drivers/block/nbd.c

   1 /*
   2  * Network block device - make block devices work over TCP
   3  *
   4  * Note that you can not swap over this thing, yet. Seems to work but
   5  * deadlocks sometimes - you can not swap over TCP in general.
   6  *
   7  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   8  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
   9  *
  10  * This file is released under GPLv2 or later.
  11  *
  12  * (part of code stolen from loop.c)
  13  */
  14
  15 #include <linux/major.h>
  16
  17 #include <linux/blkdev.h>
  18 #include <linux/module.h>
  19 #include <linux/init.h>
  20 #include <linux/sched.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/err.h>
  30 #include <linux/kernel.h>
  31 #include <linux/slab.h>
  32 #include <net/sock.h>
  33 #include <linux/net.h>
  34 #include <linux/kthread.h>
  35 #include <linux/types.h>
  36
  37 #include <asm/uaccess.h>
  38 #include <asm/types.h>
  39
  40 #include <linux/nbd.h>
  41
  42 struct nbd_device {
  43         int flags;
  44         struct socket * sock;   /* If == NULL, device is not ready, yet */
  45         int magic;
  46
  47         spinlock_t queue_lock;
  48         struct list_head queue_head;    /* Requests waiting result */
  49         struct request *active_req;
  50         wait_queue_head_t active_wq;
  51         struct list_head waiting_queue; /* Requests to be sent */
  52         wait_queue_head_t waiting_wq;
  53
  54         struct mutex tx_lock;
  55         struct gendisk *disk;
  56         int blksize;
  57         loff_t bytesize;
  58         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         struct timer_list timeout_timer;
  63         struct task_struct *task_recv;
  64         struct task_struct *task_send;
  65 };
  66
  67 #define NBD_MAGIC 0x68797548
  68
  69 static unsigned int nbds_max = 16;
  70 static struct nbd_device *nbd_dev;
  71 static int max_part;
  72
  73 /*
  74  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  75  * 1. Each NIC is essentially a synchronization point for all servers
  76  *    accessed through that NIC so there's no need to have more locks
  77  *    than NICs anyway.
  78  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  79  *    down each lock to the point where they're actually slower than just
  80  *    a single lock.
  81  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  82  */
  83 static DEFINE_SPINLOCK(nbd_lock);
  84
  85 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
  86 {
  87         return disk_to_dev(nbd->disk);
  88 }
  89
  90 static const char *nbdcmd_to_ascii(int cmd)
  91 {
  92         switch (cmd) {
  93         case  NBD_CMD_READ: return "read";
  94         case NBD_CMD_WRITE: return "write";
  95         case  NBD_CMD_DISC: return "disconnect";
  96         case NBD_CMD_FLUSH: return "flush";
  97         case  NBD_CMD_TRIM: return "trim/discard";
  98         }
  99         return "invalid";
 100 }
 101
 102 static void nbd_end_request(struct nbd_device *nbd, struct request *req)
 103 {
 104         int error = req->errors ? -EIO : 0;
 105         struct request_queue *q = req->q;
 106         unsigned long flags;
 107
 108         dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
 109                 error ? "failed" : "done");
 110
 111         spin_lock_irqsave(q->queue_lock, flags);
 112         __blk_end_request_all(req, error);
 113         spin_unlock_irqrestore(q->queue_lock, flags);
 114 }
 115
 116 /*
 117  * Forcibly shutdown the socket causing all listeners to error
 118  */
 119 static void sock_shutdown(struct nbd_device *nbd)
 120 {
 121         if (!nbd->sock)
 122                 return;
 123
 124         dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
 125         kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
 126         nbd->sock = NULL;
 127         del_timer_sync(&nbd->timeout_timer);
 128 }
 129
 130 static void nbd_xmit_timeout(unsigned long arg)
 131 {
 132         struct nbd_device *nbd = (struct nbd_device *)arg;
 133         struct task_struct *task;
 134
 135         if (list_empty(&nbd->queue_head))
 136                 return;
 137
 138         nbd->disconnect = 1;
 139
 140         task = READ_ONCE(nbd->task_recv);
 141         if (task)
 142                 force_sig(SIGKILL, task);
 143
 144         task = READ_ONCE(nbd->task_send);
 145         if (task)
 146                 force_sig(SIGKILL, nbd->task_send);
 147
 148         dev_err(nbd_to_dev(nbd), "Connection timed out, killed receiver and sender, shutting down connection\n");
 149 }
 150
 151 /*
 152  *  Send or receive packet.
