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(mm, addr, 1, FOLL_WRITE, NULL, &locked); 65 if (locked) 66 mmap_read_unlock(mm); 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 first = list_empty(&vcpu->async_pf.done); 73 list_add_tail(&apf->link, &vcpu->async_pf.done); 74 apf->vcpu = NULL; 75 spin_unlock(&vcpu->async_pf.lock); 76 77 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first) 78 kvm_arch_async_page_present_queued(vcpu); 79 80 /* 81 * apf may be freed by kvm_check_async_pf_completion() after 82 * this point 83 */ 84 85 trace_kvm_async_pf_completed(addr, cr2_or_gpa); 86 87 __kvm_vcpu_wake_up(vcpu); 88 89 mmput(mm); 90 kvm_put_kvm(vcpu->kvm); 91 } 92 93 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu) 94 { 95 spin_lock(&vcpu->async_pf.lock); 96 97 /* cancel outstanding work queue item */ 98 while (!list_empty(&vcpu->async_pf.queue)) { 99 struct kvm_async_pf *work = 100 list_first_entry(&vcpu->async_pf.queue, 101 typeof(*work), queue); 102 list_del(&work->queue); 103 104 /* 105 * We know it's present in vcpu->async_pf.done, do 106 * nothing here. 107 */ 108 if (!work->vcpu) 109 continue; 110 111 spin_unlock(&vcpu->async_pf.lock); 112 #ifdef CONFIG_KVM_ASYNC_PF_SYNC 113 flush_work(&work->work); 114 #else 115 if (cancel_work_sync(&work->work)) { 116 mmput(work->mm); 117 kvm_put_kvm(vcpu->kvm); /* == work->vcpu->kvm */ 118 kmem_cache_free(async_pf_cache, work); 119 } 120 #endif 121 spin_lock(&vcpu->async_pf.lock); 122 } 123 124 while (!list_empty(&vcpu->async_pf.done)) { 125 struct kvm_async_pf *work = 126 list_first_entry(&vcpu->async_pf.done, 127 typeof(*work), link); 128 list_del(&work->link); 129 kmem_cache_free(async_pf_cache, work); 130 } 131 spin_unlock(&vcpu->async_pf.lock); 132 133 vcpu->async_pf.queued = 0; 134 } 135 136 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu) 137 { 138 struct kvm_async_pf *work; 139 140 while (!list_empty_careful(&vcpu->async_pf.done) && 141 kvm_arch_can_dequeue_async_page_present(vcpu)) { 142 spin_lock(&vcpu->async_pf.lock); 143 work = list_first_entry(&vcpu->async_pf.done, typeof(*work), 144 link); 145 list_del(&work->link); 146 spin_unlock(&vcpu->async_pf.lock); 147 148 kvm_arch_async_page_ready(vcpu, work); 149 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC)) 150 kvm_arch_async_page_present(vcpu, work); 151 152 list_del(&work->queue); 153 vcpu->async_pf.queued--; 154 kmem_cache_free(async_pf_cache, work); 155 } 156 } 157 158 /* 159 * Try to schedule a job to handle page fault asynchronously. Returns 'true' on 160 * success, 'false' on failure (page fault has to be handled synchronously). 161 */ 162 bool kvm_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, 163 unsigned long hva, struct kvm_arch_async_pf *arch) 164 { 165 struct kvm_async_pf *work; 166 167 if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU) 168 return false; 169 170 /* Arch specific code should not do async PF in this case */ 171 if (unlikely(kvm_is_error_hva(hva))) 172 return false; 173 174 /* 175 * do alloc nowait since if we are going to sleep anyway we 176 * may as well sleep faulting in page 177 */ 178 work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT | __GFP_NOWARN); 179 if (!work) 180 return false; 181 182 work->wakeup_all = false; 183 work->vcpu = vcpu; 184 work->cr2_or_gpa = cr2_or_gpa; 185 work->addr = hva; 186 work->arch = *arch; 187 work->mm = current->mm; 188 mmget(work->mm); 189 kvm_get_kvm(work->vcpu->kvm); 190 191 INIT_WORK(&work->work, async_pf_execute); 192 193 list_add_tail(&work->queue, &vcpu->async_pf.queue); 194 vcpu->async_pf.queued++; 195 work->notpresent_injected = kvm_arch_async_page_not_present(vcpu, work); 196 197 schedule_work(&work->work); 198 199 return true; 200 } 201 202 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu) 203 { 204 struct kvm_async_pf *work; 205 bool first; 206 207 if (!list_empty_careful(&vcpu->async_pf.done)) 208 return 0; 209 210 work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC); 211 if (!work) 212 return -ENOMEM; 213 214 work->wakeup_all = true; 215 INIT_LIST_HEAD(&work->queue); /* for list_del to work */ 216 217 spin_lock(&vcpu->async_pf.lock); 218 first = list_empty(&vcpu->async_pf.done); 219 list_add_tail(&work->link, &vcpu->async_pf.done); 220 spin_unlock(&vcpu->async_pf.lock); 221 222 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first) 223 kvm_arch_async_page_present_queued(vcpu); 224 225 vcpu->async_pf.queued++; 226 return 0; 227 } 228