git.oblomov.eu Git - linux-2.6/blob - drivers/block/nbd.c

   1 /*
   2  * Network block device - make block devices work over TCP
   3  *
   4  * Note that you can not swap over this thing, yet. Seems to work but
   5  * deadlocks sometimes - you can not swap over TCP in general.
   6  *
   7  * Copyright 1997-2000, 2008 Pavel Machek <pavel@suse.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 static int max_part;
  60
  61 /*
  62  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  63  * 1. Each NIC is essentially a synchronization point for all servers
  64  *    accessed through that NIC so there's no need to have more locks
  65  *    than NICs anyway.
  66  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  67  *    down each lock to the point where they're actually slower than just
  68  *    a single lock.
  69  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  70  */
  71 static DEFINE_SPINLOCK(nbd_lock);
  72
  73 #ifndef NDEBUG
  74 static const char *ioctl_cmd_to_ascii(int cmd)
  75 {
  76         switch (cmd) {
  77         case NBD_SET_SOCK: return "set-sock";
  78         case NBD_SET_BLKSIZE: return "set-blksize";
  79         case NBD_SET_SIZE: return "set-size";
  80         case NBD_DO_IT: return "do-it";
  81         case NBD_CLEAR_SOCK: return "clear-sock";
  82         case NBD_CLEAR_QUE: return "clear-que";
  83         case NBD_PRINT_DEBUG: return "print-debug";
  84         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
  85         case NBD_DISCONNECT: return "disconnect";
  86         case BLKROSET: return "set-read-only";
  87         case BLKFLSBUF: return "flush-buffer-cache";
  88         }
  89         return "unknown";
  90 }
  91
  92 static const char *nbdcmd_to_ascii(int cmd)
  93 {
  94         switch (cmd) {
  95         case  NBD_CMD_READ: return "read";
  96         case NBD_CMD_WRITE: return "write";
  97         case  NBD_CMD_DISC: return "disconnect";
  98         }
  99         return "invalid";
 100 }
 101 #endif /* NDEBUG */
 102
 103 static void nbd_end_request(struct request *req)
 104 {
 105         int error = req->errors ? -EIO : 0;
 106         struct request_queue *q = req->q;
 107         unsigned long flags;
 108
 109         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 110                         req, error ? "failed" : "done");
 111
 112         spin_lock_irqsave(q->queue_lock, flags);
 113         __blk_end_request(req, error, req->nr_sectors << 9);
 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         if (unlikely(!sock)) {
 159                 printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
 160                        lo->disk->disk_name, (send ? "send" : "recv"));
 161                 return -EINVAL;
 162         }
 163
 164         /* Allow interception of SIGKILL only
 165          * Don't allow other signals to interrupt the transmission */
 166         siginitsetinv(&blocked, sigmask(SIGKILL));
 167         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 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_flags = msg_flags | MSG_NOSIGNAL;
 178
 179                 if (send) {
 180                         struct timer_list ti;
 181
 182                         if (lo->xmit_timeout) {
 183                                 init_timer(&ti);
 184                                 ti.function = nbd_xmit_timeout;
 185                                 ti.data = (unsigned long)current;
 186                                 ti.expires = jiffies + lo->xmit_timeout;
 187                                 add_timer(&ti);
 188                         }
 189                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 190                         if (lo->xmit_timeout)
 191                                 del_timer_sync(&ti);
 192                 } else
 193                         result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
 194
 195                 if (signal_pending(current)) {
 196                         siginfo_t info;
 197                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 198                                 task_pid_nr(current), current->comm,
 199                                 dequeue_signal_lock(current, &current->blocked, &info));
 200                         result = -EINTR;
 201                         sock_shutdown(lo, !send);
 202                         break;
 203                 }
 204
 205                 if (result <= 0) {
 206                         if (result == 0)
 207                                 result = -EPIPE; /* short read */
 208                         break;
 209                 }
 210                 size -= result;
