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 struct nbd_device {
  42         int flags;
  43         int harderror;          /* Code of hard error                   */
  44         struct socket * sock;   /* If == NULL, device is not ready, yet */
  45         int magic;
  46
  47         spinlock_t queue_lock;
  48         struct list_head queue_head;    /* Requests waiting result */
  49         struct request *active_req;
  50         wait_queue_head_t active_wq;
  51         struct list_head waiting_queue; /* Requests to be sent */
  52         wait_queue_head_t waiting_wq;
  53
  54         struct mutex tx_lock;
  55         struct gendisk *disk;
  56         int blksize;
  57         u64 bytesize;
  58         pid_t pid; /* pid of nbd-client, if attached */
  59         int xmit_timeout;
  60         int disconnect; /* a disconnect has been requested by user */
  61 };
  62
  63 #define NBD_MAGIC 0x68797548
  64
  65 #ifdef NDEBUG
  66 #define dprintk(flags, fmt...)
  67 #else /* NDEBUG */
  68 #define dprintk(flags, fmt...) do { \
  69         if (debugflags & (flags)) printk(KERN_DEBUG fmt); \
  70 } while (0)
  71 #define DBG_IOCTL       0x0004
  72 #define DBG_INIT        0x0010
  73 #define DBG_EXIT        0x0020
  74 #define DBG_BLKDEV      0x0100
  75 #define DBG_RX          0x0200
  76 #define DBG_TX          0x0400
  77 static unsigned int debugflags;
  78 #endif /* NDEBUG */
  79
  80 static unsigned int nbds_max = 16;
  81 static struct nbd_device *nbd_dev;
  82 static int max_part;
  83
  84 /*
  85  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  86  * 1. Each NIC is essentially a synchronization point for all servers
  87  *    accessed through that NIC so there's no need to have more locks
  88  *    than NICs anyway.
  89  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  90  *    down each lock to the point where they're actually slower than just
  91  *    a single lock.
  92  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  93  */
  94 static DEFINE_SPINLOCK(nbd_lock);
  95
  96 #ifndef NDEBUG
  97 static const char *ioctl_cmd_to_ascii(int cmd)
  98 {
  99         switch (cmd) {
 100         case NBD_SET_SOCK: return "set-sock";
 101         case NBD_SET_BLKSIZE: return "set-blksize";
 102         case NBD_SET_SIZE: return "set-size";
 103         case NBD_SET_TIMEOUT: return "set-timeout";
 104         case NBD_SET_FLAGS: return "set-flags";
 105         case NBD_DO_IT: return "do-it";
 106         case NBD_CLEAR_SOCK: return "clear-sock";
 107         case NBD_CLEAR_QUE: return "clear-que";
 108         case NBD_PRINT_DEBUG: return "print-debug";
 109         case NBD_SET_SIZE_BLOCKS: return "set-size-blocks";
 110         case NBD_DISCONNECT: return "disconnect";
 111         case BLKROSET: return "set-read-only";
 112         case BLKFLSBUF: return "flush-buffer-cache";
 113         }
 114         return "unknown";
 115 }
 116
 117 static const char *nbdcmd_to_ascii(int cmd)
 118 {
 119         switch (cmd) {
 120         case  NBD_CMD_READ: return "read";
 121         case NBD_CMD_WRITE: return "write";
 122         case  NBD_CMD_DISC: return "disconnect";
 123         case NBD_CMD_FLUSH: return "flush";
 124         case  NBD_CMD_TRIM: return "trim/discard";
 125         }
 126         return "invalid";
 127 }
 128 #endif /* NDEBUG */
 129
 130 static void nbd_end_request(struct request *req)
 131 {
 132         int error = req->errors ? -EIO : 0;
 133         struct request_queue *q = req->q;
 134         unsigned long flags;
 135
 136         dprintk(DBG_BLKDEV, "%s: request %p: %s\n", req->rq_disk->disk_name,
 137                         req, error ? "failed" : "done");
 138
 139         spin_lock_irqsave(q->queue_lock, flags);
 140         __blk_end_request_all(req, error);
 141         spin_unlock_irqrestore(q->queue_lock, flags);
 142 }
 143
 144 static void sock_shutdown(struct nbd_device *nbd, int lock)
 145 {
 146         /* Forcibly shutdown the socket causing all listeners
 147          * to error
 148          *
 149          * FIXME: This code is duplicated from sys_shutdown, but
 150          * there should be a more generic interface rather than
 151          * calling socket ops directly here */
 152         if (lock)
 153                 mutex_lock(&nbd->tx_lock);
 154         if (nbd->sock) {
 155                 dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
 156                 kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
 157                 nbd->sock = NULL;
 158         }
 159         if (lock)
 160                 mutex_unlock(&nbd->tx_lock);
 161 }
 162
 163 static void nbd_xmit_timeout(unsigned long arg)
 164 {
 165         struct task_struct *task = (struct task_struct *)arg;
 166
 167         printk(KERN_WARNING "nbd: killing hung xmit (%s, pid: %d)\n",
 168                 task->comm, task->pid);
 169         force_sig(SIGKILL, task);
 170 }
 171
 172 /*
 173  *  Send or receive packet.
