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/uio.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/poll.h>
  22 #include <linux/file.h>
  23 #include <linux/highmem.h>
  24 #include <linux/slab.h>
  25 #include <linux/vmalloc.h>
  26 #include <linux/kthread.h>
  27 #include <linux/cgroup.h>
  28 #include <linux/module.h>
  29 #include <linux/sort.h>
  30
  31 #include "vhost.h"
  32
  33 enum {
  34         VHOST_MEMORY_MAX_NREGIONS = 64,
  35         VHOST_MEMORY_F_LOG = 0x1,
  36 };
  37
  38 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  39 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  40
  41 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  42 static void vhost_vq_reset_user_be(struct vhost_virtqueue *vq)
  43 {
  44         vq->user_be = !virtio_legacy_is_little_endian();
  45 }
  46
  47 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  48 {
  49         struct vhost_vring_state s;
  50
  51         if (vq->private_data)
  52                 return -EBUSY;
  53
  54         if (copy_from_user(&s, argp, sizeof(s)))
  55                 return -EFAULT;
  56
  57         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  58             s.num != VHOST_VRING_BIG_ENDIAN)
  59                 return -EINVAL;
  60
  61         vq->user_be = s.num;
  62
  63         return 0;
  64 }
  65
  66 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  67                                    int __user *argp)
  68 {
  69         struct vhost_vring_state s = {
  70                 .index = idx,
  71                 .num = vq->user_be
  72         };
  73
  74         if (copy_to_user(argp, &s, sizeof(s)))
  75                 return -EFAULT;
  76
  77         return 0;
  78 }
  79
  80 static void vhost_init_is_le(struct vhost_virtqueue *vq)
  81 {
  82         /* Note for legacy virtio: user_be is initialized at reset time
  83          * according to the host endianness. If userspace does not set an
  84          * explicit endianness, the default behavior is native endian, as
  85          * expected by legacy virtio.
  86          */
  87         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
  88 }
  89 #else
  90 static void vhost_vq_reset_user_be(struct vhost_virtqueue *vq)
  91 {
  92 }
  93
  94 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  95 {
  96         return -ENOIOCTLCMD;
  97 }
  98
  99 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 100                                    int __user *argp)
 101 {
 102         return -ENOIOCTLCMD;
 103 }
 104
 105 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 106 {
 107         if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
 108                 vq->is_le = true;
 109 }
 110 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 111
 112 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 113                             poll_table *pt)
 114 {
 115         struct vhost_poll *poll;
 116
 117         poll = container_of(pt, struct vhost_poll, table);
 118         poll->wqh = wqh;
 119         add_wait_queue(wqh, &poll->wait);
 120 }
 121
 122 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
 123                              void *key)
 124 {
 125         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 126
 127         if (!((unsigned long)key & poll->mask))
 128                 return 0;
 129
 130         vhost_poll_queue(poll);
 131         return 0;
 132 }
 133
 134 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 135 {
 136         INIT_LIST_HEAD(&work->node);
 137         work->fn = fn;
 138         init_waitqueue_head(&work->done);
 139         work->flushing = 0;
 140         work->queue_seq = work->done_seq = 0;
 141 }
 142 EXPORT_SYMBOL_GPL(vhost_work_init);
 143
 144 /* Init poll structure */
 145 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 146                      unsigned long mask, struct vhost_dev *dev)
 147 {
 148         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 149         init_poll_funcptr(&poll->table, vhost_poll_func);
 150         poll->mask = mask;
 151         poll->dev = dev;
 152         poll->wqh = NULL;
 153
 154         vhost_work_init(&poll->work, fn);
 155 }
 156 EXPORT_SYMBOL_GPL(vhost_poll_init);
 157
 158 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 159  * keep a reference to a file until after vhost_poll_stop is called. */
 160 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 161 {
 162         unsigned long mask;
 163         int ret = 0;
 164
 165         if (poll->wqh)
 166                 return 0;
 167
 168         mask = file->f_op->poll(file, &poll->table);
 169         if (mask)
 170                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 171         if (mask & POLLERR) {
 172                 if (poll->wqh)
 173                         remove_wait_queue(poll->wqh, &poll->wait);
 174                 ret = -EINVAL;
 175         }
 176
 177         return ret;
 178 }
 179 EXPORT_SYMBOL_GPL(vhost_poll_start);
 180
 181 /* Stop polling a file. After this function returns, it becomes safe to drop the
 182  * file reference. You must also flush afterwards. */
 183 void vhost_poll_stop(struct vhost_poll *poll)
 184 {
 185         if (poll->wqh) {
 186                 remove_wait_queue(poll->wqh, &poll->wait);
 187                 poll->wqh = NULL;
 188         }
 189 }
 190 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 191
 192 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
 193                                 unsigned seq)
 194 {
 195         int left;
 196
 197         spin_lock_irq(&dev->work_lock);
 198         left = seq - work->done_seq;
 199         spin_unlock_irq(&dev->work_lock);
 200         return left <= 0;
 201 }
 202
 203 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 204 {
 205         unsigned seq;
 206         int flushing;
 207
 208         spin_lock_irq(&dev->work_lock);
 209         seq = work->queue_seq;
 210         work->flushing++;
 211         spin_unlock_irq(&dev->work_lock);
 212         wait_event(work->done, vhost_work_seq_done(dev, work, seq));
 213         spin_lock_irq(&dev->work_lock);
 214         flushing = --work->flushing;
 215         spin_unlock_irq(&dev->work_lock);
 216         BUG_ON(flushing < 0);
 217 }
 218 EXPORT_SYMBOL_GPL(vhost_work_flush);
 219
 220 /* Flush any work that has been scheduled. When calling this, don't hold any
 221  * locks that are also used by the callback. */
 222 void vhost_poll_flush(struct vhost_poll *poll)
 223 {
 224         vhost_work_flush(poll->dev, &poll->work);
 225 }
 226 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 227
 228 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 229 {
 230         unsigned long flags;
 231
 232         spin_lock_irqsave(&dev->work_lock, flags);
 233         if (list_empty(&work->node)) {
 234                 list_add_tail(&work->node, &dev->work_list);
 235                 work->queue_seq++;
 236                 spin_unlock_irqrestore(&dev->work_lock, flags);
 237                 wake_up_process(dev->worker);
 238         } else {
 239                 spin_unlock_irqrestore(&dev->work_lock, flags);
 240         }
 241 }
 242 EXPORT_SYMBOL_GPL(vhost_work_queue);
 243
 244 void vhost_poll_queue(struct vhost_poll *poll)
 245 {
 246         vhost_work_queue(poll->dev, &poll->work);
 247 }
 248 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 249
 250 static void vhost_vq_reset(struct vhost_dev *dev,
 251                            struct vhost_virtqueue *vq)
 252 {
 253         vq->num = 1;
 254         vq->desc = NULL;
 255         vq->avail = NULL;
 256         vq->used = NULL;
 257         vq->last_avail_idx = 0;
 258         vq->avail_idx = 0;
 259         vq->last_used_idx = 0;
 260         vq->signalled_used = 0;
 261         vq->signalled_used_valid = false;
 262         vq->used_flags = 0;
 263         vq->log_used = false;
 264         vq->log_addr = -1ull;
