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/kthread.h>
  26 #include <linux/cgroup.h>
  27 #include <linux/module.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) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  37 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  38
  39 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  40 static void vhost_vq_reset_user_be(struct vhost_virtqueue *vq)
  41 {
  42         vq->user_be = !virtio_legacy_is_little_endian();
  43 }
  44
  45 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  46 {
  47         struct vhost_vring_state s;
  48
  49         if (vq->private_data)
  50                 return -EBUSY;
  51
  52         if (copy_from_user(&s, argp, sizeof(s)))
  53                 return -EFAULT;
  54
  55         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  56             s.num != VHOST_VRING_BIG_ENDIAN)
  57                 return -EINVAL;
  58
  59         vq->user_be = s.num;
  60
  61         return 0;
  62 }
  63
  64 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  65                                    int __user *argp)
  66 {
  67         struct vhost_vring_state s = {
  68                 .index = idx,
  69                 .num = vq->user_be
  70         };
  71
  72         if (copy_to_user(argp, &s, sizeof(s)))
  73                 return -EFAULT;
  74
  75         return 0;
  76 }
  77
  78 static void vhost_init_is_le(struct vhost_virtqueue *vq)
  79 {
  80         /* Note for legacy virtio: user_be is initialized at reset time
  81          * according to the host endianness. If userspace does not set an
  82          * explicit endianness, the default behavior is native endian, as
  83          * expected by legacy virtio.
  84          */
  85         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
  86 }
  87 #else
  88 static void vhost_vq_reset_user_be(struct vhost_virtqueue *vq)
  89 {
  90 }
  91
  92 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  93 {
  94         return -ENOIOCTLCMD;
  95 }
  96
  97 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  98                                    int __user *argp)
  99 {
 100         return -ENOIOCTLCMD;
 101 }
 102
 103 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 104 {
 105         if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
 106                 vq->is_le = true;
 107 }
 108 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 109
 110 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 111                             poll_table *pt)
 112 {
 113         struct vhost_poll *poll;
 114
 115         poll = container_of(pt, struct vhost_poll, table);
 116         poll->wqh = wqh;
 117         add_wait_queue(wqh, &poll->wait);
 118 }
 119
 120 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
 121                              void *key)
 122 {
 123         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 124
 125         if (!((unsigned long)key & poll->mask))
 126                 return 0;
 127
 128         vhost_poll_queue(poll);
 129         return 0;
 130 }
 131
 132 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 133 {
 134         INIT_LIST_HEAD(&work->node);
 135         work->fn = fn;
 136         init_waitqueue_head(&work->done);
 137         work->flushing = 0;
 138         work->queue_seq = work->done_seq = 0;
 139 }
 140 EXPORT_SYMBOL_GPL(vhost_work_init);
 141
 142 /* Init poll structure */
 143 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 144                      unsigned long mask, struct vhost_dev *dev)
 145 {
 146         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 147         init_poll_funcptr(&poll->table, vhost_poll_func);
 148         poll->mask = mask;
 149         poll->dev = dev;
 150         poll->wqh = NULL;
 151
 152         vhost_work_init(&poll->work, fn);
 153 }
 154 EXPORT_SYMBOL_GPL(vhost_poll_init);
 155
 156 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 157  * keep a reference to a file until after vhost_poll_stop is called. */
 158 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 159 {
 160         unsigned long mask;
 161         int ret = 0;
 162
 163         if (poll->wqh)
 164                 return 0;
 165
 166         mask = file->f_op->poll(file, &poll->table);
 167         if (mask)
 168                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 169         if (mask & POLLERR) {
 170                 if (poll->wqh)
 171                         remove_wait_queue(poll->wqh, &poll->wait);
 172                 ret = -EINVAL;
 173         }
 174
 175         return ret;
 176 }
 177 EXPORT_SYMBOL_GPL(vhost_poll_start);
 178
 179 /* Stop polling a file. After this function returns, it becomes safe to drop the
 180  * file reference. You must also flush afterwards. */
 181 void vhost_poll_stop(struct vhost_poll *poll)
 182 {
 183         if (poll->wqh) {
 184                 remove_wait_queue(poll->wqh, &poll->wait);
 185                 poll->wqh = NULL;
 186         }
 187 }
 188 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 189
 190 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
 191                                 unsigned seq)
 192 {
 193         int left;
 194
 195         spin_lock_irq(&dev->work_lock);
 196         left = seq - work->done_seq;
 197         spin_unlock_irq(&dev->work_lock);
 198         return left <= 0;
 199 }
 200
 201 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 202 {
 203         unsigned seq;
 204         int flushing;
 205
 206         spin_lock_irq(&dev->work_lock);
 207         seq = work->queue_seq;
 208         work->flushing++;
 209         spin_unlock_irq(&dev->work_lock);
 210         wait_event(work->done, vhost_work_seq_done(dev, work, seq));
 211         spin_lock_irq(&dev->work_lock);
 212         flushing = --work->flushing;
 213         spin_unlock_irq(&dev->work_lock);
 214         BUG_ON(flushing < 0);
 215 }
 216 EXPORT_SYMBOL_GPL(vhost_work_flush);
 217
 218 /* Flush any work that has been scheduled. When calling this, don't hold any
 219  * locks that are also used by the callback. */
 220 void vhost_poll_flush(struct vhost_poll *poll)
 221 {
 222         vhost_work_flush(poll->dev, &poll->work);
 223 }
 224 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 225
 226 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 227 {
 228         unsigned long flags;
 229
 230         spin_lock_irqsave(&dev->work_lock, flags);
 231         if (list_empty(&work->node)) {
 232                 list_add_tail(&work->node, &dev->work_list);
 233                 work->queue_seq++;
 234                 spin_unlock_irqrestore(&dev->work_lock, flags);
 235                 wake_up_process(dev->worker);
 236         } else {
 237                 spin_unlock_irqrestore(&dev->work_lock, flags);
 238         }
 239 }
 240 EXPORT_SYMBOL_GPL(vhost_work_queue);
 241
 242 void vhost_poll_queue(struct vhost_poll *poll)
 243 {
 244         vhost_work_queue(poll->dev, &poll->work);
 245 }
 246 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 247
 248 static void vhost_vq_reset(struct vhost_dev *dev,
 249                            struct vhost_virtqueue *vq)
 250 {
 251         vq->num = 1;
 252         vq->desc = NULL;
 253         vq->avail = NULL;
 254         vq->used = NULL;
 255         vq->last_avail_idx = 0;
 256         vq->avail_idx = 0;
 257         vq->last_used_idx = 0;
 258         vq->signalled_used = 0;
 259         vq->signalled_used_valid = false;
 260         vq->used_flags = 0;
 261         vq->log_used = false;
 262         vq->log_addr = -1ull;
 263         vq->private_data = NULL;
 264         vq->acked_features = 0;
 265         vq->log_base = NULL;
 266         vq->error_ctx = NULL;
