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, 2008 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/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/err.h>
  30 #include <linux/kernel.h>
  31 #include <linux/slab.h>
  32 #include <net/sock.h>
  33 #include <linux/net.h>
  34 #include <linux/kthread.h>
  35
  36 #include <asm/uaccess.h>
  37 #include <asm/system.h>
  38 #include <asm/types.h>
  39
  40 #include <linux/nbd.h>
  41
  42 #define LO_MAGIC 0x68797548
  43
  44 #ifdef NDEBUG
  45 #define dprintk(flags, fmt...)
  46 #else /* NDEBUG */
  47 #define dprintk(flags, fmt...) do { \
  48         if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
  49 } while (0)
  50 #define DBG_IOCTL       0x0004
  51 #define DBG_INIT        0x0010
  52 #define DBG_EXIT        0x0020
  53 #define DBG_BLKDEV      0x0100
  54 #define DBG_RX          0x0200
  55 #define DBG_TX          0x0400
  56 static unsigned int debugflags;
  57 #endif /* NDEBUG */
  58
  59 static unsigned int nbds_max = 16;
  60 static struct nbd_device *nbd_dev;
  61 static int max_part;
  62
  63 /*
  64  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  65  * 1. Each NIC is essentially a synchronization point for all servers
  66  *    accessed through that NIC so there's no need to have more locks
  67  *    than NICs anyway.
  68  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  69  *    down each lock to the point where they're actually slower than just
  70  *    a single lock.
  71  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  72  */
  73 static DEFINE_SPINLOCK(nbd_lock);
  74
  75 #ifndef NDEBUG
  76 static const char *ioctl_cmd_to_ascii(int cmd)
  77 {
  78         switch (cmd) {
  79         case NBD_SET_SOCK: return "set-sock";
  80         case NBD_SET_BLKSIZE: return "set-blksize";
  81         case NBD_SET_SIZE: return "set-size";
  82         case NBD_DO_IT: return "do-it";
  83         case NBD_CLEAR_SOCK: return "clear-sock";
  84         case NBD_CLEAR_QUE: return "clear-que";
  85         case NBD_PRINT_DEBUG: return "print-debug";
  86         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
  87         case NBD_DISCONNECT: return "disconnect";
  88         case BLKROSET: return "set-read-only";
  89         case BLKFLSBUF: return "flush-buffer-cache";
  90         }
  91         return "unknown";
  92 }
  93
  94 static const char *nbdcmd_to_ascii(int cmd)
  95 {
  96         switch (cmd) {
  97         case  NBD_CMD_READ: return "read";
  98         case NBD_CMD_WRITE: return "write";
  99         case  NBD_CMD_DISC: return "disconnect";
 100         }
 101         return "invalid";
 102 }
 103 #endif /* NDEBUG */
 104
 105 static void nbd_end_request(struct request *req)
 106 {
 107         int error = req->errors ? -EIO : 0;
 108         struct request_queue *q = req->q;
 109         unsigned long flags;
 110
 111         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 112                         req, error ? "failed" : "done");
 113
 114         spin_lock_irqsave(q->queue_lock, flags);
 115         __blk_end_request_all(req, error);
 116         spin_unlock_irqrestore(q->queue_lock, flags);
 117 }
 118
 119 static void sock_shutdown(struct nbd_device *lo, int lock)
 120 {
 121         /* Forcibly shutdown the socket causing all listeners
 122          * to error
 123          *
 124          * FIXME: This code is duplicated from sys_shutdown, but
 125          * there should be a more generic interface rather than
 126          * calling socket ops directly here */
 127         if (lock)
 128                 mutex_lock(&lo->tx_lock);
 129         if (lo->sock) {
 130                 dev_warn(disk_to_dev(lo->disk), "shutting down socket\n");
 131                 kernel_sock_shutdown(lo->sock, SHUT_RDWR);
 132                 lo->sock = NULL;
 133         }
 134         if (lock)
 135                 mutex_unlock(&lo->tx_lock);
 136 }
 137
 138 static void nbd_xmit_timeout(unsigned long arg)
 139 {
 140         struct task_struct *task = (struct task_struct *)arg;
 141
 142         printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
 143                 task->comm, task->pid);
 144         force_sig(SIGKILL, task);
 145 }
 146
 147 /*
 148  *  Send or receive packet.
