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 static int max_iotlb_entries = 2048;
  39 module_param(max_iotlb_entries, int, 0444);
  40 MODULE_PARM_DESC(max_iotlb_entries,
  41         "Maximum number of iotlb entries. (default: 2048)");
  42
  43 enum {
  44         VHOST_MEMORY_F_LOG = 0x1,
  45 };
  46
  47 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  48 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  49
  50 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  51                      rb, __u64, __subtree_last,
  52                      START, LAST, static inline, vhost_umem_interval_tree);
  53
  54 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  55 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  56 {
  57         vq->user_be = !virtio_legacy_is_little_endian();
  58 }
  59
  60 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  61 {
  62         vq->user_be = true;
  63 }
  64
  65 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  66 {
  67         vq->user_be = false;
  68 }
  69
  70 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  71 {
  72         struct vhost_vring_state s;
  73
  74         if (vq->private_data)
  75                 return -EBUSY;
  76
  77         if (copy_from_user(&s, argp, sizeof(s)))
  78                 return -EFAULT;
  79
  80         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  81             s.num != VHOST_VRING_BIG_ENDIAN)
  82                 return -EINVAL;
  83
  84         if (s.num == VHOST_VRING_BIG_ENDIAN)
  85                 vhost_enable_cross_endian_big(vq);
  86         else
  87                 vhost_enable_cross_endian_little(vq);
  88
  89         return 0;
  90 }
  91
  92 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  93                                    int __user *argp)
  94 {
  95         struct vhost_vring_state s = {
  96                 .index = idx,
  97                 .num = vq->user_be
  98         };
  99
 100         if (copy_to_user(argp, &s, sizeof(s)))
 101                 return -EFAULT;
 102
 103         return 0;
 104 }
 105
 106 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 107 {
 108         /* Note for legacy virtio: user_be is initialized at reset time
 109          * according to the host endianness. If userspace does not set an
 110          * explicit endianness, the default behavior is native endian, as
 111          * expected by legacy virtio.
 112          */
 113         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 114 }
 115 #else
 116 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 117 {
 118 }
 119
 120 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 121 {
 122         return -ENOIOCTLCMD;
 123 }
 124
 125 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 126                                    int __user *argp)
 127 {
 128         return -ENOIOCTLCMD;
 129 }
 130
 131 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 132 {
 133         if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
 134                 vq->is_le = true;
 135 }
 136 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 137
 138 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 139 {
 140         vq->is_le = virtio_legacy_is_little_endian();
 141 }
 142
 143 struct vhost_flush_struct {
 144         struct vhost_work work;
 145         struct completion wait_event;
 146 };
 147
 148 static void vhost_flush_work(struct vhost_work *work)
 149 {
 150         struct vhost_flush_struct *s;
 151
 152         s = container_of(work, struct vhost_flush_struct, work);
 153         complete(&s->wait_event);
 154 }
 155
 156 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 157                             poll_table *pt)
 158 {
 159         struct vhost_poll *poll;
 160
 161         poll = container_of(pt, struct vhost_poll, table);
 162         poll->wqh = wqh;
 163         add_wait_queue(wqh, &poll->wait);
 164 }
 165
 166 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
 167                              void *key)
 168 {
 169         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 170
 171         if (!((unsigned long)key & poll->mask))
 172                 return 0;
 173
 174         vhost_poll_queue(poll);
 175         return 0;
 176 }
 177
 178 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 179 {
 180         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 181         work->fn = fn;
 182         init_waitqueue_head(&work->done);
 183 }
 184 EXPORT_SYMBOL_GPL(vhost_work_init);
 185
 186 /* Init poll structure */
 187 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 188                      unsigned long mask, struct vhost_dev *dev)
 189 {
 190         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 191         init_poll_funcptr(&poll->table, vhost_poll_func);
 192         poll->mask = mask;
 193         poll->dev = dev;
 194         poll->wqh = NULL;
 195
 196         vhost_work_init(&poll->work, fn);
 197 }
 198 EXPORT_SYMBOL_GPL(vhost_poll_init);
 199
 200 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 201  * keep a reference to a file until after vhost_poll_stop is called. */
 202 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 203 {
 204         unsigned long mask;
 205         int ret = 0;
 206
 207         if (poll->wqh)
 208                 return 0;
 209
 210         mask = file->f_op->poll(file, &poll->table);
 211         if (mask)
 212                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 213         if (mask & POLLERR) {
 214                 if (poll->wqh)
 215                         remove_wait_queue(poll->wqh, &poll->wait);
 216                 ret = -EINVAL;
 217         }
 218
 219         return ret;
 220 }
 221 EXPORT_SYMBOL_GPL(vhost_poll_start);
 222
 223 /* Stop polling a file. After this function returns, it becomes safe to drop the
 224  * file reference. You must also flush afterwards. */
 225 void vhost_poll_stop(struct vhost_poll *poll)
 226 {
 227         if (poll->wqh) {
 228                 remove_wait_queue(poll->wqh, &poll->wait);
 229                 poll->wqh = NULL;
 230         }
 231 }
 232 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 233
 234 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 235 {
 236         struct vhost_flush_struct flush;
 237
 238         if (dev->worker) {
 239                 init_completion(&flush.wait_event);
 240                 vhost_work_init(&flush.work, vhost_flush_work);
 241
 242                 vhost_work_queue(dev, &flush.work);
 243                 wait_for_completion(&flush.wait_event);
 244         }
 245 }
 246 EXPORT_SYMBOL_GPL(vhost_work_flush);
 247
 248 /* Flush any work that has been scheduled. When calling this, don't hold any
 249  * locks that are also used by the callback. */
 250 void vhost_poll_flush(struct vhost_poll *poll)
 251 {
 252         vhost_work_flush(poll->dev, &poll->work);
 253 }
 254 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 255
 256 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 257 {
 258         if (!dev->worker)
 259                 return;
 260
 261         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 262                 /* We can only add the work to the list after we're
 263                  * sure it was not in the list.
 264                  * test_and_set_bit() implies a memory barrier.
 265                  */
 266                 llist_add(&work->node, &dev->work_list);
 267                 wake_up_process(dev->worker);
 268         }
 269 }
 270 EXPORT_SYMBOL_GPL(vhost_work_queue);
 271
 272 /* A lockless hint for busy polling code to exit the loop */
 273 bool vhost_has_work(struct vhost_dev *dev)
 274 {
 275         return !llist_empty(&dev->work_list);
 276 }
 277 EXPORT_SYMBOL_GPL(vhost_has_work);
 278
 279 void vhost_poll_queue(struct vhost_poll *poll)
 280 {
 281         vhost_work_queue(poll->dev, &poll->work);
 282 }
 283 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 284
 285 static void vhost_vq_reset(struct vhost_dev *dev,
 286                            struct vhost_virtqueue *vq)
 287 {
 288         vq->num = 1;
 289         vq->desc = NULL;
 290         vq->avail = NULL;
 291         vq->used = NULL;
 292         vq->last_avail_idx = 0;
 293         vq->last_used_event = 0;
 294         vq->avail_idx = 0;
 295         vq->last_used_idx = 0;
 296         vq->signalled_used = 0;
 297         vq->signalled_used_valid = false;
 298         vq->used_flags = 0;
 299         vq->log_used = false;
 300         vq->log_addr = -1ull;
 301         vq->private_data = NULL;
 302         vq->acked_features = 0;
 303         vq->log_base = NULL;
 304         vq->error_ctx = NULL;
 305         vq->error = NULL;
 306         vq->kick = NULL;
 307         vq->call_ctx = NULL;
 308         vq->call = NULL;
 309         vq->log_ctx = NULL;
 310         vhost_reset_is_le(vq);
 311         vhost_disable_cross_endian(vq);
 312         vq->busyloop_timeout = 0;
 313         vq->umem = NULL;
 314         vq->iotlb = NULL;
 315 }
 316
 317 static int vhost_worker(void *data)
 318 {
 319         struct vhost_dev *dev = data;
 320         struct vhost_work *work, *work_next;
 321         struct llist_node *node;
 322         mm_segment_t oldfs = get_fs();
 323
 324         set_fs(USER_DS);
 325         use_mm(dev->mm);
 326
 327         for (;;) {
 328                 /* mb paired w/ kthread_stop */
 329                 set_current_state(TASK_INTERRUPTIBLE);
 330
 331                 if (kthread_should_stop()) {
 332                         __set_current_state(TASK_RUNNING);
 333                         break;
 334                 }
 335
 336                 node = llist_del_all(&dev->work_list);
 337                 if (!node)
 338                         schedule();
 339
 340                 node = llist_reverse_order(node);
 341                 /* make sure flag is seen after deletion */
 342                 smp_wmb();
 343                 llist_for_each_entry_safe(work, work_next, node, node) {
 344                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 345                         __set_current_state(TASK_RUNNING);
 346                         work->fn(work);
 347                         if (need_resched())
 348                                 schedule();
 349                 }
 350         }
 351         unuse_mm(dev->mm);
 352         set_fs(oldfs);
 353         return 0;
 354 }
 355
 356 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 357 {
 358         kfree(vq->indirect);
 359         vq->indirect = NULL;
 360         kfree(vq->log);
 361         vq->log = NULL;
 362         kfree(vq->heads);
 363         vq->heads = NULL;
 364 }
 365
 366 /* Helper to allocate iovec buffers for all vqs. */
 367 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 368 {
 369         struct vhost_virtqueue *vq;
 370         int i;
 371
 372         for (i = 0; i < dev->nvqs; ++i) {
