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         struct socket * sock;   /* If == NULL, device is not ready, yet */
  45         int magic;
  46
  47         spinlock_t queue_lock;
  48         struct list_head queue_head;    /* Requests waiting result */
  49         struct request *active_req;
  50         wait_queue_head_t active_wq;
  51         struct list_head waiting_queue; /* Requests to be sent */
  52         wait_queue_head_t waiting_wq;
  53
  54         struct mutex tx_lock;
  55         struct gendisk *disk;
  56         int blksize;
  57         loff_t bytesize;
  58         int xmit_timeout;
  59         int disconnect; /* a disconnect has been requested by user */
  60
  61         struct timer_list timeout_timer;
  62         struct task_struct *task_recv;
  63         struct task_struct *task_send;
  64 };
  65
  66 #define NBD_MAGIC 0x68797548
  67
  68 static unsigned int nbds_max = 16;
  69 static struct nbd_device *nbd_dev;
  70 static int max_part;
  71
  72 /*
  73  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  74  * 1. Each NIC is essentially a synchronization point for all servers
  75  *    accessed through that NIC so there's no need to have more locks
  76  *    than NICs anyway.
  77  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  78  *    down each lock to the point where they're actually slower than just
  79  *    a single lock.
  80  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  81  */
  82 static DEFINE_SPINLOCK(nbd_lock);
  83
  84 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
  85 {
  86         return disk_to_dev(nbd->disk);
  87 }
  88
  89 static const char *nbdcmd_to_ascii(int cmd)
  90 {
  91         switch (cmd) {
  92         case  NBD_CMD_READ: return "read";
  93         case NBD_CMD_WRITE: return "write";
  94         case  NBD_CMD_DISC: return "disconnect";
  95         case NBD_CMD_FLUSH: return "flush";
  96         case  NBD_CMD_TRIM: return "trim/discard";
  97         }
  98         return "invalid";
  99 }
 100
 101 static void nbd_end_request(struct nbd_device *nbd, struct request *req)
 102 {
 103         int error = req->errors ? -EIO : 0;
 104         struct request_queue *q = req->q;
 105         unsigned long flags;
 106
 107         dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
 108                 error ? "failed" : "done");
 109
 110         spin_lock_irqsave(q->queue_lock, flags);
 111         __blk_end_request_all(req, error);
 112         spin_unlock_irqrestore(q->queue_lock, flags);
 113 }
 114
 115 /*
 116  * Forcibly shutdown the socket causing all listeners to error
 117  */
 118 static void sock_shutdown(struct nbd_device *nbd)
 119 {
 120         if (!nbd->sock)
 121                 return;
 122
 123         dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
 124         kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
 125         nbd->sock = NULL;
 126         del_timer_sync(&nbd->timeout_timer);
 127 }
 128
 129 static void nbd_xmit_timeout(unsigned long arg)
 130 {
 131         struct nbd_device *nbd = (struct nbd_device *)arg;
 132         struct task_struct *task;
 133
 134         if (list_empty(&nbd->queue_head))
 135                 return;
 136
 137         nbd->disconnect = 1;
 138
 139         task = READ_ONCE(nbd->task_recv);
 140         if (task)
 141                 force_sig(SIGKILL, task);
 142
 143         task = READ_ONCE(nbd->task_send);
 144         if (task)
 145                 force_sig(SIGKILL, nbd->task_send);
 146
 147         dev_err(nbd_to_dev(nbd), "Connection timed out, killed receiver and sender, shutting down connection\n");
 148 }
 149
 150 /*
 151  *  Send or receive packet.