 265         vq->private_data = NULL;
 266         vq->acked_features = 0;
 267         vq->log_base = NULL;
 268         vq->error_ctx = NULL;
 269         vq->error = NULL;
 270         vq->kick = NULL;
 271         vq->call_ctx = NULL;
 272         vq->call = NULL;
 273         vq->log_ctx = NULL;
 274         vq->memory = NULL;
 275         vq->is_le = virtio_legacy_is_little_endian();
 276         vhost_vq_reset_user_be(vq);
 277 }
 278
 279 static int vhost_worker(void *data)
 280 {
 281         struct vhost_dev *dev = data;
 282         struct vhost_work *work = NULL;
 283         unsigned uninitialized_var(seq);
 284         mm_segment_t oldfs = get_fs();
 285
 286         set_fs(USER_DS);
 287         use_mm(dev->mm);
 288
 289         for (;;) {
 290                 /* mb paired w/ kthread_stop */
 291                 set_current_state(TASK_INTERRUPTIBLE);
 292
 293                 spin_lock_irq(&dev->work_lock);
 294                 if (work) {
 295                         work->done_seq = seq;
 296                         if (work->flushing)
 297                                 wake_up_all(&work->done);
 298                 }
 299
 300                 if (kthread_should_stop()) {
 301                         spin_unlock_irq(&dev->work_lock);
 302                         __set_current_state(TASK_RUNNING);
 303                         break;
 304                 }
 305                 if (!list_empty(&dev->work_list)) {
 306                         work = list_first_entry(&dev->work_list,
 307                                                 struct vhost_work, node);
 308                         list_del_init(&work->node);
 309                         seq = work->queue_seq;
 310                 } else
 311                         work = NULL;
 312                 spin_unlock_irq(&dev->work_lock);
 313
 314                 if (work) {
 315                         __set_current_state(TASK_RUNNING);
 316                         work->fn(work);
 317                         if (need_resched())
 318                                 schedule();
 319                 } else
 320                         schedule();
 321
 322         }
 323         unuse_mm(dev->mm);
 324         set_fs(oldfs);
 325         return 0;
 326 }
 327
 328 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 329 {
 330         kfree(vq->indirect);
 331         vq->indirect = NULL;
 332         kfree(vq->log);
 333         vq->log = NULL;
 334         kfree(vq->heads);
 335         vq->heads = NULL;
 336 }
 337
 338 /* Helper to allocate iovec buffers for all vqs. */
 339 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 340 {
 341         struct vhost_virtqueue *vq;
 342         int i;
 343
 344         for (i = 0; i < dev->nvqs; ++i) {
 345                 vq = dev->vqs[i];
 346                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 347                                        GFP_KERNEL);
 348                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 349                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 350                 if (!vq->indirect || !vq->log || !vq->heads)
 351                         goto err_nomem;
 352         }
 353         return 0;
 354
 355 err_nomem:
 356         for (; i >= 0; --i)
 357                 vhost_vq_free_iovecs(dev->vqs[i]);
 358         return -ENOMEM;
 359 }
 360
 361 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 362 {
 363         int i;
 364
 365         for (i = 0; i < dev->nvqs; ++i)
 366                 vhost_vq_free_iovecs(dev->vqs[i]);
 367 }
 368
 369 void vhost_dev_init(struct vhost_dev *dev,
 370                     struct vhost_virtqueue **vqs, int nvqs)
 371 {
 372         struct vhost_virtqueue *vq;
 373         int i;
 374
 375         dev->vqs = vqs;
 376         dev->nvqs = nvqs;
 377         mutex_init(&dev->mutex);
 378         dev->log_ctx = NULL;
 379         dev->log_file = NULL;
 380         dev->memory = NULL;
 381         dev->mm = NULL;
 382         spin_lock_init(&dev->work_lock);
 383         INIT_LIST_HEAD(&dev->work_list);
 384         dev->worker = NULL;
 385
 386         for (i = 0; i < dev->nvqs; ++i) {
 387                 vq = dev->vqs[i];
 388                 vq->log = NULL;
 389                 vq->indirect = NULL;
 390                 vq->heads = NULL;
 391                 vq->dev = dev;
 392                 mutex_init(&vq->mutex);
 393                 vhost_vq_reset(dev, vq);
 394                 if (vq->handle_kick)
 395                         vhost_poll_init(&vq->poll, vq->handle_kick,
 396                                         POLLIN, dev);
 397         }
 398 }
 399 EXPORT_SYMBOL_GPL(vhost_dev_init);
 400
 401 /* Caller should have device mutex */
 402 long vhost_dev_check_owner(struct vhost_dev *dev)
 403 {
 404         /* Are you the owner? If not, I don't think you mean to do that */
 405         return dev->mm == current->mm ? 0 : -EPERM;
 406 }
 407 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 408
 409 struct vhost_attach_cgroups_struct {
 410         struct vhost_work work;
 411         struct task_struct *owner;
 412         int ret;
 413 };
 414
 415 static void vhost_attach_cgroups_work(struct vhost_work *work)
 416 {
 417         struct vhost_attach_cgroups_struct *s;
 418
 419         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 420         s->ret = cgroup_attach_task_all(s->owner, current);
 421 }
 422
 423 static int vhost_attach_cgroups(struct vhost_dev *dev)
 424 {
 425         struct vhost_attach_cgroups_struct attach;
 426
 427         attach.owner = current;
 428         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 429         vhost_work_queue(dev, &attach.work);
 430         vhost_work_flush(dev, &attach.work);
 431         return attach.ret;
 432 }
 433
 434 /* Caller should have device mutex */
 435 bool vhost_dev_has_owner(struct vhost_dev *dev)
 436 {
 437         return dev->mm;
 438 }
 439 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 440
 441 /* Caller should have device mutex */
 442 long vhost_dev_set_owner(struct vhost_dev *dev)
 443 {
 444         struct task_struct *worker;
 445         int err;
 446
 447         /* Is there an owner already? */
 448         if (vhost_dev_has_owner(dev)) {
 449                 err = -EBUSY;
 450                 goto err_mm;
 451         }
 452
 453         /* No owner, become one */
 454         dev->mm = get_task_mm(current);
 455         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 456         if (IS_ERR(worker)) {
 457                 err = PTR_ERR(worker);
 458                 goto err_worker;
 459         }
 460
 461         dev->worker = worker;
 462         wake_up_process(worker);        /* avoid contributing to loadavg */
 463
 464         err = vhost_attach_cgroups(dev);
 465         if (err)
 466                 goto err_cgroup;
 467
 468         err = vhost_dev_alloc_iovecs(dev);
 469         if (err)
 470                 goto err_cgroup;
 471
 472         return 0;
 473 err_cgroup:
 474         kthread_stop(worker);
 475         dev->worker = NULL;
 476 err_worker:
 477         if (dev->mm)
 478                 mmput(dev->mm);
 479         dev->mm = NULL;
 480 err_mm:
 481         return err;
 482 }
 483 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 484
 485 struct vhost_memory *vhost_dev_reset_owner_prepare(void)
 486 {
 487         return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 488 }
 489 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 490
 491 /* Caller should have device mutex */
 492 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
 493 {
 494         int i;
 495
 496         vhost_dev_cleanup(dev, true);
 497
 498         /* Restore memory to default empty mapping. */
 499         memory->nregions = 0;
 500         dev->memory = memory;
 501         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 502          * VQs aren't running.