 153  */
 154 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
 155                 int msg_flags)
 156 {
 157         struct socket *sock = nbd->sock;
 158         int result;
 159         struct msghdr msg;
 160         struct kvec iov;
 161         sigset_t blocked, oldset;
 162         unsigned long pflags = current->flags;
 163
 164         if (unlikely(!sock)) {
 165                 dev_err(disk_to_dev(nbd->disk),
 166                         "Attempted %s on closed socket in sock_xmit\n",
 167                         (send ? "send" : "recv"));
 168                 return -EINVAL;
 169         }
 170
 171         /* Allow interception of SIGKILL only
 172          * Don't allow other signals to interrupt the transmission */
 173         siginitsetinv(&blocked, sigmask(SIGKILL));
 174         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 175
 176         current->flags |= PF_MEMALLOC;
 177         do {
 178                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 179                 iov.iov_base = buf;
 180                 iov.iov_len = size;
 181                 msg.msg_name = NULL;
 182                 msg.msg_namelen = 0;
 183                 msg.msg_control = NULL;
 184                 msg.msg_controllen = 0;
 185                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 186
 187                 if (send)
 188                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 189                 else
 190                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 191                                                 msg.msg_flags);
 192
 193                 if (result <= 0) {
 194                         if (result == 0)
 195                                 result = -EPIPE; /* short read */
 196                         break;
 197                 }
 198                 size -= result;
 199                 buf += result;
 200         } while (size > 0);
 201
 202         sigprocmask(SIG_SETMASK, &oldset, NULL);
 203         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 204
 205         if (!send && nbd->xmit_timeout)
 206                 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
 207
 208         return result;
 209 }
 210
 211 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
 212                 int flags)
 213 {
 214         int result;
 215         void *kaddr = kmap(bvec->bv_page);
 216         result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
 217                            bvec->bv_len, flags);
 218         kunmap(bvec->bv_page);
 219         return result;
 220 }
 221
 222 /* always call with the tx_lock held */
 223 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
 224 {
 225         int result, flags;
 226         struct nbd_request request;
 227         unsigned long size = blk_rq_bytes(req);
 228         u32 type;
 229
 230         if (req->cmd_type == REQ_TYPE_DRV_PRIV)
 231                 type = NBD_CMD_DISC;
 232         else if (req->cmd_flags & REQ_DISCARD)
 233                 type = NBD_CMD_TRIM;
 234         else if (req->cmd_flags & REQ_FLUSH)
 235                 type = NBD_CMD_FLUSH;
 236         else if (rq_data_dir(req) == WRITE)
 237                 type = NBD_CMD_WRITE;
 238         else
 239                 type = NBD_CMD_READ;
 240
 241         memset(&request, 0, sizeof(request));
 242         request.magic = htonl(NBD_REQUEST_MAGIC);
 243         request.type = htonl(type);
 244         if (type != NBD_CMD_FLUSH && type != NBD_CMD_DISC) {
 245                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 246                 request.len = htonl(size);
 247         }
 248         memcpy(request.handle, &req, sizeof(req));
 249
 250         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 251                 req, nbdcmd_to_ascii(type),
 252                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 253         result = sock_xmit(nbd, 1, &request, sizeof(request),
 254                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
 255         if (result <= 0) {
 256                 dev_err(disk_to_dev(nbd->disk),
 257                         "Send control failed (result %d)\n", result);
 258                 return -EIO;
 259         }
 260
 261         if (type == NBD_CMD_WRITE) {
 262                 struct req_iterator iter;
 263                 struct bio_vec bvec;
 264                 /*
 265                  * we are really probing at internals to determine
 266                  * whether to set MSG_MORE or not...
