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 #include <linux/debugfs.h>
  37
  38 #include <asm/uaccess.h>
  39 #include <asm/types.h>
  40
  41 #include <linux/nbd.h>
  42
  43 struct nbd_device {
  44         u32 flags;
  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         int xmit_timeout;
  60         bool 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 #if IS_ENABLED(CONFIG_DEBUG_FS)
  67         struct dentry *dbg_dir;
  68 #endif
  69 };
  70
  71 #if IS_ENABLED(CONFIG_DEBUG_FS)
  72 static struct dentry *nbd_dbg_dir;
  73 #endif
  74
  75 #define nbd_name(nbd) ((nbd)->disk->disk_name)
  76
  77 #define NBD_MAGIC 0x68797548
  78
  79 static unsigned int nbds_max = 16;
  80 static struct nbd_device *nbd_dev;
  81 static int max_part;
  82
  83 /*
  84  * Use just one lock (or at most 1 per NIC). Two arguments for this:
  85  * 1. Each NIC is essentially a synchronization point for all servers
  86  *    accessed through that NIC so there's no need to have more locks
  87  *    than NICs anyway.
  88  * 2. More locks lead to more "Dirty cache line bouncing" which will slow
  89  *    down each lock to the point where they're actually slower than just
  90  *    a single lock.
  91  * Thanks go to Jens Axboe and Al Viro for their LKML emails explaining this!
  92  */
  93 static DEFINE_SPINLOCK(nbd_lock);
  94
  95 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
  96 {
  97         return disk_to_dev(nbd->disk);
  98 }
  99
 100 static const char *nbdcmd_to_ascii(int cmd)
 101 {
 102         switch (cmd) {
 103         case  NBD_CMD_READ: return "read";
 104         case NBD_CMD_WRITE: return "write";
 105         case  NBD_CMD_DISC: return "disconnect";
 106         case NBD_CMD_FLUSH: return "flush";
 107         case  NBD_CMD_TRIM: return "trim/discard";
 108         }
 109         return "invalid";
 110 }
 111
 112 static void nbd_end_request(struct nbd_device *nbd, struct request *req)
 113 {
 114         int error = req->errors ? -EIO : 0;
 115         struct request_queue *q = req->q;
 116         unsigned long flags;
 117
 118         dev_dbg(nbd_to_dev(nbd), "request %p: %s\n", req,
 119                 error ? "failed" : "done");
 120
 121         spin_lock_irqsave(q->queue_lock, flags);
 122         __blk_end_request_all(req, error);
 123         spin_unlock_irqrestore(q->queue_lock, flags);
 124 }
 125
 126 /*
 127  * Forcibly shutdown the socket causing all listeners to error
 128  */
 129 static void sock_shutdown(struct nbd_device *nbd)
 130 {
 131         if (!nbd->sock)
 132                 return;
 133
 134         dev_warn(disk_to_dev(nbd->disk), "shutting down socket\n");
 135         kernel_sock_shutdown(nbd->sock, SHUT_RDWR);
 136         nbd->sock = NULL;
 137         del_timer_sync(&nbd->timeout_timer);
 138 }
 139
 140 static void nbd_xmit_timeout(unsigned long arg)
 141 {
 142         struct nbd_device *nbd = (struct nbd_device *)arg;
 143         struct task_struct *task;
 144
 145         if (list_empty(&nbd->queue_head))
 146                 return;
 147
 148         nbd->disconnect = true;
 149
 150         task = READ_ONCE(nbd->task_recv);
 151         if (task)
 152                 force_sig(SIGKILL, task);
 153
 154         task = READ_ONCE(nbd->task_send);
 155         if (task)
 156                 force_sig(SIGKILL, nbd->task_send);
 157
 158         dev_err(nbd_to_dev(nbd), "Connection timed out, killed receiver and sender, shutting down connection\n");
 159 }
 160
 161 /*
 162  *  Send or receive packet.
