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 bool 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 false;
 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 bool 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 false;
 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 false;
 677
 678
 679                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 680                                     node->size))
 681                         return false;
 682                 else if (log_all && !log_access_ok(log_base,
 683                                                    node->start,
 684                                                    node->size))
 685                         return false;
 686         }
 687         return true;
 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 bool 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                 bool 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 = true;
 721                 mutex_unlock(&d->vqs[i]->mutex);
 722                 if (!ok)
 723                         return false;
 724         }
 725         return true;
 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_USED);
 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 bool 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 false;
 969
 970         if ((access & VHOST_ACCESS_RO) &&
 971             !access_ok(VERIFY_READ, (void __user *)a, size))
 972                 return false;
 973         if ((access & VHOST_ACCESS_WO) &&
 974             !access_ok(VERIFY_WRITE, (void __user *)a, size))
 975                 return false;
 976         return true;
 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 bool 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 bool 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 bool 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 bool 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 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1246 {
1247         if (!vq_log_access_ok(vq, vq->log_base))
1248                 return false;
1249
1250         /* Access validation occurs at prefetch time with IOTLB */
1251         if (vq->iotlb)
1252                 return true;
1253
1254         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1255 }
1256 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1257
1258 static struct vhost_umem *vhost_umem_alloc(void)
1259 {
1260         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1261
1262         if (!umem)
1263                 return NULL;
1264
1265         umem->umem_tree = RB_ROOT_CACHED;
1266         umem->numem = 0;
1267         INIT_LIST_HEAD(&umem->umem_list);
1268
1269         return umem;
1270 }
1271
1272 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1273 {
1274         struct vhost_memory mem, *newmem;
1275         struct vhost_memory_region *region;
1276         struct vhost_umem *newumem, *oldumem;
1277         unsigned long size = offsetof(struct vhost_memory, regions);
1278         int i;
1279
1280         if (copy_from_user(&mem, m, size))
1281                 return -EFAULT;
1282         if (mem.padding)
1283                 return -EOPNOTSUPP;
1284         if (mem.nregions > max_mem_regions)
1285                 return -E2BIG;
1286         newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1287         if (!newmem)
1288                 return -ENOMEM;
1289
1290         memcpy(newmem, &mem, size);
1291         if (copy_from_user(newmem->regions, m->regions,
1292                            mem.nregions * sizeof *m->regions)) {
1293                 kvfree(newmem);
1294                 return -EFAULT;
1295         }
1296
1297         newumem = vhost_umem_alloc();
1298         if (!newumem) {
1299                 kvfree(newmem);
1300                 return -ENOMEM;
1301         }
1302
1303         for (region = newmem->regions;
1304              region < newmem->regions + mem.nregions;
1305              region++) {
1306                 if (vhost_new_umem_range(newumem,
1307                                          region->guest_phys_addr,
1308                                          region->memory_size,
1309                                          region->guest_phys_addr +
1310                                          region->memory_size - 1,
1311                                          region->userspace_addr,
1312                                          VHOST_ACCESS_RW))
1313                         goto err;
1314         }
1315
1316         if (!memory_access_ok(d, newumem, 0))
1317                 goto err;
1318
1319         oldumem = d->umem;
1320         d->umem = newumem;
1321
1322         /* All memory accesses are done under some VQ mutex. */
1323         for (i = 0; i < d->nvqs; ++i) {
1324                 mutex_lock(&d->vqs[i]->mutex);
1325                 d->vqs[i]->umem = newumem;
1326                 mutex_unlock(&d->vqs[i]->mutex);
1327         }
1328
1329         kvfree(newmem);
1330         vhost_umem_clean(oldumem);
1331         return 0;
1332
1333 err:
1334         vhost_umem_clean(newumem);
1335         kvfree(newmem);
1336         return -EFAULT;
1337 }
1338
1339 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1340 {
1341         struct file *eventfp, *filep = NULL;
1342         bool pollstart = false, pollstop = false;
1343         struct eventfd_ctx *ctx = NULL;
1344         u32 __user *idxp = argp;
1345         struct vhost_virtqueue *vq;
1346         struct vhost_vring_state s;
1347         struct vhost_vring_file f;
1348         struct vhost_vring_addr a;
1349         u32 idx;
1350         long r;
1351
1352         r = get_user(idx, idxp);
1353         if (r < 0)
1354                 return r;
1355         if (idx >= d->nvqs)
1356                 return -ENOBUFS;
1357
1358         vq = d->vqs[idx];
1359
1360         mutex_lock(&vq->mutex);
1361
1362         switch (ioctl) {
1363         case VHOST_SET_VRING_NUM:
1364                 /* Resizing ring with an active backend?
