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
20static struct kmem_cache *async_pf_cache;
21
22int kvm_async_pf_init(void)
23{
24	async_pf_cache = KMEM_CACHE(kvm_async_pf, 0);
25
26	if (!async_pf_cache)
27		return -ENOMEM;
28
29	return 0;
30}
31
32void kvm_async_pf_deinit(void)
33{
34	kmem_cache_destroy(async_pf_cache);
35	async_pf_cache = NULL;
36}
37
38void kvm_async_pf_vcpu_init(struct kvm_vcpu *vcpu)
39{
40	INIT_LIST_HEAD(&vcpu->async_pf.done);
41	INIT_LIST_HEAD(&vcpu->async_pf.queue);
42	spin_lock_init(&vcpu->async_pf.lock);
43}
44
45static void async_pf_execute(struct work_struct *work)
46{
47	struct kvm_async_pf *apf =
48		container_of(work, struct kvm_async_pf, work);
49	struct kvm_vcpu *vcpu = apf->vcpu;
50	struct mm_struct *mm = vcpu->kvm->mm;
51	unsigned long addr = apf->addr;
52	gpa_t cr2_or_gpa = apf->cr2_or_gpa;
53	int locked = 1;
54	bool first;
55
56	might_sleep();
57
58	/*
59	 * Attempt to pin the VM's host address space, and simply skip gup() if
60	 * acquiring a pin fail, i.e. if the process is exiting.  Note, KVM
61	 * holds a reference to its associated mm_struct until the very end of
62	 * kvm_destroy_vm(), i.e. the struct itself won't be freed before this
63	 * work item is fully processed.
64	 */
65	if (mmget_not_zero(mm)) {
66		mmap_read_lock(mm);
67		get_user_pages_remote(mm, addr, 1, FOLL_WRITE, NULL, &locked);
68		if (locked)
69			mmap_read_unlock(mm);
70		mmput(mm);
71	}
72
73	/*
74	 * Notify and kick the vCPU even if faulting in the page failed, e.g.
75	 * so that the vCPU can retry the fault synchronously.
76	 */
77	if (IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC))
78		kvm_arch_async_page_present(vcpu, apf);
79
80	spin_lock(&vcpu->async_pf.lock);
81	first = list_empty(&vcpu->async_pf.done);
82	list_add_tail(&apf->link, &vcpu->async_pf.done);
83	apf->vcpu = NULL;
84	spin_unlock(&vcpu->async_pf.lock);
85
86	/*
87	 * The apf struct may be freed by kvm_check_async_pf_completion() as
88	 * soon as the lock is dropped.  Nullify it to prevent improper usage.
89	 */
90	apf = NULL;
91
92	if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first)
93		kvm_arch_async_page_present_queued(vcpu);
94
95	trace_kvm_async_pf_completed(addr, cr2_or_gpa);
96
97	__kvm_vcpu_wake_up(vcpu);
98}
99
100static void kvm_flush_and_free_async_pf_work(struct kvm_async_pf *work)
101{
102	/*
103	 * The async #PF is "done", but KVM must wait for the work item itself,
104	 * i.e. async_pf_execute(), to run to completion.  If KVM is a module,
105	 * KVM must ensure *no* code owned by the KVM (the module) can be run
106	 * after the last call to module_put().  Note, flushing the work item
107	 * is always required when the item is taken off the completion queue.
108	 * E.g. even if the vCPU handles the item in the "normal" path, the VM
109	 * could be terminated before async_pf_execute() completes.
110	 *
111	 * Wake all events skip the queue and go straight done, i.e. don't
112	 * need to be flushed (but sanity check that the work wasn't queued).
113	 */
114	if (work->wakeup_all)
115		WARN_ON_ONCE(work->work.func);
116	else
117		flush_work(&work->work);
118	kmem_cache_free(async_pf_cache, work);
119}
120
121void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
122{
123	spin_lock(&vcpu->async_pf.lock);
124
125	/* cancel outstanding work queue item */
126	while (!list_empty(&vcpu->async_pf.queue)) {
127		struct kvm_async_pf *work =
128			list_first_entry(&vcpu->async_pf.queue,
129					 typeof(*work), queue);
130		list_del(&work->queue);
131
132		/*
133		 * We know it's present in vcpu->async_pf.done, do
134		 * nothing here.
