Pileus Git - ~andy/linux/blob - drivers/vhost/vhost.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Copyright (C) 2006 Rusty Russell IBM Corporation
   3  *
   4  * Author: Michael S. Tsirkin <mst@redhat.com>
   5  *
   6  * Inspiration, some code, and most witty comments come from
   7  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.
  10  *
  11  * Generic code for virtio server in host kernel.
  12  */
  13
  14 #include <linux/eventfd.h>
  15 #include <linux/vhost.h>
  16 #include <linux/virtio_net.h>
  17 #include <linux/mm.h>
  18 #include <linux/mmu_context.h>
  19 #include <linux/miscdevice.h>
  20 #include <linux/mutex.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/poll.h>
  23 #include <linux/file.h>
  24 #include <linux/highmem.h>
  25 #include <linux/slab.h>
  26 #include <linux/kthread.h>
  27 #include <linux/cgroup.h>
  28
  29 #include "vhost.h"
  30
  31 enum {
  32         VHOST_MEMORY_MAX_NREGIONS = 64,
  33         VHOST_MEMORY_F_LOG = 0x1,
  34 };
  35
  36 #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 static long vhost_dev_set_owner(struct vhost_dev *dev)
 348 {
 349         struct task_struct *worker;
 350         int err;
 351
 352         /* Is there an owner already? */
 353         if (dev->mm) {
 354                 err = -EBUSY;
 355                 goto err_mm;
 356         }
 357
 358         /* No owner, become one */
 359         dev->mm = get_task_mm(current);
 360         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 361         if (IS_ERR(worker)) {
 362                 err = PTR_ERR(worker);
 363                 goto err_worker;
 364         }
 365
 366         dev->worker = worker;
 367         wake_up_process(worker);        /* avoid contributing to loadavg */
 368
 369         err = vhost_attach_cgroups(dev);
 370         if (err)
 371                 goto err_cgroup;
 372
 373         err = vhost_dev_alloc_iovecs(dev);
 374         if (err)
 375                 goto err_cgroup;
 376
 377         return 0;
 378 err_cgroup:
 379         kthread_stop(worker);
 380         dev->worker = NULL;
 381 err_worker:
 382         if (dev->mm)
 383                 mmput(dev->mm);
 384         dev->mm = NULL;
 385 err_mm:
 386         return err;
 387 }
 388
 389 /* Caller should have device mutex */
 390 long vhost_dev_reset_owner(struct vhost_dev *dev)
 391 {
 392         struct vhost_memory *memory;
 393
 394         /* Restore memory to default empty mapping. */
 395         memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 396         if (!memory)
 397                 return -ENOMEM;
 398
 399         vhost_dev_cleanup(dev, true);
 400
 401         memory->nregions = 0;
 402         RCU_INIT_POINTER(dev->memory, memory);
 403         return 0;
 404 }
 405
 406 void vhost_dev_stop(struct vhost_dev *dev)
 407 {
 408         int i;
 409
 410         for (i = 0; i < dev->nvqs; ++i) {
 411                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 412                         vhost_poll_stop(&dev->vqs[i]->poll);
 413                         vhost_poll_flush(&dev->vqs[i]->poll);
 414                 }
 415         }
 416 }
 417
 418 /* Caller should have device mutex if and only if locked is set */
 419 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 420 {
 421         int i;
 422
 423         for (i = 0; i < dev->nvqs; ++i) {
 424                 if (dev->vqs[i]->error_ctx)
 425                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 426                 if (dev->vqs[i]->error)
 427                         fput(dev->vqs[i]->error);
 428                 if (dev->vqs[i]->kick)
 429                         fput(dev->vqs[i]->kick);
 430                 if (dev->vqs[i]->call_ctx)
 431                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 432                 if (dev->vqs[i]->call)
 433                         fput(dev->vqs[i]->call);
 434                 vhost_vq_reset(dev, dev->vqs[i]);
 435         }
 436         vhost_dev_free_iovecs(dev);
 437         if (dev->log_ctx)
 438                 eventfd_ctx_put(dev->log_ctx);
 439         dev->log_ctx = NULL;
 440         if (dev->log_file)
 441                 fput(dev->log_file);
 442         dev->log_file = NULL;
 443         /* No one will access memory at this point */
 444         kfree(rcu_dereference_protected(dev->memory,
 445                                         locked ==
 446                                                 lockdep_is_held(&dev->mutex)));
 447         RCU_INIT_POINTER(dev->memory, NULL);
 448         WARN_ON(!list_empty(&dev->work_list));
 449         if (dev->worker) {
 450                 kthread_stop(dev->worker);
 451                 dev->worker = NULL;
 452         }
 453         if (dev->mm)
 454                 mmput(dev->mm);
 455         dev->mm = NULL;
 456 }
 457
 458 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 459 {
 460         u64 a = addr / VHOST_PAGE_SIZE / 8;
 461
 462         /* Make sure 64 bit math will not overflow. */
 463         if (a > ULONG_MAX - (unsigned long)log_base ||
 464             a + (unsigned long)log_base > ULONG_MAX)
 465                 return 0;
 466
 467         return access_ok(VERIFY_WRITE, log_base + a,
 468                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 469 }
 470
 471 /* Caller should have vq mutex and device mutex. */
 472 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 473                                int log_all)
 474 {
 475         int i;
 476
 477         if (!mem)
 478                 return 0;
 479
 480         for (i = 0; i < mem->nregions; ++i) {
 481                 struct vhost_memory_region *m = mem->regions + i;
 482                 unsigned long a = m->userspace_addr;
 483                 if (m->memory_size > ULONG_MAX)