 152  */
 153 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
 154                 int msg_flags)
 155 {
 156         struct socket *sock = nbd->sock;
 157         int result;
 158         struct msghdr msg;
 159         struct kvec iov;
 160         sigset_t blocked, oldset;
 161         unsigned long pflags = current->flags;
 162
 163         if (unlikely(!sock)) {
 164                 dev_err(disk_to_dev(nbd->disk),
 165                         "Attempted %s on closed socket in sock_xmit\n",
 166                         (send ? "send" : "recv"));
 167                 return -EINVAL;
 168         }
 169
 170         /* Allow interception of SIGKILL only
 171          * Don't allow other signals to interrupt the transmission */
 172         siginitsetinv(&blocked, sigmask(SIGKILL));
 173         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 174
 175         current->flags |= PF_MEMALLOC;
 176         do {
 177                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 178                 iov.iov_base = buf;
 179                 iov.iov_len = size;
 180                 msg.msg_name = NULL;
 181                 msg.msg_namelen = 0;
 182                 msg.msg_control = NULL;
 183                 msg.msg_controllen = 0;
 184                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 185
 186                 if (send)
 187                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 188                 else
 189                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 190                                                 msg.msg_flags);
 191
 192                 if (result <= 0) {
 193                         if (result == 0)
 194                                 result = -EPIPE; /* short read */
 195                         break;
 196                 }
 197                 size -= result;
 198                 buf += result;
 199         } while (size > 0);
 200
 201         sigprocmask(SIG_SETMASK, &oldset, NULL);
 202         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 203
 204         if (!send && nbd->xmit_timeout)
 205                 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
 206
 207         return result;
 208 }
 209
 210 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
 211                 int flags)
 212 {
 213         int result;
 214         void *kaddr = kmap(bvec->bv_page);
 215         result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
 216                            bvec->bv_len, flags);
 217         kunmap(bvec->bv_page);
 218         return result;
 219 }
 220
 221 /* always call with the tx_lock held */
 222 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
 223 {
 224         int result, flags;
 225         struct nbd_request request;
 226         unsigned long size = blk_rq_bytes(req);
 227         u32 type;
 228
 229         if (req->cmd_type == REQ_TYPE_DRV_PRIV)
 230                 type = NBD_CMD_DISC;
 231         else if (req->cmd_flags & REQ_DISCARD)
 232                 type = NBD_CMD_TRIM;
 233         else if (req->cmd_flags & REQ_FLUSH)
 234                 type = NBD_CMD_FLUSH;
 235         else if (rq_data_dir(req) == WRITE)
 236                 type = NBD_CMD_WRITE;
 237         else
 238                 type = NBD_CMD_READ;
 239
 240         memset(&request, 0, sizeof(request));
 241         request.magic = htonl(NBD_REQUEST_MAGIC);
 242         request.type = htonl(type);
 243         if (type != NBD_CMD_FLUSH && type != NBD_CMD_DISC) {
 244                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 245                 request.len = htonl(size);
 246         }
 247         memcpy(request.handle, &req, sizeof(req));
 248
 249         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 250                 req, nbdcmd_to_ascii(type),
 251                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 252         result = sock_xmit(nbd, 1, &request, sizeof(request),
 253                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
 254         if (result <= 0) {
 255                 dev_err(disk_to_dev(nbd->disk),
 256                         "Send control failed (result %d)\n", result);
 257                 return -EIO;
 258         }
 259
 260         if (type == NBD_CMD_WRITE) {
 261                 struct req_iterator iter;
 262                 struct bio_vec bvec;
 263                 /*
 264                  * we are really probing at internals to determine
 265                  * whether to set MSG_MORE or not...