1365                  * You don't want to do that. */
1366                 if (vq->private_data) {
1367                         r = -EBUSY;
1368                         break;
1369                 }
1370                 if (copy_from_user(&s, argp, sizeof s)) {
1371                         r = -EFAULT;
1372                         break;
1373                 }
1374                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1375                         r = -EINVAL;
1376                         break;
1377                 }
1378                 vq->num = s.num;
1379                 break;
1380         case VHOST_SET_VRING_BASE:
1381                 /* Moving base with an active backend?
1382                  * You don't want to do that. */
1383                 if (vq->private_data) {
1384                         r = -EBUSY;
1385                         break;
1386                 }
1387                 if (copy_from_user(&s, argp, sizeof s)) {
1388                         r = -EFAULT;
1389                         break;
1390                 }
1391                 if (s.num > 0xffff) {
1392                         r = -EINVAL;
1393                         break;
1394                 }
1395                 vq->last_avail_idx = s.num;
1396                 /* Forget the cached index value. */
1397                 vq->avail_idx = vq->last_avail_idx;
1398                 break;
1399         case VHOST_GET_VRING_BASE:
1400                 s.index = idx;
1401                 s.num = vq->last_avail_idx;
1402                 if (copy_to_user(argp, &s, sizeof s))
1403                         r = -EFAULT;
1404                 break;
1405         case VHOST_SET_VRING_ADDR:
1406                 if (copy_from_user(&a, argp, sizeof a)) {
1407                         r = -EFAULT;
1408                         break;
1409                 }
1410                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1411                         r = -EOPNOTSUPP;
1412                         break;
1413                 }
1414                 /* For 32bit, verify that the top 32bits of the user
1415                    data are set to zero. */
1416                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1417                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1418                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1419                         r = -EFAULT;
1420                         break;
1421                 }
1422
1423                 /* Make sure it's safe to cast pointers to vring types. */
1424                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1425                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1426                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1427                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1428                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1429                         r = -EINVAL;
1430                         break;
1431                 }
1432
1433                 /* We only verify access here if backend is configured.
1434                  * If it is not, we don't as size might not have been setup.
1435                  * We will verify when backend is configured. */
1436                 if (vq->private_data) {
1437                         if (!vq_access_ok(vq, vq->num,
1438                                 (void __user *)(unsigned long)a.desc_user_addr,
1439                                 (void __user *)(unsigned long)a.avail_user_addr,
1440                                 (void __user *)(unsigned long)a.used_user_addr)) {
1441                                 r = -EINVAL;
1442                                 break;
1443                         }
1444
1445                         /* Also validate log access for used ring if enabled. */
1446                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1447                             !log_access_ok(vq->log_base, a.log_guest_addr,
1448                                            sizeof *vq->used +
1449                                            vq->num * sizeof *vq->used->ring)) {
1450                                 r = -EINVAL;
1451                                 break;
1452                         }
1453                 }
1454
1455                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1456                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1457                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1458                 vq->log_addr = a.log_guest_addr;
1459                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1460                 break;
1461         case VHOST_SET_VRING_KICK:
1462                 if (copy_from_user(&f, argp, sizeof f)) {
1463                         r = -EFAULT;
1464                         break;
1465                 }
1466                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1467                 if (IS_ERR(eventfp)) {
1468                         r = PTR_ERR(eventfp);
1469                         break;
1470                 }
1471                 if (eventfp != vq->kick) {
1472                         pollstop = (filep = vq->kick) != NULL;
1473                         pollstart = (vq->kick = eventfp) != NULL;
1474                 } else
1475                         filep = eventfp;
1476                 break;
1477         case VHOST_SET_VRING_CALL:
1478                 if (copy_from_user(&f, argp, sizeof f)) {
1479                         r = -EFAULT;
1480                         break;
1481                 }
1482                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1483                 if (IS_ERR(ctx)) {
1484                         r = PTR_ERR(ctx);
1485                         break;
1486                 }
1487                 swap(ctx, vq->call_ctx);
1488                 break;
1489         case VHOST_SET_VRING_ERR:
1490                 if (copy_from_user(&f, argp, sizeof f)) {
1491                         r = -EFAULT;
1492                         break;
