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 struct workqueue_struct *nbd_del_wq;
  53 static int nbd_total_devices = 0;
  54
  55 struct nbd_sock {
  56         struct socket *sock;
  57         struct mutex tx_lock;
  58         struct request *pending;
  59         int sent;
  60         bool dead;
  61         int fallback_index;
  62         int cookie;
  63 };
  64
  65 struct recv_thread_args {
  66         struct work_struct work;
  67         struct nbd_device *nbd;
  68         int index;
  69 };
  70
  71 struct link_dead_args {
  72         struct work_struct work;
  73         int index;
  74 };
  75
  76 #define NBD_RT_TIMEDOUT                 0
  77 #define NBD_RT_DISCONNECT_REQUESTED     1
  78 #define NBD_RT_DISCONNECTED             2
  79 #define NBD_RT_HAS_PID_FILE             3
  80 #define NBD_RT_HAS_CONFIG_REF           4
  81 #define NBD_RT_BOUND                    5
  82 #define NBD_RT_DISCONNECT_ON_CLOSE      6
  83 #define NBD_RT_HAS_BACKEND_FILE         7
  84
  85 #define NBD_DESTROY_ON_DISCONNECT       0
  86 #define NBD_DISCONNECT_REQUESTED        1
  87
  88 struct nbd_config {
  89         u32 flags;
  90         unsigned long runtime_flags;
  91         u64 dead_conn_timeout;
  92
  93         struct nbd_sock **socks;
  94         int num_connections;
  95         atomic_t live_connections;
  96         wait_queue_head_t conn_wait;
  97
  98         atomic_t recv_threads;
  99         wait_queue_head_t recv_wq;
 100         unsigned int blksize_bits;
 101         loff_t bytesize;
 102 #if IS_ENABLED(CONFIG_DEBUG_FS)
 103         struct dentry *dbg_dir;
 104 #endif
 105 };
 106
 107 static inline unsigned int nbd_blksize(struct nbd_config *config)
 108 {
 109         return 1u << config->blksize_bits;
 110 }
 111
 112 struct nbd_device {
 113         struct blk_mq_tag_set tag_set;
 114
 115         int index;
 116         refcount_t config_refs;
 117         refcount_t refs;
 118         struct nbd_config *config;
 119         struct mutex config_lock;
 120         struct gendisk *disk;
 121         struct workqueue_struct *recv_workq;
 122         struct work_struct remove_work;
 123
 124         struct list_head list;
 125         struct task_struct *task_recv;
 126         struct task_struct *task_setup;
 127
 128         unsigned long flags;
 129
 130         char *backend;
 131 };
 132
 133 #define NBD_CMD_REQUEUED        1
 134
 135 struct nbd_cmd {
 136         struct nbd_device *nbd;
 137         struct mutex lock;
 138         int index;
 139         int cookie;
 140         int retries;
 141         blk_status_t status;
 142         unsigned long flags;
 143         u32 cmd_cookie;
 144 };
 145
 146 #if IS_ENABLED(CONFIG_DEBUG_FS)
 147 static struct dentry *nbd_dbg_dir;
 148 #endif
 149
 150 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 151
 152 #define NBD_MAGIC 0x68797548
 153
 154 #define NBD_DEF_BLKSIZE_BITS 10
 155
 156 static unsigned int nbds_max = 16;
 157 static int max_part = 16;
 158 static int part_shift;
 159
 160 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 161 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 162 static void nbd_config_put(struct nbd_device *nbd);
 163 static void nbd_connect_reply(struct genl_info *info, int index);
 164 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 165 static void nbd_dead_link_work(struct work_struct *work);
 166 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 167
 168 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 169 {
 170         return disk_to_dev(nbd->disk);
 171 }
 172
 173 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 174 {
 175         struct request *req = blk_mq_rq_from_pdu(cmd);
 176
 177         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 178                 blk_mq_requeue_request(req, true);
 179 }
 180
 181 #define NBD_COOKIE_BITS 32
 182
 183 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 184 {
 185         struct request *req = blk_mq_rq_from_pdu(cmd);
 186         u32 tag = blk_mq_unique_tag(req);
 187         u64 cookie = cmd->cmd_cookie;
 188
 189         return (cookie << NBD_COOKIE_BITS) | tag;
 190 }
 191
 192 static u32 nbd_handle_to_tag(u64 handle)
 193 {
 194         return (u32)handle;
 195 }
 196
 197 static u32 nbd_handle_to_cookie(u64 handle)
 198 {
 199         return (u32)(handle >> NBD_COOKIE_BITS);
 200 }
 201
 202 static const char *nbdcmd_to_ascii(int cmd)
 203 {
 204         switch (cmd) {
 205         case  NBD_CMD_READ: return "read";
 206         case NBD_CMD_WRITE: return "write";
 207         case  NBD_CMD_DISC: return "disconnect";
 208         case NBD_CMD_FLUSH: return "flush";
 209         case  NBD_CMD_TRIM: return "trim/discard";
 210         }
 211         return "invalid";
 212 }
 213
 214 static ssize_t pid_show(struct device *dev,
 215                         struct device_attribute *attr, char *buf)
 216 {
 217         struct gendisk *disk = dev_to_disk(dev);
 218         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 219
 220         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 221 }
 222
 223 static const struct device_attribute pid_attr = {
 224         .attr = { .name = "pid", .mode = 0444},
 225         .show = pid_show,
 226 };
 227
 228 static ssize_t backend_show(struct device *dev,
 229                 struct device_attribute *attr, char *buf)
 230 {
 231         struct gendisk *disk = dev_to_disk(dev);
 232         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 233
 234         return sprintf(buf, "%s\n", nbd->backend ?: "");
 235 }
 236
 237 static const struct device_attribute backend_attr = {
 238         .attr = { .name = "backend", .mode = 0444},
 239         .show = backend_show,
 240 };
 241
 242 static void nbd_dev_remove(struct nbd_device *nbd)
 243 {
 244         struct gendisk *disk = nbd->disk;
 245
 246         del_gendisk(disk);
 247         blk_cleanup_disk(disk);
 248         blk_mq_free_tag_set(&nbd->tag_set);
 249
 250         /*
 251          * Remove from idr after del_gendisk() completes, so if the same ID is
 252          * reused, the following add_disk() will succeed.
 253          */
 254         mutex_lock(&nbd_index_mutex);
 255         idr_remove(&nbd_index_idr, nbd->index);
 256         mutex_unlock(&nbd_index_mutex);
 257
 258         kfree(nbd);
 259 }
 260
 261 static void nbd_dev_remove_work(struct work_struct *work)
 262 {
 263         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 264 }
 265
 266 static void nbd_put(struct nbd_device *nbd)
 267 {
 268         if (!refcount_dec_and_test(&nbd->refs))
 269                 return;
 270
 271         /* Call del_gendisk() asynchrounously to prevent deadlock */
 272         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 273                 queue_work(nbd_del_wq, &nbd->remove_work);
 274         else
 275                 nbd_dev_remove(nbd);
 276 }
 277
 278 static int nbd_disconnected(struct nbd_config *config)
 279 {
 280         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 281                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 282 }
 283
 284 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 285                                 int notify)
 286 {
 287         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 288                 struct link_dead_args *args;
 289                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 290                 if (args) {
 291                         INIT_WORK(&args->work, nbd_dead_link_work);
 292                         args->index = nbd->index;
 293                         queue_work(system_wq, &args->work);
 294                 }
 295         }
 296         if (!nsock->dead) {
 297                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 298                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 299                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 300                                                &nbd->config->runtime_flags)) {
 301                                 set_bit(NBD_RT_DISCONNECTED,
 302                                         &nbd->config->runtime_flags);
 303                                 dev_info(nbd_to_dev(nbd),
 304                                         "Disconnected due to user request.\n");
 305                         }
 306                 }
 307         }
 308         nsock->dead = true;
 309         nsock->pending = NULL;
 310         nsock->sent = 0;
 311 }
 312
 313 static void nbd_size_clear(struct nbd_device *nbd)
 314 {
 315         if (nbd->config->bytesize) {
 316                 set_capacity(nbd->disk, 0);
 317                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 318         }
 319 }
 320
 321 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 322                 loff_t blksize)
 323 {
 324         if (!blksize)
 325                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 326         if (blksize < 512 || blksize > PAGE_SIZE || !is_power_of_2(blksize))
 327                 return -EINVAL;
 328
 329         nbd->config->bytesize = bytesize;
 330         nbd->config->blksize_bits = __ffs(blksize);
 331
 332         if (!nbd->task_recv)
 333                 return 0;
 334
 335         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 336                 nbd->disk->queue->limits.discard_granularity = blksize;
 337                 nbd->disk->queue->limits.discard_alignment = blksize;
 338                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 339         }
 340         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 341         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 342
 343         if (max_part)
 344                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 345         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 346                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 347         return 0;
 348 }
 349
 350 static void nbd_complete_rq(struct request *req)
 351 {
 352         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 353
 354         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 355                 cmd->status ? "failed" : "done");
 356
 357         blk_mq_end_request(req, cmd->status);
 358 }
 359
 360 /*
 361  * Forcibly shutdown the socket causing all listeners to error
 362  */
 363 static void sock_shutdown(struct nbd_device *nbd)
 364 {
 365         struct nbd_config *config = nbd->config;
 366         int i;
 367
 368         if (config->num_connections == 0)
 369                 return;
 370         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 371                 return;
 372
 373         for (i = 0; i < config->num_connections; i++) {
 374                 struct nbd_sock *nsock = config->socks[i];
 375                 mutex_lock(&nsock->tx_lock);
 376                 nbd_mark_nsock_dead(nbd, nsock, 0);
 377                 mutex_unlock(&nsock->tx_lock);
 378         }
 379         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 380 }
 381
 382 static u32 req_to_nbd_cmd_type(struct request *req)
 383 {
 384         switch (req_op(req)) {
 385         case REQ_OP_DISCARD:
 386                 return NBD_CMD_TRIM;
 387         case REQ_OP_FLUSH:
 388                 return NBD_CMD_FLUSH;
 389         case REQ_OP_WRITE:
 390                 return NBD_CMD_WRITE;
 391         case REQ_OP_READ:
 392                 return NBD_CMD_READ;
 393         default:
 394                 return U32_MAX;
 395         }
 396 }
 397
 398 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 399                                                  bool reserved)
 400 {
 401         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 402         struct nbd_device *nbd = cmd->nbd;
 403         struct nbd_config *config;
 404
 405         if (!mutex_trylock(&cmd->lock))
 406                 return BLK_EH_RESET_TIMER;
 407
 408         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 409                 cmd->status = BLK_STS_TIMEOUT;
 410                 mutex_unlock(&cmd->lock);
 411                 goto done;
 412         }
 413         config = nbd->config;
 414
 415         if (config->num_connections > 1 ||
 416             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 417                 dev_err_ratelimited(nbd_to_dev(nbd),
 418                                     "Connection timed out, retrying (%d/%d alive)\n",
 419                                     atomic_read(&config->live_connections),
 420                                     config->num_connections);
 421                 /*
 422                  * Hooray we have more connections, requeue this IO, the submit
 423                  * path will put it on a real connection. Or if only one
 424                  * connection is configured, the submit path will wait util
 425                  * a new connection is reconfigured or util dead timeout.
