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  * (part of code stolen from loop.c)
  11  *
  12  * 97-3-25 compiled 0-th version, not yet tested it
  13  *   (it did not work, BTW) (later that day) HEY! it works!
  14  *   (bit later) hmm, not that much... 2:00am next day:
  15  *   yes, it works, but it gives something like 50kB/sec
  16  * 97-4-01 complete rewrite to make it possible for many requests at
  17  *   once to be processed
  18  * 97-4-11 Making protocol independent of endianity etc.
  19  * 97-9-13 Cosmetic changes
  20  * 98-5-13 Attempt to make 64-bit-clean on 64-bit machines
  21  * 99-1-11 Attempt to make 64-bit-clean on 32-bit machines <ankry@mif.pg.gda.pl>
  22  * 01-2-27 Fix to store proper blockcount for kernel (calculated using
  23  *   BLOCK_SIZE_BITS, not device blocksize) <aga@permonline.ru>
  24  * 01-3-11 Make nbd work with new Linux block layer code. It now supports
  25  *   plugging like all the other block devices. Also added in MSG_MORE to
  26  *   reduce number of partial TCP segments sent. <steve@chygwyn.com>
  27  * 01-12-6 Fix deadlock condition by making queue locks independent of
  28  *   the transmit lock. <steve@chygwyn.com>
  29  * 02-10-11 Allow hung xmit to be aborted via SIGKILL & various fixes.
  30  *   <Paul.Clements@SteelEye.com> <James.Bottomley@SteelEye.com>
  31  * 03-06-22 Make nbd work with new linux 2.5 block layer design. This fixes
  32  *   memory corruption from module removal and possible memory corruption
  33  *   from sending/receiving disk data. <ldl@aros.net>
  34  * 03-06-23 Cosmetic changes. <ldl@aros.net>
  35  * 03-06-23 Enhance diagnostics support. <ldl@aros.net>
  36  * 03-06-24 Remove unneeded blksize_bits field from nbd_device struct.
  37  *   <ldl@aros.net>
  38  * 03-06-24 Cleanup PARANOIA usage & code. <ldl@aros.net>
  39  * 04-02-19 Remove PARANOIA, plus various cleanups (Paul Clements)
  40  * possible FIXME: make set_sock / set_blksize / set_size / do_it one syscall
  41  * why not: would need access_ok and friends, would share yet another
  42  *          structure with userland
  43  */
  44
  45 #include <linux/major.h>
  46
  47 #include <linux/blkdev.h>
  48 #include <linux/module.h>
  49 #include <linux/init.h>
  50 #include <linux/sched.h>
  51 #include <linux/fs.h>
  52 #include <linux/bio.h>
  53 #include <linux/stat.h>
  54 #include <linux/errno.h>
  55 #include <linux/file.h>
  56 #include <linux/ioctl.h>
  57 #include <linux/compiler.h>
  58 #include <linux/err.h>
  59 #include <linux/kernel.h>
  60 #include <net/sock.h>
  61
  62 #include <linux/devfs_fs_kernel.h>
  63
  64 #include <asm/uaccess.h>
  65 #include <asm/system.h>
  66 #include <asm/types.h>
  67
  68 #include <linux/nbd.h>
  69
  70 #define LO_MAGIC 0x68797548
  71
  72 #ifdef NDEBUG
  73 #define dprintk(flags, fmt...)
  74 #else /* NDEBUG */
  75 #define dprintk(flags, fmt...) do { \
  76         if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
  77 } while (0)
  78 #define DBG_IOCTL       0x0004
  79 #define DBG_INIT        0x0010
  80 #define DBG_EXIT        0x0020
  81 #define DBG_BLKDEV      0x0100
  82 #define DBG_RX          0x0200
  83 #define DBG_TX          0x0400
  84 static unsigned int debugflags;
  85 static unsigned int nbds_max = 16;
  86 #endif /* NDEBUG */
  87
  88 static struct nbd_device nbd_dev[MAX_NBD];
  89
  90 /*
  91  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  92  * 1. Each NIC is essentially a synchronization point for all servers
  93  *    accessed through that NIC so there's no need to have more locks
  94  *    than NICs anyway.
  95  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  96  *    down each lock to the point where they're actually slower than just
  97  *    a single lock.
