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 <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 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 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 141 {
 142         unsigned long flags;
 143
 144         spin_lock_irqsave(&dev->work_lock, flags);
 145         if (list_empty(&work->node)) {
 146                 list_add_tail(&work->node, &dev->work_list);
 147                 work->queue_seq++;
 148                 wake_up_process(dev->worker);
 149         }
 150         spin_unlock_irqrestore(&dev->work_lock, flags);
 151 }
 152
 153 void vhost_poll_queue(struct vhost_poll *poll)
 154 {
 155         vhost_work_queue(poll->dev, &poll->work);
 156 }
 157
 158 static void vhost_vq_reset(struct vhost_dev *dev,
 159                            struct vhost_virtqueue *vq)
 160 {
 161         vq->num = 1;
 162         vq->desc = NULL;
 163         vq->avail = NULL;
 164         vq->used = NULL;
 165         vq->last_avail_idx = 0;
 166         vq->avail_idx = 0;
 167         vq->last_used_idx = 0;
 168         vq->signalled_used = 0;
 169         vq->signalled_used_valid = false;
 170         vq->used_flags = 0;
 171         vq->log_used = false;
 172         vq->log_addr = -1ull;
 173         vq->vhost_hlen = 0;
 174         vq->sock_hlen = 0;
 175         vq->private_data = NULL;
 176         vq->log_base = NULL;
 177         vq->error_ctx = NULL;
 178         vq->error = NULL;
 179         vq->kick = NULL;
 180         vq->call_ctx = NULL;
 181         vq->call = NULL;
 182         vq->log_ctx = NULL;
 183         vq->upend_idx = 0;
 184         vq->done_idx = 0;
 185         vq->ubufs = NULL;
 186 }
 187
 188 static int vhost_worker(void *data)
 189 {
 190         struct vhost_dev *dev = data;
 191         struct vhost_work *work = NULL;
 192         unsigned uninitialized_var(seq);
 193         mm_segment_t oldfs = get_fs();
 194
 195         set_fs(USER_DS);
 196         use_mm(dev->mm);
 197
 198         for (;;) {
 199                 /* mb paired w/ kthread_stop */
 200                 set_current_state(TASK_INTERRUPTIBLE);
 201
 202                 spin_lock_irq(&dev->work_lock);
 203                 if (work) {
 204                         work->done_seq = seq;
 205                         if (work->flushing)
 206                                 wake_up_all(&work->done);
 207                 }
 208
 209                 if (kthread_should_stop()) {
 210                         spin_unlock_irq(&dev->work_lock);
 211                         __set_current_state(TASK_RUNNING);
 212                         break;
 213                 }
 214                 if (!list_empty(&dev->work_list)) {
 215                         work = list_first_entry(&dev->work_list,
 216                                                 struct vhost_work, node);
 217                         list_del_init(&work->node);
 218                         seq = work->queue_seq;
 219                 } else
 220                         work = NULL;
 221                 spin_unlock_irq(&dev->work_lock);
 222
 223                 if (work) {
 224                         __set_current_state(TASK_RUNNING);
 225                         work->fn(work);
 226                         if (need_resched())
 227                                 schedule();
 228                 } else
 229                         schedule();
 230
 231         }
 232         unuse_mm(dev->mm);
 233         set_fs(oldfs);
 234         return 0;
 235 }
 236
 237 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 238 {
 239         kfree(vq->indirect);
 240         vq->indirect = NULL;
 241         kfree(vq->log);
 242         vq->log = NULL;
 243         kfree(vq->heads);
 244         vq->heads = NULL;
 245         kfree(vq->ubuf_info);
 246         vq->ubuf_info = NULL;
 247 }
 248
 249 void vhost_enable_zcopy(int vq)
 250 {
 251         vhost_zcopy_mask |= 0x1 << vq;
 252 }
 253
 254 /* Helper to allocate iovec buffers for all vqs. */
 255 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 256 {
 257         int i;
 258         bool zcopy;
 259
 260         for (i = 0; i < dev->nvqs; ++i) {
 261                 dev->vqs[i].indirect = kmalloc(sizeof *dev->vqs[i].indirect *
 262                                                UIO_MAXIOV, GFP_KERNEL);
 263                 dev->vqs[i].log = kmalloc(sizeof *dev->vqs[i].log * UIO_MAXIOV,
 264                                           GFP_KERNEL);
 265                 dev->vqs[i].heads = kmalloc(sizeof *dev->vqs[i].heads *
 266                                             UIO_MAXIOV, GFP_KERNEL);
 267                 zcopy = vhost_zcopy_mask & (0x1 << i);
 268                 if (zcopy)
 269                         dev->vqs[i].ubuf_info =
 270                                 kmalloc(sizeof *dev->vqs[i].ubuf_info *
 271                                         UIO_MAXIOV, GFP_KERNEL);
 272                 if (!dev->vqs[i].indirect || !dev->vqs[i].log ||
 273                         !dev->vqs[i].heads ||
 274                         (zcopy && !dev->vqs[i].ubuf_info))
 275                         goto err_nomem;
 276         }
 277         return 0;
 278
 279 err_nomem:
 280         for (; i >= 0; --i)
 281                 vhost_vq_free_iovecs(&dev->vqs[i]);
 282         return -ENOMEM;
 283 }
 284
 285 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 286 {
 287         int i;
 288
 289         for (i = 0; i < dev->nvqs; ++i)
 290                 vhost_vq_free_iovecs(&dev->vqs[i]);
 291 }
 292
 293 long vhost_dev_init(struct vhost_dev *dev,
 294                     struct vhost_virtqueue *vqs, int nvqs)
 295 {
 296         int i;
 297
 298         dev->vqs = vqs;
 299         dev->nvqs = nvqs;
 300         mutex_init(&dev->mutex);
 301         dev->log_ctx = NULL;
 302         dev->log_file = NULL;
 303         dev->memory = NULL;
 304         dev->mm = NULL;
 305         spin_lock_init(&dev->work_lock);
 306         INIT_LIST_HEAD(&dev->work_list);
 307         dev->worker = NULL;
 308
 309         for (i = 0; i < dev->nvqs; ++i) {
 310                 dev->vqs[i].log = NULL;
 311                 dev->vqs[i].indirect = NULL;
 312                 dev->vqs[i].heads = NULL;
 313                 dev->vqs[i].ubuf_info = NULL;
 314                 dev->vqs[i].dev = dev;
 315                 mutex_init(&dev->vqs[i].mutex);
 316                 vhost_vq_reset(dev, dev->vqs + i);
 317                 if (dev->vqs[i].handle_kick)
 318                         vhost_poll_init(&dev->vqs[i].poll,
 319                                         dev->vqs[i].handle_kick, POLLIN, dev);
 320         }
 321
 322         return 0;
 323 }