 163  */
 164 static int sock_xmit(struct nbd_device *nbd, int send, void *buf, int size,
 165                 int msg_flags)
 166 {
 167         struct socket *sock = nbd->sock;
 168         int result;
 169         struct msghdr msg;
 170         struct kvec iov;
 171         sigset_t blocked, oldset;
 172         unsigned long pflags = current->flags;
 173
 174         if (unlikely(!sock)) {
 175                 dev_err(disk_to_dev(nbd->disk),
 176                         "Attempted %s on closed socket in sock_xmit\n",
 177                         (send ? "send" : "recv"));
 178                 return -EINVAL;
 179         }
 180
 181         /* Allow interception of SIGKILL only
 182          * Don't allow other signals to interrupt the transmission */
 183         siginitsetinv(&blocked, sigmask(SIGKILL));
 184         sigprocmask(SIG_SETMASK, &blocked, &oldset);
 185
 186         current->flags |= PF_MEMALLOC;
 187         do {
 188                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 189                 iov.iov_base = buf;
 190                 iov.iov_len = size;
 191                 msg.msg_name = NULL;
 192                 msg.msg_namelen = 0;
 193                 msg.msg_control = NULL;
 194                 msg.msg_controllen = 0;
 195                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 196
 197                 if (send)
 198                         result = kernel_sendmsg(sock, &msg, &iov, 1, size);
 199                 else
 200                         result = kernel_recvmsg(sock, &msg, &iov, 1, size,
 201                                                 msg.msg_flags);
 202
 203                 if (result <= 0) {
 204                         if (result == 0)
 205                                 result = -EPIPE; /* short read */
 206                         break;
 207                 }
 208                 size -= result;
 209                 buf += result;
 210         } while (size > 0);
 211
 212         sigprocmask(SIG_SETMASK, &oldset, NULL);
 213         tsk_restore_flags(current, pflags, PF_MEMALLOC);
 214
 215         if (!send && nbd->xmit_timeout)
 216                 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
 217
 218         return result;
 219 }
 220
 221 static inline int sock_send_bvec(struct nbd_device *nbd, struct bio_vec *bvec,
 222                 int flags)
 223 {
 224         int result;
 225         void *kaddr = kmap(bvec->bv_page);
 226         result = sock_xmit(nbd, 1, kaddr + bvec->bv_offset,
 227                            bvec->bv_len, flags);
 228         kunmap(bvec->bv_page);
 229         return result;
 230 }
 231
 232 /* always call with the tx_lock held */
 233 static int nbd_send_req(struct nbd_device *nbd, struct request *req)
 234 {
 235         int result, flags;
 236         struct nbd_request request;
 237         unsigned long size = blk_rq_bytes(req);
 238         u32 type;
 239
 240         if (req->cmd_type == REQ_TYPE_DRV_PRIV)
 241                 type = NBD_CMD_DISC;
 242         else if (req->cmd_flags & REQ_DISCARD)
 243                 type = NBD_CMD_TRIM;
 244         else if (req->cmd_flags & REQ_FLUSH)
 245                 type = NBD_CMD_FLUSH;
 246         else if (rq_data_dir(req) == WRITE)
 247                 type = NBD_CMD_WRITE;
 248         else
 249                 type = NBD_CMD_READ;
 250
 251         memset(&request, 0, sizeof(request));
 252         request.magic = htonl(NBD_REQUEST_MAGIC);
 253         request.type = htonl(type);
 254         if (type != NBD_CMD_FLUSH && type != NBD_CMD_DISC) {
 255                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 256                 request.len = htonl(size);
 257         }
 258         memcpy(request.handle, &req, sizeof(req));
 259
 260         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 261                 req, nbdcmd_to_ascii(type),
 262                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 263         result = sock_xmit(nbd, 1, &request, sizeof(request),
 264                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0);
 265         if (result <= 0) {
 266                 dev_err(disk_to_dev(nbd->disk),
 267                         "Send control failed (result %d)\n", result);
 268                 return -EIO;
 269         }
 270
 271         if (type == NBD_CMD_WRITE) {
 272                 struct req_iterator iter;
 273                 struct bio_vec bvec;
 274                 /*
 275                  * we are really probing at internals to determine
 276                  * whether to set MSG_MORE or not...