135		 */
136		if (!work->vcpu)
137			continue;
138
139		spin_unlock(&vcpu->async_pf.lock);
140#ifdef CONFIG_KVM_ASYNC_PF_SYNC
141		flush_work(&work->work);
142#else
143		if (cancel_work_sync(&work->work))
144			kmem_cache_free(async_pf_cache, work);
145#endif
146		spin_lock(&vcpu->async_pf.lock);
147	}
148
149	while (!list_empty(&vcpu->async_pf.done)) {
150		struct kvm_async_pf *work =
151			list_first_entry(&vcpu->async_pf.done,
152					 typeof(*work), link);
153		list_del(&work->link);
154
155		spin_unlock(&vcpu->async_pf.lock);
156		kvm_flush_and_free_async_pf_work(work);
157		spin_lock(&vcpu->async_pf.lock);
158	}
159	spin_unlock(&vcpu->async_pf.lock);
160
161	vcpu->async_pf.queued = 0;
162}
163
164void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
165{
166	struct kvm_async_pf *work;
167
168	while (!list_empty_careful(&vcpu->async_pf.done) &&
169	      kvm_arch_can_dequeue_async_page_present(vcpu)) {
170		spin_lock(&vcpu->async_pf.lock);
171		work = list_first_entry(&vcpu->async_pf.done, typeof(*work),
172					      link);
173		list_del(&work->link);
174		spin_unlock(&vcpu->async_pf.lock);
175
176		kvm_arch_async_page_ready(vcpu, work);
177		if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC))
178			kvm_arch_async_page_present(vcpu, work);
179
180		list_del(&work->queue);
181		vcpu->async_pf.queued--;
182		kvm_flush_and_free_async_pf_work(work);
183	}
184}
185
186/*
187 * Try to schedule a job to handle page fault asynchronously. Returns 'true' on
188 * success, 'false' on failure (page fault has to be handled synchronously).
189 */
190bool kvm_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
191			unsigned long hva, struct kvm_arch_async_pf *arch)
192{
193	struct kvm_async_pf *work;
194
195	if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU)
196		return false;
197
198	/* Arch specific code should not do async PF in this case */
199	if (unlikely(kvm_is_error_hva(hva)))
200		return false;
201
202	/*
203	 * do alloc nowait since if we are going to sleep anyway we
204	 * may as well sleep faulting in page
205	 */
206	work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT | __GFP_NOWARN);
207	if (!work)
208		return false;
209
210	work->wakeup_all = false;
211	work->vcpu = vcpu;
212	work->cr2_or_gpa = cr2_or_gpa;
213	work->addr = hva;
214	work->arch = *arch;
215
216	INIT_WORK(&work->work, async_pf_execute);
217
218	list_add_tail(&work->queue, &vcpu->async_pf.queue);
219	vcpu->async_pf.queued++;
220	work->notpresent_injected = kvm_arch_async_page_not_present(vcpu, work);
221
222	schedule_work(&work->work);
223
224	return true;
225}
226
227int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu)
228{
229	struct kvm_async_pf *work;
230	bool first;
231
232	if (!list_empty_careful(&vcpu->async_pf.done))
233		return 0;
234
235	work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC);
236	if (!work)
237		return -ENOMEM;
238
239	work->wakeup_all = true;
240	INIT_LIST_HEAD(&work->queue); /* for list_del to work */
241
242	spin_lock(&vcpu->async_pf.lock);
243	first = list_empty(&vcpu->async_pf.done);
244	list_add_tail(&work->link, &vcpu->async_pf.done);
245	spin_unlock(&vcpu->async_pf.lock);
246
247	if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first)
248		kvm_arch_async_page_present_queued(vcpu);
249
250	vcpu->async_pf.queued++;
251	return 0;
252}
253