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 #include <linux/interval_tree_generic.h>
  31
  32 #include "vhost.h"
  33
  34 static ushort max_mem_regions = 64;
  35 module_param(max_mem_regions, ushort, 0444);
  36 MODULE_PARM_DESC(max_mem_regions,
  37         "Maximum number of memory regions in memory map. (default: 64)");
  38
  39 enum {
  40         VHOST_MEMORY_F_LOG = 0x1,
  41 };
  42
  43 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  44 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  45
  46 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  47                      rb, __u64, __subtree_last,
  48                      START, LAST, , vhost_umem_interval_tree);
  49
  50 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  51 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  52 {
  53         vq->user_be = !virtio_legacy_is_little_endian();
  54 }
  55
  56 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  57 {
  58         vq->user_be = true;
  59 }
  60
  61 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  62 {
  63         vq->user_be = false;
  64 }
  65
  66 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  67 {
  68         struct vhost_vring_state s;
  69
  70         if (vq->private_data)
  71                 return -EBUSY;
  72
  73         if (copy_from_user(&s, argp, sizeof(s)))
  74                 return -EFAULT;
  75
  76         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  77             s.num != VHOST_VRING_BIG_ENDIAN)
  78                 return -EINVAL;
  79
  80         if (s.num == VHOST_VRING_BIG_ENDIAN)
  81                 vhost_enable_cross_endian_big(vq);
  82         else
  83                 vhost_enable_cross_endian_little(vq);
  84
  85         return 0;
  86 }
  87
  88 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  89                                    int __user *argp)
  90 {
  91         struct vhost_vring_state s = {
  92                 .index = idx,
  93                 .num = vq->user_be
  94         };
  95
  96         if (copy_to_user(argp, &s, sizeof(s)))
  97                 return -EFAULT;
  98
  99         return 0;
 100 }
 101
 102 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 103 {
 104         /* Note for legacy virtio: user_be is initialized at reset time
 105          * according to the host endianness. If userspace does not set an
 106          * explicit endianness, the default behavior is native endian, as
 107          * expected by legacy virtio.
 108          */
 109         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 110 }
 111 #else
 112 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 113 {
 114 }
 115
 116 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 117 {
 118         return -ENOIOCTLCMD;
 119 }
 120
 121 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 122                                    int __user *argp)
 123 {
 124         return -ENOIOCTLCMD;
 125 }
 126
 127 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 128 {
 129         if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
 130                 vq->is_le = true;
 131 }
 132 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 133
 134 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 135 {
 136         vq->is_le = virtio_legacy_is_little_endian();
 137 }
 138
 139 struct vhost_flush_struct {
 140         struct vhost_work work;
 141         struct completion wait_event;
 142 };
 143
 144 static void vhost_flush_work(struct vhost_work *work)
 145 {
 146         struct vhost_flush_struct *s;
 147
 148         s = container_of(work, struct vhost_flush_struct, work);
 149         complete(&s->wait_event);
 150 }
 151
 152 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 153                             poll_table *pt)
 154 {
 155         struct vhost_poll *poll;
 156
 157         poll = container_of(pt, struct vhost_poll, table);
 158         poll->wqh = wqh;
 159         add_wait_queue(wqh, &poll->wait);
 160 }
 161
 162 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
 163                              void *key)
 164 {
 165         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 166
 167         if (!((unsigned long)key & poll->mask))
 168                 return 0;
 169
 170         vhost_poll_queue(poll);
 171         return 0;
 172 }
 173
 174 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 175 {
 176         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 177         work->fn = fn;
 178         init_waitqueue_head(&work->done);
 179 }
 180 EXPORT_SYMBOL_GPL(vhost_work_init);
 181
 182 /* Init poll structure */
 183 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 184                      unsigned long mask, struct vhost_dev *dev)
 185 {
 186         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 187         init_poll_funcptr(&poll->table, vhost_poll_func);
 188         poll->mask = mask;
 189         poll->dev = dev;
 190         poll->wqh = NULL;
 191
 192         vhost_work_init(&poll->work, fn);
 193 }
 194 EXPORT_SYMBOL_GPL(vhost_poll_init);
 195
 196 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 197  * keep a reference to a file until after vhost_poll_stop is called. */
 198 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 199 {
 200         unsigned long mask;
 201         int ret = 0;
 202
 203         if (poll->wqh)
 204                 return 0;
 205
 206         mask = file->f_op->poll(file, &poll->table);
 207         if (mask)
 208                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 209         if (mask & POLLERR) {
 210                 if (poll->wqh)
 211                         remove_wait_queue(poll->wqh, &poll->wait);
 212                 ret = -EINVAL;
 213         }
 214
 215         return ret;
 216 }
 217 EXPORT_SYMBOL_GPL(vhost_poll_start);
 218
 219 /* Stop polling a file. After this function returns, it becomes safe to drop the
 220  * file reference. You must also flush afterwards. */
 221 void vhost_poll_stop(struct vhost_poll *poll)
 222 {
 223         if (poll->wqh) {
 224                 remove_wait_queue(poll->wqh, &poll->wait);
 225                 poll->wqh = NULL;
 226         }
 227 }
 228 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 229
 230 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 231 {
 232         struct vhost_flush_struct flush;
 233
 234         if (dev->worker) {
 235                 init_completion(&flush.wait_event);
 236                 vhost_work_init(&flush.work, vhost_flush_work);
 237
 238                 vhost_work_queue(dev, &flush.work);
 239                 wait_for_completion(&flush.wait_event);
 240         }
 241 }
 242 EXPORT_SYMBOL_GPL(vhost_work_flush);
 243
 244 /* Flush any work that has been scheduled. When calling this, don't hold any
 245  * locks that are also used by the callback. */
 246 void vhost_poll_flush(struct vhost_poll *poll)
 247 {
 248         vhost_work_flush(poll->dev, &poll->work);
 249 }
 250 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 251
 252 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 253 {
 254         if (!dev->worker)
 255                 return;
 256
 257         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 258                 /* We can only add the work to the list after we're
 259                  * sure it was not in the list.
