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
  31 #include "vhost.h"
  32
  33 static ushort max_mem_regions = 64;
  34 module_param(max_mem_regions, ushort, 0444);
  35 MODULE_PARM_DESC(max_mem_regions,
  36         "Maximum number of memory regions in memory map. (default: 64)");
  37
  38 enum {
  39         VHOST_MEMORY_F_LOG = 0x1,
  40 };
  41
  42 #define vhost_used_event(vq) ((__virtio16 __user *)&vq->avail->ring[vq->num])
  43 #define vhost_avail_event(vq) ((__virtio16 __user *)&vq->used->ring[vq->num])
  44
  45 #ifdef CONFIG_VHOST_CROSS_ENDIAN_LEGACY
  46 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
  47 {
  48         vq->user_be = !virtio_legacy_is_little_endian();
  49 }
  50
  51 static void vhost_enable_cross_endian_big(struct vhost_virtqueue *vq)
  52 {
  53         vq->user_be = true;
  54 }
  55
  56 static void vhost_enable_cross_endian_little(struct vhost_virtqueue *vq)
  57 {
  58         vq->user_be = false;
  59 }
  60
  61 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
  62 {
  63         struct vhost_vring_state s;
  64
  65         if (vq->private_data)
  66                 return -EBUSY;
  67
  68         if (copy_from_user(&s, argp, sizeof(s)))
  69                 return -EFAULT;
  70
  71         if (s.num != VHOST_VRING_LITTLE_ENDIAN &&
  72             s.num != VHOST_VRING_BIG_ENDIAN)
  73                 return -EINVAL;
  74
  75         if (s.num == VHOST_VRING_BIG_ENDIAN)
  76                 vhost_enable_cross_endian_big(vq);
  77         else
  78                 vhost_enable_cross_endian_little(vq);
  79
  80         return 0;
  81 }
  82
  83 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
  84                                    int __user *argp)
  85 {
  86         struct vhost_vring_state s = {
  87                 .index = idx,
  88                 .num = vq->user_be
  89         };
  90
  91         if (copy_to_user(argp, &s, sizeof(s)))
  92                 return -EFAULT;
  93
  94         return 0;
  95 }
  96
  97 static void vhost_init_is_le(struct vhost_virtqueue *vq)
  98 {
  99         /* Note for legacy virtio: user_be is initialized at reset time
 100          * according to the host endianness. If userspace does not set an
 101          * explicit endianness, the default behavior is native endian, as
 102          * expected by legacy virtio.
 103          */
 104         vq->is_le = vhost_has_feature(vq, VIRTIO_F_VERSION_1) || !vq->user_be;
 105 }
 106 #else
 107 static void vhost_disable_cross_endian(struct vhost_virtqueue *vq)
 108 {
 109 }
 110
 111 static long vhost_set_vring_endian(struct vhost_virtqueue *vq, int __user *argp)
 112 {
 113         return -ENOIOCTLCMD;
 114 }
 115
 116 static long vhost_get_vring_endian(struct vhost_virtqueue *vq, u32 idx,
 117                                    int __user *argp)
 118 {
 119         return -ENOIOCTLCMD;
 120 }
 121
 122 static void vhost_init_is_le(struct vhost_virtqueue *vq)
 123 {
 124         if (vhost_has_feature(vq, VIRTIO_F_VERSION_1))
 125                 vq->is_le = true;
 126 }
 127 #endif /* CONFIG_VHOST_CROSS_ENDIAN_LEGACY */
 128
 129 static void vhost_reset_is_le(struct vhost_virtqueue *vq)
 130 {
 131         vq->is_le = virtio_legacy_is_little_endian();
 132 }
 133
 134 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
 135                             poll_table *pt)
 136 {
 137         struct vhost_poll *poll;
 138
 139         poll = container_of(pt, struct vhost_poll, table);
 140         poll->wqh = wqh;
 141         add_wait_queue(wqh, &poll->wait);
 142 }
 143
 144 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
 145                              void *key)
 146 {
 147         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
 148
 149         if (!((unsigned long)key & poll->mask))
 150                 return 0;
 151
 152         vhost_poll_queue(poll);
 153         return 0;
 154 }
 155
 156 void vhost_work_init(struct vhost_work *work, vhost_work_fn_t fn)
 157 {
 158         INIT_LIST_HEAD(&work->node);
 159         work->fn = fn;
 160         init_waitqueue_head(&work->done);
 161         work->flushing = 0;
 162         work->queue_seq = work->done_seq = 0;
 163 }
 164 EXPORT_SYMBOL_GPL(vhost_work_init);
 165
 166 /* Init poll structure */
 167 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
 168                      unsigned long mask, struct vhost_dev *dev)
 169 {
 170         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
 171         init_poll_funcptr(&poll->table, vhost_poll_func);
 172         poll->mask = mask;
 173         poll->dev = dev;
 174         poll->wqh = NULL;
 175
 176         vhost_work_init(&poll->work, fn);
 177 }
 178 EXPORT_SYMBOL_GPL(vhost_poll_init);
 179
 180 /* Start polling a file. We add ourselves to file's wait queue. The caller must
 181  * keep a reference to a file until after vhost_poll_stop is called. */
 182 int vhost_poll_start(struct vhost_poll *poll, struct file *file)
 183 {
 184         unsigned long mask;
 185         int ret = 0;
 186
 187         if (poll->wqh)
 188                 return 0;
 189
 190         mask = file->f_op->poll(file, &poll->table);
 191         if (mask)
 192                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
 193         if (mask & POLLERR) {
 194                 if (poll->wqh)
 195                         remove_wait_queue(poll->wqh, &poll->wait);
 196                 ret = -EINVAL;
 197         }
 198
 199         return ret;
 200 }
 201 EXPORT_SYMBOL_GPL(vhost_poll_start);
 202
 203 /* Stop polling a file. After this function returns, it becomes safe to drop the
 204  * file reference. You must also flush afterwards. */
 205 void vhost_poll_stop(struct vhost_poll *poll)
 206 {
 207         if (poll->wqh) {
 208                 remove_wait_queue(poll->wqh, &poll->wait);
 209                 poll->wqh = NULL;
 210         }
 211 }
 212 EXPORT_SYMBOL_GPL(vhost_poll_stop);
 213
 214 static bool vhost_work_seq_done(struct vhost_dev *dev, struct vhost_work *work,
 215                                 unsigned seq)
 216 {
 217         int left;
 218
 219         spin_lock_irq(&dev->work_lock);
 220         left = seq - work->done_seq;
 221         spin_unlock_irq(&dev->work_lock);
 222         return left <= 0;
 223 }
 224
 225 void vhost_work_flush(struct vhost_dev *dev, struct vhost_work *work)
 226 {
 227         unsigned seq;
 228         int flushing;
 229
 230         spin_lock_irq(&dev->work_lock);
 231         seq = work->queue_seq;
 232         work->flushing++;
 233         spin_unlock_irq(&dev->work_lock);
 234         wait_event(work->done, vhost_work_seq_done(dev, work, seq));
 235         spin_lock_irq(&dev->work_lock);
 236         flushing = --work->flushing;
 237         spin_unlock_irq(&dev->work_lock);
 238         BUG_ON(flushing < 0);
 239 }
 240 EXPORT_SYMBOL_GPL(vhost_work_flush);
 241
 242 /* Flush any work that has been scheduled. When calling this, don't hold any
 243  * locks that are also used by the callback. */
 244 void vhost_poll_flush(struct vhost_poll *poll)
 245 {
 246         vhost_work_flush(poll->dev, &poll->work);
 247 }
 248 EXPORT_SYMBOL_GPL(vhost_poll_flush);
 249
 250 void vhost_work_queue(struct vhost_dev *dev, struct vhost_work *work)
 251 {
 252         unsigned long flags;
 253
 254         spin_lock_irqsave(&dev->work_lock, flags);
 255         if (list_empty(&work->node)) {
 256                 list_add_tail(&work->node, &dev->work_list);
 257                 work->queue_seq++;
 258                 spin_unlock_irqrestore(&dev->work_lock, flags);
 259                 wake_up_process(dev->worker);
 260         } else {
 261                 spin_unlock_irqrestore(&dev->work_lock, flags);
 262         }
 263 }
 264 EXPORT_SYMBOL_GPL(vhost_work_queue);
 265
 266 /* A lockless hint for busy polling code to exit the loop */
 267 bool vhost_has_work(struct vhost_dev *dev)
 268 {
 269         return !list_empty(&dev->work_list);
 270 }
 271 EXPORT_SYMBOL_GPL(vhost_has_work);
 272
 273 void vhost_poll_queue(struct vhost_poll *poll)
 274 {
 275         vhost_work_queue(poll->dev, &poll->work);
 276 }
 277 EXPORT_SYMBOL_GPL(vhost_poll_queue);
 278
 279 static void vhost_vq_reset(struct vhost_dev *dev,
 280                            struct vhost_virtqueue *vq)
 281 {
 282         vq->num = 1;
 283         vq->desc = NULL;