 426                  */
 427                 if (config->socks) {
 428                         if (cmd->index < config->num_connections) {
 429                                 struct nbd_sock *nsock =
 430                                         config->socks[cmd->index];
 431                                 mutex_lock(&nsock->tx_lock);
 432                                 /* We can have multiple outstanding requests, so
 433                                  * we don't want to mark the nsock dead if we've
 434                                  * already reconnected with a new socket, so
 435                                  * only mark it dead if its the same socket we
 436                                  * were sent out on.
 437                                  */
 438                                 if (cmd->cookie == nsock->cookie)
 439                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 440                                 mutex_unlock(&nsock->tx_lock);
 441                         }
 442                         mutex_unlock(&cmd->lock);
 443                         nbd_requeue_cmd(cmd);
 444                         nbd_config_put(nbd);
 445                         return BLK_EH_DONE;
 446                 }
 447         }
 448
 449         if (!nbd->tag_set.timeout) {
 450                 /*
 451                  * Userspace sets timeout=0 to disable socket disconnection,
 452                  * so just warn and reset the timer.
 453                  */
 454                 struct nbd_sock *nsock = config->socks[cmd->index];
 455                 cmd->retries++;
 456                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 457                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 458                         (unsigned long long)blk_rq_pos(req) << 9,
 459                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 460
 461                 mutex_lock(&nsock->tx_lock);
 462                 if (cmd->cookie != nsock->cookie) {
 463                         nbd_requeue_cmd(cmd);
 464                         mutex_unlock(&nsock->tx_lock);
 465                         mutex_unlock(&cmd->lock);
 466                         nbd_config_put(nbd);
 467                         return BLK_EH_DONE;
 468                 }
 469                 mutex_unlock(&nsock->tx_lock);
 470                 mutex_unlock(&cmd->lock);
 471                 nbd_config_put(nbd);
 472                 return BLK_EH_RESET_TIMER;
 473         }
 474
 475         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 476         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 477         cmd->status = BLK_STS_IOERR;
 478         mutex_unlock(&cmd->lock);
 479         sock_shutdown(nbd);
 480         nbd_config_put(nbd);
 481 done:
 482         blk_mq_complete_request(req);
 483         return BLK_EH_DONE;
 484 }
 485
 486 /*
 487  *  Send or receive packet.
 488  */
 489 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 490                      struct iov_iter *iter, int msg_flags, int *sent)
 491 {
 492         struct nbd_config *config = nbd->config;
 493         struct socket *sock = config->socks[index]->sock;
 494         int result;
 495         struct msghdr msg;
 496         unsigned int noreclaim_flag;
 497
 498         if (unlikely(!sock)) {
 499                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 500                         "Attempted %s on closed socket in sock_xmit\n",
 501                         (send ? "send" : "recv"));
 502                 return -EINVAL;
 503         }
 504
 505         msg.msg_iter = *iter;
 506
 507         noreclaim_flag = memalloc_noreclaim_save();
 508         do {
 509                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 510                 msg.msg_name = NULL;
 511                 msg.msg_namelen = 0;
 512                 msg.msg_control = NULL;
 513                 msg.msg_controllen = 0;
 514                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 515
 516                 if (send)
 517                         result = sock_sendmsg(sock, &msg);
 518                 else
 519                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 520
 521                 if (result <= 0) {
 522                         if (result == 0)
 523                                 result = -EPIPE; /* short read */
 524                         break;
 525                 }
 526                 if (sent)
 527                         *sent += result;
 528         } while (msg_data_left(&msg));
 529
 530         memalloc_noreclaim_restore(noreclaim_flag);
 531
 532         return result;
 533 }
 534
 535 /*
 536  * Different settings for sk->sk_sndtimeo can result in different return values
 537  * if there is a signal pending when we enter sendmsg, because reasons?
 538  */
 539 static inline int was_interrupted(int result)
 540 {
 541         return result == -ERESTARTSYS || result == -EINTR;
 542 }
 543
 544 /* always call with the tx_lock held */
 545 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 546 {
 547         struct request *req = blk_mq_rq_from_pdu(cmd);
 548         struct nbd_config *config = nbd->config;
 549         struct nbd_sock *nsock = config->socks[index];
 550         int result;
 551         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 552         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 553         struct iov_iter from;
 554         unsigned long size = blk_rq_bytes(req);
 555         struct bio *bio;
 556         u64 handle;
 557         u32 type;
 558         u32 nbd_cmd_flags = 0;
 559         int sent = nsock->sent, skip = 0;
 560
 561         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 562
 563         type = req_to_nbd_cmd_type(req);
 564         if (type == U32_MAX)
 565                 return -EIO;
 566
 567         if (rq_data_dir(req) == WRITE &&
 568             (config->flags & NBD_FLAG_READ_ONLY)) {
 569                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 570                                     "Write on read-only\n");
 571                 return -EIO;
 572         }
 573
 574         if (req->cmd_flags & REQ_FUA)
 575                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 576
 577         /* We did a partial send previously, and we at least sent the whole
 578          * request struct, so just go and send the rest of the pages in the
 579          * request.
 580          */
 581         if (sent) {
 582                 if (sent >= sizeof(request)) {
 583                         skip = sent - sizeof(request);
 584
 585                         /* initialize handle for tracing purposes */
 586                         handle = nbd_cmd_handle(cmd);
 587
 588                         goto send_pages;
 589                 }
 590                 iov_iter_advance(&from, sent);
 591         } else {
 592                 cmd->cmd_cookie++;
 593         }
 594         cmd->index = index;
 595         cmd->cookie = nsock->cookie;
 596         cmd->retries = 0;
 597         request.type = htonl(type | nbd_cmd_flags);
 598         if (type != NBD_CMD_FLUSH) {
 599                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 600                 request.len = htonl(size);
 601         }
 602         handle = nbd_cmd_handle(cmd);
 603         memcpy(request.handle, &handle, sizeof(handle));
 604
 605         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 606
 607         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 608                 req, nbdcmd_to_ascii(type),
 609                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 610         result = sock_xmit(nbd, index, 1, &from,
 611                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 612         trace_nbd_header_sent(req, handle);
 613         if (result <= 0) {
 614                 if (was_interrupted(result)) {
 615                         /* If we havne't sent anything we can just return BUSY,
 616                          * however if we have sent something we need to make
 617                          * sure we only allow this req to be sent until we are
 618                          * completely done.
 619                          */
 620                         if (sent) {
 621                                 nsock->pending = req;
 622                                 nsock->sent = sent;
 623                         }
 624                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 625                         return BLK_STS_RESOURCE;
 626                 }
 627                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 628                         "Send control failed (result %d)\n", result);
 629                 return -EAGAIN;
 630         }
 631 send_pages:
 632         if (type != NBD_CMD_WRITE)
 633                 goto out;
 634
 635         bio = req->bio;
 636         while (bio) {
 637                 struct bio *next = bio->bi_next;
 638                 struct bvec_iter iter;
 639                 struct bio_vec bvec;
 640
 641                 bio_for_each_segment(bvec, bio, iter) {
 642                         bool is_last = !next && bio_iter_last(bvec, iter);
 643                         int flags = is_last ? 0 : MSG_MORE;
 644
 645                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 646                                 req, bvec.bv_len);
 647                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 648                         if (skip) {
 649                                 if (skip >= iov_iter_count(&from)) {
 650                                         skip -= iov_iter_count(&from);
 651                                         continue;
 652                                 }
 653                                 iov_iter_advance(&from, skip);
 654                                 skip = 0;
 655                         }
 656                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 657                         if (result <= 0) {
 658                                 if (was_interrupted(result)) {
 659                                         /* We've already sent the header, we
 660                                          * have no choice but to set pending and
 661                                          * return BUSY.