 267                  */
 268                 rq_for_each_segment(bvec, req, iter) {
 269                         flags = 0;
 270                         if (!rq_iter_last(bvec, iter))
 271                                 flags = MSG_MORE;
 272                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 273                                 req, bvec.bv_len);
 274                         result = sock_send_bvec(nbd, &bvec, flags);
 275                         if (result <= 0) {
 276                                 dev_err(disk_to_dev(nbd->disk),
 277                                         "Send data failed (result %d)\n",
 278                                         result);
 279                                 return -EIO;
 280                         }
 281                 }
 282         }
 283         return 0;
 284 }
 285
 286 static struct request *nbd_find_request(struct nbd_device *nbd,
 287                                         struct request *xreq)
 288 {
 289         struct request *req, *tmp;
 290         int err;
 291
 292         err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
 293         if (unlikely(err))
 294                 return ERR_PTR(err);
 295
 296         spin_lock(&nbd->queue_lock);
 297         list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
 298                 if (req != xreq)
 299                         continue;
 300                 list_del_init(&req->queuelist);
 301                 spin_unlock(&nbd->queue_lock);
 302                 return req;
 303         }
 304         spin_unlock(&nbd->queue_lock);
 305
 306         return ERR_PTR(-ENOENT);
 307 }
 308
 309 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
 310 {
 311         int result;
 312         void *kaddr = kmap(bvec->bv_page);
 313         result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 314                         MSG_WAITALL);
 315         kunmap(bvec->bv_page);
 316         return result;
 317 }
 318
 319 /* NULL returned = something went wrong, inform userspace */
 320 static struct request *nbd_read_stat(struct nbd_device *nbd)
 321 {
 322         int result;
 323         struct nbd_reply reply;
 324         struct request *req;
 325
 326         reply.magic = 0;
 327         result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
 328         if (result <= 0) {
 329                 dev_err(disk_to_dev(nbd->disk),
 330                         "Receive control failed (result %d)\n", result);
 331                 return ERR_PTR(result);
 332         }
 333
 334         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 335                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 336                                 (unsigned long)ntohl(reply.magic));
 337                 return ERR_PTR(-EPROTO);
 338         }
 339
 340         req = nbd_find_request(nbd, *(struct request **)reply.handle);
 341         if (IS_ERR(req)) {
 342                 result = PTR_ERR(req);
 343                 if (result != -ENOENT)
 344                         return ERR_PTR(result);
 345
 346                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
 347                         reply.handle);
 348                 return ERR_PTR(-EBADR);
 349         }
 350
 351         if (ntohl(reply.error)) {
 352                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 353                         ntohl(reply.error));
 354                 req->errors++;
 355                 return req;
 356         }
 357
 358         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 359         if (rq_data_dir(req) != WRITE) {
 360                 struct req_iterator iter;
 361                 struct bio_vec bvec;
 362
 363                 rq_for_each_segment(bvec, req, iter) {
 364                         result = sock_recv_bvec(nbd, &bvec);
 365                         if (result <= 0) {
 366                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 367                                         result);
 368                                 req->errors++;
 369                                 return req;
 370                         }
 371                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 372                                 req, bvec.bv_len);
 373                 }
 374         }
 375         return req;
 376 }
 377
 378 static ssize_t pid_show(struct device *dev,
 379                         struct device_attribute *attr, char *buf)
 380 {
