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 unsigned int debugflags;
  57 #endif /* NDEBUG */
  58
  59 static unsigned int nbds_max = 16;
  60 static struct nbd_device *nbd_dev;
  61 static int max_part;
  62
  63 /*
  64  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  65  * 1. Each NIC is essentially a synchronization point for all servers
  66  *    accessed through that NIC so there's no need to have more locks
  67  *    than NICs anyway.
  68  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  69  *    down each lock to the point where they're actually slower than just
  70  *    a single lock.
  71  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  72  */
  73 static DEFINE_SPINLOCK(nbd_lock);
  74
  75 #ifndef NDEBUG
  76 static const char *ioctl_cmd_to_ascii(int cmd)
  77 {
  78         switch (cmd) {
  79         case NBD_SET_SOCK: return "set-sock";
  80         case NBD_SET_BLKSIZE: return "set-blksize";
  81         case NBD_SET_SIZE: return "set-size";
  82         case NBD_DO_IT: return "do-it";
  83         case NBD_CLEAR_SOCK: return "clear-sock";
  84         case NBD_CLEAR_QUE: return "clear-que";
  85         case NBD_PRINT_DEBUG: return "print-debug";
  86         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
  87         case NBD_DISCONNECT: return "disconnect";
  88         case BLKROSET: return "set-read-only";
  89         case BLKFLSBUF: return "flush-buffer-cache";
  90         }
  91         return "unknown";
  92 }
  93
  94 static const char *nbdcmd_to_ascii(int cmd)
  95 {
  96         switch (cmd) {
  97         case  NBD_CMD_READ: return "read";
  98         case NBD_CMD_WRITE: return "write";
  99         case  NBD_CMD_DISC: return "disconnect";
 100         }
 101         return "invalid";
 102 }
 103 #endif /* NDEBUG */
 104
 105 static void nbd_end_request(struct request *req)
 106 {
 107         int error = req->errors ? -EIO : 0;
 108         struct request_queue *q = req->q;
 109         unsigned long flags;
 110
 111         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 112                         req, error ? "failed" : "done");
 113
 114         spin_lock_irqsave(q->queue_lock, flags);
 115         __blk_end_request_all(req, error);
 116         spin_unlock_irqrestore(q->queue_lock, flags);
 117 }
 118
 119 static void sock_shutdown(struct nbd_device *lo, int lock)
 120 {
 121         /* Forcibly shutdown the socket causing all listeners
 122          * to error
 123          *
 124          * FIXME: This code is duplicated from sys_shutdown, but
 125          * there should be a more generic interface rather than
 126          * calling socket ops directly here */
 127         if (lock)
 128                 mutex_lock(&lo->tx_lock);
 129         if (lo->sock) {
 130                 dev_warn(disk_to_dev(lo->disk), "shutting down socket\n");
 131                 kernel_sock_shutdown(lo->sock, SHUT_RDWR);
 132                 lo->sock = NULL;
 133         }
 134         if (lock)
 135                 mutex_unlock(&lo->tx_lock);
 136 }
 137
 138 static void nbd_xmit_timeout(unsigned long arg)
 139 {
 140         struct task_struct *task = (struct task_struct *)arg;
 141
 142         printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
 143                 task->comm, task->pid);
 144         force_sig(SIGKILL, task);
 145 }
 146
 147 /*
 148  *  Send or receive packet.
