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 <[email protected]>
   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                      struct vhost_virtqueue *vq)
 192 {
 193         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 194         init_poll_funcptr(&poll->table, vhost_poll_func);
 195         poll->mask = mask;
 196         poll->dev = dev;
 197         poll->wqh = NULL;
 198         poll->vq = vq;
 199
 200         vhost_work_init(&poll->work, fn);
 201 }
 202 EXPORT_SYMBOL_GPL(vhost_poll_init);
 203
 204 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 205  * keep a reference to a file until after vhost_poll_stop is called. */
 206 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 207 {
 208         __poll_t mask;
 209
 210         if (poll->wqh)
 211                 return 0;
 212
 213         mask = vfs_poll(file, &poll->table);
 214         if (mask)
 215                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 216         if (mask & EPOLLERR) {
 217                 vhost_poll_stop(poll);
 218                 return -EINVAL;
 219         }
 220
 221         return 0;
 222 }
 223 EXPORT_SYMBOL_GPL(vhost_poll_start);
 224
 225 /* Stop polling a file. After this function returns, it becomes safe to drop the
 226  * file reference. You must also flush afterwards. */
 227 void vhost_poll_stop(struct vhost_poll *poll)
 228 {
 229         if (poll->wqh) {
 230                 remove_wait_queue(poll->wqh, &poll->wait);
 231                 poll->wqh = NULL;
 232         }
 233 }
 234 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 235
 236 static void vhost_worker_queue(struct vhost_worker *worker,
 237                                struct vhost_work *work)
 238 {
 239         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 240                 /* We can only add the work to the list after we're
 241                  * sure it was not in the list.
 242                  * test_and_set_bit() implies a memory barrier.
 243                  */
 244                 llist_add(&work->node, &worker->work_list);
 245                 vhost_task_wake(worker->vtsk);
 246         }
 247 }
 248
 249 bool vhost_vq_work_queue(struct vhost_virtqueue *vq, struct vhost_work *work)
 250 {
 251         struct vhost_worker *worker;
 252         bool queued = false;
 253
 254         rcu_read_lock();
 255         worker = rcu_dereference(vq->worker);
 256         if (worker) {
 257                 queued = true;
 258                 vhost_worker_queue(worker, work);
 259         }
 260         rcu_read_unlock();
 261
 262         return queued;
 263 }
 264 EXPORT_SYMBOL_GPL(vhost_vq_work_queue);
 265
 266 /**
 267  * __vhost_worker_flush - flush a worker
 268  * @worker: worker to flush
 269  *
 270  * The worker's flush_mutex must be held.
 271  */
 272 static void __vhost_worker_flush(struct vhost_worker *worker)
 273 {
 274         struct vhost_flush_struct flush;
 275
 276         if (!worker->attachment_cnt || worker->killed)
 277                 return;
 278
 279         init_completion(&flush.wait_event);
 280         vhost_work_init(&flush.work, vhost_flush_work);
 281
 282         vhost_worker_queue(worker, &flush.work);
 283         /*
 284          * Drop mutex in case our worker is killed and it needs to take the
 285          * mutex to force cleanup.
 286          */
 287         mutex_unlock(&worker->mutex);
 288         wait_for_completion(&flush.wait_event);
 289         mutex_lock(&worker->mutex);
 290 }
 291
 292 static void vhost_worker_flush(struct vhost_worker *worker)
 293 {
 294         mutex_lock(&worker->mutex);
 295         __vhost_worker_flush(worker);
 296         mutex_unlock(&worker->mutex);
 297 }
 298
 299 void vhost_dev_flush(struct vhost_dev *dev)
 300 {
 301         struct vhost_worker *worker;
 302         unsigned long i;
 303
 304         xa_for_each(&dev->worker_xa, i, worker)
 305                 vhost_worker_flush(worker);
 306 }
 307 EXPORT_SYMBOL_GPL(vhost_dev_flush);
 308
 309 /* A lockless hint for busy polling code to exit the loop */
 310 bool vhost_vq_has_work(struct vhost_virtqueue *vq)
 311 {
 312         struct vhost_worker *worker;
 313         bool has_work = false;
 314
 315         rcu_read_lock();
 316         worker = rcu_dereference(vq->worker);
 317         if (worker && !llist_empty(&worker->work_list))
 318                 has_work = true;
 319         rcu_read_unlock();
 320
 321         return has_work;
 322 }
 323 EXPORT_SYMBOL_GPL(vhost_vq_has_work);
 324
 325 void vhost_poll_queue(struct vhost_poll *poll)
 326 {
 327         vhost_vq_work_queue(poll->vq, &poll->work);
 328 }
 329 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 330
 331 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 332 {
 333         int j;
 334
 335         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 336                 vq->meta_iotlb[j] = NULL;
 337 }
 338
 339 static void vhost_vq_meta_reset(struct vhost_dev *d)
 340 {
 341         int i;
 342
 343         for (i = 0; i < d->nvqs; ++i)
 344                 __vhost_vq_meta_reset(d->vqs[i]);
 345 }
 346
 347 static void vhost_vring_call_reset(struct vhost_vring_call *call_ctx)
 348 {
 349         call_ctx->ctx = NULL;
 350         memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
 351 }
 352
 353 bool vhost_vq_is_setup(struct vhost_virtqueue *vq)
 354 {
 355         return vq->avail && vq->desc && vq->used && vhost_vq_access_ok(vq);
 356 }
 357 EXPORT_SYMBOL_GPL(vhost_vq_is_setup);
 358
 359 static void vhost_vq_reset(struct vhost_dev *dev,
 360                            struct vhost_virtqueue *vq)
 361 {
 362         vq->num = 1;
 363         vq->desc = NULL;
 364         vq->avail = NULL;
 365         vq->used = NULL;
 366         vq->last_avail_idx = 0;
 367         vq->avail_idx = 0;
 368         vq->last_used_idx = 0;
 369         vq->signalled_used = 0;
 370         vq->signalled_used_valid = false;
 371         vq->used_flags = 0;
 372         vq->log_used = false;
 373         vq->log_addr = -1ull;
 374         vq->private_data = NULL;
 375         vq->acked_features = 0;
 376         vq->acked_backend_features = 0;
 377         vq->log_base = NULL;
 378         vq->error_ctx = NULL;
 379         vq->kick = NULL;
 380         vq->log_ctx = NULL;
 381         vhost_disable_cross_endian(vq);
 382         vhost_reset_is_le(vq);
 383         vq->busyloop_timeout = 0;
 384         vq->umem = NULL;
 385         vq->iotlb = NULL;
 386         rcu_assign_pointer(vq->worker, NULL);
 387         vhost_vring_call_reset(&vq->call_ctx);
 388         __vhost_vq_meta_reset(vq);
 389 }
 390
 391 static bool vhost_run_work_list(void *data)
 392 {
 393         struct vhost_worker *worker = data;
 394         struct vhost_work *work, *work_next;
 395         struct llist_node *node;
 396
 397         node = llist_del_all(&worker->work_list);
 398         if (node) {
 399                 __set_current_state(TASK_RUNNING);
 400
 401                 node = llist_reverse_order(node);
 402                 /* make sure flag is seen after deletion */
 403                 smp_wmb();
 404                 llist_for_each_entry_safe(work, work_next, node, node) {
 405                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 406                         kcov_remote_start_common(worker->kcov_handle);
 407                         work->fn(work);
 408                         kcov_remote_stop();
 409                         cond_resched();
 410                 }
 411         }
 412
 413         return !!node;
 414 }
 415
 416 static void vhost_worker_killed(void *data)
 417 {
 418         struct vhost_worker *worker = data;
 419         struct vhost_dev *dev = worker->dev;
 420         struct vhost_virtqueue *vq;
 421         int i, attach_cnt = 0;
 422
 423         mutex_lock(&worker->mutex);
 424         worker->killed = true;
 425
 426         for (i = 0; i < dev->nvqs; i++) {
 427                 vq = dev->vqs[i];
 428
 429                 mutex_lock(&vq->mutex);
 430                 if (worker ==
 431                     rcu_dereference_check(vq->worker,
 432                                           lockdep_is_held(&vq->mutex))) {
 433                         rcu_assign_pointer(vq->worker, NULL);
 434                         attach_cnt++;
 435                 }
 436                 mutex_unlock(&vq->mutex);
 437         }
 438
 439         worker->attachment_cnt -= attach_cnt;
 440         if (attach_cnt)
 441                 synchronize_rcu();
 442         /*
 443          * Finish vhost_worker_flush calls and any other works that snuck in
 444          * before the synchronize_rcu.
