drivers/vhost/vhost.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Copyright (C) 2006 Rusty Russell IBM Corporation
   3  *
   4  * Author: Michael S. Tsirkin <mst@redhat.com>
   5  *
   6  * Inspiration, some code, and most witty comments come from
   7  * Documentation/virtual/lguest/lguest.c, by Rusty Russell
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.
  10  *
  11  * Generic code for virtio server in host kernel.
  12  */
  13
  14 #include <linux/eventfd.h>
  15 #include <linux/vhost.h>
  16 #include <linux/uio.h>
  17 #include <linux/mm.h>
  18 #include <linux/mmu_context.h>
  19 #include <linux/miscdevice.h>
  20 #include <linux/mutex.h>
  21 #include <linux/poll.h>
  22 #include <linux/file.h>
  23 #include <linux/highmem.h>
  24 #include <linux/slab.h>
  25 #include <linux/vmalloc.h>
  26 #include <linux/kthread.h>
  27 #include <linux/cgroup.h>
  28 #include <linux/module.h>
  29 #include <linux/sort.h>
  30 #include <linux/sched/mm.h>
  31 #include <linux/sched/signal.h>
  32 #include <linux/interval_tree_generic.h>
  33
  34 #include "vhost.h"
  35
  36 static ushort max_mem_regions = 64;
  37 module_param(max_mem_regions, ushort, 0444);
  38 MODULE_PARM_DESC(max_mem_regions,
  39         "Maximum number of memory regions in memory map. (default: 64)");
  40 static int max_iotlb_entries = 2048;
  41 module_param(max_iotlb_entries, int, 0444);
  42 MODULE_PARM_DESC(max_iotlb_entries,
  43         "Maximum number of iotlb entries. (default: 2048)");
  44
  45 enum {
  46         VHOST_MEMORY_F_LOG = 0x1,
  47 };
  48
  49 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  50 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  51
  52 INTERVAL_TREE_DEFINE(struct vhost_umem_node,
  53                      rb, __u64, __subtree_last,
  54                      START, LAST, static inline, vhost_umem_interval_tree);
  55
  56 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  57 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  58 {
  59         vq->user_be = !virtio_legacy_is_little_endian();
  60 }
  61
  62 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  63 {
  64         vq->user_be = true;
  65 }
  66
  67 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  68 {
  69         vq->user_be = false;
  70 }
  71
  72 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  73 {
  74         struct vhost_vring_state s;
  75
  76         if (vq->private_data)
  77                 return -EBUSY;
  78
  79         if (copy_from_user(&s, argp, sizeof(s)))
  80                 return -EFAULT;
  81
  82         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  83             s.num != VHOST_VRING_BIG_ENDIAN)
  84                 return -EINVAL;
  85
  86         if (s.num == VHOST_VRING_BIG_ENDIAN)
  87                 vhost_enable_cross_endian_big(vq);
  88         else
  89                 vhost_enable_cross_endian_little(vq);
  90
  91         return 0;
  92 }
  93
  94 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  95                                    int __user *argp)
  96 {
  97         struct vhost_vring_state s = {
  98                 .index = idx,
  99                 .num = vq->user_be
 100         };
 101
 102         if (copy_to_user(argp, &s, sizeof(s)))
 103                 return -EFAULT;
 104
 105         return 0;
 106 }
 107
 108 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 109 {
 110         /* Note for legacy virtio: user_be is initialized at reset time
 111          * according to the host endianness. If userspace does not set an
 112          * explicit endianness, the default behavior is native endian, as
 113          * expected by legacy virtio.
 114          */
 115         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 116 }
 117 #else
 118 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 119 {
 120 }
 121
 122 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 123 {
 124         return -ENOIOCTLCMD;
 125 }
 126
 127 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 128                                    int __user *argp)
 129 {
 130         return -ENOIOCTLCMD;
 131 }
 132
 133 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 134 {
 135         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1)
 136                 || virtio_legacy_is_little_endian();
 137 }
 138 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 139
 140 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 141 {
 142         vhost_init_is_le(vq);
 143 }
 144
 145 struct vhost_flush_struct {
 146         struct vhost_work work;
 147         struct completion wait_event;
 148 };
 149
 150 static void vhost_flush_work(struct vhost_work *work)
 151 {
 152         struct vhost_flush_struct *s;
 153
 154         s = container_of(work, struct vhost_flush_struct, work);
 155         complete(&s->wait_event);
 156 }
 157
 158 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 159                             poll_table *pt)
 160 {
 161         struct vhost_poll *poll;
 162
 163         poll = container_of(pt, struct vhost_poll, table);
 164         poll->wqh = wqh;
 165         add_wait_queue(wqh, &poll->wait);
 166 }
 167
 168 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
 169                              void *key)
 170 {
 171         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 172
 173         if (!((unsigned long)key & poll->mask))
 174                 return 0;
 175
 176         vhost_poll_queue(poll);
 177         return 0;
 178 }
 179
 180 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 181 {
 182         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 183         work->fn = fn;
 184         init_waitqueue_head(&work->done);
 185 }
 186 EXPORT_SYMBOL_GPL(vhost_work_init);
 187
 188 /* Init poll structure */
 189 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 190                      unsigned long mask, struct vhost_dev *dev)
 191 {
 192         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 193         init_poll_funcptr(&poll->table, vhost_poll_func);
 194         poll->mask = mask;
 195         poll->dev = dev;
 196         poll->wqh = NULL;
 197
 198         vhost_work_init(&poll->work, fn);
 199 }
 200 EXPORT_SYMBOL_GPL(vhost_poll_init);
 201
 202 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 203  * keep a reference to a file until after vhost_poll_stop is called. */
 204 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 205 {
 206         unsigned long mask;
 207         int ret = 0;
 208
 209         if (poll->wqh)
 210                 return 0;
 211
 212         mask = file->f_op->poll(file, &poll->table);
 213         if (mask)
 214                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 215         if (mask & POLLERR) {
 216                 if (poll->wqh)
 217                         remove_wait_queue(poll->wqh, &poll->wait);
 218                 ret = -EINVAL;
 219         }
 220
 221         return ret;
 222 }
 223 EXPORT_SYMBOL_GPL(vhost_poll_start);
 224
 225 /* Stop polling a file. After this function returns, it becomes safe to drop the
 226  * file reference. You must also flush afterwards. */
 227 void vhost_poll_stop(struct vhost_poll *poll)
 228 {
 229         if (poll->wqh) {
 230                 remove_wait_queue(poll->wqh, &poll->wait);
 231                 poll->wqh = NULL;
 232         }
 233 }
 234 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 235
 236 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 237 {
 238         struct vhost_flush_struct flush;
 239
 240         if (dev->worker) {
 241                 init_completion(&flush.wait_event);
 242                 vhost_work_init(&flush.work, vhost_flush_work);
 243
 244                 vhost_work_queue(dev, &flush.work);
 245                 wait_for_completion(&flush.wait_event);
 246         }
 247 }
 248 EXPORT_SYMBOL_GPL(vhost_work_flush);
 249
 250 /* Flush any work that has been scheduled. When calling this, don't hold any
 251  * locks that are also used by the callback. */
 252 void vhost_poll_flush(struct vhost_poll *poll)
 253 {
 254         vhost_work_flush(poll->dev, &poll->work);
 255 }
 256 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 257
 258 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 259 {
 260         if (!dev->worker)
 261                 return;
 262
 263         if (!test_and_set_bit(VHOST_WORK_QUEUED, &work->flags)) {
 264                 /* We can only add the work to the list after we're
 265                  * sure it was not in the list.
 266                  * test_and_set_bit() implies a memory barrier.
