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_vring_call_reset(struct vhost_vring_call *call_ctx)
 302 {
 303         call_ctx->ctx = NULL;
 304         memset(&call_ctx->producer, 0x0, sizeof(struct irq_bypass_producer));
 305 }
 306
 307 static void vhost_vq_reset(struct vhost_dev *dev,
 308                            struct vhost_virtqueue *vq)
 309 {
 310         vq->num = 1;
 311         vq->desc = NULL;
 312         vq->avail = NULL;
 313         vq->used = NULL;
 314         vq->last_avail_idx = 0;
 315         vq->avail_idx = 0;
 316         vq->last_used_idx = 0;
 317         vq->signalled_used = 0;
 318         vq->signalled_used_valid = false;
 319         vq->used_flags = 0;
 320         vq->log_used = false;
 321         vq->log_addr = -1ull;
 322         vq->private_data = NULL;
 323         vq->acked_features = 0;
 324         vq->acked_backend_features = 0;
 325         vq->log_base = NULL;
 326         vq->error_ctx = NULL;
 327         vq->kick = NULL;
 328         vq->log_ctx = NULL;
 329         vhost_reset_is_le(vq);
 330         vhost_disable_cross_endian(vq);
 331         vq->busyloop_timeout = 0;
 332         vq->umem = NULL;
 333         vq->iotlb = NULL;
 334         vhost_vring_call_reset(&vq->call_ctx);
 335         __vhost_vq_meta_reset(vq);
 336 }
 337
 338 static int vhost_worker(void *data)
 339 {
 340         struct vhost_dev *dev = data;
 341         struct vhost_work *work, *work_next;
 342         struct llist_node *node;
 343
 344         kthread_use_mm(dev->mm);
 345
 346         for (;;) {
 347                 /* mb paired w/ kthread_stop */
 348                 set_current_state(TASK_INTERRUPTIBLE);
 349
 350                 if (kthread_should_stop()) {
 351                         __set_current_state(TASK_RUNNING);
 352                         break;
 353                 }
 354
 355                 node = llist_del_all(&dev->work_list);
 356                 if (!node)
 357                         schedule();
 358
 359                 node = llist_reverse_order(node);
 360                 /* make sure flag is seen after deletion */
 361                 smp_wmb();
 362                 llist_for_each_entry_safe(work, work_next, node, node) {
 363                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 364                         __set_current_state(TASK_RUNNING);
 365                         kcov_remote_start_common(dev->kcov_handle);
 366                         work->fn(work);
 367                         kcov_remote_stop();
 368                         if (need_resched())
 369                                 schedule();
 370                 }
 371         }
 372         kthread_unuse_mm(dev->mm);
 373         return 0;
 374 }
 375
 376 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 377 {
 378         kfree(vq->indirect);
 379         vq->indirect = NULL;
 380         kfree(vq->log);
 381         vq->log = NULL;
 382         kfree(vq->heads);
 383         vq->heads = NULL;
 384 }
 385
 386 /* Helper to allocate iovec buffers for all vqs. */
 387 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 388 {
 389         struct vhost_virtqueue *vq;
 390         int i;
 391
 392         for (i = 0; i < dev->nvqs; ++i) {
 393                 vq = dev->vqs[i];
 394                 vq->indirect = kmalloc_array(UIO_MAXIOV,
 395                                              sizeof(*vq->indirect),
 396                                              GFP_KERNEL);
 397                 vq->log = kmalloc_array(dev->iov_limit, sizeof(*vq->log),
 398                                         GFP_KERNEL);
 399                 vq->heads = kmalloc_array(dev->iov_limit, sizeof(*vq->heads),
 400                                           GFP_KERNEL);
 401                 if (!vq->indirect || !vq->log || !vq->heads)
 402                         goto err_nomem;
 403         }
 404         return 0;
 405
 406 err_nomem:
 407         for (; i >= 0; --i)
 408                 vhost_vq_free_iovecs(dev->vqs[i]);
 409         return -ENOMEM;
 410 }
 411
 412 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 413 {
 414         int i;
 415
 416         for (i = 0; i < dev->nvqs; ++i)
 417                 vhost_vq_free_iovecs(dev->vqs[i]);
 418 }
 419
 420 bool vhost_exceeds_weight(struct vhost_virtqueue *vq,
 421                           int pkts, int total_len)
 422 {
 423         struct vhost_dev *dev = vq->dev;
 424
 425         if ((dev->byte_weight && total_len >= dev->byte_weight) ||
 426             pkts >= dev->weight) {
 427                 vhost_poll_queue(&vq->poll);
 428                 return true;
 429         }
 430
 431         return false;
 432 }
 433 EXPORT_SYMBOL_GPL(vhost_exceeds_weight);
 434
 435 static size_t vhost_get_avail_size(struct vhost_virtqueue *vq,
 436                                    unsigned int num)
 437 {
 438         size_t event __maybe_unused =
 439                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 440
 441         return sizeof(*vq->avail) +
 442                sizeof(*vq->avail->ring) * num + event;
 443 }
 444
 445 static size_t vhost_get_used_size(struct vhost_virtqueue *vq,
 446                                   unsigned int num)
 447 {
 448         size_t event __maybe_unused =
 449                vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 450
 451         return sizeof(*vq->used) +
 452                sizeof(*vq->used->ring) * num + event;
 453 }
 454
 455 static size_t vhost_get_desc_size(struct vhost_virtqueue *vq,
 456                                   unsigned int num)
 457 {
 458         return sizeof(*vq->desc) * num;
 459 }
 460
 461 void vhost_dev_init(struct vhost_dev *dev,
 462                     struct vhost_virtqueue **vqs, int nvqs,
 463                     int iov_limit, int weight, int byte_weight,
 464                     bool use_worker,
 465                     int (*msg_handler)(struct vhost_dev *dev,
 466                                        struct vhost_iotlb_msg *msg))
 467 {
 468         struct vhost_virtqueue *vq;
 469         int i;
 470
 471         dev->vqs = vqs;
 472         dev->nvqs = nvqs;
 473         mutex_init(&dev->mutex);
 474         dev->log_ctx = NULL;
 475         dev->umem = NULL;
 476         dev->iotlb = NULL;
 477         dev->mm = NULL;
 478         dev->worker = NULL;
 479         dev->iov_limit = iov_limit;
 480         dev->weight = weight;
 481         dev->byte_weight = byte_weight;
 482         dev->use_worker = use_worker;
 483         dev->msg_handler = msg_handler;
 484         init_llist_head(&dev->work_list);
 485         init_waitqueue_head(&dev->wait);
 486         INIT_LIST_HEAD(&dev->read_list);
 487         INIT_LIST_HEAD(&dev->pending_list);
 488         spin_lock_init(&dev->iotlb_lock);
 489
 490
 491         for (i = 0; i < dev->nvqs; ++i) {
 492                 vq = dev->vqs[i];
 493                 vq->log = NULL;
 494                 vq->indirect = NULL;
 495                 vq->heads = NULL;
 496                 vq->dev = dev;
 497                 mutex_init(&vq->mutex);
 498                 vhost_vq_reset(dev, vq);
 499                 if (vq->handle_kick)
 500                         vhost_poll_init(&vq->poll, vq->handle_kick,
 501                                         EPOLLIN, dev);
 502         }
 503 }
 504 EXPORT_SYMBOL_GPL(vhost_dev_init);
 505
 506 /* Caller should have device mutex */
 507 long vhost_dev_check_owner(struct vhost_dev *dev)
 508 {
 509         /* Are you the owner? If not, I don't think you mean to do that */
 510         return dev->mm == current->mm ? 0 : -EPERM;
 511 }
 512 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 513
 514 struct vhost_attach_cgroups_struct {
 515         struct vhost_work work;
 516         struct task_struct *owner;
 517         int ret;
 518 };
 519
 520 static void vhost_attach_cgroups_work(struct vhost_work *work)
 521 {
 522         struct vhost_attach_cgroups_struct *s;
 523
 524         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 525         s->ret = cgroup_attach_task_all(s->owner, current);
 526 }
 527
 528 static int vhost_attach_cgroups(struct vhost_dev *dev)
 529 {
 530         struct vhost_attach_cgroups_struct attach;
 531
 532         attach.owner = current;
 533         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 534         vhost_work_queue(dev, &attach.work);
 535         vhost_work_flush(dev, &attach.work);
 536         return attach.ret;
 537 }
 538
 539 /* Caller should have device mutex */
 540 bool vhost_dev_has_owner(struct vhost_dev *dev)
 541 {
 542         return dev->mm;
 543 }
 544 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 545
 546 static void vhost_attach_mm(struct vhost_dev *dev)
 547 {
 548         /* No owner, become one */
 549         if (dev->use_worker) {
 550                 dev->mm = get_task_mm(current);
 551         } else {
 552                 /* vDPA device does not use worker thead, so there's
 553                  * no need to hold the address space for mm. This help
 554                  * to avoid deadlock in the case of mmap() which may
 555                  * held the refcnt of the file and depends on release
 556                  * method to remove vma.
