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 #include <linux/nospec.h>
  34
  35 #include "vhost.h"
  36
  37 static ushort max_mem_regions = 64;
  38 module_param(max_mem_regions, ushort, 0444);
  39 MODULE_PARM_DESC(max_mem_regions,
  40         "Maximum number of memory regions in memory map. (default: 64)");
  41 static int max_iotlb_entries = 2048;
  42 module_param(max_iotlb_entries, int, 0444);
  43 MODULE_PARM_DESC(max_iotlb_entries,
  44         "Maximum number of iotlb entries. (default: 2048)");
  45
  46 enum {
  47         VHOST_MEMORY_F_LOG = 0x1,
  48 };
  49
  50 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  51 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  52
  53 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  54                      rb, __u64, __subtree_last,
  55                      START, LAST, static inline, vhost_umem_interval_tree);
  56
  57 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  58 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  59 {
  60         vq->user_be = !virtio_legacy_is_little_endian();
  61 }
  62
  63 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  64 {
  65         vq->user_be = true;
  66 }
  67
  68 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  69 {
  70         vq->user_be = false;
  71 }
  72
  73 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  74 {
  75         struct vhost_vring_state s;
  76
  77         if (vq->private_data)
  78                 return -EBUSY;
  79
  80         if (copy_from_user(&s, argp, sizeof(s)))
  81                 return -EFAULT;
  82
  83         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  84             s.num != VHOST_VRING_BIG_ENDIAN)
  85                 return -EINVAL;
  86
  87         if (s.num == VHOST_VRING_BIG_ENDIAN)
  88                 vhost_enable_cross_endian_big(vq);
  89         else
  90                 vhost_enable_cross_endian_little(vq);
  91
  92         return 0;
  93 }
  94
  95 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  96                                    int __user *argp)
  97 {
  98         struct vhost_vring_state s = {
  99                 .index = idx,
 100                 .num = vq->user_be
 101         };
 102
 103         if (copy_to_user(argp, &s, sizeof(s)))
 104                 return -EFAULT;
 105
 106         return 0;
 107 }
 108
 109 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 110 {
 111         /* Note for legacy virtio: user_be is initialized at reset time
 112          * according to the host endianness. If userspace does not set an
 113          * explicit endianness, the default behavior is native endian, as
 114          * expected by legacy virtio.
 115          */
 116         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 117 }
 118 #else
 119 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 120 {
 121 }
 122
 123 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 124 {
 125         return -ENOIOCTLCMD;
 126 }
 127
 128 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 129                                    int __user *argp)
 130 {
 131         return -ENOIOCTLCMD;
 132 }
 133
 134 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 135 {
 136         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 137                 || virtio_legacy_is_little_endian();
 138 }
 139 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 140
 141 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 142 {
 143         vhost_init_is_le(vq);
 144 }
 145
 146 struct vhost_flush_struct {
 147         struct vhost_work work;
 148         struct completion wait_event;
 149 };
 150
 151 static void vhost_flush_work(struct vhost_work *work)
 152 {
 153         struct vhost_flush_struct *s;
 154
 155         s = container_of(work, struct vhost_flush_struct, work);
 156         complete(&s->wait_event);
 157 }
 158
 159 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 160                             poll_table *pt)
 161 {
 162         struct vhost_poll *poll;
 163
 164         poll = container_of(pt, struct vhost_poll, table);
 165         poll->wqh = wqh;
 166         add_wait_queue(wqh, &poll->wait);
 167 }
 168
 169 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 170                              void *key)
 171 {
 172         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 173
 174         if (!(key_to_poll(key) & poll->mask))
 175                 return 0;
 176
 177         vhost_poll_queue(poll);
 178         return 0;
 179 }
 180
 181 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 182 {
 183         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 184         work->fn = fn;
 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                      __poll_t 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         __poll_t mask;
 207         int ret = 0;
 208
 209         if (poll->wqh)
 210                 return 0;
 211
 212         mask = vfs_poll(file, &poll->table);
 213         if (mask)
 214                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 215         if (mask & EPOLLERR) {
 216                 vhost_poll_stop(poll);
 217                 ret = -EINVAL;
 218         }
 219
 220         return ret;
 221 }
 222 EXPORT_SYMBOL_GPL(vhost_poll_start);
 223
 224 /* Stop polling a file. After this function returns, it becomes safe to drop the
 225  * file reference. You must also flush afterwards. */
 226 void vhost_poll_stop(struct vhost_poll *poll)
 227 {
 228         if (poll->wqh) {
 229                 remove_wait_queue(poll->wqh, &poll->wait);
 230                 poll->wqh = NULL;
 231         }
 232 }
 233 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 234
 235 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 236 {
 237         struct vhost_flush_struct flush;
 238
 239         if (dev->worker) {
 240                 init_completion(&flush.wait_event);
 241                 vhost_work_init(&flush.work, vhost_flush_work);
 242
 243                 vhost_work_queue(dev, &flush.work);
 244                 wait_for_completion(&flush.wait_event);
 245         }
 246 }
 247 EXPORT_SYMBOL_GPL(vhost_work_flush);
 248
 249 /* Flush any work that has been scheduled. When calling this, don't hold any
 250  * locks that are also used by the callback. */
 251 void vhost_poll_flush(struct vhost_poll *poll)
 252 {
 253         vhost_work_flush(poll->dev, &poll->work);
 254 }
 255 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 256
 257 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 258 {
 259         if (!dev->worker)
 260                 return;
 261
 262         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 263                 /* We can only add the work to the list after we're
 264                  * sure it was not in the list.
 265                  * test_and_set_bit() implies a memory barrier.
 266                  */
 267                 llist_add(&work->node, &dev->work_list);
 268                 wake_up_process(dev->worker);
 269         }
 270 }
 271 EXPORT_SYMBOL_GPL(vhost_work_queue);
 272
 273 /* A lockless hint for busy polling code to exit the loop */
 274 bool vhost_has_work(struct vhost_dev *dev)
 275 {
 276         return !llist_empty(&dev->work_list);
 277 }
 278 EXPORT_SYMBOL_GPL(vhost_has_work);
 279
 280 void vhost_poll_queue(struct vhost_poll *poll)
 281 {
 282         vhost_work_queue(poll->dev, &poll->work);
 283 }
 284 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 285
 286 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 287 {
 288         int j;
 289
 290         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 291                 vq->meta_iotlb[j] = NULL;
 292 }
 293
 294 static void vhost_vq_meta_reset(struct vhost_dev *d)
 295 {
 296         int i;
 297
 298         for (i = 0; i < d->nvqs; ++i)
 299                 __vhost_vq_meta_reset(d->vqs[i]);
 300 }
 301
 302 static void vhost_vq_reset(struct vhost_dev *dev,
 303                            struct vhost_virtqueue *vq)
 304 {
 305         vq->num = 1;
 306         vq->desc = NULL;
 307         vq->avail = NULL;
 308         vq->used = NULL;
 309         vq->last_avail_idx = 0;
 310         vq->avail_idx = 0;
 311         vq->last_used_idx = 0;
 312         vq->signalled_used = 0;
 313         vq->signalled_used_valid = false;
 314         vq->used_flags = 0;
 315         vq->log_used = false;
 316         vq->log_addr = -1ull;
 317         vq->private_data = NULL;
 318         vq->acked_features = 0;
 319         vq->acked_backend_features = 0;
 320         vq->log_base = NULL;
 321         vq->error_ctx = NULL;
 322         vq->kick = NULL;
 323         vq->call_ctx = NULL;
 324         vq->log_ctx = NULL;
 325         vhost_reset_is_le(vq);
 326         vhost_disable_cross_endian(vq);
 327         vq->busyloop_timeout = 0;
 328         vq->umem = NULL;
 329         vq->iotlb = NULL;
 330         __vhost_vq_meta_reset(vq);
 331 }
 332
 333 static int vhost_worker(void *data)
 334 {
 335         struct vhost_dev *dev = data;
 336         struct vhost_work *work, *work_next;
 337         struct llist_node *node;
 338         mm_segment_t oldfs = get_fs();
 339
 340         set_fs(USER_DS);
 341         use_mm(dev->mm);
 342
 343         for (;;) {
 344                 /* mb paired w/ kthread_stop */
 345                 set_current_state(TASK_INTERRUPTIBLE);
 346
 347                 if (kthread_should_stop()) {
 348                         __set_current_state(TASK_RUNNING);
 349                         break;
 350                 }
 351
 352                 node = llist_del_all(&dev->work_list);
 353                 if (!node)
 354                         schedule();
 355
 356                 node = llist_reverse_order(node);
 357                 /* make sure flag is seen after deletion */
 358                 smp_wmb();
 359                 llist_for_each_entry_safe(work, work_next, node, node) {
 360                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 361                         __set_current_state(TASK_RUNNING);
 362                         work->fn(work);
 363                         if (need_resched())
 364                                 schedule();
 365                 }
 366         }
 367         unuse_mm(dev->mm);
 368         set_fs(oldfs);
 369         return 0;
 370 }
 371
 372 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 373 {
 374         kfree(vq->indirect);
 375         vq->indirect = NULL;
 376         kfree(vq->log);
 377         vq->log = NULL;
 378         kfree(vq->heads);
 379         vq->heads = NULL;
 380 }
 381
 382 /* Helper to allocate iovec buffers for all vqs. */
 383 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 384 {
 385         struct vhost_virtqueue *vq;
 386         int i;
 387
 388         for (i = 0; i < dev->nvqs; ++i) {
 389                 vq = dev->vqs[i];
 390                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 391                                              sizeof(*vq->indirect),
 392                                              GFP_KERNEL);
 393                 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 394                                         GFP_KERNEL);
 395                 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 396                                           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 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 417                           int pkts, int total_len)
 418 {
 419         struct vhost_dev *dev = vq->dev;
 420
 421         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 422             pkts >= dev->weight) {
 423                 vhost_poll_queue(&vq->poll);