 267                  */
 268                 llist_add(&work->node, &dev->work_list);
 269                 wake_up_process(dev->worker);
 270         }
 271 }
 272 EXPORT_SYMBOL_GPL(vhost_work_queue);
 273
 274 /* A lockless hint for busy polling code to exit the loop */
 275 bool vhost_has_work(struct vhost_dev *dev)
 276 {
 277         return !llist_empty(&dev->work_list);
 278 }
 279 EXPORT_SYMBOL_GPL(vhost_has_work);
 280
 281 void vhost_poll_queue(struct vhost_poll *poll)
 282 {
 283         vhost_work_queue(poll->dev, &poll->work);
 284 }
 285 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 286
 287 static void __vhost_vq_meta_reset(struct vhost_virtqueue *vq)
 288 {
 289         int j;
 290
 291         for (j = 0; j < VHOST_NUM_ADDRS; j++)
 292                 vq->meta_iotlb[j] = NULL;
 293 }
 294
 295 static void vhost_vq_meta_reset(struct vhost_dev *d)
 296 {
 297         int i;
 298
 299         for (i = 0; i < d->nvqs; ++i)
 300                 __vhost_vq_meta_reset(d->vqs[i]);
 301 }
 302
 303 static void vhost_vq_reset(struct vhost_dev *dev,
 304                            struct vhost_virtqueue *vq)
 305 {
 306         vq->num = 1;
 307         vq->desc = NULL;
 308         vq->avail = NULL;
 309         vq->used = NULL;
 310         vq->last_avail_idx = 0;
 311         vq->last_used_event = 0;
 312         vq->avail_idx = 0;
 313         vq->last_used_idx = 0;
 314         vq->signalled_used = 0;
 315         vq->signalled_used_valid = false;
 316         vq->used_flags = 0;
 317         vq->log_used = false;
 318         vq->log_addr = -1ull;
 319         vq->private_data = NULL;
 320         vq->acked_features = 0;
 321         vq->log_base = NULL;
 322         vq->error_ctx = NULL;
 323         vq->error = NULL;
 324         vq->kick = NULL;
 325         vq->call_ctx = NULL;
 326         vq->call = NULL;
 327         vq->log_ctx = NULL;
 328         vhost_reset_is_le(vq);
 329         vhost_disable_cross_endian(vq);
 330         vq->busyloop_timeout = 0;
 331         vq->umem = NULL;
 332         vq->iotlb = NULL;
 333         __vhost_vq_meta_reset(vq);
 334 }
 335
 336 static int vhost_worker(void *data)
 337 {
 338         struct vhost_dev *dev = data;
 339         struct vhost_work *work, *work_next;
 340         struct llist_node *node;
 341         mm_segment_t oldfs = get_fs();
 342
 343         set_fs(USER_DS);
 344         use_mm(dev->mm);
 345
 346         for (;;) {
 347                 /* mb paired w/ kthread_stop */
 348                 set_current_state(TASK_INTERRUPTIBLE);
 349
 350                 if (kthread_should_stop()) {
 351                         __set_current_state(TASK_RUNNING);
 352                         break;
 353                 }
 354
 355                 node = llist_del_all(&dev->work_list);
 356                 if (!node)
 357                         schedule();
 358
 359                 node = llist_reverse_order(node);
 360                 /* make sure flag is seen after deletion */
 361                 smp_wmb();
 362                 llist_for_each_entry_safe(work, work_next, node, node) {
 363                         clear_bit(VHOST_WORK_QUEUED, &work->flags);
 364                         __set_current_state(TASK_RUNNING);
 365                         work->fn(work);
 366                         if (need_resched())
 367                                 schedule();
 368                 }
 369         }
 370         unuse_mm(dev->mm);
 371         set_fs(oldfs);
 372         return 0;
 373 }
 374
 375 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 376 {
 377         kfree(vq->indirect);
 378         vq->indirect = NULL;
 379         kfree(vq->log);
 380         vq->log = NULL;
 381         kfree(vq->heads);
 382         vq->heads = NULL;
 383 }
 384
 385 /* Helper to allocate iovec buffers for all vqs. */
 386 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 387 {
 388         struct vhost_virtqueue *vq;
 389         int i;
 390
 391         for (i = 0; i < dev->nvqs; ++i) {
 392                 vq = dev->vqs[i];
 393                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 394                                        GFP_KERNEL);
 395                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 396                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 397                 if (!vq->indirect || !vq->log || !vq->heads)
 398                         goto err_nomem;
 399         }
 400         return 0;
 401
 402 err_nomem:
 403         for (; i >= 0; --i)
 404                 vhost_vq_free_iovecs(dev->vqs[i]);
 405         return -ENOMEM;
 406 }
 407
 408 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 409 {
 410         int i;
 411
 412         for (i = 0; i < dev->nvqs; ++i)
 413                 vhost_vq_free_iovecs(dev->vqs[i]);
 414 }
 415
 416 void vhost_dev_init(struct vhost_dev *dev,
 417                     struct vhost_virtqueue **vqs, int nvqs)
 418 {
 419         struct vhost_virtqueue *vq;
 420         int i;
 421
 422         dev->vqs = vqs;
 423         dev->nvqs = nvqs;
 424         mutex_init(&dev->mutex);
 425         dev->log_ctx = NULL;
 426         dev->log_file = NULL;
 427         dev->umem = NULL;
 428         dev->iotlb = NULL;
 429         dev->mm = NULL;
 430         dev->worker = NULL;
 431         init_llist_head(&dev->work_list);
 432         init_waitqueue_head(&dev->wait);
 433         INIT_LIST_HEAD(&dev->read_list);
 434         INIT_LIST_HEAD(&dev->pending_list);
 435         spin_lock_init(&dev->iotlb_lock);
 436
 437
 438         for (i = 0; i < dev->nvqs; ++i) {
 439                 vq = dev->vqs[i];
 440                 vq->log = NULL;
 441                 vq->indirect = NULL;
 442                 vq->heads = NULL;
 443                 vq->dev = dev;
 444                 mutex_init(&vq->mutex);
 445                 vhost_vq_reset(dev, vq);
 446                 if (vq->handle_kick)
 447                         vhost_poll_init(&vq->poll, vq->handle_kick,
 448                                         POLLIN, dev);
 449         }
 450 }
 451 EXPORT_SYMBOL_GPL(vhost_dev_init);
 452
 453 /* Caller should have device mutex */
 454 long vhost_dev_check_owner(struct vhost_dev *dev)
 455 {
 456         /* Are you the owner? If not, I don't think you mean to do that */
 457         return dev->mm == current->mm ? 0 : -EPERM;
 458 }
 459 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 460
 461 struct vhost_attach_cgroups_struct {
 462         struct vhost_work work;
 463         struct task_struct *owner;
 464         int ret;
 465 };
 466
 467 static void vhost_attach_cgroups_work(struct vhost_work *work)
 468 {
 469         struct vhost_attach_cgroups_struct *s;
 470
 471         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 472         s->ret = cgroup_attach_task_all(s->owner, current);
 473 }
 474
 475 static int vhost_attach_cgroups(struct vhost_dev *dev)
 476 {
 477         struct vhost_attach_cgroups_struct attach;
 478
 479         attach.owner = current;
 480         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 481         vhost_work_queue(dev, &attach.work);
 482         vhost_work_flush(dev, &attach.work);
 483         return attach.ret;
 484 }
 485
 486 /* Caller should have device mutex */
 487 bool vhost_dev_has_owner(struct vhost_dev *dev)
 488 {
 489         return dev->mm;
 490 }
 491 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 492
 493 /* Caller should have device mutex */
 494 long vhost_dev_set_owner(struct vhost_dev *dev)
 495 {
 496         struct task_struct *worker;
 497         int err;
 498
 499         /* Is there an owner already? */
 500         if (vhost_dev_has_owner(dev)) {
 501                 err = -EBUSY;
 502                 goto err_mm;
 503         }
 504
 505         /* No owner, become one */
 506         dev->mm = get_task_mm(current);
 507         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 508         if (IS_ERR(worker)) {
 509                 err = PTR_ERR(worker);
 510                 goto err_worker;
 511         }
 512
 513         dev->worker = worker;
 514         wake_up_process(worker);        /* avoid contributing to loadavg */
 515
 516         err = vhost_attach_cgroups(dev);
 517         if (err)
 518                 goto err_cgroup;
 519
 520         err = vhost_dev_alloc_iovecs(dev);
 521         if (err)
 522                 goto err_cgroup;
 523
 524         return 0;
 525 err_cgroup:
 526         kthread_stop(worker);
 527         dev->worker = NULL;
 528 err_worker:
 529         if (dev->mm)
 530                 mmput(dev->mm);
 531         dev->mm = NULL;
 532 err_mm:
 533         return err;
 534 }
 535 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 536
 537 static void *vhost_kvzalloc(unsigned long size)
 538 {
 539         void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 540
 541         if (!n)
 542                 n = vzalloc(size);
 543         return n;
 544 }
 545
 546 struct vhost_umem *vhost_dev_reset_owner_prepare(void)
 547 {
 548         return vhost_kvzalloc(sizeof(struct vhost_umem));
 549 }
 550 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 551
 552 /* Caller should have device mutex */
 553 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_umem *umem)
 554 {
 555         int i;
 556
 557         vhost_dev_cleanup(dev, true);
 558
 559         /* Restore memory to default empty mapping. */
 560         INIT_LIST_HEAD(&umem->umem_list);
 561         dev->umem = umem;
 562         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 563          * VQs aren't running.
 564          */
 565         for (i = 0; i < dev->nvqs; ++i)
 566                 dev->vqs[i]->umem = umem;
 567 }
 568 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 569
 570 void vhost_dev_stop(struct vhost_dev *dev)
 571 {
 572         int i;
 573
 574         for (i = 0; i < dev->nvqs; ++i) {
 575                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 576                         vhost_poll_stop(&dev->vqs[i]->poll);
 577                         vhost_poll_flush(&dev->vqs[i]->poll);
 578                 }
 579         }
 580 }
 581 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 582
 583 static void vhost_umem_free(struct vhost_umem *umem,
 584                             struct vhost_umem_node *node)
 585 {
 586         vhost_umem_interval_tree_remove(node, &umem->umem_tree);
 587         list_del(&node->link);
 588         kfree(node);
 589         umem->numem--;
 590 }
 591
 592 static void vhost_umem_clean(struct vhost_umem *umem)
 593 {
 594         struct vhost_umem_node *node, *tmp;
 595
 596         if (!umem)
 597                 return;
 598
 599         list_for_each_entry_safe(node, tmp, &umem->umem_list, link)
 600                 vhost_umem_free(umem, node);
 601
 602         kvfree(umem);
 603 }
 604
 605 static void vhost_clear_msg(struct vhost_dev *dev)
 606 {
 607         struct vhost_msg_node *node, *n;
 608
 609         spin_lock(&dev->iotlb_lock);
 610
 611         list_for_each_entry_safe(node, n, &dev->read_list, node) {
 612                 list_del(&node->node);
 613                 kfree(node);
 614         }
 615
 616         list_for_each_entry_safe(node, n, &dev->pending_list, node) {
 617                 list_del(&node->node);
 618                 kfree(node);
 619         }
 620
 621         spin_unlock(&dev->iotlb_lock);
 622 }
 623
 624 /* Caller should have device mutex if and only if locked is set */
 625 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 626 {
 627         int i;
 628
 629         for (i = 0; i < dev->nvqs; ++i) {
 630                 if (dev->vqs[i]->error_ctx)
 631                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 632                 if (dev->vqs[i]->error)
 633                         fput(dev->vqs[i]->error);
 634                 if (dev->vqs[i]->kick)
 635                         fput(dev->vqs[i]->kick);
 636                 if (dev->vqs[i]->call_ctx)
