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 (!((unsigned long)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         init_waitqueue_head(&work->done);
 185 }
 186 EXPORT_SYMBOL_GPL(vhost_work_init);
 187
 188 /* Init poll structure */
 189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 190                      unsigned long mask, struct vhost_dev *dev)
 191 {
 192         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 193         init_poll_funcptr(&poll->table, vhost_poll_func);
 194         poll->mask = mask;
 195         poll->dev = dev;
 196         poll->wqh = NULL;
 197
 198         vhost_work_init(&poll->work, fn);
 199 }
 200 EXPORT_SYMBOL_GPL(vhost_poll_init);
 201
 202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 203  * keep a reference to a file until after vhost_poll_stop is called. */
 204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 205 {
 206         unsigned long mask;
 207         int ret = 0;
 208
 209         if (poll->wqh)
 210                 return 0;
 211
 212         mask = file->f_op->poll(file, &poll->table);
 213         if (mask)
 214                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 215         if (mask & POLLERR) {
 216                 if (poll->wqh)
 217                         remove_wait_queue(poll->wqh, &poll->wait);
 218                 ret = -EINVAL;
 219         }
 220
 221         return ret;
 222 }
 223 EXPORT_SYMBOL_GPL(vhost_poll_start);
 224
 225 /* Stop polling a file. After this function returns, it becomes safe to drop the
 226  * file reference. You must also flush afterwards. */
 227 void vhost_poll_stop(struct vhost_poll *poll)
 228 {
 229         if (poll->wqh) {
 230                 remove_wait_queue(poll->wqh, &poll->wait);
 231                 poll->wqh = NULL;
 232         }
 233 }
 234 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 235
 236 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 237 {
 238         struct vhost_flush_struct flush;
 239
 240         if (dev->worker) {
 241                 init_completion(&flush.wait_event);
 242                 vhost_work_init(&flush.work, vhost_flush_work);
 243
 244                 vhost_work_queue(dev, &flush.work);
 245                 wait_for_completion(&flush.wait_event);
 246         }
 247 }
 248 EXPORT_SYMBOL_GPL(vhost_work_flush);
 249
 250 /* Flush any work that has been scheduled. When calling this, don't hold any
 251  * locks that are also used by the callback. */
 252 void vhost_poll_flush(struct vhost_poll *poll)
 253 {
 254         vhost_work_flush(poll->dev, &poll->work);
 255 }
 256 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 257
 258 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 259 {
 260         if (!dev->worker)
 261                 return;
 262
 263         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 264                 /* We can only add the work to the list after we're
 265                  * sure it was not in the list.
 266                  * test_and_set_bit() implies a memory barrier.
 267                  */
 268                 llist_add(&work->node, &dev->work_list);
 269                 wake_up_process(dev->worker);
 270         }
 271 }
 272 EXPORT_SYMBOL_GPL(vhost_work_queue);
 273
 274 /* A lockless hint for busy polling code to exit the loop */
 275 bool vhost_has_work(struct vhost_dev *dev)
 276 {
 277         return !llist_empty(&dev->work_list);
 278 }
 279 EXPORT_SYMBOL_GPL(vhost_has_work);
 280
 281 void vhost_poll_queue(struct vhost_poll *poll)
 282 {
 283         vhost_work_queue(poll->dev, &poll->work);
 284 }
 285 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 286
 287 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 288 {
 289         int j;
 290
 291         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 292                 vq->meta_iotlb[j] = NULL;
 293 }
 294
 295 static void vhost_vq_meta_reset(struct vhost_dev *d)
 296 {
 297         int i;
 298
 299         for (i = 0; i < d->nvqs; ++i)
 300                 __vhost_vq_meta_reset(d->vqs[i]);
 301 }
 302
 303 static void vhost_vq_reset(struct vhost_dev *dev,
 304                            struct vhost_virtqueue *vq)
 305 {
 306         vq->num = 1;
 307         vq->desc = NULL;
 308         vq->avail = NULL;
 309         vq->used = NULL;
 310         vq->last_avail_idx = 0;
 311         vq->last_used_event = 0;
 312         vq->avail_idx = 0;
 313         vq->last_used_idx = 0;
 314         vq->signalled_used = 0;
 315         vq->signalled_used_valid = false;
 316         vq->used_flags = 0;
 317         vq->log_used = false;
 318         vq->log_addr = -1ull;
 319         vq->private_data = NULL;
 320         vq->acked_features = 0;
 321         vq->log_base = NULL;
 322         vq->error_ctx = NULL;
 323         vq->error = NULL;
 324         vq->kick = NULL;
 325         vq->call_ctx = NULL;
 326         vq->call = NULL;
 327         vq->log_ctx = NULL;
 328         vhost_reset_is_le(vq);
 329         vhost_disable_cross_endian(vq);
 330         vq->busyloop_timeout = 0;
 331         vq->umem = NULL;
 332         vq->iotlb = NULL;
 333         __vhost_vq_meta_reset(vq);
 334 }
 335
 336 static int vhost_worker(void *data)
 337 {
 338         struct vhost_dev *dev = data;
 339         struct vhost_work *work, *work_next;
 340         struct llist_node *node;
 341         mm_segment_t oldfs = get_fs();
 342
 343         set_fs(USER_DS);
 344         use_mm(dev->mm);
 345
 346         for (;;) {
 347                 /* mb paired w/ kthread_stop */
 348                 set_current_state(TASK_INTERRUPTIBLE);
 349
 350                 if (kthread_should_stop()) {
 351                         __set_current_state(TASK_RUNNING);
 352                         break;
 353                 }
 354
 355                 node = llist_del_all(&dev->work_list);
 356                 if (!node)
 357                         schedule();
 358
 359                 node = llist_reverse_order(node);
 360                 /* make sure flag is seen after deletion */
 361                 smp_wmb();
 362                 llist_for_each_entry_safe(work, work_next, node, node) {
 363                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 364                         __set_current_state(TASK_RUNNING);
 365                         work->fn(work);
 366                         if (need_resched())
 367                                 schedule();
 368                 }
 369         }
 370         unuse_mm(dev->mm);
 371         set_fs(oldfs);
 372         return 0;
 373 }
 374
 375 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 376 {
 377         kfree(vq->indirect);
 378         vq->indirect = NULL;
 379         kfree(vq->log);
 380         vq->log = NULL;
 381         kfree(vq->heads);
 382         vq->heads = NULL;
 383 }
 384
 385 /* Helper to allocate iovec buffers for all vqs. */
 386 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 387 {
 388         struct vhost_virtqueue *vq;
 389         int i;
 390
 391         for (i = 0; i < dev->nvqs; ++i) {
 392                 vq = dev->vqs[i];
 393                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 394                                        GFP_KERNEL);
 395                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 396                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 397                 if (!vq->indirect || !vq->log || !vq->heads)
 398                         goto err_nomem;
 399         }
 400         return 0;
 401
 402 err_nomem:
 403         for (; i >= 0; --i)
 404                 vhost_vq_free_iovecs(dev->vqs[i]);
 405         return -ENOMEM;
 406 }
 407
 408 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 409 {
 410         int i;
 411
 412         for (i = 0; i < dev->nvqs; ++i)
 413                 vhost_vq_free_iovecs(dev->vqs[i]);
 414 }
 415
 416 void vhost_dev_init(struct vhost_dev *dev,
 417                     struct vhost_virtqueue **vqs, int nvqs)
 418 {
 419         struct vhost_virtqueue *vq;
 420         int i;
 421
 422         dev->vqs = vqs;
 423         dev->nvqs = nvqs;
 424         mutex_init(&dev->mutex);
 425         dev->log_ctx = NULL;
 426         dev->log_file = NULL;
 427         dev->umem = NULL;
 428         dev->iotlb = NULL;
 429         dev->mm = NULL;
 430         dev->worker = NULL;
 431         init_llist_head(&dev->work_list);
 432         init_waitqueue_head(&dev->wait);
 433         INIT_LIST_HEAD(&dev->read_list);
 434         INIT_LIST_HEAD(&dev->pending_list);
 435         spin_lock_init(&dev->iotlb_lock);
 436
 437
 438         for (i = 0; i < dev->nvqs; ++i) {
 439                 vq = dev->vqs[i];
 440                 vq->log = NULL;
 441                 vq->indirect = NULL;
 442                 vq->heads = NULL;
 443                 vq->dev = dev;
 444                 mutex_init(&vq->mutex);
 445                 vhost_vq_reset(dev, vq);
 446                 if (vq->handle_kick)
 447                         vhost_poll_init(&vq->poll, vq->handle_kick,
 448                                         POLLIN, dev);
 449         }
 450 }
 451 EXPORT_SYMBOL_GPL(vhost_dev_init);
 452
 453 /* Caller should have device mutex */
 454 long vhost_dev_check_owner(struct vhost_dev *dev)
 455 {
 456         /* Are you the owner? If not, I don't think you mean to do that */
 457         return dev->mm == current->mm ? 0 : -EPERM;
 458 }
 459 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 460
 461 struct vhost_attach_cgroups_struct {
 462         struct vhost_work work;
 463         struct task_struct *owner;
 464         int ret;
 465 };
 466
 467 static void vhost_attach_cgroups_work(struct vhost_work *work)
 468 {
 469         struct vhost_attach_cgroups_struct *s;
 470
