drivers/vhost/vhost.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Copyright (C) 2006 Rusty Russell IBM Corporation
   3  *
   4  * Author: Michael S. Tsirkin <mst@redhat.com>
   5  *
   6  * Inspiration, some code, and most witty comments come from
   7  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.
  10  *
  11  * Generic code for virtio server in host kernel.
  12  */
  13
  14 #include <linux/eventfd.h>
  15 #include <linux/vhost.h>
  16 #include <linux/virtio_net.h>
  17 #include <linux/mm.h>
  18 #include <linux/mmu_context.h>
  19 #include <linux/miscdevice.h>
  20 #include <linux/mutex.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/poll.h>
  23 #include <linux/file.h>
  24 #include <linux/highmem.h>
  25 #include <linux/slab.h>
  26 #include <linux/kthread.h>
  27 #include <linux/cgroup.h>
  28
  29 #include <linux/net.h>
  30 #include <linux/if_packet.h>
  31 #include <linux/if_arp.h>
  32
  33 #include "vhost.h"
  34
  35 enum {
  36         VHOST_MEMORY_MAX_NREGIONS = 64,
  37         VHOST_MEMORY_F_LOG = 0x1,
  38 };
  39
  40 static unsigned vhost_zcopy_mask __read_mostly;
  41
  42 #define vhost_used_event(vq) ((u16 __user *)&vq->avail->ring[vq->num])
  43 #define vhost_avail_event(vq) ((u16 __user *)&vq->used->ring[vq->num])
  44
  45 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
  46                             poll_table *pt)
  47 {
  48         struct vhost_poll *poll;
  49
  50         poll = container_of(pt, struct vhost_poll, table);
  51         poll->wqh = wqh;
  52         add_wait_queue(wqh, &poll->wait);
  53 }
  54
  55 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
  56                              void *key)
  57 {
  58         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
  59
  60         if (!((unsigned long)key & poll->mask))
  61                 return 0;
  62
  63         vhost_poll_queue(poll);
  64         return 0;
  65 }
  66
  67 static void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
  68 {
  69         INIT_LIST_HEAD(&work->node);
  70         work->fn = fn;
  71         init_waitqueue_head(&work->done);
  72         work->flushing = 0;
  73         work->queue_seq = work->done_seq = 0;
  74 }
  75
  76 /* Init poll structure */
  77 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
  78                      unsigned long mask, struct vhost_dev *dev)
  79 {
  80         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
  81         init_poll_funcptr(&poll->table, vhost_poll_func);
  82         poll->mask = mask;
  83         poll->dev = dev;
  84
  85         vhost_work_init(&poll->work, fn);
  86 }
  87
  88 /* Start polling a file. We add ourselves to file's wait queue. The caller must
  89  * keep a reference to a file until after vhost_poll_stop is called. */
  90 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
  91 {
  92         unsigned long mask;
  93
  94         mask = file->f_op->poll(file, &poll->table);
  95         if (mask)
  96                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
  97 }
  98
  99 /* Stop polling a file. After this function returns, it becomes safe to drop the
 100  * file reference. You must also flush afterwards. */
 101 void vhost_poll_stop(struct vhost_poll *poll)
 102 {
 103         remove_wait_queue(poll->wqh, &poll->wait);
 104 }
 105
 106 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
 107                                 unsigned seq)
 108 {
 109         int left;
 110
 111         spin_lock_irq(&dev->work_lock);
 112         left = seq - work->done_seq;
 113         spin_unlock_irq(&dev->work_lock);
 114         return left <= 0;
 115 }
 116
 117 static void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 118 {
 119         unsigned seq;
 120         int flushing;
 121
 122         spin_lock_irq(&dev->work_lock);
 123         seq = work->queue_seq;
 124         work->flushing++;
 125         spin_unlock_irq(&dev->work_lock);
 126         wait_event(work->done, vhost_work_seq_done(dev, work, seq));
 127         spin_lock_irq(&dev->work_lock);
 128         flushing = --work->flushing;
 129         spin_unlock_irq(&dev->work_lock);
 130         BUG_ON(flushing < 0);
 131 }
 132
 133 /* Flush any work that has been scheduled. When calling this, don't hold any
 134  * locks that are also used by the callback. */
 135 void vhost_poll_flush(struct vhost_poll *poll)
 136 {
 137         vhost_work_flush(poll->dev, &poll->work);
 138 }
 139
 140 static inline void vhost_work_queue(struct vhost_dev *dev,
 141                                     struct vhost_work *work)
 142 {
 143         unsigned long flags;
 144
 145         spin_lock_irqsave(&dev->work_lock, flags);
 146         if (list_empty(&work->node)) {
 147                 list_add_tail(&work->node, &dev->work_list);
 148                 work->queue_seq++;
 149                 wake_up_process(dev->worker);
 150         }
 151         spin_unlock_irqrestore(&dev->work_lock, flags);
 152 }
 153
 154 void vhost_poll_queue(struct vhost_poll *poll)
 155 {
 156         vhost_work_queue(poll->dev, &poll->work);
 157 }
 158
 159 static void vhost_vq_reset(struct vhost_dev *dev,
 160                            struct vhost_virtqueue *vq)
 161 {
 162         vq->num = 1;
 163         vq->desc = NULL;
 164         vq->avail = NULL;
 165         vq->used = NULL;
 166         vq->last_avail_idx = 0;
 167         vq->avail_idx = 0;
 168         vq->last_used_idx = 0;
 169         vq->signalled_used = 0;
 170         vq->signalled_used_valid = false;
 171         vq->used_flags = 0;
 172         vq->log_used = false;
 173         vq->log_addr = -1ull;
 174         vq->vhost_hlen = 0;
 175         vq->sock_hlen = 0;
 176         vq->private_data = NULL;
 177         vq->log_base = NULL;
 178         vq->error_ctx = NULL;
 179         vq->error = NULL;
 180         vq->kick = NULL;
 181         vq->call_ctx = NULL;
 182         vq->call = NULL;
 183         vq->log_ctx = NULL;
 184         vq->upend_idx = 0;
 185         vq->done_idx = 0;
 186         vq->ubufs = NULL;
 187 }
 188
 189 static int vhost_worker(void *data)
 190 {
 191         struct vhost_dev *dev = data;
 192         struct vhost_work *work = NULL;
 193         unsigned uninitialized_var(seq);
 194
 195         use_mm(dev->mm);
 196
 197         for (;;) {
 198                 /* mb paired w/ kthread_stop */
 199                 set_current_state(TASK_INTERRUPTIBLE);
 200
 201                 spin_lock_irq(&dev->work_lock);
 202                 if (work) {
 203                         work->done_seq = seq;
 204                         if (work->flushing)
 205                                 wake_up_all(&work->done);
 206                 }
 207
 208                 if (kthread_should_stop()) {
 209                         spin_unlock_irq(&dev->work_lock);
