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/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/virtio_net.h>
  17 #include <linux/mm.h>
  18 #include <linux/miscdevice.h>
  19 #include <linux/mutex.h>
  20 #include <linux/rcupdate.h>
  21 #include <linux/poll.h>
  22 #include <linux/file.h>
  23 #include <linux/highmem.h>
  24 #include <linux/slab.h>
  25 #include <linux/kthread.h>
  26 #include <linux/cgroup.h>
  27
  28 #include <linux/net.h>
  29 #include <linux/if_packet.h>
  30 #include <linux/if_arp.h>
  31
  32 #include <net/sock.h>
  33
  34 #include "vhost.h"
  35
  36 enum {
  37         VHOST_MEMORY_MAX_NREGIONS = 64,
  38         VHOST_MEMORY_F_LOG = 0x1,
  39 };
  40
  41 static void vhost_poll_func(struct file *file, wait_queue_head_t *wqh,
  42                             poll_table *pt)
  43 {
  44         struct vhost_poll *poll;
  45         poll = container_of(pt, struct vhost_poll, table);
  46
  47         poll->wqh = wqh;
  48         add_wait_queue(wqh, &poll->wait);
  49 }
  50
  51 static int vhost_poll_wakeup(wait_queue_t *wait, unsigned mode, int sync,
  52                              void *key)
  53 {
  54         struct vhost_poll *poll = container_of(wait, struct vhost_poll, wait);
  55
  56         if (!((unsigned long)key & poll->mask))
  57                 return 0;
  58
  59         vhost_poll_queue(poll);
  60         return 0;
  61 }
  62
  63 /* Init poll structure */
  64 void vhost_poll_init(struct vhost_poll *poll, vhost_work_fn_t fn,
  65                      unsigned long mask, struct vhost_dev *dev)
  66 {
  67         struct vhost_work *work = &poll->work;
  68
  69         init_waitqueue_func_entry(&poll->wait, vhost_poll_wakeup);
  70         init_poll_funcptr(&poll->table, vhost_poll_func);
  71         poll->mask = mask;
  72         poll->dev = dev;
  73
  74         INIT_LIST_HEAD(&work->node);
  75         work->fn = fn;
  76         init_waitqueue_head(&work->done);
  77         work->flushing = 0;
  78         work->queue_seq = work->done_seq = 0;
  79 }
  80
  81 /* Start polling a file. We add ourselves to file's wait queue. The caller must
  82  * keep a reference to a file until after vhost_poll_stop is called. */
  83 void vhost_poll_start(struct vhost_poll *poll, struct file *file)
  84 {
  85         unsigned long mask;
  86         mask = file->f_op->poll(file, &poll->table);
  87         if (mask)
  88                 vhost_poll_wakeup(&poll->wait, 0, 0, (void *)mask);
  89 }
  90
  91 /* Stop polling a file. After this function returns, it becomes safe to drop the
  92  * file reference. You must also flush afterwards. */
  93 void vhost_poll_stop(struct vhost_poll *poll)
  94 {
  95         remove_wait_queue(poll->wqh, &poll->wait);
  96 }
  97
  98 /* Flush any work that has been scheduled. When calling this, don't hold any
  99  * locks that are also used by the callback. */
 100 void vhost_poll_flush(struct vhost_poll *poll)
 101 {
 102         struct vhost_work *work = &poll->work;
 103         unsigned seq;
 104         int left;
 105         int flushing;
 106
 107         spin_lock_irq(&poll->dev->work_lock);
 108         seq = work->queue_seq;
 109         work->flushing++;
 110         spin_unlock_irq(&poll->dev->work_lock);
 111         wait_event(work->done, ({
 112                    spin_lock_irq(&poll->dev->work_lock);
 113                    left = seq - work->done_seq <= 0;
 114                    spin_unlock_irq(&poll->dev->work_lock);
 115                    left;
 116         }));
 117         spin_lock_irq(&poll->dev->work_lock);
 118         flushing = --work->flushing;
 119         spin_unlock_irq(&poll->dev->work_lock);
 120         BUG_ON(flushing < 0);
 121 }
 122
 123 void vhost_poll_queue(struct vhost_poll *poll)
 124 {
 125         struct vhost_dev *dev = poll->dev;
 126         struct vhost_work *work = &poll->work;
 127         unsigned long flags;
 128
 129         spin_lock_irqsave(&dev->work_lock, flags);
 130         if (list_empty(&work->node)) {
 131                 list_add_tail(&work->node, &dev->work_list);
 132                 work->queue_seq++;
 133                 wake_up_process(dev->worker);
 134         }
 135         spin_unlock_irqrestore(&dev->work_lock, flags);
 136 }
 137
 138 static void vhost_vq_reset(struct vhost_dev *dev,
 139                            struct vhost_virtqueue *vq)
 140 {
 141         vq->num = 1;
 142         vq->desc = NULL;
 143         vq->avail = NULL;
 144         vq->used = NULL;
 145         vq->last_avail_idx = 0;
 146         vq->avail_idx = 0;
 147         vq->last_used_idx = 0;
 148         vq->used_flags = 0;
 149         vq->used_flags = 0;
 150         vq->log_used = false;
 151         vq->log_addr = -1ull;
 152         vq->vhost_hlen = 0;
 153         vq->sock_hlen = 0;
 154         vq->private_data = NULL;
 155         vq->log_base = NULL;
 156         vq->error_ctx = NULL;
 157         vq->error = NULL;
 158         vq->kick = NULL;
 159         vq->call_ctx = NULL;
 160         vq->call = NULL;
 161         vq->log_ctx = NULL;
 162 }
 163
 164 static int vhost_worker(void *data)
 165 {
 166         struct vhost_dev *dev = data;
 167         struct vhost_work *work = NULL;
 168         unsigned uninitialized_var(seq);
 169
 170         for (;;) {
 171                 /* mb paired w/ kthread_stop */
 172                 set_current_state(TASK_INTERRUPTIBLE);
 173
 174                 spin_lock_irq(&dev->work_lock);
 175                 if (work) {
 176                         work->done_seq = seq;
 177                         if (work->flushing)
 178                                 wake_up_all(&work->done);
 179                 }
 180
 181                 if (kthread_should_stop()) {
 182                         spin_unlock_irq(&dev->work_lock);
 183                         __set_current_state(TASK_RUNNING);