 381         struct gendisk *disk = dev_to_disk(dev);
 382
 383         return sprintf(buf, "%ld\n",
 384                 (long) ((struct nbd_device *)disk->private_data)->pid);
 385 }
 386
 387 static struct device_attribute pid_attr = {
 388         .attr = { .name = "pid", .mode = S_IRUGO},
 389         .show = pid_show,
 390 };
 391
 392 static int nbd_do_it(struct nbd_device *nbd)
 393 {
 394         struct request *req;
 395         int ret;
 396
 397         BUG_ON(nbd->magic != NBD_MAGIC);
 398
 399         sk_set_memalloc(nbd->sock->sk);
 400         nbd->pid = task_pid_nr(current);
 401         ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 402         if (ret) {
 403                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 404                 nbd->pid = 0;
 405                 return ret;
 406         }
 407
 408         nbd->task_recv = current;
 409
 410         while (1) {
 411                 req = nbd_read_stat(nbd);
 412                 if (IS_ERR(req)) {
 413                         ret = PTR_ERR(req);
 414                         break;
 415                 }
 416
 417                 nbd_end_request(nbd, req);
 418         }
 419
 420         nbd->task_recv = NULL;
 421
 422         if (signal_pending(current)) {
 423                 siginfo_t info;
 424
 425                 ret = dequeue_signal_lock(current, &current->blocked, &info);
 426                 dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
 427                          task_pid_nr(current), current->comm, ret);
 428                 mutex_lock(&nbd->tx_lock);
 429                 sock_shutdown(nbd);
 430                 mutex_unlock(&nbd->tx_lock);
 431                 ret = -ETIMEDOUT;
 432         }
 433
 434         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 435         nbd->pid = 0;
 436         return ret;
 437 }
 438
 439 static void nbd_clear_que(struct nbd_device *nbd)
 440 {
 441         struct request *req;
 442
 443         BUG_ON(nbd->magic != NBD_MAGIC);
 444
 445         /*
 446          * Because we have set nbd->sock to NULL under the tx_lock, all
 447          * modifications to the list must have completed by now.  For
 448          * the same reason, the active_req must be NULL.
 449          *
 450          * As a consequence, we don't need to take the spin lock while
 451          * purging the list here.
 452          */
 453         BUG_ON(nbd->sock);
 454         BUG_ON(nbd->active_req);
 455
 456         while (!list_empty(&nbd->queue_head)) {
 457                 req = list_entry(nbd->queue_head.next, struct request,
 458                                  queuelist);
 459                 list_del_init(&req->queuelist);
 460                 req->errors++;
 461                 nbd_end_request(nbd, req);
 462         }
 463
 464         while (!list_empty(&nbd->waiting_queue)) {
 465                 req = list_entry(nbd->waiting_queue.next, struct request,
 466                                  queuelist);
 467                 list_del_init(&req->queuelist);
 468                 req->errors++;
 469                 nbd_end_request(nbd, req);
 470         }
 471         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 472 }
 473
 474
 475 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
 476 {
 477         if (req->cmd_type != REQ_TYPE_FS)
 478                 goto error_out;
 479
 480         if (rq_data_dir(req) == WRITE &&
 481             (nbd->flags & NBD_FLAG_READ_ONLY)) {
 482                 dev_err(disk_to_dev(nbd->disk),
 483                         "Write on read-only\n");
 484                 goto error_out;
 485         }
 486
 487         req->errors = 0;
 488
 489         mutex_lock(&nbd->tx_lock);
 490         if (unlikely(!nbd->sock)) {
 491                 mutex_unlock(&nbd->tx_lock);
 492                 dev_err(disk_to_dev(nbd->disk),
 493                         "Attempted send on closed socket\n");
 494                 goto error_out;
 495         }
 496
 497         nbd->active_req = req;
 498
 499         if (nbd->xmit_timeout && list_empty_careful(&nbd->queue_head))
 500                 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
 501
 502         if (nbd_send_req(nbd, req) != 0) {
 503                 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
 504                 req->errors++;
 505                 nbd_end_request(nbd, req);
 506         } else {
 507                 spin_lock(&nbd->queue_lock);
 508                 list_add_tail(&req->queuelist, &nbd->queue_head);
