drivers/block/nbd.c

   1 /*
   2  * Network block device - make block devices work over TCP
   3  *
   4  * Note that you can not swap over this thing, yet. Seems to work but
   5  * deadlocks sometimes - you can not swap over TCP in general.
   6  *
   7  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   8  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
   9  *
  10  * This file is released under GPLv2 or later.
  11  *
  12  * (part of code stolen from loop.c)
  13  */
  14
  15 #include <linux/major.h>
  16
  17 #include <linux/blkdev.h>
  18 #include <linux/module.h>
  19 #include <linux/init.h>
  20 #include <linux/sched.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/err.h>
  30 #include <linux/kernel.h>
  31 #include <linux/slab.h>
  32 #include <net/sock.h>
  33 #include <linux/net.h>
  34 #include <linux/kthread.h>
  35
  36 #include <asm/uaccess.h>
  37 #include <asm/types.h>
  38
  39 #include <linux/nbd.h>
  40
  41 #define NBD_MAGIC 0x68797548
  42
  43 #ifdef NDEBUG
  44 #define dprintk(flags, fmt...)
  45 #else /* NDEBUG */
  46 #define dprintk(flags, fmt...) do { \
  47         if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
  48 } while (0)
  49 #define DBG_IOCTL       0x0004
  50 #define DBG_INIT        0x0010
  51 #define DBG_EXIT        0x0020
  52 #define DBG_BLKDEV      0x0100
  53 #define DBG_RX          0x0200
  54 #define DBG_TX          0x0400
  55 static unsigned int debugflags;
  56 #endif /* NDEBUG */
  57
  58 static unsigned int nbds_max = 16;
  59 static struct nbd_device *nbd_dev;
  60 static int max_part;
  61
  62 /*
  63  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  64  * 1. Each NIC is essentially a synchronization point for all servers
  65  *    accessed through that NIC so there's no need to have more locks
  66  *    than NICs anyway.
  67  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  68  *    down each lock to the point where they're actually slower than just
  69  *    a single lock.
  70  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  71  */
  72 static DEFINE_SPINLOCK(nbd_lock);
  73
  74 #ifndef NDEBUG
  75 static const char *ioctl_cmd_to_ascii(int cmd)
  76 {
  77         switch (cmd) {
  78         case NBD_SET_SOCK: return "set-sock";
  79         case NBD_SET_BLKSIZE: return "set-blksize";
  80         case NBD_SET_SIZE: return "set-size";
  81         case NBD_SET_TIMEOUT: return "set-timeout";
  82         case NBD_SET_FLAGS: return "set-flags";
  83         case NBD_DO_IT: return "do-it";
  84         case NBD_CLEAR_SOCK: return "clear-sock";
  85         case NBD_CLEAR_QUE: return "clear-que";
  86         case NBD_PRINT_DEBUG: return "print-debug";
  87         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
  88         case NBD_DISCONNECT: return "disconnect";
  89         case BLKROSET: return "set-read-only";
  90         case BLKFLSBUF: return "flush-buffer-cache";
  91         }
  92         return "unknown";
  93 }
  94
  95 static const char *nbdcmd_to_ascii(int cmd)
  96 {
  97         switch (cmd) {
  98         case  NBD_CMD_READ: return "read";
  99         case NBD_CMD_WRITE: return "write";
 100         case  NBD_CMD_DISC: return "disconnect";
 101         case  NBD_CMD_TRIM: return "trim/discard";
 102         }
 103         return "invalid";
 104 }
 105 #endif /* NDEBUG */
 106
 107 static void nbd_end_request(struct request *req)
 108 {
 109         int error = req->errors ? -EIO : 0;
 110         struct request_queue *q = req->q;
 111         unsigned long flags;
 112
 113         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 114                         req, error ? "failed" : "done");
 115
 116         spin_lock_irqsave(q->queue_lock, flags);
 117         __blk_end_request_all(req, error);
 118         spin_unlock_irqrestore(q->queue_lock, flags);
 119 }
 120
 121 static void sock_shutdown(struct nbd_device *nbd, int lock)
 122 {
 123         /* Forcibly shutdown the socket causing all listeners
 124          * to error
 125          *
 126          * FIXME: This code is duplicated from sys_shutdown, but
 127          * there should be a more generic interface rather than
 128          * calling socket ops directly here */
 129         if (lock)
 130                 mutex_lock(&nbd->tx_lock);
 131         if (nbd->sock) {
 132                 dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
 133                 kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
 134                 nbd->sock = NULL;
 135         }
 136         if (lock)
 137                 mutex_unlock(&nbd->tx_lock);
 138 }
 139
 140 static void nbd_xmit_timeout(unsigned long arg)
 141 {
 142         struct task_struct *task = (struct task_struct *)arg;
 143
 144         printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
 145                 task->comm, task->pid);
 146         force_sig(SIGKILL, task);
 147 }
 148
 149 /*
 150  *  Send or receive packet.