 277                  */
 278                 rq_for_each_segment(bvec, req, iter) {
 279                         flags = 0;
 280                         if (!rq_iter_last(bvec, iter))
 281                                 flags = MSG_MORE;
 282                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 283                                 req, bvec.bv_len);
 284                         result = sock_send_bvec(nbd, &bvec, flags);
 285                         if (result <= 0) {
 286                                 dev_err(disk_to_dev(nbd->disk),
 287                                         "Send data failed (result %d)\n",
 288                                         result);
 289                                 return -EIO;
 290                         }
 291                 }
 292         }
 293         return 0;
 294 }
 295
 296 static struct request *nbd_find_request(struct nbd_device *nbd,
 297                                         struct request *xreq)
 298 {
 299         struct request *req, *tmp;
 300         int err;
 301
 302         err = wait_event_interruptible(nbd->active_wq, nbd->active_req != xreq);
 303         if (unlikely(err))
 304                 return ERR_PTR(err);
 305
 306         spin_lock(&nbd->queue_lock);
 307         list_for_each_entry_safe(req, tmp, &nbd->queue_head, queuelist) {
 308                 if (req != xreq)
 309                         continue;
 310                 list_del_init(&req->queuelist);
 311                 spin_unlock(&nbd->queue_lock);
 312                 return req;
 313         }
 314         spin_unlock(&nbd->queue_lock);
 315
 316         return ERR_PTR(-ENOENT);
 317 }
 318
 319 static inline int sock_recv_bvec(struct nbd_device *nbd, struct bio_vec *bvec)
 320 {
 321         int result;
 322         void *kaddr = kmap(bvec->bv_page);
 323         result = sock_xmit(nbd, 0, kaddr + bvec->bv_offset, bvec->bv_len,
 324                         MSG_WAITALL);
 325         kunmap(bvec->bv_page);
 326         return result;
 327 }
 328
 329 /* NULL returned = something went wrong, inform userspace */
 330 static struct request *nbd_read_stat(struct nbd_device *nbd)
 331 {
 332         int result;
 333         struct nbd_reply reply;
 334         struct request *req;
 335
 336         reply.magic = 0;
 337         result = sock_xmit(nbd, 0, &reply, sizeof(reply), MSG_WAITALL);
 338         if (result <= 0) {
 339                 dev_err(disk_to_dev(nbd->disk),
 340                         "Receive control failed (result %d)\n", result);
 341                 return ERR_PTR(result);
 342         }
 343
 344         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 345                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 346                                 (unsigned long)ntohl(reply.magic));
 347                 return ERR_PTR(-EPROTO);
 348         }
 349
 350         req = nbd_find_request(nbd, *(struct request **)reply.handle);
 351         if (IS_ERR(req)) {
 352                 result = PTR_ERR(req);
 353                 if (result != -ENOENT)
 354                         return ERR_PTR(result);
 355
 356                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%p)\n",
 357                         reply.handle);
 358                 return ERR_PTR(-EBADR);
 359         }
 360
 361         if (ntohl(reply.error)) {
 362                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 363                         ntohl(reply.error));
 364                 req->errors++;
 365                 return req;
 366         }
 367
 368         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 369         if (rq_data_dir(req) != WRITE) {
 370                 struct req_iterator iter;
 371                 struct bio_vec bvec;
 372
 373                 rq_for_each_segment(bvec, req, iter) {
 374                         result = sock_recv_bvec(nbd, &bvec);
 375                         if (result <= 0) {
 376                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 377                                         result);
 378                                 req->errors++;
 379                                 return req;
 380                         }
 381                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 382                                 req, bvec.bv_len);
 383                 }
 384         }
 385         return req;
 386 }
 387
 388 static ssize_t pid_show(struct device *dev,
 389                         struct device_attribute *attr, char *buf)
 390 {
 391         struct gendisk *disk = dev_to_disk(dev);
 392         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 393
 394         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 395 }
 396
 397 static struct device_attribute pid_attr = {
 398         .attr = { .name = "pid", .mode = S_IRUGO},
 399         .show = pid_show,
 400 };
 401
 402 static int nbd_thread_recv(struct nbd_device *nbd)
 403 {
 404         struct request *req;
 405         int ret;
 406
 407         BUG_ON(nbd->magic != NBD_MAGIC);
 408
 409         sk_set_memalloc(nbd->sock->sk);
 410
 411         nbd->task_recv = current;
 412
 413         ret = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
 414         if (ret) {
 415                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
 416                 nbd->task_recv = NULL;
 417                 return ret;
 418         }
 419
 420         while (1) {
 421                 req = nbd_read_stat(nbd);
 422                 if (IS_ERR(req)) {
 423                         ret = PTR_ERR(req);
 424                         break;
 425                 }
 426
 427                 nbd_end_request(nbd, req);
 428         }
 429
 430         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
 431
 432         nbd->task_recv = NULL;
 433
 434         if (signal_pending(current)) {
 435                 siginfo_t info;
 436
 437                 ret = dequeue_signal_lock(current, &current->blocked, &info);
 438                 dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
 439                          task_pid_nr(current), current->comm, ret);
 440                 mutex_lock(&nbd->tx_lock);
 441                 sock_shutdown(nbd);
 442                 mutex_unlock(&nbd->tx_lock);
 443                 ret = -ETIMEDOUT;
 444         }
 445
 446         return ret;
 447 }
 448
 449 static void nbd_clear_que(struct nbd_device *nbd)
 450 {
 451         struct request *req;
 452
 453         BUG_ON(nbd->magic != NBD_MAGIC);
 454
 455         /*
 456          * Because we have set nbd->sock to NULL under the tx_lock, all
 457          * modifications to the list must have completed by now.  For
 458          * the same reason, the active_req must be NULL.