  98  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  99  */
 100 static DEFINE_SPINLOCK(nbd_lock);
 101
 102 #ifndef NDEBUG
 103 static const char *ioctl_cmd_to_ascii(int cmd)
 104 {
 105         switch (cmd) {
 106         case NBD_SET_SOCK: return "set-sock";
 107         case NBD_SET_BLKSIZE: return "set-blksize";
 108         case NBD_SET_SIZE: return "set-size";
 109         case NBD_DO_IT: return "do-it";
 110         case NBD_CLEAR_SOCK: return "clear-sock";
 111         case NBD_CLEAR_QUE: return "clear-que";
 112         case NBD_PRINT_DEBUG: return "print-debug";
 113         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
 114         case NBD_DISCONNECT: return "disconnect";
 115         case BLKROSET: return "set-read-only";
 116         case BLKFLSBUF: return "flush-buffer-cache";
 117         }
 118         return "unknown";
 119 }
 120
 121 static const char *nbdcmd_to_ascii(int cmd)
 122 {
 123         switch (cmd) {
 124         case  NBD_CMD_READ: return "read";
 125         case NBD_CMD_WRITE: return "write";
 126         case  NBD_CMD_DISC: return "disconnect";
 127         }
 128         return "invalid";
 129 }
 130 #endif /* NDEBUG */
 131
 132 static void nbd_end_request(struct request *req)
 133 {
 134         int uptodate = (req->errors == 0) ? 1 : 0;
 135         request_queue_t *q = req->q;
 136         unsigned long flags;
 137
 138         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 139                         req, uptodate? "done": "failed");
 140
 141         spin_lock_irqsave(q->queue_lock, flags);
 142         if (!end_that_request_first(req, uptodate, req->nr_sectors)) {
 143                 end_that_request_last(req, uptodate);
 144         }
 145         spin_unlock_irqrestore(q->queue_lock, flags);
 146 }
 147
 148 /*
 149  *  Send or receive packet.
 150  */
 151 static int sock_xmit(struct socket *sock, int send, void *buf, int size,
 152                 int msg_flags)
 153 {
 154         int result;
 155         struct msghdr msg;
 156         struct kvec iov;
 157         unsigned long flags;
 158         sigset_t oldset;
 159
 160         /* Allow interception of SIGKILL only
 161          * Don't allow other signals to interrupt the transmission */
 162         spin_lock_irqsave(&current->sighand->siglock, flags);
 163         oldset = current->blocked;
 164         sigfillset(&current->blocked);
 165         sigdelsetmask(&current->blocked, sigmask(SIGKILL));
 166         recalc_sigpending();
 167         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 168
 169         do {
 170                 sock->sk->sk_allocation = GFP_NOIO;
 171                 iov.iov_base = buf;
 172                 iov.iov_len = size;
 173                 msg.msg_name = NULL;
 174                 msg.msg_namelen = 0;
 175                 msg.msg_control = NULL;
 176                 msg.msg_controllen = 0;
 177                 msg.msg_namelen = 0;
 178                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 179
 180                 if (send)
 181                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 182                 else
 183                         result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
 184
 185                 if (signal_pending(current)) {
 186                         siginfo_t info;
 187                         spin_lock_irqsave(&current->sighand->siglock, flags);
 188                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 189                                 current->pid, current->comm,
 190                                 dequeue_signal(current, &current->blocked, &info));
 191                         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 192                         result = -EINTR;
 193                         break;
 194                 }
 195
 196                 if (result <= 0) {
 197                         if (result == 0)
 198                                 result = -EPIPE; /* short read */
 199                         break;
 200                 }
 201                 size -= result;
 202                 buf += result;
 203         } while (size > 0);
 204
 205         spin_lock_irqsave(&current->sighand->siglock, flags);
 206         current->blocked = oldset;
 207         recalc_sigpending();
 208         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 209
 210         return result;
 211 }
 212
 213 static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec,
 214                 int flags)
 215 {
 216         int result;
 217         void *kaddr = kmap(bvec->bv_page);
 218         result = sock_xmit(sock, 1, kaddr + bvec->bv_offset, bvec->bv_len,
 219                         flags);
 220         kunmap(bvec->bv_page);
 221         return result;
 222 }
 223
 224 static int nbd_send_req(struct nbd_device *lo, struct request *req)
 225 {
 226         int result, i, flags;
 227         struct nbd_request request;
 228         unsigned long size = req->nr_sectors << 9;
 229         struct socket *sock = lo->sock;
 230
 231         request.magic = htonl(NBD_REQUEST_MAGIC);
 232         request.type = htonl(nbd_cmd(req));
 233         request.from = cpu_to_be64((u64) req->sector << 9);
 234         request.len = htonl(size);
 235         memcpy(request.handle, &req, sizeof(req));
 236
 237         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
 238                         lo->disk->disk_name, req,
 239                         nbdcmd_to_ascii(nbd_cmd(req)),
 240                         (unsigned long long)req->sector << 9,
 241                         req->nr_sectors << 9);
 242         result = sock_xmit(sock, 1, &request, sizeof(request),
 243                         (nbd_cmd(req) == NBD_CMD_WRITE)? MSG_MORE: 0);
 244         if (result <= 0) {
 245                 printk(KERN_ERR "%s: Send control failed (result %d)\n",
 246                                 lo->disk->disk_name, result);
 247                 goto error_out;
 248         }
 249
 250         if (nbd_cmd(req) == NBD_CMD_WRITE) {
 251                 struct bio *bio;
 252                 /*
 253                  * we are really probing at internals to determine
 254                  * whether to set MSG_MORE or not...