 174  */
 175 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
 176                 int msg_flags)
 177 {
 178         struct socket *sock = nbd->sock;
 179         int result;
 180         struct msghdr msg;
 181         struct kvec iov;
 182         sigset_t blocked, oldset;
 183         unsigned long pflags = current->flags;
 184
 185         if (unlikely(!sock)) {
 186                 dev_err(disk_to_dev(nbd->disk),
 187                         "Attempted %s on closed socket in sock_xmit\n",
 188                         (send ? "send" : "recv"));
 189                 return -EINVAL;
 190         }
 191
 192         /* Allow interception of SIGKILL only
 193          * Don't allow other signals to interrupt the transmission */
 194         siginitsetinv(&blocked, sigmask(SIGKILL));
 195         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 196
 197         current->flags |= PF_MEMALLOC;
 198         do {
 199                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 200                 iov.iov_base = buf;
 201                 iov.iov_len = size;
 202                 msg.msg_name = NULL;
 203                 msg.msg_namelen = 0;
 204                 msg.msg_control = NULL;
 205                 msg.msg_controllen = 0;
 206                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 207
 208                 if (send) {
 209                         struct timer_list ti;
 210
 211                         if (nbd->xmit_timeout) {
 212                                 init_timer(&ti);
 213                                 ti.function = nbd_xmit_timeout;
 214                                 ti.data = (unsigned long)current;
 215                                 ti.expires = jiffies + nbd->xmit_timeout;
 216                                 add_timer(&ti);
 217                         }
 218                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 219                         if (nbd->xmit_timeout)
 220                                 del_timer_sync(&ti);
 221                 } else
 222                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 223                                                 msg.msg_flags);
 224
 225                 if (signal_pending(current)) {
 226                         siginfo_t info;
 227                         printk(KERN_WARNING "nbd (pid %d: %s) got signal %d\n",
 228                                 task_pid_nr(current), current->comm,
 229                                 dequeue_signal_lock(current, &current->blocked, &info));
 230                         result = -EINTR;
 231                         sock_shutdown(nbd, !send);
 232                         break;
 233                 }
 234
 235                 if (result <= 0) {
 236                         if (result == 0)
 237                                 result = -EPIPE; /* short read */
 238                         break;
 239                 }
 240                 size -= result;
 241                 buf += result;
 242         } while (size > 0);
 243
 244         sigprocmask(SIG_SETMASK, &oldset, NULL);
 245         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 246
 247         return result;
 248 }
 249
 250 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
 251                 int flags)
 252 {
 253         int result;
 254         void *kaddr = kmap(bvec->bv_page);
 255         result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
 256                            bvec->bv_len, flags);
 257         kunmap(bvec->bv_page);
 258         return result;
 259 }
 260
 261 /* always call with the tx_lock held */
 262 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
 263 {
 264         int result, flags;
 265         struct nbd_request request;
 266         unsigned long size = blk_rq_bytes(req);
 267
 268         memset(&request, 0, sizeof(request));
