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 int init_used(struct vhost_virtqueue *vq,
 633                      struct vring_used __user *used)
 634 {
 635         int r = put_user(vq->used_flags, &used->flags);
 636
 637         if (r)
 638                 return r;
 639         vq->signalled_used_valid = false;
 640         return get_user(vq->last_used_idx, &used->idx);
 641 }
 642
 643 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
 644 {
 645         struct file *eventfp, *filep = NULL,
 646                     *pollstart = NULL, *pollstop = NULL;
 647         struct eventfd_ctx *ctx = NULL;
 648         u32 __user *idxp = argp;
 649         struct vhost_virtqueue *vq;
 650         struct vhost_vring_state s;
 651         struct vhost_vring_file f;
 652         struct vhost_vring_addr a;
 653         u32 idx;
 654         long r;
 655
 656         r = get_user(idx, idxp);
 657         if (r < 0)
 658                 return r;
 659         if (idx >= d->nvqs)
 660                 return -ENOBUFS;
 661
 662         vq = d->vqs + idx;
 663
 664         mutex_lock(&vq->mutex);
 665
 666         switch (ioctl) {
 667         case VHOST_SET_VRING_NUM:
 668                 /* Resizing ring with an active backend?
 669                  * You don't want to do that. */
 670                 if (vq->private_data) {
 671                         r = -EBUSY;
 672                         break;
 673                 }
 674                 if (copy_from_user(&s, argp, sizeof s)) {
 675                         r = -EFAULT;
 676                         break;
 677                 }
 678                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 679                         r = -EINVAL;
 680                         break;
 681                 }
 682                 vq->num = s.num;
 683                 break;
 684         case VHOST_SET_VRING_BASE:
 685                 /* Moving base with an active backend?
 686                  * You don't want to do that. */
 687                 if (vq->private_data) {
 688                         r = -EBUSY;
 689                         break;
 690                 }
 691                 if (copy_from_user(&s, argp, sizeof s)) {
 692                         r = -EFAULT;
 693                         break;
 694                 }
 695                 if (s.num > 0xffff) {
 696                         r = -EINVAL;
 697                         break;
 698                 }
 699                 vq->last_avail_idx = s.num;
 700                 /* Forget the cached index value. */
 701                 vq->avail_idx = vq->last_avail_idx;
 702                 break;
 703         case VHOST_GET_VRING_BASE:
 704                 s.index = idx;
 705                 s.num = vq->last_avail_idx;
 706                 if (copy_to_user(argp, &s, sizeof s))
 707                         r = -EFAULT;
 708                 break;
 709         case VHOST_SET_VRING_ADDR:
 710                 if (copy_from_user(&a, argp, sizeof a)) {
 711                         r = -EFAULT;
 712                         break;
 713                 }
 714                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 715                         r = -EOPNOTSUPP;
 716                         break;
 717                 }
 718                 /* For 32bit, verify that the top 32bits of the user
 719                    data are set to zero. */
 720                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 721                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 722                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 723                         r = -EFAULT;
 724                         break;
 725                 }
 726                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 727                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 728                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 729                         r = -EINVAL;
 730                         break;
 731                 }
 732
 733                 /* We only verify access here if backend is configured.
 734                  * If it is not, we don't as size might not have been setup.