 149  */
 150 static int sock_xmit(struct nbd_device *lo, int send, void *buf, int size,
 151                 int msg_flags)
 152 {
 153         struct socket *sock = lo->sock;
 154         int result;
 155         struct msghdr msg;
 156         struct kvec iov;
 157         sigset_t blocked, oldset;
 158
 159         if (unlikely(!sock)) {
 160                 dev_err(disk_to_dev(lo->disk),
 161                         "Attempted %s on closed socket in sock_xmit\n",
 162                         (send ? "send" : "recv"));
 163                 return -EINVAL;
 164         }
 165
 166         /* Allow interception of SIGKILL only
 167          * Don't allow other signals to interrupt the transmission */
 168         siginitsetinv(&blocked, sigmask(SIGKILL));
 169         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 170
 171         do {
 172                 sock->sk->sk_allocation = GFP_NOIO;
 173                 iov.iov_base = buf;
 174                 iov.iov_len = size;
 175                 msg.msg_name = NULL;
 176                 msg.msg_namelen = 0;
 177                 msg.msg_control = NULL;
 178                 msg.msg_controllen = 0;
 179                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 180
 181                 if (send) {
 182                         struct timer_list ti;
 183
 184                         if (lo->xmit_timeout) {
 185                                 init_timer(&ti);
 186                                 ti.function = nbd_xmit_timeout;
 187                                 ti.data = (unsigned long)current;
 188                                 ti.expires = jiffies + lo->xmit_timeout;
 189                                 add_timer(&ti);
 190                         }
 191                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 192                         if (lo->xmit_timeout)
 193                                 del_timer_sync(&ti);
 194                 } else
 195                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 196                                                 msg.msg_flags);
 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;
 236         struct nbd_request request;
 237         unsigned long size = blk_rq_bytes(req);
 238         struct bio *bio;
 239
 240         request.magic = htonl(NBD_REQUEST_MAGIC);
 241         request.type = htonl(nbd_cmd(req));
 242         request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 243         request.len = htonl(size);
 244         memcpy(request.handle, &req, sizeof(req));
 245
 246         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
 247                         lo->disk->disk_name, req,
 248                         nbdcmd_to_ascii(nbd_cmd(req)),
 249                         (unsigned long long)blk_rq_pos(req) << 9,
 250                         blk_rq_bytes(req));
 251         result = sock_xmit(lo, 1, &request, sizeof(request),
 252                         (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
 253         if (result <= 0) {
 254                 dev_err(disk_to_dev(lo->disk),
 255                         "Send control failed (result %d)\n", result);
 256                 goto error_out;
 257         }
 258
 259         if (nbd_cmd(req) != NBD_CMD_WRITE)
 260                 return 0;
 261
 262         bio = req->bio;
 263         while (bio) {
 264                 struct bio *next = bio->bi_next;
 265                 int i;
 266                 struct bio_vec *bvec;
 267
 268                 bio_for_each_segment(bvec, bio, i) {
 269                         bool is_last = !next && i == bio->bi_vcnt - 1;
 270                         int flags = is_last ? 0 : MSG_MORE;
 271
 272                         dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 273                                         lo->disk->disk_name, req, bvec->bv_len);
 274                         result = sock_send_bvec(lo, bvec, flags);
 275                         if (result <= 0) {
 276                                 dev_err(disk_to_dev(lo->disk),
 277                                         "Send data failed (result %d)\n",
 278                                         result);
 279                                 goto error_out;
 280                         }
 281                         /*
 282                          * The completion might already have come in,
 283                          * so break for the last one instead of letting
 284                          * the iterator do it. This prevents use-after-free
 285                          * of the bio.