 373                 vq = dev->vqs[i];
 374                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 375                                        GFP_KERNEL);
 376                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 377                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 378                 if (!vq->indirect || !vq->log || !vq->heads)
 379                         goto err_nomem;
 380         }
 381         return 0;
 382
 383 err_nomem:
 384         for (; i >= 0; --i)
 385                 vhost_vq_free_iovecs(dev->vqs[i]);
 386         return -ENOMEM;
 387 }
 388
 389 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 390 {
 391         int i;
 392
 393         for (i = 0; i < dev->nvqs; ++i)
 394                 vhost_vq_free_iovecs(dev->vqs[i]);
 395 }
 396
 397 void vhost_dev_init(struct vhost_dev *dev,
 398                     struct vhost_virtqueue **vqs, int nvqs)
 399 {
 400         struct vhost_virtqueue *vq;
 401         int i;
 402
 403         dev->vqs = vqs;
 404         dev->nvqs = nvqs;
 405         mutex_init(&dev->mutex);
 406         dev->log_ctx = NULL;
 407         dev->log_file = NULL;
 408         dev->umem = NULL;
 409         dev->iotlb = NULL;
 410         dev->mm = NULL;
 411         dev->worker = NULL;
 412         init_llist_head(&dev->work_list);
 413         init_waitqueue_head(&dev->wait);
 414         INIT_LIST_HEAD(&dev->read_list);
 415         INIT_LIST_HEAD(&dev->pending_list);
 416         spin_lock_init(&dev->iotlb_lock);
 417
 418
 419         for (i = 0; i < dev->nvqs; ++i) {
 420                 vq = dev->vqs[i];
 421                 vq->log = NULL;
 422                 vq->indirect = NULL;
 423                 vq->heads = NULL;
 424                 vq->dev = dev;
 425                 mutex_init(&vq->mutex);
 426                 vhost_vq_reset(dev, vq);
 427                 if (vq->handle_kick)
 428                         vhost_poll_init(&vq->poll, vq->handle_kick,
 429                                         POLLIN, dev);
 430         }
 431 }
 432 EXPORT_SYMBOL_GPL(vhost_dev_init);
 433
 434 /* Caller should have device mutex */
 435 long vhost_dev_check_owner(struct vhost_dev *dev)
 436 {
 437         /* Are you the owner? If not, I don't think you mean to do that */
 438         return dev->mm == current->mm ? 0 : -EPERM;
 439 }
 440 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 441
 442 struct vhost_attach_cgroups_struct {
 443         struct vhost_work work;
 444         struct task_struct *owner;
 445         int ret;
 446 };
 447
 448 static void vhost_attach_cgroups_work(struct vhost_work *work)
 449 {
 450         struct vhost_attach_cgroups_struct *s;
 451
 452         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 453         s->ret = cgroup_attach_task_all(s->owner, current);
 454 }
 455
 456 static int vhost_attach_cgroups(struct vhost_dev *dev)
 457 {
 458         struct vhost_attach_cgroups_struct attach;
 459
 460         attach.owner = current;
 461         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 462         vhost_work_queue(dev, &attach.work);
 463         vhost_work_flush(dev, &attach.work);
 464         return attach.ret;
 465 }
 466
 467 /* Caller should have device mutex */
 468 bool vhost_dev_has_owner(struct vhost_dev *dev)
 469 {
 470         return dev->mm;
 471 }
 472 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 473
 474 /* Caller should have device mutex */
 475 long vhost_dev_set_owner(struct vhost_dev *dev)
 476 {
 477         struct task_struct *worker;
 478         int err;
 479
 480         /* Is there an owner already? */
 481         if (vhost_dev_has_owner(dev)) {
 482                 err = -EBUSY;
 483                 goto err_mm;
 484         }
 485
 486         /* No owner, become one */
 487         dev->mm = get_task_mm(current);
 488         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 489         if (IS_ERR(worker)) {
 490                 err = PTR_ERR(worker);
 491                 goto err_worker;
 492         }
 493
 494         dev->worker = worker;
 495         wake_up_process(worker);        /* avoid contributing to loadavg */
 496
 497         err = vhost_attach_cgroups(dev);
 498         if (err)
 499                 goto err_cgroup;
 500
 501         err = vhost_dev_alloc_iovecs(dev);
 502         if (err)
 503                 goto err_cgroup;
 504
 505         return 0;
 506 err_cgroup:
 507         kthread_stop(worker);
 508         dev->worker = NULL;
 509 err_worker:
 510         if (dev->mm)
 511                 mmput(dev->mm);
 512         dev->mm = NULL;
 513 err_mm:
 514         return err;
 515 }
 516 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 517
 518 static void *vhost_kvzalloc(unsigned long size)
 519 {
 520         void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 521
 522         if (!n)
 523                 n = vzalloc(size);
 524         return n;
 525 }
 526
 527 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 528 {
 529         return vhost_kvzalloc(sizeof(struct vhost_umem));
 530 }
 531 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 532
 533 /* Caller should have device mutex */
 534 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 535 {
 536         int i;
 537
 538         vhost_dev_cleanup(dev, true);
 539
 540         /* Restore memory to default empty mapping. */
 541         INIT_LIST_HEAD(&umem->umem_list);
 542         dev->umem = umem;
 543         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 544          * VQs aren't running.
 545          */
 546         for (i = 0; i < dev->nvqs; ++i)
 547                 dev->vqs[i]->umem = umem;
 548 }
 549 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 550
 551 void vhost_dev_stop(struct vhost_dev *dev)
 552 {
 553         int i;
 554
 555         for (i = 0; i < dev->nvqs; ++i) {
 556                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 557                         vhost_poll_stop(&dev->vqs[i]->poll);
 558                         vhost_poll_flush(&dev->vqs[i]->poll);
 559                 }
 560         }
 561 }
 562 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 563
 564 static void vhost_umem_free(struct vhost_umem *umem,
 565                             struct vhost_umem_node *node)
 566 {
 567         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 568         list_del(&node->link);
 569         kfree(node);
 570         umem->numem--;
 571 }
 572
 573 static void vhost_umem_clean(struct vhost_umem *umem)
 574 {
 575         struct vhost_umem_node *node, *tmp;
 576
 577         if (!umem)
 578                 return;
 579
 580         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 581                 vhost_umem_free(umem, node);
 582
 583         kvfree(umem);
 584 }
 585
 586 static void vhost_clear_msg(struct vhost_dev *dev)
 587 {
 588         struct vhost_msg_node *node, *n;
 589
 590         spin_lock(&dev->iotlb_lock);
 591
 592         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 593                 list_del(&node->node);
 594                 kfree(node);
 595         }
 596
 597         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 598                 list_del(&node->node);
 599                 kfree(node);
 600         }
 601
 602         spin_unlock(&dev->iotlb_lock);
 603 }
 604
 605 /* Caller should have device mutex if and only if locked is set */
 606 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 607 {
 608         int i;
 609
 610         for (i = 0; i < dev->nvqs; ++i) {
 611                 if (dev->vqs[i]->error_ctx)
 612                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 613                 if (dev->vqs[i]->error)
 614                         fput(dev->vqs[i]->error);
 615                 if (dev->vqs[i]->kick)
 616                         fput(dev->vqs[i]->kick);
 617                 if (dev->vqs[i]->call_ctx)
 618                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 619                 if (dev->vqs[i]->call)
 620                         fput(dev->vqs[i]->call);
 621                 vhost_vq_reset(dev, dev->vqs[i]);
 622         }
 623         vhost_dev_free_iovecs(dev);
 624         if (dev->log_ctx)
 625                 eventfd_ctx_put(dev->log_ctx);
 626         dev->log_ctx = NULL;
 627         if (dev->log_file)
 628                 fput(dev->log_file);
 629         dev->log_file = NULL;
 630         /* No one will access memory at this point */
 631         vhost_umem_clean(dev->umem);
 632         dev->umem = NULL;
 633         vhost_umem_clean(dev->iotlb);
 634         dev->iotlb = NULL;
 635         vhost_clear_msg(dev);
 636         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
 637         WARN_ON(!llist_empty(&dev->work_list));
 638         if (dev->worker) {
 639                 kthread_stop(dev->worker);
 640                 dev->worker = NULL;
 641         }
 642         if (dev->mm)
 643                 mmput(dev->mm);
 644         dev->mm = NULL;
 645 }
 646 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 647
 648 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 649 {
 650         u64 a = addr / VHOST_PAGE_SIZE / 8;
 651
 652         /* Make sure 64 bit math will not overflow. */
 653         if (a > ULONG_MAX - (unsigned long)log_base ||
 654             a + (unsigned long)log_base > ULONG_MAX)
 655                 return 0;
 656
 657         return access_ok(VERIFY_WRITE, log_base + a,
 658                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 659 }
 660
 661 static bool vhost_overflow(u64 uaddr, u64 size)
 662 {
 663         /* Make sure 64 bit math will not overflow. */
 664         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 665 }
 666
 667 /* Caller should have vq mutex and device mutex. */
 668 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 669                                int log_all)
 670 {
 671         struct vhost_umem_node *node;
 672
 673         if (!umem)
 674                 return 0;
 675
 676         list_for_each_entry(node, &umem->umem_list, link) {
 677                 unsigned long a = node->userspace_addr;
 678
 679                 if (vhost_overflow(node->userspace_addr, node->size))
 680                         return 0;
 681
 682
 683                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 684                                     node->size))
 685                         return 0;
 686                 else if (log_all && !log_access_ok(log_base,
 687                                                    node->start,
 688                                                    node->size))
 689                         return 0;
 690         }
 691         return 1;
 692 }
 693
 694 /* Can we switch to this memory table? */
 695 /* Caller should have device mutex but not vq mutex */
