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_create(KVM_DEV_STRING(async_pf), 25 sizeof(struct kvm_async_pf), 26 __alignof__(struct kvm_async_pf), 27 0, NULL); 28 29 if (!async_pf_cache) 30 return -ENOMEM; 31 32 return 0; 33 } 34 35 void kvm_async_pf_deinit(void) 36 { 37 kmem_cache_destroy(async_pf_cache); 38 async_pf_cache = NULL; 39 } 40 41 void kvm_async_pf_vcpu_init(struct kvm_vcpu *vcpu) 42 { 43 INIT_LIST_HEAD(&vcpu->async_pf.done); 44 INIT_LIST_HEAD(&vcpu->async_pf.queue); 45 spin_lock_init(&vcpu->async_pf.lock); 46 } 47 48 static void async_pf_execute(struct work_struct *work) 49 { 50 struct kvm_async_pf *apf = 51 container_of(work, struct kvm_async_pf, work); 52 struct kvm_vcpu *vcpu = apf->vcpu; 53 struct mm_struct *mm = vcpu->kvm->mm; 54 unsigned long addr = apf->addr; 55 gpa_t cr2_or_gpa = apf->cr2_or_gpa; 56 int locked = 1; 57 bool first; 58 59 might_sleep(); 60 61 /* 62 * Attempt to pin the VM's host address space, and simply skip gup() if 63 * acquiring a pin fail, i.e. if the process is exiting. Note, KVM 64 * holds a reference to its associated mm_struct until the very end of 65 * kvm_destroy_vm(), i.e. the struct itself won't be freed before this 66 * work item is fully processed. 67 */ 68 if (mmget_not_zero(mm)) { 69 mmap_read_lock(mm); 70 get_user_pages_remote(mm, addr, 1, FOLL_WRITE, NULL, &locked); 71 if (locked) 72 mmap_read_unlock(mm); 73 mmput(mm); 74 } 75 76 /* 77 * Notify and kick the vCPU even if faulting in the page failed, e.g. 78 * so that the vCPU can retry the fault synchronously. 79 */ 80 if (IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC)) 81 kvm_arch_async_page_present(vcpu, apf); 82 83 spin_lock(&vcpu->async_pf.lock); 84 first = list_empty(&vcpu->async_pf.done); 85 list_add_tail(&apf->link, &vcpu->async_pf.done); 86 spin_unlock(&vcpu->async_pf.lock); 87 88 /* 89 * The apf struct may be freed by kvm_check_async_pf_completion() as 90 * soon as the lock is dropped. Nullify it to prevent improper usage. 91 */ 92 apf = NULL; 93 94 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first) 95 kvm_arch_async_page_present_queued(vcpu); 96 97 trace_kvm_async_pf_completed(addr, cr2_or_gpa); 98 99 __kvm_vcpu_wake_up(vcpu); 100 } 101 102 static void kvm_flush_and_free_async_pf_work(struct kvm_async_pf *work) 103 { 104 /* 105 * The async #PF is "done", but KVM must wait for the work item itself, 106 * i.e. async_pf_execute(), to run to completion. If KVM is a module, 107 * KVM must ensure *no* code owned by the KVM (the module) can be run 108 * after the last call to module_put(). Note, flushing the work item 109 * is always required when the item is taken off the completion queue. 110 * E.g. even if the vCPU handles the item in the "normal" path, the VM 111 * could be terminated before async_pf_execute() completes. 112 * 113 * Wake all events skip the queue and go straight done, i.e. don't 114 * need to be flushed (but sanity check that the work wasn't queued). 115 */ 116 if (work->wakeup_all) 117 WARN_ON_ONCE(work->work.func); 118 else 119 flush_work(&work->work); 120 kmem_cache_free(async_pf_cache, work); 121 } 122 123 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu) 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 #ifdef CONFIG_KVM_ASYNC_PF_SYNC 133 flush_work(&work->work); 134 #else 135 if (cancel_work_sync(&work->work)) 136 kmem_cache_free(async_pf_cache, work); 137 #endif 138 } 139 140 spin_lock(&vcpu->async_pf.lock); 141 while (!list_empty(&vcpu->async_pf.done)) { 142 struct kvm_async_pf *work = 143 list_first_entry(&vcpu->async_pf.done, 144 typeof(*work), link); 145 list_del(&work->link); 146 147 spin_unlock(&vcpu->async_pf.lock); 148 kvm_flush_and_free_async_pf_work(work); 149 spin_lock(&vcpu->async_pf.lock); 150 } 151 spin_unlock(&vcpu->async_pf.lock); 152 153 vcpu->async_pf.queued = 0; 154 } 155 156 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu) 157 { 158 struct kvm_async_pf *work; 159 160 while (!list_empty_careful(&vcpu->async_pf.done) && 161 kvm_arch_can_dequeue_async_page_present(vcpu)) { 162 spin_lock(&vcpu->async_pf.lock); 163 work = list_first_entry(&vcpu->async_pf.done, typeof(*work), 164 link); 165 list_del(&work->link); 166 spin_unlock(&vcpu->async_pf.lock); 167 168 kvm_arch_async_page_ready(vcpu, work); 169 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC)) 170 kvm_arch_async_page_present(vcpu, work); 171 172 list_del(&work->queue); 173 vcpu->async_pf.queued--; 174 kvm_flush_and_free_async_pf_work(work); 175 } 176 } 177 178 /* 179 * Try to schedule a job to handle page fault asynchronously. Returns 'true' on 180 * success, 'false' on failure (page fault has to be handled synchronously). 181 */ 182 bool kvm_setup_async_pf(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa, 183 unsigned long hva, struct kvm_arch_async_pf *arch) 184 { 185 struct kvm_async_pf *work; 186 187 if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU) 188 return false; 189 190 /* Arch specific code should not do async PF in this case */ 191 if (unlikely(kvm_is_error_hva(hva))) 192 return false; 193 194 /* 195 * do alloc nowait since if we are going to sleep anyway we 196 * may as well sleep faulting in page 197 */ 198 work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT); 199 if (!work) 200 return false; 201 202 work->wakeup_all = false; 203 work->vcpu = vcpu; 204 work->cr2_or_gpa = cr2_or_gpa; 205 work->addr = hva; 206 work->arch = *arch; 207 208 INIT_WORK(&work->work, async_pf_execute); 209 210 list_add_tail(&work->queue, &vcpu->async_pf.queue); 211 vcpu->async_pf.queued++; 212 work->notpresent_injected = kvm_arch_async_page_not_present(vcpu, work); 213 214 schedule_work(&work->work); 215 216 return true; 217 } 218 219 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu) 220 { 221 struct kvm_async_pf *work; 222 bool first; 223 224 if (!list_empty_careful(&vcpu->async_pf.done)) 225 return 0; 226 227 work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC); 228 if (!work) 229 return -ENOMEM; 230 231 work->wakeup_all = true; 232 INIT_LIST_HEAD(&work->queue); /* for list_del to work */ 233 234 spin_lock(&vcpu->async_pf.lock); 235 first = list_empty(&vcpu->async_pf.done); 236 list_add_tail(&work->link, &vcpu->async_pf.done); 237 spin_unlock(&vcpu->async_pf.lock); 238 239 if (!IS_ENABLED(CONFIG_KVM_ASYNC_PF_SYNC) && first) 240 kvm_arch_async_page_present_queued(vcpu); 241 242 vcpu->async_pf.queued++; 243 return 0; 244 } 245