 184                         return 0;
 185                 }
 186                 if (!list_empty(&dev->work_list)) {
 187                         work = list_first_entry(&dev->work_list,
 188                                                 struct vhost_work, node);
 189                         list_del_init(&work->node);
 190                         seq = work->queue_seq;
 191                 } else
 192                         work = NULL;
 193                 spin_unlock_irq(&dev->work_lock);
 194
 195                 if (work) {
 196                         __set_current_state(TASK_RUNNING);
 197                         work->fn(work);
 198                 } else
 199                         schedule();
 200
 201         }
 202 }
 203
 204 long vhost_dev_init(struct vhost_dev *dev,
 205                     struct vhost_virtqueue *vqs, int nvqs)
 206 {
 207         int i;
 208
 209         dev->vqs = vqs;
 210         dev->nvqs = nvqs;
 211         mutex_init(&dev->mutex);
 212         dev->log_ctx = NULL;
 213         dev->log_file = NULL;
 214         dev->memory = NULL;
 215         dev->mm = NULL;
 216         spin_lock_init(&dev->work_lock);
 217         INIT_LIST_HEAD(&dev->work_list);
 218         dev->worker = NULL;
 219
 220         for (i = 0; i < dev->nvqs; ++i) {
 221                 dev->vqs[i].dev = dev;
 222                 mutex_init(&dev->vqs[i].mutex);
 223                 vhost_vq_reset(dev, dev->vqs + i);
 224                 if (dev->vqs[i].handle_kick)
 225                         vhost_poll_init(&dev->vqs[i].poll,
 226                                         dev->vqs[i].handle_kick, POLLIN, dev);
 227         }
 228
 229         return 0;
 230 }
 231
 232 /* Caller should have device mutex */
 233 long vhost_dev_check_owner(struct vhost_dev *dev)
 234 {
 235         /* Are you the owner? If not, I don't think you mean to do that */
 236         return dev->mm == current->mm ? 0 : -EPERM;
 237 }
 238
 239 /* Caller should have device mutex */
 240 static long vhost_dev_set_owner(struct vhost_dev *dev)
 241 {
 242         struct task_struct *worker;
 243         int err;
 244         /* Is there an owner already? */
 245         if (dev->mm) {
 246                 err = -EBUSY;
 247                 goto err_mm;
 248         }
 249         /* No owner, become one */
 250         dev->mm = get_task_mm(current);
 251         worker = kthread_create(vhost_worker, dev, "vhost-%d", current->pid);
 252         if (IS_ERR(worker)) {
 253                 err = PTR_ERR(worker);
 254                 goto err_worker;
 255         }
 256
 257         dev->worker = worker;
 258         err = cgroup_attach_task_current_cg(worker);
 259         if (err)
 260                 goto err_cgroup;
 261         wake_up_process(worker);        /* avoid contributing to loadavg */
 262
 263         return 0;
 264 err_cgroup:
 265         kthread_stop(worker);
 266 err_worker:
 267         if (dev->mm)
 268                 mmput(dev->mm);
 269         dev->mm = NULL;
 270 err_mm:
 271         return err;
 272 }
 273
 274 /* Caller should have device mutex */
 275 long vhost_dev_reset_owner(struct vhost_dev *dev)
 276 {
 277         struct vhost_memory *memory;
 278
 279         /* Restore memory to default empty mapping. */
 280         memory = kmalloc(offsetof(struct vhost_memory, regions), GFP_KERNEL);
 281         if (!memory)
 282                 return -ENOMEM;
 283
 284         vhost_dev_cleanup(dev);
 285
 286         memory->nregions = 0;
 287         dev->memory = memory;
 288         return 0;
 289 }
 290
 291 /* Caller should have device mutex */
 292 void vhost_dev_cleanup(struct vhost_dev *dev)
 293 {
 294         int i;
 295         for (i = 0; i < dev->nvqs; ++i) {
 296                 if (dev->vqs[i].kick && dev->vqs[i].handle_kick) {
 297                         vhost_poll_stop(&dev->vqs[i].poll);
 298                         vhost_poll_flush(&dev->vqs[i].poll);
 299                 }
 300                 if (dev->vqs[i].error_ctx)
 301                         eventfd_ctx_put(dev->vqs[i].error_ctx);
 302                 if (dev->vqs[i].error)
 303                         fput(dev->vqs[i].error);
 304                 if (dev->vqs[i].kick)
 305                         fput(dev->vqs[i].kick);
 306                 if (dev->vqs[i].call_ctx)
 307                         eventfd_ctx_put(dev->vqs[i].call_ctx);
 308                 if (dev->vqs[i].call)
 309                         fput(dev->vqs[i].call);
 310                 vhost_vq_reset(dev, dev->vqs + i);
 311         }
 312         if (dev->log_ctx)
 313                 eventfd_ctx_put(dev->log_ctx);
 314         dev->log_ctx = NULL;
 315         if (dev->log_file)
 316                 fput(dev->log_file);
 317         dev->log_file = NULL;
 318         /* No one will access memory at this point */
 319         kfree(dev->memory);
 320         dev->memory = NULL;
 321         if (dev->mm)
 322                 mmput(dev->mm);
 323         dev->mm = NULL;
 324
 325         WARN_ON(!list_empty(&dev->work_list));
 326         if (dev->worker) {
 327                 kthread_stop(dev->worker);
 328                 dev->worker = NULL;
 329         }
 330 }
 331
 332 static int log_access_ok(void __user *log_base, u64 addr, unsigned long sz)
 333 {
 334         u64 a = addr / VHOST_PAGE_SIZE / 8;
 335         /* Make sure 64 bit math will not overflow. */
 336         if (a > ULONG_MAX - (unsigned long)log_base ||
 337             a + (unsigned long)log_base > ULONG_MAX)
 338                 return -EFAULT;
 339
 340         return access_ok(VERIFY_WRITE, log_base + a,
 341                          (sz + VHOST_PAGE_SIZE * 8 - 1) / VHOST_PAGE_SIZE / 8);
 342 }
 343
 344 /* Caller should have vq mutex and device mutex. */
 345 static int vq_memory_access_ok(void __user *log_base, struct vhost_memory *mem,