 324
 325 /* Caller should have device mutex */
 326 long vhost_dev_check_owner(struct vhost_dev *dev)
 327 {
 328         /* Are you the owner? If not, I don't think you mean to do that */
 329         return dev->mm == current->mm ? 0 : -EPERM;
 330 }
 331
 332 struct vhost_attach_cgroups_struct {
 333         struct vhost_work work;
 334         struct task_struct *owner;
 335         int ret;
 336 };
 337
 338 static void vhost_attach_cgroups_work(struct vhost_work *work)
 339 {
 340         struct vhost_attach_cgroups_struct *s;
 341
 342         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 343         s->ret = cgroup_attach_task_all(s->owner, current);
 344 }
 345
 346 static int vhost_attach_cgroups(struct vhost_dev *dev)
 347 {
 348         struct vhost_attach_cgroups_struct attach;
 349
 350         attach.owner = current;
 351         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 352         vhost_work_queue(dev, &attach.work);
 353         vhost_work_flush(dev, &attach.work);
 354         return attach.ret;
 355 }
 356
 357 /* Caller should have device mutex */
 358 static long vhost_dev_set_owner(struct vhost_dev *dev)
 359 {
 360         struct task_struct *worker;
 361         int err;
 362
 363         /* Is there an owner already? */
 364         if (dev->mm) {
 365                 err = -EBUSY;
 366                 goto err_mm;
 367         }
 368
 369         /* No owner, become one */
 370         dev->mm = get_task_mm(current);
 371         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 372         if (IS_ERR(worker)) {
 373                 err = PTR_ERR(worker);
 374                 goto err_worker;
 375         }
 376
 377         dev->worker = worker;
 378         wake_up_process(worker);        /* avoid contributing to loadavg */
 379
 380         err = vhost_attach_cgroups(dev);
 381         if (err)
 382                 goto err_cgroup;
 383
 384         err = vhost_dev_alloc_iovecs(dev);
 385         if (err)
 386                 goto err_cgroup;
 387
 388         return 0;
 389 err_cgroup:
 390         kthread_stop(worker);
 391         dev->worker = NULL;
 392 err_worker:
 393         if (dev->mm)
 394                 mmput(dev->mm);
 395         dev->mm = NULL;
 396 err_mm:
 397         return err;
 398 }
 399
 400 /* Caller should have device mutex */
 401 long vhost_dev_reset_owner(struct vhost_dev *dev)
 402 {
 403         struct vhost_memory *memory;
 404
 405         /* Restore memory to default empty mapping. */
 406         memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 407         if (!memory)
 408                 return -ENOMEM;
 409
 410         vhost_dev_cleanup(dev, true);
 411
 412         memory->nregions = 0;
 413         RCU_INIT_POINTER(dev->memory, memory);
 414         return 0;
 415 }
 416
 417 /* In case of DMA done not in order in lower device driver for some reason.
 418  * upend_idx is used to track end of used idx, done_idx is used to track head
 419  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 420  * guest used idx.
 421  */
 422 int vhost_zerocopy_signal_used(struct vhost_virtqueue *vq)
 423 {
 424         int i;
 425         int j = 0;
 426
 427         for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 428                 if ((vq->heads[i].len == VHOST_DMA_DONE_LEN)) {
 429                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 430                         vhost_add_used_and_signal(vq->dev, vq,
 431                                                   vq->heads[i].id, 0);
 432                         ++j;
 433                 } else
 434                         break;
 435         }
 436         if (j)
 437                 vq->done_idx = i;
 438         return j;
 439 }
 440
 441 /* Caller should have device mutex if and only if locked is set */
 442 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 443 {
 444         int i;
 445
 446         for (i = 0; i < dev->nvqs; ++i) {
 447                 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
 448                         vhost_poll_stop(&dev->vqs[i].poll);
 449                         vhost_poll_flush(&dev->vqs[i].poll);
 450                 }
 451                 /* Wait for all lower device DMAs done. */
 452                 if (dev->vqs[i].ubufs)
 453                         vhost_ubuf_put_and_wait(dev->vqs[i].ubufs);
 454
 455                 /* Signal guest as appropriate. */
 456                 vhost_zerocopy_signal_used(&dev->vqs[i]);
 457
 458                 if (dev->vqs[i].error_ctx)
 459                         eventfd_ctx_put(dev->vqs[i].error_ctx);
 460                 if (dev->vqs[i].error)
 461                         fput(dev->vqs[i].error);
 462                 if (dev->vqs[i].kick)
 463                         fput(dev->vqs[i].kick);
 464                 if (dev->vqs[i].call_ctx)
 465                         eventfd_ctx_put(dev->vqs[i].call_ctx);
 466                 if (dev->vqs[i].call)
 467                         fput(dev->vqs[i].call);
 468                 vhost_vq_reset(dev, dev->vqs + i);
 469         }
 470         vhost_dev_free_iovecs(dev);
 471         if (dev->log_ctx)
 472                 eventfd_ctx_put(dev->log_ctx);
 473         dev->log_ctx = NULL;
 474         if (dev->log_file)
 475                 fput(dev->log_file);
 476         dev->log_file = NULL;
 477         /* No one will access memory at this point */
 478         kfree(rcu_dereference_protected(dev->memory,
 479                                         locked ==
 480                                                 lockdep_is_held(&dev->mutex)));
 481         RCU_INIT_POINTER(dev->memory, NULL);
 482         WARN_ON(!list_empty(&dev->work_list));
 483         if (dev->worker) {
 484                 kthread_stop(dev->worker);
 485                 dev->worker = NULL;
 486         }
 487         if (dev->mm)
 488                 mmput(dev->mm);
 489         dev->mm = NULL;
 490 }
 491
 492 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 493 {
 494         u64 a = addr / VHOST_PAGE_SIZE / 8;
 495
 496         /* Make sure 64 bit math will not overflow. */
 497         if (a > ULONG_MAX - (unsigned long)log_base ||
 498             a + (unsigned long)log_base > ULONG_MAX)
 499                 return 0;
 500
 501         return access_ok(VERIFY_WRITE, log_base + a,
 502                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 503 }
 504
 505 /* Caller should have vq mutex and device mutex. */
 506 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 507                                int log_all)
 508 {
 509         int i;
 510
 511         if (!mem)
 512                 return 0;
