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/sched/mm.h>
  31 #include <linux/sched/signal.h>
  32 #include <linux/interval_tree_generic.h>
  33
  34 #include "vhost.h"
  35
  36 static ushort max_mem_regions = 64;
  37 module_param(max_mem_regions, ushort, 0444);
  38 MODULE_PARM_DESC(max_mem_regions,
  39         "Maximum number of memory regions in memory map. (default: 64)");
  40 static int max_iotlb_entries = 2048;
  41 module_param(max_iotlb_entries, int, 0444);
  42 MODULE_PARM_DESC(max_iotlb_entries,
  43         "Maximum number of iotlb entries. (default: 2048)");
  44
  45 enum {
  46         VHOST_MEMORY_F_LOG = 0x1,
  47 };
  48
  49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  53                      rb, __u64, __subtree_last,
  54                      START, LAST, static inline, vhost_umem_interval_tree);
  55
  56 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  57 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  58 {
  59         vq->user_be = !virtio_legacy_is_little_endian();
  60 }
  61
  62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  63 {
  64         vq->user_be = true;
  65 }
  66
  67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  68 {
  69         vq->user_be = false;
  70 }
  71
  72 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  73 {
  74         struct vhost_vring_state s;
  75
  76         if (vq->private_data)
  77                 return -EBUSY;
  78
  79         if (copy_from_user(&s, argp, sizeof(s)))
  80                 return -EFAULT;
  81
  82         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  83             s.num != VHOST_VRING_BIG_ENDIAN)
  84                 return -EINVAL;
  85
  86         if (s.num == VHOST_VRING_BIG_ENDIAN)
  87                 vhost_enable_cross_endian_big(vq);
  88         else
  89                 vhost_enable_cross_endian_little(vq);
  90
  91         return 0;
  92 }
  93
  94 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  95                                    int __user *argp)
  96 {
  97         struct vhost_vring_state s = {
  98                 .index = idx,
  99                 .num = vq->user_be
 100         };
 101
 102         if (copy_to_user(argp, &s, sizeof(s)))
 103                 return -EFAULT;
 104
 105         return 0;
 106 }
 107
 108 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 109 {
 110         /* Note for legacy virtio: user_be is initialized at reset time
 111          * according to the host endianness. If userspace does not set an
 112          * explicit endianness, the default behavior is native endian, as
 113          * expected by legacy virtio.
 114          */
 115         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 116 }
 117 #else
 118 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 119 {
 120 }
 121
 122 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 123 {
 124         return -ENOIOCTLCMD;
 125 }
 126
 127 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 128                                    int __user *argp)
 129 {
 130         return -ENOIOCTLCMD;
 131 }
 132
 133 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 134 {
 135         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 136                 || virtio_legacy_is_little_endian();
 137 }
 138 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 139
 140 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 141 {
 142         vhost_init_is_le(vq);
 143 }
 144
 145 struct vhost_flush_struct {
 146         struct vhost_work work;
 147         struct completion wait_event;
 148 };
 149
 150 static void vhost_flush_work(struct vhost_work *work)
 151 {
 152         struct vhost_flush_struct *s;
 153
 154         s = container_of(work, struct vhost_flush_struct, work);
 155         complete(&s->wait_event);
 156 }
 157
 158 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 159                             poll_table *pt)
 160 {
 161         struct vhost_poll *poll;
 162
 163         poll = container_of(pt, struct vhost_poll, table);
 164         poll->wqh = wqh;
 165         add_wait_queue(wqh, &poll->wait);
 166 }
 167
 168 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 169                              void *key)
 170 {
 171         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 172
 173         if (!(key_to_poll(key) & poll->mask))
 174                 return 0;
 175
 176         vhost_poll_queue(poll);
 177         return 0;
 178 }
 179
 180 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 181 {
 182         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 183         work->fn = fn;
 184 }
 185 EXPORT_SYMBOL_GPL(vhost_work_init);
 186
 187 /* Init poll structure */
 188 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 189                      __poll_t mask, struct vhost_dev *dev)
 190 {
 191         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 192         init_poll_funcptr(&poll->table, vhost_poll_func);
 193         poll->mask = mask;
 194         poll->dev = dev;
 195         poll->wqh = NULL;
 196
 197         vhost_work_init(&poll->work, fn);
 198 }
 199 EXPORT_SYMBOL_GPL(vhost_poll_init);
 200
 201 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 202  * keep a reference to a file until after vhost_poll_stop is called. */
 203 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 204 {
 205         __poll_t mask;
 206         int ret = 0;
 207
 208         if (poll->wqh)
 209                 return 0;
 210
 211         mask = file->f_op->poll(file, &poll->table);
 212         if (mask)
 213                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 214         if (mask & EPOLLERR) {
 215                 vhost_poll_stop(poll);
 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_meta_reset(struct vhost_virtqueue *vq)
 286 {
 287         int j;
 288
 289         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 290                 vq->meta_iotlb[j] = NULL;
 291 }
 292
 293 static void vhost_vq_meta_reset(struct vhost_dev *d)
 294 {
 295         int i;
 296
 297         for (i = 0; i < d->nvqs; ++i)
 298                 __vhost_vq_meta_reset(d->vqs[i]);
 299 }
 300
 301 static void vhost_vq_reset(struct vhost_dev *dev,
 302                            struct vhost_virtqueue *vq)
 303 {
 304         vq->num = 1;
 305         vq->desc = NULL;
 306         vq->avail = NULL;
 307         vq->used = NULL;
 308         vq->last_avail_idx = 0;
 309         vq->avail_idx = 0;
 310         vq->last_used_idx = 0;
 311         vq->signalled_used = 0;
 312         vq->signalled_used_valid = false;
 313         vq->used_flags = 0;
 314         vq->log_used = false;
 315         vq->log_addr = -1ull;
 316         vq->private_data = NULL;
 317         vq->acked_features = 0;
 318         vq->log_base = NULL;
 319         vq->error_ctx = NULL;
 320         vq->kick = NULL;
 321         vq->call_ctx = NULL;
 322         vq->log_ctx = NULL;
 323         vhost_reset_is_le(vq);
 324         vhost_disable_cross_endian(vq);
 325         vq->busyloop_timeout = 0;
 326         vq->umem = NULL;
 327         vq->iotlb = NULL;
 328         __vhost_vq_meta_reset(vq);
 329 }
 330
 331 static int vhost_worker(void *data)
 332 {
 333         struct vhost_dev *dev = data;
 334         struct vhost_work *work, *work_next;
 335         struct llist_node *node;
 336         mm_segment_t oldfs = get_fs();
 337
 338         set_fs(USER_DS);
 339         use_mm(dev->mm);
 340
 341         for (;;) {
 342                 /* mb paired w/ kthread_stop */
 343                 set_current_state(TASK_INTERRUPTIBLE);
 344
 345                 if (kthread_should_stop()) {
 346                         __set_current_state(TASK_RUNNING);
 347                         break;
 348                 }
 349
 350                 node = llist_del_all(&dev->work_list);
 351                 if (!node)
 352                         schedule();
 353
 354                 node = llist_reverse_order(node);
 355                 /* make sure flag is seen after deletion */
 356                 smp_wmb();
 357                 llist_for_each_entry_safe(work, work_next, node, node) {
 358                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 359                         __set_current_state(TASK_RUNNING);
 360                         work->fn(work);
 361                         if (need_resched())
 362                                 schedule();
 363                 }
 364         }
 365         unuse_mm(dev->mm);
 366         set_fs(oldfs);
 367         return 0;
 368 }
 369
 370 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 371 {
 372         kfree(vq->indirect);
 373         vq->indirect = NULL;
 374         kfree(vq->log);
 375         vq->log = NULL;
 376         kfree(vq->heads);
 377         vq->heads = NULL;
 378 }
 379
 380 /* Helper to allocate iovec buffers for all vqs. */
 381 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 382 {
 383         struct vhost_virtqueue *vq;
 384         int i;
 385
 386         for (i = 0; i < dev->nvqs; ++i) {
 387                 vq = dev->vqs[i];
 388                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 389                                        GFP_KERNEL);
 390                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 391                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 392                 if (!vq->indirect || !vq->log || !vq->heads)
 393                         goto err_nomem;
 394         }
 395         return 0;
 396
 397 err_nomem:
 398         for (; i >= 0; --i)
 399                 vhost_vq_free_iovecs(dev->vqs[i]);
 400         return -ENOMEM;
 401 }
 402
 403 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 404 {
 405         int i;
 406
 407         for (i = 0; i < dev->nvqs; ++i)
 408                 vhost_vq_free_iovecs(dev->vqs[i]);
 409 }
 410
 411 void vhost_dev_init(struct vhost_dev *dev,
 412                     struct vhost_virtqueue **vqs, int nvqs)
 413 {
 414         struct vhost_virtqueue *vq;
 415         int i;
 416
 417         dev->vqs = vqs;
 418         dev->nvqs = nvqs;
 419         mutex_init(&dev->mutex);
 420         dev->log_ctx = NULL;
 421         dev->umem = NULL;
 422         dev->iotlb = NULL;
 423         dev->mm = NULL;
 424         dev->worker = NULL;
 425         init_llist_head(&dev->work_list);
 426         init_waitqueue_head(&dev->wait);
 427         INIT_LIST_HEAD(&dev->read_list);
 428         INIT_LIST_HEAD(&dev->pending_list);
 429         spin_lock_init(&dev->iotlb_lock);
 430
 431
 432         for (i = 0; i < dev->nvqs; ++i) {
