xref: /openbmc/linux/virt/kvm/async_pf.c (revision a20eefae)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * kvm asynchronous fault support
4  *
5  * Copyright 2010 Red Hat, Inc.
6  *
7  * Author:
8  *      Gleb Natapov <gleb@redhat.com>
9  */
10 
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>
16 
17 #include "async_pf.h"
18 #include <trace/events/kvm.h>
19 
20 static inline void kvm_async_page_present_sync(struct kvm_vcpu *vcpu,
21 					       struct kvm_async_pf *work)
22 {
23 #ifdef CONFIG_KVM_ASYNC_PF_SYNC
24 	kvm_arch_async_page_present(vcpu, work);
25 #endif
26 }
27 static inline void kvm_async_page_present_async(struct kvm_vcpu *vcpu,
28 						struct kvm_async_pf *work)
29 {
30 #ifndef CONFIG_KVM_ASYNC_PF_SYNC
31 	kvm_arch_async_page_present(vcpu, work);
32 #endif
33 }
34 
35 static struct kmem_cache *async_pf_cache;
36 
37 int kvm_async_pf_init(void)
38 {
39 	async_pf_cache = KMEM_CACHE(kvm_async_pf, 0);
40 
41 	if (!async_pf_cache)
42 		return -ENOMEM;
43 
44 	return 0;
45 }
46 
47 void kvm_async_pf_deinit(void)
48 {
49 	kmem_cache_destroy(async_pf_cache);
50 	async_pf_cache = NULL;
51 }
52 
53 void kvm_async_pf_vcpu_init(struct kvm_vcpu *vcpu)
54 {
55 	INIT_LIST_HEAD(&vcpu->async_pf.done);
56 	INIT_LIST_HEAD(&vcpu->async_pf.queue);
57 	spin_lock_init(&vcpu->async_pf.lock);
58 }
59 
60 static void async_pf_execute(struct work_struct *work)
61 {
62 	struct kvm_async_pf *apf =
63 		container_of(work, struct kvm_async_pf, work);
64 	struct mm_struct *mm = apf->mm;
65 	struct kvm_vcpu *vcpu = apf->vcpu;
66 	unsigned long addr = apf->addr;
67 	gva_t gva = apf->gva;
68 	int locked = 1;
69 
70 	might_sleep();
71 
72 	/*
73 	 * This work is run asynchronously to the task which owns
74 	 * mm and might be done in another context, so we must
75 	 * access remotely.
76 	 */
77 	down_read(&mm->mmap_sem);
78 	get_user_pages_remote(NULL, mm, addr, 1, FOLL_WRITE, NULL, NULL,
79 			&locked);
80 	if (locked)
81 		up_read(&mm->mmap_sem);
82 
83 	kvm_async_page_present_sync(vcpu, apf);
84 
85 	spin_lock(&vcpu->async_pf.lock);
86 	list_add_tail(&apf->link, &vcpu->async_pf.done);
87 	apf->vcpu = NULL;
88 	spin_unlock(&vcpu->async_pf.lock);
89 
90 	/*
91 	 * apf may be freed by kvm_check_async_pf_completion() after
92 	 * this point
93 	 */
94 
95 	trace_kvm_async_pf_completed(addr, gva);
96 
97 	if (swq_has_sleeper(&vcpu->wq))
98 		swake_up_one(&vcpu->wq);
99 
100 	mmput(mm);
101 	kvm_put_kvm(vcpu->kvm);
102 }
103 
104 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
105 {
106 	spin_lock(&vcpu->async_pf.lock);
107 
108 	/* cancel outstanding work queue item */
109 	while (!list_empty(&vcpu->async_pf.queue)) {
110 		struct kvm_async_pf *work =
111 			list_first_entry(&vcpu->async_pf.queue,
112 					 typeof(*work), queue);
113 		list_del(&work->queue);
114 
115 		/*
116 		 * We know it's present in vcpu->async_pf.done, do
117 		 * nothing here.