1493                 }
1494                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1495                 if (IS_ERR(ctx)) {
1496                         r = PTR_ERR(ctx);
1497                         break;
1498                 }
1499                 swap(ctx, vq->error_ctx);
1500                 break;
1501         case VHOST_SET_VRING_ENDIAN:
1502                 r = vhost_set_vring_endian(vq, argp);
1503                 break;
1504         case VHOST_GET_VRING_ENDIAN:
1505                 r = vhost_get_vring_endian(vq, idx, argp);
1506                 break;
1507         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1508                 if (copy_from_user(&s, argp, sizeof(s))) {
1509                         r = -EFAULT;
1510                         break;
1511                 }
1512                 vq->busyloop_timeout = s.num;
1513                 break;
1514         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1515                 s.index = idx;
1516                 s.num = vq->busyloop_timeout;
1517                 if (copy_to_user(argp, &s, sizeof(s)))
1518                         r = -EFAULT;
1519                 break;
1520         default:
1521                 r = -ENOIOCTLCMD;
1522         }
1523
1524         if (pollstop && vq->handle_kick)
1525                 vhost_poll_stop(&vq->poll);
1526
1527         if (!IS_ERR_OR_NULL(ctx))
1528                 eventfd_ctx_put(ctx);
1529         if (filep)
1530                 fput(filep);
1531
1532         if (pollstart && vq->handle_kick)
1533                 r = vhost_poll_start(&vq->poll, vq->kick);
1534
1535         mutex_unlock(&vq->mutex);
1536
1537         if (pollstop && vq->handle_kick)
1538                 vhost_poll_flush(&vq->poll);
1539         return r;
1540 }
1541 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1542
1543 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1544 {
1545         struct vhost_umem *niotlb, *oiotlb;
1546         int i;
1547
1548         niotlb = vhost_umem_alloc();
1549         if (!niotlb)
1550                 return -ENOMEM;
1551
1552         oiotlb = d->iotlb;
1553         d->iotlb = niotlb;
1554
1555         for (i = 0; i < d->nvqs; ++i) {
1556                 mutex_lock(&d->vqs[i]->mutex);
1557                 d->vqs[i]->iotlb = niotlb;
1558                 mutex_unlock(&d->vqs[i]->mutex);
1559         }
1560
1561         vhost_umem_clean(oiotlb);
1562
1563         return 0;
1564 }
1565 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1566
1567 /* Caller must have device mutex */
1568 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1569 {
1570         struct eventfd_ctx *ctx;
1571         u64 p;
1572         long r;
1573         int i, fd;
1574
1575         /* If you are not the owner, you can become one */
1576         if (ioctl == VHOST_SET_OWNER) {
1577                 r = vhost_dev_set_owner(d);
1578                 goto done;
1579         }
1580
1581         /* You must be the owner to do anything else */
1582         r = vhost_dev_check_owner(d);
1583         if (r)
1584                 goto done;
1585
1586         switch (ioctl) {
1587         case VHOST_SET_MEM_TABLE:
1588                 r = vhost_set_memory(d, argp);
1589                 break;
1590         case VHOST_SET_LOG_BASE:
1591                 if (copy_from_user(&p, argp, sizeof p)) {
1592                         r = -EFAULT;
1593                         break;
1594                 }
1595                 if ((u64)(unsigned long)p != p) {
1596                         r = -EFAULT;
1597                         break;
1598                 }
1599                 for (i = 0; i < d->nvqs; ++i) {
1600                         struct vhost_virtqueue *vq;
1601                         void __user *base = (void __user *)(unsigned long)p;
1602                         vq = d->vqs[i];
1603                         mutex_lock(&vq->mutex);
1604                         /* If ring is inactive, will check when it's enabled. */
1605                         if (vq->private_data && !vq_log_access_ok(vq, base))
1606                                 r = -EFAULT;
1607                         else
1608                                 vq->log_base = base;
1609                         mutex_unlock(&vq->mutex);
1610                 }
1611                 break;
1612         case VHOST_SET_LOG_FD:
1613                 r = get_user(fd, (int __user *)argp);
1614                 if (r < 0)
1615                         break;
1616                 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1617                 if (IS_ERR(ctx)) {
1618                         r = PTR_ERR(ctx);
1619                         break;
1620                 }
1621                 swap(ctx, d->log_ctx);
1622                 for (i = 0; i < d->nvqs; ++i) {
1623                         mutex_lock(&d->vqs[i]->mutex);
1624                         d->vqs[i]->log_ctx = d->log_ctx;
1625                         mutex_unlock(&d->vqs[i]->mutex);
1626                 }
1627                 if (ctx)
1628                         eventfd_ctx_put(ctx);
1629                 break;
1630         default:
1631                 r = -ENOIOCTLCMD;
1632                 break;
1633         }
1634 done:
1635         return r;
1636 }
1637 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1638
1639 /* TODO: This is really inefficient.  We need something like get_user()
1640  * (instruction directly accesses the data, with an exception table entry
1641  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1642  */
1643 static int set_bit_to_user(int nr, void __user *addr)
1644 {
1645         unsigned long log = (unsigned long)addr;
1646         struct page *page;
1647         void *base;
1648         int bit = nr + (log % PAGE_SIZE) * 8;
1649         int r;
1650
1651         r = get_user_pages_fast(log, 1, 1, &page);
1652         if (r < 0)
1653                 return r;
1654         BUG_ON(r != 1);
1655         base = kmap_atomic(page);
1656         set_bit(bit, base);
1657         kunmap_atomic(base);