 255                  */
 256                 rq_for_each_bio(bio, req) {
 257                         struct bio_vec *bvec;
 258                         bio_for_each_segment(bvec, bio, i) {
 259                                 flags = 0;
 260                                 if ((i < (bio->bi_vcnt - 1)) || bio->bi_next)
 261                                         flags = MSG_MORE;
 262                                 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 263                                                 lo->disk->disk_name, req,
 264                                                 bvec->bv_len);
 265                                 result = sock_send_bvec(sock, bvec, flags);
 266                                 if (result <= 0) {
 267                                         printk(KERN_ERR "%s: Send data failed (result %d)\n",
 268                                                         lo->disk->disk_name,
 269                                                         result);
 270                                         goto error_out;
 271                                 }
 272                         }
 273                 }
 274         }
 275         return 0;
 276
 277 error_out:
 278         return 1;
 279 }
 280
 281 static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
 282 {
 283         struct request *req;
 284         struct list_head *tmp;
 285         struct request *xreq;
 286         int err;
 287
 288         memcpy(&xreq, handle, sizeof(xreq));
 289
 290         err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
 291         if (unlikely(err))
 292                 goto out;
 293
 294         spin_lock(&lo->queue_lock);
 295         list_for_each(tmp, &lo->queue_head) {
 296                 req = list_entry(tmp, struct request, queuelist);
 297                 if (req != xreq)
 298                         continue;
 299                 list_del_init(&req->queuelist);
 300                 spin_unlock(&lo->queue_lock);
 301                 return req;
 302         }
 303         spin_unlock(&lo->queue_lock);
 304
 305         err = -ENOENT;
 306
 307 out:
 308         return ERR_PTR(err);
 309 }
 310
 311 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
 312 {
 313         int result;
 314         void *kaddr = kmap(bvec->bv_page);
 315         result = sock_xmit(sock, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 316                         MSG_WAITALL);
 317         kunmap(bvec->bv_page);
 318         return result;
 319 }
 320
 321 /* NULL returned = something went wrong, inform userspace */
 322 static struct request *nbd_read_stat(struct nbd_device *lo)
 323 {
 324         int result;
 325         struct nbd_reply reply;
 326         struct request *req;
 327         struct socket *sock = lo->sock;
 328
 329         reply.magic = 0;
 330         result = sock_xmit(sock, 0, &reply, sizeof(reply), MSG_WAITALL);
 331         if (result <= 0) {
 332                 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
 333                                 lo->disk->disk_name, result);
 334                 goto harderror;
 335         }
 336         req = nbd_find_request(lo, reply.handle);
 337         if (unlikely(IS_ERR(req))) {
 338                 result = PTR_ERR(req);
 339                 if (result != -ENOENT)
 340                         goto harderror;
 341
 342                 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
 343                                 lo->disk->disk_name, reply.handle);
 344                 result = -EBADR;
 345                 goto harderror;
 346         }
 347
 348         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 349                 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
 350                                 lo->disk->disk_name,
 351                                 (unsigned long)ntohl(reply.magic));
 352                 result = -EPROTO;
 353                 goto harderror;
 354         }
 355         if (ntohl(reply.error)) {
 356                 printk(KERN_ERR "%s: Other side returned error (%d)\n",
 357                                 lo->disk->disk_name, ntohl(reply.error));
 358                 req->errors++;
 359                 return req;
 360         }
 361
 362         dprintk(DBG_RX, "%s: request %p: got reply\n",
 363                         lo->disk->disk_name, req);
 364         if (nbd_cmd(req) == NBD_CMD_READ) {
 365                 int i;
 366                 struct bio *bio;
 367                 rq_for_each_bio(bio, req) {
 368                         struct bio_vec *bvec;
 369                         bio_for_each_segment(bvec, bio, i) {
 370                                 result = sock_recv_bvec(sock, bvec);
 371                                 if (result <= 0) {
 372                                         printk(KERN_ERR "%s: Receive data failed (result %d)\n",
 373                                                         lo->disk->disk_name,
 374                                                         result);
 375                                         goto harderror;
 376                                 }
 377                                 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 378                                         lo->disk->disk_name, req, bvec->bv_len);
 379                         }
 380                 }
 381         }
 382         return req;
 383 harderror:
 384         lo->harderror = result;
 385         return NULL;
 386 }
 387
 388 static void nbd_do_it(struct nbd_device *lo)
 389 {
 390         struct request *req;
 391
 392         BUG_ON(lo->magic != LO_MAGIC);
 393
 394         while ((req = nbd_read_stat(lo)) != NULL)
 395                 nbd_end_request(req);
 396         return;
 397 }
 398
 399 static void nbd_clear_que(struct nbd_device *lo)
 400 {
 401         struct request *req;
 402
 403         BUG_ON(lo->magic != LO_MAGIC);
 404
 405         /*
 406          * Because we have set lo->sock to NULL under the tx_lock, all
 407          * modifications to the list must have completed by now.  For
 408          * the same reason, the active_req must be NULL.
 409          *
 410          * As a consequence, we don't need to take the spin lock while
 411          * purging the list here.
 412          */
 413         BUG_ON(lo->sock);
 414         BUG_ON(lo->active_req);
 415
 416         while (!list_empty(&lo->queue_head)) {
 417                 req = list_entry(lo->queue_head.next, struct request,
 418                                  queuelist);
 419                 list_del_init(&req->queuelist);
 420                 req->errors++;
 421                 nbd_end_request(req);
 422         }
 423 }
 424
 425 /*
 426  * We always wait for result of write, for now. It would be nice to make it optional
 427  * in future
 428  * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
 429  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 430  */
 431
 432 static void do_nbd_request(request_queue_t * q)
 433 {
 434         struct request *req;
 435
 436         while ((req = elv_next_request(q)) != NULL) {
 437                 struct nbd_device *lo;
 438
 439                 blkdev_dequeue_request(req);
 440                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%lx)\n",
 441                                 req->rq_disk->disk_name, req, req->flags);
 442
 443                 if (!(req->flags & REQ_CMD))
 444                         goto error_out;
 445
 446                 lo = req->rq_disk->private_data;
 447
 448                 BUG_ON(lo->magic != LO_MAGIC);
 449
 450                 nbd_cmd(req) = NBD_CMD_READ;
 451                 if (rq_data_dir(req) == WRITE) {
 452                         nbd_cmd(req) = NBD_CMD_WRITE;
 453                         if (lo->flags & NBD_READ_ONLY) {
 454                                 printk(KERN_ERR "%s: Write on read-only\n",
 455                                                 lo->disk->disk_name);
 456                                 goto error_out;
 457                         }
 458                 }
 459
 460                 req->errors = 0;
 461                 spin_unlock_irq(q->queue_lock);
 462
 463                 down(&lo->tx_lock);
 464                 if (unlikely(!lo->sock)) {
 465                         up(&lo->tx_lock);
 466                         printk(KERN_ERR "%s: Attempted send on closed socket\n",
 467                                lo->disk->disk_name);
 468                         req->errors++;
 469                         nbd_end_request(req);
 470                         spin_lock_irq(q->queue_lock);
 471                         continue;
 472                 }
 473
 474                 lo->active_req = req;
 475
 476                 if (nbd_send_req(lo, req) != 0) {
 477                         printk(KERN_ERR "%s: Request send failed\n",
 478                                         lo->disk->disk_name);
 479                         req->errors++;
 480                         nbd_end_request(req);
 481                 } else {
 482                         spin_lock(&lo->queue_lock);
 483                         list_add(&req->queuelist, &lo->queue_head);
 484                         spin_unlock(&lo->queue_lock);
 485                 }
 486
 487                 lo->active_req = NULL;
 488                 up(&lo->tx_lock);
 489                 wake_up_all(&lo->active_wq);
 490
 491                 spin_lock_irq(q->queue_lock);
 492                 continue;
 493
 494 error_out:
 495                 req->errors++;
 496                 spin_unlock(q->queue_lock);
 497                 nbd_end_request(req);
 498                 spin_lock(q->queue_lock);
 499         }
 500         return;
 501 }
 502
 503 static int nbd_ioctl(struct inode *inode, struct file *file,
 504                      unsigned int cmd, unsigned long arg)
 505 {
 506         struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
 507         int error;
 508         struct request sreq ;
 509
 510         if (!capable(CAP_SYS_ADMIN))
 511                 return -EPERM;
 512
 513         BUG_ON(lo->magic != LO_MAGIC);
 514
 515         /* Anyone capable of this syscall can do *real bad* things */
 516         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 517                         lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 518
 519         switch (cmd) {
 520         case NBD_DISCONNECT:
 521                 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
 522                 sreq.flags = REQ_SPECIAL;
 523                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 524                 /*
 525                  * Set these to sane values in case server implementation
 526                  * fails to check the request type first and also to keep
 527                  * debugging output cleaner.