 445          */
 446         vhost_run_work_list(worker);
 447         mutex_unlock(&worker->mutex);
 448 }
 449
 450 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 451 {
 452         kfree(vq->indirect);
 453         vq->indirect = NULL;
 454         kfree(vq->log);
 455         vq->log = NULL;
 456         kfree(vq->heads);
 457         vq->heads = NULL;
 458 }
 459
 460 /* Helper to allocate iovec buffers for all vqs. */
 461 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 462 {
 463         struct vhost_virtqueue *vq;
 464         int i;
 465
 466         for (i = 0; i < dev->nvqs; ++i) {
 467                 vq = dev->vqs[i];
 468                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 469                                              sizeof(*vq->indirect),
 470                                              GFP_KERNEL);
 471                 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 472                                         GFP_KERNEL);
 473                 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 474                                           GFP_KERNEL);
 475                 if (!vq->indirect || !vq->log || !vq->heads)
 476                         goto err_nomem;
 477         }
 478         return 0;
 479
 480 err_nomem:
 481         for (; i >= 0; --i)
 482                 vhost_vq_free_iovecs(dev->vqs[i]);
 483         return -ENOMEM;
 484 }
 485
 486 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 487 {
 488         int i;
 489
 490         for (i = 0; i < dev->nvqs; ++i)
 491                 vhost_vq_free_iovecs(dev->vqs[i]);
 492 }
 493
 494 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 495                           int pkts, int total_len)
 496 {
 497         struct vhost_dev *dev = vq->dev;
 498
 499         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 500             pkts >= dev->weight) {
 501                 vhost_poll_queue(&vq->poll);
 502                 return true;
 503         }
 504
 505         return false;
 506 }
 507 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 508
 509 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 510                                    unsigned int num)
 511 {
 512         size_t event __maybe_unused =
 513                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 514
 515         return size_add(struct_size(vq->avail, ring, num), event);
 516 }
 517
 518 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 519                                   unsigned int num)
 520 {
 521         size_t event __maybe_unused =
 522                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 523
 524         return size_add(struct_size(vq->used, ring, num), event);
 525 }
 526
 527 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 528                                   unsigned int num)
 529 {
 530         return sizeof(*vq->desc) * num;
 531 }
 532
 533 void vhost_dev_init(struct vhost_dev *dev,
 534                     struct vhost_virtqueue **vqs, int nvqs,
 535                     int iov_limit, int weight, int byte_weight,
 536                     bool use_worker,
 537                     int (*msg_handler)(struct vhost_dev *dev, u32 asid,
 538                                        struct vhost_iotlb_msg *msg))
 539 {
 540         struct vhost_virtqueue *vq;
 541         int i;
 542
 543         dev->vqs = vqs;
 544         dev->nvqs = nvqs;
 545         mutex_init(&dev->mutex);
 546         dev->log_ctx = NULL;
 547         dev->umem = NULL;
 548         dev->iotlb = NULL;
 549         dev->mm = NULL;
 550         dev->iov_limit = iov_limit;
 551         dev->weight = weight;
 552         dev->byte_weight = byte_weight;
 553         dev->use_worker = use_worker;
 554         dev->msg_handler = msg_handler;
 555         init_waitqueue_head(&dev->wait);
 556         INIT_LIST_HEAD(&dev->read_list);
 557         INIT_LIST_HEAD(&dev->pending_list);
 558         spin_lock_init(&dev->iotlb_lock);
 559         xa_init_flags(&dev->worker_xa, XA_FLAGS_ALLOC);
 560
 561         for (i = 0; i < dev->nvqs; ++i) {
 562                 vq = dev->vqs[i];
 563                 vq->log = NULL;
 564                 vq->indirect = NULL;
 565                 vq->heads = NULL;
 566                 vq->dev = dev;
 567                 mutex_init(&vq->mutex);
 568                 vhost_vq_reset(dev, vq);
 569                 if (vq->handle_kick)
 570                         vhost_poll_init(&vq->poll, vq->handle_kick,
 571                                         EPOLLIN, dev, vq);
 572         }
 573 }
 574 EXPORT_SYMBOL_GPL(vhost_dev_init);
 575
 576 /* Caller should have device mutex */
 577 long vhost_dev_check_owner(struct vhost_dev *dev)
 578 {
 579         /* Are you the owner? If not, I don't think you mean to do that */
 580         return dev->mm == current->mm ? 0 : -EPERM;
 581 }
 582 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 583
 584 /* Caller should have device mutex */
 585 bool vhost_dev_has_owner(struct vhost_dev *dev)
 586 {
 587         return dev->mm;
 588 }
 589 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 590
 591 static void vhost_attach_mm(struct vhost_dev *dev)
 592 {
 593         /* No owner, become one */
 594         if (dev->use_worker) {
 595                 dev->mm = get_task_mm(current);
 596         } else {
 597                 /* vDPA device does not use worker thead, so there's
 598                  * no need to hold the address space for mm. This help
 599                  * to avoid deadlock in the case of mmap() which may
 600                  * held the refcnt of the file and depends on release
 601                  * method to remove vma.
 602                  */
 603                 dev->mm = current->mm;
 604                 mmgrab(dev->mm);
 605         }
 606 }
 607
 608 static void vhost_detach_mm(struct vhost_dev *dev)
 609 {
 610         if (!dev->mm)
 611                 return;
 612
 613         if (dev->use_worker)
 614                 mmput(dev->mm);
 615         else
 616                 mmdrop(dev->mm);
 617
 618         dev->mm = NULL;
 619 }
 620
 621 static void vhost_worker_destroy(struct vhost_dev *dev,
 622                                  struct vhost_worker *worker)
 623 {
 624         if (!worker)
 625                 return;
 626
 627         WARN_ON(!llist_empty(&worker->work_list));
 628         xa_erase(&dev->worker_xa, worker->id);
 629         vhost_task_stop(worker->vtsk);
 630         kfree(worker);
 631 }
 632
 633 static void vhost_workers_free(struct vhost_dev *dev)
 634 {
 635         struct vhost_worker *worker;
 636         unsigned long i;
 637
 638         if (!dev->use_worker)
 639                 return;
 640
 641         for (i = 0; i < dev->nvqs; i++)
 642                 rcu_assign_pointer(dev->vqs[i]->worker, NULL);
 643         /*
 644          * Free the default worker we created and cleanup workers userspace
 645          * created but couldn't clean up (it forgot or crashed).
 646          */
 647         xa_for_each(&dev->worker_xa, i, worker)
 648                 vhost_worker_destroy(dev, worker);
 649         xa_destroy(&dev->worker_xa);
 650 }
 651
 652 static struct vhost_worker *vhost_worker_create(struct vhost_dev *dev)
 653 {
 654         struct vhost_worker *worker;
 655         struct vhost_task *vtsk;
 656         char name[TASK_COMM_LEN];
 657         int ret;
 658         u32 id;
 659
 660         worker = kzalloc(sizeof(*worker), GFP_KERNEL_ACCOUNT);
 661         if (!worker)
 662                 return NULL;
 663
 664         worker->dev = dev;
 665         snprintf(name, sizeof(name), "vhost-%d", current->pid);
 666
 667         vtsk = vhost_task_create(vhost_run_work_list, vhost_worker_killed,
 668                                  worker, name);
 669         if (!vtsk)
 670                 goto free_worker;
 671
 672         mutex_init(&worker->mutex);
 673         init_llist_head(&worker->work_list);
 674         worker->kcov_handle = kcov_common_handle();
 675         worker->vtsk = vtsk;
 676
 677         vhost_task_start(vtsk);
 678
 679         ret = xa_alloc(&dev->worker_xa, &id, worker, xa_limit_32b, GFP_KERNEL);
 680         if (ret < 0)
 681                 goto stop_worker;
 682         worker->id = id;
 683
 684         return worker;
 685
 686 stop_worker:
 687         vhost_task_stop(vtsk);
 688 free_worker:
 689         kfree(worker);
 690         return NULL;
 691 }
 692
 693 /* Caller must have device mutex */
 694 static void __vhost_vq_attach_worker(struct vhost_virtqueue *vq,
 695                                      struct vhost_worker *worker)
 696 {
 697         struct vhost_worker *old_worker;
 698
 699         mutex_lock(&worker->mutex);
 700         if (worker->killed) {
 701                 mutex_unlock(&worker->mutex);
 702                 return;
 703         }
 704
 705         mutex_lock(&vq->mutex);
 706
 707         old_worker = rcu_dereference_check(vq->worker,
 708                                            lockdep_is_held(&vq->mutex));
 709         rcu_assign_pointer(vq->worker, worker);
 710         worker->attachment_cnt++;
 711
 712         if (!old_worker) {
 713                 mutex_unlock(&vq->mutex);
 714                 mutex_unlock(&worker->mutex);
 715                 return;
 716         }
 717         mutex_unlock(&vq->mutex);
 718         mutex_unlock(&worker->mutex);
 719
 720         /*
 721          * Take the worker mutex to make sure we see the work queued from
 722          * device wide flushes which doesn't use RCU for execution.
 723          */
 724         mutex_lock(&old_worker->mutex);
 725         if (old_worker->killed) {
 726                 mutex_unlock(&old_worker->mutex);
 727                 return;
 728         }
 729
 730         /*
 731          * We don't want to call synchronize_rcu for every vq during setup
 732          * because it will slow down VM startup. If we haven't done
 733          * VHOST_SET_VRING_KICK and not done the driver specific
 734          * SET_ENDPOINT/RUNNUNG then we can skip the sync since there will
 735          * not be any works queued for scsi and net.
 736          */
 737         mutex_lock(&vq->mutex);
 738         if (!vhost_vq_get_backend(vq) && !vq->kick) {
 739                 mutex_unlock(&vq->mutex);
 740
 741                 old_worker->attachment_cnt--;
 742                 mutex_unlock(&old_worker->mutex);
 743                 /*
 744                  * vsock can queue anytime after VHOST_VSOCK_SET_GUEST_CID.
 745                  * Warn if it adds support for multiple workers but forgets to
 746                  * handle the early queueing case.
 747                  */
 748                 WARN_ON(!old_worker->attachment_cnt &&
 749                         !llist_empty(&old_worker->work_list));
 750                 return;
 751         }
 752         mutex_unlock(&vq->mutex);
 753
 754         /* Make sure new vq queue/flush/poll calls see the new worker */
 755         synchronize_rcu();
 756         /* Make sure whatever was queued gets run */
 757         __vhost_worker_flush(old_worker);
 758         old_worker->attachment_cnt--;
 759         mutex_unlock(&old_worker->mutex);
 760 }
 761
 762  /* Caller must have device mutex */
 763 static int vhost_vq_attach_worker(struct vhost_virtqueue *vq,
 764                                   struct vhost_vring_worker *info)
 765 {
 766         unsigned long index = info->worker_id;
 767         struct vhost_dev *dev = vq->dev;
 768         struct vhost_worker *worker;
 769
 770         if (!dev->use_worker)
 771                 return -EINVAL;
 772
 773         worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
 774         if (!worker || worker->id != info->worker_id)
 775                 return -ENODEV;
 776
 777         __vhost_vq_attach_worker(vq, worker);
 778         return 0;
 779 }
 780
 781 /* Caller must have device mutex */
 782 static int vhost_new_worker(struct vhost_dev *dev,
 783                             struct vhost_worker_state *info)
 784 {
 785         struct vhost_worker *worker;
 786
 787         worker = vhost_worker_create(dev);
 788         if (!worker)
 789                 return -ENOMEM;
 790
 791         info->worker_id = worker->id;
 792         return 0;
 793 }
 794
 795 /* Caller must have device mutex */
 796 static int vhost_free_worker(struct vhost_dev *dev,
 797                              struct vhost_worker_state *info)
 798 {
 799         unsigned long index = info->worker_id;
 800         struct vhost_worker *worker;
 801
 802         worker = xa_find(&dev->worker_xa, &index, UINT_MAX, XA_PRESENT);
 803         if (!worker || worker->id != info->worker_id)
 804                 return -ENODEV;
 805
 806         mutex_lock(&worker->mutex);
 807         if (worker->attachment_cnt || worker->killed) {
 808                 mutex_unlock(&worker->mutex);
 809                 return -EBUSY;
 810         }
 811         /*
 812          * A flush might have raced and snuck in before attachment_cnt was set
 813          * to zero. Make sure flushes are flushed from the queue before
 814          * freeing.