 471         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 472         s->ret = cgroup_attach_task_all(s->owner, current);
 473 }
 474
 475 static int vhost_attach_cgroups(struct vhost_dev *dev)
 476 {
 477         struct vhost_attach_cgroups_struct attach;
 478
 479         attach.owner = current;
 480         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 481         vhost_work_queue(dev, &attach.work);
 482         vhost_work_flush(dev, &attach.work);
 483         return attach.ret;
 484 }
 485
 486 /* Caller should have device mutex */
 487 bool vhost_dev_has_owner(struct vhost_dev *dev)
 488 {
 489         return dev->mm;
 490 }
 491 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 492
 493 /* Caller should have device mutex */
 494 long vhost_dev_set_owner(struct vhost_dev *dev)
 495 {
 496         struct task_struct *worker;
 497         int err;
 498
 499         /* Is there an owner already? */
 500         if (vhost_dev_has_owner(dev)) {
 501                 err = -EBUSY;
 502                 goto err_mm;
 503         }
 504
 505         /* No owner, become one */
 506         dev->mm = get_task_mm(current);
 507         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 508         if (IS_ERR(worker)) {
 509                 err = PTR_ERR(worker);
 510                 goto err_worker;
 511         }
 512
 513         dev->worker = worker;
 514         wake_up_process(worker);        /* avoid contributing to loadavg */
 515
 516         err = vhost_attach_cgroups(dev);
 517         if (err)
 518                 goto err_cgroup;
 519
 520         err = vhost_dev_alloc_iovecs(dev);
 521         if (err)
 522                 goto err_cgroup;
 523
 524         return 0;
 525 err_cgroup:
 526         kthread_stop(worker);
 527         dev->worker = NULL;
 528 err_worker:
 529         if (dev->mm)
 530                 mmput(dev->mm);
 531         dev->mm = NULL;
 532 err_mm:
 533         return err;
 534 }
 535 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 536
 537 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 538 {
 539         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
 540 }
 541 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 542
 543 /* Caller should have device mutex */
 544 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 545 {
 546         int i;
 547
 548         vhost_dev_cleanup(dev, true);
 549
 550         /* Restore memory to default empty mapping. */
 551         INIT_LIST_HEAD(&umem->umem_list);
 552         dev->umem = umem;
 553         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 554          * VQs aren't running.
 555          */
 556         for (i = 0; i < dev->nvqs; ++i)
 557                 dev->vqs[i]->umem = umem;
 558 }
 559 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 560
 561 void vhost_dev_stop(struct vhost_dev *dev)
 562 {
 563         int i;
 564
 565         for (i = 0; i < dev->nvqs; ++i) {
 566                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 567                         vhost_poll_stop(&dev->vqs[i]->poll);
 568                         vhost_poll_flush(&dev->vqs[i]->poll);
 569                 }
 570         }
 571 }
 572 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 573
 574 static void vhost_umem_free(struct vhost_umem *umem,
 575                             struct vhost_umem_node *node)
 576 {
 577         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 578         list_del(&node->link);
 579         kfree(node);
 580         umem->numem--;
 581 }
 582
 583 static void vhost_umem_clean(struct vhost_umem *umem)
 584 {
 585         struct vhost_umem_node *node, *tmp;
 586
 587         if (!umem)
 588                 return;
 589
 590         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 591                 vhost_umem_free(umem, node);
 592
 593         kvfree(umem);
 594 }
 595
 596 static void vhost_clear_msg(struct vhost_dev *dev)
 597 {
 598         struct vhost_msg_node *node, *n;
 599
 600         spin_lock(&dev->iotlb_lock);
 601
 602         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 603                 list_del(&node->node);
 604                 kfree(node);
 605         }
 606
 607         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 608                 list_del(&node->node);
 609                 kfree(node);
 610         }
 611
 612         spin_unlock(&dev->iotlb_lock);
 613 }
 614
 615 /* Caller should have device mutex if and only if locked is set */
 616 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 617 {
 618         int i;
 619
 620         for (i = 0; i < dev->nvqs; ++i) {
 621                 if (dev->vqs[i]->error_ctx)
 622                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 623                 if (dev->vqs[i]->error)
 624                         fput(dev->vqs[i]->error);
 625                 if (dev->vqs[i]->kick)
 626                         fput(dev->vqs[i]->kick);
 627                 if (dev->vqs[i]->call_ctx)
 628                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 629                 if (dev->vqs[i]->call)
 630                         fput(dev->vqs[i]->call);
 631                 vhost_vq_reset(dev, dev->vqs[i]);
 632         }
 633         vhost_dev_free_iovecs(dev);
 634         if (dev->log_ctx)
 635                 eventfd_ctx_put(dev->log_ctx);
 636         dev->log_ctx = NULL;
 637         if (dev->log_file)
 638                 fput(dev->log_file);
 639         dev->log_file = NULL;
 640         /* No one will access memory at this point */
 641         vhost_umem_clean(dev->umem);
 642         dev->umem = NULL;
 643         vhost_umem_clean(dev->iotlb);
 644         dev->iotlb = NULL;
 645         vhost_clear_msg(dev);
 646         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
 647         WARN_ON(!llist_empty(&dev->work_list));
 648         if (dev->worker) {
 649                 kthread_stop(dev->worker);
 650                 dev->worker = NULL;
 651         }
 652         if (dev->mm)
 653                 mmput(dev->mm);
 654         dev->mm = NULL;
 655 }
 656 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 657
 658 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 659 {
 660         u64 a = addr / VHOST_PAGE_SIZE / 8;
 661
 662         /* Make sure 64 bit math will not overflow. */
 663         if (a > ULONG_MAX - (unsigned long)log_base ||
 664             a + (unsigned long)log_base > ULONG_MAX)
 665                 return 0;
 666
 667         return access_ok(VERIFY_WRITE, log_base + a,
 668                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 669 }
 670
 671 static bool vhost_overflow(u64 uaddr, u64 size)
 672 {
 673         /* Make sure 64 bit math will not overflow. */
 674         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 675 }
 676
 677 /* Caller should have vq mutex and device mutex. */
 678 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 679                                int log_all)
 680 {
 681         struct vhost_umem_node *node;
 682
 683         if (!umem)
 684                 return 0;
 685
 686         list_for_each_entry(node, &umem->umem_list, link) {
 687                 unsigned long a = node->userspace_addr;
 688
 689                 if (vhost_overflow(node->userspace_addr, node->size))
 690                         return 0;
 691
 692
 693                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 694                                     node->size))
 695                         return 0;
 696                 else if (log_all && !log_access_ok(log_base,
 697                                                    node->start,
 698                                                    node->size))
 699                         return 0;
 700         }
 701         return 1;
 702 }
 703
 704 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 705                                                u64 addr, unsigned int size,
 706                                                int type)
 707 {
 708         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 709
 710         if (!node)
 711                 return NULL;
 712
 713         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 714 }
 715
 716 /* Can we switch to this memory table? */
 717 /* Caller should have device mutex but not vq mutex */
 718 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 719                             int log_all)
 720 {
 721         int i;
 722
 723         for (i = 0; i < d->nvqs; ++i) {
 724                 int ok;
 725                 bool log;
 726
 727                 mutex_lock(&d->vqs[i]->mutex);
 728                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 729                 /* If ring is inactive, will check when it's enabled. */
 730                 if (d->vqs[i]->private_data)
 731                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 732                                                  umem, log);
 733                 else
 734                         ok = 1;
 735                 mutex_unlock(&d->vqs[i]->mutex);
 736                 if (!ok)
 737                         return 0;
 738         }
 739         return 1;
 740 }
 741
 742 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 743                           struct iovec iov[], int iov_size, int access);
 744
 745 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 746                               const void *from, unsigned size)
 747 {
 748         int ret;
 749
 750         if (!vq->iotlb)
 751                 return __copy_to_user(to, from, size);
 752         else {
 753                 /* This function should be called after iotlb
 754                  * prefetch, which means we're sure that all vq
 755                  * could be access through iotlb. So -EAGAIN should
 756                  * not happen in this case.