 459          *
 460          * As a consequence, we don't need to take the spin lock while
 461          * purging the list here.
 462          */
 463         BUG_ON(nbd->sock);
 464         BUG_ON(nbd->active_req);
 465
 466         while (!list_empty(&nbd->queue_head)) {
 467                 req = list_entry(nbd->queue_head.next, struct request,
 468                                  queuelist);
 469                 list_del_init(&req->queuelist);
 470                 req->errors++;
 471                 nbd_end_request(nbd, req);
 472         }
 473
 474         while (!list_empty(&nbd->waiting_queue)) {
 475                 req = list_entry(nbd->waiting_queue.next, struct request,
 476                                  queuelist);
 477                 list_del_init(&req->queuelist);
 478                 req->errors++;
 479                 nbd_end_request(nbd, req);
 480         }
 481         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 482 }
 483
 484
 485 static void nbd_handle_req(struct nbd_device *nbd, struct request *req)
 486 {
 487         if (req->cmd_type != REQ_TYPE_FS)
 488                 goto error_out;
 489
 490         if (rq_data_dir(req) == WRITE &&
 491             (nbd->flags & NBD_FLAG_READ_ONLY)) {
 492                 dev_err(disk_to_dev(nbd->disk),
 493                         "Write on read-only\n");
 494                 goto error_out;
 495         }
 496
 497         req->errors = 0;
 498
 499         mutex_lock(&nbd->tx_lock);
 500         if (unlikely(!nbd->sock)) {
 501                 mutex_unlock(&nbd->tx_lock);
 502                 dev_err(disk_to_dev(nbd->disk),
 503                         "Attempted send on closed socket\n");
 504                 goto error_out;
 505         }
 506
 507         nbd->active_req = req;
 508
 509         if (nbd->xmit_timeout && list_empty_careful(&nbd->queue_head))
 510                 mod_timer(&nbd->timeout_timer, jiffies + nbd->xmit_timeout);
 511
 512         if (nbd_send_req(nbd, req) != 0) {
 513                 dev_err(disk_to_dev(nbd->disk), "Request send failed\n");
 514                 req->errors++;
 515                 nbd_end_request(nbd, req);
 516         } else {
 517                 spin_lock(&nbd->queue_lock);
 518                 list_add_tail(&req->queuelist, &nbd->queue_head);
 519                 spin_unlock(&nbd->queue_lock);
 520         }
 521
 522         nbd->active_req = NULL;
 523         mutex_unlock(&nbd->tx_lock);
 524         wake_up_all(&nbd->active_wq);
 525
 526         return;
 527
 528 error_out:
 529         req->errors++;
 530         nbd_end_request(nbd, req);
 531 }
 532
 533 static int nbd_thread_send(void *data)
 534 {
 535         struct nbd_device *nbd = data;
 536         struct request *req;
 537
 538         nbd->task_send = current;
 539
 540         set_user_nice(current, MIN_NICE);
 541         while (!kthread_should_stop() || !list_empty(&nbd->waiting_queue)) {
 542                 /* wait for something to do */
 543                 wait_event_interruptible(nbd->waiting_wq,
 544                                          kthread_should_stop() ||
 545                                          !list_empty(&nbd->waiting_queue));
 546
 547                 if (signal_pending(current)) {
 548                         siginfo_t info;
 549                         int ret;
 550
 551                         ret = dequeue_signal_lock(current, &current->blocked,
 552                                                   &info);
 553                         dev_warn(nbd_to_dev(nbd), "pid %d, %s, got signal %d\n",
 554                                  task_pid_nr(current), current->comm, ret);
 555                         mutex_lock(&nbd->tx_lock);
 556                         sock_shutdown(nbd);
 557                         mutex_unlock(&nbd->tx_lock);
 558                         break;
 559                 }
 560
 561                 /* extract request */
 562                 if (list_empty(&nbd->waiting_queue))
 563                         continue;
 564
 565                 spin_lock_irq(&nbd->queue_lock);
 566                 req = list_entry(nbd->waiting_queue.next, struct request,
 567                                  queuelist);
 568                 list_del_init(&req->queuelist);
 569                 spin_unlock_irq(&nbd->queue_lock);
 570
 571                 /* handle request */
 572                 nbd_handle_req(nbd, req);
 573         }
 574
 575         nbd->task_send = NULL;
 576
 577         return 0;
 578 }
 579
 580 /*
 581  * We always wait for result of write, for now. It would be nice to make it optional
 582  * in future
 583  * if ((rq_data_dir(req) == WRITE) && (nbd->flags & NBD_WRITE_NOCHK))
 584  *   { printk( "Warning: Ignoring result!\n"); nbd_end_request( req ); }
 585  */
 586
 587 static void nbd_request_handler(struct request_queue *q)