 433                 vq = dev->vqs[i];
 434                 vq->log = NULL;
 435                 vq->indirect = NULL;
 436                 vq->heads = NULL;
 437                 vq->dev = dev;
 438                 mutex_init(&vq->mutex);
 439                 vhost_vq_reset(dev, vq);
 440                 if (vq->handle_kick)
 441                         vhost_poll_init(&vq->poll, vq->handle_kick,
 442                                         EPOLLIN, dev);
 443         }
 444 }
 445 EXPORT_SYMBOL_GPL(vhost_dev_init);
 446
 447 /* Caller should have device mutex */
 448 long vhost_dev_check_owner(struct vhost_dev *dev)
 449 {
 450         /* Are you the owner? If not, I don't think you mean to do that */
 451         return dev->mm == current->mm ? 0 : -EPERM;
 452 }
 453 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 454
 455 struct vhost_attach_cgroups_struct {
 456         struct vhost_work work;
 457         struct task_struct *owner;
 458         int ret;
 459 };
 460
 461 static void vhost_attach_cgroups_work(struct vhost_work *work)
 462 {
 463         struct vhost_attach_cgroups_struct *s;
 464
 465         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 466         s->ret = cgroup_attach_task_all(s->owner, current);
 467 }
 468
 469 static int vhost_attach_cgroups(struct vhost_dev *dev)
 470 {
 471         struct vhost_attach_cgroups_struct attach;
 472
 473         attach.owner = current;
 474         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 475         vhost_work_queue(dev, &attach.work);
 476         vhost_work_flush(dev, &attach.work);
 477         return attach.ret;
 478 }
 479
 480 /* Caller should have device mutex */
 481 bool vhost_dev_has_owner(struct vhost_dev *dev)
 482 {
 483         return dev->mm;
 484 }
 485 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 486
 487 /* Caller should have device mutex */
 488 long vhost_dev_set_owner(struct vhost_dev *dev)
 489 {
 490         struct task_struct *worker;
 491         int err;
 492
 493         /* Is there an owner already? */
 494         if (vhost_dev_has_owner(dev)) {
 495                 err = -EBUSY;
 496                 goto err_mm;
 497         }
 498
 499         /* No owner, become one */
 500         dev->mm = get_task_mm(current);
 501         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 502         if (IS_ERR(worker)) {
 503                 err = PTR_ERR(worker);
 504                 goto err_worker;
 505         }
 506
 507         dev->worker = worker;
 508         wake_up_process(worker);        /* avoid contributing to loadavg */
 509
 510         err = vhost_attach_cgroups(dev);
 511         if (err)
 512                 goto err_cgroup;
 513
 514         err = vhost_dev_alloc_iovecs(dev);
 515         if (err)
 516                 goto err_cgroup;
 517
 518         return 0;
 519 err_cgroup:
 520         kthread_stop(worker);
 521         dev->worker = NULL;
 522 err_worker:
 523         if (dev->mm)
 524                 mmput(dev->mm);
 525         dev->mm = NULL;
 526 err_mm:
 527         return err;
 528 }
 529 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 530
 531 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 532 {
 533         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
 534 }
 535 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 536
 537 /* Caller should have device mutex */
 538 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 539 {
 540         int i;
 541
 542         vhost_dev_cleanup(dev);
 543
 544         /* Restore memory to default empty mapping. */
 545         INIT_LIST_HEAD(&umem->umem_list);
 546         dev->umem = umem;
 547         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 548          * VQs aren't running.
 549          */
 550         for (i = 0; i < dev->nvqs; ++i)
 551                 dev->vqs[i]->umem = umem;
 552 }
 553 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 554
 555 void vhost_dev_stop(struct vhost_dev *dev)
 556 {
 557         int i;
 558
 559         for (i = 0; i < dev->nvqs; ++i) {
 560                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 561                         vhost_poll_stop(&dev->vqs[i]->poll);
 562                         vhost_poll_flush(&dev->vqs[i]->poll);
 563                 }
 564         }
 565 }
 566 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 567
 568 static void vhost_umem_free(struct vhost_umem *umem,
 569                             struct vhost_umem_node *node)
 570 {
 571         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 572         list_del(&node->link);
 573         kfree(node);
 574         umem->numem--;
 575 }
 576
 577 static void vhost_umem_clean(struct vhost_umem *umem)
 578 {
 579         struct vhost_umem_node *node, *tmp;
 580
 581         if (!umem)
 582                 return;
 583
 584         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 585                 vhost_umem_free(umem, node);
 586
 587         kvfree(umem);
 588 }
 589
 590 static void vhost_clear_msg(struct vhost_dev *dev)
 591 {
 592         struct vhost_msg_node *node, *n;
 593
 594         spin_lock(&dev->iotlb_lock);
 595
 596         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 597                 list_del(&node->node);
 598                 kfree(node);
 599         }
 600
 601         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 602                 list_del(&node->node);
 603                 kfree(node);
 604         }
 605
 606         spin_unlock(&dev->iotlb_lock);
 607 }
 608
 609 void vhost_dev_cleanup(struct vhost_dev *dev)
 610 {
 611         int i;
 612
 613         for (i = 0; i < dev->nvqs; ++i) {
 614                 if (dev->vqs[i]->error_ctx)
 615                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 616                 if (dev->vqs[i]->kick)
 617                         fput(dev->vqs[i]->kick);
 618                 if (dev->vqs[i]->call_ctx)
 619                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 620                 vhost_vq_reset(dev, dev->vqs[i]);
 621         }
 622         vhost_dev_free_iovecs(dev);
 623         if (dev->log_ctx)
 624                 eventfd_ctx_put(dev->log_ctx);
 625         dev->log_ctx = NULL;
 626         /* No one will access memory at this point */
 627         vhost_umem_clean(dev->umem);
 628         dev->umem = NULL;
 629         vhost_umem_clean(dev->iotlb);
 630         dev->iotlb = NULL;
 631         vhost_clear_msg(dev);
 632         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 633         WARN_ON(!llist_empty(&dev->work_list));
 634         if (dev->worker) {
 635                 kthread_stop(dev->worker);
 636                 dev->worker = NULL;
 637         }
 638         if (dev->mm)
 639                 mmput(dev->mm);
 640         dev->mm = NULL;
 641 }
 642 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 643
 644 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 645 {
 646         u64 a = addr / VHOST_PAGE_SIZE / 8;
 647
 648         /* Make sure 64 bit math will not overflow. */
 649         if (a > ULONG_MAX - (unsigned long)log_base ||
 650             a + (unsigned long)log_base > ULONG_MAX)
 651                 return 0;
 652
 653         return access_ok(VERIFY_WRITE, log_base + a,
 654                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 655 }
 656
 657 static bool vhost_overflow(u64 uaddr, u64 size)
 658 {
 659         /* Make sure 64 bit math will not overflow. */
 660         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 661 }
 662
 663 /* Caller should have vq mutex and device mutex. */
 664 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 665                                int log_all)
 666 {
 667         struct vhost_umem_node *node;
 668
 669         if (!umem)
 670                 return 0;
 671
 672         list_for_each_entry(node, &umem->umem_list, link) {
 673                 unsigned long a = node->userspace_addr;
 674
 675                 if (vhost_overflow(node->userspace_addr, node->size))
 676                         return 0;
 677
 678
 679                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 680                                     node->size))
 681                         return 0;
 682                 else if (log_all && !log_access_ok(log_base,
 683                                                    node->start,
 684                                                    node->size))
 685                         return 0;
 686         }
 687         return 1;
 688 }
 689
 690 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 691                                                u64 addr, unsigned int size,
 692                                                int type)
 693 {
 694         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 695
 696         if (!node)
 697                 return NULL;
 698
 699         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 700 }
 701
 702 /* Can we switch to this memory table? */
 703 /* Caller should have device mutex but not vq mutex */
 704 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 705                             int log_all)
 706 {
 707         int i;
 708
 709         for (i = 0; i < d->nvqs; ++i) {
 710                 int ok;
 711                 bool log;
 712
 713                 mutex_lock(&d->vqs[i]->mutex);
 714                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 715                 /* If ring is inactive, will check when it's enabled. */
 716                 if (d->vqs[i]->private_data)
 717                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 718                                                  umem, log);
 719                 else
 720                         ok = 1;
 721                 mutex_unlock(&d->vqs[i]->mutex);
 722                 if (!ok)
 723                         return 0;
 724         }
 725         return 1;
 726 }
 727
 728 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 729                           struct iovec iov[], int iov_size, int access);
 730
 731 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 732                               const void *from, unsigned size)
 733 {
 734         int ret;
 735
 736         if (!vq->iotlb)
 737                 return __copy_to_user(to, from, size);
 738         else {
 739                 /* This function should be called after iotlb
 740                  * prefetch, which means we're sure that all vq
 741                  * could be access through iotlb. So -EAGAIN should
 742                  * not happen in this case.