 269         request.magic = htonl(NBD_REQUEST_MAGIC);
 270         request.type = htonl(nbd_cmd(req));
 271
 272         if (nbd_cmd(req) != NBD_CMD_FLUSH && nbd_cmd(req) != NBD_CMD_DISC) {
 273                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 274                 request.len = htonl(size);
 275         }
 276         memcpy(request.handle, &req, sizeof(req));
 277
 278         dprintk(DBG_TX, "%s: request %p: sending control (%s@%llu,%uB)\n",
 279                         nbd->disk->disk_name, req,
 280                         nbdcmd_to_ascii(nbd_cmd(req)),
 281                         (unsigned long long)blk_rq_pos(req) << 9,
 282                         blk_rq_bytes(req));
 283         result = sock_xmit(nbd, 1, &request, sizeof(request),
 284                         (nbd_cmd(req) == NBD_CMD_WRITE) ? MSG_MORE : 0);
 285         if (result <= 0) {
 286                 dev_err(disk_to_dev(nbd->disk),
 287                         "Send control failed (result %d)\n", result);
 288                 goto error_out;
 289         }
 290
 291         if (nbd_cmd(req) == NBD_CMD_WRITE) {
 292                 struct req_iterator iter;
 293                 struct bio_vec bvec;
 294                 /*
 295                  * we are really probing at internals to determine
 296                  * whether to set MSG_MORE or not...
 297                  */
 298                 rq_for_each_segment(bvec, req, iter) {
 299                         flags = 0;
 300                         if (!rq_iter_last(bvec, iter))
 301                                 flags = MSG_MORE;
 302                         dprintk(DBG_TX, "%s: request %p: sending %d bytes data\n",
 303                                         nbd->disk->disk_name, req, bvec.bv_len);
 304                         result = sock_send_bvec(nbd, &bvec, flags);
 305                         if (result <= 0) {
 306                                 dev_err(disk_to_dev(nbd->disk),
 307                                         "Send data failed (result %d)\n",
 308                                         result);
 309                                 goto error_out;
 310                         }
 311                 }
 312         }
 313         return 0;
 314
 315 error_out:
 316         return -EIO;
 317 }
 318
 319 static struct request *nbd_find_request(struct nbd_device *nbd,
 320                                         struct request *xreq)
 321 {
 322         struct request *req, *tmp;
 323         int err;
 324
 325         err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
 326         if (unlikely(err))
 327                 goto out;
 328
 329         spin_lock(&nbd->queue_lock);
 330         list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
 331                 if (req != xreq)
 332                         continue;
 333                 list_del_init(&req->queuelist);
 334                 spin_unlock(&nbd->queue_lock);
 335                 return req;
 336         }
 337         spin_unlock(&nbd->queue_lock);
 338
 339         err = -ENOENT;
 340
 341 out:
 342         return ERR_PTR(err);
 343 }
 344
 345 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
 346 {
 347         int result;
 348         void *kaddr = kmap(bvec->bv_page);
 349         result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 350                         MSG_WAITALL);
 351         kunmap(bvec->bv_page);
 352         return result;
 353 }
 354
 355 /* NULL returned = something went wrong, inform userspace */
 356 static struct request *nbd_read_stat(struct nbd_device *nbd)
 357 {
 358         int result;
 359         struct nbd_reply reply;
 360         struct request *req;
 361
 362         reply.magic = 0;
 363         result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
 364         if (result <= 0) {
 365                 dev_err(disk_to_dev(nbd->disk),
