drivers/vhost/vhost.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Copyright (C) 2009 Red Hat, Inc.
   3  * Copyright (C) 2006 Rusty Russell IBM Corporation
   4  *
   5  * Author: Michael S. Tsirkin <mst@redhat.com>
   6  *
   7  * Inspiration, some code, and most witty comments come from
   8  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   9  *
  10  * Generic code for virtio server in host kernel.
  11  */
  12
  13 #include <linux/eventfd.h>
  14 #include <linux/vhost.h>
  15 #include <linux/uio.h>
  16 #include <linux/mm.h>
  17 #include <linux/miscdevice.h>
  18 #include <linux/mutex.h>
  19 #include <linux/poll.h>
  20 #include <linux/file.h>
  21 #include <linux/highmem.h>
  22 #include <linux/slab.h>
  23 #include <linux/vmalloc.h>
  24 #include <linux/kthread.h>
  25 #include <linux/cgroup.h>
  26 #include <linux/module.h>
  27 #include <linux/sort.h>
  28 #include <linux/sched/mm.h>
  29 #include <linux/sched/signal.h>
  30 #include <linux/interval_tree_generic.h>
  31 #include <linux/nospec.h>
  32 #include <linux/kcov.h>
  33
  34 #include "vhost.h"
  35
  36 static ushort max_mem_regions = 64;
  37 module_param(max_mem_regions, ushort, 0444);
  38 MODULE_PARM_DESC(max_mem_regions,
  39         "Maximum number of memory regions in memory map. (default: 64)");
  40 static int max_iotlb_entries = 2048;
  41 module_param(max_iotlb_entries, int, 0444);
  42 MODULE_PARM_DESC(max_iotlb_entries,
  43         "Maximum number of iotlb entries. (default: 2048)");
  44
  45 enum {
  46         VHOST_MEMORY_F_LOG = 0x1,
  47 };
  48
  49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  53 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  54 {
  55         vq->user_be = !virtio_legacy_is_little_endian();
  56 }
  57
  58 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  59 {
  60         vq->user_be = true;
  61 }
  62
  63 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  64 {
  65         vq->user_be = false;
  66 }
  67
  68 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  69 {
  70         struct vhost_vring_state s;
  71
  72         if (vq->private_data)
  73                 return -EBUSY;
  74
  75         if (copy_from_user(&s, argp, sizeof(s)))
  76                 return -EFAULT;
  77
  78         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  79             s.num != VHOST_VRING_BIG_ENDIAN)
  80                 return -EINVAL;
  81
  82         if (s.num == VHOST_VRING_BIG_ENDIAN)
  83                 vhost_enable_cross_endian_big(vq);
  84         else
  85                 vhost_enable_cross_endian_little(vq);
  86
  87         return 0;
  88 }
  89
  90 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  91                                    int __user *argp)
  92 {
  93         struct vhost_vring_state s = {
  94                 .index = idx,
  95                 .num = vq->user_be
  96         };
  97
  98         if (copy_to_user(argp, &s, sizeof(s)))
  99                 return -EFAULT;
 100
 101         return 0;
 102 }
 103
 104 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 105 {
 106         /* Note for legacy virtio: user_be is initialized at reset time
 107          * according to the host endianness. If userspace does not set an
 108          * explicit endianness, the default behavior is native endian, as
 109          * expected by legacy virtio.
 110          */
 111         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 112 }
 113 #else
 114 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 115 {
 116 }
 117
 118 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 119 {
 120         return -ENOIOCTLCMD;
 121 }
 122
 123 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 124                                    int __user *argp)
 125 {
 126         return -ENOIOCTLCMD;
 127 }
 128
 129 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 130 {
 131         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 132                 || virtio_legacy_is_little_endian();
 133 }
 134 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 135
 136 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 137 {
 138         vhost_init_is_le(vq);
 139 }
 140
 141 struct vhost_flush_struct {
 142         struct vhost_work work;
 143         struct completion wait_event;
 144 };
 145
 146 static void vhost_flush_work(struct vhost_work *work)
 147 {
 148         struct vhost_flush_struct *s;
 149
 150         s = container_of(work, struct vhost_flush_struct, work);
 151         complete(&s->wait_event);
 152 }
 153
 154 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 155                             poll_table *pt)
 156 {
 157         struct vhost_poll *poll;
 158
 159         poll = container_of(pt, struct vhost_poll, table);
 160         poll->wqh = wqh;
 161         add_wait_queue(wqh, &poll->wait);
 162 }
 163
 164 static int vhost_poll_wakeup(wait_queue_entry_t *wait, unsigned mode, int sync,
 165                              void *key)
 166 {
 167         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 168         struct vhost_work *work = &poll->work;
 169
 170         if (!(key_to_poll(key) & poll->mask))
 171                 return 0;
 172
 173         if (!poll->dev->use_worker)
 174                 work->fn(work);
 175         else
 176                 vhost_poll_queue(poll);
 177
 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
 208         if (poll->wqh)
 209                 return 0;
 210
 211         mask = vfs_poll(file, &poll->table);
 212         if (mask)
 213                 vhost_poll_wakeup(&poll->wait, 0, 0, poll_to_key(mask));
 214         if (mask & EPOLLERR) {
 215                 vhost_poll_stop(poll);
 216                 return -EINVAL;
 217         }
 218
 219         return 0;
 220 }
 221 EXPORT_SYMBOL_GPL(vhost_poll_start);
 222
 223 /* Stop polling a file. After this function returns, it becomes safe to drop the
 224  * file reference. You must also flush afterwards. */
 225 void vhost_poll_stop(struct vhost_poll *poll)
 226 {
 227         if (poll->wqh) {
 228                 remove_wait_queue(poll->wqh, &poll->wait);
 229                 poll->wqh = NULL;
 230         }
 231 }
 232 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 233
 234 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 235 {
 236         struct vhost_flush_struct flush;
 237
 238         if (dev->worker) {
 239                 init_completion(&flush.wait_event);
 240                 vhost_work_init(&flush.work, vhost_flush_work);
 241
 242                 vhost_work_queue(dev, &flush.work);
 243                 wait_for_completion(&flush.wait_event);
 244         }
 245 }
 246 EXPORT_SYMBOL_GPL(vhost_work_flush);
 247
 248 /* Flush any work that has been scheduled. When calling this, don't hold any
 249  * locks that are also used by the callback. */
 250 void vhost_poll_flush(struct vhost_poll *poll)
 251 {
 252         vhost_work_flush(poll->dev, &poll->work);
 253 }
 254 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 255
 256 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 257 {
 258         if (!dev->worker)
 259                 return;
 260
 261         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 262                 /* We can only add the work to the list after we're
 263                  * sure it was not in the list.
 264                  * test_and_set_bit() implies a memory barrier.
 265                  */
 266                 llist_add(&work->node, &dev->work_list);
 267                 wake_up_process(dev->worker);
 268         }
 269 }
 270 EXPORT_SYMBOL_GPL(vhost_work_queue);
 271
 272 /* A lockless hint for busy polling code to exit the loop */
 273 bool vhost_has_work(struct vhost_dev *dev)
 274 {
 275         return !llist_empty(&dev->work_list);
 276 }
 277 EXPORT_SYMBOL_GPL(vhost_has_work);
 278
 279 void vhost_poll_queue(struct vhost_poll *poll)
 280 {
 281         vhost_work_queue(poll->dev, &poll->work);
 282 }
 283 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 284
 285 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 286 {
 287         int j;
 288
 289         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 290                 vq->meta_iotlb[j] = NULL;
 291 }
 292
 293 static void vhost_vq_meta_reset(struct vhost_dev *d)
 294 {
 295         int i;
 296
 297         for (i = 0; i < d->nvqs; ++i)
 298                 __vhost_vq_meta_reset(d->vqs[i]);
 299 }
 300
 301 static void vhost_vq_reset(struct vhost_dev *dev,
 302                            struct vhost_virtqueue *vq)
 303 {
 304         vq->num = 1;
 305         vq->desc = NULL;
 306         vq->avail = NULL;
 307         vq->used = NULL;
 308         vq->last_avail_idx = 0;
 309         vq->avail_idx = 0;
 310         vq->last_used_idx = 0;
 311         vq->signalled_used = 0;
 312         vq->signalled_used_valid = false;
 313         vq->used_flags = 0;
 314         vq->log_used = false;
 315         vq->log_addr = -1ull;
 316         vq->private_data = NULL;
 317         vq->acked_features = 0;
 318         vq->acked_backend_features = 0;
 319         vq->log_base = NULL;
 320         vq->error_ctx = NULL;
 321         vq->kick = NULL;
 322         vq->call_ctx = NULL;
 323         vq->log_ctx = NULL;
 324         vhost_reset_is_le(vq);
 325         vhost_disable_cross_endian(vq);
 326         vq->busyloop_timeout = 0;
 327         vq->umem = NULL;
 328         vq->iotlb = NULL;
 329         __vhost_vq_meta_reset(vq);
 330 }
 331
 332 static int vhost_worker(void *data)
 333 {
 334         struct vhost_dev *dev = data;
 335         struct vhost_work *work, *work_next;
 336         struct llist_node *node;
 337
 338         kthread_use_mm(dev->mm);
 339
 340         for (;;) {
 341                 /* mb paired w/ kthread_stop */
 342                 set_current_state(TASK_INTERRUPTIBLE);
 343
 344                 if (kthread_should_stop()) {
 345                         __set_current_state(TASK_RUNNING);
 346                         break;