 557                  */
 558                 dev->mm = current->mm;
 559                 mmgrab(dev->mm);
 560         }
 561 }
 562
 563 static void vhost_detach_mm(struct vhost_dev *dev)
 564 {
 565         if (!dev->mm)
 566                 return;
 567
 568         if (dev->use_worker)
 569                 mmput(dev->mm);
 570         else
 571                 mmdrop(dev->mm);
 572
 573         dev->mm = NULL;
 574 }
 575
 576 /* Caller should have device mutex */
 577 long vhost_dev_set_owner(struct vhost_dev *dev)
 578 {
 579         struct task_struct *worker;
 580         int err;
 581
 582         /* Is there an owner already? */
 583         if (vhost_dev_has_owner(dev)) {
 584                 err = -EBUSY;
 585                 goto err_mm;
 586         }
 587
 588         vhost_attach_mm(dev);
 589
 590         dev->kcov_handle = kcov_common_handle();
 591         if (dev->use_worker) {
 592                 worker = kthread_create(vhost_worker, dev,
 593                                         "vhost-%d", current->pid);
 594                 if (IS_ERR(worker)) {
 595                         err = PTR_ERR(worker);
 596                         goto err_worker;
 597                 }
 598
 599                 dev->worker = worker;
 600                 wake_up_process(worker); /* avoid contributing to loadavg */
 601
 602                 err = vhost_attach_cgroups(dev);
 603                 if (err)
 604                         goto err_cgroup;
 605         }
 606
 607         err = vhost_dev_alloc_iovecs(dev);
 608         if (err)
 609                 goto err_cgroup;
 610
 611         return 0;
 612 err_cgroup:
 613         if (dev->worker) {
 614                 kthread_stop(dev->worker);
 615                 dev->worker = NULL;
 616         }
 617 err_worker:
 618         vhost_detach_mm(dev);
 619         dev->kcov_handle = 0;
 620 err_mm:
 621         return err;
 622 }
 623 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 624
 625 static struct vhost_iotlb *iotlb_alloc(void)
 626 {
 627         return vhost_iotlb_alloc(max_iotlb_entries,
 628                                  VHOST_IOTLB_FLAG_RETIRE);
 629 }
 630
 631 struct vhost_iotlb *vhost_dev_reset_owner_prepare(void)
 632 {
 633         return iotlb_alloc();
 634 }
 635 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 636
 637 /* Caller should have device mutex */
 638 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_iotlb *umem)
 639 {
 640         int i;
 641
 642         vhost_dev_cleanup(dev);
 643
 644         dev->umem = umem;
 645         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 646          * VQs aren't running.
 647          */
 648         for (i = 0; i < dev->nvqs; ++i)
 649                 dev->vqs[i]->umem = umem;
 650 }
 651 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 652
 653 void vhost_dev_stop(struct vhost_dev *dev)
 654 {
 655         int i;
 656
 657         for (i = 0; i < dev->nvqs; ++i) {
 658                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 659                         vhost_poll_stop(&dev->vqs[i]->poll);
 660                         vhost_poll_flush(&dev->vqs[i]->poll);
 661                 }
 662         }
 663 }
 664 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 665
 666 static void vhost_clear_msg(struct vhost_dev *dev)
 667 {
 668         struct vhost_msg_node *node, *n;
 669
 670         spin_lock(&dev->iotlb_lock);
 671
 672         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 673                 list_del(&node->node);
 674                 kfree(node);
 675         }
 676
 677         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 678                 list_del(&node->node);
 679                 kfree(node);
 680         }
 681
 682         spin_unlock(&dev->iotlb_lock);
 683 }
 684
 685 void vhost_dev_cleanup(struct vhost_dev *dev)
 686 {
 687         int i;
 688
 689         for (i = 0; i < dev->nvqs; ++i) {
 690                 if (dev->vqs[i]->error_ctx)
 691                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 692                 if (dev->vqs[i]->kick)
 693                         fput(dev->vqs[i]->kick);
 694                 if (dev->vqs[i]->call_ctx.ctx)
 695                         eventfd_ctx_put(dev->vqs[i]->call_ctx.ctx);
 696                 vhost_vq_reset(dev, dev->vqs[i]);
 697         }
 698         vhost_dev_free_iovecs(dev);
 699         if (dev->log_ctx)
 700                 eventfd_ctx_put(dev->log_ctx);
 701         dev->log_ctx = NULL;
 702         /* No one will access memory at this point */
 703         vhost_iotlb_free(dev->umem);
 704         dev->umem = NULL;
 705         vhost_iotlb_free(dev->iotlb);
 706         dev->iotlb = NULL;
 707         vhost_clear_msg(dev);
 708         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
 709         WARN_ON(!llist_empty(&dev->work_list));
 710         if (dev->worker) {
 711                 kthread_stop(dev->worker);
 712                 dev->worker = NULL;
 713                 dev->kcov_handle = 0;
 714         }
 715         vhost_detach_mm(dev);
 716 }
 717 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 718
 719 static bool log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 720 {
 721         u64 a = addr / VHOST_PAGE_SIZE / 8;
 722
 723         /* Make sure 64 bit math will not overflow. */
 724         if (a > ULONG_MAX - (unsigned long)log_base ||
 725             a + (unsigned long)log_base > ULONG_MAX)
 726                 return false;
 727
 728         return access_ok(log_base + a,
 729                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 730 }
 731
 732 static bool vhost_overflow(u64 uaddr, u64 size)
 733 {
 734         /* Make sure 64 bit math will not overflow. */
 735         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 736 }
 737
 738 /* Caller should have vq mutex and device mutex. */
 739 static bool vq_memory_access_ok(void __user *log_base, struct vhost_iotlb *umem,
 740                                 int log_all)
 741 {
 742         struct vhost_iotlb_map *map;
 743
 744         if (!umem)
 745                 return false;
 746
 747         list_for_each_entry(map, &umem->list, link) {
 748                 unsigned long a = map->addr;
 749
 750                 if (vhost_overflow(map->addr, map->size))
 751                         return false;
 752
 753
 754                 if (!access_ok((void __user *)a, map->size))
 755                         return false;
 756                 else if (log_all && !log_access_ok(log_base,
 757                                                    map->start,
 758                                                    map->size))
 759                         return false;
 760         }
 761         return true;
 762 }
 763
 764 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 765                                                u64 addr, unsigned int size,
 766                                                int type)
 767 {
 768         const struct vhost_iotlb_map *map = vq->meta_iotlb[type];
 769
 770         if (!map)
 771                 return NULL;
 772
 773         return (void __user *)(uintptr_t)(map->addr + addr - map->start);
 774 }
 775
 776 /* Can we switch to this memory table? */
 777 /* Caller should have device mutex but not vq mutex */
 778 static bool memory_access_ok(struct vhost_dev *d, struct vhost_iotlb *umem,
 779                              int log_all)
 780 {
 781         int i;
 782
 783         for (i = 0; i < d->nvqs; ++i) {
 784                 bool ok;
 785                 bool log;
 786
 787                 mutex_lock(&d->vqs[i]->mutex);
 788                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 789                 /* If ring is inactive, will check when it's enabled. */
 790                 if (d->vqs[i]->private_data)
 791                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 792                                                  umem, log);
 793                 else
 794                         ok = true;
 795                 mutex_unlock(&d->vqs[i]->mutex);
 796                 if (!ok)
 797                         return false;
 798         }
 799         return true;
 800 }
 801
 802 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 803                           struct iovec iov[], int iov_size, int access);
 804
 805 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 806                               const void *from, unsigned size)
 807 {
 808         int ret;
 809
 810         if (!vq->iotlb)
 811                 return __copy_to_user(to, from, size);
 812         else {
 813                 /* This function should be called after iotlb
 814                  * prefetch, which means we're sure that all vq
 815                  * could be access through iotlb. So -EAGAIN should
 816                  * not happen in this case.
 817                  */
 818                 struct iov_iter t;
 819                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 820                                      (u64)(uintptr_t)to, size,
 821                                      VHOST_ADDR_USED);
 822
 823                 if (uaddr)
 824                         return __copy_to_user(uaddr, from, size);
 825
 826                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 827                                      ARRAY_SIZE(vq->iotlb_iov),
 828                                      VHOST_ACCESS_WO);
 829                 if (ret < 0)
 830                         goto out;
 831                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 832                 ret = copy_to_iter(from, size, &t);
 833                 if (ret == size)
 834                         ret = 0;
 835         }
 836 out:
 837         return ret;
 838 }
 839
 840 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 841                                 void __user *from, unsigned size)
 842 {
 843         int ret;
 844
 845         if (!vq->iotlb)
 846                 return __copy_from_user(to, from, size);
 847         else {
 848                 /* This function should be called after iotlb
 849                  * prefetch, which means we're sure that vq
 850                  * could be access through iotlb. So -EAGAIN should
 851                  * not happen in this case.
 852                  */
 853                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 854                                      (u64)(uintptr_t)from, size,
 855                                      VHOST_ADDR_DESC);
 856                 struct iov_iter f;
 857
 858                 if (uaddr)
 859                         return __copy_from_user(to, uaddr, size);
 860
 861                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 862                                      ARRAY_SIZE(vq->iotlb_iov),
 863                                      VHOST_ACCESS_RO);
 864                 if (ret < 0) {
 865                         vq_err(vq, "IOTLB translation failure: uaddr "
 866                                "%p size 0x%llx\n", from,
 867                                (unsigned long long) size);
 868                         goto out;
 869                 }
 870                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 871                 ret = copy_from_iter(to, size, &f);
 872                 if (ret == size)
 873                         ret = 0;
 874         }
 875
 876 out:
 877         return ret;
 878 }
 879
 880 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 881                                           void __user *addr, unsigned int size,
 882                                           int type)
 883 {
 884         int ret;
 885
 886         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 887                              ARRAY_SIZE(vq->iotlb_iov),
 888                              VHOST_ACCESS_RO);
 889         if (ret < 0) {
 890                 vq_err(vq, "IOTLB translation failure: uaddr "
 891                         "%p size 0x%llx\n", addr,
 892                         (unsigned long long) size);
 893                 return NULL;
 894         }
 895
 896         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 897                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 898                         "%p size 0x%llx\n", addr,
 899                         (unsigned long long) size);
 900                 return NULL;
 901         }
 902
 903         return vq->iotlb_iov[0].iov_base;
 904 }
 905
 906 /* This function should be called after iotlb
 907  * prefetch, which means we're sure that vq
 908  * could be access through iotlb. So -EAGAIN should
 909  * not happen in this case.