 267         vq->error = NULL;
 268         vq->kick = NULL;
 269         vq->call_ctx = NULL;
 270         vq->call = NULL;
 271         vq->log_ctx = NULL;
 272         vq->memory = NULL;
 273         vq->is_le = virtio_legacy_is_little_endian();
 274         vhost_vq_reset_user_be(vq);
 275 }
 276
 277 static int vhost_worker(void *data)
 278 {
 279         struct vhost_dev *dev = data;
 280         struct vhost_work *work = NULL;
 281         unsigned uninitialized_var(seq);
 282         mm_segment_t oldfs = get_fs();
 283
 284         set_fs(USER_DS);
 285         use_mm(dev->mm);
 286
 287         for (;;) {
 288                 /* mb paired w/ kthread_stop */
 289                 set_current_state(TASK_INTERRUPTIBLE);
 290
 291                 spin_lock_irq(&dev->work_lock);
 292                 if (work) {
 293                         work->done_seq = seq;
 294                         if (work->flushing)
 295                                 wake_up_all(&work->done);
 296                 }
 297
 298                 if (kthread_should_stop()) {
 299                         spin_unlock_irq(&dev->work_lock);
 300                         __set_current_state(TASK_RUNNING);
 301                         break;
 302                 }
 303                 if (!list_empty(&dev->work_list)) {
 304                         work = list_first_entry(&dev->work_list,
 305                                                 struct vhost_work, node);
 306                         list_del_init(&work->node);
 307                         seq = work->queue_seq;
 308                 } else
 309                         work = NULL;
 310                 spin_unlock_irq(&dev->work_lock);
 311
 312                 if (work) {
 313                         __set_current_state(TASK_RUNNING);
 314                         work->fn(work);
 315                         if (need_resched())
 316                                 schedule();
 317                 } else
 318                         schedule();
 319
 320         }
 321         unuse_mm(dev->mm);
 322         set_fs(oldfs);
 323         return 0;
 324 }
 325
 326 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 327 {
 328         kfree(vq->indirect);
 329         vq->indirect = NULL;
 330         kfree(vq->log);
 331         vq->log = NULL;
 332         kfree(vq->heads);
 333         vq->heads = NULL;
 334 }
 335
 336 /* Helper to allocate iovec buffers for all vqs. */
 337 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 338 {
 339         struct vhost_virtqueue *vq;
 340         int i;
 341
 342         for (i = 0; i < dev->nvqs; ++i) {
 343                 vq = dev->vqs[i];
 344                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 345                                        GFP_KERNEL);
 346                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 347                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 348                 if (!vq->indirect || !vq->log || !vq->heads)
 349                         goto err_nomem;
 350         }
 351         return 0;
 352
 353 err_nomem:
 354         for (; i >= 0; --i)
 355                 vhost_vq_free_iovecs(dev->vqs[i]);
 356         return -ENOMEM;
 357 }
 358
 359 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 360 {
 361         int i;
 362
 363         for (i = 0; i < dev->nvqs; ++i)
 364                 vhost_vq_free_iovecs(dev->vqs[i]);
 365 }
 366
 367 void vhost_dev_init(struct vhost_dev *dev,
 368                     struct vhost_virtqueue **vqs, int nvqs)
 369 {
 370         struct vhost_virtqueue *vq;
 371         int i;
 372
 373         dev->vqs = vqs;
 374         dev->nvqs = nvqs;
 375         mutex_init(&dev->mutex);
 376         dev->log_ctx = NULL;
 377         dev->log_file = NULL;
 378         dev->memory = NULL;
 379         dev->mm = NULL;
 380         spin_lock_init(&dev->work_lock);
 381         INIT_LIST_HEAD(&dev->work_list);
 382         dev->worker = NULL;
 383
 384         for (i = 0; i < dev->nvqs; ++i) {
 385                 vq = dev->vqs[i];
 386                 vq->log = NULL;
 387                 vq->indirect = NULL;
 388                 vq->heads = NULL;
 389                 vq->dev = dev;
 390                 mutex_init(&vq->mutex);
 391                 vhost_vq_reset(dev, vq);
 392                 if (vq->handle_kick)
 393                         vhost_poll_init(&vq->poll, vq->handle_kick,
 394                                         POLLIN, dev);
 395         }
 396 }
 397 EXPORT_SYMBOL_GPL(vhost_dev_init);
 398
 399 /* Caller should have device mutex */
 400 long vhost_dev_check_owner(struct vhost_dev *dev)
 401 {
 402         /* Are you the owner? If not, I don't think you mean to do that */
 403         return dev->mm == current->mm ? 0 : -EPERM;
 404 }
 405 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 406
 407 struct vhost_attach_cgroups_struct {
 408         struct vhost_work work;
 409         struct task_struct *owner;
 410         int ret;
 411 };
 412
 413 static void vhost_attach_cgroups_work(struct vhost_work *work)
 414 {
 415         struct vhost_attach_cgroups_struct *s;
 416
 417         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 418         s->ret = cgroup_attach_task_all(s->owner, current);
 419 }
 420
 421 static int vhost_attach_cgroups(struct vhost_dev *dev)
 422 {
 423         struct vhost_attach_cgroups_struct attach;
 424
 425         attach.owner = current;
 426         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 427         vhost_work_queue(dev, &attach.work);
 428         vhost_work_flush(dev, &attach.work);
 429         return attach.ret;
 430 }
 431
 432 /* Caller should have device mutex */
 433 bool vhost_dev_has_owner(struct vhost_dev *dev)
 434 {
 435         return dev->mm;
 436 }
 437 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 438
 439 /* Caller should have device mutex */
 440 long vhost_dev_set_owner(struct vhost_dev *dev)
 441 {
 442         struct task_struct *worker;
 443         int err;
 444
 445         /* Is there an owner already? */
 446         if (vhost_dev_has_owner(dev)) {
 447                 err = -EBUSY;
 448                 goto err_mm;
 449         }
 450
 451         /* No owner, become one */
 452         dev->mm = get_task_mm(current);
 453         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 454         if (IS_ERR(worker)) {
 455                 err = PTR_ERR(worker);
 456                 goto err_worker;
 457         }
 458
 459         dev->worker = worker;
 460         wake_up_process(worker);        /* avoid contributing to loadavg */
 461
 462         err = vhost_attach_cgroups(dev);
 463         if (err)
 464                 goto err_cgroup;
 465
 466         err = vhost_dev_alloc_iovecs(dev);
 467         if (err)
 468                 goto err_cgroup;
 469
 470         return 0;
 471 err_cgroup:
 472         kthread_stop(worker);
 473         dev->worker = NULL;
 474 err_worker:
 475         if (dev->mm)
 476                 mmput(dev->mm);
 477         dev->mm = NULL;
 478 err_mm:
 479         return err;
 480 }
 481 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 482
 483 struct vhost_memory *vhost_dev_reset_owner_prepare(void)
 484 {
 485         return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 486 }
 487 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 488
 489 /* Caller should have device mutex */
 490 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
 491 {
 492         int i;
 493
 494         vhost_dev_cleanup(dev, true);
 495
 496         /* Restore memory to default empty mapping. */
 497         memory->nregions = 0;
 498         dev->memory = memory;
 499         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 500          * VQs aren't running.
