2 * KVM paravirt_ops implementation
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.
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.
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.
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>
23 #include <linux/context_tracking.h>
24 #include <linux/init.h>
25 #include <linux/kernel.h>
26 #include <linux/kvm_para.h>
27 #include <linux/cpu.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 <linux/swait.h>
40 #include <asm/timer.h>
42 #include <asm/traps.h>
44 #include <asm/tlbflush.h>
46 #include <asm/apicdef.h>
47 #include <asm/hypervisor.h>
48 #include <asm/kvm_guest.h>
50 static int kvmapf = 1;
52 static int __init parse_no_kvmapf(char *arg)
58 early_param("no-kvmapf", parse_no_kvmapf);
60 static int steal_acc = 1;
61 static int __init parse_no_stealacc(char *arg)
67 early_param("no-steal-acc", parse_no_stealacc);
69 static int kvmclock_vsyscall = 1;
70 static int __init parse_no_kvmclock_vsyscall(char *arg)
72 kvmclock_vsyscall = 0;
76 early_param("no-kvmclock-vsyscall", parse_no_kvmclock_vsyscall);
78 static DEFINE_PER_CPU_DECRYPTED(struct kvm_vcpu_pv_apf_data, apf_reason) __aligned(64);
79 static DEFINE_PER_CPU_DECRYPTED(struct kvm_steal_time, steal_time) __aligned(64);
80 static int has_steal_clock = 0;
83 * No need for any "IO delay" on KVM
85 static void kvm_io_delay(void)
89 #define KVM_TASK_SLEEP_HASHBITS 8
90 #define KVM_TASK_SLEEP_HASHSIZE (1<<KVM_TASK_SLEEP_HASHBITS)
92 struct kvm_task_sleep_node {
93 struct hlist_node link;
94 struct swait_queue_head wq;
100 static struct kvm_task_sleep_head {
102 struct hlist_head list;
103 } async_pf_sleepers[KVM_TASK_SLEEP_HASHSIZE];
105 static struct kvm_task_sleep_node *_find_apf_task(struct kvm_task_sleep_head *b,
108 struct hlist_node *p;
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)
121 * @interrupt_kernel: Is this called from a routine which interrupts the kernel
122 * (other than user space)?
124 void kvm_async_pf_task_wait(u32 token, int interrupt_kernel)
126 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
127 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
128 struct kvm_task_sleep_node n, *e;
129 DECLARE_SWAITQUEUE(wait);
133 raw_spin_lock(&b->lock);
134 e = _find_apf_task(b, token);
136 /* dummy entry exist -> wake up was delivered ahead of PF */
139 raw_spin_unlock(&b->lock);
146 n.cpu = smp_processor_id();
147 n.halted = is_idle_task(current) ||
148 (IS_ENABLED(CONFIG_PREEMPT_COUNT)
149 ? preempt_count() > 1 || rcu_preempt_depth()
151 init_swait_queue_head(&n.wq);
152 hlist_add_head(&n.link, &b->list);
153 raw_spin_unlock(&b->lock);
157 prepare_to_swait(&n.wq, &wait, TASK_UNINTERRUPTIBLE);
158 if (hlist_unhashed(&n.link))
169 * We cannot reschedule. So halt.
178 finish_swait(&n.wq, &wait);
183 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wait);
185 static void apf_task_wake_one(struct kvm_task_sleep_node *n)
187 hlist_del_init(&n->link);
189 smp_send_reschedule(n->cpu);
190 else if (swq_has_sleeper(&n->wq))
194 static void apf_task_wake_all(void)
198 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++) {
199 struct hlist_node *p, *next;
200 struct kvm_task_sleep_head *b = &async_pf_sleepers[i];
201 raw_spin_lock(&b->lock);
202 hlist_for_each_safe(p, next, &b->list) {
203 struct kvm_task_sleep_node *n =
204 hlist_entry(p, typeof(*n), link);
205 if (n->cpu == smp_processor_id())
206 apf_task_wake_one(n);
208 raw_spin_unlock(&b->lock);
212 void kvm_async_pf_task_wake(u32 token)
214 u32 key = hash_32(token, KVM_TASK_SLEEP_HASHBITS);
215 struct kvm_task_sleep_head *b = &async_pf_sleepers[key];
216 struct kvm_task_sleep_node *n;
224 raw_spin_lock(&b->lock);
225 n = _find_apf_task(b, token);
228 * async PF was not yet handled.
229 * Add dummy entry for the token.