 424                 return true;
 425         }
 426
 427         return false;
 428 }
 429 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 430
 431 void vhost_dev_init(struct vhost_dev *dev,
 432                     struct vhost_virtqueue **vqs, int nvqs,
 433                     int iov_limit, int weight, int byte_weight)
 434 {
 435         struct vhost_virtqueue *vq;
 436         int i;
 437
 438         dev->vqs = vqs;
 439         dev->nvqs = nvqs;
 440         mutex_init(&dev->mutex);
 441         dev->log_ctx = NULL;
 442         dev->umem = NULL;
 443         dev->iotlb = NULL;
 444         dev->mm = NULL;
 445         dev->worker = NULL;
 446         dev->iov_limit = iov_limit;
 447         dev->weight = weight;
 448         dev->byte_weight = byte_weight;
 449         init_llist_head(&dev->work_list);
 450         init_waitqueue_head(&dev->wait);
 451         INIT_LIST_HEAD(&dev->read_list);
 452         INIT_LIST_HEAD(&dev->pending_list);
 453         spin_lock_init(&dev->iotlb_lock);
 454
 455
 456         for (i = 0; i < dev->nvqs; ++i) {
 457                 vq = dev->vqs[i];
 458                 vq->log = NULL;
 459                 vq->indirect = NULL;
 460                 vq->heads = NULL;
 461                 vq->dev = dev;
 462                 mutex_init(&vq->mutex);
 463                 vhost_vq_reset(dev, vq);
 464                 if (vq->handle_kick)
 465                         vhost_poll_init(&vq->poll, vq->handle_kick,
 466                                         EPOLLIN, dev);
 467         }
 468 }
 469 EXPORT_SYMBOL_GPL(vhost_dev_init);
 470
 471 /* Caller should have device mutex */
 472 long vhost_dev_check_owner(struct vhost_dev *dev)
 473 {
 474         /* Are you the owner? If not, I don't think you mean to do that */
 475         return dev->mm == current->mm ? 0 : -EPERM;
 476 }
 477 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 478
 479 struct vhost_attach_cgroups_struct {
 480         struct vhost_work work;
 481         struct task_struct *owner;
 482         int ret;
 483 };
 484
 485 static void vhost_attach_cgroups_work(struct vhost_work *work)
 486 {
 487         struct vhost_attach_cgroups_struct *s;
 488
 489         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 490         s->ret = cgroup_attach_task_all(s->owner, current);
 491 }
 492
 493 static int vhost_attach_cgroups(struct vhost_dev *dev)
 494 {
 495         struct vhost_attach_cgroups_struct attach;
 496
 497         attach.owner = current;
 498         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 499         vhost_work_queue(dev, &attach.work);
 500         vhost_work_flush(dev, &attach.work);
 501         return attach.ret;
 502 }
 503
 504 /* Caller should have device mutex */
 505 bool vhost_dev_has_owner(struct vhost_dev *dev)
 506 {
 507         return dev->mm;
 508 }
 509 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 510
 511 /* Caller should have device mutex */
 512 long vhost_dev_set_owner(struct vhost_dev *dev)
 513 {
 514         struct task_struct *worker;
 515         int err;
 516
 517         /* Is there an owner already? */
 518         if (vhost_dev_has_owner(dev)) {
 519                 err = -EBUSY;
 520                 goto err_mm;
 521         }
 522
 523         /* No owner, become one */
 524         dev->mm = get_task_mm(current);
 525         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 526         if (IS_ERR(worker)) {
 527                 err = PTR_ERR(worker);
 528                 goto err_worker;
 529         }
 530
 531         dev->worker = worker;
 532         wake_up_process(worker);        /* avoid contributing to loadavg */
 533
 534         err = vhost_attach_cgroups(dev);
 535         if (err)
 536                 goto err_cgroup;
 537
 538         err = vhost_dev_alloc_iovecs(dev);
 539         if (err)
 540                 goto err_cgroup;
 541
 542         return 0;
 543 err_cgroup:
 544         kthread_stop(worker);
 545         dev->worker = NULL;
 546 err_worker:
 547         if (dev->mm)
 548                 mmput(dev->mm);
 549         dev->mm = NULL;
 550 err_mm:
 551         return err;
 552 }
 553 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 554
 555 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 556 {
 557         return kvzalloc(sizeof(struct vhost_umem), GFP_KERNEL);
 558 }
 559 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 560
 561 /* Caller should have device mutex */
 562 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 563 {
 564         int i;
 565
 566         vhost_dev_cleanup(dev);
 567
 568         /* Restore memory to default empty mapping. */
 569         INIT_LIST_HEAD(&umem->umem_list);
 570         dev->umem = umem;
 571         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 572          * VQs aren't running.
 573          */
 574         for (i = 0; i < dev->nvqs; ++i)
 575                 dev->vqs[i]->umem = umem;
 576 }
 577 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 578
 579 void vhost_dev_stop(struct vhost_dev *dev)
 580 {
 581         int i;
 582
 583         for (i = 0; i < dev->nvqs; ++i) {
 584                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 585                         vhost_poll_stop(&dev->vqs[i]->poll);
 586                         vhost_poll_flush(&dev->vqs[i]->poll);
 587                 }
 588         }
 589 }
 590 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 591
 592 static void vhost_umem_free(struct vhost_umem *umem,
 593                             struct vhost_umem_node *node)
 594 {
 595         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 596         list_del(&node->link);
 597         kfree(node);
 598         umem->numem--;
 599 }
 600
 601 static void vhost_umem_clean(struct vhost_umem *umem)
 602 {
 603         struct vhost_umem_node *node, *tmp;
 604
 605         if (!umem)
 606                 return;
 607
 608         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 609                 vhost_umem_free(umem, node);
 610
 611         kvfree(umem);
 612 }
 613
 614 static void vhost_clear_msg(struct vhost_dev *dev)
 615 {
 616         struct vhost_msg_node *node, *n;
 617
 618         spin_lock(&dev->iotlb_lock);
 619
 620         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 621                 list_del(&node->node);
 622                 kfree(node);
 623         }
 624
 625         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 626                 list_del(&node->node);
 627                 kfree(node);
 628         }
 629
 630         spin_unlock(&dev->iotlb_lock);
 631 }
 632
 633 void vhost_dev_cleanup(struct vhost_dev *dev)
 634 {
 635         int i;
 636
 637         for (i = 0; i < dev->nvqs; ++i) {
 638                 if (dev->vqs[i]->error_ctx)
 639                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 640                 if (dev->vqs[i]->kick)
 641                         fput(dev->vqs[i]->kick);
 642                 if (dev->vqs[i]->call_ctx)
 643                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 644                 vhost_vq_reset(dev, dev->vqs[i]);
 645         }
 646         vhost_dev_free_iovecs(dev);
 647         if (dev->log_ctx)
 648                 eventfd_ctx_put(dev->log_ctx);
 649         dev->log_ctx = NULL;
 650         /* No one will access memory at this point */
 651         vhost_umem_clean(dev->umem);
 652         dev->umem = NULL;
 653         vhost_umem_clean(dev->iotlb);
 654         dev->iotlb = NULL;
 655         vhost_clear_msg(dev);
 656         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 657         WARN_ON(!llist_empty(&dev->work_list));
 658         if (dev->worker) {
 659                 kthread_stop(dev->worker);
 660                 dev->worker = NULL;
 661         }
 662         if (dev->mm)
 663                 mmput(dev->mm);
 664         dev->mm = NULL;
 665 }
 666 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 667
 668 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 669 {
 670         u64 a = addr / VHOST_PAGE_SIZE / 8;
 671
 672         /* Make sure 64 bit math will not overflow. */
 673         if (a > ULONG_MAX - (unsigned long)log_base ||
 674             a + (unsigned long)log_base > ULONG_MAX)
 675                 return false;
 676
 677         return access_ok(log_base + a,
 678                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 679 }
 680
 681 static bool vhost_overflow(u64 uaddr, u64 size)
 682 {
 683         /* Make sure 64 bit math will not overflow. */
 684         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 685 }
 686
 687 /* Caller should have vq mutex and device mutex. */
 688 static bool vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 689                                 int log_all)
 690 {
 691         struct vhost_umem_node *node;
 692
 693         if (!umem)
 694                 return false;
 695
 696         list_for_each_entry(node, &umem->umem_list, link) {
 697                 unsigned long a = node->userspace_addr;
 698
 699                 if (vhost_overflow(node->userspace_addr, node->size))
 700                         return false;
 701
 702
 703                 if (!access_ok((void __user *)a,
 704                                     node->size))
 705                         return false;
 706                 else if (log_all && !log_access_ok(log_base,
 707                                                    node->start,
 708                                                    node->size))
 709                         return false;
 710         }
 711         return true;
 712 }
 713
 714 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 715                                                u64 addr, unsigned int size,
 716                                                int type)
 717 {
 718         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 719
 720         if (!node)
 721                 return NULL;
 722
 723         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 724 }
 725
 726 /* Can we switch to this memory table? */
 727 /* Caller should have device mutex but not vq mutex */
 728 static bool memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 729                              int log_all)
 730 {
 731         int i;
 732
 733         for (i = 0; i < d->nvqs; ++i) {
 734                 bool ok;
 735                 bool log;
 736
 737                 mutex_lock(&d->vqs[i]->mutex);
 738                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 739                 /* If ring is inactive, will check when it's enabled. */
 740                 if (d->vqs[i]->private_data)
 741                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 742                                                  umem, log);
 743                 else
 744                         ok = true;
 745                 mutex_unlock(&d->vqs[i]->mutex);
 746                 if (!ok)
 747                         return false;
 748         }
 749         return true;
 750 }
 751
 752 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 753                           struct iovec iov[], int iov_size, int access);
 754
 755 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 756                               const void *from, unsigned size)
 757 {
 758         int ret;
 759
 760         if (!vq->iotlb)
 761                 return __copy_to_user(to, from, size);
 762         else {
 763                 /* This function should be called after iotlb
 764                  * prefetch, which means we're sure that all vq
 765                  * could be access through iotlb. So -EAGAIN should
 766                  * not happen in this case.
