drivers/vhost/vhost.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (C) 2009 Red Hat, Inc.
   3  * Copyright (C) 2006 Rusty Russell IBM Corporation
   4  *
   5  * Author: Michael S. Tsirkin <mst@redhat.com>
   6  *
   7  * Inspiration, some code, and most witty comments come from
   8  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   9  *
  10  * Generic code for virtio server in host kernel.
  11  */
  12
  13 #include <linux/eventfd.h>
  14 #include <linux/vhost.h>
  15 #include <linux/uio.h>
  16 #include <linux/mm.h>
  17 #include <linux/miscdevice.h>
  18 #include <linux/mutex.h>
  19 #include <linux/poll.h>
  20 #include <linux/file.h>
  21 #include <linux/highmem.h>
  22 #include <linux/slab.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/kthread.h>
  25 #include <linux/module.h>
  26 #include <linux/sort.h>
  27 #include <linux/sched/mm.h>
  28 #include <linux/sched/signal.h>
  29 #include <linux/sched/vhost_task.h>
  30 #include <linux/interval_tree_generic.h>
  31 #include <linux/nospec.h>
  32 #include <linux/kcov.h>
  33
  34 #include "vhost.h"
  35
  36 static ushort max_mem_regions = 64;
  37 module_param(max_mem_regions, ushort, 0444);
  38 MODULE_PARM_DESC(max_mem_regions,
  39         "Maximum number of memory regions in memory map. (default: 64)");
  40 static int max_iotlb_entries = 2048;
  41 module_param(max_iotlb_entries, int, 0444);
  42 MODULE_PARM_DESC(max_iotlb_entries,
  43         "Maximum number of iotlb entries. (default: 2048)");
  44
  45 enum {
  46         VHOST_MEMORY_F_LOG = 0x1,
  47 };
  48
  49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  53 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  54 {
  55         vq->user_be = !virtio_legacy_is_little_endian();
  56 }
  57
  58 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  59 {
  60         vq->user_be = true;
  61 }
  62
  63 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  64 {
  65         vq->user_be = false;
  66 }
  67
  68 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  69 {
  70         struct vhost_vring_state s;
  71
  72         if (vq->private_data)
  73                 return -EBUSY;
  74
  75         if (copy_from_user(&s, argp, sizeof(s)))
  76                 return -EFAULT;
  77
  78         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  79             s.num != VHOST_VRING_BIG_ENDIAN)
  80                 return -EINVAL;
  81
  82         if (s.num == VHOST_VRING_BIG_ENDIAN)
  83                 vhost_enable_cross_endian_big(vq);
  84         else
  85                 vhost_enable_cross_endian_little(vq);
  86
  87         return 0;
  88 }
  89
  90 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  91                                    int __user *argp)
  92 {
  93         struct vhost_vring_state s = {
  94                 .index = idx,
  95                 .num = vq->user_be
  96         };
  97
  98         if (copy_to_user(argp, &s, sizeof(s)))
  99                 return -EFAULT;
 100
 101         return 0;
 102 }
 103
 104 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 105 {
 106         /* Note for legacy virtio: user_be is initialized at reset time
 107          * according to the host endianness. If userspace does not set an
 108          * explicit endianness, the default behavior is native endian, as
 109          * expected by legacy virtio.
 110          */
 111         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 112 }
 113 #else
 114 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 115 {
 116 }
 117
 118 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 119 {
 120         return -ENOIOCTLCMD;
 121 }
 122
 123 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 124                                    int __user *argp)
 125 {
 126         return -ENOIOCTLCMD;
 127 }
 128
 129 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 130 {
 131         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 132                 || virtio_legacy_is_little_endian();
 133 }
 134 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 135
 136 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 137 {
 138         vhost_init_is_le(vq);
 139 }
 140
 141 struct vhost_flush_struct {
 142         struct vhost_work work;
 143         struct completion wait_event;
 144 };
 145
 146 static void vhost_flush_work(struct vhost_work *work)
 147 {
 148         struct vhost_flush_struct *s;
 149
 150         s = container_of(work, struct vhost_flush_struct, work);
 151         complete(&s->wait_event);
 152 }
 153
 154 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 155                             poll_table *pt)
 156 {
 157         struct vhost_poll *poll;
 158
 159         poll = container_of(pt, struct vhost_poll, table);
 160         poll->wqh = wqh;
 161         add_wait_queue(wqh, &poll->wait);
 162 }
 163
 164 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 165                              void *key)
 166 {
 167         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 168         struct vhost_work *work = &poll->work;
 169
 170         if (!(key_to_poll(key) & poll->mask))
 171                 return 0;
 172
 173         if (!poll->dev->use_worker)
 174                 work->fn(work);
 175         else
 176                 vhost_poll_queue(poll);
 177
 178         return 0;
 179 }
 180
 181 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 182 {
 183         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 184         work->fn = fn;
 185 }
 186 EXPORT_SYMBOL_GPL(vhost_work_init);
 187
 188 /* Init poll structure */
 189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 190                      __poll_t mask, struct vhost_dev *dev)
 191 {
 192         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 193         init_poll_funcptr(&poll->table, vhost_poll_func);
 194         poll->mask = mask;
 195         poll->dev = dev;
 196         poll->wqh = NULL;
 197
 198         vhost_work_init(&poll->work, fn);
 199 }
 200 EXPORT_SYMBOL_GPL(vhost_poll_init);
 201
 202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 203  * keep a reference to a file until after vhost_poll_stop is called. */
 204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 205 {
 206         __poll_t mask;
 207
 208         if (poll->wqh)
 209                 return 0;
 210
 211         mask = vfs_poll(file, &poll->table);
 212         if (mask)
 213                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 214         if (mask & EPOLLERR) {
 215                 vhost_poll_stop(poll);
 216                 return -EINVAL;
 217         }
 218
 219         return 0;
 220 }
 221 EXPORT_SYMBOL_GPL(vhost_poll_start);
 222
 223 /* Stop polling a file. After this function returns, it becomes safe to drop the
 224  * file reference. You must also flush afterwards. */
 225 void vhost_poll_stop(struct vhost_poll *poll)
 226 {
 227         if (poll->wqh) {
 228                 remove_wait_queue(poll->wqh, &poll->wait);
 229                 poll->wqh = NULL;
 230         }
 231 }
 232 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 233
 234 void vhost_dev_flush(struct vhost_dev *dev)
 235 {
 236         struct vhost_flush_struct flush;
 237
 238         if (dev->worker) {
 239                 init_completion(&flush.wait_event);
 240                 vhost_work_init(&flush.work, vhost_flush_work);
 241
 242                 vhost_work_queue(dev, &flush.work);
 243                 wait_for_completion(&flush.wait_event);
 244         }
 245 }
 246 EXPORT_SYMBOL_GPL(vhost_dev_flush);
 247
 248 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 249 {
 250         if (!dev->worker)
 251                 return;
 252
 253         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 254                 /* We can only add the work to the list after we're
 255                  * sure it was not in the list.
 256                  * test_and_set_bit() implies a memory barrier.