 588                 __releases(q->queue_lock) __acquires(q->queue_lock)
 589 {
 590         struct request *req;
 591
 592         while ((req = blk_fetch_request(q)) != NULL) {
 593                 struct nbd_device *nbd;
 594
 595                 spin_unlock_irq(q->queue_lock);
 596
 597                 nbd = req->rq_disk->private_data;
 598
 599                 BUG_ON(nbd->magic != NBD_MAGIC);
 600
 601                 dev_dbg(nbd_to_dev(nbd), "request %p: dequeued (flags=%x)\n",
 602                         req, req->cmd_type);
 603
 604                 if (unlikely(!nbd->sock)) {
 605                         dev_err(disk_to_dev(nbd->disk),
 606                                 "Attempted send on closed socket\n");
 607                         req->errors++;
 608                         nbd_end_request(nbd, req);
 609                         spin_lock_irq(q->queue_lock);
 610                         continue;
 611                 }
 612
 613                 spin_lock_irq(&nbd->queue_lock);
 614                 list_add_tail(&req->queuelist, &nbd->waiting_queue);
 615                 spin_unlock_irq(&nbd->queue_lock);
 616
 617                 wake_up(&nbd->waiting_wq);
 618
 619                 spin_lock_irq(q->queue_lock);
 620         }
 621 }
 622
 623 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 624 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 625
 626 /* Must be called with tx_lock held */
 627
 628 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
 629                        unsigned int cmd, unsigned long arg)
 630 {
 631         switch (cmd) {
 632         case NBD_DISCONNECT: {
 633                 struct request sreq;
 634
 635                 dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
 636                 if (!nbd->sock)
 637                         return -EINVAL;
 638
 639                 mutex_unlock(&nbd->tx_lock);
 640                 fsync_bdev(bdev);
 641                 mutex_lock(&nbd->tx_lock);
 642                 blk_rq_init(NULL, &sreq);
 643                 sreq.cmd_type = REQ_TYPE_DRV_PRIV;
 644
 645                 /* Check again after getting mutex back.  */
 646                 if (!nbd->sock)
 647                         return -EINVAL;
 648
 649                 nbd->disconnect = true;
 650
 651                 nbd_send_req(nbd, &sreq);
 652                 return 0;
 653         }
 654
 655         case NBD_CLEAR_SOCK: {
 656                 struct socket *sock = nbd->sock;
 657                 nbd->sock = NULL;
 658                 nbd_clear_que(nbd);
 659                 BUG_ON(!list_empty(&nbd->queue_head));
 660                 BUG_ON(!list_empty(&nbd->waiting_queue));
 661                 kill_bdev(bdev);
 662                 if (sock)
 663                         sockfd_put(sock);
 664                 return 0;
 665         }
 666
 667         case NBD_SET_SOCK: {
 668                 struct socket *sock;
 669                 int err;
 670                 if (nbd->sock)
 671                         return -EBUSY;
 672                 sock = sockfd_lookup(arg, &err);
 673                 if (sock) {
 674                         nbd->sock = sock;
 675                         if (max_part > 0)
 676                                 bdev->bd_invalidated = 1;
 677                         nbd->disconnect = false; /* we're connected now */
 678                         return 0;
 679                 }
 680                 return -EINVAL;
 681         }
 682
 683         case NBD_SET_BLKSIZE:
 684                 nbd->blksize = arg;
 685                 nbd->bytesize &= ~(nbd->blksize-1);
 686                 bdev->bd_inode->i_size = nbd->bytesize;
 687                 set_blocksize(bdev, nbd->blksize);
 688                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 689                 return 0;
 690
 691         case NBD_SET_SIZE:
 692                 nbd->bytesize = arg & ~(nbd->blksize-1);
 693                 bdev->bd_inode->i_size = nbd->bytesize;
 694                 set_blocksize(bdev, nbd->blksize);
 695                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 696                 return 0;
 697
 698         case NBD_SET_TIMEOUT:
 699                 nbd->xmit_timeout = arg * HZ;
 700                 if (arg)
 701                         mod_timer(&nbd->timeout_timer,
 702                                   jiffies + nbd->xmit_timeout);
 703                 else
 704                         del_timer_sync(&nbd->timeout_timer);
 705
 706                 return 0;
 707
 708         case NBD_SET_FLAGS:
 709                 nbd->flags = arg;
 710                 return 0;
 711
 712         case NBD_SET_SIZE_BLOCKS:
 713                 nbd->bytesize = ((u64) arg) * nbd->blksize;
 714                 bdev->bd_inode->i_size = nbd->bytesize;
 715                 set_blocksize(bdev, nbd->blksize);