 815          */
 816         __vhost_worker_flush(worker);
 817         mutex_unlock(&worker->mutex);
 818
 819         vhost_worker_destroy(dev, worker);
 820         return 0;
 821 }
 822
 823 static int vhost_get_vq_from_user(struct vhost_dev *dev, void __user *argp,
 824                                   struct vhost_virtqueue **vq, u32 *id)
 825 {
 826         u32 __user *idxp = argp;
 827         u32 idx;
 828         long r;
 829
 830         r = get_user(idx, idxp);
 831         if (r < 0)
 832                 return r;
 833
 834         if (idx >= dev->nvqs)
 835                 return -ENOBUFS;
 836
 837         idx = array_index_nospec(idx, dev->nvqs);
 838
 839         *vq = dev->vqs[idx];
 840         *id = idx;
 841         return 0;
 842 }
 843
 844 /* Caller must have device mutex */
 845 long vhost_worker_ioctl(struct vhost_dev *dev, unsigned int ioctl,
 846                         void __user *argp)
 847 {
 848         struct vhost_vring_worker ring_worker;
 849         struct vhost_worker_state state;
 850         struct vhost_worker *worker;
 851         struct vhost_virtqueue *vq;
 852         long ret;
 853         u32 idx;
 854
 855         if (!dev->use_worker)
 856                 return -EINVAL;
 857
 858         if (!vhost_dev_has_owner(dev))
 859                 return -EINVAL;
 860
 861         ret = vhost_dev_check_owner(dev);
 862         if (ret)
 863                 return ret;
 864
 865         switch (ioctl) {
 866         /* dev worker ioctls */
 867         case VHOST_NEW_WORKER:
 868                 ret = vhost_new_worker(dev, &state);
 869                 if (!ret && copy_to_user(argp, &state, sizeof(state)))
 870                         ret = -EFAULT;
 871                 return ret;
 872         case VHOST_FREE_WORKER:
 873                 if (copy_from_user(&state, argp, sizeof(state)))
 874                         return -EFAULT;
 875                 return vhost_free_worker(dev, &state);
 876         /* vring worker ioctls */
 877         case VHOST_ATTACH_VRING_WORKER:
 878         case VHOST_GET_VRING_WORKER:
 879                 break;
 880         default:
 881                 return -ENOIOCTLCMD;
 882         }
 883
 884         ret = vhost_get_vq_from_user(dev, argp, &vq, &idx);
 885         if (ret)
 886                 return ret;
 887
 888         switch (ioctl) {
 889         case VHOST_ATTACH_VRING_WORKER:
 890                 if (copy_from_user(&ring_worker, argp, sizeof(ring_worker))) {
 891                         ret = -EFAULT;
 892                         break;
 893                 }
 894
 895                 ret = vhost_vq_attach_worker(vq, &ring_worker);
 896                 break;
 897         case VHOST_GET_VRING_WORKER:
 898                 worker = rcu_dereference_check(vq->worker,
 899                                                lockdep_is_held(&dev->mutex));
 900                 if (!worker) {
 901                         ret = -EINVAL;
 902                         break;
 903                 }
 904
 905                 ring_worker.index = idx;
 906                 ring_worker.worker_id = worker->id;
 907
 908                 if (copy_to_user(argp, &ring_worker, sizeof(ring_worker)))
 909                         ret = -EFAULT;
 910                 break;
 911         default:
 912                 ret = -ENOIOCTLCMD;
 913                 break;
 914         }
 915
 916         return ret;
 917 }
 918 EXPORT_SYMBOL_GPL(vhost_worker_ioctl);
 919
 920 /* Caller should have device mutex */
 921 long vhost_dev_set_owner(struct vhost_dev *dev)
 922 {
 923         struct vhost_worker *worker;
 924         int err, i;
 925
 926         /* Is there an owner already? */
 927         if (vhost_dev_has_owner(dev)) {
 928                 err = -EBUSY;
 929                 goto err_mm;
 930         }
 931
 932         vhost_attach_mm(dev);
 933
 934         err = vhost_dev_alloc_iovecs(dev);
 935         if (err)
 936                 goto err_iovecs;
 937
 938         if (dev->use_worker) {
 939                 /*
 940                  * This should be done last, because vsock can queue work
 941                  * before VHOST_SET_OWNER so it simplifies the failure path
 942                  * below since we don't have to worry about vsock queueing
 943                  * while we free the worker.
 944                  */
 945                 worker = vhost_worker_create(dev);
 946                 if (!worker) {
 947                         err = -ENOMEM;
 948                         goto err_worker;
 949                 }
 950
 951                 for (i = 0; i < dev->nvqs; i++)
 952                         __vhost_vq_attach_worker(dev->vqs[i], worker);
 953         }
 954
 955         return 0;
 956
 957 err_worker:
 958         vhost_dev_free_iovecs(dev);
 959 err_iovecs:
 960         vhost_detach_mm(dev);
 961 err_mm:
 962         return err;
 963 }
 964 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 965
 966 static struct vhost_iotlb *iotlb_alloc(void)
 967 {
 968         return vhost_iotlb_alloc(max_iotlb_entries,
 969                                  VHOST_IOTLB_FLAG_RETIRE);
 970 }
 971
 972 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 973 {
 974         return iotlb_alloc();
 975 }
 976 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 977
 978 /* Caller should have device mutex */
 979 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 980 {
 981         int i;
 982
 983         vhost_dev_cleanup(dev);
 984
 985         dev->umem = umem;
 986         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 987          * VQs aren't running.
 988          */
 989         for (i = 0; i < dev->nvqs; ++i)
 990                 dev->vqs[i]->umem = umem;
 991 }
 992 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 993
 994 void vhost_dev_stop(struct vhost_dev *dev)
 995 {
 996         int i;
 997
 998         for (i = 0; i < dev->nvqs; ++i) {
 999                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick)
1000                         vhost_poll_stop(&dev->vqs[i]->poll);
1001         }
1002
1003         vhost_dev_flush(dev);
1004 }
1005 EXPORT_SYMBOL_GPL(vhost_dev_stop);
1006
1007 void vhost_clear_msg(struct vhost_dev *dev)
1008 {
1009         struct vhost_msg_node *node, *n;
1010
1011         spin_lock(&dev->iotlb_lock);
1012
1013         list_for_each_entry_safe(node, n, &dev->read_list, node) {
1014                 list_del(&node->node);
1015                 kfree(node);
1016         }
1017
1018         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
1019                 list_del(&node->node);
1020                 kfree(node);
1021         }
1022
1023         spin_unlock(&dev->iotlb_lock);
1024 }
1025 EXPORT_SYMBOL_GPL(vhost_clear_msg);
1026
1027 void vhost_dev_cleanup(struct vhost_dev *dev)
1028 {
1029         int i;
1030
1031         for (i = 0; i < dev->nvqs; ++i) {
1032                 if (dev->vqs[i]->error_ctx)
1033                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
1034                 if (dev->vqs[i]->kick)
1035                         fput(dev->vqs[i]->kick);
1036                 if (dev->vqs[i]->call_ctx.ctx)
1037                         eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
1038                 vhost_vq_reset(dev, dev->vqs[i]);
1039         }
1040         vhost_dev_free_iovecs(dev);
1041         if (dev->log_ctx)
1042                 eventfd_ctx_put(dev->log_ctx);
1043         dev->log_ctx = NULL;
1044         /* No one will access memory at this point */
1045         vhost_iotlb_free(dev->umem);
1046         dev->umem = NULL;
1047         vhost_iotlb_free(dev->iotlb);
1048         dev->iotlb = NULL;
1049         vhost_clear_msg(dev);
1050         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
1051         vhost_workers_free(dev);
1052         vhost_detach_mm(dev);
1053 }
1054 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
1055
1056 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
1057 {
1058         u64 a = addr / VHOST_PAGE_SIZE / 8;
1059
1060         /* Make sure 64 bit math will not overflow. */
1061         if (a > ULONG_MAX - (unsigned long)log_base ||
1062             a + (unsigned long)log_base > ULONG_MAX)
1063                 return false;
1064
1065         return access_ok(log_base + a,
1066                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
1067 }
1068
1069 /* Make sure 64 bit math will not overflow. */
1070 static bool vhost_overflow(u64 uaddr, u64 size)
1071 {
1072         if (uaddr > ULONG_MAX || size > ULONG_MAX)
1073                 return true;
1074
1075         if (!size)
1076                 return false;
1077
1078         return uaddr > ULONG_MAX - size + 1;
1079 }
1080
1081 /* Caller should have vq mutex and device mutex. */
1082 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
1083                                 int log_all)
1084 {
1085         struct vhost_iotlb_map *map;
1086
1087         if (!umem)
1088                 return false;
1089
1090         list_for_each_entry(map, &umem->list, link) {
1091                 unsigned long a = map->addr;
1092
1093                 if (vhost_overflow(map->addr, map->size))
1094                         return false;
1095
1096
1097                 if (!access_ok((void __user *)a, map->size))
1098                         return false;
1099                 else if (log_all && !log_access_ok(log_base,
1100                                                    map->start,
1101                                                    map->size))
1102                         return false;
1103         }
1104         return true;
1105 }
1106
1107 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
1108                                                u64 addr, unsigned int size,
1109                                                int type)
1110 {
1111         const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
1112
1113         if (!map)
1114                 return NULL;
1115
1116         return (void __user *)(uintptr_t)(map->addr + addr - map->start);
1117 }
1118
1119 /* Can we switch to this memory table? */
1120 /* Caller should have device mutex but not vq mutex */
1121 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
1122                              int log_all)
1123 {
1124         int i;
1125
1126         for (i = 0; i < d->nvqs; ++i) {
1127                 bool ok;
1128                 bool log;
1129
1130                 mutex_lock(&d->vqs[i]->mutex);
1131                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
1132                 /* If ring is inactive, will check when it's enabled. */
1133                 if (d->vqs[i]->private_data)
1134                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
1135                                                  umem, log);
1136                 else
1137                         ok = true;
1138                 mutex_unlock(&d->vqs[i]->mutex);
1139                 if (!ok)
1140                         return false;
1141         }
1142         return true;
1143 }
1144
1145 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1146                           struct iovec iov[], int iov_size, int access);
1147
1148 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
1149                               const void *from, unsigned size)
1150 {
1151         int ret;
1152
1153         if (!vq->iotlb)
1154                 return __copy_to_user(to, from, size);
1155         else {
1156                 /* This function should be called after iotlb
1157                  * prefetch, which means we're sure that all vq
1158                  * could be access through iotlb. So -EAGAIN should
1159                  * not happen in this case.
1160                  */
1161                 struct iov_iter t;
1162                 void __user *uaddr = vhost_vq_meta_fetch(vq,
1163                                      (u64)(uintptr_t)to, size,
1164                                      VHOST_ADDR_USED);
1165
1166                 if (uaddr)
1167                         return __copy_to_user(uaddr, from, size);
1168
1169                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
1170                                      ARRAY_SIZE(vq->iotlb_iov),
1171                                      VHOST_ACCESS_WO);
1172                 if (ret < 0)
1173                         goto out;
1174                 iov_iter_init(&t, ITER_DEST, vq->iotlb_iov, ret, size);
1175                 ret = copy_to_iter(from, size, &t);
1176                 if (ret == size)
1177                         ret = 0;
1178         }
1179 out:
1180         return ret;
1181 }
1182
1183 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
1184                                 void __user *from, unsigned size)
1185 {
1186         int ret;
1187
1188         if (!vq->iotlb)
1189                 return __copy_from_user(to, from, size);
1190         else {
1191                 /* This function should be called after iotlb
1192                  * prefetch, which means we're sure that vq
1193                  * could be access through iotlb. So -EAGAIN should
1194                  * not happen in this case.