 257                  */
 258                 llist_add(&work->node, &dev->worker->work_list);
 259                 vhost_task_wake(dev->worker->vtsk);
 260         }
 261 }
 262 EXPORT_SYMBOL_GPL(vhost_work_queue);
 263
 264 /* A lockless hint for busy polling code to exit the loop */
 265 bool vhost_has_work(struct vhost_dev *dev)
 266 {
 267         return dev->worker && !llist_empty(&dev->worker->work_list);
 268 }
 269 EXPORT_SYMBOL_GPL(vhost_has_work);
 270
 271 void vhost_poll_queue(struct vhost_poll *poll)
 272 {
 273         vhost_work_queue(poll->dev, &poll->work);
 274 }
 275 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 276
 277 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 278 {
 279         int j;
 280
 281         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 282                 vq->meta_iotlb[j] = NULL;
 283 }
 284
 285 static void vhost_vq_meta_reset(struct vhost_dev *d)
 286 {
 287         int i;
 288
 289         for (i = 0; i < d->nvqs; ++i)
 290                 __vhost_vq_meta_reset(d->vqs[i]);
 291 }
 292
 293 static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
 294 {
 295         call_ctx->ctx = NULL;
 296         memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
 297 }
 298
 299 bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
 300 {
 301         return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
 302 }
 303 EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
 304
 305 static void vhost_vq_reset(struct vhost_dev *dev,
 306                            struct vhost_virtqueue *vq)
 307 {
 308         vq->num = 1;
 309         vq->desc = NULL;
 310         vq->avail = NULL;
 311         vq->used = NULL;
 312         vq->last_avail_idx = 0;
 313         vq->avail_idx = 0;
 314         vq->last_used_idx = 0;
 315         vq->signalled_used = 0;
 316         vq->signalled_used_valid = false;
 317         vq->used_flags = 0;
 318         vq->log_used = false;
 319         vq->log_addr = -1ull;
 320         vq->private_data = NULL;
 321         vq->acked_features = 0;
 322         vq->acked_backend_features = 0;
 323         vq->log_base = NULL;
 324         vq->error_ctx = NULL;
 325         vq->kick = NULL;
 326         vq->log_ctx = NULL;
 327         vhost_disable_cross_endian(vq);
 328         vhost_reset_is_le(vq);
 329         vq->busyloop_timeout = 0;
 330         vq->umem = NULL;
 331         vq->iotlb = NULL;
 332         vhost_vring_call_reset(&vq->call_ctx);
 333         __vhost_vq_meta_reset(vq);
 334 }
 335
 336 static bool vhost_worker(void *data)
 337 {
 338         struct vhost_worker *worker = data;
 339         struct vhost_work *work, *work_next;
 340         struct llist_node *node;
 341
 342         node = llist_del_all(&worker->work_list);
 343         if (node) {
 344                 node = llist_reverse_order(node);
 345                 /* make sure flag is seen after deletion */
 346                 smp_wmb();
 347                 llist_for_each_entry_safe(work, work_next, node, node) {
 348                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 349                         kcov_remote_start_common(worker->kcov_handle);
 350                         work->fn(work);
 351                         kcov_remote_stop();
 352                         cond_resched();
 353                 }
 354         }
 355
 356         return !!node;
 357 }
 358
 359 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 360 {
 361         kfree(vq->indirect);
 362         vq->indirect = NULL;
 363         kfree(vq->log);
 364         vq->log = NULL;
 365         kfree(vq->heads);
 366         vq->heads = NULL;
 367 }
 368
 369 /* Helper to allocate iovec buffers for all vqs. */
 370 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 371 {
 372         struct vhost_virtqueue *vq;
 373         int i;
 374
 375         for (i = 0; i < dev->nvqs; ++i) {
 376                 vq = dev->vqs[i];
 377                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 378                                              sizeof(*vq->indirect),
 379                                              GFP_KERNEL);
 380                 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 381                                         GFP_KERNEL);
 382                 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 383                                           GFP_KERNEL);
 384                 if (!vq->indirect || !vq->log || !vq->heads)
 385                         goto err_nomem;
 386         }
 387         return 0;
 388
 389 err_nomem:
 390         for (; i >= 0; --i)
 391                 vhost_vq_free_iovecs(dev->vqs[i]);
 392         return -ENOMEM;
 393 }
 394
 395 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 396 {
 397         int i;
 398
 399         for (i = 0; i < dev->nvqs; ++i)
 400                 vhost_vq_free_iovecs(dev->vqs[i]);
 401 }
 402
 403 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 404                           int pkts, int total_len)
 405 {
 406         struct vhost_dev *dev = vq->dev;
 407
 408         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 409             pkts >= dev->weight) {
 410                 vhost_poll_queue(&vq->poll);
 411                 return true;
 412         }
 413
 414         return false;
 415 }
 416 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 417
 418 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 419                                    unsigned int num)
 420 {
 421         size_t event __maybe_unused =
 422                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 423
 424         return size_add(struct_size(vq->avail, ring, num), event);
 425 }
 426
 427 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 428                                   unsigned int num)
 429 {
 430         size_t event __maybe_unused =
 431                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 432
 433         return size_add(struct_size(vq->used, ring, num), event);
 434 }
 435
 436 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 437                                   unsigned int num)
 438 {
 439         return sizeof(*vq->desc) * num;
 440 }
 441
 442 void vhost_dev_init(struct vhost_dev *dev,
 443                     struct vhost_virtqueue **vqs, int nvqs,
 444                     int iov_limit, int weight, int byte_weight,
 445                     bool use_worker,
 446                     int (*msg_handler)(struct vhost_dev *dev, u32 asid,
 447                                        struct vhost_iotlb_msg *msg))
 448 {
 449         struct vhost_virtqueue *vq;
 450         int i;
 451
 452         dev->vqs = vqs;
 453         dev->nvqs = nvqs;
 454         mutex_init(&dev->mutex);
 455         dev->log_ctx = NULL;
 456         dev->umem = NULL;
 457         dev->iotlb = NULL;
 458         dev->mm = NULL;
 459         dev->worker = NULL;
 460         dev->iov_limit = iov_limit;
 461         dev->weight = weight;
 462         dev->byte_weight = byte_weight;
 463         dev->use_worker = use_worker;
 464         dev->msg_handler = msg_handler;
 465         init_waitqueue_head(&dev->wait);
 466         INIT_LIST_HEAD(&dev->read_list);
 467         INIT_LIST_HEAD(&dev->pending_list);
 468         spin_lock_init(&dev->iotlb_lock);
 469
 470
 471         for (i = 0; i < dev->nvqs; ++i) {
 472                 vq = dev->vqs[i];
 473                 vq->log = NULL;
 474                 vq->indirect = NULL;
 475                 vq->heads = NULL;
 476                 vq->dev = dev;
 477                 mutex_init(&vq->mutex);
 478                 vhost_vq_reset(dev, vq);
 479                 if (vq->handle_kick)
 480                         vhost_poll_init(&vq->poll, vq->handle_kick,
 481                                         EPOLLIN, dev);
 482         }
 483 }
 484 EXPORT_SYMBOL_GPL(vhost_dev_init);
 485
 486 /* Caller should have device mutex */
 487 long vhost_dev_check_owner(struct vhost_dev *dev)
 488 {
 489         /* Are you the owner? If not, I don't think you mean to do that */
 490         return dev->mm == current->mm ? 0 : -EPERM;
 491 }
 492 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 493
 494 /* Caller should have device mutex */
 495 bool vhost_dev_has_owner(struct vhost_dev *dev)
 496 {
 497         return dev->mm;
 498 }
 499 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 500
 501 static void vhost_attach_mm(struct vhost_dev *dev)
 502 {
 503         /* No owner, become one */
 504         if (dev->use_worker) {
 505                 dev->mm = get_task_mm(current);
 506         } else {
 507                 /* vDPA device does not use worker thead, so there's
 508                  * no need to hold the address space for mm. This help
 509                  * to avoid deadlock in the case of mmap() which may
 510                  * held the refcnt of the file and depends on release
 511                  * method to remove vma.
 512                  */
 513                 dev->mm = current->mm;
 514                 mmgrab(dev->mm);
 515         }
 516 }
 517
 518 static void vhost_detach_mm(struct vhost_dev *dev)
 519 {
 520         if (!dev->mm)
 521                 return;
 522
 523         if (dev->use_worker)
 524                 mmput(dev->mm);
 525         else
 526                 mmdrop(dev->mm);
 527
 528         dev->mm = NULL;
 529 }
 530
 531 static void vhost_worker_free(struct vhost_dev *dev)
 532 {
 533         struct vhost_worker *worker = dev->worker;
 534
 535         if (!worker)
 536                 return;
 537
 538         dev->worker = NULL;
 539         WARN_ON(!llist_empty(&worker->work_list));
 540         vhost_task_stop(worker->vtsk);
 541         kfree(worker);
 542 }
 543
 544 static int vhost_worker_create(struct vhost_dev *dev)
 545 {
 546         struct vhost_worker *worker;
 547         struct vhost_task *vtsk;
 548         char name[TASK_COMM_LEN];
 549         int ret;
 550
 551         worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
 552         if (!worker)
 553                 return -ENOMEM;
 554
 555         dev->worker = worker;
 556         worker->kcov_handle = kcov_common_handle();
 557         init_llist_head(&worker->work_list);
 558         snprintf(name, sizeof(name), "vhost-%d", current->pid);
 559
 560         vtsk = vhost_task_create(vhost_worker, worker, name);
 561         if (!vtsk) {
 562                 ret = -ENOMEM;
 563                 goto free_worker;
 564         }
 565
 566         worker->vtsk = vtsk;
 567         vhost_task_start(vtsk);
 568         return 0;
 569
 570 free_worker:
 571         kfree(worker);
 572         dev->worker = NULL;
 573         return ret;
 574 }
 575
 576 /* Caller should have device mutex */
 577 long vhost_dev_set_owner(struct vhost_dev *dev)
 578 {
 579         int err;
 580
 581         /* Is there an owner already? */
 582         if (vhost_dev_has_owner(dev)) {
 583                 err = -EBUSY;
 584                 goto err_mm;
 585         }
 586
 587         vhost_attach_mm(dev);
 588
 589         if (dev->use_worker) {
 590                 err = vhost_worker_create(dev);
 591                 if (err)
 592                         goto err_worker;
 593         }
 594
 595         err = vhost_dev_alloc_iovecs(dev);
 596         if (err)
 597                 goto err_iovecs;
 598
 599         return 0;
 600 err_iovecs:
 601         vhost_worker_free(dev);
 602 err_worker:
 603         vhost_detach_mm(dev);
 604 err_mm:
 605         return err;
 606 }
 607 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 608
 609 static struct vhost_iotlb *iotlb_alloc(void)
 610 {
 611         return vhost_iotlb_alloc(max_iotlb_entries,
 612                                  VHOST_IOTLB_FLAG_RETIRE);
 613 }
 614
 615 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 616 {
 617         return iotlb_alloc();
 618 }
 619 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 620
 621 /* Caller should have device mutex */
 622 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 623 {
 624         int i;
 625
 626         vhost_dev_cleanup(dev);
 627
 628         dev->umem = umem;
 629         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 630          * VQs aren't running.
