drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   9  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #include <linux/major.h>
  15
  16 #include <linux/blkdev.h>
  17 #include <linux/module.h>
  18 #include <linux/init.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/mm.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/completion.h>
  30 #include <linux/err.h>
  31 #include <linux/kernel.h>
  32 #include <linux/slab.h>
  33 #include <net/sock.h>
  34 #include <linux/net.h>
  35 #include <linux/kthread.h>
  36 #include <linux/types.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/blk-mq.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/types.h>
  42
  43 #include <linux/nbd.h>
  44 #include <linux/nbd-netlink.h>
  45 #include <net/genetlink.h>
  46
  47 #define CREATE_TRACE_POINTS
  48 #include <trace/events/nbd.h>
  49
  50 static DEFINE_IDR(nbd_index_idr);
  51 static DEFINE_MUTEX(nbd_index_mutex);
  52 static int nbd_total_devices = 0;
  53
  54 struct nbd_sock {
  55         struct socket *sock;
  56         struct mutex tx_lock;
  57         struct request *pending;
  58         int sent;
  59         bool dead;
  60         int fallback_index;
  61         int cookie;
  62 };
  63
  64 struct recv_thread_args {
  65         struct work_struct work;
  66         struct nbd_device *nbd;
  67         int index;
  68 };
  69
  70 struct link_dead_args {
  71         struct work_struct work;
  72         int index;
  73 };
  74
  75 #define NBD_RT_TIMEDOUT                 0
  76 #define NBD_RT_DISCONNECT_REQUESTED     1
  77 #define NBD_RT_DISCONNECTED             2
  78 #define NBD_RT_HAS_PID_FILE             3
  79 #define NBD_RT_HAS_CONFIG_REF           4
  80 #define NBD_RT_BOUND                    5
  81 #define NBD_RT_DISCONNECT_ON_CLOSE      6
  82
  83 #define NBD_DESTROY_ON_DISCONNECT       0
  84 #define NBD_DISCONNECT_REQUESTED        1
  85
  86 struct nbd_config {
  87         u32 flags;
  88         unsigned long runtime_flags;
  89         u64 dead_conn_timeout;
  90
  91         struct nbd_sock **socks;
  92         int num_connections;
  93         atomic_t live_connections;
  94         wait_queue_head_t conn_wait;
  95
  96         atomic_t recv_threads;
  97         wait_queue_head_t recv_wq;
  98         loff_t blksize;
  99         loff_t bytesize;
 100 #if IS_ENABLED(CONFIG_DEBUG_FS)
 101         struct dentry *dbg_dir;
 102 #endif
 103 };
 104
 105 struct nbd_device {
 106         struct blk_mq_tag_set tag_set;
 107
 108         int index;
 109         refcount_t config_refs;
 110         refcount_t refs;
 111         struct nbd_config *config;
 112         struct mutex config_lock;
 113         struct gendisk *disk;
 114         struct workqueue_struct *recv_workq;
 115
 116         struct list_head list;
 117         struct task_struct *task_recv;
 118         struct task_struct *task_setup;
 119
 120         struct completion *destroy_complete;
 121         unsigned long flags;
 122 };
 123
 124 #define NBD_CMD_REQUEUED        1
 125
 126 struct nbd_cmd {
 127         struct nbd_device *nbd;
 128         struct mutex lock;
 129         int index;
 130         int cookie;
 131         int retries;
 132         blk_status_t status;
 133         unsigned long flags;
 134         u32 cmd_cookie;
 135 };
 136
 137 #if IS_ENABLED(CONFIG_DEBUG_FS)
 138 static struct dentry *nbd_dbg_dir;
 139 #endif
 140
 141 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 142
 143 #define NBD_MAGIC 0x68797548
 144
 145 #define NBD_DEF_BLKSIZE 1024
 146
 147 static unsigned int nbds_max = 16;
 148 static int max_part = 16;
 149 static int part_shift;
 150
 151 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 152 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 153 static void nbd_config_put(struct nbd_device *nbd);
 154 static void nbd_connect_reply(struct genl_info *info, int index);
 155 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 156 static void nbd_dead_link_work(struct work_struct *work);
 157 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 158
 159 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 160 {
 161         return disk_to_dev(nbd->disk);
 162 }
 163
 164 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 165 {
 166         struct request *req = blk_mq_rq_from_pdu(cmd);
 167
 168         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 169                 blk_mq_requeue_request(req, true);
 170 }
 171
 172 #define NBD_COOKIE_BITS 32
 173
 174 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 175 {
 176         struct request *req = blk_mq_rq_from_pdu(cmd);
 177         u32 tag = blk_mq_unique_tag(req);
 178         u64 cookie = cmd->cmd_cookie;
 179
 180         return (cookie << NBD_COOKIE_BITS) | tag;
 181 }
 182
 183 static u32 nbd_handle_to_tag(u64 handle)
 184 {
 185         return (u32)handle;
 186 }
 187
 188 static u32 nbd_handle_to_cookie(u64 handle)
 189 {
 190         return (u32)(handle >> NBD_COOKIE_BITS);
 191 }
 192
 193 static const char *nbdcmd_to_ascii(int cmd)
 194 {
 195         switch (cmd) {
 196         case  NBD_CMD_READ: return "read";
 197         case NBD_CMD_WRITE: return "write";
 198         case  NBD_CMD_DISC: return "disconnect";
 199         case NBD_CMD_FLUSH: return "flush";
 200         case  NBD_CMD_TRIM: return "trim/discard";
 201         }
 202         return "invalid";
 203 }
 204
 205 static ssize_t pid_show(struct device *dev,
 206                         struct device_attribute *attr, char *buf)
 207 {
 208         struct gendisk *disk = dev_to_disk(dev);
 209         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 210
 211         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 212 }
 213
 214 static const struct device_attribute pid_attr = {
 215         .attr = { .name = "pid", .mode = 0444},
 216         .show = pid_show,
 217 };
 218
 219 static void nbd_dev_remove(struct nbd_device *nbd)
 220 {
 221         struct gendisk *disk = nbd->disk;
 222         struct request_queue *q;
 223
 224         if (disk) {
 225                 q = disk->queue;
 226                 del_gendisk(disk);
 227                 blk_cleanup_queue(q);
 228                 blk_mq_free_tag_set(&nbd->tag_set);
 229                 disk->private_data = NULL;
 230                 put_disk(disk);
 231         }
 232
 233         /*
 234          * Place this in the last just before the nbd is freed to
 235          * make sure that the disk and the related kobject are also
 236          * totally removed to avoid duplicate creation of the same
 237          * one.
 238          */
 239         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) && nbd->destroy_complete)
 240                 complete(nbd->destroy_complete);
 241
 242         kfree(nbd);
 243 }
 244
 245 static void nbd_put(struct nbd_device *nbd)
 246 {
 247         if (refcount_dec_and_mutex_lock(&nbd->refs,
 248                                         &nbd_index_mutex)) {
 249                 idr_remove(&nbd_index_idr, nbd->index);
 250                 nbd_dev_remove(nbd);
 251                 mutex_unlock(&nbd_index_mutex);
 252         }
 253 }
 254
 255 static int nbd_disconnected(struct nbd_config *config)
 256 {
 257         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 258                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 259 }
 260
 261 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 262                                 int notify)
 263 {
 264         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 265                 struct link_dead_args *args;
 266                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 267                 if (args) {
 268                         INIT_WORK(&args->work, nbd_dead_link_work);
 269                         args->index = nbd->index;
 270                         queue_work(system_wq, &args->work);
 271                 }
 272         }
 273         if (!nsock->dead) {
 274                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 275                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 276                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 277                                                &nbd->config->runtime_flags)) {
 278                                 set_bit(NBD_RT_DISCONNECTED,
 279                                         &nbd->config->runtime_flags);
 280                                 dev_info(nbd_to_dev(nbd),
 281                                         "Disconnected due to user request.\n");
 282                         }
 283                 }
 284         }
 285         nsock->dead = true;
 286         nsock->pending = NULL;
 287         nsock->sent = 0;
 288 }
 289
 290 static void nbd_size_clear(struct nbd_device *nbd)
 291 {
 292         if (nbd->config->bytesize) {
 293                 set_capacity(nbd->disk, 0);
 294                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 295         }
 296 }
 297
 298 static void nbd_size_update(struct nbd_device *nbd, bool start)
 299 {
 300         struct nbd_config *config = nbd->config;
 301         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 302         sector_t nr_sectors = config->bytesize >> 9;
 303
 304         if (config->flags & NBD_FLAG_SEND_TRIM) {
 305                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 306                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 307                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 308         }
 309         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 310         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 311         set_capacity(nbd->disk, nr_sectors);
 312         if (bdev) {
 313                 if (bdev->bd_disk) {
 314                         bd_set_nr_sectors(bdev, nr_sectors);
 315                         if (start)
 316                                 set_blocksize(bdev, config->blksize);
 317                 } else
 318                         set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 319                 bdput(bdev);
 320         }
 321         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 322 }
 323
 324 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 325                          loff_t nr_blocks)
 326 {
 327         struct nbd_config *config = nbd->config;
 328         config->blksize = blocksize;
 329         config->bytesize = blocksize * nr_blocks;
 330         if (nbd->task_recv != NULL)
 331                 nbd_size_update(nbd, false);
 332 }
 333
 334 static void nbd_complete_rq(struct request *req)
 335 {
 336         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 337
 338         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 339                 cmd->status ? "failed" : "done");
 340
 341         blk_mq_end_request(req, cmd->status);
 342 }
 343
 344 /*
 345  * Forcibly shutdown the socket causing all listeners to error
 346  */
 347 static void sock_shutdown(struct nbd_device *nbd)
 348 {
 349         struct nbd_config *config = nbd->config;
 350         int i;
 351
 352         if (config->num_connections == 0)
 353                 return;
 354         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 355                 return;
 356
 357         for (i = 0; i < config->num_connections; i++) {
 358                 struct nbd_sock *nsock = config->socks[i];
 359                 mutex_lock(&nsock->tx_lock);
 360                 nbd_mark_nsock_dead(nbd, nsock, 0);
 361                 mutex_unlock(&nsock->tx_lock);
 362         }
 363         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 364 }
 365
 366 static u32 req_to_nbd_cmd_type(struct request *req)
 367 {
 368         switch (req_op(req)) {
 369         case REQ_OP_DISCARD:
 370                 return NBD_CMD_TRIM;
 371         case REQ_OP_FLUSH:
 372                 return NBD_CMD_FLUSH;
 373         case REQ_OP_WRITE:
 374                 return NBD_CMD_WRITE;
 375         case REQ_OP_READ:
 376                 return NBD_CMD_READ;
 377         default:
 378                 return U32_MAX;
 379         }
 380 }
 381
 382 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 383                                                  bool reserved)
 384 {
 385         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 386         struct nbd_device *nbd = cmd->nbd;
 387         struct nbd_config *config;
 388
 389         if (!mutex_trylock(&cmd->lock))
 390                 return BLK_EH_RESET_TIMER;
 391
 392         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 393                 cmd->status = BLK_STS_TIMEOUT;
 394                 mutex_unlock(&cmd->lock);
 395                 goto done;
 396         }
 397         config = nbd->config;
 398
 399         if (config->num_connections > 1 ||
 400             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 401                 dev_err_ratelimited(nbd_to_dev(nbd),
 402                                     "Connection timed out, retrying (%d/%d alive)\n",
 403                                     atomic_read(&config->live_connections),
 404                                     config->num_connections);
 405                 /*
 406                  * Hooray we have more connections, requeue this IO, the submit
 407                  * path will put it on a real connection. Or if only one
 408                  * connection is configured, the submit path will wait util
 409                  * a new connection is reconfigured or util dead timeout.
