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         struct timer_list timeout_timer;
  64         struct task_struct *task_recv;
  65         struct task_struct *task_send;
  66 };
  67
  68 #define NBD_MAGIC 0x68797548
  69
  70 static unsigned int nbds_max = 16;
  71 static struct nbd_device *nbd_dev;
  72 static int max_part;
  73
  74 /*
  75  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  76  * 1. Each NIC is essentially a synchronization point for all servers
  77  *    accessed through that NIC so there's no need to have more locks
  78  *    than NICs anyway.
  79  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  80  *    down each lock to the point where they're actually slower than just
  81  *    a single lock.
  82  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  83  */
  84 static DEFINE_SPINLOCK(nbd_lock);
  85
  86 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
  87 {
  88         return disk_to_dev(nbd->disk);
  89 }
  90
  91 static const char *nbdcmd_to_ascii(int cmd)
  92 {
  93         switch (cmd) {
  94         case  NBD_CMD_READ: return "read";
  95         case NBD_CMD_WRITE: return "write";
  96         case  NBD_CMD_DISC: return "disconnect";
  97         case NBD_CMD_FLUSH: return "flush";
  98         case  NBD_CMD_TRIM: return "trim/discard";
  99         }
 100         return "invalid";
 101 }
 102
 103 static void nbd_end_request(struct nbd_device *nbd, struct request *req)
 104 {
 105         int error = req->errors ? -EIO : 0;
 106         struct request_queue *q = req->q;
 107         unsigned long flags;
 108
 109         dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
 110                 error ? "failed" : "done");
 111
 112         spin_lock_irqsave(q->queue_lock, flags);
 113         __blk_end_request_all(req, error);
 114         spin_unlock_irqrestore(q->queue_lock, flags);
 115 }
 116
 117 /*
 118  * Forcibly shutdown the socket causing all listeners to error
 119  */
 120 static void sock_shutdown(struct nbd_device *nbd, int lock)
 121 {
 122         if (lock)
 123                 mutex_lock(&nbd->tx_lock);
 124         if (nbd->sock) {
 125                 dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
 126                 kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
 127                 nbd->sock = NULL;
 128                 del_timer_sync(&nbd->timeout_timer);
 129         }
 130         if (lock)
 131                 mutex_unlock(&nbd->tx_lock);
 132 }
 133
 134 static void nbd_xmit_timeout(unsigned long arg)
 135 {
 136         struct nbd_device *nbd = (struct nbd_device *)arg;
 137         struct task_struct *task;
 138
 139         if (list_empty(&nbd->queue_head))
 140                 return;
 141
 142         nbd->disconnect = 1;
 143
 144         task = READ_ONCE(nbd->task_recv);
 145         if (task)
 146                 force_sig(SIGKILL, task);
 147
 148         task = READ_ONCE(nbd->task_send);
 149         if (task)
 150                 force_sig(SIGKILL, nbd->task_send);
 151
 152         dev_err(nbd_to_dev(nbd), "Connection timed out, killed receiver and sender, shutting down connection\n");
 153 }
 154
 155 /*
 156  *  Send or receive packet.