 484                         return 0;
 485                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 486                                     m->memory_size))
 487                         return 0;
 488                 else if (log_all && !log_access_ok(log_base,
 489                                                    m->guest_phys_addr,
 490                                                    m->memory_size))
 491                         return 0;
 492         }
 493         return 1;
 494 }
 495
 496 /* Can we switch to this memory table? */
 497 /* Caller should have device mutex but not vq mutex */
 498 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 499                             int log_all)
 500 {
 501         int i;
 502
 503         for (i = 0; i < d->nvqs; ++i) {
 504                 int ok;
 505                 mutex_lock(&d->vqs[i]->mutex);
 506                 /* If ring is inactive, will check when it's enabled. */
 507                 if (d->vqs[i]->private_data)
 508                         ok = vq_memory_access_ok(d->vqs[i]->log_base, mem,
 509                                                  log_all);
 510                 else
 511                         ok = 1;
 512                 mutex_unlock(&d->vqs[i]->mutex);
 513                 if (!ok)
 514                         return 0;
 515         }
 516         return 1;
 517 }
 518
 519 static int vq_access_ok(struct vhost_dev *d, unsigned int num,
 520                         struct vring_desc __user *desc,
 521                         struct vring_avail __user *avail,
 522                         struct vring_used __user *used)
 523 {
 524         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 525         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 526                access_ok(VERIFY_READ, avail,
 527                          sizeof *avail + num * sizeof *avail->ring + s) &&
 528                access_ok(VERIFY_WRITE, used,
 529                         sizeof *used + num * sizeof *used->ring + s);
 530 }
 531
 532 /* Can we log writes? */
 533 /* Caller should have device mutex but not vq mutex */
 534 int vhost_log_access_ok(struct vhost_dev *dev)
 535 {
 536         struct vhost_memory *mp;
 537
 538         mp = rcu_dereference_protected(dev->memory,
 539                                        lockdep_is_held(&dev->mutex));
 540         return memory_access_ok(dev, mp, 1);
 541 }
 542
 543 /* Verify access for write logging. */
 544 /* Caller should have vq mutex and device mutex */
 545 static int vq_log_access_ok(struct vhost_dev *d, struct vhost_virtqueue *vq,
 546                             void __user *log_base)
 547 {
 548         struct vhost_memory *mp;
 549         size_t s = vhost_has_feature(d, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 550
 551         mp = rcu_dereference_protected(vq->dev->memory,
 552                                        lockdep_is_held(&vq->mutex));
 553         return vq_memory_access_ok(log_base, mp,
 554                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
 555                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 556                                         sizeof *vq->used +
 557                                         vq->num * sizeof *vq->used->ring + s));
 558 }
 559
 560 /* Can we start vq? */
 561 /* Caller should have vq mutex and device mutex */
 562 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 563 {
 564         return vq_access_ok(vq->dev, vq->num, vq->desc, vq->avail, vq->used) &&
 565                 vq_log_access_ok(vq->dev, vq, vq->log_base);
 566 }
 567
 568 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 569 {
 570         struct vhost_memory mem, *newmem, *oldmem;
 571         unsigned long size = offsetof(struct vhost_memory, regions);
 572
 573         if (copy_from_user(&mem, m, size))
 574                 return -EFAULT;
 575         if (mem.padding)
 576                 return -EOPNOTSUPP;
 577         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 578                 return -E2BIG;
 579         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 580         if (!newmem)
 581                 return -ENOMEM;
 582
 583         memcpy(newmem, &mem, size);
 584         if (copy_from_user(newmem->regions, m->regions,
 585                            mem.nregions * sizeof *m->regions)) {
 586                 kfree(newmem);
 587                 return -EFAULT;
 588         }
 589
 590         if (!memory_access_ok(d, newmem,
 591                               vhost_has_feature(d, VHOST_F_LOG_ALL))) {
 592                 kfree(newmem);
 593                 return -EFAULT;
 594         }
 595         oldmem = rcu_dereference_protected(d->memory,
 596                                            lockdep_is_held(&d->mutex));
 597         rcu_assign_pointer(d->memory, newmem);
 598         synchronize_rcu();
 599         kfree(oldmem);
 600         return 0;
 601 }
 602
 603 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 604 {
 605         struct file *eventfp, *filep = NULL;
 606         bool pollstart = false, pollstop = false;
 607         struct eventfd_ctx *ctx = NULL;
 608         u32 __user *idxp = argp;
 609         struct vhost_virtqueue *vq;
 610         struct vhost_vring_state s;
 611         struct vhost_vring_file f;
 612         struct vhost_vring_addr a;
 613         u32 idx;
 614         long r;
 615
 616         r = get_user(idx, idxp);
 617         if (r < 0)
 618                 return r;
 619         if (idx >= d->nvqs)
 620                 return -ENOBUFS;
 621
 622         vq = d->vqs[idx];
 623
 624         mutex_lock(&vq->mutex);
 625
 626         switch (ioctl) {
 627         case VHOST_SET_VRING_NUM:
 628                 /* Resizing ring with an active backend?