 501          */
 502         for (i = 0; i < dev->nvqs; ++i)
 503                 dev->vqs[i]->memory = memory;
 504 }
 505 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 506
 507 void vhost_dev_stop(struct vhost_dev *dev)
 508 {
 509         int i;
 510
 511         for (i = 0; i < dev->nvqs; ++i) {
 512                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 513                         vhost_poll_stop(&dev->vqs[i]->poll);
 514                         vhost_poll_flush(&dev->vqs[i]->poll);
 515                 }
 516         }
 517 }
 518 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 519
 520 /* Caller should have device mutex if and only if locked is set */
 521 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 522 {
 523         int i;
 524
 525         for (i = 0; i < dev->nvqs; ++i) {
 526                 if (dev->vqs[i]->error_ctx)
 527                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 528                 if (dev->vqs[i]->error)
 529                         fput(dev->vqs[i]->error);
 530                 if (dev->vqs[i]->kick)
 531                         fput(dev->vqs[i]->kick);
 532                 if (dev->vqs[i]->call_ctx)
 533                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 534                 if (dev->vqs[i]->call)
 535                         fput(dev->vqs[i]->call);
 536                 vhost_vq_reset(dev, dev->vqs[i]);
 537         }
 538         vhost_dev_free_iovecs(dev);
 539         if (dev->log_ctx)
 540                 eventfd_ctx_put(dev->log_ctx);
 541         dev->log_ctx = NULL;
 542         if (dev->log_file)
 543                 fput(dev->log_file);
 544         dev->log_file = NULL;
 545         /* No one will access memory at this point */
 546         kfree(dev->memory);
 547         dev->memory = NULL;
 548         WARN_ON(!list_empty(&dev->work_list));
 549         if (dev->worker) {
 550                 kthread_stop(dev->worker);
 551                 dev->worker = NULL;
 552         }
 553         if (dev->mm)
 554                 mmput(dev->mm);
 555         dev->mm = NULL;
 556 }
 557 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 558
 559 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 560 {
 561         u64 a = addr / VHOST_PAGE_SIZE / 8;
 562
 563         /* Make sure 64 bit math will not overflow. */
 564         if (a > ULONG_MAX - (unsigned long)log_base ||
 565             a + (unsigned long)log_base > ULONG_MAX)
 566                 return 0;
 567
 568         return access_ok(VERIFY_WRITE, log_base + a,
 569                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 570 }
 571
 572 /* Caller should have vq mutex and device mutex. */
 573 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 574                                int log_all)
 575 {
 576         int i;
 577
 578         if (!mem)
 579                 return 0;
 580
 581         for (i = 0; i < mem->nregions; ++i) {
 582                 struct vhost_memory_region *m = mem->regions + i;
 583                 unsigned long a = m->userspace_addr;
 584                 if (m->memory_size > ULONG_MAX)
 585                         return 0;
 586                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 587                                     m->memory_size))
 588                         return 0;
 589                 else if (log_all && !log_access_ok(log_base,
 590                                                    m->guest_phys_addr,
 591                                                    m->memory_size))
 592                         return 0;
 593         }
 594         return 1;
 595 }
 596
 597 /* Can we switch to this memory table? */
 598 /* Caller should have device mutex but not vq mutex */
 599 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 600                             int log_all)
 601 {
 602         int i;
 603
 604         for (i = 0; i < d->nvqs; ++i) {
 605                 int ok;
 606                 bool log;
 607
 608                 mutex_lock(&d->vqs[i]->mutex);
 609                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 610                 /* If ring is inactive, will check when it's enabled. */
 611                 if (d->vqs[i]->private_data)
 612                         ok = vq_memory_access_ok(d->vqs[i]->log_base, mem, log);
 613                 else
 614                         ok = 1;
 615                 mutex_unlock(&d->vqs[i]->mutex);
 616                 if (!ok)
 617                         return 0;
 618         }
 619         return 1;
 620 }
 621
 622 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
 623                         struct vring_desc __user *desc,
 624                         struct vring_avail __user *avail,
 625                         struct vring_used __user *used)
 626 {
 627         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 628         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 629                access_ok(VERIFY_READ, avail,
 630                          sizeof *avail + num * sizeof *avail->ring + s) &&
 631                access_ok(VERIFY_WRITE, used,
 632                         sizeof *used + num * sizeof *used->ring + s);
 633 }
 634
 635 /* Can we log writes? */
 636 /* Caller should have device mutex but not vq mutex */
 637 int vhost_log_access_ok(struct vhost_dev *dev)
 638 {
 639         return memory_access_ok(dev, dev->memory, 1);
 640 }
 641 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
 642
 643 /* Verify access for write logging. */
 644 /* Caller should have vq mutex and device mutex */
 645 static int vq_log_access_ok(struct vhost_virtqueue *vq,
 646                             void __user *log_base)
 647 {
 648         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 649
 650         return vq_memory_access_ok(log_base, vq->memory,
 651                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
 652                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 653                                         sizeof *vq->used +
 654                                         vq->num * sizeof *vq->used->ring + s));
 655 }
 656
 657 /* Can we start vq? */
 658 /* Caller should have vq mutex and device mutex */
 659 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 660 {
 661         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
 662                 vq_log_access_ok(vq, vq->log_base);
 663 }
 664 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
 665
 666 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 667 {
 668         struct vhost_memory mem, *newmem, *oldmem;
 669         unsigned long size = offsetof(struct vhost_memory, regions);
 670         int i;
 671
 672         if (copy_from_user(&mem, m, size))
 673                 return -EFAULT;
 674         if (mem.padding)
 675                 return -EOPNOTSUPP;
 676         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 677                 return -E2BIG;
 678         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 679         if (!newmem)
 680                 return -ENOMEM;
 681
 682         memcpy(newmem, &mem, size);
 683         if (copy_from_user(newmem->regions, m->regions,
 684                            mem.nregions * sizeof *m->regions)) {
 685                 kfree(newmem);
 686                 return -EFAULT;
 687         }
 688
 689         if (!memory_access_ok(d, newmem, 0)) {