 210                         __set_current_state(TASK_RUNNING);
 211                         break;
 212                 }
 213                 if (!list_empty(&dev->work_list)) {
 214                         work = list_first_entry(&dev->work_list,
 215                                                 struct vhost_work, node);
 216                         list_del_init(&work->node);
 217                         seq = work->queue_seq;
 218                 } else
 219                         work = NULL;
 220                 spin_unlock_irq(&dev->work_lock);
 221
 222                 if (work) {
 223                         __set_current_state(TASK_RUNNING);
 224                         work->fn(work);
 225                 } else
 226                         schedule();
 227
 228         }
 229         unuse_mm(dev->mm);
 230         return 0;
 231 }
 232
 233 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 234 {
 235         kfree(vq->indirect);
 236         vq->indirect = NULL;
 237         kfree(vq->log);
 238         vq->log = NULL;
 239         kfree(vq->heads);
 240         vq->heads = NULL;
 241         kfree(vq->ubuf_info);
 242         vq->ubuf_info = NULL;
 243 }
 244
 245 void vhost_enable_zcopy(int vq)
 246 {
 247         vhost_zcopy_mask |= 0x1 << vq;
 248 }
 249
 250 /* Helper to allocate iovec buffers for all vqs. */
 251 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 252 {
 253         int i;
 254         bool zcopy;
 255
 256         for (i = 0; i < dev->nvqs; ++i) {
 257                 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
 258                                                UIO_MAXIOV, GFP_KERNEL);
 259                 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
 260                                           GFP_KERNEL);
 261                 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
 262                                             UIO_MAXIOV, GFP_KERNEL);
 263                 zcopy = vhost_zcopy_mask & (0x1 << i);
 264                 if (zcopy)
 265                         dev->vqs[i].ubuf_info =
 266                                 kmalloc(sizeof *dev->vqs[i].ubuf_info *
 267                                         UIO_MAXIOV, GFP_KERNEL);
 268                 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
 269                         !dev->vqs[i].heads ||
 270                         (zcopy && !dev->vqs[i].ubuf_info))
 271                         goto err_nomem;
 272         }
 273         return 0;
 274
 275 err_nomem:
 276         for (; i >= 0; --i)
 277                 vhost_vq_free_iovecs(&dev->vqs[i]);
 278         return -ENOMEM;
 279 }
 280
 281 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 282 {
 283         int i;
 284
 285         for (i = 0; i < dev->nvqs; ++i)
 286                 vhost_vq_free_iovecs(&dev->vqs[i]);
 287 }
 288
 289 long vhost_dev_init(struct vhost_dev *dev,
 290                     struct vhost_virtqueue *vqs, int nvqs)
 291 {
 292         int i;
 293
 294         dev->vqs = vqs;
 295         dev->nvqs = nvqs;
 296         mutex_init(&dev->mutex);
 297         dev->log_ctx = NULL;
 298         dev->log_file = NULL;
 299         dev->memory = NULL;
 300         dev->mm = NULL;
 301         spin_lock_init(&dev->work_lock);
 302         INIT_LIST_HEAD(&dev->work_list);
 303         dev->worker = NULL;
 304
 305         for (i = 0; i < dev->nvqs; ++i) {
 306                 dev->vqs[i].log = NULL;
 307                 dev->vqs[i].indirect = NULL;
 308                 dev->vqs[i].heads = NULL;
 309                 dev->vqs[i].ubuf_info = NULL;
 310                 dev->vqs[i].dev = dev;
 311                 mutex_init(&dev->vqs[i].mutex);
 312                 vhost_vq_reset(dev, dev->vqs + i);
 313                 if (dev->vqs[i].handle_kick)
 314                         vhost_poll_init(&dev->vqs[i].poll,
 315                                         dev->vqs[i].handle_kick, POLLIN, dev);
 316         }
 317
 318         return 0;
 319 }
 320
 321 /* Caller should have device mutex */
 322 long vhost_dev_check_owner(struct vhost_dev *dev)
 323 {
 324         /* Are you the owner? If not, I don't think you mean to do that */
 325         return dev->mm == current->mm ? 0 : -EPERM;
 326 }
 327
 328 struct vhost_attach_cgroups_struct {
 329         struct vhost_work work;
 330         struct task_struct *owner;
 331         int ret;
 332 };
 333
 334 static void vhost_attach_cgroups_work(struct vhost_work *work)
 335 {
 336         struct vhost_attach_cgroups_struct *s;
 337
 338         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 339         s->ret = cgroup_attach_task_all(s->owner, current);
 340 }
 341
 342 static int vhost_attach_cgroups(struct vhost_dev *dev)
 343 {
 344         struct vhost_attach_cgroups_struct attach;
 345
 346         attach.owner = current;
 347         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 348         vhost_work_queue(dev, &attach.work);
 349         vhost_work_flush(dev, &attach.work);
 350         return attach.ret;
 351 }
 352
 353 /* Caller should have device mutex */
 354 static long vhost_dev_set_owner(struct vhost_dev *dev)
 355 {
 356         struct task_struct *worker;
 357         int err;
 358
 359         /* Is there an owner already? */
 360         if (dev->mm) {
 361                 err = -EBUSY;
 362                 goto err_mm;
 363         }
 364
 365         /* No owner, become one */
 366         dev->mm = get_task_mm(current);
 367         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 368         if (IS_ERR(worker)) {
 369                 err = PTR_ERR(worker);
 370                 goto err_worker;
 371         }
 372
 373         dev->worker = worker;
 374         wake_up_process(worker);        /* avoid contributing to loadavg */
 375
 376         err = vhost_attach_cgroups(dev);
 377         if (err)
 378                 goto err_cgroup;
 379
 380         err = vhost_dev_alloc_iovecs(dev);
 381         if (err)
 382                 goto err_cgroup;
 383
 384         return 0;
 385 err_cgroup:
 386         kthread_stop(worker);
 387         dev->worker = NULL;
 388 err_worker:
 389         if (dev->mm)
 390                 mmput(dev->mm);
 391         dev->mm = NULL;
 392 err_mm:
 393         return err;
 394 }
 395
 396 /* Caller should have device mutex */
 397 long vhost_dev_reset_owner(struct vhost_dev *dev)
 398 {
 399         struct vhost_memory *memory;
 400
 401         /* Restore memory to default empty mapping. */
 402         memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 403         if (!memory)
 404                 return -ENOMEM;
 405
 406         vhost_dev_cleanup(dev);
 407
 408         memory->nregions = 0;
 409         RCU_INIT_POINTER(dev->memory, memory);
 410         return 0;
 411 }
 412
 413 /* In case of DMA done not in order in lower device driver for some reason.