1658         set_page_dirty_lock(page);
1659         put_page(page);
1660         return 0;
1661 }
1662
1663 static int log_write(void __user *log_base,
1664                      u64 write_address, u64 write_length)
1665 {
1666         u64 write_page = write_address / VHOST_PAGE_SIZE;
1667         int r;
1668
1669         if (!write_length)
1670                 return 0;
1671         write_length += write_address % VHOST_PAGE_SIZE;
1672         for (;;) {
1673                 u64 base = (u64)(unsigned long)log_base;
1674                 u64 log = base + write_page / 8;
1675                 int bit = write_page % 8;
1676                 if ((u64)(unsigned long)log != log)
1677                         return -EFAULT;
1678                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1679                 if (r < 0)
1680                         return r;
1681                 if (write_length <= VHOST_PAGE_SIZE)
1682                         break;
1683                 write_length -= VHOST_PAGE_SIZE;
1684                 write_page += 1;
1685         }
1686         return r;
1687 }
1688
1689 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1690                     unsigned int log_num, u64 len)
1691 {
1692         int i, r;
1693
1694         /* Make sure data written is seen before log. */
1695         smp_wmb();
1696         for (i = 0; i < log_num; ++i) {
1697                 u64 l = min(log[i].len, len);
1698                 r = log_write(vq->log_base, log[i].addr, l);
1699                 if (r < 0)
1700                         return r;
1701                 len -= l;
1702                 if (!len) {
1703                         if (vq->log_ctx)
1704                                 eventfd_signal(vq->log_ctx, 1);
1705                         return 0;
1706                 }
1707         }
1708         /* Length written exceeds what we have stored. This is a bug. */
1709         BUG();
1710         return 0;
1711 }
1712 EXPORT_SYMBOL_GPL(vhost_log_write);
1713
1714 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1715 {
1716         void __user *used;
1717         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1718                            &vq->used->flags) < 0)
1719                 return -EFAULT;
1720         if (unlikely(vq->log_used)) {
1721                 /* Make sure the flag is seen before log. */
1722                 smp_wmb();
1723                 /* Log used flag write. */
1724                 used = &vq->used->flags;
1725                 log_write(vq->log_base, vq->log_addr +
1726                           (used - (void __user *)vq->used),
1727                           sizeof vq->used->flags);
1728                 if (vq->log_ctx)
1729                         eventfd_signal(vq->log_ctx, 1);
1730         }
1731         return 0;
1732 }
1733
1734 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1735 {
1736         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1737                            vhost_avail_event(vq)))
1738                 return -EFAULT;
1739         if (unlikely(vq->log_used)) {
1740                 void __user *used;
1741                 /* Make sure the event is seen before log. */
1742                 smp_wmb();
1743                 /* Log avail event write */
1744                 used = vhost_avail_event(vq);
1745                 log_write(vq->log_base, vq->log_addr +
1746                           (used - (void __user *)vq->used),
1747                           sizeof *vhost_avail_event(vq));
1748                 if (vq->log_ctx)
1749                         eventfd_signal(vq->log_ctx, 1);
1750         }
1751         return 0;
1752 }
1753
1754 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1755 {
1756         __virtio16 last_used_idx;
1757         int r;
1758         bool is_le = vq->is_le;
1759
1760         if (!vq->private_data)
1761                 return 0;
1762
1763         vhost_init_is_le(vq);
1764
1765         r = vhost_update_used_flags(vq);
1766         if (r)
1767                 goto err;
1768         vq->signalled_used_valid = false;
1769         if (!vq->iotlb &&
1770             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1771                 r = -EFAULT;
1772                 goto err;
1773         }
1774         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1775         if (r) {
1776                 vq_err(vq, "Can't access used idx at %p\n",
1777                        &vq->used->idx);
1778                 goto err;
1779         }
1780         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1781         return 0;
1782
1783 err:
1784         vq->is_le = is_le;
1785         return r;
1786 }
1787 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1788
1789 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1790                           struct iovec iov[], int iov_size, int access)
1791 {
1792         const struct vhost_umem_node *node;
1793         struct vhost_dev *dev = vq->dev;
1794         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1795         struct iovec *_iov;
1796         u64 s = 0;
1797         int ret = 0;
1798
1799         while ((u64)len > s) {
1800                 u64 size;
1801                 if (unlikely(ret >= iov_size)) {
1802                         ret = -ENOBUFS;
1803                         break;
1804                 }
1805
1806                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1807                                                         addr, addr + len - 1);
1808                 if (node == NULL || node->start > addr) {
1809                         if (umem != dev->iotlb) {
1810                                 ret = -EFAULT;
1811                                 break;
1812                         }
1813                         ret = -EAGAIN;
1814                         break;
1815                 } else if (!(node->perm & access)) {
1816                         ret = -EPERM;
1817                         break;
1818                 }
1819
1820                 _iov = iov + ret;
1821                 size = node->size - addr + node->start;