 347                 }
 348
 349                 node = llist_del_all(&dev->work_list);
 350                 if (!node)
 351                         schedule();
 352
 353                 node = llist_reverse_order(node);
 354                 /* make sure flag is seen after deletion */
 355                 smp_wmb();
 356                 llist_for_each_entry_safe(work, work_next, node, node) {
 357                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 358                         __set_current_state(TASK_RUNNING);
 359                         kcov_remote_start_common(dev->kcov_handle);
 360                         work->fn(work);
 361                         kcov_remote_stop();
 362                         if (need_resched())
 363                                 schedule();
 364                 }
 365         }
 366         kthread_unuse_mm(dev->mm);
 367         return 0;
 368 }
 369
 370 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 371 {
 372         kfree(vq->indirect);
 373         vq->indirect = NULL;
 374         kfree(vq->log);
 375         vq->log = NULL;
 376         kfree(vq->heads);
 377         vq->heads = NULL;
 378 }
 379
 380 /* Helper to allocate iovec buffers for all vqs. */
 381 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 382 {
 383         struct vhost_virtqueue *vq;
 384         int i;
 385
 386         for (i = 0; i < dev->nvqs; ++i) {
 387                 vq = dev->vqs[i];
 388                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 389                                              sizeof(*vq->indirect),
 390                                              GFP_KERNEL);
 391                 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 392                                         GFP_KERNEL);
 393                 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 394                                           GFP_KERNEL);
 395                 if (!vq->indirect || !vq->log || !vq->heads)
 396                         goto err_nomem;
 397         }
 398         return 0;
 399
 400 err_nomem:
 401         for (; i >= 0; --i)
 402                 vhost_vq_free_iovecs(dev->vqs[i]);
 403         return -ENOMEM;
 404 }
 405
 406 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 407 {
 408         int i;
 409
 410         for (i = 0; i < dev->nvqs; ++i)
 411                 vhost_vq_free_iovecs(dev->vqs[i]);
 412 }
 413
 414 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 415                           int pkts, int total_len)
 416 {
 417         struct vhost_dev *dev = vq->dev;
 418
 419         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 420             pkts >= dev->weight) {
 421                 vhost_poll_queue(&vq->poll);
 422                 return true;
 423         }
 424
 425         return false;
 426 }
 427 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 428
 429 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 430                                    unsigned int num)
 431 {
 432         size_t event __maybe_unused =
 433                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 434
 435         return sizeof(*vq->avail) +
 436                sizeof(*vq->avail->ring) * num + event;
 437 }
 438
 439 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 440                                   unsigned int num)
 441 {
 442         size_t event __maybe_unused =
 443                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 444
 445         return sizeof(*vq->used) +
 446                sizeof(*vq->used->ring) * num + event;
 447 }
 448
 449 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 450                                   unsigned int num)
 451 {
 452         return sizeof(*vq->desc) * num;
 453 }
 454
 455 void vhost_dev_init(struct vhost_dev *dev,
 456                     struct vhost_virtqueue **vqs, int nvqs,
 457                     int iov_limit, int weight, int byte_weight,
 458                     bool use_worker,
 459                     int (*msg_handler)(struct vhost_dev *dev,
 460                                        struct vhost_iotlb_msg *msg))
 461 {
 462         struct vhost_virtqueue *vq;
 463         int i;
 464
 465         dev->vqs = vqs;
 466         dev->nvqs = nvqs;
 467         mutex_init(&dev->mutex);
 468         dev->log_ctx = NULL;
 469         dev->umem = NULL;
 470         dev->iotlb = NULL;
 471         dev->mm = NULL;
 472         dev->worker = NULL;
 473         dev->iov_limit = iov_limit;
 474         dev->weight = weight;
 475         dev->byte_weight = byte_weight;
 476         dev->use_worker = use_worker;
 477         dev->msg_handler = msg_handler;
 478         init_llist_head(&dev->work_list);
 479         init_waitqueue_head(&dev->wait);
 480         INIT_LIST_HEAD(&dev->read_list);
 481         INIT_LIST_HEAD(&dev->pending_list);
 482         spin_lock_init(&dev->iotlb_lock);
 483
 484
 485         for (i = 0; i < dev->nvqs; ++i) {
 486                 vq = dev->vqs[i];
 487                 vq->log = NULL;
 488                 vq->indirect = NULL;
 489                 vq->heads = NULL;
 490                 vq->dev = dev;
 491                 mutex_init(&vq->mutex);
 492                 vhost_vq_reset(dev, vq);
 493                 if (vq->handle_kick)
 494                         vhost_poll_init(&vq->poll, vq->handle_kick,
 495                                         EPOLLIN, dev);
 496         }
 497 }
 498 EXPORT_SYMBOL_GPL(vhost_dev_init);
 499
 500 /* Caller should have device mutex */
 501 long vhost_dev_check_owner(struct vhost_dev *dev)
 502 {
 503         /* Are you the owner? If not, I don't think you mean to do that */
 504         return dev->mm == current->mm ? 0 : -EPERM;
 505 }
 506 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 507
 508 struct vhost_attach_cgroups_struct {
 509         struct vhost_work work;
 510         struct task_struct *owner;
 511         int ret;
 512 };
 513
 514 static void vhost_attach_cgroups_work(struct vhost_work *work)
 515 {
 516         struct vhost_attach_cgroups_struct *s;
 517
 518         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 519         s->ret = cgroup_attach_task_all(s->owner, current);
 520 }
 521
 522 static int vhost_attach_cgroups(struct vhost_dev *dev)
 523 {
 524         struct vhost_attach_cgroups_struct attach;
 525
 526         attach.owner = current;
 527         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 528         vhost_work_queue(dev, &attach.work);
 529         vhost_work_flush(dev, &attach.work);
 530         return attach.ret;
 531 }
 532
 533 /* Caller should have device mutex */
 534 bool vhost_dev_has_owner(struct vhost_dev *dev)
 535 {
 536         return dev->mm;
 537 }
 538 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 539
 540 static void vhost_attach_mm(struct vhost_dev *dev)
 541 {
 542         /* No owner, become one */
 543         if (dev->use_worker) {
 544                 dev->mm = get_task_mm(current);
 545         } else {
 546                 /* vDPA device does not use worker thead, so there's
 547                  * no need to hold the address space for mm. This help
 548                  * to avoid deadlock in the case of mmap() which may
 549                  * held the refcnt of the file and depends on release
 550                  * method to remove vma.
 551                  */
 552                 dev->mm = current->mm;
 553                 mmgrab(dev->mm);
 554         }
 555 }
 556
 557 static void vhost_detach_mm(struct vhost_dev *dev)
 558 {
 559         if (!dev->mm)
 560                 return;
 561
 562         if (dev->use_worker)
 563                 mmput(dev->mm);
 564         else
 565                 mmdrop(dev->mm);
 566
 567         dev->mm = NULL;
 568 }
 569
 570 /* Caller should have device mutex */
 571 long vhost_dev_set_owner(struct vhost_dev *dev)
 572 {
 573         struct task_struct *worker;
 574         int err;
 575
 576         /* Is there an owner already? */
 577         if (vhost_dev_has_owner(dev)) {
 578                 err = -EBUSY;
 579                 goto err_mm;
 580         }
 581
 582         vhost_attach_mm(dev);
 583
 584         dev->kcov_handle = kcov_common_handle();
 585         if (dev->use_worker) {
 586                 worker = kthread_create(vhost_worker, dev,
 587                                         "vhost-%d", current->pid);
 588                 if (IS_ERR(worker)) {
 589                         err = PTR_ERR(worker);
 590                         goto err_worker;
 591                 }
 592
 593                 dev->worker = worker;
 594                 wake_up_process(worker); /* avoid contributing to loadavg */
 595
 596                 err = vhost_attach_cgroups(dev);
 597                 if (err)
 598                         goto err_cgroup;
 599         }
 600
 601         err = vhost_dev_alloc_iovecs(dev);
 602         if (err)
 603                 goto err_cgroup;
 604
 605         return 0;
 606 err_cgroup:
 607         if (dev->worker) {
 608                 kthread_stop(dev->worker);
 609                 dev->worker = NULL;
 610         }
 611 err_worker:
 612         vhost_detach_mm(dev);
 613         dev->kcov_handle = 0;
 614 err_mm:
 615         return err;
 616 }
 617 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 618
 619 static struct vhost_iotlb *iotlb_alloc(void)
 620 {
 621         return vhost_iotlb_alloc(max_iotlb_entries,
 622                                  VHOST_IOTLB_FLAG_RETIRE);
 623 }
 624
 625 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 626 {
 627         return iotlb_alloc();
 628 }
 629 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 630
 631 /* Caller should have device mutex */
 632 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 633 {
 634         int i;
 635
 636         vhost_dev_cleanup(dev);
 637
 638         dev->umem = umem;
 639         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 640          * VQs aren't running.
