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 #include <linux/types.h>
  36
  37 #include <asm/uaccess.h>
  38 #include <asm/types.h>
  39
  40 #include <linux/nbd.h>
  41
  42 struct nbd_device {
  43         int flags;
  44         int harderror;          /* Code of hard error                   */
  45         struct socket * sock;   /* If == NULL, device is not ready, yet */
  46         int magic;
  47
  48         spinlock_t queue_lock;
  49         struct list_head queue_head;    /* Requests waiting result */
  50         struct request *active_req;
  51         wait_queue_head_t active_wq;
  52         struct list_head waiting_queue; /* Requests to be sent */
  53         wait_queue_head_t waiting_wq;
  54
  55         struct mutex tx_lock;
  56         struct gendisk *disk;
  57         int blksize;
  58         loff_t bytesize;
  59         pid_t pid; /* pid of nbd-client, if attached */
  60         int xmit_timeout;
  61         int disconnect; /* a disconnect has been requested by user */
  62 };
  63
  64 #define NBD_MAGIC 0x68797548
  65
  66 static unsigned int nbds_max = 16;
  67 static struct nbd_device *nbd_dev;
  68 static int max_part;
  69
  70 /*
  71  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  72  * 1. Each NIC is essentially a synchronization point for all servers
  73  *    accessed through that NIC so there's no need to have more locks
  74  *    than NICs anyway.
  75  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  76  *    down each lock to the point where they're actually slower than just
  77  *    a single lock.
  78  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  79  */
  80 static DEFINE_SPINLOCK(nbd_lock);
  81
  82 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
  83 {
  84         return disk_to_dev(nbd->disk);
  85 }
  86
  87 static const char *nbdcmd_to_ascii(int cmd)
  88 {
  89         switch (cmd) {
  90         case  NBD_CMD_READ: return "read";
  91         case NBD_CMD_WRITE: return "write";
  92         case  NBD_CMD_DISC: return "disconnect";
  93         case NBD_CMD_FLUSH: return "flush";
  94         case  NBD_CMD_TRIM: return "trim/discard";
  95         }
  96         return "invalid";
  97 }
  98
  99 static void nbd_end_request(struct nbd_device *nbd, struct request *req)
 100 {
 101         int error = req->errors ? -EIO : 0;
 102         struct request_queue *q = req->q;
 103         unsigned long flags;
 104
 105         dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
 106                 error ? "failed" : "done");
 107
 108         spin_lock_irqsave(q->queue_lock, flags);
 109         __blk_end_request_all(req, error);
 110         spin_unlock_irqrestore(q->queue_lock, flags);
 111 }
 112
 113 /*
 114  * Forcibly shutdown the socket causing all listeners to error
 115  */
 116 static void sock_shutdown(struct nbd_device *nbd, int lock)
 117 {
 118         if (lock)
 119                 mutex_lock(&nbd->tx_lock);
 120         if (nbd->sock) {
 121                 dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
 122                 kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
 123                 nbd->sock = NULL;
 124         }
 125         if (lock)
 126                 mutex_unlock(&nbd->tx_lock);
 127 }
 128
 129 static void nbd_xmit_timeout(unsigned long arg)
 130 {
 131         struct task_struct *task = (struct task_struct *)arg;
 132
 133         printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
 134                 task->comm, task->pid);
 135         force_sig(SIGKILL, task);
 136 }
 137
 138 /*
 139  *  Send or receive packet.
