drivers/block/nbd.c

   1 /*
   2  * Network block device - make block devices work over TCP
   3  *
   4  * Note that you can not swap over this thing, yet. Seems to work but
   5  * deadlocks sometimes - you can not swap over TCP in general.
   6  *
   7  * Copyright 1997-2000 Pavel Machek <pavel@ucw.cz>
   8  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
   9  *
  10  * This file is released under GPLv2 or later.
  11  *
  12  * (part of code stolen from loop.c)
  13  */
  14
  15 #include <linux/major.h>
  16
  17 #include <linux/blkdev.h>
  18 #include <linux/module.h>
  19 #include <linux/init.h>
  20 #include <linux/sched.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/compiler.h>
  28 #include <linux/err.h>
  29 #include <linux/kernel.h>
  30 #include <net/sock.h>
  31
  32 #include <linux/devfs_fs_kernel.h>
  33
  34 #include <asm/uaccess.h>
  35 #include <asm/system.h>
  36 #include <asm/types.h>
  37
  38 #include <linux/nbd.h>
  39
  40 #define LO_MAGIC 0x68797548
  41
  42 #ifdef NDEBUG
  43 #define dprintk(flags, fmt...)
  44 #else /* NDEBUG */
  45 #define dprintk(flags, fmt...) do { \
  46         if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
  47 } while (0)
  48 #define DBG_IOCTL       0x0004
  49 #define DBG_INIT        0x0010
  50 #define DBG_EXIT        0x0020
  51 #define DBG_BLKDEV      0x0100
  52 #define DBG_RX          0x0200
  53 #define DBG_TX          0x0400
  54 static unsigned int debugflags;
  55 static unsigned int nbds_max = 16;
  56 #endif /* NDEBUG */
  57
  58 static struct nbd_device nbd_dev[MAX_NBD];
  59
  60 /*
  61  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  62  * 1. Each NIC is essentially a synchronization point for all servers
  63  *    accessed through that NIC so there's no need to have more locks
  64  *    than NICs anyway.
  65  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  66  *    down each lock to the point where they're actually slower than just
  67  *    a single lock.
  68  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  69  */
  70 static DEFINE_SPINLOCK(nbd_lock);
  71
  72 #ifndef NDEBUG
  73 static const char *ioctl_cmd_to_ascii(int cmd)
  74 {
  75         switch (cmd) {
  76         case NBD_SET_SOCK: return "set-sock";
  77         case NBD_SET_BLKSIZE: return "set-blksize";
  78         case NBD_SET_SIZE: return "set-size";
  79         case NBD_DO_IT: return "do-it";
  80         case NBD_CLEAR_SOCK: return "clear-sock";
  81         case NBD_CLEAR_QUE: return "clear-que";
  82         case NBD_PRINT_DEBUG: return "print-debug";
  83         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
  84         case NBD_DISCONNECT: return "disconnect";
  85         case BLKROSET: return "set-read-only";
  86         case BLKFLSBUF: return "flush-buffer-cache";
  87         }
  88         return "unknown";
  89 }
  90
  91 static const char *nbdcmd_to_ascii(int cmd)
  92 {
  93         switch (cmd) {
  94         case  NBD_CMD_READ: return "read";
  95         case NBD_CMD_WRITE: return "write";
  96         case  NBD_CMD_DISC: return "disconnect";
  97         }
  98         return "invalid";
  99 }
 100 #endif /* NDEBUG */
 101
 102 static void nbd_end_request(struct request *req)
 103 {
 104         int uptodate = (req->errors == 0) ? 1 : 0;
 105         request_queue_t *q = req->q;
 106         unsigned long flags;
 107
 108         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 109                         req, uptodate? "done": "failed");
 110
 111         spin_lock_irqsave(q->queue_lock, flags);
 112         if (!end_that_request_first(req, uptodate, req->nr_sectors)) {
 113                 end_that_request_last(req, uptodate);
 114         }
 115         spin_unlock_irqrestore(q->queue_lock, flags);
 116 }
 117
 118 /*
 119  *  Send or receive packet.