1195                  */
1196                 void __user *uaddr = vhost_vq_meta_fetch(vq,
1197                                      (u64)(uintptr_t)from, size,
1198                                      VHOST_ADDR_DESC);
1199                 struct iov_iter f;
1200
1201                 if (uaddr)
1202                         return __copy_from_user(to, uaddr, size);
1203
1204                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
1205                                      ARRAY_SIZE(vq->iotlb_iov),
1206                                      VHOST_ACCESS_RO);
1207                 if (ret < 0) {
1208                         vq_err(vq, "IOTLB translation failure: uaddr "
1209                                "%p size 0x%llx\n", from,
1210                                (unsigned long long) size);
1211                         goto out;
1212                 }
1213                 iov_iter_init(&f, ITER_SOURCE, vq->iotlb_iov, ret, size);
1214                 ret = copy_from_iter(to, size, &f);
1215                 if (ret == size)
1216                         ret = 0;
1217         }
1218
1219 out:
1220         return ret;
1221 }
1222
1223 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
1224                                           void __user *addr, unsigned int size,
1225                                           int type)
1226 {
1227         int ret;
1228
1229         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
1230                              ARRAY_SIZE(vq->iotlb_iov),
1231                              VHOST_ACCESS_RO);
1232         if (ret < 0) {
1233                 vq_err(vq, "IOTLB translation failure: uaddr "
1234                         "%p size 0x%llx\n", addr,
1235                         (unsigned long long) size);
1236                 return NULL;
1237         }
1238
1239         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
1240                 vq_err(vq, "Non atomic userspace memory access: uaddr "
1241                         "%p size 0x%llx\n", addr,
1242                         (unsigned long long) size);
1243                 return NULL;
1244         }
1245
1246         return vq->iotlb_iov[0].iov_base;
1247 }
1248
1249 /* This function should be called after iotlb
1250  * prefetch, which means we're sure that vq
1251  * could be access through iotlb. So -EAGAIN should
1252  * not happen in this case.
1253  */
1254 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
1255                                             void __user *addr, unsigned int size,
1256                                             int type)
1257 {
1258         void __user *uaddr = vhost_vq_meta_fetch(vq,
1259                              (u64)(uintptr_t)addr, size, type);
1260         if (uaddr)
1261                 return uaddr;
1262
1263         return __vhost_get_user_slow(vq, addr, size, type);
1264 }
1265
1266 #define vhost_put_user(vq, x, ptr)              \
1267 ({ \
1268         int ret; \
1269         if (!vq->iotlb) { \
1270                 ret = __put_user(x, ptr); \
1271         } else { \
1272                 __typeof__(ptr) to = \
1273                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
1274                                           sizeof(*ptr), VHOST_ADDR_USED); \
1275                 if (to != NULL) \
1276                         ret = __put_user(x, to); \
1277                 else \
1278                         ret = -EFAULT;  \
1279         } \
1280         ret; \
1281 })
1282
1283 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
1284 {
1285         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1286                               vhost_avail_event(vq));
1287 }
1288
1289 static inline int vhost_put_used(struct vhost_virtqueue *vq,
1290                                  struct vring_used_elem *head, int idx,
1291                                  int count)
1292 {
1293         return vhost_copy_to_user(vq, vq->used->ring + idx, head,
1294                                   count * sizeof(*head));
1295 }
1296
1297 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
1298
1299 {
1300         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1301                               &vq->used->flags);
1302 }
1303
1304 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
1305
1306 {
1307         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
1308                               &vq->used->idx);
1309 }
1310
1311 #define vhost_get_user(vq, x, ptr, type)                \
1312 ({ \
1313         int ret; \
1314         if (!vq->iotlb) { \
1315                 ret = __get_user(x, ptr); \
1316         } else { \
1317                 __typeof__(ptr) from = \
1318                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
1319                                                            sizeof(*ptr), \
1320                                                            type); \
1321                 if (from != NULL) \
1322                         ret = __get_user(x, from); \
1323                 else \
1324                         ret = -EFAULT; \
1325         } \
1326         ret; \
1327 })
1328
1329 #define vhost_get_avail(vq, x, ptr) \
1330         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
1331
1332 #define vhost_get_used(vq, x, ptr) \
1333         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
1334
1335 static void vhost_dev_lock_vqs(struct vhost_dev *d)
1336 {
1337         int i = 0;
1338         for (i = 0; i < d->nvqs; ++i)
1339                 mutex_lock_nested(&d->vqs[i]->mutex, i);
1340 }
1341
1342 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1343 {
1344         int i = 0;
1345         for (i = 0; i < d->nvqs; ++i)
1346                 mutex_unlock(&d->vqs[i]->mutex);
1347 }
1348
1349 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq)
1350 {
1351         __virtio16 idx;
1352         int r;
1353
1354         r = vhost_get_avail(vq, idx, &vq->avail->idx);
1355         if (unlikely(r < 0)) {
1356                 vq_err(vq, "Failed to access available index at %p (%d)\n",
1357                        &vq->avail->idx, r);
1358                 return r;
1359         }
1360
1361         /* Check it isn't doing very strange thing with available indexes */
1362         vq->avail_idx = vhost16_to_cpu(vq, idx);
1363         if (unlikely((u16)(vq->avail_idx - vq->last_avail_idx) > vq->num)) {
1364                 vq_err(vq, "Invalid available index change from %u to %u",
1365                        vq->last_avail_idx, vq->avail_idx);
1366                 return -EINVAL;
1367         }
1368
1369         /* We're done if there is nothing new */
1370         if (vq->avail_idx == vq->last_avail_idx)
1371                 return 0;
1372
1373         /*
1374          * We updated vq->avail_idx so we need a memory barrier between
1375          * the index read above and the caller reading avail ring entries.
1376          */
1377         smp_rmb();
1378         return 1;
1379 }
1380
1381 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1382                                        __virtio16 *head, int idx)
1383 {
1384         return vhost_get_avail(vq, *head,
1385                                &vq->avail->ring[idx & (vq->num - 1)]);
1386 }
1387
1388 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1389                                         __virtio16 *flags)
1390 {
1391         return vhost_get_avail(vq, *flags, &vq->avail->flags);
1392 }
1393
1394 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1395                                        __virtio16 *event)
1396 {
1397         return vhost_get_avail(vq, *event, vhost_used_event(vq));
1398 }
1399
1400 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1401                                      __virtio16 *idx)
1402 {
1403         return vhost_get_used(vq, *idx, &vq->used->idx);
1404 }
1405
1406 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1407                                  struct vring_desc *desc, int idx)
1408 {
1409         return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1410 }
1411
1412 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1413                                   struct vhost_iotlb_msg *msg)
1414 {
1415         struct vhost_msg_node *node, *n;
1416
1417         spin_lock(&d->iotlb_lock);
1418
1419         list_for_each_entry_safe(node, n, &d->pending_list, node) {
1420                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1421                 if (msg->iova <= vq_msg->iova &&
1422                     msg->iova + msg->size - 1 >= vq_msg->iova &&
1423                     vq_msg->type == VHOST_IOTLB_MISS) {
1424                         vhost_poll_queue(&node->vq->poll);
1425                         list_del(&node->node);
1426                         kfree(node);
1427                 }
1428         }
1429
1430         spin_unlock(&d->iotlb_lock);
1431 }
1432
1433 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1434 {
1435         unsigned long a = uaddr;
1436
1437         /* Make sure 64 bit math will not overflow. */
1438         if (vhost_overflow(uaddr, size))
1439                 return false;
1440
1441         if ((access & VHOST_ACCESS_RO) &&
1442             !access_ok((void __user *)a, size))
1443                 return false;
1444         if ((access & VHOST_ACCESS_WO) &&
1445             !access_ok((void __user *)a, size))
1446                 return false;
1447         return true;
1448 }
1449
1450 static int vhost_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1451                                    struct vhost_iotlb_msg *msg)
1452 {
1453         int ret = 0;
1454
1455         if (asid != 0)
1456                 return -EINVAL;
1457
1458         mutex_lock(&dev->mutex);
1459         vhost_dev_lock_vqs(dev);
1460         switch (msg->type) {
1461         case VHOST_IOTLB_UPDATE:
1462                 if (!dev->iotlb) {
1463                         ret = -EFAULT;
1464                         break;
1465                 }
1466                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1467                         ret = -EFAULT;
1468                         break;
1469                 }
1470                 vhost_vq_meta_reset(dev);
1471                 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1472                                           msg->iova + msg->size - 1,
1473                                           msg->uaddr, msg->perm)) {
1474                         ret = -ENOMEM;
1475                         break;
1476                 }
1477                 vhost_iotlb_notify_vq(dev, msg);
1478                 break;
1479         case VHOST_IOTLB_INVALIDATE:
1480                 if (!dev->iotlb) {
1481                         ret = -EFAULT;
1482                         break;
1483                 }
1484                 vhost_vq_meta_reset(dev);
1485                 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1486                                       msg->iova + msg->size - 1);
1487                 break;
1488         default:
1489                 ret = -EINVAL;
1490                 break;
1491         }
1492
1493         vhost_dev_unlock_vqs(dev);
1494         mutex_unlock(&dev->mutex);
1495
1496         return ret;
1497 }
1498 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1499                              struct iov_iter *from)
1500 {
1501         struct vhost_iotlb_msg msg;
1502         size_t offset;
1503         int type, ret;
1504         u32 asid = 0;
1505
1506         ret = copy_from_iter(&type, sizeof(type), from);
1507         if (ret != sizeof(type)) {
1508                 ret = -EINVAL;
1509                 goto done;
1510         }
1511
1512         switch (type) {
1513         case VHOST_IOTLB_MSG:
1514                 /* There maybe a hole after type for V1 message type,
1515                  * so skip it here.
