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  * (part of code stolen from loop.c)
  11  *
  12  * 97-3-25 compiled 0-th version, not yet tested it
  13  *   (it did not work, BTW) (later that day) HEY! it works!
  14  *   (bit later) hmm, not that much... 2:00am next day:
  15  *   yes, it works, but it gives something like 50kB/sec
  16  * 97-4-01 complete rewrite to make it possible for many requests at
  17  *   once to be processed
  18  * 97-4-11 Making protocol independent of endianity etc.
  19  * 97-9-13 Cosmetic changes
  20  * 98-5-13 Attempt to make 64-bit-clean on 64-bit machines
  21  * 99-1-11 Attempt to make 64-bit-clean on 32-bit machines <ankry@mif.pg.gda.pl>
  22  * 01-2-27 Fix to store proper blockcount for kernel (calculated using
  23  *   BLOCK_SIZE_BITS, not device blocksize) <aga@permonline.ru>
  24  * 01-3-11 Make nbd work with new Linux block layer code. It now supports
  25  *   plugging like all the other block devices. Also added in MSG_MORE to
  26  *   reduce number of partial TCP segments sent. <steve@chygwyn.com>
  27  * 01-12-6 Fix deadlock condition by making queue locks independent of
  28  *   the transmit lock. <steve@chygwyn.com>
  29  * 02-10-11 Allow hung xmit to be aborted via SIGKILL & various fixes.
  30  *   <Paul.Clements@SteelEye.com> <James.Bottomley@SteelEye.com>
  31  * 03-06-22 Make nbd work with new linux 2.5 block layer design. This fixes
  32  *   memory corruption from module removal and possible memory corruption
  33  *   from sending/receiving disk data. <ldl@aros.net>
  34  * 03-06-23 Cosmetic changes. <ldl@aros.net>
  35  * 03-06-23 Enhance diagnostics support. <ldl@aros.net>
  36  * 03-06-24 Remove unneeded blksize_bits field from nbd_device struct.
  37  *   <ldl@aros.net>
  38  * 03-06-24 Cleanup PARANOIA usage & code. <ldl@aros.net>
  39  *
  40  * possible FIXME: make set_sock / set_blksize / set_size / do_it one syscall
  41  * why not: would need verify_area and friends, would share yet another
  42  *          structure with userland
  43  */
  44
  45 #include <linux/major.h>
  46
  47 #include <linux/blkdev.h>
  48 #include <linux/module.h>
  49 #include <linux/init.h>
  50 #include <linux/sched.h>
  51 #include <linux/fs.h>
  52 #include <linux/bio.h>
  53 #include <linux/stat.h>
  54 #include <linux/errno.h>
  55 #include <linux/file.h>
  56 #include <linux/ioctl.h>
  57 #include <net/sock.h>
  58
  59 #include <linux/devfs_fs_kernel.h>
  60
  61 #include <asm/uaccess.h>
  62 #include <asm/types.h>
  63
  64 /* Define PARANOIA in linux/nbd.h to turn on extra sanity checking */
  65 #include <linux/nbd.h>
  66
  67 #ifdef PARANOIA
  68 #define LO_MAGIC 0x68797548
  69 #endif
  70
  71 #ifdef NDEBUG
  72 #define dprintk(flags, fmt...)
  73 #else /* NDEBUG */
  74 #define dprintk(flags, fmt...) do { \
  75         if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
  76 } while (0)
  77 #define DBG_IOCTL       0x0004
  78 #define DBG_INIT        0x0010
  79 #define DBG_EXIT        0x0020
  80 #define DBG_BLKDEV      0x0100
  81 #define DBG_RX          0x0200
  82 #define DBG_TX          0x0400
  83 static unsigned int debugflags;
  84 #endif /* NDEBUG */
  85
  86 static struct nbd_device nbd_dev[MAX_NBD];
  87
  88 /*
  89  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  90  * 1. Each NIC is essentially a synchronization point for all servers
  91  *    accessed through that NIC so there's no need to have more locks
  92  *    than NICs anyway.
  93  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  94  *    down each lock to the point where they're actually slower than just
  95  *    a single lock.