 637                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 638                 if (dev->vqs[i]->call)
 639                         fput(dev->vqs[i]->call);
 640                 vhost_vq_reset(dev, dev->vqs[i]);
 641         }
 642         vhost_dev_free_iovecs(dev);
 643         if (dev->log_ctx)
 644                 eventfd_ctx_put(dev->log_ctx);
 645         dev->log_ctx = NULL;
 646         if (dev->log_file)
 647                 fput(dev->log_file);
 648         dev->log_file = NULL;
 649         /* No one will access memory at this point */
 650         vhost_umem_clean(dev->umem);
 651         dev->umem = NULL;
 652         vhost_umem_clean(dev->iotlb);
 653         dev->iotlb = NULL;
 654         vhost_clear_msg(dev);
 655         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
 656         WARN_ON(!llist_empty(&dev->work_list));
 657         if (dev->worker) {
 658                 kthread_stop(dev->worker);
 659                 dev->worker = NULL;
 660         }
 661         if (dev->mm)
 662                 mmput(dev->mm);
 663         dev->mm = NULL;
 664 }
 665 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 666
 667 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 668 {
 669         u64 a = addr / VHOST_PAGE_SIZE / 8;
 670
 671         /* Make sure 64 bit math will not overflow. */
 672         if (a > ULONG_MAX - (unsigned long)log_base ||
 673             a + (unsigned long)log_base > ULONG_MAX)
 674                 return 0;
 675
 676         return access_ok(VERIFY_WRITE, log_base + a,
 677                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 678 }
 679
 680 static bool vhost_overflow(u64 uaddr, u64 size)
 681 {
 682         /* Make sure 64 bit math will not overflow. */
 683         return uaddr > ULONG_MAX || size > ULONG_MAX || uaddr > ULONG_MAX - size;
 684 }
 685
 686 /* Caller should have vq mutex and device mutex. */
 687 static int vq_memory_access_ok(void __user *log_base, struct vhost_umem *umem,
 688                                int log_all)
 689 {
 690         struct vhost_umem_node *node;
 691
 692         if (!umem)
 693                 return 0;
 694
 695         list_for_each_entry(node, &umem->umem_list, link) {
 696                 unsigned long a = node->userspace_addr;
 697
 698                 if (vhost_overflow(node->userspace_addr, node->size))
 699                         return 0;
 700
 701
 702                 if (!access_ok(VERIFY_WRITE, (void __user *)a,
 703                                     node->size))
 704                         return 0;
 705                 else if (log_all && !log_access_ok(log_base,
 706                                                    node->start,
 707                                                    node->size))
 708                         return 0;
 709         }
 710         return 1;
 711 }
 712
 713 static inline void __user *vhost_vq_meta_fetch(struct vhost_virtqueue *vq,
 714                                                u64 addr, unsigned int size,
 715                                                int type)
 716 {
 717         const struct vhost_umem_node *node = vq->meta_iotlb[type];
 718
 719         if (!node)
 720                 return NULL;
 721
 722         return (void *)(uintptr_t)(node->userspace_addr + addr - node->start);
 723 }
 724
 725 /* Can we switch to this memory table? */
 726 /* Caller should have device mutex but not vq mutex */
 727 static int memory_access_ok(struct vhost_dev *d, struct vhost_umem *umem,
 728                             int log_all)
 729 {
 730         int i;
 731
 732         for (i = 0; i < d->nvqs; ++i) {
 733                 int ok;
 734                 bool log;
 735
 736                 mutex_lock(&d->vqs[i]->mutex);
 737                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 738                 /* If ring is inactive, will check when it's enabled. */
 739                 if (d->vqs[i]->private_data)
 740                         ok = vq_memory_access_ok(d->vqs[i]->log_base,
 741                                                  umem, log);
 742                 else
 743                         ok = 1;
 744                 mutex_unlock(&d->vqs[i]->mutex);
 745                 if (!ok)
 746                         return 0;
 747         }
 748         return 1;
 749 }
 750
 751 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
 752                           struct iovec iov[], int iov_size, int access);
 753
 754 static int vhost_copy_to_user(struct vhost_virtqueue *vq, void __user *to,
 755                               const void *from, unsigned size)
 756 {
 757         int ret;
 758
 759         if (!vq->iotlb)
 760                 return __copy_to_user(to, from, size);
 761         else {
 762                 /* This function should be called after iotlb
 763                  * prefetch, which means we're sure that all vq
 764                  * could be access through iotlb. So -EAGAIN should
 765                  * not happen in this case.
 766                  */
 767                 struct iov_iter t;
 768                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 769                                      (u64)(uintptr_t)to, size,
 770                                      VHOST_ADDR_DESC);
 771
 772                 if (uaddr)
 773                         return __copy_to_user(uaddr, from, size);
 774
 775                 ret = translate_desc(vq, (u64)(uintptr_t)to, size, vq->iotlb_iov,
 776                                      ARRAY_SIZE(vq->iotlb_iov),
 777                                      VHOST_ACCESS_WO);
 778                 if (ret < 0)
 779                         goto out;
 780                 iov_iter_init(&t, WRITE, vq->iotlb_iov, ret, size);
 781                 ret = copy_to_iter(from, size, &t);
 782                 if (ret == size)
 783                         ret = 0;
 784         }
 785 out:
 786         return ret;
 787 }
 788
 789 static int vhost_copy_from_user(struct vhost_virtqueue *vq, void *to,
 790                                 void __user *from, unsigned size)
 791 {
 792         int ret;
 793
 794         if (!vq->iotlb)
 795                 return __copy_from_user(to, from, size);
 796         else {
 797                 /* This function should be called after iotlb
 798                  * prefetch, which means we're sure that vq
 799                  * could be access through iotlb. So -EAGAIN should
 800                  * not happen in this case.
 801                  */
 802                 void __user *uaddr = vhost_vq_meta_fetch(vq,
 803                                      (u64)(uintptr_t)from, size,
 804                                      VHOST_ADDR_DESC);
 805                 struct iov_iter f;
 806
 807                 if (uaddr)
 808                         return __copy_from_user(to, uaddr, size);
 809
 810                 ret = translate_desc(vq, (u64)(uintptr_t)from, size, vq->iotlb_iov,
 811                                      ARRAY_SIZE(vq->iotlb_iov),
 812                                      VHOST_ACCESS_RO);
 813                 if (ret < 0) {
 814                         vq_err(vq, "IOTLB translation failure: uaddr "
 815                                "%p size 0x%llx\n", from,
 816                                (unsigned long long) size);
 817                         goto out;
 818                 }
 819                 iov_iter_init(&f, READ, vq->iotlb_iov, ret, size);
 820                 ret = copy_from_iter(to, size, &f);
 821                 if (ret == size)
 822                         ret = 0;
 823         }
 824
 825 out:
 826         return ret;
 827 }
 828
 829 static void __user *__vhost_get_user_slow(struct vhost_virtqueue *vq,
 830                                           void __user *addr, unsigned int size,
 831                                           int type)
 832 {
 833         int ret;
 834
 835         ret = translate_desc(vq, (u64)(uintptr_t)addr, size, vq->iotlb_iov,
 836                              ARRAY_SIZE(vq->iotlb_iov),
 837                              VHOST_ACCESS_RO);
 838         if (ret < 0) {
 839                 vq_err(vq, "IOTLB translation failure: uaddr "
 840                         "%p size 0x%llx\n", addr,
 841                         (unsigned long long) size);
 842                 return NULL;
 843         }
 844
 845         if (ret != 1 || vq->iotlb_iov[0].iov_len != size) {
 846                 vq_err(vq, "Non atomic userspace memory access: uaddr "
 847                         "%p size 0x%llx\n", addr,
 848                         (unsigned long long) size);
 849                 return NULL;
 850         }
 851
 852         return vq->iotlb_iov[0].iov_base;
 853 }
 854
 855 /* This function should be called after iotlb
 856  * prefetch, which means we're sure that vq
 857  * could be access through iotlb. So -EAGAIN should
 858  * not happen in this case.
 859  */
 860 static inline void __user *__vhost_get_user(struct vhost_virtqueue *vq,
 861                                             void *addr, unsigned int size,
 862                                             int type)
 863 {
 864         void __user *uaddr = vhost_vq_meta_fetch(vq,
 865                              (u64)(uintptr_t)addr, size, type);
 866         if (uaddr)
 867                 return uaddr;
 868
 869         return __vhost_get_user_slow(vq, addr, size, type);
 870 }
 871
 872 #define vhost_put_user(vq, x, ptr)              \
 873 ({ \
 874         int ret = -EFAULT; \
 875         if (!vq->iotlb) { \
 876                 ret = __put_user(x, ptr); \
 877         } else { \
 878                 __typeof__(ptr) to = \
 879                         (__typeof__(ptr)) __vhost_get_user(vq, ptr,     \
 880                                           sizeof(*ptr), VHOST_ADDR_USED); \
 881                 if (to != NULL) \
 882                         ret = __put_user(x, to); \
 883                 else \
 884                         ret = -EFAULT;  \
 885         } \
 886         ret; \
 887 })
 888
 889 #define vhost_get_user(vq, x, ptr, type)                \
 890 ({ \
 891         int ret; \
 892         if (!vq->iotlb) { \
 893                 ret = __get_user(x, ptr); \
 894         } else { \
 895                 __typeof__(ptr) from = \
 896                         (__typeof__(ptr)) __vhost_get_user(vq, ptr, \
 897                                                            sizeof(*ptr), \
 898                                                            type); \
 899                 if (from != NULL) \
 900                         ret = __get_user(x, from); \
 901                 else \
 902                         ret = -EFAULT; \
 903         } \
 904         ret; \
 905 })
 906
 907 #define vhost_get_avail(vq, x, ptr) \
 908         vhost_get_user(vq, x, ptr, VHOST_ADDR_AVAIL)
 909
 910 #define vhost_get_used(vq, x, ptr) \
 911         vhost_get_user(vq, x, ptr, VHOST_ADDR_USED)
 912
 913 static void vhost_dev_lock_vqs(struct vhost_dev *d)
 914 {
 915         int i = 0;
 916         for (i = 0; i < d->nvqs; ++i)
 917                 mutex_lock(&d->vqs[i]->mutex);
 918 }
 919
 920 static void vhost_dev_unlock_vqs(struct vhost_dev *d)
 921 {
 922         int i = 0;
 923         for (i = 0; i < d->nvqs; ++i)
 924                 mutex_unlock(&d->vqs[i]->mutex);
 925 }
 926
 927 static int vhost_new_umem_range(struct vhost_umem *umem,
 928                                 u64 start, u64 size, u64 end,
 929                                 u64 userspace_addr, int perm)
 930 {
 931         struct vhost_umem_node *tmp, *node = kmalloc(sizeof(*node), GFP_ATOMIC);
 932
 933         if (!node)
 934                 return -ENOMEM;
 935
 936         if (umem->numem == max_iotlb_entries) {
 937                 tmp = list_first_entry(&umem->umem_list, typeof(*tmp), link);
 938                 vhost_umem_free(umem, tmp);
 939         }
 940
 941         node->start = start;
 942         node->size = size;
 943         node->last = end;