 284         vq->avail = NULL;
 285         vq->used = NULL;
 286         vq->last_avail_idx = 0;
 287         vq->avail_idx = 0;
 288         vq->last_used_idx = 0;
 289         vq->signalled_used = 0;
 290         vq->signalled_used_valid = false;
 291         vq->used_flags = 0;
 292         vq->log_used = false;
 293         vq->log_addr = -1ull;
 294         vq->private_data = NULL;
 295         vq->acked_features = 0;
 296         vq->log_base = NULL;
 297         vq->error_ctx = NULL;
 298         vq->error = NULL;
 299         vq->kick = NULL;
 300         vq->call_ctx = NULL;
 301         vq->call = NULL;
 302         vq->log_ctx = NULL;
 303         vq->memory = NULL;
 304         vhost_reset_is_le(vq);
 305         vhost_disable_cross_endian(vq);
 306         vq->busyloop_timeout = 0;
 307 }
 308
 309 static int vhost_worker(void *data)
 310 {
 311         struct vhost_dev *dev = data;
 312         struct vhost_work *work = NULL;
 313         unsigned uninitialized_var(seq);
 314         mm_segment_t oldfs = get_fs();
 315
 316         set_fs(USER_DS);
 317         use_mm(dev->mm);
 318
 319         for (;;) {
 320                 /* mb paired w/ kthread_stop */
 321                 set_current_state(TASK_INTERRUPTIBLE);
 322
 323                 spin_lock_irq(&dev->work_lock);
 324                 if (work) {
 325                         work->done_seq = seq;
 326                         if (work->flushing)
 327                                 wake_up_all(&work->done);
 328                 }
 329
 330                 if (kthread_should_stop()) {
 331                         spin_unlock_irq(&dev->work_lock);
 332                         __set_current_state(TASK_RUNNING);
 333                         break;
 334                 }
 335                 if (!list_empty(&dev->work_list)) {
 336                         work = list_first_entry(&dev->work_list,
 337                                                 struct vhost_work, node);
 338                         list_del_init(&work->node);
 339                         seq = work->queue_seq;
 340                 } else
 341                         work = NULL;
 342                 spin_unlock_irq(&dev->work_lock);
 343
 344                 if (work) {
 345                         __set_current_state(TASK_RUNNING);
 346                         work->fn(work);
 347                         if (need_resched())
 348                                 schedule();
 349                 } else
 350                         schedule();
 351
 352         }
 353         unuse_mm(dev->mm);
 354         set_fs(oldfs);
 355         return 0;
 356 }
 357
 358 static void vhost_vq_free_iovecs(struct vhost_virtqueue *vq)
 359 {
 360         kfree(vq->indirect);
 361         vq->indirect = NULL;
 362         kfree(vq->log);
 363         vq->log = NULL;
 364         kfree(vq->heads);
 365         vq->heads = NULL;
 366 }
 367
 368 /* Helper to allocate iovec buffers for all vqs. */
 369 static long vhost_dev_alloc_iovecs(struct vhost_dev *dev)
 370 {
 371         struct vhost_virtqueue *vq;
 372         int i;
 373
 374         for (i = 0; i < dev->nvqs; ++i) {
 375                 vq = dev->vqs[i];
 376                 vq->indirect = kmalloc(sizeof *vq->indirect * UIO_MAXIOV,
 377                                        GFP_KERNEL);
 378                 vq->log = kmalloc(sizeof *vq->log * UIO_MAXIOV, GFP_KERNEL);
 379                 vq->heads = kmalloc(sizeof *vq->heads * UIO_MAXIOV, GFP_KERNEL);
 380                 if (!vq->indirect || !vq->log || !vq->heads)
 381                         goto err_nomem;
 382         }
 383         return 0;
 384
 385 err_nomem:
 386         for (; i >= 0; --i)
 387                 vhost_vq_free_iovecs(dev->vqs[i]);
 388         return -ENOMEM;
 389 }
 390
 391 static void vhost_dev_free_iovecs(struct vhost_dev *dev)
 392 {
 393         int i;
 394
 395         for (i = 0; i < dev->nvqs; ++i)
 396                 vhost_vq_free_iovecs(dev->vqs[i]);
 397 }
 398
 399 void vhost_dev_init(struct vhost_dev *dev,
 400                     struct vhost_virtqueue **vqs, int nvqs)
 401 {
 402         struct vhost_virtqueue *vq;
 403         int i;
 404
 405         dev->vqs = vqs;
 406         dev->nvqs = nvqs;
 407         mutex_init(&dev->mutex);
 408         dev->log_ctx = NULL;
 409         dev->log_file = NULL;
 410         dev->memory = NULL;
 411         dev->mm = NULL;
 412         spin_lock_init(&dev->work_lock);
 413         INIT_LIST_HEAD(&dev->work_list);
 414         dev->worker = NULL;
 415
 416         for (i = 0; i < dev->nvqs; ++i) {
 417                 vq = dev->vqs[i];
 418                 vq->log = NULL;
 419                 vq->indirect = NULL;
 420                 vq->heads = NULL;
 421                 vq->dev = dev;
 422                 mutex_init(&vq->mutex);
 423                 vhost_vq_reset(dev, vq);
 424                 if (vq->handle_kick)
 425                         vhost_poll_init(&vq->poll, vq->handle_kick,
 426                                         POLLIN, dev);
 427         }
 428 }
 429 EXPORT_SYMBOL_GPL(vhost_dev_init);
 430
 431 /* Caller should have device mutex */
 432 long vhost_dev_check_owner(struct vhost_dev *dev)
 433 {
 434         /* Are you the owner? If not, I don't think you mean to do that */
 435         return dev->mm == current->mm ? 0 : -EPERM;
 436 }
 437 EXPORT_SYMBOL_GPL(vhost_dev_check_owner);
 438
 439 struct vhost_attach_cgroups_struct {
 440         struct vhost_work work;
 441         struct task_struct *owner;
 442         int ret;
 443 };
 444
 445 static void vhost_attach_cgroups_work(struct vhost_work *work)
 446 {
 447         struct vhost_attach_cgroups_struct *s;
 448
 449         s = container_of(work, struct vhost_attach_cgroups_struct, work);
 450         s->ret = cgroup_attach_task_all(s->owner, current);
 451 }
 452
 453 static int vhost_attach_cgroups(struct vhost_dev *dev)
 454 {
 455         struct vhost_attach_cgroups_struct attach;
 456
 457         attach.owner = current;
 458         vhost_work_init(&attach.work, vhost_attach_cgroups_work);
 459         vhost_work_queue(dev, &attach.work);
 460         vhost_work_flush(dev, &attach.work);
 461         return attach.ret;
 462 }
 463
 464 /* Caller should have device mutex */
 465 bool vhost_dev_has_owner(struct vhost_dev *dev)
 466 {
 467         return dev->mm;
 468 }
 469 EXPORT_SYMBOL_GPL(vhost_dev_has_owner);
 470
 471 /* Caller should have device mutex */
 472 long vhost_dev_set_owner(struct vhost_dev *dev)
 473 {
 474         struct task_struct *worker;
 475         int err;
 476
 477         /* Is there an owner already? */
 478         if (vhost_dev_has_owner(dev)) {
 479                 err = -EBUSY;
 480                 goto err_mm;
 481         }
 482
 483         /* No owner, become one */
 484         dev->mm = get_task_mm(current);
 485         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 486         if (IS_ERR(worker)) {
 487                 err = PTR_ERR(worker);
 488                 goto err_worker;
 489         }
 490
 491         dev->worker = worker;
 492         wake_up_process(worker);        /* avoid contributing to loadavg */
 493
 494         err = vhost_attach_cgroups(dev);
 495         if (err)
 496                 goto err_cgroup;
 497
 498         err = vhost_dev_alloc_iovecs(dev);
 499         if (err)
 500                 goto err_cgroup;
 501
 502         return 0;
 503 err_cgroup:
 504         kthread_stop(worker);
 505         dev->worker = NULL;
 506 err_worker:
 507         if (dev->mm)
 508                 mmput(dev->mm);
 509         dev->mm = NULL;
 510 err_mm:
 511         return err;
 512 }
 513 EXPORT_SYMBOL_GPL(vhost_dev_set_owner);
 514
 515 struct vhost_memory *vhost_dev_reset_owner_prepare(void)
 516 {
 517         return kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 518 }
 519 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner_prepare);
 520
 521 /* Caller should have device mutex */
 522 void vhost_dev_reset_owner(struct vhost_dev *dev, struct vhost_memory *memory)
 523 {
 524         int i;
 525
 526         vhost_dev_cleanup(dev, true);
 527
 528         /* Restore memory to default empty mapping. */
 529         memory->nregions = 0;
 530         dev->memory = memory;
 531         /* We don't need VQ locks below since vhost_dev_cleanup makes sure
 532          * VQs aren't running.