 509                 spin_unlock(&nbd->queue_lock);
 510         }
 511
 512         nbd->active_req = NULL;
 513         mutex_unlock(&nbd->tx_lock);
 514         wake_up_all(&nbd->active_wq);
 515
 516         return;
 517
 518 error_out:
 519         req->errors++;
 520         nbd_end_request(nbd, req);
 521 }
 522
 523 static int nbd_thread(void *data)
 524 {
 525         struct nbd_device *nbd = data;
 526         struct request *req;
 527
 528         nbd->task_send = current;
 529
 530         set_user_nice(current, MIN_NICE);
 531         while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
 532                 /* wait for something to do */
 533                 wait_event_interruptible(nbd->waiting_wq,
 534                                          kthread_should_stop() ||
 535                                          !list_empty(&nbd->waiting_queue));
 536
 537                 if (signal_pending(current)) {
 538                         siginfo_t info;
 539                         int ret;
 540
 541                         ret = dequeue_signal_lock(current, &current->blocked,
 542                                                   &info);
 543                         dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
 544                                  task_pid_nr(current), current->comm, ret);
 545                         mutex_lock(&nbd->tx_lock);
 546                         sock_shutdown(nbd);
 547                         mutex_unlock(&nbd->tx_lock);
 548                         break;
 549                 }
 550
 551                 /* extract request */
 552                 if (list_empty(&nbd->waiting_queue))
 553                         continue;
 554
 555                 spin_lock_irq(&nbd->queue_lock);
 556                 req = list_entry(nbd->waiting_queue.next, struct request,
 557                                  queuelist);
 558                 list_del_init(&req->queuelist);
 559                 spin_unlock_irq(&nbd->queue_lock);
 560
 561                 /* handle request */
 562                 nbd_handle_req(nbd, req);
 563         }
 564
 565         nbd->task_send = NULL;
 566
 567         return 0;
 568 }
 569
 570 /*
 571  * We always wait for result of write, for now. It would be nice to make it optional
 572  * in future
 573  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
 574  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 575  */
 576
 577 static void do_nbd_request(struct request_queue *q)
 578                 __releases(q->queue_lock) __acquires(q->queue_lock)
 579 {
 580         struct request *req;
 581
 582         while ((req = blk_fetch_request(q)) != NULL) {
 583                 struct nbd_device *nbd;
 584
 585                 spin_unlock_irq(q->queue_lock);
 586
 587                 nbd = req->rq_disk->private_data;
 588
 589                 BUG_ON(nbd->magic != NBD_MAGIC);
 590
 591                 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
 592                         req, req->cmd_type);
 593
 594                 if (unlikely(!nbd->sock)) {
 595                         dev_err(disk_to_dev(nbd->disk),
 596                                 "Attempted send on closed socket\n");
 597                         req->errors++;
 598                         nbd_end_request(nbd, req);
 599                         spin_lock_irq(q->queue_lock);
 600                         continue;
 601                 }
 602
 603                 spin_lock_irq(&nbd->queue_lock);
 604                 list_add_tail(&req->queuelist, &nbd->waiting_queue);
 605                 spin_unlock_irq(&nbd->queue_lock);
 606
 607                 wake_up(&nbd->waiting_wq);
 608
 609                 spin_lock_irq(q->queue_lock);
 610         }
 611 }
 612
 613 /* Must be called with tx_lock held */
 614
 615 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 616                        unsigned int cmd, unsigned long arg)
 617 {
 618         switch (cmd) {
 619         case NBD_DISCONNECT: {
 620                 struct request sreq;
 621
 622                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 623                 if (!nbd->sock)
 624                         return -EINVAL;
 625
 626                 mutex_unlock(&nbd->tx_lock);
 627                 fsync_bdev(bdev);
 628                 mutex_lock(&nbd->tx_lock);
 629                 blk_rq_init(NULL, &sreq);
 630                 sreq.cmd_type = REQ_TYPE_DRV_PRIV;
 631
 632                 /* Check again after getting mutex back.  */