 735                  * We will verify when backend is configured. */
 736                 if (vq->private_data) {
 737                         if (!vq_access_ok(d, vq->num,
 738                                 (void __user *)(unsigned long)a.desc_user_addr,
 739                                 (void __user *)(unsigned long)a.avail_user_addr,
 740                                 (void __user *)(unsigned long)a.used_user_addr)) {
 741                                 r = -EINVAL;
 742                                 break;
 743                         }
 744
 745                         /* Also validate log access for used ring if enabled. */
 746                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 747                             !log_access_ok(vq->log_base, a.log_guest_addr,
 748                                            sizeof *vq->used +
 749                                            vq->num * sizeof *vq->used->ring)) {
 750                                 r = -EINVAL;
 751                                 break;
 752                         }
 753                 }
 754
 755                 r = init_used(vq, (struct vring_used __user *)(unsigned long)
 756                               a.used_user_addr);
 757                 if (r)
 758                         break;
 759                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 760                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 761                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 762                 vq->log_addr = a.log_guest_addr;
 763                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 764                 break;
 765         case VHOST_SET_VRING_KICK:
 766                 if (copy_from_user(&f, argp, sizeof f)) {
 767                         r = -EFAULT;
 768                         break;
 769                 }
 770                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 771                 if (IS_ERR(eventfp)) {
 772                         r = PTR_ERR(eventfp);
 773                         break;
 774                 }
 775                 if (eventfp != vq->kick) {
 776                         pollstop = filep = vq->kick;
 777                         pollstart = vq->kick = eventfp;
 778                 } else
 779                         filep = eventfp;
 780                 break;
 781         case VHOST_SET_VRING_CALL:
 782                 if (copy_from_user(&f, argp, sizeof f)) {
 783                         r = -EFAULT;
 784                         break;
 785                 }
 786                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 787                 if (IS_ERR(eventfp)) {
 788                         r = PTR_ERR(eventfp);
 789                         break;
 790                 }
 791                 if (eventfp != vq->call) {
 792                         filep = vq->call;
 793                         ctx = vq->call_ctx;
 794                         vq->call = eventfp;
 795                         vq->call_ctx = eventfp ?
 796                                 eventfd_ctx_fileget(eventfp) : NULL;
 797                 } else
 798                         filep = eventfp;
 799                 break;
 800         case VHOST_SET_VRING_ERR:
 801                 if (copy_from_user(&f, argp, sizeof f)) {
 802                         r = -EFAULT;
 803                         break;
 804                 }
 805                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 806                 if (IS_ERR(eventfp)) {
 807                         r = PTR_ERR(eventfp);
 808                         break;
 809                 }
 810                 if (eventfp != vq->error) {
 811                         filep = vq->error;
 812                         vq->error = eventfp;
 813                         ctx = vq->error_ctx;
 814                         vq->error_ctx = eventfp ?
 815                                 eventfd_ctx_fileget(eventfp) : NULL;
 816                 } else
 817                         filep = eventfp;
 818                 break;
 819         default:
 820                 r = -ENOIOCTLCMD;
 821         }
 822
 823         if (pollstop && vq->handle_kick)
 824                 vhost_poll_stop(&vq->poll);
 825
 826         if (ctx)
 827                 eventfd_ctx_put(ctx);
 828         if (filep)
 829                 fput(filep);
 830
 831         if (pollstart && vq->handle_kick)
 832                 vhost_poll_start(&vq->poll, vq->kick);
 833
 834         mutex_unlock(&vq->mutex);
 835
 836         if (pollstop && vq->handle_kick)
 837                 vhost_poll_flush(&vq->poll);
 838         return r;
 839 }
 840
 841 /* Caller must have device mutex */
 842 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
 843 {
 844         void __user *argp = (void __user *)arg;
 845         struct file *eventfp, *filep = NULL;
 846         struct eventfd_ctx *ctx = NULL;
 847         u64 p;
 848         long r;
 849         int i, fd;
 850
 851         /* If you are not the owner, you can become one */
 852         if (ioctl == VHOST_SET_OWNER) {
 853                 r = vhost_dev_set_owner(d);
 854                 goto done;
 855         }
 856
 857         /* You must be the owner to do anything else */
 858         r = vhost_dev_check_owner(d);
 859         if (r)
 860                 goto done;
 861
 862         switch (ioctl) {
 863         case VHOST_SET_MEM_TABLE:
 864                 r = vhost_set_memory(d, argp);
 865                 break;
 866         case VHOST_SET_LOG_BASE:
 867                 if (copy_from_user(&p, argp, sizeof p)) {
 868                         r = -EFAULT;
 869                         break;
 870                 }
 871                 if ((u64)(unsigned long)p != p) {
 872                         r = -EFAULT;
 873                         break;
 874                 }
 875                 for (i = 0; i < d->nvqs; ++i) {
 876                         struct vhost_virtqueue *vq;
 877                         void __user *base = (void __user *)(unsigned long)p;
 878                         vq = d->vqs + i;
 879                         mutex_lock(&vq->mutex);
 880                         /* If ring is inactive, will check when it's enabled. */
 881                         if (vq->private_data && !vq_log_access_ok(d, vq, base))
 882                                 r = -EFAULT;
 883                         else
 884                                 vq->log_base = base;
 885                         mutex_unlock(&vq->mutex);
 886                 }
 887                 break;
 888         case VHOST_SET_LOG_FD:
 889                 r = get_user(fd, (int __user *)argp);
 890                 if (r < 0)
 891                         break;
 892                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 893                 if (IS_ERR(eventfp)) {
 894                         r = PTR_ERR(eventfp);
 895                         break;
 896                 }
 897                 if (eventfp != d->log_file) {
 898                         filep = d->log_file;
 899                         ctx = d->log_ctx;
 900                         d->log_ctx = eventfp ?