 629                  * You don't want to do that. */
 630                 if (vq->private_data) {
 631                         r = -EBUSY;
 632                         break;
 633                 }
 634                 if (copy_from_user(&s, argp, sizeof s)) {
 635                         r = -EFAULT;
 636                         break;
 637                 }
 638                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 639                         r = -EINVAL;
 640                         break;
 641                 }
 642                 vq->num = s.num;
 643                 break;
 644         case VHOST_SET_VRING_BASE:
 645                 /* Moving base with an active backend?
 646                  * You don't want to do that. */
 647                 if (vq->private_data) {
 648                         r = -EBUSY;
 649                         break;
 650                 }
 651                 if (copy_from_user(&s, argp, sizeof s)) {
 652                         r = -EFAULT;
 653                         break;
 654                 }
 655                 if (s.num > 0xffff) {
 656                         r = -EINVAL;
 657                         break;
 658                 }
 659                 vq->last_avail_idx = s.num;
 660                 /* Forget the cached index value. */
 661                 vq->avail_idx = vq->last_avail_idx;
 662                 break;
 663         case VHOST_GET_VRING_BASE:
 664                 s.index = idx;
 665                 s.num = vq->last_avail_idx;
 666                 if (copy_to_user(argp, &s, sizeof s))
 667                         r = -EFAULT;
 668                 break;
 669         case VHOST_SET_VRING_ADDR:
 670                 if (copy_from_user(&a, argp, sizeof a)) {
 671                         r = -EFAULT;
 672                         break;
 673                 }
 674                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 675                         r = -EOPNOTSUPP;
 676                         break;
 677                 }
 678                 /* For 32bit, verify that the top 32bits of the user
 679                    data are set to zero. */
 680                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 681                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 682                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 683                         r = -EFAULT;
 684                         break;
 685                 }
 686                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 687                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 688                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 689                         r = -EINVAL;
 690                         break;
 691                 }
 692
 693                 /* We only verify access here if backend is configured.
 694                  * If it is not, we don't as size might not have been setup.
 695                  * We will verify when backend is configured. */
 696                 if (vq->private_data) {
 697                         if (!vq_access_ok(d, vq->num,
 698                                 (void __user *)(unsigned long)a.desc_user_addr,
 699                                 (void __user *)(unsigned long)a.avail_user_addr,
 700                                 (void __user *)(unsigned long)a.used_user_addr)) {
 701                                 r = -EINVAL;
 702                                 break;
 703                         }
 704
 705                         /* Also validate log access for used ring if enabled. */
 706                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 707                             !log_access_ok(vq->log_base, a.log_guest_addr,
 708                                            sizeof *vq->used +
 709                                            vq->num * sizeof *vq->used->ring)) {
 710                                 r = -EINVAL;
 711                                 break;
 712                         }
 713                 }
 714
 715                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 716                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 717                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 718                 vq->log_addr = a.log_guest_addr;
 719                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 720                 break;
 721         case VHOST_SET_VRING_KICK:
 722                 if (copy_from_user(&f, argp, sizeof f)) {
 723                         r = -EFAULT;
 724                         break;
 725                 }
 726                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 727                 if (IS_ERR(eventfp)) {
 728                         r = PTR_ERR(eventfp);
 729                         break;
 730                 }
 731                 if (eventfp != vq->kick) {
 732                         pollstop = (filep = vq->kick) != NULL;
 733                         pollstart = (vq->kick = eventfp) != NULL;
 734                 } else
 735                         filep = eventfp;
 736                 break;
 737         case VHOST_SET_VRING_CALL:
 738                 if (copy_from_user(&f, argp, sizeof f)) {
 739                         r = -EFAULT;
 740                         break;
 741                 }
 742                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 743                 if (IS_ERR(eventfp)) {
 744                         r = PTR_ERR(eventfp);
 745                         break;
 746                 }
 747                 if (eventfp != vq->call) {
 748                         filep = vq->call;
 749                         ctx = vq->call_ctx;
 750                         vq->call = eventfp;
 751                         vq->call_ctx = eventfp ?
 752                                 eventfd_ctx_fileget(eventfp) : NULL;
 753                 } else
 754                         filep = eventfp;
 755                 break;
 756         case VHOST_SET_VRING_ERR:
 757                 if (copy_from_user(&f, argp, sizeof f)) {
 758                         r = -EFAULT;
 759                         break;
 760                 }
 761                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 762                 if (IS_ERR(eventfp)) {
 763                         r = PTR_ERR(eventfp);
 764                         break;
 765                 }
 766                 if (eventfp != vq->error) {
 767                         filep = vq->error;
 768                         vq->error = eventfp;
 769                         ctx = vq->error_ctx;
 770                         vq->error_ctx = eventfp ?