 757                  */
 758                 struct iov_iter t;
 759                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 760                                      (u64)(uintptr_t)to, size,
 761                                      VHOST_ADDR_DESC);
 762
 763                 if (uaddr)
 764                         return __copy_to_user(uaddr, from, size);
 765
 766                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 767                                      ARRAY_SIZE(vq->iotlb_iov),
 768                                      VHOST_ACCESS_WO);
 769                 if (ret < 0)
 770                         goto out;
 771                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 772                 ret = copy_to_iter(from, size, &t);
 773                 if (ret == size)
 774                         ret = 0;
 775         }
 776 out:
 777         return ret;
 778 }
 779
 780 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 781                                 void __user *from, unsigned size)
 782 {
 783         int ret;
 784
 785         if (!vq->iotlb)
 786                 return __copy_from_user(to, from, size);
 787         else {
 788                 /* This function should be called after iotlb
 789                  * prefetch, which means we're sure that vq
 790                  * could be access through iotlb. So -EAGAIN should
 791                  * not happen in this case.
 792                  */
 793                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 794                                      (u64)(uintptr_t)from, size,
 795                                      VHOST_ADDR_DESC);
 796                 struct iov_iter f;
 797
 798                 if (uaddr)
 799                         return __copy_from_user(to, uaddr, size);
 800
 801                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 802                                      ARRAY_SIZE(vq->iotlb_iov),
 803                                      VHOST_ACCESS_RO);
 804                 if (ret < 0) {
 805                         vq_err(vq, "IOTLB translation failure: uaddr "
 806                                "%p size 0x%llx\n", from,
 807                                (unsigned long long) size);
 808                         goto out;
 809                 }
 810                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 811                 ret = copy_from_iter(to, size, &f);
 812                 if (ret == size)
 813                         ret = 0;
 814         }
 815
 816 out:
 817         return ret;
 818 }
 819
 820 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 821                                           void __user *addr, unsigned int size,
 822                                           int type)
 823 {
 824         int ret;
 825
 826         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 827                              ARRAY_SIZE(vq->iotlb_iov),
 828                              VHOST_ACCESS_RO);
 829         if (ret < 0) {
 830                 vq_err(vq, "IOTLB translation failure: uaddr "
 831                         "%p size 0x%llx\n", addr,
 832                         (unsigned long long) size);
 833                 return NULL;
 834         }
 835
 836         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 837                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 838                         "%p size 0x%llx\n", addr,
 839                         (unsigned long long) size);
 840                 return NULL;
 841         }
 842
 843         return vq->iotlb_iov[0].iov_base;
 844 }
 845
 846 /* This function should be called after iotlb
 847  * prefetch, which means we're sure that vq
 848  * could be access through iotlb. So -EAGAIN should
 849  * not happen in this case.
 850  */
 851 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 852                                             void *addr, unsigned int size,
 853                                             int type)
 854 {
 855         void __user *uaddr = vhost_vq_meta_fetch(vq,
 856                              (u64)(uintptr_t)addr, size, type);
 857         if (uaddr)
 858                 return uaddr;
 859
 860         return __vhost_get_user_slow(vq, addr, size, type);
 861 }
 862
 863 #define vhost_put_user(vq, x, ptr)              \
 864 ({ \
 865         int ret = -EFAULT; \
 866         if (!vq->iotlb) { \
 867                 ret = __put_user(x, ptr); \
 868         } else { \
 869                 __typeof__(ptr) to = \
 870                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 871                                           sizeof(*ptr), VHOST_ADDR_USED); \
 872                 if (to != NULL) \
 873                         ret = __put_user(x, to); \
 874                 else \
 875                         ret = -EFAULT;  \
 876         } \
 877         ret; \
 878 })
 879
 880 #define vhost_get_user(vq, x, ptr, type)                \
 881 ({ \
 882         int ret; \
 883         if (!vq->iotlb) { \
 884                 ret = __get_user(x, ptr); \
 885         } else { \
 886                 __typeof__(ptr) from = \
 887                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 888                                                            sizeof(*ptr), \
 889                                                            type); \
 890                 if (from != NULL) \
 891                         ret = __get_user(x, from); \
 892                 else \
 893                         ret = -EFAULT; \
 894         } \
 895         ret; \
 896 })
 897
 898 #define vhost_get_avail(vq, x, ptr) \
 899         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 900
 901 #define vhost_get_used(vq, x, ptr) \
 902         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 903
 904 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 905 {
 906         int i = 0;
 907         for (i = 0; i < d->nvqs; ++i)
 908                 mutex_lock(&d->vqs[i]->mutex);
 909 }
 910
 911 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 912 {
 913         int i = 0;
 914         for (i = 0; i < d->nvqs; ++i)
 915                 mutex_unlock(&d->vqs[i]->mutex);
 916 }
 917
 918 static int vhost_new_umem_range(struct vhost_umem *umem,
 919                                 u64 start, u64 size, u64 end,
 920                                 u64 userspace_addr, int perm)
 921 {
 922         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
 923
 924         if (!node)
 925                 return -ENOMEM;
 926
 927         if (umem->numem == max_iotlb_entries) {
 928                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 929                 vhost_umem_free(umem, tmp);
 930         }
 931
 932         node->start = start;
 933         node->size = size;
 934         node->last = end;
 935         node->userspace_addr = userspace_addr;
 936         node->perm = perm;
 937         INIT_LIST_HEAD(&node->link);
 938         list_add_tail(&node->link, &umem->umem_list);
 939         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 940         umem->numem++;
 941
 942         return 0;
 943 }
 944
 945 static void vhost_del_umem_range(struct vhost_umem *umem,
 946                                  u64 start, u64 end)
 947 {
 948         struct vhost_umem_node *node;
 949
 950         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 951                                                            start, end)))
 952                 vhost_umem_free(umem, node);
 953 }
 954
 955 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 956                                   struct vhost_iotlb_msg *msg)
 957 {
 958         struct vhost_msg_node *node, *n;
 959
 960         spin_lock(&d->iotlb_lock);
 961
 962         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 963                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 964                 if (msg->iova <= vq_msg->iova &&
 965                     msg->iova + msg->size - 1 > vq_msg->iova &&
 966                     vq_msg->type == VHOST_IOTLB_MISS) {
 967                         vhost_poll_queue(&node->vq->poll);
 968                         list_del(&node->node);
 969                         kfree(node);
 970                 }
 971         }
 972
 973         spin_unlock(&d->iotlb_lock);
 974 }
 975
 976 static int umem_access_ok(u64 uaddr, u64 size, int access)
 977 {
 978         unsigned long a = uaddr;
 979
 980         /* Make sure 64 bit math will not overflow. */
 981         if (vhost_overflow(uaddr, size))
 982                 return -EFAULT;
 983
 984         if ((access & VHOST_ACCESS_RO) &&
 985             !access_ok(VERIFY_READ, (void __user *)a, size))
 986                 return -EFAULT;
 987         if ((access & VHOST_ACCESS_WO) &&
 988             !access_ok(VERIFY_WRITE, (void __user *)a, size))
 989                 return -EFAULT;
 990         return 0;
 991 }
 992
 993 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
 994                                    struct vhost_iotlb_msg *msg)
 995 {
 996         int ret = 0;
 997
 998         vhost_dev_lock_vqs(dev);
 999         switch (msg->type) {
1000         case VHOST_IOTLB_UPDATE:
1001                 if (!dev->iotlb) {
1002                         ret = -EFAULT;
1003                         break;
1004                 }
1005                 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1006                         ret = -EFAULT;
1007                         break;
1008                 }
1009                 vhost_vq_meta_reset(dev);
1010                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1011                                          msg->iova + msg->size - 1,
1012                                          msg->uaddr, msg->perm)) {
1013                         ret = -ENOMEM;
1014                         break;
1015                 }
1016                 vhost_iotlb_notify_vq(dev, msg);
1017                 break;
1018         case VHOST_IOTLB_INVALIDATE:
1019                 vhost_vq_meta_reset(dev);
1020                 vhost_del_umem_range(dev->iotlb, msg->iova,
1021                                      msg->iova + msg->size - 1);
1022                 break;
1023         default:
1024                 ret = -EINVAL;
1025                 break;
1026         }
1027
1028         vhost_dev_unlock_vqs(dev);
1029         return ret;
1030 }
1031 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1032                              struct iov_iter *from)
1033 {
1034         struct vhost_msg_node node;
1035         unsigned size = sizeof(struct vhost_msg);
1036         size_t ret;
1037         int err;
1038
1039         if (iov_iter_count(from) < size)
1040                 return 0;
1041         ret = copy_from_iter(&node.msg, size, from);
1042         if (ret != size)
1043                 goto done;
1044
1045         switch (node.msg.type) {
1046         case VHOST_IOTLB_MSG:
1047                 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1048                 if (err)
1049                         ret = err;
1050                 break;
1051         default:
1052                 ret = -EINVAL;
1053                 break;
1054         }
1055
1056 done:
1057         return ret;
1058 }
1059 EXPORT_SYMBOL(vhost_chr_write_iter);