 503          */
 504         for (i = 0; i < dev->nvqs; ++i)
 505                 dev->vqs[i]->memory = memory;
 506 }
 507 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 508
 509 void vhost_dev_stop(struct vhost_dev *dev)
 510 {
 511         int i;
 512
 513         for (i = 0; i < dev->nvqs; ++i) {
 514                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 515                         vhost_poll_stop(&dev->vqs[i]->poll);
 516                         vhost_poll_flush(&dev->vqs[i]->poll);
 517                 }
 518         }
 519 }
 520 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 521
 522 /* Caller should have device mutex if and only if locked is set */
 523 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 524 {
 525         int i;
 526
 527         for (i = 0; i < dev->nvqs; ++i) {
 528                 if (dev->vqs[i]->error_ctx)
 529                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 530                 if (dev->vqs[i]->error)
 531                         fput(dev->vqs[i]->error);
 532                 if (dev->vqs[i]->kick)
 533                         fput(dev->vqs[i]->kick);
 534                 if (dev->vqs[i]->call_ctx)
 535                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 536                 if (dev->vqs[i]->call)
 537                         fput(dev->vqs[i]->call);
 538                 vhost_vq_reset(dev, dev->vqs[i]);
 539         }
 540         vhost_dev_free_iovecs(dev);
 541         if (dev->log_ctx)
 542                 eventfd_ctx_put(dev->log_ctx);
 543         dev->log_ctx = NULL;
 544         if (dev->log_file)
 545                 fput(dev->log_file);
 546         dev->log_file = NULL;
 547         /* No one will access memory at this point */
 548         kvfree(dev->memory);
 549         dev->memory = NULL;
 550         WARN_ON(!list_empty(&dev->work_list));
 551         if (dev->worker) {
 552                 kthread_stop(dev->worker);
 553                 dev->worker = NULL;
 554         }
 555         if (dev->mm)
 556                 mmput(dev->mm);
 557         dev->mm = NULL;
 558 }
 559 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 560
 561 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 562 {
 563         u64 a = addr / VHOST_PAGE_SIZE / 8;
 564
 565         /* Make sure 64 bit math will not overflow. */
 566         if (a > ULONG_MAX - (unsigned long)log_base ||
 567             a + (unsigned long)log_base > ULONG_MAX)
 568                 return 0;
 569
 570         return access_ok(VERIFY_WRITE, log_base + a,
 571                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 572 }
 573
 574 /* Caller should have vq mutex and device mutex. */
 575 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 576                                int log_all)
 577 {
 578         int i;
 579
 580         if (!mem)
 581                 return 0;
 582
 583         for (i = 0; i < mem->nregions; ++i) {
 584                 struct vhost_memory_region *m = mem->regions + i;
 585                 unsigned long a = m->userspace_addr;
 586                 if (m->memory_size > ULONG_MAX)
 587                         return 0;
 588                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 589                                     m->memory_size))
 590                         return 0;
 591                 else if (log_all && !log_access_ok(log_base,
 592                                                    m->guest_phys_addr,
 593                                                    m->memory_size))
 594                         return 0;
 595         }
 596         return 1;
 597 }
 598
 599 /* Can we switch to this memory table? */
 600 /* Caller should have device mutex but not vq mutex */
 601 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 602                             int log_all)
 603 {
 604         int i;
 605
 606         for (i = 0; i < d->nvqs; ++i) {
 607                 int ok;
 608                 bool log;
 609
 610                 mutex_lock(&d->vqs[i]->mutex);
 611                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 612                 /* If ring is inactive, will check when it's enabled. */
 613                 if (d->vqs[i]->private_data)
 614                         ok = vq_memory_access_ok(d->vqs[i]->log_base, mem, log);
 615                 else
 616                         ok = 1;
 617                 mutex_unlock(&d->vqs[i]->mutex);
 618                 if (!ok)
 619                         return 0;
 620         }
 621         return 1;
 622 }
 623
 624 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
 625                         struct vring_desc __user *desc,
 626                         struct vring_avail __user *avail,
 627                         struct vring_used __user *used)
 628 {
 629         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 630         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 631                access_ok(VERIFY_READ, avail,
 632                          sizeof *avail + num * sizeof *avail->ring + s) &&
 633                access_ok(VERIFY_WRITE, used,
 634                         sizeof *used + num * sizeof *used->ring + s);
 635 }
 636
 637 /* Can we log writes? */
 638 /* Caller should have device mutex but not vq mutex */
 639 int vhost_log_access_ok(struct vhost_dev *dev)
 640 {
 641         return memory_access_ok(dev, dev->memory, 1);
 642 }
 643 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
 644
 645 /* Verify access for write logging. */
 646 /* Caller should have vq mutex and device mutex */
 647 static int vq_log_access_ok(struct vhost_virtqueue *vq,
 648                             void __user *log_base)
 649 {
 650         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 651
 652         return vq_memory_access_ok(log_base, vq->memory,
 653                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
 654                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 655                                         sizeof *vq->used +
 656                                         vq->num * sizeof *vq->used->ring + s));
 657 }
 658
 659 /* Can we start vq? */
 660 /* Caller should have vq mutex and device mutex */
 661 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 662 {
 663         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
 664                 vq_log_access_ok(vq, vq->log_base);
 665 }
 666 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
 667
 668 static int vhost_memory_reg_sort_cmp(const void *p1, const void *p2)
 669 {
 670         const struct vhost_memory_region *r1 = p1, *r2 = p2;
 671         if (r1->guest_phys_addr < r2->guest_phys_addr)
 672                 return 1;
 673         if (r1->guest_phys_addr > r2->guest_phys_addr)
 674                 return -1;
 675         return 0;
 676 }
 677
 678 static void *vhost_kvzalloc(unsigned long size)
 679 {
 680         void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 681
 682         if (!n) {
 683                 n = vzalloc(size);
 684                 if (!n)
 685                         return ERR_PTR(-ENOMEM);
 686         }
 687         return n;
 688 }
 689
 690 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 691 {
 692         struct vhost_memory mem, *newmem, *oldmem;
 693         unsigned long size = offsetof(struct vhost_memory, regions);
 694         int i;
 695
 696         if (copy_from_user(&mem, m, size))
 697                 return -EFAULT;
 698         if (mem.padding)
 699                 return -EOPNOTSUPP;
 700         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 701                 return -E2BIG;
 702         newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
 703         if (!newmem)
 704                 return -ENOMEM;
 705