 151  */
 152 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
 153                 int msg_flags)
 154 {
 155         struct socket *sock = nbd->sock;
 156         int result;
 157         struct msghdr msg;
 158         struct kvec iov;
 159         sigset_t blocked, oldset;
 160         unsigned long pflags = current->flags;
 161
 162         if (unlikely(!sock)) {
 163                 dev_err(disk_to_dev(nbd->disk),
 164                         "Attempted %s on closed socket in sock_xmit\n",
 165                         (send ? "send" : "recv"));
 166                 return -EINVAL;
 167         }
 168
 169         /* Allow interception of SIGKILL only
 170          * Don't allow other signals to interrupt the transmission */
 171         siginitsetinv(&blocked, sigmask(SIGKILL));
 172         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 173
 174         current->flags |= PF_MEMALLOC;
 175         do {
 176                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 177                 iov.iov_base = buf;
 178                 iov.iov_len = size;
 179                 msg.msg_name = NULL;
 180                 msg.msg_namelen = 0;
 181                 msg.msg_control = NULL;
 182                 msg.msg_controllen = 0;
 183                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 184
 185                 if (send) {
 186                         struct timer_list ti;
 187
 188                         if (nbd->xmit_timeout) {
 189                                 init_timer(&ti);
 190                                 ti.function = nbd_xmit_timeout;
 191                                 ti.data = (unsigned long)current;
 192                                 ti.expires = jiffies + nbd->xmit_timeout;
 193                                 add_timer(&ti);
 194                         }
 195                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 196                         if (nbd->xmit_timeout)
 197                                 del_timer_sync(&ti);
 198                 } else
 199                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 200                                                 msg.msg_flags);
 201
 202                 if (signal_pending(current)) {
 203                         siginfo_t info;
 204                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 205                                 task_pid_nr(current), current->comm,
 206                                 dequeue_signal_lock(current, &current->blocked, &info));
 207                         result = -EINTR;
 208                         sock_shutdown(nbd, !send);
 209                         break;
 210                 }
 211
 212                 if (result <= 0) {
 213                         if (result == 0)
 214                                 result = -EPIPE; /* short read */
 215                         break;
 216                 }
 217                 size -= result;
 218                 buf += result;
 219         } while (size > 0);
 220
 221         sigprocmask(SIG_SETMASK, &oldset, NULL);
 222         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 223
 224         return result;
 225 }
 226
 227 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
 228                 int flags)
 229 {
 230         int result;
 231         void *kaddr = kmap(bvec->bv_page);
 232         result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
 233                            bvec->bv_len, flags);
 234         kunmap(bvec->bv_page);
 235         return result;
 236 }
 237
 238 /* always call with the tx_lock held */
 239 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
 240 {
 241         int result, flags;
 242         struct nbd_request request;
 243         unsigned long size = blk_rq_bytes(req);
 244
 245         request.magic = htonl(NBD_REQUEST_MAGIC);
 246         request.type = htonl(nbd_cmd(req));
 247         request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 248         request.len = htonl(size);
 249         memcpy(request.handle, &req, sizeof(req));
 250
 251         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
 252                         nbd->disk->disk_name, req,
 253                         nbdcmd_to_ascii(nbd_cmd(req)),
 254                         (unsigned long long)blk_rq_pos(req) << 9,
 255                         blk_rq_bytes(req));
 256         result = sock_xmit(nbd, 1, &request, sizeof(request),
 257                         (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
 258         if (result <= 0) {
 259                 dev_err(disk_to_dev(nbd->disk),
 260                         "Send control failed (result %d)\n", result);
 261                 goto error_out;
 262         }
 263
 264         if (nbd_cmd(req) == NBD_CMD_WRITE) {
 265                 struct req_iterator iter;
 266                 struct bio_vec *bvec;
 267                 /*
 268                  * we are really probing at internals to determine
 269                  * whether to set MSG_MORE or not...