  96  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  97  */
  98 static spinlock_t nbd_lock = SPIN_LOCK_UNLOCKED;
  99
 100 #ifdef PARANOIA
 101 static int requests_in;
 102 static int requests_out;
 103 #endif
 104
 105 #ifndef NDEBUG
 106 static const char *ioctl_cmd_to_ascii(int cmd)
 107 {
 108         switch (cmd) {
 109         case NBD_SET_SOCK: return "set-sock";
 110         case NBD_SET_BLKSIZE: return "set-blksize";
 111         case NBD_SET_SIZE: return "set-size";
 112         case NBD_DO_IT: return "do-it";
 113         case NBD_CLEAR_SOCK: return "clear-sock";
 114         case NBD_CLEAR_QUE: return "clear-que";
 115         case NBD_PRINT_DEBUG: return "print-debug";
 116         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
 117         case NBD_DISCONNECT: return "disconnect";
 118         case BLKROSET: return "set-read-only";
 119         case BLKFLSBUF: return "flush-buffer-cache";
 120         }
 121         return "unknown";
 122 }
 123
 124 static const char *nbdcmd_to_ascii(int cmd)
 125 {
 126         switch (cmd) {
 127         case  NBD_CMD_READ: return "read";
 128         case NBD_CMD_WRITE: return "write";
 129         case  NBD_CMD_DISC: return "disconnect";
 130         }
 131         return "invalid";
 132 }
 133 #endif /* NDEBUG */
 134
 135 static void nbd_end_request(struct request *req)
 136 {
 137         int uptodate = (req->errors == 0) ? 1 : 0;
 138         request_queue_t *q = req->q;
 139         struct nbd_device *lo = req->rq_disk->private_data;
 140         unsigned long flags;
 141
 142         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 143                         req, uptodate? "done": "failed");
 144
 145         spin_lock(&lo->queue_lock);
 146         while (req->ref_count > 1) { /* still in send */
 147                 spin_unlock(&lo->queue_lock);
 148                 printk(KERN_DEBUG "%s: request %p still in use (%d), waiting\n",
 149                     lo->disk->disk_name, req, req->ref_count);
 150                 set_current_state(TASK_UNINTERRUPTIBLE);
 151                 schedule_timeout(HZ); /* wait a second */
 152                 spin_lock(&lo->queue_lock);
 153         }
 154         spin_unlock(&lo->queue_lock);
 155
 156 #ifdef PARANOIA
 157         requests_out++;
 158 #endif
 159         spin_lock_irqsave(q->queue_lock, flags);
 160         if (!end_that_request_first(req, uptodate, req->nr_sectors)) {
 161                 end_that_request_last(req);
 162         }
 163         spin_unlock_irqrestore(q->queue_lock, flags);
 164 }
 165
 166 /*
 167  *  Send or receive packet.
 168  */
 169 static int sock_xmit(struct socket *sock, int send, void *buf, int size,
 170                 int msg_flags)
 171 {
 172         mm_segment_t oldfs;
 173         int result;
 174         struct msghdr msg;
 175         struct iovec iov;
 176         unsigned long flags;
 177         sigset_t oldset;
 178
 179         oldfs = get_fs();
 180         set_fs(get_ds());
 181         /* Allow interception of SIGKILL only
 182          * Don't allow other signals to interrupt the transmission */
 183         spin_lock_irqsave(&current->sighand->siglock, flags);
 184         oldset = current->blocked;
 185         sigfillset(&current->blocked);
 186         sigdelsetmask(&current->blocked, sigmask(SIGKILL));
 187         recalc_sigpending();
 188         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 189
 190         do {
 191                 sock->sk->sk_allocation = GFP_NOIO;
 192                 iov.iov_base = buf;
 193                 iov.iov_len = size;
 194                 msg.msg_name = NULL;
 195                 msg.msg_namelen = 0;
 196                 msg.msg_iov = &iov;
 197                 msg.msg_iovlen = 1;
 198                 msg.msg_control = NULL;
 199                 msg.msg_controllen = 0;
 200                 msg.msg_namelen = 0;
 201                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 202
 203                 if (send)
 204                         result = sock_sendmsg(sock, &msg, size);
 205                 else
 206                         result = sock_recvmsg(sock, &msg, size, 0);
 207
 208                 if (signal_pending(current)) {
 209                         siginfo_t info;
 210                         spin_lock_irqsave(&current->sighand->siglock, flags);
 211                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 212                                 current->pid, current->comm,
 213                                 dequeue_signal(current, &current->blocked, &info));
 214                         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 215                         result = -EINTR;