 211                 buf += result;
 212         } while (size > 0);
 213
 214         sigprocmask(SIG_SETMASK, &oldset, NULL);
 215
 216         return result;
 217 }
 218
 219 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
 220                 int flags)
 221 {
 222         int result;
 223         void *kaddr = kmap(bvec->bv_page);
 224         result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
 225         kunmap(bvec->bv_page);
 226         return result;
 227 }
 228
 229 /* always call with the tx_lock held */
 230 static int nbd_send_req(struct nbd_device *lo, struct request *req)
 231 {
 232         int result, flags;
 233         struct nbd_request request;
 234         unsigned long size = req->nr_sectors << 9;
 235
 236         request.magic = htonl(NBD_REQUEST_MAGIC);
 237         request.type = htonl(nbd_cmd(req));
 238         request.from = cpu_to_be64((u64) req->sector << 9);
 239         request.len = htonl(size);
 240         memcpy(request.handle, &req, sizeof(req));
 241
 242         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
 243                         lo->disk->disk_name, req,
 244                         nbdcmd_to_ascii(nbd_cmd(req)),
 245                         (unsigned long long)req->sector << 9,
 246                         req->nr_sectors << 9);
 247         result = sock_xmit(lo, 1, &request, sizeof(request),
 248                         (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
 249         if (result <= 0) {
 250                 printk(KERN_ERR "%s: Send control failed (result %d)\n",
 251                                 lo->disk->disk_name, result);
 252                 goto error_out;
 253         }
 254
 255         if (nbd_cmd(req) == NBD_CMD_WRITE) {
 256                 struct req_iterator iter;
 257                 struct bio_vec *bvec;
 258                 /*
 259                  * we are really probing at internals to determine
 260                  * whether to set MSG_MORE or not...
 261                  */
 262                 rq_for_each_segment(bvec, req, iter) {
 263                         flags = 0;
 264                         if (!rq_iter_last(req, iter))
 265                                 flags = MSG_MORE;
 266                         dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 267                                         lo->disk->disk_name, req, bvec->bv_len);
 268                         result = sock_send_bvec(lo, bvec, flags);
 269                         if (result <= 0) {
 270                                 printk(KERN_ERR "%s: Send data failed (result %d)\n",
 271                                                 lo->disk->disk_name, result);
 272                                 goto error_out;
 273                         }
 274                 }
 275         }
 276         return 0;
 277
 278 error_out:
 279         return -EIO;
 280 }
 281
 282 static struct request *nbd_find_request(struct nbd_device *lo,
 283                                         struct request *xreq)
 284 {
 285         struct request *req, *tmp;
 286         int err;
 287
 288         err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
 289         if (unlikely(err))
 290                 goto out;
 291
 292         spin_lock(&lo->queue_lock);
 293         list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
 294                 if (req != xreq)
 295                         continue;
 296                 list_del_init(&req->queuelist);
 297                 spin_unlock(&lo->queue_lock);
 298                 return req;
 299         }
 300         spin_unlock(&lo->queue_lock);
 301
 302         err = -ENOENT;
 303
 304 out:
 305         return ERR_PTR(err);
 306 }
 307
 308 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
 309 {
 310         int result;
 311         void *kaddr = kmap(bvec->bv_page);
 312         result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 313                         MSG_WAITALL);
 314         kunmap(bvec->bv_page);
 315         return result;
 316 }
 317
 318 /* NULL returned = something went wrong, inform userspace */
 319 static struct request *nbd_read_stat(struct nbd_device *lo)
 320 {
 321         int result;
 322         struct nbd_reply reply;
 323         struct request *req;
 324
 325         reply.magic = 0;
 326         result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
 327         if (result <= 0) {
 328                 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
 329                                 lo->disk->disk_name, result);
 330                 goto harderror;
 331         }
 332