 641          */
 642         for (i = 0; i < dev->nvqs; ++i)
 643                 dev->vqs[i]->umem = umem;
 644 }
 645 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 646
 647 void vhost_dev_stop(struct vhost_dev *dev)
 648 {
 649         int i;
 650
 651         for (i = 0; i < dev->nvqs; ++i) {
 652                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 653                         vhost_poll_stop(&dev->vqs[i]->poll);
 654                         vhost_poll_flush(&dev->vqs[i]->poll);
 655                 }
 656         }
 657 }
 658 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 659
 660 static void vhost_clear_msg(struct vhost_dev *dev)
 661 {
 662         struct vhost_msg_node *node, *n;
 663
 664         spin_lock(&dev->iotlb_lock);
 665
 666         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 667                 list_del(&node->node);
 668                 kfree(node);
 669         }
 670
 671         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 672                 list_del(&node->node);
 673                 kfree(node);
 674         }
 675
 676         spin_unlock(&dev->iotlb_lock);
 677 }
 678
 679 void vhost_dev_cleanup(struct vhost_dev *dev)
 680 {
 681         int i;
 682
 683         for (i = 0; i < dev->nvqs; ++i) {
 684                 if (dev->vqs[i]->error_ctx)
 685                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 686                 if (dev->vqs[i]->kick)
 687                         fput(dev->vqs[i]->kick);
 688                 if (dev->vqs[i]->call_ctx)
 689                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 690                 vhost_vq_reset(dev, dev->vqs[i]);
 691         }
 692         vhost_dev_free_iovecs(dev);
 693         if (dev->log_ctx)
 694                 eventfd_ctx_put(dev->log_ctx);
 695         dev->log_ctx = NULL;
 696         /* No one will access memory at this point */
 697         vhost_iotlb_free(dev->umem);
 698         dev->umem = NULL;
 699         vhost_iotlb_free(dev->iotlb);
 700         dev->iotlb = NULL;
 701         vhost_clear_msg(dev);
 702         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 703         WARN_ON(!llist_empty(&dev->work_list));
 704         if (dev->worker) {
 705                 kthread_stop(dev->worker);
 706                 dev->worker = NULL;
 707                 dev->kcov_handle = 0;
 708         }
 709         vhost_detach_mm(dev);
 710 }
 711 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 712
 713 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 714 {
 715         u64 a = addr / VHOST_PAGE_SIZE / 8;
 716
 717         /* Make sure 64 bit math will not overflow. */
 718         if (a > ULONG_MAX - (unsigned long)log_base ||
 719             a + (unsigned long)log_base > ULONG_MAX)
 720                 return false;
 721
 722         return access_ok(log_base + a,
 723                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 724 }
 725
 726 static bool vhost_overflow(u64 uaddr, u64 size)
 727 {
 728         /* Make sure 64 bit math will not overflow. */
 729         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 730 }
 731
 732 /* Caller should have vq mutex and device mutex. */
 733 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
 734                                 int log_all)
 735 {
 736         struct vhost_iotlb_map *map;
 737
 738         if (!umem)
 739                 return false;
 740
 741         list_for_each_entry(map, &umem->list, link) {
 742                 unsigned long a = map->addr;
 743
 744                 if (vhost_overflow(map->addr, map->size))
 745                         return false;
 746
 747
 748                 if (!access_ok((void __user *)a, map->size))
 749                         return false;
 750                 else if (log_all && !log_access_ok(log_base,
 751                                                    map->start,
 752                                                    map->size))
 753                         return false;
 754         }
 755         return true;
 756 }
 757
 758 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 759                                                u64 addr, unsigned int size,
 760                                                int type)
 761 {
 762         const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
 763
 764         if (!map)
 765                 return NULL;
 766
 767         return (void __user *)(uintptr_t)(map->addr + addr - map->start);
 768 }
 769
 770 /* Can we switch to this memory table? */
 771 /* Caller should have device mutex but not vq mutex */
 772 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
 773                              int log_all)
 774 {
 775         int i;
 776
 777         for (i = 0; i < d->nvqs; ++i) {
 778                 bool ok;
 779                 bool log;
 780
 781                 mutex_lock(&d->vqs[i]->mutex);
 782                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 783                 /* If ring is inactive, will check when it's enabled. */
 784                 if (d->vqs[i]->private_data)
 785                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 786                                                  umem, log);
 787                 else
 788                         ok = true;
 789                 mutex_unlock(&d->vqs[i]->mutex);
 790                 if (!ok)
 791                         return false;
 792         }
 793         return true;
 794 }
 795
 796 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 797                           struct iovec iov[], int iov_size, int access);
 798
 799 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 800                               const void *from, unsigned size)
 801 {
 802         int ret;
 803
 804         if (!vq->iotlb)
 805                 return __copy_to_user(to, from, size);
 806         else {
 807                 /* This function should be called after iotlb
 808                  * prefetch, which means we're sure that all vq
 809                  * could be access through iotlb. So -EAGAIN should
 810                  * not happen in this case.
 811                  */
 812                 struct iov_iter t;
 813                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 814                                      (u64)(uintptr_t)to, size,
 815                                      VHOST_ADDR_USED);
 816
 817                 if (uaddr)
 818                         return __copy_to_user(uaddr, from, size);
 819
 820                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 821                                      ARRAY_SIZE(vq->iotlb_iov),
 822                                      VHOST_ACCESS_WO);
 823                 if (ret < 0)
 824                         goto out;
 825                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 826                 ret = copy_to_iter(from, size, &t);
 827                 if (ret == size)
 828                         ret = 0;
 829         }
 830 out:
 831         return ret;
 832 }
 833
 834 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 835                                 void __user *from, unsigned size)
 836 {
 837         int ret;
 838
 839         if (!vq->iotlb)
 840                 return __copy_from_user(to, from, size);
 841         else {
 842                 /* This function should be called after iotlb
 843                  * prefetch, which means we're sure that vq
 844                  * could be access through iotlb. So -EAGAIN should
 845                  * not happen in this case.
 846                  */
 847                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 848                                      (u64)(uintptr_t)from, size,
 849                                      VHOST_ADDR_DESC);
 850                 struct iov_iter f;
 851
 852                 if (uaddr)
 853                         return __copy_from_user(to, uaddr, size);
 854
 855                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 856                                      ARRAY_SIZE(vq->iotlb_iov),
 857                                      VHOST_ACCESS_RO);
 858                 if (ret < 0) {
 859                         vq_err(vq, "IOTLB translation failure: uaddr "
 860                                "%p size 0x%llx\n", from,
 861                                (unsigned long long) size);
 862                         goto out;
 863                 }
 864                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 865                 ret = copy_from_iter(to, size, &f);
 866                 if (ret == size)
 867                         ret = 0;
 868         }
 869
 870 out:
 871         return ret;
 872 }
 873
 874 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 875                                           void __user *addr, unsigned int size,
 876                                           int type)
 877 {
 878         int ret;
 879
 880         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 881                              ARRAY_SIZE(vq->iotlb_iov),
 882                              VHOST_ACCESS_RO);
 883         if (ret < 0) {
 884                 vq_err(vq, "IOTLB translation failure: uaddr "
 885                         "%p size 0x%llx\n", addr,
 886                         (unsigned long long) size);
 887                 return NULL;
 888         }
 889
 890         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 891                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 892                         "%p size 0x%llx\n", addr,
 893                         (unsigned long long) size);
 894                 return NULL;
 895         }
 896
 897         return vq->iotlb_iov[0].iov_base;
 898 }
 899
 900 /* This function should be called after iotlb
 901  * prefetch, which means we're sure that vq
 902  * could be access through iotlb. So -EAGAIN should
 903  * not happen in this case.