 690                 kfree(newmem);
 691                 return -EFAULT;
 692         }
 693         oldmem = d->memory;
 694         d->memory = newmem;
 695
 696         /* All memory accesses are done under some VQ mutex. */
 697         for (i = 0; i < d->nvqs; ++i) {
 698                 mutex_lock(&d->vqs[i]->mutex);
 699                 d->vqs[i]->memory = newmem;
 700                 mutex_unlock(&d->vqs[i]->mutex);
 701         }
 702         kfree(oldmem);
 703         return 0;
 704 }
 705
 706 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 707 {
 708         struct file *eventfp, *filep = NULL;
 709         bool pollstart = false, pollstop = false;
 710         struct eventfd_ctx *ctx = NULL;
 711         u32 __user *idxp = argp;
 712         struct vhost_virtqueue *vq;
 713         struct vhost_vring_state s;
 714         struct vhost_vring_file f;
 715         struct vhost_vring_addr a;
 716         u32 idx;
 717         long r;
 718
 719         r = get_user(idx, idxp);
 720         if (r < 0)
 721                 return r;
 722         if (idx >= d->nvqs)
 723                 return -ENOBUFS;
 724
 725         vq = d->vqs[idx];
 726
 727         mutex_lock(&vq->mutex);
 728
 729         switch (ioctl) {
 730         case VHOST_SET_VRING_NUM:
 731                 /* Resizing ring with an active backend?
 732                  * You don't want to do that. */
 733                 if (vq->private_data) {
 734                         r = -EBUSY;
 735                         break;
 736                 }
 737                 if (copy_from_user(&s, argp, sizeof s)) {
 738                         r = -EFAULT;
 739                         break;
 740                 }
 741                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 742                         r = -EINVAL;
 743                         break;
 744                 }
 745                 vq->num = s.num;
 746                 break;
 747         case VHOST_SET_VRING_BASE:
 748                 /* Moving base with an active backend?
 749                  * You don't want to do that. */
 750                 if (vq->private_data) {
 751                         r = -EBUSY;
 752                         break;
 753                 }
 754                 if (copy_from_user(&s, argp, sizeof s)) {
 755                         r = -EFAULT;
 756                         break;
 757                 }
 758                 if (s.num > 0xffff) {
 759                         r = -EINVAL;
 760                         break;
 761                 }
 762                 vq->last_avail_idx = s.num;
 763                 /* Forget the cached index value. */
 764                 vq->avail_idx = vq->last_avail_idx;
 765                 break;
 766         case VHOST_GET_VRING_BASE:
 767                 s.index = idx;
 768                 s.num = vq->last_avail_idx;
 769                 if (copy_to_user(argp, &s, sizeof s))
 770                         r = -EFAULT;
 771                 break;
 772         case VHOST_SET_VRING_ADDR:
 773                 if (copy_from_user(&a, argp, sizeof a)) {
 774                         r = -EFAULT;
 775                         break;
 776                 }
 777                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 778                         r = -EOPNOTSUPP;
 779                         break;
 780                 }
 781                 /* For 32bit, verify that the top 32bits of the user
 782                    data are set to zero. */
 783                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 784                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 785                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 786                         r = -EFAULT;
 787                         break;
 788                 }
 789
 790                 /* Make sure it's safe to cast pointers to vring types. */
 791                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
 792                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
 793                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
 794                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
 795                     (a.log_guest_addr & (sizeof(u64) - 1))) {
 796                         r = -EINVAL;
 797                         break;
 798                 }
 799
 800                 /* We only verify access here if backend is configured.
 801                  * If it is not, we don't as size might not have been setup.
 802                  * We will verify when backend is configured. */
 803                 if (vq->private_data) {
 804                         if (!vq_access_ok(vq, vq->num,
 805                                 (void __user *)(unsigned long)a.desc_user_addr,
 806                                 (void __user *)(unsigned long)a.avail_user_addr,
 807                                 (void __user *)(unsigned long)a.used_user_addr)) {
 808                                 r = -EINVAL;
 809                                 break;
 810                         }
 811
 812                         /* Also validate log access for used ring if enabled. */
 813                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 814                             !log_access_ok(vq->log_base, a.log_guest_addr,
 815                                            sizeof *vq->used +
 816                                            vq->num * sizeof *vq->used->ring)) {
 817                                 r = -EINVAL;
 818                                 break;
 819                         }
 820                 }
 821
 822                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 823                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 824                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 825                 vq->log_addr = a.log_guest_addr;
 826                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 827                 break;
 828         case VHOST_SET_VRING_KICK:
 829                 if (copy_from_user(&f, argp, sizeof f)) {
 830                         r = -EFAULT;
 831                         break;
 832                 }
 833                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 834                 if (IS_ERR(eventfp)) {
 835                         r = PTR_ERR(eventfp);
 836                         break;
 837                 }
 838                 if (eventfp != vq->kick) {
 839                         pollstop = (filep = vq->kick) != NULL;
 840                         pollstart = (vq->kick = eventfp) != NULL;
 841                 } else
 842                         filep = eventfp;
 843                 break;
 844         case VHOST_SET_VRING_CALL:
 845                 if (copy_from_user(&f, argp, sizeof f)) {
 846                         r = -EFAULT;
 847                         break;
 848                 }
 849                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 850                 if (IS_ERR(eventfp)) {
 851                         r = PTR_ERR(eventfp);
 852                         break;
 853                 }
 854                 if (eventfp != vq->call) {
 855                         filep = vq->call;
 856                         ctx = vq->call_ctx;
 857                         vq->call = eventfp;
 858                         vq->call_ctx = eventfp ?