 270                  */
 271                 rq_for_each_segment(bvec, req, iter) {
 272                         flags = 0;
 273                         if (!rq_iter_last(req, iter))
 274                                 flags = MSG_MORE;
 275                         dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 276                                         nbd->disk->disk_name, req, bvec->bv_len);
 277                         result = sock_send_bvec(nbd, bvec, flags);
 278                         if (result <= 0) {
 279                                 dev_err(disk_to_dev(nbd->disk),
 280                                         "Send data failed (result %d)\n",
 281                                         result);
 282                                 goto error_out;
 283                         }
 284                 }
 285         }
 286         return 0;
 287
 288 error_out:
 289         return -EIO;
 290 }
 291
 292 static struct request *nbd_find_request(struct nbd_device *nbd,
 293                                         struct request *xreq)
 294 {
 295         struct request *req, *tmp;
 296         int err;
 297
 298         err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
 299         if (unlikely(err))
 300                 goto out;
 301
 302         spin_lock(&nbd->queue_lock);
 303         list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
 304                 if (req != xreq)
 305                         continue;
 306                 list_del_init(&req->queuelist);
 307                 spin_unlock(&nbd->queue_lock);
 308                 return req;
 309         }
 310         spin_unlock(&nbd->queue_lock);
 311
 312         err = -ENOENT;
 313
 314 out:
 315         return ERR_PTR(err);
 316 }
 317
 318 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
 319 {
 320         int result;
 321         void *kaddr = kmap(bvec->bv_page);
 322         result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 323                         MSG_WAITALL);
 324         kunmap(bvec->bv_page);
 325         return result;
 326 }
 327
 328 /* NULL returned = something went wrong, inform userspace */
 329 static struct request *nbd_read_stat(struct nbd_device *nbd)
 330 {
 331         int result;
 332         struct nbd_reply reply;
 333         struct request *req;
 334
 335         reply.magic = 0;
 336         result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
 337         if (result <= 0) {
 338                 dev_err(disk_to_dev(nbd->disk),
 339                         "Receive control failed (result %d)\n", result);
 340                 goto harderror;
 341         }
 342
 343         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 344                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 345                                 (unsigned long)ntohl(reply.magic));
 346                 result = -EPROTO;
 347                 goto harderror;
 348         }
 349
 350         req = nbd_find_request(nbd, *(struct request **)reply.handle);
 351         if (IS_ERR(req)) {
 352                 result = PTR_ERR(req);
 353                 if (result != -ENOENT)
 354                         goto harderror;
 355
 356                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
 357                         reply.handle);
 358                 result = -EBADR;
 359                 goto harderror;
 360         }
 361
 362         if (ntohl(reply.error)) {
 363                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 364                         ntohl(reply.error));
 365                 req->errors++;
 366                 return req;
 367         }
 368
 369         dprintk(DBG_RX, "%s: request %p: got reply\n",
 370                         nbd->disk->disk_name, req);
 371         if (nbd_cmd(req) == NBD_CMD_READ) {
 372                 struct req_iterator iter;
 373                 struct bio_vec *bvec;
 374
 375                 rq_for_each_segment(bvec, req, iter) {
 376                         result = sock_recv_bvec(nbd, bvec);
 377                         if (result <= 0) {
 378                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 379                                         result);
 380                                 req->errors++;
 381                                 return req;
 382                         }
 383                         dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 384                                 nbd->disk->disk_name, req, bvec->bv_len);
