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@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/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/err.h>
  30 #include <linux/kernel.h>
  31 #include <linux/slab.h>
  32 #include <net/sock.h>
  33 #include <linux/net.h>
  34 #include <linux/kthread.h>
  35
  36 #include <asm/uaccess.h>
  37 #include <asm/system.h>
  38 #include <asm/types.h>
  39
  40 #include <linux/nbd.h>
  41
  42 #define LO_MAGIC 0x68797548
  43
  44 #ifdef NDEBUG
  45 #define dprintk(flags, fmt...)
  46 #else /* NDEBUG */
  47 #define dprintk(flags, fmt...) do { \
  48         if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
  49 } while (0)
  50 #define DBG_IOCTL       0x0004
  51 #define DBG_INIT        0x0010
  52 #define DBG_EXIT        0x0020
  53 #define DBG_BLKDEV      0x0100
  54 #define DBG_RX          0x0200
  55 #define DBG_TX          0x0400
  56 static DEFINE_MUTEX(nbd_mutex);
  57 static unsigned int debugflags;
  58 #endif /* NDEBUG */
  59
  60 static unsigned int nbds_max = 16;
  61 static struct nbd_device *nbd_dev;
  62 static int max_part;
  63
  64 /*
  65  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  66  * 1. Each NIC is essentially a synchronization point for all servers
  67  *    accessed through that NIC so there's no need to have more locks
  68  *    than NICs anyway.
  69  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  70  *    down each lock to the point where they're actually slower than just
  71  *    a single lock.
  72  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  73  */
  74 static DEFINE_SPINLOCK(nbd_lock);
  75
  76 #ifndef NDEBUG
  77 static const char *ioctl_cmd_to_ascii(int cmd)
  78 {
  79         switch (cmd) {
  80         case NBD_SET_SOCK: return "set-sock";
  81         case NBD_SET_BLKSIZE: return "set-blksize";
  82         case NBD_SET_SIZE: return "set-size";
  83         case NBD_DO_IT: return "do-it";
  84         case NBD_CLEAR_SOCK: return "clear-sock";
  85         case NBD_CLEAR_QUE: return "clear-que";
  86         case NBD_PRINT_DEBUG: return "print-debug";
  87         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
  88         case NBD_DISCONNECT: return "disconnect";
  89         case BLKROSET: return "set-read-only";
  90         case BLKFLSBUF: return "flush-buffer-cache";
  91         }
  92         return "unknown";
  93 }
  94
  95 static const char *nbdcmd_to_ascii(int cmd)
  96 {
  97         switch (cmd) {
  98         case  NBD_CMD_READ: return "read";
  99         case NBD_CMD_WRITE: return "write";
 100         case  NBD_CMD_DISC: return "disconnect";
 101         }
 102         return "invalid";
 103 }
 104 #endif /* NDEBUG */
 105
 106 static void nbd_end_request(struct request *req)
 107 {
 108         int error = req->errors ? -EIO : 0;
 109         struct request_queue *q = req->q;
 110         unsigned long flags;
 111
 112         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 113                         req, error ? "failed" : "done");
 114
 115         spin_lock_irqsave(q->queue_lock, flags);
 116         __blk_end_request_all(req, error);
 117         spin_unlock_irqrestore(q->queue_lock, flags);
 118 }
 119
 120 static void sock_shutdown(struct nbd_device *lo, int lock)
 121 {
 122         /* Forcibly shutdown the socket causing all listeners
 123          * to error
 124          *
 125          * FIXME: This code is duplicated from sys_shutdown, but
 126          * there should be a more generic interface rather than
 127          * calling socket ops directly here */
 128         if (lock)
 129                 mutex_lock(&lo->tx_lock);
 130         if (lo->sock) {
 131                 printk(KERN_WARNING "%s: shutting down socket\n",
 132                         lo->disk->disk_name);
 133                 kernel_sock_shutdown(lo->sock, SHUT_RDWR);
 134                 lo->sock = NULL;
 135         }
 136         if (lock)
 137                 mutex_unlock(&lo->tx_lock);
 138 }
 139
 140 static void nbd_xmit_timeout(unsigned long arg)
 141 {
 142         struct task_struct *task = (struct task_struct *)arg;
 143
 144         printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
 145                 task->comm, task->pid);
 146         force_sig(SIGKILL, task);
 147 }
 148
 149 /*
 150  *  Send or receive packet.
