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_setup;
 126
 127         unsigned long flags;
 128         pid_t pid; /* pid of nbd-client, if attached */
 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", nbd->pid);
 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         destroy_workqueue(nbd->recv_workq);
 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->pid)
 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
 900         if (!wait_event_timeout(config->conn_wait,
 901                                 test_bit(NBD_RT_DISCONNECTED,
 902                                          &config->runtime_flags) ||
 903                                 atomic_read(&config->live_connections) > 0,
 904                                 config->dead_conn_timeout))
 905                 return 0;
 906
 907         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
 908 }
 909
 910 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 911 {
 912         struct request *req = blk_mq_rq_from_pdu(cmd);
 913         struct nbd_device *nbd = cmd->nbd;
 914         struct nbd_config *config;
 915         struct nbd_sock *nsock;
 916         int ret;
 917
 918         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 919                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 920                                     "Socks array is empty\n");
 921                 blk_mq_start_request(req);
 922                 return -EINVAL;
 923         }
 924         config = nbd->config;
 925
 926         if (index >= config->num_connections) {
 927                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 928                                     "Attempted send on invalid socket\n");
 929                 nbd_config_put(nbd);
 930                 blk_mq_start_request(req);
 931                 return -EINVAL;
 932         }
 933         cmd->status = BLK_STS_OK;
 934 again:
 935         nsock = config->socks[index];
 936         mutex_lock(&nsock->tx_lock);
 937         if (nsock->dead) {
 938                 int old_index = index;
 939                 index = find_fallback(nbd, index);
 940                 mutex_unlock(&nsock->tx_lock);
 941                 if (index < 0) {
 942                         if (wait_for_reconnect(nbd)) {
 943                                 index = old_index;
 944                                 goto again;
 945                         }
 946                         /* All the sockets should already be down at this point,
 947                          * we just want to make sure that DISCONNECTED is set so
 948                          * any requests that come in that were queue'ed waiting
 949                          * for the reconnect timer don't trigger the timer again
 950                          * and instead just error out.
 951                          */
 952                         sock_shutdown(nbd);
 953                         nbd_config_put(nbd);
 954                         blk_mq_start_request(req);
 955                         return -EIO;
 956                 }
 957                 goto again;
 958         }
 959
 960         /* Handle the case that we have a pending request that was partially
 961          * transmitted that _has_ to be serviced first.  We need to call requeue
 962          * here so that it gets put _after_ the request that is already on the
 963          * dispatch list.
 964          */
 965         blk_mq_start_request(req);
 966         if (unlikely(nsock->pending && nsock->pending != req)) {
 967                 nbd_requeue_cmd(cmd);
 968                 ret = 0;
 969                 goto out;
 970         }
 971         /*
 972          * Some failures are related to the link going down, so anything that
 973          * returns EAGAIN can be retried on a different socket.
 974          */
 975         ret = nbd_send_cmd(nbd, cmd, index);
 976         if (ret == -EAGAIN) {
 977                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 978                                     "Request send failed, requeueing\n");
 979                 nbd_mark_nsock_dead(nbd, nsock, 1);
 980                 nbd_requeue_cmd(cmd);
 981                 ret = 0;
 982         }
 983 out:
 984         mutex_unlock(&nsock->tx_lock);
 985         nbd_config_put(nbd);
 986         return ret;
 987 }
 988
 989 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 990                         const struct blk_mq_queue_data *bd)
 991 {
 992         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 993         int ret;
 994
 995         /*
 996          * Since we look at the bio's to send the request over the network we
 997          * need to make sure the completion work doesn't mark this request done
 998          * before we are done doing our send.  This keeps us from dereferencing
 999          * freed data if we have particularly fast completions (ie we get the
1000          * completion before we exit sock_xmit on the last bvec) or in the case
1001          * that the server is misbehaving (or there was an error) before we're
1002          * done sending everything over the wire.
1003          */
1004         mutex_lock(&cmd->lock);
1005         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1006
1007         /* We can be called directly from the user space process, which means we
1008          * could possibly have signals pending so our sendmsg will fail.  In
1009          * this case we need to return that we are busy, otherwise error out as
1010          * appropriate.