 662                                          */
 663                                         nsock->pending = req;
 664                                         nsock->sent = sent;
 665                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 666                                         return BLK_STS_RESOURCE;
 667                                 }
 668                                 dev_err(disk_to_dev(nbd->disk),
 669                                         "Send data failed (result %d)\n",
 670                                         result);
 671                                 return -EAGAIN;
 672                         }
 673                         /*
 674                          * The completion might already have come in,
 675                          * so break for the last one instead of letting
 676                          * the iterator do it. This prevents use-after-free
 677                          * of the bio.
 678                          */
 679                         if (is_last)
 680                                 break;
 681                 }
 682                 bio = next;
 683         }
 684 out:
 685         trace_nbd_payload_sent(req, handle);
 686         nsock->pending = NULL;
 687         nsock->sent = 0;
 688         return 0;
 689 }
 690
 691 /* NULL returned = something went wrong, inform userspace */
 692 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 693 {
 694         struct nbd_config *config = nbd->config;
 695         int result;
 696         struct nbd_reply reply;
 697         struct nbd_cmd *cmd;
 698         struct request *req = NULL;
 699         u64 handle;
 700         u16 hwq;
 701         u32 tag;
 702         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 703         struct iov_iter to;
 704         int ret = 0;
 705
 706         reply.magic = 0;
 707         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 708         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 709         if (result <= 0) {
 710                 if (!nbd_disconnected(config))
 711                         dev_err(disk_to_dev(nbd->disk),
 712                                 "Receive control failed (result %d)\n", result);
 713                 return ERR_PTR(result);
 714         }
 715
 716         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 717                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 718                                 (unsigned long)ntohl(reply.magic));
 719                 return ERR_PTR(-EPROTO);
 720         }
 721
 722         memcpy(&handle, reply.handle, sizeof(handle));
 723         tag = nbd_handle_to_tag(handle);
 724         hwq = blk_mq_unique_tag_to_hwq(tag);
 725         if (hwq < nbd->tag_set.nr_hw_queues)
 726                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 727                                        blk_mq_unique_tag_to_tag(tag));
 728         if (!req || !blk_mq_request_started(req)) {
 729                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 730                         tag, req);
 731                 return ERR_PTR(-ENOENT);
 732         }
 733         trace_nbd_header_received(req, handle);
 734         cmd = blk_mq_rq_to_pdu(req);
 735
 736         mutex_lock(&cmd->lock);
 737         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 738                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 739                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 740                 ret = -ENOENT;
 741                 goto out;
 742         }
 743         if (cmd->status != BLK_STS_OK) {
 744                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 745                         req);
 746                 ret = -ENOENT;
 747                 goto out;
 748         }
 749         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 750                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 751                         req);
 752                 ret = -ENOENT;
 753                 goto out;
 754         }
 755         if (ntohl(reply.error)) {
 756                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 757                         ntohl(reply.error));
 758                 cmd->status = BLK_STS_IOERR;
 759                 goto out;
 760         }
 761
 762         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 763         if (rq_data_dir(req) != WRITE) {
 764                 struct req_iterator iter;
 765                 struct bio_vec bvec;
 766
 767                 rq_for_each_segment(bvec, req, iter) {
 768                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 769                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 770                         if (result <= 0) {
 771                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 772                                         result);
 773                                 /*
 774                                  * If we've disconnected, we need to make sure we
 775                                  * complete this request, otherwise error out
 776                                  * and let the timeout stuff handle resubmitting
 777                                  * this request onto another connection.
 778                                  */
 779                                 if (nbd_disconnected(config)) {
 780                                         cmd->status = BLK_STS_IOERR;
 781                                         goto out;
 782                                 }
 783                                 ret = -EIO;
 784                                 goto out;
 785                         }
 786                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 787                                 req, bvec.bv_len);
 788                 }
 789         }
 790 out:
 791         trace_nbd_payload_received(req, handle);
 792         mutex_unlock(&cmd->lock);
 793         return ret ? ERR_PTR(ret) : cmd;
 794 }
 795
 796 static void recv_work(struct work_struct *work)
 797 {
 798         struct recv_thread_args *args = container_of(work,
 799                                                      struct recv_thread_args,
 800                                                      work);
 801         struct nbd_device *nbd = args->nbd;
 802         struct nbd_config *config = nbd->config;
 803         struct nbd_cmd *cmd;
 804         struct request *rq;
 805
 806         while (1) {
 807                 cmd = nbd_read_stat(nbd, args->index);
 808                 if (IS_ERR(cmd)) {
 809                         struct nbd_sock *nsock = config->socks[args->index];
 810
 811                         mutex_lock(&nsock->tx_lock);
 812                         nbd_mark_nsock_dead(nbd, nsock, 1);
 813                         mutex_unlock(&nsock->tx_lock);
 814                         break;
 815                 }
 816
 817                 rq = blk_mq_rq_from_pdu(cmd);
 818                 if (likely(!blk_should_fake_timeout(rq->q)))
 819                         blk_mq_complete_request(rq);
 820         }
 821         nbd_config_put(nbd);
 822         atomic_dec(&config->recv_threads);
 823         wake_up(&config->recv_wq);
 824         kfree(args);
 825 }
 826
 827 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 828 {
 829         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 830
 831         /* don't abort one completed request */
 832         if (blk_mq_request_completed(req))
 833                 return true;
 834
 835         mutex_lock(&cmd->lock);
 836         cmd->status = BLK_STS_IOERR;
 837         mutex_unlock(&cmd->lock);
 838
 839         blk_mq_complete_request(req);
 840         return true;
 841 }
 842
 843 static void nbd_clear_que(struct nbd_device *nbd)
 844 {
 845         blk_mq_quiesce_queue(nbd->disk->queue);
 846         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 847         blk_mq_unquiesce_queue(nbd->disk->queue);
 848         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 849 }
 850
 851 static int find_fallback(struct nbd_device *nbd, int index)
 852 {
 853         struct nbd_config *config = nbd->config;
 854         int new_index = -1;
 855         struct nbd_sock *nsock = config->socks[index];
 856         int fallback = nsock->fallback_index;
 857
 858         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 859                 return new_index;
 860
 861         if (config->num_connections <= 1) {
 862                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 863                                     "Dead connection, failed to find a fallback\n");
 864                 return new_index;
 865         }
 866
 867         if (fallback >= 0 && fallback < config->num_connections &&
 868             !config->socks[fallback]->dead)
 869                 return fallback;
 870
 871         if (nsock->fallback_index < 0 ||
 872             nsock->fallback_index >= config->num_connections ||
 873             config->socks[nsock->fallback_index]->dead) {
 874                 int i;
 875                 for (i = 0; i < config->num_connections; i++) {
 876                         if (i == index)
 877                                 continue;
 878                         if (!config->socks[i]->dead) {
 879                                 new_index = i;
 880                                 break;
 881                         }
 882                 }
 883                 nsock->fallback_index = new_index;
 884                 if (new_index < 0) {
 885                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 886                                             "Dead connection, failed to find a fallback\n");
 887                         return new_index;
 888                 }
 889         }
 890         new_index = nsock->fallback_index;
 891         return new_index;
 892 }
 893
 894 static int wait_for_reconnect(struct nbd_device *nbd)
 895 {
 896         struct nbd_config *config = nbd->config;
 897         if (!config->dead_conn_timeout)
 898                 return 0;
 899         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 900                 return 0;
 901         return wait_event_timeout(config->conn_wait,
 902                                   atomic_read(&config->live_connections) > 0,
 903                                   config->dead_conn_timeout) > 0;
 904 }
 905
 906 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 907 {
 908         struct request *req = blk_mq_rq_from_pdu(cmd);
 909         struct nbd_device *nbd = cmd->nbd;
 910         struct nbd_config *config;
 911         struct nbd_sock *nsock;
 912         int ret;
 913
 914         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 915                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 916                                     "Socks array is empty\n");
 917                 blk_mq_start_request(req);
 918                 return -EINVAL;
 919         }
 920         config = nbd->config;
 921
 922         if (index >= config->num_connections) {
 923                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 924                                     "Attempted send on invalid socket\n");
 925                 nbd_config_put(nbd);
 926                 blk_mq_start_request(req);
 927                 return -EINVAL;
 928         }
 929         cmd->status = BLK_STS_OK;
 930 again:
 931         nsock = config->socks[index];
 932         mutex_lock(&nsock->tx_lock);
 933         if (nsock->dead) {
 934                 int old_index = index;
 935                 index = find_fallback(nbd, index);
 936                 mutex_unlock(&nsock->tx_lock);
 937                 if (index < 0) {
 938                         if (wait_for_reconnect(nbd)) {
 939                                 index = old_index;
 940                                 goto again;
 941                         }
 942                         /* All the sockets should already be down at this point,
 943                          * we just want to make sure that DISCONNECTED is set so
 944                          * any requests that come in that were queue'ed waiting
 945                          * for the reconnect timer don't trigger the timer again
 946                          * and instead just error out.