 385                 }
 386         }
 387         return req;
 388 harderror:
 389         nbd->harderror = result;
 390         return NULL;
 391 }
 392
 393 static ssize_t pid_show(struct device *dev,
 394                         struct device_attribute *attr, char *buf)
 395 {
 396         struct gendisk *disk = dev_to_disk(dev);
 397
 398         return sprintf(buf, "%ld\n",
 399                 (long) ((struct nbd_device *)disk->private_data)->pid);
 400 }
 401
 402 static struct device_attribute pid_attr = {
 403         .attr = { .name = "pid", .mode = S_IRUGO},
 404         .show = pid_show,
 405 };
 406
 407 static int nbd_do_it(struct nbd_device *nbd)
 408 {
 409         struct request *req;
 410         int ret;
 411
 412         BUG_ON(nbd->magic != NBD_MAGIC);
 413
 414         sk_set_memalloc(nbd->sock->sk);
 415         nbd->pid = task_pid_nr(current);
 416         ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 417         if (ret) {
 418                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 419                 nbd->pid = 0;
 420                 return ret;
 421         }
 422
 423         while ((req = nbd_read_stat(nbd)) != NULL)
 424                 nbd_end_request(req);
 425
 426         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 427         nbd->pid = 0;
 428         return 0;
 429 }
 430
 431 static void nbd_clear_que(struct nbd_device *nbd)
 432 {
 433         struct request *req;
 434
 435         BUG_ON(nbd->magic != NBD_MAGIC);
 436
 437         /*
 438          * Because we have set nbd->sock to NULL under the tx_lock, all
 439          * modifications to the list must have completed by now.  For
 440          * the same reason, the active_req must be NULL.
 441          *
 442          * As a consequence, we don't need to take the spin lock while
 443          * purging the list here.
 444          */
 445         BUG_ON(nbd->sock);
 446         BUG_ON(nbd->active_req);
 447
 448         while (!list_empty(&nbd->queue_head)) {
 449                 req = list_entry(nbd->queue_head.next, struct request,
 450                                  queuelist);
 451                 list_del_init(&req->queuelist);
 452                 req->errors++;
 453                 nbd_end_request(req);
 454         }
 455
 456         while (!list_empty(&nbd->waiting_queue)) {
 457                 req = list_entry(nbd->waiting_queue.next, struct request,
 458                                  queuelist);
 459                 list_del_init(&req->queuelist);
 460                 req->errors++;
 461                 nbd_end_request(req);
 462         }
 463 }
 464
 465
 466 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
 467 {
 468         if (req->cmd_type != REQ_TYPE_FS)
 469                 goto error_out;
 470
 471         nbd_cmd(req) = NBD_CMD_READ;
 472         if (rq_data_dir(req) == WRITE) {
 473                 if ((req->cmd_flags & REQ_DISCARD)) {
 474                         WARN_ON(!(nbd->flags & NBD_FLAG_SEND_TRIM));
 475                         nbd_cmd(req) = NBD_CMD_TRIM;
 476                 } else
 477                         nbd_cmd(req) = NBD_CMD_WRITE;
 478                 if (nbd->flags & NBD_FLAG_READ_ONLY) {
 479                         dev_err(disk_to_dev(nbd->disk),
 480                                 "Write on read-only\n");
 481                         goto error_out;
 482                 }
 483         }
 484
 485         req->errors = 0;
 486
 487         mutex_lock(&nbd->tx_lock);
 488         if (unlikely(!nbd->sock)) {
 489                 mutex_unlock(&nbd->tx_lock);
 490                 dev_err(disk_to_dev(nbd->disk),
 491                         "Attempted send on closed socket\n");
 492                 goto error_out;
 493         }
 494
 495         nbd->active_req = req;
 496
 497         if (nbd_send_req(nbd, req) != 0) {
 498                 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
 499                 req->errors++;
 500                 nbd_end_request(req);
 501         } else {
 502                 spin_lock(&nbd->queue_lock);
 503                 list_add_tail(&req->queuelist, &nbd->queue_head);
 504                 spin_unlock(&nbd->queue_lock);
 505         }
 506
 507         nbd->active_req = NULL;
 508         mutex_unlock(&nbd->tx_lock);
 509         wake_up_all(&nbd->active_wq);
 510
 511         return;
 512
 513 error_out:
 514         req->errors++;