1011          */
1012         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1013         if (ret < 0)
1014                 ret = BLK_STS_IOERR;
1015         else if (!ret)
1016                 ret = BLK_STS_OK;
1017         mutex_unlock(&cmd->lock);
1018
1019         return ret;
1020 }
1021
1022 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1023                                      int *err)
1024 {
1025         struct socket *sock;
1026
1027         *err = 0;
1028         sock = sockfd_lookup(fd, err);
1029         if (!sock)
1030                 return NULL;
1031
1032         if (sock->ops->shutdown == sock_no_shutdown) {
1033                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1034                 *err = -EINVAL;
1035                 sockfd_put(sock);
1036                 return NULL;
1037         }
1038
1039         return sock;
1040 }
1041
1042 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1043                           bool netlink)
1044 {
1045         struct nbd_config *config = nbd->config;
1046         struct socket *sock;
1047         struct nbd_sock **socks;
1048         struct nbd_sock *nsock;
1049         int err;
1050
1051         sock = nbd_get_socket(nbd, arg, &err);
1052         if (!sock)
1053                 return err;
1054
1055         /*
1056          * We need to make sure we don't get any errant requests while we're
1057          * reallocating the ->socks array.
1058          */
1059         blk_mq_freeze_queue(nbd->disk->queue);
1060
1061         if (!netlink && !nbd->task_setup &&
1062             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1063                 nbd->task_setup = current;
1064
1065         if (!netlink &&
1066             (nbd->task_setup != current ||
1067              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1068                 dev_err(disk_to_dev(nbd->disk),
1069                         "Device being setup by another task");
1070                 err = -EBUSY;
1071                 goto put_socket;
1072         }
1073
1074         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1075         if (!nsock) {
1076                 err = -ENOMEM;
1077                 goto put_socket;
1078         }
1079
1080         socks = krealloc(config->socks, (config->num_connections + 1) *
1081                          sizeof(struct nbd_sock *), GFP_KERNEL);
1082         if (!socks) {
1083                 kfree(nsock);
1084                 err = -ENOMEM;
1085                 goto put_socket;
1086         }
1087
1088         config->socks = socks;
1089
1090         nsock->fallback_index = -1;
1091         nsock->dead = false;
1092         mutex_init(&nsock->tx_lock);
1093         nsock->sock = sock;
1094         nsock->pending = NULL;
1095         nsock->sent = 0;
1096         nsock->cookie = 0;
1097         socks[config->num_connections++] = nsock;
1098         atomic_inc(&config->live_connections);
1099         blk_mq_unfreeze_queue(nbd->disk->queue);
1100
1101         return 0;
1102
1103 put_socket:
1104         blk_mq_unfreeze_queue(nbd->disk->queue);
1105         sockfd_put(sock);
1106         return err;
1107 }
1108
1109 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1110 {
1111         struct nbd_config *config = nbd->config;
1112         struct socket *sock, *old;
1113         struct recv_thread_args *args;
1114         int i;
1115         int err;
1116
1117         sock = nbd_get_socket(nbd, arg, &err);
1118         if (!sock)
1119                 return err;
1120
1121         args = kzalloc(sizeof(*args), GFP_KERNEL);
1122         if (!args) {
1123                 sockfd_put(sock);
1124                 return -ENOMEM;
1125         }
1126
1127         for (i = 0; i < config->num_connections; i++) {
1128                 struct nbd_sock *nsock = config->socks[i];
1129
1130                 if (!nsock->dead)
1131                         continue;
1132
1133                 mutex_lock(&nsock->tx_lock);
1134                 if (!nsock->dead) {
1135                         mutex_unlock(&nsock->tx_lock);
1136                         continue;
1137                 }
1138                 sk_set_memalloc(sock->sk);
1139                 if (nbd->tag_set.timeout)
1140                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1141                 atomic_inc(&config->recv_threads);
1142                 refcount_inc(&nbd->config_refs);
1143                 old = nsock->sock;
1144                 nsock->fallback_index = -1;
1145                 nsock->sock = sock;
1146                 nsock->dead = false;
1147                 INIT_WORK(&args->work, recv_work);
1148                 args->index = i;
1149                 args->nbd = nbd;
1150                 nsock->cookie++;
1151                 mutex_unlock(&nsock->tx_lock);
1152                 sockfd_put(old);
1153
1154                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1155
1156                 /* We take the tx_mutex in an error path in the recv_work, so we
1157                  * need to queue_work outside of the tx_mutex.