1516                  */
1517                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1518                 break;
1519         case VHOST_IOTLB_MSG_V2:
1520                 if (vhost_backend_has_feature(dev->vqs[0],
1521                                               VHOST_BACKEND_F_IOTLB_ASID)) {
1522                         ret = copy_from_iter(&asid, sizeof(asid), from);
1523                         if (ret != sizeof(asid)) {
1524                                 ret = -EINVAL;
1525                                 goto done;
1526                         }
1527                         offset = 0;
1528                 } else
1529                         offset = sizeof(__u32);
1530                 break;
1531         default:
1532                 ret = -EINVAL;
1533                 goto done;
1534         }
1535
1536         iov_iter_advance(from, offset);
1537         ret = copy_from_iter(&msg, sizeof(msg), from);
1538         if (ret != sizeof(msg)) {
1539                 ret = -EINVAL;
1540                 goto done;
1541         }
1542
1543         if (msg.type == VHOST_IOTLB_UPDATE && msg.size == 0) {
1544                 ret = -EINVAL;
1545                 goto done;
1546         }
1547
1548         if (dev->msg_handler)
1549                 ret = dev->msg_handler(dev, asid, &msg);
1550         else
1551                 ret = vhost_process_iotlb_msg(dev, asid, &msg);
1552         if (ret) {
1553                 ret = -EFAULT;
1554                 goto done;
1555         }
1556
1557         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1558               sizeof(struct vhost_msg_v2);
1559 done:
1560         return ret;
1561 }
1562 EXPORT_SYMBOL(vhost_chr_write_iter);
1563
1564 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1565                             poll_table *wait)
1566 {
1567         __poll_t mask = 0;
1568
1569         poll_wait(file, &dev->wait, wait);
1570
1571         if (!list_empty(&dev->read_list))
1572                 mask |= EPOLLIN | EPOLLRDNORM;
1573
1574         return mask;
1575 }
1576 EXPORT_SYMBOL(vhost_chr_poll);
1577
1578 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1579                             int noblock)
1580 {
1581         DEFINE_WAIT(wait);
1582         struct vhost_msg_node *node;
1583         ssize_t ret = 0;
1584         unsigned size = sizeof(struct vhost_msg);
1585
1586         if (iov_iter_count(to) < size)
1587                 return 0;
1588
1589         while (1) {
1590                 if (!noblock)
1591                         prepare_to_wait(&dev->wait, &wait,
1592                                         TASK_INTERRUPTIBLE);
1593
1594                 node = vhost_dequeue_msg(dev, &dev->read_list);
1595                 if (node)
1596                         break;
1597                 if (noblock) {
1598                         ret = -EAGAIN;
1599                         break;
1600                 }
1601                 if (signal_pending(current)) {
1602                         ret = -ERESTARTSYS;
1603                         break;
1604                 }
1605                 if (!dev->iotlb) {
1606                         ret = -EBADFD;
1607                         break;
1608                 }
1609
1610                 schedule();
1611         }
1612
1613         if (!noblock)
1614                 finish_wait(&dev->wait, &wait);
1615
1616         if (node) {
1617                 struct vhost_iotlb_msg *msg;
1618                 void *start = &node->msg;
1619
1620                 switch (node->msg.type) {
1621                 case VHOST_IOTLB_MSG:
1622                         size = sizeof(node->msg);
1623                         msg = &node->msg.iotlb;
1624                         break;
1625                 case VHOST_IOTLB_MSG_V2:
1626                         size = sizeof(node->msg_v2);
1627                         msg = &node->msg_v2.iotlb;
1628                         break;
1629                 default:
1630                         BUG();
1631                         break;
1632                 }
1633
1634                 ret = copy_to_iter(start, size, to);
1635                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1636                         kfree(node);
1637                         return ret;
1638                 }
1639                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1640         }
1641
1642         return ret;
1643 }
1644 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1645
1646 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1647 {
1648         struct vhost_dev *dev = vq->dev;
1649         struct vhost_msg_node *node;
1650         struct vhost_iotlb_msg *msg;
1651         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1652
1653         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1654         if (!node)
1655                 return -ENOMEM;
1656
1657         if (v2) {
1658                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1659                 msg = &node->msg_v2.iotlb;
1660         } else {
1661                 msg = &node->msg.iotlb;
1662         }
1663
1664         msg->type = VHOST_IOTLB_MISS;
1665         msg->iova = iova;
1666         msg->perm = access;
1667
1668         vhost_enqueue_msg(dev, &dev->read_list, node);
1669
1670         return 0;
1671 }
1672
1673 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1674                          vring_desc_t __user *desc,
1675                          vring_avail_t __user *avail,
1676                          vring_used_t __user *used)
1677
1678 {
1679         /* If an IOTLB device is present, the vring addresses are
1680          * GIOVAs. Access validation occurs at prefetch time. */
1681         if (vq->iotlb)
1682                 return true;
1683
1684         return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1685                access_ok(avail, vhost_get_avail_size(vq, num)) &&
1686                access_ok(used, vhost_get_used_size(vq, num));
1687 }
1688
1689 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1690                                  const struct vhost_iotlb_map *map,
1691                                  int type)
1692 {
1693         int access = (type == VHOST_ADDR_USED) ?
1694                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1695
1696         if (likely(map->perm & access))
1697                 vq->meta_iotlb[type] = map;
1698 }
1699
1700 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1701                             int access, u64 addr, u64 len, int type)
1702 {
1703         const struct vhost_iotlb_map *map;
1704         struct vhost_iotlb *umem = vq->iotlb;
1705         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1706
1707         if (vhost_vq_meta_fetch(vq, addr, len, type))
1708                 return true;
1709
1710         while (len > s) {
1711                 map = vhost_iotlb_itree_first(umem, addr, last);
1712                 if (map == NULL || map->start > addr) {
1713                         vhost_iotlb_miss(vq, addr, access);
1714                         return false;
1715                 } else if (!(map->perm & access)) {
1716                         /* Report the possible access violation by
1717                          * request another translation from userspace.
1718                          */
1719                         return false;
1720                 }
1721
1722                 size = map->size - addr + map->start;
1723
1724                 if (orig_addr == addr && size >= len)
1725                         vhost_vq_meta_update(vq, map, type);
1726
1727                 s += size;
1728                 addr += size;
1729         }
1730
1731         return true;
1732 }
1733
1734 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1735 {
1736         unsigned int num = vq->num;
1737
1738         if (!vq->iotlb)
1739                 return 1;
1740
1741         return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1742                                vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1743                iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1744                                vhost_get_avail_size(vq, num),
1745                                VHOST_ADDR_AVAIL) &&
1746                iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1747                                vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1748 }
1749 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1750
1751 /* Can we log writes? */
1752 /* Caller should have device mutex but not vq mutex */
1753 bool vhost_log_access_ok(struct vhost_dev *dev)
1754 {
1755         return memory_access_ok(dev, dev->umem, 1);
1756 }
1757 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1758
1759 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1760                                   void __user *log_base,
1761                                   bool log_used,
1762                                   u64 log_addr)
1763 {
1764         /* If an IOTLB device is present, log_addr is a GIOVA that
1765          * will never be logged by log_used(). */
1766         if (vq->iotlb)
1767                 return true;
1768
1769         return !log_used || log_access_ok(log_base, log_addr,
1770                                           vhost_get_used_size(vq, vq->num));
1771 }
1772
1773 /* Verify access for write logging. */
1774 /* Caller should have vq mutex and device mutex */
1775 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1776                              void __user *log_base)
1777 {
1778         return vq_memory_access_ok(log_base, vq->umem,
1779                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1780                 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1781 }
1782
1783 /* Can we start vq? */
1784 /* Caller should have vq mutex and device mutex */
1785 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1786 {
1787         if (!vq_log_access_ok(vq, vq->log_base))
1788                 return false;
1789
1790         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1791 }
1792 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1793
1794 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1795 {
1796         struct vhost_memory mem, *newmem;
1797         struct vhost_memory_region *region;
1798         struct vhost_iotlb *newumem, *oldumem;
1799         unsigned long size = offsetof(struct vhost_memory, regions);
1800         int i;
1801
1802         if (copy_from_user(&mem, m, size))
1803                 return -EFAULT;
1804         if (mem.padding)
1805                 return -EOPNOTSUPP;
1806         if (mem.nregions > max_mem_regions)
1807                 return -E2BIG;
1808         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1809                         GFP_KERNEL);
1810         if (!newmem)
1811                 return -ENOMEM;
1812
1813         memcpy(newmem, &mem, size);
1814         if (copy_from_user(newmem->regions, m->regions,
1815                            flex_array_size(newmem, regions, mem.nregions))) {
1816                 kvfree(newmem);
1817                 return -EFAULT;
1818         }
1819
1820         newumem = iotlb_alloc();
1821         if (!newumem) {
1822                 kvfree(newmem);
1823                 return -ENOMEM;
1824         }
1825
1826         for (region = newmem->regions;
1827              region < newmem->regions + mem.nregions;
1828              region++) {
1829                 if (vhost_iotlb_add_range(newumem,
1830                                           region->guest_phys_addr,
1831                                           region->guest_phys_addr +
1832                                           region->memory_size - 1,
1833                                           region->userspace_addr,
1834                                           VHOST_MAP_RW))
1835                         goto err;
1836         }
1837
1838         if (!memory_access_ok(d, newumem, 0))
1839                 goto err;
1840
1841         oldumem = d->umem;
1842         d->umem = newumem;
1843
1844         /* All memory accesses are done under some VQ mutex. */
1845         for (i = 0; i < d->nvqs; ++i) {
1846                 mutex_lock(&d->vqs[i]->mutex);
1847                 d->vqs[i]->umem = newumem;
1848                 mutex_unlock(&d->vqs[i]->mutex);
1849         }
1850
1851         kvfree(newmem);
1852         vhost_iotlb_free(oldumem);
1853         return 0;
1854
1855 err:
1856         vhost_iotlb_free(newumem);
1857         kvfree(newmem);
1858         return -EFAULT;
1859 }
1860
1861 static long vhost_vring_set_num(struct vhost_dev *d,
1862                                 struct vhost_virtqueue *vq,
1863                                 void __user *argp)
1864 {
1865         struct vhost_vring_state s;
1866
1867         /* Resizing ring with an active backend?
1868          * You don't want to do that. */
1869         if (vq->private_data)
1870                 return -EBUSY;
1871
1872         if (copy_from_user(&s, argp, sizeof s))
1873                 return -EFAULT;
1874
1875         if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1876                 return -EINVAL;
1877         vq->num = s.num;
1878
1879         return 0;
1880 }
1881
1882 static long vhost_vring_set_addr(struct vhost_dev *d,
1883                                  struct vhost_virtqueue *vq,
1884                                  void __user *argp)
1885 {
1886         struct vhost_vring_addr a;
1887
1888         if (copy_from_user(&a, argp, sizeof a))
1889                 return -EFAULT;
1890         if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1891                 return -EOPNOTSUPP;
1892
1893         /* For 32bit, verify that the top 32bits of the user
1894            data are set to zero. */
1895         if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1896             (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1897             (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1898                 return -EFAULT;
1899
1900         /* Make sure it's safe to cast pointers to vring types. */
1901         BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1902         BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1903         if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1904             (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1905             (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1906                 return -EINVAL;
1907
1908         /* We only verify access here if backend is configured.
1909          * If it is not, we don't as size might not have been setup.