 266                  */
 267                 rq_for_each_segment(bvec, req, iter) {
 268                         flags = 0;
 269                         if (!rq_iter_last(bvec, iter))
 270                                 flags = MSG_MORE;
 271                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 272                                 req, bvec.bv_len);
 273                         result = sock_send_bvec(nbd, &bvec, flags);
 274                         if (result <= 0) {
 275                                 dev_err(disk_to_dev(nbd->disk),
 276                                         "Send data failed (result %d)\n",
 277                                         result);
 278                                 return -EIO;
 279                         }
 280                 }
 281         }
 282         return 0;
 283 }
 284
 285 static struct request *nbd_find_request(struct nbd_device *nbd,
 286                                         struct request *xreq)
 287 {
 288         struct request *req, *tmp;
 289         int err;
 290
 291         err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
 292         if (unlikely(err))
 293                 return ERR_PTR(err);
 294
 295         spin_lock(&nbd->queue_lock);
 296         list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
 297                 if (req != xreq)
 298                         continue;
 299                 list_del_init(&req->queuelist);
 300                 spin_unlock(&nbd->queue_lock);
 301                 return req;
 302         }
 303         spin_unlock(&nbd->queue_lock);
 304
 305         return ERR_PTR(-ENOENT);
 306 }
 307
 308 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
 309 {
 310         int result;
 311         void *kaddr = kmap(bvec->bv_page);
 312         result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 313                         MSG_WAITALL);
 314         kunmap(bvec->bv_page);
 315         return result;
 316 }
 317
 318 /* NULL returned = something went wrong, inform userspace */
 319 static struct request *nbd_read_stat(struct nbd_device *nbd)
 320 {
 321         int result;
 322         struct nbd_reply reply;
 323         struct request *req;
 324
 325         reply.magic = 0;
 326         result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
 327         if (result <= 0) {
 328                 dev_err(disk_to_dev(nbd->disk),
 329                         "Receive control failed (result %d)\n", result);
 330                 return ERR_PTR(result);
 331         }
 332
 333         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 334                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 335                                 (unsigned long)ntohl(reply.magic));
 336                 return ERR_PTR(-EPROTO);
 337         }
 338
 339         req = nbd_find_request(nbd, *(struct request **)reply.handle);
 340         if (IS_ERR(req)) {
 341                 result = PTR_ERR(req);
 342                 if (result != -ENOENT)
 343                         return ERR_PTR(result);
 344
 345                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
 346                         reply.handle);
 347                 return ERR_PTR(-EBADR);
 348         }
 349
 350         if (ntohl(reply.error)) {
 351                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 352                         ntohl(reply.error));
 353                 req->errors++;
 354                 return req;
 355         }
 356
 357         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 358         if (rq_data_dir(req) != WRITE) {
 359                 struct req_iterator iter;
 360                 struct bio_vec bvec;
 361
 362                 rq_for_each_segment(bvec, req, iter) {
 363                         result = sock_recv_bvec(nbd, &bvec);
 364                         if (result <= 0) {
 365                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 366                                         result);
 367                                 req->errors++;
 368                                 return req;
 369                         }
 370                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 371                                 req, bvec.bv_len);
 372                 }
 373         }
 374         return req;
 375 }
 376
 377 static ssize_t pid_show(struct device *dev,
 378                         struct device_attribute *attr, char *buf)
 379 {
 380         struct gendisk *disk = dev_to_disk(dev);
 381         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 382
 383         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 384 }
 385
 386 static struct device_attribute pid_attr = {
 387         .attr = { .name = "pid", .mode = S_IRUGO},
 388         .show = pid_show,
 389 };
 390
 391 static int nbd_do_it(struct nbd_device *nbd)
 392 {
 393         struct request *req;
 394         int ret;
 395
 396         BUG_ON(nbd->magic != NBD_MAGIC);
 397
 398         sk_set_memalloc(nbd->sock->sk);
 399
 400         nbd->task_recv = current;
 401
 402         ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 403         if (ret) {
 404                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 405                 nbd->task_recv = NULL;
 406                 return ret;
 407         }
 408
 409         while (1) {
 410                 req = nbd_read_stat(nbd);
 411                 if (IS_ERR(req)) {
 412                         ret = PTR_ERR(req);
 413                         break;
 414                 }
 415
 416                 nbd_end_request(nbd, req);
 417         }
 418
 419         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 420
 421         nbd->task_recv = NULL;
 422
 423         if (signal_pending(current)) {
 424                 siginfo_t info;
 425
 426                 ret = dequeue_signal_lock(current, &current->blocked, &info);
 427                 dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
 428                          task_pid_nr(current), current->comm, ret);
 429                 mutex_lock(&nbd->tx_lock);
 430                 sock_shutdown(nbd);
 431                 mutex_unlock(&nbd->tx_lock);
 432                 ret = -ETIMEDOUT;
 433         }
 434
 435         return ret;
 436 }
 437
 438 static void nbd_clear_que(struct nbd_device *nbd)
 439 {
 440         struct request *req;
 441
 442         BUG_ON(nbd->magic != NBD_MAGIC);
 443
 444         /*
 445          * Because we have set nbd->sock to NULL under the tx_lock, all
 446          * modifications to the list must have completed by now.  For
 447          * the same reason, the active_req must be NULL.