 533          */
 534         for (i = 0; i < dev->nvqs; ++i)
 535                 dev->vqs[i]->memory = memory;
 536 }
 537 EXPORT_SYMBOL_GPL(vhost_dev_reset_owner);
 538
 539 void vhost_dev_stop(struct vhost_dev *dev)
 540 {
 541         int i;
 542
 543         for (i = 0; i < dev->nvqs; ++i) {
 544                 if (dev->vqs[i]->kick && dev->vqs[i]->handle_kick) {
 545                         vhost_poll_stop(&dev->vqs[i]->poll);
 546                         vhost_poll_flush(&dev->vqs[i]->poll);
 547                 }
 548         }
 549 }
 550 EXPORT_SYMBOL_GPL(vhost_dev_stop);
 551
 552 /* Caller should have device mutex if and only if locked is set */
 553 void vhost_dev_cleanup(struct vhost_dev *dev, bool locked)
 554 {
 555         int i;
 556
 557         for (i = 0; i < dev->nvqs; ++i) {
 558                 if (dev->vqs[i]->error_ctx)
 559                         eventfd_ctx_put(dev->vqs[i]->error_ctx);
 560                 if (dev->vqs[i]->error)
 561                         fput(dev->vqs[i]->error);
 562                 if (dev->vqs[i]->kick)
 563                         fput(dev->vqs[i]->kick);
 564                 if (dev->vqs[i]->call_ctx)
 565                         eventfd_ctx_put(dev->vqs[i]->call_ctx);
 566                 if (dev->vqs[i]->call)
 567                         fput(dev->vqs[i]->call);
 568                 vhost_vq_reset(dev, dev->vqs[i]);
 569         }
 570         vhost_dev_free_iovecs(dev);
 571         if (dev->log_ctx)
 572                 eventfd_ctx_put(dev->log_ctx);
 573         dev->log_ctx = NULL;
 574         if (dev->log_file)
 575                 fput(dev->log_file);
 576         dev->log_file = NULL;
 577         /* No one will access memory at this point */
 578         kvfree(dev->memory);
 579         dev->memory = NULL;
 580         WARN_ON(!list_empty(&dev->work_list));
 581         if (dev->worker) {
 582                 kthread_stop(dev->worker);
 583                 dev->worker = NULL;
 584         }
 585         if (dev->mm)
 586                 mmput(dev->mm);
 587         dev->mm = NULL;
 588 }
 589 EXPORT_SYMBOL_GPL(vhost_dev_cleanup);
 590
 591 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 592 {
 593         u64 a = addr / VHOST_PAGE_SIZE / 8;
 594
 595         /* Make sure 64 bit math will not overflow. */
 596         if (a > ULONG_MAX - (unsigned long)log_base ||
 597             a + (unsigned long)log_base > ULONG_MAX)
 598                 return 0;
 599
 600         return access_ok(VERIFY_WRITE, log_base + a,
 601                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 602 }
 603
 604 /* Caller should have vq mutex and device mutex. */
 605 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 606                                int log_all)
 607 {
 608         int i;
 609
 610         if (!mem)
 611                 return 0;
 612
 613         for (i = 0; i < mem->nregions; ++i) {
 614                 struct vhost_memory_region *m = mem->regions + i;
 615                 unsigned long a = m->userspace_addr;
 616                 if (m->memory_size > ULONG_MAX)
 617                         return 0;
 618                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 619                                     m->memory_size))
 620                         return 0;
 621                 else if (log_all && !log_access_ok(log_base,
 622                                                    m->guest_phys_addr,
 623                                                    m->memory_size))
 624                         return 0;
 625         }
 626         return 1;
 627 }
 628
 629 /* Can we switch to this memory table? */
 630 /* Caller should have device mutex but not vq mutex */
 631 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 632                             int log_all)
 633 {
 634         int i;
 635
 636         for (i = 0; i < d->nvqs; ++i) {
 637                 int ok;
 638                 bool log;
 639
 640                 mutex_lock(&d->vqs[i]->mutex);
 641                 log = log_all || vhost_has_feature(d->vqs[i], VHOST_F_LOG_ALL);
 642                 /* If ring is inactive, will check when it's enabled. */
 643                 if (d->vqs[i]->private_data)
 644                         ok = vq_memory_access_ok(d->vqs[i]->log_base, mem, log);
 645                 else
 646                         ok = 1;
 647                 mutex_unlock(&d->vqs[i]->mutex);
 648                 if (!ok)
 649                         return 0;
 650         }
 651         return 1;
 652 }
 653
 654 static int vq_access_ok(struct vhost_virtqueue *vq, unsigned int num,
 655                         struct vring_desc __user *desc,
 656                         struct vring_avail __user *avail,
 657                         struct vring_used __user *used)
 658 {
 659         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 660         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 661                access_ok(VERIFY_READ, avail,
 662                          sizeof *avail + num * sizeof *avail->ring + s) &&
 663                access_ok(VERIFY_WRITE, used,
 664                         sizeof *used + num * sizeof *used->ring + s);
 665 }
 666
 667 /* Can we log writes? */
 668 /* Caller should have device mutex but not vq mutex */
 669 int vhost_log_access_ok(struct vhost_dev *dev)
 670 {
 671         return memory_access_ok(dev, dev->memory, 1);
 672 }
 673 EXPORT_SYMBOL_GPL(vhost_log_access_ok);
 674
 675 /* Verify access for write logging. */
 676 /* Caller should have vq mutex and device mutex */
 677 static int vq_log_access_ok(struct vhost_virtqueue *vq,
 678                             void __user *log_base)
 679 {
 680         size_t s = vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX) ? 2 : 0;
 681
 682         return vq_memory_access_ok(log_base, vq->memory,
 683                                    vhost_has_feature(vq, VHOST_F_LOG_ALL)) &&
 684                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 685                                         sizeof *vq->used +
 686                                         vq->num * sizeof *vq->used->ring + s));
 687 }
 688
 689 /* Can we start vq? */
 690 /* Caller should have vq mutex and device mutex */
 691 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 692 {
 693         return vq_access_ok(vq, vq->num, vq->desc, vq->avail, vq->used) &&
 694                 vq_log_access_ok(vq, vq->log_base);
 695 }
 696 EXPORT_SYMBOL_GPL(vhost_vq_access_ok);
 697
 698 static int vhost_memory_reg_sort_cmp(const void *p1, const void *p2)
 699 {
 700         const struct vhost_memory_region *r1 = p1, *r2 = p2;
 701         if (r1->guest_phys_addr < r2->guest_phys_addr)
 702                 return 1;
 703         if (r1->guest_phys_addr > r2->guest_phys_addr)
 704                 return -1;
 705         return 0;
 706 }
 707
 708 static void *vhost_kvzalloc(unsigned long size)
 709 {
 710         void *n = kzalloc(size, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 711
 712         if (!n)
 713                 n = vzalloc(size);
 714         return n;
 715 }
 716
 717 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 718 {
 719         struct vhost_memory mem, *newmem, *oldmem;
 720         unsigned long size = offsetof(struct vhost_memory, regions);
 721         int i;
 722
 723         if (copy_from_user(&mem, m, size))
 724                 return -EFAULT;
 725         if (mem.padding)
 726                 return -EOPNOTSUPP;
 727         if (mem.nregions > max_mem_regions)
 728                 return -E2BIG;
 729         newmem = vhost_kvzalloc(size + mem.nregions * sizeof(*m->regions));
 730         if (!newmem)
 731                 return -ENOMEM;
 732
 733         memcpy(newmem, &mem, size);
 734         if (copy_from_user(newmem->regions, m->regions,