1910          * We will verify when backend is configured. */
1911         if (vq->private_data) {
1912                 if (!vq_access_ok(vq, vq->num,
1913                         (void __user *)(unsigned long)a.desc_user_addr,
1914                         (void __user *)(unsigned long)a.avail_user_addr,
1915                         (void __user *)(unsigned long)a.used_user_addr))
1916                         return -EINVAL;
1917
1918                 /* Also validate log access for used ring if enabled. */
1919                 if (!vq_log_used_access_ok(vq, vq->log_base,
1920                                 a.flags & (0x1 << VHOST_VRING_F_LOG),
1921                                 a.log_guest_addr))
1922                         return -EINVAL;
1923         }
1924
1925         vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1926         vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1927         vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1928         vq->log_addr = a.log_guest_addr;
1929         vq->used = (void __user *)(unsigned long)a.used_user_addr;
1930
1931         return 0;
1932 }
1933
1934 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1935                                      struct vhost_virtqueue *vq,
1936                                      unsigned int ioctl,
1937                                      void __user *argp)
1938 {
1939         long r;
1940
1941         mutex_lock(&vq->mutex);
1942
1943         switch (ioctl) {
1944         case VHOST_SET_VRING_NUM:
1945                 r = vhost_vring_set_num(d, vq, argp);
1946                 break;
1947         case VHOST_SET_VRING_ADDR:
1948                 r = vhost_vring_set_addr(d, vq, argp);
1949                 break;
1950         default:
1951                 BUG();
1952         }
1953
1954         mutex_unlock(&vq->mutex);
1955
1956         return r;
1957 }
1958 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1959 {
1960         struct file *eventfp, *filep = NULL;
1961         bool pollstart = false, pollstop = false;
1962         struct eventfd_ctx *ctx = NULL;
1963         struct vhost_virtqueue *vq;
1964         struct vhost_vring_state s;
1965         struct vhost_vring_file f;
1966         u32 idx;
1967         long r;
1968
1969         r = vhost_get_vq_from_user(d, argp, &vq, &idx);
1970         if (r < 0)
1971                 return r;
1972
1973         if (ioctl == VHOST_SET_VRING_NUM ||
1974             ioctl == VHOST_SET_VRING_ADDR) {
1975                 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1976         }
1977
1978         mutex_lock(&vq->mutex);
1979
1980         switch (ioctl) {
1981         case VHOST_SET_VRING_BASE:
1982                 /* Moving base with an active backend?
1983                  * You don't want to do that. */
1984                 if (vq->private_data) {
1985                         r = -EBUSY;
1986                         break;
1987                 }
1988                 if (copy_from_user(&s, argp, sizeof s)) {
1989                         r = -EFAULT;
1990                         break;
1991                 }
1992                 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
1993                         vq->last_avail_idx = s.num & 0xffff;
1994                         vq->last_used_idx = (s.num >> 16) & 0xffff;
1995                 } else {
1996                         if (s.num > 0xffff) {
1997                                 r = -EINVAL;
1998                                 break;
1999                         }
2000                         vq->last_avail_idx = s.num;
2001                 }
2002                 /* Forget the cached index value. */
2003                 vq->avail_idx = vq->last_avail_idx;
2004                 break;
2005         case VHOST_GET_VRING_BASE:
2006                 s.index = idx;
2007                 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED))
2008                         s.num = (u32)vq->last_avail_idx | ((u32)vq->last_used_idx << 16);
2009                 else
2010                         s.num = vq->last_avail_idx;
2011                 if (copy_to_user(argp, &s, sizeof s))
2012                         r = -EFAULT;
2013                 break;
2014         case VHOST_SET_VRING_KICK:
2015                 if (copy_from_user(&f, argp, sizeof f)) {
2016                         r = -EFAULT;
2017                         break;
2018                 }
2019                 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
2020                 if (IS_ERR(eventfp)) {
2021                         r = PTR_ERR(eventfp);
2022                         break;
2023                 }
2024                 if (eventfp != vq->kick) {
2025                         pollstop = (filep = vq->kick) != NULL;
2026                         pollstart = (vq->kick = eventfp) != NULL;
2027                 } else
2028                         filep = eventfp;
2029                 break;
2030         case VHOST_SET_VRING_CALL:
2031                 if (copy_from_user(&f, argp, sizeof f)) {
2032                         r = -EFAULT;
2033                         break;
2034                 }
2035                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
2036                 if (IS_ERR(ctx)) {
2037                         r = PTR_ERR(ctx);
2038                         break;
2039                 }
2040
2041                 swap(ctx, vq->call_ctx.ctx);
2042                 break;
2043         case VHOST_SET_VRING_ERR:
2044                 if (copy_from_user(&f, argp, sizeof f)) {
2045                         r = -EFAULT;
2046                         break;
2047                 }
2048                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
2049                 if (IS_ERR(ctx)) {
2050                         r = PTR_ERR(ctx);
2051                         break;
2052                 }
2053                 swap(ctx, vq->error_ctx);
2054                 break;
2055         case VHOST_SET_VRING_ENDIAN:
2056                 r = vhost_set_vring_endian(vq, argp);
2057                 break;
2058         case VHOST_GET_VRING_ENDIAN:
2059                 r = vhost_get_vring_endian(vq, idx, argp);
2060                 break;
2061         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
2062                 if (copy_from_user(&s, argp, sizeof(s))) {
2063                         r = -EFAULT;
2064                         break;
2065                 }
2066                 vq->busyloop_timeout = s.num;
2067                 break;
2068         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
2069                 s.index = idx;
2070                 s.num = vq->busyloop_timeout;
2071                 if (copy_to_user(argp, &s, sizeof(s)))
2072                         r = -EFAULT;
2073                 break;
2074         default:
2075                 r = -ENOIOCTLCMD;
2076         }
2077
2078         if (pollstop && vq->handle_kick)
2079                 vhost_poll_stop(&vq->poll);
2080
2081         if (!IS_ERR_OR_NULL(ctx))
2082                 eventfd_ctx_put(ctx);
2083         if (filep)
2084                 fput(filep);
2085
2086         if (pollstart && vq->handle_kick)
2087                 r = vhost_poll_start(&vq->poll, vq->kick);
2088
2089         mutex_unlock(&vq->mutex);
2090
2091         if (pollstop && vq->handle_kick)
2092                 vhost_dev_flush(vq->poll.dev);
2093         return r;
2094 }
2095 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
2096
2097 int vhost_init_device_iotlb(struct vhost_dev *d)
2098 {
2099         struct vhost_iotlb *niotlb, *oiotlb;
2100         int i;
2101
2102         niotlb = iotlb_alloc();
2103         if (!niotlb)
2104                 return -ENOMEM;
2105
2106         oiotlb = d->iotlb;
2107         d->iotlb = niotlb;
2108
2109         for (i = 0; i < d->nvqs; ++i) {
2110                 struct vhost_virtqueue *vq = d->vqs[i];
2111
2112                 mutex_lock(&vq->mutex);
2113                 vq->iotlb = niotlb;
2114                 __vhost_vq_meta_reset(vq);
2115                 mutex_unlock(&vq->mutex);
2116         }
2117
2118         vhost_iotlb_free(oiotlb);
2119
2120         return 0;
2121 }
2122 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
2123
2124 /* Caller must have device mutex */
2125 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
2126 {
2127         struct eventfd_ctx *ctx;
2128         u64 p;
2129         long r;
2130         int i, fd;
2131
2132         /* If you are not the owner, you can become one */
2133         if (ioctl == VHOST_SET_OWNER) {
2134                 r = vhost_dev_set_owner(d);
2135                 goto done;
2136         }
2137
2138         /* You must be the owner to do anything else */
2139         r = vhost_dev_check_owner(d);
2140         if (r)
2141                 goto done;
2142
2143         switch (ioctl) {
2144         case VHOST_SET_MEM_TABLE:
2145                 r = vhost_set_memory(d, argp);
2146                 break;
2147         case VHOST_SET_LOG_BASE:
2148                 if (copy_from_user(&p, argp, sizeof p)) {
2149                         r = -EFAULT;
2150                         break;
2151                 }
2152                 if ((u64)(unsigned long)p != p) {
2153                         r = -EFAULT;
2154                         break;
2155                 }
2156                 for (i = 0; i < d->nvqs; ++i) {
2157                         struct vhost_virtqueue *vq;
2158                         void __user *base = (void __user *)(unsigned long)p;
2159                         vq = d->vqs[i];
2160                         mutex_lock(&vq->mutex);
2161                         /* If ring is inactive, will check when it's enabled. */
2162                         if (vq->private_data && !vq_log_access_ok(vq, base))
2163                                 r = -EFAULT;
2164                         else
2165                                 vq->log_base = base;
2166                         mutex_unlock(&vq->mutex);
2167                 }
2168                 break;
2169         case VHOST_SET_LOG_FD:
2170                 r = get_user(fd, (int __user *)argp);
2171                 if (r < 0)
2172                         break;
2173                 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
2174                 if (IS_ERR(ctx)) {
2175                         r = PTR_ERR(ctx);
2176                         break;
2177                 }
2178                 swap(ctx, d->log_ctx);
2179                 for (i = 0; i < d->nvqs; ++i) {
2180                         mutex_lock(&d->vqs[i]->mutex);
2181                         d->vqs[i]->log_ctx = d->log_ctx;
2182                         mutex_unlock(&d->vqs[i]->mutex);
2183                 }
2184                 if (ctx)
2185                         eventfd_ctx_put(ctx);
2186                 break;
2187         default:
2188                 r = -ENOIOCTLCMD;
2189                 break;
2190         }
2191 done:
2192         return r;
2193 }
2194 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
2195
2196 /* TODO: This is really inefficient.  We need something like get_user()
2197  * (instruction directly accesses the data, with an exception table entry
2198  * returning -EFAULT). See Documentation/arch/x86/exception-tables.rst.
2199  */
2200 static int set_bit_to_user(int nr, void __user *addr)
2201 {
2202         unsigned long log = (unsigned long)addr;
2203         struct page *page;
2204         void *base;
2205         int bit = nr + (log % PAGE_SIZE) * 8;
2206         int r;
2207
2208         r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
2209         if (r < 0)
2210                 return r;
2211         BUG_ON(r != 1);
2212         base = kmap_atomic(page);
2213         set_bit(bit, base);
2214         kunmap_atomic(base);
2215         unpin_user_pages_dirty_lock(&page, 1, true);
2216         return 0;
2217 }
2218
2219 static int log_write(void __user *log_base,
2220                      u64 write_address, u64 write_length)
2221 {
2222         u64 write_page = write_address / VHOST_PAGE_SIZE;
2223         int r;
2224
2225         if (!write_length)
2226                 return 0;
2227         write_length += write_address % VHOST_PAGE_SIZE;
2228         for (;;) {
2229                 u64 base = (u64)(unsigned long)log_base;
2230                 u64 log = base + write_page / 8;
2231                 int bit = write_page % 8;
2232                 if ((u64)(unsigned long)log != log)
2233                         return -EFAULT;
2234                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
2235                 if (r < 0)
2236                         return r;
2237                 if (write_length <= VHOST_PAGE_SIZE)
2238                         break;
2239                 write_length -= VHOST_PAGE_SIZE;
2240                 write_page += 1;
2241         }
2242         return r;
2243 }
2244
2245 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
2246 {
2247         struct vhost_iotlb *umem = vq->umem;
2248         struct vhost_iotlb_map *u;
2249         u64 start, end, l, min;
2250         int r;
2251         bool hit = false;
2252
2253         while (len) {
2254                 min = len;
2255                 /* More than one GPAs can be mapped into a single HVA. So
2256                  * iterate all possible umems here to be safe.