 140  */
 141 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
 142                 int msg_flags)
 143 {
 144         struct socket *sock = nbd->sock;
 145         int result;
 146         struct msghdr msg;
 147         struct kvec iov;
 148         sigset_t blocked, oldset;
 149         unsigned long pflags = current->flags;
 150
 151         if (unlikely(!sock)) {
 152                 dev_err(disk_to_dev(nbd->disk),
 153                         "Attempted %s on closed socket in sock_xmit\n",
 154                         (send ? "send" : "recv"));
 155                 return -EINVAL;
 156         }
 157
 158         /* Allow interception of SIGKILL only
 159          * Don't allow other signals to interrupt the transmission */
 160         siginitsetinv(&blocked, sigmask(SIGKILL));
 161         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 162
 163         current->flags |= PF_MEMALLOC;
 164         do {
 165                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 166                 iov.iov_base = buf;
 167                 iov.iov_len = size;
 168                 msg.msg_name = NULL;
 169                 msg.msg_namelen = 0;
 170                 msg.msg_control = NULL;
 171                 msg.msg_controllen = 0;
 172                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 173
 174                 if (send) {
 175                         struct timer_list ti;
 176
 177                         if (nbd->xmit_timeout) {
 178                                 init_timer(&ti);
 179                                 ti.function = nbd_xmit_timeout;
 180                                 ti.data = (unsigned long)current;
 181                                 ti.expires = jiffies + nbd->xmit_timeout;
 182                                 add_timer(&ti);
 183                         }
 184                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 185                         if (nbd->xmit_timeout)
 186                                 del_timer_sync(&ti);
 187                 } else
 188                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 189                                                 msg.msg_flags);
 190
 191                 if (signal_pending(current)) {
 192                         siginfo_t info;
 193                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 194                                 task_pid_nr(current), current->comm,
 195                                 dequeue_signal_lock(current, &current->blocked, &info));
 196                         result = -EINTR;
 197                         sock_shutdown(nbd, !send);
 198                         break;
 199                 }
 200
 201                 if (result <= 0) {
 202                         if (result == 0)
 203                                 result = -EPIPE; /* short read */
 204                         break;
 205                 }
 206                 size -= result;
 207                 buf += result;
 208         } while (size > 0);
 209
 210         sigprocmask(SIG_SETMASK, &oldset, NULL);
 211         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 212
 213         return result;
 214 }
 215
 216 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
 217                 int flags)
 218 {
 219         int result;
 220         void *kaddr = kmap(bvec->bv_page);
 221         result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
 222                            bvec->bv_len, flags);
 223         kunmap(bvec->bv_page);
 224         return result;
 225 }
 226
 227 /* always call with the tx_lock held */
 228 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
 229 {
 230         int result, flags;
 231         struct nbd_request request;
 232         unsigned long size = blk_rq_bytes(req);
 233
 234         memset(&request, 0, sizeof(request));
 235         request.magic = htonl(NBD_REQUEST_MAGIC);
 236         request.type = htonl(nbd_cmd(req));
 237
 238         if (nbd_cmd(req) != NBD_CMD_FLUSH && nbd_cmd(req) != NBD_CMD_DISC) {
 239                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 240                 request.len = htonl(size);
 241         }
 242         memcpy(request.handle, &req, sizeof(req));
 243
 244         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 245                 req, nbdcmd_to_ascii(nbd_cmd(req)),
 246                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 247         result = sock_xmit(nbd, 1, &request, sizeof(request),
 248                         (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
 249         if (result <= 0) {
 250                 dev_err(disk_to_dev(nbd->disk),
 251                         "Send control failed (result %d)\n", result);
 252                 goto error_out;
 253         }
 254
 255         if (nbd_cmd(req) == NBD_CMD_WRITE) {
 256                 struct req_iterator iter;
 257                 struct bio_vec bvec;
 258                 /*
 259                  * we are really probing at internals to determine
 260                  * whether to set MSG_MORE or not...