 767                  */
 768                 struct iov_iter t;
 769                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 770                                      (u64)(uintptr_t)to, size,
 771                                      VHOST_ADDR_USED);
 772
 773                 if (uaddr)
 774                         return __copy_to_user(uaddr, from, size);
 775
 776                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 777                                      ARRAY_SIZE(vq->iotlb_iov),
 778                                      VHOST_ACCESS_WO);
 779                 if (ret < 0)
 780                         goto out;
 781                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 782                 ret = copy_to_iter(from, size, &t);
 783                 if (ret == size)
 784                         ret = 0;
 785         }
 786 out:
 787         return ret;
 788 }
 789
 790 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 791                                 void __user *from, unsigned size)
 792 {
 793         int ret;
 794
 795         if (!vq->iotlb)
 796                 return __copy_from_user(to, from, size);
 797         else {
 798                 /* This function should be called after iotlb
 799                  * prefetch, which means we're sure that vq
 800                  * could be access through iotlb. So -EAGAIN should
 801                  * not happen in this case.
 802                  */
 803                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 804                                      (u64)(uintptr_t)from, size,
 805                                      VHOST_ADDR_DESC);
 806                 struct iov_iter f;
 807
 808                 if (uaddr)
 809                         return __copy_from_user(to, uaddr, size);
 810
 811                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 812                                      ARRAY_SIZE(vq->iotlb_iov),
 813                                      VHOST_ACCESS_RO);
 814                 if (ret < 0) {
 815                         vq_err(vq, "IOTLB translation failure: uaddr "
 816                                "%p size 0x%llx\n", from,
 817                                (unsigned long long) size);
 818                         goto out;
 819                 }
 820                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 821                 ret = copy_from_iter(to, size, &f);
 822                 if (ret == size)
 823                         ret = 0;
 824         }
 825
 826 out:
 827         return ret;
 828 }
 829
 830 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 831                                           void __user *addr, unsigned int size,
 832                                           int type)
 833 {
 834         int ret;
 835
 836         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 837                              ARRAY_SIZE(vq->iotlb_iov),
 838                              VHOST_ACCESS_RO);
 839         if (ret < 0) {
 840                 vq_err(vq, "IOTLB translation failure: uaddr "
 841                         "%p size 0x%llx\n", addr,
 842                         (unsigned long long) size);
 843                 return NULL;
 844         }
 845
 846         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 847                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 848                         "%p size 0x%llx\n", addr,
 849                         (unsigned long long) size);
 850                 return NULL;
 851         }
 852
 853         return vq->iotlb_iov[0].iov_base;
 854 }
 855
 856 /* This function should be called after iotlb
 857  * prefetch, which means we're sure that vq
 858  * could be access through iotlb. So -EAGAIN should
 859  * not happen in this case.
 860  */
 861 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 862                                             void *addr, unsigned int size,
 863                                             int type)
 864 {
 865         void __user *uaddr = vhost_vq_meta_fetch(vq,
 866                              (u64)(uintptr_t)addr, size, type);
 867         if (uaddr)
 868                 return uaddr;
 869
 870         return __vhost_get_user_slow(vq, addr, size, type);
 871 }
 872
 873 #define vhost_put_user(vq, x, ptr)              \
 874 ({ \
 875         int ret = -EFAULT; \
 876         if (!vq->iotlb) { \
 877                 ret = __put_user(x, ptr); \
 878         } else { \
 879                 __typeof__(ptr) to = \
 880                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 881                                           sizeof(*ptr), VHOST_ADDR_USED); \
 882                 if (to != NULL) \
 883                         ret = __put_user(x, to); \
 884                 else \
 885                         ret = -EFAULT;  \
 886         } \
 887         ret; \
 888 })
 889
 890 #define vhost_get_user(vq, x, ptr, type)                \
 891 ({ \
 892         int ret; \
 893         if (!vq->iotlb) { \
 894                 ret = __get_user(x, ptr); \
 895         } else { \
 896                 __typeof__(ptr) from = \
 897                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 898                                                            sizeof(*ptr), \
 899                                                            type); \
 900                 if (from != NULL) \
 901                         ret = __get_user(x, from); \
 902                 else \
 903                         ret = -EFAULT; \
 904         } \
 905         ret; \
 906 })
 907
 908 #define vhost_get_avail(vq, x, ptr) \
 909         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 910
 911 #define vhost_get_used(vq, x, ptr) \
 912         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 913
 914 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 915 {
 916         int i = 0;
 917         for (i = 0; i < d->nvqs; ++i)
 918                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 919 }
 920
 921 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 922 {
 923         int i = 0;
 924         for (i = 0; i < d->nvqs; ++i)
 925                 mutex_unlock(&d->vqs[i]->mutex);
 926 }
 927
 928 static int vhost_new_umem_range(struct vhost_umem *umem,
 929                                 u64 start, u64 size, u64 end,
 930                                 u64 userspace_addr, int perm)
 931 {
 932         struct vhost_umem_node *tmp, *node;
 933
 934         if (!size)
 935                 return -EFAULT;
 936
 937         node = kmalloc(sizeof(*node), GFP_ATOMIC);
 938         if (!node)
 939                 return -ENOMEM;
 940
 941         if (umem->numem == max_iotlb_entries) {
 942                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 943                 vhost_umem_free(umem, tmp);
 944         }
 945
 946         node->start = start;
 947         node->size = size;
 948         node->last = end;
 949         node->userspace_addr = userspace_addr;
 950         node->perm = perm;
 951         INIT_LIST_HEAD(&node->link);
 952         list_add_tail(&node->link, &umem->umem_list);
 953         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 954         umem->numem++;
 955
 956         return 0;
 957 }
 958
 959 static void vhost_del_umem_range(struct vhost_umem *umem,
 960                                  u64 start, u64 end)
 961 {
 962         struct vhost_umem_node *node;
 963
 964         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 965                                                            start, end)))
 966                 vhost_umem_free(umem, node);
 967 }
 968
 969 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 970                                   struct vhost_iotlb_msg *msg)
 971 {
 972         struct vhost_msg_node *node, *n;
 973
 974         spin_lock(&d->iotlb_lock);
 975
 976         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 977                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 978                 if (msg->iova <= vq_msg->iova &&
 979                     msg->iova + msg->size - 1 >= vq_msg->iova &&
 980                     vq_msg->type == VHOST_IOTLB_MISS) {
 981                         vhost_poll_queue(&node->vq->poll);
 982                         list_del(&node->node);
 983                         kfree(node);
 984                 }
 985         }
 986
 987         spin_unlock(&d->iotlb_lock);
 988 }
 989
 990 static bool umem_access_ok(u64 uaddr, u64 size, int access)
 991 {
 992         unsigned long a = uaddr;
 993
 994         /* Make sure 64 bit math will not overflow. */
 995         if (vhost_overflow(uaddr, size))
 996                 return false;
 997
 998         if ((access & VHOST_ACCESS_RO) &&
 999             !access_ok((void __user *)a, size))
1000                 return false;
1001         if ((access & VHOST_ACCESS_WO) &&
1002             !access_ok((void __user *)a, size))
1003                 return false;
1004         return true;
1005 }
1006
1007 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1008                                    struct vhost_iotlb_msg *msg)
1009 {
1010         int ret = 0;
1011
1012         mutex_lock(&dev->mutex);
1013         vhost_dev_lock_vqs(dev);
1014         switch (msg->type) {
1015         case VHOST_IOTLB_UPDATE:
1016                 if (!dev->iotlb) {
1017                         ret = -EFAULT;
1018                         break;
1019                 }
1020                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1021                         ret = -EFAULT;
1022                         break;
1023                 }
1024                 vhost_vq_meta_reset(dev);
1025                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1026                                          msg->iova + msg->size - 1,
1027                                          msg->uaddr, msg->perm)) {
1028                         ret = -ENOMEM;
1029                         break;
1030                 }
1031                 vhost_iotlb_notify_vq(dev, msg);
1032                 break;
1033         case VHOST_IOTLB_INVALIDATE:
1034                 if (!dev->iotlb) {
1035                         ret = -EFAULT;
1036                         break;
1037                 }
1038                 vhost_vq_meta_reset(dev);
1039                 vhost_del_umem_range(dev->iotlb, msg->iova,
1040                                      msg->iova + msg->size - 1);
1041                 break;
1042         default:
1043                 ret = -EINVAL;
1044                 break;
1045         }
1046
1047         vhost_dev_unlock_vqs(dev);
1048         mutex_unlock(&dev->mutex);
1049
1050         return ret;
1051 }
1052 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1053                              struct iov_iter *from)
1054 {
1055         struct vhost_iotlb_msg msg;
1056         size_t offset;
1057         int type, ret;
1058
1059         ret = copy_from_iter(&type, sizeof(type), from);
1060         if (ret != sizeof(type)) {
1061                 ret = -EINVAL;
1062                 goto done;
1063         }
1064
1065         switch (type) {
1066         case VHOST_IOTLB_MSG:
1067                 /* There maybe a hole after type for V1 message type,
1068                  * so skip it here.