 771                                 eventfd_ctx_fileget(eventfp) : NULL;
 772                 } else
 773                         filep = eventfp;
 774                 break;
 775         default:
 776                 r = -ENOIOCTLCMD;
 777         }
 778
 779         if (pollstop && vq->handle_kick)
 780                 vhost_poll_stop(&vq->poll);
 781
 782         if (ctx)
 783                 eventfd_ctx_put(ctx);
 784         if (filep)
 785                 fput(filep);
 786
 787         if (pollstart && vq->handle_kick)
 788                 r = vhost_poll_start(&vq->poll, vq->kick);
 789
 790         mutex_unlock(&vq->mutex);
 791
 792         if (pollstop && vq->handle_kick)
 793                 vhost_poll_flush(&vq->poll);
 794         return r;
 795 }
 796
 797 /* Caller must have device mutex */
 798 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
 799 {
 800         struct file *eventfp, *filep = NULL;
 801         struct eventfd_ctx *ctx = NULL;
 802         u64 p;
 803         long r;
 804         int i, fd;
 805
 806         /* If you are not the owner, you can become one */
 807         if (ioctl == VHOST_SET_OWNER) {
 808                 r = vhost_dev_set_owner(d);
 809                 goto done;
 810         }
 811
 812         /* You must be the owner to do anything else */
 813         r = vhost_dev_check_owner(d);
 814         if (r)
 815                 goto done;
 816
 817         switch (ioctl) {
 818         case VHOST_SET_MEM_TABLE:
 819                 r = vhost_set_memory(d, argp);
 820                 break;
 821         case VHOST_SET_LOG_BASE:
 822                 if (copy_from_user(&p, argp, sizeof p)) {
 823                         r = -EFAULT;
 824                         break;
 825                 }
 826                 if ((u64)(unsigned long)p != p) {
 827                         r = -EFAULT;
 828                         break;
 829                 }
 830                 for (i = 0; i < d->nvqs; ++i) {
 831                         struct vhost_virtqueue *vq;
 832                         void __user *base = (void __user *)(unsigned long)p;
 833                         vq = d->vqs[i];
 834                         mutex_lock(&vq->mutex);
 835                         /* If ring is inactive, will check when it's enabled. */
 836                         if (vq->private_data && !vq_log_access_ok(d, vq, base))
 837                                 r = -EFAULT;
 838                         else
 839                                 vq->log_base = base;
 840                         mutex_unlock(&vq->mutex);
 841                 }
 842                 break;
 843         case VHOST_SET_LOG_FD:
 844                 r = get_user(fd, (int __user *)argp);
 845                 if (r < 0)
 846                         break;
 847                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 848                 if (IS_ERR(eventfp)) {
 849                         r = PTR_ERR(eventfp);
 850                         break;
 851                 }
 852                 if (eventfp != d->log_file) {
 853                         filep = d->log_file;
 854                         ctx = d->log_ctx;
 855                         d->log_ctx = eventfp ?
 856                                 eventfd_ctx_fileget(eventfp) : NULL;
 857                 } else
 858                         filep = eventfp;
 859                 for (i = 0; i < d->nvqs; ++i) {
 860                         mutex_lock(&d->vqs[i]->mutex);
 861                         d->vqs[i]->log_ctx = d->log_ctx;
 862                         mutex_unlock(&d->vqs[i]->mutex);
 863                 }
 864                 if (ctx)
 865                         eventfd_ctx_put(ctx);
 866                 if (filep)
 867                         fput(filep);
 868                 break;
 869         default:
 870                 r = -ENOIOCTLCMD;
 871                 break;
 872         }
 873 done:
 874         return r;
 875 }
 876
 877 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 878                                                      __u64 addr, __u32 len)
 879 {
 880         struct vhost_memory_region *reg;
 881         int i;
 882
 883         /* linear search is not brilliant, but we really have on the order of 6
 884          * regions in practice */
 885         for (i = 0; i < mem->nregions; ++i) {
 886                 reg = mem->regions + i;
 887                 if (reg->guest_phys_addr <= addr &&
 888                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 889                         return reg;
 890         }
 891         return NULL;
 892 }
 893
 894 /* TODO: This is really inefficient.  We need something like get_user()
 895  * (instruction directly accesses the data, with an exception table entry
 896  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 897  */
 898 static int set_bit_to_user(int nr, void __user *addr)
 899 {
 900         unsigned long log = (unsigned long)addr;
 901         struct page *page;
 902         void *base;
 903         int bit = nr + (log % PAGE_SIZE) * 8;
 904         int r;
 905
 906         r = get_user_pages_fast(log, 1, 1, &page);
 907         if (r < 0)
 908                 return r;
 909         BUG_ON(r != 1);
 910         base = kmap_atomic(page);
 911         set_bit(bit, base);
 912         kunmap_atomic(base);
 913         set_page_dirty_lock(page);
 914         put_page(page);
 915         return 0;
 916 }
 917
 918 static int log_write(void __user *log_base,
 919                      u64 write_address, u64 write_length)
 920 {
 921         u64 write_page = write_address / VHOST_PAGE_SIZE;
 922         int r;
 923
 924         if (!write_length)
 925                 return 0;
 926         write_length += write_address % VHOST_PAGE_SIZE;
 927         for (;;) {
 928                 u64 base = (u64)(unsigned long)log_base;
 929                 u64 log = base + write_page / 8;
 930                 int bit = write_page % 8;
 931                 if ((u64)(unsigned long)log != log)
 932                         return -EFAULT;
 933                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 934                 if (r < 0)
 935                         return r;
 936                 if (write_length <= VHOST_PAGE_SIZE)
 937                         break;