 120  */
 121 static int sock_xmit(struct socket *sock, int send, void *buf, int size,
 122                 int msg_flags)
 123 {
 124         int result;
 125         struct msghdr msg;
 126         struct kvec iov;
 127         unsigned long flags;
 128         sigset_t oldset;
 129
 130         /* Allow interception of SIGKILL only
 131          * Don't allow other signals to interrupt the transmission */
 132         spin_lock_irqsave(&current->sighand->siglock, flags);
 133         oldset = current->blocked;
 134         sigfillset(&current->blocked);
 135         sigdelsetmask(&current->blocked, sigmask(SIGKILL));
 136         recalc_sigpending();
 137         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 138
 139         do {
 140                 sock->sk->sk_allocation = GFP_NOIO;
 141                 iov.iov_base = buf;
 142                 iov.iov_len = size;
 143                 msg.msg_name = NULL;
 144                 msg.msg_namelen = 0;
 145                 msg.msg_control = NULL;
 146                 msg.msg_controllen = 0;
 147                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 148
 149                 if (send)
 150                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 151                 else
 152                         result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
 153
 154                 if (signal_pending(current)) {
 155                         siginfo_t info;
 156                         spin_lock_irqsave(&current->sighand->siglock, flags);
 157                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 158                                 current->pid, current->comm,
 159                                 dequeue_signal(current, &current->blocked, &info));
 160                         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 161                         result = -EINTR;
 162                         break;
 163                 }
 164
 165                 if (result <= 0) {
 166                         if (result == 0)
 167                                 result = -EPIPE; /* short read */
 168                         break;
 169                 }
 170                 size -= result;
 171                 buf += result;
 172         } while (size > 0);
 173
 174         spin_lock_irqsave(&current->sighand->siglock, flags);
 175         current->blocked = oldset;
 176         recalc_sigpending();
 177         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 178
 179         return result;
 180 }
 181
 182 static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec,
 183                 int flags)
 184 {
 185         int result;
 186         void *kaddr = kmap(bvec->bv_page);
 187         result = sock_xmit(sock, 1, kaddr + bvec->bv_offset, bvec->bv_len,
 188                         flags);
 189         kunmap(bvec->bv_page);
 190         return result;
 191 }
 192
 193 static int nbd_send_req(struct nbd_device *lo, struct request *req)
 194 {
 195         int result, i, flags;
 196         struct nbd_request request;
 197         unsigned long size = req->nr_sectors << 9;
 198         struct socket *sock = lo->sock;
 199
 200         request.magic = htonl(NBD_REQUEST_MAGIC);
 201         request.type = htonl(nbd_cmd(req));
 202         request.from = cpu_to_be64((u64) req->sector << 9);
 203         request.len = htonl(size);
 204         memcpy(request.handle, &req, sizeof(req));
 205
 206         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
 207                         lo->disk->disk_name, req,
 208                         nbdcmd_to_ascii(nbd_cmd(req)),
 209                         (unsigned long long)req->sector << 9,
 210                         req->nr_sectors << 9);
 211         result = sock_xmit(sock, 1, &request, sizeof(request),
 212                         (nbd_cmd(req) == NBD_CMD_WRITE)? MSG_MORE: 0);
 213         if (result <= 0) {
 214                 printk(KERN_ERR "%s: Send control failed (result %d)\n",
 215                                 lo->disk->disk_name, result);
 216                 goto error_out;
 217         }
 218
 219         if (nbd_cmd(req) == NBD_CMD_WRITE) {
 220                 struct bio *bio;
 221                 /*
 222                  * we are really probing at internals to determine
 223                  * whether to set MSG_MORE or not...