 410                  */
 411                 if (config->socks) {
 412                         if (cmd->index < config->num_connections) {
 413                                 struct nbd_sock *nsock =
 414                                         config->socks[cmd->index];
 415                                 mutex_lock(&nsock->tx_lock);
 416                                 /* We can have multiple outstanding requests, so
 417                                  * we don't want to mark the nsock dead if we've
 418                                  * already reconnected with a new socket, so
 419                                  * only mark it dead if its the same socket we
 420                                  * were sent out on.
 421                                  */
 422                                 if (cmd->cookie == nsock->cookie)
 423                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 424                                 mutex_unlock(&nsock->tx_lock);
 425                         }
 426                         mutex_unlock(&cmd->lock);
 427                         nbd_requeue_cmd(cmd);
 428                         nbd_config_put(nbd);
 429                         return BLK_EH_DONE;
 430                 }
 431         }
 432
 433         if (!nbd->tag_set.timeout) {
 434                 /*
 435                  * Userspace sets timeout=0 to disable socket disconnection,
 436                  * so just warn and reset the timer.
 437                  */
 438                 struct nbd_sock *nsock = config->socks[cmd->index];
 439                 cmd->retries++;
 440                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 441                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 442                         (unsigned long long)blk_rq_pos(req) << 9,
 443                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 444
 445                 mutex_lock(&nsock->tx_lock);
 446                 if (cmd->cookie != nsock->cookie) {
 447                         nbd_requeue_cmd(cmd);
 448                         mutex_unlock(&nsock->tx_lock);
 449                         mutex_unlock(&cmd->lock);
 450                         nbd_config_put(nbd);
 451                         return BLK_EH_DONE;
 452                 }
 453                 mutex_unlock(&nsock->tx_lock);
 454                 mutex_unlock(&cmd->lock);
 455                 nbd_config_put(nbd);
 456                 return BLK_EH_RESET_TIMER;
 457         }
 458
 459         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 460         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 461         cmd->status = BLK_STS_IOERR;
 462         mutex_unlock(&cmd->lock);
 463         sock_shutdown(nbd);
 464         nbd_config_put(nbd);
 465 done:
 466         blk_mq_complete_request(req);
 467         return BLK_EH_DONE;
 468 }
 469
 470 /*
 471  *  Send or receive packet.
 472  */
 473 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 474                      struct iov_iter *iter, int msg_flags, int *sent)
 475 {
 476         struct nbd_config *config = nbd->config;
 477         struct socket *sock = config->socks[index]->sock;
 478         int result;
 479         struct msghdr msg;
 480         unsigned int noreclaim_flag;
 481
 482         if (unlikely(!sock)) {
 483                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 484                         "Attempted %s on closed socket in sock_xmit\n",
 485                         (send ? "send" : "recv"));
 486                 return -EINVAL;
 487         }
 488
 489         msg.msg_iter = *iter;
 490
 491         noreclaim_flag = memalloc_noreclaim_save();
 492         do {
 493                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 494                 msg.msg_name = NULL;
 495                 msg.msg_namelen = 0;
 496                 msg.msg_control = NULL;
 497                 msg.msg_controllen = 0;
 498                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 499
 500                 if (send)
 501                         result = sock_sendmsg(sock, &msg);
 502                 else
 503                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 504
 505                 if (result <= 0) {
 506                         if (result == 0)
 507                                 result = -EPIPE; /* short read */
 508                         break;
 509                 }
 510                 if (sent)
 511                         *sent += result;
 512         } while (msg_data_left(&msg));
 513
 514         memalloc_noreclaim_restore(noreclaim_flag);
 515
 516         return result;
 517 }
 518
 519 /*
 520  * Different settings for sk->sk_sndtimeo can result in different return values
 521  * if there is a signal pending when we enter sendmsg, because reasons?
 522  */
 523 static inline int was_interrupted(int result)
 524 {
 525         return result == -ERESTARTSYS || result == -EINTR;
 526 }
 527
 528 /* always call with the tx_lock held */
 529 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 530 {
 531         struct request *req = blk_mq_rq_from_pdu(cmd);
 532         struct nbd_config *config = nbd->config;
 533         struct nbd_sock *nsock = config->socks[index];
 534         int result;
 535         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 536         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 537         struct iov_iter from;
 538         unsigned long size = blk_rq_bytes(req);
 539         struct bio *bio;
 540         u64 handle;
 541         u32 type;
 542         u32 nbd_cmd_flags = 0;
 543         int sent = nsock->sent, skip = 0;
 544
 545         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 546
 547         type = req_to_nbd_cmd_type(req);
 548         if (type == U32_MAX)
 549                 return -EIO;
 550
 551         if (rq_data_dir(req) == WRITE &&
 552             (config->flags & NBD_FLAG_READ_ONLY)) {
 553                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 554                                     "Write on read-only\n");
 555                 return -EIO;
 556         }
 557
 558         if (req->cmd_flags & REQ_FUA)
 559                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 560
 561         /* We did a partial send previously, and we at least sent the whole
 562          * request struct, so just go and send the rest of the pages in the
 563          * request.
 564          */
 565         if (sent) {
 566                 if (sent >= sizeof(request)) {
 567                         skip = sent - sizeof(request);
 568
 569                         /* initialize handle for tracing purposes */
 570                         handle = nbd_cmd_handle(cmd);
 571
 572                         goto send_pages;
 573                 }
 574                 iov_iter_advance(&from, sent);
 575         } else {
 576                 cmd->cmd_cookie++;
 577         }
 578         cmd->index = index;
 579         cmd->cookie = nsock->cookie;
 580         cmd->retries = 0;
 581         request.type = htonl(type | nbd_cmd_flags);
 582         if (type != NBD_CMD_FLUSH) {
 583                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 584                 request.len = htonl(size);
 585         }
 586         handle = nbd_cmd_handle(cmd);
 587         memcpy(request.handle, &handle, sizeof(handle));
 588
 589         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 590
 591         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 592                 req, nbdcmd_to_ascii(type),
 593                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 594         result = sock_xmit(nbd, index, 1, &from,
 595                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 596         trace_nbd_header_sent(req, handle);
 597         if (result <= 0) {
 598                 if (was_interrupted(result)) {
 599                         /* If we havne't sent anything we can just return BUSY,
 600                          * however if we have sent something we need to make
 601                          * sure we only allow this req to be sent until we are
 602                          * completely done.
 603                          */
 604                         if (sent) {
 605                                 nsock->pending = req;
 606                                 nsock->sent = sent;
 607                         }
 608                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 609                         return BLK_STS_RESOURCE;
 610                 }
 611                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 612                         "Send control failed (result %d)\n", result);
 613                 return -EAGAIN;
 614         }
 615 send_pages:
 616         if (type != NBD_CMD_WRITE)
 617                 goto out;
 618
 619         bio = req->bio;
 620         while (bio) {
 621                 struct bio *next = bio->bi_next;
 622                 struct bvec_iter iter;
 623                 struct bio_vec bvec;
 624
 625                 bio_for_each_segment(bvec, bio, iter) {
 626                         bool is_last = !next && bio_iter_last(bvec, iter);
 627                         int flags = is_last ? 0 : MSG_MORE;
 628
 629                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 630                                 req, bvec.bv_len);
 631                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 632                         if (skip) {
 633                                 if (skip >= iov_iter_count(&from)) {
 634                                         skip -= iov_iter_count(&from);
 635                                         continue;
 636                                 }
 637                                 iov_iter_advance(&from, skip);
 638                                 skip = 0;
 639                         }
 640                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 641                         if (result <= 0) {
 642                                 if (was_interrupted(result)) {
 643                                         /* We've already sent the header, we
 644                                          * have no choice but to set pending and
 645                                          * return BUSY.