 346                                int log_all)
 347 {
 348         int i;
 349
 350         if (!mem)
 351                 return 0;
 352
 353         for (i = 0; i < mem->nregions; ++i) {
 354                 struct vhost_memory_region *m = mem->regions + i;
 355                 unsigned long a = m->userspace_addr;
 356                 if (m->memory_size > ULONG_MAX)
 357                         return 0;
 358                 else if (!access_ok(VERIFY_WRITE, (void __user *)a,
 359                                     m->memory_size))
 360                         return 0;
 361                 else if (log_all && !log_access_ok(log_base,
 362                                                    m->guest_phys_addr,
 363                                                    m->memory_size))
 364                         return 0;
 365         }
 366         return 1;
 367 }
 368
 369 /* Can we switch to this memory table? */
 370 /* Caller should have device mutex but not vq mutex */
 371 static int memory_access_ok(struct vhost_dev *d, struct vhost_memory *mem,
 372                             int log_all)
 373 {
 374         int i;
 375         for (i = 0; i < d->nvqs; ++i) {
 376                 int ok;
 377                 mutex_lock(&d->vqs[i].mutex);
 378                 /* If ring is inactive, will check when it's enabled. */
 379                 if (d->vqs[i].private_data)
 380                         ok = vq_memory_access_ok(d->vqs[i].log_base, mem,
 381                                                  log_all);
 382                 else
 383                         ok = 1;
 384                 mutex_unlock(&d->vqs[i].mutex);
 385                 if (!ok)
 386                         return 0;
 387         }
 388         return 1;
 389 }
 390
 391 static int vq_access_ok(unsigned int num,
 392                         struct vring_desc __user *desc,
 393                         struct vring_avail __user *avail,
 394                         struct vring_used __user *used)
 395 {
 396         return access_ok(VERIFY_READ, desc, num * sizeof *desc) &&
 397                access_ok(VERIFY_READ, avail,
 398                          sizeof *avail + num * sizeof *avail->ring) &&
 399                access_ok(VERIFY_WRITE, used,
 400                         sizeof *used + num * sizeof *used->ring);
 401 }
 402
 403 /* Can we log writes? */
 404 /* Caller should have device mutex but not vq mutex */
 405 int vhost_log_access_ok(struct vhost_dev *dev)
 406 {
 407         return memory_access_ok(dev, dev->memory, 1);
 408 }
 409
 410 /* Verify access for write logging. */
 411 /* Caller should have vq mutex and device mutex */
 412 static int vq_log_access_ok(struct vhost_virtqueue *vq, void __user *log_base)
 413 {
 414         return vq_memory_access_ok(log_base, vq->dev->memory,
 415                             vhost_has_feature(vq->dev, VHOST_F_LOG_ALL)) &&
 416                 (!vq->log_used || log_access_ok(log_base, vq->log_addr,
 417                                         sizeof *vq->used +
 418                                         vq->num * sizeof *vq->used->ring));
 419 }
 420
 421 /* Can we start vq? */
 422 /* Caller should have vq mutex and device mutex */
 423 int vhost_vq_access_ok(struct vhost_virtqueue *vq)
 424 {
 425         return vq_access_ok(vq->num, vq->desc, vq->avail, vq->used) &&
 426                 vq_log_access_ok(vq, vq->log_base);
 427 }
 428
 429 static long vhost_set_memory(struct vhost_dev *d, struct vhost_memory __user *m)
 430 {
 431         struct vhost_memory mem, *newmem, *oldmem;
 432         unsigned long size = offsetof(struct vhost_memory, regions);
 433         if (copy_from_user(&mem, m, size))
 434                 return -EFAULT;
 435         if (mem.padding)
 436                 return -EOPNOTSUPP;
 437         if (mem.nregions > VHOST_MEMORY_MAX_NREGIONS)
 438                 return -E2BIG;
 439         newmem = kmalloc(size + mem.nregions * sizeof *m->regions, GFP_KERNEL);
 440         if (!newmem)
 441                 return -ENOMEM;
 442
 443         memcpy(newmem, &mem, size);
 444         if (copy_from_user(newmem->regions, m->regions,
 445                            mem.nregions * sizeof *m->regions)) {
 446                 kfree(newmem);
 447                 return -EFAULT;
 448         }
 449
 450         if (!memory_access_ok(d, newmem, vhost_has_feature(d, VHOST_F_LOG_ALL))) {
 451                 kfree(newmem);
 452                 return -EFAULT;
 453         }
 454         oldmem = d->memory;
 455         rcu_assign_pointer(d->memory, newmem);
 456         synchronize_rcu();
 457         kfree(oldmem);
 458         return 0;
 459 }
 460
 461 static int init_used(struct vhost_virtqueue *vq,
 462                      struct vring_used __user *used)
 463 {
 464         int r = put_user(vq->used_flags, &used->flags);
 465         if (r)
 466                 return r;
 467         return get_user(vq->last_used_idx, &used->idx);
 468 }
 469
 470 static long vhost_set_vring(struct vhost_dev *d, int ioctl, void __user *argp)
 471 {
 472         struct file *eventfp, *filep = NULL,
 473                     *pollstart = NULL, *pollstop = NULL;
 474         struct eventfd_ctx *ctx = NULL;
 475         u32 __user *idxp = argp;
 476         struct vhost_virtqueue *vq;
 477         struct vhost_vring_state s;
 478         struct vhost_vring_file f;
 479         struct vhost_vring_addr a;
 480         u32 idx;
 481         long r;
 482
 483         r = get_user(idx, idxp);
 484         if (r < 0)
 485                 return r;
 486         if (idx >= d->nvqs)
 487                 return -ENOBUFS;
 488
 489         vq = d->vqs + idx;
 490
 491         mutex_lock(&vq->mutex);
 492
 493         switch (ioctl) {
 494         case VHOST_SET_VRING_NUM:
 495                 /* Resizing ring with an active backend?