 938                 write_length -= VHOST_PAGE_SIZE;
 939                 write_page += 1;
 940         }
 941         return r;
 942 }
 943
 944 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 945                     unsigned int log_num, u64 len)
 946 {
 947         int i, r;
 948
 949         /* Make sure data written is seen before log. */
 950         smp_wmb();
 951         for (i = 0; i < log_num; ++i) {
 952                 u64 l = min(log[i].len, len);
 953                 r = log_write(vq->log_base, log[i].addr, l);
 954                 if (r < 0)
 955                         return r;
 956                 len -= l;
 957                 if (!len) {
 958                         if (vq->log_ctx)
 959                                 eventfd_signal(vq->log_ctx, 1);
 960                         return 0;
 961                 }
 962         }
 963         /* Length written exceeds what we have stored. This is a bug. */
 964         BUG();
 965         return 0;
 966 }
 967
 968 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
 969 {
 970         void __user *used;
 971         if (__put_user(vq->used_flags, &vq->used->flags) < 0)
 972                 return -EFAULT;
 973         if (unlikely(vq->log_used)) {
 974                 /* Make sure the flag is seen before log. */
 975                 smp_wmb();
 976                 /* Log used flag write. */
 977                 used = &vq->used->flags;
 978                 log_write(vq->log_base, vq->log_addr +
 979                           (used - (void __user *)vq->used),
 980                           sizeof vq->used->flags);
 981                 if (vq->log_ctx)
 982                         eventfd_signal(vq->log_ctx, 1);
 983         }
 984         return 0;
 985 }
 986
 987 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
 988 {
 989         if (__put_user(vq->avail_idx, vhost_avail_event(vq)))
 990                 return -EFAULT;
 991         if (unlikely(vq->log_used)) {
 992                 void __user *used;
 993                 /* Make sure the event is seen before log. */
 994                 smp_wmb();
 995                 /* Log avail event write */
 996                 used = vhost_avail_event(vq);
 997                 log_write(vq->log_base, vq->log_addr +
 998                           (used - (void __user *)vq->used),
 999                           sizeof *vhost_avail_event(vq));
1000                 if (vq->log_ctx)
1001                         eventfd_signal(vq->log_ctx, 1);
1002         }
1003         return 0;
1004 }
1005
1006 int vhost_init_used(struct vhost_virtqueue *vq)
1007 {
1008         int r;
1009         if (!vq->private_data)
1010                 return 0;
1011
1012         r = vhost_update_used_flags(vq);
1013         if (r)
1014                 return r;
1015         vq->signalled_used_valid = false;
1016         return get_user(vq->last_used_idx, &vq->used->idx);
1017 }
1018
1019 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
1020                           struct iovec iov[], int iov_size)
1021 {
1022         const struct vhost_memory_region *reg;
1023         struct vhost_memory *mem;
1024         struct iovec *_iov;
1025         u64 s = 0;
1026         int ret = 0;
1027
1028         rcu_read_lock();
1029
1030         mem = rcu_dereference(dev->memory);
1031         while ((u64)len > s) {
1032                 u64 size;
1033                 if (unlikely(ret >= iov_size)) {
1034                         ret = -ENOBUFS;
1035                         break;
1036                 }
1037                 reg = find_region(mem, addr, len);
1038                 if (unlikely(!reg)) {
1039                         ret = -EFAULT;
1040                         break;
1041                 }
1042                 _iov = iov + ret;
1043                 size = reg->memory_size - addr + reg->guest_phys_addr;
1044                 _iov->iov_len = min((u64)len - s, size);
1045                 _iov->iov_base = (void __user *)(unsigned long)
1046                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1047                 s += size;
1048                 addr += size;
1049                 ++ret;
1050         }
1051
1052         rcu_read_unlock();
1053         return ret;
1054 }
1055
1056 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1057  * function returns the next descriptor in the chain,
1058  * or -1U if we're at the end. */
1059 static unsigned next_desc(struct vring_desc *desc)
1060 {
1061         unsigned int next;
1062
1063         /* If this descriptor says it doesn't chain, we're done. */
1064         if (!(desc->flags & VRING_DESC_F_NEXT))
1065                 return -1U;
1066
1067         /* Check they're not leading us off end of descriptors. */
1068         next = desc->next;
1069         /* Make sure compiler knows to grab that: we don't want it changing! */
1070         /* We will use the result as an index in an array, so most
1071          * architectures only need a compiler barrier here. */
1072         read_barrier_depends();
1073
1074         return next;
1075 }
1076
1077 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1078                         struct iovec iov[], unsigned int iov_size,
1079                         unsigned int *out_num, unsigned int *in_num,
1080                         struct vhost_log *log, unsigned int *log_num,
1081                         struct vring_desc *indirect)
1082 {
1083         struct vring_desc desc;
1084         unsigned int i = 0, count, found = 0;
1085         int ret;
1086
1087         /* Sanity check */
1088         if (unlikely(indirect->len % sizeof desc)) {
1089                 vq_err(vq, "Invalid length in indirect descriptor: "
1090                        "len 0x%llx not multiple of 0x%zx\n",
1091                        (unsigned long long)indirect->len,
1092                        sizeof desc);
1093                 return -EINVAL;
1094         }
1095
1096         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
1097                              UIO_MAXIOV);
1098         if (unlikely(ret < 0)) {
1099                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1100                 return ret;
1101         }
1102
1103         /* We will use the result as an address to read from, so most
1104          * architectures only need a compiler barrier here. */