 333         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 334                 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
 335                                 lo->disk->disk_name,
 336                                 (unsigned long)ntohl(reply.magic));
 337                 result = -EPROTO;
 338                 goto harderror;
 339         }
 340
 341         req = nbd_find_request(lo, *(struct request **)reply.handle);
 342         if (IS_ERR(req)) {
 343                 result = PTR_ERR(req);
 344                 if (result != -ENOENT)
 345                         goto harderror;
 346
 347                 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
 348                                 lo->disk->disk_name, reply.handle);
 349                 result = -EBADR;
 350                 goto harderror;
 351         }
 352
 353         if (ntohl(reply.error)) {
 354                 printk(KERN_ERR "%s: Other side returned error (%d)\n",
 355                                 lo->disk->disk_name, ntohl(reply.error));
 356                 req->errors++;
 357                 return req;
 358         }
 359
 360         dprintk(DBG_RX, "%s: request %p: got reply\n",
 361                         lo->disk->disk_name, req);
 362         if (nbd_cmd(req) == NBD_CMD_READ) {
 363                 struct req_iterator iter;
 364                 struct bio_vec *bvec;
 365
 366                 rq_for_each_segment(bvec, req, iter) {
 367                         result = sock_recv_bvec(lo, bvec);
 368                         if (result <= 0) {
 369                                 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
 370                                                 lo->disk->disk_name, result);
 371                                 req->errors++;
 372                                 return req;
 373                         }
 374                         dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 375                                 lo->disk->disk_name, req, bvec->bv_len);
 376                 }
 377         }
 378         return req;
 379 harderror:
 380         lo->harderror = result;
 381         return NULL;
 382 }
 383
 384 static ssize_t pid_show(struct device *dev,
 385                         struct device_attribute *attr, char *buf)
 386 {
 387         struct gendisk *disk = dev_to_disk(dev);
 388
 389         return sprintf(buf, "%ld\n",
 390                 (long) ((struct nbd_device *)disk->private_data)->pid);
 391 }
 392
 393 static struct device_attribute pid_attr = {
 394         .attr = { .name = "pid", .mode = S_IRUGO},
 395         .show = pid_show,
 396 };
 397
 398 static int nbd_do_it(struct nbd_device *lo)
 399 {
 400         struct request *req;
 401         int ret;
 402
 403         BUG_ON(lo->magic != LO_MAGIC);
 404
 405         lo->pid = current->pid;
 406         ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
 407         if (ret) {
 408                 printk(KERN_ERR "nbd: sysfs_create_file failed!");
 409                 lo->pid = 0;
 410                 return ret;
 411         }
 412
 413         while ((req = nbd_read_stat(lo)) != NULL)
 414                 nbd_end_request(req);
 415
 416         sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
 417         lo->pid = 0;
 418         return 0;
 419 }
 420
 421 static void nbd_clear_que(struct nbd_device *lo)
 422 {
 423         struct request *req;
 424
 425         BUG_ON(lo->magic != LO_MAGIC);
 426
 427         /*
 428          * Because we have set lo->sock to NULL under the tx_lock, all
 429          * modifications to the list must have completed by now.  For
 430          * the same reason, the active_req must be NULL.
 431          *
 432          * As a consequence, we don't need to take the spin lock while
 433          * purging the list here.
 434          */
 435         BUG_ON(lo->sock);
 436         BUG_ON(lo->active_req);
 437
 438         while (!list_empty(&lo->queue_head)) {
 439                 req = list_entry(lo->queue_head.next, struct request,
 440                                  queuelist);
 441                 list_del_init(&req->queuelist);
 442                 req->errors++;
 443                 nbd_end_request(req);
 444         }
 445 }
 446
 447
 448 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
 449 {
 450         if (!blk_fs_request(req))
 451                 goto error_out;
 452
 453         nbd_cmd(req) = NBD_CMD_READ;
 454         if (rq_data_dir(req) == WRITE) {
 455                 nbd_cmd(req) = NBD_CMD_WRITE;
 456                 if (lo->flags & NBD_READ_ONLY) {
 457                         printk(KERN_ERR "%s: Write on read-only\n",