 260                  */
 261                 smp_mb();
 262                 llist_add(&work->node, &dev->work_list);
 263                 wake_up_process(dev->worker);
 264         }
 265 }
 266 EXPORT_SYMBOL_GPL(vhost_work_queue);
 267
 268 /* A lockless hint for busy polling code to exit the loop */
 269 bool vhost_has_work(struct vhost_dev *dev)
 270 {
 271         return !llist_empty(&dev->work_list);
 272 }
 273 EXPORT_SYMBOL_GPL(vhost_has_work);
 274
 275 void vhost_poll_queue(struct vhost_poll *poll)
 276 {
 277         vhost_work_queue(poll->dev, &poll->work);
 278 }
 279 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 280
 281 static void vhost_vq_reset(struct vhost_dev *dev,
 282                            struct vhost_virtqueue *vq)
 283 {
 284         vq->num = 1;
 285         vq->desc = NULL;
 286         vq->avail = NULL;
 287         vq->used = NULL;
 288         vq->last_avail_idx = 0;
 289         vq->avail_idx = 0;
 290         vq->last_used_idx = 0;
 291         vq->signalled_used = 0;
 292         vq->signalled_used_valid = false;
 293         vq->used_flags = 0;
 294         vq->log_used = false;
 295         vq->log_addr = -1ull;
 296         vq->private_data = NULL;
 297         vq->acked_features = 0;
 298         vq->log_base = NULL;
 299         vq->error_ctx = NULL;
 300         vq->error = NULL;
 301         vq->kick = NULL;
 302         vq->call_ctx = NULL;
 303         vq->call = NULL;
 304         vq->log_ctx = NULL;
 305         vhost_reset_is_le(vq);
 306         vhost_disable_cross_endian(vq);
 307         vq->busyloop_timeout = 0;
 308         vq->umem = NULL;
 309 }
 310
 311 static int vhost_worker(void *data)
 312 {
 313         struct vhost_dev *dev = data;
 314         struct vhost_work *work, *work_next;
 315         struct llist_node *node;
 316         mm_segment_t oldfs = get_fs();
 317
 318         set_fs(USER_DS);
 319         use_mm(dev->mm);
 320
 321         for (;;) {
 322                 /* mb paired w/ kthread_stop */
 323                 set_current_state(TASK_INTERRUPTIBLE);
 324
 325                 if (kthread_should_stop()) {
 326                         __set_current_state(TASK_RUNNING);
 327                         break;
 328                 }
 329
 330                 node = llist_del_all(&dev->work_list);
 331                 if (!node)
 332                         schedule();
 333
 334                 node = llist_reverse_order(node);
 335                 /* make sure flag is seen after deletion */
 336                 smp_wmb();
 337                 llist_for_each_entry_safe(work, work_next, node, node) {
 338                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 339                         __set_current_state(TASK_RUNNING);
 340                         work->fn(work);
 341                         if (need_resched())
 342                                 schedule();
 343                 }
 344         }
 345         unuse_mm(dev->mm);
 346         set_fs(oldfs);
 347         return 0;
 348 }
 349
 350 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 351 {
 352         kfree(vq->indirect);
 353         vq->indirect = NULL;
 354         kfree(vq->log);
 355         vq->log = NULL;
 356         kfree(vq->heads);
 357         vq->heads = NULL;
 358 }
 359
 360 /* Helper to allocate iovec buffers for all vqs. */
 361 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 362 {
 363         struct vhost_virtqueue *vq;
 364         int i;
 365
 366         for (i = 0; i < dev->nvqs; ++i) {
 367                 vq = dev->vqs[i];
 368                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 369                                        GFP_KERNEL);
 370                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 371                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 372                 if (!vq->indirect || !vq->log || !vq->heads)
 373                         goto err_nomem;
 374         }
 375         return 0;
 376
 377 err_nomem:
 378         for (; i >= 0; --i)
 379                 vhost_vq_free_iovecs(dev->vqs[i]);
 380         return -ENOMEM;
 381 }
 382
 383 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 384 {
 385         int i;
 386
 387         for (i = 0; i < dev->nvqs; ++i)
 388                 vhost_vq_free_iovecs(dev->vqs[i]);
 389 }
 390
 391 void vhost_dev_init(struct vhost_dev *dev,
 392                     struct vhost_virtqueue **vqs, int nvqs)
 393 {
 394         struct vhost_virtqueue *vq;
 395         int i;
 396
 397         dev->vqs = vqs;
 398         dev->nvqs = nvqs;
 399         mutex_init(&dev->mutex);
 400         dev->log_ctx = NULL;
 401         dev->log_file = NULL;
 402         dev->umem = NULL;
 403         dev->mm = NULL;
 404         dev->worker = NULL;
 405         init_llist_head(&dev->work_list);
 406
 407
 408         for (i = 0; i < dev->nvqs; ++i) {
 409                 vq = dev->vqs[i];
 410                 vq->log = NULL;
 411                 vq->indirect = NULL;
 412                 vq->heads = NULL;
 413                 vq->dev = dev;
 414                 mutex_init(&vq->mutex);
 415                 vhost_vq_reset(dev, vq);
 416                 if (vq->handle_kick)
 417                         vhost_poll_init(&vq->poll, vq->handle_kick,
 418                                         POLLIN, dev);
 419         }
 420 }
 421 EXPORT_SYMBOL_GPL(vhost_dev_init);
 422
 423 /* Caller should have device mutex */
 424 long vhost_dev_check_owner(struct vhost_dev *dev)
 425 {
 426         /* Are you the owner? If not, I don't think you mean to do that */
 427         return dev->mm == current->mm ? 0 : -EPERM;
 428 }
 429 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 430
 431 struct vhost_attach_cgroups_struct {
 432         struct vhost_work work;
 433         struct task_struct *owner;
 434         int ret;
 435 };
 436
 437 static void vhost_attach_cgroups_work(struct vhost_work *work)
 438 {
 439         struct vhost_attach_cgroups_struct *s;
 440
 441         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 442         s->ret = cgroup_attach_task_all(s->owner, current);
 443 }
 444
 445 static int vhost_attach_cgroups(struct vhost_dev *dev)
 446 {
 447         struct vhost_attach_cgroups_struct attach;
 448
 449         attach.owner = current;
 450         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 451         vhost_work_queue(dev, &attach.work);
 452         vhost_work_flush(dev, &attach.work);
 453         return attach.ret;
 454 }
 455
 456 /* Caller should have device mutex */
 457 bool vhost_dev_has_owner(struct vhost_dev *dev)
 458 {
 459         return dev->mm;
 460 }
 461 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 462
 463 /* Caller should have device mutex */
 464 long vhost_dev_set_owner(struct vhost_dev *dev)
 465 {
 466         struct task_struct *worker;
 467         int err;
 468
 469         /* Is there an owner already? */
 470         if (vhost_dev_has_owner(dev)) {
 471                 err = -EBUSY;
 472                 goto err_mm;
 473         }
 474
 475         /* No owner, become one */
 476         dev->mm = get_task_mm(current);
 477         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 478         if (IS_ERR(worker)) {
 479                 err = PTR_ERR(worker);
 480                 goto err_worker;
 481         }
 482
 483         dev->worker = worker;
 484         wake_up_process(worker);        /* avoid contributing to loadavg */
 485
 486         err = vhost_attach_cgroups(dev);
 487         if (err)
 488                 goto err_cgroup;
 489
 490         err = vhost_dev_alloc_iovecs(dev);
 491         if (err)
 492                 goto err_cgroup;
 493
 494         return 0;
 495 err_cgroup:
 496         kthread_stop(worker);
 497         dev->worker = NULL;
 498 err_worker:
 499         if (dev->mm)
 500                 mmput(dev->mm);
 501         dev->mm = NULL;
 502 err_mm:
 503         return err;
 504 }
 505 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 506
 507 static void *vhost_kvzalloc(unsigned long size)
 508 {
 509         void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 510
 511         if (!n)
 512                 n = vzalloc(size);
 513         return n;
 514 }
 515
 516 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 517 {
 518         return vhost_kvzalloc(sizeof(struct vhost_umem));
 519 }
 520 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 521
 522 /* Caller should have device mutex */
 523 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 524 {
 525         int i;
 526
 527         vhost_dev_cleanup(dev, true);
 528
 529         /* Restore memory to default empty mapping. */
 530         INIT_LIST_HEAD(&umem->umem_list);
 531         dev->umem = umem;
 532         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 533          * VQs aren't running.