 910  */
 911 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 912                                             void __user *addr, unsigned int size,
 913                                             int type)
 914 {
 915         void __user *uaddr = vhost_vq_meta_fetch(vq,
 916                              (u64)(uintptr_t)addr, size, type);
 917         if (uaddr)
 918                 return uaddr;
 919
 920         return __vhost_get_user_slow(vq, addr, size, type);
 921 }
 922
 923 #define vhost_put_user(vq, x, ptr)              \
 924 ({ \
 925         int ret; \
 926         if (!vq->iotlb) { \
 927                 ret = __put_user(x, ptr); \
 928         } else { \
 929                 __typeof__(ptr) to = \
 930                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 931                                           sizeof(*ptr), VHOST_ADDR_USED); \
 932                 if (to != NULL) \
 933                         ret = __put_user(x, to); \
 934                 else \
 935                         ret = -EFAULT;  \
 936         } \
 937         ret; \
 938 })
 939
 940 static inline int vhost_put_avail_event(struct vhost_virtqueue *vq)
 941 {
 942         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
 943                               vhost_avail_event(vq));
 944 }
 945
 946 static inline int vhost_put_used(struct vhost_virtqueue *vq,
 947                                  struct vring_used_elem *head, int idx,
 948                                  int count)
 949 {
 950         return vhost_copy_to_user(vq, vq->used->ring + idx, head,
 951                                   count * sizeof(*head));
 952 }
 953
 954 static inline int vhost_put_used_flags(struct vhost_virtqueue *vq)
 955
 956 {
 957         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
 958                               &vq->used->flags);
 959 }
 960
 961 static inline int vhost_put_used_idx(struct vhost_virtqueue *vq)
 962
 963 {
 964         return vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
 965                               &vq->used->idx);
 966 }
 967
 968 #define vhost_get_user(vq, x, ptr, type)                \
 969 ({ \
 970         int ret; \
 971         if (!vq->iotlb) { \
 972                 ret = __get_user(x, ptr); \
 973         } else { \
 974                 __typeof__(ptr) from = \
 975                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 976                                                            sizeof(*ptr), \
 977                                                            type); \
 978                 if (from != NULL) \
 979                         ret = __get_user(x, from); \
 980                 else \
 981                         ret = -EFAULT; \
 982         } \
 983         ret; \
 984 })
 985
 986 #define vhost_get_avail(vq, x, ptr) \
 987         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 988
 989 #define vhost_get_used(vq, x, ptr) \
 990         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 991
 992 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 993 {
 994         int i = 0;
 995         for (i = 0; i < d->nvqs; ++i)
 996                 mutex_lock_nested(&d->vqs[i]->mutex, i);
 997 }
 998
 999 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
1000 {
1001         int i = 0;
1002         for (i = 0; i < d->nvqs; ++i)
1003                 mutex_unlock(&d->vqs[i]->mutex);
1004 }
1005
1006 static inline int vhost_get_avail_idx(struct vhost_virtqueue *vq,
1007                                       __virtio16 *idx)
1008 {
1009         return vhost_get_avail(vq, *idx, &vq->avail->idx);
1010 }
1011
1012 static inline int vhost_get_avail_head(struct vhost_virtqueue *vq,
1013                                        __virtio16 *head, int idx)
1014 {
1015         return vhost_get_avail(vq, *head,
1016                                &vq->avail->ring[idx & (vq->num - 1)]);
1017 }
1018
1019 static inline int vhost_get_avail_flags(struct vhost_virtqueue *vq,
1020                                         __virtio16 *flags)
1021 {
1022         return vhost_get_avail(vq, *flags, &vq->avail->flags);
1023 }
1024
1025 static inline int vhost_get_used_event(struct vhost_virtqueue *vq,
1026                                        __virtio16 *event)
1027 {
1028         return vhost_get_avail(vq, *event, vhost_used_event(vq));
1029 }
1030
1031 static inline int vhost_get_used_idx(struct vhost_virtqueue *vq,
1032                                      __virtio16 *idx)
1033 {
1034         return vhost_get_used(vq, *idx, &vq->used->idx);
1035 }
1036
1037 static inline int vhost_get_desc(struct vhost_virtqueue *vq,
1038                                  struct vring_desc *desc, int idx)
1039 {
1040         return vhost_copy_from_user(vq, desc, vq->desc + idx, sizeof(*desc));
1041 }
1042
1043 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
1044                                   struct vhost_iotlb_msg *msg)
1045 {
1046         struct vhost_msg_node *node, *n;
1047
1048         spin_lock(&d->iotlb_lock);
1049
1050         list_for_each_entry_safe(node, n, &d->pending_list, node) {
1051                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
1052                 if (msg->iova <= vq_msg->iova &&
1053                     msg->iova + msg->size - 1 >= vq_msg->iova &&
1054                     vq_msg->type == VHOST_IOTLB_MISS) {
1055                         vhost_poll_queue(&node->vq->poll);
1056                         list_del(&node->node);
1057                         kfree(node);
1058                 }
1059         }
1060
1061         spin_unlock(&d->iotlb_lock);
1062 }
1063
1064 static bool umem_access_ok(u64 uaddr, u64 size, int access)
1065 {
1066         unsigned long a = uaddr;
1067
1068         /* Make sure 64 bit math will not overflow. */
1069         if (vhost_overflow(uaddr, size))
1070                 return false;
1071
1072         if ((access & VHOST_ACCESS_RO) &&
1073             !access_ok((void __user *)a, size))
1074                 return false;
1075         if ((access & VHOST_ACCESS_WO) &&
1076             !access_ok((void __user *)a, size))
1077                 return false;
1078         return true;
1079 }
1080
1081 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1082                                    struct vhost_iotlb_msg *msg)
1083 {
1084         int ret = 0;
1085
1086         mutex_lock(&dev->mutex);
1087         vhost_dev_lock_vqs(dev);
1088         switch (msg->type) {
1089         case VHOST_IOTLB_UPDATE:
1090                 if (!dev->iotlb) {
1091                         ret = -EFAULT;
1092                         break;
1093                 }
1094                 if (!umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1095                         ret = -EFAULT;
1096                         break;
1097                 }
1098                 vhost_vq_meta_reset(dev);
1099                 if (vhost_iotlb_add_range(dev->iotlb, msg->iova,
1100                                           msg->iova + msg->size - 1,
1101                                           msg->uaddr, msg->perm)) {
1102                         ret = -ENOMEM;
1103                         break;
1104                 }
1105                 vhost_iotlb_notify_vq(dev, msg);
1106                 break;
1107         case VHOST_IOTLB_INVALIDATE:
1108                 if (!dev->iotlb) {
1109                         ret = -EFAULT;
1110                         break;
1111                 }
1112                 vhost_vq_meta_reset(dev);
1113                 vhost_iotlb_del_range(dev->iotlb, msg->iova,
1114                                       msg->iova + msg->size - 1);
1115                 break;
1116         default:
1117                 ret = -EINVAL;
1118                 break;
1119         }
1120
1121         vhost_dev_unlock_vqs(dev);
1122         mutex_unlock(&dev->mutex);
1123
1124         return ret;
1125 }
1126 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1127                              struct iov_iter *from)
1128 {
1129         struct vhost_iotlb_msg msg;
1130         size_t offset;
1131         int type, ret;
1132
1133         ret = copy_from_iter(&type, sizeof(type), from);
1134         if (ret != sizeof(type)) {
1135                 ret = -EINVAL;
1136                 goto done;
1137         }
1138
1139         switch (type) {
1140         case VHOST_IOTLB_MSG:
1141                 /* There maybe a hole after type for V1 message type,
1142                  * so skip it here.