 458                                         lo->disk->disk_name);
 459                         goto error_out;
 460                 }
 461         }
 462
 463         req->errors = 0;
 464
 465         mutex_lock(&lo->tx_lock);
 466         if (unlikely(!lo->sock)) {
 467                 mutex_unlock(&lo->tx_lock);
 468                 printk(KERN_ERR "%s: Attempted send on closed socket\n",
 469                        lo->disk->disk_name);
 470                 goto error_out;
 471         }
 472
 473         lo->active_req = req;
 474
 475         if (nbd_send_req(lo, req) != 0) {
 476                 printk(KERN_ERR "%s: Request send failed\n",
 477                                 lo->disk->disk_name);
 478                 req->errors++;
 479                 nbd_end_request(req);
 480         } else {
 481                 spin_lock(&lo->queue_lock);
 482                 list_add(&req->queuelist, &lo->queue_head);
 483                 spin_unlock(&lo->queue_lock);
 484         }
 485
 486         lo->active_req = NULL;
 487         mutex_unlock(&lo->tx_lock);
 488         wake_up_all(&lo->active_wq);
 489
 490         return;
 491
 492 error_out:
 493         req->errors++;
 494         nbd_end_request(req);
 495 }
 496
 497 static int nbd_thread(void *data)
 498 {
 499         struct nbd_device *lo = data;
 500         struct request *req;
 501
 502         set_user_nice(current, -20);
 503         while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
 504                 /* wait for something to do */
 505                 wait_event_interruptible(lo->waiting_wq,
 506                                          kthread_should_stop() ||
 507                                          !list_empty(&lo->waiting_queue));
 508
 509                 /* extract request */
 510                 if (list_empty(&lo->waiting_queue))
 511                         continue;
 512
 513                 spin_lock_irq(&lo->queue_lock);
 514                 req = list_entry(lo->waiting_queue.next, struct request,
 515                                  queuelist);
 516                 list_del_init(&req->queuelist);
 517                 spin_unlock_irq(&lo->queue_lock);
 518
 519                 /* handle request */
 520                 nbd_handle_req(lo, req);
 521         }
 522         return 0;
 523 }
 524
 525 /*
 526  * We always wait for result of write, for now. It would be nice to make it optional
 527  * in future
 528  * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
 529  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 530  */
 531
 532 static void do_nbd_request(struct request_queue *q)
 533 {
 534         struct request *req;
 535
 536         while ((req = elv_next_request(q)) != NULL) {
 537                 struct nbd_device *lo;
 538
 539                 blkdev_dequeue_request(req);
 540
 541                 spin_unlock_irq(q->queue_lock);
 542
 543                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
 544                                 req->rq_disk->disk_name, req, req->cmd_type);
 545
 546                 lo = req->rq_disk->private_data;
 547
 548                 BUG_ON(lo->magic != LO_MAGIC);
 549
 550                 if (unlikely(!lo->sock)) {
 551                         printk(KERN_ERR "%s: Attempted send on closed socket\n",
 552                                 lo->disk->disk_name);
 553                         req->errors++;
 554                         nbd_end_request(req);
 555                         spin_lock_irq(q->queue_lock);
 556                         continue;
 557                 }
 558
 559                 spin_lock_irq(&lo->queue_lock);
 560                 list_add_tail(&req->queuelist, &lo->waiting_queue);
 561                 spin_unlock_irq(&lo->queue_lock);
 562
 563                 wake_up(&lo->waiting_wq);
 564
 565                 spin_lock_irq(q->queue_lock);
 566         }
 567 }
 568
 569 /* Must be called with tx_lock held */
 570
 571 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
 572                        unsigned int cmd, unsigned long arg)
 573 {
 574         switch (cmd) {
 575         case NBD_DISCONNECT: {
 576                 struct request sreq;
 577
 578                 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
 579
 580                 blk_rq_init(NULL, &sreq);
 581                 sreq.cmd_type = REQ_TYPE_SPECIAL;
 582                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 583                 /*
 584                  * Set these to sane values in case server implementation
 585                  * fails to check the request type first and also to keep
 586                  * debugging output cleaner.