 859                                 eventfd_ctx_fileget(eventfp) : NULL;
 860                 } else
 861                         filep = eventfp;
 862                 break;
 863         case VHOST_SET_VRING_ERR:
 864                 if (copy_from_user(&f, argp, sizeof f)) {
 865                         r = -EFAULT;
 866                         break;
 867                 }
 868                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 869                 if (IS_ERR(eventfp)) {
 870                         r = PTR_ERR(eventfp);
 871                         break;
 872                 }
 873                 if (eventfp != vq->error) {
 874                         filep = vq->error;
 875                         vq->error = eventfp;
 876                         ctx = vq->error_ctx;
 877                         vq->error_ctx = eventfp ?
 878                                 eventfd_ctx_fileget(eventfp) : NULL;
 879                 } else
 880                         filep = eventfp;
 881                 break;
 882         case VHOST_SET_VRING_ENDIAN:
 883                 r = vhost_set_vring_endian(vq, argp);
 884                 break;
 885         case VHOST_GET_VRING_ENDIAN:
 886                 r = vhost_get_vring_endian(vq, idx, argp);
 887                 break;
 888         default:
 889                 r = -ENOIOCTLCMD;
 890         }
 891
 892         if (pollstop && vq->handle_kick)
 893                 vhost_poll_stop(&vq->poll);
 894
 895         if (ctx)
 896                 eventfd_ctx_put(ctx);
 897         if (filep)
 898                 fput(filep);
 899
 900         if (pollstart && vq->handle_kick)
 901                 r = vhost_poll_start(&vq->poll, vq->kick);
 902
 903         mutex_unlock(&vq->mutex);
 904
 905         if (pollstop && vq->handle_kick)
 906                 vhost_poll_flush(&vq->poll);
 907         return r;
 908 }
 909 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
 910
 911 /* Caller must have device mutex */
 912 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
 913 {
 914         struct file *eventfp, *filep = NULL;
 915         struct eventfd_ctx *ctx = NULL;
 916         u64 p;
 917         long r;
 918         int i, fd;
 919
 920         /* If you are not the owner, you can become one */
 921         if (ioctl == VHOST_SET_OWNER) {
 922                 r = vhost_dev_set_owner(d);
 923                 goto done;
 924         }
 925
 926         /* You must be the owner to do anything else */
 927         r = vhost_dev_check_owner(d);
 928         if (r)
 929                 goto done;
 930
 931         switch (ioctl) {
 932         case VHOST_SET_MEM_TABLE:
 933                 r = vhost_set_memory(d, argp);
 934                 break;
 935         case VHOST_SET_LOG_BASE:
 936                 if (copy_from_user(&p, argp, sizeof p)) {
 937                         r = -EFAULT;
 938                         break;
 939                 }
 940                 if ((u64)(unsigned long)p != p) {
 941                         r = -EFAULT;
 942                         break;
 943                 }
 944                 for (i = 0; i < d->nvqs; ++i) {
 945                         struct vhost_virtqueue *vq;
 946                         void __user *base = (void __user *)(unsigned long)p;
 947                         vq = d->vqs[i];
 948                         mutex_lock(&vq->mutex);
 949                         /* If ring is inactive, will check when it's enabled. */
 950                         if (vq->private_data && !vq_log_access_ok(vq, base))
 951                                 r = -EFAULT;
 952                         else
 953                                 vq->log_base = base;
 954                         mutex_unlock(&vq->mutex);
 955                 }
 956                 break;
 957         case VHOST_SET_LOG_FD:
 958                 r = get_user(fd, (int __user *)argp);
 959                 if (r < 0)
 960                         break;
 961                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 962                 if (IS_ERR(eventfp)) {
 963                         r = PTR_ERR(eventfp);
 964                         break;
 965                 }
 966                 if (eventfp != d->log_file) {
 967                         filep = d->log_file;
 968                         ctx = d->log_ctx;
 969                         d->log_ctx = eventfp ?
 970                                 eventfd_ctx_fileget(eventfp) : NULL;
 971                 } else
 972                         filep = eventfp;
 973                 for (i = 0; i < d->nvqs; ++i) {
 974                         mutex_lock(&d->vqs[i]->mutex);
 975                         d->vqs[i]->log_ctx = d->log_ctx;
 976                         mutex_unlock(&d->vqs[i]->mutex);
 977                 }
 978                 if (ctx)
 979                         eventfd_ctx_put(ctx);
 980                 if (filep)
 981                         fput(filep);
 982                 break;
 983         default:
 984                 r = -ENOIOCTLCMD;
 985                 break;
 986         }
 987 done:
 988         return r;
 989 }
 990 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
 991
 992 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 993                                                      __u64 addr, __u32 len)
 994 {
 995         struct vhost_memory_region *reg;
 996         int i;
 997
 998         /* linear search is not brilliant, but we really have on the order of 6
 999          * regions in practice */
1000         for (i = 0; i < mem->nregions; ++i) {
1001                 reg = mem->regions + i;
1002                 if (reg->guest_phys_addr <= addr &&
1003                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
1004                         return reg;
1005         }
1006         return NULL;
1007 }
1008
1009 /* TODO: This is really inefficient.  We need something like get_user()
1010  * (instruction directly accesses the data, with an exception table entry
1011  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1012  */
1013 static int set_bit_to_user(int nr, void __user *addr)
1014 {
1015         unsigned long log = (unsigned long)addr;
1016         struct page *page;
1017         void *base;
1018         int bit = nr + (log % PAGE_SIZE) * 8;
1019         int r;
1020
1021         r = get_user_pages_fast(log, 1, 1, &page);
1022         if (r < 0)
1023                 return r;
1024         BUG_ON(r != 1);
1025         base = kmap_atomic(page);
1026         set_bit(bit, base);
1027         kunmap_atomic(base);
1028         set_page_dirty_lock(page);
1029         put_page(page);
1030         return 0;
1031 }
1032
1033 static int log_write(void __user *log_base,
1034                      u64 write_address, u64 write_length)
1035 {
1036         u64 write_page = write_address / VHOST_PAGE_SIZE;
1037         int r;
1038
1039         if (!write_length)
1040                 return 0;
1041         write_length += write_address % VHOST_PAGE_SIZE;
1042         for (;;) {
1043                 u64 base = (u64)(unsigned long)log_base;
1044                 u64 log = base + write_page / 8;
1045                 int bit = write_page % 8;
1046                 if ((u64)(unsigned long)log != log)
1047                         return -EFAULT;
1048                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1049                 if (r < 0)
1050                         return r;
1051                 if (write_length <= VHOST_PAGE_SIZE)
1052                         break;
1053                 write_length -= VHOST_PAGE_SIZE;
1054                 write_page += 1;
1055         }
1056         return r;
1057 }
1058
1059 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1060                     unsigned int log_num, u64 len)
1061 {
1062         int i, r;
1063
1064         /* Make sure data written is seen before log. */
1065         smp_wmb();
1066         for (i = 0; i < log_num; ++i) {
1067                 u64 l = min(log[i].len, len);
1068                 r = log_write(vq->log_base, log[i].addr, l);
1069                 if (r < 0)
1070                         return r;
1071                 len -= l;
1072                 if (!len) {
1073                         if (vq->log_ctx)
1074                                 eventfd_signal(vq->log_ctx, 1);
1075                         return 0;
1076                 }
1077         }
1078         /* Length written exceeds what we have stored. This is a bug. */
1079         BUG();
1080         return 0;
1081 }
1082 EXPORT_SYMBOL_GPL(vhost_log_write);
1083
1084 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1085 {
1086         void __user *used;
1087         if (__put_user(cpu_to_vhost16(vq, vq->used_flags), &vq->used->flags) < 0)
1088                 return -EFAULT;
1089         if (unlikely(vq->log_used)) {
1090                 /* Make sure the flag is seen before log. */