 151  */
 152 static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
 153                 int msg_flags)
 154 {
 155         struct socket *sock = lo->sock;
 156         int result;
 157         struct msghdr msg;
 158         struct kvec iov;
 159         sigset_t blocked, oldset;
 160
 161         if (unlikely(!sock)) {
 162                 printk(KERN_ERR "%s: Attempted %s on closed socket in sock_xmit\n",
 163                        lo->disk->disk_name, (send ? "send" : "recv"));
 164                 return -EINVAL;
 165         }
 166
 167         /* Allow interception of SIGKILL only
 168          * Don't allow other signals to interrupt the transmission */
 169         siginitsetinv(&blocked, sigmask(SIGKILL));
 170         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 171
 172         do {
 173                 sock->sk->sk_allocation = GFP_NOIO;
 174                 iov.iov_base = buf;
 175                 iov.iov_len = size;
 176                 msg.msg_name = NULL;
 177                 msg.msg_namelen = 0;
 178                 msg.msg_control = NULL;
 179                 msg.msg_controllen = 0;
 180                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 181
 182                 if (send) {
 183                         struct timer_list ti;
 184
 185                         if (lo->xmit_timeout) {
 186                                 init_timer(&ti);
 187                                 ti.function = nbd_xmit_timeout;
 188                                 ti.data = (unsigned long)current;
 189                                 ti.expires = jiffies + lo->xmit_timeout;
 190                                 add_timer(&ti);
 191                         }
 192                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 193                         if (lo->xmit_timeout)
 194                                 del_timer_sync(&ti);
 195                 } else
 196                         result = kernel_recvmsg(sock, &msg, &iov, 1, size, 0);
 197
 198                 if (signal_pending(current)) {
 199                         siginfo_t info;
 200                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 201                                 task_pid_nr(current), current->comm,
 202                                 dequeue_signal_lock(current, &current->blocked, &info));
 203                         result = -EINTR;
 204                         sock_shutdown(lo, !send);
 205                         break;
 206                 }
 207
 208                 if (result <= 0) {
 209                         if (result == 0)
 210                                 result = -EPIPE; /* short read */
 211                         break;
 212                 }
 213                 size -= result;
 214                 buf += result;
 215         } while (size > 0);
 216
 217         sigprocmask(SIG_SETMASK, &oldset, NULL);
 218
 219         return result;
 220 }
 221
 222 static inline int sock_send_bvec(struct nbd_device *lo, struct bio_vec *bvec,
 223                 int flags)
 224 {
 225         int result;
 226         void *kaddr = kmap(bvec->bv_page);
 227         result = sock_xmit(lo, 1, kaddr + bvec->bv_offset, bvec->bv_len, flags);
 228         kunmap(bvec->bv_page);
 229         return result;
 230 }
 231
 232 /* always call with the tx_lock held */
 233 static int nbd_send_req(struct nbd_device *lo, struct request *req)
 234 {
 235         int result, flags;
 236         struct nbd_request request;
 237         unsigned long size = blk_rq_bytes(req);
 238
 239         request.magic = htonl(NBD_REQUEST_MAGIC);
 240         request.type = htonl(nbd_cmd(req));
 241         request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 242         request.len = htonl(size);
 243         memcpy(request.handle, &req, sizeof(req));
 244
 245         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
 246                         lo->disk->disk_name, req,
 247                         nbdcmd_to_ascii(nbd_cmd(req)),
 248                         (unsigned long long)blk_rq_pos(req) << 9,
 249                         blk_rq_bytes(req));
 250         result = sock_xmit(lo, 1, &request, sizeof(request),
 251                         (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
 252         if (result <= 0) {
 253                 printk(KERN_ERR "%s: Send control failed (result %d)\n",
 254                                 lo->disk->disk_name, result);
 255                 goto error_out;
 256         }
 257
 258         if (nbd_cmd(req) == NBD_CMD_WRITE) {
 259                 struct req_iterator iter;
 260                 struct bio_vec *bvec;
 261                 /*
 262                  * we are really probing at internals to determine
 263                  * whether to set MSG_MORE or not...