 904  */
 905 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 906                                             void __user *addr, unsigned int size,
 907                                             int type)
 908 {
 909         void __user *uaddr = vhost_vq_meta_fetch(vq,
 910                              (u64)(uintptr_t)addr, size, type);
 911         if (uaddr)
 912                 return uaddr;
 913
 914         return __vhost_get_user_slow(vq, addr, size, type);
 915 }
 916
 917 #define vhost_put_user(vq, x, ptr)              \
 918 ({ \
 919         int ret; \
 920         if (!vq->iotlb) { \
 921                 ret = __put_user(x, ptr); \
 922         } else { \
 923                 __typeof__(ptr) to = \
 924                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 925                                           sizeof(*ptr), VHOST_ADDR_USED); \
 926                 if (to != NULL) \
 927                         ret = __put_user(x, to); \
 928                 else \
 929                         ret = -EFAULT;  \
 930         } \
 931         ret; \
 932 })
 933
 934 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 935 {
 936         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 937                               vhost_avail_event(vq));
 938 }
 939
 940 static inline int vhost_put_used(struct vhost_virtqueue *vq,
 941                                  struct vring_used_elem *head, int idx,
 942                                  int count)
 943 {
 944         return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 945                                   count * sizeof(*head));
 946 }
 947
 948 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 949
 950 {
 951         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 952                               &vq->used->flags);
 953 }
 954
 955 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 956
 957 {
 958         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 959                               &vq->used->idx);
 960 }
 961
 962 #define vhost_get_user(vq, x, ptr, type)                \
 963 ({ \
 964         int ret; \
 965         if (!vq->iotlb) { \
 966                 ret = __get_user(x, ptr); \
 967         } else { \
 968                 __typeof__(ptr) from = \
 969                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 970                                                            sizeof(*ptr), \
 971                                                            type); \
 972                 if (from != NULL) \
 973                         ret = __get_user(x, from); \
 974                 else \
 975                         ret = -EFAULT; \
 976         } \
 977         ret; \
 978 })
 979
 980 #define vhost_get_avail(vq, x, ptr) \
 981         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 982
 983 #define vhost_get_used(vq, x, ptr) \
 984         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 985
 986 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 987 {
 988         int i = 0;
 989         for (i = 0; i < d->nvqs; ++i)
 990                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 991 }
 992
 993 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 994 {
 995         int i = 0;
 996         for (i = 0; i < d->nvqs; ++i)
 997                 mutex_unlock(&d->vqs[i]->mutex);
 998 }
 999
1000 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1001                                       __virtio16 *idx)
1002 {
1003         return vhost_get_avail(vq, *idx, &vq->avail->idx);
1004 }
1005
1006 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1007                                        __virtio16 *head, int idx)
1008 {
1009         return vhost_get_avail(vq, *head,
1010                                &vq->avail->ring[idx & (vq->num - 1)]);
1011 }
1012
1013 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1014                                         __virtio16 *flags)
1015 {
1016         return vhost_get_avail(vq, *flags, &vq->avail->flags);
1017 }
1018
1019 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1020                                        __virtio16 *event)
1021 {
1022         return vhost_get_avail(vq, *event, vhost_used_event(vq));
1023 }
1024
1025 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1026                                      __virtio16 *idx)
1027 {
1028         return vhost_get_used(vq, *idx, &vq->used->idx);
1029 }
1030
1031 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1032                                  struct vring_desc *desc, int idx)
1033 {
1034         return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1035 }
1036
1037 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1038                                   struct vhost_iotlb_msg *msg)
1039 {
1040         struct vhost_msg_node *node, *n;
1041
1042         spin_lock(&d->iotlb_lock);
1043
1044         list_for_each_entry_safe(node, n, &d->pending_list, node) {
1045                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1046                 if (msg->iova <= vq_msg->iova &&
1047                     msg->iova + msg->size - 1 >= vq_msg->iova &&
1048                     vq_msg->type == VHOST_IOTLB_MISS) {
1049                         vhost_poll_queue(&node->vq->poll);
1050                         list_del(&node->node);
1051                         kfree(node);
1052                 }
1053         }
1054
1055         spin_unlock(&d->iotlb_lock);
1056 }
1057
1058 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1059 {
1060         unsigned long a = uaddr;
1061
1062         /* Make sure 64 bit math will not overflow. */
1063         if (vhost_overflow(uaddr, size))
1064                 return false;
1065
1066         if ((access & VHOST_ACCESS_RO) &&
1067             !access_ok((void __user *)a, size))
1068                 return false;
1069         if ((access & VHOST_ACCESS_WO) &&
1070             !access_ok((void __user *)a, size))
1071                 return false;
1072         return true;
1073 }
1074
1075 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1076                                    struct vhost_iotlb_msg *msg)
1077 {
1078         int ret = 0;
1079
1080         mutex_lock(&dev->mutex);
1081         vhost_dev_lock_vqs(dev);
1082         switch (msg->type) {
1083         case VHOST_IOTLB_UPDATE:
1084                 if (!dev->iotlb) {
1085                         ret = -EFAULT;
1086                         break;
1087                 }
1088                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1089                         ret = -EFAULT;
1090                         break;
1091                 }
1092                 vhost_vq_meta_reset(dev);
1093                 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1094                                           msg->iova + msg->size - 1,
1095                                           msg->uaddr, msg->perm)) {
1096                         ret = -ENOMEM;
1097                         break;
1098                 }
1099                 vhost_iotlb_notify_vq(dev, msg);
1100                 break;
1101         case VHOST_IOTLB_INVALIDATE:
1102                 if (!dev->iotlb) {
1103                         ret = -EFAULT;
1104                         break;
1105                 }
1106                 vhost_vq_meta_reset(dev);
1107                 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1108                                       msg->iova + msg->size - 1);
1109                 break;
1110         default:
1111                 ret = -EINVAL;
1112                 break;
1113         }
1114
1115         vhost_dev_unlock_vqs(dev);
1116         mutex_unlock(&dev->mutex);
1117
1118         return ret;
1119 }
1120 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1121                              struct iov_iter *from)
1122 {
1123         struct vhost_iotlb_msg msg;
1124         size_t offset;
1125         int type, ret;
1126
1127         ret = copy_from_iter(&type, sizeof(type), from);
1128         if (ret != sizeof(type)) {
1129                 ret = -EINVAL;
1130                 goto done;
1131         }
1132
1133         switch (type) {
1134         case VHOST_IOTLB_MSG:
1135                 /* There maybe a hole after type for V1 message type,
1136                  * so skip it here.
1137                  */
1138                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1139                 break;
1140         case VHOST_IOTLB_MSG_V2:
1141                 offset = sizeof(__u32);
1142                 break;
1143         default:
1144                 ret = -EINVAL;
1145                 goto done;
1146         }
1147
1148         iov_iter_advance(from, offset);
1149         ret = copy_from_iter(&msg, sizeof(msg), from);
1150         if (ret != sizeof(msg)) {
1151                 ret = -EINVAL;
1152                 goto done;
1153         }
1154
1155         if (dev->msg_handler)
1156                 ret = dev->msg_handler(dev, &msg);
1157         else
1158                 ret = vhost_process_iotlb_msg(dev, &msg);
1159         if (ret) {
1160                 ret = -EFAULT;
1161                 goto done;
1162         }
1163
1164         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1165               sizeof(struct vhost_msg_v2);
1166 done:
1167         return ret;
1168 }
1169 EXPORT_SYMBOL(vhost_chr_write_iter);
1170
1171 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1172                             poll_table *wait)
1173 {
1174         __poll_t mask = 0;
1175
1176         poll_wait(file, &dev->wait, wait);
1177
1178         if (!list_empty(&dev->read_list))
1179                 mask |= EPOLLIN | EPOLLRDNORM;
1180
1181         return mask;
1182 }
1183 EXPORT_SYMBOL(vhost_chr_poll);
1184
1185 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1186                             int noblock)
1187 {
1188         DEFINE_WAIT(wait);
1189         struct vhost_msg_node *node;
1190         ssize_t ret = 0;
1191         unsigned size = sizeof(struct vhost_msg);
1192
1193         if (iov_iter_count(to) < size)
1194                 return 0;
1195
1196         while (1) {
1197                 if (!noblock)
1198                         prepare_to_wait(&dev->wait, &wait,
1199                                         TASK_INTERRUPTIBLE);
1200
1201                 node = vhost_dequeue_msg(dev, &dev->read_list);
1202                 if (node)
1203                         break;
1204                 if (noblock) {
1205                         ret = -EAGAIN;
1206                         break;
1207                 }
1208                 if (signal_pending(current)) {
1209                         ret = -ERESTARTSYS;
1210                         break;
1211                 }
1212                 if (!dev->iotlb) {
1213                         ret = -EBADFD;
1214                         break;
1215                 }
1216
1217                 schedule();
1218         }
1219
1220         if (!noblock)
1221                 finish_wait(&dev->wait, &wait);
1222
1223         if (node) {
1224                 struct vhost_iotlb_msg *msg;
1225                 void *start = &node->msg;
1226
1227                 switch (node->msg.type) {
1228                 case VHOST_IOTLB_MSG:
1229                         size = sizeof(node->msg);
1230                         msg = &node->msg.iotlb;
1231                         break;
1232                 case VHOST_IOTLB_MSG_V2:
1233                         size = sizeof(node->msg_v2);
1234                         msg = &node->msg_v2.iotlb;
1235                         break;
1236                 default:
1237                         BUG();
1238                         break;
1239                 }
1240
1241                 ret = copy_to_iter(start, size, to);
1242                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1243                         kfree(node);
1244                         return ret;
1245                 }
1246                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1247         }
1248
1249         return ret;
1250 }
1251 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1252
1253 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1254 {
1255         struct vhost_dev *dev = vq->dev;
1256         struct vhost_msg_node *node;
1257         struct vhost_iotlb_msg *msg;
1258         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1259
1260         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1261         if (!node)
1262                 return -ENOMEM;
1263
1264         if (v2) {
1265                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1266                 msg = &node->msg_v2.iotlb;
1267         } else {
1268                 msg = &node->msg.iotlb;
1269         }
1270
1271         msg->type = VHOST_IOTLB_MISS;
1272         msg->iova = iova;
1273         msg->perm = access;
1274
1275         vhost_enqueue_msg(dev, &dev->read_list, node);
1276
1277         return 0;
1278 }
1279
1280 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1281                          vring_desc_t __user *desc,
1282                          vring_avail_t __user *avail,
1283                          vring_used_t __user *used)
1284
1285 {
1286         return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1287                access_ok(avail, vhost_get_avail_size(vq, num)) &&
1288                access_ok(used, vhost_get_used_size(vq, num));
1289 }
1290
1291 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1292                                  const struct vhost_iotlb_map *map,
1293                                  int type)
1294 {
1295         int access = (type == VHOST_ADDR_USED) ?
1296                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1297
1298         if (likely(map->perm & access))
1299                 vq->meta_iotlb[type] = map;
1300 }
1301
1302 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1303                             int access, u64 addr, u64 len, int type)
1304 {
1305         const struct vhost_iotlb_map *map;
1306         struct vhost_iotlb *umem = vq->iotlb;
1307         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1308
1309         if (vhost_vq_meta_fetch(vq, addr, len, type))
1310                 return true;
1311
1312         while (len > s) {
1313                 map = vhost_iotlb_itree_first(umem, addr, last);
1314                 if (map == NULL || map->start > addr) {
1315                         vhost_iotlb_miss(vq, addr, access);
1316                         return false;
1317                 } else if (!(map->perm & access)) {
1318                         /* Report the possible access violation by
1319                          * request another translation from userspace.