 735                            mem.nregions * sizeof *m->regions)) {
 736                 kvfree(newmem);
 737                 return -EFAULT;
 738         }
 739         sort(newmem->regions, newmem->nregions, sizeof(*newmem->regions),
 740                 vhost_memory_reg_sort_cmp, NULL);
 741
 742         if (!memory_access_ok(d, newmem, 0)) {
 743                 kvfree(newmem);
 744                 return -EFAULT;
 745         }
 746         oldmem = d->memory;
 747         d->memory = newmem;
 748
 749         /* All memory accesses are done under some VQ mutex. */
 750         for (i = 0; i < d->nvqs; ++i) {
 751                 mutex_lock(&d->vqs[i]->mutex);
 752                 d->vqs[i]->memory = newmem;
 753                 mutex_unlock(&d->vqs[i]->mutex);
 754         }
 755         kvfree(oldmem);
 756         return 0;
 757 }
 758
 759 long vhost_vring_ioctl(struct vhost_dev *d, int ioctl, void __user *argp)
 760 {
 761         struct file *eventfp, *filep = NULL;
 762         bool pollstart = false, pollstop = false;
 763         struct eventfd_ctx *ctx = NULL;
 764         u32 __user *idxp = argp;
 765         struct vhost_virtqueue *vq;
 766         struct vhost_vring_state s;
 767         struct vhost_vring_file f;
 768         struct vhost_vring_addr a;
 769         u32 idx;
 770         long r;
 771
 772         r = get_user(idx, idxp);
 773         if (r < 0)
 774                 return r;
 775         if (idx >= d->nvqs)
 776                 return -ENOBUFS;
 777
 778         vq = d->vqs[idx];
 779
 780         mutex_lock(&vq->mutex);
 781
 782         switch (ioctl) {
 783         case VHOST_SET_VRING_NUM:
 784                 /* Resizing ring with an active backend?
 785                  * You don't want to do that. */
 786                 if (vq->private_data) {
 787                         r = -EBUSY;
 788                         break;
 789                 }
 790                 if (copy_from_user(&s, argp, sizeof s)) {
 791                         r = -EFAULT;
 792                         break;
 793                 }
 794                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 795                         r = -EINVAL;
 796                         break;
 797                 }
 798                 vq->num = s.num;
 799                 break;
 800         case VHOST_SET_VRING_BASE:
 801                 /* Moving base with an active backend?
 802                  * You don't want to do that. */
 803                 if (vq->private_data) {
 804                         r = -EBUSY;
 805                         break;
 806                 }
 807                 if (copy_from_user(&s, argp, sizeof s)) {
 808                         r = -EFAULT;
 809                         break;
 810                 }
 811                 if (s.num > 0xffff) {
 812                         r = -EINVAL;
 813                         break;
 814                 }
 815                 vq->last_avail_idx = s.num;
 816                 /* Forget the cached index value. */
 817                 vq->avail_idx = vq->last_avail_idx;
 818                 break;
 819         case VHOST_GET_VRING_BASE:
 820                 s.index = idx;
 821                 s.num = vq->last_avail_idx;
 822                 if (copy_to_user(argp, &s, sizeof s))
 823                         r = -EFAULT;
 824                 break;
 825         case VHOST_SET_VRING_ADDR:
 826                 if (copy_from_user(&a, argp, sizeof a)) {
 827                         r = -EFAULT;
 828                         break;
 829                 }
 830                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 831                         r = -EOPNOTSUPP;
 832                         break;
 833                 }
 834                 /* For 32bit, verify that the top 32bits of the user
 835                    data are set to zero. */
 836                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 837                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 838                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 839                         r = -EFAULT;
 840                         break;
 841                 }
 842
 843                 /* Make sure it's safe to cast pointers to vring types. */
 844                 BUILD_BUG_ON(__alignof__ *vq->avail > VRING_AVAIL_ALIGN_SIZE);
 845                 BUILD_BUG_ON(__alignof__ *vq->used > VRING_USED_ALIGN_SIZE);
 846                 if ((a.avail_user_addr & (VRING_AVAIL_ALIGN_SIZE - 1)) ||
 847                     (a.used_user_addr & (VRING_USED_ALIGN_SIZE - 1)) ||
 848                     (a.log_guest_addr & (VRING_USED_ALIGN_SIZE - 1))) {
 849                         r = -EINVAL;
 850                         break;
 851                 }
 852
 853                 /* We only verify access here if backend is configured.
 854                  * If it is not, we don't as size might not have been setup.
 855                  * We will verify when backend is configured. */
 856                 if (vq->private_data) {
 857                         if (!vq_access_ok(vq, vq->num,
 858                                 (void __user *)(unsigned long)a.desc_user_addr,
 859                                 (void __user *)(unsigned long)a.avail_user_addr,
 860                                 (void __user *)(unsigned long)a.used_user_addr)) {
 861                                 r = -EINVAL;
 862                                 break;
 863                         }
 864
 865                         /* Also validate log access for used ring if enabled. */
 866                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 867                             !log_access_ok(vq->log_base, a.log_guest_addr,
 868                                            sizeof *vq->used +
 869                                            vq->num * sizeof *vq->used->ring)) {
 870                                 r = -EINVAL;
 871                                 break;
 872                         }
 873                 }
 874
 875                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 876                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 877                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 878                 vq->log_addr = a.log_guest_addr;
 879                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 880                 break;
 881         case VHOST_SET_VRING_KICK:
 882                 if (copy_from_user(&f, argp, sizeof f)) {
 883                         r = -EFAULT;
 884                         break;
 885                 }
 886                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 887                 if (IS_ERR(eventfp)) {
 888                         r = PTR_ERR(eventfp);
 889                         break;
 890                 }
 891                 if (eventfp != vq->kick) {
 892                         pollstop = (filep = vq->kick) != NULL;
 893                         pollstart = (vq->kick = eventfp) != NULL;
 894                 } else
 895                         filep = eventfp;
 896                 break;
 897         case VHOST_SET_VRING_CALL:
 898                 if (copy_from_user(&f, argp, sizeof f)) {
 899                         r = -EFAULT;
 900                         break;
 901                 }
 902                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 903                 if (IS_ERR(eventfp)) {
 904                         r = PTR_ERR(eventfp);
 905                         break;
 906                 }
 907                 if (eventfp != vq->call) {
 908                         filep = vq->call;
 909                         ctx = vq->call_ctx;
 910                         vq->call = eventfp;
 911                         vq->call_ctx = eventfp ?