 157  */
 158 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
 159                 int msg_flags)
 160 {
 161         struct socket *sock = nbd->sock;
 162         int result;
 163         struct msghdr msg;
 164         struct kvec iov;
 165         sigset_t blocked, oldset;
 166         unsigned long pflags = current->flags;
 167
 168         if (unlikely(!sock)) {
 169                 dev_err(disk_to_dev(nbd->disk),
 170                         "Attempted %s on closed socket in sock_xmit\n",
 171                         (send ? "send" : "recv"));
 172                 return -EINVAL;
 173         }
 174
 175         /* Allow interception of SIGKILL only
 176          * Don't allow other signals to interrupt the transmission */
 177         siginitsetinv(&blocked, sigmask(SIGKILL));
 178         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 179
 180         current->flags |= PF_MEMALLOC;
 181         do {
 182                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 183                 iov.iov_base = buf;
 184                 iov.iov_len = size;
 185                 msg.msg_name = NULL;
 186                 msg.msg_namelen = 0;
 187                 msg.msg_control = NULL;
 188                 msg.msg_controllen = 0;
 189                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 190
 191                 if (send)
 192                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 193                 else
 194                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 195                                                 msg.msg_flags);
 196
 197                 if (result <= 0) {
 198                         if (result == 0)
 199                                 result = -EPIPE; /* short read */
 200                         break;
 201                 }
 202                 size -= result;
 203                 buf += result;
 204         } while (size > 0);
 205
 206         sigprocmask(SIG_SETMASK, &oldset, NULL);
 207         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 208
 209         if (!send && nbd->xmit_timeout)
 210                 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
 211
 212         return result;
 213 }
 214
 215 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
 216                 int flags)
 217 {
 218         int result;
 219         void *kaddr = kmap(bvec->bv_page);
 220         result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
 221                            bvec->bv_len, flags);
 222         kunmap(bvec->bv_page);
 223         return result;
 224 }
 225
 226 /* always call with the tx_lock held */
 227 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
 228 {
 229         int result, flags;
 230         struct nbd_request request;
 231         unsigned long size = blk_rq_bytes(req);
 232         u32 type;
 233
 234         if (req->cmd_type == REQ_TYPE_DRV_PRIV)
 235                 type = NBD_CMD_DISC;
 236         else if (req->cmd_flags & REQ_DISCARD)
 237                 type = NBD_CMD_TRIM;
 238         else if (req->cmd_flags & REQ_FLUSH)
 239                 type = NBD_CMD_FLUSH;
 240         else if (rq_data_dir(req) == WRITE)
 241                 type = NBD_CMD_WRITE;
 242         else
 243                 type = NBD_CMD_READ;
 244
 245         memset(&request, 0, sizeof(request));
 246         request.magic = htonl(NBD_REQUEST_MAGIC);
 247         request.type = htonl(type);
 248         if (type != NBD_CMD_FLUSH && type != NBD_CMD_DISC) {
 249                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 250                 request.len = htonl(size);
 251         }
 252         memcpy(request.handle, &req, sizeof(req));
 253
 254         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 255                 req, nbdcmd_to_ascii(type),
 256                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 257         result = sock_xmit(nbd, 1, &request, sizeof(request),
 258                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
 259         if (result <= 0) {
 260                 dev_err(disk_to_dev(nbd->disk),
 261                         "Send control failed (result %d)\n", result);
 262                 return -EIO;
 263         }
 264
 265         if (type == NBD_CMD_WRITE) {
 266                 struct req_iterator iter;
 267                 struct bio_vec bvec;
 268                 /*
 269                  * we are really probing at internals to determine
 270                  * whether to set MSG_MORE or not...