 366                         "Receive control failed (result %d)\n", result);
 367                 goto harderror;
 368         }
 369
 370         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 371                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 372                                 (unsigned long)ntohl(reply.magic));
 373                 result = -EPROTO;
 374                 goto harderror;
 375         }
 376
 377         req = nbd_find_request(nbd, *(struct request **)reply.handle);
 378         if (IS_ERR(req)) {
 379                 result = PTR_ERR(req);
 380                 if (result != -ENOENT)
 381                         goto harderror;
 382
 383                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
 384                         reply.handle);
 385                 result = -EBADR;
 386                 goto harderror;
 387         }
 388
 389         if (ntohl(reply.error)) {
 390                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 391                         ntohl(reply.error));
 392                 req->errors++;
 393                 return req;
 394         }
 395
 396         dprintk(DBG_RX, "%s: request %p: got reply\n",
 397                         nbd->disk->disk_name, req);
 398         if (nbd_cmd(req) == NBD_CMD_READ) {
 399                 struct req_iterator iter;
 400                 struct bio_vec bvec;
 401
 402                 rq_for_each_segment(bvec, req, iter) {
 403                         result = sock_recv_bvec(nbd, &bvec);
 404                         if (result <= 0) {
 405                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 406                                         result);
 407                                 req->errors++;
 408                                 return req;
 409                         }
 410                         dprintk(DBG_RX, "%s: request %p: got %d bytes data\n",
 411                                 nbd->disk->disk_name, req, bvec.bv_len);
 412                 }
 413         }
 414         return req;
 415 harderror:
 416         nbd->harderror = result;
 417         return NULL;
 418 }
 419
 420 static ssize_t pid_show(struct device *dev,
 421                         struct device_attribute *attr, char *buf)
 422 {
 423         struct gendisk *disk = dev_to_disk(dev);
 424
 425         return sprintf(buf, "%ld\n",
 426                 (long) ((struct nbd_device *)disk->private_data)->pid);
 427 }
 428
 429 static struct device_attribute pid_attr = {
 430         .attr = { .name = "pid", .mode = S_IRUGO},
 431         .show = pid_show,
 432 };
 433
 434 static int nbd_do_it(struct nbd_device *nbd)
 435 {
 436         struct request *req;
 437         int ret;
 438
 439         BUG_ON(nbd->magic != NBD_MAGIC);
 440
 441         sk_set_memalloc(nbd->sock->sk);
 442         nbd->pid = task_pid_nr(current);
 443         ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 444         if (ret) {
 445                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 446                 nbd->pid = 0;
 447                 return ret;
 448         }
 449
 450         while ((req = nbd_read_stat(nbd)) != NULL)
 451                 nbd_end_request(req);
 452
 453         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 454         nbd->pid = 0;
 455         return 0;
 456 }
 457
 458 static void nbd_clear_que(struct nbd_device *nbd)
 459 {
 460         struct request *req;
 461
 462         BUG_ON(nbd->magic != NBD_MAGIC);
 463
 464         /*
 465          * Because we have set nbd->sock to NULL under the tx_lock, all
 466          * modifications to the list must have completed by now.  For
 467          * the same reason, the active_req must be NULL.
 468          *
 469          * As a consequence, we don't need to take the spin lock while
 470          * purging the list here.