 261                  */
 262                 rq_for_each_segment(bvec, req, iter) {
 263                         flags = 0;
 264                         if (!rq_iter_last(bvec, iter))
 265                                 flags = MSG_MORE;
 266                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 267                                 req, bvec.bv_len);
 268                         result = sock_send_bvec(nbd, &bvec, flags);
 269                         if (result <= 0) {
 270                                 dev_err(disk_to_dev(nbd->disk),
 271                                         "Send data failed (result %d)\n",
 272                                         result);
 273                                 goto error_out;
 274                         }
 275                 }
 276         }
 277         return 0;
 278
 279 error_out:
 280         return -EIO;
 281 }
 282
 283 static struct request *nbd_find_request(struct nbd_device *nbd,
 284                                         struct request *xreq)
 285 {
 286         struct request *req, *tmp;
 287         int err;
 288
 289         err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
 290         if (unlikely(err))
 291                 goto out;
 292
 293         spin_lock(&nbd->queue_lock);
 294         list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
 295                 if (req != xreq)
 296                         continue;
 297                 list_del_init(&req->queuelist);
 298                 spin_unlock(&nbd->queue_lock);
 299                 return req;
 300         }
 301         spin_unlock(&nbd->queue_lock);
 302
 303         err = -ENOENT;
 304
 305 out:
 306         return ERR_PTR(err);
 307 }
 308
 309 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
 310 {
 311         int result;
 312         void *kaddr = kmap(bvec->bv_page);
 313         result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 314                         MSG_WAITALL);
 315         kunmap(bvec->bv_page);
 316         return result;
 317 }
 318
 319 /* NULL returned = something went wrong, inform userspace */
 320 static struct request *nbd_read_stat(struct nbd_device *nbd)
 321 {
 322         int result;
 323         struct nbd_reply reply;
 324         struct request *req;
 325
 326         reply.magic = 0;
 327         result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
 328         if (result <= 0) {
 329                 dev_err(disk_to_dev(nbd->disk),
 330                         "Receive control failed (result %d)\n", result);
 331                 goto harderror;
 332         }
 333
 334         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 335                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 336                                 (unsigned long)ntohl(reply.magic));
 337                 result = -EPROTO;
 338                 goto harderror;
 339         }
 340
 341         req = nbd_find_request(nbd, *(struct request **)reply.handle);
 342         if (IS_ERR(req)) {
 343                 result = PTR_ERR(req);
 344                 if (result != -ENOENT)
 345                         goto harderror;
 346
 347                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
 348                         reply.handle);
 349                 result = -EBADR;
 350                 goto harderror;
 351         }
 352
 353         if (ntohl(reply.error)) {
 354                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 355                         ntohl(reply.error));
 356                 req->errors++;
 357                 return req;
 358         }
 359
 360         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 361         if (nbd_cmd(req) == NBD_CMD_READ) {
 362                 struct req_iterator iter;
 363                 struct bio_vec bvec;
 364
 365                 rq_for_each_segment(bvec, req, iter) {
 366                         result = sock_recv_bvec(nbd, &bvec);
 367                         if (result <= 0) {
 368                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 369                                         result);
 370                                 req->errors++;
 371                                 return req;
 372                         }
 373                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 374                                 req, bvec.bv_len);
 375                 }
 376         }
 377         return req;
 378 harderror:
 379         nbd->harderror = result;
 380         return NULL;
 381 }
 382
 383 static ssize_t pid_show(struct device *dev,
 384                         struct device_attribute *attr, char *buf)
 385 {
 386         struct gendisk *disk = dev_to_disk(dev);
 387
 388         return sprintf(buf, "%ld\n",
 389                 (long) ((struct nbd_device *)disk->private_data)->pid);
 390 }
 391
 392 static struct device_attribute pid_attr = {
 393         .attr = { .name = "pid", .mode = S_IRUGO},
 394         .show = pid_show,
 395 };
 396
 397 static int nbd_do_it(struct nbd_device *nbd)
 398 {
 399         struct request *req;
 400         int ret;
 401
 402         BUG_ON(nbd->magic != NBD_MAGIC);
 403
 404         sk_set_memalloc(nbd->sock->sk);
 405         nbd->pid = task_pid_nr(current);
 406         ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 407         if (ret) {
 408                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 409                 nbd->pid = 0;
 410                 return ret;
 411         }
 412
 413         while ((req = nbd_read_stat(nbd)) != NULL)
 414                 nbd_end_request(nbd, req);
 415
 416         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 417         nbd->pid = 0;
 418         return 0;
 419 }
 420
 421 static void nbd_clear_que(struct nbd_device *nbd)
 422 {
 423         struct request *req;
 424
 425         BUG_ON(nbd->magic != NBD_MAGIC);
 426
 427         /*
 428          * Because we have set nbd->sock to NULL under the tx_lock, all
 429          * modifications to the list must have completed by now.  For
 430          * the same reason, the active_req must be NULL.