 448          *
 449          * As a consequence, we don't need to take the spin lock while
 450          * purging the list here.
 451          */
 452         BUG_ON(nbd->sock);
 453         BUG_ON(nbd->active_req);
 454
 455         while (!list_empty(&nbd->queue_head)) {
 456                 req = list_entry(nbd->queue_head.next, struct request,
 457                                  queuelist);
 458                 list_del_init(&req->queuelist);
 459                 req->errors++;
 460                 nbd_end_request(nbd, req);
 461         }
 462
 463         while (!list_empty(&nbd->waiting_queue)) {
 464                 req = list_entry(nbd->waiting_queue.next, struct request,
 465                                  queuelist);
 466                 list_del_init(&req->queuelist);
 467                 req->errors++;
 468                 nbd_end_request(nbd, req);
 469         }
 470         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 471 }
 472
 473
 474 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
 475 {
 476         if (req->cmd_type != REQ_TYPE_FS)
 477                 goto error_out;
 478
 479         if (rq_data_dir(req) == WRITE &&
 480             (nbd->flags & NBD_FLAG_READ_ONLY)) {
 481                 dev_err(disk_to_dev(nbd->disk),
 482                         "Write on read-only\n");
 483                 goto error_out;
 484         }
 485
 486         req->errors = 0;
 487
 488         mutex_lock(&nbd->tx_lock);
 489         if (unlikely(!nbd->sock)) {
 490                 mutex_unlock(&nbd->tx_lock);
 491                 dev_err(disk_to_dev(nbd->disk),
 492                         "Attempted send on closed socket\n");
 493                 goto error_out;
 494         }
 495
 496         nbd->active_req = req;
 497
 498         if (nbd->xmit_timeout && list_empty_careful(&nbd->queue_head))
 499                 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
 500
 501         if (nbd_send_req(nbd, req) != 0) {
 502                 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
 503                 req->errors++;
 504                 nbd_end_request(nbd, req);
 505         } else {
 506                 spin_lock(&nbd->queue_lock);
 507                 list_add_tail(&req->queuelist, &nbd->queue_head);
 508                 spin_unlock(&nbd->queue_lock);
 509         }
 510
 511         nbd->active_req = NULL;
 512         mutex_unlock(&nbd->tx_lock);
 513         wake_up_all(&nbd->active_wq);
 514
 515         return;
 516
 517 error_out:
 518         req->errors++;
 519         nbd_end_request(nbd, req);
 520 }
 521
 522 static int nbd_thread(void *data)
 523 {
 524         struct nbd_device *nbd = data;
 525         struct request *req;
 526
 527         nbd->task_send = current;
 528
 529         set_user_nice(current, MIN_NICE);
 530         while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
 531                 /* wait for something to do */
 532                 wait_event_interruptible(nbd->waiting_wq,
 533                                          kthread_should_stop() ||
 534                                          !list_empty(&nbd->waiting_queue));
 535
 536                 if (signal_pending(current)) {
 537                         siginfo_t info;
 538                         int ret;
 539
 540                         ret = dequeue_signal_lock(current, &current->blocked,
 541                                                   &info);
 542                         dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
 543                                  task_pid_nr(current), current->comm, ret);
 544                         mutex_lock(&nbd->tx_lock);
 545                         sock_shutdown(nbd);
 546                         mutex_unlock(&nbd->tx_lock);
 547                         break;
 548                 }
 549
 550                 /* extract request */
 551                 if (list_empty(&nbd->waiting_queue))
 552                         continue;
 553
 554                 spin_lock_irq(&nbd->queue_lock);
 555                 req = list_entry(nbd->waiting_queue.next, struct request,
 556                                  queuelist);
 557                 list_del_init(&req->queuelist);
 558                 spin_unlock_irq(&nbd->queue_lock);
 559
 560                 /* handle request */
 561                 nbd_handle_req(nbd, req);
 562         }
 563
 564         nbd->task_send = NULL;
 565
 566         return 0;
 567 }
 568
 569 /*
 570  * We always wait for result of write, for now. It would be nice to make it optional
 571  * in future
 572  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
 573  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 574  */