1069                  */
1070                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1071                 break;
1072         case VHOST_IOTLB_MSG_V2:
1073                 offset = sizeof(__u32);
1074                 break;
1075         default:
1076                 ret = -EINVAL;
1077                 goto done;
1078         }
1079
1080         iov_iter_advance(from, offset);
1081         ret = copy_from_iter(&msg, sizeof(msg), from);
1082         if (ret != sizeof(msg)) {
1083                 ret = -EINVAL;
1084                 goto done;
1085         }
1086         if (vhost_process_iotlb_msg(dev, &msg)) {
1087                 ret = -EFAULT;
1088                 goto done;
1089         }
1090
1091         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1092               sizeof(struct vhost_msg_v2);
1093 done:
1094         return ret;
1095 }
1096 EXPORT_SYMBOL(vhost_chr_write_iter);
1097
1098 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1099                             poll_table *wait)
1100 {
1101         __poll_t mask = 0;
1102
1103         poll_wait(file, &dev->wait, wait);
1104
1105         if (!list_empty(&dev->read_list))
1106                 mask |= EPOLLIN | EPOLLRDNORM;
1107
1108         return mask;
1109 }
1110 EXPORT_SYMBOL(vhost_chr_poll);
1111
1112 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1113                             int noblock)
1114 {
1115         DEFINE_WAIT(wait);
1116         struct vhost_msg_node *node;
1117         ssize_t ret = 0;
1118         unsigned size = sizeof(struct vhost_msg);
1119
1120         if (iov_iter_count(to) < size)
1121                 return 0;
1122
1123         while (1) {
1124                 if (!noblock)
1125                         prepare_to_wait(&dev->wait, &wait,
1126                                         TASK_INTERRUPTIBLE);
1127
1128                 node = vhost_dequeue_msg(dev, &dev->read_list);
1129                 if (node)
1130                         break;
1131                 if (noblock) {
1132                         ret = -EAGAIN;
1133                         break;
1134                 }
1135                 if (signal_pending(current)) {
1136                         ret = -ERESTARTSYS;
1137                         break;
1138                 }
1139                 if (!dev->iotlb) {
1140                         ret = -EBADFD;
1141                         break;
1142                 }
1143
1144                 schedule();
1145         }
1146
1147         if (!noblock)
1148                 finish_wait(&dev->wait, &wait);
1149
1150         if (node) {
1151                 struct vhost_iotlb_msg *msg;
1152                 void *start = &node->msg;
1153
1154                 switch (node->msg.type) {
1155                 case VHOST_IOTLB_MSG:
1156                         size = sizeof(node->msg);
1157                         msg = &node->msg.iotlb;
1158                         break;
1159                 case VHOST_IOTLB_MSG_V2:
1160                         size = sizeof(node->msg_v2);
1161                         msg = &node->msg_v2.iotlb;
1162                         break;
1163                 default:
1164                         BUG();
1165                         break;
1166                 }
1167
1168                 ret = copy_to_iter(start, size, to);
1169                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1170                         kfree(node);
1171                         return ret;
1172                 }
1173                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1174         }
1175
1176         return ret;
1177 }
1178 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1179
1180 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1181 {
1182         struct vhost_dev *dev = vq->dev;
1183         struct vhost_msg_node *node;
1184         struct vhost_iotlb_msg *msg;
1185         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1186
1187         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1188         if (!node)
1189                 return -ENOMEM;
1190
1191         if (v2) {
1192                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1193                 msg = &node->msg_v2.iotlb;
1194         } else {
1195                 msg = &node->msg.iotlb;
1196         }
1197
1198         msg->type = VHOST_IOTLB_MISS;
1199         msg->iova = iova;
1200         msg->perm = access;
1201
1202         vhost_enqueue_msg(dev, &dev->read_list, node);
1203
1204         return 0;
1205 }
1206
1207 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1208                          struct vring_desc __user *desc,
1209                          struct vring_avail __user *avail,
1210                          struct vring_used __user *used)
1211
1212 {
1213         size_t s __maybe_unused = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1214
1215         return access_ok(desc, num * sizeof *desc) &&
1216                access_ok(avail,
1217                          sizeof *avail + num * sizeof *avail->ring + s) &&
1218                access_ok(used,
1219                         sizeof *used + num * sizeof *used->ring + s);
1220 }
1221
1222 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1223                                  const struct vhost_umem_node *node,
1224                                  int type)
1225 {
1226         int access = (type == VHOST_ADDR_USED) ?
1227                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1228
1229         if (likely(node->perm & access))
1230                 vq->meta_iotlb[type] = node;
1231 }
1232
1233 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1234                             int access, u64 addr, u64 len, int type)
1235 {
1236         const struct vhost_umem_node *node;
1237         struct vhost_umem *umem = vq->iotlb;
1238         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1239
1240         if (vhost_vq_meta_fetch(vq, addr, len, type))
1241                 return true;
1242
1243         while (len > s) {
1244                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1245                                                            addr,
1246                                                            last);
1247                 if (node == NULL || node->start > addr) {
1248                         vhost_iotlb_miss(vq, addr, access);
1249                         return false;
1250                 } else if (!(node->perm & access)) {
1251                         /* Report the possible access violation by
1252                          * request another translation from userspace.