 414  * upend_idx is used to track end of used idx, done_idx is used to track head
 415  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 416  * guest used idx.
 417  */
 418 int vhost_zerocopy_signal_used(struct vhost_virtqueue *vq)
 419 {
 420         int i;
 421         int j = 0;
 422
 423         for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 424                 if ((vq->heads[i].len == VHOST_DMA_DONE_LEN)) {
 425                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 426                         vhost_add_used_and_signal(vq->dev, vq,
 427                                                   vq->heads[i].id, 0);
 428                         ++j;
 429                 } else
 430                         break;
 431         }
 432         if (j)
 433                 vq->done_idx = i;
 434         return j;
 435 }
 436
 437 /* Caller should have device mutex */
 438 void vhost_dev_cleanup(struct vhost_dev *dev)
 439 {
 440         int i;
 441
 442         for (i = 0; i < dev->nvqs; ++i) {
 443                 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
 444                         vhost_poll_stop(&dev->vqs[i].poll);
 445                         vhost_poll_flush(&dev->vqs[i].poll);
 446                 }
 447                 /* Wait for all lower device DMAs done. */
 448                 if (dev->vqs[i].ubufs)
 449                         vhost_ubuf_put_and_wait(dev->vqs[i].ubufs);
 450
 451                 /* Signal guest as appropriate. */
 452                 vhost_zerocopy_signal_used(&dev->vqs[i]);
 453
 454                 if (dev->vqs[i].error_ctx)
 455                         eventfd_ctx_put(dev->vqs[i].error_ctx);
 456                 if (dev->vqs[i].error)
 457                         fput(dev->vqs[i].error);
 458                 if (dev->vqs[i].kick)
 459                         fput(dev->vqs[i].kick);
 460                 if (dev->vqs[i].call_ctx)
 461                         eventfd_ctx_put(dev->vqs[i].call_ctx);
 462                 if (dev->vqs[i].call)
 463                         fput(dev->vqs[i].call);
 464                 vhost_vq_reset(dev, dev->vqs + i);
 465         }
 466         vhost_dev_free_iovecs(dev);
 467         if (dev->log_ctx)
 468                 eventfd_ctx_put(dev->log_ctx);
 469         dev->log_ctx = NULL;
 470         if (dev->log_file)
 471                 fput(dev->log_file);
 472         dev->log_file = NULL;
 473         /* No one will access memory at this point */
 474         kfree(rcu_dereference_protected(dev->memory,
 475                                         lockdep_is_held(&dev->mutex)));
 476         RCU_INIT_POINTER(dev->memory, NULL);
 477         WARN_ON(!list_empty(&dev->work_list));
 478         if (dev->worker) {
 479                 kthread_stop(dev->worker);
 480                 dev->worker = NULL;
 481         }
 482         if (dev->mm)
 483                 mmput(dev->mm);
 484         dev->mm = NULL;
 485 }
 486
 487 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 488 {
 489         u64 a = addr / VHOST_PAGE_SIZE / 8;
 490
 491         /* Make sure 64 bit math will not overflow. */
 492         if (a > ULONG_MAX - (unsigned long)log_base ||
 493             a + (unsigned long)log_base > ULONG_MAX)
 494                 return 0;
 495
 496         return access_ok(VERIFY_WRITE, log_base + a,
 497                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 498 }
 499
 500 /* Caller should have vq mutex and device mutex. */
 501 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 502                                int log_all)
 503 {
 504         int i;
 505
 506         if (!mem)
 507                 return 0;
 508
 509         for (i = 0; i < mem->nregions; ++i) {
 510                 struct vhost_memory_region *m = mem->regions + i;
 511                 unsigned long a = m->userspace_addr;
 512                 if (m->memory_size > ULONG_MAX)
 513                         return 0;
 514                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 515                                     m->memory_size))
 516                         return 0;
 517                 else if (log_all && !log_access_ok(log_base,
 518                                                    m->guest_phys_addr,
 519                                                    m->memory_size))
 520                         return 0;
 521         }
 522         return 1;
 523 }
 524
 525 /* Can we switch to this memory table? */
 526 /* Caller should have device mutex but not vq mutex */
 527 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 528                             int log_all)
 529 {
 530         int i;
 531
 532         for (i = 0; i < d->nvqs; ++i) {
 533                 int ok;
 534                 mutex_lock(&d->vqs[i].mutex);
 535                 /* If ring is inactive, will check when it's enabled. */
 536                 if (d->vqs[i].private_data)
 537                         ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
 538                                                  log_all);
 539                 else
 540                         ok = 1;
 541                 mutex_unlock(&d->vqs[i].mutex);
 542                 if (!ok)
 543                         return 0;
 544         }
 545         return 1;
 546 }
 547
 548 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
 549                         struct vring_desc __user *desc,
 550                         struct vring_avail __user *avail,
 551                         struct vring_used __user *used)
 552 {
 553         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 554         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 555                access_ok(VERIFY_READ, avail,
 556                          sizeof *avail + num * sizeof *avail->ring + s) &&
 557                access_ok(VERIFY_WRITE, used,
 558                         sizeof *used + num * sizeof *used->ring + s);
 559 }
 560
 561 /* Can we log writes? */
 562 /* Caller should have device mutex but not vq mutex */
 563 int vhost_log_access_ok(struct vhost_dev *dev)
 564 {
 565         struct vhost_memory *mp;
 566
 567         mp = rcu_dereference_protected(dev->memory,
 568                                        lockdep_is_held(&dev->mutex));
 569         return memory_access_ok(dev, mp, 1);
 570 }
 571
 572 /* Verify access for write logging. */
 573 /* Caller should have vq mutex and device mutex */
 574 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
 575                             void __user *log_base)
 576 {
 577         struct vhost_memory *mp;
 578         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 579
 580         mp = rcu_dereference_protected(vq->dev->memory,
 581                                        lockdep_is_held(&vq->mutex));
 582         return vq_memory_access_ok(log_base, mp,
 583                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
 584                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 585                                         sizeof *vq->used +
 586                                         vq->num * sizeof *vq->used->ring + s));
 587 }
 588
 589 /* Can we start vq? */
 590 /* Caller should have vq mutex and device mutex */
 591 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 592 {
 593         return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
 594                 vq_log_access_ok(vq->dev, vq, vq->log_base);
 595 }
 596
 597 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 598 {
 599         struct vhost_memory mem, *newmem, *oldmem;
 600         unsigned long size = offsetof(struct vhost_memory, regions);
 601
 602         if (copy_from_user(&mem, m, size))
 603                 return -EFAULT;
 604         if (mem.padding)
 605                 return -EOPNOTSUPP;
 606         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 607                 return -E2BIG;
 608         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 609         if (!newmem)
 610                 return -ENOMEM;
 611
 612         memcpy(newmem, &mem, size);
 613         if (copy_from_user(newmem->regions, m->regions,