 947                          */
 948                         sock_shutdown(nbd);
 949                         nbd_config_put(nbd);
 950                         blk_mq_start_request(req);
 951                         return -EIO;
 952                 }
 953                 goto again;
 954         }
 955
 956         /* Handle the case that we have a pending request that was partially
 957          * transmitted that _has_ to be serviced first.  We need to call requeue
 958          * here so that it gets put _after_ the request that is already on the
 959          * dispatch list.
 960          */
 961         blk_mq_start_request(req);
 962         if (unlikely(nsock->pending && nsock->pending != req)) {
 963                 nbd_requeue_cmd(cmd);
 964                 ret = 0;
 965                 goto out;
 966         }
 967         /*
 968          * Some failures are related to the link going down, so anything that
 969          * returns EAGAIN can be retried on a different socket.
 970          */
 971         ret = nbd_send_cmd(nbd, cmd, index);
 972         if (ret == -EAGAIN) {
 973                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 974                                     "Request send failed, requeueing\n");
 975                 nbd_mark_nsock_dead(nbd, nsock, 1);
 976                 nbd_requeue_cmd(cmd);
 977                 ret = 0;
 978         }
 979 out:
 980         mutex_unlock(&nsock->tx_lock);
 981         nbd_config_put(nbd);
 982         return ret;
 983 }
 984
 985 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 986                         const struct blk_mq_queue_data *bd)
 987 {
 988         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 989         int ret;
 990
 991         /*
 992          * Since we look at the bio's to send the request over the network we
 993          * need to make sure the completion work doesn't mark this request done
 994          * before we are done doing our send.  This keeps us from dereferencing
 995          * freed data if we have particularly fast completions (ie we get the
 996          * completion before we exit sock_xmit on the last bvec) or in the case
 997          * that the server is misbehaving (or there was an error) before we're
 998          * done sending everything over the wire.
 999          */
1000         mutex_lock(&cmd->lock);
1001         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1002
1003         /* We can be called directly from the user space process, which means we
1004          * could possibly have signals pending so our sendmsg will fail.  In
1005          * this case we need to return that we are busy, otherwise error out as
1006          * appropriate.
1007          */
1008         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1009         if (ret < 0)
1010                 ret = BLK_STS_IOERR;
1011         else if (!ret)
1012                 ret = BLK_STS_OK;
1013         mutex_unlock(&cmd->lock);
1014
1015         return ret;
1016 }
1017
1018 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1019                                      int *err)
1020 {
1021         struct socket *sock;
1022
1023         *err = 0;
1024         sock = sockfd_lookup(fd, err);
1025         if (!sock)
1026                 return NULL;
1027
1028         if (sock->ops->shutdown == sock_no_shutdown) {
1029                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1030                 *err = -EINVAL;
1031                 sockfd_put(sock);
1032                 return NULL;
1033         }
1034
1035         return sock;
1036 }
1037
1038 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1039                           bool netlink)
1040 {
1041         struct nbd_config *config = nbd->config;
1042         struct socket *sock;
1043         struct nbd_sock **socks;
1044         struct nbd_sock *nsock;
1045         int err;
1046
1047         sock = nbd_get_socket(nbd, arg, &err);
1048         if (!sock)
1049                 return err;
1050
1051         /*
1052          * We need to make sure we don't get any errant requests while we're
1053          * reallocating the ->socks array.
1054          */
1055         blk_mq_freeze_queue(nbd->disk->queue);
1056
1057         if (!netlink && !nbd->task_setup &&
1058             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1059                 nbd->task_setup = current;
1060
1061         if (!netlink &&
1062             (nbd->task_setup != current ||
1063              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1064                 dev_err(disk_to_dev(nbd->disk),
1065                         "Device being setup by another task");
1066                 err = -EBUSY;
1067                 goto put_socket;
1068         }
1069
1070         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1071         if (!nsock) {
1072                 err = -ENOMEM;
1073                 goto put_socket;
1074         }
1075
1076         socks = krealloc(config->socks, (config->num_connections + 1) *
1077                          sizeof(struct nbd_sock *), GFP_KERNEL);
1078         if (!socks) {
1079                 kfree(nsock);
1080                 err = -ENOMEM;
1081                 goto put_socket;
1082         }
1083
1084         config->socks = socks;
1085
1086         nsock->fallback_index = -1;
1087         nsock->dead = false;
1088         mutex_init(&nsock->tx_lock);
1089         nsock->sock = sock;
1090         nsock->pending = NULL;
1091         nsock->sent = 0;
1092         nsock->cookie = 0;
1093         socks[config->num_connections++] = nsock;
1094         atomic_inc(&config->live_connections);
1095         blk_mq_unfreeze_queue(nbd->disk->queue);
1096
1097         return 0;
1098
1099 put_socket:
1100         blk_mq_unfreeze_queue(nbd->disk->queue);
1101         sockfd_put(sock);
1102         return err;
1103 }
1104
1105 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1106 {
1107         struct nbd_config *config = nbd->config;
1108         struct socket *sock, *old;
1109         struct recv_thread_args *args;
1110         int i;
1111         int err;
1112
1113         sock = nbd_get_socket(nbd, arg, &err);
1114         if (!sock)
1115                 return err;
1116
1117         args = kzalloc(sizeof(*args), GFP_KERNEL);
1118         if (!args) {
1119                 sockfd_put(sock);
1120                 return -ENOMEM;
1121         }
1122
1123         for (i = 0; i < config->num_connections; i++) {
1124                 struct nbd_sock *nsock = config->socks[i];
1125
1126                 if (!nsock->dead)
1127                         continue;
1128
1129                 mutex_lock(&nsock->tx_lock);
1130                 if (!nsock->dead) {
1131                         mutex_unlock(&nsock->tx_lock);
1132                         continue;
1133                 }
1134                 sk_set_memalloc(sock->sk);
1135                 if (nbd->tag_set.timeout)
1136                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1137                 atomic_inc(&config->recv_threads);
1138                 refcount_inc(&nbd->config_refs);
1139                 old = nsock->sock;
1140                 nsock->fallback_index = -1;
1141                 nsock->sock = sock;
1142                 nsock->dead = false;
1143                 INIT_WORK(&args->work, recv_work);
1144                 args->index = i;
1145                 args->nbd = nbd;
1146                 nsock->cookie++;
1147                 mutex_unlock(&nsock->tx_lock);
1148                 sockfd_put(old);
1149
1150                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1151
1152                 /* We take the tx_mutex in an error path in the recv_work, so we
1153                  * need to queue_work outside of the tx_mutex.