 264                  */
 265                 rq_for_each_segment(bvec, req, iter) {
 266                         flags = 0;
 267                         if (!rq_iter_last(req, iter))
 268                                 flags = MSG_MORE;
 269                         dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 270                                         lo->disk->disk_name, req, bvec->bv_len);
 271                         result = sock_send_bvec(lo, bvec, flags);
 272                         if (result <= 0) {
 273                                 printk(KERN_ERR "%s: Send data failed (result %d)\n",
 274                                                 lo->disk->disk_name, result);
 275                                 goto error_out;
 276                         }
 277                 }
 278         }
 279         return 0;
 280
 281 error_out:
 282         return -EIO;
 283 }
 284
 285 static struct request *nbd_find_request(struct nbd_device *lo,
 286                                         struct request *xreq)
 287 {
 288         struct request *req, *tmp;
 289         int err;
 290
 291         err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
 292         if (unlikely(err))
 293                 goto out;
 294
 295         spin_lock(&lo->queue_lock);
 296         list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
 297                 if (req != xreq)
 298                         continue;
 299                 list_del_init(&req->queuelist);
 300                 spin_unlock(&lo->queue_lock);
 301                 return req;
 302         }
 303         spin_unlock(&lo->queue_lock);
 304
 305         err = -ENOENT;
 306
 307 out:
 308         return ERR_PTR(err);
 309 }
 310
 311 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
 312 {
 313         int result;
 314         void *kaddr = kmap(bvec->bv_page);
 315         result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 316                         MSG_WAITALL);
 317         kunmap(bvec->bv_page);
 318         return result;
 319 }
 320
 321 /* NULL returned = something went wrong, inform userspace */
 322 static struct request *nbd_read_stat(struct nbd_device *lo)
 323 {
 324         int result;
 325         struct nbd_reply reply;
 326         struct request *req;
 327
 328         reply.magic = 0;
 329         result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
 330         if (result <= 0) {
 331                 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
 332                                 lo->disk->disk_name, result);
 333                 goto harderror;
 334         }
 335
 336         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 337                 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
 338                                 lo->disk->disk_name,
 339                                 (unsigned long)ntohl(reply.magic));
 340                 result = -EPROTO;
 341                 goto harderror;
 342         }
 343
 344         req = nbd_find_request(lo, *(struct request **)reply.handle);
 345         if (IS_ERR(req)) {
 346                 result = PTR_ERR(req);
 347                 if (result != -ENOENT)
 348                         goto harderror;
 349
 350                 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
 351                                 lo->disk->disk_name, reply.handle);
 352                 result = -EBADR;
 353                 goto harderror;
 354         }
 355
 356         if (ntohl(reply.error)) {
 357                 printk(KERN_ERR "%s: Other side returned error (%d)\n",
 358                                 lo->disk->disk_name, ntohl(reply.error));
 359                 req->errors++;
 360                 return req;
 361         }
 362
 363         dprintk(DBG_RX, "%s: request %p: got reply\n",
 364                         lo->disk->disk_name, req);
 365         if (nbd_cmd(req) == NBD_CMD_READ) {
 366                 struct req_iterator iter;
 367                 struct bio_vec *bvec;
 368
 369                 rq_for_each_segment(bvec, req, iter) {
 370                         result = sock_recv_bvec(lo, bvec);
 371                         if (result <= 0) {
 372                                 printk(KERN_ERR "%s: Receive data failed (result %d)\n",
 373                                                 lo->disk->disk_name, result);