 614                            mem.nregions * sizeof *m->regions)) {
 615                 kfree(newmem);
 616                 return -EFAULT;
 617         }
 618
 619         if (!memory_access_ok(d, newmem,
 620                               vhost_has_feature(d, VHOST_F_LOG_ALL))) {
 621                 kfree(newmem);
 622                 return -EFAULT;
 623         }
 624         oldmem = rcu_dereference_protected(d->memory,
 625                                            lockdep_is_held(&d->mutex));
 626         rcu_assign_pointer(d->memory, newmem);
 627         synchronize_rcu();
 628         kfree(oldmem);
 629         return 0;
 630 }
 631
 632 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
 633 {
 634         struct file *eventfp, *filep = NULL,
 635                     *pollstart = NULL, *pollstop = NULL;
 636         struct eventfd_ctx *ctx = NULL;
 637         u32 __user *idxp = argp;
 638         struct vhost_virtqueue *vq;
 639         struct vhost_vring_state s;
 640         struct vhost_vring_file f;
 641         struct vhost_vring_addr a;
 642         u32 idx;
 643         long r;
 644
 645         r = get_user(idx, idxp);
 646         if (r < 0)
 647                 return r;
 648         if (idx >= d->nvqs)
 649                 return -ENOBUFS;
 650
 651         vq = d->vqs + idx;
 652
 653         mutex_lock(&vq->mutex);
 654
 655         switch (ioctl) {
 656         case VHOST_SET_VRING_NUM:
 657                 /* Resizing ring with an active backend?
 658                  * You don't want to do that. */
 659                 if (vq->private_data) {
 660                         r = -EBUSY;
 661                         break;
 662                 }
 663                 if (copy_from_user(&s, argp, sizeof s)) {
 664                         r = -EFAULT;
 665                         break;
 666                 }
 667                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 668                         r = -EINVAL;
 669                         break;
 670                 }
 671                 vq->num = s.num;
 672                 break;
 673         case VHOST_SET_VRING_BASE:
 674                 /* Moving base with an active backend?
 675                  * You don't want to do that. */
 676                 if (vq->private_data) {
 677                         r = -EBUSY;
 678                         break;
 679                 }
 680                 if (copy_from_user(&s, argp, sizeof s)) {
 681                         r = -EFAULT;
 682                         break;
 683                 }
 684                 if (s.num > 0xffff) {
 685                         r = -EINVAL;
 686                         break;
 687                 }
 688                 vq->last_avail_idx = s.num;
 689                 /* Forget the cached index value. */
 690                 vq->avail_idx = vq->last_avail_idx;
 691                 break;
 692         case VHOST_GET_VRING_BASE:
 693                 s.index = idx;
 694                 s.num = vq->last_avail_idx;
 695                 if (copy_to_user(argp, &s, sizeof s))
 696                         r = -EFAULT;
 697                 break;
 698         case VHOST_SET_VRING_ADDR:
 699                 if (copy_from_user(&a, argp, sizeof a)) {
 700                         r = -EFAULT;
 701                         break;
 702                 }
 703                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 704                         r = -EOPNOTSUPP;
 705                         break;
 706                 }
 707                 /* For 32bit, verify that the top 32bits of the user
 708                    data are set to zero. */
 709                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 710                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 711                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 712                         r = -EFAULT;
 713                         break;
 714                 }
 715                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 716                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 717                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 718                         r = -EINVAL;
 719                         break;
 720                 }
 721
 722                 /* We only verify access here if backend is configured.
 723                  * If it is not, we don't as size might not have been setup.
 724                  * We will verify when backend is configured. */
 725                 if (vq->private_data) {
 726                         if (!vq_access_ok(d, vq->num,
 727                                 (void __user *)(unsigned long)a.desc_user_addr,
 728                                 (void __user *)(unsigned long)a.avail_user_addr,
 729                                 (void __user *)(unsigned long)a.used_user_addr)) {
 730                                 r = -EINVAL;
 731                                 break;
 732                         }
 733
 734                         /* Also validate log access for used ring if enabled. */
 735                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 736                             !log_access_ok(vq->log_base, a.log_guest_addr,
 737                                            sizeof *vq->used +
 738                                            vq->num * sizeof *vq->used->ring)) {
 739                                 r = -EINVAL;
 740                                 break;
 741                         }
 742                 }
 743
 744                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 745                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 746                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 747                 vq->log_addr = a.log_guest_addr;
 748                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 749                 break;
 750         case VHOST_SET_VRING_KICK:
 751                 if (copy_from_user(&f, argp, sizeof f)) {
 752                         r = -EFAULT;
 753                         break;
 754                 }
 755                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 756                 if (IS_ERR(eventfp)) {
 757                         r = PTR_ERR(eventfp);
 758                         break;
 759                 }
 760                 if (eventfp != vq->kick) {
 761                         pollstop = filep = vq->kick;
 762                         pollstart = vq->kick = eventfp;
 763                 } else
 764                         filep = eventfp;
 765                 break;
 766         case VHOST_SET_VRING_CALL:
 767                 if (copy_from_user(&f, argp, sizeof f)) {
 768                         r = -EFAULT;
 769                         break;
 770                 }
 771                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 772                 if (IS_ERR(eventfp)) {
 773                         r = PTR_ERR(eventfp);
 774                         break;
 775                 }
 776                 if (eventfp != vq->call) {
 777                         filep = vq->call;
 778                         ctx = vq->call_ctx;
 779                         vq->call = eventfp;
 780                         vq->call_ctx = eventfp ?
 781                                 eventfd_ctx_fileget(eventfp) : NULL;
 782                 } else
 783                         filep = eventfp;
 784                 break;
 785         case VHOST_SET_VRING_ERR:
 786                 if (copy_from_user(&f, argp, sizeof f)) {
 787                         r = -EFAULT;
 788                         break;
 789                 }
 790                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 791                 if (IS_ERR(eventfp)) {
 792                         r = PTR_ERR(eventfp);
 793                         break;
 794                 }
 795                 if (eventfp != vq->error) {
 796                         filep = vq->error;
 797                         vq->error = eventfp;
 798                         ctx = vq->error_ctx;
 799                         vq->error_ctx = eventfp ?