 431          *
 432          * As a consequence, we don't need to take the spin lock while
 433          * purging the list here.
 434          */
 435         BUG_ON(nbd->sock);
 436         BUG_ON(nbd->active_req);
 437
 438         while (!list_empty(&nbd->queue_head)) {
 439                 req = list_entry(nbd->queue_head.next, struct request,
 440                                  queuelist);
 441                 list_del_init(&req->queuelist);
 442                 req->errors++;
 443                 nbd_end_request(nbd, req);
 444         }
 445
 446         while (!list_empty(&nbd->waiting_queue)) {
 447                 req = list_entry(nbd->waiting_queue.next, struct request,
 448                                  queuelist);
 449                 list_del_init(&req->queuelist);
 450                 req->errors++;
 451                 nbd_end_request(nbd, req);
 452         }
 453 }
 454
 455
 456 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
 457 {
 458         if (req->cmd_type != REQ_TYPE_FS)
 459                 goto error_out;
 460
 461         nbd_cmd(req) = NBD_CMD_READ;
 462         if (rq_data_dir(req) == WRITE) {
 463                 if ((req->cmd_flags & REQ_DISCARD)) {
 464                         WARN_ON(!(nbd->flags & NBD_FLAG_SEND_TRIM));
 465                         nbd_cmd(req) = NBD_CMD_TRIM;
 466                 } else
 467                         nbd_cmd(req) = NBD_CMD_WRITE;
 468                 if (nbd->flags & NBD_FLAG_READ_ONLY) {
 469                         dev_err(disk_to_dev(nbd->disk),
 470                                 "Write on read-only\n");
 471                         goto error_out;
 472                 }
 473         }
 474
 475         if (req->cmd_flags & REQ_FLUSH) {
 476                 BUG_ON(unlikely(blk_rq_sectors(req)));
 477                 nbd_cmd(req) = NBD_CMD_FLUSH;
 478         }
 479
 480         req->errors = 0;
 481
 482         mutex_lock(&nbd->tx_lock);
 483         if (unlikely(!nbd->sock)) {
 484                 mutex_unlock(&nbd->tx_lock);
 485                 dev_err(disk_to_dev(nbd->disk),
 486                         "Attempted send on closed socket\n");
 487                 goto error_out;
 488         }
 489
 490         nbd->active_req = req;
 491
 492         if (nbd_send_req(nbd, req) != 0) {
 493                 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
 494                 req->errors++;
 495                 nbd_end_request(nbd, req);
 496         } else {
 497                 spin_lock(&nbd->queue_lock);
 498                 list_add_tail(&req->queuelist, &nbd->queue_head);
 499                 spin_unlock(&nbd->queue_lock);
 500         }
 501
 502         nbd->active_req = NULL;
 503         mutex_unlock(&nbd->tx_lock);
 504         wake_up_all(&nbd->active_wq);
 505
 506         return;
 507
 508 error_out:
 509         req->errors++;
 510         nbd_end_request(nbd, req);
 511 }
 512
 513 static int nbd_thread(void *data)
 514 {
 515         struct nbd_device *nbd = data;
 516         struct request *req;
 517
 518         set_user_nice(current, MIN_NICE);
 519         while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
 520                 /* wait for something to do */
 521                 wait_event_interruptible(nbd->waiting_wq,
 522                                          kthread_should_stop() ||
 523                                          !list_empty(&nbd->waiting_queue));
 524
 525                 /* extract request */
 526                 if (list_empty(&nbd->waiting_queue))
 527                         continue;
 528
 529                 spin_lock_irq(&nbd->queue_lock);
 530                 req = list_entry(nbd->waiting_queue.next, struct request,
 531                                  queuelist);
 532                 list_del_init(&req->queuelist);
 533                 spin_unlock_irq(&nbd->queue_lock);
 534
 535                 /* handle request */
 536                 nbd_handle_req(nbd, req);
 537         }
 538         return 0;
 539 }
 540
 541 /*
 542  * We always wait for result of write, for now. It would be nice to make it optional
 543  * in future
 544  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
 545  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 546  */
 547
 548 static void do_nbd_request(struct request_queue *q)
 549                 __releases(q->queue_lock) __acquires(q->queue_lock)
 550 {
 551         struct request *req;
 552
 553         while ((req = blk_fetch_request(q)) != NULL) {