 286                          */
 287                         if (is_last)
 288                                 break;
 289                 }
 290                 bio = next;
 291         }
 292         return 0;
 293
 294 error_out:
 295         return -EIO;
 296 }
 297
 298 static struct request *nbd_find_request(struct nbd_device *lo,
 299                                         struct request *xreq)
 300 {
 301         struct request *req, *tmp;
 302         int err;
 303
 304         err = wait_event_interruptible(lo->active_wq, lo->active_req != xreq);
 305         if (unlikely(err))
 306                 goto out;
 307
 308         spin_lock(&lo->queue_lock);
 309         list_for_each_entry_safe(req, tmp, &lo->queue_head, queuelist) {
 310                 if (req != xreq)
 311                         continue;
 312                 list_del_init(&req->queuelist);
 313                 spin_unlock(&lo->queue_lock);
 314                 return req;
 315         }
 316         spin_unlock(&lo->queue_lock);
 317
 318         err = -ENOENT;
 319
 320 out:
 321         return ERR_PTR(err);
 322 }
 323
 324 static inline int sock_recv_bvec(struct nbd_device *lo, struct bio_vec *bvec)
 325 {
 326         int result;
 327         void *kaddr = kmap(bvec->bv_page);
 328         result = sock_xmit(lo, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 329                         MSG_WAITALL);
 330         kunmap(bvec->bv_page);
 331         return result;
 332 }
 333
 334 /* NULL returned = something went wrong, inform userspace */
 335 static struct request *nbd_read_stat(struct nbd_device *lo)
 336 {
 337         int result;
 338         struct nbd_reply reply;
 339         struct request *req;
 340
 341         reply.magic = 0;
 342         result = sock_xmit(lo, 0, &reply, sizeof(reply), MSG_WAITALL);
 343         if (result <= 0) {
 344                 dev_err(disk_to_dev(lo->disk),
 345                         "Receive control failed (result %d)\n", result);
 346                 goto harderror;
 347         }
 348
 349         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 350                 dev_err(disk_to_dev(lo->disk), "Wrong magic (0x%lx)\n",
 351                                 (unsigned long)ntohl(reply.magic));
 352                 result = -EPROTO;
 353                 goto harderror;
 354         }
 355
 356         req = nbd_find_request(lo, *(struct request **)reply.handle);
 357         if (IS_ERR(req)) {
 358                 result = PTR_ERR(req);
 359                 if (result != -ENOENT)
 360                         goto harderror;
 361
 362                 dev_err(disk_to_dev(lo->disk), "Unexpected reply (%p)\n",
 363                         reply.handle);
 364                 result = -EBADR;
 365                 goto harderror;
 366         }
 367
 368         if (ntohl(reply.error)) {
 369                 dev_err(disk_to_dev(lo->disk), "Other side returned error (%d)\n",
 370                         ntohl(reply.error));
 371                 req->errors++;
 372                 return req;
 373         }
 374
 375         dprintk(DBG_RX, "%s: request %p: got reply\n",
 376                         lo->disk->disk_name, req);
 377         if (nbd_cmd(req) == NBD_CMD_READ) {
 378                 struct req_iterator iter;
 379                 struct bio_vec *bvec;
 380
 381                 rq_for_each_segment(bvec, req, iter) {
 382                         result = sock_recv_bvec(lo, bvec);
 383                         if (result <= 0) {
 384                                 dev_err(disk_to_dev(lo->disk), "Receive data failed (result %d)\n",
 385                                         result);
 386                                 req->errors++;
 387                                 return req;
 388                         }
 389                         dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 390                                 lo->disk->disk_name, req, bvec->bv_len);
 391                 }
 392         }
 393         return req;
 394 harderror:
 395         lo->harderror = result;
 396         return NULL;
 397 }
 398
 399 static ssize_t pid_show(struct device *dev,
 400                         struct device_attribute *attr, char *buf)
 401 {
 402         struct gendisk *disk = dev_to_disk(dev);
 403
 404         return sprintf(buf, "%ld\n",
 405                 (long) ((struct nbd_device *)disk->private_data)->pid);
 406 }
 407
 408 static struct device_attribute pid_attr = {
 409         .attr = { .name = "pid", .mode = S_IRUGO},
 410         .show = pid_show,
 411 };
 412
 413 static int nbd_do_it(struct nbd_device *lo)
 414 {
 415         struct request *req;
 416         int ret;
 417
 418         BUG_ON(lo->magic != LO_MAGIC);
 419
 420         lo->pid = task_pid_nr(current);
 421         ret = device_create_file(disk_to_dev(lo->disk), &pid_attr);
 422         if (ret) {
 423                 dev_err(disk_to_dev(lo->disk), "device_create_file failed!\n");
 424                 lo->pid = 0;
 425                 return ret;
 426         }
 427
 428         while ((req = nbd_read_stat(lo)) != NULL)
 429                 nbd_end_request(req);
 430
 431         device_remove_file(disk_to_dev(lo->disk), &pid_attr);
 432         lo->pid = 0;
 433         return 0;
 434 }
 435
 436 static void nbd_clear_que(struct nbd_device *lo)
 437 {
 438         struct request *req;
 439
 440         BUG_ON(lo->magic != LO_MAGIC);
 441
 442         /*
 443          * Because we have set lo->sock to NULL under the tx_lock, all
 444          * modifications to the list must have completed by now.  For
 445          * the same reason, the active_req must be NULL.