 216                         break;
 217                 }
 218
 219                 if (result <= 0) {
 220 #ifdef PARANOIA
 221                         printk(KERN_ERR "nbd: %s - sock=%p at buf=%p, size=%d returned %d.\n",
 222                                send? "send": "receive", sock, buf, size, result);
 223 #endif
 224                         break;
 225                 }
 226                 size -= result;
 227                 buf += result;
 228         } while (size > 0);
 229
 230         spin_lock_irqsave(&current->sighand->siglock, flags);
 231         current->blocked = oldset;
 232         recalc_sigpending();
 233         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 234
 235         set_fs(oldfs);
 236         return result;
 237 }
 238
 239 static inline int sock_send_bvec(struct socket *sock, struct bio_vec *bvec,
 240                 int flags)
 241 {
 242         int result;
 243         void *kaddr = kmap(bvec->bv_page);
 244         result = sock_xmit(sock, 1, kaddr + bvec->bv_offset, bvec->bv_len,
 245                         flags);
 246         kunmap(bvec->bv_page);
 247         return result;
 248 }
 249
 250 void nbd_send_req(struct nbd_device *lo, struct request *req)
 251 {
 252         int result, i, flags;
 253         struct nbd_request request;
 254         unsigned long size = req->nr_sectors << 9;
 255         struct socket *sock = lo->sock;
 256
 257         request.magic = htonl(NBD_REQUEST_MAGIC);
 258         request.type = htonl(nbd_cmd(req));
 259         request.from = cpu_to_be64((u64) req->sector << 9);
 260         request.len = htonl(size);
 261         memcpy(request.handle, &req, sizeof(req));
 262
 263         down(&lo->tx_lock);
 264
 265         if (!sock || !lo->sock) {
 266                 printk(KERN_ERR "%s: Attempted send on closed socket\n",
 267                                 lo->disk->disk_name);
 268                 goto error_out;
 269         }
 270
 271         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%luB)\n",
 272                         lo->disk->disk_name, req,
 273                         nbdcmd_to_ascii(nbd_cmd(req)),
 274                         (unsigned long long)req->sector << 9,
 275                         req->nr_sectors << 9);
 276         result = sock_xmit(sock, 1, &request, sizeof(request),
 277                         (nbd_cmd(req) == NBD_CMD_WRITE)? MSG_MORE: 0);
 278         if (result <= 0) {
 279                 printk(KERN_ERR "%s: Send control failed (result %d)\n",
 280                                 lo->disk->disk_name, result);
 281                 goto error_out;
 282         }
 283
 284         if (nbd_cmd(req) == NBD_CMD_WRITE) {
 285                 struct bio *bio;
 286                 /*
 287                  * we are really probing at internals to determine
 288                  * whether to set MSG_MORE or not...
 289                  */
 290                 rq_for_each_bio(bio, req) {
 291                         struct bio_vec *bvec;
 292                         bio_for_each_segment(bvec, bio, i) {
 293                                 flags = 0;
 294                                 if ((i < (bio->bi_vcnt - 1)) || bio->bi_next)
 295                                         flags = MSG_MORE;
 296                                 dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 297                                                 lo->disk->disk_name, req,
 298                                                 bvec->bv_len);
 299                                 result = sock_send_bvec(sock, bvec, flags);
 300                                 if (result <= 0) {
 301                                         printk(KERN_ERR "%s: Send data failed (result %d)\n",
 302                                                         lo->disk->disk_name,
 303                                                         result);
 304                                         goto error_out;
 305                                 }
 306                         }
 307                 }
 308         }
 309         up(&lo->tx_lock);
 310         return;
 311
 312       error_out:
 313         up(&lo->tx_lock);
 314         req->errors++;
 315 }
 316
 317 static struct request *nbd_find_request(struct nbd_device *lo, char *handle)
 318 {
 319         struct request *req;
 320         struct list_head *tmp;
 321         struct request *xreq;
 322
 323         memcpy(&xreq, handle, sizeof(xreq));
 324
 325         spin_lock(&lo->queue_lock);
 326         list_for_each(tmp, &lo->queue_head) {
 327                 req = list_entry(tmp, struct request, queuelist);
 328                 if (req != xreq)
 329                         continue;