 901                                 eventfd_ctx_fileget(eventfp) : NULL;
 902                 } else
 903                         filep = eventfp;
 904                 for (i = 0; i < d->nvqs; ++i) {
 905                         mutex_lock(&d->vqs[i].mutex);
 906                         d->vqs[i].log_ctx = d->log_ctx;
 907                         mutex_unlock(&d->vqs[i].mutex);
 908                 }
 909                 if (ctx)
 910                         eventfd_ctx_put(ctx);
 911                 if (filep)
 912                         fput(filep);
 913                 break;
 914         default:
 915                 r = vhost_set_vring(d, ioctl, argp);
 916                 break;
 917         }
 918 done:
 919         return r;
 920 }
 921
 922 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 923                                                      __u64 addr, __u32 len)
 924 {
 925         struct vhost_memory_region *reg;
 926         int i;
 927
 928         /* linear search is not brilliant, but we really have on the order of 6
 929          * regions in practice */
 930         for (i = 0; i < mem->nregions; ++i) {
 931                 reg = mem->regions + i;
 932                 if (reg->guest_phys_addr <= addr &&
 933                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 934                         return reg;
 935         }
 936         return NULL;
 937 }
 938
 939 /* TODO: This is really inefficient.  We need something like get_user()
 940  * (instruction directly accesses the data, with an exception table entry
 941  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 942  */
 943 static int set_bit_to_user(int nr, void __user *addr)
 944 {
 945         unsigned long log = (unsigned long)addr;
 946         struct page *page;
 947         void *base;
 948         int bit = nr + (log % PAGE_SIZE) * 8;
 949         int r;
 950
 951         r = get_user_pages_fast(log, 1, 1, &page);
 952         if (r < 0)
 953                 return r;
 954         BUG_ON(r != 1);
 955         base = kmap_atomic(page, KM_USER0);
 956         set_bit(bit, base);
 957         kunmap_atomic(base, KM_USER0);
 958         set_page_dirty_lock(page);
 959         put_page(page);
 960         return 0;
 961 }
 962
 963 static int log_write(void __user *log_base,
 964                      u64 write_address, u64 write_length)
 965 {
 966         u64 write_page = write_address / VHOST_PAGE_SIZE;
 967         int r;
 968
 969         if (!write_length)
 970                 return 0;
 971         write_length += write_address % VHOST_PAGE_SIZE;
 972         for (;;) {
 973                 u64 base = (u64)(unsigned long)log_base;
 974                 u64 log = base + write_page / 8;
 975                 int bit = write_page % 8;
 976                 if ((u64)(unsigned long)log != log)
 977                         return -EFAULT;
 978                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 979                 if (r < 0)
 980                         return r;
 981                 if (write_length <= VHOST_PAGE_SIZE)
 982                         break;
 983                 write_length -= VHOST_PAGE_SIZE;
 984                 write_page += 1;
 985         }
 986         return r;
 987 }
 988
 989 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 990                     unsigned int log_num, u64 len)
 991 {
 992         int i, r;
 993
 994         /* Make sure data written is seen before log. */
 995         smp_wmb();
 996         for (i = 0; i < log_num; ++i) {
 997                 u64 l = min(log[i].len, len);
 998                 r = log_write(vq->log_base, log[i].addr, l);
 999                 if (r < 0)
1000                         return r;
1001                 len -= l;
1002                 if (!len) {
1003                         if (vq->log_ctx)
1004                                 eventfd_signal(vq->log_ctx, 1);
1005                         return 0;
1006                 }
1007         }
1008         /* Length written exceeds what we have stored. This is a bug. */
1009         BUG();
1010         return 0;
1011 }
1012
1013 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1014                           struct iovec iov[], int iov_size)
1015 {
1016         const struct vhost_memory_region *reg;
1017         struct vhost_memory *mem;
1018         struct iovec *_iov;
1019         u64 s = 0;
1020         int ret = 0;
1021
1022         rcu_read_lock();
1023
1024         mem = rcu_dereference(dev->memory);
1025         while ((u64)len > s) {
1026                 u64 size;
1027                 if (unlikely(ret >= iov_size)) {
1028                         ret = -ENOBUFS;
1029                         break;
1030                 }
1031                 reg = find_region(mem, addr, len);
1032                 if (unlikely(!reg)) {