 224                  */
 225                 rq_for_each_bio(bio, req) {
 226                         struct bio_vec *bvec;
 227                         bio_for_each_segment(bvec, bio, i) {
 228                                 flags = 0;
 229                                 if ((i < (bio->bi_vcnt - 1)) || bio->bi_next)
 230                                         flags = MSG_MORE;
 231                                 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 232                                                 lo->disk->disk_name, req,
 233                                                 bvec->bv_len);
 234                                 result = sock_send_bvec(sock, bvec, flags);
 235                                 if (result <= 0) {
 236                                         printk(KERN_ERR "%s: Send data failed (result %d)\n",
 237                                                         lo->disk->disk_name,
 238                                                         result);
 239                                         goto error_out;
 240                                 }
 241                         }
 242                 }
 243         }
 244         return 0;
 245
 246 error_out:
 247         return 1;
 248 }
 249
 250 static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
 251 {
 252         struct request *req;
 253         struct list_head *tmp;
 254         struct request *xreq;
 255         int err;
 256
 257         memcpy(&xreq, handle, sizeof(xreq));
 258
 259         err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
 260         if (unlikely(err))
 261                 goto out;
 262
 263         spin_lock(&lo->queue_lock);
 264         list_for_each(tmp, &lo->queue_head) {
 265                 req = list_entry(tmp, struct request, queuelist);
 266                 if (req != xreq)
 267                         continue;
 268                 list_del_init(&req->queuelist);
 269                 spin_unlock(&lo->queue_lock);
 270                 return req;
 271         }
 272         spin_unlock(&lo->queue_lock);
 273
 274         err = -ENOENT;
 275
 276 out:
 277         return ERR_PTR(err);
 278 }
 279
 280 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
 281 {
 282         int result;
 283         void *kaddr = kmap(bvec->bv_page);
 284         result = sock_xmit(sock, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 285                         MSG_WAITALL);
 286         kunmap(bvec->bv_page);
 287         return result;
 288 }
 289
 290 /* NULL returned = something went wrong, inform userspace */
 291 static struct request *nbd_read_stat(struct nbd_device *lo)
 292 {
 293         int result;
 294         struct nbd_reply reply;
 295         struct request *req;
 296         struct socket *sock = lo->sock;
 297
 298         reply.magic = 0;
 299         result = sock_xmit(sock, 0, &reply, sizeof(reply), MSG_WAITALL);
 300         if (result <= 0) {
 301                 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
 302                                 lo->disk->disk_name, result);
 303                 goto harderror;
 304         }
 305         req = nbd_find_request(lo, reply.handle);
 306         if (unlikely(IS_ERR(req))) {
 307                 result = PTR_ERR(req);
 308                 if (result != -ENOENT)
 309                         goto harderror;
 310
 311                 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
 312                                 lo->disk->disk_name, reply.handle);
 313                 result = -EBADR;
 314                 goto harderror;
 315         }
 316
 317         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 318                 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
 319                                 lo->disk->disk_name,
 320                                 (unsigned long)ntohl(reply.magic));
 321                 result = -EPROTO;
 322                 goto harderror;
 323         }
 324         if (ntohl(reply.error)) {
 325                 printk(KERN_ERR "%s: Other side returned error (%d)\n",
 326                                 lo->disk->disk_name, ntohl(reply.error));
 327                 req->errors++;
 328                 return req;
 329         }
 330
 331         dprintk(DBG_RX, "%s: request %p: got reply\n",
 332                         lo->disk->disk_name, req);
 333         if (nbd_cmd(req) == NBD_CMD_READ) {
 334                 int i;
 335                 struct bio *bio;
 336                 rq_for_each_bio(bio, req) {
 337                         struct bio_vec *bvec;
 338                         bio_for_each_segment(bvec, bio, i) {
 339                                 result = sock_recv_bvec(sock, bvec);
 340                                 if (result <= 0) {
 341                                         printk(KERN_ERR "%s: Receive data failed (result %d)\n",
 342                                                         lo->disk->disk_name,
 343                                                         result);
 344                                         goto harderror;
 345                                 }
 346                                 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 347                                         lo->disk->disk_name, req, bvec->bv_len);
 348                         }
 349                 }
 350         }
 351         return req;
 352 harderror:
 353         lo->harderror = result;
 354         return NULL;
 355 }
 356
 357 static void nbd_do_it(struct nbd_device *lo)
 358 {
 359         struct request *req;
 360
 361         BUG_ON(lo->magic != LO_MAGIC);
 362
 363         while ((req = nbd_read_stat(lo)) != NULL)
 364                 nbd_end_request(req);
 365         return;
 366 }
 367
 368 static void nbd_clear_que(struct nbd_device *lo)
 369 {
 370         struct request *req;
 371
 372         BUG_ON(lo->magic != LO_MAGIC);
 373
 374         /*
 375          * Because we have set lo->sock to NULL under the tx_lock, all
 376          * modifications to the list must have completed by now.  For
 377          * the same reason, the active_req must be NULL.