1154                  */
1155                 queue_work(nbd->recv_workq, &args->work);
1156
1157                 atomic_inc(&config->live_connections);
1158                 wake_up(&config->conn_wait);
1159                 return 0;
1160         }
1161         sockfd_put(sock);
1162         kfree(args);
1163         return -ENOSPC;
1164 }
1165
1166 static void nbd_bdev_reset(struct block_device *bdev)
1167 {
1168         if (bdev->bd_openers > 1)
1169                 return;
1170         set_capacity(bdev->bd_disk, 0);
1171 }
1172
1173 static void nbd_parse_flags(struct nbd_device *nbd)
1174 {
1175         struct nbd_config *config = nbd->config;
1176         if (config->flags & NBD_FLAG_READ_ONLY)
1177                 set_disk_ro(nbd->disk, true);
1178         else
1179                 set_disk_ro(nbd->disk, false);
1180         if (config->flags & NBD_FLAG_SEND_TRIM)
1181                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1182         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1183                 if (config->flags & NBD_FLAG_SEND_FUA)
1184                         blk_queue_write_cache(nbd->disk->queue, true, true);
1185                 else
1186                         blk_queue_write_cache(nbd->disk->queue, true, false);
1187         }
1188         else
1189                 blk_queue_write_cache(nbd->disk->queue, false, false);
1190 }
1191
1192 static void send_disconnects(struct nbd_device *nbd)
1193 {
1194         struct nbd_config *config = nbd->config;
1195         struct nbd_request request = {
1196                 .magic = htonl(NBD_REQUEST_MAGIC),
1197                 .type = htonl(NBD_CMD_DISC),
1198         };
1199         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1200         struct iov_iter from;
1201         int i, ret;
1202
1203         for (i = 0; i < config->num_connections; i++) {
1204                 struct nbd_sock *nsock = config->socks[i];
1205
1206                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1207                 mutex_lock(&nsock->tx_lock);
1208                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1209                 if (ret <= 0)
1210                         dev_err(disk_to_dev(nbd->disk),
1211                                 "Send disconnect failed %d\n", ret);
1212                 mutex_unlock(&nsock->tx_lock);
1213         }
1214 }
1215
1216 static int nbd_disconnect(struct nbd_device *nbd)
1217 {
1218         struct nbd_config *config = nbd->config;
1219
1220         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1221         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1222         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1223         send_disconnects(nbd);
1224         return 0;
1225 }
1226
1227 static void nbd_clear_sock(struct nbd_device *nbd)
1228 {
1229         sock_shutdown(nbd);
1230         nbd_clear_que(nbd);
1231         nbd->task_setup = NULL;
1232 }
1233
1234 static void nbd_config_put(struct nbd_device *nbd)
1235 {
1236         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1237                                         &nbd->config_lock)) {
1238                 struct nbd_config *config = nbd->config;
1239                 nbd_dev_dbg_close(nbd);
1240                 nbd_size_clear(nbd);
1241                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1242                                        &config->runtime_flags))
1243                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1244                 nbd->task_recv = NULL;
1245                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1246                                        &config->runtime_flags)) {
1247                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1248                         kfree(nbd->backend);
1249                         nbd->backend = NULL;
1250                 }
1251                 nbd_clear_sock(nbd);
1252                 if (config->num_connections) {
1253                         int i;
1254                         for (i = 0; i < config->num_connections; i++) {
1255                                 sockfd_put(config->socks[i]->sock);
1256                                 kfree(config->socks[i]);
1257                         }
1258                         kfree(config->socks);
1259                 }
1260                 kfree(nbd->config);
1261                 nbd->config = NULL;
1262
1263                 if (nbd->recv_workq)
1264                         destroy_workqueue(nbd->recv_workq);
1265                 nbd->recv_workq = NULL;
1266
1267                 nbd->tag_set.timeout = 0;
1268                 nbd->disk->queue->limits.discard_granularity = 0;
1269                 nbd->disk->queue->limits.discard_alignment = 0;
1270                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1271                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1272
1273                 mutex_unlock(&nbd->config_lock);
1274                 nbd_put(nbd);
1275                 module_put(THIS_MODULE);
1276         }
1277 }
1278
1279 static int nbd_start_device(struct nbd_device *nbd)
1280 {
1281         struct nbd_config *config = nbd->config;
1282         int num_connections = config->num_connections;
1283         int error = 0, i;
1284
1285         if (nbd->task_recv)
1286                 return -EBUSY;
1287         if (!config->socks)
1288                 return -EINVAL;
1289         if (num_connections > 1 &&
1290             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1291                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1292                 return -EINVAL;
1293         }
1294
1295         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1296                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1297                                           WQ_UNBOUND, 0, nbd->index);
1298         if (!nbd->recv_workq) {
1299                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1300                 return -ENOMEM;
1301         }
1302
1303         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1304         nbd->task_recv = current;
1305
1306         nbd_parse_flags(nbd);
1307
1308         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1309         if (error) {
1310                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1311                 return error;
1312         }
1313         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1314
1315         nbd_dev_dbg_init(nbd);
1316         for (i = 0; i < num_connections; i++) {
1317                 struct recv_thread_args *args;
1318
1319                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1320                 if (!args) {
1321                         sock_shutdown(nbd);
1322                         /*
1323                          * If num_connections is m (2 < m),
1324                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1325                          * But NO.(n + 1) failed. We still have n recv threads.
1326                          * So, add flush_workqueue here to prevent recv threads
1327                          * dropping the last config_refs and trying to destroy
1328                          * the workqueue from inside the workqueue.
1329                          */
1330                         if (i)
1331                                 flush_workqueue(nbd->recv_workq);
1332                         return -ENOMEM;
1333                 }
1334                 sk_set_memalloc(config->socks[i]->sock->sk);
1335                 if (nbd->tag_set.timeout)
1336                         config->socks[i]->sock->sk->sk_sndtimeo =
1337                                 nbd->tag_set.timeout;
1338                 atomic_inc(&config->recv_threads);
1339                 refcount_inc(&nbd->config_refs);
1340                 INIT_WORK(&args->work, recv_work);
1341                 args->nbd = nbd;
1342                 args->index = i;
1343                 queue_work(nbd->recv_workq, &args->work);
1344         }
1345         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1346 }
1347
1348 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1349 {
1350         struct nbd_config *config = nbd->config;
1351         int ret;
1352
1353         ret = nbd_start_device(nbd);
1354         if (ret)
1355                 return ret;
1356
1357         if (max_part)
1358                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1359         mutex_unlock(&nbd->config_lock);
1360         ret = wait_event_interruptible(config->recv_wq,
1361                                          atomic_read(&config->recv_threads) == 0);
1362         if (ret)
1363                 sock_shutdown(nbd);
1364         flush_workqueue(nbd->recv_workq);
1365
1366         mutex_lock(&nbd->config_lock);
1367         nbd_bdev_reset(bdev);
1368         /* user requested, ignore socket errors */
1369         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1370                 ret = 0;
1371         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1372                 ret = -ETIMEDOUT;
1373         return ret;
1374 }
1375
1376 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1377                                  struct block_device *bdev)
1378 {
1379         sock_shutdown(nbd);
1380         __invalidate_device(bdev, true);
1381         nbd_bdev_reset(bdev);
1382         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1383                                &nbd->config->runtime_flags))
1384                 nbd_config_put(nbd);
1385 }
1386
1387 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1388 {
1389         nbd->tag_set.timeout = timeout * HZ;
1390         if (timeout)
1391                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1392         else
1393                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1394 }
1395
1396 /* Must be called with config_lock held */
1397 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1398                        unsigned int cmd, unsigned long arg)
1399 {
1400         struct nbd_config *config = nbd->config;
1401         loff_t bytesize;
1402
1403         switch (cmd) {
1404         case NBD_DISCONNECT:
1405                 return nbd_disconnect(nbd);
1406         case NBD_CLEAR_SOCK:
1407                 nbd_clear_sock_ioctl(nbd, bdev);
1408                 return 0;
1409         case NBD_SET_SOCK:
1410                 return nbd_add_socket(nbd, arg, false);
1411         case NBD_SET_BLKSIZE:
1412                 return nbd_set_size(nbd, config->bytesize, arg);
1413         case NBD_SET_SIZE:
1414                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1415         case NBD_SET_SIZE_BLOCKS:
1416                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1417                         return -EINVAL;
1418                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1419         case NBD_SET_TIMEOUT:
1420                 nbd_set_cmd_timeout(nbd, arg);
1421                 return 0;
1422
1423         case NBD_SET_FLAGS:
1424                 config->flags = arg;
1425                 return 0;
1426         case NBD_DO_IT:
1427                 return nbd_start_device_ioctl(nbd, bdev);
1428         case NBD_CLEAR_QUE:
1429                 /*
1430                  * This is for compatibility only.  The queue is always cleared
1431                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1432                  */
1433                 return 0;
1434         case NBD_PRINT_DEBUG:
1435                 /*
1436                  * For compatibility only, we no longer keep a list of
1437                  * outstanding requests.
1438                  */
1439                 return 0;
1440         }
1441         return -ENOTTY;
1442 }
1443
1444 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1445                      unsigned int cmd, unsigned long arg)
1446 {
1447         struct nbd_device *nbd = bdev->bd_disk->private_data;
1448         struct nbd_config *config = nbd->config;
1449         int error = -EINVAL;
1450
1451         if (!capable(CAP_SYS_ADMIN))
1452                 return -EPERM;
1453
1454         /* The block layer will pass back some non-nbd ioctls in case we have
1455          * special handling for them, but we don't so just return an error.
