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