 374                                 req->errors++;
 375                                 return req;
 376                         }
 377                         dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 378                                 lo->disk->disk_name, req, bvec->bv_len);
 379                 }
 380         }
 381         return req;
 382 harderror:
 383         lo->harderror = result;
 384         return NULL;
 385 }
 386
 387 static ssize_t pid_show(struct device *dev,
 388                         struct device_attribute *attr, char *buf)
 389 {
 390         struct gendisk *disk = dev_to_disk(dev);
 391
 392         return sprintf(buf, "%ld\n",
 393                 (long) ((struct nbd_device *)disk->private_data)->pid);
 394 }
 395
 396 static struct device_attribute pid_attr = {
 397         .attr = { .name = "pid", .mode = S_IRUGO},
 398         .show = pid_show,
 399 };
 400
 401 static int nbd_do_it(struct nbd_device *lo)
 402 {
 403         struct request *req;
 404         int ret;
 405
 406         BUG_ON(lo->magic != LO_MAGIC);
 407
 408         lo->pid = current->pid;
 409         ret = sysfs_create_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
 410         if (ret) {
 411                 printk(KERN_ERR "nbd: sysfs_create_file failed!");
 412                 lo->pid = 0;
 413                 return ret;
 414         }
 415
 416         while ((req = nbd_read_stat(lo)) != NULL)
 417                 nbd_end_request(req);
 418
 419         sysfs_remove_file(&disk_to_dev(lo->disk)->kobj, &pid_attr.attr);
 420         lo->pid = 0;
 421         return 0;
 422 }
 423
 424 static void nbd_clear_que(struct nbd_device *lo)
 425 {
 426         struct request *req;
 427
 428         BUG_ON(lo->magic != LO_MAGIC);
 429
 430         /*
 431          * Because we have set lo->sock to NULL under the tx_lock, all
 432          * modifications to the list must have completed by now.  For
 433          * the same reason, the active_req must be NULL.
 434          *
 435          * As a consequence, we don't need to take the spin lock while
 436          * purging the list here.
 437          */
 438         BUG_ON(lo->sock);
 439         BUG_ON(lo->active_req);
 440
 441         while (!list_empty(&lo->queue_head)) {
 442                 req = list_entry(lo->queue_head.next, struct request,
 443                                  queuelist);
 444                 list_del_init(&req->queuelist);
 445                 req->errors++;
 446                 nbd_end_request(req);
 447         }
 448 }
 449
 450
 451 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
 452 {
 453         if (req->cmd_type != REQ_TYPE_FS)
 454                 goto error_out;
 455
 456         nbd_cmd(req) = NBD_CMD_READ;
 457         if (rq_data_dir(req) == WRITE) {
 458                 nbd_cmd(req) = NBD_CMD_WRITE;
 459                 if (lo->flags & NBD_READ_ONLY) {
 460                         printk(KERN_ERR "%s: Write on read-only\n",
 461                                         lo->disk->disk_name);
 462                         goto error_out;
 463                 }
 464         }
 465
 466         req->errors = 0;
 467
 468         mutex_lock(&lo->tx_lock);
 469         if (unlikely(!lo->sock)) {
 470                 mutex_unlock(&lo->tx_lock);
 471                 printk(KERN_ERR "%s: Attempted send on closed socket\n",
 472                        lo->disk->disk_name);
 473                 goto error_out;
 474         }
 475
 476         lo->active_req = req;
 477
 478         if (nbd_send_req(lo, req) != 0) {
 479                 printk(KERN_ERR "%s: Request send failed\n",
 480                                 lo->disk->disk_name);
 481                 req->errors++;
 482                 nbd_end_request(req);
 483         } else {
 484                 spin_lock(&lo->queue_lock);
 485                 list_add(&req->queuelist, &lo->queue_head);
 486                 spin_unlock(&lo->queue_lock);
 487         }
 488
 489         lo->active_req = NULL;
 490         mutex_unlock(&lo->tx_lock);
 491         wake_up_all(&lo->active_wq);
 492
 493         return;
 494
 495 error_out:
 496         req->errors++;
 497         nbd_end_request(req);
 498 }
 499
 500 static int nbd_thread(void *data)
 501 {
 502         struct nbd_device *lo = data;
 503         struct request *req;
 504
 505         set_user_nice(current, -20);