 696 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 697                             int log_all)
 698 {
 699         int i;
 700
 701         for (i = 0; i < d->nvqs; ++i) {
 702                 int ok;
 703                 bool log;
 704
 705                 mutex_lock(&d->vqs[i]->mutex);
 706                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 707                 /* If ring is inactive, will check when it's enabled. */
 708                 if (d->vqs[i]->private_data)
 709                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 710                                                  umem, log);
 711                 else
 712                         ok = 1;
 713                 mutex_unlock(&d->vqs[i]->mutex);
 714                 if (!ok)
 715                         return 0;
 716         }
 717         return 1;
 718 }
 719
 720 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 721                           struct iovec iov[], int iov_size, int access);
 722
 723 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 724                               const void *from, unsigned size)
 725 {
 726         int ret;
 727
 728         if (!vq->iotlb)
 729                 return __copy_to_user(to, from, size);
 730         else {
 731                 /* This function should be called after iotlb
 732                  * prefetch, which means we're sure that all vq
 733                  * could be access through iotlb. So -EAGAIN should
 734                  * not happen in this case.
 735                  */
 736                 /* TODO: more fast path */
 737                 struct iov_iter t;
 738                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 739                                      ARRAY_SIZE(vq->iotlb_iov),
 740                                      VHOST_ACCESS_WO);
 741                 if (ret < 0)
 742                         goto out;
 743                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 744                 ret = copy_to_iter(from, size, &t);
 745                 if (ret == size)
 746                         ret = 0;
 747         }
 748 out:
 749         return ret;
 750 }
 751
 752 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 753                                 void __user *from, unsigned size)
 754 {
 755         int ret;
 756
 757         if (!vq->iotlb)
 758                 return __copy_from_user(to, from, size);
 759         else {
 760                 /* This function should be called after iotlb
 761                  * prefetch, which means we're sure that vq
 762                  * could be access through iotlb. So -EAGAIN should
 763                  * not happen in this case.
 764                  */
 765                 /* TODO: more fast path */
 766                 struct iov_iter f;
 767                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 768                                      ARRAY_SIZE(vq->iotlb_iov),
 769                                      VHOST_ACCESS_RO);
 770                 if (ret < 0) {
 771                         vq_err(vq, "IOTLB translation failure: uaddr "
 772                                "%p size 0x%llx\n", from,
 773                                (unsigned long long) size);
 774                         goto out;
 775                 }
 776                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 777                 ret = copy_from_iter(to, size, &f);
 778                 if (ret == size)
 779                         ret = 0;
 780         }
 781
 782 out:
 783         return ret;
 784 }
 785
 786 static void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 787                                      void __user *addr, unsigned size)
 788 {
 789         int ret;
 790
 791         /* This function should be called after iotlb
 792          * prefetch, which means we're sure that vq
 793          * could be access through iotlb. So -EAGAIN should
 794          * not happen in this case.
 795          */
 796         /* TODO: more fast path */
 797         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 798                              ARRAY_SIZE(vq->iotlb_iov),
 799                              VHOST_ACCESS_RO);
 800         if (ret < 0) {
 801                 vq_err(vq, "IOTLB translation failure: uaddr "
 802                         "%p size 0x%llx\n", addr,
 803                         (unsigned long long) size);
 804                 return NULL;
 805         }
 806
 807         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 808                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 809                         "%p size 0x%llx\n", addr,
 810                         (unsigned long long) size);
 811                 return NULL;
 812         }
 813
 814         return vq->iotlb_iov[0].iov_base;
 815 }
 816
 817 #define vhost_put_user(vq, x, ptr) \
 818 ({ \
 819         int ret = -EFAULT; \
 820         if (!vq->iotlb) { \
 821                 ret = __put_user(x, ptr); \
 822         } else { \
 823                 __typeof__(ptr) to = \
 824                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, sizeof(*ptr)); \
 825                 if (to != NULL) \
 826                         ret = __put_user(x, to); \
 827                 else \
 828                         ret = -EFAULT;  \
 829         } \
 830         ret; \
 831 })
 832
 833 #define vhost_get_user(vq, x, ptr) \
 834 ({ \
 835         int ret; \
 836         if (!vq->iotlb) { \
 837                 ret = __get_user(x, ptr); \
 838         } else { \
 839                 __typeof__(ptr) from = \
 840                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, sizeof(*ptr)); \
 841                 if (from != NULL) \
 842                         ret = __get_user(x, from); \
 843                 else \
 844                         ret = -EFAULT; \
 845         } \
 846         ret; \
 847 })
 848
 849 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 850 {
 851         int i = 0;
 852         for (i = 0; i < d->nvqs; ++i)
 853                 mutex_lock(&d->vqs[i]->mutex);
 854 }
 855
 856 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 857 {
 858         int i = 0;
 859         for (i = 0; i < d->nvqs; ++i)
 860                 mutex_unlock(&d->vqs[i]->mutex);
 861 }
 862
 863 static int vhost_new_umem_range(struct vhost_umem *umem,
 864                                 u64 start, u64 size, u64 end,
 865                                 u64 userspace_addr, int perm)
 866 {
 867         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
 868
 869         if (!node)
 870                 return -ENOMEM;
 871
 872         if (umem->numem == max_iotlb_entries) {
 873                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 874                 vhost_umem_free(umem, tmp);
 875         }
 876
 877         node->start = start;
 878         node->size = size;
 879         node->last = end;
 880         node->userspace_addr = userspace_addr;
 881         node->perm = perm;
 882         INIT_LIST_HEAD(&node->link);
 883         list_add_tail(&node->link, &umem->umem_list);
 884         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 885         umem->numem++;
 886
 887         return 0;
 888 }
 889
 890 static void vhost_del_umem_range(struct vhost_umem *umem,
 891                                  u64 start, u64 end)
 892 {
 893         struct vhost_umem_node *node;
 894
 895         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 896                                                            start, end)))
 897                 vhost_umem_free(umem, node);
 898 }
 899
 900 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 901                                   struct vhost_iotlb_msg *msg)
 902 {
 903         struct vhost_msg_node *node, *n;
 904
 905         spin_lock(&d->iotlb_lock);
 906
 907         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 908                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 909                 if (msg->iova <= vq_msg->iova &&
 910                     msg->iova + msg->size - 1 > vq_msg->iova &&
 911                     vq_msg->type == VHOST_IOTLB_MISS) {
 912                         vhost_poll_queue(&node->vq->poll);
 913                         list_del(&node->node);
 914                         kfree(node);
 915                 }
 916         }
 917
 918         spin_unlock(&d->iotlb_lock);
 919 }
 920
 921 static int umem_access_ok(u64 uaddr, u64 size, int access)
 922 {
 923         unsigned long a = uaddr;
 924
 925         /* Make sure 64 bit math will not overflow. */
 926         if (vhost_overflow(uaddr, size))
 927                 return -EFAULT;
 928
 929         if ((access & VHOST_ACCESS_RO) &&
 930             !access_ok(VERIFY_READ, (void __user *)a, size))
 931                 return -EFAULT;
 932         if ((access & VHOST_ACCESS_WO) &&
 933             !access_ok(VERIFY_WRITE, (void __user *)a, size))
 934                 return -EFAULT;
 935         return 0;
 936 }
 937
 938 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
 939                                    struct vhost_iotlb_msg *msg)
 940 {
 941         int ret = 0;
 942
 943         vhost_dev_lock_vqs(dev);
 944         switch (msg->type) {
 945         case VHOST_IOTLB_UPDATE:
 946                 if (!dev->iotlb) {
 947                         ret = -EFAULT;
 948                         break;
 949                 }
 950                 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
 951                         ret = -EFAULT;
 952                         break;
 953                 }
 954                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
 955                                          msg->iova + msg->size - 1,
 956                                          msg->uaddr, msg->perm)) {
 957                         ret = -ENOMEM;
 958                         break;
 959                 }
 960                 vhost_iotlb_notify_vq(dev, msg);
 961                 break;
 962         case VHOST_IOTLB_INVALIDATE:
 963                 vhost_del_umem_range(dev->iotlb, msg->iova,
 964                                      msg->iova + msg->size - 1);
 965                 break;
 966         default:
 967                 ret = -EINVAL;
 968                 break;
 969         }
 970
 971         vhost_dev_unlock_vqs(dev);
 972         return ret;
 973 }
 974 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
 975                              struct iov_iter *from)
 976 {
 977         struct vhost_msg_node node;
 978         unsigned size = sizeof(struct vhost_msg);
 979         size_t ret;
 980         int err;
 981
 982         if (iov_iter_count(from) < size)
 983                 return 0;
 984         ret = copy_from_iter(&node.msg, size, from);
 985         if (ret != size)
 986                 goto done;
 987
 988         switch (node.msg.type) {
