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