 378          *
 379          * As a consequence, we don't need to take the spin lock while
 380          * purging the list here.
 381          */
 382         BUG_ON(lo->sock);
 383         BUG_ON(lo->active_req);
 384
 385         while (!list_empty(&lo->queue_head)) {
 386                 req = list_entry(lo->queue_head.next, struct request,
 387                                  queuelist);
 388                 list_del_init(&req->queuelist);
 389                 req->errors++;
 390                 nbd_end_request(req);
 391         }
 392 }
 393
 394 /*
 395  * We always wait for result of write, for now. It would be nice to make it optional
 396  * in future
 397  * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
 398  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 399  */
 400
 401 static void do_nbd_request(request_queue_t * q)
 402 {
 403         struct request *req;
 404
 405         while ((req = elv_next_request(q)) != NULL) {
 406                 struct nbd_device *lo;
 407
 408                 blkdev_dequeue_request(req);
 409                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%lx)\n",
 410                                 req->rq_disk->disk_name, req, req->flags);
 411
 412                 if (!(req->flags & REQ_CMD))
 413                         goto error_out;
 414
 415                 lo = req->rq_disk->private_data;
 416
 417                 BUG_ON(lo->magic != LO_MAGIC);
 418
 419                 nbd_cmd(req) = NBD_CMD_READ;
 420                 if (rq_data_dir(req) == WRITE) {
 421                         nbd_cmd(req) = NBD_CMD_WRITE;
 422                         if (lo->flags & NBD_READ_ONLY) {
 423                                 printk(KERN_ERR "%s: Write on read-only\n",
 424                                                 lo->disk->disk_name);
 425                                 goto error_out;
 426                         }
 427                 }
 428
 429                 req->errors = 0;
 430                 spin_unlock_irq(q->queue_lock);
 431
 432                 mutex_lock(&lo->tx_lock);
 433                 if (unlikely(!lo->sock)) {
 434                         mutex_unlock(&lo->tx_lock);
 435                         printk(KERN_ERR "%s: Attempted send on closed socket\n",
 436                                lo->disk->disk_name);
 437                         req->errors++;
 438                         nbd_end_request(req);
 439                         spin_lock_irq(q->queue_lock);
 440                         continue;
 441                 }
 442
 443                 lo->active_req = req;
 444
 445                 if (nbd_send_req(lo, req) != 0) {
 446                         printk(KERN_ERR "%s: Request send failed\n",
 447                                         lo->disk->disk_name);
 448                         req->errors++;
 449                         nbd_end_request(req);
 450                 } else {
 451                         spin_lock(&lo->queue_lock);
 452                         list_add(&req->queuelist, &lo->queue_head);
 453                         spin_unlock(&lo->queue_lock);
 454                 }
 455
 456                 lo->active_req = NULL;
 457                 mutex_unlock(&lo->tx_lock);
 458                 wake_up_all(&lo->active_wq);
 459
 460                 spin_lock_irq(q->queue_lock);
 461                 continue;
 462
 463 error_out:
 464                 req->errors++;
 465                 spin_unlock(q->queue_lock);
 466                 nbd_end_request(req);
 467                 spin_lock(q->queue_lock);
 468         }
 469         return;
 470 }
 471
 472 static int nbd_ioctl(struct inode *inode, struct file *file,
 473                      unsigned int cmd, unsigned long arg)
 474 {
 475         struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
 476         int error;
 477         struct request sreq ;
 478
 479         if (!capable(CAP_SYS_ADMIN))
 480                 return -EPERM;
 481
 482         BUG_ON(lo->magic != LO_MAGIC);
 483
 484         /* Anyone capable of this syscall can do *real bad* things */
 485         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 486                         lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 487
 488         switch (cmd) {
 489         case NBD_DISCONNECT:
 490                 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
 491                 sreq.flags = REQ_SPECIAL;
 492                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 493                 /*
 494                  * Set these to sane values in case server implementation
 495                  * fails to check the request type first and also to keep
 496                  * debugging output cleaner.