1143                  */
1144                 offset = offsetof(struct vhost_msg, iotlb) - sizeof(int);
1145                 break;
1146         case VHOST_IOTLB_MSG_V2:
1147                 offset = sizeof(__u32);
1148                 break;
1149         default:
1150                 ret = -EINVAL;
1151                 goto done;
1152         }
1153
1154         iov_iter_advance(from, offset);
1155         ret = copy_from_iter(&msg, sizeof(msg), from);
1156         if (ret != sizeof(msg)) {
1157                 ret = -EINVAL;
1158                 goto done;
1159         }
1160
1161         if (dev->msg_handler)
1162                 ret = dev->msg_handler(dev, &msg);
1163         else
1164                 ret = vhost_process_iotlb_msg(dev, &msg);
1165         if (ret) {
1166                 ret = -EFAULT;
1167                 goto done;
1168         }
1169
1170         ret = (type == VHOST_IOTLB_MSG) ? sizeof(struct vhost_msg) :
1171               sizeof(struct vhost_msg_v2);
1172 done:
1173         return ret;
1174 }
1175 EXPORT_SYMBOL(vhost_chr_write_iter);
1176
1177 __poll_t vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1178                             poll_table *wait)
1179 {
1180         __poll_t mask = 0;
1181
1182         poll_wait(file, &dev->wait, wait);
1183
1184         if (!list_empty(&dev->read_list))
1185                 mask |= EPOLLIN | EPOLLRDNORM;
1186
1187         return mask;
1188 }
1189 EXPORT_SYMBOL(vhost_chr_poll);
1190
1191 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1192                             int noblock)
1193 {
1194         DEFINE_WAIT(wait);
1195         struct vhost_msg_node *node;
1196         ssize_t ret = 0;
1197         unsigned size = sizeof(struct vhost_msg);
1198
1199         if (iov_iter_count(to) < size)
1200                 return 0;
1201
1202         while (1) {
1203                 if (!noblock)
1204                         prepare_to_wait(&dev->wait, &wait,
1205                                         TASK_INTERRUPTIBLE);
1206
1207                 node = vhost_dequeue_msg(dev, &dev->read_list);
1208                 if (node)
1209                         break;
1210                 if (noblock) {
1211                         ret = -EAGAIN;
1212                         break;
1213                 }
1214                 if (signal_pending(current)) {
1215                         ret = -ERESTARTSYS;
1216                         break;
1217                 }
1218                 if (!dev->iotlb) {
1219                         ret = -EBADFD;
1220                         break;
1221                 }
1222
1223                 schedule();
1224         }
1225
1226         if (!noblock)
1227                 finish_wait(&dev->wait, &wait);
1228
1229         if (node) {
1230                 struct vhost_iotlb_msg *msg;
1231                 void *start = &node->msg;
1232
1233                 switch (node->msg.type) {
1234                 case VHOST_IOTLB_MSG:
1235                         size = sizeof(node->msg);
1236                         msg = &node->msg.iotlb;
1237                         break;
1238                 case VHOST_IOTLB_MSG_V2:
1239                         size = sizeof(node->msg_v2);
1240                         msg = &node->msg_v2.iotlb;
1241                         break;
1242                 default:
1243                         BUG();
1244                         break;
1245                 }
1246
1247                 ret = copy_to_iter(start, size, to);
1248                 if (ret != size || msg->type != VHOST_IOTLB_MISS) {
1249                         kfree(node);
1250                         return ret;
1251                 }
1252                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1253         }
1254
1255         return ret;
1256 }
1257 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1258
1259 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1260 {
1261         struct vhost_dev *dev = vq->dev;
1262         struct vhost_msg_node *node;
1263         struct vhost_iotlb_msg *msg;
1264         bool v2 = vhost_backend_has_feature(vq, VHOST_BACKEND_F_IOTLB_MSG_V2);
1265
1266         node = vhost_new_msg(vq, v2 ? VHOST_IOTLB_MSG_V2 : VHOST_IOTLB_MSG);
1267         if (!node)
1268                 return -ENOMEM;
1269
1270         if (v2) {
1271                 node->msg_v2.type = VHOST_IOTLB_MSG_V2;
1272                 msg = &node->msg_v2.iotlb;
1273         } else {
1274                 msg = &node->msg.iotlb;
1275         }
1276
1277         msg->type = VHOST_IOTLB_MISS;
1278         msg->iova = iova;
1279         msg->perm = access;
1280
1281         vhost_enqueue_msg(dev, &dev->read_list, node);
1282
1283         return 0;
1284 }
1285
1286 static bool vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1287                          vring_desc_t __user *desc,
1288                          vring_avail_t __user *avail,
1289                          vring_used_t __user *used)
1290
1291 {
1292         /* If an IOTLB device is present, the vring addresses are
1293          * GIOVAs. Access validation occurs at prefetch time. */
1294         if (vq->iotlb)
1295                 return true;
1296
1297         return access_ok(desc, vhost_get_desc_size(vq, num)) &&
1298                access_ok(avail, vhost_get_avail_size(vq, num)) &&
1299                access_ok(used, vhost_get_used_size(vq, num));
1300 }
1301
1302 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1303                                  const struct vhost_iotlb_map *map,
1304                                  int type)
1305 {
1306         int access = (type == VHOST_ADDR_USED) ?
1307                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1308
1309         if (likely(map->perm & access))
1310                 vq->meta_iotlb[type] = map;
1311 }
1312
1313 static bool iotlb_access_ok(struct vhost_virtqueue *vq,
1314                             int access, u64 addr, u64 len, int type)
1315 {
1316         const struct vhost_iotlb_map *map;
1317         struct vhost_iotlb *umem = vq->iotlb;
1318         u64 s = 0, size, orig_addr = addr, last = addr + len - 1;
1319
1320         if (vhost_vq_meta_fetch(vq, addr, len, type))
1321                 return true;
1322
1323         while (len > s) {
1324                 map = vhost_iotlb_itree_first(umem, addr, last);
1325                 if (map == NULL || map->start > addr) {
1326                         vhost_iotlb_miss(vq, addr, access);
1327                         return false;
1328                 } else if (!(map->perm & access)) {
1329                         /* Report the possible access violation by
1330                          * request another translation from userspace.
1331                          */
1332                         return false;
1333                 }
1334
1335                 size = map->size - addr + map->start;
1336
1337                 if (orig_addr == addr && size >= len)
1338                         vhost_vq_meta_update(vq, map, type);
1339
1340                 s += size;
1341                 addr += size;
1342         }
1343
1344         return true;
1345 }
1346
1347 int vq_meta_prefetch(struct vhost_virtqueue *vq)
1348 {
1349         unsigned int num = vq->num;
1350
1351         if (!vq->iotlb)
1352                 return 1;
1353
1354         return iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->desc,
1355                                vhost_get_desc_size(vq, num), VHOST_ADDR_DESC) &&
1356                iotlb_access_ok(vq, VHOST_MAP_RO, (u64)(uintptr_t)vq->avail,
1357                                vhost_get_avail_size(vq, num),
1358                                VHOST_ADDR_AVAIL) &&
1359                iotlb_access_ok(vq, VHOST_MAP_WO, (u64)(uintptr_t)vq->used,
1360                                vhost_get_used_size(vq, num), VHOST_ADDR_USED);
1361 }
1362 EXPORT_SYMBOL_GPL(vq_meta_prefetch);
1363
1364 /* Can we log writes? */
1365 /* Caller should have device mutex but not vq mutex */
1366 bool vhost_log_access_ok(struct vhost_dev *dev)
1367 {
1368         return memory_access_ok(dev, dev->umem, 1);
1369 }
1370 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1371
1372 static bool vq_log_used_access_ok(struct vhost_virtqueue *vq,
1373                                   void __user *log_base,
1374                                   bool log_used,
1375                                   u64 log_addr)
1376 {
1377         /* If an IOTLB device is present, log_addr is a GIOVA that
1378          * will never be logged by log_used(). */
1379         if (vq->iotlb)
1380                 return true;
1381
1382         return !log_used || log_access_ok(log_base, log_addr,
1383                                           vhost_get_used_size(vq, vq->num));
1384 }
1385
1386 /* Verify access for write logging. */
1387 /* Caller should have vq mutex and device mutex */
1388 static bool vq_log_access_ok(struct vhost_virtqueue *vq,
1389                              void __user *log_base)
1390 {
1391         return vq_memory_access_ok(log_base, vq->umem,
1392                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1393                 vq_log_used_access_ok(vq, log_base, vq->log_used, vq->log_addr);
1394 }
1395
1396 /* Can we start vq? */
1397 /* Caller should have vq mutex and device mutex */
1398 bool vhost_vq_access_ok(struct vhost_virtqueue *vq)
1399 {
1400         if (!vq_log_access_ok(vq, vq->log_base))
1401                 return false;
1402
1403         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used);
1404 }
1405 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1406
1407 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1408 {
1409         struct vhost_memory mem, *newmem;
1410         struct vhost_memory_region *region;
1411         struct vhost_iotlb *newumem, *oldumem;
1412         unsigned long size = offsetof(struct vhost_memory, regions);
1413         int i;
1414
1415         if (copy_from_user(&mem, m, size))
1416                 return -EFAULT;
1417         if (mem.padding)
1418                 return -EOPNOTSUPP;
1419         if (mem.nregions > max_mem_regions)
1420                 return -E2BIG;
1421         newmem = kvzalloc(struct_size(newmem, regions, mem.nregions),
1422                         GFP_KERNEL);
1423         if (!newmem)
1424                 return -ENOMEM;
1425
1426         memcpy(newmem, &mem, size);
1427         if (copy_from_user(newmem->regions, m->regions,
1428                            flex_array_size(newmem, regions, mem.nregions))) {
1429                 kvfree(newmem);
1430                 return -EFAULT;
1431         }
1432
1433         newumem = iotlb_alloc();
1434         if (!newumem) {
1435                 kvfree(newmem);
1436                 return -ENOMEM;
1437         }
1438
1439         for (region = newmem->regions;
1440              region < newmem->regions + mem.nregions;
1441              region++) {
1442                 if (vhost_iotlb_add_range(newumem,
1443                                           region->guest_phys_addr,
1444                                           region->guest_phys_addr +
1445                                           region->memory_size - 1,
1446                                           region->userspace_addr,
1447                                           VHOST_MAP_RW))
1448                         goto err;
1449         }
1450
1451         if (!memory_access_ok(d, newumem, 0))
1452                 goto err;
1453
1454         oldumem = d->umem;
1455         d->umem = newumem;
1456
1457         /* All memory accesses are done under some VQ mutex. */
1458         for (i = 0; i < d->nvqs; ++i) {
1459                 mutex_lock(&d->vqs[i]->mutex);
1460                 d->vqs[i]->umem = newumem;
1461                 mutex_unlock(&d->vqs[i]->mutex);
1462         }
1463
1464         kvfree(newmem);
1465         vhost_iotlb_free(oldumem);
1466         return 0;
1467
1468 err:
1469         vhost_iotlb_free(newumem);
1470         kvfree(newmem);
1471         return -EFAULT;
1472 }
1473
1474 static long vhost_vring_set_num(struct vhost_dev *d,
1475                                 struct vhost_virtqueue *vq,
1476                                 void __user *argp)
1477 {
1478         struct vhost_vring_state s;
1479
1480         /* Resizing ring with an active backend?