 631          */
 632         for (i = 0; i < dev->nvqs; ++i)
 633                 dev->vqs[i]->umem = umem;
 634 }
 635 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 636
 637 void vhost_dev_stop(struct vhost_dev *dev)
 638 {
 639         int i;
 640
 641         for (i = 0; i < dev->nvqs; ++i) {
 642                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
 643                         vhost_poll_stop(&dev->vqs[i]->poll);
 644         }
 645
 646         vhost_dev_flush(dev);
 647 }
 648 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 649
 650 void vhost_clear_msg(struct vhost_dev *dev)
 651 {
 652         struct vhost_msg_node *node, *n;
 653
 654         spin_lock(&dev->iotlb_lock);
 655
 656         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 657                 list_del(&node->node);
 658                 kfree(node);
 659         }
 660
 661         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 662                 list_del(&node->node);
 663                 kfree(node);
 664         }
 665
 666         spin_unlock(&dev->iotlb_lock);
 667 }
 668 EXPORT_SYMBOL_GPL(vhost_clear_msg);
 669
 670 void vhost_dev_cleanup(struct vhost_dev *dev)
 671 {
 672         int i;
 673
 674         for (i = 0; i < dev->nvqs; ++i) {
 675                 if (dev->vqs[i]->error_ctx)
 676                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 677                 if (dev->vqs[i]->kick)
 678                         fput(dev->vqs[i]->kick);
 679                 if (dev->vqs[i]->call_ctx.ctx)
 680                         eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
 681                 vhost_vq_reset(dev, dev->vqs[i]);
 682         }
 683         vhost_dev_free_iovecs(dev);
 684         if (dev->log_ctx)
 685                 eventfd_ctx_put(dev->log_ctx);
 686         dev->log_ctx = NULL;
 687         /* No one will access memory at this point */
 688         vhost_iotlb_free(dev->umem);
 689         dev->umem = NULL;
 690         vhost_iotlb_free(dev->iotlb);
 691         dev->iotlb = NULL;
 692         vhost_clear_msg(dev);
 693         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 694         vhost_worker_free(dev);
 695         vhost_detach_mm(dev);
 696 }
 697 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 698
 699 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 700 {
 701         u64 a = addr / VHOST_PAGE_SIZE / 8;
 702
 703         /* Make sure 64 bit math will not overflow. */
 704         if (a > ULONG_MAX - (unsigned long)log_base ||
 705             a + (unsigned long)log_base > ULONG_MAX)
 706                 return false;
 707
 708         return access_ok(log_base + a,
 709                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 710 }
 711
 712 /* Make sure 64 bit math will not overflow. */
 713 static bool vhost_overflow(u64 uaddr, u64 size)
 714 {
 715         if (uaddr > ULONG_MAX || size > ULONG_MAX)
 716                 return true;
 717
 718         if (!size)
 719                 return false;
 720
 721         return uaddr > ULONG_MAX - size + 1;
 722 }
 723
 724 /* Caller should have vq mutex and device mutex. */
 725 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
 726                                 int log_all)
 727 {
 728         struct vhost_iotlb_map *map;
 729
 730         if (!umem)
 731                 return false;
 732
 733         list_for_each_entry(map, &umem->list, link) {
 734                 unsigned long a = map->addr;
 735
 736                 if (vhost_overflow(map->addr, map->size))
 737                         return false;
 738
 739
 740                 if (!access_ok((void __user *)a, map->size))
 741                         return false;
 742                 else if (log_all && !log_access_ok(log_base,
 743                                                    map->start,
 744                                                    map->size))
 745                         return false;
 746         }
 747         return true;
 748 }
 749
 750 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 751                                                u64 addr, unsigned int size,
 752                                                int type)
 753 {
 754         const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
 755
 756         if (!map)
 757                 return NULL;
 758
 759         return (void __user *)(uintptr_t)(map->addr + addr - map->start);
 760 }
 761
 762 /* Can we switch to this memory table? */
 763 /* Caller should have device mutex but not vq mutex */
 764 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
 765                              int log_all)
 766 {
 767         int i;
 768
 769         for (i = 0; i < d->nvqs; ++i) {
 770                 bool ok;
 771                 bool log;
 772
 773                 mutex_lock(&d->vqs[i]->mutex);
 774                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 775                 /* If ring is inactive, will check when it's enabled. */
 776                 if (d->vqs[i]->private_data)
 777                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 778                                                  umem, log);
 779                 else
 780                         ok = true;
 781                 mutex_unlock(&d->vqs[i]->mutex);
 782                 if (!ok)
 783                         return false;
 784         }
 785         return true;
 786 }
 787
 788 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 789                           struct iovec iov[], int iov_size, int access);
 790
 791 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 792                               const void *from, unsigned size)
 793 {
 794         int ret;
 795
 796         if (!vq->iotlb)
 797                 return __copy_to_user(to, from, size);
 798         else {
 799                 /* This function should be called after iotlb
 800                  * prefetch, which means we're sure that all vq
 801                  * could be access through iotlb. So -EAGAIN should
 802                  * not happen in this case.
 803                  */
 804                 struct iov_iter t;
 805                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 806                                      (u64)(uintptr_t)to, size,
 807                                      VHOST_ADDR_USED);
 808
 809                 if (uaddr)
 810                         return __copy_to_user(uaddr, from, size);
 811
 812                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 813                                      ARRAY_SIZE(vq->iotlb_iov),
 814                                      VHOST_ACCESS_WO);
 815                 if (ret < 0)
 816                         goto out;
 817                 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
 818                 ret = copy_to_iter(from, size, &t);
 819                 if (ret == size)
 820                         ret = 0;
 821         }
 822 out:
 823         return ret;
 824 }
 825
 826 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 827                                 void __user *from, unsigned size)
 828 {
 829         int ret;
 830
 831         if (!vq->iotlb)
 832                 return __copy_from_user(to, from, size);
 833         else {
 834                 /* This function should be called after iotlb
 835                  * prefetch, which means we're sure that vq
 836                  * could be access through iotlb. So -EAGAIN should
 837                  * not happen in this case.
 838                  */
 839                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 840                                      (u64)(uintptr_t)from, size,
 841                                      VHOST_ADDR_DESC);
 842                 struct iov_iter f;
 843
 844                 if (uaddr)
 845                         return __copy_from_user(to, uaddr, size);
 846
 847                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 848                                      ARRAY_SIZE(vq->iotlb_iov),
 849                                      VHOST_ACCESS_RO);
 850                 if (ret < 0) {
 851                         vq_err(vq, "IOTLB translation failure: uaddr "
 852                                "%p size 0x%llx\n", from,
 853                                (unsigned long long) size);
 854                         goto out;
 855                 }
 856                 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
 857                 ret = copy_from_iter(to, size, &f);
 858                 if (ret == size)
 859                         ret = 0;
 860         }
 861
 862 out:
 863         return ret;
 864 }
 865
 866 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 867                                           void __user *addr, unsigned int size,
 868                                           int type)
 869 {
 870         int ret;
 871
 872         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 873                              ARRAY_SIZE(vq->iotlb_iov),
 874                              VHOST_ACCESS_RO);
 875         if (ret < 0) {
 876                 vq_err(vq, "IOTLB translation failure: uaddr "
 877                         "%p size 0x%llx\n", addr,
 878                         (unsigned long long) size);
 879                 return NULL;
 880         }
 881
 882         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 883                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 884                         "%p size 0x%llx\n", addr,
 885                         (unsigned long long) size);
 886                 return NULL;
 887         }
 888
 889         return vq->iotlb_iov[0].iov_base;
 890 }
 891
 892 /* This function should be called after iotlb
 893  * prefetch, which means we're sure that vq
 894  * could be access through iotlb. So -EAGAIN should
 895  * not happen in this case.
 896  */
 897 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 898                                             void __user *addr, unsigned int size,
 899                                             int type)
 900 {
 901         void __user *uaddr = vhost_vq_meta_fetch(vq,
 902                              (u64)(uintptr_t)addr, size, type);
 903         if (uaddr)
 904                 return uaddr;
 905
 906         return __vhost_get_user_slow(vq, addr, size, type);
 907 }
 908
 909 #define vhost_put_user(vq, x, ptr)              \
 910 ({ \
 911         int ret; \
 912         if (!vq->iotlb) { \
 913                 ret = __put_user(x, ptr); \
 914         } else { \
 915                 __typeof__(ptr) to = \
 916                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 917                                           sizeof(*ptr), VHOST_ADDR_USED); \
 918                 if (to != NULL) \
 919                         ret = __put_user(x, to); \
 920                 else \
 921                         ret = -EFAULT;  \
 922         } \
 923         ret; \
 924 })
 925
 926 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 927 {
 928         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 929                               vhost_avail_event(vq));
 930 }
 931
 932 static inline int vhost_put_used(struct vhost_virtqueue *vq,
 933                                  struct vring_used_elem *head, int idx,
 934                                  int count)
 935 {
 936         return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 937                                   count * sizeof(*head));
 938 }
 939
 940 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 941
 942 {
 943         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 944                               &vq->used->flags);
 945 }
 946
 947 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 948
 949 {
 950         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 951                               &vq->used->idx);
 952 }
 953
 954 #define vhost_get_user(vq, x, ptr, type)                \
 955 ({ \
 956         int ret; \
 957         if (!vq->iotlb) { \
 958                 ret = __get_user(x, ptr); \
 959         } else { \
 960                 __typeof__(ptr) from = \
 961                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 962                                                            sizeof(*ptr), \
 963                                                            type); \
 964                 if (from != NULL) \
 965                         ret = __get_user(x, from); \
 966                 else \
 967                         ret = -EFAULT; \
 968         } \
 969         ret; \
 970 })
 971
 972 #define vhost_get_avail(vq, x, ptr) \
 973         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 974
 975 #define vhost_get_used(vq, x, ptr) \
 976         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 977