 496                  * You don't want to do that. */
 497                 if (vq->private_data) {
 498                         r = -EBUSY;
 499                         break;
 500                 }
 501                 if (copy_from_user(&s, argp, sizeof s)) {
 502                         r = -EFAULT;
 503                         break;
 504                 }
 505                 if (!s.num || s.num > 0xffff || (s.num & (s.num - 1))) {
 506                         r = -EINVAL;
 507                         break;
 508                 }
 509                 vq->num = s.num;
 510                 break;
 511         case VHOST_SET_VRING_BASE:
 512                 /* Moving base with an active backend?
 513                  * You don't want to do that. */
 514                 if (vq->private_data) {
 515                         r = -EBUSY;
 516                         break;
 517                 }
 518                 if (copy_from_user(&s, argp, sizeof s)) {
 519                         r = -EFAULT;
 520                         break;
 521                 }
 522                 if (s.num > 0xffff) {
 523                         r = -EINVAL;
 524                         break;
 525                 }
 526                 vq->last_avail_idx = s.num;
 527                 /* Forget the cached index value. */
 528                 vq->avail_idx = vq->last_avail_idx;
 529                 break;
 530         case VHOST_GET_VRING_BASE:
 531                 s.index = idx;
 532                 s.num = vq->last_avail_idx;
 533                 if (copy_to_user(argp, &s, sizeof s))
 534                         r = -EFAULT;
 535                 break;
 536         case VHOST_SET_VRING_ADDR:
 537                 if (copy_from_user(&a, argp, sizeof a)) {
 538                         r = -EFAULT;
 539                         break;
 540                 }
 541                 if (a.flags & ~(0x1 << VHOST_VRING_F_LOG)) {
 542                         r = -EOPNOTSUPP;
 543                         break;
 544                 }
 545                 /* For 32bit, verify that the top 32bits of the user
 546                    data are set to zero. */
 547                 if ((u64)(unsigned long)a.desc_user_addr != a.desc_user_addr ||
 548                     (u64)(unsigned long)a.used_user_addr != a.used_user_addr ||
 549                     (u64)(unsigned long)a.avail_user_addr != a.avail_user_addr) {
 550                         r = -EFAULT;
 551                         break;
 552                 }
 553                 if ((a.avail_user_addr & (sizeof *vq->avail->ring - 1)) ||
 554                     (a.used_user_addr & (sizeof *vq->used->ring - 1)) ||
 555                     (a.log_guest_addr & (sizeof *vq->used->ring - 1))) {
 556                         r = -EINVAL;
 557                         break;
 558                 }
 559
 560                 /* We only verify access here if backend is configured.
 561                  * If it is not, we don't as size might not have been setup.
 562                  * We will verify when backend is configured. */
 563                 if (vq->private_data) {
 564                         if (!vq_access_ok(vq->num,
 565                                 (void __user *)(unsigned long)a.desc_user_addr,
 566                                 (void __user *)(unsigned long)a.avail_user_addr,
 567                                 (void __user *)(unsigned long)a.used_user_addr)) {
 568                                 r = -EINVAL;
 569                                 break;
 570                         }
 571
 572                         /* Also validate log access for used ring if enabled. */
 573                         if ((a.flags & (0x1 << VHOST_VRING_F_LOG)) &&
 574                             !log_access_ok(vq->log_base, a.log_guest_addr,
 575                                            sizeof *vq->used +
 576                                            vq->num * sizeof *vq->used->ring)) {
 577                                 r = -EINVAL;
 578                                 break;
 579                         }
 580                 }
 581
 582                 r = init_used(vq, (struct vring_used __user *)(unsigned long)
 583                               a.used_user_addr);
 584                 if (r)
 585                         break;
 586                 vq->log_used = !!(a.flags & (0x1 << VHOST_VRING_F_LOG));
 587                 vq->desc = (void __user *)(unsigned long)a.desc_user_addr;
 588                 vq->avail = (void __user *)(unsigned long)a.avail_user_addr;
 589                 vq->log_addr = a.log_guest_addr;
 590                 vq->used = (void __user *)(unsigned long)a.used_user_addr;
 591                 break;
 592         case VHOST_SET_VRING_KICK:
 593                 if (copy_from_user(&f, argp, sizeof f)) {
 594                         r = -EFAULT;
 595                         break;
 596                 }
 597                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 598                 if (IS_ERR(eventfp)) {
 599                         r = PTR_ERR(eventfp);
 600                         break;
 601                 }
 602                 if (eventfp != vq->kick) {
 603                         pollstop = filep = vq->kick;
 604                         pollstart = vq->kick = eventfp;
 605                 } else
 606                         filep = eventfp;
 607                 break;
 608         case VHOST_SET_VRING_CALL:
 609                 if (copy_from_user(&f, argp, sizeof f)) {
 610                         r = -EFAULT;
 611                         break;
 612                 }
 613                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 614                 if (IS_ERR(eventfp)) {
 615                         r = PTR_ERR(eventfp);
 616                         break;
 617                 }
 618                 if (eventfp != vq->call) {
 619                         filep = vq->call;
 620                         ctx = vq->call_ctx;
 621                         vq->call = eventfp;
 622                         vq->call_ctx = eventfp ?
 623                                 eventfd_ctx_fileget(eventfp) : NULL;
 624                 } else
 625                         filep = eventfp;
 626                 break;
 627         case VHOST_SET_VRING_ERR:
 628                 if (copy_from_user(&f, argp, sizeof f)) {
 629                         r = -EFAULT;
 630                         break;
 631                 }
 632                 eventfp = f.fd == -1 ? NULL : eventfd_fget(f.fd);
 633                 if (IS_ERR(eventfp)) {
 634                         r = PTR_ERR(eventfp);
 635                         break;
 636                 }
 637                 if (eventfp != vq->error) {
 638                         filep = vq->error;
 639                         vq->error = eventfp;
 640                         ctx = vq->error_ctx;
 641                         vq->error_ctx = eventfp ?