 716                 set_capacity(nbd->disk, nbd->bytesize >> 9);
 717                 return 0;
 718
 719         case NBD_DO_IT: {
 720                 struct task_struct *thread;
 721                 struct socket *sock;
 722                 int error;
 723
 724                 if (nbd->task_recv)
 725                         return -EBUSY;
 726                 if (!nbd->sock)
 727                         return -EINVAL;
 728
 729                 mutex_unlock(&nbd->tx_lock);
 730
 731                 if (nbd->flags & NBD_FLAG_READ_ONLY)
 732                         set_device_ro(bdev, true);
 733                 if (nbd->flags & NBD_FLAG_SEND_TRIM)
 734                         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD,
 735                                 nbd->disk->queue);
 736                 if (nbd->flags & NBD_FLAG_SEND_FLUSH)
 737                         blk_queue_flush(nbd->disk->queue, REQ_FLUSH);
 738                 else
 739                         blk_queue_flush(nbd->disk->queue, 0);
 740
 741                 thread = kthread_run(nbd_thread_send, nbd, "%s",
 742                                      nbd_name(nbd));
 743                 if (IS_ERR(thread)) {
 744                         mutex_lock(&nbd->tx_lock);
 745                         return PTR_ERR(thread);
 746                 }
 747
 748                 nbd_dev_dbg_init(nbd);
 749                 error = nbd_thread_recv(nbd);
 750                 nbd_dev_dbg_close(nbd);
 751                 kthread_stop(thread);
 752
 753                 mutex_lock(&nbd->tx_lock);
 754
 755                 sock_shutdown(nbd);
 756                 sock = nbd->sock;
 757                 nbd->sock = NULL;
 758                 nbd_clear_que(nbd);
 759                 kill_bdev(bdev);
 760                 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 761                 set_device_ro(bdev, false);
 762                 if (sock)
 763                         sockfd_put(sock);
 764                 nbd->flags = 0;
 765                 nbd->bytesize = 0;
 766                 bdev->bd_inode->i_size = 0;
 767                 set_capacity(nbd->disk, 0);
 768                 if (max_part > 0)
 769                         blkdev_reread_part(bdev);
 770                 if (nbd->disconnect) /* user requested, ignore socket errors */
 771                         return 0;
 772                 return error;
 773         }
 774
 775         case NBD_CLEAR_QUE:
 776                 /*
 777                  * This is for compatibility only.  The queue is always cleared
 778                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
 779                  */
 780                 return 0;
 781
 782         case NBD_PRINT_DEBUG:
 783                 dev_info(disk_to_dev(nbd->disk),
 784                         "next = %p, prev = %p, head = %p\n",
 785                         nbd->queue_head.next, nbd->queue_head.prev,
 786                         &nbd->queue_head);
 787                 return 0;
 788         }
 789         return -ENOTTY;
 790 }
 791
 792 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
 793                      unsigned int cmd, unsigned long arg)
 794 {
 795         struct nbd_device *nbd = bdev->bd_disk->private_data;
 796         int error;
 797
 798         if (!capable(CAP_SYS_ADMIN))
 799                 return -EPERM;
 800
 801         BUG_ON(nbd->magic != NBD_MAGIC);
 802
 803         mutex_lock(&nbd->tx_lock);
 804         error = __nbd_ioctl(bdev, nbd, cmd, arg);
 805         mutex_unlock(&nbd->tx_lock);
 806
 807         return error;
 808 }
 809
 810 static const struct block_device_operations nbd_fops =
 811 {
 812         .owner =        THIS_MODULE,
 813         .ioctl =        nbd_ioctl,
 814 };
 815
 816 #if IS_ENABLED(CONFIG_DEBUG_FS)
 817
 818 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
 819 {
 820         struct nbd_device *nbd = s->private;
 821
 822         if (nbd->task_recv)
 823                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
 824         if (nbd->task_send)
 825                 seq_printf(s, "send: %d\n", task_pid_nr(nbd->task_send));
 826
 827         return 0;
 828 }
 829
 830 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
 831 {
 832         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
 833 }
 834
 835 static const struct file_operations nbd_dbg_tasks_ops = {
 836         .open = nbd_dbg_tasks_open,
 837         .read = seq_read,
 838         .llseek = seq_lseek,
 839         .release = single_release,
 840 };
 841
 842 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
 843 {
 844         struct nbd_device *nbd = s->private;
 845         u32 flags = nbd->flags;
 846
 847         seq_printf(s, "Hex: 0x%08x\n\n", flags);
 848
 849         seq_puts(s, "Known flags:\n");
 850
 851         if (flags & NBD_FLAG_HAS_FLAGS)