 330                 list_del_init(&req->queuelist);
 331                 spin_unlock(&lo->queue_lock);
 332                 return req;
 333         }
 334         spin_unlock(&lo->queue_lock);
 335         return NULL;
 336 }
 337
 338 static inline int sock_recv_bvec(struct socket *sock, struct bio_vec *bvec)
 339 {
 340         int result;
 341         void *kaddr = kmap(bvec->bv_page);
 342         result = sock_xmit(sock, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 343                         MSG_WAITALL);
 344         kunmap(bvec->bv_page);
 345         return result;
 346 }
 347
 348 /* NULL returned = something went wrong, inform userspace */
 349 struct request *nbd_read_stat(struct nbd_device *lo)
 350 {
 351         int result;
 352         struct nbd_reply reply;
 353         struct request *req;
 354         struct socket *sock = lo->sock;
 355
 356         reply.magic = 0;
 357         result = sock_xmit(sock, 0, &reply, sizeof(reply), MSG_WAITALL);
 358         if (result <= 0) {
 359                 printk(KERN_ERR "%s: Receive control failed (result %d)\n",
 360                                 lo->disk->disk_name, result);
 361                 lo->harderror = result;
 362                 return NULL;
 363         }
 364         req = nbd_find_request(lo, reply.handle);
 365         if (req == NULL) {
 366                 printk(KERN_ERR "%s: Unexpected reply (%p)\n",
 367                                 lo->disk->disk_name, reply.handle);
 368                 lo->harderror = result;
 369                 return NULL;
 370         }
 371
 372         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 373                 printk(KERN_ERR "%s: Wrong magic (0x%lx)\n",
 374                                 lo->disk->disk_name,
 375                                 (unsigned long)ntohl(reply.magic));
 376                 lo->harderror = result;
 377                 return NULL;
 378         }
 379         if (ntohl(reply.error)) {
 380                 printk(KERN_ERR "%s: Other side returned error (%d)\n",
 381                                 lo->disk->disk_name, ntohl(reply.error));
 382                 req->errors++;
 383                 return req;
 384         }
 385
 386         dprintk(DBG_RX, "%s: request %p: got reply\n",
 387                         lo->disk->disk_name, req);
 388         if (nbd_cmd(req) == NBD_CMD_READ) {
 389                 int i;
 390                 struct bio *bio;
 391                 rq_for_each_bio(bio, req) {
 392                         struct bio_vec *bvec;
 393                         bio_for_each_segment(bvec, bio, i) {
 394                                 result = sock_recv_bvec(sock, bvec);
 395                                 if (result <= 0) {
 396                                         printk(KERN_ERR "%s: Receive data failed (result %d)\n",
 397                                                         lo->disk->disk_name,
 398                                                         result);
 399                                         lo->harderror = result;
 400                                         return NULL;
 401                                 }
 402                                 dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 403                                         lo->disk->disk_name, req, bvec->bv_len);
 404                         }
 405                 }
 406         }
 407         return req;
 408 }
 409
 410 void nbd_do_it(struct nbd_device *lo)
 411 {
 412         struct request *req;
 413
 414 #ifdef PARANOIA
 415         BUG_ON(lo->magic != LO_MAGIC);
 416 #endif
 417         while ((req = nbd_read_stat(lo)) != NULL)
 418                 nbd_end_request(req);
 419         printk(KERN_NOTICE "%s: req should never be null\n",
 420                         lo->disk->disk_name);
 421         return;
 422 }
 423
 424 void nbd_clear_que(struct nbd_device *lo)
 425 {
 426         struct request *req;
 427
 428 #ifdef PARANOIA
 429         BUG_ON(lo->magic != LO_MAGIC);
 430 #endif
 431
 432         do {
 433                 req = NULL;
 434                 spin_lock(&lo->queue_lock);
 435                 if (!list_empty(&lo->queue_head)) {
 436                         req = list_entry(lo->queue_head.next, struct request, queuelist);
 437                         list_del_init(&req->queuelist);
 438                 }
 439                 spin_unlock(&lo->queue_lock);
 440                 if (req) {
 441                         req->errors++;
 442                         nbd_end_request(req);
 443                 }
 444         } while (req);
 445 }
 446
 447 /*
 448  * We always wait for result of write, for now. It would be nice to make it optional
 449  * in future