 912                                 eventfd_ctx_fileget(eventfp) : NULL;
 913                 } else
 914                         filep = eventfp;
 915                 break;
 916         case VHOST_SET_VRING_ERR:
 917                 if (copy_from_user(&f, argp, sizeof f)) {
 918                         r = -EFAULT;
 919                         break;
 920                 }
 921                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 922                 if (IS_ERR(eventfp)) {
 923                         r = PTR_ERR(eventfp);
 924                         break;
 925                 }
 926                 if (eventfp != vq->error) {
 927                         filep = vq->error;
 928                         vq->error = eventfp;
 929                         ctx = vq->error_ctx;
 930                         vq->error_ctx = eventfp ?
 931                                 eventfd_ctx_fileget(eventfp) : NULL;
 932                 } else
 933                         filep = eventfp;
 934                 break;
 935         case VHOST_SET_VRING_ENDIAN:
 936                 r = vhost_set_vring_endian(vq, argp);
 937                 break;
 938         case VHOST_GET_VRING_ENDIAN:
 939                 r = vhost_get_vring_endian(vq, idx, argp);
 940                 break;
 941         case VHOST_SET_VRING_BUSYLOOP_TIMEOUT:
 942                 if (copy_from_user(&s, argp, sizeof(s))) {
 943                         r = -EFAULT;
 944                         break;
 945                 }
 946                 vq->busyloop_timeout = s.num;
 947                 break;
 948         case VHOST_GET_VRING_BUSYLOOP_TIMEOUT:
 949                 s.index = idx;
 950                 s.num = vq->busyloop_timeout;
 951                 if (copy_to_user(argp, &s, sizeof(s)))
 952                         r = -EFAULT;
 953                 break;
 954         default:
 955                 r = -ENOIOCTLCMD;
 956         }
 957
 958         if (pollstop && vq->handle_kick)
 959                 vhost_poll_stop(&vq->poll);
 960
 961         if (ctx)
 962                 eventfd_ctx_put(ctx);
 963         if (filep)
 964                 fput(filep);
 965
 966         if (pollstart && vq->handle_kick)
 967                 r = vhost_poll_start(&vq->poll, vq->kick);
 968
 969         mutex_unlock(&vq->mutex);
 970
 971         if (pollstop && vq->handle_kick)
 972                 vhost_poll_flush(&vq->poll);
 973         return r;
 974 }
 975 EXPORT_SYMBOL_GPL(vhost_vring_ioctl);
 976
 977 /* Caller must have device mutex */
 978 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, void __user *argp)
 979 {
 980         struct file *eventfp, *filep = NULL;
 981         struct eventfd_ctx *ctx = NULL;
 982         u64 p;
 983         long r;
 984         int i, fd;
 985
 986         /* If you are not the owner, you can become one */
 987         if (ioctl == VHOST_SET_OWNER) {
 988                 r = vhost_dev_set_owner(d);
 989                 goto done;
 990         }
 991
 992         /* You must be the owner to do anything else */
 993         r = vhost_dev_check_owner(d);
 994         if (r)
 995                 goto done;
 996
 997         switch (ioctl) {
 998         case VHOST_SET_MEM_TABLE:
 999                 r = vhost_set_memory(d, argp);
1000                 break;
1001         case VHOST_SET_LOG_BASE:
1002                 if (copy_from_user(&p, argp, sizeof p)) {
1003                         r = -EFAULT;
1004                         break;
1005                 }
1006                 if ((u64)(unsigned long)p != p) {
1007                         r = -EFAULT;
1008                         break;
1009                 }
1010                 for (i = 0; i < d->nvqs; ++i) {
1011                         struct vhost_virtqueue *vq;
1012                         void __user *base = (void __user *)(unsigned long)p;
1013                         vq = d->vqs[i];
1014                         mutex_lock(&vq->mutex);
1015                         /* If ring is inactive, will check when it's enabled. */
1016                         if (vq->private_data && !vq_log_access_ok(vq, base))
1017                                 r = -EFAULT;
1018                         else
1019                                 vq->log_base = base;
1020                         mutex_unlock(&vq->mutex);
1021                 }
1022                 break;
1023         case VHOST_SET_LOG_FD:
1024                 r = get_user(fd, (int __user *)argp);
1025                 if (r < 0)
1026                         break;
1027                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
1028                 if (IS_ERR(eventfp)) {
1029                         r = PTR_ERR(eventfp);
1030                         break;
1031                 }
1032                 if (eventfp != d->log_file) {
1033                         filep = d->log_file;
1034                         d->log_file = eventfp;
1035                         ctx = d->log_ctx;
1036                         d->log_ctx = eventfp ?
1037                                 eventfd_ctx_fileget(eventfp) : NULL;
1038                 } else
1039                         filep = eventfp;
1040                 for (i = 0; i < d->nvqs; ++i) {
1041                         mutex_lock(&d->vqs[i]->mutex);
1042                         d->vqs[i]->log_ctx = d->log_ctx;
1043                         mutex_unlock(&d->vqs[i]->mutex);
1044                 }
1045                 if (ctx)
1046                         eventfd_ctx_put(ctx);
1047                 if (filep)
1048                         fput(filep);
1049                 break;
1050         default:
1051                 r = -ENOIOCTLCMD;
1052                 break;
1053         }
1054 done:
1055         return r;
1056 }
1057 EXPORT_SYMBOL_GPL(vhost_dev_ioctl);
1058
1059 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
1060                                                      __u64 addr, __u32 len)
1061 {
1062         const struct vhost_memory_region *reg;
1063         int start = 0, end = mem->nregions;
1064
1065         while (start < end) {
1066                 int slot = start + (end - start) / 2;
1067                 reg = mem->regions + slot;
1068                 if (addr >= reg->guest_phys_addr)
1069                         end = slot;
1070                 else
1071                         start = slot + 1;
1072         }
1073
1074         reg = mem->regions + start;
1075         if (addr >= reg->guest_phys_addr &&
1076                 reg->guest_phys_addr + reg->memory_size > addr)
1077                 return reg;
1078         return NULL;
1079 }
1080
1081 /* TODO: This is really inefficient.  We need something like get_user()
1082  * (instruction directly accesses the data, with an exception table entry
1083  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
1084  */
1085 static int set_bit_to_user(int nr, void __user *addr)
1086 {
1087         unsigned long log = (unsigned long)addr;
1088         struct page *page;
1089         void *base;
1090         int bit = nr + (log % PAGE_SIZE) * 8;
1091         int r;
1092
1093         r = get_user_pages_fast(log, 1, 1, &page);
1094         if (r < 0)
1095                 return r;
1096         BUG_ON(r != 1);
1097         base = kmap_atomic(page);
1098         set_bit(bit, base);
1099         kunmap_atomic(base);
1100         set_page_dirty_lock(page);
1101         put_page(page);
1102         return 0;
1103 }
1104
1105 static int log_write(void __user *log_base,
1106                      u64 write_address, u64 write_length)
1107 {
1108         u64 write_page = write_address / VHOST_PAGE_SIZE;
1109         int r;
1110
1111         if (!write_length)
1112                 return 0;
1113         write_length += write_address % VHOST_PAGE_SIZE;
1114         for (;;) {
1115                 u64 base = (u64)(unsigned long)log_base;
1116                 u64 log = base + write_page / 8;
1117                 int bit = write_page % 8;
1118                 if ((u64)(unsigned long)log != log)
1119                         return -EFAULT;
1120                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
1121                 if (r < 0)
1122                         return r;
1123                 if (write_length <= VHOST_PAGE_SIZE)
1124                         break;
1125                 write_length -= VHOST_PAGE_SIZE;
1126                 write_page += 1;
1127         }
1128         return r;
1129 }
1130
1131 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
1132                     unsigned int log_num, u64 len)
1133 {
1134         int i, r;
1135
1136         /* Make sure data written is seen before log. */
1137         smp_wmb();
1138         for (i = 0; i < log_num; ++i) {
1139                 u64 l = min(log[i].len, len);
1140                 r = log_write(vq->log_base, log[i].addr, l);
1141                 if (r < 0)
1142                         return r;
1143                 len -= l;
1144                 if (!len) {
1145                         if (vq->log_ctx)
1146                                 eventfd_signal(vq->log_ctx, 1);
1147                         return 0;
1148                 }
1149         }
1150         /* Length written exceeds what we have stored. This is a bug. */
1151         BUG();
1152         return 0;
1153 }
1154 EXPORT_SYMBOL_GPL(vhost_log_write);
1155
1156 static int vhost_update_used_flags(struct vhost_virtqueue *vq)
1157 {
1158         void __user *used;