1481          * You don't want to do that. */
1482         if (vq->private_data)
1483                 return -EBUSY;
1484
1485         if (copy_from_user(&s, argp, sizeof s))
1486                 return -EFAULT;
1487
1488         if (!s.num || s.num > 0xffff || (s.num & (s.num - 1)))
1489                 return -EINVAL;
1490         vq->num = s.num;
1491
1492         return 0;
1493 }
1494
1495 static long vhost_vring_set_addr(struct vhost_dev *d,
1496                                  struct vhost_virtqueue *vq,
1497                                  void __user *argp)
1498 {
1499         struct vhost_vring_addr a;
1500
1501         if (copy_from_user(&a, argp, sizeof a))
1502                 return -EFAULT;
1503         if (a.flags & ~(0x1 << VHOST_VRING_F_LOG))
1504                 return -EOPNOTSUPP;
1505
1506         /* For 32bit, verify that the top 32bits of the user
1507            data are set to zero. */
1508         if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1509             (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1510             (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr)
1511                 return -EFAULT;
1512
1513         /* Make sure it's safe to cast pointers to vring types. */
1514         BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1515         BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1516         if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1517             (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1518             (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1)))
1519                 return -EINVAL;
1520
1521         /* We only verify access here if backend is configured.
1522          * If it is not, we don't as size might not have been setup.
1523          * We will verify when backend is configured. */
1524         if (vq->private_data) {
1525                 if (!vq_access_ok(vq, vq->num,
1526                         (void __user *)(unsigned long)a.desc_user_addr,
1527                         (void __user *)(unsigned long)a.avail_user_addr,
1528                         (void __user *)(unsigned long)a.used_user_addr))
1529                         return -EINVAL;
1530
1531                 /* Also validate log access for used ring if enabled. */
1532                 if (!vq_log_used_access_ok(vq, vq->log_base,
1533                                 a.flags & (0x1 << VHOST_VRING_F_LOG),
1534                                 a.log_guest_addr))
1535                         return -EINVAL;
1536         }
1537
1538         vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1539         vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1540         vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1541         vq->log_addr = a.log_guest_addr;
1542         vq->used = (void __user *)(unsigned long)a.used_user_addr;
1543
1544         return 0;
1545 }
1546
1547 static long vhost_vring_set_num_addr(struct vhost_dev *d,
1548                                      struct vhost_virtqueue *vq,
1549                                      unsigned int ioctl,
1550                                      void __user *argp)
1551 {
1552         long r;
1553
1554         mutex_lock(&vq->mutex);
1555
1556         switch (ioctl) {
1557         case VHOST_SET_VRING_NUM:
1558                 r = vhost_vring_set_num(d, vq, argp);
1559                 break;
1560         case VHOST_SET_VRING_ADDR:
1561                 r = vhost_vring_set_addr(d, vq, argp);
1562                 break;
1563         default:
1564                 BUG();
1565         }
1566
1567         mutex_unlock(&vq->mutex);
1568
1569         return r;
1570 }
1571 long vhost_vring_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1572 {
1573         struct file *eventfp, *filep = NULL;
1574         bool pollstart = false, pollstop = false;
1575         struct eventfd_ctx *ctx = NULL;
1576         u32 __user *idxp = argp;
1577         struct vhost_virtqueue *vq;
1578         struct vhost_vring_state s;
1579         struct vhost_vring_file f;
1580         u32 idx;
1581         long r;
1582
1583         r = get_user(idx, idxp);
1584         if (r < 0)
1585                 return r;
1586         if (idx >= d->nvqs)
1587                 return -ENOBUFS;
1588
1589         idx = array_index_nospec(idx, d->nvqs);
1590         vq = d->vqs[idx];
1591
1592         if (ioctl == VHOST_SET_VRING_NUM ||
1593             ioctl == VHOST_SET_VRING_ADDR) {
1594                 return vhost_vring_set_num_addr(d, vq, ioctl, argp);
1595         }
1596
1597         mutex_lock(&vq->mutex);
1598
1599         switch (ioctl) {
1600         case VHOST_SET_VRING_BASE:
1601                 /* Moving base with an active backend?
1602                  * You don't want to do that. */
1603                 if (vq->private_data) {
1604                         r = -EBUSY;
1605                         break;
1606                 }
1607                 if (copy_from_user(&s, argp, sizeof s)) {
1608                         r = -EFAULT;
1609                         break;
1610                 }
1611                 if (s.num > 0xffff) {
1612                         r = -EINVAL;
1613                         break;
1614                 }
1615                 vq->last_avail_idx = s.num;
1616                 /* Forget the cached index value. */
1617                 vq->avail_idx = vq->last_avail_idx;
1618                 break;
1619         case VHOST_GET_VRING_BASE:
1620                 s.index = idx;
1621                 s.num = vq->last_avail_idx;
1622                 if (copy_to_user(argp, &s, sizeof s))
1623                         r = -EFAULT;
1624                 break;
1625         case VHOST_SET_VRING_KICK:
1626                 if (copy_from_user(&f, argp, sizeof f)) {
1627                         r = -EFAULT;
1628                         break;
1629                 }
1630                 eventfp = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_fget(f.fd);
1631                 if (IS_ERR(eventfp)) {
1632                         r = PTR_ERR(eventfp);
1633                         break;
1634                 }
1635                 if (eventfp != vq->kick) {
1636                         pollstop = (filep = vq->kick) != NULL;
1637                         pollstart = (vq->kick = eventfp) != NULL;
1638                 } else
1639                         filep = eventfp;
1640                 break;
1641         case VHOST_SET_VRING_CALL:
1642                 if (copy_from_user(&f, argp, sizeof f)) {
1643                         r = -EFAULT;
1644                         break;
1645                 }
1646                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1647                 if (IS_ERR(ctx)) {
1648                         r = PTR_ERR(ctx);
1649                         break;
1650                 }
1651
1652                 swap(ctx, vq->call_ctx.ctx);
1653                 break;
1654         case VHOST_SET_VRING_ERR:
1655                 if (copy_from_user(&f, argp, sizeof f)) {
1656                         r = -EFAULT;
1657                         break;
1658                 }
1659                 ctx = f.fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(f.fd);
1660                 if (IS_ERR(ctx)) {
1661                         r = PTR_ERR(ctx);
1662                         break;
1663                 }
1664                 swap(ctx, vq->error_ctx);
1665                 break;
1666         case VHOST_SET_VRING_ENDIAN:
1667                 r = vhost_set_vring_endian(vq, argp);
1668                 break;
1669         case VHOST_GET_VRING_ENDIAN:
1670                 r = vhost_get_vring_endian(vq, idx, argp);
1671                 break;
1672         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1673                 if (copy_from_user(&s, argp, sizeof(s))) {
1674                         r = -EFAULT;
1675                         break;
1676                 }
1677                 vq->busyloop_timeout = s.num;
1678                 break;
1679         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1680                 s.index = idx;
1681                 s.num = vq->busyloop_timeout;
1682                 if (copy_to_user(argp, &s, sizeof(s)))
1683                         r = -EFAULT;
1684                 break;
1685         default:
1686                 r = -ENOIOCTLCMD;
1687         }
1688
1689         if (pollstop && vq->handle_kick)
1690                 vhost_poll_stop(&vq->poll);
1691
1692         if (!IS_ERR_OR_NULL(ctx))
1693                 eventfd_ctx_put(ctx);
1694         if (filep)
1695                 fput(filep);
1696
1697         if (pollstart && vq->handle_kick)
1698                 r = vhost_poll_start(&vq->poll, vq->kick);
1699
1700         mutex_unlock(&vq->mutex);
1701
1702         if (pollstop && vq->handle_kick)
1703                 vhost_poll_flush(&vq->poll);
1704         return r;
1705 }
1706 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1707
1708 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1709 {
1710         struct vhost_iotlb *niotlb, *oiotlb;
1711         int i;
1712
1713         niotlb = iotlb_alloc();
1714         if (!niotlb)
1715                 return -ENOMEM;
1716
1717         oiotlb = d->iotlb;
1718         d->iotlb = niotlb;
1719
1720         for (i = 0; i < d->nvqs; ++i) {
1721                 struct vhost_virtqueue *vq = d->vqs[i];
1722
1723                 mutex_lock(&vq->mutex);
1724                 vq->iotlb = niotlb;
1725                 __vhost_vq_meta_reset(vq);
1726                 mutex_unlock(&vq->mutex);
1727         }
1728
1729         vhost_iotlb_free(oiotlb);
1730
1731         return 0;
1732 }
1733 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1734
1735 /* Caller must have device mutex */
1736 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1737 {
1738         struct eventfd_ctx *ctx;
1739         u64 p;
1740         long r;
1741         int i, fd;
1742
1743         /* If you are not the owner, you can become one */
1744         if (ioctl == VHOST_SET_OWNER) {
1745                 r = vhost_dev_set_owner(d);
1746                 goto done;
1747         }
1748
1749         /* You must be the owner to do anything else */
1750         r = vhost_dev_check_owner(d);
1751         if (r)
1752                 goto done;
1753
1754         switch (ioctl) {
1755         case VHOST_SET_MEM_TABLE:
1756                 r = vhost_set_memory(d, argp);
1757                 break;
1758         case VHOST_SET_LOG_BASE:
1759                 if (copy_from_user(&p, argp, sizeof p)) {
1760                         r = -EFAULT;
1761                         break;
1762                 }
1763                 if ((u64)(unsigned long)p != p) {
1764                         r = -EFAULT;
1765                         break;
1766                 }
1767                 for (i = 0; i < d->nvqs; ++i) {
1768                         struct vhost_virtqueue *vq;
1769                         void __user *base = (void __user *)(unsigned long)p;
1770                         vq = d->vqs[i];
1771                         mutex_lock(&vq->mutex);
1772                         /* If ring is inactive, will check when it's enabled. */
1773                         if (vq->private_data && !vq_log_access_ok(vq, base))
1774                                 r = -EFAULT;
1775                         else
1776                                 vq->log_base = base;
1777                         mutex_unlock(&vq->mutex);
1778                 }
1779                 break;
1780         case VHOST_SET_LOG_FD:
1781                 r = get_user(fd, (int __user *)argp);
1782                 if (r < 0)
1783                         break;
1784                 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
1785                 if (IS_ERR(ctx)) {
1786                         r = PTR_ERR(ctx);
1787                         break;
1788                 }
1789                 swap(ctx, d->log_ctx);
1790                 for (i = 0; i < d->nvqs; ++i) {
1791                         mutex_lock(&d->vqs[i]->mutex);
1792                         d->vqs[i]->log_ctx = d->log_ctx;
1793                         mutex_unlock(&d->vqs[i]->mutex);
1794                 }
1795                 if (ctx)
1796                         eventfd_ctx_put(ctx);
1797                 break;
1798         default:
1799                 r = -ENOIOCTLCMD;
1800                 break;
1801         }
1802 done:
1803         return r;
1804 }
1805 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1806
1807 /* TODO: This is really inefficient.  We need something like get_user()
1808  * (instruction directly accesses the data, with an exception table entry
1809  * returning -EFAULT). See Documentation/x86/exception-tables.rst.