 989         case VHOST_IOTLB_MSG:
 990                 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
 991                 if (err)
 992                         ret = err;
 993                 break;
 994         default:
 995                 ret = -EINVAL;
 996                 break;
 997         }
 998
 999 done:
1000         return ret;
1001 }
1002 EXPORT_SYMBOL(vhost_chr_write_iter);
1003
1004 unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1005                             poll_table *wait)
1006 {
1007         unsigned int mask = 0;
1008
1009         poll_wait(file, &dev->wait, wait);
1010
1011         if (!list_empty(&dev->read_list))
1012                 mask |= POLLIN | POLLRDNORM;
1013
1014         return mask;
1015 }
1016 EXPORT_SYMBOL(vhost_chr_poll);
1017
1018 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1019                             int noblock)
1020 {
1021         DEFINE_WAIT(wait);
1022         struct vhost_msg_node *node;
1023         ssize_t ret = 0;
1024         unsigned size = sizeof(struct vhost_msg);
1025
1026         if (iov_iter_count(to) < size)
1027                 return 0;
1028
1029         while (1) {
1030                 if (!noblock)
1031                         prepare_to_wait(&dev->wait, &wait,
1032                                         TASK_INTERRUPTIBLE);
1033
1034                 node = vhost_dequeue_msg(dev, &dev->read_list);
1035                 if (node)
1036                         break;
1037                 if (noblock) {
1038                         ret = -EAGAIN;
1039                         break;
1040                 }
1041                 if (signal_pending(current)) {
1042                         ret = -ERESTARTSYS;
1043                         break;
1044                 }
1045                 if (!dev->iotlb) {
1046                         ret = -EBADFD;
1047                         break;
1048                 }
1049
1050                 schedule();
1051         }
1052
1053         if (!noblock)
1054                 finish_wait(&dev->wait, &wait);
1055
1056         if (node) {
1057                 ret = copy_to_iter(&node->msg, size, to);
1058
1059                 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1060                         kfree(node);
1061                         return ret;
1062                 }
1063
1064                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1065         }
1066
1067         return ret;
1068 }
1069 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1070
1071 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1072 {
1073         struct vhost_dev *dev = vq->dev;
1074         struct vhost_msg_node *node;
1075         struct vhost_iotlb_msg *msg;
1076
1077         node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1078         if (!node)
1079                 return -ENOMEM;
1080
1081         msg = &node->msg.iotlb;
1082         msg->type = VHOST_IOTLB_MISS;
1083         msg->iova = iova;
1084         msg->perm = access;
1085
1086         vhost_enqueue_msg(dev, &dev->read_list, node);
1087
1088         return 0;
1089 }
1090
1091 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1092                         struct vring_desc __user *desc,
1093                         struct vring_avail __user *avail,
1094                         struct vring_used __user *used)
1095
1096 {
1097         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1098
1099         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1100                access_ok(VERIFY_READ, avail,
1101                          sizeof *avail + num * sizeof *avail->ring + s) &&
1102                access_ok(VERIFY_WRITE, used,
1103                         sizeof *used + num * sizeof *used->ring + s);
1104 }
1105
1106 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1107                            int access, u64 addr, u64 len)
1108 {
1109         const struct vhost_umem_node *node;
1110         struct vhost_umem *umem = vq->iotlb;
1111         u64 s = 0, size;
1112
1113         while (len > s) {
1114                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1115                                                            addr,
1116                                                            addr + len - 1);
1117                 if (node == NULL || node->start > addr) {
1118                         vhost_iotlb_miss(vq, addr, access);
1119                         return false;
1120                 } else if (!(node->perm & access)) {
1121                         /* Report the possible access violation by
1122                          * request another translation from userspace.
1123                          */
1124                         return false;
1125                 }
1126
1127                 size = node->size - addr + node->start;
1128                 s += size;
1129                 addr += size;
1130         }
1131
1132         return true;
1133 }
1134
1135 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1136 {
1137         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1138         unsigned int num = vq->num;
1139
1140         if (!vq->iotlb)
1141                 return 1;
1142
1143         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1144                                num * sizeof *vq->desc) &&
1145                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1146                                sizeof *vq->avail +
1147                                num * sizeof *vq->avail->ring + s) &&
1148                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1149                                sizeof *vq->used +
1150                                num * sizeof *vq->used->ring + s);
1151 }
1152 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1153
1154 /* Can we log writes? */
1155 /* Caller should have device mutex but not vq mutex */
1156 int vhost_log_access_ok(struct vhost_dev *dev)
1157 {
1158         return memory_access_ok(dev, dev->umem, 1);
1159 }
1160 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1161
1162 /* Verify access for write logging. */
1163 /* Caller should have vq mutex and device mutex */
1164 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1165                             void __user *log_base)
1166 {
1167         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1168
1169         return vq_memory_access_ok(log_base, vq->umem,
1170                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1171                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1172                                         sizeof *vq->used +
1173                                         vq->num * sizeof *vq->used->ring + s));
1174 }
1175
1176 /* Can we start vq? */
1177 /* Caller should have vq mutex and device mutex */
1178 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1179 {
1180         if (vq->iotlb) {
1181                 /* When device IOTLB was used, the access validation
1182                  * will be validated during prefetching.
1183                  */
1184                 return 1;
1185         }
1186         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1187                 vq_log_access_ok(vq, vq->log_base);
1188 }
1189 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1190
1191 static struct vhost_umem *vhost_umem_alloc(void)
1192 {
1193         struct vhost_umem *umem = vhost_kvzalloc(sizeof(*umem));
1194
1195         if (!umem)
1196                 return NULL;
1197
1198         umem->umem_tree = RB_ROOT;
1199         umem->numem = 0;
1200         INIT_LIST_HEAD(&umem->umem_list);
1201
1202         return umem;
1203 }
1204
1205 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1206 {
1207         struct vhost_memory mem, *newmem;
1208         struct vhost_memory_region *region;
1209         struct vhost_umem *newumem, *oldumem;
1210         unsigned long size = offsetof(struct vhost_memory, regions);
1211         int i;
1212
1213         if (copy_from_user(&mem, m, size))
1214                 return -EFAULT;
1215         if (mem.padding)
1216                 return -EOPNOTSUPP;
1217         if (mem.nregions > max_mem_regions)
1218                 return -E2BIG;
1219         newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
1220         if (!newmem)
1221                 return -ENOMEM;
1222
1223         memcpy(newmem, &mem, size);
1224         if (copy_from_user(newmem->regions, m->regions,
1225                            mem.nregions * sizeof *m->regions)) {
1226                 kvfree(newmem);
1227                 return -EFAULT;
1228         }
1229
1230         newumem = vhost_umem_alloc();
1231         if (!newumem) {
1232                 kvfree(newmem);
1233                 return -ENOMEM;
1234         }
1235
1236         for (region = newmem->regions;
1237              region < newmem->regions + mem.nregions;
1238              region++) {
1239                 if (vhost_new_umem_range(newumem,
1240                                          region->guest_phys_addr,
1241                                          region->memory_size,
1242                                          region->guest_phys_addr +
1243                                          region->memory_size - 1,
1244                                          region->userspace_addr,
1245                                          VHOST_ACCESS_RW))
1246                         goto err;
1247         }
1248
1249         if (!memory_access_ok(d, newumem, 0))
1250                 goto err;
1251
1252         oldumem = d->umem;
1253         d->umem = newumem;
1254
1255         /* All memory accesses are done under some VQ mutex. */
1256         for (i = 0; i < d->nvqs; ++i) {
1257                 mutex_lock(&d->vqs[i]->mutex);
1258                 d->vqs[i]->umem = newumem;
1259                 mutex_unlock(&d->vqs[i]->mutex);
1260         }
1261
1262         kvfree(newmem);
1263         vhost_umem_clean(oldumem);
1264         return 0;
1265
1266 err:
1267         vhost_umem_clean(newumem);
1268         kvfree(newmem);
1269         return -EFAULT;
1270 }
1271
1272 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1273 {
1274         struct file *eventfp, *filep = NULL;
1275         bool pollstart = false, pollstop = false;
1276         struct eventfd_ctx *ctx = NULL;
1277         u32 __user *idxp = argp;
1278         struct vhost_virtqueue *vq;
1279         struct vhost_vring_state s;
1280         struct vhost_vring_file f;
1281         struct vhost_vring_addr a;
1282         u32 idx;
1283         long r;
1284
1285         r = get_user(idx, idxp);
1286         if (r < 0)
1287                 return r;
1288         if (idx >= d->nvqs)
1289                 return -ENOBUFS;
1290
1291         vq = d->vqs[idx];
1292
1293         mutex_lock(&vq->mutex);
1294
1295         switch (ioctl) {
1296         case VHOST_SET_VRING_NUM:
1297                 /* Resizing ring with an active backend?