2257                  */
2258                 list_for_each_entry(u, &umem->list, link) {
2259                         if (u->addr > hva - 1 + len ||
2260                             u->addr - 1 + u->size < hva)
2261                                 continue;
2262                         start = max(u->addr, hva);
2263                         end = min(u->addr - 1 + u->size, hva - 1 + len);
2264                         l = end - start + 1;
2265                         r = log_write(vq->log_base,
2266                                       u->start + start - u->addr,
2267                                       l);
2268                         if (r < 0)
2269                                 return r;
2270                         hit = true;
2271                         min = min(l, min);
2272                 }
2273
2274                 if (!hit)
2275                         return -EFAULT;
2276
2277                 len -= min;
2278                 hva += min;
2279         }
2280
2281         return 0;
2282 }
2283
2284 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
2285 {
2286         struct iovec *iov = vq->log_iov;
2287         int i, ret;
2288
2289         if (!vq->iotlb)
2290                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
2291
2292         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
2293                              len, iov, 64, VHOST_ACCESS_WO);
2294         if (ret < 0)
2295                 return ret;
2296
2297         for (i = 0; i < ret; i++) {
2298                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
2299                                     iov[i].iov_len);
2300                 if (ret)
2301                         return ret;
2302         }
2303
2304         return 0;
2305 }
2306
2307 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
2308                     unsigned int log_num, u64 len, struct iovec *iov, int count)
2309 {
2310         int i, r;
2311
2312         /* Make sure data written is seen before log. */
2313         smp_wmb();
2314
2315         if (vq->iotlb) {
2316                 for (i = 0; i < count; i++) {
2317                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
2318                                           iov[i].iov_len);
2319                         if (r < 0)
2320                                 return r;
2321                 }
2322                 return 0;
2323         }
2324
2325         for (i = 0; i < log_num; ++i) {
2326                 u64 l = min(log[i].len, len);
2327                 r = log_write(vq->log_base, log[i].addr, l);
2328                 if (r < 0)
2329                         return r;
2330                 len -= l;
2331                 if (!len) {
2332                         if (vq->log_ctx)
2333                                 eventfd_signal(vq->log_ctx);
2334                         return 0;
2335                 }
2336         }
2337         /* Length written exceeds what we have stored. This is a bug. */
2338         BUG();
2339         return 0;
2340 }
2341 EXPORT_SYMBOL_GPL(vhost_log_write);
2342
2343 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
2344 {
2345         void __user *used;
2346         if (vhost_put_used_flags(vq))
2347                 return -EFAULT;
2348         if (unlikely(vq->log_used)) {
2349                 /* Make sure the flag is seen before log. */
2350                 smp_wmb();
2351                 /* Log used flag write. */
2352                 used = &vq->used->flags;
2353                 log_used(vq, (used - (void __user *)vq->used),
2354                          sizeof vq->used->flags);
2355                 if (vq->log_ctx)
2356                         eventfd_signal(vq->log_ctx);
2357         }
2358         return 0;
2359 }
2360
2361 static int vhost_update_avail_event(struct vhost_virtqueue *vq)
2362 {
2363         if (vhost_put_avail_event(vq))
2364                 return -EFAULT;
2365         if (unlikely(vq->log_used)) {
2366                 void __user *used;
2367                 /* Make sure the event is seen before log. */
2368                 smp_wmb();
2369                 /* Log avail event write */
2370                 used = vhost_avail_event(vq);
2371                 log_used(vq, (used - (void __user *)vq->used),
2372                          sizeof *vhost_avail_event(vq));
2373                 if (vq->log_ctx)
2374                         eventfd_signal(vq->log_ctx);
2375         }
2376         return 0;
2377 }
2378
2379 int vhost_vq_init_access(struct vhost_virtqueue *vq)
2380 {
2381         __virtio16 last_used_idx;
2382         int r;
2383         bool is_le = vq->is_le;
2384
2385         if (!vq->private_data)
2386                 return 0;
2387
2388         vhost_init_is_le(vq);
2389
2390         r = vhost_update_used_flags(vq);
2391         if (r)
2392                 goto err;
2393         vq->signalled_used_valid = false;
2394         if (!vq->iotlb &&
2395             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2396                 r = -EFAULT;
2397                 goto err;
2398         }
2399         r = vhost_get_used_idx(vq, &last_used_idx);
2400         if (r) {
2401                 vq_err(vq, "Can't access used idx at %p\n",
2402                        &vq->used->idx);
2403                 goto err;
2404         }
2405         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2406         return 0;
2407
2408 err:
2409         vq->is_le = is_le;
2410         return r;
2411 }
2412 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2413
2414 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2415                           struct iovec iov[], int iov_size, int access)
2416 {
2417         const struct vhost_iotlb_map *map;
2418         struct vhost_dev *dev = vq->dev;
2419         struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2420         struct iovec *_iov;
2421         u64 s = 0, last = addr + len - 1;
2422         int ret = 0;
2423
2424         while ((u64)len > s) {
2425                 u64 size;
2426                 if (unlikely(ret >= iov_size)) {
2427                         ret = -ENOBUFS;
2428                         break;
2429                 }
2430
2431                 map = vhost_iotlb_itree_first(umem, addr, last);
2432                 if (map == NULL || map->start > addr) {
2433                         if (umem != dev->iotlb) {
2434                                 ret = -EFAULT;
2435                                 break;
2436                         }
2437                         ret = -EAGAIN;
2438                         break;
2439                 } else if (!(map->perm & access)) {
2440                         ret = -EPERM;
2441                         break;
2442                 }
2443
2444                 _iov = iov + ret;
2445                 size = map->size - addr + map->start;
2446                 _iov->iov_len = min((u64)len - s, size);
2447                 _iov->iov_base = (void __user *)(unsigned long)
2448                                  (map->addr + addr - map->start);
2449                 s += size;
2450                 addr += size;
2451                 ++ret;
2452         }
2453
2454         if (ret == -EAGAIN)
2455                 vhost_iotlb_miss(vq, addr, access);
2456         return ret;
2457 }
2458
2459 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
2460  * function returns the next descriptor in the chain,
2461  * or -1U if we're at the end. */
2462 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2463 {
2464         unsigned int next;
2465
2466         /* If this descriptor says it doesn't chain, we're done. */
2467         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2468                 return -1U;
2469
2470         /* Check they're not leading us off end of descriptors. */
2471         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2472         return next;
2473 }
2474
2475 static int get_indirect(struct vhost_virtqueue *vq,
2476                         struct iovec iov[], unsigned int iov_size,
2477                         unsigned int *out_num, unsigned int *in_num,
2478                         struct vhost_log *log, unsigned int *log_num,
2479                         struct vring_desc *indirect)
2480 {
2481         struct vring_desc desc;
2482         unsigned int i = 0, count, found = 0;
2483         u32 len = vhost32_to_cpu(vq, indirect->len);
2484         struct iov_iter from;
2485         int ret, access;
2486
2487         /* Sanity check */
2488         if (unlikely(len % sizeof desc)) {
2489                 vq_err(vq, "Invalid length in indirect descriptor: "
2490                        "len 0x%llx not multiple of 0x%zx\n",
2491                        (unsigned long long)len,
2492                        sizeof desc);
2493                 return -EINVAL;
2494         }
2495
2496         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2497                              UIO_MAXIOV, VHOST_ACCESS_RO);
2498         if (unlikely(ret < 0)) {
2499                 if (ret != -EAGAIN)
2500                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2501                 return ret;
2502         }
2503         iov_iter_init(&from, ITER_SOURCE, vq->indirect, ret, len);
2504         count = len / sizeof desc;
2505         /* Buffers are chained via a 16 bit next field, so
2506          * we can have at most 2^16 of these. */
2507         if (unlikely(count > USHRT_MAX + 1)) {
2508                 vq_err(vq, "Indirect buffer length too big: %d\n",
2509                        indirect->len);
2510                 return -E2BIG;
2511         }
2512
2513         do {
2514                 unsigned iov_count = *in_num + *out_num;
2515                 if (unlikely(++found > count)) {
2516                         vq_err(vq, "Loop detected: last one at %u "
2517                                "indirect size %u\n",
2518                                i, count);
2519                         return -EINVAL;
2520                 }
2521                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2522                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2523                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2524                         return -EINVAL;
2525                 }
2526                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2527                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2528                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2529                         return -EINVAL;
2530                 }
2531
2532                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2533                         access = VHOST_ACCESS_WO;
2534                 else
2535                         access = VHOST_ACCESS_RO;
2536
2537                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2538                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2539                                      iov_size - iov_count, access);
2540                 if (unlikely(ret < 0)) {
2541                         if (ret != -EAGAIN)
2542                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2543                                         ret, i);
2544                         return ret;
2545                 }
2546                 /* If this is an input descriptor, increment that count. */
2547                 if (access == VHOST_ACCESS_WO) {
2548                         *in_num += ret;
2549                         if (unlikely(log && ret)) {
2550                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2551                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2552                                 ++*log_num;
2553                         }
2554                 } else {
2555                         /* If it's an output descriptor, they're all supposed
2556                          * to come before any input descriptors. */
2557                         if (unlikely(*in_num)) {
2558                                 vq_err(vq, "Indirect descriptor "
2559                                        "has out after in: idx %d\n", i);
2560                                 return -EINVAL;
2561                         }
2562                         *out_num += ret;
2563                 }
2564         } while ((i = next_desc(vq, &desc)) != -1);
2565         return 0;
2566 }
2567
2568 /* This looks in the virtqueue and for the first available buffer, and converts
2569  * it to an iovec for convenient access.  Since descriptors consist of some
2570  * number of output then some number of input descriptors, it's actually two
2571  * iovecs, but we pack them into one and note how many of each there were.