1456          */
1457         if (_IOC_TYPE(cmd) != 0xab)
1458                 return -EINVAL;
1459
1460         mutex_lock(&nbd->config_lock);
1461
1462         /* Don't allow ioctl operations on a nbd device that was created with
1463          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1464          */
1465         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1466             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1467                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1468         else
1469                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1470         mutex_unlock(&nbd->config_lock);
1471         return error;
1472 }
1473
1474 static struct nbd_config *nbd_alloc_config(void)
1475 {
1476         struct nbd_config *config;
1477
1478         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1479         if (!config)
1480                 return NULL;
1481         atomic_set(&config->recv_threads, 0);
1482         init_waitqueue_head(&config->recv_wq);
1483         init_waitqueue_head(&config->conn_wait);
1484         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1485         atomic_set(&config->live_connections, 0);
1486         try_module_get(THIS_MODULE);
1487         return config;
1488 }
1489
1490 static int nbd_open(struct block_device *bdev, fmode_t mode)
1491 {
1492         struct nbd_device *nbd;
1493         int ret = 0;
1494
1495         mutex_lock(&nbd_index_mutex);
1496         nbd = bdev->bd_disk->private_data;
1497         if (!nbd) {
1498                 ret = -ENXIO;
1499                 goto out;
1500         }
1501         if (!refcount_inc_not_zero(&nbd->refs)) {
1502                 ret = -ENXIO;
1503                 goto out;
1504         }
1505         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1506                 struct nbd_config *config;
1507
1508                 mutex_lock(&nbd->config_lock);
1509                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1510                         mutex_unlock(&nbd->config_lock);
1511                         goto out;
1512                 }
1513                 config = nbd->config = nbd_alloc_config();
1514                 if (!config) {
1515                         ret = -ENOMEM;
1516                         mutex_unlock(&nbd->config_lock);
1517                         goto out;
1518                 }
1519                 refcount_set(&nbd->config_refs, 1);
1520                 refcount_inc(&nbd->refs);
1521                 mutex_unlock(&nbd->config_lock);
1522                 if (max_part)
1523                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1524         } else if (nbd_disconnected(nbd->config)) {
1525                 if (max_part)
1526                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1527         }
1528 out:
1529         mutex_unlock(&nbd_index_mutex);
1530         return ret;
1531 }
1532
1533 static void nbd_release(struct gendisk *disk, fmode_t mode)
1534 {
1535         struct nbd_device *nbd = disk->private_data;
1536
1537         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1538                         disk->part0->bd_openers == 0)
1539                 nbd_disconnect_and_put(nbd);
1540
1541         nbd_config_put(nbd);
1542         nbd_put(nbd);
1543 }
1544
1545 static const struct block_device_operations nbd_fops =
1546 {
1547         .owner =        THIS_MODULE,
1548         .open =         nbd_open,
1549         .release =      nbd_release,
1550         .ioctl =        nbd_ioctl,
1551         .compat_ioctl = nbd_ioctl,
1552 };
1553
1554 #if IS_ENABLED(CONFIG_DEBUG_FS)
1555
1556 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1557 {
1558         struct nbd_device *nbd = s->private;
1559
1560         if (nbd->task_recv)
1561                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1562
1563         return 0;
1564 }
1565
1566 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1567
1568 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1569 {
1570         struct nbd_device *nbd = s->private;
1571         u32 flags = nbd->config->flags;
1572
1573         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1574
1575         seq_puts(s, "Known flags:\n");
1576
1577         if (flags & NBD_FLAG_HAS_FLAGS)
1578                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1579         if (flags & NBD_FLAG_READ_ONLY)
1580                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1581         if (flags & NBD_FLAG_SEND_FLUSH)
1582                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1583         if (flags & NBD_FLAG_SEND_FUA)
1584                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1585         if (flags & NBD_FLAG_SEND_TRIM)
1586                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1587
1588         return 0;
1589 }
1590
1591 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1592
1593 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1594 {
1595         struct dentry *dir;
1596         struct nbd_config *config = nbd->config;
1597
1598         if (!nbd_dbg_dir)
1599                 return -EIO;
1600
1601         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1602         if (!dir) {
1603                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1604                         nbd_name(nbd));
1605                 return -EIO;
1606         }
1607         config->dbg_dir = dir;
1608
1609         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1610         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1611         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1612         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1613         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1614
1615         return 0;
1616 }
1617
1618 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1619 {
1620         debugfs_remove_recursive(nbd->config->dbg_dir);
1621 }
1622
1623 static int nbd_dbg_init(void)
1624 {
1625         struct dentry *dbg_dir;
1626
1627         dbg_dir = debugfs_create_dir("nbd", NULL);
1628         if (!dbg_dir)
1629                 return -EIO;
1630
1631         nbd_dbg_dir = dbg_dir;
1632
1633         return 0;
1634 }
1635
1636 static void nbd_dbg_close(void)
1637 {
1638         debugfs_remove_recursive(nbd_dbg_dir);
1639 }
1640
1641 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1642
1643 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1644 {
1645         return 0;
1646 }
1647
1648 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1649 {
1650 }
1651
1652 static int nbd_dbg_init(void)
1653 {
1654         return 0;
1655 }
1656
1657 static void nbd_dbg_close(void)
1658 {
1659 }
1660
1661 #endif
1662
1663 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1664                             unsigned int hctx_idx, unsigned int numa_node)
1665 {
1666         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1667         cmd->nbd = set->driver_data;
1668         cmd->flags = 0;
1669         mutex_init(&cmd->lock);
1670         return 0;
1671 }
1672
1673 static const struct blk_mq_ops nbd_mq_ops = {
1674         .queue_rq       = nbd_queue_rq,
1675         .complete       = nbd_complete_rq,
1676         .init_request   = nbd_init_request,
1677         .timeout        = nbd_xmit_timeout,
1678 };
1679
1680 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1681 {
1682         struct nbd_device *nbd;
1683         struct gendisk *disk;
1684         int err = -ENOMEM;
1685
1686         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1687         if (!nbd)
1688                 goto out;
1689
1690         nbd->tag_set.ops = &nbd_mq_ops;
1691         nbd->tag_set.nr_hw_queues = 1;
1692         nbd->tag_set.queue_depth = 128;
1693         nbd->tag_set.numa_node = NUMA_NO_NODE;
1694         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1695         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1696                 BLK_MQ_F_BLOCKING;
1697         nbd->tag_set.driver_data = nbd;
1698         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1699         nbd->backend = NULL;
1700
1701         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1702         if (err)
1703                 goto out_free_nbd;
1704
1705         mutex_lock(&nbd_index_mutex);
1706         if (index >= 0) {
1707                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1708                                 GFP_KERNEL);
1709                 if (err == -ENOSPC)
1710                         err = -EEXIST;
1711         } else {
1712                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1713                 if (err >= 0)
1714                         index = err;
1715         }
1716         nbd->index = index;
1717         mutex_unlock(&nbd_index_mutex);
1718         if (err < 0)
1719                 goto out_free_tags;
1720
1721         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1722         if (IS_ERR(disk)) {
1723                 err = PTR_ERR(disk);
1724                 goto out_free_idr;
1725         }
1726         nbd->disk = disk;
1727
1728         /*
1729          * Tell the block layer that we are not a rotational device
1730          */
1731         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1732         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1733         disk->queue->limits.discard_granularity = 0;
1734         disk->queue->limits.discard_alignment = 0;
1735         blk_queue_max_discard_sectors(disk->queue, 0);
1736         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1737         blk_queue_max_segments(disk->queue, USHRT_MAX);
1738         blk_queue_max_hw_sectors(disk->queue, 65536);
1739         disk->queue->limits.max_sectors = 256;
1740
1741         mutex_init(&nbd->config_lock);
1742         refcount_set(&nbd->config_refs, 0);
1743         /*
1744          * Start out with a zero references to keep other threads from using
1745          * this device until it is fully initialized.
1746          */
1747         refcount_set(&nbd->refs, 0);
1748         INIT_LIST_HEAD(&nbd->list);
1749         disk->major = NBD_MAJOR;
1750
1751         /* Too big first_minor can cause duplicate creation of
1752          * sysfs files/links, since first_minor will be truncated to
1753          * byte in __device_add_disk().
1754          */
1755         disk->first_minor = index << part_shift;
1756         if (disk->first_minor > 0xff) {
1757                 err = -EINVAL;
1758                 goto out_free_idr;
1759         }
1760
1761         disk->minors = 1 << part_shift;
1762         disk->fops = &nbd_fops;
1763         disk->private_data = nbd;
1764         sprintf(disk->disk_name, "nbd%d", index);
1765         add_disk(disk);
1766
1767         /*
1768          * Now publish the device.
1769          */
1770         refcount_set(&nbd->refs, refs);
1771         nbd_total_devices++;
1772         return nbd;
1773
1774 out_free_idr:
1775         mutex_lock(&nbd_index_mutex);
1776         idr_remove(&nbd_index_idr, index);
1777         mutex_unlock(&nbd_index_mutex);
1778 out_free_tags:
1779         blk_mq_free_tag_set(&nbd->tag_set);
1780 out_free_nbd:
1781         kfree(nbd);
1782 out:
1783         return ERR_PTR(err);
1784 }
1785
1786 static struct nbd_device *nbd_find_get_unused(void)
1787 {
1788         struct nbd_device *nbd;
1789         int id;
1790
1791         lockdep_assert_held(&nbd_index_mutex);
1792
1793         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1794                 if (refcount_read(&nbd->config_refs) ||
1795                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1796                         continue;
1797                 if (refcount_inc_not_zero(&nbd->refs))
1798                         return nbd;
1799         }
1800
1801         return NULL;
1802 }
1803
1804 /* Netlink interface. */
1805 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1806         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1807         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1808         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1809         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1810         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1811         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1812         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1813         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1814         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1815         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1816 };
1817
1818 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1819         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1820 };
1821
1822 /* We don't use this right now since we don't parse the incoming list, but we
1823  * still want it here so userspace knows what to expect.