1060
1061 unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1062                             poll_table *wait)
1063 {
1064         unsigned int mask = 0;
1065
1066         poll_wait(file, &dev->wait, wait);
1067
1068         if (!list_empty(&dev->read_list))
1069                 mask |= POLLIN | POLLRDNORM;
1070
1071         return mask;
1072 }
1073 EXPORT_SYMBOL(vhost_chr_poll);
1074
1075 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1076                             int noblock)
1077 {
1078         DEFINE_WAIT(wait);
1079         struct vhost_msg_node *node;
1080         ssize_t ret = 0;
1081         unsigned size = sizeof(struct vhost_msg);
1082
1083         if (iov_iter_count(to) < size)
1084                 return 0;
1085
1086         while (1) {
1087                 if (!noblock)
1088                         prepare_to_wait(&dev->wait, &wait,
1089                                         TASK_INTERRUPTIBLE);
1090
1091                 node = vhost_dequeue_msg(dev, &dev->read_list);
1092                 if (node)
1093                         break;
1094                 if (noblock) {
1095                         ret = -EAGAIN;
1096                         break;
1097                 }
1098                 if (signal_pending(current)) {
1099                         ret = -ERESTARTSYS;
1100                         break;
1101                 }
1102                 if (!dev->iotlb) {
1103                         ret = -EBADFD;
1104                         break;
1105                 }
1106
1107                 schedule();
1108         }
1109
1110         if (!noblock)
1111                 finish_wait(&dev->wait, &wait);
1112
1113         if (node) {
1114                 ret = copy_to_iter(&node->msg, size, to);
1115
1116                 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1117                         kfree(node);
1118                         return ret;
1119                 }
1120
1121                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1122         }
1123
1124         return ret;
1125 }
1126 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1127
1128 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1129 {
1130         struct vhost_dev *dev = vq->dev;
1131         struct vhost_msg_node *node;
1132         struct vhost_iotlb_msg *msg;
1133
1134         node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1135         if (!node)
1136                 return -ENOMEM;
1137
1138         msg = &node->msg.iotlb;
1139         msg->type = VHOST_IOTLB_MISS;
1140         msg->iova = iova;
1141         msg->perm = access;
1142
1143         vhost_enqueue_msg(dev, &dev->read_list, node);
1144
1145         return 0;
1146 }
1147
1148 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1149                         struct vring_desc __user *desc,
1150                         struct vring_avail __user *avail,
1151                         struct vring_used __user *used)
1152
1153 {
1154         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1155
1156         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1157                access_ok(VERIFY_READ, avail,
1158                          sizeof *avail + num * sizeof *avail->ring + s) &&
1159                access_ok(VERIFY_WRITE, used,
1160                         sizeof *used + num * sizeof *used->ring + s);
1161 }
1162
1163 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1164                                  const struct vhost_umem_node *node,
1165                                  int type)
1166 {
1167         int access = (type == VHOST_ADDR_USED) ?
1168                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1169
1170         if (likely(node->perm & access))
1171                 vq->meta_iotlb[type] = node;
1172 }
1173
1174 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1175                            int access, u64 addr, u64 len, int type)
1176 {
1177         const struct vhost_umem_node *node;
1178         struct vhost_umem *umem = vq->iotlb;
1179         u64 s = 0, size, orig_addr = addr;
1180
1181         if (vhost_vq_meta_fetch(vq, addr, len, type))
1182                 return true;
1183
1184         while (len > s) {
1185                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1186                                                            addr,
1187                                                            addr + len - 1);
1188                 if (node == NULL || node->start > addr) {
1189                         vhost_iotlb_miss(vq, addr, access);
1190                         return false;
1191                 } else if (!(node->perm & access)) {
1192                         /* Report the possible access violation by
1193                          * request another translation from userspace.
1194                          */
1195                         return false;
1196                 }
1197
1198                 size = node->size - addr + node->start;
1199
1200                 if (orig_addr == addr && size >= len)
1201                         vhost_vq_meta_update(vq, node, type);
1202
1203                 s += size;
1204                 addr += size;
1205         }
1206
1207         return true;
1208 }
1209
1210 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1211 {
1212         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1213         unsigned int num = vq->num;
1214
1215         if (!vq->iotlb)
1216                 return 1;
1217
1218         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1219                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1220                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1221                                sizeof *vq->avail +
1222                                num * sizeof(*vq->avail->ring) + s,
1223                                VHOST_ADDR_AVAIL) &&
1224                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1225                                sizeof *vq->used +
1226                                num * sizeof(*vq->used->ring) + s,
1227                                VHOST_ADDR_USED);
1228 }
1229 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1230
1231 /* Can we log writes? */
1232 /* Caller should have device mutex but not vq mutex */
1233 int vhost_log_access_ok(struct vhost_dev *dev)
1234 {
1235         return memory_access_ok(dev, dev->umem, 1);
1236 }
1237 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1238
1239 /* Verify access for write logging. */
1240 /* Caller should have vq mutex and device mutex */
1241 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1242                             void __user *log_base)
1243 {
1244         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1245
1246         return vq_memory_access_ok(log_base, vq->umem,
1247                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1248                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1249                                         sizeof *vq->used +
1250                                         vq->num * sizeof *vq->used->ring + s));
1251 }
1252
1253 /* Can we start vq? */
1254 /* Caller should have vq mutex and device mutex */
1255 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1256 {
1257         if (vq->iotlb) {
1258                 /* When device IOTLB was used, the access validation
1259                  * will be validated during prefetching.
1260                  */
1261                 return 1;
1262         }
1263         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1264                 vq_log_access_ok(vq, vq->log_base);
1265 }
1266 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1267
1268 static struct vhost_umem *vhost_umem_alloc(void)
1269 {
1270         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1271
1272         if (!umem)
1273                 return NULL;
1274
1275         umem->umem_tree = RB_ROOT;
1276         umem->numem = 0;
1277         INIT_LIST_HEAD(&umem->umem_list);
1278
1279         return umem;
1280 }
1281
1282 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1283 {
1284         struct vhost_memory mem, *newmem;
1285         struct vhost_memory_region *region;
1286         struct vhost_umem *newumem, *oldumem;
1287         unsigned long size = offsetof(struct vhost_memory, regions);
1288         int i;
1289
1290         if (copy_from_user(&mem, m, size))
1291                 return -EFAULT;
1292         if (mem.padding)
1293                 return -EOPNOTSUPP;
1294         if (mem.nregions > max_mem_regions)
1295                 return -E2BIG;
1296         newmem = kvzalloc(size + mem.nregions * sizeof(*m->regions), GFP_KERNEL);
1297         if (!newmem)
1298                 return -ENOMEM;
1299
1300         memcpy(newmem, &mem, size);
1301         if (copy_from_user(newmem->regions, m->regions,
1302                            mem.nregions * sizeof *m->regions)) {
1303                 kvfree(newmem);
1304                 return -EFAULT;
1305         }
1306
1307         newumem = vhost_umem_alloc();
1308         if (!newumem) {
1309                 kvfree(newmem);
1310                 return -ENOMEM;
1311         }
1312
1313         for (region = newmem->regions;
1314              region < newmem->regions + mem.nregions;
1315              region++) {
1316                 if (vhost_new_umem_range(newumem,
1317                                          region->guest_phys_addr,
1318                                          region->memory_size,
1319                                          region->guest_phys_addr +
1320                                          region->memory_size - 1,
1321                                          region->userspace_addr,
1322                                          VHOST_ACCESS_RW))
1323                         goto err;
1324         }
1325
1326         if (!memory_access_ok(d, newumem, 0))
1327                 goto err;
1328
1329         oldumem = d->umem;
1330         d->umem = newumem;
1331
1332         /* All memory accesses are done under some VQ mutex. */
1333         for (i = 0; i < d->nvqs; ++i) {
1334                 mutex_lock(&d->vqs[i]->mutex);
1335                 d->vqs[i]->umem = newumem;
1336                 mutex_unlock(&d->vqs[i]->mutex);
1337         }
1338
1339         kvfree(newmem);
1340         vhost_umem_clean(oldumem);
1341         return 0;
1342
1343 err:
1344         vhost_umem_clean(newumem);
1345         kvfree(newmem);
1346         return -EFAULT;
1347 }
1348
1349 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1350 {
1351         struct file *eventfp, *filep = NULL;
1352         bool pollstart = false, pollstop = false;
1353         struct eventfd_ctx *ctx = NULL;
1354         u32 __user *idxp = argp;
1355         struct vhost_virtqueue *vq;
1356         struct vhost_vring_state s;
1357         struct vhost_vring_file f;
1358         struct vhost_vring_addr a;
1359         u32 idx;
1360         long r;
1361
1362         r = get_user(idx, idxp);
1363         if (r < 0)
1364                 return r;
1365         if (idx >= d->nvqs)
1366                 return -ENOBUFS;
1367
1368         vq = d->vqs[idx];
1369
1370         mutex_lock(&vq->mutex);
1371
1372         switch (ioctl) {
1373         case VHOST_SET_VRING_NUM:
1374                 /* Resizing ring with an active backend?