1253                          */
1254                         return false;
1255                 }
1256
1257                 size = node->size - addr + node->start;
1258
1259                 if (orig_addr == addr && size >= len)
1260                         vhost_vq_meta_update(vq, node, type);
1261
1262                 s += size;
1263                 addr += size;
1264         }
1265
1266         return true;
1267 }
1268
1269 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1270 {
1271         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1272         unsigned int num = vq->num;
1273
1274         if (!vq->iotlb)
1275                 return 1;
1276
1277         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1278                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1279                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1280                                sizeof *vq->avail +
1281                                num * sizeof(*vq->avail->ring) + s,
1282                                VHOST_ADDR_AVAIL) &&
1283                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1284                                sizeof *vq->used +
1285                                num * sizeof(*vq->used->ring) + s,
1286                                VHOST_ADDR_USED);
1287 }
1288 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1289
1290 /* Can we log writes? */
1291 /* Caller should have device mutex but not vq mutex */
1292 bool vhost_log_access_ok(struct vhost_dev *dev)
1293 {
1294         return memory_access_ok(dev, dev->umem, 1);
1295 }
1296 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1297
1298 /* Verify access for write logging. */
1299 /* Caller should have vq mutex and device mutex */
1300 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1301                              void __user *log_base)
1302 {
1303         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1304
1305         return vq_memory_access_ok(log_base, vq->umem,
1306                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1307                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1308                                         sizeof *vq->used +
1309                                         vq->num * sizeof *vq->used->ring + s));
1310 }
1311
1312 /* Can we start vq? */
1313 /* Caller should have vq mutex and device mutex */
1314 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1315 {
1316         if (!vq_log_access_ok(vq, vq->log_base))
1317                 return false;
1318
1319         /* Access validation occurs at prefetch time with IOTLB */
1320         if (vq->iotlb)
1321                 return true;
1322
1323         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1324 }
1325 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1326
1327 static struct vhost_umem *vhost_umem_alloc(void)
1328 {
1329         struct vhost_umem *umem = kvzalloc(sizeof(*umem), GFP_KERNEL);
1330
1331         if (!umem)
1332                 return NULL;
1333
1334         umem->umem_tree = RB_ROOT_CACHED;
1335         umem->numem = 0;
1336         INIT_LIST_HEAD(&umem->umem_list);
1337
1338         return umem;
1339 }
1340
1341 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1342 {
1343         struct vhost_memory mem, *newmem;
1344         struct vhost_memory_region *region;
1345         struct vhost_umem *newumem, *oldumem;
1346         unsigned long size = offsetof(struct vhost_memory, regions);
1347         int i;
1348
1349         if (copy_from_user(&mem, m, size))
1350                 return -EFAULT;
1351         if (mem.padding)
1352                 return -EOPNOTSUPP;
1353         if (mem.nregions > max_mem_regions)
1354                 return -E2BIG;
1355         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1356                         GFP_KERNEL);
1357         if (!newmem)
1358                 return -ENOMEM;
1359
1360         memcpy(newmem, &mem, size);
1361         if (copy_from_user(newmem->regions, m->regions,
1362                            mem.nregions * sizeof *m->regions)) {
1363                 kvfree(newmem);
1364                 return -EFAULT;
1365         }
1366
1367         newumem = vhost_umem_alloc();
1368         if (!newumem) {
1369                 kvfree(newmem);
1370                 return -ENOMEM;
1371         }
1372
1373         for (region = newmem->regions;
1374              region < newmem->regions + mem.nregions;
1375              region++) {
1376                 if (vhost_new_umem_range(newumem,
1377                                          region->guest_phys_addr,
1378                                          region->memory_size,
1379                                          region->guest_phys_addr +
1380                                          region->memory_size - 1,
1381                                          region->userspace_addr,
1382                                          VHOST_ACCESS_RW))
1383                         goto err;
1384         }
1385
1386         if (!memory_access_ok(d, newumem, 0))
1387                 goto err;
1388
1389         oldumem = d->umem;
1390         d->umem = newumem;
1391
1392         /* All memory accesses are done under some VQ mutex. */
1393         for (i = 0; i < d->nvqs; ++i) {
1394                 mutex_lock(&d->vqs[i]->mutex);
1395                 d->vqs[i]->umem = newumem;
1396                 mutex_unlock(&d->vqs[i]->mutex);
1397         }
1398
1399         kvfree(newmem);
1400         vhost_umem_clean(oldumem);
1401         return 0;
1402
1403 err:
1404         vhost_umem_clean(newumem);
1405         kvfree(newmem);
1406         return -EFAULT;
1407 }
1408
1409 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1410 {
1411         struct file *eventfp, *filep = NULL;
1412         bool pollstart = false, pollstop = false;
1413         struct eventfd_ctx *ctx = NULL;
1414         u32 __user *idxp = argp;
1415         struct vhost_virtqueue *vq;
1416         struct vhost_vring_state s;
1417         struct vhost_vring_file f;
1418         struct vhost_vring_addr a;
1419         u32 idx;
1420         long r;
1421
1422         r = get_user(idx, idxp);
1423         if (r < 0)
1424                 return r;
1425         if (idx >= d->nvqs)
1426                 return -ENOBUFS;
1427
1428         idx = array_index_nospec(idx, d->nvqs);
1429         vq = d->vqs[idx];
1430
1431         mutex_lock(&vq->mutex);
1432
1433         switch (ioctl) {
1434         case VHOST_SET_VRING_NUM:
1435                 /* Resizing ring with an active backend?
1436                  * You don't want to do that. */
1437                 if (vq->private_data) {
1438                         r = -EBUSY;
1439                         break;
1440                 }
1441                 if (copy_from_user(&s, argp, sizeof s)) {
1442                         r = -EFAULT;
1443                         break;
1444                 }
1445                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1446                         r = -EINVAL;
1447                         break;
1448                 }
1449                 vq->num = s.num;
1450                 break;
1451         case VHOST_SET_VRING_BASE:
1452                 /* Moving base with an active backend?
1453                  * You don't want to do that. */
1454                 if (vq->private_data) {
1455                         r = -EBUSY;
1456                         break;
1457                 }
1458                 if (copy_from_user(&s, argp, sizeof s)) {
1459                         r = -EFAULT;
1460                         break;
1461                 }
1462                 if (s.num > 0xffff) {
1463                         r = -EINVAL;
1464                         break;
1465                 }
1466                 vq->last_avail_idx = s.num;
1467                 /* Forget the cached index value. */
1468                 vq->avail_idx = vq->last_avail_idx;
1469                 break;
1470         case VHOST_GET_VRING_BASE:
1471                 s.index = idx;
1472                 s.num = vq->last_avail_idx;
1473                 if (copy_to_user(argp, &s, sizeof s))
1474                         r = -EFAULT;
1475                 break;
1476         case VHOST_SET_VRING_ADDR:
1477                 if (copy_from_user(&a, argp, sizeof a)) {
1478                         r = -EFAULT;
1479                         break;
1480                 }
1481                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1482                         r = -EOPNOTSUPP;
1483                         break;
1484                 }
1485                 /* For 32bit, verify that the top 32bits of the user
1486                    data are set to zero. */
1487                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1488                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1489                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1490                         r = -EFAULT;
1491                         break;
1492                 }
1493
1494                 /* Make sure it's safe to cast pointers to vring types. */
1495                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1496                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1497                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1498                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1499                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1500                         r = -EINVAL;
1501                         break;
1502                 }
1503
1504                 /* We only verify access here if backend is configured.
1505                  * If it is not, we don't as size might not have been setup.
1506                  * We will verify when backend is configured. */
1507                 if (vq->private_data) {
1508                         if (!vq_access_ok(vq, vq->num,
1509                                 (void __user *)(unsigned long)a.desc_user_addr,
1510                                 (void __user *)(unsigned long)a.avail_user_addr,
1511                                 (void __user *)(unsigned long)a.used_user_addr)) {
1512                                 r = -EINVAL;
1513                                 break;
1514                         }
1515
1516                         /* Also validate log access for used ring if enabled. */
1517                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1518                             !log_access_ok(vq->log_base, a.log_guest_addr,
1519                                            sizeof *vq->used +
1520                                            vq->num * sizeof *vq->used->ring)) {
1521                                 r = -EINVAL;
1522                                 break;
1523                         }
1524                 }
1525
1526                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1527                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1528                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1529                 vq->log_addr = a.log_guest_addr;
1530                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1531                 break;
1532         case VHOST_SET_VRING_KICK:
1533                 if (copy_from_user(&f, argp, sizeof f)) {
1534                         r = -EFAULT;
1535                         break;
1536                 }
1537                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1538                 if (IS_ERR(eventfp)) {
1539                         r = PTR_ERR(eventfp);
1540                         break;
1541                 }
1542                 if (eventfp != vq->kick) {
1543                         pollstop = (filep = vq->kick) != NULL;