 706         memcpy(newmem, &mem, size);
 707         if (copy_from_user(newmem->regions, m->regions,
 708                            mem.nregions * sizeof *m->regions)) {
 709                 kvfree(newmem);
 710                 return -EFAULT;
 711         }
 712         sort(newmem->regions, newmem->nregions, sizeof(*newmem->regions),
 713                 vhost_memory_reg_sort_cmp, NULL);
 714
 715         if (!memory_access_ok(d, newmem, 0)) {
 716                 kvfree(newmem);
 717                 return -EFAULT;
 718         }
 719         oldmem = d->memory;
 720         d->memory = newmem;
 721
 722         /* All memory accesses are done under some VQ mutex. */
 723         for (i = 0; i < d->nvqs; ++i) {
 724                 mutex_lock(&d->vqs[i]->mutex);
 725                 d->vqs[i]->memory = newmem;
 726                 mutex_unlock(&d->vqs[i]->mutex);
 727         }
 728         kvfree(oldmem);
 729         return 0;
 730 }
 731
 732 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 733 {
 734         struct file *eventfp, *filep = NULL;
 735         bool pollstart = false, pollstop = false;
 736         struct eventfd_ctx *ctx = NULL;
 737         u32 __user *idxp = argp;
 738         struct vhost_virtqueue *vq;
 739         struct vhost_vring_state s;
 740         struct vhost_vring_file f;
 741         struct vhost_vring_addr a;
 742         u32 idx;
 743         long r;
 744
 745         r = get_user(idx, idxp);
 746         if (r < 0)
 747                 return r;
 748         if (idx >= d->nvqs)
 749                 return -ENOBUFS;
 750
 751         vq = d->vqs[idx];
 752
 753         mutex_lock(&vq->mutex);
 754
 755         switch (ioctl) {
 756         case VHOST_SET_VRING_NUM:
 757                 /* Resizing ring with an active backend?
 758                  * You don't want to do that. */
 759                 if (vq->private_data) {
 760                         r = -EBUSY;
 761                         break;
 762                 }
 763                 if (copy_from_user(&s, argp, sizeof s)) {
 764                         r = -EFAULT;
 765                         break;
 766                 }
 767                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 768                         r = -EINVAL;
 769                         break;
 770                 }
 771                 vq->num = s.num;
 772                 break;
 773         case VHOST_SET_VRING_BASE:
 774                 /* Moving base with an active backend?
 775                  * You don't want to do that. */
 776                 if (vq->private_data) {
 777                         r = -EBUSY;
 778                         break;
 779                 }
 780                 if (copy_from_user(&s, argp, sizeof s)) {
 781                         r = -EFAULT;
 782                         break;
 783                 }
 784                 if (s.num > 0xffff) {
 785                         r = -EINVAL;
 786                         break;
 787                 }
 788                 vq->last_avail_idx = s.num;
 789                 /* Forget the cached index value. */
 790                 vq->avail_idx = vq->last_avail_idx;
 791                 break;
 792         case VHOST_GET_VRING_BASE:
 793                 s.index = idx;
 794                 s.num = vq->last_avail_idx;
 795                 if (copy_to_user(argp, &s, sizeof s))
 796                         r = -EFAULT;
 797                 break;
 798         case VHOST_SET_VRING_ADDR:
 799                 if (copy_from_user(&a, argp, sizeof a)) {
 800                         r = -EFAULT;
 801                         break;
 802                 }
 803                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 804                         r = -EOPNOTSUPP;
 805                         break;
 806                 }
 807                 /* For 32bit, verify that the top 32bits of the user
 808                    data are set to zero. */
 809                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 810                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 811                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 812                         r = -EFAULT;
 813                         break;
 814                 }
 815
 816                 /* Make sure it's safe to cast pointers to vring types. */
 817                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
 818                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
 819                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
 820                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
 821                     (a.log_guest_addr & (sizeof(u64) - 1))) {
 822                         r = -EINVAL;
 823                         break;
 824                 }
 825
 826                 /* We only verify access here if backend is configured.
 827                  * If it is not, we don't as size might not have been setup.
 828                  * We will verify when backend is configured. */
 829                 if (vq->private_data) {
 830                         if (!vq_access_ok(vq, vq->num,
 831                                 (void __user *)(unsigned long)a.desc_user_addr,
 832                                 (void __user *)(unsigned long)a.avail_user_addr,
 833                                 (void __user *)(unsigned long)a.used_user_addr)) {
 834                                 r = -EINVAL;
 835                                 break;
 836                         }
 837
 838                         /* Also validate log access for used ring if enabled. */
 839                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 840                             !log_access_ok(vq->log_base, a.log_guest_addr,
 841                                            sizeof *vq->used +
 842                                            vq->num * sizeof *vq->used->ring)) {
 843                                 r = -EINVAL;
 844                                 break;
 845                         }
 846                 }
 847
 848                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 849                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 850                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 851                 vq->log_addr = a.log_guest_addr;
 852                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 853                 break;
 854         case VHOST_SET_VRING_KICK:
 855                 if (copy_from_user(&f, argp, sizeof f)) {
 856                         r = -EFAULT;
 857                         break;
 858                 }
 859                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 860                 if (IS_ERR(eventfp)) {
 861                         r = PTR_ERR(eventfp);
 862                         break;
 863                 }
 864                 if (eventfp != vq->kick) {
 865                         pollstop = (filep = vq->kick) != NULL;
 866                         pollstart = (vq->kick = eventfp) != NULL;
 867                 } else
 868                         filep = eventfp;
 869                 break;
 870         case VHOST_SET_VRING_CALL:
 871                 if (copy_from_user(&f, argp, sizeof f)) {
 872                         r = -EFAULT;
 873                         break;
 874                 }
 875                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 876                 if (IS_ERR(eventfp)) {
 877                         r = PTR_ERR(eventfp);
 878                         break;
 879                 }
 880                 if (eventfp != vq->call) {
 881                         filep = vq->call;
 882                         ctx = vq->call_ctx;
 883                         vq->call = eventfp;
 884                         vq->call_ctx = eventfp ?