1105         read_barrier_depends();
1106
1107         count = indirect->len / sizeof desc;
1108         /* Buffers are chained via a 16 bit next field, so
1109          * we can have at most 2^16 of these. */
1110         if (unlikely(count > USHRT_MAX + 1)) {
1111                 vq_err(vq, "Indirect buffer length too big: %d\n",
1112                        indirect->len);
1113                 return -E2BIG;
1114         }
1115
1116         do {
1117                 unsigned iov_count = *in_num + *out_num;
1118                 if (unlikely(++found > count)) {
1119                         vq_err(vq, "Loop detected: last one at %u "
1120                                "indirect size %u\n",
1121                                i, count);
1122                         return -EINVAL;
1123                 }
1124                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc,
1125                                               vq->indirect, sizeof desc))) {
1126                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1127                                i, (size_t)indirect->addr + i * sizeof desc);
1128                         return -EINVAL;
1129                 }
1130                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
1131                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1132                                i, (size_t)indirect->addr + i * sizeof desc);
1133                         return -EINVAL;
1134                 }
1135
1136                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1137                                      iov_size - iov_count);
1138                 if (unlikely(ret < 0)) {
1139                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1140                                ret, i);
1141                         return ret;
1142                 }
1143                 /* If this is an input descriptor, increment that count. */
1144                 if (desc.flags & VRING_DESC_F_WRITE) {
1145                         *in_num += ret;
1146                         if (unlikely(log)) {
1147                                 log[*log_num].addr = desc.addr;
1148                                 log[*log_num].len = desc.len;
1149                                 ++*log_num;
1150                         }
1151                 } else {
1152                         /* If it's an output descriptor, they're all supposed
1153                          * to come before any input descriptors. */
1154                         if (unlikely(*in_num)) {
1155                                 vq_err(vq, "Indirect descriptor "
1156                                        "has out after in: idx %d\n", i);
1157                                 return -EINVAL;
1158                         }
1159                         *out_num += ret;
1160                 }
1161         } while ((i = next_desc(&desc)) != -1);
1162         return 0;
1163 }
1164
1165 /* This looks in the virtqueue and for the first available buffer, and converts
1166  * it to an iovec for convenient access.  Since descriptors consist of some
1167  * number of output then some number of input descriptors, it's actually two
1168  * iovecs, but we pack them into one and note how many of each there were.
1169  *
1170  * This function returns the descriptor number found, or vq->num (which is
1171  * never a valid descriptor number) if none was found.  A negative code is
1172  * returned on error. */
1173 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
1174                       struct iovec iov[], unsigned int iov_size,
1175                       unsigned int *out_num, unsigned int *in_num,
1176                       struct vhost_log *log, unsigned int *log_num)
1177 {
1178         struct vring_desc desc;
1179         unsigned int i, head, found = 0;
1180         u16 last_avail_idx;
1181         int ret;
1182
1183         /* Check it isn't doing very strange things with descriptor numbers. */
1184         last_avail_idx = vq->last_avail_idx;
1185         if (unlikely(__get_user(vq->avail_idx, &vq->avail->idx))) {
1186                 vq_err(vq, "Failed to access avail idx at %p\n",
1187                        &vq->avail->idx);
1188                 return -EFAULT;
1189         }
1190
1191         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1192                 vq_err(vq, "Guest moved used index from %u to %u",
1193                        last_avail_idx, vq->avail_idx);
1194                 return -EFAULT;
1195         }
1196
1197         /* If there's nothing new since last we looked, return invalid. */
1198         if (vq->avail_idx == last_avail_idx)
1199                 return vq->num;
1200
1201         /* Only get avail ring entries after they have been exposed by guest. */
1202         smp_rmb();
1203
1204         /* Grab the next descriptor number they're advertising, and increment
1205          * the index we've seen. */
1206         if (unlikely(__get_user(head,
1207                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1208                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1209                        last_avail_idx,
1210                        &vq->avail->ring[last_avail_idx % vq->num]);
1211                 return -EFAULT;
1212         }
1213
1214         /* If their number is silly, that's an error. */
1215         if (unlikely(head >= vq->num)) {
1216                 vq_err(vq, "Guest says index %u > %u is available",
1217                        head, vq->num);
1218                 return -EINVAL;
1219         }
1220
1221         /* When we start there are none of either input nor output. */
1222         *out_num = *in_num = 0;
1223         if (unlikely(log))
1224                 *log_num = 0;
1225
1226         i = head;
1227         do {
1228                 unsigned iov_count = *in_num + *out_num;
1229                 if (unlikely(i >= vq->num)) {
1230                         vq_err(vq, "Desc index is %u > %u, head = %u",
1231                                i, vq->num, head);
1232                         return -EINVAL;
1233                 }
1234                 if (unlikely(++found > vq->num)) {
1235                         vq_err(vq, "Loop detected: last one at %u "
1236                                "vq size %u head %u\n",
1237                                i, vq->num, head);