 534          */
 535         for (i = 0; i < dev->nvqs; ++i)
 536                 dev->vqs[i]->umem = umem;
 537 }
 538 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 539
 540 void vhost_dev_stop(struct vhost_dev *dev)
 541 {
 542         int i;
 543
 544         for (i = 0; i < dev->nvqs; ++i) {
 545                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 546                         vhost_poll_stop(&dev->vqs[i]->poll);
 547                         vhost_poll_flush(&dev->vqs[i]->poll);
 548                 }
 549         }
 550 }
 551 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 552
 553 static void vhost_umem_clean(struct vhost_umem *umem)
 554 {
 555         struct vhost_umem_node *node, *tmp;
 556
 557         if (!umem)
 558                 return;
 559
 560         list_for_each_entry_safe(node, tmp, &umem->umem_list, link) {
 561                 vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 562                 list_del(&node->link);
 563                 kvfree(node);
 564         }
 565         kvfree(umem);
 566 }
 567
 568 /* Caller should have device mutex if and only if locked is set */
 569 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 570 {
 571         int i;
 572
 573         for (i = 0; i < dev->nvqs; ++i) {
 574                 if (dev->vqs[i]->error_ctx)
 575                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 576                 if (dev->vqs[i]->error)
 577                         fput(dev->vqs[i]->error);
 578                 if (dev->vqs[i]->kick)
 579                         fput(dev->vqs[i]->kick);
 580                 if (dev->vqs[i]->call_ctx)
 581                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 582                 if (dev->vqs[i]->call)
 583                         fput(dev->vqs[i]->call);
 584                 vhost_vq_reset(dev, dev->vqs[i]);
 585         }
 586         vhost_dev_free_iovecs(dev);
 587         if (dev->log_ctx)
 588                 eventfd_ctx_put(dev->log_ctx);
 589         dev->log_ctx = NULL;
 590         if (dev->log_file)
 591                 fput(dev->log_file);
 592         dev->log_file = NULL;
 593         /* No one will access memory at this point */
 594         vhost_umem_clean(dev->umem);
 595         dev->umem = NULL;
 596         WARN_ON(!llist_empty(&dev->work_list));
 597         if (dev->worker) {
 598                 kthread_stop(dev->worker);
 599                 dev->worker = NULL;
 600         }
 601         if (dev->mm)
 602                 mmput(dev->mm);
 603         dev->mm = NULL;
 604 }
 605 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 606
 607 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 608 {
 609         u64 a = addr / VHOST_PAGE_SIZE / 8;
 610
 611         /* Make sure 64 bit math will not overflow. */
 612         if (a > ULONG_MAX - (unsigned long)log_base ||
 613             a + (unsigned long)log_base > ULONG_MAX)
 614                 return 0;
 615
 616         return access_ok(VERIFY_WRITE, log_base + a,
 617                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 618 }
 619
 620 /* Caller should have vq mutex and device mutex. */
 621 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 622                                int log_all)
 623 {
 624         struct vhost_umem_node *node;
 625
 626         if (!umem)
 627                 return 0;
 628
 629         list_for_each_entry(node, &umem->umem_list, link) {
 630                 unsigned long a = node->userspace_addr;
 631
 632                 if (node->size > ULONG_MAX)
 633                         return 0;
 634                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 635                                     node->size))
 636                         return 0;
 637                 else if (log_all && !log_access_ok(log_base,
 638                                                    node->start,
 639                                                    node->size))
 640                         return 0;
 641         }
 642         return 1;
 643 }
 644
 645 /* Can we switch to this memory table? */
 646 /* Caller should have device mutex but not vq mutex */
 647 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 648                             int log_all)
 649 {
 650         int i;
 651
 652         for (i = 0; i < d->nvqs; ++i) {
 653                 int ok;
 654                 bool log;
 655
 656                 mutex_lock(&d->vqs[i]->mutex);
 657                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 658                 /* If ring is inactive, will check when it's enabled. */
 659                 if (d->vqs[i]->private_data)
 660                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 661                                                  umem, log);
 662                 else
 663                         ok = 1;
 664                 mutex_unlock(&d->vqs[i]->mutex);
 665                 if (!ok)
 666                         return 0;
 667         }
 668         return 1;
 669 }
 670
 671 #define vhost_put_user(vq, x, ptr)  __put_user(x, ptr)
 672
 673 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void *to,
 674                               const void *from, unsigned size)
 675 {
 676         return __copy_to_user(to, from, size);
 677 }
 678
 679 #define vhost_get_user(vq, x, ptr) __get_user(x, ptr)
 680
 681 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 682                                 void *from, unsigned size)
 683 {
 684         return __copy_from_user(to, from, size);
 685 }
 686
 687 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
 688                         struct vring_desc __user *desc,
 689                         struct vring_avail __user *avail,
 690                         struct vring_used __user *used)
 691 {
 692         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 693         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 694                access_ok(VERIFY_READ, avail,
 695                          sizeof *avail + num * sizeof *avail->ring + s) &&
 696                access_ok(VERIFY_WRITE, used,
 697                         sizeof *used + num * sizeof *used->ring + s);
 698 }
 699
 700 /* Can we log writes? */
 701 /* Caller should have device mutex but not vq mutex */
 702 int vhost_log_access_ok(struct vhost_dev *dev)
 703 {
 704         return memory_access_ok(dev, dev->umem, 1);
 705 }
 706 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
 707
 708 /* Verify access for write logging. */
 709 /* Caller should have vq mutex and device mutex */
 710 static int vq_log_access_ok(struct vhost_virtqueue *vq,
 711                             void __user *log_base)
 712 {
 713         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 714
 715         return vq_memory_access_ok(log_base, vq->umem,
 716                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
 717                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 718                                         sizeof *vq->used +
 719                                         vq->num * sizeof *vq->used->ring + s));
 720 }
 721
 722 /* Can we start vq? */
 723 /* Caller should have vq mutex and device mutex */
 724 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 725 {
 726         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
 727                 vq_log_access_ok(vq, vq->log_base);
 728 }
 729 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
 730
 731 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 732 {
 733         struct vhost_memory mem, *newmem;
 734         struct vhost_memory_region *region;
 735         struct vhost_umem_node *node;
 736         struct vhost_umem *newumem, *oldumem;
 737         unsigned long size = offsetof(struct vhost_memory, regions);
 738         int i;
 739
 740         if (copy_from_user(&mem, m, size))
 741                 return -EFAULT;
 742         if (mem.padding)
 743                 return -EOPNOTSUPP;
 744         if (mem.nregions > max_mem_regions)
 745                 return -E2BIG;
 746         newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
 747         if (!newmem)
 748                 return -ENOMEM;
 749
 750         memcpy(newmem, &mem, size);
 751         if (copy_from_user(newmem->regions, m->regions,
 752                            mem.nregions * sizeof *m->regions)) {
 753                 kvfree(newmem);
 754                 return -EFAULT;
 755         }
 756
 757         newumem = vhost_kvzalloc(sizeof(*newumem));