 978 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 979 {
 980         int i = 0;
 981         for (i = 0; i < d->nvqs; ++i)
 982                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 983 }
 984
 985 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 986 {
 987         int i = 0;
 988         for (i = 0; i < d->nvqs; ++i)
 989                 mutex_unlock(&d->vqs[i]->mutex);
 990 }
 991
 992 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
 993                                       __virtio16 *idx)
 994 {
 995         return vhost_get_avail(vq, *idx, &vq->avail->idx);
 996 }
 997
 998 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
 999                                        __virtio16 *head, int idx)
1000 {
1001         return vhost_get_avail(vq, *head,
1002                                &vq->avail->ring[idx & (vq->num - 1)]);
1003 }
1004
1005 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1006                                         __virtio16 *flags)
1007 {
1008         return vhost_get_avail(vq, *flags, &vq->avail->flags);
1009 }
1010
1011 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1012                                        __virtio16 *event)
1013 {
1014         return vhost_get_avail(vq, *event, vhost_used_event(vq));
1015 }
1016
1017 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1018                                      __virtio16 *idx)
1019 {
1020         return vhost_get_used(vq, *idx, &vq->used->idx);
1021 }
1022
1023 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1024                                  struct vring_desc *desc, int idx)
1025 {
1026         return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1027 }
1028
1029 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1030                                   struct vhost_iotlb_msg *msg)
1031 {
1032         struct vhost_msg_node *node, *n;
1033
1034         spin_lock(&d->iotlb_lock);
1035
1036         list_for_each_entry_safe(node, n, &d->pending_list, node) {
1037                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1038                 if (msg->iova <= vq_msg->iova &&
1039                     msg->iova + msg->size - 1 >= vq_msg->iova &&
1040                     vq_msg->type == VHOST_IOTLB_MISS) {
1041                         vhost_poll_queue(&node->vq->poll);
1042                         list_del(&node->node);
1043                         kfree(node);
1044                 }
1045         }
1046
1047         spin_unlock(&d->iotlb_lock);
1048 }
1049
1050 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1051 {
1052         unsigned long a = uaddr;
1053
1054         /* Make sure 64 bit math will not overflow. */
1055         if (vhost_overflow(uaddr, size))
1056                 return false;
1057
1058         if ((access & VHOST_ACCESS_RO) &&
1059             !access_ok((void __user *)a, size))
1060                 return false;
1061         if ((access & VHOST_ACCESS_WO) &&
1062             !access_ok((void __user *)a, size))
1063                 return false;
1064         return true;
1065 }
1066
1067 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1068                                    struct vhost_iotlb_msg *msg)
1069 {
1070         int ret = 0;
1071
1072         if (asid != 0)
1073                 return -EINVAL;
1074
1075         mutex_lock(&dev->mutex);
1076         vhost_dev_lock_vqs(dev);
1077         switch (msg->type) {
1078         case VHOST_IOTLB_UPDATE:
1079                 if (!dev->iotlb) {
1080                         ret = -EFAULT;
1081                         break;
1082                 }
1083                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1084                         ret = -EFAULT;
1085                         break;
1086                 }
1087                 vhost_vq_meta_reset(dev);
1088                 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1089                                           msg->iova + msg->size - 1,
1090                                           msg->uaddr, msg->perm)) {
1091                         ret = -ENOMEM;
1092                         break;
1093                 }
1094                 vhost_iotlb_notify_vq(dev, msg);
1095                 break;
1096         case VHOST_IOTLB_INVALIDATE:
1097                 if (!dev->iotlb) {
1098                         ret = -EFAULT;
1099                         break;
1100                 }
1101                 vhost_vq_meta_reset(dev);
1102                 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1103                                       msg->iova + msg->size - 1);
1104                 break;
1105         default:
1106                 ret = -EINVAL;
1107                 break;
1108         }
1109
1110         vhost_dev_unlock_vqs(dev);
1111         mutex_unlock(&dev->mutex);
1112
1113         return ret;
1114 }
1115 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1116                              struct iov_iter *from)
1117 {
1118         struct vhost_iotlb_msg msg;
1119         size_t offset;
1120         int type, ret;
1121         u32 asid = 0;
1122
1123         ret = copy_from_iter(&type, sizeof(type), from);
1124         if (ret != sizeof(type)) {
1125                 ret = -EINVAL;
1126                 goto done;
1127         }
1128
1129         switch (type) {
1130         case VHOST_IOTLB_MSG:
1131                 /* There maybe a hole after type for V1 message type,
1132                  * so skip it here.
1133                  */
1134                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1135                 break;
1136         case VHOST_IOTLB_MSG_V2:
1137                 if (vhost_backend_has_feature(dev->vqs[0],
1138                                               VHOST_BACKEND_F_IOTLB_ASID)) {
1139                         ret = copy_from_iter(&asid, sizeof(asid), from);
1140                         if (ret != sizeof(asid)) {
1141                                 ret = -EINVAL;
1142                                 goto done;
1143                         }
1144                         offset = 0;
1145                 } else
1146                         offset = sizeof(__u32);
1147                 break;
1148         default:
1149                 ret = -EINVAL;
1150                 goto done;
1151         }
1152
1153         iov_iter_advance(from, offset);
1154         ret = copy_from_iter(&msg, sizeof(msg), from);
1155         if (ret != sizeof(msg)) {
1156                 ret = -EINVAL;
1157                 goto done;
1158         }
1159
1160         if ((msg.type == VHOST_IOTLB_UPDATE ||
1161              msg.type == VHOST_IOTLB_INVALIDATE) &&
1162              msg.size == 0) {
1163                 ret = -EINVAL;
1164                 goto done;
1165         }
1166
1167         if (dev->msg_handler)
1168                 ret = dev->msg_handler(dev, asid, &msg);
1169         else
1170                 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1171         if (ret) {
1172                 ret = -EFAULT;
1173                 goto done;
1174         }
1175
1176         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1177               sizeof(struct vhost_msg_v2);
1178 done:
1179         return ret;
1180 }
1181 EXPORT_SYMBOL(vhost_chr_write_iter);
1182
1183 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1184                             poll_table *wait)
1185 {
1186         __poll_t mask = 0;
1187
1188         poll_wait(file, &dev->wait, wait);
1189
1190         if (!list_empty(&dev->read_list))
1191                 mask |= EPOLLIN | EPOLLRDNORM;
1192
1193         return mask;
1194 }
1195 EXPORT_SYMBOL(vhost_chr_poll);
1196
1197 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1198                             int noblock)
1199 {
1200         DEFINE_WAIT(wait);
1201         struct vhost_msg_node *node;
1202         ssize_t ret = 0;
1203         unsigned size = sizeof(struct vhost_msg);
1204
1205         if (iov_iter_count(to) < size)
1206                 return 0;
1207
1208         while (1) {
1209                 if (!noblock)
1210                         prepare_to_wait(&dev->wait, &wait,
1211                                         TASK_INTERRUPTIBLE);
1212
1213                 node = vhost_dequeue_msg(dev, &dev->read_list);
1214                 if (node)
1215                         break;
1216                 if (noblock) {
1217                         ret = -EAGAIN;
1218                         break;
1219                 }
1220                 if (signal_pending(current)) {
1221                         ret = -ERESTARTSYS;
1222                         break;
1223                 }
1224                 if (!dev->iotlb) {
1225                         ret = -EBADFD;
1226                         break;
1227                 }
1228
1229                 schedule();
1230         }
1231
1232         if (!noblock)
1233                 finish_wait(&dev->wait, &wait);
1234
1235         if (node) {
1236                 struct vhost_iotlb_msg *msg;
1237                 void *start = &node->msg;
1238
1239                 switch (node->msg.type) {
1240                 case VHOST_IOTLB_MSG:
1241                         size = sizeof(node->msg);
1242                         msg = &node->msg.iotlb;
1243                         break;
1244                 case VHOST_IOTLB_MSG_V2:
1245                         size = sizeof(node->msg_v2);
1246                         msg = &node->msg_v2.iotlb;
1247                         break;
1248                 default:
1249                         BUG();
1250                         break;
1251                 }
1252
1253                 ret = copy_to_iter(start, size, to);
1254                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1255                         kfree(node);
1256                         return ret;
1257                 }
1258                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1259         }
1260
1261         return ret;
1262 }
1263 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1264
1265 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1266 {
1267         struct vhost_dev *dev = vq->dev;
1268         struct vhost_msg_node *node;
1269         struct vhost_iotlb_msg *msg;
1270         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1271
1272         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1273         if (!node)
1274                 return -ENOMEM;
1275
1276         if (v2) {
1277                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1278                 msg = &node->msg_v2.iotlb;
1279         } else {
1280                 msg = &node->msg.iotlb;
1281         }
1282
1283         msg->type = VHOST_IOTLB_MISS;
1284         msg->iova = iova;
1285         msg->perm = access;
1286
1287         vhost_enqueue_msg(dev, &dev->read_list, node);
1288
1289         return 0;
1290 }
1291
1292 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1293                          vring_desc_t __user *desc,
1294                          vring_avail_t __user *avail,
1295                          vring_used_t __user *used)
1296
1297 {
1298         /* If an IOTLB device is present, the vring addresses are
1299          * GIOVAs. Access validation occurs at prefetch time. */
1300         if (vq->iotlb)
1301                 return true;
1302
1303         return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1304                access_ok(avail, vhost_get_avail_size(vq, num)) &&
1305                access_ok(used, vhost_get_used_size(vq, num));
1306 }
1307
1308 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1309                                  const struct vhost_iotlb_map *map,
1310                                  int type)
1311 {
1312         int access = (type == VHOST_ADDR_USED) ?
1313                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1314
1315         if (likely(map->perm & access))
1316                 vq->meta_iotlb[type] = map;
1317 }
1318
1319 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1320                             int access, u64 addr, u64 len, int type)
1321 {
1322         const struct vhost_iotlb_map *map;
1323         struct vhost_iotlb *umem = vq->iotlb;
1324         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1325
1326         if (vhost_vq_meta_fetch(vq, addr, len, type))
1327                 return true;
1328
1329         while (len > s) {
1330                 map = vhost_iotlb_itree_first(umem, addr, last);
1331                 if (map == NULL || map->start > addr) {
1332                         vhost_iotlb_miss(vq, addr, access);
1333                         return false;
1334                 } else if (!(map->perm & access)) {
1335                         /* Report the possible access violation by
1336                          * request another translation from userspace.