1320                          */
1321                         return false;
1322                 }
1323
1324                 size = map->size - addr + map->start;
1325
1326                 if (orig_addr == addr && size >= len)
1327                         vhost_vq_meta_update(vq, map, type);
1328
1329                 s += size;
1330                 addr += size;
1331         }
1332
1333         return true;
1334 }
1335
1336 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1337 {
1338         unsigned int num = vq->num;
1339
1340         if (!vq->iotlb)
1341                 return 1;
1342
1343         return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1344                                vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1345                iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1346                                vhost_get_avail_size(vq, num),
1347                                VHOST_ADDR_AVAIL) &&
1348                iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1349                                vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1350 }
1351 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1352
1353 /* Can we log writes? */
1354 /* Caller should have device mutex but not vq mutex */
1355 bool vhost_log_access_ok(struct vhost_dev *dev)
1356 {
1357         return memory_access_ok(dev, dev->umem, 1);
1358 }
1359 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1360
1361 /* Verify access for write logging. */
1362 /* Caller should have vq mutex and device mutex */
1363 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1364                              void __user *log_base)
1365 {
1366         return vq_memory_access_ok(log_base, vq->umem,
1367                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1368                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1369                                   vhost_get_used_size(vq, vq->num)));
1370 }
1371
1372 /* Can we start vq? */
1373 /* Caller should have vq mutex and device mutex */
1374 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1375 {
1376         if (!vq_log_access_ok(vq, vq->log_base))
1377                 return false;
1378
1379         /* Access validation occurs at prefetch time with IOTLB */
1380         if (vq->iotlb)
1381                 return true;
1382
1383         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1384 }
1385 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1386
1387 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1388 {
1389         struct vhost_memory mem, *newmem;
1390         struct vhost_memory_region *region;
1391         struct vhost_iotlb *newumem, *oldumem;
1392         unsigned long size = offsetof(struct vhost_memory, regions);
1393         int i;
1394
1395         if (copy_from_user(&mem, m, size))
1396                 return -EFAULT;
1397         if (mem.padding)
1398                 return -EOPNOTSUPP;
1399         if (mem.nregions > max_mem_regions)
1400                 return -E2BIG;
1401         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1402                         GFP_KERNEL);
1403         if (!newmem)
1404                 return -ENOMEM;
1405
1406         memcpy(newmem, &mem, size);
1407         if (copy_from_user(newmem->regions, m->regions,
1408                            mem.nregions * sizeof *m->regions)) {
1409                 kvfree(newmem);
1410                 return -EFAULT;
1411         }
1412
1413         newumem = iotlb_alloc();
1414         if (!newumem) {
1415                 kvfree(newmem);
1416                 return -ENOMEM;
1417         }
1418
1419         for (region = newmem->regions;
1420              region < newmem->regions + mem.nregions;
1421              region++) {
1422                 if (vhost_iotlb_add_range(newumem,
1423                                           region->guest_phys_addr,
1424                                           region->guest_phys_addr +
1425                                           region->memory_size - 1,
1426                                           region->userspace_addr,
1427                                           VHOST_MAP_RW))
1428                         goto err;
1429         }
1430
1431         if (!memory_access_ok(d, newumem, 0))
1432                 goto err;
1433
1434         oldumem = d->umem;
1435         d->umem = newumem;
1436
1437         /* All memory accesses are done under some VQ mutex. */
1438         for (i = 0; i < d->nvqs; ++i) {
1439                 mutex_lock(&d->vqs[i]->mutex);
1440                 d->vqs[i]->umem = newumem;
1441                 mutex_unlock(&d->vqs[i]->mutex);
1442         }
1443
1444         kvfree(newmem);
1445         vhost_iotlb_free(oldumem);
1446         return 0;
1447
1448 err:
1449         vhost_iotlb_free(newumem);
1450         kvfree(newmem);
1451         return -EFAULT;
1452 }
1453
1454 static long vhost_vring_set_num(struct vhost_dev *d,
1455                                 struct vhost_virtqueue *vq,
1456                                 void __user *argp)
1457 {
1458         struct vhost_vring_state s;
1459
1460         /* Resizing ring with an active backend?
1461          * You don't want to do that. */
1462         if (vq->private_data)
1463                 return -EBUSY;
1464
1465         if (copy_from_user(&s, argp, sizeof s))
1466                 return -EFAULT;
1467
1468         if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1469                 return -EINVAL;
1470         vq->num = s.num;
1471
1472         return 0;
1473 }
1474
1475 static long vhost_vring_set_addr(struct vhost_dev *d,
1476                                  struct vhost_virtqueue *vq,
1477                                  void __user *argp)
1478 {
1479         struct vhost_vring_addr a;
1480
1481         if (copy_from_user(&a, argp, sizeof a))
1482                 return -EFAULT;
1483         if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1484                 return -EOPNOTSUPP;
1485
1486         /* For 32bit, verify that the top 32bits of the user
1487            data are set to zero. */
1488         if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1489             (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1490             (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1491                 return -EFAULT;
1492
1493         /* Make sure it's safe to cast pointers to vring types. */
1494         BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1495         BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1496         if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1497             (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1498             (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1499                 return -EINVAL;
1500
1501         /* We only verify access here if backend is configured.
1502          * If it is not, we don't as size might not have been setup.
1503          * We will verify when backend is configured. */
1504         if (vq->private_data) {
1505                 if (!vq_access_ok(vq, vq->num,
1506                         (void __user *)(unsigned long)a.desc_user_addr,
1507                         (void __user *)(unsigned long)a.avail_user_addr,
1508                         (void __user *)(unsigned long)a.used_user_addr))
1509                         return -EINVAL;
1510
1511                 /* Also validate log access for used ring if enabled. */
1512                 if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1513                         !log_access_ok(vq->log_base, a.log_guest_addr,
1514                                 sizeof *vq->used +
1515                                 vq->num * sizeof *vq->used->ring))
1516                         return -EINVAL;
1517         }
1518
1519         vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1520         vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1521         vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1522         vq->log_addr = a.log_guest_addr;
1523         vq->used = (void __user *)(unsigned long)a.used_user_addr;
1524
1525         return 0;
1526 }
1527
1528 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1529                                      struct vhost_virtqueue *vq,
1530                                      unsigned int ioctl,
1531                                      void __user *argp)
1532 {
1533         long r;
1534
1535         mutex_lock(&vq->mutex);
1536
1537         switch (ioctl) {
1538         case VHOST_SET_VRING_NUM:
1539                 r = vhost_vring_set_num(d, vq, argp);
1540                 break;
1541         case VHOST_SET_VRING_ADDR:
1542                 r = vhost_vring_set_addr(d, vq, argp);
1543                 break;
1544         default:
1545                 BUG();
1546         }
1547
1548         mutex_unlock(&vq->mutex);
1549
1550         return r;
1551 }
1552 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1553 {
1554         struct file *eventfp, *filep = NULL;
1555         bool pollstart = false, pollstop = false;
1556         struct eventfd_ctx *ctx = NULL;
1557         u32 __user *idxp = argp;
1558         struct vhost_virtqueue *vq;
1559         struct vhost_vring_state s;
1560         struct vhost_vring_file f;
1561         u32 idx;
1562         long r;
1563
1564         r = get_user(idx, idxp);
1565         if (r < 0)
1566                 return r;
1567         if (idx >= d->nvqs)
1568                 return -ENOBUFS;
1569
1570         idx = array_index_nospec(idx, d->nvqs);
1571         vq = d->vqs[idx];
1572
1573         if (ioctl == VHOST_SET_VRING_NUM ||
1574             ioctl == VHOST_SET_VRING_ADDR) {
1575                 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1576         }
1577
1578         mutex_lock(&vq->mutex);
1579
1580         switch (ioctl) {
1581         case VHOST_SET_VRING_BASE:
1582                 /* Moving base with an active backend?