 554                 struct nbd_device *nbd;
 555
 556                 spin_unlock_irq(q->queue_lock);
 557
 558                 nbd = req->rq_disk->private_data;
 559
 560                 BUG_ON(nbd->magic != NBD_MAGIC);
 561
 562                 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
 563                         req, req->cmd_type);
 564
 565                 if (unlikely(!nbd->sock)) {
 566                         dev_err(disk_to_dev(nbd->disk),
 567                                 "Attempted send on closed socket\n");
 568                         req->errors++;
 569                         nbd_end_request(nbd, req);
 570                         spin_lock_irq(q->queue_lock);
 571                         continue;
 572                 }
 573
 574                 spin_lock_irq(&nbd->queue_lock);
 575                 list_add_tail(&req->queuelist, &nbd->waiting_queue);
 576                 spin_unlock_irq(&nbd->queue_lock);
 577
 578                 wake_up(&nbd->waiting_wq);
 579
 580                 spin_lock_irq(q->queue_lock);
 581         }
 582 }
 583
 584 /* Must be called with tx_lock held */
 585
 586 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 587                        unsigned int cmd, unsigned long arg)
 588 {
 589         switch (cmd) {
 590         case NBD_DISCONNECT: {
 591                 struct request sreq;
 592
 593                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 594                 if (!nbd->sock)
 595                         return -EINVAL;
 596
 597                 mutex_unlock(&nbd->tx_lock);
 598                 fsync_bdev(bdev);
 599                 mutex_lock(&nbd->tx_lock);
 600                 blk_rq_init(NULL, &sreq);
 601                 sreq.cmd_type = REQ_TYPE_SPECIAL;
 602                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 603
 604                 /* Check again after getting mutex back.  */
 605                 if (!nbd->sock)
 606                         return -EINVAL;
 607
 608                 nbd->disconnect = 1;
 609
 610                 nbd_send_req(nbd, &sreq);
 611                 return 0;
 612         }
 613
 614         case NBD_CLEAR_SOCK: {
 615                 struct socket *sock = nbd->sock;
 616                 nbd->sock = NULL;
 617                 nbd_clear_que(nbd);
 618                 BUG_ON(!list_empty(&nbd->queue_head));
 619                 BUG_ON(!list_empty(&nbd->waiting_queue));
 620                 kill_bdev(bdev);
 621                 if (sock)
 622                         sockfd_put(sock);
 623                 return 0;
 624         }
 625
 626         case NBD_SET_SOCK: {
 627                 struct socket *sock;
 628                 int err;
 629                 if (nbd->sock)
 630                         return -EBUSY;
 631                 sock = sockfd_lookup(arg, &err);
 632                 if (sock) {
 633                         nbd->sock = sock;
 634                         if (max_part > 0)
 635                                 bdev->bd_invalidated = 1;
 636                         nbd->disconnect = 0; /* we're connected now */
 637                         return 0;
 638                 }
 639                 return -EINVAL;
 640         }
 641
 642         case NBD_SET_BLKSIZE:
 643                 nbd->blksize = arg;
 644                 nbd->bytesize &= ~(nbd->blksize-1);
 645                 bdev->bd_inode->i_size = nbd->bytesize;
 646                 set_blocksize(bdev, nbd->blksize);
 647                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 648                 return 0;
 649
 650         case NBD_SET_SIZE:
 651                 nbd->bytesize = arg & ~(nbd->blksize-1);
 652                 bdev->bd_inode->i_size = nbd->bytesize;
 653                 set_blocksize(bdev, nbd->blksize);
 654                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 655                 return 0;
 656
 657         case NBD_SET_TIMEOUT:
 658                 nbd->xmit_timeout = arg * HZ;
 659                 return 0;
 660
 661         case NBD_SET_FLAGS:
 662                 nbd->flags = arg;
 663                 return 0;
 664
 665         case NBD_SET_SIZE_BLOCKS:
 666                 nbd->bytesize = ((u64) arg) * nbd->blksize;
 667                 bdev->bd_inode->i_size = nbd->bytesize;
 668                 set_blocksize(bdev, nbd->blksize);
 669                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 670                 return 0;
 671
 672         case NBD_DO_IT: {
 673                 struct task_struct *thread;