 646                                          */
 647                                         nsock->pending = req;
 648                                         nsock->sent = sent;
 649                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 650                                         return BLK_STS_RESOURCE;
 651                                 }
 652                                 dev_err(disk_to_dev(nbd->disk),
 653                                         "Send data failed (result %d)\n",
 654                                         result);
 655                                 return -EAGAIN;
 656                         }
 657                         /*
 658                          * The completion might already have come in,
 659                          * so break for the last one instead of letting
 660                          * the iterator do it. This prevents use-after-free
 661                          * of the bio.
 662                          */
 663                         if (is_last)
 664                                 break;
 665                 }
 666                 bio = next;
 667         }
 668 out:
 669         trace_nbd_payload_sent(req, handle);
 670         nsock->pending = NULL;
 671         nsock->sent = 0;
 672         return 0;
 673 }
 674
 675 /* NULL returned = something went wrong, inform userspace */
 676 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 677 {
 678         struct nbd_config *config = nbd->config;
 679         int result;
 680         struct nbd_reply reply;
 681         struct nbd_cmd *cmd;
 682         struct request *req = NULL;
 683         u64 handle;
 684         u16 hwq;
 685         u32 tag;
 686         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 687         struct iov_iter to;
 688         int ret = 0;
 689
 690         reply.magic = 0;
 691         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 692         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 693         if (result <= 0) {
 694                 if (!nbd_disconnected(config))
 695                         dev_err(disk_to_dev(nbd->disk),
 696                                 "Receive control failed (result %d)\n", result);
 697                 return ERR_PTR(result);
 698         }
 699
 700         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 701                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 702                                 (unsigned long)ntohl(reply.magic));
 703                 return ERR_PTR(-EPROTO);
 704         }
 705
 706         memcpy(&handle, reply.handle, sizeof(handle));
 707         tag = nbd_handle_to_tag(handle);
 708         hwq = blk_mq_unique_tag_to_hwq(tag);
 709         if (hwq < nbd->tag_set.nr_hw_queues)
 710                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 711                                        blk_mq_unique_tag_to_tag(tag));
 712         if (!req || !blk_mq_request_started(req)) {
 713                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 714                         tag, req);
 715                 return ERR_PTR(-ENOENT);
 716         }
 717         trace_nbd_header_received(req, handle);
 718         cmd = blk_mq_rq_to_pdu(req);
 719
 720         mutex_lock(&cmd->lock);
 721         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 722                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 723                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 724                 ret = -ENOENT;
 725                 goto out;
 726         }
 727         if (cmd->status != BLK_STS_OK) {
 728                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 729                         req);
 730                 ret = -ENOENT;
 731                 goto out;
 732         }
 733         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 734                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 735                         req);
 736                 ret = -ENOENT;
 737                 goto out;
 738         }
 739         if (ntohl(reply.error)) {
 740                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 741                         ntohl(reply.error));
 742                 cmd->status = BLK_STS_IOERR;
 743                 goto out;
 744         }
 745
 746         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 747         if (rq_data_dir(req) != WRITE) {
 748                 struct req_iterator iter;
 749                 struct bio_vec bvec;
 750
 751                 rq_for_each_segment(bvec, req, iter) {
 752                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 753                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 754                         if (result <= 0) {
 755                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 756                                         result);
 757                                 /*
 758                                  * If we've disconnected, we need to make sure we
 759                                  * complete this request, otherwise error out
 760                                  * and let the timeout stuff handle resubmitting
 761                                  * this request onto another connection.
 762                                  */
 763                                 if (nbd_disconnected(config)) {
 764                                         cmd->status = BLK_STS_IOERR;
 765                                         goto out;
 766                                 }
 767                                 ret = -EIO;
 768                                 goto out;
 769                         }
 770                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 771                                 req, bvec.bv_len);
 772                 }
 773         }
 774 out:
 775         trace_nbd_payload_received(req, handle);
 776         mutex_unlock(&cmd->lock);
 777         return ret ? ERR_PTR(ret) : cmd;
 778 }
 779
 780 static void recv_work(struct work_struct *work)
 781 {
 782         struct recv_thread_args *args = container_of(work,
 783                                                      struct recv_thread_args,
 784                                                      work);
 785         struct nbd_device *nbd = args->nbd;
 786         struct nbd_config *config = nbd->config;
 787         struct nbd_cmd *cmd;
 788         struct request *rq;
 789
 790         while (1) {
 791                 cmd = nbd_read_stat(nbd, args->index);
 792                 if (IS_ERR(cmd)) {
 793                         struct nbd_sock *nsock = config->socks[args->index];
 794
 795                         mutex_lock(&nsock->tx_lock);
 796                         nbd_mark_nsock_dead(nbd, nsock, 1);
 797                         mutex_unlock(&nsock->tx_lock);
 798                         break;
 799                 }
 800
 801                 rq = blk_mq_rq_from_pdu(cmd);
 802                 if (likely(!blk_should_fake_timeout(rq->q)))
 803                         blk_mq_complete_request(rq);
 804         }
 805         nbd_config_put(nbd);
 806         atomic_dec(&config->recv_threads);
 807         wake_up(&config->recv_wq);
 808         kfree(args);
 809 }
 810
 811 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 812 {
 813         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 814
 815         /* don't abort one completed request */
 816         if (blk_mq_request_completed(req))
 817                 return true;
 818
 819         mutex_lock(&cmd->lock);
 820         cmd->status = BLK_STS_IOERR;
 821         mutex_unlock(&cmd->lock);
 822
 823         blk_mq_complete_request(req);
 824         return true;
 825 }
 826
 827 static void nbd_clear_que(struct nbd_device *nbd)
 828 {
 829         blk_mq_quiesce_queue(nbd->disk->queue);
 830         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 831         blk_mq_unquiesce_queue(nbd->disk->queue);
 832         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 833 }
 834
 835 static int find_fallback(struct nbd_device *nbd, int index)
 836 {
 837         struct nbd_config *config = nbd->config;
 838         int new_index = -1;
 839         struct nbd_sock *nsock = config->socks[index];
 840         int fallback = nsock->fallback_index;
 841
 842         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 843                 return new_index;
 844
 845         if (config->num_connections <= 1) {
 846                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 847                                     "Dead connection, failed to find a fallback\n");
 848                 return new_index;
 849         }
 850
 851         if (fallback >= 0 && fallback < config->num_connections &&
 852             !config->socks[fallback]->dead)
 853                 return fallback;
 854
 855         if (nsock->fallback_index < 0 ||
 856             nsock->fallback_index >= config->num_connections ||
 857             config->socks[nsock->fallback_index]->dead) {
 858                 int i;
 859                 for (i = 0; i < config->num_connections; i++) {
 860                         if (i == index)
 861                                 continue;
 862                         if (!config->socks[i]->dead) {
 863                                 new_index = i;
 864                                 break;
 865                         }
 866                 }
 867                 nsock->fallback_index = new_index;
 868                 if (new_index < 0) {
 869                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 870                                             "Dead connection, failed to find a fallback\n");
 871                         return new_index;
 872                 }
 873         }
 874         new_index = nsock->fallback_index;
 875         return new_index;
 876 }
 877
 878 static int wait_for_reconnect(struct nbd_device *nbd)
 879 {
 880         struct nbd_config *config = nbd->config;
 881         if (!config->dead_conn_timeout)
 882                 return 0;
 883
 884         if (!wait_event_timeout(config->conn_wait,
 885                                 test_bit(NBD_RT_DISCONNECTED,
 886                                          &config->runtime_flags) ||
 887                                 atomic_read(&config->live_connections) > 0,
 888                                 config->dead_conn_timeout))
 889                 return 0;
 890
 891         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
 892 }
 893
 894 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 895 {
 896         struct request *req = blk_mq_rq_from_pdu(cmd);
 897         struct nbd_device *nbd = cmd->nbd;
 898         struct nbd_config *config;
 899         struct nbd_sock *nsock;
 900         int ret;
 901
 902         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 903                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 904                                     "Socks array is empty\n");
 905                 blk_mq_start_request(req);
 906                 return -EINVAL;
 907         }
 908         config = nbd->config;
 909
 910         if (index >= config->num_connections) {
 911                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 912                                     "Attempted send on invalid socket\n");
 913                 nbd_config_put(nbd);
 914                 blk_mq_start_request(req);
 915                 return -EINVAL;
 916         }
 917         cmd->status = BLK_STS_OK;
 918 again:
 919         nsock = config->socks[index];
 920         mutex_lock(&nsock->tx_lock);
 921         if (nsock->dead) {
 922                 int old_index = index;
 923                 index = find_fallback(nbd, index);
 924                 mutex_unlock(&nsock->tx_lock);
 925                 if (index < 0) {
 926                         if (wait_for_reconnect(nbd)) {
 927                                 index = old_index;
 928                                 goto again;
 929                         }
 930                         /* All the sockets should already be down at this point,
 931                          * we just want to make sure that DISCONNECTED is set so
 932                          * any requests that come in that were queue'ed waiting
 933                          * for the reconnect timer don't trigger the timer again
 934                          * and instead just error out.