1822                 _iov->iov_len = min((u64)len - s, size);
1823                 _iov->iov_base = (void __user *)(unsigned long)
1824                         (node->userspace_addr + addr - node->start);
1825                 s += size;
1826                 addr += size;
1827                 ++ret;
1828         }
1829
1830         if (ret == -EAGAIN)
1831                 vhost_iotlb_miss(vq, addr, access);
1832         return ret;
1833 }
1834
1835 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1836  * function returns the next descriptor in the chain,
1837  * or -1U if we're at the end. */
1838 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1839 {
1840         unsigned int next;
1841
1842         /* If this descriptor says it doesn't chain, we're done. */
1843         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1844                 return -1U;
1845
1846         /* Check they're not leading us off end of descriptors. */
1847         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1848         return next;
1849 }
1850
1851 static int get_indirect(struct vhost_virtqueue *vq,
1852                         struct iovec iov[], unsigned int iov_size,
1853                         unsigned int *out_num, unsigned int *in_num,
1854                         struct vhost_log *log, unsigned int *log_num,
1855                         struct vring_desc *indirect)
1856 {
1857         struct vring_desc desc;
1858         unsigned int i = 0, count, found = 0;
1859         u32 len = vhost32_to_cpu(vq, indirect->len);
1860         struct iov_iter from;
1861         int ret, access;
1862
1863         /* Sanity check */
1864         if (unlikely(len % sizeof desc)) {
1865                 vq_err(vq, "Invalid length in indirect descriptor: "
1866                        "len 0x%llx not multiple of 0x%zx\n",
1867                        (unsigned long long)len,
1868                        sizeof desc);
1869                 return -EINVAL;
1870         }
1871
1872         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1873                              UIO_MAXIOV, VHOST_ACCESS_RO);
1874         if (unlikely(ret < 0)) {
1875                 if (ret != -EAGAIN)
1876                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1877                 return ret;
1878         }
1879         iov_iter_init(&from, READ, vq->indirect, ret, len);
1880
1881         /* We will use the result as an address to read from, so most
1882          * architectures only need a compiler barrier here. */
1883         read_barrier_depends();
1884
1885         count = len / sizeof desc;
1886         /* Buffers are chained via a 16 bit next field, so
1887          * we can have at most 2^16 of these. */
1888         if (unlikely(count > USHRT_MAX + 1)) {
1889                 vq_err(vq, "Indirect buffer length too big: %d\n",
1890                        indirect->len);
1891                 return -E2BIG;
1892         }
1893
1894         do {
1895                 unsigned iov_count = *in_num + *out_num;
1896                 if (unlikely(++found > count)) {
1897                         vq_err(vq, "Loop detected: last one at %u "
1898                                "indirect size %u\n",
1899                                i, count);
1900                         return -EINVAL;
1901                 }
1902                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1903                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1904                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1905                         return -EINVAL;
1906                 }
1907                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1908                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1909                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1910                         return -EINVAL;
1911                 }
1912
1913                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1914                         access = VHOST_ACCESS_WO;
1915                 else
1916                         access = VHOST_ACCESS_RO;
1917
1918                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1919                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1920                                      iov_size - iov_count, access);
1921                 if (unlikely(ret < 0)) {
1922                         if (ret != -EAGAIN)
1923                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1924                                         ret, i);
1925                         return ret;
1926                 }
1927                 /* If this is an input descriptor, increment that count. */
1928                 if (access == VHOST_ACCESS_WO) {
1929                         *in_num += ret;
1930                         if (unlikely(log)) {
1931                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1932                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1933                                 ++*log_num;
1934                         }
1935                 } else {
1936                         /* If it's an output descriptor, they're all supposed
1937                          * to come before any input descriptors. */
1938                         if (unlikely(*in_num)) {
1939                                 vq_err(vq, "Indirect descriptor "
1940                                        "has out after in: idx %d\n", i);
1941                                 return -EINVAL;
1942                         }
1943                         *out_num += ret;
1944                 }
1945         } while ((i = next_desc(vq, &desc)) != -1);
1946         return 0;
1947 }
1948
1949 /* This looks in the virtqueue and for the first available buffer, and converts
1950  * it to an iovec for convenient access.  Since descriptors consist of some
1951  * number of output then some number of input descriptors, it's actually two
1952  * iovecs, but we pack them into one and note how many of each there were.