1337                          */
1338                         return false;
1339                 }
1340
1341                 size = map->size - addr + map->start;
1342
1343                 if (orig_addr == addr && size >= len)
1344                         vhost_vq_meta_update(vq, map, type);
1345
1346                 s += size;
1347                 addr += size;
1348         }
1349
1350         return true;
1351 }
1352
1353 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1354 {
1355         unsigned int num = vq->num;
1356
1357         if (!vq->iotlb)
1358                 return 1;
1359
1360         return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1361                                vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1362                iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1363                                vhost_get_avail_size(vq, num),
1364                                VHOST_ADDR_AVAIL) &&
1365                iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1366                                vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1367 }
1368 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1369
1370 /* Can we log writes? */
1371 /* Caller should have device mutex but not vq mutex */
1372 bool vhost_log_access_ok(struct vhost_dev *dev)
1373 {
1374         return memory_access_ok(dev, dev->umem, 1);
1375 }
1376 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1377
1378 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1379                                   void __user *log_base,
1380                                   bool log_used,
1381                                   u64 log_addr)
1382 {
1383         /* If an IOTLB device is present, log_addr is a GIOVA that
1384          * will never be logged by log_used(). */
1385         if (vq->iotlb)
1386                 return true;
1387
1388         return !log_used || log_access_ok(log_base, log_addr,
1389                                           vhost_get_used_size(vq, vq->num));
1390 }
1391
1392 /* Verify access for write logging. */
1393 /* Caller should have vq mutex and device mutex */
1394 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1395                              void __user *log_base)
1396 {
1397         return vq_memory_access_ok(log_base, vq->umem,
1398                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1399                 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1400 }
1401
1402 /* Can we start vq? */
1403 /* Caller should have vq mutex and device mutex */
1404 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1405 {
1406         if (!vq_log_access_ok(vq, vq->log_base))
1407                 return false;
1408
1409         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1410 }
1411 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1412
1413 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1414 {
1415         struct vhost_memory mem, *newmem;
1416         struct vhost_memory_region *region;
1417         struct vhost_iotlb *newumem, *oldumem;
1418         unsigned long size = offsetof(struct vhost_memory, regions);
1419         int i;
1420
1421         if (copy_from_user(&mem, m, size))
1422                 return -EFAULT;
1423         if (mem.padding)
1424                 return -EOPNOTSUPP;
1425         if (mem.nregions > max_mem_regions)
1426                 return -E2BIG;
1427         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1428                         GFP_KERNEL);
1429         if (!newmem)
1430                 return -ENOMEM;
1431
1432         memcpy(newmem, &mem, size);
1433         if (copy_from_user(newmem->regions, m->regions,
1434                            flex_array_size(newmem, regions, mem.nregions))) {
1435                 kvfree(newmem);
1436                 return -EFAULT;
1437         }
1438
1439         newumem = iotlb_alloc();
1440         if (!newumem) {
1441                 kvfree(newmem);
1442                 return -ENOMEM;
1443         }
1444
1445         for (region = newmem->regions;
1446              region < newmem->regions + mem.nregions;
1447              region++) {
1448                 if (vhost_iotlb_add_range(newumem,
1449                                           region->guest_phys_addr,
1450                                           region->guest_phys_addr +
1451                                           region->memory_size - 1,
1452                                           region->userspace_addr,
1453                                           VHOST_MAP_RW))
1454                         goto err;
1455         }
1456
1457         if (!memory_access_ok(d, newumem, 0))
1458                 goto err;
1459
1460         oldumem = d->umem;
1461         d->umem = newumem;
1462
1463         /* All memory accesses are done under some VQ mutex. */
1464         for (i = 0; i < d->nvqs; ++i) {
1465                 mutex_lock(&d->vqs[i]->mutex);
1466                 d->vqs[i]->umem = newumem;
1467                 mutex_unlock(&d->vqs[i]->mutex);
1468         }
1469
1470         kvfree(newmem);
1471         vhost_iotlb_free(oldumem);
1472         return 0;
1473
1474 err:
1475         vhost_iotlb_free(newumem);
1476         kvfree(newmem);
1477         return -EFAULT;
1478 }
1479
1480 static long vhost_vring_set_num(struct vhost_dev *d,
1481                                 struct vhost_virtqueue *vq,
1482                                 void __user *argp)
1483 {
1484         struct vhost_vring_state s;
1485
1486         /* Resizing ring with an active backend?
1487          * You don't want to do that. */
1488         if (vq->private_data)
1489                 return -EBUSY;
1490
1491         if (copy_from_user(&s, argp, sizeof s))
1492                 return -EFAULT;
1493
1494         if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1495                 return -EINVAL;
1496         vq->num = s.num;
1497
1498         return 0;
1499 }
1500
1501 static long vhost_vring_set_addr(struct vhost_dev *d,
1502                                  struct vhost_virtqueue *vq,
1503                                  void __user *argp)
1504 {
1505         struct vhost_vring_addr a;
1506
1507         if (copy_from_user(&a, argp, sizeof a))
1508                 return -EFAULT;
1509         if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1510                 return -EOPNOTSUPP;
1511
1512         /* For 32bit, verify that the top 32bits of the user
1513            data are set to zero. */
1514         if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1515             (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1516             (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1517                 return -EFAULT;
1518
1519         /* Make sure it's safe to cast pointers to vring types. */
1520         BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1521         BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1522         if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1523             (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1524             (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1525                 return -EINVAL;
1526
1527         /* We only verify access here if backend is configured.
1528          * If it is not, we don't as size might not have been setup.
1529          * We will verify when backend is configured. */
1530         if (vq->private_data) {
1531                 if (!vq_access_ok(vq, vq->num,
1532                         (void __user *)(unsigned long)a.desc_user_addr,
1533                         (void __user *)(unsigned long)a.avail_user_addr,
1534                         (void __user *)(unsigned long)a.used_user_addr))
1535                         return -EINVAL;
1536
1537                 /* Also validate log access for used ring if enabled. */
1538                 if (!vq_log_used_access_ok(vq, vq->log_base,
1539                                 a.flags & (0x1 << VHOST_VRING_F_LOG),
1540                                 a.log_guest_addr))
1541                         return -EINVAL;
1542         }
1543
1544         vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1545         vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1546         vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1547         vq->log_addr = a.log_guest_addr;
1548         vq->used = (void __user *)(unsigned long)a.used_user_addr;
1549
1550         return 0;
1551 }
1552
1553 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1554                                      struct vhost_virtqueue *vq,
1555                                      unsigned int ioctl,
1556                                      void __user *argp)
1557 {
1558         long r;
1559
1560         mutex_lock(&vq->mutex);
1561
1562         switch (ioctl) {
1563         case VHOST_SET_VRING_NUM:
1564                 r = vhost_vring_set_num(d, vq, argp);
1565                 break;
1566         case VHOST_SET_VRING_ADDR:
1567                 r = vhost_vring_set_addr(d, vq, argp);
1568                 break;
1569         default:
1570                 BUG();
1571         }
1572
1573         mutex_unlock(&vq->mutex);
1574
1575         return r;
1576 }
1577 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1578 {
1579         struct file *eventfp, *filep = NULL;
1580         bool pollstart = false, pollstop = false;
1581         struct eventfd_ctx *ctx = NULL;
1582         u32 __user *idxp = argp;
1583         struct vhost_virtqueue *vq;
1584         struct vhost_vring_state s;
1585         struct vhost_vring_file f;
1586         u32 idx;
1587         long r;
1588
1589         r = get_user(idx, idxp);
1590         if (r < 0)
1591                 return r;
1592         if (idx >= d->nvqs)
1593                 return -ENOBUFS;
1594
1595         idx = array_index_nospec(idx, d->nvqs);
1596         vq = d->vqs[idx];
1597
1598         if (ioctl == VHOST_SET_VRING_NUM ||
1599             ioctl == VHOST_SET_VRING_ADDR) {
1600                 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1601         }
1602
1603         mutex_lock(&vq->mutex);
1604
1605         switch (ioctl) {
1606         case VHOST_SET_VRING_BASE:
1607                 /* Moving base with an active backend?