 149  */
 150 static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
 151                 int msg_flags)
 152 {
 153         struct socket *sock = lo->sock;
 154         int result;
 155         struct msghdr msg;
 156         struct kvec iov;
 157         sigset_t blocked, oldset;
 158
 159         if (unlikely(!sock)) {
 160                 dev_err(disk_to_dev(lo->disk),
 161                         "Attempted %s on closed socket in sock_xmit\n",
 162                         (send ? "send" : "recv"));
 163                 return -EINVAL;
 164         }
 165
 166         /* Allow interception of SIGKILL only
 167          * Don't allow other signals to interrupt the transmission */
 168         siginitsetinv(&blocked, sigmask(SIGKILL));
 169         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 170
 171         do {
 172                 sock->sk->sk_allocation = GFP_NOIO;
 173                 iov.iov_base = buf;
 174                 iov.iov_len = size;
 175                 msg.msg_name = NULL;
 176                 msg.msg_namelen = 0;
 177                 msg.msg_control = NULL;
 178                 msg.msg_controllen = 0;
 179                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 180
 181                 if (send) {
 182                         struct timer_list ti;
 183
 184                         if (lo->xmit_timeout) {
 185                                 init_timer(&ti);
 186                                 ti.function = nbd_xmit_timeout;
 187                                 ti.data = (unsigned long)current;
 188                                 ti.expires = jiffies + lo->xmit_timeout;
 189                                 add_timer(&ti);
 190                         }
 191                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 192                         if (lo->xmit_timeout)
 193                                 del_timer_sync(&ti);
 194                 } else
 195                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 196                                                 msg.msg_flags);
 197
 198                 if (signal_pending(current)) {
 199                         siginfo_t info;
 200                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 201                                 task_pid_nr(current), current->comm,
 202                                 dequeue_signal_lock(current, &current->blocked, &info));
 203                         result = -EINTR;
 204                         sock_shutdown(lo, !send);
 205                         break;
 206                 }
 207
 208                 if (result <= 0) {
 209                         if (result == 0)
 210                                 result = -EPIPE; /* short read */
 211                         break;
 212                 }
 213                 size -= result;
 214                 buf += result;
 215         } while (size > 0);
 216
 217         sigprocmask(SIG_SETMASK, &oldset, NULL);
 218
 219         return result;
 220 }
 221
 222 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
 223                 int flags)
 224 {
 225         int result;
 226         void *kaddr = kmap(bvec->bv_page);
 227         result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
 228         kunmap(bvec->bv_page);
 229         return result;
 230 }
 231
 232 /* always call with the tx_lock held */
 233 static int nbd_send_req(struct nbd_device *lo, struct request *req)
 234 {
 235         int result, flags;
 236         struct nbd_request request;
 237         unsigned long size = blk_rq_bytes(req);
 238         struct bio *bio;
 239
 240         request.magic = htonl(NBD_REQUEST_MAGIC);
 241         request.type = htonl(nbd_cmd(req));
 242         request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 243         request.len = htonl(size);
 244         memcpy(request.handle, &req, sizeof(req));
 245
 246         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
 247                         lo->disk->disk_name, req,
 248                         nbdcmd_to_ascii(nbd_cmd(req)),
 249                         (unsigned long long)blk_rq_pos(req) << 9,
 250                         blk_rq_bytes(req));
 251         result = sock_xmit(lo, 1, &request, sizeof(request),
 252                         (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
 253         if (result <= 0) {
 254                 dev_err(disk_to_dev(lo->disk),
 255                         "Send control failed (result %d)\n", result);
 256                 goto error_out;
 257         }
 258
 259         if (nbd_cmd(req) != NBD_CMD_WRITE)
 260                 return 0;
 261
 262         flags = 0;
 263         bio = req->bio;
 264         while (bio) {
 265                 struct bio *next = bio->bi_next;
 266                 int i;
 267                 struct bio_vec *bvec;
 268
 269                 bio_for_each_segment(bvec, bio, i) {
 270                         bool is_last = !next && i == bio->bi_vcnt - 1;
 271
 272                         if (is_last)
 273                                 flags = MSG_MORE;
 274                         dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 275                                         lo->disk->disk_name, req, bvec->bv_len);
 276                         result = sock_send_bvec(lo, bvec, flags);
 277                         if (result <= 0) {
 278                                 dev_err(disk_to_dev(lo->disk),
 279                                         "Send data failed (result %d)\n",
 280                                         result);
 281                                 goto error_out;
 282                         }
 283                         /*
 284                          * The completion might already have come in,
 285                          * so break for the last one instead of letting
 286                          * the iterator do it. This prevents use-after-free
 287                          * of the bio.