1375                  * You don't want to do that. */
1376                 if (vq->private_data) {
1377                         r = -EBUSY;
1378                         break;
1379                 }
1380                 if (copy_from_user(&s, argp, sizeof s)) {
1381                         r = -EFAULT;
1382                         break;
1383                 }
1384                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1385                         r = -EINVAL;
1386                         break;
1387                 }
1388                 vq->num = s.num;
1389                 break;
1390         case VHOST_SET_VRING_BASE:
1391                 /* Moving base with an active backend?
1392                  * You don't want to do that. */
1393                 if (vq->private_data) {
1394                         r = -EBUSY;
1395                         break;
1396                 }
1397                 if (copy_from_user(&s, argp, sizeof s)) {
1398                         r = -EFAULT;
1399                         break;
1400                 }
1401                 if (s.num > 0xffff) {
1402                         r = -EINVAL;
1403                         break;
1404                 }
1405                 vq->last_avail_idx = vq->last_used_event = s.num;
1406                 /* Forget the cached index value. */
1407                 vq->avail_idx = vq->last_avail_idx;
1408                 break;
1409         case VHOST_GET_VRING_BASE:
1410                 s.index = idx;
1411                 s.num = vq->last_avail_idx;
1412                 if (copy_to_user(argp, &s, sizeof s))
1413                         r = -EFAULT;
1414                 break;
1415         case VHOST_SET_VRING_ADDR:
1416                 if (copy_from_user(&a, argp, sizeof a)) {
1417                         r = -EFAULT;
1418                         break;
1419                 }
1420                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1421                         r = -EOPNOTSUPP;
1422                         break;
1423                 }
1424                 /* For 32bit, verify that the top 32bits of the user
1425                    data are set to zero. */
1426                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1427                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1428                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1429                         r = -EFAULT;
1430                         break;
1431                 }
1432
1433                 /* Make sure it's safe to cast pointers to vring types. */
1434                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1435                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1436                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1437                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1438                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1439                         r = -EINVAL;
1440                         break;
1441                 }
1442
1443                 /* We only verify access here if backend is configured.
1444                  * If it is not, we don't as size might not have been setup.
1445                  * We will verify when backend is configured. */
1446                 if (vq->private_data) {
1447                         if (!vq_access_ok(vq, vq->num,
1448                                 (void __user *)(unsigned long)a.desc_user_addr,
1449                                 (void __user *)(unsigned long)a.avail_user_addr,
1450                                 (void __user *)(unsigned long)a.used_user_addr)) {
1451                                 r = -EINVAL;
1452                                 break;
1453                         }
1454
1455                         /* Also validate log access for used ring if enabled. */
1456                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1457                             !log_access_ok(vq->log_base, a.log_guest_addr,
1458                                            sizeof *vq->used +
1459                                            vq->num * sizeof *vq->used->ring)) {
1460                                 r = -EINVAL;
1461                                 break;
1462                         }
1463                 }
1464
1465                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1466                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1467                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1468                 vq->log_addr = a.log_guest_addr;
1469                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1470                 break;
1471         case VHOST_SET_VRING_KICK:
1472                 if (copy_from_user(&f, argp, sizeof f)) {
1473                         r = -EFAULT;
1474                         break;
1475                 }
1476                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1477                 if (IS_ERR(eventfp)) {
1478                         r = PTR_ERR(eventfp);
1479                         break;
1480                 }
1481                 if (eventfp != vq->kick) {
1482                         pollstop = (filep = vq->kick) != NULL;
1483                         pollstart = (vq->kick = eventfp) != NULL;
1484                 } else
1485                         filep = eventfp;
1486                 break;
1487         case VHOST_SET_VRING_CALL:
1488                 if (copy_from_user(&f, argp, sizeof f)) {
1489                         r = -EFAULT;
1490                         break;
1491                 }
1492                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1493                 if (IS_ERR(eventfp)) {
1494                         r = PTR_ERR(eventfp);
1495                         break;
1496                 }
1497                 if (eventfp != vq->call) {
1498                         filep = vq->call;
1499                         ctx = vq->call_ctx;
1500                         vq->call = eventfp;
1501                         vq->call_ctx = eventfp ?
1502                                 eventfd_ctx_fileget(eventfp) : NULL;
1503                 } else
1504                         filep = eventfp;
1505                 break;
1506         case VHOST_SET_VRING_ERR:
1507                 if (copy_from_user(&f, argp, sizeof f)) {
1508                         r = -EFAULT;
1509                         break;
1510                 }
1511                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1512                 if (IS_ERR(eventfp)) {
1513                         r = PTR_ERR(eventfp);
1514                         break;
1515                 }
1516                 if (eventfp != vq->error) {
1517                         filep = vq->error;
1518                         vq->error = eventfp;
1519                         ctx = vq->error_ctx;
1520                         vq->error_ctx = eventfp ?
1521                                 eventfd_ctx_fileget(eventfp) : NULL;
1522                 } else
1523                         filep = eventfp;
1524                 break;
1525         case VHOST_SET_VRING_ENDIAN:
1526                 r = vhost_set_vring_endian(vq, argp);
1527                 break;
1528         case VHOST_GET_VRING_ENDIAN:
1529                 r = vhost_get_vring_endian(vq, idx, argp);
1530                 break;
1531         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1532                 if (copy_from_user(&s, argp, sizeof(s))) {
1533                         r = -EFAULT;
1534                         break;
1535                 }
1536                 vq->busyloop_timeout = s.num;
1537                 break;
1538         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1539                 s.index = idx;
1540                 s.num = vq->busyloop_timeout;
1541                 if (copy_to_user(argp, &s, sizeof(s)))
1542                         r = -EFAULT;
1543                 break;
1544         default:
1545                 r = -ENOIOCTLCMD;
1546         }
1547
1548         if (pollstop && vq->handle_kick)
1549                 vhost_poll_stop(&vq->poll);
1550
1551         if (ctx)
1552                 eventfd_ctx_put(ctx);
1553         if (filep)
1554                 fput(filep);
1555
1556         if (pollstart && vq->handle_kick)
1557                 r = vhost_poll_start(&vq->poll, vq->kick);
1558
1559         mutex_unlock(&vq->mutex);
1560
1561         if (pollstop && vq->handle_kick)
1562                 vhost_poll_flush(&vq->poll);
1563         return r;
1564 }
1565 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1566
1567 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1568 {
1569         struct vhost_umem *niotlb, *oiotlb;
1570         int i;
1571
1572         niotlb = vhost_umem_alloc();
1573         if (!niotlb)
1574                 return -ENOMEM;
1575
1576         oiotlb = d->iotlb;
1577         d->iotlb = niotlb;
1578
1579         for (i = 0; i < d->nvqs; ++i) {
1580                 mutex_lock(&d->vqs[i]->mutex);
1581                 d->vqs[i]->iotlb = niotlb;
1582                 mutex_unlock(&d->vqs[i]->mutex);
1583         }
1584
1585         vhost_umem_clean(oiotlb);
1586
1587         return 0;
1588 }
1589 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1590
1591 /* Caller must have device mutex */
1592 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1593 {
1594         struct file *eventfp, *filep = NULL;
1595         struct eventfd_ctx *ctx = NULL;
1596         u64 p;
1597         long r;
1598         int i, fd;
1599
1600         /* If you are not the owner, you can become one */
1601         if (ioctl == VHOST_SET_OWNER) {
1602                 r = vhost_dev_set_owner(d);
1603                 goto done;
1604         }
1605
1606         /* You must be the owner to do anything else */
1607         r = vhost_dev_check_owner(d);
1608         if (r)
1609                 goto done;
1610
1611         switch (ioctl) {
1612         case VHOST_SET_MEM_TABLE:
1613                 r = vhost_set_memory(d, argp);
1614                 break;
1615         case VHOST_SET_LOG_BASE:
1616                 if (copy_from_user(&p, argp, sizeof p)) {
1617                         r = -EFAULT;
1618                         break;
1619                 }
1620                 if ((u64)(unsigned long)p != p) {
1621                         r = -EFAULT;
1622                         break;
1623                 }
1624                 for (i = 0; i < d->nvqs; ++i) {
1625                         struct vhost_virtqueue *vq;
1626                         void __user *base = (void __user *)(unsigned long)p;
1627                         vq = d->vqs[i];
1628                         mutex_lock(&vq->mutex);
1629                         /* If ring is inactive, will check when it's enabled. */
1630                         if (vq->private_data && !vq_log_access_ok(vq, base))
1631                                 r = -EFAULT;
1632                         else
1633                                 vq->log_base = base;
1634                         mutex_unlock(&vq->mutex);
1635                 }
1636                 break;
1637         case VHOST_SET_LOG_FD:
1638                 r = get_user(fd, (int __user *)argp);
1639                 if (r < 0)
1640                         break;
1641                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1642                 if (IS_ERR(eventfp)) {
1643                         r = PTR_ERR(eventfp);
1644                         break;
1645                 }
1646                 if (eventfp != d->log_file) {
1647                         filep = d->log_file;
1648                         d->log_file = eventfp;
1649                         ctx = d->log_ctx;
1650                         d->log_ctx = eventfp ?