 528                  */
 529                 sreq.sector = 0;
 530                 sreq.nr_sectors = 0;
 531                 if (!lo->sock)
 532                         return -EINVAL;
 533                 nbd_send_req(lo, &sreq);
 534                 return 0;
 535
 536         case NBD_CLEAR_SOCK:
 537                 error = 0;
 538                 down(&lo->tx_lock);
 539                 lo->sock = NULL;
 540                 up(&lo->tx_lock);
 541                 file = lo->file;
 542                 lo->file = NULL;
 543                 nbd_clear_que(lo);
 544                 BUG_ON(!list_empty(&lo->queue_head));
 545                 if (file)
 546                         fput(file);
 547                 return error;
 548         case NBD_SET_SOCK:
 549                 if (lo->file)
 550                         return -EBUSY;
 551                 error = -EINVAL;
 552                 file = fget(arg);
 553                 if (file) {
 554                         inode = file->f_dentry->d_inode;
 555                         if (S_ISSOCK(inode->i_mode)) {
 556                                 lo->file = file;
 557                                 lo->sock = SOCKET_I(inode);
 558                                 error = 0;
 559                         } else {
 560                                 fput(file);
 561                         }
 562                 }
 563                 return error;
 564         case NBD_SET_BLKSIZE:
 565                 lo->blksize = arg;
 566                 lo->bytesize &= ~(lo->blksize-1);
 567                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 568                 set_blocksize(inode->i_bdev, lo->blksize);
 569                 set_capacity(lo->disk, lo->bytesize >> 9);
 570                 return 0;
 571         case NBD_SET_SIZE:
 572                 lo->bytesize = arg & ~(lo->blksize-1);
 573                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 574                 set_blocksize(inode->i_bdev, lo->blksize);
 575                 set_capacity(lo->disk, lo->bytesize >> 9);
 576                 return 0;
 577         case NBD_SET_SIZE_BLOCKS:
 578                 lo->bytesize = ((u64) arg) * lo->blksize;
 579                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 580                 set_blocksize(inode->i_bdev, lo->blksize);
 581                 set_capacity(lo->disk, lo->bytesize >> 9);
 582                 return 0;
 583         case NBD_DO_IT:
 584                 if (!lo->file)
 585                         return -EINVAL;
 586                 nbd_do_it(lo);
 587                 /* on return tidy up in case we have a signal */
 588                 /* Forcibly shutdown the socket causing all listeners
 589                  * to error
 590                  *
 591                  * FIXME: This code is duplicated from sys_shutdown, but
 592                  * there should be a more generic interface rather than
 593                  * calling socket ops directly here */
 594                 down(&lo->tx_lock);
 595                 if (lo->sock) {
 596                         printk(KERN_WARNING "%s: shutting down socket\n",
 597                                 lo->disk->disk_name);
 598                         lo->sock->ops->shutdown(lo->sock,
 599                                 SEND_SHUTDOWN|RCV_SHUTDOWN);
 600                         lo->sock = NULL;
 601                 }
 602                 up(&lo->tx_lock);
 603                 file = lo->file;
 604                 lo->file = NULL;
 605                 nbd_clear_que(lo);
 606                 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
 607                 if (file)
 608                         fput(file);
 609                 return lo->harderror;
 610         case NBD_CLEAR_QUE:
 611                 /*
 612                  * This is for compatibility only.  The queue is always cleared
 613                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 614                  */
 615                 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
 616                 return 0;
 617         case NBD_PRINT_DEBUG:
 618                 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
 619                         inode->i_bdev->bd_disk->disk_name,
 620                         lo->queue_head.next, lo->queue_head.prev,