 935                          */
 936                         sock_shutdown(nbd);
 937                         nbd_config_put(nbd);
 938                         blk_mq_start_request(req);
 939                         return -EIO;
 940                 }
 941                 goto again;
 942         }
 943
 944         /* Handle the case that we have a pending request that was partially
 945          * transmitted that _has_ to be serviced first.  We need to call requeue
 946          * here so that it gets put _after_ the request that is already on the
 947          * dispatch list.
 948          */
 949         blk_mq_start_request(req);
 950         if (unlikely(nsock->pending && nsock->pending != req)) {
 951                 nbd_requeue_cmd(cmd);
 952                 ret = 0;
 953                 goto out;
 954         }
 955         /*
 956          * Some failures are related to the link going down, so anything that
 957          * returns EAGAIN can be retried on a different socket.
 958          */
 959         ret = nbd_send_cmd(nbd, cmd, index);
 960         if (ret == -EAGAIN) {
 961                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 962                                     "Request send failed, requeueing\n");
 963                 nbd_mark_nsock_dead(nbd, nsock, 1);
 964                 nbd_requeue_cmd(cmd);
 965                 ret = 0;
 966         }
 967 out:
 968         mutex_unlock(&nsock->tx_lock);
 969         nbd_config_put(nbd);
 970         return ret;
 971 }
 972
 973 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 974                         const struct blk_mq_queue_data *bd)
 975 {
 976         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 977         int ret;
 978
 979         /*
 980          * Since we look at the bio's to send the request over the network we
 981          * need to make sure the completion work doesn't mark this request done
 982          * before we are done doing our send.  This keeps us from dereferencing
 983          * freed data if we have particularly fast completions (ie we get the
 984          * completion before we exit sock_xmit on the last bvec) or in the case
 985          * that the server is misbehaving (or there was an error) before we're
 986          * done sending everything over the wire.
 987          */
 988         mutex_lock(&cmd->lock);
 989         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 990
 991         /* We can be called directly from the user space process, which means we
 992          * could possibly have signals pending so our sendmsg will fail.  In
 993          * this case we need to return that we are busy, otherwise error out as
 994          * appropriate.
 995          */
 996         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 997         if (ret < 0)
 998                 ret = BLK_STS_IOERR;
 999         else if (!ret)
1000                 ret = BLK_STS_OK;
1001         mutex_unlock(&cmd->lock);
1002
1003         return ret;
1004 }
1005
1006 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1007                                      int *err)
1008 {
1009         struct socket *sock;
1010
1011         *err = 0;
1012         sock = sockfd_lookup(fd, err);
1013         if (!sock)
1014                 return NULL;
1015
1016         if (sock->ops->shutdown == sock_no_shutdown) {
1017                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1018                 *err = -EINVAL;
1019                 sockfd_put(sock);
1020                 return NULL;
1021         }
1022
1023         return sock;
1024 }
1025
1026 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1027                           bool netlink)
1028 {
1029         struct nbd_config *config = nbd->config;
1030         struct socket *sock;
1031         struct nbd_sock **socks;
1032         struct nbd_sock *nsock;
1033         int err;
1034
1035         sock = nbd_get_socket(nbd, arg, &err);
1036         if (!sock)
1037                 return err;
1038
1039         /*
1040          * We need to make sure we don't get any errant requests while we're
1041          * reallocating the ->socks array.
1042          */
1043         blk_mq_freeze_queue(nbd->disk->queue);
1044
1045         if (!netlink && !nbd->task_setup &&
1046             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1047                 nbd->task_setup = current;
1048
1049         if (!netlink &&
1050             (nbd->task_setup != current ||
1051              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1052                 dev_err(disk_to_dev(nbd->disk),
1053                         "Device being setup by another task");
1054                 err = -EBUSY;
1055                 goto put_socket;
1056         }
1057
1058         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1059         if (!nsock) {
1060                 err = -ENOMEM;
1061                 goto put_socket;
1062         }
1063
1064         socks = krealloc(config->socks, (config->num_connections + 1) *
1065                          sizeof(struct nbd_sock *), GFP_KERNEL);
1066         if (!socks) {
1067                 kfree(nsock);
1068                 err = -ENOMEM;
1069                 goto put_socket;
1070         }
1071
1072         config->socks = socks;
1073
1074         nsock->fallback_index = -1;
1075         nsock->dead = false;
1076         mutex_init(&nsock->tx_lock);
1077         nsock->sock = sock;
1078         nsock->pending = NULL;
1079         nsock->sent = 0;
1080         nsock->cookie = 0;
1081         socks[config->num_connections++] = nsock;
1082         atomic_inc(&config->live_connections);
1083         blk_mq_unfreeze_queue(nbd->disk->queue);
1084
1085         return 0;
1086
1087 put_socket:
1088         blk_mq_unfreeze_queue(nbd->disk->queue);
1089         sockfd_put(sock);
1090         return err;
1091 }
1092
1093 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1094 {
1095         struct nbd_config *config = nbd->config;
1096         struct socket *sock, *old;
1097         struct recv_thread_args *args;
1098         int i;
1099         int err;
1100
1101         sock = nbd_get_socket(nbd, arg, &err);
1102         if (!sock)
1103                 return err;
1104
1105         args = kzalloc(sizeof(*args), GFP_KERNEL);
1106         if (!args) {
1107                 sockfd_put(sock);
1108                 return -ENOMEM;
1109         }
1110
1111         for (i = 0; i < config->num_connections; i++) {
1112                 struct nbd_sock *nsock = config->socks[i];
1113
1114                 if (!nsock->dead)
1115                         continue;
1116
1117                 mutex_lock(&nsock->tx_lock);
1118                 if (!nsock->dead) {
1119                         mutex_unlock(&nsock->tx_lock);
1120                         continue;
1121                 }
1122                 sk_set_memalloc(sock->sk);
1123                 if (nbd->tag_set.timeout)
1124                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1125                 atomic_inc(&config->recv_threads);
1126                 refcount_inc(&nbd->config_refs);
1127                 old = nsock->sock;
1128                 nsock->fallback_index = -1;
1129                 nsock->sock = sock;
1130                 nsock->dead = false;
1131                 INIT_WORK(&args->work, recv_work);
1132                 args->index = i;
1133                 args->nbd = nbd;
1134                 nsock->cookie++;
1135                 mutex_unlock(&nsock->tx_lock);
1136                 sockfd_put(old);
1137
1138                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1139
1140                 /* We take the tx_mutex in an error path in the recv_work, so we
1141                  * need to queue_work outside of the tx_mutex.
1142                  */
1143                 queue_work(nbd->recv_workq, &args->work);
1144
1145                 atomic_inc(&config->live_connections);
1146                 wake_up(&config->conn_wait);
1147                 return 0;
1148         }
1149         sockfd_put(sock);
1150         kfree(args);
1151         return -ENOSPC;
1152 }
1153
1154 static void nbd_bdev_reset(struct block_device *bdev)
1155 {
1156         if (bdev->bd_openers > 1)
1157                 return;
1158         bd_set_nr_sectors(bdev, 0);
1159 }
1160
1161 static void nbd_parse_flags(struct nbd_device *nbd)
1162 {
1163         struct nbd_config *config = nbd->config;
1164         if (config->flags & NBD_FLAG_READ_ONLY)
1165                 set_disk_ro(nbd->disk, true);
1166         else
1167                 set_disk_ro(nbd->disk, false);
1168         if (config->flags & NBD_FLAG_SEND_TRIM)
1169                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1170         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1171                 if (config->flags & NBD_FLAG_SEND_FUA)
1172                         blk_queue_write_cache(nbd->disk->queue, true, true);
1173                 else
1174                         blk_queue_write_cache(nbd->disk->queue, true, false);
1175         }
1176         else
1177                 blk_queue_write_cache(nbd->disk->queue, false, false);
1178 }
1179
1180 static void send_disconnects(struct nbd_device *nbd)
1181 {
1182         struct nbd_config *config = nbd->config;
1183         struct nbd_request request = {
1184                 .magic = htonl(NBD_REQUEST_MAGIC),
1185                 .type = htonl(NBD_CMD_DISC),
1186         };
1187         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1188         struct iov_iter from;
1189         int i, ret;
1190
1191         for (i = 0; i < config->num_connections; i++) {
1192                 struct nbd_sock *nsock = config->socks[i];
1193
1194                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1195                 mutex_lock(&nsock->tx_lock);
1196                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1197                 if (ret <= 0)
1198                         dev_err(disk_to_dev(nbd->disk),
1199                                 "Send disconnect failed %d\n", ret);
1200                 mutex_unlock(&nsock->tx_lock);
1201         }
1202 }
1203
1204 static int nbd_disconnect(struct nbd_device *nbd)
1205 {
1206         struct nbd_config *config = nbd->config;
1207
1208         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1209         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1210         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1211         send_disconnects(nbd);
1212         return 0;
1213 }
1214
1215 static void nbd_clear_sock(struct nbd_device *nbd)
1216 {
1217         sock_shutdown(nbd);
1218         nbd_clear_que(nbd);
1219         nbd->task_setup = NULL;
1220 }
1221
1222 static void nbd_config_put(struct nbd_device *nbd)
1223 {
1224         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1225                                         &nbd->config_lock)) {
1226                 struct nbd_config *config = nbd->config;
1227                 nbd_dev_dbg_close(nbd);
1228                 nbd_size_clear(nbd);
1229                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1230                                        &config->runtime_flags))
1231                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1232                 nbd->task_recv = NULL;
1233                 nbd_clear_sock(nbd);
1234                 if (config->num_connections) {
1235                         int i;
1236                         for (i = 0; i < config->num_connections; i++) {
1237                                 sockfd_put(config->socks[i]->sock);
1238                                 kfree(config->socks[i]);
1239                         }
1240                         kfree(config->socks);
1241                 }
1242                 kfree(nbd->config);
1243                 nbd->config = NULL;
1244
1245                 if (nbd->recv_workq)
1246                         destroy_workqueue(nbd->recv_workq);
1247                 nbd->recv_workq = NULL;
1248
1249                 nbd->tag_set.timeout = 0;
1250                 nbd->disk->queue->limits.discard_granularity = 0;
1251                 nbd->disk->queue->limits.discard_alignment = 0;
1252                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1253                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1254
1255                 mutex_unlock(&nbd->config_lock);
1256                 nbd_put(nbd);
1257                 module_put(THIS_MODULE);
1258         }
1259 }
1260
1261 static int nbd_start_device(struct nbd_device *nbd)
1262 {
1263         struct nbd_config *config = nbd->config;
1264         int num_connections = config->num_connections;
1265         int error = 0, i;
1266
1267         if (nbd->task_recv)
1268                 return -EBUSY;
1269         if (!config->socks)
1270                 return -EINVAL;
1271         if (num_connections > 1 &&
1272             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1273                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1274                 return -EINVAL;
1275         }
1276
1277         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1278                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1279                                           WQ_UNBOUND, 0, nbd->index);
1280         if (!nbd->recv_workq) {
1281                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1282                 return -ENOMEM;
1283         }
1284
1285         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1286         nbd->task_recv = current;
1287
1288         nbd_parse_flags(nbd);
1289
1290         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1291         if (error) {
1292                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1293                 return error;
1294         }
1295         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1296
1297         nbd_dev_dbg_init(nbd);
1298         for (i = 0; i < num_connections; i++) {
1299                 struct recv_thread_args *args;
1300
1301                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1302                 if (!args) {
1303                         sock_shutdown(nbd);
1304                         /*
1305                          * If num_connections is m (2 < m),
1306                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1307                          * But NO.(n + 1) failed. We still have n recv threads.