1810  */
1811 static int set_bit_to_user(int nr, void __user *addr)
1812 {
1813         unsigned long log = (unsigned long)addr;
1814         struct page *page;
1815         void *base;
1816         int bit = nr + (log % PAGE_SIZE) * 8;
1817         int r;
1818
1819         r = pin_user_pages_fast(log, 1, FOLL_WRITE, &page);
1820         if (r < 0)
1821                 return r;
1822         BUG_ON(r != 1);
1823         base = kmap_atomic(page);
1824         set_bit(bit, base);
1825         kunmap_atomic(base);
1826         unpin_user_pages_dirty_lock(&page, 1, true);
1827         return 0;
1828 }
1829
1830 static int log_write(void __user *log_base,
1831                      u64 write_address, u64 write_length)
1832 {
1833         u64 write_page = write_address / VHOST_PAGE_SIZE;
1834         int r;
1835
1836         if (!write_length)
1837                 return 0;
1838         write_length += write_address % VHOST_PAGE_SIZE;
1839         for (;;) {
1840                 u64 base = (u64)(unsigned long)log_base;
1841                 u64 log = base + write_page / 8;
1842                 int bit = write_page % 8;
1843                 if ((u64)(unsigned long)log != log)
1844                         return -EFAULT;
1845                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1846                 if (r < 0)
1847                         return r;
1848                 if (write_length <= VHOST_PAGE_SIZE)
1849                         break;
1850                 write_length -= VHOST_PAGE_SIZE;
1851                 write_page += 1;
1852         }
1853         return r;
1854 }
1855
1856 static int log_write_hva(struct vhost_virtqueue *vq, u64 hva, u64 len)
1857 {
1858         struct vhost_iotlb *umem = vq->umem;
1859         struct vhost_iotlb_map *u;
1860         u64 start, end, l, min;
1861         int r;
1862         bool hit = false;
1863
1864         while (len) {
1865                 min = len;
1866                 /* More than one GPAs can be mapped into a single HVA. So
1867                  * iterate all possible umems here to be safe.
1868                  */
1869                 list_for_each_entry(u, &umem->list, link) {
1870                         if (u->addr > hva - 1 + len ||
1871                             u->addr - 1 + u->size < hva)
1872                                 continue;
1873                         start = max(u->addr, hva);
1874                         end = min(u->addr - 1 + u->size, hva - 1 + len);
1875                         l = end - start + 1;
1876                         r = log_write(vq->log_base,
1877                                       u->start + start - u->addr,
1878                                       l);
1879                         if (r < 0)
1880                                 return r;
1881                         hit = true;
1882                         min = min(l, min);
1883                 }
1884
1885                 if (!hit)
1886                         return -EFAULT;
1887
1888                 len -= min;
1889                 hva += min;
1890         }
1891
1892         return 0;
1893 }
1894
1895 static int log_used(struct vhost_virtqueue *vq, u64 used_offset, u64 len)
1896 {
1897         struct iovec *iov = vq->log_iov;
1898         int i, ret;
1899
1900         if (!vq->iotlb)
1901                 return log_write(vq->log_base, vq->log_addr + used_offset, len);
1902
1903         ret = translate_desc(vq, (uintptr_t)vq->used + used_offset,
1904                              len, iov, 64, VHOST_ACCESS_WO);
1905         if (ret < 0)
1906                 return ret;
1907
1908         for (i = 0; i < ret; i++) {
1909                 ret = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1910                                     iov[i].iov_len);
1911                 if (ret)
1912                         return ret;
1913         }
1914
1915         return 0;
1916 }
1917
1918 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1919                     unsigned int log_num, u64 len, struct iovec *iov, int count)
1920 {
1921         int i, r;
1922
1923         /* Make sure data written is seen before log. */
1924         smp_wmb();
1925
1926         if (vq->iotlb) {
1927                 for (i = 0; i < count; i++) {
1928                         r = log_write_hva(vq, (uintptr_t)iov[i].iov_base,
1929                                           iov[i].iov_len);
1930                         if (r < 0)
1931                                 return r;
1932                 }
1933                 return 0;
1934         }
1935
1936         for (i = 0; i < log_num; ++i) {
1937                 u64 l = min(log[i].len, len);
1938                 r = log_write(vq->log_base, log[i].addr, l);
1939                 if (r < 0)
1940                         return r;
1941                 len -= l;
1942                 if (!len) {
1943                         if (vq->log_ctx)
1944                                 eventfd_signal(vq->log_ctx, 1);
1945                         return 0;
1946                 }
1947         }
1948         /* Length written exceeds what we have stored. This is a bug. */
1949         BUG();
1950         return 0;
1951 }
1952 EXPORT_SYMBOL_GPL(vhost_log_write);
1953
1954 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1955 {
1956         void __user *used;
1957         if (vhost_put_used_flags(vq))
1958                 return -EFAULT;
1959         if (unlikely(vq->log_used)) {
1960                 /* Make sure the flag is seen before log. */
1961                 smp_wmb();
1962                 /* Log used flag write. */
1963                 used = &vq->used->flags;
1964                 log_used(vq, (used - (void __user *)vq->used),
1965                          sizeof vq->used->flags);
1966                 if (vq->log_ctx)
1967                         eventfd_signal(vq->log_ctx, 1);
1968         }
1969         return 0;
1970 }
1971
1972 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1973 {
1974         if (vhost_put_avail_event(vq))
1975                 return -EFAULT;
1976         if (unlikely(vq->log_used)) {
1977                 void __user *used;
1978                 /* Make sure the event is seen before log. */
1979                 smp_wmb();
1980                 /* Log avail event write */
1981                 used = vhost_avail_event(vq);
1982                 log_used(vq, (used - (void __user *)vq->used),
1983                          sizeof *vhost_avail_event(vq));
1984                 if (vq->log_ctx)
1985                         eventfd_signal(vq->log_ctx, 1);
1986         }
1987         return 0;
1988 }
1989
1990 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1991 {
1992         __virtio16 last_used_idx;
1993         int r;
1994         bool is_le = vq->is_le;
1995
1996         if (!vq->private_data)
1997                 return 0;
1998
1999         vhost_init_is_le(vq);
2000
2001         r = vhost_update_used_flags(vq);
2002         if (r)
2003                 goto err;
2004         vq->signalled_used_valid = false;
2005         if (!vq->iotlb &&
2006             !access_ok(&vq->used->idx, sizeof vq->used->idx)) {
2007                 r = -EFAULT;
2008                 goto err;
2009         }
2010         r = vhost_get_used_idx(vq, &last_used_idx);
2011         if (r) {
2012                 vq_err(vq, "Can't access used idx at %p\n",
2013                        &vq->used->idx);
2014                 goto err;
2015         }
2016         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
2017         return 0;
2018
2019 err:
2020         vq->is_le = is_le;
2021         return r;
2022 }
2023 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
2024
2025 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
2026                           struct iovec iov[], int iov_size, int access)
2027 {
2028         const struct vhost_iotlb_map *map;
2029         struct vhost_dev *dev = vq->dev;
2030         struct vhost_iotlb *umem = dev->iotlb ? dev->iotlb : dev->umem;
2031         struct iovec *_iov;
2032         u64 s = 0;
2033         int ret = 0;
2034
2035         while ((u64)len > s) {
2036                 u64 size;
2037                 if (unlikely(ret >= iov_size)) {
2038                         ret = -ENOBUFS;
2039                         break;
2040                 }
2041
2042                 map = vhost_iotlb_itree_first(umem, addr, addr + len - 1);
2043                 if (map == NULL || map->start > addr) {
2044                         if (umem != dev->iotlb) {
2045                                 ret = -EFAULT;
2046                                 break;
2047                         }
2048                         ret = -EAGAIN;
2049                         break;
2050                 } else if (!(map->perm & access)) {
2051                         ret = -EPERM;
2052                         break;
2053                 }
2054
2055                 _iov = iov + ret;
2056                 size = map->size - addr + map->start;
2057                 _iov->iov_len = min((u64)len - s, size);
2058                 _iov->iov_base = (void __user *)(unsigned long)
2059                                  (map->addr + addr - map->start);
2060                 s += size;
2061                 addr += size;
2062                 ++ret;
2063         }
2064
2065         if (ret == -EAGAIN)
2066                 vhost_iotlb_miss(vq, addr, access);
2067         return ret;
2068 }
2069
2070 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
2071  * function returns the next descriptor in the chain,
2072  * or -1U if we're at the end. */
2073 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
2074 {
2075         unsigned int next;
2076
2077         /* If this descriptor says it doesn't chain, we're done. */
2078         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
2079                 return -1U;
2080
2081         /* Check they're not leading us off end of descriptors. */
2082         next = vhost16_to_cpu(vq, READ_ONCE(desc->next));
2083         return next;
2084 }
2085
2086 static int get_indirect(struct vhost_virtqueue *vq,
2087                         struct iovec iov[], unsigned int iov_size,
2088                         unsigned int *out_num, unsigned int *in_num,
2089                         struct vhost_log *log, unsigned int *log_num,
2090                         struct vring_desc *indirect)
2091 {
2092         struct vring_desc desc;
2093         unsigned int i = 0, count, found = 0;
2094         u32 len = vhost32_to_cpu(vq, indirect->len);
2095         struct iov_iter from;
2096         int ret, access;
2097
2098         /* Sanity check */
2099         if (unlikely(len % sizeof desc)) {
2100                 vq_err(vq, "Invalid length in indirect descriptor: "
2101                        "len 0x%llx not multiple of 0x%zx\n",
2102                        (unsigned long long)len,
2103                        sizeof desc);
2104                 return -EINVAL;
2105         }
2106
2107         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
