virt/kvm/async_pf.c

   1 /*
   2  * kvm asynchronous fault support
   3  *
   4  * Copyright 2010 Red Hat, Inc.
   5  *
   6  * Author:
   7  *      Gleb Natapov <gleb@redhat.com>
   8  *
   9  * This file is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License
  11  * as published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with this program; if not, write to the Free Software Foundation,
  20  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
  21  */
  22
  23 #include <linux/kvm_host.h>
  24 #include <linux/slab.h>
  25 #include <linux/module.h>
  26 #include <linux/mmu_context.h>
  27
  28 #include "async_pf.h"
  29 #include <trace/events/kvm.h>
  30
  31 static inline void kvm_async_page_present_sync(struct kvm_vcpu *vcpu,
  32                                                struct kvm_async_pf *work)
  33 {
  34 #ifdef CONFIG_KVM_ASYNC_PF_SYNC
  35         kvm_arch_async_page_present(vcpu, work);
  36 #endif
  37 }
  38 static inline void kvm_async_page_present_async(struct kvm_vcpu *vcpu,
  39                                                 struct kvm_async_pf *work)
  40 {
  41 #ifndef CONFIG_KVM_ASYNC_PF_SYNC
  42         kvm_arch_async_page_present(vcpu, work);
  43 #endif
  44 }
  45
  46 static struct kmem_cache *async_pf_cache;
  47
  48 int kvm_async_pf_init(void)
  49 {
  50         async_pf_cache = KMEM_CACHE(kvm_async_pf, 0);
  51
  52         if (!async_pf_cache)
  53                 return -ENOMEM;
  54
  55         return 0;
  56 }
  57
  58 void kvm_async_pf_deinit(void)
  59 {
  60         kmem_cache_destroy(async_pf_cache);
  61         async_pf_cache = NULL;
  62 }
  63
  64 void kvm_async_pf_vcpu_init(struct kvm_vcpu *vcpu)
  65 {
  66         INIT_LIST_HEAD(&vcpu->async_pf.done);
  67         INIT_LIST_HEAD(&vcpu->async_pf.queue);
  68         spin_lock_init(&vcpu->async_pf.lock);
  69 }
  70
  71 static void async_pf_execute(struct work_struct *work)
  72 {
  73         struct kvm_async_pf *apf =
  74                 container_of(work, struct kvm_async_pf, work);
  75         struct mm_struct *mm = apf->mm;
  76         struct kvm_vcpu *vcpu = apf->vcpu;
  77         unsigned long addr = apf->addr;
  78         gva_t gva = apf->gva;
  79
  80         might_sleep();
  81
  82         /*
  83          * This work is run asynchromously to the task which owns
  84          * mm and might be done in another context, so we must
  85          * use FOLL_REMOTE.
  86          */
  87         __get_user_pages_unlocked(NULL, mm, addr, 1, NULL,
  88                         FOLL_WRITE | FOLL_REMOTE);
  89
  90         kvm_async_page_present_sync(vcpu, apf);
  91
  92         spin_lock(&vcpu->async_pf.lock);
  93         list_add_tail(&apf->link, &vcpu->async_pf.done);
  94         spin_unlock(&vcpu->async_pf.lock);
  95
  96         /*
  97          * apf may be freed by kvm_check_async_pf_completion() after
  98          * this point
  99          */
 100
 101         trace_kvm_async_pf_completed(addr, gva);
 102
 103         /*
 104          * This memory barrier pairs with prepare_to_wait's set_current_state()
 105          */
 106         smp_mb();
 107         if (swait_active(&vcpu->wq))
 108                 swake_up(&vcpu->wq);
 109
 110         mmput(mm);
 111         kvm_put_kvm(vcpu->kvm);
 112 }
 113
 114 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
 115 {
 116         /* cancel outstanding work queue item */
 117         while (!list_empty(&vcpu->async_pf.queue)) {
 118                 struct kvm_async_pf *work =
 119                         list_first_entry(&vcpu->async_pf.queue,
 120                                          typeof(*work), queue);
 121                 list_del(&work->queue);
 122
 123 #ifdef CONFIG_KVM_ASYNC_PF_SYNC
 124                 flush_work(&work->work);
 125 #else
 126                 if (cancel_work_sync(&work->work)) {
 127                         mmput(work->mm);