1158                  */
1159                 queue_work(nbd->recv_workq, &args->work);
1160
1161                 atomic_inc(&config->live_connections);
1162                 wake_up(&config->conn_wait);
1163                 return 0;
1164         }
1165         sockfd_put(sock);
1166         kfree(args);
1167         return -ENOSPC;
1168 }
1169
1170 static void nbd_bdev_reset(struct block_device *bdev)
1171 {
1172         if (bdev->bd_openers > 1)
1173                 return;
1174         set_capacity(bdev->bd_disk, 0);
1175 }
1176
1177 static void nbd_parse_flags(struct nbd_device *nbd)
1178 {
1179         struct nbd_config *config = nbd->config;
1180         if (config->flags & NBD_FLAG_READ_ONLY)
1181                 set_disk_ro(nbd->disk, true);
1182         else
1183                 set_disk_ro(nbd->disk, false);
1184         if (config->flags & NBD_FLAG_SEND_TRIM)
1185                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1186         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1187                 if (config->flags & NBD_FLAG_SEND_FUA)
1188                         blk_queue_write_cache(nbd->disk->queue, true, true);
1189                 else
1190                         blk_queue_write_cache(nbd->disk->queue, true, false);
1191         }
1192         else
1193                 blk_queue_write_cache(nbd->disk->queue, false, false);
1194 }
1195
1196 static void send_disconnects(struct nbd_device *nbd)
1197 {
1198         struct nbd_config *config = nbd->config;
1199         struct nbd_request request = {
1200                 .magic = htonl(NBD_REQUEST_MAGIC),
1201                 .type = htonl(NBD_CMD_DISC),
1202         };
1203         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1204         struct iov_iter from;
1205         int i, ret;
1206
1207         for (i = 0; i < config->num_connections; i++) {
1208                 struct nbd_sock *nsock = config->socks[i];
1209
1210                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1211                 mutex_lock(&nsock->tx_lock);
1212                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1213                 if (ret <= 0)
1214                         dev_err(disk_to_dev(nbd->disk),
1215                                 "Send disconnect failed %d\n", ret);
1216                 mutex_unlock(&nsock->tx_lock);
1217         }
1218 }
1219
1220 static int nbd_disconnect(struct nbd_device *nbd)
1221 {
1222         struct nbd_config *config = nbd->config;
1223
1224         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1225         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1226         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1227         send_disconnects(nbd);
1228         return 0;
1229 }
1230
1231 static void nbd_clear_sock(struct nbd_device *nbd)
1232 {
1233         sock_shutdown(nbd);
1234         nbd_clear_que(nbd);
1235         nbd->task_setup = NULL;
1236 }
1237
1238 static void nbd_config_put(struct nbd_device *nbd)
1239 {
1240         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1241                                         &nbd->config_lock)) {
1242                 struct nbd_config *config = nbd->config;
1243                 nbd_dev_dbg_close(nbd);
1244                 nbd_size_clear(nbd);
1245                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1246                                        &config->runtime_flags))
1247                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1248                 nbd->pid = 0;
1249                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1250                                        &config->runtime_flags)) {
1251                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1252                         kfree(nbd->backend);
1253                         nbd->backend = NULL;
1254                 }
1255                 nbd_clear_sock(nbd);
1256                 if (config->num_connections) {
1257                         int i;
1258                         for (i = 0; i < config->num_connections; i++) {
1259                                 sockfd_put(config->socks[i]->sock);
1260                                 kfree(config->socks[i]);
1261                         }
1262                         kfree(config->socks);
1263                 }
1264                 kfree(nbd->config);
1265                 nbd->config = 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->pid)
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         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1296         nbd->pid = task_pid_nr(current);
1297
1298         nbd_parse_flags(nbd);
1299
1300         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1301         if (error) {
1302                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1303                 return error;
1304         }
1305         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1306
1307         nbd_dev_dbg_init(nbd);
1308         for (i = 0; i < num_connections; i++) {
1309                 struct recv_thread_args *args;
1310
1311                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1312                 if (!args) {
1313                         sock_shutdown(nbd);
1314                         /*
1315                          * If num_connections is m (2 < m),
1316                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1317                          * But NO.(n + 1) failed. We still have n recv threads.