1308                          * So, add flush_workqueue here to prevent recv threads
1309                          * dropping the last config_refs and trying to destroy
1310                          * the workqueue from inside the workqueue.
1311                          */
1312                         if (i)
1313                                 flush_workqueue(nbd->recv_workq);
1314                         return -ENOMEM;
1315                 }
1316                 sk_set_memalloc(config->socks[i]->sock->sk);
1317                 if (nbd->tag_set.timeout)
1318                         config->socks[i]->sock->sk->sk_sndtimeo =
1319                                 nbd->tag_set.timeout;
1320                 atomic_inc(&config->recv_threads);
1321                 refcount_inc(&nbd->config_refs);
1322                 INIT_WORK(&args->work, recv_work);
1323                 args->nbd = nbd;
1324                 args->index = i;
1325                 queue_work(nbd->recv_workq, &args->work);
1326         }
1327         nbd_size_update(nbd, true);
1328         return error;
1329 }
1330
1331 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1332 {
1333         struct nbd_config *config = nbd->config;
1334         int ret;
1335
1336         ret = nbd_start_device(nbd);
1337         if (ret)
1338                 return ret;
1339
1340         if (max_part)
1341                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1342         mutex_unlock(&nbd->config_lock);
1343         ret = wait_event_interruptible(config->recv_wq,
1344                                          atomic_read(&config->recv_threads) == 0);
1345         if (ret) {
1346                 sock_shutdown(nbd);
1347                 nbd_clear_que(nbd);
1348         }
1349
1350         flush_workqueue(nbd->recv_workq);
1351         mutex_lock(&nbd->config_lock);
1352         nbd_bdev_reset(bdev);
1353         /* user requested, ignore socket errors */
1354         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1355                 ret = 0;
1356         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1357                 ret = -ETIMEDOUT;
1358         return ret;
1359 }
1360
1361 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1362                                  struct block_device *bdev)
1363 {
1364         nbd_clear_sock(nbd);
1365         __invalidate_device(bdev, true);
1366         nbd_bdev_reset(bdev);
1367         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1368                                &nbd->config->runtime_flags))
1369                 nbd_config_put(nbd);
1370 }
1371
1372 static bool nbd_is_valid_blksize(unsigned long blksize)
1373 {
1374         if (!blksize || !is_power_of_2(blksize) || blksize < 512 ||
1375             blksize > PAGE_SIZE)
1376                 return false;
1377         return true;
1378 }
1379
1380 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1381 {
1382         nbd->tag_set.timeout = timeout * HZ;
1383         if (timeout)
1384                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1385         else
1386                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1387 }
1388
1389 /* Must be called with config_lock held */
1390 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1391                        unsigned int cmd, unsigned long arg)
1392 {
1393         struct nbd_config *config = nbd->config;
1394
1395         switch (cmd) {
1396         case NBD_DISCONNECT:
1397                 return nbd_disconnect(nbd);
1398         case NBD_CLEAR_SOCK:
1399                 nbd_clear_sock_ioctl(nbd, bdev);
1400                 return 0;
1401         case NBD_SET_SOCK:
1402                 return nbd_add_socket(nbd, arg, false);
1403         case NBD_SET_BLKSIZE:
1404                 if (!arg)
1405                         arg = NBD_DEF_BLKSIZE;
1406                 if (!nbd_is_valid_blksize(arg))
1407                         return -EINVAL;
1408                 nbd_size_set(nbd, arg,
1409                              div_s64(config->bytesize, arg));
1410                 return 0;
1411         case NBD_SET_SIZE:
1412                 nbd_size_set(nbd, config->blksize,
1413                              div_s64(arg, config->blksize));
1414                 return 0;
1415         case NBD_SET_SIZE_BLOCKS:
1416                 nbd_size_set(nbd, config->blksize, arg);
1417                 return 0;
1418         case NBD_SET_TIMEOUT:
1419                 nbd_set_cmd_timeout(nbd, arg);
1420                 return 0;
1421
1422         case NBD_SET_FLAGS:
1423                 config->flags = arg;
1424                 return 0;
1425         case NBD_DO_IT:
1426                 return nbd_start_device_ioctl(nbd, bdev);
1427         case NBD_CLEAR_QUE:
1428                 /*
1429                  * This is for compatibility only.  The queue is always cleared
1430                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1431                  */
1432                 return 0;
1433         case NBD_PRINT_DEBUG:
1434                 /*
1435                  * For compatibility only, we no longer keep a list of
1436                  * outstanding requests.
1437                  */
1438                 return 0;
1439         }
1440         return -ENOTTY;
1441 }
1442
1443 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1444                      unsigned int cmd, unsigned long arg)
1445 {
1446         struct nbd_device *nbd = bdev->bd_disk->private_data;
1447         struct nbd_config *config = nbd->config;
1448         int error = -EINVAL;
1449
1450         if (!capable(CAP_SYS_ADMIN))
1451                 return -EPERM;
1452
1453         /* The block layer will pass back some non-nbd ioctls in case we have
1454          * special handling for them, but we don't so just return an error.