 513
 514         for (i = 0; i < mem->nregions; ++i) {
 515                 struct vhost_memory_region *m = mem->regions + i;
 516                 unsigned long a = m->userspace_addr;
 517                 if (m->memory_size > ULONG_MAX)
 518                         return 0;
 519                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 520                                     m->memory_size))
 521                         return 0;
 522                 else if (log_all && !log_access_ok(log_base,
 523                                                    m->guest_phys_addr,
 524                                                    m->memory_size))
 525                         return 0;
 526         }
 527         return 1;
 528 }
 529
 530 /* Can we switch to this memory table? */
 531 /* Caller should have device mutex but not vq mutex */
 532 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 533                             int log_all)
 534 {
 535         int i;
 536
 537         for (i = 0; i < d->nvqs; ++i) {
 538                 int ok;
 539                 mutex_lock(&d->vqs[i].mutex);
 540                 /* If ring is inactive, will check when it's enabled. */
 541                 if (d->vqs[i].private_data)
 542                         ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
 543                                                  log_all);
 544                 else
 545                         ok = 1;
 546                 mutex_unlock(&d->vqs[i].mutex);
 547                 if (!ok)
 548                         return 0;
 549         }
 550         return 1;
 551 }
 552
 553 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
 554                         struct vring_desc __user *desc,
 555                         struct vring_avail __user *avail,
 556                         struct vring_used __user *used)
 557 {
 558         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 559         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 560                access_ok(VERIFY_READ, avail,
 561                          sizeof *avail + num * sizeof *avail->ring + s) &&
 562                access_ok(VERIFY_WRITE, used,
 563                         sizeof *used + num * sizeof *used->ring + s);
 564 }
 565
 566 /* Can we log writes? */
 567 /* Caller should have device mutex but not vq mutex */
 568 int vhost_log_access_ok(struct vhost_dev *dev)
 569 {
 570         struct vhost_memory *mp;
 571
 572         mp = rcu_dereference_protected(dev->memory,
 573                                        lockdep_is_held(&dev->mutex));
 574         return memory_access_ok(dev, mp, 1);
 575 }
 576
 577 /* Verify access for write logging. */
 578 /* Caller should have vq mutex and device mutex */
 579 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
 580                             void __user *log_base)
 581 {
 582         struct vhost_memory *mp;
 583         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 584
 585         mp = rcu_dereference_protected(vq->dev->memory,
 586                                        lockdep_is_held(&vq->mutex));
 587         return vq_memory_access_ok(log_base, mp,
 588                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
 589                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 590                                         sizeof *vq->used +
 591                                         vq->num * sizeof *vq->used->ring + s));
 592 }
 593
 594 /* Can we start vq? */
 595 /* Caller should have vq mutex and device mutex */
 596 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 597 {
 598         return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
 599                 vq_log_access_ok(vq->dev, vq, vq->log_base);
 600 }
 601
 602 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 603 {
 604         struct vhost_memory mem, *newmem, *oldmem;
 605         unsigned long size = offsetof(struct vhost_memory, regions);
 606
 607         if (copy_from_user(&mem, m, size))
 608                 return -EFAULT;
 609         if (mem.padding)
 610                 return -EOPNOTSUPP;
 611         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 612                 return -E2BIG;
 613         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 614         if (!newmem)
 615                 return -ENOMEM;
 616
 617         memcpy(newmem, &mem, size);
 618         if (copy_from_user(newmem->regions, m->regions,
 619                            mem.nregions * sizeof *m->regions)) {
 620                 kfree(newmem);
 621                 return -EFAULT;
 622         }
 623
 624         if (!memory_access_ok(d, newmem,
 625                               vhost_has_feature(d, VHOST_F_LOG_ALL))) {
 626                 kfree(newmem);
 627                 return -EFAULT;
 628         }
 629         oldmem = rcu_dereference_protected(d->memory,
 630                                            lockdep_is_held(&d->mutex));
 631         rcu_assign_pointer(d->memory, newmem);
 632         synchronize_rcu();
 633         kfree(oldmem);
 634         return 0;
 635 }
 636
 637 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
 638 {
 639         struct file *eventfp, *filep = NULL;
 640         bool pollstart = false, pollstop = false;
 641         struct eventfd_ctx *ctx = NULL;
 642         u32 __user *idxp = argp;
 643         struct vhost_virtqueue *vq;
 644         struct vhost_vring_state s;
 645         struct vhost_vring_file f;
 646         struct vhost_vring_addr a;
 647         u32 idx;
 648         long r;
 649
 650         r = get_user(idx, idxp);
 651         if (r < 0)
 652                 return r;
 653         if (idx >= d->nvqs)
 654                 return -ENOBUFS;
 655
 656         vq = d->vqs + idx;
 657
 658         mutex_lock(&vq->mutex);
 659
 660         switch (ioctl) {
 661         case VHOST_SET_VRING_NUM:
 662                 /* Resizing ring with an active backend?
 663                  * You don't want to do that. */
 664                 if (vq->private_data) {
 665                         r = -EBUSY;
 666                         break;
 667                 }
 668                 if (copy_from_user(&s, argp, sizeof s)) {
 669                         r = -EFAULT;
 670                         break;
 671                 }
 672                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 673                         r = -EINVAL;
 674                         break;
 675                 }
 676                 vq->num = s.num;
 677                 break;
 678         case VHOST_SET_VRING_BASE:
 679                 /* Moving base with an active backend?