1651                                 eventfd_ctx_fileget(eventfp) : NULL;
1652                 } else
1653                         filep = eventfp;
1654                 for (i = 0; i < d->nvqs; ++i) {
1655                         mutex_lock(&d->vqs[i]->mutex);
1656                         d->vqs[i]->log_ctx = d->log_ctx;
1657                         mutex_unlock(&d->vqs[i]->mutex);
1658                 }
1659                 if (ctx)
1660                         eventfd_ctx_put(ctx);
1661                 if (filep)
1662                         fput(filep);
1663                 break;
1664         default:
1665                 r = -ENOIOCTLCMD;
1666                 break;
1667         }
1668 done:
1669         return r;
1670 }
1671 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1672
1673 /* TODO: This is really inefficient.  We need something like get_user()
1674  * (instruction directly accesses the data, with an exception table entry
1675  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1676  */
1677 static int set_bit_to_user(int nr, void __user *addr)
1678 {
1679         unsigned long log = (unsigned long)addr;
1680         struct page *page;
1681         void *base;
1682         int bit = nr + (log % PAGE_SIZE) * 8;
1683         int r;
1684
1685         r = get_user_pages_fast(log, 1, 1, &page);
1686         if (r < 0)
1687                 return r;
1688         BUG_ON(r != 1);
1689         base = kmap_atomic(page);
1690         set_bit(bit, base);
1691         kunmap_atomic(base);
1692         set_page_dirty_lock(page);
1693         put_page(page);
1694         return 0;
1695 }
1696
1697 static int log_write(void __user *log_base,
1698                      u64 write_address, u64 write_length)
1699 {
1700         u64 write_page = write_address / VHOST_PAGE_SIZE;
1701         int r;
1702
1703         if (!write_length)
1704                 return 0;
1705         write_length += write_address % VHOST_PAGE_SIZE;
1706         for (;;) {
1707                 u64 base = (u64)(unsigned long)log_base;
1708                 u64 log = base + write_page / 8;
1709                 int bit = write_page % 8;
1710                 if ((u64)(unsigned long)log != log)
1711                         return -EFAULT;
1712                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1713                 if (r < 0)
1714                         return r;
1715                 if (write_length <= VHOST_PAGE_SIZE)
1716                         break;
1717                 write_length -= VHOST_PAGE_SIZE;
1718                 write_page += 1;
1719         }
1720         return r;
1721 }
1722
1723 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1724                     unsigned int log_num, u64 len)
1725 {
1726         int i, r;
1727
1728         /* Make sure data written is seen before log. */
1729         smp_wmb();
1730         for (i = 0; i < log_num; ++i) {
1731                 u64 l = min(log[i].len, len);
1732                 r = log_write(vq->log_base, log[i].addr, l);
1733                 if (r < 0)
1734                         return r;
1735                 len -= l;
1736                 if (!len) {
1737                         if (vq->log_ctx)
1738                                 eventfd_signal(vq->log_ctx, 1);
1739                         return 0;
1740                 }
1741         }
1742         /* Length written exceeds what we have stored. This is a bug. */
1743         BUG();
1744         return 0;
1745 }
1746 EXPORT_SYMBOL_GPL(vhost_log_write);
1747
1748 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1749 {
1750         void __user *used;
1751         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1752                            &vq->used->flags) < 0)
1753                 return -EFAULT;
1754         if (unlikely(vq->log_used)) {
1755                 /* Make sure the flag is seen before log. */
1756                 smp_wmb();
1757                 /* Log used flag write. */
1758                 used = &vq->used->flags;
1759                 log_write(vq->log_base, vq->log_addr +
1760                           (used - (void __user *)vq->used),
1761                           sizeof vq->used->flags);
1762                 if (vq->log_ctx)
1763                         eventfd_signal(vq->log_ctx, 1);
1764         }
1765         return 0;
1766 }
1767
1768 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1769 {
1770         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1771                            vhost_avail_event(vq)))
1772                 return -EFAULT;
1773         if (unlikely(vq->log_used)) {
1774                 void __user *used;
1775                 /* Make sure the event is seen before log. */
1776                 smp_wmb();
1777                 /* Log avail event write */
1778                 used = vhost_avail_event(vq);
1779                 log_write(vq->log_base, vq->log_addr +
1780                           (used - (void __user *)vq->used),
1781                           sizeof *vhost_avail_event(vq));
1782                 if (vq->log_ctx)
1783                         eventfd_signal(vq->log_ctx, 1);
1784         }
1785         return 0;
1786 }
1787
1788 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1789 {
1790         __virtio16 last_used_idx;
1791         int r;
1792         bool is_le = vq->is_le;
1793
1794         if (!vq->private_data)
1795                 return 0;
1796
1797         vhost_init_is_le(vq);
1798
1799         r = vhost_update_used_flags(vq);
1800         if (r)
1801                 goto err;
1802         vq->signalled_used_valid = false;
1803         if (!vq->iotlb &&
1804             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1805                 r = -EFAULT;
1806                 goto err;
1807         }
1808         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1809         if (r) {
1810                 vq_err(vq, "Can't access used idx at %p\n",
1811                        &vq->used->idx);
1812                 goto err;
1813         }
1814         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1815         return 0;
1816
1817 err:
1818         vq->is_le = is_le;
1819         return r;
1820 }
1821 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1822
1823 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1824                           struct iovec iov[], int iov_size, int access)
1825 {
1826         const struct vhost_umem_node *node;
1827         struct vhost_dev *dev = vq->dev;
1828         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1829         struct iovec *_iov;
1830         u64 s = 0;
1831         int ret = 0;
1832
1833         while ((u64)len > s) {
1834                 u64 size;
1835                 if (unlikely(ret >= iov_size)) {
1836                         ret = -ENOBUFS;
1837                         break;
1838                 }
1839
1840                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1841                                                         addr, addr + len - 1);
1842                 if (node == NULL || node->start > addr) {
1843                         if (umem != dev->iotlb) {
1844                                 ret = -EFAULT;
1845                                 break;
1846                         }
1847                         ret = -EAGAIN;
1848                         break;
1849                 } else if (!(node->perm & access)) {
1850                         ret = -EPERM;
1851                         break;
1852                 }
1853
1854                 _iov = iov + ret;
1855                 size = node->size - addr + node->start;
1856                 _iov->iov_len = min((u64)len - s, size);
1857                 _iov->iov_base = (void __user *)(unsigned long)
1858                         (node->userspace_addr + addr - node->start);
1859                 s += size;
1860                 addr += size;
1861                 ++ret;
1862         }
1863
1864         if (ret == -EAGAIN)
1865                 vhost_iotlb_miss(vq, addr, access);
1866         return ret;
1867 }
1868
1869 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1870  * function returns the next descriptor in the chain,
1871  * or -1U if we're at the end. */
1872 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1873 {
1874         unsigned int next;
1875
1876         /* If this descriptor says it doesn't chain, we're done. */
1877         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1878                 return -1U;
1879
1880         /* Check they're not leading us off end of descriptors. */
1881         next = vhost16_to_cpu(vq, desc->next);
1882         /* Make sure compiler knows to grab that: we don't want it changing! */
1883         /* We will use the result as an index in an array, so most
1884          * architectures only need a compiler barrier here. */
1885         read_barrier_depends();
1886
1887         return next;
1888 }
1889
1890 static int get_indirect(struct vhost_virtqueue *vq,
1891                         struct iovec iov[], unsigned int iov_size,
1892                         unsigned int *out_num, unsigned int *in_num,
1893                         struct vhost_log *log, unsigned int *log_num,
1894                         struct vring_desc *indirect)
1895 {
1896         struct vring_desc desc;
1897         unsigned int i = 0, count, found = 0;
1898         u32 len = vhost32_to_cpu(vq, indirect->len);
1899         struct iov_iter from;
1900         int ret, access;
1901
1902         /* Sanity check */
1903         if (unlikely(len % sizeof desc)) {
1904                 vq_err(vq, "Invalid length in indirect descriptor: "
1905                        "len 0x%llx not multiple of 0x%zx\n",
1906                        (unsigned long long)len,
1907                        sizeof desc);
1908                 return -EINVAL;
1909         }
1910
1911         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1912                              UIO_MAXIOV, VHOST_ACCESS_RO);
1913         if (unlikely(ret < 0)) {
1914                 if (ret != -EAGAIN)
1915                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1916                 return ret;
1917         }
1918         iov_iter_init(&from, READ, vq->indirect, ret, len);
1919
1920         /* We will use the result as an address to read from, so most
1921          * architectures only need a compiler barrier here. */
1922         read_barrier_depends();
1923
1924         count = len / sizeof desc;
1925         /* Buffers are chained via a 16 bit next field, so
1926          * we can have at most 2^16 of these. */
1927         if (unlikely(count > USHRT_MAX + 1)) {
1928                 vq_err(vq, "Indirect buffer length too big: %d\n",
1929                        indirect->len);
1930                 return -E2BIG;
1931         }
1932
1933         do {
1934                 unsigned iov_count = *in_num + *out_num;
1935                 if (unlikely(++found > count)) {