1091                 smp_wmb();
1092                 /* Log used flag write. */
1093                 used = &vq->used->flags;
1094                 log_write(vq->log_base, vq->log_addr +
1095                           (used - (void __user *)vq->used),
1096                           sizeof vq->used->flags);
1097                 if (vq->log_ctx)
1098                         eventfd_signal(vq->log_ctx, 1);
1099         }
1100         return 0;
1101 }
1102
1103 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1104 {
1105         if (__put_user(cpu_to_vhost16(vq, vq->avail_idx), vhost_avail_event(vq)))
1106                 return -EFAULT;
1107         if (unlikely(vq->log_used)) {
1108                 void __user *used;
1109                 /* Make sure the event is seen before log. */
1110                 smp_wmb();
1111                 /* Log avail event write */
1112                 used = vhost_avail_event(vq);
1113                 log_write(vq->log_base, vq->log_addr +
1114                           (used - (void __user *)vq->used),
1115                           sizeof *vhost_avail_event(vq));
1116                 if (vq->log_ctx)
1117                         eventfd_signal(vq->log_ctx, 1);
1118         }
1119         return 0;
1120 }
1121
1122 int vhost_init_used(struct vhost_virtqueue *vq)
1123 {
1124         __virtio16 last_used_idx;
1125         int r;
1126         if (!vq->private_data) {
1127                 vq->is_le = virtio_legacy_is_little_endian();
1128                 return 0;
1129         }
1130
1131         vhost_init_is_le(vq);
1132
1133         r = vhost_update_used_flags(vq);
1134         if (r)
1135                 return r;
1136         vq->signalled_used_valid = false;
1137         if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx))
1138                 return -EFAULT;
1139         r = __get_user(last_used_idx, &vq->used->idx);
1140         if (r)
1141                 return r;
1142         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1143         return 0;
1144 }
1145 EXPORT_SYMBOL_GPL(vhost_init_used);
1146
1147 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1148                           struct iovec iov[], int iov_size)
1149 {
1150         const struct vhost_memory_region *reg;
1151         struct vhost_memory *mem;
1152         struct iovec *_iov;
1153         u64 s = 0;
1154         int ret = 0;
1155
1156         mem = vq->memory;
1157         while ((u64)len > s) {
1158                 u64 size;
1159                 if (unlikely(ret >= iov_size)) {
1160                         ret = -ENOBUFS;
1161                         break;
1162                 }
1163                 reg = find_region(mem, addr, len);
1164                 if (unlikely(!reg)) {
1165                         ret = -EFAULT;
1166                         break;
1167                 }
1168                 _iov = iov + ret;
1169                 size = reg->memory_size - addr + reg->guest_phys_addr;
1170                 _iov->iov_len = min((u64)len - s, size);
1171                 _iov->iov_base = (void __user *)(unsigned long)
1172                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1173                 s += size;
1174                 addr += size;
1175                 ++ret;
1176         }
1177
1178         return ret;
1179 }
1180
1181 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1182  * function returns the next descriptor in the chain,
1183  * or -1U if we're at the end. */
1184 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1185 {
1186         unsigned int next;
1187
1188         /* If this descriptor says it doesn't chain, we're done. */
1189         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1190                 return -1U;
1191
1192         /* Check they're not leading us off end of descriptors. */
1193         next = vhost16_to_cpu(vq, desc->next);
1194         /* Make sure compiler knows to grab that: we don't want it changing! */
1195         /* We will use the result as an index in an array, so most
1196          * architectures only need a compiler barrier here. */
1197         read_barrier_depends();
1198
1199         return next;
1200 }
1201
1202 static int get_indirect(struct vhost_virtqueue *vq,
1203                         struct iovec iov[], unsigned int iov_size,
1204                         unsigned int *out_num, unsigned int *in_num,
1205                         struct vhost_log *log, unsigned int *log_num,
1206                         struct vring_desc *indirect)
1207 {
1208         struct vring_desc desc;
1209         unsigned int i = 0, count, found = 0;
1210         u32 len = vhost32_to_cpu(vq, indirect->len);
1211         struct iov_iter from;
1212         int ret;
1213
1214         /* Sanity check */
1215         if (unlikely(len % sizeof desc)) {
1216                 vq_err(vq, "Invalid length in indirect descriptor: "
1217                        "len 0x%llx not multiple of 0x%zx\n",
1218                        (unsigned long long)len,
1219                        sizeof desc);
1220                 return -EINVAL;
1221         }
1222
1223         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1224                              UIO_MAXIOV);
1225         if (unlikely(ret < 0)) {
1226                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1227                 return ret;
1228         }
1229         iov_iter_init(&from, READ, vq->indirect, ret, len);
1230
1231         /* We will use the result as an address to read from, so most
1232          * architectures only need a compiler barrier here. */
1233         read_barrier_depends();
1234
1235         count = len / sizeof desc;
1236         /* Buffers are chained via a 16 bit next field, so
1237          * we can have at most 2^16 of these. */
1238         if (unlikely(count > USHRT_MAX + 1)) {
1239                 vq_err(vq, "Indirect buffer length too big: %d\n",
1240                        indirect->len);
1241                 return -E2BIG;
1242         }
1243
1244         do {
1245                 unsigned iov_count = *in_num + *out_num;
1246                 if (unlikely(++found > count)) {
1247                         vq_err(vq, "Loop detected: last one at %u "
1248                                "indirect size %u\n",
1249                                i, count);
1250                         return -EINVAL;
1251                 }
1252                 if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1253                              sizeof(desc))) {
1254                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1255                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1256                         return -EINVAL;
1257                 }
1258                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1259                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1260                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1261                         return -EINVAL;
1262                 }
1263
1264                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1265                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1266                                      iov_size - iov_count);
1267                 if (unlikely(ret < 0)) {
1268                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1269                                ret, i);
1270                         return ret;
1271                 }
1272                 /* If this is an input descriptor, increment that count. */
1273                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1274                         *in_num += ret;
1275                         if (unlikely(log)) {
1276                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1277                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1278                                 ++*log_num;
1279                         }
1280                 } else {
1281                         /* If it's an output descriptor, they're all supposed
1282                          * to come before any input descriptors. */
1283                         if (unlikely(*in_num)) {
1284                                 vq_err(vq, "Indirect descriptor "
1285                                        "has out after in: idx %d\n", i);
1286                                 return -EINVAL;
1287                         }
1288                         *out_num += ret;
1289                 }
1290         } while ((i = next_desc(vq, &desc)) != -1);
1291         return 0;
1292 }
1293
1294 /* This looks in the virtqueue and for the first available buffer, and converts
1295  * it to an iovec for convenient access.  Since descriptors consist of some
1296  * number of output then some number of input descriptors, it's actually two
1297  * iovecs, but we pack them into one and note how many of each there were.