 471          */
 472         BUG_ON(nbd->sock);
 473         BUG_ON(nbd->active_req);
 474
 475         while (!list_empty(&nbd->queue_head)) {
 476                 req = list_entry(nbd->queue_head.next, struct request,
 477                                  queuelist);
 478                 list_del_init(&req->queuelist);
 479                 req->errors++;
 480                 nbd_end_request(req);
 481         }
 482
 483         while (!list_empty(&nbd->waiting_queue)) {
 484                 req = list_entry(nbd->waiting_queue.next, struct request,
 485                                  queuelist);
 486                 list_del_init(&req->queuelist);
 487                 req->errors++;
 488                 nbd_end_request(req);
 489         }
 490 }
 491
 492
 493 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
 494 {
 495         if (req->cmd_type != REQ_TYPE_FS)
 496                 goto error_out;
 497
 498         nbd_cmd(req) = NBD_CMD_READ;
 499         if (rq_data_dir(req) == WRITE) {
 500                 if ((req->cmd_flags & REQ_DISCARD)) {
 501                         WARN_ON(!(nbd->flags & NBD_FLAG_SEND_TRIM));
 502                         nbd_cmd(req) = NBD_CMD_TRIM;
 503                 } else
 504                         nbd_cmd(req) = NBD_CMD_WRITE;
 505                 if (nbd->flags & NBD_FLAG_READ_ONLY) {
 506                         dev_err(disk_to_dev(nbd->disk),
 507                                 "Write on read-only\n");
 508                         goto error_out;
 509                 }
 510         }
 511
 512         if (req->cmd_flags & REQ_FLUSH) {
 513                 BUG_ON(unlikely(blk_rq_sectors(req)));
 514                 nbd_cmd(req) = NBD_CMD_FLUSH;
 515         }
 516
 517         req->errors = 0;
 518
 519         mutex_lock(&nbd->tx_lock);
 520         if (unlikely(!nbd->sock)) {
 521                 mutex_unlock(&nbd->tx_lock);
 522                 dev_err(disk_to_dev(nbd->disk),
 523                         "Attempted send on closed socket\n");
 524                 goto error_out;
 525         }
 526
 527         nbd->active_req = req;
 528
 529         if (nbd_send_req(nbd, req) != 0) {
 530                 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
 531                 req->errors++;
 532                 nbd_end_request(req);
 533         } else {
 534                 spin_lock(&nbd->queue_lock);
 535                 list_add_tail(&req->queuelist, &nbd->queue_head);
 536                 spin_unlock(&nbd->queue_lock);
 537         }
 538
 539         nbd->active_req = NULL;
 540         mutex_unlock(&nbd->tx_lock);
 541         wake_up_all(&nbd->active_wq);
 542
 543         return;
 544
 545 error_out:
 546         req->errors++;
 547         nbd_end_request(req);
 548 }
 549
 550 static int nbd_thread(void *data)
 551 {
 552         struct nbd_device *nbd = data;
 553         struct request *req;
 554
 555         set_user_nice(current, MIN_NICE);
 556         while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
 557                 /* wait for something to do */
 558                 wait_event_interruptible(nbd->waiting_wq,
 559                                          kthread_should_stop() ||
 560                                          !list_empty(&nbd->waiting_queue));
 561
 562                 /* extract request */
 563                 if (list_empty(&nbd->waiting_queue))
 564                         continue;
 565
 566                 spin_lock_irq(&nbd->queue_lock);
 567                 req = list_entry(nbd->waiting_queue.next, struct request,
 568                                  queuelist);
 569                 list_del_init(&req->queuelist);
 570                 spin_unlock_irq(&nbd->queue_lock);
 571
 572                 /* handle request */
 573                 nbd_handle_req(nbd, req);
 574         }
 575         return 0;
 576 }
 577
 578 /*
 579  * We always wait for result of write, for now. It would be nice to make it optional
 580  * in future
 581  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
 582  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 583  */
 584
 585 static void do_nbd_request(struct request_queue *q)
 586                 __releases(q->queue_lock) __acquires(q->queue_lock)
 587 {
 588         struct request *req;
 589
 590         while ((req = blk_fetch_request(q)) != NULL) {
 591                 struct nbd_device *nbd;
 592
 593                 spin_unlock_irq(q->queue_lock);
 594
 595                 dprintk(DBG_BLKDEV, "%s: request %p: dequeued (flags=%x)\n",
 596                                 req->rq_disk->disk_name, req, req->cmd_type);