1608                  * You don't want to do that. */
1609                 if (vq->private_data) {
1610                         r = -EBUSY;
1611                         break;
1612                 }
1613                 if (copy_from_user(&s, argp, sizeof s)) {
1614                         r = -EFAULT;
1615                         break;
1616                 }
1617                 if (s.num > 0xffff) {
1618                         r = -EINVAL;
1619                         break;
1620                 }
1621                 vq->last_avail_idx = s.num;
1622                 /* Forget the cached index value. */
1623                 vq->avail_idx = vq->last_avail_idx;
1624                 break;
1625         case VHOST_GET_VRING_BASE:
1626                 s.index = idx;
1627                 s.num = vq->last_avail_idx;
1628                 if (copy_to_user(argp, &s, sizeof s))
1629                         r = -EFAULT;
1630                 break;
1631         case VHOST_SET_VRING_KICK:
1632                 if (copy_from_user(&f, argp, sizeof f)) {
1633                         r = -EFAULT;
1634                         break;
1635                 }
1636                 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1637                 if (IS_ERR(eventfp)) {
1638                         r = PTR_ERR(eventfp);
1639                         break;
1640                 }
1641                 if (eventfp != vq->kick) {
1642                         pollstop = (filep = vq->kick) != NULL;
1643                         pollstart = (vq->kick = eventfp) != NULL;
1644                 } else
1645                         filep = eventfp;
1646                 break;
1647         case VHOST_SET_VRING_CALL:
1648                 if (copy_from_user(&f, argp, sizeof f)) {
1649                         r = -EFAULT;
1650                         break;
1651                 }
1652                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1653                 if (IS_ERR(ctx)) {
1654                         r = PTR_ERR(ctx);
1655                         break;
1656                 }
1657
1658                 swap(ctx, vq->call_ctx.ctx);
1659                 break;
1660         case VHOST_SET_VRING_ERR:
1661                 if (copy_from_user(&f, argp, sizeof f)) {
1662                         r = -EFAULT;
1663                         break;
1664                 }
1665                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1666                 if (IS_ERR(ctx)) {
1667                         r = PTR_ERR(ctx);
1668                         break;
1669                 }
1670                 swap(ctx, vq->error_ctx);
1671                 break;
1672         case VHOST_SET_VRING_ENDIAN:
1673                 r = vhost_set_vring_endian(vq, argp);
1674                 break;
1675         case VHOST_GET_VRING_ENDIAN:
1676                 r = vhost_get_vring_endian(vq, idx, argp);
1677                 break;
1678         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1679                 if (copy_from_user(&s, argp, sizeof(s))) {
1680                         r = -EFAULT;
1681                         break;
1682                 }
1683                 vq->busyloop_timeout = s.num;
1684                 break;
1685         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1686                 s.index = idx;
1687                 s.num = vq->busyloop_timeout;
1688                 if (copy_to_user(argp, &s, sizeof(s)))
1689                         r = -EFAULT;
1690                 break;
1691         default:
1692                 r = -ENOIOCTLCMD;
1693         }
1694
1695         if (pollstop && vq->handle_kick)
1696                 vhost_poll_stop(&vq->poll);
1697
1698         if (!IS_ERR_OR_NULL(ctx))
1699                 eventfd_ctx_put(ctx);
1700         if (filep)
1701                 fput(filep);
1702
1703         if (pollstart && vq->handle_kick)
1704                 r = vhost_poll_start(&vq->poll, vq->kick);
1705
1706         mutex_unlock(&vq->mutex);
1707
1708         if (pollstop && vq->handle_kick)
1709                 vhost_dev_flush(vq->poll.dev);
1710         return r;
1711 }
1712 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1713
1714 int vhost_init_device_iotlb(struct vhost_dev *d)
1715 {
1716         struct vhost_iotlb *niotlb, *oiotlb;
1717         int i;
1718
1719         niotlb = iotlb_alloc();
1720         if (!niotlb)
1721                 return -ENOMEM;
1722
1723         oiotlb = d->iotlb;
1724         d->iotlb = niotlb;
1725
1726         for (i = 0; i < d->nvqs; ++i) {
1727                 struct vhost_virtqueue *vq = d->vqs[i];
1728
1729                 mutex_lock(&vq->mutex);
1730                 vq->iotlb = niotlb;
1731                 __vhost_vq_meta_reset(vq);
1732                 mutex_unlock(&vq->mutex);
1733         }
1734
1735         vhost_iotlb_free(oiotlb);
1736
1737         return 0;
1738 }
1739 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1740
1741 /* Caller must have device mutex */
1742 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1743 {
1744         struct eventfd_ctx *ctx;
1745         u64 p;
1746         long r;
1747         int i, fd;
1748
1749         /* If you are not the owner, you can become one */
1750         if (ioctl == VHOST_SET_OWNER) {
1751                 r = vhost_dev_set_owner(d);
1752                 goto done;
1753         }
1754
1755         /* You must be the owner to do anything else */
1756         r = vhost_dev_check_owner(d);
1757         if (r)
1758                 goto done;
1759
1760         switch (ioctl) {
1761         case VHOST_SET_MEM_TABLE:
1762                 r = vhost_set_memory(d, argp);
1763                 break;
1764         case VHOST_SET_LOG_BASE:
1765                 if (copy_from_user(&p, argp, sizeof p)) {
1766                         r = -EFAULT;
1767                         break;
1768                 }
1769                 if ((u64)(unsigned long)p != p) {
1770                         r = -EFAULT;
1771                         break;
1772                 }
1773                 for (i = 0; i < d->nvqs; ++i) {
1774                         struct vhost_virtqueue *vq;
1775                         void __user *base = (void __user *)(unsigned long)p;
1776                         vq = d->vqs[i];
1777                         mutex_lock(&vq->mutex);
1778                         /* If ring is inactive, will check when it's enabled. */
1779                         if (vq->private_data && !vq_log_access_ok(vq, base))
1780                                 r = -EFAULT;
1781                         else
1782                                 vq->log_base = base;
1783                         mutex_unlock(&vq->mutex);
1784                 }
1785                 break;
1786         case VHOST_SET_LOG_FD:
1787                 r = get_user(fd, (int __user *)argp);
1788                 if (r < 0)
1789                         break;
1790                 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1791                 if (IS_ERR(ctx)) {
1792                         r = PTR_ERR(ctx);
1793                         break;
1794                 }
1795                 swap(ctx, d->log_ctx);
1796                 for (i = 0; i < d->nvqs; ++i) {
1797                         mutex_lock(&d->vqs[i]->mutex);
1798                         d->vqs[i]->log_ctx = d->log_ctx;
1799                         mutex_unlock(&d->vqs[i]->mutex);
1800                 }
1801                 if (ctx)
1802                         eventfd_ctx_put(ctx);
1803                 break;
1804         default:
1805                 r = -ENOIOCTLCMD;
1806                 break;
1807         }
1808 done:
1809         return r;
1810 }
1811 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1812
1813 /* TODO: This is really inefficient.  We need something like get_user()
1814  * (instruction directly accesses the data, with an exception table entry
1815  * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
1816  */
1817 static int set_bit_to_user(int nr, void __user *addr)
1818 {
1819         unsigned long log = (unsigned long)addr;
1820         struct page *page;
1821         void *base;
1822         int bit = nr + (log % PAGE_SIZE) * 8;
1823         int r;
1824
1825         r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1826         if (r < 0)
1827                 return r;
1828         BUG_ON(r != 1);
1829         base = kmap_atomic(page);
1830         set_bit(bit, base);
1831         kunmap_atomic(base);
1832         unpin_user_pages_dirty_lock(&page, 1, true);
1833         return 0;
1834 }
1835
1836 static int log_write(void __user *log_base,
1837                      u64 write_address, u64 write_length)
1838 {
1839         u64 write_page = write_address / VHOST_PAGE_SIZE;
1840         int r;
1841
1842         if (!write_length)
1843                 return 0;
1844         write_length += write_address % VHOST_PAGE_SIZE;
1845         for (;;) {
1846                 u64 base = (u64)(unsigned long)log_base;
1847                 u64 log = base + write_page / 8;
1848                 int bit = write_page % 8;
1849                 if ((u64)(unsigned long)log != log)
1850                         return -EFAULT;
1851                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1852                 if (r < 0)
1853                         return r;
1854                 if (write_length <= VHOST_PAGE_SIZE)
1855                         break;
1856                 write_length -= VHOST_PAGE_SIZE;
1857                 write_page += 1;
1858         }
1859         return r;
1860 }
1861
1862 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1863 {
1864         struct vhost_iotlb *umem = vq->umem;
1865         struct vhost_iotlb_map *u;
1866         u64 start, end, l, min;
1867         int r;
1868         bool hit = false;
1869
1870         while (len) {
1871                 min = len;
1872                 /* More than one GPAs can be mapped into a single HVA. So
1873                  * iterate all possible umems here to be safe.
1874                  */
1875                 list_for_each_entry(u, &umem->list, link) {
1876                         if (u->addr > hva - 1 + len ||
1877                             u->addr - 1 + u->size < hva)
1878                                 continue;
1879                         start = max(u->addr, hva);
1880                         end = min(u->addr - 1 + u->size, hva - 1 + len);
1881                         l = end - start + 1;
1882                         r = log_write(vq->log_base,
1883                                       u->start + start - u->addr,
1884                                       l);
1885                         if (r < 0)
1886                                 return r;
1887                         hit = true;
1888                         min = min(l, min);
1889                 }
1890
1891                 if (!hit)
1892                         return -EFAULT;
1893
1894                 len -= min;
1895                 hva += min;
1896         }
1897
1898         return 0;
1899 }
1900
1901 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1902 {
1903         struct iovec *iov = vq->log_iov;
1904         int i, ret;
1905
1906         if (!vq->iotlb)
1907                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1908
1909         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1910                              len, iov, 64, VHOST_ACCESS_WO);
1911         if (ret < 0)
1912                 return ret;
1913
1914         for (i = 0; i < ret; i++) {
1915                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1916                                     iov[i].iov_len);
1917                 if (ret)
1918                         return ret;
1919         }
1920
1921         return 0;
1922 }
1923
1924 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1925                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1926 {
1927         int i, r;
1928
1929         /* Make sure data written is seen before log. */
1930         smp_wmb();
1931
1932         if (vq->iotlb) {
1933                 for (i = 0; i < count; i++) {
1934                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1935                                           iov[i].iov_len);
1936                         if (r < 0)