231 n = kzalloc(sizeof(*n), GFP_ATOMIC);
234 * Allocation failed! Busy wait while other cpu
237 raw_spin_unlock(&b->lock);
242 n->cpu = smp_processor_id();
243 init_swait_queue_head(&n->wq);
244 hlist_add_head(&n->link, &b->list);
246 apf_task_wake_one(n);
247 raw_spin_unlock(&b->lock);
250 EXPORT_SYMBOL_GPL(kvm_async_pf_task_wake);
252 u32 kvm_read_and_reset_pf_reason(void)
256 if (__this_cpu_read(apf_reason.enabled)) {
257 reason = __this_cpu_read(apf_reason.reason);
258 __this_cpu_write(apf_reason.reason, 0);
263 EXPORT_SYMBOL_GPL(kvm_read_and_reset_pf_reason);
264 NOKPROBE_SYMBOL(kvm_read_and_reset_pf_reason);
267 do_async_page_fault(struct pt_regs *regs, unsigned long error_code)
269 enum ctx_state prev_state;
271 switch (kvm_read_and_reset_pf_reason()) {
273 do_page_fault(regs, error_code);
275 case KVM_PV_REASON_PAGE_NOT_PRESENT:
276 /* page is swapped out by the host. */
277 prev_state = exception_enter();
278 kvm_async_pf_task_wait((u32)read_cr2(), !user_mode(regs));
279 exception_exit(prev_state);
281 case KVM_PV_REASON_PAGE_READY:
283 kvm_async_pf_task_wake((u32)read_cr2());
288 NOKPROBE_SYMBOL(do_async_page_fault);
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_cpu_ops.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("kvm-stealtime: cpu %d, msr %llx\n",
312 cpu, (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_available())
336 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF) && kvmapf) {
337 u64 pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
339 #ifdef CONFIG_PREEMPT
340 pa |= KVM_ASYNC_PF_SEND_ALWAYS;
342 pa |= KVM_ASYNC_PF_ENABLED;
344 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
345 pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;
347 wrmsrl(MSR_KVM_ASYNC_PF_EN, pa);
348 __this_cpu_write(apf_reason.enabled, 1);
349 printk(KERN_INFO"KVM setup async PF for cpu %d\n",
353 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI)) {
355 /* Size alignment is implied but just to make it explicit. */
356 BUILD_BUG_ON(__alignof__(kvm_apic_eoi) < 4);
357 __this_cpu_write(kvm_apic_eoi, 0);
358 pa = slow_virt_to_phys(this_cpu_ptr(&kvm_apic_eoi))
360 wrmsrl(MSR_KVM_PV_EOI_EN, pa);
364 kvm_register_steal_time();
367 static void kvm_pv_disable_apf(void)
369 if (!__this_cpu_read(apf_reason.enabled))
372 wrmsrl(MSR_KVM_ASYNC_PF_EN, 0);
373 __this_cpu_write(apf_reason.enabled, 0);
375 printk(KERN_INFO"Unregister pv shared memory for cpu %d\n",
379 static void kvm_pv_guest_cpu_reboot(void *unused)
382 * We disable PV EOI before we load a new kernel by kexec,
383 * since MSR_KVM_PV_EOI_EN stores a pointer into old kernel's memory.
384 * New kernel can re-enable when it boots.
386 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
387 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
388 kvm_pv_disable_apf();
389 kvm_disable_steal_time();
392 static int kvm_pv_reboot_notify(struct notifier_block *nb,
393 unsigned long code, void *unused)
395 if (code == SYS_RESTART)
396 on_each_cpu(kvm_pv_guest_cpu_reboot, NULL, 1);
400 static struct notifier_block kvm_pv_reboot_nb = {
401 .notifier_call = kvm_pv_reboot_notify,
404 static u64 kvm_steal_clock(int cpu)
407 struct kvm_steal_time *src;
410 src = &per_cpu(steal_time, cpu);
412 version = src->version;
416 } while ((version & 1) || (version != src->version));
421 void kvm_disable_steal_time(void)
423 if (!has_steal_clock)
426 wrmsr(MSR_KVM_STEAL_TIME, 0, 0);
429 static inline void __set_percpu_decrypted(void *ptr, unsigned long size)
431 early_set_memory_decrypted((unsigned long) ptr, size);
435 * Iterate through all possible CPUs and map the memory region pointed
436 * by apf_reason, steal_time and kvm_apic_eoi as decrypted at once.