1318                          * So, add flush_workqueue here to prevent recv threads
1319                          * dropping the last config_refs and trying to destroy
1320                          * the workqueue from inside the workqueue.
1321                          */
1322                         if (i)
1323                                 flush_workqueue(nbd->recv_workq);
1324                         return -ENOMEM;
1325                 }
1326                 sk_set_memalloc(config->socks[i]->sock->sk);
1327                 if (nbd->tag_set.timeout)
1328                         config->socks[i]->sock->sk->sk_sndtimeo =
1329                                 nbd->tag_set.timeout;
1330                 atomic_inc(&config->recv_threads);
1331                 refcount_inc(&nbd->config_refs);
1332                 INIT_WORK(&args->work, recv_work);
1333                 args->nbd = nbd;
1334                 args->index = i;
1335                 queue_work(nbd->recv_workq, &args->work);
1336         }
1337         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1338 }
1339
1340 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1341 {
1342         struct nbd_config *config = nbd->config;
1343         int ret;
1344
1345         ret = nbd_start_device(nbd);
1346         if (ret)
1347                 return ret;
1348
1349         if (max_part)
1350                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1351         mutex_unlock(&nbd->config_lock);
1352         ret = wait_event_interruptible(config->recv_wq,
1353                                          atomic_read(&config->recv_threads) == 0);
1354         if (ret)
1355                 sock_shutdown(nbd);
1356         flush_workqueue(nbd->recv_workq);
1357
1358         mutex_lock(&nbd->config_lock);
1359         nbd_bdev_reset(bdev);
1360         /* user requested, ignore socket errors */
1361         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1362                 ret = 0;
1363         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1364                 ret = -ETIMEDOUT;
1365         return ret;
1366 }
1367
1368 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1369                                  struct block_device *bdev)
1370 {
1371         nbd_clear_sock(nbd);
1372         __invalidate_device(bdev, true);
1373         nbd_bdev_reset(bdev);
1374         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1375                                &nbd->config->runtime_flags))
1376                 nbd_config_put(nbd);
1377 }
1378
1379 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1380 {
1381         nbd->tag_set.timeout = timeout * HZ;
1382         if (timeout)
1383                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1384         else
1385                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1386 }
1387
1388 /* Must be called with config_lock held */
1389 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1390                        unsigned int cmd, unsigned long arg)
1391 {
1392         struct nbd_config *config = nbd->config;
1393         loff_t bytesize;
1394
1395         switch (cmd) {
1396         case NBD_DISCONNECT:
1397                 return nbd_disconnect(nbd);
1398         case NBD_CLEAR_SOCK:
1399                 nbd_clear_sock_ioctl(nbd, bdev);
1400                 return 0;
1401         case NBD_SET_SOCK:
1402                 return nbd_add_socket(nbd, arg, false);
1403         case NBD_SET_BLKSIZE:
1404                 return nbd_set_size(nbd, config->bytesize, arg);
1405         case NBD_SET_SIZE:
1406                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1407         case NBD_SET_SIZE_BLOCKS:
1408                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1409                         return -EINVAL;
1410                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1411         case NBD_SET_TIMEOUT:
1412                 nbd_set_cmd_timeout(nbd, arg);
1413                 return 0;
1414
1415         case NBD_SET_FLAGS:
1416                 config->flags = arg;
1417                 return 0;
1418         case NBD_DO_IT:
1419                 return nbd_start_device_ioctl(nbd, bdev);
1420         case NBD_CLEAR_QUE:
1421                 /*
1422                  * This is for compatibility only.  The queue is always cleared
1423                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1424                  */
1425                 return 0;
1426         case NBD_PRINT_DEBUG:
1427                 /*
1428                  * For compatibility only, we no longer keep a list of
1429                  * outstanding requests.