 515         nbd_end_request(req);
 516 }
 517
 518 static int nbd_thread(void *data)
 519 {
 520         struct nbd_device *nbd = data;
 521         struct request *req;
 522
 523         set_user_nice(current, -20);
 524         while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
 525                 /* wait for something to do */
 526                 wait_event_interruptible(nbd->waiting_wq,
 527                                          kthread_should_stop() ||
 528                                          !list_empty(&nbd->waiting_queue));
 529
 530                 /* extract request */
 531                 if (list_empty(&nbd->waiting_queue))
 532                         continue;
 533
 534                 spin_lock_irq(&nbd->queue_lock);
 535                 req = list_entry(nbd->waiting_queue.next, struct request,
 536                                  queuelist);
 537                 list_del_init(&req->queuelist);
 538                 spin_unlock_irq(&nbd->queue_lock);
 539
 540                 /* handle request */
 541                 nbd_handle_req(nbd, req);
 542         }
 543         return 0;
 544 }
 545
 546 /*
 547  * We always wait for result of write, for now. It would be nice to make it optional
 548  * in future
 549  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
 550  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 551  */
 552
 553 static void do_nbd_request(struct request_queue *q)
 554 {
 555         struct request *req;
 556
 557         while ((req = blk_fetch_request(q)) != NULL) {
 558                 struct nbd_device *nbd;
 559
 560                 spin_unlock_irq(q->queue_lock);
 561
 562                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
 563                                 req->rq_disk->disk_name, req, req->cmd_type);
 564
 565                 nbd = req->rq_disk->private_data;
 566
 567                 BUG_ON(nbd->magic != NBD_MAGIC);
 568
 569                 if (unlikely(!nbd->sock)) {
 570                         dev_err(disk_to_dev(nbd->disk),
 571                                 "Attempted send on closed socket\n");
 572                         req->errors++;
 573                         nbd_end_request(req);
 574                         spin_lock_irq(q->queue_lock);
 575                         continue;
 576                 }
 577
 578                 spin_lock_irq(&nbd->queue_lock);
 579                 list_add_tail(&req->queuelist, &nbd->waiting_queue);
 580                 spin_unlock_irq(&nbd->queue_lock);
 581
 582                 wake_up(&nbd->waiting_wq);
 583
 584                 spin_lock_irq(q->queue_lock);
 585         }
 586 }
 587
 588 /* Must be called with tx_lock held */
 589
 590 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 591                        unsigned int cmd, unsigned long arg)
 592 {
 593         switch (cmd) {
 594         case NBD_DISCONNECT: {
 595                 struct request sreq;
 596
 597                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 598
 599                 blk_rq_init(NULL, &sreq);
 600                 sreq.cmd_type = REQ_TYPE_SPECIAL;
 601                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 602                 if (!nbd->sock)
 603                         return -EINVAL;
 604                 nbd_send_req(nbd, &sreq);
 605                 return 0;
 606         }
 607
 608         case NBD_CLEAR_SOCK: {
 609                 struct file *file;
 610
 611                 nbd->sock = NULL;
 612                 file = nbd->file;
 613                 nbd->file = NULL;
 614                 nbd_clear_que(nbd);
 615                 BUG_ON(!list_empty(&nbd->queue_head));
 616                 BUG_ON(!list_empty(&nbd->waiting_queue));
 617                 if (file)
 618                         fput(file);
 619                 return 0;
 620         }
 621
 622         case NBD_SET_SOCK: {
 623                 struct file *file;
 624                 if (nbd->file)
 625                         return -EBUSY;
 626                 file = fget(arg);
 627                 if (file) {
 628                         struct inode *inode = file_inode(file);
 629                         if (S_ISSOCK(inode->i_mode)) {
 630                                 nbd->file = file;
 631                                 nbd->sock = SOCKET_I(inode);
 632                                 if (max_part > 0)
 633                                         bdev->bd_invalidated = 1;
 634                                 return 0;
