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