118 		 */
119 		if (!work->vcpu)
120 			continue;
121 
122 		spin_unlock(&vcpu->async_pf.lock);
123 #ifdef CONFIG_KVM_ASYNC_PF_SYNC
124 		flush_work(&work->work);
125 #else
126 		if (cancel_work_sync(&work->work)) {
127 			mmput(work->mm);
128 			kvm_put_kvm(vcpu->kvm); /* == work->vcpu->kvm */
129 			kmem_cache_free(async_pf_cache, work);
130 		}
131 #endif
132 		spin_lock(&vcpu->async_pf.lock);
133 	}
134 
135 	while (!list_empty(&vcpu->async_pf.done)) {
136 		struct kvm_async_pf *work =
137 			list_first_entry(&vcpu->async_pf.done,
138 					 typeof(*work), link);
139 		list_del(&work->link);
140 		kmem_cache_free(async_pf_cache, work);
141 	}
142 	spin_unlock(&vcpu->async_pf.lock);
143 
144 	vcpu->async_pf.queued = 0;
145 }
146 
147 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
148 {
149 	struct kvm_async_pf *work;
150 
151 	while (!list_empty_careful(&vcpu->async_pf.done) &&
152 	      kvm_arch_can_inject_async_page_present(vcpu)) {
153 		spin_lock(&vcpu->async_pf.lock);
154 		work = list_first_entry(&vcpu->async_pf.done, typeof(*work),
155 					      link);
156 		list_del(&work->link);
157 		spin_unlock(&vcpu->async_pf.lock);
158 
159 		kvm_arch_async_page_ready(vcpu, work);
160 		kvm_async_page_present_async(vcpu, work);
161 
162 		list_del(&work->queue);
163 		vcpu->async_pf.queued--;
164 		kmem_cache_free(async_pf_cache, work);
165 	}
166 }
167 
168 int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, unsigned long hva,
169 		       struct kvm_arch_async_pf *arch)
170 {
171 	struct kvm_async_pf *work;
172 
173 	if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU)
174 		return 0;
175 
176 	/* setup delayed work */
177 
178 	/*
179 	 * do alloc nowait since if we are going to sleep anyway we
180 	 * may as well sleep faulting in page
181 	 */
182 	work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT | __GFP_NOWARN);
183 	if (!work)
184 		return 0;
185 
186 	work->wakeup_all = false;
187 	work->vcpu = vcpu;
188 	work->gva = gva;
189 	work->addr = hva;
190 	work->arch = *arch;
191 	work->mm = current->mm;
192 	mmget(work->mm);
193 	kvm_get_kvm(work->vcpu->kvm);
194 
195 	/* this can't really happen otherwise gfn_to_pfn_async
196 	   would succeed */
197 	if (unlikely(kvm_is_error_hva(work->addr)))
198 		goto retry_sync;
199 
200 	INIT_WORK(&work->work, async_pf_execute);
201 	if (!schedule_work(&work->work))
202 		goto retry_sync;
203 
204 	list_add_tail(&work->queue, &vcpu->async_pf.queue);
205 	vcpu->async_pf.queued++;
206 	kvm_arch_async_page_not_present(vcpu, work);
207 	return 1;
208 retry_sync:
209 	kvm_put_kvm(work->vcpu->kvm);
210 	mmput(work->mm);
211 	kmem_cache_free(async_pf_cache, work);
212 	return 0;
213 }
214 
215 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu)
216 {
217 	struct kvm_async_pf *work;
218 
219 	if (!list_empty_careful(&vcpu->async_pf.done))
220 		return 0;
221 
222 	work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC);
223 	if (!work)
224 		return -ENOMEM;
225 
226 	work->wakeup_all = true;
227 	INIT_LIST_HEAD(&work->queue); /* for list_del to work */
228 
229 	spin_lock(&vcpu->async_pf.lock);
230 	list_add_tail(&work->link, &vcpu->async_pf.done);
231 	spin_unlock(&vcpu->async_pf.lock);
232 
233 	vcpu->async_pf.queued++;
234 	return 0;
235 }
236