2108                              UIO_MAXIOV, VHOST_ACCESS_RO);
2109         if (unlikely(ret < 0)) {
2110                 if (ret != -EAGAIN)
2111                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
2112                 return ret;
2113         }
2114         iov_iter_init(&from, READ, vq->indirect, ret, len);
2115         count = len / sizeof desc;
2116         /* Buffers are chained via a 16 bit next field, so
2117          * we can have at most 2^16 of these. */
2118         if (unlikely(count > USHRT_MAX + 1)) {
2119                 vq_err(vq, "Indirect buffer length too big: %d\n",
2120                        indirect->len);
2121                 return -E2BIG;
2122         }
2123
2124         do {
2125                 unsigned iov_count = *in_num + *out_num;
2126                 if (unlikely(++found > count)) {
2127                         vq_err(vq, "Loop detected: last one at %u "
2128                                "indirect size %u\n",
2129                                i, count);
2130                         return -EINVAL;
2131                 }
2132                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
2133                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
2134                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2135                         return -EINVAL;
2136                 }
2137                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
2138                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
2139                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
2140                         return -EINVAL;
2141                 }
2142
2143                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2144                         access = VHOST_ACCESS_WO;
2145                 else
2146                         access = VHOST_ACCESS_RO;
2147
2148                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2149                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2150                                      iov_size - iov_count, access);
2151                 if (unlikely(ret < 0)) {
2152                         if (ret != -EAGAIN)
2153                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
2154                                         ret, i);
2155                         return ret;
2156                 }
2157                 /* If this is an input descriptor, increment that count. */
2158                 if (access == VHOST_ACCESS_WO) {
2159                         *in_num += ret;
2160                         if (unlikely(log && ret)) {
2161                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2162                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2163                                 ++*log_num;
2164                         }
2165                 } else {
2166                         /* If it's an output descriptor, they're all supposed
2167                          * to come before any input descriptors. */
2168                         if (unlikely(*in_num)) {
2169                                 vq_err(vq, "Indirect descriptor "
2170                                        "has out after in: idx %d\n", i);
2171                                 return -EINVAL;
2172                         }
2173                         *out_num += ret;
2174                 }
2175         } while ((i = next_desc(vq, &desc)) != -1);
2176         return 0;
2177 }
2178
2179 /* This looks in the virtqueue and for the first available buffer, and converts
2180  * it to an iovec for convenient access.  Since descriptors consist of some
2181  * number of output then some number of input descriptors, it's actually two
2182  * iovecs, but we pack them into one and note how many of each there were.
2183  *
2184  * This function returns the descriptor number found, or vq->num (which is
2185  * never a valid descriptor number) if none was found.  A negative code is
2186  * returned on error. */
2187 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2188                       struct iovec iov[], unsigned int iov_size,
2189                       unsigned int *out_num, unsigned int *in_num,
2190                       struct vhost_log *log, unsigned int *log_num)
2191 {
2192         struct vring_desc desc;
2193         unsigned int i, head, found = 0;
2194         u16 last_avail_idx;
2195         __virtio16 avail_idx;
2196         __virtio16 ring_head;
2197         int ret, access;
2198
2199         /* Check it isn't doing very strange things with descriptor numbers. */
2200         last_avail_idx = vq->last_avail_idx;
2201
2202         if (vq->avail_idx == vq->last_avail_idx) {
2203                 if (unlikely(vhost_get_avail_idx(vq, &avail_idx))) {
2204                         vq_err(vq, "Failed to access avail idx at %p\n",
2205                                 &vq->avail->idx);
2206                         return -EFAULT;
2207                 }
2208                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2209
2210                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2211                         vq_err(vq, "Guest moved used index from %u to %u",
2212                                 last_avail_idx, vq->avail_idx);
2213                         return -EFAULT;
2214                 }
2215
2216                 /* If there's nothing new since last we looked, return
2217                  * invalid.
2218                  */
2219                 if (vq->avail_idx == last_avail_idx)
2220                         return vq->num;
2221
2222                 /* Only get avail ring entries after they have been
2223                  * exposed by guest.
2224                  */
2225                 smp_rmb();
2226         }
2227
2228         /* Grab the next descriptor number they're advertising, and increment
2229          * the index we've seen. */
2230         if (unlikely(vhost_get_avail_head(vq, &ring_head, last_avail_idx))) {
2231                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2232                        last_avail_idx,
2233                        &vq->avail->ring[last_avail_idx % vq->num]);
2234                 return -EFAULT;
2235         }
2236
2237         head = vhost16_to_cpu(vq, ring_head);
2238
2239         /* If their number is silly, that's an error. */
2240         if (unlikely(head >= vq->num)) {
2241                 vq_err(vq, "Guest says index %u > %u is available",
2242                        head, vq->num);
2243                 return -EINVAL;
2244         }
2245
2246         /* When we start there are none of either input nor output. */
2247         *out_num = *in_num = 0;
2248         if (unlikely(log))
2249                 *log_num = 0;
2250
2251         i = head;
2252         do {
2253                 unsigned iov_count = *in_num + *out_num;
2254                 if (unlikely(i >= vq->num)) {
2255                         vq_err(vq, "Desc index is %u > %u, head = %u",
2256                                i, vq->num, head);
2257                         return -EINVAL;
2258                 }
2259                 if (unlikely(++found > vq->num)) {
2260                         vq_err(vq, "Loop detected: last one at %u "
2261                                "vq size %u head %u\n",
2262                                i, vq->num, head);
2263                         return -EINVAL;
2264                 }
2265                 ret = vhost_get_desc(vq, &desc, i);
2266                 if (unlikely(ret)) {
2267                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2268                                i, vq->desc + i);
2269                         return -EFAULT;
2270                 }
2271                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2272                         ret = get_indirect(vq, iov, iov_size,
2273                                            out_num, in_num,
2274                                            log, log_num, &desc);
2275                         if (unlikely(ret < 0)) {
2276                                 if (ret != -EAGAIN)
2277                                         vq_err(vq, "Failure detected "
2278                                                 "in indirect descriptor at idx %d\n", i);
2279                                 return ret;
2280                         }
2281                         continue;
2282                 }
2283
2284                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2285                         access = VHOST_ACCESS_WO;
2286                 else
2287                         access = VHOST_ACCESS_RO;
2288                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2289                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2290                                      iov_size - iov_count, access);
2291                 if (unlikely(ret < 0)) {
2292                         if (ret != -EAGAIN)
2293                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2294                                         ret, i);
2295                         return ret;
2296                 }
2297                 if (access == VHOST_ACCESS_WO) {
2298                         /* If this is an input descriptor,
2299                          * increment that count. */
2300                         *in_num += ret;
2301                         if (unlikely(log && ret)) {
2302                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2303                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2304                                 ++*log_num;
2305                         }
2306                 } else {
2307                         /* If it's an output descriptor, they're all supposed
2308                          * to come before any input descriptors. */
2309                         if (unlikely(*in_num)) {
2310                                 vq_err(vq, "Descriptor has out after in: "
2311                                        "idx %d\n", i);
2312                                 return -EINVAL;
2313                         }
2314                         *out_num += ret;
2315                 }
2316         } while ((i = next_desc(vq, &desc)) != -1);
2317
2318         /* On success, increment avail index. */
2319         vq->last_avail_idx++;
2320
2321         /* Assume notifications from guest are disabled at this point,
2322          * if they aren't we would need to update avail_event index. */
2323         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2324         return head;
2325 }
2326 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2327
2328 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2329 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2330 {
2331         vq->last_avail_idx -= n;
2332 }
2333 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2334
2335 /* After we've used one of their buffers, we tell them about it.  We'll then
2336  * want to notify the guest, using eventfd. */