 743                  */
 744                 struct iov_iter t;
 745                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 746                                      (u64)(uintptr_t)to, size,
 747                                      VHOST_ADDR_DESC);
 748
 749                 if (uaddr)
 750                         return __copy_to_user(uaddr, from, size);
 751
 752                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 753                                      ARRAY_SIZE(vq->iotlb_iov),
 754                                      VHOST_ACCESS_WO);
 755                 if (ret < 0)
 756                         goto out;
 757                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 758                 ret = copy_to_iter(from, size, &t);
 759                 if (ret == size)
 760                         ret = 0;
 761         }
 762 out:
 763         return ret;
 764 }
 765
 766 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 767                                 void __user *from, unsigned size)
 768 {
 769         int ret;
 770
 771         if (!vq->iotlb)
 772                 return __copy_from_user(to, from, size);
 773         else {
 774                 /* This function should be called after iotlb
 775                  * prefetch, which means we're sure that vq
 776                  * could be access through iotlb. So -EAGAIN should
 777                  * not happen in this case.
 778                  */
 779                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 780                                      (u64)(uintptr_t)from, size,
 781                                      VHOST_ADDR_DESC);
 782                 struct iov_iter f;
 783
 784                 if (uaddr)
 785                         return __copy_from_user(to, uaddr, size);
 786
 787                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 788                                      ARRAY_SIZE(vq->iotlb_iov),
 789                                      VHOST_ACCESS_RO);
 790                 if (ret < 0) {
 791                         vq_err(vq, "IOTLB translation failure: uaddr "
 792                                "%p size 0x%llx\n", from,
 793                                (unsigned long long) size);
 794                         goto out;
 795                 }
 796                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 797                 ret = copy_from_iter(to, size, &f);
 798                 if (ret == size)
 799                         ret = 0;
 800         }
 801
 802 out:
 803         return ret;
 804 }
 805
 806 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 807                                           void __user *addr, unsigned int size,
 808                                           int type)
 809 {
 810         int ret;
 811
 812         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 813                              ARRAY_SIZE(vq->iotlb_iov),
 814                              VHOST_ACCESS_RO);
 815         if (ret < 0) {
 816                 vq_err(vq, "IOTLB translation failure: uaddr "
 817                         "%p size 0x%llx\n", addr,
 818                         (unsigned long long) size);
 819                 return NULL;
 820         }
 821
 822         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 823                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 824                         "%p size 0x%llx\n", addr,
 825                         (unsigned long long) size);
 826                 return NULL;
 827         }
 828
 829         return vq->iotlb_iov[0].iov_base;
 830 }
 831
 832 /* This function should be called after iotlb
 833  * prefetch, which means we're sure that vq
 834  * could be access through iotlb. So -EAGAIN should
 835  * not happen in this case.
 836  */
 837 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 838                                             void *addr, unsigned int size,
 839                                             int type)
 840 {
 841         void __user *uaddr = vhost_vq_meta_fetch(vq,
 842                              (u64)(uintptr_t)addr, size, type);
 843         if (uaddr)
 844                 return uaddr;
 845
 846         return __vhost_get_user_slow(vq, addr, size, type);
 847 }
 848
 849 #define vhost_put_user(vq, x, ptr)              \
 850 ({ \
 851         int ret = -EFAULT; \
 852         if (!vq->iotlb) { \
 853                 ret = __put_user(x, ptr); \
 854         } else { \
 855                 __typeof__(ptr) to = \
 856                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 857                                           sizeof(*ptr), VHOST_ADDR_USED); \
 858                 if (to != NULL) \
 859                         ret = __put_user(x, to); \
 860                 else \
 861                         ret = -EFAULT;  \
 862         } \
 863         ret; \
 864 })
 865
 866 #define vhost_get_user(vq, x, ptr, type)                \
 867 ({ \
 868         int ret; \
 869         if (!vq->iotlb) { \
 870                 ret = __get_user(x, ptr); \
 871         } else { \
 872                 __typeof__(ptr) from = \
 873                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 874                                                            sizeof(*ptr), \
 875                                                            type); \
 876                 if (from != NULL) \
 877                         ret = __get_user(x, from); \
 878                 else \
 879                         ret = -EFAULT; \
 880         } \
 881         ret; \
 882 })
 883
 884 #define vhost_get_avail(vq, x, ptr) \
 885         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 886
 887 #define vhost_get_used(vq, x, ptr) \
 888         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 889
 890 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 891 {
 892         int i = 0;
 893         for (i = 0; i < d->nvqs; ++i)
 894                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 895 }
 896
 897 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 898 {
 899         int i = 0;
 900         for (i = 0; i < d->nvqs; ++i)
 901                 mutex_unlock(&d->vqs[i]->mutex);
 902 }
 903
 904 static int vhost_new_umem_range(struct vhost_umem *umem,
 905                                 u64 start, u64 size, u64 end,
 906                                 u64 userspace_addr, int perm)
 907 {
 908         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
 909
 910         if (!node)
 911                 return -ENOMEM;
 912
 913         if (umem->numem == max_iotlb_entries) {
 914                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 915                 vhost_umem_free(umem, tmp);
 916         }
 917
 918         node->start = start;
 919         node->size = size;
 920         node->last = end;
 921         node->userspace_addr = userspace_addr;
 922         node->perm = perm;
 923         INIT_LIST_HEAD(&node->link);
 924         list_add_tail(&node->link, &umem->umem_list);
 925         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 926         umem->numem++;
 927
 928         return 0;
 929 }
 930
 931 static void vhost_del_umem_range(struct vhost_umem *umem,
 932                                  u64 start, u64 end)
 933 {
 934         struct vhost_umem_node *node;
 935
 936         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 937                                                            start, end)))
 938                 vhost_umem_free(umem, node);
 939 }
 940
 941 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 942                                   struct vhost_iotlb_msg *msg)
 943 {
 944         struct vhost_msg_node *node, *n;
 945
 946         spin_lock(&d->iotlb_lock);
 947
 948         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 949                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 950                 if (msg->iova <= vq_msg->iova &&
 951                     msg->iova + msg->size - 1 > vq_msg->iova &&
 952                     vq_msg->type == VHOST_IOTLB_MISS) {
 953                         vhost_poll_queue(&node->vq->poll);
 954                         list_del(&node->node);
 955                         kfree(node);
 956                 }
 957         }
 958
 959         spin_unlock(&d->iotlb_lock);
 960 }
 961
 962 static int umem_access_ok(u64 uaddr, u64 size, int access)
 963 {
 964         unsigned long a = uaddr;
 965
 966         /* Make sure 64 bit math will not overflow. */
 967         if (vhost_overflow(uaddr, size))
 968                 return -EFAULT;
 969
 970         if ((access & VHOST_ACCESS_RO) &&
 971             !access_ok(VERIFY_READ, (void __user *)a, size))
 972                 return -EFAULT;
 973         if ((access & VHOST_ACCESS_WO) &&
 974             !access_ok(VERIFY_WRITE, (void __user *)a, size))
 975                 return -EFAULT;
 976         return 0;
 977 }
 978
 979 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
 980                                    struct vhost_iotlb_msg *msg)
 981 {
 982         int ret = 0;
 983
 984         vhost_dev_lock_vqs(dev);
 985         switch (msg->type) {
 986         case VHOST_IOTLB_UPDATE:
 987                 if (!dev->iotlb) {
 988                         ret = -EFAULT;
 989                         break;
 990                 }
 991                 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
 992                         ret = -EFAULT;
 993                         break;
 994                 }
 995                 vhost_vq_meta_reset(dev);
 996                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
 997                                          msg->iova + msg->size - 1,
 998                                          msg->uaddr, msg->perm)) {
 999                         ret = -ENOMEM;
1000                         break;
1001                 }
1002                 vhost_iotlb_notify_vq(dev, msg);
1003                 break;
1004         case VHOST_IOTLB_INVALIDATE:
1005                 if (!dev->iotlb) {
1006                         ret = -EFAULT;
1007                         break;
1008                 }
1009                 vhost_vq_meta_reset(dev);
1010                 vhost_del_umem_range(dev->iotlb, msg->iova,
1011                                      msg->iova + msg->size - 1);
1012                 break;
1013         default:
1014                 ret = -EINVAL;
1015                 break;
1016         }
1017
1018         vhost_dev_unlock_vqs(dev);
1019         return ret;
1020 }
1021 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1022                              struct iov_iter *from)
1023 {
1024         struct vhost_msg_node node;
1025         unsigned size = sizeof(struct vhost_msg);
1026         size_t ret;
1027         int err;
1028
1029         if (iov_iter_count(from) < size)
1030                 return 0;
1031         ret = copy_from_iter(&node.msg, size, from);
1032         if (ret != size)
1033                 goto done;
1034
1035         switch (node.msg.type) {
1036         case VHOST_IOTLB_MSG:
1037                 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1038                 if (err)
1039                         ret = err;
1040                 break;
1041         default:
1042                 ret = -EINVAL;
1043                 break;
1044         }
1045
1046 done:
1047         return ret;
1048 }
1049 EXPORT_SYMBOL(vhost_chr_write_iter);
1050
1051 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1052                             poll_table *wait)
1053 {
1054         __poll_t mask = 0;
1055
1056         poll_wait(file, &dev->wait, wait);
1057
1058         if (!list_empty(&dev->read_list))
1059                 mask |= EPOLLIN | EPOLLRDNORM;
1060
1061         return mask;
1062 }
1063 EXPORT_SYMBOL(vhost_chr_poll);
1064
1065 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1066                             int noblock)
1067 {
1068         DEFINE_WAIT(wait);
1069         struct vhost_msg_node *node;
1070         ssize_t ret = 0;
1071         unsigned size = sizeof(struct vhost_msg);
1072
1073         if (iov_iter_count(to) < size)
1074                 return 0;
1075
1076         while (1) {
1077                 if (!noblock)
1078                         prepare_to_wait(&dev->wait, &wait,
1079                                         TASK_INTERRUPTIBLE);
1080
1081                 node = vhost_dequeue_msg(dev, &dev->read_list);
1082                 if (node)
1083                         break;
1084                 if (noblock) {
1085                         ret = -EAGAIN;
1086                         break;
1087                 }
1088                 if (signal_pending(current)) {
1089                         ret = -ERESTARTSYS;
1090                         break;
1091                 }
1092                 if (!dev->iotlb) {
1093                         ret = -EBADFD;
1094                         break;
1095                 }
1096
1097                 schedule();
1098         }
1099
1100         if (!noblock)
1101                 finish_wait(&dev->wait, &wait);
1102
1103         if (node) {
1104                 ret = copy_to_iter(&node->msg, size, to);
1105
1106                 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1107                         kfree(node);
1108                         return ret;
1109                 }
1110
1111                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1112         }
1113
1114         return ret;
1115 }
1116 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1117
1118 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1119 {
1120         struct vhost_dev *dev = vq->dev;
1121         struct vhost_msg_node *node;
1122         struct vhost_iotlb_msg *msg;
1123
1124         node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1125         if (!node)
1126                 return -ENOMEM;
1127
1128         msg = &node->msg.iotlb;
1129         msg->type = VHOST_IOTLB_MISS;
1130         msg->iova = iova;
1131         msg->perm = access;
1132
1133         vhost_enqueue_msg(dev, &dev->read_list, node);
1134
1135         return 0;
1136 }
1137
1138 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1139                         struct vring_desc __user *desc,
1140                         struct vring_avail __user *avail,
1141                         struct vring_used __user *used)
1142
1143 {
1144         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1145
1146         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1147                access_ok(VERIFY_READ, avail,
1148                          sizeof *avail + num * sizeof *avail->ring + s) &&
1149                access_ok(VERIFY_WRITE, used,
1150                         sizeof *used + num * sizeof *used->ring + s);
1151 }
1152
1153 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1154                                  const struct vhost_umem_node *node,
1155                                  int type)
1156 {
1157         int access = (type == VHOST_ADDR_USED) ?
1158                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1159
1160         if (likely(node->perm & access))
1161                 vq->meta_iotlb[type] = node;
1162 }
1163
1164 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1165                            int access, u64 addr, u64 len, int type)
1166 {
1167         const struct vhost_umem_node *node;
1168         struct vhost_umem *umem = vq->iotlb;
1169         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1170
1171         if (vhost_vq_meta_fetch(vq, addr, len, type))
1172                 return true;
1173
1174         while (len > s) {
1175                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1176                                                            addr,
1177                                                            last);
1178                 if (node == NULL || node->start > addr) {
1179                         vhost_iotlb_miss(vq, addr, access);
1180                         return false;
1181                 } else if (!(node->perm & access)) {
1182                         /* Report the possible access violation by
1183                          * request another translation from userspace.
1184                          */
1185                         return false;
1186                 }
1187
1188                 size = node->size - addr + node->start;
1189
1190                 if (orig_addr == addr && size >= len)
1191                         vhost_vq_meta_update(vq, node, type);
1192
1193                 s += size;
1194                 addr += size;
1195         }
1196
1197         return true;
1198 }
1199
1200 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1201 {
1202         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1203         unsigned int num = vq->num;
1204
1205         if (!vq->iotlb)
1206                 return 1;
1207
1208         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1209                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1210                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1211                                sizeof *vq->avail +
1212                                num * sizeof(*vq->avail->ring) + s,
1213                                VHOST_ADDR_AVAIL) &&
1214                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1215                                sizeof *vq->used +
1216                                num * sizeof(*vq->used->ring) + s,
1217                                VHOST_ADDR_USED);
1218 }
1219 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1220
1221 /* Can we log writes? */
1222 /* Caller should have device mutex but not vq mutex */
1223 int vhost_log_access_ok(struct vhost_dev *dev)
1224 {
1225         return memory_access_ok(dev, dev->umem, 1);
1226 }
1227 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1228
1229 /* Verify access for write logging. */
1230 /* Caller should have vq mutex and device mutex */
1231 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1232                             void __user *log_base)
1233 {
1234         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1235
1236         return vq_memory_access_ok(log_base, vq->umem,
1237                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1238                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1239                                         sizeof *vq->used +
1240                                         vq->num * sizeof *vq->used->ring + s));
1241 }
1242
1243 /* Can we start vq? */
1244 /* Caller should have vq mutex and device mutex */
1245 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1246 {
1247         int ret = vq_log_access_ok(vq, vq->log_base);
1248
1249         if (ret || vq->iotlb)
1250                 return ret;
1251
1252         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1253 }
1254 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1255
1256 static struct vhost_umem *vhost_umem_alloc(void)
1257 {
1258         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1259
1260         if (!umem)
1261                 return NULL;
1262
1263         umem->umem_tree = RB_ROOT_CACHED;
1264         umem->numem = 0;
1265         INIT_LIST_HEAD(&umem->umem_list);
1266
1267         return umem;
1268 }
1269
1270 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1271 {
1272         struct vhost_memory mem, *newmem;
1273         struct vhost_memory_region *region;
1274         struct vhost_umem *newumem, *oldumem;
1275         unsigned long size = offsetof(struct vhost_memory, regions);
1276         int i;
1277
1278         if (copy_from_user(&mem, m, size))
1279                 return -EFAULT;
1280         if (mem.padding)
1281                 return -EOPNOTSUPP;
1282         if (mem.nregions > max_mem_regions)
1283                 return -E2BIG;
1284         newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1285         if (!newmem)
1286                 return -ENOMEM;
1287
1288         memcpy(newmem, &mem, size);
1289         if (copy_from_user(newmem->regions, m->regions,
1290                            mem.nregions * sizeof *m->regions)) {
1291                 kvfree(newmem);
1292                 return -EFAULT;
1293         }
1294
1295         newumem = vhost_umem_alloc();
1296         if (!newumem) {
1297                 kvfree(newmem);
1298                 return -ENOMEM;
1299         }
1300
1301         for (region = newmem->regions;
1302              region < newmem->regions + mem.nregions;
1303              region++) {
1304                 if (vhost_new_umem_range(newumem,
1305                                          region->guest_phys_addr,
1306                                          region->memory_size,
1307                                          region->guest_phys_addr +
1308                                          region->memory_size - 1,
1309                                          region->userspace_addr,
1310                                          VHOST_ACCESS_RW))
1311                         goto err;
1312         }
1313
1314         if (!memory_access_ok(d, newumem, 0))
1315                 goto err;
1316
1317         oldumem = d->umem;
1318         d->umem = newumem;
1319
1320         /* All memory accesses are done under some VQ mutex. */
1321         for (i = 0; i < d->nvqs; ++i) {
1322                 mutex_lock(&d->vqs[i]->mutex);
1323                 d->vqs[i]->umem = newumem;
1324                 mutex_unlock(&d->vqs[i]->mutex);
1325         }
1326
1327         kvfree(newmem);
1328         vhost_umem_clean(oldumem);
1329         return 0;
1330
1331 err:
1332         vhost_umem_clean(newumem);
1333         kvfree(newmem);
1334         return -EFAULT;
1335 }
1336
1337 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1338 {
1339         struct file *eventfp, *filep = NULL;
1340         bool pollstart = false, pollstop = false;
1341         struct eventfd_ctx *ctx = NULL;
1342         u32 __user *idxp = argp;
1343         struct vhost_virtqueue *vq;
1344         struct vhost_vring_state s;
1345         struct vhost_vring_file f;
1346         struct vhost_vring_addr a;
1347         u32 idx;
1348         long r;
1349
1350         r = get_user(idx, idxp);
1351         if (r < 0)
1352                 return r;
1353         if (idx >= d->nvqs)
1354                 return -ENOBUFS;
1355
1356         vq = d->vqs[idx];
1357
1358         mutex_lock(&vq->mutex);
1359
1360         switch (ioctl) {
1361         case VHOST_SET_VRING_NUM:
1362                 /* Resizing ring with an active backend?