1455          */
1456         if (_IOC_TYPE(cmd) != 0xab)
1457                 return -EINVAL;
1458
1459         mutex_lock(&nbd->config_lock);
1460
1461         /* Don't allow ioctl operations on a nbd device that was created with
1462          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1463          */
1464         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1465             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1466                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1467         else
1468                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1469         mutex_unlock(&nbd->config_lock);
1470         return error;
1471 }
1472
1473 static struct nbd_config *nbd_alloc_config(void)
1474 {
1475         struct nbd_config *config;
1476
1477         if (!try_module_get(THIS_MODULE))
1478                 return ERR_PTR(-ENODEV);
1479
1480         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1481         if (!config) {
1482                 module_put(THIS_MODULE);
1483                 return ERR_PTR(-ENOMEM);
1484         }
1485
1486         atomic_set(&config->recv_threads, 0);
1487         init_waitqueue_head(&config->recv_wq);
1488         init_waitqueue_head(&config->conn_wait);
1489         config->blksize = NBD_DEF_BLKSIZE;
1490         atomic_set(&config->live_connections, 0);
1491         return config;
1492 }
1493
1494 static int nbd_open(struct block_device *bdev, fmode_t mode)
1495 {
1496         struct nbd_device *nbd;
1497         int ret = 0;
1498
1499         mutex_lock(&nbd_index_mutex);
1500         nbd = bdev->bd_disk->private_data;
1501         if (!nbd) {
1502                 ret = -ENXIO;
1503                 goto out;
1504         }
1505         if (!refcount_inc_not_zero(&nbd->refs)) {
1506                 ret = -ENXIO;
1507                 goto out;
1508         }
1509         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1510                 struct nbd_config *config;
1511
1512                 mutex_lock(&nbd->config_lock);
1513                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1514                         mutex_unlock(&nbd->config_lock);
1515                         goto out;
1516                 }
1517                 config = nbd_alloc_config();
1518                 if (IS_ERR(config)) {
1519                         ret = PTR_ERR(config);
1520                         mutex_unlock(&nbd->config_lock);
1521                         goto out;
1522                 }
1523                 nbd->config = config;
1524                 refcount_set(&nbd->config_refs, 1);
1525                 refcount_inc(&nbd->refs);
1526                 mutex_unlock(&nbd->config_lock);
1527                 set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1528         } else if (nbd_disconnected(nbd->config)) {
1529                 set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1530         }
1531 out:
1532         mutex_unlock(&nbd_index_mutex);
1533         return ret;
1534 }
1535
1536 static void nbd_release(struct gendisk *disk, fmode_t mode)
1537 {
1538         struct nbd_device *nbd = disk->private_data;
1539         struct block_device *bdev = bdget_disk(disk, 0);
1540
1541         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1542                         bdev->bd_openers == 0)
1543                 nbd_disconnect_and_put(nbd);
1544         bdput(bdev);
1545
1546         nbd_config_put(nbd);
1547         nbd_put(nbd);
1548 }
1549
1550 static const struct block_device_operations nbd_fops =
1551 {
1552         .owner =        THIS_MODULE,
1553         .open =         nbd_open,
1554         .release =      nbd_release,
1555         .ioctl =        nbd_ioctl,
1556         .compat_ioctl = nbd_ioctl,
1557 };
1558
1559 #if IS_ENABLED(CONFIG_DEBUG_FS)
1560
1561 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1562 {
1563         struct nbd_device *nbd = s->private;
1564
1565         if (nbd->task_recv)
1566                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1567
1568         return 0;
1569 }
1570
1571 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1572 {
1573         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1574 }
1575
1576 static const struct file_operations nbd_dbg_tasks_ops = {
1577         .open = nbd_dbg_tasks_open,
1578         .read = seq_read,
1579         .llseek = seq_lseek,
1580         .release = single_release,
1581 };
1582
1583 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1584 {
1585         struct nbd_device *nbd = s->private;
1586         u32 flags = nbd->config->flags;
1587
1588         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1589
1590         seq_puts(s, "Known flags:\n");
1591
1592         if (flags & NBD_FLAG_HAS_FLAGS)
1593                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1594         if (flags & NBD_FLAG_READ_ONLY)
1595                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1596         if (flags & NBD_FLAG_SEND_FLUSH)
1597                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1598         if (flags & NBD_FLAG_SEND_FUA)
1599                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1600         if (flags & NBD_FLAG_SEND_TRIM)
1601                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1602
1603         return 0;
1604 }
1605
1606 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1607 {
1608         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1609 }
1610
1611 static const struct file_operations nbd_dbg_flags_ops = {
1612         .open = nbd_dbg_flags_open,
1613         .read = seq_read,
1614         .llseek = seq_lseek,
1615         .release = single_release,
1616 };
1617
1618 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1619 {
1620         struct dentry *dir;
1621         struct nbd_config *config = nbd->config;
1622
1623         if (!nbd_dbg_dir)
1624                 return -EIO;
1625
1626         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1627         if (IS_ERR(dir)) {
1628                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1629                         nbd_name(nbd));
1630                 return -EIO;
1631         }
1632         config->dbg_dir = dir;
1633
1634         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1635         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1636         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1637         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1638         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1639
1640         return 0;
1641 }
1642
1643 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1644 {
1645         debugfs_remove_recursive(nbd->config->dbg_dir);
1646 }
1647
1648 static int nbd_dbg_init(void)
1649 {
1650         struct dentry *dbg_dir;
1651
1652         dbg_dir = debugfs_create_dir("nbd", NULL);
1653         if (IS_ERR(dbg_dir))
1654                 return -EIO;
1655
1656         nbd_dbg_dir = dbg_dir;
1657
1658         return 0;
1659 }
1660
1661 static void nbd_dbg_close(void)
1662 {
1663         debugfs_remove_recursive(nbd_dbg_dir);
1664 }
1665
1666 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1667
1668 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1669 {
1670         return 0;
1671 }
1672
1673 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1674 {
1675 }
1676
1677 static int nbd_dbg_init(void)
1678 {
1679         return 0;
1680 }
1681
1682 static void nbd_dbg_close(void)
1683 {
1684 }
1685
1686 #endif
1687
1688 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1689                             unsigned int hctx_idx, unsigned int numa_node)
1690 {
1691         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1692         cmd->nbd = set->driver_data;
1693         cmd->flags = 0;
1694         mutex_init(&cmd->lock);
1695         return 0;
1696 }
1697
1698 static const struct blk_mq_ops nbd_mq_ops = {
1699         .queue_rq       = nbd_queue_rq,
1700         .complete       = nbd_complete_rq,
1701         .init_request   = nbd_init_request,
1702         .timeout        = nbd_xmit_timeout,
1703 };
1704
1705 static int nbd_dev_add(int index)
1706 {
1707         struct nbd_device *nbd;
1708         struct gendisk *disk;
1709         struct request_queue *q;
1710         int err = -ENOMEM;
1711
1712         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1713         if (!nbd)
1714                 goto out;
1715
1716         disk = alloc_disk(1 << part_shift);
1717         if (!disk)
1718                 goto out_free_nbd;
1719
1720         if (index >= 0) {
1721                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1722                                 GFP_KERNEL);
1723                 if (err == -ENOSPC)
1724                         err = -EEXIST;
1725         } else {
1726                 err = idr_alloc(&nbd_index_idr, nbd, 0,
1727                                 (MINORMASK >> part_shift) + 1, GFP_KERNEL);
1728                 if (err >= 0)
1729                         index = err;
1730         }
1731         if (err < 0)
1732                 goto out_free_disk;
1733
1734         nbd->index = index;
1735         nbd->disk = disk;
1736         nbd->tag_set.ops = &nbd_mq_ops;
1737         nbd->tag_set.nr_hw_queues = 1;
1738         nbd->tag_set.queue_depth = 128;
1739         nbd->tag_set.numa_node = NUMA_NO_NODE;
1740         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1741         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1742                 BLK_MQ_F_BLOCKING;
1743         nbd->tag_set.driver_data = nbd;
1744         nbd->destroy_complete = NULL;
1745
1746         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1747         if (err)
1748                 goto out_free_idr;
1749
1750         q = blk_mq_init_queue(&nbd->tag_set);
1751         if (IS_ERR(q)) {
1752                 err = PTR_ERR(q);
1753                 goto out_free_tags;
1754         }
1755         disk->queue = q;
1756
1757         /*
1758          * Tell the block layer that we are not a rotational device
1759          */
1760         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1761         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1762         disk->queue->limits.discard_granularity = 0;
1763         disk->queue->limits.discard_alignment = 0;
1764         blk_queue_max_discard_sectors(disk->queue, 0);
1765         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1766         blk_queue_max_segments(disk->queue, USHRT_MAX);
1767         blk_queue_max_hw_sectors(disk->queue, 65536);
1768         disk->queue->limits.max_sectors = 256;
1769
1770         mutex_init(&nbd->config_lock);
1771         refcount_set(&nbd->config_refs, 0);
1772         refcount_set(&nbd->refs, 1);
1773         INIT_LIST_HEAD(&nbd->list);
1774         disk->major = NBD_MAJOR;
1775         disk->first_minor = index << part_shift;
1776         disk->fops = &nbd_fops;
1777         disk->private_data = nbd;
1778         sprintf(disk->disk_name, "nbd%d", index);
1779         add_disk(disk);
1780         nbd_total_devices++;
1781         return index;
1782
1783 out_free_tags:
1784         blk_mq_free_tag_set(&nbd->tag_set);
1785 out_free_idr:
1786         idr_remove(&nbd_index_idr, index);
1787 out_free_disk:
1788         put_disk(disk);
1789 out_free_nbd:
1790         kfree(nbd);
1791 out:
1792         return err;
1793 }
1794
1795 static int find_free_cb(int id, void *ptr, void *data)
1796 {
1797         struct nbd_device *nbd = ptr;
1798         struct nbd_device **found = data;
1799
1800         if (!refcount_read(&nbd->config_refs)) {
1801                 *found = nbd;
1802                 return 1;
1803         }
1804         return 0;
1805 }
1806
1807 /* Netlink interface. */
1808 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1809         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1810         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1811         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1812         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1813         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1814         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1815         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1816         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1817         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1818 };
1819
1820 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1821         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1822 };
1823
1824 /* We don't use this right now since we don't parse the incoming list, but we
1825  * still want it here so userspace knows what to expect.
1826  */
1827 static const struct nla_policy __attribute__((unused))
1828 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1829         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1830         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1831 };
1832
1833 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1834 {
1835         struct nbd_config *config = nbd->config;
1836         u64 bsize = config->blksize;
1837         u64 bytes = config->bytesize;
1838
1839         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1840                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1841
1842         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1843                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1844                 if (!bsize)
1845                         bsize = NBD_DEF_BLKSIZE;
1846                 if (!nbd_is_valid_blksize(bsize)) {
1847                         printk(KERN_ERR "Invalid block size %llu\n", bsize);
1848                         return -EINVAL;
1849                 }
1850         }
1851
1852         if (bytes != config->bytesize || bsize != config->blksize)
1853                 nbd_size_set(nbd, bsize, div64_u64(bytes, bsize));
1854         return 0;
1855 }
1856
1857 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1858 {
1859         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1860         struct nbd_device *nbd = NULL;
1861         struct nbd_config *config;
1862         int index = -1;
1863         int ret;
1864         bool put_dev = false;
1865
1866         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1867                 return -EPERM;
1868
1869         if (info->attrs[NBD_ATTR_INDEX]) {
1870                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1871
1872                 /*
1873                  * Too big first_minor can cause duplicate creation of
1874                  * sysfs files/links, since index << part_shift might overflow, or
1875                  * MKDEV() expect that the max bits of first_minor is 20.
