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