1937                                 return r;
1938                 }
1939                 return 0;
1940         }
1941
1942         for (i = 0; i < log_num; ++i) {
1943                 u64 l = min(log[i].len, len);
1944                 r = log_write(vq->log_base, log[i].addr, l);
1945                 if (r < 0)
1946                         return r;
1947                 len -= l;
1948                 if (!len) {
1949                         if (vq->log_ctx)
1950                                 eventfd_signal(vq->log_ctx, 1);
1951                         return 0;
1952                 }
1953         }
1954         /* Length written exceeds what we have stored. This is a bug. */
1955         BUG();
1956         return 0;
1957 }
1958 EXPORT_SYMBOL_GPL(vhost_log_write);
1959
1960 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1961 {
1962         void __user *used;
1963         if (vhost_put_used_flags(vq))
1964                 return -EFAULT;
1965         if (unlikely(vq->log_used)) {
1966                 /* Make sure the flag is seen before log. */
1967                 smp_wmb();
1968                 /* Log used flag write. */
1969                 used = &vq->used->flags;
1970                 log_used(vq, (used - (void __user *)vq->used),
1971                          sizeof vq->used->flags);
1972                 if (vq->log_ctx)
1973                         eventfd_signal(vq->log_ctx, 1);
1974         }
1975         return 0;
1976 }
1977
1978 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
1979 {
1980         if (vhost_put_avail_event(vq))
1981                 return -EFAULT;
1982         if (unlikely(vq->log_used)) {
1983                 void __user *used;
1984                 /* Make sure the event is seen before log. */
1985                 smp_wmb();
1986                 /* Log avail event write */
1987                 used = vhost_avail_event(vq);
1988                 log_used(vq, (used - (void __user *)vq->used),
1989                          sizeof *vhost_avail_event(vq));
1990                 if (vq->log_ctx)
1991                         eventfd_signal(vq->log_ctx, 1);
1992         }
1993         return 0;
1994 }
1995
1996 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1997 {
1998         __virtio16 last_used_idx;
1999         int r;
2000         bool is_le = vq->is_le;
2001
2002         if (!vq->private_data)
2003                 return 0;
2004
2005         vhost_init_is_le(vq);
2006
2007         r = vhost_update_used_flags(vq);
2008         if (r)
2009                 goto err;
2010         vq->signalled_used_valid = false;
2011         if (!vq->iotlb &&
2012             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2013                 r = -EFAULT;
2014                 goto err;
2015         }
2016         r = vhost_get_used_idx(vq, &last_used_idx);
2017         if (r) {
2018                 vq_err(vq, "Can't access used idx at %p\n",
2019                        &vq->used->idx);
2020                 goto err;
2021         }
2022         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2023         return 0;
2024
2025 err:
2026         vq->is_le = is_le;
2027         return r;
2028 }
2029 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2030
2031 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2032                           struct iovec iov[], int iov_size, int access)
2033 {
2034         const struct vhost_iotlb_map *map;
2035         struct vhost_dev *dev = vq->dev;
2036         struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2037         struct iovec *_iov;
2038         u64 s = 0, last = addr + len - 1;
2039         int ret = 0;
2040
2041         while ((u64)len > s) {
2042                 u64 size;
2043                 if (unlikely(ret >= iov_size)) {
2044                         ret = -ENOBUFS;
2045                         break;
2046                 }
2047
2048                 map = vhost_iotlb_itree_first(umem, addr, last);
2049                 if (map == NULL || map->start > addr) {
2050                         if (umem != dev->iotlb) {
2051                                 ret = -EFAULT;
2052                                 break;
2053                         }
2054                         ret = -EAGAIN;
2055                         break;
2056                 } else if (!(map->perm & access)) {
2057                         ret = -EPERM;
2058                         break;
2059                 }
2060
2061                 _iov = iov + ret;
2062                 size = map->size - addr + map->start;
2063                 _iov->iov_len = min((u64)len - s, size);
2064                 _iov->iov_base = (void __user *)(unsigned long)
2065                                  (map->addr + addr - map->start);
2066                 s += size;
2067                 addr += size;
2068                 ++ret;
2069         }
2070
2071         if (ret == -EAGAIN)
2072                 vhost_iotlb_miss(vq, addr, access);
2073         return ret;
2074 }
2075
2076 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
2077  * function returns the next descriptor in the chain,
2078  * or -1U if we're at the end. */
2079 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2080 {
2081         unsigned int next;
2082
2083         /* If this descriptor says it doesn't chain, we're done. */
2084         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2085                 return -1U;
2086
2087         /* Check they're not leading us off end of descriptors. */
2088         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2089         return next;
2090 }
2091
2092 static int get_indirect(struct vhost_virtqueue *vq,
2093                         struct iovec iov[], unsigned int iov_size,
2094                         unsigned int *out_num, unsigned int *in_num,
2095                         struct vhost_log *log, unsigned int *log_num,
2096                         struct vring_desc *indirect)
2097 {
2098         struct vring_desc desc;
2099         unsigned int i = 0, count, found = 0;
2100         u32 len = vhost32_to_cpu(vq, indirect->len);
2101         struct iov_iter from;
2102         int ret, access;
2103
2104         /* Sanity check */
2105         if (unlikely(len % sizeof desc)) {
2106                 vq_err(vq, "Invalid length in indirect descriptor: "
2107                        "len 0x%llx not multiple of 0x%zx\n",
2108                        (unsigned long long)len,
2109                        sizeof desc);
2110                 return -EINVAL;
2111         }
2112
2113         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2114                              UIO_MAXIOV, VHOST_ACCESS_RO);
2115         if (unlikely(ret < 0)) {
2116                 if (ret != -EAGAIN)
2117                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2118                 return ret;
2119         }
2120         iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2121         count = len / sizeof desc;
2122         /* Buffers are chained via a 16 bit next field, so
2123          * we can have at most 2^16 of these. */
2124         if (unlikely(count > USHRT_MAX + 1)) {
2125                 vq_err(vq, "Indirect buffer length too big: %d\n",
2126                        indirect->len);
2127                 return -E2BIG;
2128         }
2129
2130         do {
2131                 unsigned iov_count = *in_num + *out_num;
2132                 if (unlikely(++found > count)) {
2133                         vq_err(vq, "Loop detected: last one at %u "
2134                                "indirect size %u\n",
2135                                i, count);
2136                         return -EINVAL;
2137                 }
2138                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2139                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2140                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2141                         return -EINVAL;
2142                 }
2143                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2144                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2145                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2146                         return -EINVAL;
2147                 }
2148
2149                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2150                         access = VHOST_ACCESS_WO;
2151                 else
2152                         access = VHOST_ACCESS_RO;
2153
2154                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2155                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2156                                      iov_size - iov_count, access);
2157                 if (unlikely(ret < 0)) {
2158                         if (ret != -EAGAIN)
2159                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2160                                         ret, i);
2161                         return ret;
2162                 }
2163                 /* If this is an input descriptor, increment that count. */
2164                 if (access == VHOST_ACCESS_WO) {
2165                         *in_num += ret;
2166                         if (unlikely(log && ret)) {
2167                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2168                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2169                                 ++*log_num;
2170                         }
2171                 } else {
2172                         /* If it's an output descriptor, they're all supposed
2173                          * to come before any input descriptors. */
2174                         if (unlikely(*in_num)) {
2175                                 vq_err(vq, "Indirect descriptor "
2176                                        "has out after in: idx %d\n", i);
2177                                 return -EINVAL;
2178                         }
2179                         *out_num += ret;
2180                 }
2181         } while ((i = next_desc(vq, &desc)) != -1);
2182         return 0;
2183 }
2184
2185 /* This looks in the virtqueue and for the first available buffer, and converts
2186  * it to an iovec for convenient access.  Since descriptors consist of some
2187  * number of output then some number of input descriptors, it's actually two
2188  * iovecs, but we pack them into one and note how many of each there were.
2189  *
2190  * This function returns the descriptor number found, or vq->num (which is
2191  * never a valid descriptor number) if none was found.  A negative code is
2192  * returned on error. */
2193 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2194                       struct iovec iov[], unsigned int iov_size,
2195                       unsigned int *out_num, unsigned int *in_num,
2196                       struct vhost_log *log, unsigned int *log_num)
2197 {
2198         struct vring_desc desc;
2199         unsigned int i, head, found = 0;
2200         u16 last_avail_idx;
2201         __virtio16 avail_idx;
2202         __virtio16 ring_head;
2203         int ret, access;
2204
2205         /* Check it isn't doing very strange things with descriptor numbers. */
2206         last_avail_idx = vq->last_avail_idx;
2207
2208         if (vq->avail_idx == vq->last_avail_idx) {
2209                 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2210                         vq_err(vq, "Failed to access avail idx at %p\n",
2211                                 &vq->avail->idx);
2212                         return -EFAULT;
2213                 }
2214                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2215
2216                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2217                         vq_err(vq, "Guest moved used index from %u to %u",
2218                                 last_avail_idx, vq->avail_idx);
2219                         return -EFAULT;
2220                 }
2221
2222                 /* If there's nothing new since last we looked, return
2223                  * invalid.
2224                  */
2225                 if (vq->avail_idx == last_avail_idx)
2226                         return vq->num;
2227
2228                 /* Only get avail ring entries after they have been
2229                  * exposed by guest.
