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 struct kmem_cache *async_pf_cache; 21 22 int 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 32 void kvm_async_pf_deinit(void) 33 { 34 kmem_cache_destroy(async_pf_cache); 35 async_pf_cache = NULL; 36 } 37 38 void 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 45 static 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 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; 53 int locked = 1; 54 55 might_sleep(); 56 57 /* 58 * This work is run asynchronously to the task which owns 59 * mm and might be done in another context, so we must 60 * access remotely. 61 */ 62 down_read(&mm->mmap_sem); 63 get_user_pages_remote(NULL, mm, addr, 1, FOLL_WRITE, NULL, NULL, 64 &locked); 65 if (locked) 66 up_read(&mm->mmap_sem); 67 68 if (IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC)) 69 kvm_arch_async_page_present(vcpu, apf); 70 71 spin_lock(&vcpu->async_pf.lock); 72 list_add_tail(&apf->link, &vcpu->async_pf.done); 73 apf->vcpu = NULL; 74 spin_unlock(&vcpu->async_pf.lock); 75 76 /* 77 * apf may be freed by kvm_check_async_pf_completion() after 78 * this point 79 */ 80 81 trace_kvm_async_pf_completed(addr, cr2_or_gpa); 82 83 rcuwait_wake_up(&vcpu->wait); 84 85 mmput(mm); 86 kvm_put_kvm(vcpu->kvm); 87 } 88 89 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu) 90 { 91 spin_lock(&vcpu->async_pf.lock); 92 93 /* cancel outstanding work queue item */ 94 while (!list_empty(&vcpu->async_pf.queue)) { 95 struct kvm_async_pf *work = 96 list_first_entry(&vcpu->async_pf.queue, 97 typeof(*work), queue); 98 list_del(&work->queue); 99 100 /* 101 * We know it's present in vcpu->async_pf.done, do 102 * nothing here. 103 */ 104 if (!work->vcpu) 105 continue; 106 107 spin_unlock(&vcpu->async_pf.lock); 108 #ifdef CONFIG_KVM_ASYNC_PF_SYNC 109 flush_work(&work->work); 110 #else 111 if (cancel_work_sync(&work->work)) { 112 mmput(work->mm); 113 kvm_put_kvm(vcpu->kvm); /* == work->vcpu->kvm */ 114 kmem_cache_free(async_pf_cache, work); 115 } 116 #endif 117 spin_lock(&vcpu->async_pf.lock); 118 } 119 120 while (!list_empty(&vcpu->async_pf.done)) { 121 struct kvm_async_pf *work = 122 list_first_entry(&vcpu->async_pf.done, 123 typeof(*work), link); 124 list_del(&work->link); 125 kmem_cache_free(async_pf_cache, work); 126 } 127 spin_unlock(&vcpu->async_pf.lock); 128 129 vcpu->async_pf.queued = 0; 130 } 131 132 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu) 133 { 134 struct kvm_async_pf *work; 135 136 while (!list_empty_careful(&vcpu->async_pf.done) && 137 kvm_arch_can_dequeue_async_page_present(vcpu)) { 138 spin_lock(&vcpu->async_pf.lock); 139 work = list_first_entry(&vcpu->async_pf.done, typeof(*work), 140 link); 141 list_del(&work->link); 142 spin_unlock(&vcpu->async_pf.lock); 143 144 kvm_arch_async_page_ready(vcpu, work); 145 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC)) 146 kvm_arch_async_page_present(vcpu, work); 147 148 list_del(&work->queue); 149 vcpu->async_pf.queued--; 150 kmem_cache_free(async_pf_cache, work); 151 } 152 } 153 154 int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, 155 unsigned long hva, struct kvm_arch_async_pf *arch) 156 { 157 struct kvm_async_pf *work; 158 159 if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU) 160 return 0; 161 162 /* setup delayed work */ 163 164 /* 165 * do alloc nowait since if we are going to sleep anyway we 166 * may as well sleep faulting in page 167 */ 168 work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT | __GFP_NOWARN); 169 if (!work) 170 return 0; 171 172 work->wakeup_all = false; 173 work->vcpu = vcpu; 174 work->cr2_or_gpa = cr2_or_gpa; 175 work->addr = hva; 176 work->arch = *arch; 177 work->mm = current->mm; 178 mmget(work->mm); 179 kvm_get_kvm(work->vcpu->kvm); 180 181 /* this can't really happen otherwise gfn_to_pfn_async 182 would succeed */ 183 if (unlikely(kvm_is_error_hva(work->addr))) 184 goto retry_sync; 185 186 INIT_WORK(&work->work, async_pf_execute); 187 if (!schedule_work(&work->work)) 188 goto retry_sync; 189 190 list_add_tail(&work->queue, &vcpu->async_pf.queue); 191 vcpu->async_pf.queued++; 192 kvm_arch_async_page_not_present(vcpu, work); 193 return 1; 194 retry_sync: 195 kvm_put_kvm(work->vcpu->kvm); 196 mmput(work->mm); 197 kmem_cache_free(async_pf_cache, work); 198 return 0; 199 } 200 201 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu) 202 { 203 struct kvm_async_pf *work; 204 205 if (!list_empty_careful(&vcpu->async_pf.done)) 206 return 0; 207 208 work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC); 209 if (!work) 210 return -ENOMEM; 211 212 work->wakeup_all = true; 213 INIT_LIST_HEAD(&work->queue); /* for list_del to work */ 214 215 spin_lock(&vcpu->async_pf.lock); 216 list_add_tail(&work->link, &vcpu->async_pf.done); 217 spin_unlock(&vcpu->async_pf.lock); 218 219 vcpu->async_pf.queued++; 220 return 0; 221 } 222