 885                                 eventfd_ctx_fileget(eventfp) : NULL;
 886                 } else
 887                         filep = eventfp;
 888                 break;
 889         case VHOST_SET_VRING_ERR:
 890                 if (copy_from_user(&f, argp, sizeof f)) {
 891                         r = -EFAULT;
 892                         break;
 893                 }
 894                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 895                 if (IS_ERR(eventfp)) {
 896                         r = PTR_ERR(eventfp);
 897                         break;
 898                 }
 899                 if (eventfp != vq->error) {
 900                         filep = vq->error;
 901                         vq->error = eventfp;
 902                         ctx = vq->error_ctx;
 903                         vq->error_ctx = eventfp ?
 904                                 eventfd_ctx_fileget(eventfp) : NULL;
 905                 } else
 906                         filep = eventfp;
 907                 break;
 908         case VHOST_SET_VRING_ENDIAN:
 909                 r = vhost_set_vring_endian(vq, argp);
 910                 break;
 911         case VHOST_GET_VRING_ENDIAN:
 912                 r = vhost_get_vring_endian(vq, idx, argp);
 913                 break;
 914         default:
 915                 r = -ENOIOCTLCMD;
 916         }
 917
 918         if (pollstop && vq->handle_kick)
 919                 vhost_poll_stop(&vq->poll);
 920
 921         if (ctx)
 922                 eventfd_ctx_put(ctx);
 923         if (filep)
 924                 fput(filep);
 925
 926         if (pollstart && vq->handle_kick)
 927                 r = vhost_poll_start(&vq->poll, vq->kick);
 928
 929         mutex_unlock(&vq->mutex);
 930
 931         if (pollstop && vq->handle_kick)
 932                 vhost_poll_flush(&vq->poll);
 933         return r;
 934 }
 935 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
 936
 937 /* Caller must have device mutex */
 938 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
 939 {
 940         struct file *eventfp, *filep = NULL;
 941         struct eventfd_ctx *ctx = NULL;
 942         u64 p;
 943         long r;
 944         int i, fd;
 945
 946         /* If you are not the owner, you can become one */
 947         if (ioctl == VHOST_SET_OWNER) {
 948                 r = vhost_dev_set_owner(d);
 949                 goto done;
 950         }
 951
 952         /* You must be the owner to do anything else */
 953         r = vhost_dev_check_owner(d);
 954         if (r)
 955                 goto done;
 956
 957         switch (ioctl) {
 958         case VHOST_SET_MEM_TABLE:
 959                 r = vhost_set_memory(d, argp);
 960                 break;
 961         case VHOST_SET_LOG_BASE:
 962                 if (copy_from_user(&p, argp, sizeof p)) {
 963                         r = -EFAULT;
 964                         break;
 965                 }
 966                 if ((u64)(unsigned long)p != p) {
 967                         r = -EFAULT;
 968                         break;
 969                 }
 970                 for (i = 0; i < d->nvqs; ++i) {
 971                         struct vhost_virtqueue *vq;
 972                         void __user *base = (void __user *)(unsigned long)p;
 973                         vq = d->vqs[i];
 974                         mutex_lock(&vq->mutex);
 975                         /* If ring is inactive, will check when it's enabled. */
 976                         if (vq->private_data && !vq_log_access_ok(vq, base))
 977                                 r = -EFAULT;
 978                         else
 979                                 vq->log_base = base;
 980                         mutex_unlock(&vq->mutex);
 981                 }
 982                 break;
 983         case VHOST_SET_LOG_FD:
 984                 r = get_user(fd, (int __user *)argp);
 985                 if (r < 0)
 986                         break;
 987                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 988                 if (IS_ERR(eventfp)) {
 989                         r = PTR_ERR(eventfp);
 990                         break;
 991                 }
 992                 if (eventfp != d->log_file) {
 993                         filep = d->log_file;
 994                         ctx = d->log_ctx;
 995                         d->log_ctx = eventfp ?
 996                                 eventfd_ctx_fileget(eventfp) : NULL;
 997                 } else
 998                         filep = eventfp;
 999                 for (i = 0; i < d->nvqs; ++i) {
1000                         mutex_lock(&d->vqs[i]->mutex);
1001                         d->vqs[i]->log_ctx = d->log_ctx;
1002                         mutex_unlock(&d->vqs[i]->mutex);
1003                 }
1004                 if (ctx)
1005                         eventfd_ctx_put(ctx);
1006                 if (filep)
1007                         fput(filep);
1008                 break;
1009         default:
1010                 r = -ENOIOCTLCMD;
1011                 break;
1012         }
1013 done:
1014         return r;
1015 }
1016 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1017
1018 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
1019                                                      __u64 addr, __u32 len)
1020 {
1021         const struct vhost_memory_region *reg;
1022         int start = 0, end = mem->nregions;
1023
1024         while (start < end) {
1025                 int slot = start + (end - start) / 2;
1026                 reg = mem->regions + slot;
1027                 if (addr >= reg->guest_phys_addr)
1028                         end = slot;
1029                 else
1030                         start = slot + 1;
1031         }
1032
1033         reg = mem->regions + start;
1034         if (addr >= reg->guest_phys_addr &&
1035                 reg->guest_phys_addr + reg->memory_size > addr)
1036                 return reg;
1037         return NULL;
1038 }
1039
1040 /* TODO: This is really inefficient.  We need something like get_user()
1041  * (instruction directly accesses the data, with an exception table entry
1042  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1043  */
1044 static int set_bit_to_user(int nr, void __user *addr)
1045 {
1046         unsigned long log = (unsigned long)addr;
1047         struct page *page;
1048         void *base;
1049         int bit = nr + (log % PAGE_SIZE) * 8;
1050         int r;
1051
1052         r = get_user_pages_fast(log, 1, 1, &page);
1053         if (r < 0)
1054                 return r;
1055         BUG_ON(r != 1);
1056         base = kmap_atomic(page);
1057         set_bit(bit, base);
1058         kunmap_atomic(base);
1059         set_page_dirty_lock(page);
1060         put_page(page);
1061         return 0;
1062 }
1063
1064 static int log_write(void __user *log_base,
1065                      u64 write_address, u64 write_length)
1066 {
1067         u64 write_page = write_address / VHOST_PAGE_SIZE;
1068         int r;
1069
1070         if (!write_length)
1071                 return 0;
1072         write_length += write_address % VHOST_PAGE_SIZE;
1073         for (;;) {
1074                 u64 base = (u64)(unsigned long)log_base;
1075                 u64 log = base + write_page / 8;
1076                 int bit = write_page % 8;
1077                 if ((u64)(unsigned long)log != log)
1078                         return -EFAULT;
1079                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1080                 if (r < 0)
1081                         return r;
1082                 if (write_length <= VHOST_PAGE_SIZE)
1083                         break;
1084                 write_length -= VHOST_PAGE_SIZE;
1085                 write_page += 1;
1086         }
1087         return r;
1088 }
1089
1090 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1091                     unsigned int log_num, u64 len)
1092 {
1093         int i, r;
1094
1095         /* Make sure data written is seen before log. */
1096         smp_wmb();
1097         for (i = 0; i < log_num; ++i) {
1098                 u64 l = min(log[i].len, len);
1099                 r = log_write(vq->log_base, log[i].addr, l);
1100                 if (r < 0)
1101                         return r;
1102                 len -= l;
1103                 if (!len) {
1104                         if (vq->log_ctx)
1105                                 eventfd_signal(vq->log_ctx, 1);
1106                         return 0;
1107                 }
1108         }
1109         /* Length written exceeds what we have stored. This is a bug. */
1110         BUG();
1111         return 0;
1112 }
1113 EXPORT_SYMBOL_GPL(vhost_log_write);
1114
1115 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1116 {
1117         void __user *used;
1118         if (__put_user(cpu_to_vhost16(vq, vq->used_flags), &vq->used->flags) < 0)
1119                 return -EFAULT;
1120         if (unlikely(vq->log_used)) {