 680                  * You don't want to do that. */
 681                 if (vq->private_data) {
 682                         r = -EBUSY;
 683                         break;
 684                 }
 685                 if (copy_from_user(&s, argp, sizeof s)) {
 686                         r = -EFAULT;
 687                         break;
 688                 }
 689                 if (s.num > 0xffff) {
 690                         r = -EINVAL;
 691                         break;
 692                 }
 693                 vq->last_avail_idx = s.num;
 694                 /* Forget the cached index value. */
 695                 vq->avail_idx = vq->last_avail_idx;
 696                 break;
 697         case VHOST_GET_VRING_BASE:
 698                 s.index = idx;
 699                 s.num = vq->last_avail_idx;
 700                 if (copy_to_user(argp, &s, sizeof s))
 701                         r = -EFAULT;
 702                 break;
 703         case VHOST_SET_VRING_ADDR:
 704                 if (copy_from_user(&a, argp, sizeof a)) {
 705                         r = -EFAULT;
 706                         break;
 707                 }
 708                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 709                         r = -EOPNOTSUPP;
 710                         break;
 711                 }
 712                 /* For 32bit, verify that the top 32bits of the user
 713                    data are set to zero. */
 714                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 715                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 716                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 717                         r = -EFAULT;
 718                         break;
 719                 }
 720                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 721                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 722                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 723                         r = -EINVAL;
 724                         break;
 725                 }
 726
 727                 /* We only verify access here if backend is configured.
 728                  * If it is not, we don't as size might not have been setup.
 729                  * We will verify when backend is configured. */
 730                 if (vq->private_data) {
 731                         if (!vq_access_ok(d, vq->num,
 732                                 (void __user *)(unsigned long)a.desc_user_addr,
 733                                 (void __user *)(unsigned long)a.avail_user_addr,
 734                                 (void __user *)(unsigned long)a.used_user_addr)) {
 735                                 r = -EINVAL;
 736                                 break;
 737                         }
 738
 739                         /* Also validate log access for used ring if enabled. */
 740                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 741                             !log_access_ok(vq->log_base, a.log_guest_addr,
 742                                            sizeof *vq->used +
 743                                            vq->num * sizeof *vq->used->ring)) {
 744                                 r = -EINVAL;
 745                                 break;
 746                         }
 747                 }
 748
 749                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 750                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 751                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 752                 vq->log_addr = a.log_guest_addr;
 753                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 754                 break;
 755         case VHOST_SET_VRING_KICK:
 756                 if (copy_from_user(&f, argp, sizeof f)) {
 757                         r = -EFAULT;
 758                         break;
 759                 }
 760                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 761                 if (IS_ERR(eventfp)) {
 762                         r = PTR_ERR(eventfp);
 763                         break;
 764                 }
 765                 if (eventfp != vq->kick) {
 766                         pollstop = (filep = vq->kick) != NULL;
 767                         pollstart = (vq->kick = eventfp) != NULL;
 768                 } else
 769                         filep = eventfp;
 770                 break;
 771         case VHOST_SET_VRING_CALL:
 772                 if (copy_from_user(&f, argp, sizeof f)) {
 773                         r = -EFAULT;
 774                         break;
 775                 }
 776                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 777                 if (IS_ERR(eventfp)) {
 778                         r = PTR_ERR(eventfp);
 779                         break;
 780                 }
 781                 if (eventfp != vq->call) {
 782                         filep = vq->call;
 783                         ctx = vq->call_ctx;
 784                         vq->call = eventfp;
 785                         vq->call_ctx = eventfp ?
 786                                 eventfd_ctx_fileget(eventfp) : NULL;
 787                 } else
 788                         filep = eventfp;
 789                 break;
 790         case VHOST_SET_VRING_ERR:
 791                 if (copy_from_user(&f, argp, sizeof f)) {
 792                         r = -EFAULT;
 793                         break;
 794                 }
 795                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 796                 if (IS_ERR(eventfp)) {
 797                         r = PTR_ERR(eventfp);
 798                         break;
 799                 }
 800                 if (eventfp != vq->error) {
 801                         filep = vq->error;
 802                         vq->error = eventfp;
 803                         ctx = vq->error_ctx;
 804                         vq->error_ctx = eventfp ?
 805                                 eventfd_ctx_fileget(eventfp) : NULL;
 806                 } else
 807                         filep = eventfp;
 808                 break;
 809         default:
 810                 r = -ENOIOCTLCMD;
 811         }
 812
 813         if (pollstop && vq->handle_kick)
 814                 vhost_poll_stop(&vq->poll);
 815
 816         if (ctx)
 817                 eventfd_ctx_put(ctx);
 818         if (filep)
 819                 fput(filep);
 820
 821         if (pollstart && vq->handle_kick)
 822                 vhost_poll_start(&vq->poll, vq->kick);
 823
 824         mutex_unlock(&vq->mutex);
 825
 826         if (pollstop && vq->handle_kick)
 827                 vhost_poll_flush(&vq->poll);
 828         return r;
 829 }
 830
 831 /* Caller must have device mutex */
 832 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
 833 {
 834         void __user *argp = (void __user *)arg;
 835         struct file *eventfp, *filep = NULL;
 836         struct eventfd_ctx *ctx = NULL;
 837         u64 p;
 838         long r;
 839         int i, fd;
 840
 841         /* If you are not the owner, you can become one */
 842         if (ioctl == VHOST_SET_OWNER) {
 843                 r = vhost_dev_set_owner(d);
 844                 goto done;
 845         }
 846
 847         /* You must be the owner to do anything else */
 848         r = vhost_dev_check_owner(d);
 849         if (r)
 850                 goto done;
 851
 852         switch (ioctl) {
 853         case VHOST_SET_MEM_TABLE:
 854                 r = vhost_set_memory(d, argp);
 855                 break;
 856         case VHOST_SET_LOG_BASE:
 857                 if (copy_from_user(&p, argp, sizeof p)) {
 858                         r = -EFAULT;
 859                         break;
 860                 }
 861                 if ((u64)(unsigned long)p != p) {
 862                         r = -EFAULT;
 863                         break;
 864                 }
 865                 for (i = 0; i < d->nvqs; ++i) {
 866                         struct vhost_virtqueue *vq;
 867                         void __user *base = (void __user *)(unsigned long)p;
 868                         vq = d->vqs + i;
 869                         mutex_lock(&vq->mutex);
 870                         /* If ring is inactive, will check when it's enabled. */
 871                         if (vq->private_data && !vq_log_access_ok(d, vq, base))
 872                                 r = -EFAULT;
 873                         else
 874                                 vq->log_base = base;
 875                         mutex_unlock(&vq->mutex);
 876                 }
 877                 break;
 878         case VHOST_SET_LOG_FD:
 879                 r = get_user(fd, (int __user *)argp);
 880                 if (r < 0)
 881                         break;
 882                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 883                 if (IS_ERR(eventfp)) {
 884                         r = PTR_ERR(eventfp);
 885                         break;
 886                 }
 887                 if (eventfp != d->log_file) {
 888                         filep = d->log_file;
 889                         ctx = d->log_ctx;
 890                         d->log_ctx = eventfp ?