 587                  */
 588                 sreq.sector = 0;
 589                 sreq.nr_sectors = 0;
 590                 if (!lo->sock)
 591                         return -EINVAL;
 592                 nbd_send_req(lo, &sreq);
 593                 return 0;
 594         }
 595
 596         case NBD_CLEAR_SOCK: {
 597                 struct file *file;
 598
 599                 lo->sock = NULL;
 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 0;
 607         }
 608
 609         case NBD_SET_SOCK: {
 610                 struct file *file;
 611                 if (lo->file)
 612                         return -EBUSY;
 613                 file = fget(arg);
 614                 if (file) {
 615                         struct inode *inode = file->f_path.dentry->d_inode;
 616                         if (S_ISSOCK(inode->i_mode)) {
 617                                 lo->file = file;
 618                                 lo->sock = SOCKET_I(inode);
 619                                 if (max_part > 0)
 620                                         bdev->bd_invalidated = 1;
 621                                 return 0;
 622                         } else {
 623                                 fput(file);
 624                         }
 625                 }
 626                 return -EINVAL;
 627         }
 628
 629         case NBD_SET_BLKSIZE:
 630                 lo->blksize = arg;
 631                 lo->bytesize &= ~(lo->blksize-1);
 632                 bdev->bd_inode->i_size = lo->bytesize;
 633                 set_blocksize(bdev, lo->blksize);
 634                 set_capacity(lo->disk, lo->bytesize >> 9);
 635                 return 0;
 636
 637         case NBD_SET_SIZE:
 638                 lo->bytesize = arg & ~(lo->blksize-1);
 639                 bdev->bd_inode->i_size = lo->bytesize;
 640                 set_blocksize(bdev, lo->blksize);
 641                 set_capacity(lo->disk, lo->bytesize >> 9);
 642                 return 0;
 643
 644         case NBD_SET_TIMEOUT:
 645                 lo->xmit_timeout = arg * HZ;
 646                 return 0;
 647
 648         case NBD_SET_SIZE_BLOCKS:
 649                 lo->bytesize = ((u64) arg) * lo->blksize;
 650                 bdev->bd_inode->i_size = lo->bytesize;
 651                 set_blocksize(bdev, lo->blksize);
 652                 set_capacity(lo->disk, lo->bytesize >> 9);
 653                 return 0;
 654
 655         case NBD_DO_IT: {
 656                 struct task_struct *thread;
 657                 struct file *file;
 658                 int error;
 659
 660                 if (lo->pid)
 661                         return -EBUSY;
 662                 if (!lo->file)
 663                         return -EINVAL;
 664
 665                 mutex_unlock(&lo->tx_lock);
 666
 667                 thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
 668                 if (IS_ERR(thread)) {
 669                         mutex_lock(&lo->tx_lock);
 670                         return PTR_ERR(thread);
 671                 }
 672                 wake_up_process(thread);
 673                 error = nbd_do_it(lo);
 674                 kthread_stop(thread);
 675
 676                 mutex_lock(&lo->tx_lock);
 677                 if (error)
 678                         return error;
 679                 sock_shutdown(lo, 0);
 680                 file = lo->file;
 681                 lo->file = NULL;
 682                 nbd_clear_que(lo);
 683                 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
 684                 if (file)
 685                         fput(file);
 686                 lo->bytesize = 0;
 687                 bdev->bd_inode->i_size = 0;
 688                 set_capacity(lo->disk, 0);
 689                 if (max_part > 0)
 690                         ioctl_by_bdev(bdev, BLKRRPART, 0);
 691                 return lo->harderror;
 692         }
 693
 694         case NBD_CLEAR_QUE:
 695                 /*
 696                  * This is for compatibility only.  The queue is always cleared
 697                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 698                  */
 699                 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
 700                 return 0;
 701
 702         case NBD_PRINT_DEBUG:
 703                 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
 704                         bdev->bd_disk->disk_name,
 705                         lo->queue_head.next, lo->queue_head.prev,
 706                         &lo->queue_head);
 707                 return 0;
 708         }
 709         return -ENOTTY;
 710 }
 711
 712 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 713                      unsigned int cmd, unsigned long arg)
 714 {
 715         struct nbd_device *lo = bdev->bd_disk->private_data;
 716         int error;
 717
 718         if (!capable(CAP_SYS_ADMIN))
 719                 return -EPERM;
 720
 721         BUG_ON(lo->magic != LO_MAGIC);