2230                  */
2231                 smp_rmb();
2232         }
2233
2234         /* Grab the next descriptor number they're advertising, and increment
2235          * the index we've seen. */
2236         if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2237                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2238                        last_avail_idx,
2239                        &vq->avail->ring[last_avail_idx % vq->num]);
2240                 return -EFAULT;
2241         }
2242
2243         head = vhost16_to_cpu(vq, ring_head);
2244
2245         /* If their number is silly, that's an error. */
2246         if (unlikely(head >= vq->num)) {
2247                 vq_err(vq, "Guest says index %u > %u is available",
2248                        head, vq->num);
2249                 return -EINVAL;
2250         }
2251
2252         /* When we start there are none of either input nor output. */
2253         *out_num = *in_num = 0;
2254         if (unlikely(log))
2255                 *log_num = 0;
2256
2257         i = head;
2258         do {
2259                 unsigned iov_count = *in_num + *out_num;
2260                 if (unlikely(i >= vq->num)) {
2261                         vq_err(vq, "Desc index is %u > %u, head = %u",
2262                                i, vq->num, head);
2263                         return -EINVAL;
2264                 }
2265                 if (unlikely(++found > vq->num)) {
2266                         vq_err(vq, "Loop detected: last one at %u "
2267                                "vq size %u head %u\n",
2268                                i, vq->num, head);
2269                         return -EINVAL;
2270                 }
2271                 ret = vhost_get_desc(vq, &desc, i);
2272                 if (unlikely(ret)) {
2273                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2274                                i, vq->desc + i);
2275                         return -EFAULT;
2276                 }
2277                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2278                         ret = get_indirect(vq, iov, iov_size,
2279                                            out_num, in_num,
2280                                            log, log_num, &desc);
2281                         if (unlikely(ret < 0)) {
2282                                 if (ret != -EAGAIN)
2283                                         vq_err(vq, "Failure detected "
2284                                                 "in indirect descriptor at idx %d\n", i);
2285                                 return ret;
2286                         }
2287                         continue;
2288                 }
2289
2290                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2291                         access = VHOST_ACCESS_WO;
2292                 else
2293                         access = VHOST_ACCESS_RO;
2294                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2295                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2296                                      iov_size - iov_count, access);
2297                 if (unlikely(ret < 0)) {
2298                         if (ret != -EAGAIN)
2299                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2300                                         ret, i);
2301                         return ret;
2302                 }
2303                 if (access == VHOST_ACCESS_WO) {
2304                         /* If this is an input descriptor,
2305                          * increment that count. */
2306                         *in_num += ret;
2307                         if (unlikely(log && ret)) {
2308                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2309                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2310                                 ++*log_num;
2311                         }
2312                 } else {
2313                         /* If it's an output descriptor, they're all supposed
2314                          * to come before any input descriptors. */
2315                         if (unlikely(*in_num)) {
2316                                 vq_err(vq, "Descriptor has out after in: "
2317                                        "idx %d\n", i);
2318                                 return -EINVAL;
2319                         }
2320                         *out_num += ret;
2321                 }
2322         } while ((i = next_desc(vq, &desc)) != -1);
2323
2324         /* On success, increment avail index. */
2325         vq->last_avail_idx++;
2326
2327         /* Assume notifications from guest are disabled at this point,
2328          * if they aren't we would need to update avail_event index. */
2329         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2330         return head;
2331 }
2332 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2333
2334 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2335 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2336 {
2337         vq->last_avail_idx -= n;
2338 }
2339 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2340
2341 /* After we've used one of their buffers, we tell them about it.  We'll then
2342  * want to notify the guest, using eventfd. */
2343 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2344 {
2345         struct vring_used_elem heads = {
2346                 cpu_to_vhost32(vq, head),
2347                 cpu_to_vhost32(vq, len)
2348         };
2349
2350         return vhost_add_used_n(vq, &heads, 1);
2351 }
2352 EXPORT_SYMBOL_GPL(vhost_add_used);
2353
2354 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2355                             struct vring_used_elem *heads,
2356                             unsigned count)
2357 {
2358         vring_used_elem_t __user *used;
2359         u16 old, new;
2360         int start;
2361
2362         start = vq->last_used_idx & (vq->num - 1);
2363         used = vq->used->ring + start;
2364         if (vhost_put_used(vq, heads, start, count)) {
2365                 vq_err(vq, "Failed to write used");
2366                 return -EFAULT;
2367         }
2368         if (unlikely(vq->log_used)) {
2369                 /* Make sure data is seen before log. */
2370                 smp_wmb();
2371                 /* Log used ring entry write. */
2372                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2373                          count * sizeof *used);
2374         }
2375         old = vq->last_used_idx;
2376         new = (vq->last_used_idx += count);
2377         /* If the driver never bothers to signal in a very long while,
2378          * used index might wrap around. If that happens, invalidate
2379          * signalled_used index we stored. TODO: make sure driver
2380          * signals at least once in 2^16 and remove this. */
2381         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2382                 vq->signalled_used_valid = false;
2383         return 0;
2384 }
2385
2386 /* After we've used one of their buffers, we tell them about it.  We'll then
2387  * want to notify the guest, using eventfd. */
2388 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2389                      unsigned count)
2390 {
2391         int start, n, r;
2392
2393         start = vq->last_used_idx & (vq->num - 1);
2394         n = vq->num - start;
2395         if (n < count) {
2396                 r = __vhost_add_used_n(vq, heads, n);
2397                 if (r < 0)
2398                         return r;
2399                 heads += n;
2400                 count -= n;
2401         }
2402         r = __vhost_add_used_n(vq, heads, count);
2403
2404         /* Make sure buffer is written before we update index. */
2405         smp_wmb();
2406         if (vhost_put_used_idx(vq)) {
2407                 vq_err(vq, "Failed to increment used idx");
2408                 return -EFAULT;
2409         }
2410         if (unlikely(vq->log_used)) {
2411                 /* Make sure used idx is seen before log. */
2412                 smp_wmb();
2413                 /* Log used index update. */
2414                 log_used(vq, offsetof(struct vring_used, idx),
2415                          sizeof vq->used->idx);
2416                 if (vq->log_ctx)
2417                         eventfd_signal(vq->log_ctx, 1);
2418         }
2419         return r;
2420 }
2421 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2422
2423 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2424 {
2425         __u16 old, new;
2426         __virtio16 event;
2427         bool v;
2428         /* Flush out used index updates. This is paired
2429          * with the barrier that the Guest executes when enabling
2430          * interrupts. */
2431         smp_mb();
2432
2433         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2434             unlikely(vq->avail_idx == vq->last_avail_idx))
2435                 return true;
2436
2437         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2438                 __virtio16 flags;
2439                 if (vhost_get_avail_flags(vq, &flags)) {
2440                         vq_err(vq, "Failed to get flags");
2441                         return true;
2442                 }
2443                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2444         }
2445         old = vq->signalled_used;
2446         v = vq->signalled_used_valid;
2447         new = vq->signalled_used = vq->last_used_idx;
2448         vq->signalled_used_valid = true;
2449
2450         if (unlikely(!v))
2451                 return true;
2452
2453         if (vhost_get_used_event(vq, &event)) {
2454                 vq_err(vq, "Failed to get used event idx");
2455                 return true;
2456         }
2457         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2458 }
2459
2460 /* This actually signals the guest, using eventfd. */
2461 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2462 {
2463         /* Signal the Guest tell them we used something up. */
2464         if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2465                 eventfd_signal(vq->call_ctx.ctx, 1);
2466 }
2467 EXPORT_SYMBOL_GPL(vhost_signal);
2468
2469 /* And here's the combo meal deal.  Supersize me! */
2470 void vhost_add_used_and_signal(struct vhost_dev *dev,
2471                                struct vhost_virtqueue *vq,
2472                                unsigned int head, int len)
2473 {
2474         vhost_add_used(vq, head, len);
2475         vhost_signal(dev, vq);
2476 }
2477 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2478
2479 /* multi-buffer version of vhost_add_used_and_signal */
2480 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2481                                  struct vhost_virtqueue *vq,
2482                                  struct vring_used_elem *heads, unsigned count)
2483 {
2484         vhost_add_used_n(vq, heads, count);
2485         vhost_signal(dev, vq);
2486 }
2487 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2488
2489 /* return true if we're sure that avaiable ring is empty */
2490 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2491 {
2492         __virtio16 avail_idx;
2493         int r;
2494
2495         if (vq->avail_idx != vq->last_avail_idx)
2496                 return false;
2497
2498         r = vhost_get_avail_idx(vq, &avail_idx);
2499         if (unlikely(r))
2500                 return false;
2501         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2502
2503         return vq->avail_idx == vq->last_avail_idx;
2504 }
2505 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2506
2507 /* OK, now we need to know about added descriptors. */
2508 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2509 {
2510         __virtio16 avail_idx;
2511         int r;
2512
2513         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2514                 return false;
2515         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2516         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2517                 r = vhost_update_used_flags(vq);
2518                 if (r) {
2519                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2520                                &vq->used->flags, r);
2521                         return false;
2522                 }
2523         } else {
2524                 r = vhost_update_avail_event(vq);
2525                 if (r) {
2526                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2527                                vhost_avail_event(vq), r);
2528                         return false;
2529                 }
2530         }
2531         /* They could have slipped one in as we were doing that: make
2532          * sure it's written, then check again. */
2533         smp_mb();
2534         r = vhost_get_avail_idx(vq, &avail_idx);
2535         if (r) {
2536                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2537                        &vq->avail->idx, r);
2538                 return false;
2539         }
2540         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2541
2542         return vq->avail_idx != vq->last_avail_idx;
2543 }
2544 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2545
2546 /* We don't need to be notified again. */
2547 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2548 {
2549         int r;
2550
2551         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2552                 return;
2553         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2554         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2555                 r = vhost_update_used_flags(vq);
2556                 if (r)
2557                         vq_err(vq, "Failed to disable notification at %p: %d\n",
2558                                &vq->used->flags, r);
2559         }
2560 }
2561 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2562
2563 /* Create a new message. */
2564 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2565 {
2566         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2567         if (!node)
2568                 return NULL;
2569
2570         /* Make sure all padding within the structure is initialized. */
2571         memset(&node->msg, 0, sizeof node->msg);
2572         node->vq = vq;
2573         node->msg.type = type;
2574         return node;
2575 }
2576 EXPORT_SYMBOL_GPL(vhost_new_msg);
2577
2578 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2579                        struct vhost_msg_node *node)
2580 {
2581         spin_lock(&dev->iotlb_lock);
2582         list_add_tail(&node->node, head);
2583         spin_unlock(&dev->iotlb_lock);
2584
2585         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2586 }
2587 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2588
2589 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2590                                          struct list_head *head)
2591 {
2592         struct vhost_msg_node *node = NULL;
2593
2594         spin_lock(&dev->iotlb_lock);
2595         if (!list_empty(head)) {
2596                 node = list_first_entry(head, struct vhost_msg_node,
2597                                         node);
2598                 list_del(&node->node);
2599         }
2600         spin_unlock(&dev->iotlb_lock);
2601
2602         return node;
2603 }
2604 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2605
2606 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2607 {
2608         struct vhost_virtqueue *vq;
2609         int i;
2610
2611         mutex_lock(&dev->mutex);
2612         for (i = 0; i < dev->nvqs; ++i) {
2613                 vq = dev->vqs[i];
2614                 mutex_lock(&vq->mutex);
2615                 vq->acked_backend_features = features;
2616                 mutex_unlock(&vq->mutex);
2617         }
2618         mutex_unlock(&dev->mutex);
2619 }
2620 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2621
2622 static int __init vhost_init(void)
2623 {
2624         return 0;
2625 }
2626
2627 static void __exit vhost_exit(void)
2628 {
2629 }
2630
2631 module_init(vhost_init);
2632 module_exit(vhost_exit);
2633
2634 MODULE_VERSION("0.0.1");
2635 MODULE_LICENSE("GPL v2");
2636 MODULE_AUTHOR("Michael S. Tsirkin");
2637 MODULE_DESCRIPTION("Host kernel accelerator for virtio");