 446          *
 447          * As a consequence, we don't need to take the spin lock while
 448          * purging the list here.
 449          */
 450         BUG_ON(lo->sock);
 451         BUG_ON(lo->active_req);
 452
 453         while (!list_empty(&lo->queue_head)) {
 454                 req = list_entry(lo->queue_head.next, struct request,
 455                                  queuelist);
 456                 list_del_init(&req->queuelist);
 457                 req->errors++;
 458                 nbd_end_request(req);
 459         }
 460
 461         while (!list_empty(&lo->waiting_queue)) {
 462                 req = list_entry(lo->waiting_queue.next, struct request,
 463                                  queuelist);
 464                 list_del_init(&req->queuelist);
 465                 req->errors++;
 466                 nbd_end_request(req);
 467         }
 468 }
 469
 470
 471 static void nbd_handle_req(struct nbd_device *lo, struct request *req)
 472 {
 473         if (req->cmd_type != REQ_TYPE_FS)
 474                 goto error_out;
 475
 476         nbd_cmd(req) = NBD_CMD_READ;
 477         if (rq_data_dir(req) == WRITE) {
 478                 nbd_cmd(req) = NBD_CMD_WRITE;
 479                 if (lo->flags & NBD_READ_ONLY) {
 480                         dev_err(disk_to_dev(lo->disk),
 481                                 "Write on read-only\n");
 482                         goto error_out;
 483                 }
 484         }
 485
 486         req->errors = 0;
 487
 488         mutex_lock(&lo->tx_lock);
 489         if (unlikely(!lo->sock)) {
 490                 mutex_unlock(&lo->tx_lock);
 491                 dev_err(disk_to_dev(lo->disk),
 492                         "Attempted send on closed socket\n");
 493                 goto error_out;
 494         }
 495
 496         lo->active_req = req;
 497
 498         if (nbd_send_req(lo, req) != 0) {
 499                 dev_err(disk_to_dev(lo->disk), "Request send failed\n");
 500                 req->errors++;
 501                 nbd_end_request(req);
 502         } else {
 503                 spin_lock(&lo->queue_lock);
 504                 list_add(&req->queuelist, &lo->queue_head);
 505                 spin_unlock(&lo->queue_lock);
 506         }
 507
 508         lo->active_req = NULL;
 509         mutex_unlock(&lo->tx_lock);
 510         wake_up_all(&lo->active_wq);
 511
 512         return;
 513
 514 error_out:
 515         req->errors++;
 516         nbd_end_request(req);
 517 }
 518
 519 static int nbd_thread(void *data)
 520 {
 521         struct nbd_device *lo = data;
 522         struct request *req;
 523
 524         set_user_nice(current, -20);
 525         while (!kthread_should_stop() || !list_empty(&lo->waiting_queue)) {
 526                 /* wait for something to do */
 527                 wait_event_interruptible(lo->waiting_wq,
 528                                          kthread_should_stop() ||
 529                                          !list_empty(&lo->waiting_queue));
 530
 531                 /* extract request */
 532                 if (list_empty(&lo->waiting_queue))
 533                         continue;
 534
 535                 spin_lock_irq(&lo->queue_lock);
 536                 req = list_entry(lo->waiting_queue.next, struct request,
 537                                  queuelist);
 538                 list_del_init(&req->queuelist);
 539                 spin_unlock_irq(&lo->queue_lock);
 540
 541                 /* handle request */
 542                 nbd_handle_req(lo, req);
 543         }
 544         return 0;
 545 }
 546
 547 /*
 548  * We always wait for result of write, for now. It would be nice to make it optional
 549  * in future
 550  * if ((rq_data_dir(req) == WRITE) && (lo->flags & NBD_WRITE_NOCHK))
 551  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 552  */
 553
 554 static void do_nbd_request(struct request_queue *q)
 555 {
 556         struct request *req;
 557
 558         while ((req = blk_fetch_request(q)) != NULL) {
 559                 struct nbd_device *lo;
 560
 561                 spin_unlock_irq(q->queue_lock);
 562
 563                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
 564                                 req->rq_disk->disk_name, req, req->cmd_type);
 565