2337 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2338 {
2339         struct vring_used_elem heads = {
2340                 cpu_to_vhost32(vq, head),
2341                 cpu_to_vhost32(vq, len)
2342         };
2343
2344         return vhost_add_used_n(vq, &heads, 1);
2345 }
2346 EXPORT_SYMBOL_GPL(vhost_add_used);
2347
2348 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2349                             struct vring_used_elem *heads,
2350                             unsigned count)
2351 {
2352         vring_used_elem_t __user *used;
2353         u16 old, new;
2354         int start;
2355
2356         start = vq->last_used_idx & (vq->num - 1);
2357         used = vq->used->ring + start;
2358         if (vhost_put_used(vq, heads, start, count)) {
2359                 vq_err(vq, "Failed to write used");
2360                 return -EFAULT;
2361         }
2362         if (unlikely(vq->log_used)) {
2363                 /* Make sure data is seen before log. */
2364                 smp_wmb();
2365                 /* Log used ring entry write. */
2366                 log_used(vq, ((void __user *)used - (void __user *)vq->used),
2367                          count * sizeof *used);
2368         }
2369         old = vq->last_used_idx;
2370         new = (vq->last_used_idx += count);
2371         /* If the driver never bothers to signal in a very long while,
2372          * used index might wrap around. If that happens, invalidate
2373          * signalled_used index we stored. TODO: make sure driver
2374          * signals at least once in 2^16 and remove this. */
2375         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2376                 vq->signalled_used_valid = false;
2377         return 0;
2378 }
2379
2380 /* After we've used one of their buffers, we tell them about it.  We'll then
2381  * want to notify the guest, using eventfd. */
2382 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2383                      unsigned count)
2384 {
2385         int start, n, r;
2386
2387         start = vq->last_used_idx & (vq->num - 1);
2388         n = vq->num - start;
2389         if (n < count) {
2390                 r = __vhost_add_used_n(vq, heads, n);
2391                 if (r < 0)
2392                         return r;
2393                 heads += n;
2394                 count -= n;
2395         }
2396         r = __vhost_add_used_n(vq, heads, count);
2397
2398         /* Make sure buffer is written before we update index. */
2399         smp_wmb();
2400         if (vhost_put_used_idx(vq)) {
2401                 vq_err(vq, "Failed to increment used idx");
2402                 return -EFAULT;
2403         }
2404         if (unlikely(vq->log_used)) {
2405                 /* Make sure used idx is seen before log. */
2406                 smp_wmb();
2407                 /* Log used index update. */
2408                 log_used(vq, offsetof(struct vring_used, idx),
2409                          sizeof vq->used->idx);
2410                 if (vq->log_ctx)
2411                         eventfd_signal(vq->log_ctx, 1);
2412         }
2413         return r;
2414 }
2415 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2416
2417 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2418 {
2419         __u16 old, new;
2420         __virtio16 event;
2421         bool v;
2422         /* Flush out used index updates. This is paired
2423          * with the barrier that the Guest executes when enabling
2424          * interrupts. */
2425         smp_mb();
2426
2427         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2428             unlikely(vq->avail_idx == vq->last_avail_idx))
2429                 return true;
2430
2431         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2432                 __virtio16 flags;
2433                 if (vhost_get_avail_flags(vq, &flags)) {
2434                         vq_err(vq, "Failed to get flags");
2435                         return true;
2436                 }
2437                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2438         }
2439         old = vq->signalled_used;
2440         v = vq->signalled_used_valid;
2441         new = vq->signalled_used = vq->last_used_idx;
2442         vq->signalled_used_valid = true;
2443
2444         if (unlikely(!v))
2445                 return true;
2446
2447         if (vhost_get_used_event(vq, &event)) {
2448                 vq_err(vq, "Failed to get used event idx");
2449                 return true;
2450         }
2451         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
2452 }
2453
2454 /* This actually signals the guest, using eventfd. */
2455 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2456 {
2457         /* Signal the Guest tell them we used something up. */
2458         if (vq->call_ctx.ctx && vhost_notify(dev, vq))
2459                 eventfd_signal(vq->call_ctx.ctx, 1);
2460 }
2461 EXPORT_SYMBOL_GPL(vhost_signal);
2462
2463 /* And here's the combo meal deal.  Supersize me! */
2464 void vhost_add_used_and_signal(struct vhost_dev *dev,
2465                                struct vhost_virtqueue *vq,
2466                                unsigned int head, int len)
2467 {
2468         vhost_add_used(vq, head, len);
2469         vhost_signal(dev, vq);
2470 }
2471 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2472
2473 /* multi-buffer version of vhost_add_used_and_signal */
2474 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2475                                  struct vhost_virtqueue *vq,
2476                                  struct vring_used_elem *heads, unsigned count)
2477 {
2478         vhost_add_used_n(vq, heads, count);
2479         vhost_signal(dev, vq);
2480 }
2481 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2482
2483 /* return true if we're sure that avaiable ring is empty */
2484 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2485 {
2486         __virtio16 avail_idx;
2487         int r;
2488
2489         if (vq->avail_idx != vq->last_avail_idx)
2490                 return false;
2491
2492         r = vhost_get_avail_idx(vq, &avail_idx);
2493         if (unlikely(r))
2494                 return false;
2495         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2496
2497         return vq->avail_idx == vq->last_avail_idx;
2498 }
2499 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2500
2501 /* OK, now we need to know about added descriptors. */
2502 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2503 {
2504         __virtio16 avail_idx;
2505         int r;
2506
2507         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2508                 return false;
2509         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2510         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2511                 r = vhost_update_used_flags(vq);
2512                 if (r) {
2513                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2514                                &vq->used->flags, r);
2515                         return false;
2516                 }
2517         } else {
2518                 r = vhost_update_avail_event(vq, vq->avail_idx);
2519                 if (r) {
2520                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2521                                vhost_avail_event(vq), r);
2522                         return false;
2523                 }
2524         }
2525         /* They could have slipped one in as we were doing that: make
2526          * sure it's written, then check again. */
2527         smp_mb();
2528         r = vhost_get_avail_idx(vq, &avail_idx);
2529         if (r) {
2530                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2531                        &vq->avail->idx, r);
2532                 return false;
2533         }
2534
2535         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2536 }
2537 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2538
2539 /* We don't need to be notified again. */
2540 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2541 {
2542         int r;
2543
2544         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2545                 return;
2546         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2547         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2548                 r = vhost_update_used_flags(vq);
2549                 if (r)
2550                         vq_err(vq, "Failed to disable notification at %p: %d\n",
2551                                &vq->used->flags, r);
2552         }
2553 }
2554 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2555
2556 /* Create a new message. */
2557 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2558 {
2559         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2560         if (!node)
2561                 return NULL;
2562
2563         /* Make sure all padding within the structure is initialized. */
2564         memset(&node->msg, 0, sizeof node->msg);
2565         node->vq = vq;
2566         node->msg.type = type;
2567         return node;
2568 }
2569 EXPORT_SYMBOL_GPL(vhost_new_msg);
2570
2571 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2572                        struct vhost_msg_node *node)
2573 {
2574         spin_lock(&dev->iotlb_lock);
2575         list_add_tail(&node->node, head);
2576         spin_unlock(&dev->iotlb_lock);
2577
2578         wake_up_interruptible_poll(&dev->wait, EPOLLIN | EPOLLRDNORM);
2579 }
2580 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2581
2582 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2583                                          struct list_head *head)
2584 {
2585         struct vhost_msg_node *node = NULL;
2586
2587         spin_lock(&dev->iotlb_lock);
2588         if (!list_empty(head)) {
2589                 node = list_first_entry(head, struct vhost_msg_node,
2590                                         node);
2591                 list_del(&node->node);
2592         }
2593         spin_unlock(&dev->iotlb_lock);
2594
2595         return node;
2596 }
2597 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2598
2599 void vhost_set_backend_features(struct vhost_dev *dev, u64 features)
2600 {
2601         struct vhost_virtqueue *vq;
2602         int i;
2603
2604         mutex_lock(&dev->mutex);
2605         for (i = 0; i < dev->nvqs; ++i) {
2606                 vq = dev->vqs[i];
2607                 mutex_lock(&vq->mutex);
2608                 vq->acked_backend_features = features;
2609                 mutex_unlock(&vq->mutex);
2610         }
2611         mutex_unlock(&dev->mutex);
2612 }
2613 EXPORT_SYMBOL_GPL(vhost_set_backend_features);
2614
2615 static int __init vhost_init(void)
2616 {
2617         return 0;
2618 }
2619
2620 static void __exit vhost_exit(void)
2621 {
2622 }
2623
2624 module_init(vhost_init);
2625 module_exit(vhost_exit);
2626
2627 MODULE_VERSION("0.0.1");
2628 MODULE_LICENSE("GPL v2");
2629 MODULE_AUTHOR("Michael S. Tsirkin");
2630 MODULE_DESCRIPTION("Host kernel accelerator for virtio");