1298  *
1299  * This function returns the descriptor number found, or vq->num (which is
1300  * never a valid descriptor number) if none was found.  A negative code is
1301  * returned on error. */
1302 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1303                       struct iovec iov[], unsigned int iov_size,
1304                       unsigned int *out_num, unsigned int *in_num,
1305                       struct vhost_log *log, unsigned int *log_num)
1306 {
1307         struct vring_desc desc;
1308         unsigned int i, head, found = 0;
1309         u16 last_avail_idx;
1310         __virtio16 avail_idx;
1311         __virtio16 ring_head;
1312         int ret;
1313
1314         /* Check it isn't doing very strange things with descriptor numbers. */
1315         last_avail_idx = vq->last_avail_idx;
1316         if (unlikely(__get_user(avail_idx, &vq->avail->idx))) {
1317                 vq_err(vq, "Failed to access avail idx at %p\n",
1318                        &vq->avail->idx);
1319                 return -EFAULT;
1320         }
1321         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1322
1323         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1324                 vq_err(vq, "Guest moved used index from %u to %u",
1325                        last_avail_idx, vq->avail_idx);
1326                 return -EFAULT;
1327         }
1328
1329         /* If there's nothing new since last we looked, return invalid. */
1330         if (vq->avail_idx == last_avail_idx)
1331                 return vq->num;
1332
1333         /* Only get avail ring entries after they have been exposed by guest. */
1334         smp_rmb();
1335
1336         /* Grab the next descriptor number they're advertising, and increment
1337          * the index we've seen. */
1338         if (unlikely(__get_user(ring_head,
1339                                 &vq->avail->ring[last_avail_idx % vq->num]))) {
1340                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1341                        last_avail_idx,
1342                        &vq->avail->ring[last_avail_idx % vq->num]);
1343                 return -EFAULT;
1344         }
1345
1346         head = vhost16_to_cpu(vq, ring_head);
1347
1348         /* If their number is silly, that's an error. */
1349         if (unlikely(head >= vq->num)) {
1350                 vq_err(vq, "Guest says index %u > %u is available",
1351                        head, vq->num);
1352                 return -EINVAL;
1353         }
1354
1355         /* When we start there are none of either input nor output. */
1356         *out_num = *in_num = 0;
1357         if (unlikely(log))
1358                 *log_num = 0;
1359
1360         i = head;
1361         do {
1362                 unsigned iov_count = *in_num + *out_num;
1363                 if (unlikely(i >= vq->num)) {
1364                         vq_err(vq, "Desc index is %u > %u, head = %u",
1365                                i, vq->num, head);
1366                         return -EINVAL;
1367                 }
1368                 if (unlikely(++found > vq->num)) {
1369                         vq_err(vq, "Loop detected: last one at %u "
1370                                "vq size %u head %u\n",
1371                                i, vq->num, head);
1372                         return -EINVAL;
1373                 }
1374                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1375                 if (unlikely(ret)) {
1376                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1377                                i, vq->desc + i);
1378                         return -EFAULT;
1379                 }
1380                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1381                         ret = get_indirect(vq, iov, iov_size,
1382                                            out_num, in_num,
1383                                            log, log_num, &desc);
1384                         if (unlikely(ret < 0)) {
1385                                 vq_err(vq, "Failure detected "
1386                                        "in indirect descriptor at idx %d\n", i);
1387                                 return ret;
1388                         }
1389                         continue;
1390                 }
1391
1392                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1393                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1394                                      iov_size - iov_count);
1395                 if (unlikely(ret < 0)) {
1396                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1397                                ret, i);
1398                         return ret;
1399                 }
1400                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1401                         /* If this is an input descriptor,
1402                          * increment that count. */
1403                         *in_num += ret;
1404                         if (unlikely(log)) {
1405                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1406                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1407                                 ++*log_num;
1408                         }
1409                 } else {
1410                         /* If it's an output descriptor, they're all supposed
1411                          * to come before any input descriptors. */
1412                         if (unlikely(*in_num)) {
1413                                 vq_err(vq, "Descriptor has out after in: "
1414                                        "idx %d\n", i);
1415                                 return -EINVAL;
1416                         }
1417                         *out_num += ret;
1418                 }
1419         } while ((i = next_desc(vq, &desc)) != -1);
1420
1421         /* On success, increment avail index. */
1422         vq->last_avail_idx++;
1423
1424         /* Assume notifications from guest are disabled at this point,
1425          * if they aren't we would need to update avail_event index. */
1426         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1427         return head;
1428 }
1429 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1430
1431 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1432 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1433 {
1434         vq->last_avail_idx -= n;
1435 }
1436 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1437
1438 /* After we've used one of their buffers, we tell them about it.  We'll then
1439  * want to notify the guest, using eventfd. */
1440 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1441 {
1442         struct vring_used_elem heads = {
1443                 cpu_to_vhost32(vq, head),
1444                 cpu_to_vhost32(vq, len)
1445         };
1446
1447         return vhost_add_used_n(vq, &heads, 1);
1448 }
1449 EXPORT_SYMBOL_GPL(vhost_add_used);
1450
1451 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1452                             struct vring_used_elem *heads,
1453                             unsigned count)
1454 {
1455         struct vring_used_elem __user *used;
1456         u16 old, new;
1457         int start;
1458
1459         start = vq->last_used_idx % vq->num;
1460         used = vq->used->ring + start;
1461         if (count == 1) {
1462                 if (__put_user(heads[0].id, &used->id)) {
1463                         vq_err(vq, "Failed to write used id");
1464                         return -EFAULT;
1465                 }
1466                 if (__put_user(heads[0].len, &used->len)) {
1467                         vq_err(vq, "Failed to write used len");
1468                         return -EFAULT;
1469                 }