 506         while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
 507                 /* wait for something to do */
 508                 wait_event_interruptible(lo->waiting_wq,
 509                                          kthread_should_stop() ||
 510                                          !list_empty(&lo->waiting_queue));
 511
 512                 /* extract request */
 513                 if (list_empty(&lo->waiting_queue))
 514                         continue;
 515
 516                 spin_lock_irq(&lo->queue_lock);
 517                 req = list_entry(lo->waiting_queue.next, struct request,
 518                                  queuelist);
 519                 list_del_init(&req->queuelist);
 520                 spin_unlock_irq(&lo->queue_lock);
 521
 522                 /* handle request */
 523                 nbd_handle_req(lo, req);
 524         }
 525         return 0;
 526 }
 527
 528 /*
 529  * We always wait for result of write, for now. It would be nice to make it optional
 530  * in future
 531  * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
 532  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 533  */
 534
 535 static void do_nbd_request(struct request_queue *q)
 536 {
 537         struct request *req;
 538
 539         while ((req = blk_fetch_request(q)) != NULL) {
 540                 struct nbd_device *lo;
 541
 542                 spin_unlock_irq(q->queue_lock);
 543
 544                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
 545                                 req->rq_disk->disk_name, req, req->cmd_type);
 546
 547                 lo = req->rq_disk->private_data;
 548
 549                 BUG_ON(lo->magic != LO_MAGIC);
 550
 551                 if (unlikely(!lo->sock)) {
 552                         printk(KERN_ERR "%s: Attempted send on closed socket\n",
 553                                 lo->disk->disk_name);
 554                         req->errors++;
 555                         nbd_end_request(req);
 556                         spin_lock_irq(q->queue_lock);
 557                         continue;
 558                 }
 559
 560                 spin_lock_irq(&lo->queue_lock);
 561                 list_add_tail(&req->queuelist, &lo->waiting_queue);
 562                 spin_unlock_irq(&lo->queue_lock);
 563
 564                 wake_up(&lo->waiting_wq);
 565
 566                 spin_lock_irq(q->queue_lock);
 567         }
 568 }
 569
 570 /* Must be called with tx_lock held */
 571
 572 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
 573                        unsigned int cmd, unsigned long arg)
 574 {
 575         switch (cmd) {
 576         case NBD_DISCONNECT: {
 577                 struct request sreq;
 578
 579                 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
 580
 581                 blk_rq_init(NULL, &sreq);
 582                 sreq.cmd_type = REQ_TYPE_SPECIAL;
 583                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 584                 if (!lo->sock)
 585                         return -EINVAL;
 586                 nbd_send_req(lo, &sreq);
 587                 return 0;
 588         }
 589
 590         case NBD_CLEAR_SOCK: {
 591                 struct file *file;
 592
 593                 lo->sock = NULL;
 594                 file = lo->file;
 595                 lo->file = NULL;
 596                 nbd_clear_que(lo);
 597                 BUG_ON(!list_empty(&lo->queue_head));
 598                 if (file)
 599                         fput(file);
 600                 return 0;
 601         }
 602
 603         case NBD_SET_SOCK: {
 604                 struct file *file;
 605                 if (lo->file)
 606                         return -EBUSY;
 607                 file = fget(arg);
 608                 if (file) {
 609                         struct inode *inode = file->f_path.dentry->d_inode;
 610                         if (S_ISSOCK(inode->i_mode)) {
 611                                 lo->file = file;
 612                                 lo->sock = SOCKET_I(inode);
 613                                 if (max_part > 0)
 614                                         bdev->bd_invalidated = 1;
 615                                 return 0;
 616                         } else {
 617                                 fput(file);
 618                         }
 619                 }