 800                                 eventfd_ctx_fileget(eventfp) : NULL;
 801                 } else
 802                         filep = eventfp;
 803                 break;
 804         default:
 805                 r = -ENOIOCTLCMD;
 806         }
 807
 808         if (pollstop && vq->handle_kick)
 809                 vhost_poll_stop(&vq->poll);
 810
 811         if (ctx)
 812                 eventfd_ctx_put(ctx);
 813         if (filep)
 814                 fput(filep);
 815
 816         if (pollstart && vq->handle_kick)
 817                 vhost_poll_start(&vq->poll, vq->kick);
 818
 819         mutex_unlock(&vq->mutex);
 820
 821         if (pollstop && vq->handle_kick)
 822                 vhost_poll_flush(&vq->poll);
 823         return r;
 824 }
 825
 826 /* Caller must have device mutex */
 827 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
 828 {
 829         void __user *argp = (void __user *)arg;
 830         struct file *eventfp, *filep = NULL;
 831         struct eventfd_ctx *ctx = NULL;
 832         u64 p;
 833         long r;
 834         int i, fd;
 835
 836         /* If you are not the owner, you can become one */
 837         if (ioctl == VHOST_SET_OWNER) {
 838                 r = vhost_dev_set_owner(d);
 839                 goto done;
 840         }
 841
 842         /* You must be the owner to do anything else */
 843         r = vhost_dev_check_owner(d);
 844         if (r)
 845                 goto done;
 846
 847         switch (ioctl) {
 848         case VHOST_SET_MEM_TABLE:
 849                 r = vhost_set_memory(d, argp);
 850                 break;
 851         case VHOST_SET_LOG_BASE:
 852                 if (copy_from_user(&p, argp, sizeof p)) {
 853                         r = -EFAULT;
 854                         break;
 855                 }
 856                 if ((u64)(unsigned long)p != p) {
 857                         r = -EFAULT;
 858                         break;
 859                 }
 860                 for (i = 0; i < d->nvqs; ++i) {
 861                         struct vhost_virtqueue *vq;
 862                         void __user *base = (void __user *)(unsigned long)p;
 863                         vq = d->vqs + i;
 864                         mutex_lock(&vq->mutex);
 865                         /* If ring is inactive, will check when it's enabled. */
 866                         if (vq->private_data && !vq_log_access_ok(d, vq, base))
 867                                 r = -EFAULT;
 868                         else
 869                                 vq->log_base = base;
 870                         mutex_unlock(&vq->mutex);
 871                 }
 872                 break;
 873         case VHOST_SET_LOG_FD:
 874                 r = get_user(fd, (int __user *)argp);
 875                 if (r < 0)
 876                         break;
 877                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 878                 if (IS_ERR(eventfp)) {
 879                         r = PTR_ERR(eventfp);
 880                         break;
 881                 }
 882                 if (eventfp != d->log_file) {
 883                         filep = d->log_file;
 884                         ctx = d->log_ctx;
 885                         d->log_ctx = eventfp ?
 886                                 eventfd_ctx_fileget(eventfp) : NULL;
 887                 } else
 888                         filep = eventfp;
 889                 for (i = 0; i < d->nvqs; ++i) {
 890                         mutex_lock(&d->vqs[i].mutex);
 891                         d->vqs[i].log_ctx = d->log_ctx;
 892                         mutex_unlock(&d->vqs[i].mutex);
 893                 }
 894                 if (ctx)
 895                         eventfd_ctx_put(ctx);
 896                 if (filep)
 897                         fput(filep);
 898                 break;
 899         default:
 900                 r = vhost_set_vring(d, ioctl, argp);
 901                 break;
 902         }
 903 done:
 904         return r;
 905 }
 906
 907 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 908                                                      __u64 addr, __u32 len)
 909 {
 910         struct vhost_memory_region *reg;
 911         int i;
 912
 913         /* linear search is not brilliant, but we really have on the order of 6
 914          * regions in practice */
 915         for (i = 0; i < mem->nregions; ++i) {
 916                 reg = mem->regions + i;
 917                 if (reg->guest_phys_addr <= addr &&
 918                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 919                         return reg;
 920         }
 921         return NULL;
 922 }
 923
 924 /* TODO: This is really inefficient.  We need something like get_user()
 925  * (instruction directly accesses the data, with an exception table entry
 926  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 927  */
 928 static int set_bit_to_user(int nr, void __user *addr)
 929 {
 930         unsigned long log = (unsigned long)addr;
 931         struct page *page;
 932         void *base;
 933         int bit = nr + (log % PAGE_SIZE) * 8;
 934         int r;
 935
 936         r = get_user_pages_fast(log, 1, 1, &page);
 937         if (r < 0)
 938                 return r;
 939         BUG_ON(r != 1);
 940         base = kmap_atomic(page, KM_USER0);
 941         set_bit(bit, base);
 942         kunmap_atomic(base, KM_USER0);
 943         set_page_dirty_lock(page);
 944         put_page(page);
 945         return 0;
 946 }
 947
 948 static int log_write(void __user *log_base,
 949                      u64 write_address, u64 write_length)
 950 {
 951         u64 write_page = write_address / VHOST_PAGE_SIZE;
 952         int r;
 953
 954         if (!write_length)
 955                 return 0;
 956         write_length += write_address % VHOST_PAGE_SIZE;
 957         for (;;) {
 958                 u64 base = (u64)(unsigned long)log_base;
 959                 u64 log = base + write_page / 8;
 960                 int bit = write_page % 8;
 961                 if ((u64)(unsigned long)log != log)
 962                         return -EFAULT;
 963                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 964                 if (r < 0)
 965                         return r;
 966                 if (write_length <= VHOST_PAGE_SIZE)
 967                         break;
 968                 write_length -= VHOST_PAGE_SIZE;
 969                 write_page += 1;
 970         }
 971         return r;
 972 }
 973
 974 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 975                     unsigned int log_num, u64 len)
 976 {
 977         int i, r;
 978
 979         /* Make sure data written is seen before log. */
 980         smp_wmb();
 981         for (i = 0; i < log_num; ++i) {
 982                 u64 l = min(log[i].len, len);
 983                 r = log_write(vq->log_base, log[i].addr, l);
 984                 if (r < 0)
 985                         return r;
 986                 len -= l;
 987                 if (!len) {
 988                         if (vq->log_ctx)
 989                                 eventfd_signal(vq->log_ctx, 1);
 990                         return 0;
 991                 }
 992         }
 993         /* Length written exceeds what we have stored. This is a bug. */
 994         BUG();
 995         return 0;
 996 }
 997
 998 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
 999 {
1000         void __user *used;
1001         if (__put_user(vq->used_flags, &vq->used->flags) < 0)
1002                 return -EFAULT;
1003         if (unlikely(vq->log_used)) {
1004                 /* Make sure the flag is seen before log. */
1005                 smp_wmb();
1006                 /* Log used flag write. */
1007                 used = &vq->used->flags;
1008                 log_write(vq->log_base, vq->log_addr +
1009                           (used - (void __user *)vq->used),
1010                           sizeof vq->used->flags);