 722
 723         /* Anyone capable of this syscall can do *real bad* things */
 724         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 725                         lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 726
 727         mutex_lock(&lo->tx_lock);
 728         error = __nbd_ioctl(bdev, lo, cmd, arg);
 729         mutex_unlock(&lo->tx_lock);
 730
 731         return error;
 732 }
 733
 734 static struct block_device_operations nbd_fops =
 735 {
 736         .owner =        THIS_MODULE,
 737         .locked_ioctl = nbd_ioctl,
 738 };
 739
 740 /*
 741  * And here should be modules and kernel interface
 742  *  (Just smiley confuses emacs :-)
 743  */
 744
 745 static int __init nbd_init(void)
 746 {
 747         int err = -ENOMEM;
 748         int i;
 749         int part_shift;
 750
 751         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 752
 753         if (max_part < 0) {
 754                 printk(KERN_CRIT "nbd: max_part must be >= 0\n");
 755                 return -EINVAL;
 756         }
 757
 758         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
 759         if (!nbd_dev)
 760                 return -ENOMEM;
 761
 762         part_shift = 0;
 763         if (max_part > 0)
 764                 part_shift = fls(max_part);
 765
 766         for (i = 0; i < nbds_max; i++) {
 767                 struct gendisk *disk = alloc_disk(1 << part_shift);
 768                 if (!disk)
 769                         goto out;
 770                 nbd_dev[i].disk = disk;
 771                 /*
 772                  * The new linux 2.5 block layer implementation requires
 773                  * every gendisk to have its very own request_queue struct.
 774                  * These structs are big so we dynamically allocate them.
 775                  */
 776                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 777                 if (!disk->queue) {
 778                         put_disk(disk);
 779                         goto out;
 780                 }
 781                 /*
 782                  * Tell the block layer that we are not a rotational device
 783                  */
 784                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
 785         }
 786
 787         if (register_blkdev(NBD_MAJOR, "nbd")) {
 788                 err = -EIO;
 789                 goto out;
 790         }
 791
 792         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 793         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 794
 795         for (i = 0; i < nbds_max; i++) {
 796                 struct gendisk *disk = nbd_dev[i].disk;
 797                 nbd_dev[i].file = NULL;
 798                 nbd_dev[i].magic = LO_MAGIC;
 799                 nbd_dev[i].flags = 0;
 800                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
 801                 spin_lock_init(&nbd_dev[i].queue_lock);
 802                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 803                 mutex_init(&nbd_dev[i].tx_lock);
 804                 init_waitqueue_head(&nbd_dev[i].active_wq);
 805                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
 806                 nbd_dev[i].blksize = 1024;
 807                 nbd_dev[i].bytesize = 0;
 808                 disk->major = NBD_MAJOR;
 809                 disk->first_minor = i << part_shift;
 810                 disk->fops = &nbd_fops;
 811                 disk->private_data = &nbd_dev[i];
 812                 sprintf(disk->disk_name, "nbd%d", i);
 813                 set_capacity(disk, 0);
 814                 add_disk(disk);
 815         }
 816
 817         return 0;
 818 out:
 819         while (i--) {
 820                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 821                 put_disk(nbd_dev[i].disk);
 822         }
 823         kfree(nbd_dev);
 824         return err;
 825 }
 826
 827 static void __exit nbd_cleanup(void)
 828 {
 829         int i;
 830         for (i = 0; i < nbds_max; i++) {
 831                 struct gendisk *disk = nbd_dev[i].disk;
 832                 nbd_dev[i].magic = 0;
 833                 if (disk) {
 834                         del_gendisk(disk);
 835                         blk_cleanup_queue(disk->queue);
 836                         put_disk(disk);
 837                 }
 838         }
 839         unregister_blkdev(NBD_MAJOR, "nbd");
 840         kfree(nbd_dev);
 841         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 842 }
 843
 844 module_init(nbd_init);
 845 module_exit(nbd_cleanup);
 846
 847 MODULE_DESCRIPTION("Network Block Device");
 848 MODULE_LICENSE("GPL");
 849
 850 module_param(nbds_max, int, 0444);
 851 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
 852 module_param(max_part, int, 0444);
 853 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
 854 #ifndef NDEBUG
 855 module_param(debugflags, int, 0644);
 856 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 857 #endif