2572  *
2573  * This function returns the descriptor number found, or vq->num (which is
2574  * never a valid descriptor number) if none was found.  A negative code is
2575  * returned on error. */
2576 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2577                       struct iovec iov[], unsigned int iov_size,
2578                       unsigned int *out_num, unsigned int *in_num,
2579                       struct vhost_log *log, unsigned int *log_num)
2580 {
2581         struct vring_desc desc;
2582         unsigned int i, head, found = 0;
2583         u16 last_avail_idx = vq->last_avail_idx;
2584         __virtio16 ring_head;
2585         int ret, access;
2586
2587         if (vq->avail_idx == vq->last_avail_idx) {
2588                 ret = vhost_get_avail_idx(vq);
2589                 if (unlikely(ret < 0))
2590                         return ret;
2591
2592                 if (!ret)
2593                         return vq->num;
2594         }
2595
2596         /* Grab the next descriptor number they're advertising, and increment
2597          * the index we've seen. */
2598         if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2599                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2600                        last_avail_idx,
2601                        &vq->avail->ring[last_avail_idx % vq->num]);
2602                 return -EFAULT;
2603         }
2604
2605         head = vhost16_to_cpu(vq, ring_head);
2606
2607         /* If their number is silly, that's an error. */
2608         if (unlikely(head >= vq->num)) {
2609                 vq_err(vq, "Guest says index %u > %u is available",
2610                        head, vq->num);
2611                 return -EINVAL;
2612         }
2613
2614         /* When we start there are none of either input nor output. */
2615         *out_num = *in_num = 0;
2616         if (unlikely(log))
2617                 *log_num = 0;
2618
2619         i = head;
2620         do {
2621                 unsigned iov_count = *in_num + *out_num;
2622                 if (unlikely(i >= vq->num)) {
2623                         vq_err(vq, "Desc index is %u > %u, head = %u",
2624                                i, vq->num, head);
2625                         return -EINVAL;
2626                 }
2627                 if (unlikely(++found > vq->num)) {
2628                         vq_err(vq, "Loop detected: last one at %u "
2629                                "vq size %u head %u\n",
2630                                i, vq->num, head);
2631                         return -EINVAL;
2632                 }
2633                 ret = vhost_get_desc(vq, &desc, i);
2634                 if (unlikely(ret)) {
2635                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2636                                i, vq->desc + i);
2637                         return -EFAULT;
2638                 }
2639                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2640                         ret = get_indirect(vq, iov, iov_size,
2641                                            out_num, in_num,
2642                                            log, log_num, &desc);
2643                         if (unlikely(ret < 0)) {
2644                                 if (ret != -EAGAIN)
2645                                         vq_err(vq, "Failure detected "
2646                                                 "in indirect descriptor at idx %d\n", i);
2647                                 return ret;
2648                         }
2649                         continue;
2650                 }
2651
2652                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2653                         access = VHOST_ACCESS_WO;
2654                 else
2655                         access = VHOST_ACCESS_RO;
2656                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2657                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2658                                      iov_size - iov_count, access);
2659                 if (unlikely(ret < 0)) {
2660                         if (ret != -EAGAIN)
2661                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2662                                         ret, i);
2663                         return ret;
2664                 }
2665                 if (access == VHOST_ACCESS_WO) {
2666                         /* If this is an input descriptor,
2667                          * increment that count. */
2668                         *in_num += ret;
2669                         if (unlikely(log && ret)) {
2670                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2671                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2672                                 ++*log_num;
2673                         }
2674                 } else {
2675                         /* If it's an output descriptor, they're all supposed
2676                          * to come before any input descriptors. */
2677                         if (unlikely(*in_num)) {
2678                                 vq_err(vq, "Descriptor has out after in: "
2679                                        "idx %d\n", i);
2680                                 return -EINVAL;
2681                         }
2682                         *out_num += ret;
2683                 }
2684         } while ((i = next_desc(vq, &desc)) != -1);
2685
2686         /* On success, increment avail index. */
2687         vq->last_avail_idx++;
2688
2689         /* Assume notifications from guest are disabled at this point,
2690          * if they aren't we would need to update avail_event index. */
2691         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2692         return head;
2693 }
2694 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2695
2696 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2697 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2698 {
2699         vq->last_avail_idx -= n;
2700 }
2701 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2702
2703 /* After we've used one of their buffers, we tell them about it.  We'll then
2704  * want to notify the guest, using eventfd. */
2705 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2706 {
2707         struct vring_used_elem heads = {
2708                 cpu_to_vhost32(vq, head),
2709                 cpu_to_vhost32(vq, len)
2710         };
2711
2712         return vhost_add_used_n(vq, &heads, 1);
2713 }
2714 EXPORT_SYMBOL_GPL(vhost_add_used);
2715
2716 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2717                             struct vring_used_elem *heads,
2718                             unsigned count)
2719 {
2720         vring_used_elem_t __user *used;
2721         u16 old, new;
2722         int start;
2723
2724         start = vq->last_used_idx & (vq->num - 1);
2725         used = vq->used->ring + start;
2726         if (vhost_put_used(vq, heads, start, count)) {
2727                 vq_err(vq, "Failed to write used");
2728                 return -EFAULT;
2729         }
2730         if (unlikely(vq->log_used)) {
2731                 /* Make sure data is seen before log. */
2732                 smp_wmb();
2733                 /* Log used ring entry write. */
2734                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2735                          count * sizeof *used);
2736         }
2737         old = vq->last_used_idx;
2738         new = (vq->last_used_idx += count);
2739         /* If the driver never bothers to signal in a very long while,
2740          * used index might wrap around. If that happens, invalidate
2741          * signalled_used index we stored. TODO: make sure driver
2742          * signals at least once in 2^16 and remove this. */
2743         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2744                 vq->signalled_used_valid = false;
2745         return 0;
2746 }
2747
2748 /* After we've used one of their buffers, we tell them about it.  We'll then
2749  * want to notify the guest, using eventfd. */
2750 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2751                      unsigned count)
2752 {
2753         int start, n, r;
2754
2755         start = vq->last_used_idx & (vq->num - 1);
2756         n = vq->num - start;
2757         if (n < count) {
2758                 r = __vhost_add_used_n(vq, heads, n);
2759                 if (r < 0)
2760                         return r;
2761                 heads += n;
2762                 count -= n;
2763         }
2764         r = __vhost_add_used_n(vq, heads, count);
2765
2766         /* Make sure buffer is written before we update index. */
2767         smp_wmb();
2768         if (vhost_put_used_idx(vq)) {
2769                 vq_err(vq, "Failed to increment used idx");
2770                 return -EFAULT;
2771         }
2772         if (unlikely(vq->log_used)) {
2773                 /* Make sure used idx is seen before log. */
2774                 smp_wmb();
2775                 /* Log used index update. */
2776                 log_used(vq, offsetof(struct vring_used, idx),
2777                          sizeof vq->used->idx);
2778                 if (vq->log_ctx)
2779                         eventfd_signal(vq->log_ctx);
2780         }
2781         return r;
2782 }
2783 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2784
2785 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2786 {
2787         __u16 old, new;
2788         __virtio16 event;
2789         bool v;
2790         /* Flush out used index updates. This is paired
2791          * with the barrier that the Guest executes when enabling
2792          * interrupts. */
2793         smp_mb();
2794
2795         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2796             unlikely(vq->avail_idx == vq->last_avail_idx))
2797                 return true;
2798
2799         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2800                 __virtio16 flags;
2801                 if (vhost_get_avail_flags(vq, &flags)) {
2802                         vq_err(vq, "Failed to get flags");
2803                         return true;
2804                 }
2805                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2806         }
2807         old = vq->signalled_used;
2808         v = vq->signalled_used_valid;
2809         new = vq->signalled_used = vq->last_used_idx;
2810         vq->signalled_used_valid = true;
2811
2812         if (unlikely(!v))
2813                 return true;
2814
2815         if (vhost_get_used_event(vq, &event)) {
2816                 vq_err(vq, "Failed to get used event idx");
2817                 return true;
2818         }
2819         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2820 }
2821
2822 /* This actually signals the guest, using eventfd. */
2823 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2824 {
2825         /* Signal the Guest tell them we used something up. */
2826         if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2827                 eventfd_signal(vq->call_ctx.ctx);
2828 }
2829 EXPORT_SYMBOL_GPL(vhost_signal);
2830
2831 /* And here's the combo meal deal.  Supersize me! */
2832 void vhost_add_used_and_signal(struct vhost_dev *dev,
2833                                struct vhost_virtqueue *vq,
2834                                unsigned int head, int len)
2835 {
2836         vhost_add_used(vq, head, len);
2837         vhost_signal(dev, vq);
2838 }
2839 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2840
2841 /* multi-buffer version of vhost_add_used_and_signal */
2842 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2843                                  struct vhost_virtqueue *vq,
2844                                  struct vring_used_elem *heads, unsigned count)
2845 {
2846         vhost_add_used_n(vq, heads, count);
2847         vhost_signal(dev, vq);
2848 }
2849 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2850
2851 /* return true if we're sure that avaiable ring is empty */
2852 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2853 {
2854         int r;
2855
2856         if (vq->avail_idx != vq->last_avail_idx)
2857                 return false;
2858
2859         r = vhost_get_avail_idx(vq);
2860
2861         /* Note: we treat error as non-empty here */
2862         return r == 0;
2863 }
2864 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2865
2866 /* OK, now we need to know about added descriptors. */
2867 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2868 {
2869         int r;
2870
2871         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2872                 return false;
2873         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2874         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2875                 r = vhost_update_used_flags(vq);
2876                 if (r) {
2877                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2878                                &vq->used->flags, r);
2879                         return false;
2880                 }
2881         } else {
2882                 r = vhost_update_avail_event(vq);
2883                 if (r) {
2884                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2885                                vhost_avail_event(vq), r);
2886                         return false;
2887                 }
2888         }
2889         /* They could have slipped one in as we were doing that: make
2890          * sure it's written, then check again. */
2891         smp_mb();
2892
2893         r = vhost_get_avail_idx(vq);
2894         /* Note: we treat error as empty here */
2895         if (unlikely(r < 0))
2896                 return false;
2897
2898         return r;
2899 }
2900 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2901
2902 /* We don't need to be notified again. */
2903 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2904 {
2905         int r;
2906
2907         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2908                 return;
2909         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2910         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2911                 r = vhost_update_used_flags(vq);
2912                 if (r)
2913                         vq_err(vq, "Failed to disable notification at %p: %d\n",
2914                                &vq->used->flags, r);
2915         }
2916 }
2917 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2918
2919 /* Create a new message. */
2920 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2921 {
2922         /* Make sure all padding within the structure is initialized. */
2923         struct vhost_msg_node *node = kzalloc(sizeof(*node), GFP_KERNEL);
2924         if (!node)
2925                 return NULL;
2926
2927         node->vq = vq;
2928         node->msg.type = type;
2929         return node;
2930 }
2931 EXPORT_SYMBOL_GPL(vhost_new_msg);
2932
2933 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2934                        struct vhost_msg_node *node)
2935 {
2936         spin_lock(&dev->iotlb_lock);
2937         list_add_tail(&node->node, head);
2938         spin_unlock(&dev->iotlb_lock);
2939
2940         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2941 }
2942 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2943
2944 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2945                                          struct list_head *head)
2946 {
2947         struct vhost_msg_node *node = NULL;
2948
2949         spin_lock(&dev->iotlb_lock);
2950         if (!list_empty(head)) {
2951                 node = list_first_entry(head, struct vhost_msg_node,
2952                                         node);
2953                 list_del(&node->node);
2954         }
2955         spin_unlock(&dev->iotlb_lock);
2956
2957         return node;
2958 }
2959 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2960
2961 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2962 {
2963         struct vhost_virtqueue *vq;
2964         int i;
2965
2966         mutex_lock(&dev->mutex);
2967         for (i = 0; i < dev->nvqs; ++i) {
2968                 vq = dev->vqs[i];
2969                 mutex_lock(&vq->mutex);
2970                 vq->acked_backend_features = features;
2971                 mutex_unlock(&vq->mutex);
2972         }
2973         mutex_unlock(&dev->mutex);
2974 }
2975 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2976
2977 static int __init vhost_init(void)
2978 {
2979         return 0;
2980 }
2981
2982 static void __exit vhost_exit(void)
2983 {
2984 }
2985
2986 module_init(vhost_init);
2987 module_exit(vhost_exit);
2988
2989 MODULE_VERSION("0.0.1");
2990 MODULE_LICENSE("GPL v2");
2991 MODULE_AUTHOR("Michael S. Tsirkin");
2992 MODULE_DESCRIPTION("Host kernel accelerator for virtio");