1824  */
1825 static const struct nla_policy __attribute__((unused))
1826 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1827         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1828         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1829 };
1830
1831 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1832 {
1833         struct nbd_config *config = nbd->config;
1834         u64 bsize = nbd_blksize(config);
1835         u64 bytes = config->bytesize;
1836
1837         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1838                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1839
1840         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1841                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1842
1843         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1844                 return nbd_set_size(nbd, bytes, bsize);
1845         return 0;
1846 }
1847
1848 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1849 {
1850         struct nbd_device *nbd;
1851         struct nbd_config *config;
1852         int index = -1;
1853         int ret;
1854         bool put_dev = false;
1855
1856         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1857                 return -EPERM;
1858
1859         if (info->attrs[NBD_ATTR_INDEX])
1860                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1861         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1862                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1863                 return -EINVAL;
1864         }
1865         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1866                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1867                 return -EINVAL;
1868         }
1869 again:
1870         mutex_lock(&nbd_index_mutex);
1871         if (index == -1) {
1872                 nbd = nbd_find_get_unused();
1873         } else {
1874                 nbd = idr_find(&nbd_index_idr, index);
1875                 if (nbd) {
1876                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1877                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
1878                             !refcount_inc_not_zero(&nbd->refs)) {
1879                                 mutex_unlock(&nbd_index_mutex);
1880                                 pr_err("nbd: device at index %d is going down\n",
1881                                         index);
1882                                 return -EINVAL;
1883                         }
1884                 }
1885         }
1886         mutex_unlock(&nbd_index_mutex);
1887
1888         if (!nbd) {
1889                 nbd = nbd_dev_add(index, 2);
1890                 if (IS_ERR(nbd)) {
1891                         pr_err("nbd: failed to add new device\n");
1892                         return PTR_ERR(nbd);
1893                 }
1894         }
1895
1896         mutex_lock(&nbd->config_lock);
1897         if (refcount_read(&nbd->config_refs)) {
1898                 mutex_unlock(&nbd->config_lock);
1899                 nbd_put(nbd);
1900                 if (index == -1)
1901                         goto again;
1902                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1903                 return -EBUSY;
1904         }
1905         if (WARN_ON(nbd->config)) {
1906                 mutex_unlock(&nbd->config_lock);
1907                 nbd_put(nbd);
1908                 return -EINVAL;
1909         }
1910         config = nbd->config = nbd_alloc_config();
1911         if (!nbd->config) {
1912                 mutex_unlock(&nbd->config_lock);
1913                 nbd_put(nbd);
1914                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1915                 return -ENOMEM;
1916         }
1917         refcount_set(&nbd->config_refs, 1);
1918         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1919
1920         ret = nbd_genl_size_set(info, nbd);
1921         if (ret)
1922                 goto out;
1923
1924         if (info->attrs[NBD_ATTR_TIMEOUT])
1925                 nbd_set_cmd_timeout(nbd,
1926                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1927         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1928                 config->dead_conn_timeout =
1929                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1930                 config->dead_conn_timeout *= HZ;
1931         }
1932         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1933                 config->flags =
1934                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1935         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1936                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1937                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1938                         /*
1939                          * We have 1 ref to keep the device around, and then 1
1940                          * ref for our current operation here, which will be
1941                          * inherited by the config.  If we already have
1942                          * DESTROY_ON_DISCONNECT set then we know we don't have
1943                          * that extra ref already held so we don't need the
1944                          * put_dev.
1945                          */
1946                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1947                                               &nbd->flags))
1948                                 put_dev = true;
1949                 } else {
1950                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1951                                                &nbd->flags))
1952                                 refcount_inc(&nbd->refs);
1953                 }
1954                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1955                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1956                                 &config->runtime_flags);
1957                 }
1958         }
1959
1960         if (info->attrs[NBD_ATTR_SOCKETS]) {
1961                 struct nlattr *attr;
1962                 int rem, fd;
1963
1964                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1965                                     rem) {
1966                         struct nlattr *socks[NBD_SOCK_MAX+1];
1967
1968                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1969                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1970                                 ret = -EINVAL;
1971                                 goto out;
1972                         }
1973                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1974                                                           attr,
1975                                                           nbd_sock_policy,
1976                                                           info->extack);
1977                         if (ret != 0) {
1978                                 printk(KERN_ERR "nbd: error processing sock list\n");
1979                                 ret = -EINVAL;
1980                                 goto out;
1981                         }
1982                         if (!socks[NBD_SOCK_FD])
1983                                 continue;
1984                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1985                         ret = nbd_add_socket(nbd, fd, true);
1986                         if (ret)
1987                                 goto out;
1988                 }
1989         }
1990         ret = nbd_start_device(nbd);
1991         if (ret)
1992                 goto out;
1993         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
1994                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
1995                                           GFP_KERNEL);
1996                 if (!nbd->backend) {
1997                         ret = -ENOMEM;
1998                         goto out;
1999                 }
2000         }
2001         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2002         if (ret) {
2003                 dev_err(disk_to_dev(nbd->disk),
2004                         "device_create_file failed for backend!\n");
2005                 goto out;
2006         }
2007         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2008 out:
2009         mutex_unlock(&nbd->config_lock);
2010         if (!ret) {
2011                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2012                 refcount_inc(&nbd->config_refs);
2013                 nbd_connect_reply(info, nbd->index);
2014         }
2015         nbd_config_put(nbd);
2016         if (put_dev)
2017                 nbd_put(nbd);
2018         return ret;
2019 }
2020
2021 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2022 {
2023         mutex_lock(&nbd->config_lock);
2024         nbd_disconnect(nbd);
2025         sock_shutdown(nbd);
2026         /*
2027          * Make sure recv thread has finished, so it does not drop the last
2028          * config ref and try to destroy the workqueue from inside the work
2029          * queue. And this also ensure that we can safely call nbd_clear_que()
2030          * to cancel the inflight I/Os.
2031          */
2032         if (nbd->recv_workq)
2033                 flush_workqueue(nbd->recv_workq);
2034         nbd_clear_que(nbd);
2035         nbd->task_setup = NULL;
2036         mutex_unlock(&nbd->config_lock);
2037
2038         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2039                                &nbd->config->runtime_flags))
2040                 nbd_config_put(nbd);
2041 }
2042
2043 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2044 {
2045         struct nbd_device *nbd;
2046         int index;
2047
2048         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2049                 return -EPERM;
2050
2051         if (!info->attrs[NBD_ATTR_INDEX]) {
2052                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2053                 return -EINVAL;
2054         }
2055         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2056         mutex_lock(&nbd_index_mutex);
2057         nbd = idr_find(&nbd_index_idr, index);
2058         if (!nbd) {
2059                 mutex_unlock(&nbd_index_mutex);
2060                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2061                        index);
2062                 return -EINVAL;
2063         }
2064         if (!refcount_inc_not_zero(&nbd->refs)) {
2065                 mutex_unlock(&nbd_index_mutex);
2066                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2067                        index);
2068                 return -EINVAL;
2069         }
2070         mutex_unlock(&nbd_index_mutex);
2071         if (!refcount_inc_not_zero(&nbd->config_refs))
2072                 goto put_nbd;
2073         nbd_disconnect_and_put(nbd);
2074         nbd_config_put(nbd);
2075 put_nbd:
2076         nbd_put(nbd);
2077         return 0;
2078 }
2079
2080 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2081 {
2082         struct nbd_device *nbd = NULL;
2083         struct nbd_config *config;
2084         int index;
2085         int ret = 0;
2086         bool put_dev = false;
2087
2088         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2089                 return -EPERM;
2090
2091         if (!info->attrs[NBD_ATTR_INDEX]) {
2092                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2093                 return -EINVAL;
2094         }
2095         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2096         mutex_lock(&nbd_index_mutex);
2097         nbd = idr_find(&nbd_index_idr, index);
2098         if (!nbd) {
2099                 mutex_unlock(&nbd_index_mutex);
2100                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2101                        index);
2102                 return -EINVAL;
2103         }
2104         if (nbd->backend) {
2105                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2106                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2107                                        nbd->backend)) {
2108                                 mutex_unlock(&nbd_index_mutex);
2109                                 dev_err(nbd_to_dev(nbd),
2110                                         "backend image doesn't match with %s\n",
2111                                         nbd->backend);
2112                                 return -EINVAL;
2113                         }
2114                 } else {
2115                         mutex_unlock(&nbd_index_mutex);
2116                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2117                         return -EINVAL;
2118                 }
2119         }
2120         if (!refcount_inc_not_zero(&nbd->refs)) {
2121                 mutex_unlock(&nbd_index_mutex);
2122                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2123                        index);
2124                 return -EINVAL;
2125         }
2126         mutex_unlock(&nbd_index_mutex);
2127
2128         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2129                 dev_err(nbd_to_dev(nbd),
2130                         "not configured, cannot reconfigure\n");
2131                 nbd_put(nbd);
2132                 return -EINVAL;
2133         }
2134
2135         mutex_lock(&nbd->config_lock);
2136         config = nbd->config;
2137         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2138             !nbd->task_recv) {
2139                 dev_err(nbd_to_dev(nbd),
2140                         "not configured, cannot reconfigure\n");
2141                 ret = -EINVAL;
2142                 goto out;
2143         }
2144
2145         ret = nbd_genl_size_set(info, nbd);
2146         if (ret)
2147                 goto out;
2148
2149         if (info->attrs[NBD_ATTR_TIMEOUT])
2150                 nbd_set_cmd_timeout(nbd,
2151                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2152         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2153                 config->dead_conn_timeout =
2154                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2155                 config->dead_conn_timeout *= HZ;
2156         }
2157         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2158                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2159                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2160                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2161                                               &nbd->flags))
2162                                 put_dev = true;
2163                 } else {
2164                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2165                                                &nbd->flags))
2166                                 refcount_inc(&nbd->refs);
2167                 }
2168
2169                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2170                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2171                                         &config->runtime_flags);
2172                 } else {
2173                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2174                                         &config->runtime_flags);
2175                 }
2176         }
2177
2178         if (info->attrs[NBD_ATTR_SOCKETS]) {
2179                 struct nlattr *attr;
2180                 int rem, fd;
2181
2182                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2183                                     rem) {
2184                         struct nlattr *socks[NBD_SOCK_MAX+1];
2185
2186                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2187                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2188                                 ret = -EINVAL;
2189                                 goto out;
2190                         }
2191                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2192                                                           attr,
2193                                                           nbd_sock_policy,
2194                                                           info->extack);
2195                         if (ret != 0) {
2196                                 printk(KERN_ERR "nbd: error processing sock list\n");
2197                                 ret = -EINVAL;
2198                                 goto out;
2199                         }
2200                         if (!socks[NBD_SOCK_FD])
2201                                 continue;
2202                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2203                         ret = nbd_reconnect_socket(nbd, fd);
2204                         if (ret) {
2205                                 if (ret == -ENOSPC)
2206                                         ret = 0;
2207                                 goto out;
2208                         }
2209                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2210                 }
2211         }
2212 out:
2213         mutex_unlock(&nbd->config_lock);
2214         nbd_config_put(nbd);
2215         nbd_put(nbd);
2216         if (put_dev)
2217                 nbd_put(nbd);
2218         return ret;
2219 }
2220
2221 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2222         {
2223                 .cmd    = NBD_CMD_CONNECT,
2224                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2225                 .doit   = nbd_genl_connect,
2226         },
2227         {
2228                 .cmd    = NBD_CMD_DISCONNECT,
2229                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2230                 .doit   = nbd_genl_disconnect,
2231         },
2232         {
2233                 .cmd    = NBD_CMD_RECONFIGURE,
2234                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2235                 .doit   = nbd_genl_reconfigure,
2236         },
2237         {
2238                 .cmd    = NBD_CMD_STATUS,
2239                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2240                 .doit   = nbd_genl_status,
2241         },
2242 };
2243
2244 static const struct genl_multicast_group nbd_mcast_grps[] = {
2245         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2246 };
2247
2248 static struct genl_family nbd_genl_family __ro_after_init = {
2249         .hdrsize        = 0,
2250         .name           = NBD_GENL_FAMILY_NAME,
2251         .version        = NBD_GENL_VERSION,
2252         .module         = THIS_MODULE,
2253         .small_ops      = nbd_connect_genl_ops,
2254         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2255         .maxattr        = NBD_ATTR_MAX,
2256         .policy = nbd_attr_policy,
2257         .mcgrps         = nbd_mcast_grps,
2258         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2259 };
2260
2261 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2262 {
2263         struct nlattr *dev_opt;
2264         u8 connected = 0;
2265         int ret;
2266
2267         /* This is a little racey, but for status it's ok.  The
2268          * reason we don't take a ref here is because we can't
2269          * take a ref in the index == -1 case as we would need
2270          * to put under the nbd_index_mutex, which could
2271          * deadlock if we are configured to remove ourselves
2272          * once we're disconnected.