 288                          */
 289                         if (is_last)
 290                                 break;
 291                 }
 292                 bio = next;
 293         }
 294         return 0;
 295
 296 error_out:
 297         return -EIO;
 298 }
 299
 300 static struct request *nbd_find_request(struct nbd_device *lo,
 301                                         struct request *xreq)
 302 {
 303         struct request *req, *tmp;
 304         int err;
 305
 306         err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
 307         if (unlikely(err))
 308                 goto out;
 309
 310         spin_lock(&lo->queue_lock);
 311         list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
 312                 if (req != xreq)
 313                         continue;
 314                 list_del_init(&req->queuelist);
 315                 spin_unlock(&lo->queue_lock);
 316                 return req;
 317         }
 318         spin_unlock(&lo->queue_lock);
 319
 320         err = -ENOENT;
 321
 322 out:
 323         return ERR_PTR(err);
 324 }
 325
 326 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
 327 {
 328         int result;
 329         void *kaddr = kmap(bvec->bv_page);
 330         result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 331                         MSG_WAITALL);
 332         kunmap(bvec->bv_page);
 333         return result;
 334 }
 335
 336 /* NULL returned = something went wrong, inform userspace */
 337 static struct request *nbd_read_stat(struct nbd_device *lo)
 338 {
 339         int result;
 340         struct nbd_reply reply;
 341         struct request *req;
 342
 343         reply.magic = 0;
 344         result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
 345         if (result <= 0) {
 346                 dev_err(disk_to_dev(lo->disk),
 347                         "Receive control failed (result %d)\n", result);
 348                 goto harderror;
 349         }
 350
 351         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 352                 dev_err(disk_to_dev(lo->disk), "Wrong magic (0x%lx)\n",
 353                                 (unsigned long)ntohl(reply.magic));
 354                 result = -EPROTO;
 355                 goto harderror;
 356         }
 357
 358         req = nbd_find_request(lo, *(struct request **)reply.handle);
 359         if (IS_ERR(req)) {
 360                 result = PTR_ERR(req);
 361                 if (result != -ENOENT)
 362                         goto harderror;
 363
 364                 dev_err(disk_to_dev(lo->disk), "Unexpected reply (%p)\n",
 365                         reply.handle);
 366                 result = -EBADR;
 367                 goto harderror;
 368         }
 369
 370         if (ntohl(reply.error)) {
 371                 dev_err(disk_to_dev(lo->disk), "Other side returned error (%d)\n",
 372                         ntohl(reply.error));
 373                 req->errors++;
 374                 return req;
 375         }
 376
 377         dprintk(DBG_RX, "%s: request %p: got reply\n",
 378                         lo->disk->disk_name, req);
 379         if (nbd_cmd(req) == NBD_CMD_READ) {
 380                 struct req_iterator iter;
 381                 struct bio_vec *bvec;
 382
 383                 rq_for_each_segment(bvec, req, iter) {
 384                         result = sock_recv_bvec(lo, bvec);
 385                         if (result <= 0) {
 386                                 dev_err(disk_to_dev(lo->disk), "Receive data failed (result %d)\n",
 387                                         result);
 388                                 req->errors++;
 389                                 return req;
 390                         }
 391                         dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 392                                 lo->disk->disk_name, req, bvec->bv_len);
 393                 }
 394         }
 395         return req;
 396 harderror:
 397         lo->harderror = result;
 398         return NULL;
 399 }
 400
 401 static ssize_t pid_show(struct device *dev,
 402                         struct device_attribute *attr, char *buf)
 403 {
 404         struct gendisk *disk = dev_to_disk(dev);
 405
 406         return sprintf(buf, "%ld\n",
 407                 (long) ((struct nbd_device *)disk->private_data)->pid);
 408 }
 409
 410 static struct device_attribute pid_attr = {
 411         .attr = { .name = "pid", .mode = S_IRUGO},
 412         .show = pid_show,
 413 };
 414
 415 static int nbd_do_it(struct nbd_device *lo)
 416 {
 417         struct request *req;