 597
 598                 nbd = req->rq_disk->private_data;
 599
 600                 BUG_ON(nbd->magic != NBD_MAGIC);
 601
 602                 if (unlikely(!nbd->sock)) {
 603                         dev_err(disk_to_dev(nbd->disk),
 604                                 "Attempted send on closed socket\n");
 605                         req->errors++;
 606                         nbd_end_request(req);
 607                         spin_lock_irq(q->queue_lock);
 608                         continue;
 609                 }
 610
 611                 spin_lock_irq(&nbd->queue_lock);
 612                 list_add_tail(&req->queuelist, &nbd->waiting_queue);
 613                 spin_unlock_irq(&nbd->queue_lock);
 614
 615                 wake_up(&nbd->waiting_wq);
 616
 617                 spin_lock_irq(q->queue_lock);
 618         }
 619 }
 620
 621 /* Must be called with tx_lock held */
 622
 623 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 624                        unsigned int cmd, unsigned long arg)
 625 {
 626         switch (cmd) {
 627         case NBD_DISCONNECT: {
 628                 struct request sreq;
 629
 630                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 631                 if (!nbd->sock)
 632                         return -EINVAL;
 633
 634                 mutex_unlock(&nbd->tx_lock);
 635                 fsync_bdev(bdev);
 636                 mutex_lock(&nbd->tx_lock);
 637                 blk_rq_init(NULL, &sreq);
 638                 sreq.cmd_type = REQ_TYPE_SPECIAL;
 639                 nbd_cmd(&sreq) = NBD_CMD_DISC;
 640
 641                 /* Check again after getting mutex back.  */
 642                 if (!nbd->sock)
 643                         return -EINVAL;
 644
 645                 nbd->disconnect = 1;
 646
 647                 nbd_send_req(nbd, &sreq);
 648                 return 0;
 649         }
 650
 651         case NBD_CLEAR_SOCK: {
 652                 struct socket *sock = nbd->sock;
 653                 nbd->sock = NULL;
 654                 nbd_clear_que(nbd);
 655                 BUG_ON(!list_empty(&nbd->queue_head));
 656                 BUG_ON(!list_empty(&nbd->waiting_queue));
 657                 kill_bdev(bdev);
 658                 if (sock)
 659                         sockfd_put(sock);
 660                 return 0;
 661         }
 662
 663         case NBD_SET_SOCK: {
 664                 struct socket *sock;
 665                 int err;
 666                 if (nbd->sock)
 667                         return -EBUSY;
 668                 sock = sockfd_lookup(arg, &err);
 669                 if (sock) {
 670                         nbd->sock = sock;
 671                         if (max_part > 0)
 672                                 bdev->bd_invalidated = 1;
 673                         nbd->disconnect = 0; /* we're connected now */
 674                         return 0;
 675                 }
 676                 return -EINVAL;
 677         }
 678
 679         case NBD_SET_BLKSIZE:
 680                 nbd->blksize = arg;
 681                 nbd->bytesize &= ~(nbd->blksize-1);
 682                 bdev->bd_inode->i_size = nbd->bytesize;
 683                 set_blocksize(bdev, nbd->blksize);
 684                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 685                 return 0;
 686
 687         case NBD_SET_SIZE:
 688                 nbd->bytesize = arg & ~(nbd->blksize-1);
 689                 bdev->bd_inode->i_size = nbd->bytesize;
 690                 set_blocksize(bdev, nbd->blksize);
 691                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 692                 return 0;
 693
 694         case NBD_SET_TIMEOUT:
 695                 nbd->xmit_timeout = arg * HZ;
 696                 return 0;
 697
 698         case NBD_SET_FLAGS:
 699                 nbd->flags = arg;
 700                 return 0;
 701
 702         case NBD_SET_SIZE_BLOCKS:
 703                 nbd->bytesize = ((u64) arg) * nbd->blksize;
 704                 bdev->bd_inode->i_size = nbd->bytesize;
 705                 set_blocksize(bdev, nbd->blksize);
 706                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 707                 return 0;
 708
 709         case NBD_DO_IT: {
 710                 struct task_struct *thread;
 711                 struct socket *sock;
 712                 int error;
 713
 714                 if (nbd->pid)
 715                         return -EBUSY;
 716                 if (!nbd->sock)
 717                         return -EINVAL;
 718
 719                 mutex_unlock(&nbd->tx_lock);
 720