 642                                 eventfd_ctx_fileget(eventfp) : NULL;
 643                 } else
 644                         filep = eventfp;
 645                 break;
 646         default:
 647                 r = -ENOIOCTLCMD;
 648         }
 649
 650         if (pollstop && vq->handle_kick)
 651                 vhost_poll_stop(&vq->poll);
 652
 653         if (ctx)
 654                 eventfd_ctx_put(ctx);
 655         if (filep)
 656                 fput(filep);
 657
 658         if (pollstart && vq->handle_kick)
 659                 vhost_poll_start(&vq->poll, vq->kick);
 660
 661         mutex_unlock(&vq->mutex);
 662
 663         if (pollstop && vq->handle_kick)
 664                 vhost_poll_flush(&vq->poll);
 665         return r;
 666 }
 667
 668 /* Caller must have device mutex */
 669 long vhost_dev_ioctl(struct vhost_dev *d, unsigned int ioctl, unsigned long arg)
 670 {
 671         void __user *argp = (void __user *)arg;
 672         struct file *eventfp, *filep = NULL;
 673         struct eventfd_ctx *ctx = NULL;
 674         u64 p;
 675         long r;
 676         int i, fd;
 677
 678         /* If you are not the owner, you can become one */
 679         if (ioctl == VHOST_SET_OWNER) {
 680                 r = vhost_dev_set_owner(d);
 681                 goto done;
 682         }
 683
 684         /* You must be the owner to do anything else */
 685         r = vhost_dev_check_owner(d);
 686         if (r)
 687                 goto done;
 688
 689         switch (ioctl) {
 690         case VHOST_SET_MEM_TABLE:
 691                 r = vhost_set_memory(d, argp);
 692                 break;
 693         case VHOST_SET_LOG_BASE:
 694                 if (copy_from_user(&p, argp, sizeof p)) {
 695                         r = -EFAULT;
 696                         break;
 697                 }
 698                 if ((u64)(unsigned long)p != p) {
 699                         r = -EFAULT;
 700                         break;
 701                 }
 702                 for (i = 0; i < d->nvqs; ++i) {
 703                         struct vhost_virtqueue *vq;
 704                         void __user *base = (void __user *)(unsigned long)p;
 705                         vq = d->vqs + i;
 706                         mutex_lock(&vq->mutex);
 707                         /* If ring is inactive, will check when it's enabled. */
 708                         if (vq->private_data && !vq_log_access_ok(vq, base))
 709                                 r = -EFAULT;
 710                         else
 711                                 vq->log_base = base;
 712                         mutex_unlock(&vq->mutex);
 713                 }
 714                 break;
 715         case VHOST_SET_LOG_FD:
 716                 r = get_user(fd, (int __user *)argp);
 717                 if (r < 0)
 718                         break;
 719                 eventfp = fd == -1 ? NULL : eventfd_fget(fd);
 720                 if (IS_ERR(eventfp)) {
 721                         r = PTR_ERR(eventfp);
 722                         break;
 723                 }
 724                 if (eventfp != d->log_file) {
 725                         filep = d->log_file;
 726                         ctx = d->log_ctx;
 727                         d->log_ctx = eventfp ?
 728                                 eventfd_ctx_fileget(eventfp) : NULL;
 729                 } else
 730                         filep = eventfp;
 731                 for (i = 0; i < d->nvqs; ++i) {
 732                         mutex_lock(&d->vqs[i].mutex);
 733                         d->vqs[i].log_ctx = d->log_ctx;
 734                         mutex_unlock(&d->vqs[i].mutex);
 735                 }
 736                 if (ctx)
 737                         eventfd_ctx_put(ctx);
 738                 if (filep)
 739                         fput(filep);
 740                 break;
 741         default:
 742                 r = vhost_set_vring(d, ioctl, argp);
 743                 break;
 744         }
 745 done:
 746         return r;
 747 }
 748
 749 static const struct vhost_memory_region *find_region(struct vhost_memory *mem,
 750                                                      __u64 addr, __u32 len)
 751 {
 752         struct vhost_memory_region *reg;
 753         int i;
 754         /* linear search is not brilliant, but we really have on the order of 6
 755          * regions in practice */
 756         for (i = 0; i < mem->nregions; ++i) {
 757                 reg = mem->regions + i;
 758                 if (reg->guest_phys_addr <= addr &&
 759                     reg->guest_phys_addr + reg->memory_size - 1 >= addr)
 760                         return reg;
 761         }
 762         return NULL;
 763 }
 764
 765 /* TODO: This is really inefficient.  We need something like get_user()
 766  * (instruction directly accesses the data, with an exception table entry
 767  * returning -EFAULT). See Documentation/x86/exception-tables.txt.
 768  */
 769 static int set_bit_to_user(int nr, void __user *addr)
 770 {
 771         unsigned long log = (unsigned long)addr;
 772         struct page *page;
 773         void *base;
 774         int bit = nr + (log % PAGE_SIZE) * 8;
 775         int r;
 776         r = get_user_pages_fast(log, 1, 1, &page);
 777         if (r < 0)
 778                 return r;
 779         BUG_ON(r != 1);
 780         base = kmap_atomic(page, KM_USER0);
 781         set_bit(bit, base);
 782         kunmap_atomic(base, KM_USER0);
 783         set_page_dirty_lock(page);
 784         put_page(page);
 785         return 0;
 786 }
 787
 788 static int log_write(void __user *log_base,
 789                      u64 write_address, u64 write_length)
 790 {
 791         int r;
 792         if (!write_length)
 793                 return 0;
 794         write_address /= VHOST_PAGE_SIZE;
 795         for (;;) {
 796                 u64 base = (u64)(unsigned long)log_base;
 797                 u64 log = base + write_address / 8;
 798                 int bit = write_address % 8;
 799                 if ((u64)(unsigned long)log != log)
 800                         return -EFAULT;
 801                 r = set_bit_to_user(bit, (void __user *)(unsigned long)log);
 802                 if (r < 0)
 803                         return r;
 804                 if (write_length <= VHOST_PAGE_SIZE)
 805                         break;
 806                 write_length -= VHOST_PAGE_SIZE;
 807                 write_address += VHOST_PAGE_SIZE;
 808         }
 809         return r;
 810 }
 811
 812 int vhost_log_write(struct vhost_virtqueue *vq, struct vhost_log *log,
 813                     unsigned int log_num, u64 len)
 814 {
 815         int i, r;
 816
 817         /* Make sure data written is seen before log. */
 818         smp_wmb();
 819         for (i = 0; i < log_num; ++i) {
 820                 u64 l = min(log[i].len, len);
 821                 r = log_write(vq->log_base, log[i].addr, l);
 822                 if (r < 0)
 823                         return r;
 824                 len -= l;
 825                 if (!len)
