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