 944         node->userspace_addr = userspace_addr;
 945         node->perm = perm;
 946         INIT_LIST_HEAD(&node->link);
 947         list_add_tail(&node->link, &umem->umem_list);
 948         vhost_umem_interval_tree_insert(node, &umem->umem_tree);
 949         umem->numem++;
 950
 951         return 0;
 952 }
 953
 954 static void vhost_del_umem_range(struct vhost_umem *umem,
 955                                  u64 start, u64 end)
 956 {
 957         struct vhost_umem_node *node;
 958
 959         while ((node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
 960                                                            start, end)))
 961                 vhost_umem_free(umem, node);
 962 }
 963
 964 static void vhost_iotlb_notify_vq(struct vhost_dev *d,
 965                                   struct vhost_iotlb_msg *msg)
 966 {
 967         struct vhost_msg_node *node, *n;
 968
 969         spin_lock(&d->iotlb_lock);
 970
 971         list_for_each_entry_safe(node, n, &d->pending_list, node) {
 972                 struct vhost_iotlb_msg *vq_msg = &node->msg.iotlb;
 973                 if (msg->iova <= vq_msg->iova &&
 974                     msg->iova + msg->size - 1 > vq_msg->iova &&
 975                     vq_msg->type == VHOST_IOTLB_MISS) {
 976                         vhost_poll_queue(&node->vq->poll);
 977                         list_del(&node->node);
 978                         kfree(node);
 979                 }
 980         }
 981
 982         spin_unlock(&d->iotlb_lock);
 983 }
 984
 985 static int umem_access_ok(u64 uaddr, u64 size, int access)
 986 {
 987         unsigned long a = uaddr;
 988
 989         /* Make sure 64 bit math will not overflow. */
 990         if (vhost_overflow(uaddr, size))
 991                 return -EFAULT;
 992
 993         if ((access & VHOST_ACCESS_RO) &&
 994             !access_ok(VERIFY_READ, (void __user *)a, size))
 995                 return -EFAULT;
 996         if ((access & VHOST_ACCESS_WO) &&
 997             !access_ok(VERIFY_WRITE, (void __user *)a, size))
 998                 return -EFAULT;
 999         return 0;
1000 }
1001
1002 static int vhost_process_iotlb_msg(struct vhost_dev *dev,
1003                                    struct vhost_iotlb_msg *msg)
1004 {
1005         int ret = 0;
1006
1007         vhost_dev_lock_vqs(dev);
1008         switch (msg->type) {
1009         case VHOST_IOTLB_UPDATE:
1010                 if (!dev->iotlb) {
1011                         ret = -EFAULT;
1012                         break;
1013                 }
1014                 if (umem_access_ok(msg->uaddr, msg->size, msg->perm)) {
1015                         ret = -EFAULT;
1016                         break;
1017                 }
1018                 vhost_vq_meta_reset(dev);
1019                 if (vhost_new_umem_range(dev->iotlb, msg->iova, msg->size,
1020                                          msg->iova + msg->size - 1,
1021                                          msg->uaddr, msg->perm)) {
1022                         ret = -ENOMEM;
1023                         break;
1024                 }
1025                 vhost_iotlb_notify_vq(dev, msg);
1026                 break;
1027         case VHOST_IOTLB_INVALIDATE:
1028                 vhost_vq_meta_reset(dev);
1029                 vhost_del_umem_range(dev->iotlb, msg->iova,
1030                                      msg->iova + msg->size - 1);
1031                 break;
1032         default:
1033                 ret = -EINVAL;
1034                 break;
1035         }
1036
1037         vhost_dev_unlock_vqs(dev);
1038         return ret;
1039 }
1040 ssize_t vhost_chr_write_iter(struct vhost_dev *dev,
1041                              struct iov_iter *from)
1042 {
1043         struct vhost_msg_node node;
1044         unsigned size = sizeof(struct vhost_msg);
1045         size_t ret;
1046         int err;
1047
1048         if (iov_iter_count(from) < size)
1049                 return 0;
1050         ret = copy_from_iter(&node.msg, size, from);
1051         if (ret != size)
1052                 goto done;
1053
1054         switch (node.msg.type) {
1055         case VHOST_IOTLB_MSG:
1056                 err = vhost_process_iotlb_msg(dev, &node.msg.iotlb);
1057                 if (err)
1058                         ret = err;
1059                 break;
1060         default:
1061                 ret = -EINVAL;
1062                 break;
1063         }
1064
1065 done:
1066         return ret;
1067 }
1068 EXPORT_SYMBOL(vhost_chr_write_iter);
1069
1070 unsigned int vhost_chr_poll(struct file *file, struct vhost_dev *dev,
1071                             poll_table *wait)
1072 {
1073         unsigned int mask = 0;
1074
1075         poll_wait(file, &dev->wait, wait);
1076
1077         if (!list_empty(&dev->read_list))
1078                 mask |= POLLIN | POLLRDNORM;
1079
1080         return mask;
1081 }
1082 EXPORT_SYMBOL(vhost_chr_poll);
1083
1084 ssize_t vhost_chr_read_iter(struct vhost_dev *dev, struct iov_iter *to,
1085                             int noblock)
1086 {
1087         DEFINE_WAIT(wait);
1088         struct vhost_msg_node *node;
1089         ssize_t ret = 0;
1090         unsigned size = sizeof(struct vhost_msg);
1091
1092         if (iov_iter_count(to) < size)
1093                 return 0;
1094
1095         while (1) {
1096                 if (!noblock)
1097                         prepare_to_wait(&dev->wait, &wait,
1098                                         TASK_INTERRUPTIBLE);
1099
1100                 node = vhost_dequeue_msg(dev, &dev->read_list);
1101                 if (node)
1102                         break;
1103                 if (noblock) {
1104                         ret = -EAGAIN;
1105                         break;
1106                 }
1107                 if (signal_pending(current)) {
1108                         ret = -ERESTARTSYS;
1109                         break;
1110                 }
1111                 if (!dev->iotlb) {
1112                         ret = -EBADFD;
1113                         break;
1114                 }
1115
1116                 schedule();
1117         }
1118
1119         if (!noblock)
1120                 finish_wait(&dev->wait, &wait);
1121
1122         if (node) {
1123                 ret = copy_to_iter(&node->msg, size, to);
1124
1125                 if (ret != size || node->msg.type != VHOST_IOTLB_MISS) {
1126                         kfree(node);
1127                         return ret;
1128                 }
1129
1130                 vhost_enqueue_msg(dev, &dev->pending_list, node);
1131         }
1132
1133         return ret;
1134 }
1135 EXPORT_SYMBOL_GPL(vhost_chr_read_iter);
1136
1137 static int vhost_iotlb_miss(struct vhost_virtqueue *vq, u64 iova, int access)
1138 {
1139         struct vhost_dev *dev = vq->dev;
1140         struct vhost_msg_node *node;
1141         struct vhost_iotlb_msg *msg;
1142
1143         node = vhost_new_msg(vq, VHOST_IOTLB_MISS);
1144         if (!node)
1145                 return -ENOMEM;
1146
1147         msg = &node->msg.iotlb;
1148         msg->type = VHOST_IOTLB_MISS;
1149         msg->iova = iova;
1150         msg->perm = access;
1151
1152         vhost_enqueue_msg(dev, &dev->read_list, node);
1153
1154         return 0;
1155 }
1156
1157 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
1158                         struct vring_desc __user *desc,
1159                         struct vring_avail __user *avail,
1160                         struct vring_used __user *used)
1161
1162 {
1163         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1164
1165         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
1166                access_ok(VERIFY_READ, avail,
1167                          sizeof *avail + num * sizeof *avail->ring + s) &&
1168                access_ok(VERIFY_WRITE, used,
1169                         sizeof *used + num * sizeof *used->ring + s);
1170 }
1171
1172 static void vhost_vq_meta_update(struct vhost_virtqueue *vq,
1173                                  const struct vhost_umem_node *node,
1174                                  int type)
1175 {
1176         int access = (type == VHOST_ADDR_USED) ?
1177                      VHOST_ACCESS_WO : VHOST_ACCESS_RO;
1178
1179         if (likely(node->perm & access))
1180                 vq->meta_iotlb[type] = node;
1181 }
1182
1183 static int iotlb_access_ok(struct vhost_virtqueue *vq,
1184                            int access, u64 addr, u64 len, int type)
1185 {
1186         const struct vhost_umem_node *node;
1187         struct vhost_umem *umem = vq->iotlb;
1188         u64 s = 0, size, orig_addr = addr;
1189
1190         if (vhost_vq_meta_fetch(vq, addr, len, type))
1191                 return true;
1192
1193         while (len > s) {
1194                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1195                                                            addr,
1196                                                            addr + len - 1);
1197                 if (node == NULL || node->start > addr) {
1198                         vhost_iotlb_miss(vq, addr, access);
1199                         return false;
1200                 } else if (!(node->perm & access)) {
1201                         /* Report the possible access violation by
1202                          * request another translation from userspace.
1203                          */
1204                         return false;
1205                 }
1206
1207                 size = node->size - addr + node->start;
1208
1209                 if (orig_addr == addr && size >= len)
1210                         vhost_vq_meta_update(vq, node, type);
1211
1212                 s += size;
1213                 addr += size;
1214         }
1215
1216         return true;
1217 }
1218
1219 int vq_iotlb_prefetch(struct vhost_virtqueue *vq)
1220 {
1221         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1222         unsigned int num = vq->num;
1223
1224         if (!vq->iotlb)
1225                 return 1;
1226
1227         return iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->desc,
1228                                num * sizeof(*vq->desc), VHOST_ADDR_DESC) &&
1229                iotlb_access_ok(vq, VHOST_ACCESS_RO, (u64)(uintptr_t)vq->avail,
1230                                sizeof *vq->avail +
1231                                num * sizeof(*vq->avail->ring) + s,
1232                                VHOST_ADDR_AVAIL) &&
1233                iotlb_access_ok(vq, VHOST_ACCESS_WO, (u64)(uintptr_t)vq->used,
1234                                sizeof *vq->used +
1235                                num * sizeof(*vq->used->ring) + s,
1236                                VHOST_ADDR_USED);
1237 }
1238 EXPORT_SYMBOL_GPL(vq_iotlb_prefetch);
1239
1240 /* Can we log writes? */
1241 /* Caller should have device mutex but not vq mutex */
1242 int vhost_log_access_ok(struct vhost_dev *dev)
1243 {
1244         return memory_access_ok(dev, dev->umem, 1);
1245 }
1246 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
1247
1248 /* Verify access for write logging. */
1249 /* Caller should have vq mutex and device mutex */
1250 static int vq_log_access_ok(struct vhost_virtqueue *vq,
1251                             void __user *log_base)
1252 {
1253         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
1254
1255         return vq_memory_access_ok(log_base, vq->umem,
1256                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
1257                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
1258                                         sizeof *vq->used +
1259                                         vq->num * sizeof *vq->used->ring + s));
1260 }
1261
1262 /* Can we start vq? */
1263 /* Caller should have vq mutex and device mutex */
1264 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
1265 {
1266         if (vq->iotlb) {
1267                 /* When device IOTLB was used, the access validation
1268                  * will be validated during prefetching.