1033                         ret = -EFAULT;
1034                         break;
1035                 }
1036                 _iov = iov + ret;
1037                 size = reg->memory_size - addr + reg->guest_phys_addr;
1038                 _iov->iov_len = min((u64)len, size);
1039                 _iov->iov_base = (void __user *)(unsigned long)
1040                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1041                 s += size;
1042                 addr += size;
1043                 ++ret;
1044         }
1045
1046         rcu_read_unlock();
1047         return ret;
1048 }
1049
1050 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1051  * function returns the next descriptor in the chain,
1052  * or -1U if we're at the end. */
1053 static unsigned next_desc(struct vring_desc *desc)
1054 {
1055         unsigned int next;
1056
1057         /* If this descriptor says it doesn't chain, we're done. */
1058         if (!(desc->flags & VRING_DESC_F_NEXT))
1059                 return -1U;
1060
1061         /* Check they're not leading us off end of descriptors. */
1062         next = desc->next;
1063         /* Make sure compiler knows to grab that: we don't want it changing! */
1064         /* We will use the result as an index in an array, so most
1065          * architectures only need a compiler barrier here. */
1066         read_barrier_depends();
1067
1068         return next;
1069 }
1070
1071 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1072                         struct iovec iov[], unsigned int iov_size,
1073                         unsigned int *out_num, unsigned int *in_num,
1074                         struct vhost_log *log, unsigned int *log_num,
1075                         struct vring_desc *indirect)
1076 {
1077         struct vring_desc desc;
1078         unsigned int i = 0, count, found = 0;
1079         int ret;
1080
1081         /* Sanity check */
1082         if (unlikely(indirect->len % sizeof desc)) {
1083                 vq_err(vq, "Invalid length in indirect descriptor: "
1084                        "len 0x%llx not multiple of 0x%zx\n",
1085                        (unsigned long long)indirect->len,
1086                        sizeof desc);
1087                 return -EINVAL;
1088         }
1089
1090         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1091                              UIO_MAXIOV);
1092         if (unlikely(ret < 0)) {
1093                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1094                 return ret;
1095         }
1096
1097         /* We will use the result as an address to read from, so most
1098          * architectures only need a compiler barrier here. */
1099         read_barrier_depends();
1100
1101         count = indirect->len / sizeof desc;
1102         /* Buffers are chained via a 16 bit next field, so
1103          * we can have at most 2^16 of these. */
1104         if (unlikely(count > USHRT_MAX + 1)) {
1105                 vq_err(vq, "Indirect buffer length too big: %d\n",
1106                        indirect->len);
1107                 return -E2BIG;
1108         }
1109
1110         do {
1111                 unsigned iov_count = *in_num + *out_num;
1112                 if (unlikely(++found > count)) {
1113                         vq_err(vq, "Loop detected: last one at %u "
1114                                "indirect size %u\n",
1115                                i, count);
1116                         return -EINVAL;
1117                 }
1118                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1119                                               vq->indirect, sizeof desc))) {
1120                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1121                                i, (size_t)indirect->addr + i * sizeof desc);
1122                         return -EINVAL;
1123                 }
1124                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1125                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1126                                i, (size_t)indirect->addr + i * sizeof desc);
1127                         return -EINVAL;
1128                 }
1129
1130                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1131                                      iov_size - iov_count);
1132                 if (unlikely(ret < 0)) {
1133                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1134                                ret, i);
1135                         return ret;
1136                 }
1137                 /* If this is an input descriptor, increment that count. */
1138                 if (desc.flags & VRING_DESC_F_WRITE) {
1139                         *in_num += ret;
1140                         if (unlikely(log)) {