 891                                 eventfd_ctx_fileget(eventfp) : NULL;
 892                 } else
 893                         filep = eventfp;
 894                 for (i = 0; i < d->nvqs; ++i) {
 895                         mutex_lock(&d->vqs[i].mutex);
 896                         d->vqs[i].log_ctx = d->log_ctx;
 897                         mutex_unlock(&d->vqs[i].mutex);
 898                 }
 899                 if (ctx)
 900                         eventfd_ctx_put(ctx);
 901                 if (filep)
 902                         fput(filep);
 903                 break;
 904         default:
 905                 r = vhost_set_vring(d, ioctl, argp);
 906                 break;
 907         }
 908 done:
 909         return r;
 910 }
 911
 912 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 913                                                      __u64 addr, __u32 len)
 914 {
 915         struct vhost_memory_region *reg;
 916         int i;
 917
 918         /* linear search is not brilliant, but we really have on the order of 6
 919          * regions in practice */
 920         for (i = 0; i < mem->nregions; ++i) {
 921                 reg = mem->regions + i;
 922                 if (reg->guest_phys_addr <= addr &&
 923                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 924                         return reg;
 925         }
 926         return NULL;
 927 }
 928
 929 /* TODO: This is really inefficient.  We need something like get_user()
 930  * (instruction directly accesses the data, with an exception table entry
 931  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 932  */
 933 static int set_bit_to_user(int nr, void __user *addr)
 934 {
 935         unsigned long log = (unsigned long)addr;
 936         struct page *page;
 937         void *base;
 938         int bit = nr + (log % PAGE_SIZE) * 8;
 939         int r;
 940
 941         r = get_user_pages_fast(log, 1, 1, &page);
 942         if (r < 0)
 943                 return r;
 944         BUG_ON(r != 1);
 945         base = kmap_atomic(page);
 946         set_bit(bit, base);
 947         kunmap_atomic(base);
 948         set_page_dirty_lock(page);
 949         put_page(page);
 950         return 0;
 951 }
 952
 953 static int log_write(void __user *log_base,
 954                      u64 write_address, u64 write_length)
 955 {
 956         u64 write_page = write_address / VHOST_PAGE_SIZE;
 957         int r;
 958
 959         if (!write_length)
 960                 return 0;
 961         write_length += write_address % VHOST_PAGE_SIZE;
 962         for (;;) {
 963                 u64 base = (u64)(unsigned long)log_base;
 964                 u64 log = base + write_page / 8;
 965                 int bit = write_page % 8;
 966                 if ((u64)(unsigned long)log != log)
 967                         return -EFAULT;
 968                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 969                 if (r < 0)
 970                         return r;
 971                 if (write_length <= VHOST_PAGE_SIZE)
 972                         break;
 973                 write_length -= VHOST_PAGE_SIZE;
 974                 write_page += 1;
 975         }
 976         return r;
 977 }
 978
 979 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 980                     unsigned int log_num, u64 len)
 981 {
 982         int i, r;
 983
 984         /* Make sure data written is seen before log. */
 985         smp_wmb();
 986         for (i = 0; i < log_num; ++i) {
 987                 u64 l = min(log[i].len, len);
 988                 r = log_write(vq->log_base, log[i].addr, l);
 989                 if (r < 0)
 990                         return r;
 991                 len -= l;
 992                 if (!len) {
 993                         if (vq->log_ctx)
 994                                 eventfd_signal(vq->log_ctx, 1);
 995                         return 0;
 996                 }
 997         }
 998         /* Length written exceeds what we have stored. This is a bug. */
 999         BUG();
1000         return 0;
1001 }
1002
1003 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1004 {
1005         void __user *used;
1006         if (__put_user(vq->used_flags, &vq->used->flags) < 0)
1007                 return -EFAULT;
1008         if (unlikely(vq->log_used)) {
1009                 /* Make sure the flag is seen before log. */
1010                 smp_wmb();
1011                 /* Log used flag write. */
1012                 used = &vq->used->flags;
1013                 log_write(vq->log_base, vq->log_addr +
1014                           (used - (void __user *)vq->used),
1015                           sizeof vq->used->flags);
1016                 if (vq->log_ctx)
1017                         eventfd_signal(vq->log_ctx, 1);
1018         }
1019         return 0;
1020 }
1021
1022 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1023 {
1024         if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
1025                 return -EFAULT;
1026         if (unlikely(vq->log_used)) {
1027                 void __user *used;
1028                 /* Make sure the event is seen before log. */
1029                 smp_wmb();
1030                 /* Log avail event write */
1031                 used = vhost_avail_event(vq);
1032                 log_write(vq->log_base, vq->log_addr +
1033                           (used - (void __user *)vq->used),
1034                           sizeof *vhost_avail_event(vq));
1035                 if (vq->log_ctx)
1036                         eventfd_signal(vq->log_ctx, 1);
1037         }
1038         return 0;
1039 }
1040
1041 int vhost_init_used(struct vhost_virtqueue *vq)
1042 {
1043         int r;
1044         if (!vq->private_data)
1045                 return 0;
1046
1047         r = vhost_update_used_flags(vq);
1048         if (r)
1049                 return r;
1050         vq->signalled_used_valid = false;
1051         return get_user(vq->last_used_idx, &vq->used->idx);
1052 }
1053
1054 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1055                           struct iovec iov[], int iov_size)
1056 {
1057         const struct vhost_memory_region *reg;
1058         struct vhost_memory *mem;
1059         struct iovec *_iov;
1060         u64 s = 0;
1061         int ret = 0;
1062
1063         rcu_read_lock();