1011                 if (vq->log_ctx)
1012                         eventfd_signal(vq->log_ctx, 1);
1013         }
1014         return 0;
1015 }
1016
1017 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1018 {
1019         if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1020                 return -EFAULT;
1021         if (unlikely(vq->log_used)) {
1022                 void __user *used;
1023                 /* Make sure the event is seen before log. */
1024                 smp_wmb();
1025                 /* Log avail event write */
1026                 used = vhost_avail_event(vq);
1027                 log_write(vq->log_base, vq->log_addr +
1028                           (used - (void __user *)vq->used),
1029                           sizeof *vhost_avail_event(vq));
1030                 if (vq->log_ctx)
1031                         eventfd_signal(vq->log_ctx, 1);
1032         }
1033         return 0;
1034 }
1035
1036 int vhost_init_used(struct vhost_virtqueue *vq)
1037 {
1038         int r;
1039         if (!vq->private_data)
1040                 return 0;
1041
1042         r = vhost_update_used_flags(vq);
1043         if (r)
1044                 return r;
1045         vq->signalled_used_valid = false;
1046         return get_user(vq->last_used_idx, &vq->used->idx);
1047 }
1048
1049 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1050                           struct iovec iov[], int iov_size)
1051 {
1052         const struct vhost_memory_region *reg;
1053         struct vhost_memory *mem;
1054         struct iovec *_iov;
1055         u64 s = 0;
1056         int ret = 0;
1057
1058         rcu_read_lock();
1059
1060         mem = rcu_dereference(dev->memory);
1061         while ((u64)len > s) {
1062                 u64 size;
1063                 if (unlikely(ret >= iov_size)) {
1064                         ret = -ENOBUFS;
1065                         break;
1066                 }
1067                 reg = find_region(mem, addr, len);
1068                 if (unlikely(!reg)) {
1069                         ret = -EFAULT;
1070                         break;
1071                 }
1072                 _iov = iov + ret;
1073                 size = reg->memory_size - addr + reg->guest_phys_addr;
1074                 _iov->iov_len = min((u64)len, size);
1075                 _iov->iov_base = (void __user *)(unsigned long)
1076                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1077                 s += size;
1078                 addr += size;
1079                 ++ret;
1080         }
1081
1082         rcu_read_unlock();
1083         return ret;
1084 }
1085
1086 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1087  * function returns the next descriptor in the chain,
1088  * or -1U if we're at the end. */
1089 static unsigned next_desc(struct vring_desc *desc)
1090 {
1091         unsigned int next;
1092
1093         /* If this descriptor says it doesn't chain, we're done. */
1094         if (!(desc->flags & VRING_DESC_F_NEXT))
1095                 return -1U;
1096
1097         /* Check they're not leading us off end of descriptors. */
1098         next = desc->next;
1099         /* Make sure compiler knows to grab that: we don't want it changing! */
1100         /* We will use the result as an index in an array, so most
1101          * architectures only need a compiler barrier here. */
1102         read_barrier_depends();
1103
1104         return next;
1105 }
1106
1107 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1108                         struct iovec iov[], unsigned int iov_size,
1109                         unsigned int *out_num, unsigned int *in_num,
1110                         struct vhost_log *log, unsigned int *log_num,
1111                         struct vring_desc *indirect)
1112 {
1113         struct vring_desc desc;
1114         unsigned int i = 0, count, found = 0;
1115         int ret;
1116
1117         /* Sanity check */
1118         if (unlikely(indirect->len % sizeof desc)) {
1119                 vq_err(vq, "Invalid length in indirect descriptor: "
1120                        "len 0x%llx not multiple of 0x%zx\n",
1121                        (unsigned long long)indirect->len,
1122                        sizeof desc);
1123                 return -EINVAL;
1124         }
1125
1126         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1127                              UIO_MAXIOV);
1128         if (unlikely(ret < 0)) {
1129                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1130                 return ret;
1131         }
1132
1133         /* We will use the result as an address to read from, so most
1134          * architectures only need a compiler barrier here. */
1135         read_barrier_depends();
1136
1137         count = indirect->len / sizeof desc;
1138         /* Buffers are chained via a 16 bit next field, so
1139          * we can have at most 2^16 of these. */
1140         if (unlikely(count > USHRT_MAX + 1)) {
1141                 vq_err(vq, "Indirect buffer length too big: %d\n",
1142                        indirect->len);
1143                 return -E2BIG;
1144         }
1145
1146         do {
1147                 unsigned iov_count = *in_num + *out_num;
1148                 if (unlikely(++found > count)) {
1149                         vq_err(vq, "Loop detected: last one at %u "
1150                                "indirect size %u\n",
1151                                i, count);
1152                         return -EINVAL;
1153                 }
1154                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1155                                               vq->indirect, sizeof desc))) {
1156                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1157                                i, (size_t)indirect->addr + i * sizeof desc);
1158                         return -EINVAL;
1159                 }
1160                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1161                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1162                                i, (size_t)indirect->addr + i * sizeof desc);
1163                         return -EINVAL;
1164                 }
1165
1166                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1167                                      iov_size - iov_count);
1168                 if (unlikely(ret < 0)) {
1169                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1170                                ret, i);
1171                         return ret;
1172                 }
1173                 /* If this is an input descriptor, increment that count. */
1174                 if (desc.flags & VRING_DESC_F_WRITE) {
1175                         *in_num += ret;
1176                         if (unlikely(log)) {
1177                                 log[*log_num].addr = desc.addr;
1178                                 log[*log_num].len = desc.len;
1179                                 ++*log_num;
1180                         }
1181                 } else {
1182                         /* If it's an output descriptor, they're all supposed
1183                          * to come before any input descriptors. */
1184                         if (unlikely(*in_num)) {
1185                                 vq_err(vq, "Indirect descriptor "
1186                                        "has out after in: idx %d\n", i);
1187                                 return -EINVAL;
1188                         }
1189                         *out_num += ret;
1190                 }
1191         } while ((i = next_desc(&desc)) != -1);
1192         return 0;
1193 }
1194
1195 /* This looks in the virtqueue and for the first available buffer, and converts
1196  * it to an iovec for convenient access.  Since descriptors consist of some
1197  * number of output then some number of input descriptors, it's actually two
1198  * iovecs, but we pack them into one and note how many of each there were.