 450  * if ((req->cmd == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
 451  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 452  */
 453
 454 static void do_nbd_request(request_queue_t * q)
 455 {
 456         struct request *req;
 457
 458         while ((req = elv_next_request(q)) != NULL) {
 459                 struct nbd_device *lo;
 460
 461                 blkdev_dequeue_request(req);
 462                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%lx)\n",
 463                                 req->rq_disk->disk_name, req, req->flags);
 464
 465                 if (!(req->flags & REQ_CMD))
 466                         goto error_out;
 467
 468                 lo = req->rq_disk->private_data;
 469 #ifdef PARANOIA
 470                 BUG_ON(lo->magic != LO_MAGIC);
 471 #endif
 472                 if (!lo->file) {
 473                         printk(KERN_ERR "%s: Request when not-ready\n",
 474                                         lo->disk->disk_name);
 475                         goto error_out;
 476                 }
 477                 nbd_cmd(req) = NBD_CMD_READ;
 478                 if (rq_data_dir(req) == WRITE) {
 479                         nbd_cmd(req) = NBD_CMD_WRITE;
 480                         if (lo->flags & NBD_READ_ONLY) {
 481                                 printk(KERN_ERR "%s: Write on read-only\n",
 482                                                 lo->disk->disk_name);
 483                                 goto error_out;
 484                         }
 485                 }
 486 #ifdef PARANOIA
 487                 requests_in++;
 488 #endif
 489
 490                 req->errors = 0;
 491                 spin_unlock_irq(q->queue_lock);
 492
 493                 spin_lock(&lo->queue_lock);
 494
 495                 if (!lo->file) {
 496                         spin_unlock(&lo->queue_lock);
 497                         printk(KERN_ERR "%s: failed between accept and semaphore, file lost\n",
 498                                         lo->disk->disk_name);
 499                         req->errors++;
 500                         nbd_end_request(req);
 501                         spin_lock_irq(q->queue_lock);
 502                         continue;
 503                 }
 504
 505                 list_add(&req->queuelist, &lo->queue_head);
 506                 req->ref_count++; /* make sure req does not get freed */
 507                 spin_unlock(&lo->queue_lock);
 508
 509                 nbd_send_req(lo, req);
 510
 511                 if (req->errors) {
 512                         printk(KERN_ERR "%s: Request send failed\n",
 513                                         lo->disk->disk_name);
 514                         spin_lock(&lo->queue_lock);
 515                         list_del_init(&req->queuelist);
 516                         req->ref_count--;
 517                         spin_unlock(&lo->queue_lock);
 518                         nbd_end_request(req);
 519                         spin_lock_irq(q->queue_lock);
 520                         continue;
 521                 }
 522
 523                 spin_lock(&lo->queue_lock);
 524                 req->ref_count--;
 525                 spin_unlock(&lo->queue_lock);
 526                 spin_lock_irq(q->queue_lock);
 527                 continue;
 528
 529               error_out:
 530                 req->errors++;
 531                 spin_unlock(q->queue_lock);
 532                 nbd_end_request(req);
 533                 spin_lock(q->queue_lock);
 534         }
 535         return;
 536 }
 537
 538 static int nbd_ioctl(struct inode *inode, struct file *file,
 539                      unsigned int cmd, unsigned long arg)
 540 {
 541         struct nbd_device *lo = inode->i_bdev->bd_disk->private_data;
 542         int error;
 543         struct request sreq ;
 544
 545         if (!capable(CAP_SYS_ADMIN))
 546                 return -EPERM;
 547 #ifdef PARANOIA
 548         BUG_ON(lo->magic != LO_MAGIC);
 549 #endif
 550         /* Anyone capable of this syscall can do *real bad* things */
 551         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 552                         lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 553
 554         switch (cmd) {
 555         case NBD_DISCONNECT:
 556                 printk(KERN_INFO "%s: NBD_DISCONNECT\n", lo->disk->disk_name);
 557                 sreq.flags = REQ_SPECIAL;
 558                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 559                 /*
 560                  * Set these to sane values in case server implementation
 561                  * fails to check the request type first and also to keep
 562                  * debugging output cleaner.
