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 bool first; 55 56 might_sleep(); 57 58 /* 59 * This work is run asynchronously to the task which owns 60 * mm and might be done in another context, so we must 61 * access remotely. 62 */ 63 mmap_read_lock(mm); 64 get_user_pages_remote(NULL, mm, addr, 1, FOLL_WRITE, NULL, NULL, 65 &locked); 66 if (locked) 67 mmap_read_unlock(mm); 68 69 if (IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC)) 70 kvm_arch_async_page_present(vcpu, apf); 71 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); 75 apf->vcpu = NULL; 76 spin_unlock(&vcpu->async_pf.lock); 77 78 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first) 79 kvm_arch_async_page_present_queued(vcpu); 80 81 /* 82 * apf may be freed by kvm_check_async_pf_completion() after 83 * this point 84 */ 85 86 trace_kvm_async_pf_completed(addr, cr2_or_gpa); 87 88 rcuwait_wake_up(&vcpu->wait); 89 90 mmput(mm); 91 kvm_put_kvm(vcpu->kvm); 92 } 93 94 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu) 95 { 96 spin_lock(&vcpu->async_pf.lock); 97 98 /* cancel outstanding work queue item */ 99 while (!list_empty(&vcpu->async_pf.queue)) { 100 struct kvm_async_pf *work = 101 list_first_entry(&vcpu->async_pf.queue, 102 typeof(*work), queue); 103 list_del(&work->queue); 104 105 /* 106 * We know it's present in vcpu->async_pf.done, do 107 * nothing here. 108 */ 109 if (!work->vcpu) 110 continue; 111 112 spin_unlock(&vcpu->async_pf.lock); 113 #ifdef CONFIG_KVM_ASYNC_PF_SYNC 114 flush_work(&work->work); 115 #else 116 if (cancel_work_sync(&work->work)) { 117 mmput(work->mm); 118 kvm_put_kvm(vcpu->kvm); /* == work->vcpu->kvm */ 119 kmem_cache_free(async_pf_cache, work); 120 } 121 #endif 122 spin_lock(&vcpu->async_pf.lock); 123 } 124 125 while (!list_empty(&vcpu->async_pf.done)) { 126 struct kvm_async_pf *work = 127 list_first_entry(&vcpu->async_pf.done, 128 typeof(*work), link); 129 list_del(&work->link); 130 kmem_cache_free(async_pf_cache, work); 131 } 132 spin_unlock(&vcpu->async_pf.lock); 133 134 vcpu->async_pf.queued = 0; 135 } 136 137 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu) 138 { 139 struct kvm_async_pf *work; 140 141 while (!list_empty_careful(&vcpu->async_pf.done) && 142 kvm_arch_can_dequeue_async_page_present(vcpu)) { 143 spin_lock(&vcpu->async_pf.lock); 144 work = list_first_entry(&vcpu->async_pf.done, typeof(*work), 145 link); 146 list_del(&work->link); 147 spin_unlock(&vcpu->async_pf.lock); 148 149 kvm_arch_async_page_ready(vcpu, work); 150 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC)) 151 kvm_arch_async_page_present(vcpu, work); 152 153 list_del(&work->queue); 154 vcpu->async_pf.queued--; 155 kmem_cache_free(async_pf_cache, work); 156 } 157 } 158 159 /* 160 * Try to schedule a job to handle page fault asynchronously. Returns 'true' on 161 * success, 'false' on failure (page fault has to be handled synchronously). 162 */ 163 bool kvm_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, 164 unsigned long hva, struct kvm_arch_async_pf *arch) 165 { 166 struct kvm_async_pf *work; 167 168 if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU) 169 return false; 170 171 /* Arch specific code should not do async PF in this case */ 172 if (unlikely(kvm_is_error_hva(hva))) 173 return false; 174 175 /* 176 * do alloc nowait since if we are going to sleep anyway we 177 * may as well sleep faulting in page 178 */ 179 work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT | __GFP_NOWARN); 180 if (!work) 181 return false; 182 183 work->wakeup_all = false; 184 work->vcpu = vcpu; 185 work->cr2_or_gpa = cr2_or_gpa; 186 work->addr = hva; 187 work->arch = *arch; 188 work->mm = current->mm; 189 mmget(work->mm); 190 kvm_get_kvm(work->vcpu->kvm); 191 192 INIT_WORK(&work->work, async_pf_execute); 193 194 list_add_tail(&work->queue, &vcpu->async_pf.queue); 195 vcpu->async_pf.queued++; 196 work->notpresent_injected = kvm_arch_async_page_not_present(vcpu, work); 197 198 schedule_work(&work->work); 199 200 return true; 201 } 202 203 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu) 204 { 205 struct kvm_async_pf *work; 206 bool first; 207 208 if (!list_empty_careful(&vcpu->async_pf.done)) 209 return 0; 210 211 work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC); 212 if (!work) 213 return -ENOMEM; 214 215 work->wakeup_all = true; 216 INIT_LIST_HEAD(&work->queue); /* for list_del to work */ 217 218 spin_lock(&vcpu->async_pf.lock); 219 first = list_empty(&vcpu->async_pf.done); 220 list_add_tail(&work->link, &vcpu->async_pf.done); 221 spin_unlock(&vcpu->async_pf.lock); 222 223 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first) 224 kvm_arch_async_page_present_queued(vcpu); 225 226 vcpu->async_pf.queued++; 227 return 0; 228 } 229