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 if (swq_has_sleeper(&vcpu->wq)) 84 swake_up_one(&vcpu->wq); 85 86 mmput(mm); 87 kvm_put_kvm(vcpu->kvm); 88 } 89 90 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu) 91 { 92 spin_lock(&vcpu->async_pf.lock); 93 94 /* cancel outstanding work queue item */ 95 while (!list_empty(&vcpu->async_pf.queue)) { 96 struct kvm_async_pf *work = 97 list_first_entry(&vcpu->async_pf.queue, 98 typeof(*work), queue); 99 list_del(&work->queue); 100 101 /* 102 * We know it's present in vcpu->async_pf.done, do 103 * nothing here. 104 */ 105 if (!work->vcpu) 106 continue; 107 108 spin_unlock(&vcpu->async_pf.lock); 109 #ifdef CONFIG_KVM_ASYNC_PF_SYNC 110 flush_work(&work->work); 111 #else 112 if (cancel_work_sync(&work->work)) { 113 mmput(work->mm); 114 kvm_put_kvm(vcpu->kvm); /* == work->vcpu->kvm */ 115 kmem_cache_free(async_pf_cache, work); 116 } 117 #endif 118 spin_lock(&vcpu->async_pf.lock); 119 } 120 121 while (!list_empty(&vcpu->async_pf.done)) { 122 struct kvm_async_pf *work = 123 list_first_entry(&vcpu->async_pf.done, 124 typeof(*work), link); 125 list_del(&work->link); 126 kmem_cache_free(async_pf_cache, work); 127 } 128 spin_unlock(&vcpu->async_pf.lock); 129 130 vcpu->async_pf.queued = 0; 131 } 132 133 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu) 134 { 135 struct kvm_async_pf *work; 136 137 while (!list_empty_careful(&vcpu->async_pf.done) && 138 kvm_arch_can_inject_async_page_present(vcpu)) { 139 spin_lock(&vcpu->async_pf.lock); 140 work = list_first_entry(&vcpu->async_pf.done, typeof(*work), 141 link); 142 list_del(&work->link); 143 spin_unlock(&vcpu->async_pf.lock); 144 145 kvm_arch_async_page_ready(vcpu, work); 146 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC)) 147 kvm_arch_async_page_present(vcpu, work); 148 149 list_del(&work->queue); 150 vcpu->async_pf.queued--; 151 kmem_cache_free(async_pf_cache, work); 152 } 153 } 154 155 int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, 156 unsigned long hva, struct kvm_arch_async_pf *arch) 157 { 158 struct kvm_async_pf *work; 159 160 if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU) 161 return 0; 162 163 /* setup delayed work */ 164 165 /* 166 * do alloc nowait since if we are going to sleep anyway we 167 * may as well sleep faulting in page 168 */ 169 work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT | __GFP_NOWARN); 170 if (!work) 171 return 0; 172 173 work->wakeup_all = false; 174 work->vcpu = vcpu; 175 work->cr2_or_gpa = cr2_or_gpa; 176 work->addr = hva; 177 work->arch = *arch; 178 work->mm = current->mm; 179 mmget(work->mm); 180 kvm_get_kvm(work->vcpu->kvm); 181 182 /* this can't really happen otherwise gfn_to_pfn_async 183 would succeed */ 184 if (unlikely(kvm_is_error_hva(work->addr))) 185 goto retry_sync; 186 187 INIT_WORK(&work->work, async_pf_execute); 188 if (!schedule_work(&work->work)) 189 goto retry_sync; 190 191 list_add_tail(&work->queue, &vcpu->async_pf.queue); 192 vcpu->async_pf.queued++; 193 kvm_arch_async_page_not_present(vcpu, work); 194 return 1; 195 retry_sync: 196 kvm_put_kvm(work->vcpu->kvm); 197 mmput(work->mm); 198 kmem_cache_free(async_pf_cache, work); 199 return 0; 200 } 201 202 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu) 203 { 204 struct kvm_async_pf *work; 205 206 if (!list_empty_careful(&vcpu->async_pf.done)) 207 return 0; 208 209 work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC); 210 if (!work) 211 return -ENOMEM; 212 213 work->wakeup_all = true; 214 INIT_LIST_HEAD(&work->queue); /* for list_del to work */ 215 216 spin_lock(&vcpu->async_pf.lock); 217 list_add_tail(&work->link, &vcpu->async_pf.done); 218 spin_unlock(&vcpu->async_pf.lock); 219 220 vcpu->async_pf.queued++; 221 return 0; 222 } 223