 721                 if (nbd->flags & NBD_FLAG_READ_ONLY)
 722                         set_device_ro(bdev, true);
 723                 if (nbd->flags & NBD_FLAG_SEND_TRIM)
 724                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
 725                                 nbd->disk->queue);
 726                 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
 727                         blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
 728                 else
 729                         blk_queue_flush(nbd->disk->queue, 0);
 730
 731                 thread = kthread_create(nbd_thread, nbd, "%s",
 732                                         nbd->disk->disk_name);
 733                 if (IS_ERR(thread)) {
 734                         mutex_lock(&nbd->tx_lock);
 735                         return PTR_ERR(thread);
 736                 }
 737                 wake_up_process(thread);
 738                 error = nbd_do_it(nbd);
 739                 kthread_stop(thread);
 740
 741                 mutex_lock(&nbd->tx_lock);
 742                 if (error)
 743                         return error;
 744                 sock_shutdown(nbd, 0);
 745                 sock = nbd->sock;
 746                 nbd->sock = NULL;
 747                 nbd_clear_que(nbd);
 748                 dev_warn(disk_to_dev(nbd->disk), "queue cleared\n");
 749                 kill_bdev(bdev);
 750                 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 751                 set_device_ro(bdev, false);
 752                 if (sock)
 753                         sockfd_put(sock);
 754                 nbd->flags = 0;
 755                 nbd->bytesize = 0;
 756                 bdev->bd_inode->i_size = 0;
 757                 set_capacity(nbd->disk, 0);
 758                 if (max_part > 0)
 759                         ioctl_by_bdev(bdev, BLKRRPART, 0);
 760                 if (nbd->disconnect) /* user requested, ignore socket errors */
 761                         return 0;
 762                 return nbd->harderror;
 763         }
 764
 765         case NBD_CLEAR_QUE:
 766                 /*
 767                  * This is for compatibility only.  The queue is always cleared
 768                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 769                  */
 770                 return 0;
 771
 772         case NBD_PRINT_DEBUG:
 773                 dev_info(disk_to_dev(nbd->disk),
 774                         "next = %p, prev = %p, head = %p\n",
 775                         nbd->queue_head.next, nbd->queue_head.prev,
 776                         &nbd->queue_head);
 777                 return 0;
 778         }
 779         return -ENOTTY;
 780 }
 781
 782 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 783                      unsigned int cmd, unsigned long arg)
 784 {
 785         struct nbd_device *nbd = bdev->bd_disk->private_data;
 786         int error;
 787
 788         if (!capable(CAP_SYS_ADMIN))
 789                 return -EPERM;
 790
 791         BUG_ON(nbd->magic != NBD_MAGIC);
 792
 793         /* Anyone capable of this syscall can do *real bad* things */
 794         dprintk(DBG_IOCTL, "%s: nbd_ioctl cmd=%s(0x%x) arg=%lu\n",
 795                 nbd->disk->disk_name, ioctl_cmd_to_ascii(cmd), cmd, arg);
 796
 797         mutex_lock(&nbd->tx_lock);
 798         error = __nbd_ioctl(bdev, nbd, cmd, arg);
 799         mutex_unlock(&nbd->tx_lock);
 800
 801         return error;
 802 }
 803
 804 static const struct block_device_operations nbd_fops =
 805 {
 806         .owner =        THIS_MODULE,
 807         .ioctl =        nbd_ioctl,
 808 };
 809
 810 /*
 811  * And here should be modules and kernel interface
 812  *  (Just smiley confuses emacs :-)
 813  */
 814
 815 static int __init nbd_init(void)
 816 {
 817         int err = -ENOMEM;
 818         int i;
 819         int part_shift;
 820
 821         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 822
 823         if (max_part < 0) {
 824                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
 825                 return -EINVAL;
 826         }
 827
 828         part_shift = 0;
 829         if (max_part > 0) {
 830                 part_shift = fls(max_part);
 831
 832                 /*
 833                  * Adjust max_part according to part_shift as it is exported
 834                  * to user space so that user can know the max number of
 835                  * partition kernel should be able to manage.
 836                  *
 837                  * Note that -1 is required because partition 0 is reserved
 838                  * for the whole disk.