1363                  * You don't want to do that. */
1364                 if (vq->private_data) {
1365                         r = -EBUSY;
1366                         break;
1367                 }
1368                 if (copy_from_user(&s, argp, sizeof s)) {
1369                         r = -EFAULT;
1370                         break;
1371                 }
1372                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1373                         r = -EINVAL;
1374                         break;
1375                 }
1376                 vq->num = s.num;
1377                 break;
1378         case VHOST_SET_VRING_BASE:
1379                 /* Moving base with an active backend?
1380                  * You don't want to do that. */
1381                 if (vq->private_data) {
1382                         r = -EBUSY;
1383                         break;
1384                 }
1385                 if (copy_from_user(&s, argp, sizeof s)) {
1386                         r = -EFAULT;
1387                         break;
1388                 }
1389                 if (s.num > 0xffff) {
1390                         r = -EINVAL;
1391                         break;
1392                 }
1393                 vq->last_avail_idx = s.num;
1394                 /* Forget the cached index value. */
1395                 vq->avail_idx = vq->last_avail_idx;
1396                 break;
1397         case VHOST_GET_VRING_BASE:
1398                 s.index = idx;
1399                 s.num = vq->last_avail_idx;
1400                 if (copy_to_user(argp, &s, sizeof s))
1401                         r = -EFAULT;
1402                 break;
1403         case VHOST_SET_VRING_ADDR:
1404                 if (copy_from_user(&a, argp, sizeof a)) {
1405                         r = -EFAULT;
1406                         break;
1407                 }
1408                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1409                         r = -EOPNOTSUPP;
1410                         break;
1411                 }
1412                 /* For 32bit, verify that the top 32bits of the user
1413                    data are set to zero. */
1414                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1415                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1416                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1417                         r = -EFAULT;
1418                         break;
1419                 }
1420
1421                 /* Make sure it's safe to cast pointers to vring types. */
1422                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1423                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1424                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1425                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1426                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1427                         r = -EINVAL;
1428                         break;
1429                 }
1430
1431                 /* We only verify access here if backend is configured.
1432                  * If it is not, we don't as size might not have been setup.
1433                  * We will verify when backend is configured. */
1434                 if (vq->private_data) {
1435                         if (!vq_access_ok(vq, vq->num,
1436                                 (void __user *)(unsigned long)a.desc_user_addr,
1437                                 (void __user *)(unsigned long)a.avail_user_addr,
1438                                 (void __user *)(unsigned long)a.used_user_addr)) {
1439                                 r = -EINVAL;
1440                                 break;
1441                         }
1442
1443                         /* Also validate log access for used ring if enabled. */
1444                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1445                             !log_access_ok(vq->log_base, a.log_guest_addr,
1446                                            sizeof *vq->used +
1447                                            vq->num * sizeof *vq->used->ring)) {
1448                                 r = -EINVAL;
1449                                 break;
1450                         }
1451                 }
1452
1453                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1454                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1455                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1456                 vq->log_addr = a.log_guest_addr;
1457                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1458                 break;
1459         case VHOST_SET_VRING_KICK:
1460                 if (copy_from_user(&f, argp, sizeof f)) {
1461                         r = -EFAULT;
1462                         break;
1463                 }
1464                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1465                 if (IS_ERR(eventfp)) {
1466                         r = PTR_ERR(eventfp);
1467                         break;
1468                 }
1469                 if (eventfp != vq->kick) {
1470                         pollstop = (filep = vq->kick) != NULL;
1471                         pollstart = (vq->kick = eventfp) != NULL;
1472                 } else
1473                         filep = eventfp;
1474                 break;
1475         case VHOST_SET_VRING_CALL:
1476                 if (copy_from_user(&f, argp, sizeof f)) {
1477                         r = -EFAULT;
1478                         break;
1479                 }
1480                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1481                 if (IS_ERR(ctx)) {
1482                         r = PTR_ERR(ctx);
1483                         break;
1484                 }
1485                 swap(ctx, vq->call_ctx);
1486                 break;
1487         case VHOST_SET_VRING_ERR:
1488                 if (copy_from_user(&f, argp, sizeof f)) {
1489                         r = -EFAULT;
1490                         break;
1491                 }
1492                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1493                 if (IS_ERR(ctx)) {
1494                         r = PTR_ERR(ctx);
1495                         break;
1496                 }
1497                 swap(ctx, vq->error_ctx);
1498                 break;
1499         case VHOST_SET_VRING_ENDIAN:
1500                 r = vhost_set_vring_endian(vq, argp);
1501                 break;
1502         case VHOST_GET_VRING_ENDIAN:
1503                 r = vhost_get_vring_endian(vq, idx, argp);
1504                 break;
1505         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1506                 if (copy_from_user(&s, argp, sizeof(s))) {
1507                         r = -EFAULT;
1508                         break;
1509                 }
1510                 vq->busyloop_timeout = s.num;
1511                 break;
1512         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1513                 s.index = idx;
1514                 s.num = vq->busyloop_timeout;
1515                 if (copy_to_user(argp, &s, sizeof(s)))
1516                         r = -EFAULT;
1517                 break;
1518         default:
1519                 r = -ENOIOCTLCMD;
1520         }
1521
1522         if (pollstop && vq->handle_kick)
1523                 vhost_poll_stop(&vq->poll);
1524
1525         if (!IS_ERR_OR_NULL(ctx))
1526                 eventfd_ctx_put(ctx);
1527         if (filep)
1528                 fput(filep);
1529
1530         if (pollstart && vq->handle_kick)
1531                 r = vhost_poll_start(&vq->poll, vq->kick);
1532
1533         mutex_unlock(&vq->mutex);
1534
1535         if (pollstop && vq->handle_kick)
1536                 vhost_poll_flush(&vq->poll);
1537         return r;
1538 }
1539 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1540
1541 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1542 {
1543         struct vhost_umem *niotlb, *oiotlb;
1544         int i;
1545
1546         niotlb = vhost_umem_alloc();
1547         if (!niotlb)
1548                 return -ENOMEM;
1549
1550         oiotlb = d->iotlb;
1551         d->iotlb = niotlb;
1552
1553         for (i = 0; i < d->nvqs; ++i) {
1554                 mutex_lock(&d->vqs[i]->mutex);
1555                 d->vqs[i]->iotlb = niotlb;
1556                 mutex_unlock(&d->vqs[i]->mutex);
1557         }
1558
1559         vhost_umem_clean(oiotlb);
1560
1561         return 0;
1562 }
1563 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1564
1565 /* Caller must have device mutex */
1566 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1567 {
1568         struct eventfd_ctx *ctx;
1569         u64 p;
1570         long r;
1571         int i, fd;
1572
1573         /* If you are not the owner, you can become one */
1574         if (ioctl == VHOST_SET_OWNER) {
1575                 r = vhost_dev_set_owner(d);
1576                 goto done;
1577         }
1578
1579         /* You must be the owner to do anything else */
1580         r = vhost_dev_check_owner(d);
1581         if (r)
1582                 goto done;
1583
1584         switch (ioctl) {
1585         case VHOST_SET_MEM_TABLE:
1586                 r = vhost_set_memory(d, argp);
1587                 break;
1588         case VHOST_SET_LOG_BASE:
1589                 if (copy_from_user(&p, argp, sizeof p)) {
1590                         r = -EFAULT;
1591                         break;
1592                 }
1593                 if ((u64)(unsigned long)p != p) {
1594                         r = -EFAULT;
1595                         break;
1596                 }
1597                 for (i = 0; i < d->nvqs; ++i) {
1598                         struct vhost_virtqueue *vq;
1599                         void __user *base = (void __user *)(unsigned long)p;
1600                         vq = d->vqs[i];
1601                         mutex_lock(&vq->mutex);
1602                         /* If ring is inactive, will check when it's enabled. */
1603                         if (vq->private_data && !vq_log_access_ok(vq, base))
1604                                 r = -EFAULT;
1605                         else
1606                                 vq->log_base = base;
1607                         mutex_unlock(&vq->mutex);
1608                 }
1609                 break;
1610         case VHOST_SET_LOG_FD:
1611                 r = get_user(fd, (int __user *)argp);
1612                 if (r < 0)
1613                         break;
1614                 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1615                 if (IS_ERR(ctx)) {