 271                  */
 272                 rq_for_each_segment(bvec, req, iter) {
 273                         flags = 0;
 274                         if (!rq_iter_last(bvec, iter))
 275                                 flags = MSG_MORE;
 276                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 277                                 req, bvec.bv_len);
 278                         result = sock_send_bvec(nbd, &bvec, flags);
 279                         if (result <= 0) {
 280                                 dev_err(disk_to_dev(nbd->disk),
 281                                         "Send data failed (result %d)\n",
 282                                         result);
 283                                 return -EIO;
 284                         }
 285                 }
 286         }
 287         return 0;
 288 }
 289
 290 static struct request *nbd_find_request(struct nbd_device *nbd,
 291                                         struct request *xreq)
 292 {
 293         struct request *req, *tmp;
 294         int err;
 295
 296         err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
 297         if (unlikely(err))
 298                 return ERR_PTR(err);
 299
 300         spin_lock(&nbd->queue_lock);
 301         list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
 302                 if (req != xreq)
 303                         continue;
 304                 list_del_init(&req->queuelist);
 305                 spin_unlock(&nbd->queue_lock);
 306                 return req;
 307         }
 308         spin_unlock(&nbd->queue_lock);
 309
 310         return ERR_PTR(-ENOENT);
 311 }
 312
 313 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
 314 {
 315         int result;
 316         void *kaddr = kmap(bvec->bv_page);
 317         result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 318                         MSG_WAITALL);
 319         kunmap(bvec->bv_page);
 320         return result;
 321 }
 322
 323 /* NULL returned = something went wrong, inform userspace */
 324 static struct request *nbd_read_stat(struct nbd_device *nbd)
 325 {
 326         int result;
 327         struct nbd_reply reply;
 328         struct request *req;
 329
 330         reply.magic = 0;
 331         result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
 332         if (result <= 0) {
 333                 dev_err(disk_to_dev(nbd->disk),
 334                         "Receive control failed (result %d)\n", result);
 335                 goto harderror;
 336         }
 337
 338         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 339                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 340                                 (unsigned long)ntohl(reply.magic));
 341                 result = -EPROTO;
 342                 goto harderror;
 343         }
 344
 345         req = nbd_find_request(nbd, *(struct request **)reply.handle);
 346         if (IS_ERR(req)) {
 347                 result = PTR_ERR(req);
 348                 if (result != -ENOENT)
 349                         goto harderror;
 350
 351                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
 352                         reply.handle);
 353                 result = -EBADR;
 354                 goto harderror;
 355         }
 356
 357         if (ntohl(reply.error)) {
 358                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 359                         ntohl(reply.error));
 360                 req->errors++;
 361                 return req;
 362         }
 363
 364         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 365         if (rq_data_dir(req) != WRITE) {
 366                 struct req_iterator iter;
 367                 struct bio_vec bvec;
 368
 369                 rq_for_each_segment(bvec, req, iter) {
 370                         result = sock_recv_bvec(nbd, &bvec);
 371                         if (result <= 0) {
 372                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 373                                         result);
 374                                 req->errors++;
 375                                 return req;
 376                         }
 377                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 378                                 req, bvec.bv_len);
 379                 }
 380         }
 381         return req;
 382 harderror:
 383         nbd->harderror = result;
 384         return NULL;
 385 }
 386
 387 static ssize_t pid_show(struct device *dev,
 388                         struct device_attribute *attr, char *buf)
 389 {
 390         struct gendisk *disk = dev_to_disk(dev);
 391
 392         return sprintf(buf, "%ld\n",
 393                 (long) ((struct nbd_device *)disk->private_data)->pid);
 394 }
 395
 396 static struct device_attribute pid_attr = {
 397         .attr = { .name = "pid", .mode = S_IRUGO},
 398         .show = pid_show,
 399 };
 400
 401 static int nbd_do_it(struct nbd_device *nbd)
 402 {
 403         struct request *req;
 404         int ret;
 405
 406         BUG_ON(nbd->magic != NBD_MAGIC);
 407
 408         sk_set_memalloc(nbd->sock->sk);
 409         nbd->pid = task_pid_nr(current);
 410         ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 411         if (ret) {
 412                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 413                 nbd->pid = 0;
 414                 return ret;
 415         }
 416
 417         nbd->task_recv = current;
 418
 419         while ((req = nbd_read_stat(nbd)) != NULL)
 420                 nbd_end_request(nbd, req);
 421
 422         nbd->task_recv = NULL;
 423
 424         if (signal_pending(current)) {
 425                 siginfo_t info;
 426
 427                 ret = dequeue_signal_lock(current, &current->blocked, &info);
 428                 dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
 429                          task_pid_nr(current), current->comm, ret);
 430                 sock_shutdown(nbd, 1);
 431                 ret = -ETIMEDOUT;
 432         }
 433
 434         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 435         nbd->pid = 0;
 436         return ret;
 437 }
 438
 439 static void nbd_clear_que(struct nbd_device *nbd)
 440 {
 441         struct request *req;
 442
 443         BUG_ON(nbd->magic != NBD_MAGIC);
 444
 445         /*
 446          * Because we have set nbd->sock to NULL under the tx_lock, all
 447          * modifications to the list must have completed by now.  For
 448          * the same reason, the active_req must be NULL.
