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 struct kmem_cache *async_pf_cache;
  32
  33 int kvm_async_pf_init(void)
  34 {
  35         async_pf_cache = KMEM_CACHE(kvm_async_pf, 0);
  36
  37         if (!async_pf_cache)
  38                 return -ENOMEM;
  39
  40         return 0;
  41 }
  42
  43 void kvm_async_pf_deinit(void)
  44 {
  45         if (async_pf_cache)
  46                 kmem_cache_destroy(async_pf_cache);
  47         async_pf_cache = NULL;
  48 }
  49
  50 void kvm_async_pf_vcpu_init(struct kvm_vcpu *vcpu)
  51 {
  52         INIT_LIST_HEAD(&vcpu->async_pf.done);
  53         INIT_LIST_HEAD(&vcpu->async_pf.queue);
  54         spin_lock_init(&vcpu->async_pf.lock);
  55 }
  56
  57 static void async_pf_execute(struct work_struct *work)
  58 {
  59         struct kvm_async_pf *apf =
  60                 container_of(work, struct kvm_async_pf, work);
  61         struct mm_struct *mm = apf->mm;
  62         struct kvm_vcpu *vcpu = apf->vcpu;
  63         unsigned long addr = apf->addr;
  64         gva_t gva = apf->gva;
  65
  66         might_sleep();
  67
  68         use_mm(mm);
  69         down_read(&mm->mmap_sem);
  70         get_user_pages(current, mm, addr, 1, 1, 0, NULL, NULL);
  71         up_read(&mm->mmap_sem);
  72         unuse_mm(mm);
  73
  74         spin_lock(&vcpu->async_pf.lock);
  75         list_add_tail(&apf->link, &vcpu->async_pf.done);
  76         spin_unlock(&vcpu->async_pf.lock);
  77
  78         /*
  79          * apf may be freed by kvm_check_async_pf_completion() after
  80          * this point
  81          */
  82
  83         trace_kvm_async_pf_completed(addr, gva);
  84
  85         if (waitqueue_active(&vcpu->wq))
  86                 wake_up_interruptible(&vcpu->wq);
  87
  88         mmdrop(mm);
  89         kvm_put_kvm(vcpu->kvm);
  90 }
  91
  92 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
  93 {
  94         /* cancel outstanding work queue item */
  95         while (!list_empty(&vcpu->async_pf.queue)) {
  96                 struct kvm_async_pf *work =
  97                         list_entry(vcpu->async_pf.queue.next,
  98                                    typeof(*work), queue);
  99                 list_del(&work->queue);
 100                 if (cancel_work_sync(&work->work)) {
 101                         mmdrop(work->mm);
 102                         kvm_put_kvm(vcpu->kvm); /* == work->vcpu->kvm */
 103                         kmem_cache_free(async_pf_cache, work);
 104                 }
 105         }
 106
 107         spin_lock(&vcpu->async_pf.lock);
 108         while (!list_empty(&vcpu->async_pf.done)) {
 109                 struct kvm_async_pf *work =
 110                         list_entry(vcpu->async_pf.done.next,
 111                                    typeof(*work), link);
 112                 list_del(&work->link);
 113                 kmem_cache_free(async_pf_cache, work);
 114         }
 115         spin_unlock(&vcpu->async_pf.lock);
 116
 117         vcpu->async_pf.queued = 0;
 118 }
 119
 120 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
 121 {
 122         struct kvm_async_pf *work;
 123
 124         while (!list_empty_careful(&vcpu->async_pf.done) &&
 125               kvm_arch_can_inject_async_page_present(vcpu)) {
 126                 spin_lock(&vcpu->async_pf.lock);
 127                 work = list_first_entry(&vcpu->async_pf.done, typeof(*work),
 128                                               link);
 129                 list_del(&work->link);
 130                 spin_unlock(&vcpu->async_pf.lock);
 131
 132                 kvm_arch_async_page_ready(vcpu, work);
 133                 kvm_arch_async_page_present(vcpu, work);
 134
 135                 list_del(&work->queue);
 136                 vcpu->async_pf.queued--;
 137                 kmem_cache_free(async_pf_cache, work);
 138         }
 139 }
 140
 141 int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, gfn_t gfn,
 142                        struct kvm_arch_async_pf *arch)
 143 {
 144         struct kvm_async_pf *work;
 145
 146         if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU)
 147                 return 0;
 148
 149         /* setup delayed work */
 150
 151         /*
 152          * do alloc nowait since if we are going to sleep anyway we
 153          * may as well sleep faulting in page
 154          */
 155         work = kmem_cache_zalloc(async_pf_cache, GFP_NOWAIT);
 156         if (!work)
 157                 return 0;
 158
 159         work->wakeup_all = false;
 160         work->vcpu = vcpu;
 161         work->gva = gva;
 162         work->addr = gfn_to_hva(vcpu->kvm, gfn);
 163         work->arch = *arch;
 164         work->mm = current->mm;
 165         atomic_inc(&work->mm->mm_count);
 166         kvm_get_kvm(work->vcpu->kvm);
 167
 168         /* this can't really happen otherwise gfn_to_pfn_async
 169            would succeed */
 170         if (unlikely(kvm_is_error_hva(work->addr)))
 171                 goto retry_sync;
 172
 173         INIT_WORK(&work->work, async_pf_execute);
 174         if (!schedule_work(&work->work))
 175                 goto retry_sync;
 176
 177         list_add_tail(&work->queue, &vcpu->async_pf.queue);
 178         vcpu->async_pf.queued++;
 179         kvm_arch_async_page_not_present(vcpu, work);
 180         return 1;
 181 retry_sync:
 182         kvm_put_kvm(work->vcpu->kvm);
 183         mmdrop(work->mm);
 184         kmem_cache_free(async_pf_cache, work);
 185         return 0;
 186 }
 187
 188 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu)
 189 {
 190         struct kvm_async_pf *work;
 191
 192         if (!list_empty_careful(&vcpu->async_pf.done))
 193                 return 0;
 194
 195         work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC);
 196         if (!work)
 197                 return -ENOMEM;
 198
 199         work->wakeup_all = true;
 200         INIT_LIST_HEAD(&work->queue); /* for list_del to work */
 201
 202         spin_lock(&vcpu->async_pf.lock);
 203         list_add_tail(&work->link, &vcpu->async_pf.done);
 204         spin_unlock(&vcpu->async_pf.lock);
 205
 206         vcpu->async_pf.queued++;
 207         return 0;
 208 }