 633                 if (!nbd->sock)
 634                         return -EINVAL;
 635
 636                 nbd->disconnect = 1;
 637
 638                 nbd_send_req(nbd, &sreq);
 639                 return 0;
 640         }
 641
 642         case NBD_CLEAR_SOCK: {
 643                 struct socket *sock = nbd->sock;
 644                 nbd->sock = NULL;
 645                 nbd_clear_que(nbd);
 646                 BUG_ON(!list_empty(&nbd->queue_head));
 647                 BUG_ON(!list_empty(&nbd->waiting_queue));
 648                 kill_bdev(bdev);
 649                 if (sock)
 650                         sockfd_put(sock);
 651                 return 0;
 652         }
 653
 654         case NBD_SET_SOCK: {
 655                 struct socket *sock;
 656                 int err;
 657                 if (nbd->sock)
 658                         return -EBUSY;
 659                 sock = sockfd_lookup(arg, &err);
 660                 if (sock) {
 661                         nbd->sock = sock;
 662                         if (max_part > 0)
 663                                 bdev->bd_invalidated = 1;
 664                         nbd->disconnect = 0; /* we're connected now */
 665                         return 0;
 666                 }
 667                 return -EINVAL;
 668         }
 669
 670         case NBD_SET_BLKSIZE:
 671                 nbd->blksize = arg;
 672                 nbd->bytesize &= ~(nbd->blksize-1);
 673                 bdev->bd_inode->i_size = nbd->bytesize;
 674                 set_blocksize(bdev, nbd->blksize);
 675                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 676                 return 0;
 677
 678         case NBD_SET_SIZE:
 679                 nbd->bytesize = arg & ~(nbd->blksize-1);
 680                 bdev->bd_inode->i_size = nbd->bytesize;
 681                 set_blocksize(bdev, nbd->blksize);
 682                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 683                 return 0;
 684
 685         case NBD_SET_TIMEOUT:
 686                 nbd->xmit_timeout = arg * HZ;
 687                 if (arg)
 688                         mod_timer(&nbd->timeout_timer,
 689                                   jiffies + nbd->xmit_timeout);
 690                 else
 691                         del_timer_sync(&nbd->timeout_timer);
 692
 693                 return 0;
 694
 695         case NBD_SET_FLAGS:
 696                 nbd->flags = arg;
 697                 return 0;
 698
 699         case NBD_SET_SIZE_BLOCKS:
 700                 nbd->bytesize = ((u64) arg) * nbd->blksize;
 701                 bdev->bd_inode->i_size = nbd->bytesize;
 702                 set_blocksize(bdev, nbd->blksize);
 703                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 704                 return 0;
 705
 706         case NBD_DO_IT: {
 707                 struct task_struct *thread;
 708                 struct socket *sock;
 709                 int error;
 710
 711                 if (nbd->pid)
 712                         return -EBUSY;
 713                 if (!nbd->sock)
 714                         return -EINVAL;
 715
 716                 mutex_unlock(&nbd->tx_lock);
 717
 718                 if (nbd->flags & NBD_FLAG_READ_ONLY)
 719                         set_device_ro(bdev, true);
 720                 if (nbd->flags & NBD_FLAG_SEND_TRIM)
 721                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
 722                                 nbd->disk->queue);
 723                 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
 724                         blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
 725                 else
 726                         blk_queue_flush(nbd->disk->queue, 0);
 727
 728                 thread = kthread_run(nbd_thread, nbd, "%s",
 729                                      nbd->disk->disk_name);
 730                 if (IS_ERR(thread)) {
 731                         mutex_lock(&nbd->tx_lock);
 732                         return PTR_ERR(thread);
 733                 }
 734
 735                 error = nbd_do_it(nbd);
 736                 kthread_stop(thread);
 737
 738                 mutex_lock(&nbd->tx_lock);
 739
 740                 sock_shutdown(nbd);
 741                 sock = nbd->sock;
 742                 nbd->sock = NULL;
 743                 nbd_clear_que(nbd);
 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 error;
 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)");