1199  *
1200  * This function returns the descriptor number found, or vq->num (which is
1201  * never a valid descriptor number) if none was found.  A negative code is
1202  * returned on error. */
1203 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1204                       struct iovec iov[], unsigned int iov_size,
1205                       unsigned int *out_num, unsigned int *in_num,
1206                       struct vhost_log *log, unsigned int *log_num)
1207 {
1208         struct vring_desc desc;
1209         unsigned int i, head, found = 0;
1210         u16 last_avail_idx;
1211         int ret;
1212
1213         /* Check it isn't doing very strange things with descriptor numbers. */
1214         last_avail_idx = vq->last_avail_idx;
1215         if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1216                 vq_err(vq, "Failed to access avail idx at %p\n",
1217                        &vq->avail->idx);
1218                 return -EFAULT;
1219         }
1220
1221         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1222                 vq_err(vq, "Guest moved used index from %u to %u",
1223                        last_avail_idx, vq->avail_idx);
1224                 return -EFAULT;
1225         }
1226
1227         /* If there's nothing new since last we looked, return invalid. */
1228         if (vq->avail_idx == last_avail_idx)
1229                 return vq->num;
1230
1231         /* Only get avail ring entries after they have been exposed by guest. */
1232         smp_rmb();
1233
1234         /* Grab the next descriptor number they're advertising, and increment
1235          * the index we've seen. */
1236         if (unlikely(__get_user(head,
1237                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1238                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1239                        last_avail_idx,
1240                        &vq->avail->ring[last_avail_idx % vq->num]);
1241                 return -EFAULT;
1242         }
1243
1244         /* If their number is silly, that's an error. */
1245         if (unlikely(head >= vq->num)) {
1246                 vq_err(vq, "Guest says index %u > %u is available",
1247                        head, vq->num);
1248                 return -EINVAL;
1249         }
1250
1251         /* When we start there are none of either input nor output. */
1252         *out_num = *in_num = 0;
1253         if (unlikely(log))
1254                 *log_num = 0;
1255
1256         i = head;
1257         do {
1258                 unsigned iov_count = *in_num + *out_num;
1259                 if (unlikely(i >= vq->num)) {
1260                         vq_err(vq, "Desc index is %u > %u, head = %u",
1261                                i, vq->num, head);
1262                         return -EINVAL;
1263                 }
1264                 if (unlikely(++found > vq->num)) {
1265                         vq_err(vq, "Loop detected: last one at %u "
1266                                "vq size %u head %u\n",
1267                                i, vq->num, head);
1268                         return -EINVAL;
1269                 }
1270                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1271                 if (unlikely(ret)) {
1272                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1273                                i, vq->desc + i);
1274                         return -EFAULT;
1275                 }
1276                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1277                         ret = get_indirect(dev, vq, iov, iov_size,
1278                                            out_num, in_num,
1279                                            log, log_num, &desc);
1280                         if (unlikely(ret < 0)) {
1281                                 vq_err(vq, "Failure detected "
1282                                        "in indirect descriptor at idx %d\n", i);
1283                                 return ret;
1284                         }
1285                         continue;
1286                 }
1287
1288                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1289                                      iov_size - iov_count);
1290                 if (unlikely(ret < 0)) {
1291                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1292                                ret, i);
1293                         return ret;
1294                 }
1295                 if (desc.flags & VRING_DESC_F_WRITE) {
1296                         /* If this is an input descriptor,
1297                          * increment that count. */
1298                         *in_num += ret;
1299                         if (unlikely(log)) {
1300                                 log[*log_num].addr = desc.addr;
1301                                 log[*log_num].len = desc.len;
1302                                 ++*log_num;
1303                         }
1304                 } else {
1305                         /* If it's an output descriptor, they're all supposed
1306                          * to come before any input descriptors. */
1307                         if (unlikely(*in_num)) {
1308                                 vq_err(vq, "Descriptor has out after in: "
1309                                        "idx %d\n", i);
1310                                 return -EINVAL;
1311                         }
1312                         *out_num += ret;
1313                 }
1314         } while ((i = next_desc(&desc)) != -1);
1315
1316         /* On success, increment avail index. */
1317         vq->last_avail_idx++;
1318
1319         /* Assume notifications from guest are disabled at this point,
1320          * if they aren't we would need to update avail_event index. */
1321         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1322         return head;
1323 }
1324
1325 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1326 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1327 {
1328         vq->last_avail_idx -= n;
1329 }
1330
1331 /* After we've used one of their buffers, we tell them about it.  We'll then
1332  * want to notify the guest, using eventfd. */
1333 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1334 {
1335         struct vring_used_elem __user *used;
1336
1337         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1338          * next entry in that used ring. */
1339         used = &vq->used->ring[vq->last_used_idx % vq->num];
1340         if (__put_user(head, &used->id)) {
1341                 vq_err(vq, "Failed to write used id");
1342                 return -EFAULT;
1343         }
1344         if (__put_user(len, &used->len)) {
1345                 vq_err(vq, "Failed to write used len");
1346                 return -EFAULT;
1347         }
1348         /* Make sure buffer is written before we update index. */
1349         smp_wmb();
1350         if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1351                 vq_err(vq, "Failed to increment used idx");
1352                 return -EFAULT;
1353         }
1354         if (unlikely(vq->log_used)) {
1355                 /* Make sure data is seen before log. */
1356                 smp_wmb();
1357                 /* Log used ring entry write. */
1358                 log_write(vq->log_base,
1359                           vq->log_addr +
1360                            ((void __user *)used - (void __user *)vq->used),
1361                           sizeof *used);
1362                 /* Log used index update. */
1363                 log_write(vq->log_base,
1364                           vq->log_addr + offsetof(struct vring_used, idx),
1365                           sizeof vq->used->idx);
1366                 if (vq->log_ctx)
1367                         eventfd_signal(vq->log_ctx, 1);
1368         }
1369         vq->last_used_idx++;
1370         /* If the driver never bothers to signal in a very long while,
1371          * used index might wrap around. If that happens, invalidate
1372          * signalled_used index we stored. TODO: make sure driver
1373          * signals at least once in 2^16 and remove this. */
1374         if (unlikely(vq->last_used_idx == vq->signalled_used))
1375                 vq->signalled_used_valid = false;
1376         return 0;
1377 }
1378
1379 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1380                             struct vring_used_elem *heads,
1381                             unsigned count)
1382 {
1383         struct vring_used_elem __user *used;
1384         u16 old, new;
1385         int start;
1386
1387         start = vq->last_used_idx % vq->num;