1269                  */
1270                 return 1;
1271         }
1272         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
1273                 vq_log_access_ok(vq, vq->log_base);
1274 }
1275 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
1276
1277 static struct vhost_umem *vhost_umem_alloc(void)
1278 {
1279         struct vhost_umem *umem = vhost_kvzalloc(sizeof(*umem));
1280
1281         if (!umem)
1282                 return NULL;
1283
1284         umem->umem_tree = RB_ROOT;
1285         umem->numem = 0;
1286         INIT_LIST_HEAD(&umem->umem_list);
1287
1288         return umem;
1289 }
1290
1291 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
1292 {
1293         struct vhost_memory mem, *newmem;
1294         struct vhost_memory_region *region;
1295         struct vhost_umem *newumem, *oldumem;
1296         unsigned long size = offsetof(struct vhost_memory, regions);
1297         int i;
1298
1299         if (copy_from_user(&mem, m, size))
1300                 return -EFAULT;
1301         if (mem.padding)
1302                 return -EOPNOTSUPP;
1303         if (mem.nregions > max_mem_regions)
1304                 return -E2BIG;
1305         newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
1306         if (!newmem)
1307                 return -ENOMEM;
1308
1309         memcpy(newmem, &mem, size);
1310         if (copy_from_user(newmem->regions, m->regions,
1311                            mem.nregions * sizeof *m->regions)) {
1312                 kvfree(newmem);
1313                 return -EFAULT;
1314         }
1315
1316         newumem = vhost_umem_alloc();
1317         if (!newumem) {
1318                 kvfree(newmem);
1319                 return -ENOMEM;
1320         }
1321
1322         for (region = newmem->regions;
1323              region < newmem->regions + mem.nregions;
1324              region++) {
1325                 if (vhost_new_umem_range(newumem,
1326                                          region->guest_phys_addr,
1327                                          region->memory_size,
1328                                          region->guest_phys_addr +
1329                                          region->memory_size - 1,
1330                                          region->userspace_addr,
1331                                          VHOST_ACCESS_RW))
1332                         goto err;
1333         }
1334
1335         if (!memory_access_ok(d, newumem, 0))
1336                 goto err;
1337
1338         oldumem = d->umem;
1339         d->umem = newumem;
1340
1341         /* All memory accesses are done under some VQ mutex. */
1342         for (i = 0; i < d->nvqs; ++i) {
1343                 mutex_lock(&d->vqs[i]->mutex);
1344                 d->vqs[i]->umem = newumem;
1345                 mutex_unlock(&d->vqs[i]->mutex);
1346         }
1347
1348         kvfree(newmem);
1349         vhost_umem_clean(oldumem);
1350         return 0;
1351
1352 err:
1353         vhost_umem_clean(newumem);
1354         kvfree(newmem);
1355         return -EFAULT;
1356 }
1357
1358 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
1359 {
1360         struct file *eventfp, *filep = NULL;
1361         bool pollstart = false, pollstop = false;
1362         struct eventfd_ctx *ctx = NULL;
1363         u32 __user *idxp = argp;
1364         struct vhost_virtqueue *vq;
1365         struct vhost_vring_state s;
1366         struct vhost_vring_file f;
1367         struct vhost_vring_addr a;
1368         u32 idx;
1369         long r;
1370
1371         r = get_user(idx, idxp);
1372         if (r < 0)
1373                 return r;
1374         if (idx >= d->nvqs)
1375                 return -ENOBUFS;
1376
1377         vq = d->vqs[idx];
1378
1379         mutex_lock(&vq->mutex);
1380
1381         switch (ioctl) {
1382         case VHOST_SET_VRING_NUM:
1383                 /* Resizing ring with an active backend?
1384                  * You don't want to do that. */
1385                 if (vq->private_data) {
1386                         r = -EBUSY;
1387                         break;
1388                 }
1389                 if (copy_from_user(&s, argp, sizeof s)) {
1390                         r = -EFAULT;
1391                         break;
1392                 }
1393                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
1394                         r = -EINVAL;
1395                         break;
1396                 }
1397                 vq->num = s.num;
1398                 break;
1399         case VHOST_SET_VRING_BASE:
1400                 /* Moving base with an active backend?
1401                  * You don't want to do that. */
1402                 if (vq->private_data) {
1403                         r = -EBUSY;
1404                         break;
1405                 }
1406                 if (copy_from_user(&s, argp, sizeof s)) {
1407                         r = -EFAULT;
1408                         break;
1409                 }
1410                 if (s.num > 0xffff) {
1411                         r = -EINVAL;
1412                         break;
1413                 }
1414                 vq->last_avail_idx = vq->last_used_event = s.num;
1415                 /* Forget the cached index value. */
1416                 vq->avail_idx = vq->last_avail_idx;
1417                 break;
1418         case VHOST_GET_VRING_BASE:
1419                 s.index = idx;
1420                 s.num = vq->last_avail_idx;
1421                 if (copy_to_user(argp, &s, sizeof s))
1422                         r = -EFAULT;
1423                 break;
1424         case VHOST_SET_VRING_ADDR:
1425                 if (copy_from_user(&a, argp, sizeof a)) {
1426                         r = -EFAULT;
1427                         break;
1428                 }
1429                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
1430                         r = -EOPNOTSUPP;
1431                         break;
1432                 }
1433                 /* For 32bit, verify that the top 32bits of the user
1434                    data are set to zero. */
1435                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
1436                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
1437                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
1438                         r = -EFAULT;
1439                         break;
1440                 }
1441
1442                 /* Make sure it's safe to cast pointers to vring types. */
1443                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
1444                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
1445                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
1446                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
1447                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
1448                         r = -EINVAL;
1449                         break;
1450                 }
1451
1452                 /* We only verify access here if backend is configured.
1453                  * If it is not, we don't as size might not have been setup.
1454                  * We will verify when backend is configured. */
1455                 if (vq->private_data) {
1456                         if (!vq_access_ok(vq, vq->num,
1457                                 (void __user *)(unsigned long)a.desc_user_addr,
1458                                 (void __user *)(unsigned long)a.avail_user_addr,
1459                                 (void __user *)(unsigned long)a.used_user_addr)) {
1460                                 r = -EINVAL;
1461                                 break;
1462                         }
1463
1464                         /* Also validate log access for used ring if enabled. */
1465                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
1466                             !log_access_ok(vq->log_base, a.log_guest_addr,
1467                                            sizeof *vq->used +
1468                                            vq->num * sizeof *vq->used->ring)) {
1469                                 r = -EINVAL;
1470                                 break;
1471                         }
1472                 }
1473
1474                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
1475                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
1476                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
1477                 vq->log_addr = a.log_guest_addr;
1478                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
1479                 break;
1480         case VHOST_SET_VRING_KICK:
1481                 if (copy_from_user(&f, argp, sizeof f)) {
1482                         r = -EFAULT;
1483                         break;
1484                 }
1485                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1486                 if (IS_ERR(eventfp)) {
1487                         r = PTR_ERR(eventfp);
1488                         break;
1489                 }
1490                 if (eventfp != vq->kick) {
1491                         pollstop = (filep = vq->kick) != NULL;
1492                         pollstart = (vq->kick = eventfp) != NULL;
1493                 } else
1494                         filep = eventfp;
1495                 break;
1496         case VHOST_SET_VRING_CALL:
1497                 if (copy_from_user(&f, argp, sizeof f)) {
1498                         r = -EFAULT;
1499                         break;
1500                 }
1501                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1502                 if (IS_ERR(eventfp)) {
1503                         r = PTR_ERR(eventfp);
1504                         break;
1505                 }
1506                 if (eventfp != vq->call) {
1507                         filep = vq->call;
1508                         ctx = vq->call_ctx;
1509                         vq->call = eventfp;
1510                         vq->call_ctx = eventfp ?
1511                                 eventfd_ctx_fileget(eventfp) : NULL;
1512                 } else
1513                         filep = eventfp;
1514                 break;
1515         case VHOST_SET_VRING_ERR:
1516                 if (copy_from_user(&f, argp, sizeof f)) {
1517                         r = -EFAULT;
1518                         break;
1519                 }
1520                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
1521                 if (IS_ERR(eventfp)) {
1522                         r = PTR_ERR(eventfp);
1523                         break;
1524                 }
1525                 if (eventfp != vq->error) {
1526                         filep = vq->error;
1527                         vq->error = eventfp;
1528                         ctx = vq->error_ctx;
1529                         vq->error_ctx = eventfp ?