 758         if (!newumem) {
 759                 kvfree(newmem);
 760                 return -ENOMEM;
 761         }
 762
 763         newumem->umem_tree = RB_ROOT;
 764         INIT_LIST_HEAD(&newumem->umem_list);
 765
 766         for (region = newmem->regions;
 767              region < newmem->regions + mem.nregions;
 768              region++) {
 769                 node = vhost_kvzalloc(sizeof(*node));
 770                 if (!node)
 771                         goto err;
 772                 node->start = region->guest_phys_addr;
 773                 node->size = region->memory_size;
 774                 node->last = node->start + node->size - 1;
 775                 node->userspace_addr = region->userspace_addr;
 776                 INIT_LIST_HEAD(&node->link);
 777                 list_add_tail(&node->link, &newumem->umem_list);
 778                 vhost_umem_interval_tree_insert(node, &newumem->umem_tree);
 779         }
 780
 781         if (!memory_access_ok(d, newumem, 0))
 782                 goto err;
 783
 784         oldumem = d->umem;
 785         d->umem = newumem;
 786
 787         /* All memory accesses are done under some VQ mutex. */
 788         for (i = 0; i < d->nvqs; ++i) {
 789                 mutex_lock(&d->vqs[i]->mutex);
 790                 d->vqs[i]->umem = newumem;
 791                 mutex_unlock(&d->vqs[i]->mutex);
 792         }
 793
 794         kvfree(newmem);
 795         vhost_umem_clean(oldumem);
 796         return 0;
 797
 798 err:
 799         vhost_umem_clean(newumem);
 800         kvfree(newmem);
 801         return -EFAULT;
 802 }
 803
 804 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 805 {
 806         struct file *eventfp, *filep = NULL;
 807         bool pollstart = false, pollstop = false;
 808         struct eventfd_ctx *ctx = NULL;
 809         u32 __user *idxp = argp;
 810         struct vhost_virtqueue *vq;
 811         struct vhost_vring_state s;
 812         struct vhost_vring_file f;
 813         struct vhost_vring_addr a;
 814         u32 idx;
 815         long r;
 816
 817         r = get_user(idx, idxp);
 818         if (r < 0)
 819                 return r;
 820         if (idx >= d->nvqs)
 821                 return -ENOBUFS;
 822
 823         vq = d->vqs[idx];
 824
 825         mutex_lock(&vq->mutex);
 826
 827         switch (ioctl) {
 828         case VHOST_SET_VRING_NUM:
 829                 /* Resizing ring with an active backend?
 830                  * You don't want to do that. */
 831                 if (vq->private_data) {
 832                         r = -EBUSY;
 833                         break;
 834                 }
 835                 if (copy_from_user(&s, argp, sizeof s)) {
 836                         r = -EFAULT;
 837                         break;
 838                 }
 839                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 840                         r = -EINVAL;
 841                         break;
 842                 }
 843                 vq->num = s.num;
 844                 break;
 845         case VHOST_SET_VRING_BASE:
 846                 /* Moving base with an active backend?
 847                  * You don't want to do that. */
 848                 if (vq->private_data) {
 849                         r = -EBUSY;
 850                         break;
 851                 }
 852                 if (copy_from_user(&s, argp, sizeof s)) {
 853                         r = -EFAULT;
 854                         break;
 855                 }
 856                 if (s.num > 0xffff) {
 857                         r = -EINVAL;
 858                         break;
 859                 }
 860                 vq->last_avail_idx = s.num;
 861                 /* Forget the cached index value. */
 862                 vq->avail_idx = vq->last_avail_idx;
 863                 break;
 864         case VHOST_GET_VRING_BASE:
 865                 s.index = idx;
 866                 s.num = vq->last_avail_idx;
 867                 if (copy_to_user(argp, &s, sizeof s))
 868                         r = -EFAULT;
 869                 break;
 870         case VHOST_SET_VRING_ADDR:
 871                 if (copy_from_user(&a, argp, sizeof a)) {
 872                         r = -EFAULT;
 873                         break;
 874                 }
 875                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 876                         r = -EOPNOTSUPP;
 877                         break;
 878                 }
 879                 /* For 32bit, verify that the top 32bits of the user
 880                    data are set to zero. */
 881                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 882                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 883                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 884                         r = -EFAULT;
 885                         break;
 886                 }
 887
 888                 /* Make sure it's safe to cast pointers to vring types. */
 889                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
 890                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
 891                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
 892                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
 893                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
 894                         r = -EINVAL;
 895                         break;
 896                 }
 897
 898                 /* We only verify access here if backend is configured.
 899                  * If it is not, we don't as size might not have been setup.
 900                  * We will verify when backend is configured. */
 901                 if (vq->private_data) {
 902                         if (!vq_access_ok(vq, vq->num,
 903                                 (void __user *)(unsigned long)a.desc_user_addr,
 904                                 (void __user *)(unsigned long)a.avail_user_addr,
 905                                 (void __user *)(unsigned long)a.used_user_addr)) {
 906                                 r = -EINVAL;
 907                                 break;
 908                         }
 909
 910                         /* Also validate log access for used ring if enabled. */
 911                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 912                             !log_access_ok(vq->log_base, a.log_guest_addr,
 913                                            sizeof *vq->used +
 914                                            vq->num * sizeof *vq->used->ring)) {
 915                                 r = -EINVAL;
 916                                 break;
 917                         }
 918                 }
 919
 920                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 921                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 922                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 923                 vq->log_addr = a.log_guest_addr;
 924                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 925                 break;
 926         case VHOST_SET_VRING_KICK:
 927                 if (copy_from_user(&f, argp, sizeof f)) {
 928                         r = -EFAULT;
 929                         break;
 930                 }
 931                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 932                 if (IS_ERR(eventfp)) {
 933                         r = PTR_ERR(eventfp);
 934                         break;
 935                 }
 936                 if (eventfp != vq->kick) {
 937                         pollstop = (filep = vq->kick) != NULL;
 938                         pollstart = (vq->kick = eventfp) != NULL;
 939                 } else
 940                         filep = eventfp;
 941                 break;
 942         case VHOST_SET_VRING_CALL:
 943                 if (copy_from_user(&f, argp, sizeof f)) {
 944                         r = -EFAULT;
 945                         break;
 946                 }
 947                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 948                 if (IS_ERR(eventfp)) {
 949                         r = PTR_ERR(eventfp);
 950                         break;
 951                 }
 952                 if (eventfp != vq->call) {
 953                         filep = vq->call;
 954                         ctx = vq->call_ctx;
 955                         vq->call = eventfp;
 956                         vq->call_ctx = eventfp ?