 497                  */
 498                 sreq.sector = 0;
 499                 sreq.nr_sectors = 0;
 500                 if (!lo->sock)
 501                         return -EINVAL;
 502                 nbd_send_req(lo, &sreq);
 503                 return 0;
 504
 505         case NBD_CLEAR_SOCK:
 506                 error = 0;
 507                 mutex_lock(&lo->tx_lock);
 508                 lo->sock = NULL;
 509                 mutex_unlock(&lo->tx_lock);
 510                 file = lo->file;
 511                 lo->file = NULL;
 512                 nbd_clear_que(lo);
 513                 BUG_ON(!list_empty(&lo->queue_head));
 514                 if (file)
 515                         fput(file);
 516                 return error;
 517         case NBD_SET_SOCK:
 518                 if (lo->file)
 519                         return -EBUSY;
 520                 error = -EINVAL;
 521                 file = fget(arg);
 522                 if (file) {
 523                         inode = file->f_dentry->d_inode;
 524                         if (S_ISSOCK(inode->i_mode)) {
 525                                 lo->file = file;
 526                                 lo->sock = SOCKET_I(inode);
 527                                 error = 0;
 528                         } else {
 529                                 fput(file);
 530                         }
 531                 }
 532                 return error;
 533         case NBD_SET_BLKSIZE:
 534                 lo->blksize = arg;
 535                 lo->bytesize &= ~(lo->blksize-1);
 536                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 537                 set_blocksize(inode->i_bdev, lo->blksize);
 538                 set_capacity(lo->disk, lo->bytesize >> 9);
 539                 return 0;
 540         case NBD_SET_SIZE:
 541                 lo->bytesize = arg & ~(lo->blksize-1);
 542                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 543                 set_blocksize(inode->i_bdev, lo->blksize);
 544                 set_capacity(lo->disk, lo->bytesize >> 9);
 545                 return 0;
 546         case NBD_SET_SIZE_BLOCKS:
 547                 lo->bytesize = ((u64) arg) * lo->blksize;
 548                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 549                 set_blocksize(inode->i_bdev, lo->blksize);
 550                 set_capacity(lo->disk, lo->bytesize >> 9);
 551                 return 0;
 552         case NBD_DO_IT:
 553                 if (!lo->file)
 554                         return -EINVAL;
 555                 nbd_do_it(lo);
 556                 /* on return tidy up in case we have a signal */
 557                 /* Forcibly shutdown the socket causing all listeners
 558                  * to error
 559                  *
 560                  * FIXME: This code is duplicated from sys_shutdown, but
 561                  * there should be a more generic interface rather than
 562                  * calling socket ops directly here */
 563                 mutex_lock(&lo->tx_lock);
 564                 if (lo->sock) {
 565                         printk(KERN_WARNING "%s: shutting down socket\n",
 566                                 lo->disk->disk_name);
 567                         lo->sock->ops->shutdown(lo->sock,
 568                                 SEND_SHUTDOWN|RCV_SHUTDOWN);
 569                         lo->sock = NULL;
 570                 }
 571                 mutex_unlock(&lo->tx_lock);
 572                 file = lo->file;
 573                 lo->file = NULL;
 574                 nbd_clear_que(lo);
 575                 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
 576                 if (file)
 577                         fput(file);
 578                 return lo->harderror;
 579         case NBD_CLEAR_QUE:
 580                 /*
 581                  * This is for compatibility only.  The queue is always cleared
 582                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 583                  */
 584                 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
 585                 return 0;
 586         case NBD_PRINT_DEBUG:
 587                 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
 588                         inode->i_bdev->bd_disk->disk_name,