2273          */
2274         if (refcount_read(&nbd->config_refs))
2275                 connected = 1;
2276         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2277         if (!dev_opt)
2278                 return -EMSGSIZE;
2279         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2280         if (ret)
2281                 return -EMSGSIZE;
2282         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2283                          connected);
2284         if (ret)
2285                 return -EMSGSIZE;
2286         nla_nest_end(reply, dev_opt);
2287         return 0;
2288 }
2289
2290 static int status_cb(int id, void *ptr, void *data)
2291 {
2292         struct nbd_device *nbd = ptr;
2293         return populate_nbd_status(nbd, (struct sk_buff *)data);
2294 }
2295
2296 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2297 {
2298         struct nlattr *dev_list;
2299         struct sk_buff *reply;
2300         void *reply_head;
2301         size_t msg_size;
2302         int index = -1;
2303         int ret = -ENOMEM;
2304
2305         if (info->attrs[NBD_ATTR_INDEX])
2306                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2307
2308         mutex_lock(&nbd_index_mutex);
2309
2310         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2311                                   nla_attr_size(sizeof(u8)));
2312         msg_size *= (index == -1) ? nbd_total_devices : 1;
2313
2314         reply = genlmsg_new(msg_size, GFP_KERNEL);
2315         if (!reply)
2316                 goto out;
2317         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2318                                        NBD_CMD_STATUS);
2319         if (!reply_head) {
2320                 nlmsg_free(reply);
2321                 goto out;
2322         }
2323
2324         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2325         if (index == -1) {
2326                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2327                 if (ret) {
2328                         nlmsg_free(reply);
2329                         goto out;
2330                 }
2331         } else {
2332                 struct nbd_device *nbd;
2333                 nbd = idr_find(&nbd_index_idr, index);
2334                 if (nbd) {
2335                         ret = populate_nbd_status(nbd, reply);
2336                         if (ret) {
2337                                 nlmsg_free(reply);
2338                                 goto out;
2339                         }
2340                 }
2341         }
2342         nla_nest_end(reply, dev_list);
2343         genlmsg_end(reply, reply_head);
2344         ret = genlmsg_reply(reply, info);
2345 out:
2346         mutex_unlock(&nbd_index_mutex);
2347         return ret;
2348 }
2349
2350 static void nbd_connect_reply(struct genl_info *info, int index)
2351 {
2352         struct sk_buff *skb;
2353         void *msg_head;
2354         int ret;
2355
2356         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2357         if (!skb)
2358                 return;
2359         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2360                                      NBD_CMD_CONNECT);
2361         if (!msg_head) {
2362                 nlmsg_free(skb);
2363                 return;
2364         }
2365         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2366         if (ret) {
2367                 nlmsg_free(skb);
2368                 return;
2369         }
2370         genlmsg_end(skb, msg_head);
2371         genlmsg_reply(skb, info);
2372 }
2373
2374 static void nbd_mcast_index(int index)
2375 {
2376         struct sk_buff *skb;
2377         void *msg_head;
2378         int ret;
2379
2380         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2381         if (!skb)
2382                 return;
2383         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2384                                      NBD_CMD_LINK_DEAD);
2385         if (!msg_head) {
2386                 nlmsg_free(skb);
2387                 return;
2388         }
2389         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2390         if (ret) {
2391                 nlmsg_free(skb);
2392                 return;
2393         }
2394         genlmsg_end(skb, msg_head);
2395         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2396 }
2397
2398 static void nbd_dead_link_work(struct work_struct *work)
2399 {
2400         struct link_dead_args *args = container_of(work, struct link_dead_args,
2401                                                    work);
2402         nbd_mcast_index(args->index);
2403         kfree(args);
2404 }
2405
2406 static int __init nbd_init(void)
2407 {
2408         int i;
2409
2410         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2411
2412         if (max_part < 0) {
2413                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2414                 return -EINVAL;
2415         }
2416
2417         part_shift = 0;
2418         if (max_part > 0) {
2419                 part_shift = fls(max_part);
2420
2421                 /*
2422                  * Adjust max_part according to part_shift as it is exported
2423                  * to user space so that user can know the max number of
2424                  * partition kernel should be able to manage.
2425                  *
2426                  * Note that -1 is required because partition 0 is reserved
2427                  * for the whole disk.
2428                  */
2429                 max_part = (1UL << part_shift) - 1;
2430         }
2431
2432         if ((1UL << part_shift) > DISK_MAX_PARTS)
2433                 return -EINVAL;
2434
2435         if (nbds_max > 1UL << (MINORBITS - part_shift))
2436                 return -EINVAL;
2437
2438         if (register_blkdev(NBD_MAJOR, "nbd"))
2439                 return -EIO;
2440
2441         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2442         if (!nbd_del_wq) {
2443                 unregister_blkdev(NBD_MAJOR, "nbd");
2444                 return -ENOMEM;
2445         }
2446
2447         if (genl_register_family(&nbd_genl_family)) {
2448                 destroy_workqueue(nbd_del_wq);
2449                 unregister_blkdev(NBD_MAJOR, "nbd");
2450                 return -EINVAL;
2451         }
2452         nbd_dbg_init();
2453
2454         for (i = 0; i < nbds_max; i++)
2455                 nbd_dev_add(i, 1);
2456         return 0;
2457 }
2458
2459 static int nbd_exit_cb(int id, void *ptr, void *data)
2460 {
2461         struct list_head *list = (struct list_head *)data;
2462         struct nbd_device *nbd = ptr;
2463
2464         /* Skip nbd that is being removed asynchronously */
2465         if (refcount_read(&nbd->refs))
2466                 list_add_tail(&nbd->list, list);
2467
2468         return 0;
2469 }
2470
2471 static void __exit nbd_cleanup(void)
2472 {
2473         struct nbd_device *nbd;
2474         LIST_HEAD(del_list);
2475
2476         nbd_dbg_close();
2477
2478         mutex_lock(&nbd_index_mutex);
2479         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2480         mutex_unlock(&nbd_index_mutex);
2481
2482         while (!list_empty(&del_list)) {
2483                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2484                 list_del_init(&nbd->list);
2485                 if (refcount_read(&nbd->refs) != 1)
2486                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2487                 nbd_put(nbd);
2488         }
2489
2490         /* Also wait for nbd_dev_remove_work() completes */
2491         destroy_workqueue(nbd_del_wq);
2492
2493         idr_destroy(&nbd_index_idr);
2494         genl_unregister_family(&nbd_genl_family);
2495         unregister_blkdev(NBD_MAJOR, "nbd");
2496 }
2497
2498 module_init(nbd_init);
2499 module_exit(nbd_cleanup);
2500
2501 MODULE_DESCRIPTION("Network Block Device");
2502 MODULE_LICENSE("GPL");
2503
2504 module_param(nbds_max, int, 0444);
2505 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2506 module_param(max_part, int, 0444);
2507 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");