1121                 /* Make sure the flag is seen before log. */
1122                 smp_wmb();
1123                 /* Log used flag write. */
1124                 used = &vq->used->flags;
1125                 log_write(vq->log_base, vq->log_addr +
1126                           (used - (void __user *)vq->used),
1127                           sizeof vq->used->flags);
1128                 if (vq->log_ctx)
1129                         eventfd_signal(vq->log_ctx, 1);
1130         }
1131         return 0;
1132 }
1133
1134 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1135 {
1136         if (__put_user(cpu_to_vhost16(vq, vq->avail_idx), vhost_avail_event(vq)))
1137                 return -EFAULT;
1138         if (unlikely(vq->log_used)) {
1139                 void __user *used;
1140                 /* Make sure the event is seen before log. */
1141                 smp_wmb();
1142                 /* Log avail event write */
1143                 used = vhost_avail_event(vq);
1144                 log_write(vq->log_base, vq->log_addr +
1145                           (used - (void __user *)vq->used),
1146                           sizeof *vhost_avail_event(vq));
1147                 if (vq->log_ctx)
1148                         eventfd_signal(vq->log_ctx, 1);
1149         }
1150         return 0;
1151 }
1152
1153 int vhost_init_used(struct vhost_virtqueue *vq)
1154 {
1155         __virtio16 last_used_idx;
1156         int r;
1157         if (!vq->private_data) {
1158                 vq->is_le = virtio_legacy_is_little_endian();
1159                 return 0;
1160         }
1161
1162         vhost_init_is_le(vq);
1163
1164         r = vhost_update_used_flags(vq);
1165         if (r)
1166                 return r;
1167         vq->signalled_used_valid = false;
1168         if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx))
1169                 return -EFAULT;
1170         r = __get_user(last_used_idx, &vq->used->idx);
1171         if (r)
1172                 return r;
1173         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1174         return 0;
1175 }
1176 EXPORT_SYMBOL_GPL(vhost_init_used);
1177
1178 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1179                           struct iovec iov[], int iov_size)
1180 {
1181         const struct vhost_memory_region *reg;
1182         struct vhost_memory *mem;
1183         struct iovec *_iov;
1184         u64 s = 0;
1185         int ret = 0;
1186
1187         mem = vq->memory;
1188         while ((u64)len > s) {
1189                 u64 size;
1190                 if (unlikely(ret >= iov_size)) {
1191                         ret = -ENOBUFS;
1192                         break;
1193                 }
1194                 reg = find_region(mem, addr, len);
1195                 if (unlikely(!reg)) {
1196                         ret = -EFAULT;
1197                         break;
1198                 }
1199                 _iov = iov + ret;
1200                 size = reg->memory_size - addr + reg->guest_phys_addr;
1201                 _iov->iov_len = min((u64)len - s, size);
1202                 _iov->iov_base = (void __user *)(unsigned long)
1203                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1204                 s += size;
1205                 addr += size;
1206                 ++ret;
1207         }
1208
1209         return ret;
1210 }
1211
1212 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1213  * function returns the next descriptor in the chain,
1214  * or -1U if we're at the end. */
1215 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1216 {
1217         unsigned int next;
1218
1219         /* If this descriptor says it doesn't chain, we're done. */
1220         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1221                 return -1U;
1222
1223         /* Check they're not leading us off end of descriptors. */
1224         next = vhost16_to_cpu(vq, desc->next);
1225         /* Make sure compiler knows to grab that: we don't want it changing! */
1226         /* We will use the result as an index in an array, so most
1227          * architectures only need a compiler barrier here. */
1228         read_barrier_depends();
1229
1230         return next;
1231 }
1232
1233 static int get_indirect(struct vhost_virtqueue *vq,
1234                         struct iovec iov[], unsigned int iov_size,
1235                         unsigned int *out_num, unsigned int *in_num,
1236                         struct vhost_log *log, unsigned int *log_num,
1237                         struct vring_desc *indirect)
1238 {
1239         struct vring_desc desc;
1240         unsigned int i = 0, count, found = 0;
1241         u32 len = vhost32_to_cpu(vq, indirect->len);
1242         struct iov_iter from;
1243         int ret;
1244
1245         /* Sanity check */
1246         if (unlikely(len % sizeof desc)) {
1247                 vq_err(vq, "Invalid length in indirect descriptor: "
1248                        "len 0x%llx not multiple of 0x%zx\n",
1249                        (unsigned long long)len,
1250                        sizeof desc);
1251                 return -EINVAL;
1252         }
1253
1254         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1255                              UIO_MAXIOV);
1256         if (unlikely(ret < 0)) {
1257                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1258                 return ret;
1259         }
1260         iov_iter_init(&from, READ, vq->indirect, ret, len);
1261
1262         /* We will use the result as an address to read from, so most
1263          * architectures only need a compiler barrier here. */
1264         read_barrier_depends();
1265
1266         count = len / sizeof desc;
1267         /* Buffers are chained via a 16 bit next field, so
1268          * we can have at most 2^16 of these. */
1269         if (unlikely(count > USHRT_MAX + 1)) {
1270                 vq_err(vq, "Indirect buffer length too big: %d\n",
1271                        indirect->len);
1272                 return -E2BIG;
1273         }
1274
1275         do {
1276                 unsigned iov_count = *in_num + *out_num;
1277                 if (unlikely(++found > count)) {
1278                         vq_err(vq, "Loop detected: last one at %u "
1279                                "indirect size %u\n",
1280                                i, count);
1281                         return -EINVAL;
1282                 }
1283                 if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1284                              sizeof(desc))) {
1285                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1286                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1287                         return -EINVAL;
1288                 }
1289                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1290                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1291                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1292                         return -EINVAL;
1293                 }
1294
1295                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1296                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1297                                      iov_size - iov_count);
1298                 if (unlikely(ret < 0)) {
1299                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1300                                ret, i);
1301                         return ret;
1302                 }
1303                 /* If this is an input descriptor, increment that count. */
1304                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1305                         *in_num += ret;
1306                         if (unlikely(log)) {
1307                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1308                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1309                                 ++*log_num;
1310                         }
1311                 } else {
1312                         /* If it's an output descriptor, they're all supposed
1313                          * to come before any input descriptors. */
1314                         if (unlikely(*in_num)) {
1315                                 vq_err(vq, "Indirect descriptor "
1316                                        "has out after in: idx %d\n", i);
1317                                 return -EINVAL;
1318                         }
1319                         *out_num += ret;
1320                 }
1321         } while ((i = next_desc(vq, &desc)) != -1);
1322         return 0;
1323 }
1324
1325 /* This looks in the virtqueue and for the first available buffer, and converts
1326  * it to an iovec for convenient access.  Since descriptors consist of some
1327  * number of output then some number of input descriptors, it's actually two
1328  * iovecs, but we pack them into one and note how many of each there were.