1470         } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1471                 vq_err(vq, "Failed to write used");
1472                 return -EFAULT;
1473         }
1474         if (unlikely(vq->log_used)) {
1475                 /* Make sure data is seen before log. */
1476                 smp_wmb();
1477                 /* Log used ring entry write. */
1478                 log_write(vq->log_base,
1479                           vq->log_addr +
1480                            ((void __user *)used - (void __user *)vq->used),
1481                           count * sizeof *used);
1482         }
1483         old = vq->last_used_idx;
1484         new = (vq->last_used_idx += count);
1485         /* If the driver never bothers to signal in a very long while,
1486          * used index might wrap around. If that happens, invalidate
1487          * signalled_used index we stored. TODO: make sure driver
1488          * signals at least once in 2^16 and remove this. */
1489         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1490                 vq->signalled_used_valid = false;
1491         return 0;
1492 }
1493
1494 /* After we've used one of their buffers, we tell them about it.  We'll then
1495  * want to notify the guest, using eventfd. */
1496 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1497                      unsigned count)
1498 {
1499         int start, n, r;
1500
1501         start = vq->last_used_idx % vq->num;
1502         n = vq->num - start;
1503         if (n < count) {
1504                 r = __vhost_add_used_n(vq, heads, n);
1505                 if (r < 0)
1506                         return r;
1507                 heads += n;
1508                 count -= n;
1509         }
1510         r = __vhost_add_used_n(vq, heads, count);
1511
1512         /* Make sure buffer is written before we update index. */
1513         smp_wmb();
1514         if (__put_user(cpu_to_vhost16(vq, vq->last_used_idx), &vq->used->idx)) {
1515                 vq_err(vq, "Failed to increment used idx");
1516                 return -EFAULT;
1517         }
1518         if (unlikely(vq->log_used)) {
1519                 /* Log used index update. */
1520                 log_write(vq->log_base,
1521                           vq->log_addr + offsetof(struct vring_used, idx),
1522                           sizeof vq->used->idx);
1523                 if (vq->log_ctx)
1524                         eventfd_signal(vq->log_ctx, 1);
1525         }
1526         return r;
1527 }
1528 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1529
1530 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1531 {
1532         __u16 old, new;
1533         __virtio16 event;
1534         bool v;
1535         /* Flush out used index updates. This is paired
1536          * with the barrier that the Guest executes when enabling
1537          * interrupts. */
1538         smp_mb();
1539
1540         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1541             unlikely(vq->avail_idx == vq->last_avail_idx))
1542                 return true;
1543
1544         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1545                 __virtio16 flags;
1546                 if (__get_user(flags, &vq->avail->flags)) {
1547                         vq_err(vq, "Failed to get flags");
1548                         return true;
1549                 }
1550                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1551         }
1552         old = vq->signalled_used;
1553         v = vq->signalled_used_valid;
1554         new = vq->signalled_used = vq->last_used_idx;
1555         vq->signalled_used_valid = true;
1556
1557         if (unlikely(!v))
1558                 return true;
1559
1560         if (__get_user(event, vhost_used_event(vq))) {
1561                 vq_err(vq, "Failed to get used event idx");
1562                 return true;
1563         }
1564         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1565 }
1566
1567 /* This actually signals the guest, using eventfd. */
1568 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1569 {
1570         /* Signal the Guest tell them we used something up. */
1571         if (vq->call_ctx && vhost_notify(dev, vq))
1572                 eventfd_signal(vq->call_ctx, 1);
1573 }
1574 EXPORT_SYMBOL_GPL(vhost_signal);
1575
1576 /* And here's the combo meal deal.  Supersize me! */
1577 void vhost_add_used_and_signal(struct vhost_dev *dev,
1578                                struct vhost_virtqueue *vq,
1579                                unsigned int head, int len)
1580 {
1581         vhost_add_used(vq, head, len);
1582         vhost_signal(dev, vq);
1583 }
1584 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1585
1586 /* multi-buffer version of vhost_add_used_and_signal */
1587 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1588                                  struct vhost_virtqueue *vq,
1589                                  struct vring_used_elem *heads, unsigned count)
1590 {
1591         vhost_add_used_n(vq, heads, count);
1592         vhost_signal(dev, vq);
1593 }
1594 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1595
1596 /* OK, now we need to know about added descriptors. */
1597 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1598 {
1599         __virtio16 avail_idx;
1600         int r;
1601
1602         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1603                 return false;
1604         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1605         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1606                 r = vhost_update_used_flags(vq);
1607                 if (r) {
1608                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1609                                &vq->used->flags, r);
1610                         return false;
1611                 }
1612         } else {
1613                 r = vhost_update_avail_event(vq, vq->avail_idx);
1614                 if (r) {
1615                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1616                                vhost_avail_event(vq), r);
1617                         return false;
1618                 }
1619         }
1620         /* They could have slipped one in as we were doing that: make
1621          * sure it's written, then check again. */
1622         smp_mb();
1623         r = __get_user(avail_idx, &vq->avail->idx);
1624         if (r) {
1625                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1626                        &vq->avail->idx, r);
1627                 return false;
1628         }
1629
1630         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1631 }
1632 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1633
1634 /* We don't need to be notified again. */
1635 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1636 {
1637         int r;
1638
1639         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1640                 return;
1641         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1642         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1643                 r = vhost_update_used_flags(vq);
1644                 if (r)
1645                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1646                                &vq->used->flags, r);
1647         }
1648 }
1649 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1650
1651 static int __init vhost_init(void)
1652 {
1653         return 0;
1654 }
1655
1656 static void __exit vhost_exit(void)
1657 {
1658 }
1659
1660 module_init(vhost_init);
1661 module_exit(vhost_exit);
1662
1663 MODULE_VERSION("0.0.1");
1664 MODULE_LICENSE("GPL v2");
1665 MODULE_AUTHOR("Michael S. Tsirkin");
1666 MODULE_DESCRIPTION("Host kernel accelerator for virtio");