 418         int ret;
 419
 420         BUG_ON(lo->magic != LO_MAGIC);
 421
 422         lo->pid = task_pid_nr(current);
 423         ret = device_create_file(disk_to_dev(lo->disk), &pid_attr);
 424         if (ret) {
 425                 dev_err(disk_to_dev(lo->disk), "device_create_file failed!\n");
 426                 lo->pid = 0;
 427                 return ret;
 428         }
 429
 430         while ((req = nbd_read_stat(lo)) != NULL)
 431                 nbd_end_request(req);
 432
 433         device_remove_file(disk_to_dev(lo->disk), &pid_attr);
 434         lo->pid = 0;
 435         return 0;
 436 }
 437
 438 static void nbd_clear_que(struct nbd_device *lo)
 439 {
 440         struct request *req;
 441
 442         BUG_ON(lo->magic != LO_MAGIC);
 443
 444         /*
 445          * Because we have set lo->sock to NULL under the tx_lock, all
 446          * modifications to the list must have completed by now.  For
 447          * the same reason, the active_req must be NULL.
 448          *
 449          * As a consequence, we don't need to take the spin lock while
 450          * purging the list here.
 451          */
 452         BUG_ON(lo->sock);
 453         BUG_ON(lo->active_req);
 454
 455         while (!list_empty(&lo->queue_head)) {
 456                 req = list_entry(lo->queue_head.next, struct request,
 457                                  queuelist);
 458                 list_del_init(&req->queuelist);
 459                 req->errors++;
 460                 nbd_end_request(req);
 461         }
 462
 463         while (!list_empty(&lo->waiting_queue)) {
 464                 req = list_entry(lo->waiting_queue.next, struct request,
 465                                  queuelist);
 466                 list_del_init(&req->queuelist);
 467                 req->errors++;
 468                 nbd_end_request(req);
 469         }
 470 }
 471
 472
 473 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
 474 {
 475         if (req->cmd_type != REQ_TYPE_FS)
 476                 goto error_out;
 477
 478         nbd_cmd(req) = NBD_CMD_READ;
 479         if (rq_data_dir(req) == WRITE) {
 480                 nbd_cmd(req) = NBD_CMD_WRITE;
 481                 if (lo->flags & NBD_READ_ONLY) {
 482                         dev_err(disk_to_dev(lo->disk),
 483                                 "Write on read-only\n");
 484                         goto error_out;
 485                 }
 486         }
 487
 488         req->errors = 0;
 489
 490         mutex_lock(&lo->tx_lock);
 491         if (unlikely(!lo->sock)) {
 492                 mutex_unlock(&lo->tx_lock);
 493                 dev_err(disk_to_dev(lo->disk),
 494                         "Attempted send on closed socket\n");
 495                 goto error_out;
 496         }
 497
 498         lo->active_req = req;
 499
 500         if (nbd_send_req(lo, req) != 0) {
 501                 dev_err(disk_to_dev(lo->disk), "Request send failed\n");
 502                 req->errors++;
 503                 nbd_end_request(req);
 504         } else {
 505                 spin_lock(&lo->queue_lock);
 506                 list_add(&req->queuelist, &lo->queue_head);
 507                 spin_unlock(&lo->queue_lock);
 508         }
 509
 510         lo->active_req = NULL;
 511         mutex_unlock(&lo->tx_lock);
 512         wake_up_all(&lo->active_wq);
 513
 514         return;
 515
 516 error_out:
 517         req->errors++;
 518         nbd_end_request(req);
 519 }
 520
 521 static int nbd_thread(void *data)
 522 {
 523         struct nbd_device *lo = data;
 524         struct request *req;
 525
 526         set_user_nice(current, -20);
 527         while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
 528                 /* wait for something to do */
 529                 wait_event_interruptible(lo->waiting_wq,
 530                                          kthread_should_stop() ||
 531                                          !list_empty(&lo->waiting_queue));
 532
 533                 /* extract request */
 534                 if (list_empty(&lo->waiting_queue))
 535                         continue;
 536
 537                 spin_lock_irq(&lo->queue_lock);
 538                 req = list_entry(lo->waiting_queue.next, struct request,
 539                                  queuelist);
 540                 list_del_init(&req->queuelist);
 541                 spin_unlock_irq(&lo->queue_lock);
 542
 543                 /* handle request */
 544                 nbd_handle_req(lo, req);
 545         }
 546         return 0;
 547 }
 548
 549 /*
 550  * We always wait for result of write, for now. It would be nice to make it optional