 957                                 eventfd_ctx_fileget(eventfp) : NULL;
 958                 } else
 959                         filep = eventfp;
 960                 break;
 961         case VHOST_SET_VRING_ERR:
 962                 if (copy_from_user(&f, argp, sizeof f)) {
 963                         r = -EFAULT;
 964                         break;
 965                 }
 966                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 967                 if (IS_ERR(eventfp)) {
 968                         r = PTR_ERR(eventfp);
 969                         break;
 970                 }
 971                 if (eventfp != vq->error) {
 972                         filep = vq->error;
 973                         vq->error = eventfp;
 974                         ctx = vq->error_ctx;
 975                         vq->error_ctx = eventfp ?
 976                                 eventfd_ctx_fileget(eventfp) : NULL;
 977                 } else
 978                         filep = eventfp;
 979                 break;
 980         case VHOST_SET_VRING_ENDIAN:
 981                 r = vhost_set_vring_endian(vq, argp);
 982                 break;
 983         case VHOST_GET_VRING_ENDIAN:
 984                 r = vhost_get_vring_endian(vq, idx, argp);
 985                 break;
 986         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
 987                 if (copy_from_user(&s, argp, sizeof(s))) {
 988                         r = -EFAULT;
 989                         break;
 990                 }
 991                 vq->busyloop_timeout = s.num;
 992                 break;
 993         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
 994                 s.index = idx;
 995                 s.num = vq->busyloop_timeout;
 996                 if (copy_to_user(argp, &s, sizeof(s)))
 997                         r = -EFAULT;
 998                 break;
 999         default:
1000                 r = -ENOIOCTLCMD;
1001         }
1002
1003         if (pollstop && vq->handle_kick)
1004                 vhost_poll_stop(&vq->poll);
1005
1006         if (ctx)
1007                 eventfd_ctx_put(ctx);
1008         if (filep)
1009                 fput(filep);
1010
1011         if (pollstart && vq->handle_kick)
1012                 r = vhost_poll_start(&vq->poll, vq->kick);
1013
1014         mutex_unlock(&vq->mutex);
1015
1016         if (pollstop && vq->handle_kick)
1017                 vhost_poll_flush(&vq->poll);
1018         return r;
1019 }
1020 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1021
1022 /* Caller must have device mutex */
1023 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1024 {
1025         struct file *eventfp, *filep = NULL;
1026         struct eventfd_ctx *ctx = NULL;
1027         u64 p;
1028         long r;
1029         int i, fd;
1030
1031         /* If you are not the owner, you can become one */
1032         if (ioctl == VHOST_SET_OWNER) {
1033                 r = vhost_dev_set_owner(d);
1034                 goto done;
1035         }
1036
1037         /* You must be the owner to do anything else */
1038         r = vhost_dev_check_owner(d);
1039         if (r)
1040                 goto done;
1041
1042         switch (ioctl) {
1043         case VHOST_SET_MEM_TABLE:
1044                 r = vhost_set_memory(d, argp);
1045                 break;
1046         case VHOST_SET_LOG_BASE:
1047                 if (copy_from_user(&p, argp, sizeof p)) {
1048                         r = -EFAULT;
1049                         break;
1050                 }
1051                 if ((u64)(unsigned long)p != p) {
1052                         r = -EFAULT;
1053                         break;
1054                 }
1055                 for (i = 0; i < d->nvqs; ++i) {
1056                         struct vhost_virtqueue *vq;
1057                         void __user *base = (void __user *)(unsigned long)p;
1058                         vq = d->vqs[i];
1059                         mutex_lock(&vq->mutex);
1060                         /* If ring is inactive, will check when it's enabled. */
1061                         if (vq->private_data && !vq_log_access_ok(vq, base))
1062                                 r = -EFAULT;
1063                         else
1064                                 vq->log_base = base;
1065                         mutex_unlock(&vq->mutex);
1066                 }
1067                 break;
1068         case VHOST_SET_LOG_FD:
1069                 r = get_user(fd, (int __user *)argp);
1070                 if (r < 0)
1071                         break;
1072                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1073                 if (IS_ERR(eventfp)) {
1074                         r = PTR_ERR(eventfp);
1075                         break;
1076                 }
1077                 if (eventfp != d->log_file) {
1078                         filep = d->log_file;
1079                         d->log_file = eventfp;
1080                         ctx = d->log_ctx;
1081                         d->log_ctx = eventfp ?