 566                 lo = req->rq_disk->private_data;
 567
 568                 BUG_ON(lo->magic != LO_MAGIC);
 569
 570                 if (unlikely(!lo->sock)) {
 571                         dev_err(disk_to_dev(lo->disk),
 572                                 "Attempted send on closed socket\n");
 573                         req->errors++;
 574                         nbd_end_request(req);
 575                         spin_lock_irq(q->queue_lock);
 576                         continue;
 577                 }
 578
 579                 spin_lock_irq(&lo->queue_lock);
 580                 list_add_tail(&req->queuelist, &lo->waiting_queue);
 581                 spin_unlock_irq(&lo->queue_lock);
 582
 583                 wake_up(&lo->waiting_wq);
 584
 585                 spin_lock_irq(q->queue_lock);
 586         }
 587 }
 588
 589 /* Must be called with tx_lock held */
 590
 591 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *lo,
 592                        unsigned int cmd, unsigned long arg)
 593 {
 594         switch (cmd) {
 595         case NBD_DISCONNECT: {
 596                 struct request sreq;
 597
 598                 dev_info(disk_to_dev(lo->disk), "NBD_DISCONNECT\n");
 599                 if (!lo->sock)
 600                         return -EINVAL;
 601
 602                 mutex_unlock(&lo->tx_lock);
 603                 fsync_bdev(bdev);
 604                 mutex_lock(&lo->tx_lock);
 605                 blk_rq_init(NULL, &sreq);
 606                 sreq.cmd_type = REQ_TYPE_SPECIAL;
 607                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 608
 609                 /* Check again after getting mutex back.  */
 610                 if (!lo->sock)
 611                         return -EINVAL;
 612
 613                 lo->disconnect = 1;
 614
 615                 nbd_send_req(lo, &sreq);
 616                 return 0;
 617         }
 618
 619         case NBD_CLEAR_SOCK: {
 620                 struct file *file;
 621
 622                 lo->sock = NULL;
 623                 file = lo->file;
 624                 lo->file = NULL;
 625                 nbd_clear_que(lo);
 626                 BUG_ON(!list_empty(&lo->queue_head));
 627                 BUG_ON(!list_empty(&lo->waiting_queue));
 628                 kill_bdev(bdev);
 629                 if (file)
 630                         fput(file);
 631                 return 0;
 632         }
 633
 634         case NBD_SET_SOCK: {
 635                 struct file *file;
 636                 if (lo->file)
 637                         return -EBUSY;
 638                 file = fget(arg);
 639                 if (file) {
 640                         struct inode *inode = file->f_path.dentry->d_inode;
 641                         if (S_ISSOCK(inode->i_mode)) {
 642                                 lo->file = file;
 643                                 lo->sock = SOCKET_I(inode);
 644                                 if (max_part > 0)
 645                                         bdev->bd_invalidated = 1;
 646                                 lo->disconnect = 0; /* we're connected now */
 647                                 return 0;
 648                         } else {
 649                                 fput(file);
 650                         }
 651                 }
 652                 return -EINVAL;
 653         }
 654
 655         case NBD_SET_BLKSIZE:
 656                 lo->blksize = arg;
 657                 lo->bytesize &= ~(lo->blksize-1);
 658                 bdev->bd_inode->i_size = lo->bytesize;
 659                 set_blocksize(bdev, lo->blksize);
 660                 set_capacity(lo->disk, lo->bytesize >> 9);
 661                 return 0;
 662
 663         case NBD_SET_SIZE:
 664                 lo->bytesize = arg & ~(lo->blksize-1);
 665                 bdev->bd_inode->i_size = lo->bytesize;
 666                 set_blocksize(bdev, lo->blksize);
 667                 set_capacity(lo->disk, lo->bytesize >> 9);
 668                 return 0;
 669
 670         case NBD_SET_TIMEOUT:
 671                 lo->xmit_timeout = arg * HZ;
 672                 return 0;
 673
 674         case NBD_SET_SIZE_BLOCKS:
 675                 lo->bytesize = ((u64) arg) * lo->blksize;
 676                 bdev->bd_inode->i_size = lo->bytesize;
 677                 set_blocksize(bdev, lo->blksize);