1583                  * You don't want to do that. */
1584                 if (vq->private_data) {
1585                         r = -EBUSY;
1586                         break;
1587                 }
1588                 if (copy_from_user(&s, argp, sizeof s)) {
1589                         r = -EFAULT;
1590                         break;
1591                 }
1592                 if (s.num > 0xffff) {
1593                         r = -EINVAL;
1594                         break;
1595                 }
1596                 vq->last_avail_idx = s.num;
1597                 /* Forget the cached index value. */
1598                 vq->avail_idx = vq->last_avail_idx;
1599                 break;
1600         case VHOST_GET_VRING_BASE:
1601                 s.index = idx;
1602                 s.num = vq->last_avail_idx;
1603                 if (copy_to_user(argp, &s, sizeof s))
1604                         r = -EFAULT;
1605                 break;
1606         case VHOST_SET_VRING_KICK:
1607                 if (copy_from_user(&f, argp, sizeof f)) {
1608                         r = -EFAULT;
1609                         break;
1610                 }
1611                 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1612                 if (IS_ERR(eventfp)) {
1613                         r = PTR_ERR(eventfp);
1614                         break;
1615                 }
1616                 if (eventfp != vq->kick) {
1617                         pollstop = (filep = vq->kick) != NULL;
1618                         pollstart = (vq->kick = eventfp) != NULL;
1619                 } else
1620                         filep = eventfp;
1621                 break;
1622         case VHOST_SET_VRING_CALL:
1623                 if (copy_from_user(&f, argp, sizeof f)) {
1624                         r = -EFAULT;
1625                         break;
1626                 }
1627                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1628                 if (IS_ERR(ctx)) {
1629                         r = PTR_ERR(ctx);
1630                         break;
1631                 }
1632                 swap(ctx, vq->call_ctx);
1633                 break;
1634         case VHOST_SET_VRING_ERR:
1635                 if (copy_from_user(&f, argp, sizeof f)) {
1636                         r = -EFAULT;
1637                         break;
1638                 }
1639                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1640                 if (IS_ERR(ctx)) {
1641                         r = PTR_ERR(ctx);
1642                         break;
1643                 }
1644                 swap(ctx, vq->error_ctx);
1645                 break;
1646         case VHOST_SET_VRING_ENDIAN:
1647                 r = vhost_set_vring_endian(vq, argp);
1648                 break;
1649         case VHOST_GET_VRING_ENDIAN:
1650                 r = vhost_get_vring_endian(vq, idx, argp);
1651                 break;
1652         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1653                 if (copy_from_user(&s, argp, sizeof(s))) {
1654                         r = -EFAULT;
1655                         break;
1656                 }
1657                 vq->busyloop_timeout = s.num;
1658                 break;
1659         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1660                 s.index = idx;
1661                 s.num = vq->busyloop_timeout;
1662                 if (copy_to_user(argp, &s, sizeof(s)))
1663                         r = -EFAULT;
1664                 break;
1665         default:
1666                 r = -ENOIOCTLCMD;
1667         }
1668
1669         if (pollstop && vq->handle_kick)
1670                 vhost_poll_stop(&vq->poll);
1671
1672         if (!IS_ERR_OR_NULL(ctx))
1673                 eventfd_ctx_put(ctx);
1674         if (filep)
1675                 fput(filep);
1676
1677         if (pollstart && vq->handle_kick)
1678                 r = vhost_poll_start(&vq->poll, vq->kick);
1679
1680         mutex_unlock(&vq->mutex);
1681
1682         if (pollstop && vq->handle_kick)
1683                 vhost_poll_flush(&vq->poll);
1684         return r;
1685 }
1686 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1687
1688 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1689 {
1690         struct vhost_iotlb *niotlb, *oiotlb;
1691         int i;
1692
1693         niotlb = iotlb_alloc();
1694         if (!niotlb)
1695                 return -ENOMEM;
1696
1697         oiotlb = d->iotlb;
1698         d->iotlb = niotlb;
1699
1700         for (i = 0; i < d->nvqs; ++i) {
1701                 struct vhost_virtqueue *vq = d->vqs[i];
1702
1703                 mutex_lock(&vq->mutex);
1704                 vq->iotlb = niotlb;
1705                 __vhost_vq_meta_reset(vq);
1706                 mutex_unlock(&vq->mutex);
1707         }
1708
1709         vhost_iotlb_free(oiotlb);
1710
1711         return 0;
1712 }
1713 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1714
1715 /* Caller must have device mutex */
1716 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1717 {
1718         struct eventfd_ctx *ctx;
1719         u64 p;
1720         long r;
1721         int i, fd;
1722
1723         /* If you are not the owner, you can become one */
1724         if (ioctl == VHOST_SET_OWNER) {
1725                 r = vhost_dev_set_owner(d);
1726                 goto done;
1727         }
1728
1729         /* You must be the owner to do anything else */
1730         r = vhost_dev_check_owner(d);
1731         if (r)
1732                 goto done;
1733
1734         switch (ioctl) {
1735         case VHOST_SET_MEM_TABLE:
1736                 r = vhost_set_memory(d, argp);
1737                 break;
1738         case VHOST_SET_LOG_BASE:
1739                 if (copy_from_user(&p, argp, sizeof p)) {
1740                         r = -EFAULT;
1741                         break;
1742                 }
1743                 if ((u64)(unsigned long)p != p) {
1744                         r = -EFAULT;
1745                         break;
1746                 }
1747                 for (i = 0; i < d->nvqs; ++i) {
1748                         struct vhost_virtqueue *vq;
1749                         void __user *base = (void __user *)(unsigned long)p;
1750                         vq = d->vqs[i];
1751                         mutex_lock(&vq->mutex);
1752                         /* If ring is inactive, will check when it's enabled. */
1753                         if (vq->private_data && !vq_log_access_ok(vq, base))
1754                                 r = -EFAULT;
1755                         else
1756                                 vq->log_base = base;
1757                         mutex_unlock(&vq->mutex);
1758                 }
1759                 break;
1760         case VHOST_SET_LOG_FD:
1761                 r = get_user(fd, (int __user *)argp);
1762                 if (r < 0)
1763                         break;
1764                 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1765                 if (IS_ERR(ctx)) {
1766                         r = PTR_ERR(ctx);
1767                         break;
1768                 }
1769                 swap(ctx, d->log_ctx);
1770                 for (i = 0; i < d->nvqs; ++i) {
1771                         mutex_lock(&d->vqs[i]->mutex);
1772                         d->vqs[i]->log_ctx = d->log_ctx;
1773                         mutex_unlock(&d->vqs[i]->mutex);
1774                 }
1775                 if (ctx)
1776                         eventfd_ctx_put(ctx);
1777                 break;
1778         default:
1779                 r = -ENOIOCTLCMD;
1780                 break;
1781         }
1782 done:
1783         return r;
1784 }
1785 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1786
1787 /* TODO: This is really inefficient.  We need something like get_user()
1788  * (instruction directly accesses the data, with an exception table entry
1789  * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1790  */
1791 static int set_bit_to_user(int nr, void __user *addr)
1792 {
1793         unsigned long log = (unsigned long)addr;
1794         struct page *page;
1795         void *base;
1796         int bit = nr + (log % PAGE_SIZE) * 8;
1797         int r;
1798
1799         r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1800         if (r < 0)
1801                 return r;
1802         BUG_ON(r != 1);
1803         base = kmap_atomic(page);
1804         set_bit(bit, base);
1805         kunmap_atomic(base);
1806         unpin_user_pages_dirty_lock(&page, 1, true);
1807         return 0;
1808 }
1809
1810 static int log_write(void __user *log_base,
1811                      u64 write_address, u64 write_length)
1812 {
1813         u64 write_page = write_address / VHOST_PAGE_SIZE;
1814         int r;
1815
1816         if (!write_length)
1817                 return 0;
1818         write_length += write_address % VHOST_PAGE_SIZE;
1819         for (;;) {
1820                 u64 base = (u64)(unsigned long)log_base;
1821                 u64 log = base + write_page / 8;
1822                 int bit = write_page % 8;
1823                 if ((u64)(unsigned long)log != log)
1824                         return -EFAULT;
1825                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1826                 if (r < 0)
1827                         return r;
1828                 if (write_length <= VHOST_PAGE_SIZE)
1829                         break;
1830                 write_length -= VHOST_PAGE_SIZE;
1831                 write_page += 1;
1832         }
1833         return r;
1834 }
1835
1836 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1837 {
1838         struct vhost_iotlb *umem = vq->umem;
1839         struct vhost_iotlb_map *u;
1840         u64 start, end, l, min;
1841         int r;
1842         bool hit = false;
1843
1844         while (len) {
1845                 min = len;
1846                 /* More than one GPAs can be mapped into a single HVA. So
1847                  * iterate all possible umems here to be safe.
1848                  */
1849                 list_for_each_entry(u, &umem->list, link) {
1850                         if (u->addr > hva - 1 + len ||
1851                             u->addr - 1 + u->size < hva)
1852                                 continue;
1853                         start = max(u->addr, hva);
1854                         end = min(u->addr - 1 + u->size, hva - 1 + len);
1855                         l = end - start + 1;
1856                         r = log_write(vq->log_base,
1857                                       u->start + start - u->addr,
1858                                       l);
1859                         if (r < 0)
1860                                 return r;
1861                         hit = true;
1862                         min = min(l, min);
1863                 }
1864
1865                 if (!hit)
1866                         return -EFAULT;
1867
1868                 len -= min;
1869                 hva += min;
1870         }
1871
1872         return 0;
1873 }
1874
1875 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1876 {
1877         struct iovec iov[64];
1878         int i, ret;
1879
1880         if (!vq->iotlb)
1881                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1882
1883         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1884                              len, iov, 64, VHOST_ACCESS_WO);
1885         if (ret < 0)
1886                 return ret;
1887
1888         for (i = 0; i < ret; i++) {
1889                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1890                                     iov[i].iov_len);
1891                 if (ret)
1892                         return ret;
1893         }
1894
1895         return 0;
1896 }
1897
1898 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1899                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1900 {
1901         int i, r;
1902
1903         /* Make sure data written is seen before log. */
1904         smp_wmb();
1905
1906         if (vq->iotlb) {
1907                 for (i = 0; i < count; i++) {
1908                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1909                                           iov[i].iov_len);
1910                         if (r < 0)
1911                                 return r;
1912                 }
1913                 return 0;
