1 // SPDX-License-Identifier: GPL-2.0-only
3 * kvm asynchronous fault support
5 * Copyright 2010 Red Hat, Inc.
8 * Gleb Natapov <gleb@redhat.com>
11 #include <linux/kvm_host.h>
12 #include <linux/slab.h>
13 #include <linux/module.h>
14 #include <linux/mmu_context.h>
15 #include <linux/sched/mm.h>
18 #include <trace/events/kvm.h>
20 static struct kmem_cache *async_pf_cache;
22 int kvm_async_pf_init(void)
24 async_pf_cache = KMEM_CACHE(kvm_async_pf, 0);
32 void kvm_async_pf_deinit(void)
34 kmem_cache_destroy(async_pf_cache);
35 async_pf_cache = NULL;
38 void kvm_async_pf_vcpu_init(struct kvm_vcpu *vcpu)
40 INIT_LIST_HEAD(&vcpu->async_pf.done);
41 INIT_LIST_HEAD(&vcpu->async_pf.queue);
42 spin_lock_init(&vcpu->async_pf.lock);
45 static void async_pf_execute(struct work_struct *work)
47 struct kvm_async_pf *apf =
48 container_of(work, struct kvm_async_pf, work);
49 struct mm_struct *mm = apf->mm;
50 struct kvm_vcpu *vcpu = apf->vcpu;
51 unsigned long addr = apf->addr;
52 gpa_t cr2_or_gpa = apf->cr2_or_gpa;
59 * This work is run asynchronously to the task which owns
60 * mm and might be done in another context, so we must
64 get_user_pages_remote(mm, addr, 1, FOLL_WRITE, NULL, NULL,
69 if (IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC))
70 kvm_arch_async_page_present(vcpu, apf);
72 spin_lock(&vcpu->async_pf.lock);
73 first = list_empty(&vcpu->async_pf.done);
74 list_add_tail(&apf->link, &vcpu->async_pf.done);
76 spin_unlock(&vcpu->async_pf.lock);
78 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first)
79 kvm_arch_async_page_present_queued(vcpu);
82 * apf may be freed by kvm_check_async_pf_completion() after
86 trace_kvm_async_pf_completed(addr, cr2_or_gpa);
88 rcuwait_wake_up(&vcpu->wait);
93 static void kvm_flush_and_free_async_pf_work(struct kvm_async_pf *work)
96 * The async #PF is "done", but KVM must wait for the work item itself,
97 * i.e. async_pf_execute(), to run to completion. If KVM is a module,
98 * KVM must ensure *no* code owned by the KVM (the module) can be run
99 * after the last call to module_put(). Note, flushing the work item
100 * is always required when the item is taken off the completion queue.
101 * E.g. even if the vCPU handles the item in the "normal" path, the VM
102 * could be terminated before async_pf_execute() completes.
104 * Wake all events skip the queue and go straight done, i.e. don't
105 * need to be flushed (but sanity check that the work wasn't queued).
107 if (work->wakeup_all)
108 WARN_ON_ONCE(work->work.func);
110 flush_work(&work->work);
111 kmem_cache_free(async_pf_cache, work);
114 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
116 spin_lock(&vcpu->async_pf.lock);
118 /* cancel outstanding work queue item */
119 while (!list_empty(&vcpu->async_pf.queue)) {
120 struct kvm_async_pf *work =
121 list_first_entry(&vcpu->async_pf.queue,
122 typeof(*work), queue);
123 list_del(&work->queue);
126 * We know it's present in vcpu->async_pf.done, do
132 spin_unlock(&vcpu->async_pf.lock);
133 #ifdef CONFIG_KVM_ASYNC_PF_SYNC
134 flush_work(&work->work);
136 if (cancel_work_sync(&work->work)) {
138 kmem_cache_free(async_pf_cache, work);
141 spin_lock(&vcpu->async_pf.lock);
144 while (!list_empty(&vcpu->async_pf.done)) {
145 struct kvm_async_pf *work =
146 list_first_entry(&vcpu->async_pf.done,
147 typeof(*work), link);
148 list_del(&work->link);
150 spin_unlock(&vcpu->async_pf.lock);
151 kvm_flush_and_free_async_pf_work(work);
152 spin_lock(&vcpu->async_pf.lock);
154 spin_unlock(&vcpu->async_pf.lock);
156 vcpu->async_pf.queued = 0;
159 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
161 struct kvm_async_pf *work;
163 while (!list_empty_careful(&vcpu->async_pf.done) &&
164 kvm_arch_can_dequeue_async_page_present(vcpu)) {
165 spin_lock(&vcpu->async_pf.lock);
166 work = list_first_entry(&vcpu->async_pf.done, typeof(*work),
168 list_del(&work->link);
169 spin_unlock(&vcpu->async_pf.lock);
171 kvm_arch_async_page_ready(vcpu, work);
172 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC))
173 kvm_arch_async_page_present(vcpu, work);
175 list_del(&work->queue);
176 vcpu->async_pf.queued--;
177 kvm_flush_and_free_async_pf_work(work);
182 * Try to schedule a job to handle page fault asynchronously. Returns 'true' on
183 * success, 'false' on failure (page fault has to be handled synchronously).
185 bool kvm_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
186 unsigned long hva, struct kvm_arch_async_pf *arch)
188 struct kvm_async_pf *work;
190 if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU)
193 /* Arch specific code should not do async PF in this case */
194 if (unlikely(kvm_is_error_hva(hva)))
198 * do alloc nowait since if we are going to sleep anyway we
199 * may as well sleep faulting in page
201 work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT | __GFP_NOWARN);
205 work->wakeup_all = false;
207 work->cr2_or_gpa = cr2_or_gpa;
210 work->mm = current->mm;
213 INIT_WORK(&work->work, async_pf_execute);
215 list_add_tail(&work->queue, &vcpu->async_pf.queue);
216 vcpu->async_pf.queued++;
217 work->notpresent_injected = kvm_arch_async_page_not_present(vcpu, work);
219 schedule_work(&work->work);
224 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu)
226 struct kvm_async_pf *work;
229 if (!list_empty_careful(&vcpu->async_pf.done))
232 work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC);
236 work->wakeup_all = true;
237 INIT_LIST_HEAD(&work->queue); /* for list_del to work */
239 spin_lock(&vcpu->async_pf.lock);
240 first = list_empty(&vcpu->async_pf.done);
241 list_add_tail(&work->link, &vcpu->async_pf.done);
242 spin_unlock(&vcpu->async_pf.lock);
244 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first)
245 kvm_arch_async_page_present_queued(vcpu);
247 vcpu->async_pf.queued++;