1159         if (__put_user(cpu_to_vhost16(vq, vq->used_flags), &vq->used->flags) < 0)
1160                 return -EFAULT;
1161         if (unlikely(vq->log_used)) {
1162                 /* Make sure the flag is seen before log. */
1163                 smp_wmb();
1164                 /* Log used flag write. */
1165                 used = &vq->used->flags;
1166                 log_write(vq->log_base, vq->log_addr +
1167                           (used - (void __user *)vq->used),
1168                           sizeof vq->used->flags);
1169                 if (vq->log_ctx)
1170                         eventfd_signal(vq->log_ctx, 1);
1171         }
1172         return 0;
1173 }
1174
1175 static int vhost_update_avail_event(struct vhost_virtqueue *vq, u16 avail_event)
1176 {
1177         if (__put_user(cpu_to_vhost16(vq, vq->avail_idx), vhost_avail_event(vq)))
1178                 return -EFAULT;
1179         if (unlikely(vq->log_used)) {
1180                 void __user *used;
1181                 /* Make sure the event is seen before log. */
1182                 smp_wmb();
1183                 /* Log avail event write */
1184                 used = vhost_avail_event(vq);
1185                 log_write(vq->log_base, vq->log_addr +
1186                           (used - (void __user *)vq->used),
1187                           sizeof *vhost_avail_event(vq));
1188                 if (vq->log_ctx)
1189                         eventfd_signal(vq->log_ctx, 1);
1190         }
1191         return 0;
1192 }
1193
1194 int vhost_vq_init_access(struct vhost_virtqueue *vq)
1195 {
1196         __virtio16 last_used_idx;
1197         int r;
1198         bool is_le = vq->is_le;
1199
1200         if (!vq->private_data) {
1201                 vhost_reset_is_le(vq);
1202                 return 0;
1203         }
1204
1205         vhost_init_is_le(vq);
1206
1207         r = vhost_update_used_flags(vq);
1208         if (r)
1209                 goto err;
1210         vq->signalled_used_valid = false;
1211         if (!access_ok(VERIFY_READ, &vq->used->idx, sizeof vq->used->idx)) {
1212                 r = -EFAULT;
1213                 goto err;
1214         }
1215         r = __get_user(last_used_idx, &vq->used->idx);
1216         if (r)
1217                 goto err;
1218         vq->last_used_idx = vhost16_to_cpu(vq, last_used_idx);
1219         return 0;
1220 err:
1221         vq->is_le = is_le;
1222         return r;
1223 }
1224 EXPORT_SYMBOL_GPL(vhost_vq_init_access);
1225
1226 static int translate_desc(struct vhost_virtqueue *vq, u64 addr, u32 len,
1227                           struct iovec iov[], int iov_size)
1228 {
1229         const struct vhost_memory_region *reg;
1230         struct vhost_memory *mem;
1231         struct iovec *_iov;
1232         u64 s = 0;
1233         int ret = 0;
1234
1235         mem = vq->memory;
1236         while ((u64)len > s) {
1237                 u64 size;
1238                 if (unlikely(ret >= iov_size)) {
1239                         ret = -ENOBUFS;
1240                         break;
1241                 }
1242                 reg = find_region(mem, addr, len);
1243                 if (unlikely(!reg)) {
1244                         ret = -EFAULT;
1245                         break;
1246                 }
1247                 _iov = iov + ret;
1248                 size = reg->memory_size - addr + reg->guest_phys_addr;
1249                 _iov->iov_len = min((u64)len - s, size);
1250                 _iov->iov_base = (void __user *)(unsigned long)
1251                         (reg->userspace_addr + addr - reg->guest_phys_addr);
1252                 s += size;
1253                 addr += size;
1254                 ++ret;
1255         }
1256
1257         return ret;
1258 }
1259
1260 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
1261  * function returns the next descriptor in the chain,
1262  * or -1U if we're at the end. */
1263 static unsigned next_desc(struct vhost_virtqueue *vq, struct vring_desc *desc)
1264 {
1265         unsigned int next;
1266
1267         /* If this descriptor says it doesn't chain, we're done. */
1268         if (!(desc->flags & cpu_to_vhost16(vq, VRING_DESC_F_NEXT)))
1269                 return -1U;
1270
1271         /* Check they're not leading us off end of descriptors. */
1272         next = vhost16_to_cpu(vq, desc->next);
1273         /* Make sure compiler knows to grab that: we don't want it changing! */
1274         /* We will use the result as an index in an array, so most
1275          * architectures only need a compiler barrier here. */
1276         read_barrier_depends();
1277
1278         return next;
1279 }
1280
1281 static int get_indirect(struct vhost_virtqueue *vq,
1282                         struct iovec iov[], unsigned int iov_size,
1283                         unsigned int *out_num, unsigned int *in_num,
1284                         struct vhost_log *log, unsigned int *log_num,
1285                         struct vring_desc *indirect)
1286 {
1287         struct vring_desc desc;
1288         unsigned int i = 0, count, found = 0;
1289         u32 len = vhost32_to_cpu(vq, indirect->len);
1290         struct iov_iter from;
1291         int ret;
1292
1293         /* Sanity check */
1294         if (unlikely(len % sizeof desc)) {
1295                 vq_err(vq, "Invalid length in indirect descriptor: "
1296                        "len 0x%llx not multiple of 0x%zx\n",
1297                        (unsigned long long)len,
1298                        sizeof desc);
1299                 return -EINVAL;
1300         }
1301
1302         ret = translate_desc(vq, vhost64_to_cpu(vq, indirect->addr), len, vq->indirect,
1303                              UIO_MAXIOV);
1304         if (unlikely(ret < 0)) {
1305                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
1306                 return ret;
1307         }
1308         iov_iter_init(&from, READ, vq->indirect, ret, len);
1309
1310         /* We will use the result as an address to read from, so most
1311          * architectures only need a compiler barrier here. */
1312         read_barrier_depends();
1313
1314         count = len / sizeof desc;
1315         /* Buffers are chained via a 16 bit next field, so
1316          * we can have at most 2^16 of these. */
1317         if (unlikely(count > USHRT_MAX + 1)) {
1318                 vq_err(vq, "Indirect buffer length too big: %d\n",
1319                        indirect->len);
1320                 return -E2BIG;
1321         }
1322
1323         do {
1324                 unsigned iov_count = *in_num + *out_num;
1325                 if (unlikely(++found > count)) {
1326                         vq_err(vq, "Loop detected: last one at %u "
1327                                "indirect size %u\n",
1328                                i, count);
1329                         return -EINVAL;
1330                 }
1331                 if (unlikely(copy_from_iter(&desc, sizeof(desc), &from) !=
1332                              sizeof(desc))) {
1333                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
1334                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1335                         return -EINVAL;
1336                 }
1337                 if (unlikely(desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT))) {
1338                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
1339                                i, (size_t)vhost64_to_cpu(vq, indirect->addr) + i * sizeof desc);
1340                         return -EINVAL;
1341                 }
1342
1343                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1344                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1345                                      iov_size - iov_count);
1346                 if (unlikely(ret < 0)) {
1347                         vq_err(vq, "Translation failure %d indirect idx %d\n",
1348                                ret, i);
1349                         return ret;
1350                 }
1351                 /* If this is an input descriptor, increment that count. */
1352                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1353                         *in_num += ret;
1354                         if (unlikely(log)) {
1355                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1356                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1357                                 ++*log_num;
1358                         }
1359                 } else {
1360                         /* If it's an output descriptor, they're all supposed
1361                          * to come before any input descriptors. */
1362                         if (unlikely(*in_num)) {
1363                                 vq_err(vq, "Indirect descriptor "
1364                                        "has out after in: idx %d\n", i);
1365                                 return -EINVAL;
1366                         }
1367                         *out_num += ret;
1368                 }
1369         } while ((i = next_desc(vq, &desc)) != -1);
1370         return 0;
1371 }
1372
1373 /* This looks in the virtqueue and for the first available buffer, and converts
1374  * it to an iovec for convenient access.  Since descriptors consist of some
1375  * number of output then some number of input descriptors, it's actually two
1376  * iovecs, but we pack them into one and note how many of each there were.