 551  * in future
 552  * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
 553  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 554  */
 555
 556 static void do_nbd_request(struct request_queue *q)
 557 {
 558         struct request *req;
 559
 560         while ((req = blk_fetch_request(q)) != NULL) {
 561                 struct nbd_device *lo;
 562
 563                 spin_unlock_irq(q->queue_lock);
 564
 565                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
 566                                 req->rq_disk->disk_name, req, req->cmd_type);
 567
 568                 lo = req->rq_disk->private_data;
 569
 570                 BUG_ON(lo->magic != LO_MAGIC);
 571
 572                 if (unlikely(!lo->sock)) {
 573                         dev_err(disk_to_dev(lo->disk),
 574                                 "Attempted send on closed socket\n");
 575                         req->errors++;
 576                         nbd_end_request(req);
 577                         spin_lock_irq(q->queue_lock);
 578                         continue;
 579                 }
 580
 581                 spin_lock_irq(&lo->queue_lock);
 582                 list_add_tail(&req->queuelist, &lo->waiting_queue);
 583                 spin_unlock_irq(&lo->queue_lock);
 584
 585                 wake_up(&lo->waiting_wq);
 586
 587                 spin_lock_irq(q->queue_lock);
 588         }
 589 }
 590
 591 /* Must be called with tx_lock held */
 592
 593 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
 594                        unsigned int cmd, unsigned long arg)
 595 {
 596         switch (cmd) {
 597         case NBD_DISCONNECT: {
 598                 struct request sreq;
 599
 600                 dev_info(disk_to_dev(lo->disk), "NBD_DISCONNECT\n");
 601                 if (!lo->sock)
 602                         return -EINVAL;
 603
 604                 mutex_unlock(&lo->tx_lock);
 605                 fsync_bdev(bdev);
 606                 mutex_lock(&lo->tx_lock);
 607                 blk_rq_init(NULL, &sreq);
 608                 sreq.cmd_type = REQ_TYPE_SPECIAL;
 609                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 610
 611                 /* Check again after getting mutex back.  */
 612                 if (!lo->sock)
 613                         return -EINVAL;
 614
 615                 lo->disconnect = 1;
 616
 617                 nbd_send_req(lo, &sreq);
 618                 return 0;
 619         }
 620
 621         case NBD_CLEAR_SOCK: {
 622                 struct file *file;
 623
 624                 lo->sock = NULL;
 625                 file = lo->file;
 626                 lo->file = NULL;
 627                 nbd_clear_que(lo);
 628                 BUG_ON(!list_empty(&lo->queue_head));
 629                 BUG_ON(!list_empty(&lo->waiting_queue));
 630                 kill_bdev(bdev);
 631                 if (file)
 632                         fput(file);
 633                 return 0;
 634         }
 635
 636         case NBD_SET_SOCK: {
 637                 struct file *file;
 638                 if (lo->file)
 639                         return -EBUSY;
 640                 file = fget(arg);
 641                 if (file) {
 642                         struct inode *inode = file->f_path.dentry->d_inode;
 643                         if (S_ISSOCK(inode->i_mode)) {
 644                                 lo->file = file;
 645                                 lo->sock = SOCKET_I(inode);
 646                                 if (max_part > 0)
 647                                         bdev->bd_invalidated = 1;
 648                                 lo->disconnect = 0; /* we're connected now */
 649                                 return 0;
 650                         } else {
 651                                 fput(file);
 652                         }
 653                 }
 654                 return -EINVAL;
 655         }
 656
 657         case NBD_SET_BLKSIZE:
 658                 lo->blksize = arg;
 659                 lo->bytesize &= ~(lo->blksize-1);
 660                 bdev->bd_inode->i_size = lo->bytesize;
 661                 set_blocksize(bdev, lo->blksize);
 662                 set_capacity(lo->disk, lo->bytesize >> 9);
 663                 return 0;
 664
 665         case NBD_SET_SIZE:
 666                 lo->bytesize = arg & ~(lo->blksize-1);
 667                 bdev->bd_inode->i_size = lo->bytesize;
 668                 set_blocksize(bdev, lo->blksize);
 669                 set_capacity(lo->disk, lo->bytesize >> 9);
 670                 return 0;
 671