1064
1065         mem = rcu_dereference(dev->memory);
1066         while ((u64)len > s) {
1067                 u64 size;
1068                 if (unlikely(ret >= iov_size)) {
1069                         ret = -ENOBUFS;
1070                         break;
1071                 }
1072                 reg = find_region(mem, addr, len);
1073                 if (unlikely(!reg)) {
1074                         ret = -EFAULT;
1075                         break;
1076                 }
1077                 _iov = iov + ret;
1078                 size = reg->memory_size - addr + reg->guest_phys_addr;
1079                 _iov->iov_len = min((u64)len, size);
1080                 _iov->iov_base = (void __user *)(unsigned long)
1081                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1082                 s += size;
1083                 addr += size;
1084                 ++ret;
1085         }
1086
1087         rcu_read_unlock();
1088         return ret;
1089 }
1090
1091 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1092  * function returns the next descriptor in the chain,
1093  * or -1U if we're at the end. */
1094 static unsigned next_desc(struct vring_desc *desc)
1095 {
1096         unsigned int next;
1097
1098         /* If this descriptor says it doesn't chain, we're done. */
1099         if (!(desc->flags & VRING_DESC_F_NEXT))
1100                 return -1U;
1101
1102         /* Check they're not leading us off end of descriptors. */
1103         next = desc->next;
1104         /* Make sure compiler knows to grab that: we don't want it changing! */
1105         /* We will use the result as an index in an array, so most
1106          * architectures only need a compiler barrier here. */
1107         read_barrier_depends();
1108
1109         return next;
1110 }
1111
1112 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1113                         struct iovec iov[], unsigned int iov_size,
1114                         unsigned int *out_num, unsigned int *in_num,
1115                         struct vhost_log *log, unsigned int *log_num,
1116                         struct vring_desc *indirect)
1117 {
1118         struct vring_desc desc;
1119         unsigned int i = 0, count, found = 0;
1120         int ret;
1121
1122         /* Sanity check */
1123         if (unlikely(indirect->len % sizeof desc)) {
1124                 vq_err(vq, "Invalid length in indirect descriptor: "
1125                        "len 0x%llx not multiple of 0x%zx\n",
1126                        (unsigned long long)indirect->len,
1127                        sizeof desc);
1128                 return -EINVAL;
1129         }
1130
1131         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1132                              UIO_MAXIOV);
1133         if (unlikely(ret < 0)) {
1134                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1135                 return ret;
1136         }
1137
1138         /* We will use the result as an address to read from, so most
1139          * architectures only need a compiler barrier here. */
1140         read_barrier_depends();
1141
1142         count = indirect->len / sizeof desc;
1143         /* Buffers are chained via a 16 bit next field, so
1144          * we can have at most 2^16 of these. */
1145         if (unlikely(count > USHRT_MAX + 1)) {
1146                 vq_err(vq, "Indirect buffer length too big: %d\n",
1147                        indirect->len);
1148                 return -E2BIG;
1149         }
1150
1151         do {
1152                 unsigned iov_count = *in_num + *out_num;
1153                 if (unlikely(++found > count)) {
1154                         vq_err(vq, "Loop detected: last one at %u "
1155                                "indirect size %u\n",
1156                                i, count);
1157                         return -EINVAL;
1158                 }
1159                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1160                                               vq->indirect, sizeof desc))) {
1161                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1162                                i, (size_t)indirect->addr + i * sizeof desc);
1163                         return -EINVAL;
1164                 }
1165                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1166                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1167                                i, (size_t)indirect->addr + i * sizeof desc);
1168                         return -EINVAL;
1169                 }
1170
1171                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1172                                      iov_size - iov_count);
1173                 if (unlikely(ret < 0)) {
1174                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1175                                ret, i);
1176                         return ret;
1177                 }
1178                 /* If this is an input descriptor, increment that count. */
1179                 if (desc.flags & VRING_DESC_F_WRITE) {
1180                         *in_num += ret;
1181                         if (unlikely(log)) {
1182                                 log[*log_num].addr = desc.addr;
1183                                 log[*log_num].len = desc.len;
1184                                 ++*log_num;
1185                         }
1186                 } else {
1187                         /* If it's an output descriptor, they're all supposed
1188                          * to come before any input descriptors. */
1189                         if (unlikely(*in_num)) {
1190                                 vq_err(vq, "Indirect descriptor "
1191                                        "has out after in: idx %d\n", i);
1192                                 return -EINVAL;
1193                         }
1194                         *out_num += ret;
1195                 }
1196         } while ((i = next_desc(&desc)) != -1);
1197         return 0;
1198 }
1199
1200 /* This looks in the virtqueue and for the first available buffer, and converts
1201  * it to an iovec for convenient access.  Since descriptors consist of some
1202  * number of output then some number of input descriptors, it's actually two
1203  * iovecs, but we pack them into one and note how many of each there were.