1298                  * You don't want to do that. */
1299                 if (vq->private_data) {
1300                         r = -EBUSY;
1301                         break;
1302                 }
1303                 if (copy_from_user(&s, argp, sizeof s)) {
1304                         r = -EFAULT;
1305                         break;
1306                 }
1307                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1308                         r = -EINVAL;
1309                         break;
1310                 }
1311                 vq->num = s.num;
1312                 break;
1313         case VHOST_SET_VRING_BASE:
1314                 /* Moving base with an active backend?
1315                  * You don't want to do that. */
1316                 if (vq->private_data) {
1317                         r = -EBUSY;
1318                         break;
1319                 }
1320                 if (copy_from_user(&s, argp, sizeof s)) {
1321                         r = -EFAULT;
1322                         break;
1323                 }
1324                 if (s.num > 0xffff) {
1325                         r = -EINVAL;
1326                         break;
1327                 }
1328                 vq->last_avail_idx = vq->last_used_event = s.num;
1329                 /* Forget the cached index value. */
1330                 vq->avail_idx = vq->last_avail_idx;
1331                 break;
1332         case VHOST_GET_VRING_BASE:
1333                 s.index = idx;
1334                 s.num = vq->last_avail_idx;
1335                 if (copy_to_user(argp, &s, sizeof s))
1336                         r = -EFAULT;
1337                 break;
1338         case VHOST_SET_VRING_ADDR:
1339                 if (copy_from_user(&a, argp, sizeof a)) {
1340                         r = -EFAULT;
1341                         break;
1342                 }
1343                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1344                         r = -EOPNOTSUPP;
1345                         break;
1346                 }
1347                 /* For 32bit, verify that the top 32bits of the user
1348                    data are set to zero. */
1349                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1350                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1351                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1352                         r = -EFAULT;
1353                         break;
1354                 }
1355
1356                 /* Make sure it's safe to cast pointers to vring types. */
1357                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1358                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1359                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1360                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1361                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1362                         r = -EINVAL;
1363                         break;
1364                 }
1365
1366                 /* We only verify access here if backend is configured.
1367                  * If it is not, we don't as size might not have been setup.
1368                  * We will verify when backend is configured. */
1369                 if (vq->private_data) {
1370                         if (!vq_access_ok(vq, vq->num,
1371                                 (void __user *)(unsigned long)a.desc_user_addr,
1372                                 (void __user *)(unsigned long)a.avail_user_addr,
1373                                 (void __user *)(unsigned long)a.used_user_addr)) {
1374                                 r = -EINVAL;
1375                                 break;
1376                         }
1377
1378                         /* Also validate log access for used ring if enabled. */
1379                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1380                             !log_access_ok(vq->log_base, a.log_guest_addr,
1381                                            sizeof *vq->used +
1382                                            vq->num * sizeof *vq->used->ring)) {
1383                                 r = -EINVAL;
1384                                 break;
1385                         }
1386                 }
1387
1388                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1389                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1390                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1391                 vq->log_addr = a.log_guest_addr;
1392                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1393                 break;
1394         case VHOST_SET_VRING_KICK:
1395                 if (copy_from_user(&f, argp, sizeof f)) {
1396                         r = -EFAULT;
1397                         break;
1398                 }
1399                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1400                 if (IS_ERR(eventfp)) {
1401                         r = PTR_ERR(eventfp);
1402                         break;
1403                 }
1404                 if (eventfp != vq->kick) {
1405                         pollstop = (filep = vq->kick) != NULL;
1406                         pollstart = (vq->kick = eventfp) != NULL;
1407                 } else
1408                         filep = eventfp;
1409                 break;
1410         case VHOST_SET_VRING_CALL:
1411                 if (copy_from_user(&f, argp, sizeof f)) {
1412                         r = -EFAULT;
1413                         break;
1414                 }
1415                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1416                 if (IS_ERR(eventfp)) {
1417                         r = PTR_ERR(eventfp);
1418                         break;
1419                 }
1420                 if (eventfp != vq->call) {
1421                         filep = vq->call;
1422                         ctx = vq->call_ctx;
1423                         vq->call = eventfp;
1424                         vq->call_ctx = eventfp ?
1425                                 eventfd_ctx_fileget(eventfp) : NULL;
1426                 } else
1427                         filep = eventfp;
1428                 break;
1429         case VHOST_SET_VRING_ERR:
1430                 if (copy_from_user(&f, argp, sizeof f)) {
1431                         r = -EFAULT;
1432                         break;
1433                 }
1434                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1435                 if (IS_ERR(eventfp)) {
1436                         r = PTR_ERR(eventfp);
1437                         break;
1438                 }
1439                 if (eventfp != vq->error) {
1440                         filep = vq->error;
1441                         vq->error = eventfp;
1442                         ctx = vq->error_ctx;
1443                         vq->error_ctx = eventfp ?
1444                                 eventfd_ctx_fileget(eventfp) : NULL;
1445                 } else
1446                         filep = eventfp;
1447                 break;
1448         case VHOST_SET_VRING_ENDIAN:
1449                 r = vhost_set_vring_endian(vq, argp);
1450                 break;
1451         case VHOST_GET_VRING_ENDIAN:
1452                 r = vhost_get_vring_endian(vq, idx, argp);
1453                 break;
1454         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1455                 if (copy_from_user(&s, argp, sizeof(s))) {
1456                         r = -EFAULT;
1457                         break;
1458                 }
1459                 vq->busyloop_timeout = s.num;
1460                 break;
1461         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1462                 s.index = idx;
1463                 s.num = vq->busyloop_timeout;
1464                 if (copy_to_user(argp, &s, sizeof(s)))
1465                         r = -EFAULT;
1466                 break;
1467         default:
1468                 r = -ENOIOCTLCMD;
1469         }
1470
1471         if (pollstop && vq->handle_kick)
1472                 vhost_poll_stop(&vq->poll);
1473
1474         if (ctx)
1475                 eventfd_ctx_put(ctx);
1476         if (filep)
1477                 fput(filep);
1478
1479         if (pollstart && vq->handle_kick)
1480                 r = vhost_poll_start(&vq->poll, vq->kick);
1481
1482         mutex_unlock(&vq->mutex);
1483
1484         if (pollstop && vq->handle_kick)
1485                 vhost_poll_flush(&vq->poll);
1486         return r;
1487 }
1488 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1489
1490 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1491 {
1492         struct vhost_umem *niotlb, *oiotlb;
1493         int i;
1494
1495         niotlb = vhost_umem_alloc();
1496         if (!niotlb)
1497                 return -ENOMEM;
1498
1499         oiotlb = d->iotlb;
1500         d->iotlb = niotlb;
1501
1502         for (i = 0; i < d->nvqs; ++i) {
1503                 mutex_lock(&d->vqs[i]->mutex);
1504                 d->vqs[i]->iotlb = niotlb;
1505                 mutex_unlock(&d->vqs[i]->mutex);
1506         }
1507
1508         vhost_umem_clean(oiotlb);
1509
1510         return 0;
1511 }
1512 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1513
1514 /* Caller must have device mutex */
1515 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1516 {
1517         struct file *eventfp, *filep = NULL;
1518         struct eventfd_ctx *ctx = NULL;
1519         u64 p;
1520         long r;
1521         int i, fd;
1522
1523         /* If you are not the owner, you can become one */
1524         if (ioctl == VHOST_SET_OWNER) {
1525                 r = vhost_dev_set_owner(d);
1526                 goto done;
1527         }
1528
1529         /* You must be the owner to do anything else */
1530         r = vhost_dev_check_owner(d);
1531         if (r)
1532                 goto done;
1533
1534         switch (ioctl) {
1535         case VHOST_SET_MEM_TABLE:
1536                 r = vhost_set_memory(d, argp);
1537                 break;
1538         case VHOST_SET_LOG_BASE:
1539                 if (copy_from_user(&p, argp, sizeof p)) {
1540                         r = -EFAULT;
1541                         break;
1542                 }
1543                 if ((u64)(unsigned long)p != p) {
1544                         r = -EFAULT;
1545                         break;
1546                 }
1547                 for (i = 0; i < d->nvqs; ++i) {
1548                         struct vhost_virtqueue *vq;
1549                         void __user *base = (void __user *)(unsigned long)p;
1550                         vq = d->vqs[i];
1551                         mutex_lock(&vq->mutex);
1552                         /* If ring is inactive, will check when it's enabled. */
1553                         if (vq->private_data && !vq_log_access_ok(vq, base))
1554                                 r = -EFAULT;
1555                         else
1556                                 vq->log_base = base;
1557                         mutex_unlock(&vq->mutex);
1558                 }
1559                 break;
1560         case VHOST_SET_LOG_FD:
1561                 r = get_user(fd, (int __user *)argp);
1562                 if (r < 0)
1563                         break;
1564                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1565                 if (IS_ERR(eventfp)) {
1566                         r = PTR_ERR(eventfp);
1567                         break;
1568                 }
1569                 if (eventfp != d->log_file) {
1570                         filep = d->log_file;
1571                         d->log_file = eventfp;
1572                         ctx = d->log_ctx;
1573                         d->log_ctx = eventfp ?