1377  *
1378  * This function returns the descriptor number found, or vq->num (which is
1379  * never a valid descriptor number) if none was found.  A negative code is
1380  * returned on error. */
1381 int vhost_get_vq_desc(struct vhost_virtqueue *vq,
1382                       struct iovec iov[], unsigned int iov_size,
1383                       unsigned int *out_num, unsigned int *in_num,
1384                       struct vhost_log *log, unsigned int *log_num)
1385 {
1386         struct vring_desc desc;
1387         unsigned int i, head, found = 0;
1388         u16 last_avail_idx;
1389         __virtio16 avail_idx;
1390         __virtio16 ring_head;
1391         int ret;
1392
1393         /* Check it isn't doing very strange things with descriptor numbers. */
1394         last_avail_idx = vq->last_avail_idx;
1395         if (unlikely(__get_user(avail_idx, &vq->avail->idx))) {
1396                 vq_err(vq, "Failed to access avail idx at %p\n",
1397                        &vq->avail->idx);
1398                 return -EFAULT;
1399         }
1400         vq->avail_idx = vhost16_to_cpu(vq, avail_idx);
1401
1402         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1403                 vq_err(vq, "Guest moved used index from %u to %u",
1404                        last_avail_idx, vq->avail_idx);
1405                 return -EFAULT;
1406         }
1407
1408         /* If there's nothing new since last we looked, return invalid. */
1409         if (vq->avail_idx == last_avail_idx)
1410                 return vq->num;
1411
1412         /* Only get avail ring entries after they have been exposed by guest. */
1413         smp_rmb();
1414
1415         /* Grab the next descriptor number they're advertising, and increment
1416          * the index we've seen. */
1417         if (unlikely(__get_user(ring_head,
1418                                 &vq->avail->ring[last_avail_idx & (vq->num - 1)]))) {
1419                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1420                        last_avail_idx,
1421                        &vq->avail->ring[last_avail_idx % vq->num]);
1422                 return -EFAULT;
1423         }
1424
1425         head = vhost16_to_cpu(vq, ring_head);
1426
1427         /* If their number is silly, that's an error. */
1428         if (unlikely(head >= vq->num)) {
1429                 vq_err(vq, "Guest says index %u > %u is available",
1430                        head, vq->num);
1431                 return -EINVAL;
1432         }
1433
1434         /* When we start there are none of either input nor output. */
1435         *out_num = *in_num = 0;
1436         if (unlikely(log))
1437                 *log_num = 0;
1438
1439         i = head;
1440         do {
1441                 unsigned iov_count = *in_num + *out_num;
1442                 if (unlikely(i >= vq->num)) {
1443                         vq_err(vq, "Desc index is %u > %u, head = %u",
1444                                i, vq->num, head);
1445                         return -EINVAL;
1446                 }
1447                 if (unlikely(++found > vq->num)) {
1448                         vq_err(vq, "Loop detected: last one at %u "
1449                                "vq size %u head %u\n",
1450                                i, vq->num, head);
1451                         return -EINVAL;
1452                 }
1453                 ret = __copy_from_user(&desc, vq->desc + i, sizeof desc);
1454                 if (unlikely(ret)) {
1455                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1456                                i, vq->desc + i);
1457                         return -EFAULT;
1458                 }
1459                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_INDIRECT)) {
1460                         ret = get_indirect(vq, iov, iov_size,
1461                                            out_num, in_num,
1462                                            log, log_num, &desc);
1463                         if (unlikely(ret < 0)) {
1464                                 vq_err(vq, "Failure detected "
1465                                        "in indirect descriptor at idx %d\n", i);
1466                                 return ret;
1467                         }
1468                         continue;
1469                 }
1470
1471                 ret = translate_desc(vq, vhost64_to_cpu(vq, desc.addr),
1472                                      vhost32_to_cpu(vq, desc.len), iov + iov_count,
1473                                      iov_size - iov_count);
1474                 if (unlikely(ret < 0)) {
1475                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1476                                ret, i);
1477                         return ret;
1478                 }
1479                 if (desc.flags & cpu_to_vhost16(vq, VRING_DESC_F_WRITE)) {
1480                         /* If this is an input descriptor,
1481                          * increment that count. */
1482                         *in_num += ret;
1483                         if (unlikely(log)) {
1484                                 log[*log_num].addr = vhost64_to_cpu(vq, desc.addr);
1485                                 log[*log_num].len = vhost32_to_cpu(vq, desc.len);
1486                                 ++*log_num;
1487                         }
1488                 } else {
1489                         /* If it's an output descriptor, they're all supposed
1490                          * to come before any input descriptors. */
1491                         if (unlikely(*in_num)) {
1492                                 vq_err(vq, "Descriptor has out after in: "
1493                                        "idx %d\n", i);
1494                                 return -EINVAL;
1495                         }
1496                         *out_num += ret;
1497                 }
1498         } while ((i = next_desc(vq, &desc)) != -1);
1499
1500         /* On success, increment avail index. */
1501         vq->last_avail_idx++;
1502
1503         /* Assume notifications from guest are disabled at this point,
1504          * if they aren't we would need to update avail_event index. */
1505         BUG_ON(!(vq->used_flags & VRING_USED_F_NO_NOTIFY));
1506         return head;
1507 }
1508 EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
1509
1510 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1511 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1512 {
1513         vq->last_avail_idx -= n;
1514 }
1515 EXPORT_SYMBOL_GPL(vhost_discard_vq_desc);
1516
1517 /* After we've used one of their buffers, we tell them about it.  We'll then
1518  * want to notify the guest, using eventfd. */
1519 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1520 {
1521         struct vring_used_elem heads = {
1522                 cpu_to_vhost32(vq, head),
1523                 cpu_to_vhost32(vq, len)
1524         };
1525
1526         return vhost_add_used_n(vq, &heads, 1);
1527 }
1528 EXPORT_SYMBOL_GPL(vhost_add_used);
1529
1530 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1531                             struct vring_used_elem *heads,
1532                             unsigned count)
1533 {
1534         struct vring_used_elem __user *used;
1535         u16 old, new;
1536         int start;
1537
1538         start = vq->last_used_idx & (vq->num - 1);
1539         used = vq->used->ring + start;
1540         if (count == 1) {
1541                 if (__put_user(heads[0].id, &used->id)) {
1542                         vq_err(vq, "Failed to write used id");
1543                         return -EFAULT;
1544                 }
1545                 if (__put_user(heads[0].len, &used->len)) {
1546                         vq_err(vq, "Failed to write used len");
1547                         return -EFAULT;
1548                 }
1549         } else if (__copy_to_user(used, heads, count * sizeof *used)) {
1550                 vq_err(vq, "Failed to write used");
1551                 return -EFAULT;
1552         }
1553         if (unlikely(vq->log_used)) {