1141                                 log[*log_num].addr = desc.addr;
1142                                 log[*log_num].len = desc.len;
1143                                 ++*log_num;
1144                         }
1145                 } else {
1146                         /* If it's an output descriptor, they're all supposed
1147                          * to come before any input descriptors. */
1148                         if (unlikely(*in_num)) {
1149                                 vq_err(vq, "Indirect descriptor "
1150                                        "has out after in: idx %d\n", i);
1151                                 return -EINVAL;
1152                         }
1153                         *out_num += ret;
1154                 }
1155         } while ((i = next_desc(&desc)) != -1);
1156         return 0;
1157 }
1158
1159 /* This looks in the virtqueue and for the first available buffer, and converts
1160  * it to an iovec for convenient access.  Since descriptors consist of some
1161  * number of output then some number of input descriptors, it's actually two
1162  * iovecs, but we pack them into one and note how many of each there were.
1163  *
1164  * This function returns the descriptor number found, or vq->num (which is
1165  * never a valid descriptor number) if none was found.  A negative code is
1166  * returned on error. */
1167 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1168                       struct iovec iov[], unsigned int iov_size,
1169                       unsigned int *out_num, unsigned int *in_num,
1170                       struct vhost_log *log, unsigned int *log_num)
1171 {
1172         struct vring_desc desc;
1173         unsigned int i, head, found = 0;
1174         u16 last_avail_idx;
1175         int ret;
1176
1177         /* Check it isn't doing very strange things with descriptor numbers. */
1178         last_avail_idx = vq->last_avail_idx;
1179         if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1180                 vq_err(vq, "Failed to access avail idx at %p\n",
1181                        &vq->avail->idx);
1182                 return -EFAULT;
1183         }
1184
1185         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1186                 vq_err(vq, "Guest moved used index from %u to %u",
1187                        last_avail_idx, vq->avail_idx);
1188                 return -EFAULT;
1189         }
1190
1191         /* If there's nothing new since last we looked, return invalid. */
1192         if (vq->avail_idx == last_avail_idx)
1193                 return vq->num;
1194
1195         /* Only get avail ring entries after they have been exposed by guest. */
1196         smp_rmb();
1197
1198         /* Grab the next descriptor number they're advertising, and increment
1199          * the index we've seen. */
1200         if (unlikely(__get_user(head,
1201                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1202                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1203                        last_avail_idx,
1204                        &vq->avail->ring[last_avail_idx % vq->num]);
1205                 return -EFAULT;
1206         }
1207
1208         /* If their number is silly, that's an error. */
1209         if (unlikely(head >= vq->num)) {
1210                 vq_err(vq, "Guest says index %u > %u is available",
1211                        head, vq->num);
1212                 return -EINVAL;
1213         }
1214
1215         /* When we start there are none of either input nor output. */
1216         *out_num = *in_num = 0;
1217         if (unlikely(log))
1218                 *log_num = 0;
1219
1220         i = head;
1221         do {
1222                 unsigned iov_count = *in_num + *out_num;
1223                 if (unlikely(i >= vq->num)) {
1224                         vq_err(vq, "Desc index is %u > %u, head = %u",
1225                                i, vq->num, head);
1226                         return -EINVAL;
1227                 }
1228                 if (unlikely(++found > vq->num)) {
1229                         vq_err(vq, "Loop detected: last one at %u "
1230                                "vq size %u head %u\n",
1231                                i, vq->num, head);
1232                         return -EINVAL;
1233                 }
1234                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1235                 if (unlikely(ret)) {
1236                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1237                                i, vq->desc + i);
1238                         return -EFAULT;
1239                 }
1240                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1241                         ret = get_indirect(dev, vq, iov, iov_size,
1242                                            out_num, in_num,