 563                  */
 564                 sreq.sector = 0;
 565                 sreq.nr_sectors = 0;
 566                 if (!lo->sock)
 567                         return -EINVAL;
 568                 nbd_send_req(lo, &sreq);
 569                 return 0;
 570
 571         case NBD_CLEAR_SOCK:
 572                 error = 0;
 573                 down(&lo->tx_lock);
 574                 lo->sock = NULL;
 575                 up(&lo->tx_lock);
 576                 spin_lock(&lo->queue_lock);
 577                 file = lo->file;
 578                 lo->file = NULL;
 579                 spin_unlock(&lo->queue_lock);
 580                 nbd_clear_que(lo);
 581                 spin_lock(&lo->queue_lock);
 582                 if (!list_empty(&lo->queue_head)) {
 583                         printk(KERN_ERR "nbd: disconnect: some requests are in progress -> please try again.\n");
 584                         error = -EBUSY;
 585                 }
 586                 spin_unlock(&lo->queue_lock);
 587                 if (file)
 588                         fput(file);
 589                 return error;
 590         case NBD_SET_SOCK:
 591                 if (lo->file)
 592                         return -EBUSY;
 593                 error = -EINVAL;
 594                 file = fget(arg);
 595                 if (file) {
 596                         inode = file->f_dentry->d_inode;
 597                         if (inode->i_sock) {
 598                                 lo->file = file;
 599                                 lo->sock = SOCKET_I(inode);
 600                                 error = 0;
 601                         } else {
 602                                 fput(file);
 603                         }
 604                 }
 605                 return error;
 606         case NBD_SET_BLKSIZE:
 607                 lo->blksize = arg;
 608                 lo->bytesize &= ~(lo->blksize-1);
 609                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 610                 set_blocksize(inode->i_bdev, lo->blksize);
 611                 set_capacity(lo->disk, lo->bytesize >> 9);
 612                 return 0;
 613         case NBD_SET_SIZE:
 614                 lo->bytesize = arg & ~(lo->blksize-1);
 615                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 616                 set_blocksize(inode->i_bdev, lo->blksize);
 617                 set_capacity(lo->disk, lo->bytesize >> 9);
 618                 return 0;
 619         case NBD_SET_SIZE_BLOCKS:
 620                 lo->bytesize = ((u64) arg) * lo->blksize;
 621                 inode->i_bdev->bd_inode->i_size = lo->bytesize;
 622                 set_blocksize(inode->i_bdev, lo->blksize);
 623                 set_capacity(lo->disk, lo->bytesize >> 9);
 624                 return 0;
 625         case NBD_DO_IT:
 626                 if (!lo->file)
 627                         return -EINVAL;
 628                 nbd_do_it(lo);
 629                 /* on return tidy up in case we have a signal */
 630                 /* Forcibly shutdown the socket causing all listeners
 631                  * to error
 632                  *
 633                  * FIXME: This code is duplicated from sys_shutdown, but
 634                  * there should be a more generic interface rather than
 635                  * calling socket ops directly here */
 636                 down(&lo->tx_lock);
 637                 if (lo->sock) {
 638                         printk(KERN_WARNING "%s: shutting down socket\n",
 639                                 lo->disk->disk_name);
 640                         lo->sock->ops->shutdown(lo->sock,
 641                                 SEND_SHUTDOWN|RCV_SHUTDOWN);
 642                         lo->sock = NULL;
 643                 }
 644                 up(&lo->tx_lock);
 645                 spin_lock(&lo->queue_lock);
 646                 file = lo->file;
 647                 lo->file = NULL;
 648                 spin_unlock(&lo->queue_lock);
 649                 nbd_clear_que(lo);
 650                 printk(KERN_WARNING "%s: queue cleared\n", lo->disk->disk_name);
 651                 if (file)
 652                         fput(file);
 653                 return lo->harderror;
 654         case NBD_CLEAR_QUE:
 655                 down(&lo->tx_lock);
 656                 if (lo->sock) {
 657                         up(&lo->tx_lock);
 658                         return 0; /* probably should be error, but that would
 659                                    * break "nbd-client -d", so just return 0 */
 660                 }
 661                 up(&lo->tx_lock);