1554                 /* Make sure data is seen before log. */
1555                 smp_wmb();
1556                 /* Log used ring entry write. */
1557                 log_write(vq->log_base,
1558                           vq->log_addr +
1559                            ((void __user *)used - (void __user *)vq->used),
1560                           count * sizeof *used);
1561         }
1562         old = vq->last_used_idx;
1563         new = (vq->last_used_idx += count);
1564         /* If the driver never bothers to signal in a very long while,
1565          * used index might wrap around. If that happens, invalidate
1566          * signalled_used index we stored. TODO: make sure driver
1567          * signals at least once in 2^16 and remove this. */
1568         if (unlikely((u16)(new - vq->signalled_used) < (u16)(new - old)))
1569                 vq->signalled_used_valid = false;
1570         return 0;
1571 }
1572
1573 /* After we've used one of their buffers, we tell them about it.  We'll then
1574  * want to notify the guest, using eventfd. */
1575 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1576                      unsigned count)
1577 {
1578         int start, n, r;
1579
1580         start = vq->last_used_idx & (vq->num - 1);
1581         n = vq->num - start;
1582         if (n < count) {
1583                 r = __vhost_add_used_n(vq, heads, n);
1584                 if (r < 0)
1585                         return r;
1586                 heads += n;
1587                 count -= n;
1588         }
1589         r = __vhost_add_used_n(vq, heads, count);
1590
1591         /* Make sure buffer is written before we update index. */
1592         smp_wmb();
1593         if (__put_user(cpu_to_vhost16(vq, vq->last_used_idx), &vq->used->idx)) {
1594                 vq_err(vq, "Failed to increment used idx");
1595                 return -EFAULT;
1596         }
1597         if (unlikely(vq->log_used)) {
1598                 /* Log used index update. */
1599                 log_write(vq->log_base,
1600                           vq->log_addr + offsetof(struct vring_used, idx),
1601                           sizeof vq->used->idx);
1602                 if (vq->log_ctx)
1603                         eventfd_signal(vq->log_ctx, 1);
1604         }
1605         return r;
1606 }
1607 EXPORT_SYMBOL_GPL(vhost_add_used_n);
1608
1609 static bool vhost_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1610 {
1611         __u16 old, new;
1612         __virtio16 event;
1613         bool v;
1614         /* Flush out used index updates. This is paired
1615          * with the barrier that the Guest executes when enabling
1616          * interrupts. */
1617         smp_mb();
1618
1619         if (vhost_has_feature(vq, VIRTIO_F_NOTIFY_ON_EMPTY) &&
1620             unlikely(vq->avail_idx == vq->last_avail_idx))
1621                 return true;
1622
1623         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1624                 __virtio16 flags;
1625                 if (__get_user(flags, &vq->avail->flags)) {
1626                         vq_err(vq, "Failed to get flags");
1627                         return true;
1628                 }
1629                 return !(flags & cpu_to_vhost16(vq, VRING_AVAIL_F_NO_INTERRUPT));
1630         }
1631         old = vq->signalled_used;
1632         v = vq->signalled_used_valid;
1633         new = vq->signalled_used = vq->last_used_idx;
1634         vq->signalled_used_valid = true;
1635
1636         if (unlikely(!v))
1637                 return true;
1638
1639         if (__get_user(event, vhost_used_event(vq))) {
1640                 vq_err(vq, "Failed to get used event idx");
1641                 return true;
1642         }
1643         return vring_need_event(vhost16_to_cpu(vq, event), new, old);
1644 }
1645
1646 /* This actually signals the guest, using eventfd. */
1647 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1648 {
1649         /* Signal the Guest tell them we used something up. */
1650         if (vq->call_ctx && vhost_notify(dev, vq))
1651                 eventfd_signal(vq->call_ctx, 1);
1652 }
1653 EXPORT_SYMBOL_GPL(vhost_signal);
1654
1655 /* And here's the combo meal deal.  Supersize me! */
1656 void vhost_add_used_and_signal(struct vhost_dev *dev,
1657                                struct vhost_virtqueue *vq,
1658                                unsigned int head, int len)
1659 {
1660         vhost_add_used(vq, head, len);
1661         vhost_signal(dev, vq);
1662 }
1663 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal);
1664
1665 /* multi-buffer version of vhost_add_used_and_signal */
1666 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1667                                  struct vhost_virtqueue *vq,
1668                                  struct vring_used_elem *heads, unsigned count)
1669 {
1670         vhost_add_used_n(vq, heads, count);
1671         vhost_signal(dev, vq);
1672 }
1673 EXPORT_SYMBOL_GPL(vhost_add_used_and_signal_n);
1674
1675 /* return true if we're sure that avaiable ring is empty */
1676 bool vhost_vq_avail_empty(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1677 {
1678         __virtio16 avail_idx;
1679         int r;
1680
1681         r = __get_user(avail_idx, &vq->avail->idx);
1682         if (r)
1683                 return false;
1684
1685         return vhost16_to_cpu(vq, avail_idx) == vq->avail_idx;
1686 }
1687 EXPORT_SYMBOL_GPL(vhost_vq_avail_empty);
1688
1689 /* OK, now we need to know about added descriptors. */
1690 bool vhost_enable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1691 {
1692         __virtio16 avail_idx;
1693         int r;
1694
1695         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1696                 return false;
1697         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1698         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1699                 r = vhost_update_used_flags(vq);
1700                 if (r) {
1701                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1702                                &vq->used->flags, r);
1703                         return false;
1704                 }
1705         } else {
1706                 r = vhost_update_avail_event(vq, vq->avail_idx);
1707                 if (r) {
1708                         vq_err(vq, "Failed to update avail event index at %p: %d\n",
1709                                vhost_avail_event(vq), r);
1710                         return false;
1711                 }
1712         }
1713         /* They could have slipped one in as we were doing that: make
1714          * sure it's written, then check again. */
1715         smp_mb();
1716         r = __get_user(avail_idx, &vq->avail->idx);
1717         if (r) {
1718                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1719                        &vq->avail->idx, r);
1720                 return false;
1721         }
1722
1723         return vhost16_to_cpu(vq, avail_idx) != vq->avail_idx;
1724 }
1725 EXPORT_SYMBOL_GPL(vhost_enable_notify);
1726
1727 /* We don't need to be notified again. */
1728 void vhost_disable_notify(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1729 {
1730         int r;
1731
1732         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1733                 return;
1734         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1735         if (!vhost_has_feature(vq, VIRTIO_RING_F_EVENT_IDX)) {
1736                 r = vhost_update_used_flags(vq);
1737                 if (r)
1738                         vq_err(vq, "Failed to enable notification at %p: %d\n",
1739                                &vq->used->flags, r);
1740         }
1741 }
1742 EXPORT_SYMBOL_GPL(vhost_disable_notify);
1743
1744 static int __init vhost_init(void)
1745 {
1746         return 0;
1747 }
1748
1749 static void __exit vhost_exit(void)
1750 {
1751 }
1752
1753 module_init(vhost_init);
1754 module_exit(vhost_exit);
1755
1756 MODULE_VERSION("0.0.1");
1757 MODULE_LICENSE("GPL v2");
1758 MODULE_AUTHOR("Michael S. Tsirkin");
1759 MODULE_DESCRIPTION("Host kernel accelerator for virtio");