1430                  */
1431                 return 0;
1432         }
1433         return -ENOTTY;
1434 }
1435
1436 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1437                      unsigned int cmd, unsigned long arg)
1438 {
1439         struct nbd_device *nbd = bdev->bd_disk->private_data;
1440         struct nbd_config *config = nbd->config;
1441         int error = -EINVAL;
1442
1443         if (!capable(CAP_SYS_ADMIN))
1444                 return -EPERM;
1445
1446         /* The block layer will pass back some non-nbd ioctls in case we have
1447          * special handling for them, but we don't so just return an error.
1448          */
1449         if (_IOC_TYPE(cmd) != 0xab)
1450                 return -EINVAL;
1451
1452         mutex_lock(&nbd->config_lock);
1453
1454         /* Don't allow ioctl operations on a nbd device that was created with
1455          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1456          */
1457         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1458             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1459                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1460         else
1461                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1462         mutex_unlock(&nbd->config_lock);
1463         return error;
1464 }
1465
1466 static struct nbd_config *nbd_alloc_config(void)
1467 {
1468         struct nbd_config *config;
1469
1470         if (!try_module_get(THIS_MODULE))
1471                 return ERR_PTR(-ENODEV);
1472
1473         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1474         if (!config) {
1475                 module_put(THIS_MODULE);
1476                 return ERR_PTR(-ENOMEM);
1477         }
1478
1479         atomic_set(&config->recv_threads, 0);
1480         init_waitqueue_head(&config->recv_wq);
1481         init_waitqueue_head(&config->conn_wait);
1482         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1483         atomic_set(&config->live_connections, 0);
1484         return config;
1485 }
1486
1487 static int nbd_open(struct block_device *bdev, fmode_t mode)
1488 {
1489         struct nbd_device *nbd;
1490         int ret = 0;
1491
1492         mutex_lock(&nbd_index_mutex);
1493         nbd = bdev->bd_disk->private_data;
1494         if (!nbd) {
1495                 ret = -ENXIO;
1496                 goto out;
1497         }
1498         if (!refcount_inc_not_zero(&nbd->refs)) {
1499                 ret = -ENXIO;
1500                 goto out;
1501         }
1502         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1503                 struct nbd_config *config;
1504
1505                 mutex_lock(&nbd->config_lock);
1506                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1507                         mutex_unlock(&nbd->config_lock);
1508                         goto out;
1509                 }
1510                 config = nbd_alloc_config();
1511                 if (IS_ERR(config)) {
1512                         ret = PTR_ERR(config);
1513                         mutex_unlock(&nbd->config_lock);
1514                         goto out;
1515                 }
1516                 nbd->config = config;
1517                 refcount_set(&nbd->config_refs, 1);
1518                 refcount_inc(&nbd->refs);
1519                 mutex_unlock(&nbd->config_lock);
1520                 if (max_part)
1521                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1522         } else if (nbd_disconnected(nbd->config)) {
1523                 if (max_part)
1524                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1525         }
1526 out:
1527         mutex_unlock(&nbd_index_mutex);
1528         return ret;
1529 }
1530
1531 static void nbd_release(struct gendisk *disk, fmode_t mode)
1532 {
1533         struct nbd_device *nbd = disk->private_data;
1534
1535         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1536                         disk->part0->bd_openers == 0)
1537                 nbd_disconnect_and_put(nbd);
1538
1539         nbd_config_put(nbd);
1540         nbd_put(nbd);
1541 }
1542
1543 static const struct block_device_operations nbd_fops =
1544 {
1545         .owner =        THIS_MODULE,
1546         .open =         nbd_open,
1547         .release =      nbd_release,
1548         .ioctl =        nbd_ioctl,
1549         .compat_ioctl = nbd_ioctl,
1550 };
1551
1552 #if IS_ENABLED(CONFIG_DEBUG_FS)
1553
1554 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1555 {
1556         struct nbd_device *nbd = s->private;
1557
1558         if (nbd->pid)
1559                 seq_printf(s, "recv: %d\n", nbd->pid);
1560
1561         return 0;
1562 }
1563
1564 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1565
1566 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1567 {
1568         struct nbd_device *nbd = s->private;
1569         u32 flags = nbd->config->flags;
1570
1571         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1572
1573         seq_puts(s, "Known flags:\n");
1574
1575         if (flags & NBD_FLAG_HAS_FLAGS)
1576                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1577         if (flags & NBD_FLAG_READ_ONLY)
1578                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1579         if (flags & NBD_FLAG_SEND_FLUSH)