 826                         return 0;
 827         }
 828         if (vq->log_ctx)
 829                 eventfd_signal(vq->log_ctx, 1);
 830         /* Length written exceeds what we have stored. This is a bug. */
 831         BUG();
 832         return 0;
 833 }
 834
 835 static int translate_desc(struct vhost_dev *dev, u64 addr, u32 len,
 836                           struct iovec iov[], int iov_size)
 837 {
 838         const struct vhost_memory_region *reg;
 839         struct vhost_memory *mem;
 840         struct iovec *_iov;
 841         u64 s = 0;
 842         int ret = 0;
 843
 844         rcu_read_lock();
 845
 846         mem = rcu_dereference(dev->memory);
 847         while ((u64)len > s) {
 848                 u64 size;
 849                 if (unlikely(ret >= iov_size)) {
 850                         ret = -ENOBUFS;
 851                         break;
 852                 }
 853                 reg = find_region(mem, addr, len);
 854                 if (unlikely(!reg)) {
 855                         ret = -EFAULT;
 856                         break;
 857                 }
 858                 _iov = iov + ret;
 859                 size = reg->memory_size - addr + reg->guest_phys_addr;
 860                 _iov->iov_len = min((u64)len, size);
 861                 _iov->iov_base = (void __user *)(unsigned long)
 862                         (reg->userspace_addr + addr - reg->guest_phys_addr);
 863                 s += size;
 864                 addr += size;
 865                 ++ret;
 866         }
 867
 868         rcu_read_unlock();
 869         return ret;
 870 }
 871
 872 /* Each buffer in the virtqueues is actually a chain of descriptors.  This
 873  * function returns the next descriptor in the chain,
 874  * or -1U if we're at the end. */
 875 static unsigned next_desc(struct vring_desc *desc)
 876 {
 877         unsigned int next;
 878
 879         /* If this descriptor says it doesn't chain, we're done. */
 880         if (!(desc->flags & VRING_DESC_F_NEXT))
 881                 return -1U;
 882
 883         /* Check they're not leading us off end of descriptors. */
 884         next = desc->next;
 885         /* Make sure compiler knows to grab that: we don't want it changing! */
 886         /* We will use the result as an index in an array, so most
 887          * architectures only need a compiler barrier here. */
 888         read_barrier_depends();
 889
 890         return next;
 891 }
 892
 893 static int get_indirect(struct vhost_dev *dev, struct vhost_virtqueue *vq,
 894                         struct iovec iov[], unsigned int iov_size,
 895                         unsigned int *out_num, unsigned int *in_num,
 896                         struct vhost_log *log, unsigned int *log_num,
 897                         struct vring_desc *indirect)
 898 {
 899         struct vring_desc desc;
 900         unsigned int i = 0, count, found = 0;
 901         int ret;
 902
 903         /* Sanity check */
 904         if (unlikely(indirect->len % sizeof desc)) {
 905                 vq_err(vq, "Invalid length in indirect descriptor: "
 906                        "len 0x%llx not multiple of 0x%zx\n",
 907                        (unsigned long long)indirect->len,
 908                        sizeof desc);
 909                 return -EINVAL;
 910         }
 911
 912         ret = translate_desc(dev, indirect->addr, indirect->len, vq->indirect,
 913                              ARRAY_SIZE(vq->indirect));
 914         if (unlikely(ret < 0)) {
 915                 vq_err(vq, "Translation failure %d in indirect.\n", ret);
 916                 return ret;
 917         }
 918
 919         /* We will use the result as an address to read from, so most
 920          * architectures only need a compiler barrier here. */
 921         read_barrier_depends();
 922
 923         count = indirect->len / sizeof desc;
 924         /* Buffers are chained via a 16 bit next field, so
 925          * we can have at most 2^16 of these. */
 926         if (unlikely(count > USHRT_MAX + 1)) {
 927                 vq_err(vq, "Indirect buffer length too big: %d\n",
 928                        indirect->len);
 929                 return -E2BIG;
 930         }
 931
 932         do {
 933                 unsigned iov_count = *in_num + *out_num;
 934                 if (unlikely(++found > count)) {
 935                         vq_err(vq, "Loop detected: last one at %u "
 936                                "indirect size %u\n",
 937                                i, count);
 938                         return -EINVAL;
 939                 }
 940                 if (unlikely(memcpy_fromiovec((unsigned char *)&desc, vq->indirect,
 941                                               sizeof desc))) {
 942                         vq_err(vq, "Failed indirect descriptor: idx %d, %zx\n",
 943                                i, (size_t)indirect->addr + i * sizeof desc);
 944                         return -EINVAL;
 945                 }
 946                 if (unlikely(desc.flags & VRING_DESC_F_INDIRECT)) {
 947                         vq_err(vq, "Nested indirect descriptor: idx %d, %zx\n",
 948                                i, (size_t)indirect->addr + i * sizeof desc);
 949                         return -EINVAL;
 950                 }
 951
 952                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
 953                                      iov_size - iov_count);
 954                 if (unlikely(ret < 0)) {
 955                         vq_err(vq, "Translation failure %d indirect idx %d\n",
 956                                ret, i);
 957                         return ret;
 958                 }
 959                 /* If this is an input descriptor, increment that count. */
 960                 if (desc.flags & VRING_DESC_F_WRITE) {
 961                         *in_num += ret;
 962                         if (unlikely(log)) {
 963                                 log[*log_num].addr = desc.addr;
 964                                 log[*log_num].len = desc.len;
 965                                 ++*log_num;
 966                         }
 967                 } else {
 968                         /* If it's an output descriptor, they're all supposed
 969                          * to come before any input descriptors. */
 970                         if (unlikely(*in_num)) {
 971                                 vq_err(vq, "Indirect descriptor "
 972                                        "has out after in: idx %d\n", i);
 973                                 return -EINVAL;
 974                         }
 975                         *out_num += ret;
 976                 }
 977         } while ((i = next_desc(&desc)) != -1);
 978         return 0;
 979 }
 980
 981 /* This looks in the virtqueue and for the first available buffer, and converts
 982  * it to an iovec for convenient access.  Since descriptors consist of some
 983  * number of output then some number of input descriptors, it's actually two
 984  * iovecs, but we pack them into one and note how many of each there were.
