_mem_enc_dec_hypercall((unsigned long)__start_bss_decrypted,
971 __end_bss_decrypted - __start_bss_decrypted, 0);
974 * If not booted using EFI, enable Live migration support.
976 if (!efi_enabled(EFI_BOOT))
977 wrmsrl(MSR_KVM_MIGRATION_CONTROL,
978 KVM_MIGRATION_READY);
981 x86_platform.apic_post_init = kvm_apic_init;
984 #if defined(CONFIG_AMD_MEM_ENCRYPT)
985 static void kvm_sev_es_hcall_prepare(struct ghcb *ghcb, struct pt_regs *regs)
987 /* RAX and CPL are already in the GHCB */
988 ghcb_set_rbx(ghcb, regs->bx);
989 ghcb_set_rcx(ghcb, regs->cx);
990 ghcb_set_rdx(ghcb, regs->dx);
991 ghcb_set_rsi(ghcb, regs->si);
994 static bool kvm_sev_es_hcall_finish(struct ghcb *ghcb, struct pt_regs *regs)
996 /* No checking of the return state needed */
1001 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
1003 .detect = kvm_detect,
1004 .type = X86_HYPER_KVM,
1005 .init.guest_late_init = kvm_guest_init,
1006 .init.x2apic_available = kvm_para_available,
1007 .init.msi_ext_dest_id = kvm_msi_ext_dest_id,
1008 .init.init_platform = kvm_init_platform,
1009 #if defined(CONFIG_AMD_MEM_ENCRYPT)
1010 .runtime.sev_es_hcall_prepare = kvm_sev_es_hcall_prepare,
1011 .runtime.sev_es_hcall_finish = kvm_sev_es_hcall_finish,
1015 static __init int activate_jump_labels(void)
1017 if (has_steal_clock) {
1018 static_key_slow_inc(¶virt_steal_enabled);
1020 static_key_slow_inc(¶virt_steal_rq_enabled);
1025 arch_initcall(activate_jump_labels);
1027 #ifdef CONFIG_PARAVIRT_SPINLOCKS
1029 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
1030 static void kvm_kick_cpu(int cpu)
1032 unsigned long flags = 0;
1035 apicid = per_cpu(x86_cpu_to_apicid, cpu);
1036 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
1039 #include <asm/qspinlock.h>
1041 static void kvm_wait(u8 *ptr, u8 val)
1047 * halt until it's our turn and kicked. Note that we do safe halt
1048 * for irq enabled case to avoid hang when lock info is overwritten
1049 * in irq spinlock slowpath and no spurious interrupt occur to save us.
1051 if (irqs_disabled()) {
1052 if (READ_ONCE(*ptr) == val)
1055 local_irq_disable();
1057 /* safe_halt() will enable IRQ */
1058 if (READ_ONCE(*ptr) == val)
1066 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
1068 void __init kvm_spinlock_init(void)
1071 * In case host doesn't support KVM_FEATURE_PV_UNHALT there is still an
1072 * advantage of keeping virt_spin_lock_key enabled: virt_spin_lock() is
1073 * preferred over native qspinlock when vCPU is preempted.
1075 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT)) {
1076 pr_info("PV spinlocks disabled, no host support\n");
1081 * Disable PV spinlocks and use native qspinlock when dedicated pCPUs
1084 if (kvm_para_has_hint(KVM_HINTS_REALTIME)) {
1085 pr_info("PV spinlocks disabled with KVM_HINTS_REALTIME hints\n");
1089 if (num_possible_cpus() == 1) {
1090 pr_info("PV spinlocks disabled, single CPU\n");
1095 pr_info("PV spinlocks disabled, forced by \"nopvspin\" parameter\n");
1099 pr_info("PV spinlocks enabled\n");
1101 __pv_init_lock_hash();
1102 pv_ops.lock.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
1103 pv_ops.lock.queued_spin_unlock =
1104 PV_CALLEE_SAVE(__pv_queued_spin_unlock);
1105 pv_ops.lock.wait = kvm_wait;
1106 pv_ops.lock.kick = kvm_kick_cpu;
1109 * When PV spinlock is enabled which is preferred over
1110 * virt_spin_lock(), virt_spin_lock_key's value is meaningless.
1111 * Just disable it anyway.
1114 static_branch_disable(&virt_spin_lock_key);
1117 #endif /* CONFIG_PARAVIRT_SPINLOCKS */
1119 #ifdef CONFIG_ARCH_CPUIDLE_HALTPOLL
1121 static void kvm_disable_host_haltpoll(void *i)
1123 wrmsrl(MSR_KVM_POLL_CONTROL, 0);
1126 static void kvm_enable_host_haltpoll(void *i)
1128 wrmsrl(MSR_KVM_POLL_CONTROL, 1);
1131 void arch_haltpoll_enable(unsigned int cpu)
1133 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL)) {
1134 pr_err_once("host does not support poll control\n");
1135 pr_err_once("host upgrade recommended\n");
1139 /* Enable guest halt poll disables host halt poll */
1140 smp_call_function_single(cpu, kvm_disable_host_haltpoll, NULL, 1);
1142 EXPORT_SYMBOL_GPL(arch_haltpoll_enable);
1144 void arch_haltpoll_disable(unsigned int cpu)
1146 if (!kvm_para_has_feature(KVM_FEATURE_POLL_CONTROL))
1149 /* Disable guest halt poll enables host halt poll */
1150 smp_call_function_single(cpu, kvm_enable_host_haltpoll, NULL, 1);
1152 EXPORT_SYMBOL_GPL(arch_haltpoll_disable);