1204  *
1205  * This function returns the descriptor number found, or vq->num (which is
1206  * never a valid descriptor number) if none was found.  A negative code is
1207  * returned on error. */
1208 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1209                       struct iovec iov[], unsigned int iov_size,
1210                       unsigned int *out_num, unsigned int *in_num,
1211                       struct vhost_log *log, unsigned int *log_num)
1212 {
1213         struct vring_desc desc;
1214         unsigned int i, head, found = 0;
1215         u16 last_avail_idx;
1216         int ret;
1217
1218         /* Check it isn't doing very strange things with descriptor numbers. */
1219         last_avail_idx = vq->last_avail_idx;
1220         if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1221                 vq_err(vq, "Failed to access avail idx at %p\n",
1222                        &vq->avail->idx);
1223                 return -EFAULT;
1224         }
1225
1226         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1227                 vq_err(vq, "Guest moved used index from %u to %u",
1228                        last_avail_idx, vq->avail_idx);
1229                 return -EFAULT;
1230         }
1231
1232         /* If there's nothing new since last we looked, return invalid. */
1233         if (vq->avail_idx == last_avail_idx)
1234                 return vq->num;
1235
1236         /* Only get avail ring entries after they have been exposed by guest. */
1237         smp_rmb();
1238
1239         /* Grab the next descriptor number they're advertising, and increment
1240          * the index we've seen. */
1241         if (unlikely(__get_user(head,
1242                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1243                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1244                        last_avail_idx,
1245                        &vq->avail->ring[last_avail_idx % vq->num]);
1246                 return -EFAULT;
1247         }
1248
1249         /* If their number is silly, that's an error. */
1250         if (unlikely(head >= vq->num)) {
1251                 vq_err(vq, "Guest says index %u > %u is available",
1252                        head, vq->num);
1253                 return -EINVAL;
1254         }
1255
1256         /* When we start there are none of either input nor output. */
1257         *out_num = *in_num = 0;
1258         if (unlikely(log))
1259                 *log_num = 0;
1260
1261         i = head;
1262         do {
1263                 unsigned iov_count = *in_num + *out_num;
1264                 if (unlikely(i >= vq->num)) {
1265                         vq_err(vq, "Desc index is %u > %u, head = %u",
1266                                i, vq->num, head);
1267                         return -EINVAL;
1268                 }
1269                 if (unlikely(++found > vq->num)) {
1270                         vq_err(vq, "Loop detected: last one at %u "
1271                                "vq size %u head %u\n",
1272                                i, vq->num, head);
1273                         return -EINVAL;
1274                 }
1275                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1276                 if (unlikely(ret)) {
1277                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1278                                i, vq->desc + i);
1279                         return -EFAULT;
1280                 }
1281                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1282                         ret = get_indirect(dev, vq, iov, iov_size,
1283                                            out_num, in_num,
1284                                            log, log_num, &desc);
1285                         if (unlikely(ret < 0)) {
1286                                 vq_err(vq, "Failure detected "
1287                                        "in indirect descriptor at idx %d\n", i);
1288                                 return ret;
1289                         }
1290                         continue;
1291                 }
1292
1293                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1294                                      iov_size - iov_count);
1295                 if (unlikely(ret < 0)) {
1296                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1297                                ret, i);
1298                         return ret;
1299                 }
1300                 if (desc.flags & VRING_DESC_F_WRITE) {
1301                         /* If this is an input descriptor,
1302                          * increment that count. */
1303                         *in_num += ret;
1304                         if (unlikely(log)) {
1305                                 log[*log_num].addr = desc.addr;
1306                                 log[*log_num].len = desc.len;
1307                                 ++*log_num;
1308                         }
1309                 } else {
1310                         /* If it's an output descriptor, they're all supposed
1311                          * to come before any input descriptors. */
1312                         if (unlikely(*in_num)) {
1313                                 vq_err(vq, "Descriptor has out after in: "
1314                                        "idx %d\n", i);
1315                                 return -EINVAL;
1316                         }
1317                         *out_num += ret;
1318                 }
1319         } while ((i = next_desc(&desc)) != -1);
1320
1321         /* On success, increment avail index. */
1322         vq->last_avail_idx++;
1323
1324         /* Assume notifications from guest are disabled at this point,
1325          * if they aren't we would need to update avail_event index. */
1326         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1327         return head;
1328 }
1329
1330 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1331 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1332 {
1333         vq->last_avail_idx -= n;
1334 }
1335
1336 /* After we've used one of their buffers, we tell them about it.  We'll then
1337  * want to notify the guest, using eventfd. */
1338 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1339 {
1340         struct vring_used_elem __user *used;
1341
1342         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1343          * next entry in that used ring. */
1344         used = &vq->used->ring[vq->last_used_idx % vq->num];
1345         if (__put_user(head, &used->id)) {
1346                 vq_err(vq, "Failed to write used id");
1347                 return -EFAULT;
1348         }
1349         if (__put_user(len, &used->len)) {
1350                 vq_err(vq, "Failed to write used len");
1351                 return -EFAULT;
1352         }
1353         /* Make sure buffer is written before we update index. */
1354         smp_wmb();
1355         if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1356                 vq_err(vq, "Failed to increment used idx");
1357                 return -EFAULT;
1358         }
1359         if (unlikely(vq->log_used)) {
1360                 /* Make sure data is seen before log. */
1361                 smp_wmb();
1362                 /* Log used ring entry write. */
1363                 log_write(vq->log_base,
1364                           vq->log_addr +
1365                            ((void __user *)used - (void __user *)vq->used),
1366                           sizeof *used);
1367                 /* Log used index update. */
1368                 log_write(vq->log_base,
1369                           vq->log_addr + offsetof(struct vring_used, idx),
1370                           sizeof vq->used->idx);
1371                 if (vq->log_ctx)
1372                         eventfd_signal(vq->log_ctx, 1);
1373         }
1374         vq->last_used_idx++;
1375         /* If the driver never bothers to signal in a very long while,
1376          * used index might wrap around. If that happens, invalidate
1377          * signalled_used index we stored. TODO: make sure driver
1378          * signals at least once in 2^16 and remove this. */
1379         if (unlikely(vq->last_used_idx == vq->signalled_used))
1380                 vq->signalled_used_valid = false;
1381         return 0;
1382 }
1383
1384 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1385                             struct vring_used_elem *heads,
1386                             unsigned count)
1387 {
1388         struct vring_used_elem __user *used;
1389         u16 old, new;
1390         int start;
1391
1392         start = vq->last_used_idx % vq->num;