1616                         r = PTR_ERR(ctx);
1617                         break;
1618                 }
1619                 swap(ctx, d->log_ctx);
1620                 for (i = 0; i < d->nvqs; ++i) {
1621                         mutex_lock(&d->vqs[i]->mutex);
1622                         d->vqs[i]->log_ctx = d->log_ctx;
1623                         mutex_unlock(&d->vqs[i]->mutex);
1624                 }
1625                 if (ctx)
1626                         eventfd_ctx_put(ctx);
1627                 break;
1628         default:
1629                 r = -ENOIOCTLCMD;
1630                 break;
1631         }
1632 done:
1633         return r;
1634 }
1635 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1636
1637 /* TODO: This is really inefficient.  We need something like get_user()
1638  * (instruction directly accesses the data, with an exception table entry
1639  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1640  */
1641 static int set_bit_to_user(int nr, void __user *addr)
1642 {
1643         unsigned long log = (unsigned long)addr;
1644         struct page *page;
1645         void *base;
1646         int bit = nr + (log % PAGE_SIZE) * 8;
1647         int r;
1648
1649         r = get_user_pages_fast(log, 1, 1, &page);
1650         if (r < 0)
1651                 return r;
1652         BUG_ON(r != 1);
1653         base = kmap_atomic(page);
1654         set_bit(bit, base);
1655         kunmap_atomic(base);
1656         set_page_dirty_lock(page);
1657         put_page(page);
1658         return 0;
1659 }
1660
1661 static int log_write(void __user *log_base,
1662                      u64 write_address, u64 write_length)
1663 {
1664         u64 write_page = write_address / VHOST_PAGE_SIZE;
1665         int r;
1666
1667         if (!write_length)
1668                 return 0;
1669         write_length += write_address % VHOST_PAGE_SIZE;
1670         for (;;) {
1671                 u64 base = (u64)(unsigned long)log_base;
1672                 u64 log = base + write_page / 8;
1673                 int bit = write_page % 8;
1674                 if ((u64)(unsigned long)log != log)
1675                         return -EFAULT;
1676                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1677                 if (r < 0)
1678                         return r;
1679                 if (write_length <= VHOST_PAGE_SIZE)
1680                         break;
1681                 write_length -= VHOST_PAGE_SIZE;
1682                 write_page += 1;
1683         }
1684         return r;
1685 }
1686
1687 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1688                     unsigned int log_num, u64 len)
1689 {
1690         int i, r;
1691
1692         /* Make sure data written is seen before log. */
1693         smp_wmb();
1694         for (i = 0; i < log_num; ++i) {
1695                 u64 l = min(log[i].len, len);
1696                 r = log_write(vq->log_base, log[i].addr, l);
1697                 if (r < 0)
1698                         return r;
1699                 len -= l;
1700                 if (!len) {
1701                         if (vq->log_ctx)
1702                                 eventfd_signal(vq->log_ctx, 1);
1703                         return 0;
1704                 }
1705         }
1706         /* Length written exceeds what we have stored. This is a bug. */
1707         BUG();
1708         return 0;
1709 }
1710 EXPORT_SYMBOL_GPL(vhost_log_write);
1711
1712 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1713 {
1714         void __user *used;
1715         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1716                            &vq->used->flags) < 0)
1717                 return -EFAULT;
1718         if (unlikely(vq->log_used)) {
1719                 /* Make sure the flag is seen before log. */
1720                 smp_wmb();
1721                 /* Log used flag write. */
1722                 used = &vq->used->flags;
1723                 log_write(vq->log_base, vq->log_addr +
1724                           (used - (void __user *)vq->used),
1725                           sizeof vq->used->flags);
1726                 if (vq->log_ctx)
1727                         eventfd_signal(vq->log_ctx, 1);
1728         }
1729         return 0;
1730 }
1731
1732 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1733 {
1734         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1735                            vhost_avail_event(vq)))
1736                 return -EFAULT;
1737         if (unlikely(vq->log_used)) {
1738                 void __user *used;
1739                 /* Make sure the event is seen before log. */
1740                 smp_wmb();
1741                 /* Log avail event write */
1742                 used = vhost_avail_event(vq);
1743                 log_write(vq->log_base, vq->log_addr +
1744                           (used - (void __user *)vq->used),
1745                           sizeof *vhost_avail_event(vq));
1746                 if (vq->log_ctx)
1747                         eventfd_signal(vq->log_ctx, 1);
1748         }
1749         return 0;
1750 }
1751
1752 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1753 {
1754         __virtio16 last_used_idx;
1755         int r;
1756         bool is_le = vq->is_le;
1757
1758         if (!vq->private_data)
1759                 return 0;
1760
1761         vhost_init_is_le(vq);
1762
1763         r = vhost_update_used_flags(vq);
1764         if (r)
1765                 goto err;
1766         vq->signalled_used_valid = false;
1767         if (!vq->iotlb &&
1768             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1769                 r = -EFAULT;
1770                 goto err;
1771         }
1772         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1773         if (r) {
1774                 vq_err(vq, "Can't access used idx at %p\n",
1775                        &vq->used->idx);
1776                 goto err;
1777         }
1778         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1779         return 0;
1780
1781 err:
1782         vq->is_le = is_le;
1783         return r;
1784 }
1785 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1786
1787 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1788                           struct iovec iov[], int iov_size, int access)
1789 {
1790         const struct vhost_umem_node *node;
1791         struct vhost_dev *dev = vq->dev;
1792         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1793         struct iovec *_iov;
1794         u64 s = 0;
1795         int ret = 0;
1796
1797         while ((u64)len > s) {
1798                 u64 size;
1799                 if (unlikely(ret >= iov_size)) {
1800                         ret = -ENOBUFS;
1801                         break;
1802                 }
1803
1804                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1805                                                         addr, addr + len - 1);
1806                 if (node == NULL || node->start > addr) {
1807                         if (umem != dev->iotlb) {
1808                                 ret = -EFAULT;
1809                                 break;
1810                         }
1811                         ret = -EAGAIN;
1812                         break;
1813                 } else if (!(node->perm & access)) {
1814                         ret = -EPERM;
1815                         break;
1816                 }
1817
1818                 _iov = iov + ret;
1819                 size = node->size - addr + node->start;
1820                 _iov->iov_len = min((u64)len - s, size);
1821                 _iov->iov_base = (void __user *)(unsigned long)
1822                         (node->userspace_addr + addr - node->start);
1823                 s += size;
1824                 addr += size;
1825                 ++ret;
1826         }
1827
1828         if (ret == -EAGAIN)
1829                 vhost_iotlb_miss(vq, addr, access);
1830         return ret;
1831 }
1832
1833 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1834  * function returns the next descriptor in the chain,
1835  * or -1U if we're at the end. */
1836 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1837 {
1838         unsigned int next;
1839
1840         /* If this descriptor says it doesn't chain, we're done. */
1841         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1842                 return -1U;
1843
1844         /* Check they're not leading us off end of descriptors. */
1845         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1846         return next;
1847 }
1848
1849 static int get_indirect(struct vhost_virtqueue *vq,
1850                         struct iovec iov[], unsigned int iov_size,
1851                         unsigned int *out_num, unsigned int *in_num,
1852                         struct vhost_log *log, unsigned int *log_num,
1853                         struct vring_desc *indirect)
1854 {
1855         struct vring_desc desc;
1856         unsigned int i = 0, count, found = 0;
1857         u32 len = vhost32_to_cpu(vq, indirect->len);
1858         struct iov_iter from;
1859         int ret, access;
1860
1861         /* Sanity check */
1862         if (unlikely(len % sizeof desc)) {
1863                 vq_err(vq, "Invalid length in indirect descriptor: "
1864                        "len 0x%llx not multiple of 0x%zx\n",
1865                        (unsigned long long)len,
1866                        sizeof desc);
1867                 return -EINVAL;
1868         }
1869
1870         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1871                              UIO_MAXIOV, VHOST_ACCESS_RO);
1872         if (unlikely(ret < 0)) {
1873                 if (ret != -EAGAIN)
1874                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1875                 return ret;
1876         }
1877         iov_iter_init(&from, READ, vq->indirect, ret, len);
1878
1879         /* We will use the result as an address to read from, so most
1880          * architectures only need a compiler barrier here. */
1881         read_barrier_depends();
1882
1883         count = len / sizeof desc;
1884         /* Buffers are chained via a 16 bit next field, so
1885          * we can have at most 2^16 of these. */
1886         if (unlikely(count > USHRT_MAX + 1)) {
1887                 vq_err(vq, "Indirect buffer length too big: %d\n",
1888                        indirect->len);
1889                 return -E2BIG;
1890         }
1891
1892         do {
1893                 unsigned iov_count = *in_num + *out_num;