438 * Note: we iterate through all possible CPUs to ensure that CPUs
439 * hotplugged will have their per-cpu variable already mapped as
442 static void __init sev_map_percpu_data(void)
449 for_each_possible_cpu(cpu) {
450 __set_percpu_decrypted(&per_cpu(apf_reason, cpu), sizeof(apf_reason));
451 __set_percpu_decrypted(&per_cpu(steal_time, cpu), sizeof(steal_time));
452 __set_percpu_decrypted(&per_cpu(kvm_apic_eoi, cpu), sizeof(kvm_apic_eoi));
457 static void __init kvm_smp_prepare_cpus(unsigned int max_cpus)
459 native_smp_prepare_cpus(max_cpus);
460 if (kvm_para_has_hint(KVM_HINTS_DEDICATED))
461 static_branch_disable(&virt_spin_lock_key);
464 static void __init kvm_smp_prepare_boot_cpu(void)
467 * Map the per-cpu variables as decrypted before kvm_guest_cpu_init()
468 * shares the guest physical address with the hypervisor.
470 sev_map_percpu_data();
472 kvm_guest_cpu_init();
473 native_smp_prepare_boot_cpu();
477 static void kvm_guest_cpu_offline(void)
479 kvm_disable_steal_time();
480 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
481 wrmsrl(MSR_KVM_PV_EOI_EN, 0);
482 kvm_pv_disable_apf();
486 static int kvm_cpu_online(unsigned int cpu)
489 kvm_guest_cpu_init();
494 static int kvm_cpu_down_prepare(unsigned int cpu)
497 kvm_guest_cpu_offline();
503 static void __init kvm_apf_trap_init(void)
505 update_intr_gate(X86_TRAP_PF, async_page_fault);
508 static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
510 static void kvm_flush_tlb_others(const struct cpumask *cpumask,
511 const struct flush_tlb_info *info)
515 struct kvm_steal_time *src;
516 struct cpumask *flushmask = this_cpu_cpumask_var_ptr(__pv_tlb_mask);
518 cpumask_copy(flushmask, cpumask);
520 * We have to call flush only on online vCPUs. And
521 * queue flush_on_enter for pre-empted vCPUs
523 for_each_cpu(cpu, flushmask) {
524 src = &per_cpu(steal_time, cpu);
525 state = READ_ONCE(src->preempted);
526 if ((state & KVM_VCPU_PREEMPTED)) {
527 if (try_cmpxchg(&src->preempted, &state,
528 state | KVM_VCPU_FLUSH_TLB))
529 __cpumask_clear_cpu(cpu, flushmask);
533 native_flush_tlb_others(flushmask, info);
536 static void __init kvm_guest_init(void)
540 if (!kvm_para_available())
543 paravirt_ops_setup();
544 register_reboot_notifier(&kvm_pv_reboot_nb);
545 for (i = 0; i < KVM_TASK_SLEEP_HASHSIZE; i++)
546 raw_spin_lock_init(&async_pf_sleepers[i].lock);
547 if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF))
548 x86_init.irqs.trap_init = kvm_apf_trap_init;
550 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
552 pv_time_ops.steal_clock = kvm_steal_clock;
555 if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
556 !kvm_para_has_hint(KVM_HINTS_DEDICATED) &&
557 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME))
558 pv_mmu_ops.flush_tlb_others = kvm_flush_tlb_others;
560 if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
561 apic_set_eoi_write(kvm_guest_apic_eoi_write);
563 if (kvmclock_vsyscall)
564 kvm_setup_vsyscall_timeinfo();
567 smp_ops.smp_prepare_cpus = kvm_smp_prepare_cpus;
568 smp_ops.smp_prepare_boot_cpu = kvm_smp_prepare_boot_cpu;
569 if (cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, "x86/kvm:online",
570 kvm_cpu_online, kvm_cpu_down_prepare) < 0)
571 pr_err("kvm_guest: Failed to install cpu hotplug callbacks\n");
573 sev_map_percpu_data();
574 kvm_guest_cpu_init();
578 * Hard lockup detection is enabled by default. Disable it, as guests
579 * can get false positives too easily, for example if the host is
582 hardlockup_detector_disable();
585 static noinline uint32_t __kvm_cpuid_base(void)
587 if (boot_cpu_data.cpuid_level < 0)
588 return 0; /* So we don't blow up on old processors */