 852                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
 853         if (flags & NBD_FLAG_READ_ONLY)
 854                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
 855         if (flags & NBD_FLAG_SEND_FLUSH)
 856                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
 857         if (flags & NBD_FLAG_SEND_TRIM)
 858                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
 859
 860         return 0;
 861 }
 862
 863 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
 864 {
 865         return single_open(file, nbd_dbg_flags_show, inode->i_private);
 866 }
 867
 868 static const struct file_operations nbd_dbg_flags_ops = {
 869         .open = nbd_dbg_flags_open,
 870         .read = seq_read,
 871         .llseek = seq_lseek,
 872         .release = single_release,
 873 };
 874
 875 static int nbd_dev_dbg_init(struct nbd_device *nbd)
 876 {
 877         struct dentry *dir;
 878         struct dentry *f;
 879
 880         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
 881         if (IS_ERR_OR_NULL(dir)) {
 882                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s' (%ld)\n",
 883                         nbd_name(nbd), PTR_ERR(dir));
 884                 return PTR_ERR(dir);
 885         }
 886         nbd->dbg_dir = dir;
 887
 888         f = debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
 889         if (IS_ERR_OR_NULL(f)) {
 890                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs file 'tasks', %ld\n",
 891                         PTR_ERR(f));
 892                 return PTR_ERR(f);
 893         }
 894
 895         f = debugfs_create_u64("size_bytes", 0444, dir, &nbd->bytesize);
 896         if (IS_ERR_OR_NULL(f)) {
 897                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs file 'size_bytes', %ld\n",
 898                         PTR_ERR(f));
 899                 return PTR_ERR(f);
 900         }
 901
 902         f = debugfs_create_u32("timeout", 0444, dir, &nbd->xmit_timeout);
 903         if (IS_ERR_OR_NULL(f)) {
 904                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs file 'timeout', %ld\n",
 905                         PTR_ERR(f));
 906                 return PTR_ERR(f);
 907         }
 908
 909         f = debugfs_create_u32("blocksize", 0444, dir, &nbd->blksize);
 910         if (IS_ERR_OR_NULL(f)) {
 911                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs file 'blocksize', %ld\n",
 912                         PTR_ERR(f));
 913                 return PTR_ERR(f);
 914         }
 915
 916         f = debugfs_create_file("flags", 0444, dir, &nbd, &nbd_dbg_flags_ops);
 917         if (IS_ERR_OR_NULL(f)) {
 918                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs file 'flags', %ld\n",
 919                         PTR_ERR(f));
 920                 return PTR_ERR(f);
 921         }
 922
 923         return 0;
 924 }
 925
 926 static void nbd_dev_dbg_close(struct nbd_device *nbd)
 927 {
 928         debugfs_remove_recursive(nbd->dbg_dir);
 929 }
 930
 931 static int nbd_dbg_init(void)
 932 {
 933         struct dentry *dbg_dir;
 934
 935         dbg_dir = debugfs_create_dir("nbd", NULL);
 936         if (IS_ERR(dbg_dir))
 937                 return PTR_ERR(dbg_dir);
 938
 939         nbd_dbg_dir = dbg_dir;
 940
 941         return 0;
 942 }
 943
 944 static void nbd_dbg_close(void)
 945 {
 946         debugfs_remove_recursive(nbd_dbg_dir);
 947 }
 948
 949 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
 950
 951 static int nbd_dev_dbg_init(struct nbd_device *nbd)
 952 {
 953         return 0;
 954 }
 955
 956 static void nbd_dev_dbg_close(struct nbd_device *nbd)
 957 {
 958 }
 959
 960 static int nbd_dbg_init(void)
 961 {
 962         return 0;
 963 }
 964
 965 static void nbd_dbg_close(void)
 966 {
 967 }
 968
 969 #endif
 970
 971 /*
 972  * And here should be modules and kernel interface
 973  *  (Just smiley confuses emacs :-)
 974  */
 975
 976 static int __init nbd_init(void)
 977 {
 978         int err = -ENOMEM;
 979         int i;
 980         int part_shift;
 981
 982         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
 983
 984         if (max_part < 0) {
 985                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
 986                 return -EINVAL;
 987         }
 988
 989         part_shift = 0;
 990         if (max_part > 0) {
 991                 part_shift = fls(max_part);
 992
 993                 /*
 994                  * Adjust max_part according to part_shift as it is exported
 995                  * to user space so that user can know the max number of
 996                  * partition kernel should be able to manage.
 997                  *
 998                  * Note that -1 is required because partition 0 is reserved
 999                  * for the whole disk.