 678                 set_capacity(lo->disk, lo->bytesize >> 9);
 679                 return 0;
 680
 681         case NBD_DO_IT: {
 682                 struct task_struct *thread;
 683                 struct file *file;
 684                 int error;
 685
 686                 if (lo->pid)
 687                         return -EBUSY;
 688                 if (!lo->file)
 689                         return -EINVAL;
 690
 691                 mutex_unlock(&lo->tx_lock);
 692
 693                 thread = kthread_create(nbd_thread, lo, "%s",
 694                                         lo->disk->disk_name);
 695                 if (IS_ERR(thread)) {
 696                         mutex_lock(&lo->tx_lock);
 697                         return PTR_ERR(thread);
 698                 }
 699                 wake_up_process(thread);
 700                 error = nbd_do_it(lo);
 701                 kthread_stop(thread);
 702
 703                 mutex_lock(&lo->tx_lock);
 704                 if (error)
 705                         return error;
 706                 sock_shutdown(lo, 0);
 707                 file = lo->file;
 708                 lo->file = NULL;
 709                 nbd_clear_que(lo);
 710                 dev_warn(disk_to_dev(lo->disk), "queue cleared\n");
 711                 kill_bdev(bdev);
 712                 if (file)
 713                         fput(file);
 714                 lo->bytesize = 0;
 715                 bdev->bd_inode->i_size = 0;
 716                 set_capacity(lo->disk, 0);
 717                 if (max_part > 0)
 718                         ioctl_by_bdev(bdev, BLKRRPART, 0);
 719                 if (lo->disconnect) /* user requested, ignore socket errors */
 720                         return 0;
 721                 return lo->harderror;
 722         }
 723
 724         case NBD_CLEAR_QUE:
 725                 /*
 726                  * This is for compatibility only.  The queue is always cleared
 727                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 728                  */
 729                 BUG_ON(!lo->sock && !list_empty(&lo->queue_head));
 730                 return 0;
 731
 732         case NBD_PRINT_DEBUG:
 733                 dev_info(disk_to_dev(lo->disk),
 734                         "next = %p, prev = %p, head = %p\n",
 735                         lo->queue_head.next, lo->queue_head.prev,
 736                         &lo->queue_head);
 737                 return 0;
 738         }
 739         return -ENOTTY;
 740 }
 741
 742 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 743                      unsigned int cmd, unsigned long arg)
 744 {
 745         struct nbd_device *lo = bdev->bd_disk->private_data;
 746         int error;
 747
 748         if (!capable(CAP_SYS_ADMIN))
 749                 return -EPERM;
 750
 751         BUG_ON(lo->magic != LO_MAGIC);
 752
 753         /* Anyone capable of this syscall can do *real bad* things */
 754         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 755                         lo->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 756
 757         mutex_lock(&lo->tx_lock);
 758         error = __nbd_ioctl(bdev, lo, cmd, arg);
 759         mutex_unlock(&lo->tx_lock);
 760
 761         return error;
 762 }
 763
 764 static const struct block_device_operations nbd_fops =
 765 {
 766         .owner =        THIS_MODULE,
 767         .ioctl =        nbd_ioctl,
 768 };
 769
 770 /*
 771  * And here should be modules and kernel interface
 772  *  (Just smiley confuses emacs :-)
 773  */
 774
 775 static int __init nbd_init(void)
 776 {
 777         int err = -ENOMEM;
 778         int i;
 779         int part_shift;
 780
 781         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 782
 783         if (max_part < 0) {
 784                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
 785                 return -EINVAL;
 786         }
 787
 788         part_shift = 0;
 789         if (max_part > 0) {
 790                 part_shift = fls(max_part);
 791
 792                 /*
 793                  * Adjust max_part according to part_shift as it is exported
 794                  * to user space so that user can know the max number of
 795                  * partition kernel should be able to manage.