1544                         pollstart = (vq->kick = eventfp) != NULL;
1545                 } else
1546                         filep = eventfp;
1547                 break;
1548         case VHOST_SET_VRING_CALL:
1549                 if (copy_from_user(&f, argp, sizeof f)) {
1550                         r = -EFAULT;
1551                         break;
1552                 }
1553                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1554                 if (IS_ERR(ctx)) {
1555                         r = PTR_ERR(ctx);
1556                         break;
1557                 }
1558                 swap(ctx, vq->call_ctx);
1559                 break;
1560         case VHOST_SET_VRING_ERR:
1561                 if (copy_from_user(&f, argp, sizeof f)) {
1562                         r = -EFAULT;
1563                         break;
1564                 }
1565                 ctx = f.fd == -1 ? NULL : eventfd_ctx_fdget(f.fd);
1566                 if (IS_ERR(ctx)) {
1567                         r = PTR_ERR(ctx);
1568                         break;
1569                 }
1570                 swap(ctx, vq->error_ctx);
1571                 break;
1572         case VHOST_SET_VRING_ENDIAN:
1573                 r = vhost_set_vring_endian(vq, argp);
1574                 break;
1575         case VHOST_GET_VRING_ENDIAN:
1576                 r = vhost_get_vring_endian(vq, idx, argp);
1577                 break;
1578         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1579                 if (copy_from_user(&s, argp, sizeof(s))) {
1580                         r = -EFAULT;
1581                         break;
1582                 }
1583                 vq->busyloop_timeout = s.num;
1584                 break;
1585         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1586                 s.index = idx;
1587                 s.num = vq->busyloop_timeout;
1588                 if (copy_to_user(argp, &s, sizeof(s)))
1589                         r = -EFAULT;
1590                 break;
1591         default:
1592                 r = -ENOIOCTLCMD;
1593         }
1594
1595         if (pollstop && vq->handle_kick)
1596                 vhost_poll_stop(&vq->poll);
1597
1598         if (!IS_ERR_OR_NULL(ctx))
1599                 eventfd_ctx_put(ctx);
1600         if (filep)
1601                 fput(filep);
1602
1603         if (pollstart && vq->handle_kick)
1604                 r = vhost_poll_start(&vq->poll, vq->kick);
1605
1606         mutex_unlock(&vq->mutex);
1607
1608         if (pollstop && vq->handle_kick)
1609                 vhost_poll_flush(&vq->poll);
1610         return r;
1611 }
1612 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1613
1614 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1615 {
1616         struct vhost_umem *niotlb, *oiotlb;
1617         int i;
1618
1619         niotlb = vhost_umem_alloc();
1620         if (!niotlb)
1621                 return -ENOMEM;
1622
1623         oiotlb = d->iotlb;
1624         d->iotlb = niotlb;
1625
1626         for (i = 0; i < d->nvqs; ++i) {
1627                 struct vhost_virtqueue *vq = d->vqs[i];
1628
1629                 mutex_lock(&vq->mutex);
1630                 vq->iotlb = niotlb;
1631                 __vhost_vq_meta_reset(vq);
1632                 mutex_unlock(&vq->mutex);
1633         }
1634
1635         vhost_umem_clean(oiotlb);
1636
1637         return 0;
1638 }
1639 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1640
1641 /* Caller must have device mutex */
1642 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1643 {
1644         struct eventfd_ctx *ctx;
1645         u64 p;
1646         long r;
1647         int i, fd;
1648
1649         /* If you are not the owner, you can become one */
1650         if (ioctl == VHOST_SET_OWNER) {
1651                 r = vhost_dev_set_owner(d);
1652                 goto done;
1653         }
1654
1655         /* You must be the owner to do anything else */
1656         r = vhost_dev_check_owner(d);
1657         if (r)
1658                 goto done;
1659
1660         switch (ioctl) {
1661         case VHOST_SET_MEM_TABLE:
1662                 r = vhost_set_memory(d, argp);
1663                 break;
1664         case VHOST_SET_LOG_BASE:
1665                 if (copy_from_user(&p, argp, sizeof p)) {
1666                         r = -EFAULT;
1667                         break;
1668                 }
1669                 if ((u64)(unsigned long)p != p) {
1670                         r = -EFAULT;
1671                         break;
1672                 }
1673                 for (i = 0; i < d->nvqs; ++i) {
1674                         struct vhost_virtqueue *vq;
1675                         void __user *base = (void __user *)(unsigned long)p;
1676                         vq = d->vqs[i];
1677                         mutex_lock(&vq->mutex);
1678                         /* If ring is inactive, will check when it's enabled. */
1679                         if (vq->private_data && !vq_log_access_ok(vq, base))
1680                                 r = -EFAULT;
1681                         else
1682                                 vq->log_base = base;
1683                         mutex_unlock(&vq->mutex);
1684                 }
1685                 break;
1686         case VHOST_SET_LOG_FD:
1687                 r = get_user(fd, (int __user *)argp);
1688                 if (r < 0)
1689                         break;
1690                 ctx = fd == -1 ? NULL : eventfd_ctx_fdget(fd);
1691                 if (IS_ERR(ctx)) {
1692                         r = PTR_ERR(ctx);
1693                         break;
1694                 }
1695                 swap(ctx, d->log_ctx);
1696                 for (i = 0; i < d->nvqs; ++i) {
1697                         mutex_lock(&d->vqs[i]->mutex);
1698                         d->vqs[i]->log_ctx = d->log_ctx;
1699                         mutex_unlock(&d->vqs[i]->mutex);
1700                 }
1701                 if (ctx)
1702                         eventfd_ctx_put(ctx);
1703                 break;
1704         default:
1705                 r = -ENOIOCTLCMD;
1706                 break;
1707         }
1708 done:
1709         return r;
1710 }
1711 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1712
1713 /* TODO: This is really inefficient.  We need something like get_user()
1714  * (instruction directly accesses the data, with an exception table entry
1715  * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1716  */
1717 static int set_bit_to_user(int nr, void __user *addr)
1718 {
1719         unsigned long log = (unsigned long)addr;
1720         struct page *page;
1721         void *base;
1722         int bit = nr + (log % PAGE_SIZE) * 8;
1723         int r;
1724
1725         r = get_user_pages_fast(log, 1, FOLL_WRITE, &page);
1726         if (r < 0)
1727                 return r;
1728         BUG_ON(r != 1);
1729         base = kmap_atomic(page);
1730         set_bit(bit, base);
1731         kunmap_atomic(base);
1732         set_page_dirty_lock(page);
1733         put_page(page);
1734         return 0;
1735 }
1736
1737 static int log_write(void __user *log_base,
1738                      u64 write_address, u64 write_length)
1739 {
1740         u64 write_page = write_address / VHOST_PAGE_SIZE;
1741         int r;
1742
1743         if (!write_length)
1744                 return 0;
1745         write_length += write_address % VHOST_PAGE_SIZE;
1746         for (;;) {
1747                 u64 base = (u64)(unsigned long)log_base;
1748                 u64 log = base + write_page / 8;
1749                 int bit = write_page % 8;
1750                 if ((u64)(unsigned long)log != log)
1751                         return -EFAULT;
1752                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1753                 if (r < 0)
1754                         return r;
1755                 if (write_length <= VHOST_PAGE_SIZE)
1756                         break;
1757                 write_length -= VHOST_PAGE_SIZE;
1758                 write_page += 1;
1759         }
1760         return r;
1761 }
1762
1763 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1764 {
1765         struct vhost_umem *umem = vq->umem;
1766         struct vhost_umem_node *u;
1767         u64 start, end, l, min;
1768         int r;
1769         bool hit = false;
1770
1771         while (len) {
1772                 min = len;
1773                 /* More than one GPAs can be mapped into a single HVA. So
1774                  * iterate all possible umems here to be safe.
1775                  */
1776                 list_for_each_entry(u, &umem->umem_list, link) {
1777                         if (u->userspace_addr > hva - 1 + len ||
1778                             u->userspace_addr - 1 + u->size < hva)
1779                                 continue;
1780                         start = max(u->userspace_addr, hva);
1781                         end = min(u->userspace_addr - 1 + u->size,
1782                                   hva - 1 + len);
1783                         l = end - start + 1;
1784                         r = log_write(vq->log_base,
1785                                       u->start + start - u->userspace_addr,
1786                                       l);
1787                         if (r < 0)
1788                                 return r;
1789                         hit = true;
1790                         min = min(l, min);
1791                 }
1792
1793                 if (!hit)
1794                         return -EFAULT;
1795
1796                 len -= min;
1797                 hva += min;
1798         }
1799
1800         return 0;
1801 }
1802
1803 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1804 {
1805         struct iovec iov[64];
1806         int i, ret;
1807
1808         if (!vq->iotlb)
1809                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1810
1811         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1812                              len, iov, 64, VHOST_ACCESS_WO);
1813         if (ret < 0)
1814                 return ret;
1815
1816         for (i = 0; i < ret; i++) {
1817                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1818                                     iov[i].iov_len);
1819                 if (ret)
1820                         return ret;
1821         }
1822
1823         return 0;
1824 }
1825
1826 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1827                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1828 {
1829         int i, r;
1830
1831         /* Make sure data written is seen before log. */
1832         smp_wmb();
1833
1834         if (vq->iotlb) {
1835                 for (i = 0; i < count; i++) {
1836                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1837                                           iov[i].iov_len);
1838                         if (r < 0)
1839                                 return r;
1840                 }
1841                 return 0;
1842         }
1843
1844         for (i = 0; i < log_num; ++i) {
1845                 u64 l = min(log[i].len, len);
1846                 r = log_write(vq->log_base, log[i].addr, l);
1847                 if (r < 0)
1848                         return r;
1849                 len -= l;
1850                 if (!len) {
1851                         if (vq->log_ctx)
1852                                 eventfd_signal(vq->log_ctx, 1);
1853                         return 0;
1854                 }
1855         }
1856         /* Length written exceeds what we have stored. This is a bug. */
1857         BUG();
1858         return 0;
1859 }
1860 EXPORT_SYMBOL_GPL(vhost_log_write);
1861
1862 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1863 {
1864         void __user *used;