1530                                 eventfd_ctx_fileget(eventfp) : NULL;
1531                 } else
1532                         filep = eventfp;
1533                 break;
1534         case VHOST_SET_VRING_ENDIAN:
1535                 r = vhost_set_vring_endian(vq, argp);
1536                 break;
1537         case VHOST_GET_VRING_ENDIAN:
1538                 r = vhost_get_vring_endian(vq, idx, argp);
1539                 break;
1540         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
1541                 if (copy_from_user(&s, argp, sizeof(s))) {
1542                         r = -EFAULT;
1543                         break;
1544                 }
1545                 vq->busyloop_timeout = s.num;
1546                 break;
1547         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
1548                 s.index = idx;
1549                 s.num = vq->busyloop_timeout;
1550                 if (copy_to_user(argp, &s, sizeof(s)))
1551                         r = -EFAULT;
1552                 break;
1553         default:
1554                 r = -ENOIOCTLCMD;
1555         }
1556
1557         if (pollstop && vq->handle_kick)
1558                 vhost_poll_stop(&vq->poll);
1559
1560         if (ctx)
1561                 eventfd_ctx_put(ctx);
1562         if (filep)
1563                 fput(filep);
1564
1565         if (pollstart && vq->handle_kick)
1566                 r = vhost_poll_start(&vq->poll, vq->kick);
1567
1568         mutex_unlock(&vq->mutex);
1569
1570         if (pollstop && vq->handle_kick)
1571                 vhost_poll_flush(&vq->poll);
1572         return r;
1573 }
1574 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
1575
1576 int vhost_init_device_iotlb(struct vhost_dev *d, bool enabled)
1577 {
1578         struct vhost_umem *niotlb, *oiotlb;
1579         int i;
1580
1581         niotlb = vhost_umem_alloc();
1582         if (!niotlb)
1583                 return -ENOMEM;
1584
1585         oiotlb = d->iotlb;
1586         d->iotlb = niotlb;
1587
1588         for (i = 0; i < d->nvqs; ++i) {
1589                 mutex_lock(&d->vqs[i]->mutex);
1590                 d->vqs[i]->iotlb = niotlb;
1591                 mutex_unlock(&d->vqs[i]->mutex);
1592         }
1593
1594         vhost_umem_clean(oiotlb);
1595
1596         return 0;
1597 }
1598 EXPORT_SYMBOL_GPL(vhost_init_device_iotlb);
1599
1600 /* Caller must have device mutex */
1601 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
1602 {
1603         struct file *eventfp, *filep = NULL;
1604         struct eventfd_ctx *ctx = NULL;
1605         u64 p;
1606         long r;
1607         int i, fd;
1608
1609         /* If you are not the owner, you can become one */
1610         if (ioctl == VHOST_SET_OWNER) {
1611                 r = vhost_dev_set_owner(d);
1612                 goto done;
1613         }
1614
1615         /* You must be the owner to do anything else */
1616         r = vhost_dev_check_owner(d);
1617         if (r)
1618                 goto done;
1619
1620         switch (ioctl) {
1621         case VHOST_SET_MEM_TABLE:
1622                 r = vhost_set_memory(d, argp);
1623                 break;
1624         case VHOST_SET_LOG_BASE:
1625                 if (copy_from_user(&p, argp, sizeof p)) {
1626                         r = -EFAULT;
1627                         break;
1628                 }
1629                 if ((u64)(unsigned long)p != p) {
1630                         r = -EFAULT;
1631                         break;
1632                 }
1633                 for (i = 0; i < d->nvqs; ++i) {
1634                         struct vhost_virtqueue *vq;
1635                         void __user *base = (void __user *)(unsigned long)p;
1636                         vq = d->vqs[i];
1637                         mutex_lock(&vq->mutex);
1638                         /* If ring is inactive, will check when it's enabled. */
1639                         if (vq->private_data && !vq_log_access_ok(vq, base))
1640                                 r = -EFAULT;
1641                         else
1642                                 vq->log_base = base;
1643                         mutex_unlock(&vq->mutex);
1644                 }
1645                 break;
1646         case VHOST_SET_LOG_FD:
1647                 r = get_user(fd, (int __user *)argp);
1648                 if (r < 0)
1649                         break;
1650                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1651                 if (IS_ERR(eventfp)) {
1652                         r = PTR_ERR(eventfp);
1653                         break;
1654                 }
1655                 if (eventfp != d->log_file) {
1656                         filep = d->log_file;
1657                         d->log_file = eventfp;
1658                         ctx = d->log_ctx;
1659                         d->log_ctx = eventfp ?
1660                                 eventfd_ctx_fileget(eventfp) : NULL;
1661                 } else
1662                         filep = eventfp;
1663                 for (i = 0; i < d->nvqs; ++i) {
1664                         mutex_lock(&d->vqs[i]->mutex);
1665                         d->vqs[i]->log_ctx = d->log_ctx;
1666                         mutex_unlock(&d->vqs[i]->mutex);
1667                 }
1668                 if (ctx)
1669                         eventfd_ctx_put(ctx);
1670                 if (filep)
1671                         fput(filep);
1672                 break;
1673         default:
1674                 r = -ENOIOCTLCMD;
1675                 break;
1676         }
1677 done:
1678         return r;
1679 }
1680 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1681
1682 /* TODO: This is really inefficient.  We need something like get_user()
1683  * (instruction directly accesses the data, with an exception table entry
1684  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1685  */
1686 static int set_bit_to_user(int nr, void __user *addr)
1687 {
1688         unsigned long log = (unsigned long)addr;
1689         struct page *page;
1690         void *base;
1691         int bit = nr + (log % PAGE_SIZE) * 8;
1692         int r;
1693
1694         r = get_user_pages_fast(log, 1, 1, &page);
1695         if (r < 0)
1696                 return r;
1697         BUG_ON(r != 1);
1698         base = kmap_atomic(page);
1699         set_bit(bit, base);
1700         kunmap_atomic(base);
1701         set_page_dirty_lock(page);
1702         put_page(page);
1703         return 0;
1704 }
1705
1706 static int log_write(void __user *log_base,
1707                      u64 write_address, u64 write_length)
1708 {
1709         u64 write_page = write_address / VHOST_PAGE_SIZE;
1710         int r;
1711
1712         if (!write_length)
1713                 return 0;
1714         write_length += write_address % VHOST_PAGE_SIZE;
1715         for (;;) {
1716                 u64 base = (u64)(unsigned long)log_base;
1717                 u64 log = base + write_page / 8;
1718                 int bit = write_page % 8;
1719                 if ((u64)(unsigned long)log != log)
1720                         return -EFAULT;
1721                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1722                 if (r < 0)
1723                         return r;
1724                 if (write_length <= VHOST_PAGE_SIZE)
1725                         break;
1726                 write_length -= VHOST_PAGE_SIZE;
1727                 write_page += 1;
1728         }
1729         return r;
1730 }
1731
1732 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1733                     unsigned int log_num, u64 len)
1734 {
1735         int i, r;
1736
1737         /* Make sure data written is seen before log. */
1738         smp_wmb();
1739         for (i = 0; i < log_num; ++i) {
1740                 u64 l = min(log[i].len, len);
1741                 r = log_write(vq->log_base, log[i].addr, l);
1742                 if (r < 0)
1743                         return r;
1744                 len -= l;
1745                 if (!len) {
1746                         if (vq->log_ctx)
1747                                 eventfd_signal(vq->log_ctx, 1);
1748                         return 0;
1749                 }
1750         }
1751         /* Length written exceeds what we have stored. This is a bug. */
1752         BUG();
1753         return 0;
1754 }
1755 EXPORT_SYMBOL_GPL(vhost_log_write);
1756
1757 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1758 {
1759         void __user *used;
1760         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->used_flags),
1761                            &vq->used->flags) < 0)
1762                 return -EFAULT;
1763         if (unlikely(vq->log_used)) {
1764                 /* Make sure the flag is seen before log. */
1765                 smp_wmb();
1766                 /* Log used flag write. */
1767                 used = &vq->used->flags;
1768                 log_write(vq->log_base, vq->log_addr +
1769                           (used - (void __user *)vq->used),
1770                           sizeof vq->used->flags);
1771                 if (vq->log_ctx)
1772                         eventfd_signal(vq->log_ctx, 1);
1773         }
1774         return 0;
1775 }
1776
1777 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1778 {
1779         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->avail_idx),
1780                            vhost_avail_event(vq)))
1781                 return -EFAULT;
1782         if (unlikely(vq->log_used)) {
1783                 void __user *used;
1784                 /* Make sure the event is seen before log. */
1785                 smp_wmb();
1786                 /* Log avail event write */
1787                 used = vhost_avail_event(vq);
1788                 log_write(vq->log_base, vq->log_addr +
1789                           (used - (void __user *)vq->used),
1790                           sizeof *vhost_avail_event(vq));
1791                 if (vq->log_ctx)
1792                         eventfd_signal(vq->log_ctx, 1);
1793         }
1794         return 0;
1795 }
1796
1797 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1798 {
1799         __virtio16 last_used_idx;
1800         int r;
1801         bool is_le = vq->is_le;
1802
1803         if (!vq->private_data)
1804                 return 0;
1805
1806         vhost_init_is_le(vq);
1807
1808         r = vhost_update_used_flags(vq);
1809         if (r)
1810                 goto err;
1811         vq->signalled_used_valid = false;
1812         if (!vq->iotlb &&
1813             !access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1814                 r = -EFAULT;
1815                 goto err;
1816         }
1817         r = vhost_get_used(vq, last_used_idx, &vq->used->idx);
1818         if (r) {
1819                 vq_err(vq, "Can't access used idx at %p\n",
1820                        &vq->used->idx);
1821                 goto err;
1822         }
1823         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1824         return 0;
1825
1826 err:
1827         vq->is_le = is_le;
1828         return r;
1829 }
1830 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1831
1832 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1833                           struct iovec iov[], int iov_size, int access)
1834 {
1835         const struct vhost_umem_node *node;
1836         struct vhost_dev *dev = vq->dev;
1837         struct vhost_umem *umem = dev->iotlb ? dev->iotlb : dev->umem;
1838         struct iovec *_iov;
1839         u64 s = 0;
1840         int ret = 0;
1841
1842         while ((u64)len > s) {
1843                 u64 size;
1844                 if (unlikely(ret >= iov_size)) {
1845                         ret = -ENOBUFS;
1846                         break;
1847                 }
1848
1849                 node = vhost_umem_interval_tree_iter_first(&umem->umem_tree,
1850                                                         addr, addr + len - 1);
1851                 if (node == NULL || node->start > addr) {
1852                         if (umem != dev->iotlb) {
1853                                 ret = -EFAULT;
1854                                 break;
1855                         }
1856                         ret = -EAGAIN;
1857                         break;
1858                 } else if (!(node->perm & access)) {
1859                         ret = -EPERM;
1860                         break;
1861                 }
1862
1863                 _iov = iov + ret;
1864                 size = node->size - addr + node->start;
1865                 _iov->iov_len = min((u64)len - s, size);
1866                 _iov->iov_base = (void __user *)(unsigned long)