1000                  */
1001                 max_part = (1UL << part_shift) - 1;
1002         }
1003
1004         if ((1UL << part_shift) > DISK_MAX_PARTS)
1005                 return -EINVAL;
1006
1007         if (nbds_max > 1UL << (MINORBITS - part_shift))
1008                 return -EINVAL;
1009
1010         nbd_dev = kcalloc(nbds_max, sizeof(*nbd_dev), GFP_KERNEL);
1011         if (!nbd_dev)
1012                 return -ENOMEM;
1013
1014         for (i = 0; i < nbds_max; i++) {
1015                 struct gendisk *disk = alloc_disk(1 << part_shift);
1016                 if (!disk)
1017                         goto out;
1018                 nbd_dev[i].disk = disk;
1019                 /*
1020                  * The new linux 2.5 block layer implementation requires
1021                  * every gendisk to have its very own request_queue struct.
1022                  * These structs are big so we dynamically allocate them.
1023                  */
1024                 disk->queue = blk_init_queue(nbd_request_handler, &nbd_lock);
1025                 if (!disk->queue) {
1026                         put_disk(disk);
1027                         goto out;
1028                 }
1029                 /*
1030                  * Tell the block layer that we are not a rotational device
1031                  */
1032                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, disk->queue);
1033                 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1034                 disk->queue->limits.discard_granularity = 512;
1035                 blk_queue_max_discard_sectors(disk->queue, UINT_MAX);
1036                 disk->queue->limits.discard_zeroes_data = 0;
1037                 blk_queue_max_hw_sectors(disk->queue, 65536);
1038                 disk->queue->limits.max_sectors = 256;
1039         }
1040
1041         if (register_blkdev(NBD_MAJOR, "nbd")) {
1042                 err = -EIO;
1043                 goto out;
1044         }
1045
1046         printk(KERN_INFO "nbd: registered device at major %d\n", NBD_MAJOR);
1047
1048         nbd_dbg_init();
1049
1050         for (i = 0; i < nbds_max; i++) {
1051                 struct gendisk *disk = nbd_dev[i].disk;
1052                 nbd_dev[i].magic = NBD_MAGIC;
1053                 INIT_LIST_HEAD(&nbd_dev[i].waiting_queue);
1054                 spin_lock_init(&nbd_dev[i].queue_lock);
1055                 INIT_LIST_HEAD(&nbd_dev[i].queue_head);
1056                 mutex_init(&nbd_dev[i].tx_lock);
1057                 init_timer(&nbd_dev[i].timeout_timer);
1058                 nbd_dev[i].timeout_timer.function = nbd_xmit_timeout;
1059                 nbd_dev[i].timeout_timer.data = (unsigned long)&nbd_dev[i];
1060                 init_waitqueue_head(&nbd_dev[i].active_wq);
1061                 init_waitqueue_head(&nbd_dev[i].waiting_wq);
1062                 nbd_dev[i].blksize = 1024;
1063                 nbd_dev[i].bytesize = 0;
1064                 disk->major = NBD_MAJOR;
1065                 disk->first_minor = i << part_shift;
1066                 disk->fops = &nbd_fops;
1067                 disk->private_data = &nbd_dev[i];
1068                 sprintf(disk->disk_name, "nbd%d", i);
1069                 set_capacity(disk, 0);
1070                 add_disk(disk);
1071         }
1072
1073         return 0;
1074 out:
1075         while (i--) {
1076                 blk_cleanup_queue(nbd_dev[i].disk->queue);
1077                 put_disk(nbd_dev[i].disk);
1078         }
1079         kfree(nbd_dev);
1080         return err;
1081 }
1082
1083 static void __exit nbd_cleanup(void)
1084 {
1085         int i;
1086
1087         nbd_dbg_close();
1088
1089         for (i = 0; i < nbds_max; i++) {
1090                 struct gendisk *disk = nbd_dev[i].disk;
1091                 nbd_dev[i].magic = 0;
1092                 if (disk) {
1093                         del_gendisk(disk);
1094                         blk_cleanup_queue(disk->queue);
1095                         put_disk(disk);
1096                 }
1097         }
1098         unregister_blkdev(NBD_MAJOR, "nbd");
1099         kfree(nbd_dev);
1100         printk(KERN_INFO "nbd: unregistered device at major %d\n", NBD_MAJOR);
1101 }
1102
1103 module_init(nbd_init);
1104 module_exit(nbd_cleanup);
1105
1106 MODULE_DESCRIPTION("Network Block Device");
1107 MODULE_LICENSE("GPL");
1108
1109 module_param(nbds_max, int, 0444);
1110 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
1111 module_param(max_part, int, 0444);
1112 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 0)");