1914         }
1915
1916         for (i = 0; i < log_num; ++i) {
1917                 u64 l = min(log[i].len, len);
1918                 r = log_write(vq->log_base, log[i].addr, l);
1919                 if (r < 0)
1920                         return r;
1921                 len -= l;
1922                 if (!len) {
1923                         if (vq->log_ctx)
1924                                 eventfd_signal(vq->log_ctx, 1);
1925                         return 0;
1926                 }
1927         }
1928         /* Length written exceeds what we have stored. This is a bug. */
1929         BUG();
1930         return 0;
1931 }
1932 EXPORT_SYMBOL_GPL(vhost_log_write);
1933
1934 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1935 {
1936         void __user *used;
1937         if (vhost_put_used_flags(vq))
1938                 return -EFAULT;
1939         if (unlikely(vq->log_used)) {
1940                 /* Make sure the flag is seen before log. */
1941                 smp_wmb();
1942                 /* Log used flag write. */
1943                 used = &vq->used->flags;
1944                 log_used(vq, (used - (void __user *)vq->used),
1945                          sizeof vq->used->flags);
1946                 if (vq->log_ctx)
1947                         eventfd_signal(vq->log_ctx, 1);
1948         }
1949         return 0;
1950 }
1951
1952 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1953 {
1954         if (vhost_put_avail_event(vq))
1955                 return -EFAULT;
1956         if (unlikely(vq->log_used)) {
1957                 void __user *used;
1958                 /* Make sure the event is seen before log. */
1959                 smp_wmb();
1960                 /* Log avail event write */
1961                 used = vhost_avail_event(vq);
1962                 log_used(vq, (used - (void __user *)vq->used),
1963                          sizeof *vhost_avail_event(vq));
1964                 if (vq->log_ctx)
1965                         eventfd_signal(vq->log_ctx, 1);
1966         }
1967         return 0;
1968 }
1969
1970 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1971 {
1972         __virtio16 last_used_idx;
1973         int r;
1974         bool is_le = vq->is_le;
1975
1976         if (!vq->private_data)
1977                 return 0;
1978
1979         vhost_init_is_le(vq);
1980
1981         r = vhost_update_used_flags(vq);
1982         if (r)
1983                 goto err;
1984         vq->signalled_used_valid = false;
1985         if (!vq->iotlb &&
1986             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
1987                 r = -EFAULT;
1988                 goto err;
1989         }
1990         r = vhost_get_used_idx(vq, &last_used_idx);
1991         if (r) {
1992                 vq_err(vq, "Can't access used idx at %p\n",
1993                        &vq->used->idx);
1994                 goto err;
1995         }
1996         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1997         return 0;
1998
1999 err:
2000         vq->is_le = is_le;
2001         return r;
2002 }
2003 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2004
2005 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2006                           struct iovec iov[], int iov_size, int access)
2007 {
2008         const struct vhost_iotlb_map *map;
2009         struct vhost_dev *dev = vq->dev;
2010         struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2011         struct iovec *_iov;
2012         u64 s = 0;
2013         int ret = 0;
2014
2015         while ((u64)len > s) {
2016                 u64 size;
2017                 if (unlikely(ret >= iov_size)) {
2018                         ret = -ENOBUFS;
2019                         break;
2020                 }
2021
2022                 map = vhost_iotlb_itree_first(umem, addr, addr + len - 1);
2023                 if (map == NULL || map->start > addr) {
2024                         if (umem != dev->iotlb) {
2025                                 ret = -EFAULT;
2026                                 break;
2027                         }
2028                         ret = -EAGAIN;
2029                         break;
2030                 } else if (!(map->perm & access)) {
2031                         ret = -EPERM;
2032                         break;
2033                 }
2034
2035                 _iov = iov + ret;
2036                 size = map->size - addr + map->start;
2037                 _iov->iov_len = min((u64)len - s, size);
2038                 _iov->iov_base = (void __user *)(unsigned long)
2039                                  (map->addr + addr - map->start);
2040                 s += size;
2041                 addr += size;
2042                 ++ret;
2043         }
2044
2045         if (ret == -EAGAIN)
2046                 vhost_iotlb_miss(vq, addr, access);
2047         return ret;
2048 }
2049
2050 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
2051  * function returns the next descriptor in the chain,
2052  * or -1U if we're at the end. */
2053 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2054 {
2055         unsigned int next;
2056
2057         /* If this descriptor says it doesn't chain, we're done. */
2058         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2059                 return -1U;
2060
2061         /* Check they're not leading us off end of descriptors. */
2062         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2063         return next;
2064 }
2065
2066 static int get_indirect(struct vhost_virtqueue *vq,
2067                         struct iovec iov[], unsigned int iov_size,
2068                         unsigned int *out_num, unsigned int *in_num,
2069                         struct vhost_log *log, unsigned int *log_num,
2070                         struct vring_desc *indirect)
2071 {
2072         struct vring_desc desc;
2073         unsigned int i = 0, count, found = 0;
2074         u32 len = vhost32_to_cpu(vq, indirect->len);
2075         struct iov_iter from;
2076         int ret, access;
2077
2078         /* Sanity check */
2079         if (unlikely(len % sizeof desc)) {
2080                 vq_err(vq, "Invalid length in indirect descriptor: "
2081                        "len 0x%llx not multiple of 0x%zx\n",
2082                        (unsigned long long)len,
2083                        sizeof desc);
2084                 return -EINVAL;
2085         }
2086
2087         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2088                              UIO_MAXIOV, VHOST_ACCESS_RO);
2089         if (unlikely(ret < 0)) {
2090                 if (ret != -EAGAIN)
2091                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2092                 return ret;
2093         }
2094         iov_iter_init(&from, READ, vq->indirect, ret, len);
2095
2096         /* We will use the result as an address to read from, so most
2097          * architectures only need a compiler barrier here. */
2098         read_barrier_depends();
2099
2100         count = len / sizeof desc;
2101         /* Buffers are chained via a 16 bit next field, so
2102          * we can have at most 2^16 of these. */
2103         if (unlikely(count > USHRT_MAX + 1)) {
2104                 vq_err(vq, "Indirect buffer length too big: %d\n",
2105                        indirect->len);
2106                 return -E2BIG;
2107         }
2108
2109         do {
2110                 unsigned iov_count = *in_num + *out_num;
2111                 if (unlikely(++found > count)) {
2112                         vq_err(vq, "Loop detected: last one at %u "
2113                                "indirect size %u\n",
2114                                i, count);
2115                         return -EINVAL;
2116                 }
2117                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2118                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2119                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2120                         return -EINVAL;
2121                 }
2122                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2123                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2124                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2125                         return -EINVAL;
2126                 }
2127
2128                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2129                         access = VHOST_ACCESS_WO;
2130                 else
2131                         access = VHOST_ACCESS_RO;
2132
2133                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2134                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2135                                      iov_size - iov_count, access);
2136                 if (unlikely(ret < 0)) {
2137                         if (ret != -EAGAIN)
2138                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2139                                         ret, i);
2140                         return ret;
2141                 }
2142                 /* If this is an input descriptor, increment that count. */
2143                 if (access == VHOST_ACCESS_WO) {
2144                         *in_num += ret;
2145                         if (unlikely(log && ret)) {
2146                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2147                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2148                                 ++*log_num;
2149                         }
2150                 } else {
2151                         /* If it's an output descriptor, they're all supposed
2152                          * to come before any input descriptors. */
2153                         if (unlikely(*in_num)) {
2154                                 vq_err(vq, "Indirect descriptor "
2155                                        "has out after in: idx %d\n", i);
2156                                 return -EINVAL;
2157                         }
2158                         *out_num += ret;
2159                 }
2160         } while ((i = next_desc(vq, &desc)) != -1);
2161         return 0;
2162 }
2163
2164 /* This looks in the virtqueue and for the first available buffer, and converts
2165  * it to an iovec for convenient access.  Since descriptors consist of some
2166  * number of output then some number of input descriptors, it's actually two
2167  * iovecs, but we pack them into one and note how many of each there were.
2168  *
2169  * This function returns the descriptor number found, or vq->num (which is
2170  * never a valid descriptor number) if none was found.  A negative code is
2171  * returned on error. */
2172 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2173                       struct iovec iov[], unsigned int iov_size,
2174                       unsigned int *out_num, unsigned int *in_num,
2175                       struct vhost_log *log, unsigned int *log_num)
2176 {
2177         struct vring_desc desc;
2178         unsigned int i, head, found = 0;
2179         u16 last_avail_idx;
2180         __virtio16 avail_idx;
2181         __virtio16 ring_head;
2182         int ret, access;
2183
2184         /* Check it isn't doing very strange things with descriptor numbers. */
2185         last_avail_idx = vq->last_avail_idx;
2186
2187         if (vq->avail_idx == vq->last_avail_idx) {
2188                 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2189                         vq_err(vq, "Failed to access avail idx at %p\n",
2190                                 &vq->avail->idx);
2191                         return -EFAULT;
2192                 }
2193                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2194
2195                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2196                         vq_err(vq, "Guest moved used index from %u to %u",
2197                                 last_avail_idx, vq->avail_idx);
2198                         return -EFAULT;
2199                 }
2200
2201                 /* If there's nothing new since last we looked, return
2202                  * invalid.
2203                  */
2204                 if (vq->avail_idx == last_avail_idx)
2205                         return vq->num;
2206
2207                 /* Only get avail ring entries after they have been
2208                  * exposed by guest.