 985  *
 986  * This function returns the descriptor number found, or vq->num (which is
 987  * never a valid descriptor number) if none was found.  A negative code is
 988  * returned on error. */
 989 int vhost_get_vq_desc(struct vhost_dev *dev, struct vhost_virtqueue *vq,
 990                       struct iovec iov[], unsigned int iov_size,
 991                       unsigned int *out_num, unsigned int *in_num,
 992                       struct vhost_log *log, unsigned int *log_num)
 993 {
 994         struct vring_desc desc;
 995         unsigned int i, head, found = 0;
 996         u16 last_avail_idx;
 997         int ret;
 998
 999         /* Check it isn't doing very strange things with descriptor numbers. */
1000         last_avail_idx = vq->last_avail_idx;
1001         if (unlikely(get_user(vq->avail_idx, &vq->avail->idx))) {
1002                 vq_err(vq, "Failed to access avail idx at %p\n",
1003                        &vq->avail->idx);
1004                 return -EFAULT;
1005         }
1006
1007         if (unlikely((u16)(vq->avail_idx - last_avail_idx) > vq->num)) {
1008                 vq_err(vq, "Guest moved used index from %u to %u",
1009                        last_avail_idx, vq->avail_idx);
1010                 return -EFAULT;
1011         }
1012
1013         /* If there's nothing new since last we looked, return invalid. */
1014         if (vq->avail_idx == last_avail_idx)
1015                 return vq->num;
1016
1017         /* Only get avail ring entries after they have been exposed by guest. */
1018         smp_rmb();
1019
1020         /* Grab the next descriptor number they're advertising, and increment
1021          * the index we've seen. */
1022         if (unlikely(get_user(head,
1023                               &vq->avail->ring[last_avail_idx % vq->num]))) {
1024                 vq_err(vq, "Failed to read head: idx %d address %p\n",
1025                        last_avail_idx,
1026                        &vq->avail->ring[last_avail_idx % vq->num]);
1027                 return -EFAULT;
1028         }
1029
1030         /* If their number is silly, that's an error. */
1031         if (unlikely(head >= vq->num)) {
1032                 vq_err(vq, "Guest says index %u > %u is available",
1033                        head, vq->num);
1034                 return -EINVAL;
1035         }
1036
1037         /* When we start there are none of either input nor output. */
1038         *out_num = *in_num = 0;
1039         if (unlikely(log))
1040                 *log_num = 0;
1041
1042         i = head;
1043         do {
1044                 unsigned iov_count = *in_num + *out_num;
1045                 if (unlikely(i >= vq->num)) {
1046                         vq_err(vq, "Desc index is %u > %u, head = %u",
1047                                i, vq->num, head);
1048                         return -EINVAL;
1049                 }
1050                 if (unlikely(++found > vq->num)) {
1051                         vq_err(vq, "Loop detected: last one at %u "
1052                                "vq size %u head %u\n",
1053                                i, vq->num, head);
1054                         return -EINVAL;
1055                 }
1056                 ret = copy_from_user(&desc, vq->desc + i, sizeof desc);
1057                 if (unlikely(ret)) {
1058                         vq_err(vq, "Failed to get descriptor: idx %d addr %p\n",
1059                                i, vq->desc + i);
1060                         return -EFAULT;
1061                 }
1062                 if (desc.flags & VRING_DESC_F_INDIRECT) {
1063                         ret = get_indirect(dev, vq, iov, iov_size,
1064                                            out_num, in_num,
1065                                            log, log_num, &desc);
1066                         if (unlikely(ret < 0)) {
1067                                 vq_err(vq, "Failure detected "
1068                                        "in indirect descriptor at idx %d\n", i);
1069                                 return ret;
1070                         }
1071                         continue;
1072                 }
1073
1074                 ret = translate_desc(dev, desc.addr, desc.len, iov + iov_count,
1075                                      iov_size - iov_count);
1076                 if (unlikely(ret < 0)) {
1077                         vq_err(vq, "Translation failure %d descriptor idx %d\n",
1078                                ret, i);
1079                         return ret;
1080                 }
1081                 if (desc.flags & VRING_DESC_F_WRITE) {
1082                         /* If this is an input descriptor,
1083                          * increment that count. */
1084                         *in_num += ret;
1085                         if (unlikely(log)) {
1086                                 log[*log_num].addr = desc.addr;
1087                                 log[*log_num].len = desc.len;
1088                                 ++*log_num;
1089                         }
1090                 } else {
1091                         /* If it's an output descriptor, they're all supposed
1092                          * to come before any input descriptors. */
1093                         if (unlikely(*in_num)) {
1094                                 vq_err(vq, "Descriptor has out after in: "
1095                                        "idx %d\n", i);
1096                                 return -EINVAL;
1097                         }
1098                         *out_num += ret;
1099                 }
1100         } while ((i = next_desc(&desc)) != -1);
1101
1102         /* On success, increment avail index. */
1103         vq->last_avail_idx++;
1104         return head;
1105 }
1106
1107 /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
1108 void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
1109 {
1110         vq->last_avail_idx -= n;
1111 }
1112
1113 /* After we've used one of their buffers, we tell them about it.  We'll then
1114  * want to notify the guest, using eventfd. */
1115 int vhost_add_used(struct vhost_virtqueue *vq, unsigned int head, int len)
1116 {
1117         struct vring_used_elem __user *used;
1118
1119         /* The virtqueue contains a ring of used buffers.  Get a pointer to the
1120          * next entry in that used ring. */
1121         used = &vq->used->ring[vq->last_used_idx % vq->num];
1122         if (put_user(head, &used->id)) {
1123                 vq_err(vq, "Failed to write used id");
1124                 return -EFAULT;
1125         }
1126         if (put_user(len, &used->len)) {
1127                 vq_err(vq, "Failed to write used len");
1128                 return -EFAULT;
1129         }