590 if (boot_cpu_has(X86_FEATURE_HYPERVISOR))
591 return hypervisor_cpuid_base("KVMKVMKVM\0\0\0", 0);
596 static inline uint32_t kvm_cpuid_base(void)
598 static int kvm_cpuid_base = -1;
600 if (kvm_cpuid_base == -1)
601 kvm_cpuid_base = __kvm_cpuid_base();
603 return kvm_cpuid_base;
606 bool kvm_para_available(void)
608 return kvm_cpuid_base() != 0;
610 EXPORT_SYMBOL_GPL(kvm_para_available);
612 unsigned int kvm_arch_para_features(void)
614 return cpuid_eax(kvm_cpuid_base() | KVM_CPUID_FEATURES);
617 unsigned int kvm_arch_para_hints(void)
619 return cpuid_edx(kvm_cpuid_base() | KVM_CPUID_FEATURES);
622 static uint32_t __init kvm_detect(void)
624 return kvm_cpuid_base();
627 const __initconst struct hypervisor_x86 x86_hyper_kvm = {
629 .detect = kvm_detect,
630 .type = X86_HYPER_KVM,
631 .init.guest_late_init = kvm_guest_init,
632 .init.x2apic_available = kvm_para_available,
635 static __init int activate_jump_labels(void)
637 if (has_steal_clock) {
638 static_key_slow_inc(¶virt_steal_enabled);
640 static_key_slow_inc(¶virt_steal_rq_enabled);
645 arch_initcall(activate_jump_labels);
647 static __init int kvm_setup_pv_tlb_flush(void)
651 if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
652 !kvm_para_has_hint(KVM_HINTS_DEDICATED) &&
653 kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
654 for_each_possible_cpu(cpu) {
655 zalloc_cpumask_var_node(per_cpu_ptr(&__pv_tlb_mask, cpu),
656 GFP_KERNEL, cpu_to_node(cpu));
658 pr_info("KVM setup pv remote TLB flush\n");
663 arch_initcall(kvm_setup_pv_tlb_flush);
665 #ifdef CONFIG_PARAVIRT_SPINLOCKS
667 /* Kick a cpu by its apicid. Used to wake up a halted vcpu */
668 static void kvm_kick_cpu(int cpu)
671 unsigned long flags = 0;
673 apicid = per_cpu(x86_cpu_to_apicid, cpu);
674 kvm_hypercall2(KVM_HC_KICK_CPU, flags, apicid);
677 #include <asm/qspinlock.h>
679 static void kvm_wait(u8 *ptr, u8 val)
686 local_irq_save(flags);
688 if (READ_ONCE(*ptr) != val)
692 * halt until it's our turn and kicked. Note that we do safe halt
693 * for irq enabled case to avoid hang when lock info is overwritten
694 * in irq spinlock slowpath and no spurious interrupt occur to save us.
696 if (arch_irqs_disabled_flags(flags))
702 local_irq_restore(flags);
706 __visible bool __kvm_vcpu_is_preempted(long cpu)
708 struct kvm_steal_time *src = &per_cpu(steal_time, cpu);
710 return !!(src->preempted & KVM_VCPU_PREEMPTED);
712 PV_CALLEE_SAVE_REGS_THUNK(__kvm_vcpu_is_preempted);
716 #include <asm/asm-offsets.h>
718 extern bool __raw_callee_save___kvm_vcpu_is_preempted(long);
721 * Hand-optimize version for x86-64 to avoid 8 64-bit register saving and
722 * restoring to/from the stack.
725 ".pushsection .text;"
726 ".global __raw_callee_save___kvm_vcpu_is_preempted;"
727 ".type __raw_callee_save___kvm_vcpu_is_preempted, @function;"
728 "__raw_callee_save___kvm_vcpu_is_preempted:"
729 "movq __per_cpu_offset(,%rdi,8), %rax;"
730 "cmpb $0, " __stringify(KVM_STEAL_TIME_preempted) "+steal_time(%rax);"
738 * Setup pv_lock_ops to exploit KVM_FEATURE_PV_UNHALT if present.
740 void __init kvm_spinlock_init(void)
742 if (!kvm_para_available())
744 /* Does host kernel support KVM_FEATURE_PV_UNHALT? */
745 if (!kvm_para_has_feature(KVM_FEATURE_PV_UNHALT))
748 if (kvm_para_has_hint(KVM_HINTS_DEDICATED))
751 __pv_init_lock_hash();
752 pv_lock_ops.queued_spin_lock_slowpath = __pv_queued_spin_lock_slowpath;
753 pv_lock_ops.queued_spin_unlock = PV_CALLEE_SAVE(__pv_queued_spin_unlock);
754 pv_lock_ops.wait = kvm_wait;
755 pv_lock_ops.kick = kvm_kick_cpu;
757 if (kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
758 pv_lock_ops.vcpu_is_preempted =
759 PV_CALLEE_SAVE(__kvm_vcpu_is_preempted);
763 #endif /* CONFIG_PARAVIRT_SPINLOCKS */