2209                  */
2210                 smp_rmb();
2211         }
2212
2213         /* Grab the next descriptor number they're advertising, and increment
2214          * the index we've seen. */
2215         if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2216                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2217                        last_avail_idx,
2218                        &vq->avail->ring[last_avail_idx % vq->num]);
2219                 return -EFAULT;
2220         }
2221
2222         head = vhost16_to_cpu(vq, ring_head);
2223
2224         /* If their number is silly, that's an error. */
2225         if (unlikely(head >= vq->num)) {
2226                 vq_err(vq, "Guest says index %u > %u is available",
2227                        head, vq->num);
2228                 return -EINVAL;
2229         }
2230
2231         /* When we start there are none of either input nor output. */
2232         *out_num = *in_num = 0;
2233         if (unlikely(log))
2234                 *log_num = 0;
2235
2236         i = head;
2237         do {
2238                 unsigned iov_count = *in_num + *out_num;
2239                 if (unlikely(i >= vq->num)) {
2240                         vq_err(vq, "Desc index is %u > %u, head = %u",
2241                                i, vq->num, head);
2242                         return -EINVAL;
2243                 }
2244                 if (unlikely(++found > vq->num)) {
2245                         vq_err(vq, "Loop detected: last one at %u "
2246                                "vq size %u head %u\n",
2247                                i, vq->num, head);
2248                         return -EINVAL;
2249                 }
2250                 ret = vhost_get_desc(vq, &desc, i);
2251                 if (unlikely(ret)) {
2252                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2253                                i, vq->desc + i);
2254                         return -EFAULT;
2255                 }
2256                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2257                         ret = get_indirect(vq, iov, iov_size,
2258                                            out_num, in_num,
2259                                            log, log_num, &desc);
2260                         if (unlikely(ret < 0)) {
2261                                 if (ret != -EAGAIN)
2262                                         vq_err(vq, "Failure detected "
2263                                                 "in indirect descriptor at idx %d\n", i);
2264                                 return ret;
2265                         }
2266                         continue;
2267                 }
2268
2269                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2270                         access = VHOST_ACCESS_WO;
2271                 else
2272                         access = VHOST_ACCESS_RO;
2273                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2274                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2275                                      iov_size - iov_count, access);
2276                 if (unlikely(ret < 0)) {
2277                         if (ret != -EAGAIN)
2278                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2279                                         ret, i);
2280                         return ret;
2281                 }
2282                 if (access == VHOST_ACCESS_WO) {
2283                         /* If this is an input descriptor,
2284                          * increment that count. */
2285                         *in_num += ret;
2286                         if (unlikely(log && ret)) {
2287                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2288                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2289                                 ++*log_num;
2290                         }
2291                 } else {
2292                         /* If it's an output descriptor, they're all supposed
2293                          * to come before any input descriptors. */
2294                         if (unlikely(*in_num)) {
2295                                 vq_err(vq, "Descriptor has out after in: "
2296                                        "idx %d\n", i);
2297                                 return -EINVAL;
2298                         }
2299                         *out_num += ret;
2300                 }
2301         } while ((i = next_desc(vq, &desc)) != -1);
2302
2303         /* On success, increment avail index. */
2304         vq->last_avail_idx++;
2305
2306         /* Assume notifications from guest are disabled at this point,
2307          * if they aren't we would need to update avail_event index. */
2308         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2309         return head;
2310 }
2311 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2312
2313 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2314 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2315 {
2316         vq->last_avail_idx -= n;
2317 }
2318 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2319
2320 /* After we've used one of their buffers, we tell them about it.  We'll then
2321  * want to notify the guest, using eventfd. */
2322 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2323 {
2324         struct vring_used_elem heads = {
2325                 cpu_to_vhost32(vq, head),
2326                 cpu_to_vhost32(vq, len)
2327         };
2328
2329         return vhost_add_used_n(vq, &heads, 1);
2330 }
2331 EXPORT_SYMBOL_GPL(vhost_add_used);
2332
2333 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2334                             struct vring_used_elem *heads,
2335                             unsigned count)
2336 {
2337         vring_used_elem_t __user *used;
2338         u16 old, new;
2339         int start;
2340
2341         start = vq->last_used_idx & (vq->num - 1);
2342         used = vq->used->ring + start;
2343         if (vhost_put_used(vq, heads, start, count)) {
2344                 vq_err(vq, "Failed to write used");
2345                 return -EFAULT;
2346         }
2347         if (unlikely(vq->log_used)) {
2348                 /* Make sure data is seen before log. */
2349                 smp_wmb();
2350                 /* Log used ring entry write. */
2351                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2352                          count * sizeof *used);
2353         }
2354         old = vq->last_used_idx;
2355         new = (vq->last_used_idx += count);
2356         /* If the driver never bothers to signal in a very long while,
2357          * used index might wrap around. If that happens, invalidate
2358          * signalled_used index we stored. TODO: make sure driver
2359          * signals at least once in 2^16 and remove this. */
2360         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2361                 vq->signalled_used_valid = false;
2362         return 0;
2363 }
2364
2365 /* After we've used one of their buffers, we tell them about it.  We'll then
2366  * want to notify the guest, using eventfd. */
2367 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2368                      unsigned count)
2369 {
2370         int start, n, r;
2371
2372         start = vq->last_used_idx & (vq->num - 1);
2373         n = vq->num - start;
2374         if (n < count) {
2375                 r = __vhost_add_used_n(vq, heads, n);
2376                 if (r < 0)
2377                         return r;
2378                 heads += n;
2379                 count -= n;
2380         }
2381         r = __vhost_add_used_n(vq, heads, count);
2382
2383         /* Make sure buffer is written before we update index. */
2384         smp_wmb();
2385         if (vhost_put_used_idx(vq)) {
2386                 vq_err(vq, "Failed to increment used idx");
2387                 return -EFAULT;
2388         }
2389         if (unlikely(vq->log_used)) {
2390                 /* Make sure used idx is seen before log. */
2391                 smp_wmb();
2392                 /* Log used index update. */
2393                 log_used(vq, offsetof(struct vring_used, idx),
2394                          sizeof vq->used->idx);
2395                 if (vq->log_ctx)
2396                         eventfd_signal(vq->log_ctx, 1);
2397         }
2398         return r;
2399 }
2400 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2401
2402 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2403 {
2404         __u16 old, new;
2405         __virtio16 event;
2406         bool v;
2407         /* Flush out used index updates. This is paired
2408          * with the barrier that the Guest executes when enabling
2409          * interrupts. */
2410         smp_mb();
2411
2412         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2413             unlikely(vq->avail_idx == vq->last_avail_idx))
2414                 return true;
2415
2416         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2417                 __virtio16 flags;
2418                 if (vhost_get_avail_flags(vq, &flags)) {
2419                         vq_err(vq, "Failed to get flags");
2420                         return true;
2421                 }
2422                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2423         }
2424         old = vq->signalled_used;
2425         v = vq->signalled_used_valid;
2426         new = vq->signalled_used = vq->last_used_idx;
2427         vq->signalled_used_valid = true;
2428
2429         if (unlikely(!v))
2430                 return true;
2431
2432         if (vhost_get_used_event(vq, &event)) {
2433                 vq_err(vq, "Failed to get used event idx");
2434                 return true;
2435         }
2436         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2437 }
2438
2439 /* This actually signals the guest, using eventfd. */
2440 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2441 {
2442         /* Signal the Guest tell them we used something up. */
2443         if (vq->call_ctx && vhost_notify(dev, vq))
2444                 eventfd_signal(vq->call_ctx, 1);
2445 }
2446 EXPORT_SYMBOL_GPL(vhost_signal);
2447
2448 /* And here's the combo meal deal.  Supersize me! */
2449 void vhost_add_used_and_signal(struct vhost_dev *dev,
2450                                struct vhost_virtqueue *vq,
2451                                unsigned int head, int len)
2452 {
2453         vhost_add_used(vq, head, len);
2454         vhost_signal(dev, vq);
2455 }
2456 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2457
2458 /* multi-buffer version of vhost_add_used_and_signal */
2459 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2460                                  struct vhost_virtqueue *vq,
2461                                  struct vring_used_elem *heads, unsigned count)
2462 {
2463         vhost_add_used_n(vq, heads, count);
2464         vhost_signal(dev, vq);
2465 }
2466 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2467
2468 /* return true if we're sure that avaiable ring is empty */
2469 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2470 {
2471         __virtio16 avail_idx;
2472         int r;
2473
2474         if (vq->avail_idx != vq->last_avail_idx)
2475                 return false;
2476
2477         r = vhost_get_avail_idx(vq, &avail_idx);
2478         if (unlikely(r))
2479                 return false;
2480         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2481
2482         return vq->avail_idx == vq->last_avail_idx;
2483 }
2484 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2485
2486 /* OK, now we need to know about added descriptors. */
2487 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2488 {
2489         __virtio16 avail_idx;
2490         int r;
2491
2492         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2493                 return false;
2494         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2495         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2496                 r = vhost_update_used_flags(vq);
2497                 if (r) {
2498                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2499                                &vq->used->flags, r);
2500                         return false;
2501                 }
2502         } else {
2503                 r = vhost_update_avail_event(vq, vq->avail_idx);
2504                 if (r) {
2505                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2506                                vhost_avail_event(vq), r);
2507                         return false;
2508                 }
2509         }
2510         /* They could have slipped one in as we were doing that: make
2511          * sure it's written, then check again. */
2512         smp_mb();
2513         r = vhost_get_avail_idx(vq, &avail_idx);
2514         if (r) {
2515                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2516                        &vq->avail->idx, r);
2517                 return false;
2518         }
2519
2520         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2521 }
2522 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2523
2524 /* We don't need to be notified again. */
2525 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2526 {
2527         int r;
2528
2529         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2530                 return;
2531         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2532         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2533                 r = vhost_update_used_flags(vq);
2534                 if (r)
2535                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2536                                &vq->used->flags, r);
2537         }
2538 }
2539 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2540
2541 /* Create a new message. */
2542 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2543 {
2544         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2545         if (!node)
2546                 return NULL;
2547
2548         /* Make sure all padding within the structure is initialized. */
2549         memset(&node->msg, 0, sizeof node->msg);
2550         node->vq = vq;
2551         node->msg.type = type;
2552         return node;
2553 }
2554 EXPORT_SYMBOL_GPL(vhost_new_msg);
2555
2556 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2557                        struct vhost_msg_node *node)
2558 {
2559         spin_lock(&dev->iotlb_lock);
2560         list_add_tail(&node->node, head);
2561         spin_unlock(&dev->iotlb_lock);
2562
2563         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2564 }
2565 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2566
2567 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2568                                          struct list_head *head)
2569 {
2570         struct vhost_msg_node *node = NULL;
2571
2572         spin_lock(&dev->iotlb_lock);
2573         if (!list_empty(head)) {
2574                 node = list_first_entry(head, struct vhost_msg_node,
2575                                         node);
2576                 list_del(&node->node);
2577         }
2578         spin_unlock(&dev->iotlb_lock);
2579
2580         return node;
2581 }
2582 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2583
2584
2585 static int __init vhost_init(void)
2586 {
2587         return 0;
2588 }
2589
2590 static void __exit vhost_exit(void)
2591 {
2592 }
2593
2594 module_init(vhost_init);
2595 module_exit(vhost_exit);
2596
2597 MODULE_VERSION("0.0.1");
2598 MODULE_LICENSE("GPL v2");
2599 MODULE_AUTHOR("Michael S. Tsirkin");
2600 MODULE_DESCRIPTION("Host kernel accelerator for virtio");