1082                                 eventfd_ctx_fileget(eventfp) : NULL;
1083                 } else
1084                         filep = eventfp;
1085                 for (i = 0; i < d->nvqs; ++i) {
1086                         mutex_lock(&d->vqs[i]->mutex);
1087                         d->vqs[i]->log_ctx = d->log_ctx;
1088                         mutex_unlock(&d->vqs[i]->mutex);
1089                 }
1090                 if (ctx)
1091                         eventfd_ctx_put(ctx);
1092                 if (filep)
1093                         fput(filep);
1094                 break;
1095         default:
1096                 r = -ENOIOCTLCMD;
1097                 break;
1098         }
1099 done:
1100         return r;
1101 }
1102 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1103
1104 /* TODO: This is really inefficient.  We need something like get_user()
1105  * (instruction directly accesses the data, with an exception table entry
1106  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1107  */
1108 static int set_bit_to_user(int nr, void __user *addr)
1109 {
1110         unsigned long log = (unsigned long)addr;
1111         struct page *page;
1112         void *base;
1113         int bit = nr + (log % PAGE_SIZE) * 8;
1114         int r;
1115
1116         r = get_user_pages_fast(log, 1, 1, &page);
1117         if (r < 0)
1118                 return r;
1119         BUG_ON(r != 1);
1120         base = kmap_atomic(page);
1121         set_bit(bit, base);
1122         kunmap_atomic(base);
1123         set_page_dirty_lock(page);
1124         put_page(page);
1125         return 0;
1126 }
1127
1128 static int log_write(void __user *log_base,
1129                      u64 write_address, u64 write_length)
1130 {
1131         u64 write_page = write_address / VHOST_PAGE_SIZE;
1132         int r;
1133
1134         if (!write_length)
1135                 return 0;
1136         write_length += write_address % VHOST_PAGE_SIZE;
1137         for (;;) {
1138                 u64 base = (u64)(unsigned long)log_base;
1139                 u64 log = base + write_page / 8;
1140                 int bit = write_page % 8;
1141                 if ((u64)(unsigned long)log != log)
1142                         return -EFAULT;
1143                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1144                 if (r < 0)
1145                         return r;
1146                 if (write_length <= VHOST_PAGE_SIZE)
1147                         break;
1148                 write_length -= VHOST_PAGE_SIZE;
1149                 write_page += 1;
1150         }
1151         return r;
1152 }
1153
1154 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1155                     unsigned int log_num, u64 len)
1156 {
1157         int i, r;
1158
1159         /* Make sure data written is seen before log. */
1160         smp_wmb();
1161         for (i = 0; i < log_num; ++i) {
1162                 u64 l = min(log[i].len, len);
1163                 r = log_write(vq->log_base, log[i].addr, l);
1164                 if (r < 0)
1165                         return r;
1166                 len -= l;
1167                 if (!len) {
1168                         if (vq->log_ctx)
1169                                 eventfd_signal(vq->log_ctx, 1);
1170                         return 0;
1171                 }
1172         }
1173         /* Length written exceeds what we have stored. This is a bug. */
1174         BUG();
1175         return 0;
1176 }
1177 EXPORT_SYMBOL_GPL(vhost_log_write);
1178
1179 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1180 {
1181         void __user *used;
1182         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1183                            &vq->used->flags) < 0)
1184                 return -EFAULT;
1185         if (unlikely(vq->log_used)) {
1186                 /* Make sure the flag is seen before log. */
1187                 smp_wmb();
1188                 /* Log used flag write. */
1189                 used = &vq->used->flags;
1190                 log_write(vq->log_base, vq->log_addr +
1191                           (used - (void __user *)vq->used),
1192                           sizeof vq->used->flags);
1193                 if (vq->log_ctx)
1194                         eventfd_signal(vq->log_ctx, 1);
1195         }
1196         return 0;
1197 }
1198
1199 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1200 {
1201         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1202                            vhost_avail_event(vq)))
1203                 return -EFAULT;
1204         if (unlikely(vq->log_used)) {
1205                 void __user *used;
1206                 /* Make sure the event is seen before log. */
1207                 smp_wmb();
1208                 /* Log avail event write */
1209                 used = vhost_avail_event(vq);
1210                 log_write(vq->log_base, vq->log_addr +
1211                           (used - (void __user *)vq->used),
1212                           sizeof *vhost_avail_event(vq));
1213                 if (vq->log_ctx)
1214                         eventfd_signal(vq->log_ctx, 1);
1215         }
1216         return 0;
1217 }
1218
1219 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1220 {
1221         __virtio16 last_used_idx;
1222         int r;
1223         bool is_le = vq->is_le;
1224
1225         if (!vq->private_data) {
1226                 vhost_reset_is_le(vq);
1227                 return 0;
1228         }
1229
1230         vhost_init_is_le(vq);
1231
1232         r = vhost_update_used_flags(vq);
1233         if (r)
1234                 goto err;
1235         vq->signalled_used_valid = false;
1236         if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1237                 r = -EFAULT;
1238                 goto err;
1239         }
1240         r = vhost_get_user(vq, last_used_idx, &vq->used->idx);
1241         if (r)
1242                 goto err;
1243         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1244         return 0;
1245 err:
1246         vq->is_le = is_le;
1247         return r;
1248 }
1249 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1250
1251 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1252                           struct iovec iov[], int iov_size)
1253 {
1254         const struct vhost_umem_node *node;
1255         struct vhost_umem *umem = vq->umem;
1256         struct iovec *_iov;
1257         u64 s = 0;
1258         int ret = 0;
1259
1260         while ((u64)len > s) {
1261                 u64 size;
1262                 if (unlikely(ret >= iov_size)) {
1263                         ret = -ENOBUFS;
1264                         break;
1265                 }
1266                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1267                                                         addr, addr + len - 1);
1268                 if (node == NULL || node->start > addr) {
1269                         ret = -EFAULT;
1270                         break;
1271                 }
1272                 _iov = iov + ret;
1273                 size = node->size - addr + node->start;
1274                 _iov->iov_len = min((u64)len - s, size);
1275                 _iov->iov_base = (void __user *)(unsigned long)
1276                         (node->userspace_addr + addr - node->start);
1277                 s += size;
1278                 addr += size;
1279                 ++ret;
1280         }
1281
1282         return ret;
1283 }
1284
1285 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1286  * function returns the next descriptor in the chain,
1287  * or -1U if we're at the end. */
1288 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1289 {
1290         unsigned int next;
1291
1292         /* If this descriptor says it doesn't chain, we're done. */
1293         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1294                 return -1U;
1295
1296         /* Check they're not leading us off end of descriptors. */
1297         next = vhost16_to_cpu(vq, desc->next);
1298         /* Make sure compiler knows to grab that: we don't want it changing! */
1299         /* We will use the result as an index in an array, so most
1300          * architectures only need a compiler barrier here. */
1301         read_barrier_depends();
1302
1303         return next;
1304 }
1305
1306 static int get_indirect(struct vhost_virtqueue *vq,
1307                         struct iovec iov[], unsigned int iov_size,
1308                         unsigned int *out_num, unsigned int *in_num,
1309                         struct vhost_log *log, unsigned int *log_num,
1310                         struct vring_desc *indirect)
1311 {
1312         struct vring_desc desc;
1313         unsigned int i = 0, count, found = 0;
1314         u32 len = vhost32_to_cpu(vq, indirect->len);
1315         struct iov_iter from;
1316         int ret;
1317
1318         /* Sanity check */
1319         if (unlikely(len % sizeof desc)) {
1320                 vq_err(vq, "Invalid length in indirect descriptor: "
1321                        "len 0x%llx not multiple of 0x%zx\n",
1322                        (unsigned long long)len,
1323                        sizeof desc);
1324                 return -EINVAL;
1325         }
1326
1327         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1328                              UIO_MAXIOV);
1329         if (unlikely(ret < 0)) {
1330                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1331                 return ret;
1332         }
1333         iov_iter_init(&from, READ, vq->indirect, ret, len);
1334
1335         /* We will use the result as an address to read from, so most
1336          * architectures only need a compiler barrier here. */
1337         read_barrier_depends();
1338
1339         count = len / sizeof desc;
1340         /* Buffers are chained via a 16 bit next field, so
1341          * we can have at most 2^16 of these. */
1342         if (unlikely(count > USHRT_MAX + 1)) {
1343                 vq_err(vq, "Indirect buffer length too big: %d\n",
1344                        indirect->len);
1345                 return -E2BIG;
1346         }
1347
1348         do {
1349                 unsigned iov_count = *in_num + *out_num;
1350                 if (unlikely(++found > count)) {
1351                         vq_err(vq, "Loop detected: last one at %u "
1352                                "indirect size %u\n",
1353                                i, count);
1354                         return -EINVAL;
1355                 }
1356                 if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1357                              sizeof(desc))) {
1358                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1359                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1360                         return -EINVAL;
1361                 }
1362                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1363                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1364                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1365                         return -EINVAL;
1366                 }
1367
1368                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1369                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1370                                      iov_size - iov_count);
1371                 if (unlikely(ret < 0)) {
1372                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1373                                ret, i);
1374                         return ret;
1375                 }
1376                 /* If this is an input descriptor, increment that count. */
1377                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1378                         *in_num += ret;
1379                         if (unlikely(log)) {
1380                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1381                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1382                                 ++*log_num;
1383                         }
1384                 } else {
1385                         /* If it's an output descriptor, they're all supposed
1386                          * to come before any input descriptors. */
1387                         if (unlikely(*in_num)) {
1388                                 vq_err(vq, "Indirect descriptor "
1389                                        "has out after in: idx %d\n", i);
1390                                 return -EINVAL;
1391                         }
1392                         *out_num += ret;
1393                 }
1394         } while ((i = next_desc(vq, &desc)) != -1);
1395         return 0;
1396 }
1397
1398 /* This looks in the virtqueue and for the first available buffer, and converts
1399  * it to an iovec for convenient access.  Since descriptors consist of some
1400  * number of output then some number of input descriptors, it's actually two
1401  * iovecs, but we pack them into one and note how many of each there were.