1243                                            log, log_num, &desc);
1244                         if (unlikely(ret < 0)) {
1245                                 vq_err(vq, "Failure detected "
1246                                        "in indirect descriptor at idx %d\n", i);
1247                                 return ret;
1248                         }
1249                         continue;
1250                 }
1251
1252                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1253                                      iov_size - iov_count);
1254                 if (unlikely(ret < 0)) {
1255                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1256                                ret, i);
1257                         return ret;
1258                 }
1259                 if (desc.flags & VRING_DESC_F_WRITE) {
1260                         /* If this is an input descriptor,
1261                          * increment that count. */
1262                         *in_num += ret;
1263                         if (unlikely(log)) {
1264                                 log[*log_num].addr = desc.addr;
1265                                 log[*log_num].len = desc.len;
1266                                 ++*log_num;
1267                         }
1268                 } else {
1269                         /* If it's an output descriptor, they're all supposed
1270                          * to come before any input descriptors. */
1271                         if (unlikely(*in_num)) {
1272                                 vq_err(vq, "Descriptor has out after in: "
1273                                        "idx %d\n", i);
1274                                 return -EINVAL;
1275                         }
1276                         *out_num += ret;
1277                 }
1278         } while ((i = next_desc(&desc)) != -1);
1279
1280         /* On success, increment avail index. */
1281         vq->last_avail_idx++;
1282
1283         /* Assume notifications from guest are disabled at this point,
1284          * if they aren't we would need to update avail_event index. */
1285         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1286         return head;
1287 }
1288
1289 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1290 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1291 {
1292         vq->last_avail_idx -= n;
1293 }
1294
1295 /* After we've used one of their buffers, we tell them about it.  We'll then
1296  * want to notify the guest, using eventfd. */
1297 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1298 {
1299         struct vring_used_elem __user *used;
1300
1301         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1302          * next entry in that used ring. */
1303         used = &vq->used->ring[vq->last_used_idx % vq->num];
1304         if (__put_user(head, &used->id)) {
1305                 vq_err(vq, "Failed to write used id");
1306                 return -EFAULT;
1307         }
1308         if (__put_user(len, &used->len)) {
1309                 vq_err(vq, "Failed to write used len");
1310                 return -EFAULT;
1311         }
1312         /* Make sure buffer is written before we update index. */
1313         smp_wmb();
1314         if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1315                 vq_err(vq, "Failed to increment used idx");
1316                 return -EFAULT;
1317         }
1318         if (unlikely(vq->log_used)) {
1319                 /* Make sure data is seen before log. */
1320                 smp_wmb();
1321                 /* Log used ring entry write. */
1322                 log_write(vq->log_base,
1323                           vq->log_addr +
1324                            ((void __user *)used - (void __user *)vq->used),
1325                           sizeof *used);
1326                 /* Log used index update. */
1327                 log_write(vq->log_base,
1328                           vq->log_addr + offsetof(struct vring_used, idx),
1329                           sizeof vq->used->idx);
1330                 if (vq->log_ctx)
1331                         eventfd_signal(vq->log_ctx, 1);
1332         }
1333         vq->last_used_idx++;
1334         /* If the driver never bothers to signal in a very long while,
1335          * used index might wrap around. If that happens, invalidate
1336          * signalled_used index we stored. TODO: make sure driver
1337          * signals at least once in 2^16 and remove this. */
1338         if (unlikely(vq->last_used_idx == vq->signalled_used))
1339                 vq->signalled_used_valid = false;
1340         return 0;
1341 }
1342
1343 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1344                             struct vring_used_elem *heads,
1345                             unsigned count)