 635                         } else {
 636                                 fput(file);
 637                         }
 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 file *file;
 675                 int error;
 676
 677                 if (nbd->pid)
 678                         return -EBUSY;
 679                 if (!nbd->file)
 680                         return -EINVAL;
 681
 682                 mutex_unlock(&nbd->tx_lock);
 683
 684                 if (nbd->flags & NBD_FLAG_SEND_TRIM)
 685                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
 686                                 nbd->disk->queue);
 687
 688                 thread = kthread_create(nbd_thread, nbd, nbd->disk->disk_name);
 689                 if (IS_ERR(thread)) {
 690                         mutex_lock(&nbd->tx_lock);
 691                         return PTR_ERR(thread);
 692                 }
 693                 wake_up_process(thread);
 694                 error = nbd_do_it(nbd);
 695                 kthread_stop(thread);
 696
 697                 mutex_lock(&nbd->tx_lock);
 698                 if (error)
 699                         return error;
 700                 sock_shutdown(nbd, 0);
 701                 file = nbd->file;
 702                 nbd->file = NULL;
 703                 nbd_clear_que(nbd);
 704                 dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
 705                 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 706                 if (file)
 707                         fput(file);
 708                 nbd->bytesize = 0;
 709                 bdev->bd_inode->i_size = 0;
 710                 set_capacity(nbd->disk, 0);
 711                 if (max_part > 0)
 712                         ioctl_by_bdev(bdev, BLKRRPART, 0);
 713                 return nbd->harderror;
 714         }
 715
 716         case NBD_CLEAR_QUE:
 717                 /*
 718                  * This is for compatibility only.  The queue is always cleared
 719                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 720                  */
 721                 BUG_ON(!nbd->sock && !list_empty(&nbd->queue_head));
 722                 return 0;
 723
 724         case NBD_PRINT_DEBUG:
 725                 dev_info(disk_to_dev(nbd->disk),
 726                         "next = %p, prev = %p, head = %p\n",
 727                         nbd->queue_head.next, nbd->queue_head.prev,
 728                         &nbd->queue_head);
 729                 return 0;
 730         }
 731         return -ENOTTY;
 732 }
 733
 734 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 735                      unsigned int cmd, unsigned long arg)
 736 {
 737         struct nbd_device *nbd = bdev->bd_disk->private_data;
 738         int error;
 739
 740         if (!capable(CAP_SYS_ADMIN))
 741                 return -EPERM;
 742
 743         BUG_ON(nbd->magic != NBD_MAGIC);
 744
 745         /* Anyone capable of this syscall can do *real bad* things */
 746         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 747                 nbd->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 748
 749         mutex_lock(&nbd->tx_lock);
 750         error = __nbd_ioctl(bdev, nbd, cmd, arg);
 751         mutex_unlock(&nbd->tx_lock);
 752
 753         return error;
 754 }
 755
 756 static const struct block_device_operations nbd_fops =
 757 {
 758         .owner =        THIS_MODULE,
 759         .ioctl =        nbd_ioctl,
 760 };
 761
 762 /*
 763  * And here should be modules and kernel interface
 764  *  (Just smiley confuses emacs :-)
 765  */
 766
 767 static int __init nbd_init(void)
 768 {
 769         int err = -ENOMEM;
 770         int i;
 771         int part_shift;
 772
 773         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 774
 775         if (max_part < 0) {
 776                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
 777                 return -EINVAL;
 778         }
 779
 780         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
 781         if (!nbd_dev)
 782                 return -ENOMEM;
 783
 784         part_shift = 0;
 785         if (max_part > 0) {
 786                 part_shift = fls(max_part);
 787
 788                 /*
 789                  * Adjust max_part according to part_shift as it is exported
 790                  * to user space so that user can know the max number of
 791                  * partition kernel should be able to manage.