1402  *
1403  * This function returns the descriptor number found, or vq->num (which is
1404  * never a valid descriptor number) if none was found.  A negative code is
1405  * returned on error. */
1406 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1407                       struct iovec iov[], unsigned int iov_size,
1408                       unsigned int *out_num, unsigned int *in_num,
1409                       struct vhost_log *log, unsigned int *log_num)
1410 {
1411         struct vring_desc desc;
1412         unsigned int i, head, found = 0;
1413         u16 last_avail_idx;
1414         __virtio16 avail_idx;
1415         __virtio16 ring_head;
1416         int ret;
1417
1418         /* Check it isn't doing very strange things with descriptor numbers. */
1419         last_avail_idx = vq->last_avail_idx;
1420         if (unlikely(vhost_get_user(vq, avail_idx, &vq->avail->idx))) {
1421                 vq_err(vq, "Failed to access avail idx at %p\n",
1422                        &vq->avail->idx);
1423                 return -EFAULT;
1424         }
1425         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1426
1427         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1428                 vq_err(vq, "Guest moved used index from %u to %u",
1429                        last_avail_idx, vq->avail_idx);
1430                 return -EFAULT;
1431         }
1432
1433         /* If there's nothing new since last we looked, return invalid. */
1434         if (vq->avail_idx == last_avail_idx)
1435                 return vq->num;
1436
1437         /* Only get avail ring entries after they have been exposed by guest. */
1438         smp_rmb();
1439
1440         /* Grab the next descriptor number they're advertising, and increment
1441          * the index we've seen. */
1442         if (unlikely(vhost_get_user(vq, ring_head,
1443                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
1444                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1445                        last_avail_idx,
1446                        &vq->avail->ring[last_avail_idx % vq->num]);
1447                 return -EFAULT;
1448         }
1449
1450         head = vhost16_to_cpu(vq, ring_head);
1451
1452         /* If their number is silly, that's an error. */
1453         if (unlikely(head >= vq->num)) {
1454                 vq_err(vq, "Guest says index %u > %u is available",
1455                        head, vq->num);
1456                 return -EINVAL;
1457         }
1458
1459         /* When we start there are none of either input nor output. */
1460         *out_num = *in_num = 0;
1461         if (unlikely(log))
1462                 *log_num = 0;
1463
1464         i = head;
1465         do {
1466                 unsigned iov_count = *in_num + *out_num;
1467                 if (unlikely(i >= vq->num)) {
1468                         vq_err(vq, "Desc index is %u > %u, head = %u",
1469                                i, vq->num, head);
1470                         return -EINVAL;
1471                 }
1472                 if (unlikely(++found > vq->num)) {
1473                         vq_err(vq, "Loop detected: last one at %u "
1474                                "vq size %u head %u\n",
1475                                i, vq->num, head);
1476                         return -EINVAL;
1477                 }
1478                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
1479                                            sizeof desc);
1480                 if (unlikely(ret)) {
1481                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1482                                i, vq->desc + i);
1483                         return -EFAULT;
1484                 }
1485                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1486                         ret = get_indirect(vq, iov, iov_size,
1487                                            out_num, in_num,
1488                                            log, log_num, &desc);
1489                         if (unlikely(ret < 0)) {
1490                                 vq_err(vq, "Failure detected "
1491                                        "in indirect descriptor at idx %d\n", i);
1492                                 return ret;
1493                         }
1494                         continue;
1495                 }
1496
1497                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1498                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1499                                      iov_size - iov_count);
1500                 if (unlikely(ret < 0)) {
1501                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1502                                ret, i);
1503                         return ret;
1504                 }
1505                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1506                         /* If this is an input descriptor,
1507                          * increment that count. */
1508                         *in_num += ret;
1509                         if (unlikely(log)) {
1510                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1511                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1512                                 ++*log_num;
1513                         }
1514                 } else {
1515                         /* If it's an output descriptor, they're all supposed
1516                          * to come before any input descriptors. */
1517                         if (unlikely(*in_num)) {
1518                                 vq_err(vq, "Descriptor has out after in: "
1519                                        "idx %d\n", i);
1520                                 return -EINVAL;
1521                         }
1522                         *out_num += ret;
1523                 }
1524         } while ((i = next_desc(vq, &desc)) != -1);
1525
1526         /* On success, increment avail index. */
1527         vq->last_avail_idx++;
1528
1529         /* Assume notifications from guest are disabled at this point,
1530          * if they aren't we would need to update avail_event index. */
1531         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1532         return head;
1533 }
1534 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1535
1536 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1537 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1538 {
1539         vq->last_avail_idx -= n;
1540 }
1541 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1542
1543 /* After we've used one of their buffers, we tell them about it.  We'll then
1544  * want to notify the guest, using eventfd. */
1545 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1546 {
1547         struct vring_used_elem heads = {
1548                 cpu_to_vhost32(vq, head),
1549                 cpu_to_vhost32(vq, len)
1550         };
1551
1552         return vhost_add_used_n(vq, &heads, 1);
1553 }
1554 EXPORT_SYMBOL_GPL(vhost_add_used);
1555
1556 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1557                             struct vring_used_elem *heads,
1558                             unsigned count)
1559 {
1560         struct vring_used_elem __user *used;
1561         u16 old, new;
1562         int start;
1563
1564         start = vq->last_used_idx & (vq->num - 1);
1565         used = vq->used->ring + start;
1566         if (count == 1) {
1567                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
1568                         vq_err(vq, "Failed to write used id");
1569                         return -EFAULT;
1570                 }
1571                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
1572                         vq_err(vq, "Failed to write used len");
1573                         return -EFAULT;
1574                 }
1575         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
1576                 vq_err(vq, "Failed to write used");
1577                 return -EFAULT;
1578         }
1579         if (unlikely(vq->log_used)) {