1867                         (node->userspace_addr + addr - node->start);
1868                 s += size;
1869                 addr += size;
1870                 ++ret;
1871         }
1872
1873         if (ret == -EAGAIN)
1874                 vhost_iotlb_miss(vq, addr, access);
1875         return ret;
1876 }
1877
1878 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1879  * function returns the next descriptor in the chain,
1880  * or -1U if we're at the end. */
1881 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1882 {
1883         unsigned int next;
1884
1885         /* If this descriptor says it doesn't chain, we're done. */
1886         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1887                 return -1U;
1888
1889         /* Check they're not leading us off end of descriptors. */
1890         next = vhost16_to_cpu(vq, desc->next);
1891         /* Make sure compiler knows to grab that: we don't want it changing! */
1892         /* We will use the result as an index in an array, so most
1893          * architectures only need a compiler barrier here. */
1894         read_barrier_depends();
1895
1896         return next;
1897 }
1898
1899 static int get_indirect(struct vhost_virtqueue *vq,
1900                         struct iovec iov[], unsigned int iov_size,
1901                         unsigned int *out_num, unsigned int *in_num,
1902                         struct vhost_log *log, unsigned int *log_num,
1903                         struct vring_desc *indirect)
1904 {
1905         struct vring_desc desc;
1906         unsigned int i = 0, count, found = 0;
1907         u32 len = vhost32_to_cpu(vq, indirect->len);
1908         struct iov_iter from;
1909         int ret, access;
1910
1911         /* Sanity check */
1912         if (unlikely(len % sizeof desc)) {
1913                 vq_err(vq, "Invalid length in indirect descriptor: "
1914                        "len 0x%llx not multiple of 0x%zx\n",
1915                        (unsigned long long)len,
1916                        sizeof desc);
1917                 return -EINVAL;
1918         }
1919
1920         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1921                              UIO_MAXIOV, VHOST_ACCESS_RO);
1922         if (unlikely(ret < 0)) {
1923                 if (ret != -EAGAIN)
1924                         vq_err(vq, "Translation failure %d in indirect.\n", ret);
1925                 return ret;
1926         }
1927         iov_iter_init(&from, READ, vq->indirect, ret, len);
1928
1929         /* We will use the result as an address to read from, so most
1930          * architectures only need a compiler barrier here. */
1931         read_barrier_depends();
1932
1933         count = len / sizeof desc;
1934         /* Buffers are chained via a 16 bit next field, so
1935          * we can have at most 2^16 of these. */
1936         if (unlikely(count > USHRT_MAX + 1)) {
1937                 vq_err(vq, "Indirect buffer length too big: %d\n",
1938                        indirect->len);
1939                 return -E2BIG;
1940         }
1941
1942         do {
1943                 unsigned iov_count = *in_num + *out_num;
1944                 if (unlikely(++found > count)) {
1945                         vq_err(vq, "Loop detected: last one at %u "
1946                                "indirect size %u\n",
1947                                i, count);
1948                         return -EINVAL;
1949                 }
1950                 if (unlikely(!copy_from_iter_full(&desc, sizeof(desc), &from))) {
1951                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1952                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1953                         return -EINVAL;
1954                 }
1955                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1956                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1957                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1958                         return -EINVAL;
1959                 }
1960
1961                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
1962                         access = VHOST_ACCESS_WO;
1963                 else
1964                         access = VHOST_ACCESS_RO;
1965
1966                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1967                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1968                                      iov_size - iov_count, access);
1969                 if (unlikely(ret < 0)) {
1970                         if (ret != -EAGAIN)
1971                                 vq_err(vq, "Translation failure %d indirect idx %d\n",
1972                                         ret, i);
1973                         return ret;
1974                 }
1975                 /* If this is an input descriptor, increment that count. */
1976                 if (access == VHOST_ACCESS_WO) {
1977                         *in_num += ret;
1978                         if (unlikely(log)) {
1979                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1980                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1981                                 ++*log_num;
1982                         }
1983                 } else {
1984                         /* If it's an output descriptor, they're all supposed
1985                          * to come before any input descriptors. */
1986                         if (unlikely(*in_num)) {
1987                                 vq_err(vq, "Indirect descriptor "
1988                                        "has out after in: idx %d\n", i);
1989                                 return -EINVAL;
1990                         }
1991                         *out_num += ret;
1992                 }
1993         } while ((i = next_desc(vq, &desc)) != -1);
1994         return 0;
1995 }
1996
1997 /* This looks in the virtqueue and for the first available buffer, and converts
1998  * it to an iovec for convenient access.  Since descriptors consist of some
1999  * number of output then some number of input descriptors, it's actually two
2000  * iovecs, but we pack them into one and note how many of each there were.
2001  *
2002  * This function returns the descriptor number found, or vq->num (which is
2003  * never a valid descriptor number) if none was found.  A negative code is
2004  * returned on error. */
2005 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
2006                       struct iovec iov[], unsigned int iov_size,
2007                       unsigned int *out_num, unsigned int *in_num,
2008                       struct vhost_log *log, unsigned int *log_num)
2009 {
2010         struct vring_desc desc;
2011         unsigned int i, head, found = 0;
2012         u16 last_avail_idx;
2013         __virtio16 avail_idx;
2014         __virtio16 ring_head;
2015         int ret, access;
2016
2017         /* Check it isn't doing very strange things with descriptor numbers. */
2018         last_avail_idx = vq->last_avail_idx;
2019
2020         if (vq->avail_idx == vq->last_avail_idx) {
2021                 if (unlikely(vhost_get_avail(vq, avail_idx, &vq->avail->idx))) {
2022                         vq_err(vq, "Failed to access avail idx at %p\n",
2023                                 &vq->avail->idx);
2024                         return -EFAULT;
2025                 }
2026                 vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2027
2028                 if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
2029                         vq_err(vq, "Guest moved used index from %u to %u",
2030                                 last_avail_idx, vq->avail_idx);
2031                         return -EFAULT;
2032                 }
2033
2034                 /* If there's nothing new since last we looked, return
2035                  * invalid.
2036                  */
2037                 if (vq->avail_idx == last_avail_idx)
2038                         return vq->num;
2039
2040                 /* Only get avail ring entries after they have been
2041                  * exposed by guest.
2042                  */
2043                 smp_rmb();
2044         }
2045
2046         /* Grab the next descriptor number they're advertising, and increment
2047          * the index we've seen. */
2048         if (unlikely(vhost_get_avail(vq, ring_head,
2049                      &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
2050                 vq_err(vq, "Failed to read head: idx %d address %p\n",
2051                        last_avail_idx,
2052                        &vq->avail->ring[last_avail_idx % vq->num]);
2053                 return -EFAULT;
2054         }
2055
2056         head = vhost16_to_cpu(vq, ring_head);
2057
2058         /* If their number is silly, that's an error. */
2059         if (unlikely(head >= vq->num)) {
2060                 vq_err(vq, "Guest says index %u > %u is available",
2061                        head, vq->num);
2062                 return -EINVAL;
2063         }
2064
2065         /* When we start there are none of either input nor output. */
2066         *out_num = *in_num = 0;
2067         if (unlikely(log))
2068                 *log_num = 0;
2069
2070         i = head;
2071         do {
2072                 unsigned iov_count = *in_num + *out_num;
2073                 if (unlikely(i >= vq->num)) {
2074                         vq_err(vq, "Desc index is %u > %u, head = %u",
2075                                i, vq->num, head);
2076                         return -EINVAL;
2077                 }
2078                 if (unlikely(++found > vq->num)) {
2079                         vq_err(vq, "Loop detected: last one at %u "
2080                                "vq size %u head %u\n",
2081                                i, vq->num, head);
2082                         return -EINVAL;
2083                 }
2084                 ret = vhost_copy_from_user(vq, &desc, vq->desc + i,
2085                                            sizeof desc);
2086                 if (unlikely(ret)) {
2087                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
2088                                i, vq->desc + i);
2089                         return -EFAULT;
2090                 }
2091                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
2092                         ret = get_indirect(vq, iov, iov_size,
2093                                            out_num, in_num,
2094                                            log, log_num, &desc);
2095                         if (unlikely(ret < 0)) {
2096                                 if (ret != -EAGAIN)
2097                                         vq_err(vq, "Failure detected "
2098                                                 "in indirect descriptor at idx %d\n", i);
2099                                 return ret;
2100                         }
2101                         continue;
2102                 }
2103
2104                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE))
2105                         access = VHOST_ACCESS_WO;
2106                 else
2107                         access = VHOST_ACCESS_RO;
2108                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
2109                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
2110                                      iov_size - iov_count, access);
2111                 if (unlikely(ret < 0)) {
2112                         if (ret != -EAGAIN)
2113                                 vq_err(vq, "Translation failure %d descriptor idx %d\n",
2114                                         ret, i);
2115                         return ret;
2116                 }
2117                 if (access == VHOST_ACCESS_WO) {
2118                         /* If this is an input descriptor,
2119                          * increment that count. */
2120                         *in_num += ret;
2121                         if (unlikely(log)) {
2122                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
2123                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
2124                                 ++*log_num;
2125                         }
2126                 } else {
2127                         /* If it's an output descriptor, they're all supposed
2128                          * to come before any input descriptors. */
2129                         if (unlikely(*in_num)) {
2130                                 vq_err(vq, "Descriptor has out after in: "
2131                                        "idx %d\n", i);
2132                                 return -EINVAL;
2133                         }