1329  *
1330  * This function returns the descriptor number found, or vq->num (which is
1331  * never a valid descriptor number) if none was found.  A negative code is
1332  * returned on error. */
1333 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1334                       struct iovec iov[], unsigned int iov_size,
1335                       unsigned int *out_num, unsigned int *in_num,
1336                       struct vhost_log *log, unsigned int *log_num)
1337 {
1338         struct vring_desc desc;
1339         unsigned int i, head, found = 0;
1340         u16 last_avail_idx;
1341         __virtio16 avail_idx;
1342         __virtio16 ring_head;
1343         int ret;
1344
1345         /* Check it isn't doing very strange things with descriptor numbers. */
1346         last_avail_idx = vq->last_avail_idx;
1347         if (unlikely(__get_user(avail_idx, &vq->avail->idx))) {
1348                 vq_err(vq, "Failed to access avail idx at %p\n",
1349                        &vq->avail->idx);
1350                 return -EFAULT;
1351         }
1352         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1353
1354         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1355                 vq_err(vq, "Guest moved used index from %u to %u",
1356                        last_avail_idx, vq->avail_idx);
1357                 return -EFAULT;
1358         }
1359
1360         /* If there's nothing new since last we looked, return invalid. */
1361         if (vq->avail_idx == last_avail_idx)
1362                 return vq->num;
1363
1364         /* Only get avail ring entries after they have been exposed by guest. */
1365         smp_rmb();
1366
1367         /* Grab the next descriptor number they're advertising, and increment
1368          * the index we've seen. */
1369         if (unlikely(__get_user(ring_head,
1370                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1371                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1372                        last_avail_idx,
1373                        &vq->avail->ring[last_avail_idx % vq->num]);
1374                 return -EFAULT;
1375         }
1376
1377         head = vhost16_to_cpu(vq, ring_head);
1378
1379         /* If their number is silly, that's an error. */
1380         if (unlikely(head >= vq->num)) {
1381                 vq_err(vq, "Guest says index %u > %u is available",
1382                        head, vq->num);
1383                 return -EINVAL;
1384         }
1385
1386         /* When we start there are none of either input nor output. */
1387         *out_num = *in_num = 0;
1388         if (unlikely(log))
1389                 *log_num = 0;
1390
1391         i = head;
1392         do {
1393                 unsigned iov_count = *in_num + *out_num;
1394                 if (unlikely(i >= vq->num)) {
1395                         vq_err(vq, "Desc index is %u > %u, head = %u",
1396                                i, vq->num, head);
1397                         return -EINVAL;
1398                 }
1399                 if (unlikely(++found > vq->num)) {
1400                         vq_err(vq, "Loop detected: last one at %u "
1401                                "vq size %u head %u\n",
1402                                i, vq->num, head);
1403                         return -EINVAL;
1404                 }
1405                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1406                 if (unlikely(ret)) {
1407                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1408                                i, vq->desc + i);
1409                         return -EFAULT;
1410                 }
1411                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1412                         ret = get_indirect(vq, iov, iov_size,
1413                                            out_num, in_num,
1414                                            log, log_num, &desc);
1415                         if (unlikely(ret < 0)) {
1416                                 vq_err(vq, "Failure detected "
1417                                        "in indirect descriptor at idx %d\n", i);
1418                                 return ret;
1419                         }
1420                         continue;
1421                 }
1422
1423                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1424                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1425                                      iov_size - iov_count);
1426                 if (unlikely(ret < 0)) {
1427                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1428                                ret, i);
1429                         return ret;
1430                 }
1431                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1432                         /* If this is an input descriptor,
1433                          * increment that count. */
1434                         *in_num += ret;
1435                         if (unlikely(log)) {
1436                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1437                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1438                                 ++*log_num;
1439                         }
1440                 } else {
1441                         /* If it's an output descriptor, they're all supposed
1442                          * to come before any input descriptors. */
1443                         if (unlikely(*in_num)) {
1444                                 vq_err(vq, "Descriptor has out after in: "
1445                                        "idx %d\n", i);
1446                                 return -EINVAL;
1447                         }
1448                         *out_num += ret;
1449                 }
1450         } while ((i = next_desc(vq, &desc)) != -1);
1451
1452         /* On success, increment avail index. */
1453         vq->last_avail_idx++;
1454
1455         /* Assume notifications from guest are disabled at this point,
1456          * if they aren't we would need to update avail_event index. */
1457         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1458         return head;
1459 }
1460 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1461
1462 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1463 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1464 {
1465         vq->last_avail_idx -= n;
1466 }
1467 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1468
1469 /* After we've used one of their buffers, we tell them about it.  We'll then
1470  * want to notify the guest, using eventfd. */
1471 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1472 {
1473         struct vring_used_elem heads = {
1474                 cpu_to_vhost32(vq, head),
1475                 cpu_to_vhost32(vq, len)
1476         };
1477
1478         return vhost_add_used_n(vq, &heads, 1);
1479 }
1480 EXPORT_SYMBOL_GPL(vhost_add_used);
1481
1482 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1483                             struct vring_used_elem *heads,
1484                             unsigned count)
1485 {
1486         struct vring_used_elem __user *used;
1487         u16 old, new;
1488         int start;
1489
1490         start = vq->last_used_idx % vq->num;
1491         used = vq->used->ring + start;
1492         if (count == 1) {
1493                 if (__put_user(heads[0].id, &used->id)) {
1494                         vq_err(vq, "Failed to write used id");
1495                         return -EFAULT;
1496                 }
1497                 if (__put_user(heads[0].len, &used->len)) {
1498                         vq_err(vq, "Failed to write used len");
1499                         return -EFAULT;
1500                 }