1953  *
1954  * This function returns the descriptor number found, or vq->num (which is
1955  * never a valid descriptor number) if none was found.  A negative code is
1956  * returned on error. */
1957 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1958                       struct iovec iov[], unsigned int iov_size,
1959                       unsigned int *out_num, unsigned int *in_num,
1960                       struct vhost_log *log, unsigned int *log_num)
1961 {
1962         struct vring_desc desc;
1963         unsigned int i, head, found = 0;
1964         u16 last_avail_idx;
1965         __virtio16 avail_idx;
1966         __virtio16 ring_head;
1967         int ret, access;
1968
1969         /* Check it isn't doing very strange things with descriptor numbers. */
1970         last_avail_idx = vq->last_avail_idx;
1971
1972         if (vq->avail_idx == vq->last_avail_idx) {
1973                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
1974                         vq_err(vq, "Failed to access avail idx at %p\n",
1975                                 &vq->avail->idx);
1976                         return -EFAULT;
1977                 }
1978                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1979
1980                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1981                         vq_err(vq, "Guest moved used index from %u to %u",
1982                                 last_avail_idx, vq->avail_idx);
1983                         return -EFAULT;
1984                 }
1985
1986                 /* If there's nothing new since last we looked, return
1987                  * invalid.
1988                  */
1989                 if (vq->avail_idx == last_avail_idx)
1990                         return vq->num;
1991
1992                 /* Only get avail ring entries after they have been
1993                  * exposed by guest.
1994                  */
1995                 smp_rmb();
1996         }
1997
1998         /* Grab the next descriptor number they're advertising, and increment
1999          * the index we've seen. */
2000         if (unlikely(vhost_get_avail(vq, ring_head,
2001                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2002                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2003                        last_avail_idx,
2004                        &vq->avail->ring[last_avail_idx % vq->num]);
2005                 return -EFAULT;
2006         }
2007
2008         head = vhost16_to_cpu(vq, ring_head);
2009
2010         /* If their number is silly, that's an error. */
2011         if (unlikely(head >= vq->num)) {
2012                 vq_err(vq, "Guest says index %u > %u is available",
2013                        head, vq->num);
2014                 return -EINVAL;
2015         }
2016
2017         /* When we start there are none of either input nor output. */
2018         *out_num = *in_num = 0;
2019         if (unlikely(log))
2020                 *log_num = 0;
2021
2022         i = head;
2023         do {
2024                 unsigned iov_count = *in_num + *out_num;
2025                 if (unlikely(i >= vq->num)) {
2026                         vq_err(vq, "Desc index is %u > %u, head = %u",
2027                                i, vq->num, head);
2028                         return -EINVAL;
2029                 }
2030                 if (unlikely(++found > vq->num)) {
2031                         vq_err(vq, "Loop detected: last one at %u "
2032                                "vq size %u head %u\n",
2033                                i, vq->num, head);
2034                         return -EINVAL;
2035                 }
2036                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2037                                            sizeof desc);
2038                 if (unlikely(ret)) {
2039                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2040                                i, vq->desc + i);
2041                         return -EFAULT;
2042                 }
2043                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2044                         ret = get_indirect(vq, iov, iov_size,
2045                                            out_num, in_num,
2046                                            log, log_num, &desc);
2047                         if (unlikely(ret < 0)) {
2048                                 if (ret != -EAGAIN)
2049                                         vq_err(vq, "Failure detected "
2050                                                 "in indirect descriptor at idx %d\n", i);
2051                                 return ret;
2052                         }
2053                         continue;
2054                 }
2055
2056                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2057                         access = VHOST_ACCESS_WO;
2058                 else
2059                         access = VHOST_ACCESS_RO;
2060                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2061                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2062                                      iov_size - iov_count, access);
2063                 if (unlikely(ret < 0)) {
2064                         if (ret != -EAGAIN)
2065                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2066                                         ret, i);
2067                         return ret;
2068                 }
2069                 if (access == VHOST_ACCESS_WO) {
2070                         /* If this is an input descriptor,
2071                          * increment that count. */
2072                         *in_num += ret;
2073                         if (unlikely(log)) {
2074                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2075                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2076                                 ++*log_num;
2077                         }
2078                 } else {
2079                         /* If it's an output descriptor, they're all supposed
2080                          * to come before any input descriptors. */