1936                         vq_err(vq, "Loop detected: last one at %u "
1937                                "indirect size %u\n",
1938                                i, count);
1939                         return -EINVAL;
1940                 }
1941                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1942                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1943                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1944                         return -EINVAL;
1945                 }
1946                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1947                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1948                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1949                         return -EINVAL;
1950                 }
1951
1952                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1953                         access = VHOST_ACCESS_WO;
1954                 else
1955                         access = VHOST_ACCESS_RO;
1956
1957                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1958                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1959                                      iov_size - iov_count, access);
1960                 if (unlikely(ret < 0)) {
1961                         if (ret != -EAGAIN)
1962                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1963                                         ret, i);
1964                         return ret;
1965                 }
1966                 /* If this is an input descriptor, increment that count. */
1967                 if (access == VHOST_ACCESS_WO) {
1968                         *in_num += ret;
1969                         if (unlikely(log)) {
1970                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1971                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1972                                 ++*log_num;
1973                         }
1974                 } else {
1975                         /* If it's an output descriptor, they're all supposed
1976                          * to come before any input descriptors. */
1977                         if (unlikely(*in_num)) {
1978                                 vq_err(vq, "Indirect descriptor "
1979                                        "has out after in: idx %d\n", i);
1980                                 return -EINVAL;
1981                         }
1982                         *out_num += ret;
1983                 }
1984         } while ((i = next_desc(vq, &desc)) != -1);
1985         return 0;
1986 }
1987
1988 /* This looks in the virtqueue and for the first available buffer, and converts
1989  * it to an iovec for convenient access.  Since descriptors consist of some
1990  * number of output then some number of input descriptors, it's actually two
1991  * iovecs, but we pack them into one and note how many of each there were.
1992  *
1993  * This function returns the descriptor number found, or vq->num (which is
1994  * never a valid descriptor number) if none was found.  A negative code is
1995  * returned on error. */
1996 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1997                       struct iovec iov[], unsigned int iov_size,
1998                       unsigned int *out_num, unsigned int *in_num,
1999                       struct vhost_log *log, unsigned int *log_num)
2000 {
2001         struct vring_desc desc;
2002         unsigned int i, head, found = 0;
2003         u16 last_avail_idx;
2004         __virtio16 avail_idx;
2005         __virtio16 ring_head;
2006         int ret, access;
2007
2008         /* Check it isn't doing very strange things with descriptor numbers. */
2009         last_avail_idx = vq->last_avail_idx;
2010
2011         if (vq->avail_idx == vq->last_avail_idx) {
2012                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2013                         vq_err(vq, "Failed to access avail idx at %p\n",
2014                                 &vq->avail->idx);
2015                         return -EFAULT;
2016                 }
2017                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2018
2019                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2020                         vq_err(vq, "Guest moved used index from %u to %u",
2021                                 last_avail_idx, vq->avail_idx);
2022                         return -EFAULT;
2023                 }
2024
2025                 /* If there's nothing new since last we looked, return
2026                  * invalid.
2027                  */
2028                 if (vq->avail_idx == last_avail_idx)
2029                         return vq->num;
2030
2031                 /* Only get avail ring entries after they have been
2032                  * exposed by guest.
2033                  */
2034                 smp_rmb();
2035         }
2036
2037         /* Grab the next descriptor number they're advertising, and increment
2038          * the index we've seen. */
2039         if (unlikely(vhost_get_avail(vq, ring_head,
2040                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2041                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2042                        last_avail_idx,
2043                        &vq->avail->ring[last_avail_idx % vq->num]);
2044                 return -EFAULT;
2045         }
2046
2047         head = vhost16_to_cpu(vq, ring_head);
2048
2049         /* If their number is silly, that's an error. */
2050         if (unlikely(head >= vq->num)) {
2051                 vq_err(vq, "Guest says index %u > %u is available",
2052                        head, vq->num);
2053                 return -EINVAL;
2054         }
2055
2056         /* When we start there are none of either input nor output. */
2057         *out_num = *in_num = 0;
2058         if (unlikely(log))
2059                 *log_num = 0;
2060
2061         i = head;
2062         do {
2063                 unsigned iov_count = *in_num + *out_num;
2064                 if (unlikely(i >= vq->num)) {
2065                         vq_err(vq, "Desc index is %u > %u, head = %u",
2066                                i, vq->num, head);
2067                         return -EINVAL;
2068                 }
2069                 if (unlikely(++found > vq->num)) {
2070                         vq_err(vq, "Loop detected: last one at %u "
2071                                "vq size %u head %u\n",
2072                                i, vq->num, head);
2073                         return -EINVAL;
2074                 }
2075                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2076                                            sizeof desc);
2077                 if (unlikely(ret)) {
2078                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2079                                i, vq->desc + i);
2080                         return -EFAULT;
2081                 }
2082                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2083                         ret = get_indirect(vq, iov, iov_size,
2084                                            out_num, in_num,
2085                                            log, log_num, &desc);
2086                         if (unlikely(ret < 0)) {
2087                                 if (ret != -EAGAIN)
2088                                         vq_err(vq, "Failure detected "
2089                                                 "in indirect descriptor at idx %d\n", i);
2090                                 return ret;
2091                         }
2092                         continue;
2093                 }
2094
2095                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2096                         access = VHOST_ACCESS_WO;
2097                 else
2098                         access = VHOST_ACCESS_RO;
2099                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2100                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2101                                      iov_size - iov_count, access);
2102                 if (unlikely(ret < 0)) {
2103                         if (ret != -EAGAIN)
2104                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2105                                         ret, i);
2106                         return ret;
2107                 }
2108                 if (access == VHOST_ACCESS_WO) {
2109                         /* If this is an input descriptor,
2110                          * increment that count. */
2111                         *in_num += ret;
2112                         if (unlikely(log)) {
2113                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2114                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2115                                 ++*log_num;
2116                         }
2117                 } else {
2118                         /* If it's an output descriptor, they're all supposed
2119                          * to come before any input descriptors. */
2120                         if (unlikely(*in_num)) {
2121                                 vq_err(vq, "Descriptor has out after in: "
2122                                        "idx %d\n", i);
2123                                 return -EINVAL;
2124                         }
2125                         *out_num += ret;
2126                 }
2127         } while ((i = next_desc(vq, &desc)) != -1);
2128
2129         /* On success, increment avail index. */
2130         vq->last_avail_idx++;
2131
2132         /* Assume notifications from guest are disabled at this point,
2133          * if they aren't we would need to update avail_event index. */
2134         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2135         return head;
2136 }
2137 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2138
2139 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2140 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2141 {
2142         vq->last_avail_idx -= n;
2143 }
2144 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2145
2146 /* After we've used one of their buffers, we tell them about it.  We'll then
2147  * want to notify the guest, using eventfd. */
2148 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2149 {
2150         struct vring_used_elem heads = {
2151                 cpu_to_vhost32(vq, head),
2152                 cpu_to_vhost32(vq, len)
2153         };
2154
2155         return vhost_add_used_n(vq, &heads, 1);
2156 }
2157 EXPORT_SYMBOL_GPL(vhost_add_used);
2158
2159 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2160                             struct vring_used_elem *heads,
2161                             unsigned count)
2162 {
2163         struct vring_used_elem __user *used;
2164         u16 old, new;
2165         int start;
2166
2167         start = vq->last_used_idx & (vq->num - 1);
2168         used = vq->used->ring + start;