 575
 576 static void do_nbd_request(struct request_queue *q)
 577                 __releases(q->queue_lock) __acquires(q->queue_lock)
 578 {
 579         struct request *req;
 580
 581         while ((req = blk_fetch_request(q)) != NULL) {
 582                 struct nbd_device *nbd;
 583
 584                 spin_unlock_irq(q->queue_lock);
 585
 586                 nbd = req->rq_disk->private_data;
 587
 588                 BUG_ON(nbd->magic != NBD_MAGIC);
 589
 590                 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
 591                         req, req->cmd_type);
 592
 593                 if (unlikely(!nbd->sock)) {
 594                         dev_err(disk_to_dev(nbd->disk),
 595                                 "Attempted send on closed socket\n");
 596                         req->errors++;
 597                         nbd_end_request(nbd, req);
 598                         spin_lock_irq(q->queue_lock);
 599                         continue;
 600                 }
 601
 602                 spin_lock_irq(&nbd->queue_lock);
 603                 list_add_tail(&req->queuelist, &nbd->waiting_queue);
 604                 spin_unlock_irq(&nbd->queue_lock);
 605
 606                 wake_up(&nbd->waiting_wq);
 607
 608                 spin_lock_irq(q->queue_lock);
 609         }
 610 }
 611
 612 /* Must be called with tx_lock held */
 613
 614 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 615                        unsigned int cmd, unsigned long arg)
 616 {
 617         switch (cmd) {
 618         case NBD_DISCONNECT: {
 619                 struct request sreq;
 620
 621                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 622                 if (!nbd->sock)
 623                         return -EINVAL;
 624
 625                 mutex_unlock(&nbd->tx_lock);
 626                 fsync_bdev(bdev);
 627                 mutex_lock(&nbd->tx_lock);
 628                 blk_rq_init(NULL, &sreq);
 629                 sreq.cmd_type = REQ_TYPE_DRV_PRIV;
 630
 631                 /* Check again after getting mutex back.  */
 632                 if (!nbd->sock)
 633                         return -EINVAL;
 634
 635                 nbd->disconnect = 1;
 636
 637                 nbd_send_req(nbd, &sreq);
 638                 return 0;
 639         }
 640
 641         case NBD_CLEAR_SOCK: {
 642                 struct socket *sock = nbd->sock;
 643                 nbd->sock = NULL;
 644                 nbd_clear_que(nbd);
 645                 BUG_ON(!list_empty(&nbd->queue_head));
 646                 BUG_ON(!list_empty(&nbd->waiting_queue));
 647                 kill_bdev(bdev);
 648                 if (sock)
 649                         sockfd_put(sock);
 650                 return 0;
 651         }
 652
 653         case NBD_SET_SOCK: {
 654                 struct socket *sock;
 655                 int err;
 656                 if (nbd->sock)
 657                         return -EBUSY;
 658                 sock = sockfd_lookup(arg, &err);
 659                 if (sock) {
 660                         nbd->sock = sock;
 661                         if (max_part > 0)
 662                                 bdev->bd_invalidated = 1;
 663                         nbd->disconnect = 0; /* we're connected now */
 664                         return 0;
 665                 }
 666                 return -EINVAL;
 667         }
 668
 669         case NBD_SET_BLKSIZE:
 670                 nbd->blksize = arg;
 671                 nbd->bytesize &= ~(nbd->blksize-1);
 672                 bdev->bd_inode->i_size = nbd->bytesize;
 673                 set_blocksize(bdev, nbd->blksize);
 674                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 675                 return 0;
 676
 677         case NBD_SET_SIZE:
 678                 nbd->bytesize = arg & ~(nbd->blksize-1);
 679                 bdev->bd_inode->i_size = nbd->bytesize;
 680                 set_blocksize(bdev, nbd->blksize);
 681                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 682                 return 0;
 683
 684         case NBD_SET_TIMEOUT:
 685                 nbd->xmit_timeout = arg * HZ;
 686                 if (arg)
 687                         mod_timer(&nbd->timeout_timer,
 688                                   jiffies + nbd->xmit_timeout);
 689                 else
 690                         del_timer_sync(&nbd->timeout_timer);
 691
 692                 return 0;
 693
 694         case NBD_SET_FLAGS:
 695                 nbd->flags = arg;
 696                 return 0;
 697
 698         case NBD_SET_SIZE_BLOCKS:
 699                 nbd->bytesize = ((u64) arg) * nbd->blksize;