 674                 struct socket *sock;
 675                 int error;
 676
 677                 if (nbd->pid)
 678                         return -EBUSY;
 679                 if (!nbd->sock)
 680                         return -EINVAL;
 681
 682                 mutex_unlock(&nbd->tx_lock);
 683
 684                 if (nbd->flags & NBD_FLAG_READ_ONLY)
 685                         set_device_ro(bdev, true);
 686                 if (nbd->flags & NBD_FLAG_SEND_TRIM)
 687                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
 688                                 nbd->disk->queue);
 689                 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
 690                         blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
 691                 else
 692                         blk_queue_flush(nbd->disk->queue, 0);
 693
 694                 thread = kthread_run(nbd_thread, nbd, "%s",
 695                                      nbd->disk->disk_name);
 696                 if (IS_ERR(thread)) {
 697                         mutex_lock(&nbd->tx_lock);
 698                         return PTR_ERR(thread);
 699                 }
 700
 701                 error = nbd_do_it(nbd);
 702                 kthread_stop(thread);
 703
 704                 mutex_lock(&nbd->tx_lock);
 705                 if (error)
 706                         return error;
 707                 sock_shutdown(nbd, 0);
 708                 sock = nbd->sock;
 709                 nbd->sock = NULL;
 710                 nbd_clear_que(nbd);
 711                 dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
 712                 kill_bdev(bdev);
 713                 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 714                 set_device_ro(bdev, false);
 715                 if (sock)
 716                         sockfd_put(sock);
 717                 nbd->flags = 0;
 718                 nbd->bytesize = 0;
 719                 bdev->bd_inode->i_size = 0;
 720                 set_capacity(nbd->disk, 0);
 721                 if (max_part > 0)
 722                         ioctl_by_bdev(bdev, BLKRRPART, 0);
 723                 if (nbd->disconnect) /* user requested, ignore socket errors */
 724                         return 0;
 725                 return nbd->harderror;
 726         }
 727
 728         case NBD_CLEAR_QUE:
 729                 /*
 730                  * This is for compatibility only.  The queue is always cleared
 731                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 732                  */
 733                 return 0;
 734
 735         case NBD_PRINT_DEBUG:
 736                 dev_info(disk_to_dev(nbd->disk),
 737                         "next = %p, prev = %p, head = %p\n",
 738                         nbd->queue_head.next, nbd->queue_head.prev,
 739                         &nbd->queue_head);
 740                 return 0;
 741         }
 742         return -ENOTTY;
 743 }
 744
 745 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 746                      unsigned int cmd, unsigned long arg)
 747 {
 748         struct nbd_device *nbd = bdev->bd_disk->private_data;
 749         int error;
 750
 751         if (!capable(CAP_SYS_ADMIN))
 752                 return -EPERM;
 753
 754         BUG_ON(nbd->magic != NBD_MAGIC);
 755
 756         mutex_lock(&nbd->tx_lock);
 757         error = __nbd_ioctl(bdev, nbd, cmd, arg);
 758         mutex_unlock(&nbd->tx_lock);
 759
 760         return error;
 761 }
 762
 763 static const struct block_device_operations nbd_fops =
 764 {
 765         .owner =        THIS_MODULE,
 766         .ioctl =        nbd_ioctl,
 767 };
 768
 769 /*
 770  * And here should be modules and kernel interface
 771  *  (Just smiley confuses emacs :-)
 772  */
 773
 774 static int __init nbd_init(void)
 775 {
 776         int err = -ENOMEM;
 777         int i;
 778         int part_shift;
 779
 780         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 781
 782         if (max_part < 0) {
 783                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
 784                 return -EINVAL;
 785         }
 786
 787         part_shift = 0;
 788         if (max_part > 0) {
 789                 part_shift = fls(max_part);
 790
 791                 /*
 792                  * Adjust max_part according to part_shift as it is exported
 793                  * to user space so that user can know the max number of
 794                  * partition kernel should be able to manage.