1574                                 eventfd_ctx_fileget(eventfp) : NULL;
1575                 } else
1576                         filep = eventfp;
1577                 for (i = 0; i < d->nvqs; ++i) {
1578                         mutex_lock(&d->vqs[i]->mutex);
1579                         d->vqs[i]->log_ctx = d->log_ctx;
1580                         mutex_unlock(&d->vqs[i]->mutex);
1581                 }
1582                 if (ctx)
1583                         eventfd_ctx_put(ctx);
1584                 if (filep)
1585                         fput(filep);
1586                 break;
1587         default:
1588                 r = -ENOIOCTLCMD;
1589                 break;
1590         }
1591 done:
1592         return r;
1593 }
1594 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1595
1596 /* TODO: This is really inefficient.  We need something like get_user()
1597  * (instruction directly accesses the data, with an exception table entry
1598  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1599  */
1600 static int set_bit_to_user(int nr, void __user *addr)
1601 {
1602         unsigned long log = (unsigned long)addr;
1603         struct page *page;
1604         void *base;
1605         int bit = nr + (log % PAGE_SIZE) * 8;
1606         int r;
1607
1608         r = get_user_pages_fast(log, 1, 1, &page);
1609         if (r < 0)
1610                 return r;
1611         BUG_ON(r != 1);
1612         base = kmap_atomic(page);
1613         set_bit(bit, base);
1614         kunmap_atomic(base);
1615         set_page_dirty_lock(page);
1616         put_page(page);
1617         return 0;
1618 }
1619
1620 static int log_write(void __user *log_base,
1621                      u64 write_address, u64 write_length)
1622 {
1623         u64 write_page = write_address / VHOST_PAGE_SIZE;
1624         int r;
1625
1626         if (!write_length)
1627                 return 0;
1628         write_length += write_address % VHOST_PAGE_SIZE;
1629         for (;;) {
1630                 u64 base = (u64)(unsigned long)log_base;
1631                 u64 log = base + write_page / 8;
1632                 int bit = write_page % 8;
1633                 if ((u64)(unsigned long)log != log)
1634                         return -EFAULT;
1635                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1636                 if (r < 0)
1637                         return r;
1638                 if (write_length <= VHOST_PAGE_SIZE)
1639                         break;
1640                 write_length -= VHOST_PAGE_SIZE;
1641                 write_page += 1;
1642         }
1643         return r;
1644 }
1645
1646 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1647                     unsigned int log_num, u64 len)
1648 {
1649         int i, r;
1650
1651         /* Make sure data written is seen before log. */
1652         smp_wmb();
1653         for (i = 0; i < log_num; ++i) {
1654                 u64 l = min(log[i].len, len);
1655                 r = log_write(vq->log_base, log[i].addr, l);
1656                 if (r < 0)
1657                         return r;
1658                 len -= l;
1659                 if (!len) {
1660                         if (vq->log_ctx)
1661                                 eventfd_signal(vq->log_ctx, 1);
1662                         return 0;
1663                 }
1664         }
1665         /* Length written exceeds what we have stored. This is a bug. */
1666         BUG();
1667         return 0;
1668 }
1669 EXPORT_SYMBOL_GPL(vhost_log_write);
1670
1671 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1672 {
1673         void __user *used;
1674         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1675                            &vq->used->flags) < 0)
1676                 return -EFAULT;
1677         if (unlikely(vq->log_used)) {
1678                 /* Make sure the flag is seen before log. */
1679                 smp_wmb();
1680                 /* Log used flag write. */
1681                 used = &vq->used->flags;
1682                 log_write(vq->log_base, vq->log_addr +
1683                           (used - (void __user *)vq->used),
1684                           sizeof vq->used->flags);
1685                 if (vq->log_ctx)
1686                         eventfd_signal(vq->log_ctx, 1);
1687         }
1688         return 0;
1689 }
1690
1691 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1692 {
1693         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1694                            vhost_avail_event(vq)))
1695                 return -EFAULT;
1696         if (unlikely(vq->log_used)) {
1697                 void __user *used;
1698                 /* Make sure the event is seen before log. */
1699                 smp_wmb();
1700                 /* Log avail event write */
1701                 used = vhost_avail_event(vq);
1702                 log_write(vq->log_base, vq->log_addr +
1703                           (used - (void __user *)vq->used),
1704                           sizeof *vhost_avail_event(vq));
1705                 if (vq->log_ctx)
1706                         eventfd_signal(vq->log_ctx, 1);
1707         }
1708         return 0;
1709 }
1710
1711 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1712 {
1713         __virtio16 last_used_idx;
1714         int r;
1715         bool is_le = vq->is_le;
1716
1717         if (!vq->private_data) {
1718                 vhost_reset_is_le(vq);
1719                 return 0;
1720         }
1721
1722         vhost_init_is_le(vq);
1723
1724         r = vhost_update_used_flags(vq);
1725         if (r)
1726                 goto err;
1727         vq->signalled_used_valid = false;
1728         if (!vq->iotlb &&
1729             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1730                 r = -EFAULT;
1731                 goto err;
1732         }
1733         r = vhost_get_user(vq, last_used_idx, &vq->used->idx);
1734         if (r) {
1735                 vq_err(vq, "Can't access used idx at %p\n",
1736                        &vq->used->idx);
1737                 goto err;
1738         }
1739         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1740         return 0;
1741
1742 err:
1743         vq->is_le = is_le;
1744         return r;
1745 }
1746 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1747
1748 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1749                           struct iovec iov[], int iov_size, int access)
1750 {
1751         const struct vhost_umem_node *node;
1752         struct vhost_dev *dev = vq->dev;
1753         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1754         struct iovec *_iov;
1755         u64 s = 0;
1756         int ret = 0;
1757
1758         while ((u64)len > s) {
1759                 u64 size;
1760                 if (unlikely(ret >= iov_size)) {
1761                         ret = -ENOBUFS;
1762                         break;
1763                 }
1764
1765                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1766                                                         addr, addr + len - 1);
1767                 if (node == NULL || node->start > addr) {
1768                         if (umem != dev->iotlb) {
1769                                 ret = -EFAULT;
1770                                 break;
1771                         }
1772                         ret = -EAGAIN;
1773                         break;
1774                 } else if (!(node->perm & access)) {
1775                         ret = -EPERM;
1776                         break;
1777                 }
1778
1779                 _iov = iov + ret;
1780                 size = node->size - addr + node->start;
1781                 _iov->iov_len = min((u64)len - s, size);
1782                 _iov->iov_base = (void __user *)(unsigned long)
1783                         (node->userspace_addr + addr - node->start);
1784                 s += size;
1785                 addr += size;
1786                 ++ret;
1787         }
1788
1789         if (ret == -EAGAIN)
1790                 vhost_iotlb_miss(vq, addr, access);
1791         return ret;
1792 }
1793
1794 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1795  * function returns the next descriptor in the chain,
1796  * or -1U if we're at the end. */
1797 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1798 {
1799         unsigned int next;
1800
1801         /* If this descriptor says it doesn't chain, we're done. */
1802         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1803                 return -1U;
1804
1805         /* Check they're not leading us off end of descriptors. */
1806         next = vhost16_to_cpu(vq, desc->next);
1807         /* Make sure compiler knows to grab that: we don't want it changing! */
1808         /* We will use the result as an index in an array, so most
1809          * architectures only need a compiler barrier here. */
1810         read_barrier_depends();
1811
1812         return next;
1813 }
1814
1815 static int get_indirect(struct vhost_virtqueue *vq,
1816                         struct iovec iov[], unsigned int iov_size,
1817                         unsigned int *out_num, unsigned int *in_num,
1818                         struct vhost_log *log, unsigned int *log_num,
1819                         struct vring_desc *indirect)
1820 {
1821         struct vring_desc desc;
1822         unsigned int i = 0, count, found = 0;
1823         u32 len = vhost32_to_cpu(vq, indirect->len);
1824         struct iov_iter from;
1825         int ret, access;
1826
1827         /* Sanity check */
1828         if (unlikely(len % sizeof desc)) {
1829                 vq_err(vq, "Invalid length in indirect descriptor: "
1830                        "len 0x%llx not multiple of 0x%zx\n",
1831                        (unsigned long long)len,
1832                        sizeof desc);
1833                 return -EINVAL;
1834         }
1835
1836         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1837                              UIO_MAXIOV, VHOST_ACCESS_RO);
1838         if (unlikely(ret < 0)) {
1839                 if (ret != -EAGAIN)
1840                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1841                 return ret;
1842         }
1843         iov_iter_init(&from, READ, vq->indirect, ret, len);
1844
1845         /* We will use the result as an address to read from, so most
1846          * architectures only need a compiler barrier here. */
1847         read_barrier_depends();
1848
1849         count = len / sizeof desc;
1850         /* Buffers are chained via a 16 bit next field, so
1851          * we can have at most 2^16 of these. */
1852         if (unlikely(count > USHRT_MAX + 1)) {
1853                 vq_err(vq, "Indirect buffer length too big: %d\n",
1854                        indirect->len);
