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