virt/kvm/async_pf.c

   1 /*
   2  * kvm asynchronous fault support
   3  *
   4  * Copyright 2010 Red Hat, Inc.
   5  *
   6  * Author:
   7  *      Gleb Natapov <[email protected]>
   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 #include <linux/sched/mm.h>
  28
  29 #include "async_pf.h"
  30 #include <trace/events/kvm.h>
  31
  32 static inline void kvm_async_page_present_sync(struct kvm_vcpu *vcpu,
  33                                                struct kvm_async_pf *work)
  34 {
  35 #ifdef CONFIG_KVM_ASYNC_PF_SYNC
  36         kvm_arch_async_page_present(vcpu, work);
  37 #endif
  38 }
  39 static inline void kvm_async_page_present_async(struct kvm_vcpu *vcpu,
  40                                                 struct kvm_async_pf *work)
  41 {
  42 #ifndef CONFIG_KVM_ASYNC_PF_SYNC
  43         kvm_arch_async_page_present(vcpu, work);
  44 #endif
  45 }
  46
  47 static struct kmem_cache *async_pf_cache;
  48
  49 int kvm_async_pf_init(void)
  50 {
  51         async_pf_cache = KMEM_CACHE(kvm_async_pf, 0);
  52
  53         if (!async_pf_cache)
  54                 return -ENOMEM;
  55
  56         return 0;
  57 }
  58
  59 void kvm_async_pf_deinit(void)
  60 {
  61         kmem_cache_destroy(async_pf_cache);
  62         async_pf_cache = NULL;
  63 }
  64
  65 void kvm_async_pf_vcpu_init(struct kvm_vcpu *vcpu)
  66 {
  67         INIT_LIST_HEAD(&vcpu->async_pf.done);
  68         INIT_LIST_HEAD(&vcpu->async_pf.queue);
  69         spin_lock_init(&vcpu->async_pf.lock);
  70 }
  71
  72 static void async_pf_execute(struct work_struct *work)
  73 {
  74         struct kvm_async_pf *apf =
  75                 container_of(work, struct kvm_async_pf, work);
  76         struct mm_struct *mm = apf->mm;
  77         struct kvm_vcpu *vcpu = apf->vcpu;
  78         unsigned long addr = apf->addr;
  79         gva_t gva = apf->gva;
  80         int locked = 1;
  81
  82         might_sleep();
  83
  84         /*
  85          * This work is run asynchromously to the task which owns
  86          * mm and might be done in another context, so we must
  87          * access remotely.
  88          */
  89         down_read(&mm->mmap_sem);
  90         get_user_pages_remote(NULL, mm, addr, 1, FOLL_WRITE, NULL, NULL,
  91                         &locked);
  92         if (locked)
  93                 up_read(&mm->mmap_sem);
  94
  95         kvm_async_page_present_sync(vcpu, apf);
  96
  97         spin_lock(&vcpu->async_pf.lock);
  98         list_add_tail(&apf->link, &vcpu->async_pf.done);
  99         apf->vcpu = NULL;
 100         spin_unlock(&vcpu->async_pf.lock);
 101
 102         /*
 103          * apf may be freed by kvm_check_async_pf_completion() after
 104          * this point
 105          */
 106
 107         trace_kvm_async_pf_completed(addr, gva);
 108
 109         /*
 110          * This memory barrier pairs with prepare_to_wait's set_current_state()
 111          */
 112         smp_mb();
 113         if (swait_active(&vcpu->wq))
 114                 swake_up(&vcpu->wq);
 115
 116         mmput(mm);
 117         kvm_put_kvm(vcpu->kvm);
 118 }
 119
 120 void kvm_clear_async_pf_completion_queue(struct kvm_vcpu *vcpu)
 121 {
 122         spin_lock(&vcpu->async_pf.lock);
 123
 124         /* cancel outstanding work queue item */
 125         while (!list_empty(&vcpu->async_pf.queue)) {
 126                 struct kvm_async_pf *work =
 127                         list_first_entry(&vcpu->async_pf.queue,
 128                                          typeof(*work), queue);
 129                 list_del(&work->queue);
 130
 131                 /*
 132                  * We know it's present in vcpu->async_pf.done, do
 133                  * nothing here.