1393         used = vq->used->ring + start;
1394         if (__copy_to_user(used, heads, count * sizeof *used)) {
1395                 vq_err(vq, "Failed to write used");
1396                 return -EFAULT;
1397         }
1398         if (unlikely(vq->log_used)) {
1399                 /* Make sure data is seen before log. */
1400                 smp_wmb();
1401                 /* Log used ring entry write. */
1402                 log_write(vq->log_base,
1403                           vq->log_addr +
1404                            ((void __user *)used - (void __user *)vq->used),
1405                           count * sizeof *used);
1406         }
1407         old = vq->last_used_idx;
1408         new = (vq->last_used_idx += count);
1409         /* If the driver never bothers to signal in a very long while,
1410          * used index might wrap around. If that happens, invalidate
1411          * signalled_used index we stored. TODO: make sure driver
1412          * signals at least once in 2^16 and remove this. */
1413         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1414                 vq->signalled_used_valid = false;
1415         return 0;
1416 }
1417
1418 /* After we've used one of their buffers, we tell them about it.  We'll then
1419  * want to notify the guest, using eventfd. */
1420 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1421                      unsigned count)
1422 {
1423         int start, n, r;
1424
1425         start = vq->last_used_idx % vq->num;
1426         n = vq->num - start;
1427         if (n < count) {
1428                 r = __vhost_add_used_n(vq, heads, n);
1429                 if (r < 0)
1430                         return r;
1431                 heads += n;
1432                 count -= n;
1433         }
1434         r = __vhost_add_used_n(vq, heads, count);
1435
1436         /* Make sure buffer is written before we update index. */
1437         smp_wmb();
1438         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1439                 vq_err(vq, "Failed to increment used idx");
1440                 return -EFAULT;
1441         }
1442         if (unlikely(vq->log_used)) {
1443                 /* Log used index update. */
1444                 log_write(vq->log_base,
1445                           vq->log_addr + offsetof(struct vring_used, idx),
1446                           sizeof vq->used->idx);
1447                 if (vq->log_ctx)
1448                         eventfd_signal(vq->log_ctx, 1);
1449         }
1450         return r;
1451 }
1452
1453 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1454 {
1455         __u16 old, new, event;
1456         bool v;
1457         /* Flush out used index updates. This is paired
1458          * with the barrier that the Guest executes when enabling
1459          * interrupts. */
1460         smp_mb();
1461
1462         if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1463             unlikely(vq->avail_idx == vq->last_avail_idx))
1464                 return true;
1465
1466         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1467                 __u16 flags;
1468                 if (__get_user(flags, &vq->avail->flags)) {
1469                         vq_err(vq, "Failed to get flags");
1470                         return true;
1471                 }
1472                 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1473         }
1474         old = vq->signalled_used;
1475         v = vq->signalled_used_valid;
1476         new = vq->signalled_used = vq->last_used_idx;
1477         vq->signalled_used_valid = true;
1478
1479         if (unlikely(!v))
1480                 return true;
1481
1482         if (get_user(event, vhost_used_event(vq))) {
1483                 vq_err(vq, "Failed to get used event idx");
1484                 return true;
1485         }
1486         return vring_need_event(event, new, old);
1487 }
1488
1489 /* This actually signals the guest, using eventfd. */
1490 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1491 {
1492         /* Signal the Guest tell them we used something up. */
1493         if (vq->call_ctx && vhost_notify(dev, vq))
1494                 eventfd_signal(vq->call_ctx, 1);
1495 }
1496
1497 /* And here's the combo meal deal.  Supersize me! */
1498 void vhost_add_used_and_signal(struct vhost_dev *dev,
1499                                struct vhost_virtqueue *vq,
1500                                unsigned int head, int len)
1501 {
1502         vhost_add_used(vq, head, len);
1503         vhost_signal(dev, vq);
1504 }
1505
1506 /* multi-buffer version of vhost_add_used_and_signal */
1507 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1508                                  struct vhost_virtqueue *vq,
1509                                  struct vring_used_elem *heads, unsigned count)
1510 {
1511         vhost_add_used_n(vq, heads, count);
1512         vhost_signal(dev, vq);
1513 }
1514
1515 /* OK, now we need to know about added descriptors. */
1516 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1517 {
1518         u16 avail_idx;
1519         int r;
1520
1521         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1522                 return false;
1523         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1524         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1525                 r = vhost_update_used_flags(vq);
1526                 if (r) {
1527                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1528                                &vq->used->flags, r);
1529                         return false;
1530                 }
1531         } else {
1532                 r = vhost_update_avail_event(vq, vq->avail_idx);
1533                 if (r) {
1534                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1535                                vhost_avail_event(vq), r);
1536                         return false;
1537                 }
1538         }
1539         /* They could have slipped one in as we were doing that: make
1540          * sure it's written, then check again. */
1541         smp_mb();
1542         r = __get_user(avail_idx, &vq->avail->idx);
1543         if (r) {
1544                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1545                        &vq->avail->idx, r);
1546                 return false;
1547         }
1548
1549         return avail_idx != vq->avail_idx;
1550 }
1551
1552 /* We don't need to be notified again. */
1553 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1554 {
1555         int r;
1556
1557         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1558                 return;
1559         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1560         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1561                 r = vhost_update_used_flags(vq);
1562                 if (r)
1563                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1564                                &vq->used->flags, r);
1565         }
1566 }
1567
1568 static void vhost_zerocopy_done_signal(struct kref *kref)
1569 {
1570         struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
1571                                                     kref);
1572         wake_up(&ubufs->wait);
1573 }
1574
1575 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
1576                                         bool zcopy)
1577 {
1578         struct vhost_ubuf_ref *ubufs;
1579         /* No zero copy backend? Nothing to count. */
1580         if (!zcopy)
1581                 return NULL;
1582         ubufs = kmalloc(sizeof *ubufs, GFP_KERNEL);
1583         if (!ubufs)
1584                 return ERR_PTR(-ENOMEM);
1585         kref_init(&ubufs->kref);
1586         init_waitqueue_head(&ubufs->wait);
1587         ubufs->vq = vq;
1588         return ubufs;
1589 }
1590
1591 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
1592 {
1593         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1594 }
1595
1596 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
1597 {
1598         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1599         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
1600         kfree(ubufs);
1601 }
1602
1603 void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool status)
1604 {
1605         struct vhost_ubuf_ref *ubufs = ubuf->ctx;
1606         struct vhost_virtqueue *vq = ubufs->vq;
1607
1608         vhost_poll_queue(&vq->poll);
1609         /* set len to mark this desc buffers done DMA */
1610         vq->heads[ubuf->desc].len = VHOST_DMA_DONE_LEN;
1611         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
1612 }