2081                         if (unlikely(*in_num)) {
2082                                 vq_err(vq, "Descriptor has out after in: "
2083                                        "idx %d\n", i);
2084                                 return -EINVAL;
2085                         }
2086                         *out_num += ret;
2087                 }
2088         } while ((i = next_desc(vq, &desc)) != -1);
2089
2090         /* On success, increment avail index. */
2091         vq->last_avail_idx++;
2092
2093         /* Assume notifications from guest are disabled at this point,
2094          * if they aren't we would need to update avail_event index. */
2095         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2096         return head;
2097 }
2098 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2099
2100 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2101 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2102 {
2103         vq->last_avail_idx -= n;
2104 }
2105 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2106
2107 /* After we've used one of their buffers, we tell them about it.  We'll then
2108  * want to notify the guest, using eventfd. */
2109 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2110 {
2111         struct vring_used_elem heads = {
2112                 cpu_to_vhost32(vq, head),
2113                 cpu_to_vhost32(vq, len)
2114         };
2115
2116         return vhost_add_used_n(vq, &heads, 1);
2117 }
2118 EXPORT_SYMBOL_GPL(vhost_add_used);
2119
2120 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2121                             struct vring_used_elem *heads,
2122                             unsigned count)
2123 {
2124         struct vring_used_elem __user *used;
2125         u16 old, new;
2126         int start;
2127
2128         start = vq->last_used_idx & (vq->num - 1);
2129         used = vq->used->ring + start;
2130         if (count == 1) {
2131                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2132                         vq_err(vq, "Failed to write used id");
2133                         return -EFAULT;
2134                 }
2135                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2136                         vq_err(vq, "Failed to write used len");
2137                         return -EFAULT;
2138                 }
2139         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2140                 vq_err(vq, "Failed to write used");
2141                 return -EFAULT;
2142         }
2143         if (unlikely(vq->log_used)) {
2144                 /* Make sure data is seen before log. */
2145                 smp_wmb();
2146                 /* Log used ring entry write. */
2147                 log_write(vq->log_base,
2148                           vq->log_addr +
2149                            ((void __user *)used - (void __user *)vq->used),
2150                           count * sizeof *used);
2151         }
2152         old = vq->last_used_idx;
2153         new = (vq->last_used_idx += count);
2154         /* If the driver never bothers to signal in a very long while,
2155          * used index might wrap around. If that happens, invalidate
2156          * signalled_used index we stored. TODO: make sure driver
2157          * signals at least once in 2^16 and remove this. */
2158         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2159                 vq->signalled_used_valid = false;
2160         return 0;
2161 }
2162
2163 /* After we've used one of their buffers, we tell them about it.  We'll then
2164  * want to notify the guest, using eventfd. */
2165 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2166                      unsigned count)
2167 {
2168         int start, n, r;
2169
2170         start = vq->last_used_idx & (vq->num - 1);
2171         n = vq->num - start;
2172         if (n < count) {
2173                 r = __vhost_add_used_n(vq, heads, n);
2174                 if (r < 0)
2175                         return r;
2176                 heads += n;
2177                 count -= n;
2178         }
2179         r = __vhost_add_used_n(vq, heads, count);
2180
2181         /* Make sure buffer is written before we update index. */
2182         smp_wmb();
2183         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2184                            &vq->used->idx)) {
2185                 vq_err(vq, "Failed to increment used idx");
2186                 return -EFAULT;
2187         }
2188         if (unlikely(vq->log_used)) {
2189                 /* Log used index update. */
2190                 log_write(vq->log_base,
2191                           vq->log_addr + offsetof(struct vring_used, idx),
2192                           sizeof vq->used->idx);
2193                 if (vq->log_ctx)
2194                         eventfd_signal(vq->log_ctx, 1);
2195         }
2196         return r;
2197 }
2198 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2199
2200 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2201 {
2202         __u16 old, new;
2203         __virtio16 event;
2204         bool v;
2205         /* Flush out used index updates. This is paired
2206          * with the barrier that the Guest executes when enabling
2207          * interrupts. */
2208         smp_mb();
2209
2210         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2211             unlikely(vq->avail_idx == vq->last_avail_idx))
2212                 return true;
2213
2214         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2215                 __virtio16 flags;
2216                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2217                         vq_err(vq, "Failed to get flags");
2218                         return true;
2219                 }
2220                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2221         }
2222         old = vq->signalled_used;
2223         v = vq->signalled_used_valid;
2224         new = vq->signalled_used = vq->last_used_idx;
2225         vq->signalled_used_valid = true;
2226
