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