 672         case NBD_SET_TIMEOUT:
 673                 lo->xmit_timeout = arg * HZ;
 674                 return 0;
 675
 676         case NBD_SET_SIZE_BLOCKS:
 677                 lo->bytesize = ((u64) arg) * lo->blksize;
 678                 bdev->bd_inode->i_size = lo->bytesize;
 679                 set_blocksize(bdev, lo->blksize);
 680                 set_capacity(lo->disk, lo->bytesize >> 9);
 681                 return 0;
 682
 683         case NBD_DO_IT: {
 684                 struct task_struct *thread;
 685                 struct file *file;
 686                 int error;
 687
 688                 if (lo->pid)
 689                         return -EBUSY;
 690                 if (!lo->file)
 691                         return -EINVAL;
 692
 693                 mutex_unlock(&lo->tx_lock);
 694
 695                 thread = kthread_create(nbd_thread, lo, "%s",
 696                                         lo->disk->disk_name);
 697                 if (IS_ERR(thread)) {
 698                         mutex_lock(&lo->tx_lock);
 699                         return PTR_ERR(thread);
 700                 }
 701                 wake_up_process(thread);
 702                 error = nbd_do_it(lo);
 703                 kthread_stop(thread);
 704
 705                 mutex_lock(&lo->tx_lock);
 706                 if (error)
 707                         return error;
 708                 sock_shutdown(lo, 0);
 709                 file = lo->file;
 710                 lo->file = NULL;
 711                 nbd_clear_que(lo);
 712                 dev_warn(disk_to_dev(lo->disk), "queue cleared\n");
 713                 kill_bdev(bdev);
 714                 if (file)
 715                         fput(file);
 716                 lo->bytesize = 0;
 717                 bdev->bd_inode->i_size = 0;
 718                 set_capacity(lo->disk, 0);
 719                 if (max_part > 0)
 720                         ioctl_by_bdev(bdev, BLKRRPART, 0);
 721                 if (lo->disconnect) /* user requested, ignore socket errors */
 722                         return 0;
 723                 return lo->harderror;
 724         }
 725
 726         case NBD_CLEAR_QUE:
 727                 /*
 728                  * This is for compatibility only.  The queue is always cleared
 729                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 730                  */
 731                 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
 732                 return 0;
 733
 734         case NBD_PRINT_DEBUG:
 735                 dev_info(disk_to_dev(lo->disk),
 736                         "next = %p, prev = %p, head = %p\n",
 737                         lo->queue_head.next, lo->queue_head.prev,
 738                         &lo->queue_head);
 739                 return 0;
 740         }
 741         return -ENOTTY;
 742 }
 743
 744 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 745                      unsigned int cmd, unsigned long arg)
 746 {
 747         struct nbd_device *lo = bdev->bd_disk->private_data;
 748         int error;
 749
 750         if (!capable(CAP_SYS_ADMIN))
 751                 return -EPERM;
 752
 753         BUG_ON(lo->magic != LO_MAGIC);
 754
 755         /* Anyone capable of this syscall can do *real bad* things */
 756         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 757                         lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 758
 759         mutex_lock(&lo->tx_lock);
 760         error = __nbd_ioctl(bdev, lo, cmd, arg);
 761         mutex_unlock(&lo->tx_lock);
 762
 763         return error;
 764 }
 765
 766 static const struct block_device_operations nbd_fops =
 767 {
 768         .owner =        THIS_MODULE,
 769         .ioctl =        nbd_ioctl,
 770 };
 771
 772 /*
 773  * And here should be modules and kernel interface
 774  *  (Just smiley confuses emacs :-)
 775  */
 776
 777 static int __init nbd_init(void)
 778 {
 779         int err = -ENOMEM;
 780         int i;
 781         int part_shift;
 782
 783         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 784
 785         if (max_part < 0) {
 786                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
 787                 return -EINVAL;
 788         }
 789
 790         part_shift = 0;
 791         if (max_part > 0) {
 792                 part_shift = fls(max_part);
 793
 794                 /*
 795                  * Adjust max_part according to part_shift as it is exported
 796                  * to user space so that user can know the max number of
 797                  * partition kernel should be able to manage.