 449          *
 450          * As a consequence, we don't need to take the spin lock while
 451          * purging the list here.
 452          */
 453         BUG_ON(nbd->sock);
 454         BUG_ON(nbd->active_req);
 455
 456         while (!list_empty(&nbd->queue_head)) {
 457                 req = list_entry(nbd->queue_head.next, struct request,
 458                                  queuelist);
 459                 list_del_init(&req->queuelist);
 460                 req->errors++;
 461                 nbd_end_request(nbd, req);
 462         }
 463
 464         while (!list_empty(&nbd->waiting_queue)) {
 465                 req = list_entry(nbd->waiting_queue.next, struct request,
 466                                  queuelist);
 467                 list_del_init(&req->queuelist);
 468                 req->errors++;
 469                 nbd_end_request(nbd, req);
 470         }
 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                         sock_shutdown(nbd, 1);
 545                         break;
 546                 }
 547
 548                 /* extract request */
 549                 if (list_empty(&nbd->waiting_queue))
 550                         continue;
 551
 552                 spin_lock_irq(&nbd->queue_lock);
 553                 req = list_entry(nbd->waiting_queue.next, struct request,
 554                                  queuelist);
 555                 list_del_init(&req->queuelist);
 556                 spin_unlock_irq(&nbd->queue_lock);
 557
 558                 /* handle request */
 559                 nbd_handle_req(nbd, req);
 560         }
 561
 562         nbd->task_send = NULL;
 563
 564         return 0;
 565 }
 566
 567 /*
 568  * We always wait for result of write, for now. It would be nice to make it optional
 569  * in future
 570  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
 571  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 572  */
 573
 574 static void do_nbd_request(struct request_queue *q)
 575                 __releases(q->queue_lock) __acquires(q->queue_lock)
 576 {
 577         struct request *req;
 578
 579         while ((req = blk_fetch_request(q)) != NULL) {
 580                 struct nbd_device *nbd;
 581
 582                 spin_unlock_irq(q->queue_lock);
 583
 584                 nbd = req->rq_disk->private_data;
 585
 586                 BUG_ON(nbd->magic != NBD_MAGIC);
 587
 588                 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
 589                         req, req->cmd_type);
 590
 591                 if (unlikely(!nbd->sock)) {
 592                         dev_err(disk_to_dev(nbd->disk),
 593                                 "Attempted send on closed socket\n");
 594                         req->errors++;
 595                         nbd_end_request(nbd, req);
 596                         spin_lock_irq(q->queue_lock);
 597                         continue;
 598                 }
 599
 600                 spin_lock_irq(&nbd->queue_lock);
 601                 list_add_tail(&req->queuelist, &nbd->waiting_queue);
 602                 spin_unlock_irq(&nbd->queue_lock);
 603
 604                 wake_up(&nbd->waiting_wq);
 605
 606                 spin_lock_irq(q->queue_lock);
 607         }
 608 }
 609
 610 /* Must be called with tx_lock held */
 611
 612 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 613                        unsigned int cmd, unsigned long arg)
 614 {
 615         switch (cmd) {
 616         case NBD_DISCONNECT: {
 617                 struct request sreq;
 618
 619                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 620                 if (!nbd->sock)
 621                         return -EINVAL;
 622
 623                 mutex_unlock(&nbd->tx_lock);
 624                 fsync_bdev(bdev);
 625                 mutex_lock(&nbd->tx_lock);
 626                 blk_rq_init(NULL, &sreq);
 627                 sreq.cmd_type = REQ_TYPE_DRV_PRIV;
 628
 629                 /* Check again after getting mutex back.  */
 630                 if (!nbd->sock)
 631                         return -EINVAL;
 632
 633                 nbd->disconnect = 1;
 634
 635                 nbd_send_req(nbd, &sreq);
 636                 return 0;
 637         }
 638
 639         case NBD_CLEAR_SOCK: {
 640                 struct socket *sock = nbd->sock;