 792                  *
 793                  * Note that -1 is required because partition 0 is reserved
 794                  * for the whole disk.
 795                  */
 796                 max_part = (1UL << part_shift) - 1;
 797         }
 798
 799         if ((1UL << part_shift) > DISK_MAX_PARTS)
 800                 return -EINVAL;
 801
 802         if (nbds_max > 1UL << (MINORBITS - part_shift))
 803                 return -EINVAL;
 804
 805         for (i = 0; i < nbds_max; i++) {
 806                 struct gendisk *disk = alloc_disk(1 << part_shift);
 807                 if (!disk)
 808                         goto out;
 809                 nbd_dev[i].disk = disk;
 810                 /*
 811                  * The new linux 2.5 block layer implementation requires
 812                  * every gendisk to have its very own request_queue struct.
 813                  * These structs are big so we dynamically allocate them.
 814                  */
 815                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 816                 if (!disk->queue) {
 817                         put_disk(disk);
 818                         goto out;
 819                 }
 820                 /*
 821                  * Tell the block layer that we are not a rotational device
 822                  */
 823                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
 824                 disk->queue->limits.discard_granularity = 512;
 825                 disk->queue->limits.max_discard_sectors = UINT_MAX;
 826                 disk->queue->limits.discard_zeroes_data = 0;
 827         }
 828
 829         if (register_blkdev(NBD_MAJOR, "nbd")) {
 830                 err = -EIO;
 831                 goto out;
 832         }
 833
 834         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 835         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 836
 837         for (i = 0; i < nbds_max; i++) {
 838                 struct gendisk *disk = nbd_dev[i].disk;
 839                 nbd_dev[i].file = NULL;
 840                 nbd_dev[i].magic = NBD_MAGIC;
 841                 nbd_dev[i].flags = 0;
 842                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
 843                 spin_lock_init(&nbd_dev[i].queue_lock);
 844                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 845                 mutex_init(&nbd_dev[i].tx_lock);
 846                 init_waitqueue_head(&nbd_dev[i].active_wq);
 847                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
 848                 nbd_dev[i].blksize = 1024;
 849                 nbd_dev[i].bytesize = 0;
 850                 disk->major = NBD_MAJOR;
 851                 disk->first_minor = i << part_shift;
 852                 disk->fops = &nbd_fops;
 853                 disk->private_data = &nbd_dev[i];
 854                 sprintf(disk->disk_name, "nbd%d", i);
 855                 set_capacity(disk, 0);
 856                 add_disk(disk);
 857         }
 858
 859         return 0;
 860 out:
 861         while (i--) {
 862                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 863                 put_disk(nbd_dev[i].disk);
 864         }
 865         kfree(nbd_dev);
 866         return err;
 867 }
 868
 869 static void __exit nbd_cleanup(void)
 870 {
 871         int i;
 872         for (i = 0; i < nbds_max; i++) {
 873                 struct gendisk *disk = nbd_dev[i].disk;
 874                 nbd_dev[i].magic = 0;
 875                 if (disk) {
 876                         del_gendisk(disk);
 877                         blk_cleanup_queue(disk->queue);
 878                         put_disk(disk);
 879                 }
 880         }
 881         unregister_blkdev(NBD_MAJOR, "nbd");
 882         kfree(nbd_dev);
 883         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 884 }
 885
 886 module_init(nbd_init);
 887 module_exit(nbd_cleanup);
 888
 889 MODULE_DESCRIPTION("Network Block Device");
 890 MODULE_LICENSE("GPL");
 891
 892 module_param(nbds_max, int, 0444);
 893 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
 894 module_param(max_part, int, 0444);
 895 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
 896 #ifndef NDEBUG
 897 module_param(debugflags, int, 0644);
 898 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 899 #endif