1865         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1866                            &vq->used->flags) < 0)
1867                 return -EFAULT;
1868         if (unlikely(vq->log_used)) {
1869                 /* Make sure the flag is seen before log. */
1870                 smp_wmb();
1871                 /* Log used flag write. */
1872                 used = &vq->used->flags;
1873                 log_used(vq, (used - (void __user *)vq->used),
1874                          sizeof vq->used->flags);
1875                 if (vq->log_ctx)
1876                         eventfd_signal(vq->log_ctx, 1);
1877         }
1878         return 0;
1879 }
1880
1881 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1882 {
1883         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1884                            vhost_avail_event(vq)))
1885                 return -EFAULT;
1886         if (unlikely(vq->log_used)) {
1887                 void __user *used;
1888                 /* Make sure the event is seen before log. */
1889                 smp_wmb();
1890                 /* Log avail event write */
1891                 used = vhost_avail_event(vq);
1892                 log_used(vq, (used - (void __user *)vq->used),
1893                          sizeof *vhost_avail_event(vq));
1894                 if (vq->log_ctx)
1895                         eventfd_signal(vq->log_ctx, 1);
1896         }
1897         return 0;
1898 }
1899
1900 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1901 {
1902         __virtio16 last_used_idx;
1903         int r;
1904         bool is_le = vq->is_le;
1905
1906         if (!vq->private_data)
1907                 return 0;
1908
1909         vhost_init_is_le(vq);
1910
1911         r = vhost_update_used_flags(vq);
1912         if (r)
1913                 goto err;
1914         vq->signalled_used_valid = false;
1915         if (!vq->iotlb &&
1916             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
1917                 r = -EFAULT;
1918                 goto err;
1919         }
1920         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1921         if (r) {
1922                 vq_err(vq, "Can't access used idx at %p\n",
1923                        &vq->used->idx);
1924                 goto err;
1925         }
1926         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1927         return 0;
1928
1929 err:
1930         vq->is_le = is_le;
1931         return r;
1932 }
1933 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1934
1935 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1936                           struct iovec iov[], int iov_size, int access)
1937 {
1938         const struct vhost_umem_node *node;
1939         struct vhost_dev *dev = vq->dev;
1940         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1941         struct iovec *_iov;
1942         u64 s = 0;
1943         int ret = 0;
1944
1945         while ((u64)len > s) {
1946                 u64 size;
1947                 if (unlikely(ret >= iov_size)) {
1948                         ret = -ENOBUFS;
1949                         break;
1950                 }
1951
1952                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1953                                                         addr, addr + len - 1);
1954                 if (node == NULL || node->start > addr) {
1955                         if (umem != dev->iotlb) {
1956                                 ret = -EFAULT;
1957                                 break;
1958                         }
1959                         ret = -EAGAIN;
1960                         break;
1961                 } else if (!(node->perm & access)) {
1962                         ret = -EPERM;
1963                         break;
1964                 }
1965
1966                 _iov = iov + ret;
1967                 size = node->size - addr + node->start;
1968                 _iov->iov_len = min((u64)len - s, size);
1969                 _iov->iov_base = (void __user *)(unsigned long)
1970                         (node->userspace_addr + addr - node->start);
1971                 s += size;
1972                 addr += size;
1973                 ++ret;
1974         }
1975
1976         if (ret == -EAGAIN)
1977                 vhost_iotlb_miss(vq, addr, access);
1978         return ret;
1979 }
1980
1981 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1982  * function returns the next descriptor in the chain,
1983  * or -1U if we're at the end. */
1984 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1985 {
1986         unsigned int next;
1987
1988         /* If this descriptor says it doesn't chain, we're done. */
1989         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1990                 return -1U;
1991
1992         /* Check they're not leading us off end of descriptors. */
1993         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
1994         return next;
1995 }
1996
1997 static int get_indirect(struct vhost_virtqueue *vq,
1998                         struct iovec iov[], unsigned int iov_size,
1999                         unsigned int *out_num, unsigned int *in_num,
2000                         struct vhost_log *log, unsigned int *log_num,
2001                         struct vring_desc *indirect)
2002 {
2003         struct vring_desc desc;
2004         unsigned int i = 0, count, found = 0;
2005         u32 len = vhost32_to_cpu(vq, indirect->len);
2006         struct iov_iter from;
2007         int ret, access;
2008
2009         /* Sanity check */
2010         if (unlikely(len % sizeof desc)) {
2011                 vq_err(vq, "Invalid length in indirect descriptor: "
2012                        "len 0x%llx not multiple of 0x%zx\n",
2013                        (unsigned long long)len,
2014                        sizeof desc);
2015                 return -EINVAL;
2016         }
2017
2018         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2019                              UIO_MAXIOV, VHOST_ACCESS_RO);
2020         if (unlikely(ret < 0)) {
2021                 if (ret != -EAGAIN)
2022                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2023                 return ret;
2024         }
2025         iov_iter_init(&from, READ, vq->indirect, ret, len);
2026
2027         /* We will use the result as an address to read from, so most
2028          * architectures only need a compiler barrier here. */
2029         read_barrier_depends();
2030
2031         count = len / sizeof desc;
2032         /* Buffers are chained via a 16 bit next field, so
2033          * we can have at most 2^16 of these. */
2034         if (unlikely(count > USHRT_MAX + 1)) {
2035                 vq_err(vq, "Indirect buffer length too big: %d\n",
2036                        indirect->len);
2037                 return -E2BIG;
2038         }
2039
2040         do {
2041                 unsigned iov_count = *in_num + *out_num;
2042                 if (unlikely(++found > count)) {
2043                         vq_err(vq, "Loop detected: last one at %u "
2044                                "indirect size %u\n",
2045                                i, count);
2046                         return -EINVAL;
2047                 }
2048                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2049                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2050                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2051                         return -EINVAL;
2052                 }
2053                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2054                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2055                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2056                         return -EINVAL;
2057                 }
2058
2059                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2060                         access = VHOST_ACCESS_WO;
2061                 else
2062                         access = VHOST_ACCESS_RO;
2063
2064                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2065                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2066                                      iov_size - iov_count, access);
2067                 if (unlikely(ret < 0)) {
2068                         if (ret != -EAGAIN)
2069                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2070                                         ret, i);
2071                         return ret;
2072                 }
2073                 /* If this is an input descriptor, increment that count. */
2074                 if (access == VHOST_ACCESS_WO) {
2075                         *in_num += ret;
2076                         if (unlikely(log)) {
2077                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2078                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2079                                 ++*log_num;
2080                         }
2081                 } else {
2082                         /* If it's an output descriptor, they're all supposed
2083                          * to come before any input descriptors. */
2084                         if (unlikely(*in_num)) {
2085                                 vq_err(vq, "Indirect descriptor "
2086                                        "has out after in: idx %d\n", i);
2087                                 return -EINVAL;
2088                         }
2089                         *out_num += ret;
2090                 }
2091         } while ((i = next_desc(vq, &desc)) != -1);
2092         return 0;
2093 }
2094
2095 /* This looks in the virtqueue and for the first available buffer, and converts
2096  * it to an iovec for convenient access.  Since descriptors consist of some
2097  * number of output then some number of input descriptors, it's actually two
2098  * iovecs, but we pack them into one and note how many of each there were.
2099  *
2100  * This function returns the descriptor number found, or vq->num (which is
2101  * never a valid descriptor number) if none was found.  A negative code is
2102  * returned on error. */
2103 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2104                       struct iovec iov[], unsigned int iov_size,
2105                       unsigned int *out_num, unsigned int *in_num,
2106                       struct vhost_log *log, unsigned int *log_num)
2107 {
2108         struct vring_desc desc;
2109         unsigned int i, head, found = 0;
2110         u16 last_avail_idx;
2111         __virtio16 avail_idx;
2112         __virtio16 ring_head;
2113         int ret, access;
2114
2115         /* Check it isn't doing very strange things with descriptor numbers. */
2116         last_avail_idx = vq->last_avail_idx;
2117
2118         if (vq->avail_idx == vq->last_avail_idx) {
2119                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2120                         vq_err(vq, "Failed to access avail idx at %p\n",
2121                                 &vq->avail->idx);
2122                         return -EFAULT;
2123                 }
2124                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2125
2126                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2127                         vq_err(vq, "Guest moved used index from %u to %u",
2128                                 last_avail_idx, vq->avail_idx);
2129                         return -EFAULT;
2130                 }
2131
2132                 /* If there's nothing new since last we looked, return
2133                  * invalid.
2134                  */
2135                 if (vq->avail_idx == last_avail_idx)
2136                         return vq->num;
2137
2138                 /* Only get avail ring entries after they have been
2139                  * exposed by guest.