 839                  */
 840                 max_part = (1UL << part_shift) - 1;
 841         }
 842
 843         if ((1UL << part_shift) > DISK_MAX_PARTS)
 844                 return -EINVAL;
 845
 846         if (nbds_max > 1UL << (MINORBITS - part_shift))
 847                 return -EINVAL;
 848
 849         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
 850         if (!nbd_dev)
 851                 return -ENOMEM;
 852
 853         for (i = 0; i < nbds_max; i++) {
 854                 struct gendisk *disk = alloc_disk(1 << part_shift);
 855                 if (!disk)
 856                         goto out;
 857                 nbd_dev[i].disk = disk;
 858                 /*
 859                  * The new linux 2.5 block layer implementation requires
 860                  * every gendisk to have its very own request_queue struct.
 861                  * These structs are big so we dynamically allocate them.
 862                  */
 863                 disk->queue = blk_init_queue(do_nbd_request, &nbd_lock);
 864                 if (!disk->queue) {
 865                         put_disk(disk);
 866                         goto out;
 867                 }
 868                 /*
 869                  * Tell the block layer that we are not a rotational device
 870                  */
 871                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
 872                 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
 873                 disk->queue->limits.discard_granularity = 512;
 874                 disk->queue->limits.max_discard_sectors = UINT_MAX;
 875                 disk->queue->limits.discard_zeroes_data = 0;
 876                 blk_queue_max_hw_sectors(disk->queue, 65536);
 877                 disk->queue->limits.max_sectors = 256;
 878         }
 879
 880         if (register_blkdev(NBD_MAJOR, "nbd")) {
 881                 err = -EIO;
 882                 goto out;
 883         }
 884
 885         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
 886         dprintk(DBG_INIT, "nbd: debugflags=0x%x\n", debugflags);
 887
 888         for (i = 0; i < nbds_max; i++) {
 889                 struct gendisk *disk = nbd_dev[i].disk;
 890                 nbd_dev[i].magic = NBD_MAGIC;
 891                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
 892                 spin_lock_init(&nbd_dev[i].queue_lock);
 893                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
 894                 mutex_init(&nbd_dev[i].tx_lock);
 895                 init_waitqueue_head(&nbd_dev[i].active_wq);
 896                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
 897                 nbd_dev[i].blksize = 1024;
 898                 nbd_dev[i].bytesize = 0;
 899                 disk->major = NBD_MAJOR;
 900                 disk->first_minor = i << part_shift;
 901                 disk->fops = &nbd_fops;
 902                 disk->private_data = &nbd_dev[i];
 903                 sprintf(disk->disk_name, "nbd%d", i);
 904                 set_capacity(disk, 0);
 905                 add_disk(disk);
 906         }
 907
 908         return 0;
 909 out:
 910         while (i--) {
 911                 blk_cleanup_queue(nbd_dev[i].disk->queue);
 912                 put_disk(nbd_dev[i].disk);
 913         }
 914         kfree(nbd_dev);
 915         return err;
 916 }
 917
 918 static void __exit nbd_cleanup(void)
 919 {
 920         int i;
 921         for (i = 0; i < nbds_max; i++) {
 922                 struct gendisk *disk = nbd_dev[i].disk;
 923                 nbd_dev[i].magic = 0;
 924                 if (disk) {
 925                         del_gendisk(disk);
 926                         blk_cleanup_queue(disk->queue);
 927                         put_disk(disk);
 928                 }
 929         }
 930         unregister_blkdev(NBD_MAJOR, "nbd");
 931         kfree(nbd_dev);
 932         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
 933 }
 934
 935 module_init(nbd_init);
 936 module_exit(nbd_cleanup);
 937
 938 MODULE_DESCRIPTION("Network Block Device");
 939 MODULE_LICENSE("GPL");
 940
 941 module_param(nbds_max, int, 0444);
 942 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
 943 module_param(max_part, int, 0444);
 944 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");
 945 #ifndef NDEBUG
 946 module_param(debugflags, int, 0644);
 947 MODULE_PARM_DESC(debugflags, "flags for controlling debug output");
 948 #endif