2169         if (count == 1) {
2170                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2171                         vq_err(vq, "Failed to write used id");
2172                         return -EFAULT;
2173                 }
2174                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2175                         vq_err(vq, "Failed to write used len");
2176                         return -EFAULT;
2177                 }
2178         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2179                 vq_err(vq, "Failed to write used");
2180                 return -EFAULT;
2181         }
2182         if (unlikely(vq->log_used)) {
2183                 /* Make sure data is seen before log. */
2184                 smp_wmb();
2185                 /* Log used ring entry write. */
2186                 log_write(vq->log_base,
2187                           vq->log_addr +
2188                            ((void __user *)used - (void __user *)vq->used),
2189                           count * sizeof *used);
2190         }
2191         old = vq->last_used_idx;
2192         new = (vq->last_used_idx += count);
2193         /* If the driver never bothers to signal in a very long while,
2194          * used index might wrap around. If that happens, invalidate
2195          * signalled_used index we stored. TODO: make sure driver
2196          * signals at least once in 2^16 and remove this. */
2197         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2198                 vq->signalled_used_valid = false;
2199         return 0;
2200 }
2201
2202 /* After we've used one of their buffers, we tell them about it.  We'll then
2203  * want to notify the guest, using eventfd. */
2204 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2205                      unsigned count)
2206 {
2207         int start, n, r;
2208
2209         start = vq->last_used_idx & (vq->num - 1);
2210         n = vq->num - start;
2211         if (n < count) {
2212                 r = __vhost_add_used_n(vq, heads, n);
2213                 if (r < 0)
2214                         return r;
2215                 heads += n;
2216                 count -= n;
2217         }
2218         r = __vhost_add_used_n(vq, heads, count);
2219
2220         /* Make sure buffer is written before we update index. */
2221         smp_wmb();
2222         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2223                            &vq->used->idx)) {
2224                 vq_err(vq, "Failed to increment used idx");
2225                 return -EFAULT;
2226         }
2227         if (unlikely(vq->log_used)) {
2228                 /* Log used index update. */
2229                 log_write(vq->log_base,
2230                           vq->log_addr + offsetof(struct vring_used, idx),
2231                           sizeof vq->used->idx);
2232                 if (vq->log_ctx)
2233                         eventfd_signal(vq->log_ctx, 1);
2234         }
2235         return r;
2236 }
2237 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2238
2239 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2240 {
2241         __u16 old, new;
2242         __virtio16 event;
2243         bool v;
2244
2245         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2246             unlikely(vq->avail_idx == vq->last_avail_idx))
2247                 return true;
2248
2249         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2250                 __virtio16 flags;
2251                 /* Flush out used index updates. This is paired
2252                  * with the barrier that the Guest executes when enabling
2253                  * interrupts. */
2254                 smp_mb();
2255                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2256                         vq_err(vq, "Failed to get flags");
2257                         return true;
2258                 }
2259                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2260         }
2261         old = vq->signalled_used;
2262         v = vq->signalled_used_valid;
2263         new = vq->signalled_used = vq->last_used_idx;
2264         vq->signalled_used_valid = true;
2265
2266         if (unlikely(!v))
2267                 return true;
2268
2269         /* We're sure if the following conditions are met, there's no
2270          * need to notify guest:
2271          * 1) cached used event is ahead of new
2272          * 2) old to new updating does not cross cached used event. */
2273         if (vring_need_event(vq->last_used_event, new + vq->num, new) &&
2274             !vring_need_event(vq->last_used_event, new, old))
2275                 return false;
2276
2277         /* Flush out used index updates. This is paired
2278          * with the barrier that the Guest executes when enabling
2279          * interrupts. */
2280         smp_mb();
2281
2282         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2283                 vq_err(vq, "Failed to get used event idx");
2284                 return true;
2285         }
2286         vq->last_used_event = vhost16_to_cpu(vq, event);
2287
2288         return vring_need_event(vq->last_used_event, new, old);
2289 }
2290
2291 /* This actually signals the guest, using eventfd. */
2292 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2293 {
2294         /* Signal the Guest tell them we used something up. */
2295         if (vq->call_ctx && vhost_notify(dev, vq))
2296                 eventfd_signal(vq->call_ctx, 1);
2297 }
2298 EXPORT_SYMBOL_GPL(vhost_signal);
2299
2300 /* And here's the combo meal deal.  Supersize me! */
2301 void vhost_add_used_and_signal(struct vhost_dev *dev,
2302                                struct vhost_virtqueue *vq,
2303                                unsigned int head, int len)
2304 {
2305         vhost_add_used(vq, head, len);
2306         vhost_signal(dev, vq);
2307 }
2308 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2309
2310 /* multi-buffer version of vhost_add_used_and_signal */
2311 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2312                                  struct vhost_virtqueue *vq,
2313                                  struct vring_used_elem *heads, unsigned count)
2314 {
2315         vhost_add_used_n(vq, heads, count);
2316         vhost_signal(dev, vq);
2317 }
2318 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2319
2320 /* return true if we're sure that avaiable ring is empty */
2321 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2322 {
2323         __virtio16 avail_idx;
2324         int r;
2325
2326         if (vq->avail_idx != vq->last_avail_idx)
2327                 return false;
2328
2329         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2330         if (unlikely(r))
2331                 return false;
2332         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2333
2334         return vq->avail_idx == vq->last_avail_idx;
2335 }
2336 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2337
2338 /* OK, now we need to know about added descriptors. */
2339 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2340 {
2341         __virtio16 avail_idx;
2342         int r;
2343
2344         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2345                 return false;
2346         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2347         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2348                 r = vhost_update_used_flags(vq);
2349                 if (r) {
2350                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2351                                &vq->used->flags, r);
2352                         return false;
2353                 }
2354         } else {
2355                 r = vhost_update_avail_event(vq, vq->avail_idx);
2356                 if (r) {
2357                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2358                                vhost_avail_event(vq), r);
2359                         return false;
2360                 }
2361         }
2362         /* They could have slipped one in as we were doing that: make
2363          * sure it's written, then check again. */
2364         smp_mb();
2365         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2366         if (r) {
2367                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2368                        &vq->avail->idx, r);
2369                 return false;
2370         }
2371
2372         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2373 }
2374 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2375
2376 /* We don't need to be notified again. */
2377 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2378 {
2379         int r;
2380
2381         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2382                 return;
2383         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2384         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2385                 r = vhost_update_used_flags(vq);
2386                 if (r)
2387                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2388                                &vq->used->flags, r);
2389         }
2390 }
2391 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2392
2393 /* Create a new message. */
2394 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2395 {
2396         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2397         if (!node)
2398                 return NULL;
2399         node->vq = vq;
2400         node->msg.type = type;
2401         return node;
2402 }
2403 EXPORT_SYMBOL_GPL(vhost_new_msg);
2404
2405 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2406                        struct vhost_msg_node *node)
2407 {
2408         spin_lock(&dev->iotlb_lock);
2409         list_add_tail(&node->node, head);
2410         spin_unlock(&dev->iotlb_lock);
2411
2412         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2413 }
2414 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2415
2416 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2417                                          struct list_head *head)
2418 {
2419         struct vhost_msg_node *node = NULL;
2420
2421         spin_lock(&dev->iotlb_lock);
2422         if (!list_empty(head)) {
2423                 node = list_first_entry(head, struct vhost_msg_node,
2424                                         node);
2425                 list_del(&node->node);
2426         }
2427         spin_unlock(&dev->iotlb_lock);
2428
2429         return node;
2430 }
2431 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2432
2433
2434 static int __init vhost_init(void)
2435 {
2436         return 0;
2437 }
2438
2439 static void __exit vhost_exit(void)
2440 {
2441 }
2442
2443 module_init(vhost_init);
2444 module_exit(vhost_exit);
2445
2446 MODULE_VERSION("0.0.1");
2447 MODULE_LICENSE("GPL v2");
2448 MODULE_AUTHOR("Michael S. Tsirkin");
2449 MODULE_DESCRIPTION("Host kernel accelerator for virtio");