1876                  */
1877                 if (index < 0 || index > MINORMASK >> part_shift) {
1878                         printk(KERN_ERR "nbd: illegal input index %d\n", index);
1879                         return -EINVAL;
1880                 }
1881         }
1882         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1883                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1884                 return -EINVAL;
1885         }
1886         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1887                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1888                 return -EINVAL;
1889         }
1890 again:
1891         mutex_lock(&nbd_index_mutex);
1892         if (index == -1) {
1893                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1894                 if (ret == 0) {
1895                         int new_index;
1896                         new_index = nbd_dev_add(-1);
1897                         if (new_index < 0) {
1898                                 mutex_unlock(&nbd_index_mutex);
1899                                 printk(KERN_ERR "nbd: failed to add new device\n");
1900                                 return new_index;
1901                         }
1902                         nbd = idr_find(&nbd_index_idr, new_index);
1903                 }
1904         } else {
1905                 nbd = idr_find(&nbd_index_idr, index);
1906                 if (!nbd) {
1907                         ret = nbd_dev_add(index);
1908                         if (ret < 0) {
1909                                 mutex_unlock(&nbd_index_mutex);
1910                                 printk(KERN_ERR "nbd: failed to add new device\n");
1911                                 return ret;
1912                         }
1913                         nbd = idr_find(&nbd_index_idr, index);
1914                 }
1915         }
1916         if (!nbd) {
1917                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1918                        index);
1919                 mutex_unlock(&nbd_index_mutex);
1920                 return -EINVAL;
1921         }
1922
1923         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1924             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1925                 nbd->destroy_complete = &destroy_complete;
1926                 mutex_unlock(&nbd_index_mutex);
1927
1928                 /* Wait untill the the nbd stuff is totally destroyed */
1929                 wait_for_completion(&destroy_complete);
1930                 goto again;
1931         }
1932
1933         if (!refcount_inc_not_zero(&nbd->refs)) {
1934                 mutex_unlock(&nbd_index_mutex);
1935                 if (index == -1)
1936                         goto again;
1937                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1938                        index);
1939                 return -EINVAL;
1940         }
1941         mutex_unlock(&nbd_index_mutex);
1942
1943         mutex_lock(&nbd->config_lock);
1944         if (refcount_read(&nbd->config_refs)) {
1945                 mutex_unlock(&nbd->config_lock);
1946                 nbd_put(nbd);
1947                 if (index == -1)
1948                         goto again;
1949                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1950                 return -EBUSY;
1951         }
1952         if (WARN_ON(nbd->config)) {
1953                 mutex_unlock(&nbd->config_lock);
1954                 nbd_put(nbd);
1955                 return -EINVAL;
1956         }
1957         config = nbd_alloc_config();
1958         if (IS_ERR(config)) {
1959                 mutex_unlock(&nbd->config_lock);
1960                 nbd_put(nbd);
1961                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1962                 return PTR_ERR(config);
1963         }
1964         nbd->config = config;
1965         refcount_set(&nbd->config_refs, 1);
1966         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1967
1968         ret = nbd_genl_size_set(info, nbd);
1969         if (ret)
1970                 goto out;
1971
1972         if (info->attrs[NBD_ATTR_TIMEOUT])
1973                 nbd_set_cmd_timeout(nbd,
1974                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1975         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1976                 config->dead_conn_timeout =
1977                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1978                 config->dead_conn_timeout *= HZ;
1979         }
1980         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1981                 config->flags =
1982                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1983         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1984                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1985                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1986                         /*
1987                          * We have 1 ref to keep the device around, and then 1
1988                          * ref for our current operation here, which will be
1989                          * inherited by the config.  If we already have
1990                          * DESTROY_ON_DISCONNECT set then we know we don't have
1991                          * that extra ref already held so we don't need the
1992                          * put_dev.
1993                          */
1994                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1995                                               &nbd->flags))
1996                                 put_dev = true;
1997                 } else {
1998                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1999                                                &nbd->flags))
2000                                 refcount_inc(&nbd->refs);
2001                 }
2002                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2003                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2004                                 &config->runtime_flags);
2005                 }
2006         }
2007
2008         if (info->attrs[NBD_ATTR_SOCKETS]) {
2009                 struct nlattr *attr;
2010                 int rem, fd;
2011
2012                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2013                                     rem) {
2014                         struct nlattr *socks[NBD_SOCK_MAX+1];
2015
2016                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2017                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2018                                 ret = -EINVAL;
2019                                 goto out;
2020                         }
2021                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2022                                                           attr,
2023                                                           nbd_sock_policy,
2024                                                           info->extack);
2025                         if (ret != 0) {
2026                                 printk(KERN_ERR "nbd: error processing sock list\n");
2027                                 ret = -EINVAL;
2028                                 goto out;
2029                         }
2030                         if (!socks[NBD_SOCK_FD])
2031                                 continue;
2032                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2033                         ret = nbd_add_socket(nbd, fd, true);
2034                         if (ret)
2035                                 goto out;
2036                 }
2037         }
2038         ret = nbd_start_device(nbd);
2039 out:
2040         mutex_unlock(&nbd->config_lock);
2041         if (!ret) {
2042                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2043                 refcount_inc(&nbd->config_refs);
2044                 nbd_connect_reply(info, nbd->index);
2045         }
2046         nbd_config_put(nbd);
2047         if (put_dev)
2048                 nbd_put(nbd);
2049         return ret;
2050 }
2051
2052 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2053 {
2054         mutex_lock(&nbd->config_lock);
2055         nbd_disconnect(nbd);
2056         sock_shutdown(nbd);
2057         wake_up(&nbd->config->conn_wait);
2058         /*
2059          * Make sure recv thread has finished, so it does not drop the last
2060          * config ref and try to destroy the workqueue from inside the work
2061          * queue. And this also ensure that we can safely call nbd_clear_que()
2062          * to cancel the inflight I/Os.
2063          */
2064         if (nbd->recv_workq)
2065                 flush_workqueue(nbd->recv_workq);
2066         nbd_clear_que(nbd);
2067         nbd->task_setup = NULL;
2068         mutex_unlock(&nbd->config_lock);
2069
2070         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2071                                &nbd->config->runtime_flags))
2072                 nbd_config_put(nbd);
2073 }
2074
2075 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2076 {
2077         struct nbd_device *nbd;
2078         int index;
2079
2080         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2081                 return -EPERM;
2082
2083         if (!info->attrs[NBD_ATTR_INDEX]) {
2084                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2085                 return -EINVAL;
2086         }
2087         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2088         mutex_lock(&nbd_index_mutex);
2089         nbd = idr_find(&nbd_index_idr, index);
2090         if (!nbd) {
2091                 mutex_unlock(&nbd_index_mutex);
2092                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2093                        index);
2094                 return -EINVAL;
2095         }
2096         if (!refcount_inc_not_zero(&nbd->refs)) {
2097                 mutex_unlock(&nbd_index_mutex);
2098                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2099                        index);
2100                 return -EINVAL;
2101         }
2102         mutex_unlock(&nbd_index_mutex);
2103         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2104                 nbd_put(nbd);
2105                 return 0;
2106         }
2107         nbd_disconnect_and_put(nbd);
2108         nbd_config_put(nbd);
2109         nbd_put(nbd);
2110         return 0;
2111 }
2112
2113 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2114 {
2115         struct nbd_device *nbd = NULL;
2116         struct nbd_config *config;
2117         int index;
2118         int ret = 0;
2119         bool put_dev = false;
2120
2121         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2122                 return -EPERM;
2123
2124         if (!info->attrs[NBD_ATTR_INDEX]) {
2125                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2126                 return -EINVAL;
2127         }
2128         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2129         mutex_lock(&nbd_index_mutex);
2130         nbd = idr_find(&nbd_index_idr, index);
2131         if (!nbd) {
2132                 mutex_unlock(&nbd_index_mutex);
2133                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2134                        index);
2135                 return -EINVAL;
2136         }
2137         if (!refcount_inc_not_zero(&nbd->refs)) {
2138                 mutex_unlock(&nbd_index_mutex);
2139                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2140                        index);
2141                 return -EINVAL;
2142         }
2143         mutex_unlock(&nbd_index_mutex);
2144
2145         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2146                 dev_err(nbd_to_dev(nbd),
2147                         "not configured, cannot reconfigure\n");
2148                 nbd_put(nbd);
2149                 return -EINVAL;
2150         }
2151
2152         mutex_lock(&nbd->config_lock);
2153         config = nbd->config;
2154         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2155             !nbd->task_recv) {
2156                 dev_err(nbd_to_dev(nbd),
2157                         "not configured, cannot reconfigure\n");
2158                 ret = -EINVAL;
2159                 goto out;
2160         }
2161
2162         ret = nbd_genl_size_set(info, nbd);
2163         if (ret)
2164                 goto out;
2165
2166         if (info->attrs[NBD_ATTR_TIMEOUT])
2167                 nbd_set_cmd_timeout(nbd,
2168                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2169         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2170                 config->dead_conn_timeout =
2171                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2172                 config->dead_conn_timeout *= HZ;
2173         }
2174         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2175                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2176                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2177                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2178                                               &nbd->flags))
2179                                 put_dev = true;
2180                 } else {
2181                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2182                                                &nbd->flags))
2183                                 refcount_inc(&nbd->refs);
2184                 }
2185
2186                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2187                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2188                                         &config->runtime_flags);
2189                 } else {
2190                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2191                                         &config->runtime_flags);
2192                 }
2193         }
2194
2195         if (info->attrs[NBD_ATTR_SOCKETS]) {
2196                 struct nlattr *attr;
2197                 int rem, fd;
2198
2199                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2200                                     rem) {
2201                         struct nlattr *socks[NBD_SOCK_MAX+1];
2202
2203                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2204                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2205                                 ret = -EINVAL;
2206                                 goto out;
2207                         }
2208                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2209                                                           attr,
2210                                                           nbd_sock_policy,
2211                                                           info->extack);
2212                         if (ret != 0) {
2213                                 printk(KERN_ERR "nbd: error processing sock list\n");
2214                                 ret = -EINVAL;
2215                                 goto out;
2216                         }
2217                         if (!socks[NBD_SOCK_FD])
2218                                 continue;
2219                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2220                         ret = nbd_reconnect_socket(nbd, fd);
2221                         if (ret) {
2222                                 if (ret == -ENOSPC)
2223                                         ret = 0;
2224                                 goto out;
2225                         }
2226                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2227                 }
2228         }
2229 out:
2230         mutex_unlock(&nbd->config_lock);
2231         nbd_config_put(nbd);
2232         nbd_put(nbd);
2233         if (put_dev)
2234                 nbd_put(nbd);
2235         return ret;
2236 }
2237
2238 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2239         {
2240                 .cmd    = NBD_CMD_CONNECT,
2241                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2242                 .doit   = nbd_genl_connect,
2243         },
2244         {
2245                 .cmd    = NBD_CMD_DISCONNECT,
2246                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2247                 .doit   = nbd_genl_disconnect,
2248         },
2249         {
2250                 .cmd    = NBD_CMD_RECONFIGURE,
2251                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2252                 .doit   = nbd_genl_reconfigure,
2253         },
2254         {
2255                 .cmd    = NBD_CMD_STATUS,
2256                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2257                 .doit   = nbd_genl_status,
2258         },
2259 };
2260
2261 static const struct genl_multicast_group nbd_mcast_grps[] = {
2262         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2263 };
2264
2265 static struct genl_family nbd_genl_family __ro_after_init = {
2266         .hdrsize        = 0,
2267         .name           = NBD_GENL_FAMILY_NAME,
2268         .version        = NBD_GENL_VERSION,
2269         .module         = THIS_MODULE,
2270         .small_ops      = nbd_connect_genl_ops,
2271         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2272         .maxattr        = NBD_ATTR_MAX,
2273         .policy = nbd_attr_policy,
2274         .mcgrps         = nbd_mcast_grps,
2275         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2276 };
2277
2278 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2279 {
2280         struct nlattr *dev_opt;
2281         u8 connected = 0;
2282         int ret;
2283
2284         /* This is a little racey, but for status it's ok.  The
2285          * reason we don't take a ref here is because we can't
2286          * take a ref in the index == -1 case as we would need
2287          * to put under the nbd_index_mutex, which could
2288          * deadlock if we are configured to remove ourselves
2289          * once we're disconnected.