1130         /* Make sure buffer is written before we update index. */
1131         smp_wmb();
1132         if (put_user(vq->last_used_idx + 1, &vq->used->idx)) {
1133                 vq_err(vq, "Failed to increment used idx");
1134                 return -EFAULT;
1135         }
1136         if (unlikely(vq->log_used)) {
1137                 /* Make sure data is seen before log. */
1138                 smp_wmb();
1139                 /* Log used ring entry write. */
1140                 log_write(vq->log_base,
1141                           vq->log_addr +
1142                            ((void __user *)used - (void __user *)vq->used),
1143                           sizeof *used);
1144                 /* Log used index update. */
1145                 log_write(vq->log_base,
1146                           vq->log_addr + offsetof(struct vring_used, idx),
1147                           sizeof vq->used->idx);
1148                 if (vq->log_ctx)
1149                         eventfd_signal(vq->log_ctx, 1);
1150         }
1151         vq->last_used_idx++;
1152         return 0;
1153 }
1154
1155 static int __vhost_add_used_n(struct vhost_virtqueue *vq,
1156                             struct vring_used_elem *heads,
1157                             unsigned count)
1158 {
1159         struct vring_used_elem __user *used;
1160         int start;
1161
1162         start = vq->last_used_idx % vq->num;
1163         used = vq->used->ring + start;
1164         if (copy_to_user(used, heads, count * sizeof *used)) {
1165                 vq_err(vq, "Failed to write used");
1166                 return -EFAULT;
1167         }
1168         if (unlikely(vq->log_used)) {
1169                 /* Make sure data is seen before log. */
1170                 smp_wmb();
1171                 /* Log used ring entry write. */
1172                 log_write(vq->log_base,
1173                           vq->log_addr +
1174                            ((void __user *)used - (void __user *)vq->used),
1175                           count * sizeof *used);
1176         }
1177         vq->last_used_idx += count;
1178         return 0;
1179 }
1180
1181 /* After we've used one of their buffers, we tell them about it.  We'll then
1182  * want to notify the guest, using eventfd. */
1183 int vhost_add_used_n(struct vhost_virtqueue *vq, struct vring_used_elem *heads,
1184                      unsigned count)
1185 {
1186         int start, n, r;
1187
1188         start = vq->last_used_idx % vq->num;
1189         n = vq->num - start;
1190         if (n < count) {
1191                 r = __vhost_add_used_n(vq, heads, n);
1192                 if (r < 0)
1193                         return r;
1194                 heads += n;
1195                 count -= n;
1196         }
1197         r = __vhost_add_used_n(vq, heads, count);
1198
1199         /* Make sure buffer is written before we update index. */
1200         smp_wmb();
1201         if (put_user(vq->last_used_idx, &vq->used->idx)) {
1202                 vq_err(vq, "Failed to increment used idx");
1203                 return -EFAULT;
1204         }
1205         if (unlikely(vq->log_used)) {
1206                 /* Log used index update. */
1207                 log_write(vq->log_base,
1208                           vq->log_addr + offsetof(struct vring_used, idx),
1209                           sizeof vq->used->idx);
1210                 if (vq->log_ctx)
1211                         eventfd_signal(vq->log_ctx, 1);
1212         }
1213         return r;
1214 }
1215
1216 /* This actually signals the guest, using eventfd. */
1217 void vhost_signal(struct vhost_dev *dev, struct vhost_virtqueue *vq)
1218 {
1219         __u16 flags;
1220         /* Flush out used index updates. This is paired
1221          * with the barrier that the Guest executes when enabling
1222          * interrupts. */
1223         smp_mb();
1224
1225         if (get_user(flags, &vq->avail->flags)) {
1226                 vq_err(vq, "Failed to get flags");
1227                 return;
1228         }
1229
1230         /* If they don't want an interrupt, don't signal, unless empty. */
1231         if ((flags & VRING_AVAIL_F_NO_INTERRUPT) &&
1232             (vq->avail_idx != vq->last_avail_idx ||
1233              !vhost_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY)))
1234                 return;
1235
1236         /* Signal the Guest tell them we used something up. */
1237         if (vq->call_ctx)
1238                 eventfd_signal(vq->call_ctx, 1);
1239 }
1240
1241 /* And here's the combo meal deal.  Supersize me! */
1242 void vhost_add_used_and_signal(struct vhost_dev *dev,
1243                                struct vhost_virtqueue *vq,
1244                                unsigned int head, int len)
1245 {
1246         vhost_add_used(vq, head, len);
1247         vhost_signal(dev, vq);
1248 }
1249
1250 /* multi-buffer version of vhost_add_used_and_signal */
1251 void vhost_add_used_and_signal_n(struct vhost_dev *dev,
1252                                  struct vhost_virtqueue *vq,
1253                                  struct vring_used_elem *heads, unsigned count)
1254 {
1255         vhost_add_used_n(vq, heads, count);
1256         vhost_signal(dev, vq);
1257 }
1258
1259 /* OK, now we need to know about added descriptors. */
1260 bool vhost_enable_notify(struct vhost_virtqueue *vq)
1261 {
1262         u16 avail_idx;
1263         int r;
1264         if (!(vq->used_flags & VRING_USED_F_NO_NOTIFY))
1265                 return false;
1266         vq->used_flags &= ~VRING_USED_F_NO_NOTIFY;
1267         r = put_user(vq->used_flags, &vq->used->flags);
1268         if (r) {
1269                 vq_err(vq, "Failed to enable notification at %p: %d\n",
1270                        &vq->used->flags, r);
1271                 return false;
1272         }
1273         /* They could have slipped one in as we were doing that: make
1274          * sure it's written, then check again. */
1275         smp_mb();
1276         r = get_user(avail_idx, &vq->avail->idx);
1277         if (r) {
1278                 vq_err(vq, "Failed to check avail idx at %p: %d\n",
1279                        &vq->avail->idx, r);
1280                 return false;
1281         }
1282
1283         return avail_idx != vq->avail_idx;
1284 }
1285
1286 /* We don't need to be notified again. */
1287 void vhost_disable_notify(struct vhost_virtqueue *vq)
1288 {
1289         int r;
1290         if (vq->used_flags & VRING_USED_F_NO_NOTIFY)
1291                 return;
1292         vq->used_flags |= VRING_USED_F_NO_NOTIFY;
1293         r = put_user(vq->used_flags, &vq->used->flags);
1294         if (r)
1295                 vq_err(vq, "Failed to enable notification at %p: %d\n",
1296                        &vq->used->flags, r);
1297 }