1580                 /* Make sure data is seen before log. */
1581                 smp_wmb();
1582                 /* Log used ring entry write. */
1583                 log_write(vq->log_base,
1584                           vq->log_addr +
1585                            ((void __user *)used - (void __user *)vq->used),
1586                           count * sizeof *used);
1587         }
1588         old = vq->last_used_idx;
1589         new = (vq->last_used_idx += count);
1590         /* If the driver never bothers to signal in a very long while,
1591          * used index might wrap around. If that happens, invalidate
1592          * signalled_used index we stored. TODO: make sure driver
1593          * signals at least once in 2^16 and remove this. */
1594         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1595                 vq->signalled_used_valid = false;
1596         return 0;
1597 }
1598
1599 /* After we've used one of their buffers, we tell them about it.  We'll then
1600  * want to notify the guest, using eventfd. */
1601 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1602                      unsigned count)
1603 {
1604         int start, n, r;
1605
1606         start = vq->last_used_idx & (vq->num - 1);
1607         n = vq->num - start;
1608         if (n < count) {
1609                 r = __vhost_add_used_n(vq, heads, n);
1610                 if (r < 0)
1611                         return r;
1612                 heads += n;
1613                 count -= n;
1614         }
1615         r = __vhost_add_used_n(vq, heads, count);
1616
1617         /* Make sure buffer is written before we update index. */
1618         smp_wmb();
1619         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
1620                            &vq->used->idx)) {
1621                 vq_err(vq, "Failed to increment used idx");
1622                 return -EFAULT;
1623         }
1624         if (unlikely(vq->log_used)) {
1625                 /* Log used index update. */
1626                 log_write(vq->log_base,
1627                           vq->log_addr + offsetof(struct vring_used, idx),
1628                           sizeof vq->used->idx);
1629                 if (vq->log_ctx)
1630                         eventfd_signal(vq->log_ctx, 1);
1631         }
1632         return r;
1633 }
1634 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1635
1636 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1637 {
1638         __u16 old, new;
1639         __virtio16 event;
1640         bool v;
1641         /* Flush out used index updates. This is paired
1642          * with the barrier that the Guest executes when enabling
1643          * interrupts. */
1644         smp_mb();
1645
1646         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1647             unlikely(vq->avail_idx == vq->last_avail_idx))
1648                 return true;
1649
1650         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1651                 __virtio16 flags;
1652                 if (vhost_get_user(vq, flags, &vq->avail->flags)) {
1653                         vq_err(vq, "Failed to get flags");
1654                         return true;
1655                 }
1656                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1657         }
1658         old = vq->signalled_used;
1659         v = vq->signalled_used_valid;
1660         new = vq->signalled_used = vq->last_used_idx;
1661         vq->signalled_used_valid = true;
1662
1663         if (unlikely(!v))
1664                 return true;
1665
1666         if (vhost_get_user(vq, event, vhost_used_event(vq))) {
1667                 vq_err(vq, "Failed to get used event idx");
1668                 return true;
1669         }
1670         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1671 }
1672
1673 /* This actually signals the guest, using eventfd. */
1674 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1675 {
1676         /* Signal the Guest tell them we used something up. */
1677         if (vq->call_ctx && vhost_notify(dev, vq))
1678                 eventfd_signal(vq->call_ctx, 1);
1679 }
1680 EXPORT_SYMBOL_GPL(vhost_signal);
1681
1682 /* And here's the combo meal deal.  Supersize me! */
1683 void vhost_add_used_and_signal(struct vhost_dev *dev,
1684                                struct vhost_virtqueue *vq,
1685                                unsigned int head, int len)
1686 {
1687         vhost_add_used(vq, head, len);
1688         vhost_signal(dev, vq);
1689 }
1690 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1691
1692 /* multi-buffer version of vhost_add_used_and_signal */
1693 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1694                                  struct vhost_virtqueue *vq,
1695                                  struct vring_used_elem *heads, unsigned count)
1696 {
1697         vhost_add_used_n(vq, heads, count);
1698         vhost_signal(dev, vq);
1699 }
1700 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1701
1702 /* return true if we're sure that avaiable ring is empty */
1703 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1704 {
1705         __virtio16 avail_idx;
1706         int r;
1707
1708         r = vhost_get_user(vq, avail_idx, &vq->avail->idx);
1709         if (r)
1710                 return false;
1711
1712         return vhost16_to_cpu(vq, avail_idx) == vq->avail_idx;
1713 }
1714 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
1715
1716 /* OK, now we need to know about added descriptors. */
1717 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1718 {
1719         __virtio16 avail_idx;
1720         int r;
1721
1722         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1723                 return false;
1724         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1725         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1726                 r = vhost_update_used_flags(vq);
1727                 if (r) {
1728                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1729                                &vq->used->flags, r);
1730                         return false;
1731                 }
1732         } else {
1733                 r = vhost_update_avail_event(vq, vq->avail_idx);
1734                 if (r) {
1735                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1736                                vhost_avail_event(vq), r);
1737                         return false;
1738                 }
1739         }
1740         /* They could have slipped one in as we were doing that: make
1741          * sure it's written, then check again. */
1742         smp_mb();
1743         r = vhost_get_user(vq, avail_idx, &vq->avail->idx);
1744         if (r) {
1745                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1746                        &vq->avail->idx, r);
1747                 return false;
1748         }
1749
1750         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1751 }
1752 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1753
1754 /* We don't need to be notified again. */
1755 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1756 {
1757         int r;
1758
1759         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1760                 return;
1761         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1762         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1763                 r = vhost_update_used_flags(vq);
1764                 if (r)
1765                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1766                                &vq->used->flags, r);
1767         }
1768 }
1769 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1770
1771 static int __init vhost_init(void)
1772 {
1773         return 0;
1774 }
1775
1776 static void __exit vhost_exit(void)
1777 {
1778 }
1779
1780 module_init(vhost_init);
1781 module_exit(vhost_exit);
1782
1783 MODULE_VERSION("0.0.1");
1784 MODULE_LICENSE("GPL v2");
1785 MODULE_AUTHOR("Michael S. Tsirkin");
1786 MODULE_DESCRIPTION("Host kernel accelerator for virtio");