1894                 if (unlikely(++found > count)) {
1895                         vq_err(vq, "Loop detected: last one at %u "
1896                                "indirect size %u\n",
1897                                i, count);
1898                         return -EINVAL;
1899                 }
1900                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1901                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1902                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1903                         return -EINVAL;
1904                 }
1905                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1906                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1907                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1908                         return -EINVAL;
1909                 }
1910
1911                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1912                         access = VHOST_ACCESS_WO;
1913                 else
1914                         access = VHOST_ACCESS_RO;
1915
1916                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1917                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1918                                      iov_size - iov_count, access);
1919                 if (unlikely(ret < 0)) {
1920                         if (ret != -EAGAIN)
1921                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1922                                         ret, i);
1923                         return ret;
1924                 }
1925                 /* If this is an input descriptor, increment that count. */
1926                 if (access == VHOST_ACCESS_WO) {
1927                         *in_num += ret;
1928                         if (unlikely(log)) {
1929                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1930                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1931                                 ++*log_num;
1932                         }
1933                 } else {
1934                         /* If it's an output descriptor, they're all supposed
1935                          * to come before any input descriptors. */
1936                         if (unlikely(*in_num)) {
1937                                 vq_err(vq, "Indirect descriptor "
1938                                        "has out after in: idx %d\n", i);
1939                                 return -EINVAL;
1940                         }
1941                         *out_num += ret;
1942                 }
1943         } while ((i = next_desc(vq, &desc)) != -1);
1944         return 0;
1945 }
1946
1947 /* This looks in the virtqueue and for the first available buffer, and converts
1948  * it to an iovec for convenient access.  Since descriptors consist of some
1949  * number of output then some number of input descriptors, it's actually two
1950  * iovecs, but we pack them into one and note how many of each there were.
1951  *
1952  * This function returns the descriptor number found, or vq->num (which is
1953  * never a valid descriptor number) if none was found.  A negative code is
1954  * returned on error. */
1955 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1956                       struct iovec iov[], unsigned int iov_size,
1957                       unsigned int *out_num, unsigned int *in_num,
1958                       struct vhost_log *log, unsigned int *log_num)
1959 {
1960         struct vring_desc desc;
1961         unsigned int i, head, found = 0;
1962         u16 last_avail_idx;
1963         __virtio16 avail_idx;
1964         __virtio16 ring_head;
1965         int ret, access;
1966
1967         /* Check it isn't doing very strange things with descriptor numbers. */
1968         last_avail_idx = vq->last_avail_idx;
1969
1970         if (vq->avail_idx == vq->last_avail_idx) {
1971                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
1972                         vq_err(vq, "Failed to access avail idx at %p\n",
1973                                 &vq->avail->idx);
1974                         return -EFAULT;
1975                 }
1976                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1977
1978                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1979                         vq_err(vq, "Guest moved used index from %u to %u",
1980                                 last_avail_idx, vq->avail_idx);
1981                         return -EFAULT;
1982                 }
1983
1984                 /* If there's nothing new since last we looked, return
1985                  * invalid.
1986                  */
1987                 if (vq->avail_idx == last_avail_idx)
1988                         return vq->num;
1989
1990                 /* Only get avail ring entries after they have been
1991                  * exposed by guest.
1992                  */
1993                 smp_rmb();
1994         }
1995
1996         /* Grab the next descriptor number they're advertising, and increment
1997          * the index we've seen. */
1998         if (unlikely(vhost_get_avail(vq, ring_head,
1999                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2000                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2001                        last_avail_idx,
2002                        &vq->avail->ring[last_avail_idx % vq->num]);
2003                 return -EFAULT;
2004         }
2005
2006         head = vhost16_to_cpu(vq, ring_head);
2007
2008         /* If their number is silly, that's an error. */
2009         if (unlikely(head >= vq->num)) {
2010                 vq_err(vq, "Guest says index %u > %u is available",
2011                        head, vq->num);
2012                 return -EINVAL;
2013         }
2014
2015         /* When we start there are none of either input nor output. */
2016         *out_num = *in_num = 0;
2017         if (unlikely(log))
2018                 *log_num = 0;
2019
2020         i = head;
2021         do {
2022                 unsigned iov_count = *in_num + *out_num;
2023                 if (unlikely(i >= vq->num)) {
2024                         vq_err(vq, "Desc index is %u > %u, head = %u",
2025                                i, vq->num, head);
2026                         return -EINVAL;
2027                 }
2028                 if (unlikely(++found > vq->num)) {
2029                         vq_err(vq, "Loop detected: last one at %u "
2030                                "vq size %u head %u\n",
2031                                i, vq->num, head);
2032                         return -EINVAL;
2033                 }
2034                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2035                                            sizeof desc);
2036                 if (unlikely(ret)) {
2037                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2038                                i, vq->desc + i);
2039                         return -EFAULT;
2040                 }
2041                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2042                         ret = get_indirect(vq, iov, iov_size,
2043                                            out_num, in_num,
2044                                            log, log_num, &desc);
2045                         if (unlikely(ret < 0)) {
2046                                 if (ret != -EAGAIN)
2047                                         vq_err(vq, "Failure detected "
2048                                                 "in indirect descriptor at idx %d\n", i);
2049                                 return ret;
2050                         }
2051                         continue;
2052                 }
2053
2054                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2055                         access = VHOST_ACCESS_WO;
2056                 else
2057                         access = VHOST_ACCESS_RO;
2058                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2059                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2060                                      iov_size - iov_count, access);
2061                 if (unlikely(ret < 0)) {
2062                         if (ret != -EAGAIN)
2063                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2064                                         ret, i);
2065                         return ret;
2066                 }
2067                 if (access == VHOST_ACCESS_WO) {
2068                         /* If this is an input descriptor,
2069                          * increment that count. */
2070                         *in_num += ret;
2071                         if (unlikely(log)) {
2072                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2073                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2074                                 ++*log_num;
2075                         }
2076                 } else {
2077                         /* If it's an output descriptor, they're all supposed
2078                          * to come before any input descriptors. */
2079                         if (unlikely(*in_num)) {
2080                                 vq_err(vq, "Descriptor has out after in: "
2081                                        "idx %d\n", i);
2082                                 return -EINVAL;
2083                         }
2084                         *out_num += ret;
2085                 }
2086         } while ((i = next_desc(vq, &desc)) != -1);
2087
2088         /* On success, increment avail index. */
2089         vq->last_avail_idx++;
2090
2091         /* Assume notifications from guest are disabled at this point,
2092          * if they aren't we would need to update avail_event index. */
2093         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2094         return head;
2095 }
2096 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2097
2098 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2099 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2100 {
2101         vq->last_avail_idx -= n;
2102 }
2103 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2104
2105 /* After we've used one of their buffers, we tell them about it.  We'll then
2106  * want to notify the guest, using eventfd. */
2107 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2108 {
2109         struct vring_used_elem heads = {
2110                 cpu_to_vhost32(vq, head),
2111                 cpu_to_vhost32(vq, len)
2112         };
2113
2114         return vhost_add_used_n(vq, &heads, 1);
2115 }
2116 EXPORT_SYMBOL_GPL(vhost_add_used);
2117