 589                         lo->queue_head.next, lo->queue_head.prev,
 590                         &lo->queue_head);
 591                 return 0;
 592         }
 593         return -EINVAL;
 594 }
 595
 596 static struct block_device_operations nbd_fops =
 597 {
 598         .owner =        THIS_MODULE,
 599         .ioctl =        nbd_ioctl,
 600 };
 601
 602 /*
 603  * And here should be modules and kernel interface
 604  *  (Just smiley confuses emacs :-)
 605  */
 606
 607 static int __init nbd_init(void)
 608 {
 609         int err = -ENOMEM;
 610         int i;
 611
 612         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 613
 614         if (nbds_max > MAX_NBD) {
 615                 printk(KERN_CRIT "nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD,
 616                                 nbds_max);
 617                 return -EINVAL;
 618         }
 619
 620         for (i = 0; i < nbds_max; i++) {
 621                 struct gendisk *disk = alloc_disk(1);
 622                 if (!disk)
 623                         goto out;
 624                 nbd_dev[i].disk = disk;
 625                 /*
 626                  * The new linux 2.5 block layer implementation requires
 627                  * every gendisk to have its very own request_queue struct.
 628                  * These structs are big so we dynamically allocate them.
 629                  */
 630                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 631                 if (!disk->queue) {
 632                         put_disk(disk);
 633                         goto out;
 634                 }
 635         }
 636
 637         if (register_blkdev(NBD_MAJOR, "nbd")) {
 638                 err = -EIO;
 639                 goto out;
 640         }
 641
 642         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 643         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 644
 645         devfs_mk_dir("nbd");
 646         for (i = 0; i < nbds_max; i++) {
 647                 struct gendisk *disk = nbd_dev[i].disk;
 648                 nbd_dev[i].file = NULL;
 649                 nbd_dev[i].magic = LO_MAGIC;
 650                 nbd_dev[i].flags = 0;
 651                 spin_lock_init(&nbd_dev[i].queue_lock);
 652                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 653                 mutex_init(&nbd_dev[i].tx_lock);
 654                 init_waitqueue_head(&nbd_dev[i].active_wq);
 655                 nbd_dev[i].blksize = 1024;
 656                 nbd_dev[i].bytesize = 0x7ffffc00ULL << 10; /* 2TB */
 657                 disk->major = NBD_MAJOR;
 658                 disk->first_minor = i;
 659                 disk->fops = &nbd_fops;
 660                 disk->private_data = &nbd_dev[i];
 661                 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
 662                 sprintf(disk->disk_name, "nbd%d", i);
 663                 sprintf(disk->devfs_name, "nbd/%d", i);
 664                 set_capacity(disk, 0x7ffffc00ULL << 1); /* 2 TB */
 665                 add_disk(disk);
 666         }
 667
 668         return 0;
 669 out:
 670         while (i--) {
 671                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 672                 put_disk(nbd_dev[i].disk);
 673         }
 674         return err;
 675 }
 676
 677 static void __exit nbd_cleanup(void)
 678 {
 679         int i;
 680         for (i = 0; i < nbds_max; i++) {
 681                 struct gendisk *disk = nbd_dev[i].disk;
 682                 nbd_dev[i].magic = 0;
 683                 if (disk) {
 684                         del_gendisk(disk);
 685                         blk_cleanup_queue(disk->queue);
 686                         put_disk(disk);
 687                 }
 688         }
 689         devfs_remove("nbd");
 690         unregister_blkdev(NBD_MAJOR, "nbd");
 691         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 692 }
 693
 694 module_init(nbd_init);
 695 module_exit(nbd_cleanup);
 696
 697 MODULE_DESCRIPTION("Network Block Device");
 698 MODULE_LICENSE("GPL");
 699
 700 module_param(nbds_max, int, 0444);
 701 MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize.");
 702 #ifndef NDEBUG
 703 module_param(debugflags, int, 0644);
 704 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 705 #endif