 621                         &lo->queue_head);
 622                 return 0;
 623         }
 624         return -EINVAL;
 625 }
 626
 627 static struct block_device_operations nbd_fops =
 628 {
 629         .owner =        THIS_MODULE,
 630         .ioctl =        nbd_ioctl,
 631 };
 632
 633 /*
 634  * And here should be modules and kernel interface
 635  *  (Just smiley confuses emacs :-)
 636  */
 637
 638 static int __init nbd_init(void)
 639 {
 640         int err = -ENOMEM;
 641         int i;
 642
 643         if (sizeof(struct nbd_request) != 28) {
 644                 printk(KERN_CRIT "nbd: sizeof nbd_request needs to be 28 in order to work!\n" );
 645                 return -EIO;
 646         }
 647
 648         if (nbds_max > MAX_NBD) {
 649                 printk(KERN_CRIT "nbd: cannot allocate more than %u nbds; %u requested.\n", MAX_NBD,
 650                                 nbds_max);
 651                 return -EINVAL;
 652         }
 653
 654         for (i = 0; i < nbds_max; i++) {
 655                 struct gendisk *disk = alloc_disk(1);
 656                 if (!disk)
 657                         goto out;
 658                 nbd_dev[i].disk = disk;
 659                 /*
 660                  * The new linux 2.5 block layer implementation requires
 661                  * every gendisk to have its very own request_queue struct.
 662                  * These structs are big so we dynamically allocate them.
 663                  */
 664                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 665                 if (!disk->queue) {
 666                         put_disk(disk);
 667                         goto out;
 668                 }
 669         }
 670
 671         if (register_blkdev(NBD_MAJOR, "nbd")) {
 672                 err = -EIO;
 673                 goto out;
 674         }
 675
 676         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 677         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 678
 679         devfs_mk_dir("nbd");
 680         for (i = 0; i < nbds_max; i++) {
 681                 struct gendisk *disk = nbd_dev[i].disk;
 682                 nbd_dev[i].file = NULL;
 683                 nbd_dev[i].magic = LO_MAGIC;
 684                 nbd_dev[i].flags = 0;
 685                 spin_lock_init(&nbd_dev[i].queue_lock);
 686                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 687                 init_MUTEX(&nbd_dev[i].tx_lock);
 688                 init_waitqueue_head(&nbd_dev[i].active_wq);
 689                 nbd_dev[i].blksize = 1024;
 690                 nbd_dev[i].bytesize = 0x7ffffc00ULL << 10; /* 2TB */
 691                 disk->major = NBD_MAJOR;
 692                 disk->first_minor = i;
 693                 disk->fops = &nbd_fops;
 694                 disk->private_data = &nbd_dev[i];
 695                 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
 696                 sprintf(disk->disk_name, "nbd%d", i);
 697                 sprintf(disk->devfs_name, "nbd/%d", i);
 698                 set_capacity(disk, 0x7ffffc00ULL << 1); /* 2 TB */
 699                 add_disk(disk);
 700         }
 701
 702         return 0;
 703 out:
 704         while (i--) {
 705                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 706                 put_disk(nbd_dev[i].disk);
 707         }
 708         return err;
 709 }
 710
 711 static void __exit nbd_cleanup(void)
 712 {
 713         int i;
 714         for (i = 0; i < nbds_max; i++) {
 715                 struct gendisk *disk = nbd_dev[i].disk;
 716                 nbd_dev[i].magic = 0;
 717                 if (disk) {
 718                         del_gendisk(disk);
 719                         blk_cleanup_queue(disk->queue);
 720                         put_disk(disk);
 721                 }
 722         }
 723         devfs_remove("nbd");
 724         unregister_blkdev(NBD_MAJOR, "nbd");
 725         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 726 }
 727
 728 module_init(nbd_init);
 729 module_exit(nbd_cleanup);
 730
 731 MODULE_DESCRIPTION("Network Block Device");
 732 MODULE_LICENSE("GPL");
 733
 734 module_param(nbds_max, int, 0444);
 735 MODULE_PARM_DESC(nbds_max, "How many network block devices to initialize.");
 736 #ifndef NDEBUG
 737 module_param(debugflags, int, 0644);
 738 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 739 #endif