2118 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2119                             struct vring_used_elem *heads,
2120                             unsigned count)
2121 {
2122         struct vring_used_elem __user *used;
2123         u16 old, new;
2124         int start;
2125
2126         start = vq->last_used_idx & (vq->num - 1);
2127         used = vq->used->ring + start;
2128         if (count == 1) {
2129                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2130                         vq_err(vq, "Failed to write used id");
2131                         return -EFAULT;
2132                 }
2133                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2134                         vq_err(vq, "Failed to write used len");
2135                         return -EFAULT;
2136                 }
2137         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2138                 vq_err(vq, "Failed to write used");
2139                 return -EFAULT;
2140         }
2141         if (unlikely(vq->log_used)) {
2142                 /* Make sure data is seen before log. */
2143                 smp_wmb();
2144                 /* Log used ring entry write. */
2145                 log_write(vq->log_base,
2146                           vq->log_addr +
2147                            ((void __user *)used - (void __user *)vq->used),
2148                           count * sizeof *used);
2149         }
2150         old = vq->last_used_idx;
2151         new = (vq->last_used_idx += count);
2152         /* If the driver never bothers to signal in a very long while,
2153          * used index might wrap around. If that happens, invalidate
2154          * signalled_used index we stored. TODO: make sure driver
2155          * signals at least once in 2^16 and remove this. */
2156         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2157                 vq->signalled_used_valid = false;
2158         return 0;
2159 }
2160
2161 /* After we've used one of their buffers, we tell them about it.  We'll then
2162  * want to notify the guest, using eventfd. */
2163 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2164                      unsigned count)
2165 {
2166         int start, n, r;
2167
2168         start = vq->last_used_idx & (vq->num - 1);
2169         n = vq->num - start;
2170         if (n < count) {
2171                 r = __vhost_add_used_n(vq, heads, n);
2172                 if (r < 0)
2173                         return r;
2174                 heads += n;
2175                 count -= n;
2176         }
2177         r = __vhost_add_used_n(vq, heads, count);
2178
2179         /* Make sure buffer is written before we update index. */
2180         smp_wmb();
2181         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2182                            &vq->used->idx)) {
2183                 vq_err(vq, "Failed to increment used idx");
2184                 return -EFAULT;
2185         }
2186         if (unlikely(vq->log_used)) {
2187                 /* Log used index update. */
2188                 log_write(vq->log_base,
2189                           vq->log_addr + offsetof(struct vring_used, idx),
2190                           sizeof vq->used->idx);
2191                 if (vq->log_ctx)
2192                         eventfd_signal(vq->log_ctx, 1);
2193         }
2194         return r;
2195 }
2196 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2197
2198 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2199 {
2200         __u16 old, new;
2201         __virtio16 event;
2202         bool v;
2203         /* Flush out used index updates. This is paired
2204          * with the barrier that the Guest executes when enabling
2205          * interrupts. */
2206         smp_mb();
2207
2208         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2209             unlikely(vq->avail_idx == vq->last_avail_idx))
2210                 return true;
2211
2212         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2213                 __virtio16 flags;
2214                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2215                         vq_err(vq, "Failed to get flags");
2216                         return true;
2217                 }
2218                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2219         }
2220         old = vq->signalled_used;
2221         v = vq->signalled_used_valid;
2222         new = vq->signalled_used = vq->last_used_idx;
2223         vq->signalled_used_valid = true;
2224
2225         if (unlikely(!v))
2226                 return true;
2227
2228         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2229                 vq_err(vq, "Failed to get used event idx");
2230                 return true;
2231         }
2232         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2233 }
2234
2235 /* This actually signals the guest, using eventfd. */
2236 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2237 {
2238         /* Signal the Guest tell them we used something up. */
2239         if (vq->call_ctx && vhost_notify(dev, vq))
2240                 eventfd_signal(vq->call_ctx, 1);
2241 }
2242 EXPORT_SYMBOL_GPL(vhost_signal);
2243
2244 /* And here's the combo meal deal.  Supersize me! */
2245 void vhost_add_used_and_signal(struct vhost_dev *dev,
2246                                struct vhost_virtqueue *vq,
2247                                unsigned int head, int len)
2248 {
2249         vhost_add_used(vq, head, len);
2250         vhost_signal(dev, vq);
2251 }
2252 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2253
2254 /* multi-buffer version of vhost_add_used_and_signal */
2255 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2256                                  struct vhost_virtqueue *vq,
2257                                  struct vring_used_elem *heads, unsigned count)
2258 {
2259         vhost_add_used_n(vq, heads, count);
2260         vhost_signal(dev, vq);
2261 }
2262 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2263
2264 /* return true if we're sure that avaiable ring is empty */
2265 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2266 {
2267         __virtio16 avail_idx;
2268         int r;
2269
2270         if (vq->avail_idx != vq->last_avail_idx)
2271                 return false;
2272
2273         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2274         if (unlikely(r))
2275                 return false;
2276         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2277
2278         return vq->avail_idx == vq->last_avail_idx;
2279 }
2280 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2281
2282 /* OK, now we need to know about added descriptors. */
2283 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2284 {
2285         __virtio16 avail_idx;
2286         int r;
2287
2288         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2289                 return false;
2290         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2291         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2292                 r = vhost_update_used_flags(vq);
2293                 if (r) {
2294                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2295                                &vq->used->flags, r);
2296                         return false;
2297                 }
2298         } else {
2299                 r = vhost_update_avail_event(vq, vq->avail_idx);
2300                 if (r) {
2301                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2302                                vhost_avail_event(vq), r);
2303                         return false;
2304                 }
2305         }
2306         /* They could have slipped one in as we were doing that: make
2307          * sure it's written, then check again. */
2308         smp_mb();
2309         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2310         if (r) {
2311                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2312                        &vq->avail->idx, r);
2313                 return false;
2314         }
2315
2316         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2317 }
2318 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2319
2320 /* We don't need to be notified again. */
2321 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2322 {
2323         int r;
2324
2325         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2326                 return;
2327         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2328         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2329                 r = vhost_update_used_flags(vq);
2330                 if (r)
2331                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2332                                &vq->used->flags, r);
2333         }
2334 }
2335 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2336
2337 /* Create a new message. */
2338 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2339 {
2340         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2341         if (!node)
2342                 return NULL;
2343         node->vq = vq;
2344         node->msg.type = type;
2345         return node;
2346 }
2347 EXPORT_SYMBOL_GPL(vhost_new_msg);
2348
2349 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2350                        struct vhost_msg_node *node)
2351 {
2352         spin_lock(&dev->iotlb_lock);
2353         list_add_tail(&node->node, head);
2354         spin_unlock(&dev->iotlb_lock);
2355
2356         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2357 }
2358 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2359
2360 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2361                                          struct list_head *head)
2362 {
2363         struct vhost_msg_node *node = NULL;
2364
2365         spin_lock(&dev->iotlb_lock);
2366         if (!list_empty(head)) {
2367                 node = list_first_entry(head, struct vhost_msg_node,
2368                                         node);
2369                 list_del(&node->node);
2370         }
2371         spin_unlock(&dev->iotlb_lock);
2372
2373         return node;
2374 }
2375 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2376
2377
2378 static int __init vhost_init(void)
2379 {
2380         return 0;
2381 }
2382
2383 static void __exit vhost_exit(void)
2384 {
2385 }
2386
2387 module_init(vhost_init);
2388 module_exit(vhost_exit);
2389
2390 MODULE_VERSION("0.0.1");
2391 MODULE_LICENSE("GPL v2");
2392 MODULE_AUTHOR("Michael S. Tsirkin");
2393 MODULE_DESCRIPTION("Host kernel accelerator for virtio");