1238                         return -EINVAL;
1239                 }
1240                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1241                 if (unlikely(ret)) {
1242                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1243                                i, vq->desc + i);
1244                         return -EFAULT;
1245                 }
1246                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1247                         ret = get_indirect(dev, vq, iov, iov_size,
1248                                            out_num, in_num,
1249                                            log, log_num, &desc);
1250                         if (unlikely(ret < 0)) {
1251                                 vq_err(vq, "Failure detected "
1252                                        "in indirect descriptor at idx %d\n", i);
1253                                 return ret;
1254                         }
1255                         continue;
1256                 }
1257
1258                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1259                                      iov_size - iov_count);
1260                 if (unlikely(ret < 0)) {
1261                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1262                                ret, i);
1263                         return ret;
1264                 }
1265                 if (desc.flags & VRING_DESC_F_WRITE) {
1266                         /* If this is an input descriptor,
1267                          * increment that count. */
1268                         *in_num += ret;
1269                         if (unlikely(log)) {
1270                                 log[*log_num].addr = desc.addr;
1271                                 log[*log_num].len = desc.len;
1272                                 ++*log_num;
1273                         }
1274                 } else {
1275                         /* If it's an output descriptor, they're all supposed
1276                          * to come before any input descriptors. */
1277                         if (unlikely(*in_num)) {
1278                                 vq_err(vq, "Descriptor has out after in: "
1279                                        "idx %d\n", i);
1280                                 return -EINVAL;
1281                         }
1282                         *out_num += ret;
1283                 }
1284         } while ((i = next_desc(&desc)) != -1);
1285
1286         /* On success, increment avail index. */
1287         vq->last_avail_idx++;
1288
1289         /* Assume notifications from guest are disabled at this point,
1290          * if they aren't we would need to update avail_event index. */
1291         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1292         return head;
1293 }
1294
1295 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1296 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1297 {
1298         vq->last_avail_idx -= n;
1299 }
1300
1301 /* After we've used one of their buffers, we tell them about it.  We'll then
1302  * want to notify the guest, using eventfd. */
1303 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1304 {
1305         struct vring_used_elem __user *used;
1306
1307         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1308          * next entry in that used ring. */
1309         used = &vq->used->ring[vq->last_used_idx % vq->num];
1310         if (__put_user(head, &used->id)) {
1311                 vq_err(vq, "Failed to write used id");
1312                 return -EFAULT;
1313         }
1314         if (__put_user(len, &used->len)) {
1315                 vq_err(vq, "Failed to write used len");
1316                 return -EFAULT;
1317         }
1318         /* Make sure buffer is written before we update index. */
1319         smp_wmb();
1320         if (__put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1321                 vq_err(vq, "Failed to increment used idx");
1322                 return -EFAULT;
1323         }
1324         if (unlikely(vq->log_used)) {
1325                 /* Make sure data is seen before log. */
1326                 smp_wmb();
1327                 /* Log used ring entry write. */
1328                 log_write(vq->log_base,
1329                           vq->log_addr +
1330                            ((void __user *)used - (void __user *)vq->used),
1331                           sizeof *used);
1332                 /* Log used index update. */
1333                 log_write(vq->log_base,
1334                           vq->log_addr + offsetof(struct vring_used, idx),
1335                           sizeof vq->used->idx);
1336                 if (vq->log_ctx)
1337                         eventfd_signal(vq->log_ctx, 1);
1338         }
1339         vq->last_used_idx++;
1340         /* If the driver never bothers to signal in a very long while,
1341          * used index might wrap around. If that happens, invalidate
1342          * signalled_used index we stored. TODO: make sure driver
1343          * signals at least once in 2^16 and remove this. */
1344         if (unlikely(vq->last_used_idx == vq->signalled_used))
1345                 vq->signalled_used_valid = false;
1346         return 0;
1347 }
1348
1349 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1350                             struct vring_used_elem *heads,
1351                             unsigned count)
1352 {
1353         struct vring_used_elem __user *used;
1354         u16 old, new;
1355         int start;
1356
1357         start = vq->last_used_idx % vq->num;
1358         used = vq->used->ring + start;
1359         if (__copy_to_user(used, heads, count * sizeof *used)) {
1360                 vq_err(vq, "Failed to write used");
1361                 return -EFAULT;
1362         }
1363         if (unlikely(vq->log_used)) {
1364                 /* Make sure data is seen before log. */
1365                 smp_wmb();
1366                 /* Log used ring entry write. */
1367                 log_write(vq->log_base,
1368                           vq->log_addr +
1369                            ((void __user *)used - (void __user *)vq->used),
1370                           count * sizeof *used);
1371         }
1372         old = vq->last_used_idx;
1373         new = (vq->last_used_idx += count);