 798                  *
 799                  * Note that -1 is required because partition 0 is reserved
 800                  * for the whole disk.
 801                  */
 802                 max_part = (1UL << part_shift) - 1;
 803         }
 804
 805         if ((1UL << part_shift) > DISK_MAX_PARTS)
 806                 return -EINVAL;
 807
 808         if (nbds_max > 1UL << (MINORBITS - part_shift))
 809                 return -EINVAL;
 810
 811         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
 812         if (!nbd_dev)
 813                 return -ENOMEM;
 814
 815         for (i = 0; i < nbds_max; i++) {
 816                 struct gendisk *disk = alloc_disk(1 << part_shift);
 817                 if (!disk)
 818                         goto out;
 819                 nbd_dev[i].disk = disk;
 820                 /*
 821                  * The new linux 2.5 block layer implementation requires
 822                  * every gendisk to have its very own request_queue struct.
 823                  * These structs are big so we dynamically allocate them.
 824                  */
 825                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 826                 if (!disk->queue) {
 827                         put_disk(disk);
 828                         goto out;
 829                 }
 830                 /*
 831                  * Tell the block layer that we are not a rotational device
 832                  */
 833                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
 834         }
 835
 836         if (register_blkdev(NBD_MAJOR, "nbd")) {
 837                 err = -EIO;
 838                 goto out;
 839         }
 840
 841         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 842         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 843
 844         for (i = 0; i < nbds_max; i++) {
 845                 struct gendisk *disk = nbd_dev[i].disk;
 846                 nbd_dev[i].file = NULL;
 847                 nbd_dev[i].magic = LO_MAGIC;
 848                 nbd_dev[i].flags = 0;
 849                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
 850                 spin_lock_init(&nbd_dev[i].queue_lock);
 851                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 852                 mutex_init(&nbd_dev[i].tx_lock);
 853                 init_waitqueue_head(&nbd_dev[i].active_wq);
 854                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
 855                 nbd_dev[i].blksize = 1024;
 856                 nbd_dev[i].bytesize = 0;
 857                 disk->major = NBD_MAJOR;
 858                 disk->first_minor = i << part_shift;
 859                 disk->fops = &nbd_fops;
 860                 disk->private_data = &nbd_dev[i];
 861                 sprintf(disk->disk_name, "nbd%d", i);
 862                 set_capacity(disk, 0);
 863                 add_disk(disk);
 864         }
 865
 866         return 0;
 867 out:
 868         while (i--) {
 869                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 870                 put_disk(nbd_dev[i].disk);
 871         }
 872         kfree(nbd_dev);
 873         return err;
 874 }
 875
 876 static void __exit nbd_cleanup(void)
 877 {
 878         int i;
 879         for (i = 0; i < nbds_max; i++) {
 880                 struct gendisk *disk = nbd_dev[i].disk;
 881                 nbd_dev[i].magic = 0;
 882                 if (disk) {
 883                         del_gendisk(disk);
 884                         blk_cleanup_queue(disk->queue);
 885                         put_disk(disk);
 886                 }
 887         }
 888         unregister_blkdev(NBD_MAJOR, "nbd");
 889         kfree(nbd_dev);
 890         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 891 }
 892
 893 module_init(nbd_init);
 894 module_exit(nbd_cleanup);
 895
 896 MODULE_DESCRIPTION("Network Block Device");
 897 MODULE_LICENSE("GPL");
 898
 899 module_param(nbds_max, int, 0444);
 900 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
 901 module_param(max_part, int, 0444);
 902 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
 903 #ifndef NDEBUG
 904 module_param(debugflags, int, 0644);
 905 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 906 #endif