2227         if (unlikely(!v))
2228                 return true;
2229
2230         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2231                 vq_err(vq, "Failed to get used event idx");
2232                 return true;
2233         }
2234         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2235 }
2236
2237 /* This actually signals the guest, using eventfd. */
2238 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2239 {
2240         /* Signal the Guest tell them we used something up. */
2241         if (vq->call_ctx && vhost_notify(dev, vq))
2242                 eventfd_signal(vq->call_ctx, 1);
2243 }
2244 EXPORT_SYMBOL_GPL(vhost_signal);
2245
2246 /* And here's the combo meal deal.  Supersize me! */
2247 void vhost_add_used_and_signal(struct vhost_dev *dev,
2248                                struct vhost_virtqueue *vq,
2249                                unsigned int head, int len)
2250 {
2251         vhost_add_used(vq, head, len);
2252         vhost_signal(dev, vq);
2253 }
2254 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2255
2256 /* multi-buffer version of vhost_add_used_and_signal */
2257 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2258                                  struct vhost_virtqueue *vq,
2259                                  struct vring_used_elem *heads, unsigned count)
2260 {
2261         vhost_add_used_n(vq, heads, count);
2262         vhost_signal(dev, vq);
2263 }
2264 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2265
2266 /* return true if we're sure that avaiable ring is empty */
2267 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2268 {
2269         __virtio16 avail_idx;
2270         int r;
2271
2272         if (vq->avail_idx != vq->last_avail_idx)
2273                 return false;
2274
2275         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2276         if (unlikely(r))
2277                 return false;
2278         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2279
2280         return vq->avail_idx == vq->last_avail_idx;
2281 }
2282 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2283
2284 /* OK, now we need to know about added descriptors. */
2285 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2286 {
2287         __virtio16 avail_idx;
2288         int r;
2289
2290         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2291                 return false;
2292         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2293         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2294                 r = vhost_update_used_flags(vq);
2295                 if (r) {
2296                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2297                                &vq->used->flags, r);
2298                         return false;
2299                 }
2300         } else {
2301                 r = vhost_update_avail_event(vq, vq->avail_idx);
2302                 if (r) {
2303                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2304                                vhost_avail_event(vq), r);
2305                         return false;
2306                 }
2307         }
2308         /* They could have slipped one in as we were doing that: make
2309          * sure it's written, then check again. */
2310         smp_mb();
2311         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2312         if (r) {
2313                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2314                        &vq->avail->idx, r);
2315                 return false;
2316         }
2317
2318         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2319 }
2320 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2321
2322 /* We don't need to be notified again. */
2323 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2324 {
2325         int r;
2326
2327         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2328                 return;
2329         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2330         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2331                 r = vhost_update_used_flags(vq);
2332                 if (r)
2333                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2334                                &vq->used->flags, r);
2335         }
2336 }
2337 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2338
2339 /* Create a new message. */
2340 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2341 {
2342         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2343         if (!node)
2344                 return NULL;
2345         node->vq = vq;
2346         node->msg.type = type;
2347         return node;
2348 }
2349 EXPORT_SYMBOL_GPL(vhost_new_msg);
2350
2351 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2352                        struct vhost_msg_node *node)
2353 {
2354         spin_lock(&dev->iotlb_lock);
2355         list_add_tail(&node->node, head);
2356         spin_unlock(&dev->iotlb_lock);
2357
2358         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2359 }
2360 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2361
2362 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2363                                          struct list_head *head)
2364 {
2365         struct vhost_msg_node *node = NULL;
2366
2367         spin_lock(&dev->iotlb_lock);
2368         if (!list_empty(head)) {
2369                 node = list_first_entry(head, struct vhost_msg_node,
2370                                         node);
2371                 list_del(&node->node);
2372         }
2373         spin_unlock(&dev->iotlb_lock);
2374
2375         return node;
2376 }
2377 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2378
2379
2380 static int __init vhost_init(void)
2381 {
2382         return 0;
2383 }
2384
2385 static void __exit vhost_exit(void)
2386 {
2387 }
2388
2389 module_init(vhost_init);
2390 module_exit(vhost_exit);
2391
2392 MODULE_VERSION("0.0.1");
2393 MODULE_LICENSE("GPL v2");
2394 MODULE_AUTHOR("Michael S. Tsirkin");
2395 MODULE_DESCRIPTION("Host kernel accelerator for virtio");