1388         used = vq->used->ring + start;
1389         if (__copy_to_user(used, heads, count * sizeof *used)) {
1390                 vq_err(vq, "Failed to write used");
1391                 return -EFAULT;
1392         }
1393         if (unlikely(vq->log_used)) {
1394                 /* Make sure data is seen before log. */
1395                 smp_wmb();
1396                 /* Log used ring entry write. */
1397                 log_write(vq->log_base,
1398                           vq->log_addr +
1399                            ((void __user *)used - (void __user *)vq->used),
1400                           count * sizeof *used);
1401         }
1402         old = vq->last_used_idx;
1403         new = (vq->last_used_idx += count);
1404         /* If the driver never bothers to signal in a very long while,
1405          * used index might wrap around. If that happens, invalidate
1406          * signalled_used index we stored. TODO: make sure driver
1407          * signals at least once in 2^16 and remove this. */
1408         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1409                 vq->signalled_used_valid = false;
1410         return 0;
1411 }
1412
1413 /* After we've used one of their buffers, we tell them about it.  We'll then
1414  * want to notify the guest, using eventfd. */
1415 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1416                      unsigned count)
1417 {
1418         int start, n, r;
1419
1420         start = vq->last_used_idx % vq->num;
1421         n = vq->num - start;
1422         if (n < count) {
1423                 r = __vhost_add_used_n(vq, heads, n);
1424                 if (r < 0)
1425                         return r;
1426                 heads += n;
1427                 count -= n;
1428         }
1429         r = __vhost_add_used_n(vq, heads, count);
1430
1431         /* Make sure buffer is written before we update index. */
1432         smp_wmb();
1433         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1434                 vq_err(vq, "Failed to increment used idx");
1435                 return -EFAULT;
1436         }
1437         if (unlikely(vq->log_used)) {
1438                 /* Log used index update. */
1439                 log_write(vq->log_base,
1440                           vq->log_addr + offsetof(struct vring_used, idx),
1441                           sizeof vq->used->idx);
1442                 if (vq->log_ctx)
1443                         eventfd_signal(vq->log_ctx, 1);
1444         }
1445         return r;
1446 }
1447
1448 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1449 {
1450         __u16 old, new, event;
1451         bool v;
1452         /* Flush out used index updates. This is paired
1453          * with the barrier that the Guest executes when enabling
1454          * interrupts. */
1455         smp_mb();
1456
1457         if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1458             unlikely(vq->avail_idx == vq->last_avail_idx))
1459                 return true;
1460
1461         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1462                 __u16 flags;
1463                 if (__get_user(flags, &vq->avail->flags)) {
1464                         vq_err(vq, "Failed to get flags");
1465                         return true;
1466                 }
1467                 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1468         }
1469         old = vq->signalled_used;
1470         v = vq->signalled_used_valid;
1471         new = vq->signalled_used = vq->last_used_idx;
1472         vq->signalled_used_valid = true;
1473
1474         if (unlikely(!v))
1475                 return true;
1476
1477         if (get_user(event, vhost_used_event(vq))) {
1478                 vq_err(vq, "Failed to get used event idx");
1479                 return true;
1480         }
1481         return vring_need_event(event, new, old);
1482 }
1483
1484 /* This actually signals the guest, using eventfd. */
1485 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1486 {
1487         /* Signal the Guest tell them we used something up. */
1488         if (vq->call_ctx && vhost_notify(dev, vq))
1489                 eventfd_signal(vq->call_ctx, 1);
1490 }
1491
1492 /* And here's the combo meal deal.  Supersize me! */
1493 void vhost_add_used_and_signal(struct vhost_dev *dev,
1494                                struct vhost_virtqueue *vq,
1495                                unsigned int head, int len)
1496 {
1497         vhost_add_used(vq, head, len);
1498         vhost_signal(dev, vq);
1499 }
1500
1501 /* multi-buffer version of vhost_add_used_and_signal */
1502 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1503                                  struct vhost_virtqueue *vq,
1504                                  struct vring_used_elem *heads, unsigned count)
1505 {
1506         vhost_add_used_n(vq, heads, count);
1507         vhost_signal(dev, vq);
1508 }
1509
1510 /* OK, now we need to know about added descriptors. */
1511 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1512 {
1513         u16 avail_idx;
1514         int r;
1515
1516         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1517                 return false;
1518         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1519         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1520                 r = vhost_update_used_flags(vq);
1521                 if (r) {
1522                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1523                                &vq->used->flags, r);
1524                         return false;
1525                 }
1526         } else {
1527                 r = vhost_update_avail_event(vq, vq->avail_idx);
1528                 if (r) {
1529                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1530                                vhost_avail_event(vq), r);
1531                         return false;
1532                 }
1533         }
1534         /* They could have slipped one in as we were doing that: make
1535          * sure it's written, then check again. */
1536         smp_mb();
1537         r = __get_user(avail_idx, &vq->avail->idx);
1538         if (r) {
1539                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1540                        &vq->avail->idx, r);
1541                 return false;
1542         }
1543
1544         return avail_idx != vq->avail_idx;
1545 }
1546
1547 /* We don't need to be notified again. */
1548 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1549 {
1550         int r;
1551
1552         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1553                 return;
1554         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1555         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1556                 r = vhost_update_used_flags(vq);
1557                 if (r)
1558                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1559                                &vq->used->flags, r);
1560         }
1561 }
1562
1563 static void vhost_zerocopy_done_signal(struct kref *kref)
1564 {
1565         struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1566                                                     kref);
1567         wake_up(&ubufs->wait);
1568 }
1569
1570 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1571                                         bool zcopy)
1572 {
1573         struct vhost_ubuf_ref *ubufs;
1574         /* No zero copy backend? Nothing to count. */
1575         if (!zcopy)
1576                 return NULL;
1577         ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1578         if (!ubufs)
1579                 return ERR_PTR(-ENOMEM);
1580         kref_init(&ubufs->kref);
1581         init_waitqueue_head(&ubufs->wait);
1582         ubufs->vq = vq;
1583         return ubufs;
1584 }
1585
1586 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1587 {
1588         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1589 }
1590
1591 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1592 {
1593         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1594         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
1595         kfree(ubufs);
1596 }
1597
1598 void vhost_zerocopy_callback(void *arg)
1599 {
1600         struct ubuf_info *ubuf = arg;
1601         struct vhost_ubuf_ref *ubufs = ubuf->arg;
1602         struct vhost_virtqueue *vq = ubufs->vq;
1603
1604         /* set len = 1 to mark this desc buffers done DMA */
1605         vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
1606         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1607 }