1855                 return -E2BIG;
1856         }
1857
1858         do {
1859                 unsigned iov_count = *in_num + *out_num;
1860                 if (unlikely(++found > count)) {
1861                         vq_err(vq, "Loop detected: last one at %u "
1862                                "indirect size %u\n",
1863                                i, count);
1864                         return -EINVAL;
1865                 }
1866                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1867                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1868                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1869                         return -EINVAL;
1870                 }
1871                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1872                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1873                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1874                         return -EINVAL;
1875                 }
1876
1877                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1878                         access = VHOST_ACCESS_WO;
1879                 else
1880                         access = VHOST_ACCESS_RO;
1881
1882                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1883                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1884                                      iov_size - iov_count, access);
1885                 if (unlikely(ret < 0)) {
1886                         if (ret != -EAGAIN)
1887                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1888                                         ret, i);
1889                         return ret;
1890                 }
1891                 /* If this is an input descriptor, increment that count. */
1892                 if (access == VHOST_ACCESS_WO) {
1893                         *in_num += ret;
1894                         if (unlikely(log)) {
1895                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1896                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1897                                 ++*log_num;
1898                         }
1899                 } else {
1900                         /* If it's an output descriptor, they're all supposed
1901                          * to come before any input descriptors. */
1902                         if (unlikely(*in_num)) {
1903                                 vq_err(vq, "Indirect descriptor "
1904                                        "has out after in: idx %d\n", i);
1905                                 return -EINVAL;
1906                         }
1907                         *out_num += ret;
1908                 }
1909         } while ((i = next_desc(vq, &desc)) != -1);
1910         return 0;
1911 }
1912
1913 /* This looks in the virtqueue and for the first available buffer, and converts
1914  * it to an iovec for convenient access.  Since descriptors consist of some
1915  * number of output then some number of input descriptors, it's actually two
1916  * iovecs, but we pack them into one and note how many of each there were.
1917  *
1918  * This function returns the descriptor number found, or vq->num (which is
1919  * never a valid descriptor number) if none was found.  A negative code is
1920  * returned on error. */
1921 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1922                       struct iovec iov[], unsigned int iov_size,
1923                       unsigned int *out_num, unsigned int *in_num,
1924                       struct vhost_log *log, unsigned int *log_num)
1925 {
1926         struct vring_desc desc;
1927         unsigned int i, head, found = 0;
1928         u16 last_avail_idx;
1929         __virtio16 avail_idx;
1930         __virtio16 ring_head;
1931         int ret, access;
1932
1933         /* Check it isn't doing very strange things with descriptor numbers. */
1934         last_avail_idx = vq->last_avail_idx;
1935         if (unlikely(vhost_get_user(vq, avail_idx, &vq->avail->idx))) {
1936                 vq_err(vq, "Failed to access avail idx at %p\n",
1937                        &vq->avail->idx);
1938                 return -EFAULT;
1939         }
1940         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1941
1942         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1943                 vq_err(vq, "Guest moved used index from %u to %u",
1944                        last_avail_idx, vq->avail_idx);
1945                 return -EFAULT;
1946         }
1947
1948         /* If there's nothing new since last we looked, return invalid. */
1949         if (vq->avail_idx == last_avail_idx)
1950                 return vq->num;
1951
1952         /* Only get avail ring entries after they have been exposed by guest. */
1953         smp_rmb();
1954
1955         /* Grab the next descriptor number they're advertising, and increment
1956          * the index we've seen. */
1957         if (unlikely(vhost_get_user(vq, ring_head,
1958                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
1959                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1960                        last_avail_idx,
1961                        &vq->avail->ring[last_avail_idx % vq->num]);
1962                 return -EFAULT;
1963         }
1964
1965         head = vhost16_to_cpu(vq, ring_head);
1966
1967         /* If their number is silly, that's an error. */
1968         if (unlikely(head >= vq->num)) {
1969                 vq_err(vq, "Guest says index %u > %u is available",
1970                        head, vq->num);
1971                 return -EINVAL;
1972         }
1973
1974         /* When we start there are none of either input nor output. */
1975         *out_num = *in_num = 0;
1976         if (unlikely(log))
1977                 *log_num = 0;
1978
1979         i = head;
1980         do {
1981                 unsigned iov_count = *in_num + *out_num;
1982                 if (unlikely(i >= vq->num)) {
1983                         vq_err(vq, "Desc index is %u > %u, head = %u",
1984                                i, vq->num, head);
1985                         return -EINVAL;
1986                 }
1987                 if (unlikely(++found > vq->num)) {
1988                         vq_err(vq, "Loop detected: last one at %u "
1989                                "vq size %u head %u\n",
1990                                i, vq->num, head);
1991                         return -EINVAL;
1992                 }
1993                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
1994                                            sizeof desc);
1995                 if (unlikely(ret)) {
1996                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1997                                i, vq->desc + i);
1998                         return -EFAULT;
1999                 }
2000                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2001                         ret = get_indirect(vq, iov, iov_size,
2002                                            out_num, in_num,
2003                                            log, log_num, &desc);
2004                         if (unlikely(ret < 0)) {
2005                                 if (ret != -EAGAIN)
2006                                         vq_err(vq, "Failure detected "
2007                                                 "in indirect descriptor at idx %d\n", i);
2008                                 return ret;
2009                         }
2010                         continue;
2011                 }
2012
2013                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2014                         access = VHOST_ACCESS_WO;
2015                 else
2016                         access = VHOST_ACCESS_RO;
2017                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2018                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2019                                      iov_size - iov_count, access);
2020                 if (unlikely(ret < 0)) {
2021                         if (ret != -EAGAIN)
2022                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2023                                         ret, i);
2024                         return ret;
2025                 }
2026                 if (access == VHOST_ACCESS_WO) {
2027                         /* If this is an input descriptor,
2028                          * increment that count. */
2029                         *in_num += ret;
2030                         if (unlikely(log)) {
2031                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2032                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2033                                 ++*log_num;
2034                         }
2035                 } else {
2036                         /* If it's an output descriptor, they're all supposed
2037                          * to come before any input descriptors. */
2038                         if (unlikely(*in_num)) {
2039                                 vq_err(vq, "Descriptor has out after in: "
2040                                        "idx %d\n", i);
2041                                 return -EINVAL;
2042                         }
2043                         *out_num += ret;
2044                 }
2045         } while ((i = next_desc(vq, &desc)) != -1);
2046
2047         /* On success, increment avail index. */
2048         vq->last_avail_idx++;
2049
2050         /* Assume notifications from guest are disabled at this point,
2051          * if they aren't we would need to update avail_event index. */
2052         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2053         return head;
2054 }
2055 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2056
2057 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2058 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2059 {
2060         vq->last_avail_idx -= n;
2061 }
2062 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2063
2064 /* After we've used one of their buffers, we tell them about it.  We'll then
2065  * want to notify the guest, using eventfd. */
2066 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2067 {
2068         struct vring_used_elem heads = {
2069                 cpu_to_vhost32(vq, head),
2070                 cpu_to_vhost32(vq, len)
2071         };
2072
2073         return vhost_add_used_n(vq, &heads, 1);
2074 }
2075 EXPORT_SYMBOL_GPL(vhost_add_used);
2076
2077 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2078                             struct vring_used_elem *heads,
2079                             unsigned count)
2080 {
2081         struct vring_used_elem __user *used;
2082         u16 old, new;
2083         int start;
2084
2085         start = vq->last_used_idx & (vq->num - 1);