 700                 bdev->bd_inode->i_size = nbd->bytesize;
 701                 set_blocksize(bdev, nbd->blksize);
 702                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 703                 return 0;
 704
 705         case NBD_DO_IT: {
 706                 struct task_struct *thread;
 707                 struct socket *sock;
 708                 int error;
 709
 710                 if (nbd->task_recv)
 711                         return -EBUSY;
 712                 if (!nbd->sock)
 713                         return -EINVAL;
 714
 715                 mutex_unlock(&nbd->tx_lock);
 716
 717                 if (nbd->flags & NBD_FLAG_READ_ONLY)
 718                         set_device_ro(bdev, true);
 719                 if (nbd->flags & NBD_FLAG_SEND_TRIM)
 720                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
 721                                 nbd->disk->queue);
 722                 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
 723                         blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
 724                 else
 725                         blk_queue_flush(nbd->disk->queue, 0);
 726
 727                 thread = kthread_run(nbd_thread, nbd, "%s",
 728                                      nbd->disk->disk_name);
 729                 if (IS_ERR(thread)) {
 730                         mutex_lock(&nbd->tx_lock);
 731                         return PTR_ERR(thread);
 732                 }
 733
 734                 error = nbd_do_it(nbd);
 735                 kthread_stop(thread);
 736
 737                 mutex_lock(&nbd->tx_lock);
 738
 739                 sock_shutdown(nbd);
 740                 sock = nbd->sock;
 741                 nbd->sock = NULL;
 742                 nbd_clear_que(nbd);
 743                 kill_bdev(bdev);
 744                 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 745                 set_device_ro(bdev, false);
 746                 if (sock)
 747                         sockfd_put(sock);
 748                 nbd->flags = 0;
 749                 nbd->bytesize = 0;
 750                 bdev->bd_inode->i_size = 0;
 751                 set_capacity(nbd->disk, 0);
 752                 if (max_part > 0)
 753                         blkdev_reread_part(bdev);
 754                 if (nbd->disconnect) /* user requested, ignore socket errors */
 755                         return 0;
 756                 return error;
 757         }
 758
 759         case NBD_CLEAR_QUE:
 760                 /*
 761                  * This is for compatibility only.  The queue is always cleared
 762                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 763                  */
 764                 return 0;
 765
 766         case NBD_PRINT_DEBUG:
 767                 dev_info(disk_to_dev(nbd->disk),
 768                         "next = %p, prev = %p, head = %p\n",
 769                         nbd->queue_head.next, nbd->queue_head.prev,
 770                         &nbd->queue_head);
 771                 return 0;
 772         }
 773         return -ENOTTY;
 774 }
 775
 776 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 777                      unsigned int cmd, unsigned long arg)
 778 {
 779         struct nbd_device *nbd = bdev->bd_disk->private_data;
 780         int error;
 781
 782         if (!capable(CAP_SYS_ADMIN))
 783                 return -EPERM;
 784
 785         BUG_ON(nbd->magic != NBD_MAGIC);
 786
 787         mutex_lock(&nbd->tx_lock);
 788         error = __nbd_ioctl(bdev, nbd, cmd, arg);
 789         mutex_unlock(&nbd->tx_lock);
 790
 791         return error;
 792 }
 793
 794 static const struct block_device_operations nbd_fops =
 795 {
 796         .owner =        THIS_MODULE,
 797         .ioctl =        nbd_ioctl,
 798 };
 799
 800 /*
 801  * And here should be modules and kernel interface
 802  *  (Just smiley confuses emacs :-)
 803  */
 804
 805 static int __init nbd_init(void)
 806 {
 807         int err = -ENOMEM;
 808         int i;
 809         int part_shift;
 810
 811         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 812
 813         if (max_part < 0) {
 814                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
 815                 return -EINVAL;
 816         }
 817
 818         part_shift = 0;
 819         if (max_part > 0) {
 820                 part_shift = fls(max_part);
 821
 822                 /*
 823                  * Adjust max_part according to part_shift as it is exported
 824                  * to user space so that user can know the max number of
 825                  * partition kernel should be able to manage.