 662                 nbd_clear_que(lo);
 663                 return 0;
 664         case NBD_PRINT_DEBUG:
 665 #ifdef PARANOIA
 666                 printk(KERN_INFO "%s: next = %p, prev = %p. Global: in %d, out %d\n",
 667                         inode->i_bdev->bd_disk->disk_name, lo->queue_head.next,
 668                         lo->queue_head.prev, requests_in, requests_out);
 669 #else
 670                 printk(KERN_INFO "%s: next = %p, prev = %p\n",
 671                         inode->i_bdev->bd_disk->disk_name,
 672                         lo->queue_head.next, lo->queue_head.prev);
 673 #endif
 674                 return 0;
 675         }
 676         return -EINVAL;
 677 }
 678
 679 static struct block_device_operations nbd_fops =
 680 {
 681         .owner =        THIS_MODULE,
 682         .ioctl =        nbd_ioctl,
 683 };
 684
 685 /*
 686  * And here should be modules and kernel interface
 687  *  (Just smiley confuses emacs :-)
 688  */
 689
 690 static int __init nbd_init(void)
 691 {
 692         int err = -ENOMEM;
 693         int i;
 694
 695 #ifdef PARANOIA
 696         if (sizeof(struct nbd_request) != 28) {
 697                 printk(KERN_CRIT "nbd: Sizeof nbd_request needs to be 28 in order to work!\n" );
 698                 return -EIO;
 699         }
 700 #endif
 701
 702         for (i = 0; i < MAX_NBD; i++) {
 703                 struct gendisk *disk = alloc_disk(1);
 704                 if (!disk)
 705                         goto out;
 706                 nbd_dev[i].disk = disk;
 707                 /*
 708                  * The new linux 2.5 block layer implementation requires
 709                  * every gendisk to have its very own request_queue struct.
 710                  * These structs are big so we dynamically allocate them.
 711                  */
 712                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 713                 if (!disk->queue) {
 714                         put_disk(disk);
 715                         goto out;
 716                 }
 717         }
 718
 719         if (register_blkdev(NBD_MAJOR, "nbd")) {
 720                 err = -EIO;
 721                 goto out;
 722         }
 723
 724         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 725         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 726
 727         devfs_mk_dir("nbd");
 728         for (i = 0; i < MAX_NBD; i++) {
 729                 struct gendisk *disk = nbd_dev[i].disk;
 730                 nbd_dev[i].file = NULL;
 731 #ifdef PARANOIA
 732                 nbd_dev[i].magic = LO_MAGIC;
 733 #endif
 734                 nbd_dev[i].flags = 0;
 735                 spin_lock_init(&nbd_dev[i].queue_lock);
 736                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 737                 init_MUTEX(&nbd_dev[i].tx_lock);
 738                 nbd_dev[i].blksize = 1024;
 739                 nbd_dev[i].bytesize = ((u64)0x7ffffc00) << 10; /* 2TB */
 740                 disk->major = NBD_MAJOR;
 741                 disk->first_minor = i;
 742                 disk->fops = &nbd_fops;
 743                 disk->private_data = &nbd_dev[i];
 744                 disk->flags |= GENHD_FL_SUPPRESS_PARTITION_INFO;
 745                 sprintf(disk->disk_name, "nbd%d", i);
 746                 sprintf(disk->devfs_name, "nbd/%d", i);
 747                 set_capacity(disk, 0x3ffffe);
 748                 add_disk(disk);
 749         }
 750
 751         return 0;
 752 out:
 753         while (i--) {
 754                 if (nbd_dev[i].disk->queue)
 755                         blk_cleanup_queue(nbd_dev[i].disk->queue);
 756                 put_disk(nbd_dev[i].disk);
 757         }
 758         return err;
 759 }
 760
 761 static void __exit nbd_cleanup(void)
 762 {
 763         int i;
 764         for (i = 0; i < MAX_NBD; i++) {
 765                 struct gendisk *disk = nbd_dev[i].disk;
 766                 if (disk) {
 767                         if (disk->queue)
 768                                 blk_cleanup_queue(disk->queue);
 769                         del_gendisk(disk);
 770                         put_disk(disk);
 771                 }
 772         }
 773         devfs_remove("nbd");
 774         unregister_blkdev(NBD_MAJOR, "nbd");
 775         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 776 }
 777
 778 module_init(nbd_init);
 779 module_exit(nbd_cleanup);
 780
 781 MODULE_DESCRIPTION("Network Block Device");
 782 MODULE_LICENSE("GPL");
 783
 784 #ifndef NDEBUG
 785 MODULE_PARM(debugflags, "i");
 786 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 787 #endif