2086         used = vq->used->ring + start;
2087         if (count == 1) {
2088                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2089                         vq_err(vq, "Failed to write used id");
2090                         return -EFAULT;
2091                 }
2092                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2093                         vq_err(vq, "Failed to write used len");
2094                         return -EFAULT;
2095                 }
2096         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2097                 vq_err(vq, "Failed to write used");
2098                 return -EFAULT;
2099         }
2100         if (unlikely(vq->log_used)) {
2101                 /* Make sure data is seen before log. */
2102                 smp_wmb();
2103                 /* Log used ring entry write. */
2104                 log_write(vq->log_base,
2105                           vq->log_addr +
2106                            ((void __user *)used - (void __user *)vq->used),
2107                           count * sizeof *used);
2108         }
2109         old = vq->last_used_idx;
2110         new = (vq->last_used_idx += count);
2111         /* If the driver never bothers to signal in a very long while,
2112          * used index might wrap around. If that happens, invalidate
2113          * signalled_used index we stored. TODO: make sure driver
2114          * signals at least once in 2^16 and remove this. */
2115         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2116                 vq->signalled_used_valid = false;
2117         return 0;
2118 }
2119
2120 /* After we've used one of their buffers, we tell them about it.  We'll then
2121  * want to notify the guest, using eventfd. */
2122 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2123                      unsigned count)
2124 {
2125         int start, n, r;
2126
2127         start = vq->last_used_idx & (vq->num - 1);
2128         n = vq->num - start;
2129         if (n < count) {
2130                 r = __vhost_add_used_n(vq, heads, n);
2131                 if (r < 0)
2132                         return r;
2133                 heads += n;
2134                 count -= n;
2135         }
2136         r = __vhost_add_used_n(vq, heads, count);
2137
2138         /* Make sure buffer is written before we update index. */
2139         smp_wmb();
2140         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2141                            &vq->used->idx)) {
2142                 vq_err(vq, "Failed to increment used idx");
2143                 return -EFAULT;
2144         }
2145         if (unlikely(vq->log_used)) {
2146                 /* Log used index update. */
2147                 log_write(vq->log_base,
2148                           vq->log_addr + offsetof(struct vring_used, idx),
2149                           sizeof vq->used->idx);
2150                 if (vq->log_ctx)
2151                         eventfd_signal(vq->log_ctx, 1);
2152         }
2153         return r;
2154 }
2155 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2156
2157 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2158 {
2159         __u16 old, new;
2160         __virtio16 event;
2161         bool v;
2162
2163         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2164             unlikely(vq->avail_idx == vq->last_avail_idx))
2165                 return true;
2166
2167         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2168                 __virtio16 flags;
2169                 /* Flush out used index updates. This is paired
2170                  * with the barrier that the Guest executes when enabling
2171                  * interrupts. */
2172                 smp_mb();
2173                 if (vhost_get_user(vq, flags, &vq->avail->flags)) {
2174                         vq_err(vq, "Failed to get flags");
2175                         return true;
2176                 }
2177                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2178         }
2179         old = vq->signalled_used;
2180         v = vq->signalled_used_valid;
2181         new = vq->signalled_used = vq->last_used_idx;
2182         vq->signalled_used_valid = true;
2183
2184         if (unlikely(!v))
2185                 return true;
2186
2187         /* We're sure if the following conditions are met, there's no
2188          * need to notify guest:
2189          * 1) cached used event is ahead of new
2190          * 2) old to new updating does not cross cached used event. */
2191         if (vring_need_event(vq->last_used_event, new + vq->num, new) &&
2192             !vring_need_event(vq->last_used_event, new, old))
2193                 return false;
2194
2195         /* Flush out used index updates. This is paired
2196          * with the barrier that the Guest executes when enabling
2197          * interrupts. */
2198         smp_mb();
2199
2200         if (vhost_get_user(vq, event, vhost_used_event(vq))) {
2201                 vq_err(vq, "Failed to get used event idx");
2202                 return true;
2203         }
2204         vq->last_used_event = vhost16_to_cpu(vq, event);
2205
2206         return vring_need_event(vq->last_used_event, new, old);
2207 }
2208
2209 /* This actually signals the guest, using eventfd. */
2210 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2211 {
2212         /* Signal the Guest tell them we used something up. */
2213         if (vq->call_ctx && vhost_notify(dev, vq))
2214                 eventfd_signal(vq->call_ctx, 1);
2215 }
2216 EXPORT_SYMBOL_GPL(vhost_signal);
2217
2218 /* And here's the combo meal deal.  Supersize me! */
2219 void vhost_add_used_and_signal(struct vhost_dev *dev,
2220                                struct vhost_virtqueue *vq,
2221                                unsigned int head, int len)
2222 {
2223         vhost_add_used(vq, head, len);
2224         vhost_signal(dev, vq);
2225 }
2226 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2227
2228 /* multi-buffer version of vhost_add_used_and_signal */
2229 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2230                                  struct vhost_virtqueue *vq,
2231                                  struct vring_used_elem *heads, unsigned count)
2232 {
2233         vhost_add_used_n(vq, heads, count);
2234         vhost_signal(dev, vq);
2235 }
2236 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2237
2238 /* return true if we're sure that avaiable ring is empty */
2239 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2240 {
2241         __virtio16 avail_idx;
2242         int r;
2243
2244         r = vhost_get_user(vq, avail_idx, &vq->avail->idx);
2245         if (r)
2246                 return false;
2247
2248         return vhost16_to_cpu(vq, avail_idx) == vq->avail_idx;
2249 }
2250 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2251
2252 /* OK, now we need to know about added descriptors. */
2253 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2254 {
2255         __virtio16 avail_idx;
2256         int r;
2257
2258         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2259                 return false;
2260         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2261         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2262                 r = vhost_update_used_flags(vq);
2263                 if (r) {
2264                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2265                                &vq->used->flags, r);
2266                         return false;
2267                 }
2268         } else {
2269                 r = vhost_update_avail_event(vq, vq->avail_idx);
2270                 if (r) {
2271                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2272                                vhost_avail_event(vq), r);
2273                         return false;
2274                 }
2275         }
2276         /* They could have slipped one in as we were doing that: make
2277          * sure it's written, then check again. */
2278         smp_mb();
2279         r = vhost_get_user(vq, avail_idx, &vq->avail->idx);
2280         if (r) {
2281                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2282                        &vq->avail->idx, r);
2283                 return false;
2284         }
2285
2286         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2287 }
2288 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2289
2290 /* We don't need to be notified again. */
2291 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2292 {
2293         int r;
2294
2295         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2296                 return;
2297         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2298         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2299                 r = vhost_update_used_flags(vq);
2300                 if (r)
2301                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2302                                &vq->used->flags, r);
2303         }
2304 }
2305 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2306
2307 /* Create a new message. */
2308 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2309 {
2310         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2311         if (!node)
2312                 return NULL;
2313         node->vq = vq;
2314         node->msg.type = type;
2315         return node;
2316 }
2317 EXPORT_SYMBOL_GPL(vhost_new_msg);
2318
2319 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2320                        struct vhost_msg_node *node)
2321 {
2322         spin_lock(&dev->iotlb_lock);
2323         list_add_tail(&node->node, head);
2324         spin_unlock(&dev->iotlb_lock);
2325
2326         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2327 }
2328 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2329
2330 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2331                                          struct list_head *head)
2332 {
2333         struct vhost_msg_node *node = NULL;
2334
2335         spin_lock(&dev->iotlb_lock);
2336         if (!list_empty(head)) {
2337                 node = list_first_entry(head, struct vhost_msg_node,
2338                                         node);
2339                 list_del(&node->node);
2340         }
2341         spin_unlock(&dev->iotlb_lock);
2342
2343         return node;
2344 }
2345 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2346
2347
2348 static int __init vhost_init(void)
2349 {
2350         return 0;
2351 }
2352
2353 static void __exit vhost_exit(void)
2354 {
2355 }
2356
2357 module_init(vhost_init);
2358 module_exit(vhost_exit);
2359
2360 MODULE_VERSION("0.0.1");
2361 MODULE_LICENSE("GPL v2");
2362 MODULE_AUTHOR("Michael S. Tsirkin");
2363 MODULE_DESCRIPTION("Host kernel accelerator for virtio");