 620                 return -EINVAL;
 621         }
 622
 623         case NBD_SET_BLKSIZE:
 624                 lo->blksize = arg;
 625                 lo->bytesize &= ~(lo->blksize-1);
 626                 bdev->bd_inode->i_size = lo->bytesize;
 627                 set_blocksize(bdev, lo->blksize);
 628                 set_capacity(lo->disk, lo->bytesize >> 9);
 629                 return 0;
 630
 631         case NBD_SET_SIZE:
 632                 lo->bytesize = arg & ~(lo->blksize-1);
 633                 bdev->bd_inode->i_size = lo->bytesize;
 634                 set_blocksize(bdev, lo->blksize);
 635                 set_capacity(lo->disk, lo->bytesize >> 9);
 636                 return 0;
 637
 638         case NBD_SET_TIMEOUT:
 639                 lo->xmit_timeout = arg * HZ;
 640                 return 0;
 641
 642         case NBD_SET_SIZE_BLOCKS:
 643                 lo->bytesize = ((u64) arg) * lo->blksize;
 644                 bdev->bd_inode->i_size = lo->bytesize;
 645                 set_blocksize(bdev, lo->blksize);
 646                 set_capacity(lo->disk, lo->bytesize >> 9);
 647                 return 0;
 648
 649         case NBD_DO_IT: {
 650                 struct task_struct *thread;
 651                 struct file *file;
 652                 int error;
 653
 654                 if (lo->pid)
 655                         return -EBUSY;
 656                 if (!lo->file)
 657                         return -EINVAL;
 658
 659                 mutex_unlock(&lo->tx_lock);
 660
 661                 thread = kthread_create(nbd_thread, lo, lo->disk->disk_name);
 662                 if (IS_ERR(thread)) {
 663                         mutex_lock(&lo->tx_lock);
 664                         return PTR_ERR(thread);
 665                 }
 666                 wake_up_process(thread);
 667                 error = nbd_do_it(lo);
 668                 kthread_stop(thread);
 669
 670                 mutex_lock(&lo->tx_lock);
 671                 if (error)
 672                         return error;
 673                 sock_shutdown(lo, 0);
 674                 file = lo->file;
 675                 lo->file = NULL;
 676                 nbd_clear_que(lo);
 677                 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
 678                 if (file)
 679                         fput(file);
 680                 lo->bytesize = 0;
 681                 bdev->bd_inode->i_size = 0;
 682                 set_capacity(lo->disk, 0);
 683                 if (max_part > 0)
 684                         ioctl_by_bdev(bdev, BLKRRPART, 0);
 685                 return lo->harderror;
 686         }
 687
 688         case NBD_CLEAR_QUE:
 689                 /*
 690                  * This is for compatibility only.  The queue is always cleared
 691                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 692                  */
 693                 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
 694                 return 0;
 695
 696         case NBD_PRINT_DEBUG:
 697                 printk(KERN_INFO "%s: next = %p, prev = %p, head = %p\n",
 698                         bdev->bd_disk->disk_name,
 699                         lo->queue_head.next, lo->queue_head.prev,
 700                         &lo->queue_head);
 701                 return 0;
 702         }
 703         return -ENOTTY;
 704 }
 705
 706 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 707                      unsigned int cmd, unsigned long arg)
 708 {
 709         struct nbd_device *lo = bdev->bd_disk->private_data;
 710         int error;
 711
 712         if (!capable(CAP_SYS_ADMIN))
 713                 return -EPERM;
 714
 715         BUG_ON(lo->magic != LO_MAGIC);
 716
 717         /* Anyone capable of this syscall can do *real bad* things */
 718         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 719                         lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 720
 721         mutex_lock(&nbd_mutex);
 722         mutex_lock(&lo->tx_lock);
 723         error = __nbd_ioctl(bdev, lo, cmd, arg);
 724         mutex_unlock(&lo->tx_lock);
 725         mutex_unlock(&nbd_mutex);
 726
 727         return error;
 728 }
 729
 730 static const struct block_device_operations nbd_fops =
 731 {
 732         .owner =        THIS_MODULE,
 733         .ioctl =        nbd_ioctl,
 734 };
 735
 736 /*
