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