1374         /* If the driver never bothers to signal in a very long while,
1375          * used index might wrap around. If that happens, invalidate
1376          * signalled_used index we stored. TODO: make sure driver
1377          * signals at least once in 2^16 and remove this. */
1378         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1379                 vq->signalled_used_valid = false;
1380         return 0;
1381 }
1382
1383 /* After we've used one of their buffers, we tell them about it.  We'll then
1384  * want to notify the guest, using eventfd. */
1385 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1386                      unsigned count)
1387 {
1388         int start, n, r;
1389
1390         start = vq->last_used_idx % vq->num;
1391         n = vq->num - start;
1392         if (n < count) {
1393                 r = __vhost_add_used_n(vq, heads, n);
1394                 if (r < 0)
1395                         return r;
1396                 heads += n;
1397                 count -= n;
1398         }
1399         r = __vhost_add_used_n(vq, heads, count);
1400
1401         /* Make sure buffer is written before we update index. */
1402         smp_wmb();
1403         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1404                 vq_err(vq, "Failed to increment used idx");
1405                 return -EFAULT;
1406         }
1407         if (unlikely(vq->log_used)) {
1408                 /* Log used index update. */
1409                 log_write(vq->log_base,
1410                           vq->log_addr + offsetof(struct vring_used, idx),
1411                           sizeof vq->used->idx);
1412                 if (vq->log_ctx)
1413                         eventfd_signal(vq->log_ctx, 1);
1414         }
1415         return r;
1416 }
1417
1418 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1419 {
1420         __u16 old, new, event;
1421         bool v;
1422         /* Flush out used index updates. This is paired
1423          * with the barrier that the Guest executes when enabling
1424          * interrupts. */
1425         smp_mb();
1426
1427         if (vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1428             unlikely(vq->avail_idx == vq->last_avail_idx))
1429                 return true;
1430
1431         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1432                 __u16 flags;
1433                 if (__get_user(flags, &vq->avail->flags)) {
1434                         vq_err(vq, "Failed to get flags");
1435                         return true;
1436                 }
1437                 return !(flags & VRING_AVAIL_F_NO_INTERRUPT);
1438         }
1439         old = vq->signalled_used;
1440         v = vq->signalled_used_valid;
1441         new = vq->signalled_used = vq->last_used_idx;
1442         vq->signalled_used_valid = true;
1443
1444         if (unlikely(!v))
1445                 return true;
1446
1447         if (get_user(event, vhost_used_event(vq))) {
1448                 vq_err(vq, "Failed to get used event idx");
1449                 return true;
1450         }
1451         return vring_need_event(event, new, old);
1452 }
1453
1454 /* This actually signals the guest, using eventfd. */
1455 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1456 {
1457         /* Signal the Guest tell them we used something up. */
1458         if (vq->call_ctx && vhost_notify(dev, vq))
1459                 eventfd_signal(vq->call_ctx, 1);
1460 }
1461
1462 /* And here's the combo meal deal.  Supersize me! */
1463 void vhost_add_used_and_signal(struct vhost_dev *dev,
1464                                struct vhost_virtqueue *vq,
1465                                unsigned int head, int len)
1466 {
1467         vhost_add_used(vq, head, len);
1468         vhost_signal(dev, vq);
1469 }
1470
1471 /* multi-buffer version of vhost_add_used_and_signal */
1472 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1473                                  struct vhost_virtqueue *vq,
1474                                  struct vring_used_elem *heads, unsigned count)
1475 {
1476         vhost_add_used_n(vq, heads, count);
1477         vhost_signal(dev, vq);
1478 }
1479
1480 /* OK, now we need to know about added descriptors. */
1481 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1482 {
1483         u16 avail_idx;
1484         int r;
1485
1486         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1487                 return false;
1488         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1489         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1490                 r = vhost_update_used_flags(vq);
1491                 if (r) {
1492                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1493                                &vq->used->flags, r);
1494                         return false;
1495                 }
1496         } else {
1497                 r = vhost_update_avail_event(vq, vq->avail_idx);
1498                 if (r) {
1499                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1500                                vhost_avail_event(vq), r);
1501                         return false;
1502                 }
1503         }
1504         /* They could have slipped one in as we were doing that: make
1505          * sure it's written, then check again. */
1506         smp_mb();
1507         r = __get_user(avail_idx, &vq->avail->idx);
1508         if (r) {
1509                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1510                        &vq->avail->idx, r);
1511                 return false;
1512         }
1513
1514         return avail_idx != vq->avail_idx;
1515 }
1516
1517 /* We don't need to be notified again. */
1518 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1519 {
1520         int r;
1521
1522         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1523                 return;
1524         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1525         if (!vhost_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) {
1526                 r = vhost_update_used_flags(vq);
1527                 if (r)
1528                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1529                                &vq->used->flags, r);
1530         }
1531 }