 128                         kvm_put_kvm(vcpu->kvm); /* == work->vcpu->kvm */
 129                         kmem_cache_free(async_pf_cache, work);
 130                 }
 131 #endif
 132         }
 133
 134         spin_lock(&vcpu->async_pf.lock);
 135         while (!list_empty(&vcpu->async_pf.done)) {
 136                 struct kvm_async_pf *work =
 137                         list_first_entry(&vcpu->async_pf.done,
 138                                          typeof(*work), link);
 139                 list_del(&work->link);
 140                 kmem_cache_free(async_pf_cache, work);
 141         }
 142         spin_unlock(&vcpu->async_pf.lock);
 143
 144         vcpu->async_pf.queued = 0;
 145 }
 146
 147 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
 148 {
 149         struct kvm_async_pf *work;
 150
 151         while (!list_empty_careful(&vcpu->async_pf.done) &&
 152               kvm_arch_can_inject_async_page_present(vcpu)) {
 153                 spin_lock(&vcpu->async_pf.lock);
 154                 work = list_first_entry(&vcpu->async_pf.done, typeof(*work),
 155                                               link);
 156                 list_del(&work->link);
 157                 spin_unlock(&vcpu->async_pf.lock);
 158
 159                 kvm_arch_async_page_ready(vcpu, work);
 160                 kvm_async_page_present_async(vcpu, work);
 161
 162                 list_del(&work->queue);
 163                 vcpu->async_pf.queued--;
 164                 kmem_cache_free(async_pf_cache, work);
 165         }
 166 }
 167
 168 int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, unsigned long hva,
 169                        struct kvm_arch_async_pf *arch)
 170 {
 171         struct kvm_async_pf *work;
 172
 173         if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU)
 174                 return 0;
 175
 176         /* setup delayed work */
 177
 178         /*
 179          * do alloc nowait since if we are going to sleep anyway we
 180          * may as well sleep faulting in page
 181          */
 182         work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT | __GFP_NOWARN);
 183         if (!work)
 184                 return 0;
 185
 186         work->wakeup_all = false;
 187         work->vcpu = vcpu;
 188         work->gva = gva;
 189         work->addr = hva;
 190         work->arch = *arch;
 191         work->mm = current->mm;
 192         atomic_inc(&work->mm->mm_users);
 193         kvm_get_kvm(work->vcpu->kvm);
 194
 195         /* this can't really happen otherwise gfn_to_pfn_async
 196            would succeed */
 197         if (unlikely(kvm_is_error_hva(work->addr)))
 198                 goto retry_sync;
 199
 200         INIT_WORK(&work->work, async_pf_execute);
 201         if (!schedule_work(&work->work))
 202                 goto retry_sync;
 203
 204         list_add_tail(&work->queue, &vcpu->async_pf.queue);
 205         vcpu->async_pf.queued++;
 206         kvm_arch_async_page_not_present(vcpu, work);
 207         return 1;
 208 retry_sync:
 209         kvm_put_kvm(work->vcpu->kvm);
 210         mmput(work->mm);
 211         kmem_cache_free(async_pf_cache, work);
 212         return 0;
 213 }
 214
 215 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu)
 216 {
 217         struct kvm_async_pf *work;
 218
 219         if (!list_empty_careful(&vcpu->async_pf.done))
 220                 return 0;
 221
 222         work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC);
 223         if (!work)
 224                 return -ENOMEM;
 225
 226         work->wakeup_all = true;
 227         INIT_LIST_HEAD(&work->queue); /* for list_del to work */
 228
 229         spin_lock(&vcpu->async_pf.lock);
 230         list_add_tail(&work->link, &vcpu->async_pf.done);
 231         spin_unlock(&vcpu->async_pf.lock);
 232
 233         vcpu->async_pf.queued++;
 234         return 0;
 235 }