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