1346 {
1347         struct vring_used_elem __user *used;
1348         u16 old, new;
1349         int start;
1350
1351         start = vq->last_used_idx % vq->num;
1352         used = vq->used->ring + start;
1353         if (__copy_to_user(used, heads, count * sizeof *used)) {
1354                 vq_err(vq, "Failed to write used");
1355                 return -EFAULT;
1356         }
1357         if (unlikely(vq->log_used)) {
1358                 /* Make sure data is seen before log. */
1359                 smp_wmb();
1360                 /* Log used ring entry write. */
1361                 log_write(vq->log_base,
1362                           vq->log_addr +
1363                            ((void __user *)used - (void __user *)vq->used),
1364                           count * sizeof *used);
1365         }
1366         old = vq->last_used_idx;
1367         new = (vq->last_used_idx += count);
1368         /* If the driver never bothers to signal in a very long while,
1369          * used index might wrap around. If that happens, invalidate
1370          * signalled_used index we stored. TODO: make sure driver
1371          * signals at least once in 2^16 and remove this. */
1372         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1373                 vq->signalled_used_valid = false;
1374         return 0;
1375 }
1376
1377 /* After we've used one of their buffers, we tell them about it.  We'll then
1378  * want to notify the guest, using eventfd. */
1379 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1380                      unsigned count)
1381 {
1382         int start, n, r;
1383
1384         start = vq->last_used_idx % vq->num;
1385         n = vq->num - start;
1386         if (n < count) {
1387                 r = __vhost_add_used_n(vq, heads, n);
1388                 if (r < 0)
1389                         return r;
1390                 heads += n;
1391                 count -= n;
1392         }
1393         r = __vhost_add_used_n(vq, heads, count);
1394
1395         /* Make sure buffer is written before we update index. */
1396         smp_wmb();
1397         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1398                 vq_err(vq, "Failed to increment used idx");
1399                 return -EFAULT;
1400         }
1401         if (unlikely(vq->log_used)) {
1402                 /* Log used index update. */
1403                 log_write(vq->log_base,
1404                           vq->log_addr + offsetof(struct vring_used, idx),
1405                           sizeof vq->used->idx);
1406                 if (vq->log_ctx)
1407                         eventfd_signal(vq->log_ctx, 1);
1408         }
1409         return r;
1410 }
1411
1412 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1413 {
1414         __u16 old, new, event;
1415         bool v;
1416         /* Flush out used index updates. This is paired
1417          * with the barrier that the Guest executes when enabling
1418          * interrupts. */
1419         smp_mb();
1420
1421         if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1422             unlikely(vq->avail_idx == vq->last_avail_idx))
1423                 return true;
1424
1425         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1426                 __u16 flags;
1427                 if (__get_user(flags, &vq->avail->flags)) {
1428                         vq_err(vq, "Failed to get flags");
1429                         return true;
1430                 }
1431                 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1432         }
1433         old = vq->signalled_used;
1434         v = vq->signalled_used_valid;
1435         new = vq->signalled_used = vq->last_used_idx;
1436         vq->signalled_used_valid = true;
1437
1438         if (unlikely(!v))
1439                 return true;
1440
1441         if (get_user(event, vhost_used_event(vq))) {
1442                 vq_err(vq, "Failed to get used event idx");
1443                 return true;
1444         }
1445         return vring_need_event(event, new, old);
1446 }
1447
1448 /* This actually signals the guest, using eventfd. */
1449 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1450 {
1451         /* Signal the Guest tell them we used something up. */
1452         if (vq->call_ctx && vhost_notify(dev, vq))
1453                 eventfd_signal(vq->call_ctx, 1);
1454 }
1455
1456 /* And here's the combo meal deal.  Supersize me! */
1457 void vhost_add_used_and_signal(struct vhost_dev *dev,
1458                                struct vhost_virtqueue *vq,
1459                                unsigned int head, int len)
1460 {
1461         vhost_add_used(vq, head, len);
1462         vhost_signal(dev, vq);
1463 }
1464
1465 /* multi-buffer version of vhost_add_used_and_signal */