2134                         *out_num += ret;
2135                 }
2136         } while ((i = next_desc(vq, &desc)) != -1);
2137
2138         /* On success, increment avail index. */
2139         vq->last_avail_idx++;
2140
2141         /* Assume notifications from guest are disabled at this point,
2142          * if they aren't we would need to update avail_event index. */
2143         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
2144         return head;
2145 }
2146 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
2147
2148 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
2149 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
2150 {
2151         vq->last_avail_idx -= n;
2152 }
2153 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
2154
2155 /* After we've used one of their buffers, we tell them about it.  We'll then
2156  * want to notify the guest, using eventfd. */
2157 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
2158 {
2159         struct vring_used_elem heads = {
2160                 cpu_to_vhost32(vq, head),
2161                 cpu_to_vhost32(vq, len)
2162         };
2163
2164         return vhost_add_used_n(vq, &heads, 1);
2165 }
2166 EXPORT_SYMBOL_GPL(vhost_add_used);
2167
2168 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
2169                             struct vring_used_elem *heads,
2170                             unsigned count)
2171 {
2172         struct vring_used_elem __user *used;
2173         u16 old, new;
2174         int start;
2175
2176         start = vq->last_used_idx & (vq->num - 1);
2177         used = vq->used->ring + start;
2178         if (count == 1) {
2179                 if (vhost_put_user(vq, heads[0].id, &used->id)) {
2180                         vq_err(vq, "Failed to write used id");
2181                         return -EFAULT;
2182                 }
2183                 if (vhost_put_user(vq, heads[0].len, &used->len)) {
2184                         vq_err(vq, "Failed to write used len");
2185                         return -EFAULT;
2186                 }
2187         } else if (vhost_copy_to_user(vq, used, heads, count * sizeof *used)) {
2188                 vq_err(vq, "Failed to write used");
2189                 return -EFAULT;
2190         }
2191         if (unlikely(vq->log_used)) {
2192                 /* Make sure data is seen before log. */
2193                 smp_wmb();
2194                 /* Log used ring entry write. */
2195                 log_write(vq->log_base,
2196                           vq->log_addr +
2197                            ((void __user *)used - (void __user *)vq->used),
2198                           count * sizeof *used);
2199         }
2200         old = vq->last_used_idx;
2201         new = (vq->last_used_idx += count);
2202         /* If the driver never bothers to signal in a very long while,
2203          * used index might wrap around. If that happens, invalidate
2204          * signalled_used index we stored. TODO: make sure driver
2205          * signals at least once in 2^16 and remove this. */
2206         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
2207                 vq->signalled_used_valid = false;
2208         return 0;
2209 }
2210
2211 /* After we've used one of their buffers, we tell them about it.  We'll then
2212  * want to notify the guest, using eventfd. */
2213 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
2214                      unsigned count)
2215 {
2216         int start, n, r;
2217
2218         start = vq->last_used_idx & (vq->num - 1);
2219         n = vq->num - start;
2220         if (n < count) {
2221                 r = __vhost_add_used_n(vq, heads, n);
2222                 if (r < 0)
2223                         return r;
2224                 heads += n;
2225                 count -= n;
2226         }
2227         r = __vhost_add_used_n(vq, heads, count);
2228
2229         /* Make sure buffer is written before we update index. */
2230         smp_wmb();
2231         if (vhost_put_user(vq, cpu_to_vhost16(vq, vq->last_used_idx),
2232                            &vq->used->idx)) {
2233                 vq_err(vq, "Failed to increment used idx");
2234                 return -EFAULT;
2235         }
2236         if (unlikely(vq->log_used)) {
2237                 /* Log used index update. */
2238                 log_write(vq->log_base,
2239                           vq->log_addr + offsetof(struct vring_used, idx),
2240                           sizeof vq->used->idx);
2241                 if (vq->log_ctx)
2242                         eventfd_signal(vq->log_ctx, 1);
2243         }
2244         return r;
2245 }
2246 EXPORT_SYMBOL_GPL(vhost_add_used_n);
2247
2248 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2249 {
2250         __u16 old, new;
2251         __virtio16 event;
2252         bool v;
2253
2254         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
2255             unlikely(vq->avail_idx == vq->last_avail_idx))
2256                 return true;
2257
2258         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2259                 __virtio16 flags;
2260                 /* Flush out used index updates. This is paired
2261                  * with the barrier that the Guest executes when enabling
2262                  * interrupts. */
2263                 smp_mb();
2264                 if (vhost_get_avail(vq, flags, &vq->avail->flags)) {
2265                         vq_err(vq, "Failed to get flags");
2266                         return true;
2267                 }
2268                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
2269         }
2270         old = vq->signalled_used;
2271         v = vq->signalled_used_valid;
2272         new = vq->signalled_used = vq->last_used_idx;
2273         vq->signalled_used_valid = true;
2274
2275         if (unlikely(!v))
2276                 return true;
2277
2278         /* We're sure if the following conditions are met, there's no
2279          * need to notify guest:
2280          * 1) cached used event is ahead of new
2281          * 2) old to new updating does not cross cached used event. */
2282         if (vring_need_event(vq->last_used_event, new + vq->num, new) &&
2283             !vring_need_event(vq->last_used_event, new, old))
2284                 return false;
2285
2286         /* Flush out used index updates. This is paired
2287          * with the barrier that the Guest executes when enabling
2288          * interrupts. */
2289         smp_mb();
2290
2291         if (vhost_get_avail(vq, event, vhost_used_event(vq))) {
2292                 vq_err(vq, "Failed to get used event idx");
2293                 return true;
2294         }
2295         vq->last_used_event = vhost16_to_cpu(vq, event);
2296
2297         return vring_need_event(vq->last_used_event, new, old);
2298 }
2299
2300 /* This actually signals the guest, using eventfd. */
2301 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2302 {
2303         /* Signal the Guest tell them we used something up. */
2304         if (vq->call_ctx && vhost_notify(dev, vq))
2305                 eventfd_signal(vq->call_ctx, 1);
2306 }
2307 EXPORT_SYMBOL_GPL(vhost_signal);
2308
2309 /* And here's the combo meal deal.  Supersize me! */
2310 void vhost_add_used_and_signal(struct vhost_dev *dev,
2311                                struct vhost_virtqueue *vq,
2312                                unsigned int head, int len)
2313 {
2314         vhost_add_used(vq, head, len);
2315         vhost_signal(dev, vq);
2316 }
2317 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
2318
2319 /* multi-buffer version of vhost_add_used_and_signal */
2320 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
2321                                  struct vhost_virtqueue *vq,
2322                                  struct vring_used_elem *heads, unsigned count)
2323 {
2324         vhost_add_used_n(vq, heads, count);
2325         vhost_signal(dev, vq);
2326 }
2327 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
2328
2329 /* return true if we're sure that avaiable ring is empty */
2330 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2331 {
2332         __virtio16 avail_idx;
2333         int r;
2334
2335         if (vq->avail_idx != vq->last_avail_idx)
2336                 return false;
2337
2338         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2339         if (unlikely(r))
2340                 return false;
2341         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
2342
2343         return vq->avail_idx == vq->last_avail_idx;
2344 }
2345 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
2346
2347 /* OK, now we need to know about added descriptors. */
2348 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2349 {
2350         __virtio16 avail_idx;
2351         int r;
2352
2353         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
2354                 return false;
2355         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
2356         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2357                 r = vhost_update_used_flags(vq);
2358                 if (r) {
2359                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2360                                &vq->used->flags, r);
2361                         return false;
2362                 }
2363         } else {
2364                 r = vhost_update_avail_event(vq, vq->avail_idx);
2365                 if (r) {
2366                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
2367                                vhost_avail_event(vq), r);
2368                         return false;
2369                 }
2370         }
2371         /* They could have slipped one in as we were doing that: make
2372          * sure it's written, then check again. */
2373         smp_mb();
2374         r = vhost_get_avail(vq, avail_idx, &vq->avail->idx);
2375         if (r) {
2376                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
2377                        &vq->avail->idx, r);
2378                 return false;
2379         }
2380
2381         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
2382 }
2383 EXPORT_SYMBOL_GPL(vhost_enable_notify);
2384
2385 /* We don't need to be notified again. */
2386 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
2387 {
2388         int r;
2389
2390         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
2391                 return;
2392         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
2393         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
2394                 r = vhost_update_used_flags(vq);
2395                 if (r)
2396                         vq_err(vq, "Failed to enable notification at %p: %d\n",
2397                                &vq->used->flags, r);
2398         }
2399 }
2400 EXPORT_SYMBOL_GPL(vhost_disable_notify);
2401
2402 /* Create a new message. */
2403 struct vhost_msg_node *vhost_new_msg(struct vhost_virtqueue *vq, int type)
2404 {
2405         struct vhost_msg_node *node = kmalloc(sizeof *node, GFP_KERNEL);
2406         if (!node)
2407                 return NULL;
2408         node->vq = vq;
2409         node->msg.type = type;
2410         return node;
2411 }
2412 EXPORT_SYMBOL_GPL(vhost_new_msg);
2413
2414 void vhost_enqueue_msg(struct vhost_dev *dev, struct list_head *head,
2415                        struct vhost_msg_node *node)
2416 {
2417         spin_lock(&dev->iotlb_lock);
2418         list_add_tail(&node->node, head);
2419         spin_unlock(&dev->iotlb_lock);
2420
2421         wake_up_interruptible_poll(&dev->wait, POLLIN | POLLRDNORM);
2422 }
2423 EXPORT_SYMBOL_GPL(vhost_enqueue_msg);
2424
2425 struct vhost_msg_node *vhost_dequeue_msg(struct vhost_dev *dev,
2426                                          struct list_head *head)
2427 {
2428         struct vhost_msg_node *node = NULL;
2429
2430         spin_lock(&dev->iotlb_lock);
2431         if (!list_empty(head)) {
2432                 node = list_first_entry(head, struct vhost_msg_node,
2433                                         node);
2434                 list_del(&node->node);
2435         }
2436         spin_unlock(&dev->iotlb_lock);
2437
2438         return node;
2439 }
2440 EXPORT_SYMBOL_GPL(vhost_dequeue_msg);
2441
2442
2443 static int __init vhost_init(void)
2444 {
2445         return 0;
2446 }
2447
2448 static void __exit vhost_exit(void)
2449 {
2450 }
2451
2452 module_init(vhost_init);
2453 module_exit(vhost_exit);
2454
2455 MODULE_VERSION("0.0.1");
2456 MODULE_LICENSE("GPL v2");
2457 MODULE_AUTHOR("Michael S. Tsirkin");
2458 MODULE_DESCRIPTION("Host kernel accelerator for virtio");