 826                  *
 827                  * Note that -1 is required because partition 0 is reserved
 828                  * for the whole disk.
 829                  */
 830                 max_part = (1UL << part_shift) - 1;
 831         }
 832
 833         if ((1UL << part_shift) > DISK_MAX_PARTS)
 834                 return -EINVAL;
 835
 836         if (nbds_max > 1UL << (MINORBITS - part_shift))
 837                 return -EINVAL;
 838
 839         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
 840         if (!nbd_dev)
 841                 return -ENOMEM;
 842
 843         for (i = 0; i < nbds_max; i++) {
 844                 struct gendisk *disk = alloc_disk(1 << part_shift);
 845                 if (!disk)
 846                         goto out;
 847                 nbd_dev[i].disk = disk;
 848                 /*
 849                  * The new linux 2.5 block layer implementation requires
 850                  * every gendisk to have its very own request_queue struct.
 851                  * These structs are big so we dynamically allocate them.
 852                  */
 853                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 854                 if (!disk->queue) {
 855                         put_disk(disk);
 856                         goto out;
 857                 }
 858                 /*
 859                  * Tell the block layer that we are not a rotational device
 860                  */
 861                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
 862                 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
 863                 disk->queue->limits.discard_granularity = 512;
 864                 blk_queue_max_discard_sectors(disk->queue, UINT_MAX);
 865                 disk->queue->limits.discard_zeroes_data = 0;
 866                 blk_queue_max_hw_sectors(disk->queue, 65536);
 867                 disk->queue->limits.max_sectors = 256;
 868         }
 869
 870         if (register_blkdev(NBD_MAJOR, "nbd")) {
 871                 err = -EIO;
 872                 goto out;
 873         }
 874
 875         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 876
 877         for (i = 0; i < nbds_max; i++) {
 878                 struct gendisk *disk = nbd_dev[i].disk;
 879                 nbd_dev[i].magic = NBD_MAGIC;
 880                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
 881                 spin_lock_init(&nbd_dev[i].queue_lock);
 882                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 883                 mutex_init(&nbd_dev[i].tx_lock);
 884                 init_timer(&nbd_dev[i].timeout_timer);
 885                 nbd_dev[i].timeout_timer.function = nbd_xmit_timeout;
 886                 nbd_dev[i].timeout_timer.data = (unsigned long)&nbd_dev[i];
 887                 init_waitqueue_head(&nbd_dev[i].active_wq);
 888                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
 889                 nbd_dev[i].blksize = 1024;
 890                 nbd_dev[i].bytesize = 0;
 891                 disk->major = NBD_MAJOR;
 892                 disk->first_minor = i << part_shift;
 893                 disk->fops = &nbd_fops;
 894                 disk->private_data = &nbd_dev[i];
 895                 sprintf(disk->disk_name, "nbd%d", i);
 896                 set_capacity(disk, 0);
 897                 add_disk(disk);
 898         }
 899
 900         return 0;
 901 out:
 902         while (i--) {
 903                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 904                 put_disk(nbd_dev[i].disk);
 905         }
 906         kfree(nbd_dev);
 907         return err;
 908 }
 909
 910 static void __exit nbd_cleanup(void)
 911 {
 912         int i;
 913         for (i = 0; i < nbds_max; i++) {
 914                 struct gendisk *disk = nbd_dev[i].disk;
 915                 nbd_dev[i].magic = 0;
 916                 if (disk) {
 917                         del_gendisk(disk);
 918                         blk_cleanup_queue(disk->queue);
 919                         put_disk(disk);
 920                 }
 921         }
 922         unregister_blkdev(NBD_MAJOR, "nbd");
 923         kfree(nbd_dev);
 924         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 925 }
 926
 927 module_init(nbd_init);
 928 module_exit(nbd_cleanup);
 929
 930 MODULE_DESCRIPTION("Network Block Device");
 931 MODULE_LICENSE("GPL");
 932
 933 module_param(nbds_max, int, 0444);
 934 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
 935 module_param(max_part, int, 0444);
 936 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");