1501         } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1502                 vq_err(vq, "Failed to write used");
1503                 return -EFAULT;
1504         }
1505         if (unlikely(vq->log_used)) {
1506                 /* Make sure data is seen before log. */
1507                 smp_wmb();
1508                 /* Log used ring entry write. */
1509                 log_write(vq->log_base,
1510                           vq->log_addr +
1511                            ((void __user *)used - (void __user *)vq->used),
1512                           count * sizeof *used);
1513         }
1514         old = vq->last_used_idx;
1515         new = (vq->last_used_idx += count);
1516         /* If the driver never bothers to signal in a very long while,
1517          * used index might wrap around. If that happens, invalidate
1518          * signalled_used index we stored. TODO: make sure driver
1519          * signals at least once in 2^16 and remove this. */
1520         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1521                 vq->signalled_used_valid = false;
1522         return 0;
1523 }
1524
1525 /* After we've used one of their buffers, we tell them about it.  We'll then
1526  * want to notify the guest, using eventfd. */
1527 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1528                      unsigned count)
1529 {
1530         int start, n, r;
1531
1532         start = vq->last_used_idx % vq->num;
1533         n = vq->num - start;
1534         if (n < count) {
1535                 r = __vhost_add_used_n(vq, heads, n);
1536                 if (r < 0)
1537                         return r;
1538                 heads += n;
1539                 count -= n;
1540         }
1541         r = __vhost_add_used_n(vq, heads, count);
1542
1543         /* Make sure buffer is written before we update index. */
1544         smp_wmb();
1545         if (__put_user(cpu_to_vhost16(vq, vq->last_used_idx), &vq->used->idx)) {
1546                 vq_err(vq, "Failed to increment used idx");
1547                 return -EFAULT;
1548         }
1549         if (unlikely(vq->log_used)) {
1550                 /* Log used index update. */
1551                 log_write(vq->log_base,
1552                           vq->log_addr + offsetof(struct vring_used, idx),
1553                           sizeof vq->used->idx);
1554                 if (vq->log_ctx)
1555                         eventfd_signal(vq->log_ctx, 1);
1556         }
1557         return r;
1558 }
1559 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1560
1561 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1562 {
1563         __u16 old, new;
1564         __virtio16 event;
1565         bool v;
1566         /* Flush out used index updates. This is paired
1567          * with the barrier that the Guest executes when enabling
1568          * interrupts. */
1569         smp_mb();
1570
1571         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1572             unlikely(vq->avail_idx == vq->last_avail_idx))
1573                 return true;
1574
1575         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1576                 __virtio16 flags;
1577                 if (__get_user(flags, &vq->avail->flags)) {
1578                         vq_err(vq, "Failed to get flags");
1579                         return true;
1580                 }
1581                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1582         }
1583         old = vq->signalled_used;
1584         v = vq->signalled_used_valid;
1585         new = vq->signalled_used = vq->last_used_idx;
1586         vq->signalled_used_valid = true;
1587
1588         if (unlikely(!v))
1589                 return true;
1590
1591         if (__get_user(event, vhost_used_event(vq))) {
1592                 vq_err(vq, "Failed to get used event idx");
1593                 return true;
1594         }
1595         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1596 }
1597
1598 /* This actually signals the guest, using eventfd. */
1599 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1600 {
1601         /* Signal the Guest tell them we used something up. */
1602         if (vq->call_ctx && vhost_notify(dev, vq))
1603                 eventfd_signal(vq->call_ctx, 1);
1604 }
1605 EXPORT_SYMBOL_GPL(vhost_signal);
1606
1607 /* And here's the combo meal deal.  Supersize me! */
1608 void vhost_add_used_and_signal(struct vhost_dev *dev,
1609                                struct vhost_virtqueue *vq,
1610                                unsigned int head, int len)
1611 {
1612         vhost_add_used(vq, head, len);
1613         vhost_signal(dev, vq);
1614 }
1615 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1616
1617 /* multi-buffer version of vhost_add_used_and_signal */
1618 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1619                                  struct vhost_virtqueue *vq,
1620                                  struct vring_used_elem *heads, unsigned count)
1621 {
1622         vhost_add_used_n(vq, heads, count);
1623         vhost_signal(dev, vq);
1624 }
1625 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1626
1627 /* OK, now we need to know about added descriptors. */
1628 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1629 {
1630         __virtio16 avail_idx;
1631         int r;
1632
1633         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1634                 return false;
1635         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1636         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1637                 r = vhost_update_used_flags(vq);
1638                 if (r) {
1639                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1640                                &vq->used->flags, r);
1641                         return false;
1642                 }
1643         } else {
1644                 r = vhost_update_avail_event(vq, vq->avail_idx);
1645                 if (r) {
1646                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1647                                vhost_avail_event(vq), r);
1648                         return false;
1649                 }
1650         }
1651         /* They could have slipped one in as we were doing that: make
1652          * sure it's written, then check again. */
1653         smp_mb();
1654         r = __get_user(avail_idx, &vq->avail->idx);
1655         if (r) {
1656                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1657                        &vq->avail->idx, r);
1658                 return false;
1659         }
1660
1661         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1662 }
1663 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1664
1665 /* We don't need to be notified again. */
1666 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1667 {
1668         int r;
1669
1670         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1671                 return;
1672         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1673         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1674                 r = vhost_update_used_flags(vq);
1675                 if (r)
1676                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1677                                &vq->used->flags, r);
1678         }
1679 }
1680 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1681
1682 static int __init vhost_init(void)
1683 {
1684         return 0;
1685 }
1686
1687 static void __exit vhost_exit(void)
1688 {
1689 }
1690
1691 module_init(vhost_init);
1692 module_exit(vhost_exit);
1693
1694 MODULE_VERSION("0.0.1");
1695 MODULE_LICENSE("GPL v2");
1696 MODULE_AUTHOR("Michael S. Tsirkin");
1697 MODULE_DESCRIPTION("Host kernel accelerator for virtio");