2290          */
2291         if (refcount_read(&nbd->config_refs))
2292                 connected = 1;
2293         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2294         if (!dev_opt)
2295                 return -EMSGSIZE;
2296         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2297         if (ret)
2298                 return -EMSGSIZE;
2299         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2300                          connected);
2301         if (ret)
2302                 return -EMSGSIZE;
2303         nla_nest_end(reply, dev_opt);
2304         return 0;
2305 }
2306
2307 static int status_cb(int id, void *ptr, void *data)
2308 {
2309         struct nbd_device *nbd = ptr;
2310         return populate_nbd_status(nbd, (struct sk_buff *)data);
2311 }
2312
2313 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2314 {
2315         struct nlattr *dev_list;
2316         struct sk_buff *reply;
2317         void *reply_head;
2318         size_t msg_size;
2319         int index = -1;
2320         int ret = -ENOMEM;
2321
2322         if (info->attrs[NBD_ATTR_INDEX])
2323                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2324
2325         mutex_lock(&nbd_index_mutex);
2326
2327         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2328                                   nla_attr_size(sizeof(u8)));
2329         msg_size *= (index == -1) ? nbd_total_devices : 1;
2330
2331         reply = genlmsg_new(msg_size, GFP_KERNEL);
2332         if (!reply)
2333                 goto out;
2334         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2335                                        NBD_CMD_STATUS);
2336         if (!reply_head) {
2337                 nlmsg_free(reply);
2338                 goto out;
2339         }
2340
2341         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2342         if (!dev_list) {
2343                 nlmsg_free(reply);
2344                 ret = -EMSGSIZE;
2345                 goto out;
2346         }
2347
2348         if (index == -1) {
2349                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2350                 if (ret) {
2351                         nlmsg_free(reply);
2352                         goto out;
2353                 }
2354         } else {
2355                 struct nbd_device *nbd;
2356                 nbd = idr_find(&nbd_index_idr, index);
2357                 if (nbd) {
2358                         ret = populate_nbd_status(nbd, reply);
2359                         if (ret) {
2360                                 nlmsg_free(reply);
2361                                 goto out;
2362                         }
2363                 }
2364         }
2365         nla_nest_end(reply, dev_list);
2366         genlmsg_end(reply, reply_head);
2367         ret = genlmsg_reply(reply, info);
2368 out:
2369         mutex_unlock(&nbd_index_mutex);
2370         return ret;
2371 }
2372
2373 static void nbd_connect_reply(struct genl_info *info, int index)
2374 {
2375         struct sk_buff *skb;
2376         void *msg_head;
2377         int ret;
2378
2379         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2380         if (!skb)
2381                 return;
2382         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2383                                      NBD_CMD_CONNECT);
2384         if (!msg_head) {
2385                 nlmsg_free(skb);
2386                 return;
2387         }
2388         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2389         if (ret) {
2390                 nlmsg_free(skb);
2391                 return;
2392         }
2393         genlmsg_end(skb, msg_head);
2394         genlmsg_reply(skb, info);
2395 }
2396
2397 static void nbd_mcast_index(int index)
2398 {
2399         struct sk_buff *skb;
2400         void *msg_head;
2401         int ret;
2402
2403         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2404         if (!skb)
2405                 return;
2406         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2407                                      NBD_CMD_LINK_DEAD);
2408         if (!msg_head) {
2409                 nlmsg_free(skb);
2410                 return;
2411         }
2412         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2413         if (ret) {
2414                 nlmsg_free(skb);
2415                 return;
2416         }
2417         genlmsg_end(skb, msg_head);
2418         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2419 }
2420
2421 static void nbd_dead_link_work(struct work_struct *work)
2422 {
2423         struct link_dead_args *args = container_of(work, struct link_dead_args,
2424                                                    work);
2425         nbd_mcast_index(args->index);
2426         kfree(args);
2427 }
2428
2429 static int __init nbd_init(void)
2430 {
2431         int i;
2432
2433         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2434
2435         if (max_part < 0) {
2436                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2437                 return -EINVAL;
2438         }
2439
2440         part_shift = 0;
2441         if (max_part > 0) {
2442                 part_shift = fls(max_part);
2443
2444                 /*
2445                  * Adjust max_part according to part_shift as it is exported
2446                  * to user space so that user can know the max number of
2447                  * partition kernel should be able to manage.
2448                  *
2449                  * Note that -1 is required because partition 0 is reserved
2450                  * for the whole disk.
2451                  */
2452                 max_part = (1UL << part_shift) - 1;
2453         }
2454
2455         if ((1UL << part_shift) > DISK_MAX_PARTS)
2456                 return -EINVAL;
2457
2458         if (nbds_max > 1UL << (MINORBITS - part_shift))
2459                 return -EINVAL;
2460
2461         if (register_blkdev(NBD_MAJOR, "nbd"))
2462                 return -EIO;
2463
2464         if (genl_register_family(&nbd_genl_family)) {
2465                 unregister_blkdev(NBD_MAJOR, "nbd");
2466                 return -EINVAL;
2467         }
2468         nbd_dbg_init();
2469
2470         mutex_lock(&nbd_index_mutex);
2471         for (i = 0; i < nbds_max; i++)
2472                 nbd_dev_add(i);
2473         mutex_unlock(&nbd_index_mutex);
2474         return 0;
2475 }
2476
2477 static int nbd_exit_cb(int id, void *ptr, void *data)
2478 {
2479         struct list_head *list = (struct list_head *)data;
2480         struct nbd_device *nbd = ptr;
2481
2482         list_add_tail(&nbd->list, list);
2483         return 0;
2484 }
2485
2486 static void __exit nbd_cleanup(void)
2487 {
2488         struct nbd_device *nbd;
2489         LIST_HEAD(del_list);
2490
2491         /*
2492          * Unregister netlink interface prior to waiting
2493          * for the completion of netlink commands.
2494          */
2495         genl_unregister_family(&nbd_genl_family);
2496
2497         nbd_dbg_close();
2498
2499         mutex_lock(&nbd_index_mutex);
2500         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2501         mutex_unlock(&nbd_index_mutex);
2502
2503         while (!list_empty(&del_list)) {
2504                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2505                 list_del_init(&nbd->list);
2506                 if (refcount_read(&nbd->config_refs))
2507                         printk(KERN_ERR "nbd: possibly leaking nbd_config (ref %d)\n",
2508                                         refcount_read(&nbd->config_refs));
2509                 if (refcount_read(&nbd->refs) != 1)
2510                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2511                 nbd_put(nbd);
2512         }
2513
2514         idr_destroy(&nbd_index_idr);
2515         unregister_blkdev(NBD_MAJOR, "nbd");
2516 }
2517
2518 module_init(nbd_init);
2519 module_exit(nbd_cleanup);
2520
2521 MODULE_DESCRIPTION("Network Block Device");
2522 MODULE_LICENSE("GPL");
2523
2524 module_param(nbds_max, int, 0444);
2525 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2526 module_param(max_part, int, 0444);
2527 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");