 134                  */
 135                 if (!work->vcpu)
 136                         continue;
 137
 138                 spin_unlock(&vcpu->async_pf.lock);
 139 #ifdef CONFIG_KVM_ASYNC_PF_SYNC
 140                 flush_work(&work->work);
 141 #else
 142                 if (cancel_work_sync(&work->work)) {
 143                         mmput(work->mm);
 144                         kvm_put_kvm(vcpu->kvm); /* == work->vcpu->kvm */
 145                         kmem_cache_free(async_pf_cache, work);
 146                 }
 147 #endif
 148                 spin_lock(&vcpu->async_pf.lock);
 149         }
 150
 151         while (!list_empty(&vcpu->async_pf.done)) {
 152                 struct kvm_async_pf *work =
 153                         list_first_entry(&vcpu->async_pf.done,
 154                                          typeof(*work), link);
 155                 list_del(&work->link);
 156                 kmem_cache_free(async_pf_cache, work);
 157         }
 158         spin_unlock(&vcpu->async_pf.lock);
 159
 160         vcpu->async_pf.queued = 0;
 161 }
 162
 163 void kvm_check_async_pf_completion(struct kvm_vcpu *vcpu)
 164 {
 165         struct kvm_async_pf *work;
 166
 167         while (!list_empty_careful(&vcpu->async_pf.done) &&
 168               kvm_arch_can_inject_async_page_present(vcpu)) {
 169                 spin_lock(&vcpu->async_pf.lock);
 170                 work = list_first_entry(&vcpu->async_pf.done, typeof(*work),
 171                                               link);
 172                 list_del(&work->link);
 173                 spin_unlock(&vcpu->async_pf.lock);
 174
 175                 kvm_arch_async_page_ready(vcpu, work);
 176                 kvm_async_page_present_async(vcpu, work);
 177
 178                 list_del(&work->queue);
 179                 vcpu->async_pf.queued--;
 180                 kmem_cache_free(async_pf_cache, work);
 181         }
 182 }
 183
 184 int kvm_setup_async_pf(struct kvm_vcpu *vcpu, gva_t gva, unsigned long hva,
 185                        struct kvm_arch_async_pf *arch)
 186 {
 187         struct kvm_async_pf *work;
 188
 189         if (vcpu->async_pf.queued >= ASYNC_PF_PER_VCPU)
 190                 return 0;
 191
 192         /* setup delayed work */
 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 | __GFP_NOWARN);
 199         if (!work)
 200                 return 0;
 201
 202         work->wakeup_all = false;
 203         work->vcpu = vcpu;
 204         work->gva = gva;
 205         work->addr = hva;
 206         work->arch = *arch;
 207         work->mm = current->mm;
 208         mmget(work->mm);
 209         kvm_get_kvm(work->vcpu->kvm);
 210
 211         /* this can't really happen otherwise gfn_to_pfn_async
 212            would succeed */
 213         if (unlikely(kvm_is_error_hva(work->addr)))
 214                 goto retry_sync;
 215
 216         INIT_WORK(&work->work, async_pf_execute);
 217         if (!schedule_work(&work->work))
 218                 goto retry_sync;
 219
 220         list_add_tail(&work->queue, &vcpu->async_pf.queue);
 221         vcpu->async_pf.queued++;
 222         kvm_arch_async_page_not_present(vcpu, work);
 223         return 1;
 224 retry_sync:
 225         kvm_put_kvm(work->vcpu->kvm);
 226         mmput(work->mm);
 227         kmem_cache_free(async_pf_cache, work);
 228         return 0;
 229 }
 230
 231 int kvm_async_pf_wakeup_all(struct kvm_vcpu *vcpu)
 232 {
 233         struct kvm_async_pf *work;
 234
 235         if (!list_empty_careful(&vcpu->async_pf.done))
 236                 return 0;
 237
 238         work = kmem_cache_zalloc(async_pf_cache, GFP_ATOMIC);
 239         if (!work)
 240                 return -ENOMEM;
 241
 242         work->wakeup_all = true;
 243         INIT_LIST_HEAD(&work->queue); /* for list_del to work */
 244
 245         spin_lock(&vcpu->async_pf.lock);
 246         list_add_tail(&work->link, &vcpu->async_pf.done);
 247         spin_unlock(&vcpu->async_pf.lock);
 248
 249         vcpu->async_pf.queued++;
 250         return 0;
 251 }