 796                  *
 797                  * Note that -1 is required because partition 0 is reserved
 798                  * for the whole disk.
 799                  */
 800                 max_part = (1UL << part_shift) - 1;
 801         }
 802
 803         if ((1UL << part_shift) > DISK_MAX_PARTS)
 804                 return -EINVAL;
 805
 806         if (nbds_max > 1UL << (MINORBITS - part_shift))
 807                 return -EINVAL;
 808
 809         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
 810         if (!nbd_dev)
 811                 return -ENOMEM;
 812
 813         for (i = 0; i < nbds_max; i++) {
 814                 struct gendisk *disk = alloc_disk(1 << part_shift);
 815                 if (!disk)
 816                         goto out;
 817                 nbd_dev[i].disk = disk;
 818                 /*
 819                  * The new linux 2.5 block layer implementation requires
 820                  * every gendisk to have its very own request_queue struct.
 821                  * These structs are big so we dynamically allocate them.
 822                  */
 823                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 824                 if (!disk->queue) {
 825                         put_disk(disk);
 826                         goto out;
 827                 }
 828                 /*
 829                  * Tell the block layer that we are not a rotational device
 830                  */
 831                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
 832         }
 833
 834         if (register_blkdev(NBD_MAJOR, "nbd")) {
 835                 err = -EIO;
 836                 goto out;
 837         }
 838
 839         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 840         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 841
 842         for (i = 0; i < nbds_max; i++) {
 843                 struct gendisk *disk = nbd_dev[i].disk;
 844                 nbd_dev[i].file = NULL;
 845                 nbd_dev[i].magic = LO_MAGIC;
 846                 nbd_dev[i].flags = 0;
 847                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
 848                 spin_lock_init(&nbd_dev[i].queue_lock);
 849                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 850                 mutex_init(&nbd_dev[i].tx_lock);
 851                 init_waitqueue_head(&nbd_dev[i].active_wq);
 852                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
 853                 nbd_dev[i].blksize = 1024;
 854                 nbd_dev[i].bytesize = 0;
 855                 disk->major = NBD_MAJOR;
 856                 disk->first_minor = i << part_shift;
 857                 disk->fops = &nbd_fops;
 858                 disk->private_data = &nbd_dev[i];
 859                 sprintf(disk->disk_name, "nbd%d", i);
 860                 set_capacity(disk, 0);
 861                 add_disk(disk);
 862         }
 863
 864         return 0;
 865 out:
 866         while (i--) {
 867                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 868                 put_disk(nbd_dev[i].disk);
 869         }
 870         kfree(nbd_dev);
 871         return err;
 872 }
 873
 874 static void __exit nbd_cleanup(void)
 875 {
 876         int i;
 877         for (i = 0; i < nbds_max; i++) {
 878                 struct gendisk *disk = nbd_dev[i].disk;
 879                 nbd_dev[i].magic = 0;
 880                 if (disk) {
 881                         del_gendisk(disk);
 882                         blk_cleanup_queue(disk->queue);
 883                         put_disk(disk);
 884                 }
 885         }
 886         unregister_blkdev(NBD_MAJOR, "nbd");
 887         kfree(nbd_dev);
 888         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 889 }
 890
 891 module_init(nbd_init);
 892 module_exit(nbd_cleanup);
 893
 894 MODULE_DESCRIPTION("Network Block Device");
 895 MODULE_LICENSE("GPL");
 896
 897 module_param(nbds_max, int, 0444);
 898 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
 899 module_param(max_part, int, 0444);
 900 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
 901 #ifndef NDEBUG
 902 module_param(debugflags, int, 0644);
 903 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 904 #endif