 641                 nbd->sock = NULL;
 642                 nbd_clear_que(nbd);
 643                 BUG_ON(!list_empty(&nbd->queue_head));
 644                 BUG_ON(!list_empty(&nbd->waiting_queue));
 645                 kill_bdev(bdev);
 646                 if (sock)
 647                         sockfd_put(sock);
 648                 return 0;
 649         }
 650
 651         case NBD_SET_SOCK: {
 652                 struct socket *sock;
 653                 int err;
 654                 if (nbd->sock)
 655                         return -EBUSY;
 656                 sock = sockfd_lookup(arg, &err);
 657                 if (sock) {
 658                         nbd->sock = sock;
 659                         if (max_part > 0)
 660                                 bdev->bd_invalidated = 1;
 661                         nbd->disconnect = 0; /* we're connected now */
 662                         return 0;
 663                 }
 664                 return -EINVAL;
 665         }
 666
 667         case NBD_SET_BLKSIZE:
 668                 nbd->blksize = arg;
 669                 nbd->bytesize &= ~(nbd->blksize-1);
 670                 bdev->bd_inode->i_size = nbd->bytesize;
 671                 set_blocksize(bdev, nbd->blksize);
 672                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 673                 return 0;
 674
 675         case NBD_SET_SIZE:
 676                 nbd->bytesize = arg & ~(nbd->blksize-1);
 677                 bdev->bd_inode->i_size = nbd->bytesize;
 678                 set_blocksize(bdev, nbd->blksize);
 679                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 680                 return 0;
 681
 682         case NBD_SET_TIMEOUT:
 683                 nbd->xmit_timeout = arg * HZ;
 684                 if (arg)
 685                         mod_timer(&nbd->timeout_timer,
 686                                   jiffies + nbd->xmit_timeout);
 687                 else
 688                         del_timer_sync(&nbd->timeout_timer);
 689
 690                 return 0;
 691
 692         case NBD_SET_FLAGS:
 693                 nbd->flags = arg;
 694                 return 0;
 695
 696         case NBD_SET_SIZE_BLOCKS:
 697                 nbd->bytesize = ((u64) arg) * nbd->blksize;
 698                 bdev->bd_inode->i_size = nbd->bytesize;
 699                 set_blocksize(bdev, nbd->blksize);
 700                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 701                 return 0;
 702
 703         case NBD_DO_IT: {
 704                 struct task_struct *thread;
 705                 struct socket *sock;
 706                 int error;
 707
 708                 if (nbd->pid)
 709                         return -EBUSY;
 710                 if (!nbd->sock)
 711                         return -EINVAL;
 712
 713                 mutex_unlock(&nbd->tx_lock);
 714
 715                 if (nbd->flags & NBD_FLAG_READ_ONLY)
 716                         set_device_ro(bdev, true);
 717                 if (nbd->flags & NBD_FLAG_SEND_TRIM)
 718                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
 719                                 nbd->disk->queue);
 720                 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
 721                         blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
 722                 else
 723                         blk_queue_flush(nbd->disk->queue, 0);
 724
 725                 thread = kthread_run(nbd_thread, nbd, "%s",
 726                                      nbd->disk->disk_name);
 727                 if (IS_ERR(thread)) {
 728                         mutex_lock(&nbd->tx_lock);
 729                         return PTR_ERR(thread);
 730                 }
 731
 732                 error = nbd_do_it(nbd);
 733                 kthread_stop(thread);
 734
 735                 mutex_lock(&nbd->tx_lock);
 736                 if (error)
 737                         return error;
 738                 sock_shutdown(nbd, 0);
 739                 sock = nbd->sock;
 740                 nbd->sock = NULL;
 741                 nbd_clear_que(nbd);
 742                 dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
 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 nbd->harderror;
 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)");