 737  * And here should be modules and kernel interface
 738  *  (Just smiley confuses emacs :-)
 739  */
 740
 741 static int __init nbd_init(void)
 742 {
 743         int err = -ENOMEM;
 744         int i;
 745         int part_shift;
 746
 747         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 748
 749         if (max_part < 0) {
 750                 printk(KERN_CRIT "nbd: max_part must be >= 0\n");
 751                 return -EINVAL;
 752         }
 753
 754         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
 755         if (!nbd_dev)
 756                 return -ENOMEM;
 757
 758         part_shift = 0;
 759         if (max_part > 0)
 760                 part_shift = fls(max_part);
 761
 762         for (i = 0; i < nbds_max; i++) {
 763                 struct gendisk *disk = alloc_disk(1 << part_shift);
 764                 if (!disk)
 765                         goto out;
 766                 nbd_dev[i].disk = disk;
 767                 /*
 768                  * The new linux 2.5 block layer implementation requires
 769                  * every gendisk to have its very own request_queue struct.
 770                  * These structs are big so we dynamically allocate them.
 771                  */
 772                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 773                 if (!disk->queue) {
 774                         put_disk(disk);
 775                         goto out;
 776                 }
 777                 /*
 778                  * Tell the block layer that we are not a rotational device
 779                  */
 780                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
 781         }
 782
 783         if (register_blkdev(NBD_MAJOR, "nbd")) {
 784                 err = -EIO;
 785                 goto out;
 786         }
 787
 788         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 789         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 790
 791         for (i = 0; i < nbds_max; i++) {
 792                 struct gendisk *disk = nbd_dev[i].disk;
 793                 nbd_dev[i].file = NULL;
 794                 nbd_dev[i].magic = LO_MAGIC;
 795                 nbd_dev[i].flags = 0;
 796                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
 797                 spin_lock_init(&nbd_dev[i].queue_lock);
 798                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 799                 mutex_init(&nbd_dev[i].tx_lock);
 800                 init_waitqueue_head(&nbd_dev[i].active_wq);
 801                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
 802                 nbd_dev[i].blksize = 1024;
 803                 nbd_dev[i].bytesize = 0;
 804                 disk->major = NBD_MAJOR;
 805                 disk->first_minor = i << part_shift;
 806                 disk->fops = &nbd_fops;
 807                 disk->private_data = &nbd_dev[i];
 808                 sprintf(disk->disk_name, "nbd%d", i);
 809                 set_capacity(disk, 0);
 810                 add_disk(disk);
 811         }
 812
 813         return 0;
 814 out:
 815         while (i--) {
 816                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 817                 put_disk(nbd_dev[i].disk);
 818         }
 819         kfree(nbd_dev);
 820         return err;
 821 }
 822
 823 static void __exit nbd_cleanup(void)
 824 {
 825         int i;
 826         for (i = 0; i < nbds_max; i++) {
 827                 struct gendisk *disk = nbd_dev[i].disk;
 828                 nbd_dev[i].magic = 0;
 829                 if (disk) {
 830                         del_gendisk(disk);
 831                         blk_cleanup_queue(disk->queue);
 832                         put_disk(disk);
 833                 }
 834         }
 835         unregister_blkdev(NBD_MAJOR, "nbd");
 836         kfree(nbd_dev);
 837         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 838 }
 839
 840 module_init(nbd_init);
 841 module_exit(nbd_cleanup);
 842
 843 MODULE_DESCRIPTION("Network Block Device");
 844 MODULE_LICENSE("GPL");
 845
 846 module_param(nbds_max, int, 0444);
 847 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
 848 module_param(max_part, int, 0444);
 849 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
 850 #ifndef NDEBUG
 851 module_param(debugflags, int, 0644);
 852 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 853 #endif