1466 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1467                                  struct vhost_virtqueue *vq,
1468                                  struct vring_used_elem *heads, unsigned count)
1469 {
1470         vhost_add_used_n(vq, heads, count);
1471         vhost_signal(dev, vq);
1472 }
1473
1474 /* OK, now we need to know about added descriptors. */
1475 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1476 {
1477         u16 avail_idx;
1478         int r;
1479
1480         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1481                 return false;
1482         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1483         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1484                 r = put_user(vq->used_flags, &vq->used->flags);
1485                 if (r) {
1486                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1487                                &vq->used->flags, r);
1488                         return false;
1489                 }
1490         } else {
1491                 r = put_user(vq->avail_idx, vhost_avail_event(vq));
1492                 if (r) {
1493                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1494                                vhost_avail_event(vq), r);
1495                         return false;
1496                 }
1497         }
1498         if (unlikely(vq->log_used)) {
1499                 void __user *used;
1500                 /* Make sure data is seen before log. */
1501                 smp_wmb();
1502                 used = vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX) ?
1503                         &vq->used->flags : vhost_avail_event(vq);
1504                 /* Log used flags or event index entry write. Both are 16 bit
1505                  * fields. */
1506                 log_write(vq->log_base, vq->log_addr +
1507                            (used - (void __user *)vq->used),
1508                           sizeof(u16));
1509                 if (vq->log_ctx)
1510                         eventfd_signal(vq->log_ctx, 1);
1511         }
1512         /* They could have slipped one in as we were doing that: make
1513          * sure it's written, then check again. */
1514         smp_mb();
1515         r = __get_user(avail_idx, &vq->avail->idx);
1516         if (r) {
1517                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1518                        &vq->avail->idx, r);
1519                 return false;
1520         }
1521
1522         return avail_idx != vq->avail_idx;
1523 }
1524
1525 /* We don't need to be notified again. */
1526 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1527 {
1528         int r;
1529
1530         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1531                 return;
1532         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1533         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1534                 r = put_user(vq->used_flags, &vq->used->flags);
1535                 if (r)
1536                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1537                                &vq->used->flags, r);
1538         }
1539 }
1540
1541 static void vhost_zerocopy_done_signal(struct kref *kref)
1542 {
1543         struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1544                                                     kref);
1545         wake_up(&ubufs->wait);
1546 }
1547
1548 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1549                                         bool zcopy)
1550 {
1551         struct vhost_ubuf_ref *ubufs;
1552         /* No zero copy backend? Nothing to count. */
1553         if (!zcopy)
1554                 return NULL;
1555         ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1556         if (!ubufs)
1557                 return ERR_PTR(-ENOMEM);
1558         kref_init(&ubufs->kref);
1559         kref_get(&ubufs->kref);
1560         init_waitqueue_head(&ubufs->wait);
1561         ubufs->vq = vq;
1562         return ubufs;
1563 }
1564
1565 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1566 {
1567         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1568 }
1569
1570 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1571 {
1572         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1573         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
1574         kfree(ubufs);
1575 }
1576
1577 void vhost_zerocopy_callback(void *arg)
1578 {
1579         struct ubuf_info *ubuf = arg;
1580         struct vhost_ubuf_ref *ubufs = ubuf->arg;
1581         struct vhost_virtqueue *vq = ubufs->vq;
1582
1583         /* set len = 1 to mark this desc buffers done DMA */
1584         vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
1585         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1586 }