 795                  *
 796                  * Note that -1 is required because partition 0 is reserved
 797                  * for the whole disk.
 798                  */
 799                 max_part = (1UL << part_shift) - 1;
 800         }
 801
 802         if ((1UL << part_shift) > DISK_MAX_PARTS)
 803                 return -EINVAL;
 804
 805         if (nbds_max > 1UL << (MINORBITS - part_shift))
 806                 return -EINVAL;
 807
 808         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
 809         if (!nbd_dev)
 810                 return -ENOMEM;
 811
 812         for (i = 0; i < nbds_max; i++) {
 813                 struct gendisk *disk = alloc_disk(1 << part_shift);
 814                 if (!disk)
 815                         goto out;
 816                 nbd_dev[i].disk = disk;
 817                 /*
 818                  * The new linux 2.5 block layer implementation requires
 819                  * every gendisk to have its very own request_queue struct.
 820                  * These structs are big so we dynamically allocate them.
 821                  */
 822                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 823                 if (!disk->queue) {
 824                         put_disk(disk);
 825                         goto out;
 826                 }
 827                 /*
 828                  * Tell the block layer that we are not a rotational device
 829                  */
 830                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
 831                 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
 832                 disk->queue->limits.discard_granularity = 512;
 833                 disk->queue->limits.max_discard_sectors = UINT_MAX;
 834                 disk->queue->limits.discard_zeroes_data = 0;
 835                 blk_queue_max_hw_sectors(disk->queue, 65536);
 836                 disk->queue->limits.max_sectors = 256;
 837         }
 838
 839         if (register_blkdev(NBD_MAJOR, "nbd")) {
 840                 err = -EIO;
 841                 goto out;
 842         }
 843
 844         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 845
 846         for (i = 0; i < nbds_max; i++) {
 847                 struct gendisk *disk = nbd_dev[i].disk;
 848                 nbd_dev[i].magic = NBD_MAGIC;
 849                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
 850                 spin_lock_init(&nbd_dev[i].queue_lock);
 851                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 852                 mutex_init(&nbd_dev[i].tx_lock);
 853                 init_waitqueue_head(&nbd_dev[i].active_wq);
 854                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
 855                 nbd_dev[i].blksize = 1024;
 856                 nbd_dev[i].bytesize = 0;
 857                 disk->major = NBD_MAJOR;
 858                 disk->first_minor = i << part_shift;
 859                 disk->fops = &nbd_fops;
 860                 disk->private_data = &nbd_dev[i];
 861                 sprintf(disk->disk_name, "nbd%d", i);
 862                 set_capacity(disk, 0);
 863                 add_disk(disk);
 864         }
 865
 866         return 0;
 867 out:
 868         while (i--) {
 869                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 870                 put_disk(nbd_dev[i].disk);
 871         }
 872         kfree(nbd_dev);
 873         return err;
 874 }
 875
 876 static void __exit nbd_cleanup(void)
 877 {
 878         int i;
 879         for (i = 0; i < nbds_max; i++) {
 880                 struct gendisk *disk = nbd_dev[i].disk;
 881                 nbd_dev[i].magic = 0;
 882                 if (disk) {
 883                         del_gendisk(disk);
 884                         blk_cleanup_queue(disk->queue);
 885                         put_disk(disk);
 886                 }
 887         }
 888         unregister_blkdev(NBD_MAJOR, "nbd");
 889         kfree(nbd_dev);
 890         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 891 }
 892
 893 module_init(nbd_init);
 894 module_exit(nbd_cleanup);
 895
 896 MODULE_DESCRIPTION("Network Block Device");
 897 MODULE_LICENSE("GPL");
 898
 899 module_param(nbds_max, int, 0444);
 900 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
 901 module_param(max_part, int, 0444);
 902 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");