2140                  */
2141                 smp_rmb();
2142         }
2143
2144         /* Grab the next descriptor number they're advertising, and increment
2145          * the index we've seen. */
2146         if (unlikely(vhost_get_avail(vq, ring_head,
2147                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2148                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2149                        last_avail_idx,
2150                        &vq->avail->ring[last_avail_idx % vq->num]);
2151                 return -EFAULT;
2152         }
2153
2154         head = vhost16_to_cpu(vq, ring_head);
2155
2156         /* If their number is silly, that's an error. */
2157         if (unlikely(head >= vq->num)) {
2158                 vq_err(vq, "Guest says index %u > %u is available",
2159                        head, vq->num);
2160                 return -EINVAL;
2161         }
2162
2163         /* When we start there are none of either input nor output. */
2164         *out_num = *in_num = 0;
2165         if (unlikely(log))
2166                 *log_num = 0;
2167
2168         i = head;
2169         do {
2170                 unsigned iov_count = *in_num + *out_num;
2171                 if (unlikely(i >= vq->num)) {
2172                         vq_err(vq, "Desc index is %u > %u, head = %u",
2173                                i, vq->num, head);
2174                         return -EINVAL;
2175                 }
2176                 if (unlikely(++found > vq->num)) {
2177                         vq_err(vq, "Loop detected: last one at %u "
2178                                "vq size %u head %u\n",
2179                                i, vq->num, head);
2180                         return -EINVAL;
2181                 }
2182                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2183                                            sizeof desc);
2184                 if (unlikely(ret)) {
2185                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2186                                i, vq->desc + i);
2187                         return -EFAULT;
2188                 }
2189                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2190                         ret = get_indirect(vq, iov, iov_size,
2191                                            out_num, in_num,
2192                                            log, log_num, &desc);
2193                         if (unlikely(ret < 0)) {
2194                                 if (ret != -EAGAIN)
2195                                         vq_err(vq, "Failure detected "
2196                                                 "in indirect descriptor at idx %d\n", i);
2197                                 return ret;
2198                         }
2199                         continue;
2200                 }
2201
2202                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2203                         access = VHOST_ACCESS_WO;
2204                 else
2205                         access = VHOST_ACCESS_RO;
2206                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2207                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2208                                      iov_size - iov_count, access);
2209                 if (unlikely(ret < 0)) {
2210                         if (ret != -EAGAIN)
2211                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2212                                         ret, i);
2213                         return ret;
2214                 }
2215                 if (access == VHOST_ACCESS_WO) {
2216                         /* If this is an input descriptor,
2217                          * increment that count. */
2218                         *in_num += ret;
2219                         if (unlikely(log)) {
2220                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2221                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2222                                 ++*log_num;
2223                         }
2224                 } else {
2225                         /* If it's an output descriptor, they're all supposed
2226                          * to come before any input descriptors. */
2227                         if (unlikely(*in_num)) {
2228                                 vq_err(vq, "Descriptor has out after in: "
2229                                        "idx %d\n", i);
2230                                 return -EINVAL;
2231                         }
2232                         *out_num += ret;
2233                 }
2234         } while ((i = next_desc(vq, &desc)) != -1);
2235
2236         /* On success, increment avail index. */
2237         vq->last_avail_idx++;
2238
2239         /* Assume notifications from guest are disabled at this point,
2240          * if they aren't we would need to update avail_event index. */
2241         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2242         return head;
2243 }
2244 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2245
2246 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2247 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2248 {
2249         vq->last_avail_idx -= n;
2250 }
2251 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2252
2253 /* After we've used one of their buffers, we tell them about it.  We'll then
2254  * want to notify the guest, using eventfd. */
2255 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2256 {
2257         struct vring_used_elem heads = {
2258                 cpu_to_vhost32(vq, head),
2259                 cpu_to_vhost32(vq, len)
2260         };
2261
2262         return vhost_add_used_n(vq, &heads, 1);
2263 }
2264 EXPORT_SYMBOL_GPL(vhost_add_used);
2265
2266 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2267                             struct vring_used_elem *heads,
2268                             unsigned count)
2269 {
2270         struct vring_used_elem __user *used;
2271         u16 old, new;
2272         int start;
2273
2274         start = vq->last_used_idx & (vq->num - 1);
2275         used = vq->used->ring + start;
2276         if (count == 1) {
2277                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2278                         vq_err(vq, "Failed to write used id");
2279                         return -EFAULT;
2280                 }
2281                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2282                         vq_err(vq, "Failed to write used len");
2283                         return -EFAULT;
2284                 }
2285         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2286                 vq_err(vq, "Failed to write used");
2287                 return -EFAULT;
2288         }
2289         if (unlikely(vq->log_used)) {
2290                 /* Make sure data is seen before log. */
2291                 smp_wmb();
2292                 /* Log used ring entry write. */
2293                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2294                          count * sizeof *used);
2295         }
2296         old = vq->last_used_idx;
2297         new = (vq->last_used_idx += count);
2298         /* If the driver never bothers to signal in a very long while,
2299          * used index might wrap around. If that happens, invalidate
2300          * signalled_used index we stored. TODO: make sure driver
2301          * signals at least once in 2^16 and remove this. */
2302         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2303                 vq->signalled_used_valid = false;
2304         return 0;
2305 }
2306
2307 /* After we've used one of their buffers, we tell them about it.  We'll then
2308  * want to notify the guest, using eventfd. */
2309 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2310                      unsigned count)
2311 {
2312         int start, n, r;
2313
2314         start = vq->last_used_idx & (vq->num - 1);
2315         n = vq->num - start;
2316         if (n < count) {
2317                 r = __vhost_add_used_n(vq, heads, n);
2318                 if (r < 0)
2319                         return r;
2320                 heads += n;
2321                 count -= n;
2322         }
2323         r = __vhost_add_used_n(vq, heads, count);
2324
2325         /* Make sure buffer is written before we update index. */
2326         smp_wmb();
2327         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2328                            &vq->used->idx)) {
2329                 vq_err(vq, "Failed to increment used idx");
2330                 return -EFAULT;
2331         }
2332         if (unlikely(vq->log_used)) {
2333                 /* Make sure used idx is seen before log. */
2334                 smp_wmb();
2335                 /* Log used index update. */
2336                 log_used(vq, offsetof(struct vring_used, idx),
2337                          sizeof vq->used->idx);
2338                 if (vq->log_ctx)
2339                         eventfd_signal(vq->log_ctx, 1);
2340         }
2341         return r;
2342 }
2343 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2344
2345 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2346 {
2347         __u16 old, new;
2348         __virtio16 event;
2349         bool v;
2350         /* Flush out used index updates. This is paired
2351          * with the barrier that the Guest executes when enabling
2352          * interrupts. */
2353         smp_mb();
2354
2355         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2356             unlikely(vq->avail_idx == vq->last_avail_idx))
2357                 return true;
2358
2359         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2360                 __virtio16 flags;
2361                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2362                         vq_err(vq, "Failed to get flags");
2363                         return true;
2364                 }
2365                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2366         }
2367         old = vq->signalled_used;
2368         v = vq->signalled_used_valid;
2369         new = vq->signalled_used = vq->last_used_idx;
2370         vq->signalled_used_valid = true;
2371
2372         if (unlikely(!v))
2373                 return true;
2374
2375         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2376                 vq_err(vq, "Failed to get used event idx");
2377                 return true;
2378         }
2379         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2380 }
2381
2382 /* This actually signals the guest, using eventfd. */
2383 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2384 {
2385         /* Signal the Guest tell them we used something up. */
2386         if (vq->call_ctx && vhost_notify(dev, vq))
2387                 eventfd_signal(vq->call_ctx, 1);
2388 }
2389 EXPORT_SYMBOL_GPL(vhost_signal);
2390
2391 /* And here's the combo meal deal.  Supersize me! */
2392 void vhost_add_used_and_signal(struct vhost_dev *dev,
2393                                struct vhost_virtqueue *vq,
2394                                unsigned int head, int len)
2395 {
2396         vhost_add_used(vq, head, len);
2397         vhost_signal(dev, vq);
2398 }
2399 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2400
2401 /* multi-buffer version of vhost_add_used_and_signal */
2402 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2403                                  struct vhost_virtqueue *vq,
2404                                  struct vring_used_elem *heads, unsigned count)
2405 {
2406         vhost_add_used_n(vq, heads, count);
2407         vhost_signal(dev, vq);
2408 }
2409 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2410
2411 /* return true if we're sure that avaiable ring is empty */
2412 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2413 {
2414         __virtio16 avail_idx;
2415         int r;
2416
2417         if (vq->avail_idx != vq->last_avail_idx)
2418                 return false;
2419
2420         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2421         if (unlikely(r))
2422                 return false;
2423         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2424
2425         return vq->avail_idx == vq->last_avail_idx;
2426 }
2427 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2428
2429 /* OK, now we need to know about added descriptors. */
2430 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2431 {
2432         __virtio16 avail_idx;
2433         int r;
2434
2435         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2436                 return false;
2437         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2438         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2439                 r = vhost_update_used_flags(vq);
2440                 if (r) {
2441                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2442                                &vq->used->flags, r);
2443                         return false;
2444                 }
2445         } else {
2446                 r = vhost_update_avail_event(vq, vq->avail_idx);
2447                 if (r) {
2448                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2449                                vhost_avail_event(vq), r);
2450                         return false;
2451                 }
2452         }
2453         /* They could have slipped one in as we were doing that: make
2454          * sure it's written, then check again. */
2455         smp_mb();
2456         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2457         if (r) {
2458                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2459                        &vq->avail->idx, r);
2460                 return false;
2461         }
2462
2463         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2464 }
2465 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2466
2467 /* We don't need to be notified again. */
2468 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2469 {
2470         int r;
2471
2472         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2473                 return;
2474         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2475         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2476                 r = vhost_update_used_flags(vq);
2477                 if (r)
2478                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2479                                &vq->used->flags, r);
2480         }
2481 }
2482 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2483
2484 /* Create a new message. */
2485 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2486 {
2487         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2488         if (!node)
2489                 return NULL;
2490
2491         /* Make sure all padding within the structure is initialized. */
2492         memset(&node->msg, 0, sizeof node->msg);
2493         node->vq = vq;
2494         node->msg.type = type;
2495         return node;
2496 }
2497 EXPORT_SYMBOL_GPL(vhost_new_msg);
2498
2499 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2500                        struct vhost_msg_node *node)
2501 {
2502         spin_lock(&dev->iotlb_lock);
2503         list_add_tail(&node->node, head);
2504         spin_unlock(&dev->iotlb_lock);
2505
2506         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2507 }
2508 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2509
2510 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2511                                          struct list_head *head)
2512 {
2513         struct vhost_msg_node *node = NULL;
2514
2515         spin_lock(&dev->iotlb_lock);
2516         if (!list_empty(head)) {
2517                 node = list_first_entry(head, struct vhost_msg_node,
2518                                         node);
2519                 list_del(&node->node);
2520         }
2521         spin_unlock(&dev->iotlb_lock);
2522
2523         return node;
2524 }
2525 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2526
2527
2528 static int __init vhost_init(void)
2529 {
2530         return 0;
2531 }
2532
2533 static void __exit vhost_exit(void)
2534 {
2535 }
2536
2537 module_init(vhost_init);
2538 module_exit(vhost_exit);
2539
2540 MODULE_VERSION("0.0.1");
2541 MODULE_LICENSE("GPL v2");
2542 MODULE_AUTHOR("Michael S. Tsirkin");
2543 MODULE_DESCRIPTION("Host kernel accelerator for virtio");