1580                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1581         if (flags & NBD_FLAG_SEND_FUA)
1582                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1583         if (flags & NBD_FLAG_SEND_TRIM)
1584                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1585
1586         return 0;
1587 }
1588
1589 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1590
1591 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1592 {
1593         struct dentry *dir;
1594         struct nbd_config *config = nbd->config;
1595
1596         if (!nbd_dbg_dir)
1597                 return -EIO;
1598
1599         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1600         if (!dir) {
1601                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1602                         nbd_name(nbd));
1603                 return -EIO;
1604         }
1605         config->dbg_dir = dir;
1606
1607         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1608         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1609         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1610         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1611         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1612
1613         return 0;
1614 }
1615
1616 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1617 {
1618         debugfs_remove_recursive(nbd->config->dbg_dir);
1619 }
1620
1621 static int nbd_dbg_init(void)
1622 {
1623         struct dentry *dbg_dir;
1624
1625         dbg_dir = debugfs_create_dir("nbd", NULL);
1626         if (!dbg_dir)
1627                 return -EIO;
1628
1629         nbd_dbg_dir = dbg_dir;
1630
1631         return 0;
1632 }
1633
1634 static void nbd_dbg_close(void)
1635 {
1636         debugfs_remove_recursive(nbd_dbg_dir);
1637 }
1638
1639 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1640
1641 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1642 {
1643         return 0;
1644 }
1645
1646 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1647 {
1648 }
1649
1650 static int nbd_dbg_init(void)
1651 {
1652         return 0;
1653 }
1654
1655 static void nbd_dbg_close(void)
1656 {
1657 }
1658
1659 #endif
1660
1661 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1662                             unsigned int hctx_idx, unsigned int numa_node)
1663 {
1664         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1665         cmd->nbd = set->driver_data;
1666         cmd->flags = 0;
1667         mutex_init(&cmd->lock);
1668         return 0;
1669 }
1670
1671 static const struct blk_mq_ops nbd_mq_ops = {
1672         .queue_rq       = nbd_queue_rq,
1673         .complete       = nbd_complete_rq,
1674         .init_request   = nbd_init_request,
1675         .timeout        = nbd_xmit_timeout,
1676 };
1677
1678 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1679 {
1680         struct nbd_device *nbd;
1681         struct gendisk *disk;
1682         int err = -ENOMEM;
1683
1684         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1685         if (!nbd)
1686                 goto out;
1687
1688         nbd->tag_set.ops = &nbd_mq_ops;
1689         nbd->tag_set.nr_hw_queues = 1;
1690         nbd->tag_set.queue_depth = 128;
1691         nbd->tag_set.numa_node = NUMA_NO_NODE;
1692         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1693         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1694                 BLK_MQ_F_BLOCKING;
1695         nbd->tag_set.driver_data = nbd;
1696         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1697         nbd->backend = NULL;
1698
1699         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1700         if (err)
1701                 goto out_free_nbd;
1702
1703         mutex_lock(&nbd_index_mutex);
1704         if (index >= 0) {
1705                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1706                                 GFP_KERNEL);
1707                 if (err == -ENOSPC)
1708                         err = -EEXIST;
1709         } else {
1710                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1711                 if (err >= 0)
1712                         index = err;
1713         }
1714         nbd->index = index;
1715         mutex_unlock(&nbd_index_mutex);
1716         if (err < 0)
1717                 goto out_free_tags;
1718
1719         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1720         if (IS_ERR(disk)) {
1721                 err = PTR_ERR(disk);
1722                 goto out_free_idr;
1723         }
1724         nbd->disk = disk;
1725
1726         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1727                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1728                                           WQ_UNBOUND, 0, nbd->index);
1729         if (!nbd->recv_workq) {
1730                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1731                 err = -ENOMEM;
1732                 goto out_err_disk;
1733         }
1734
1735         /*
1736          * Tell the block layer that we are not a rotational device
1737          */
1738         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1739         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1740         disk->queue->limits.discard_granularity = 0;
1741         disk->queue->limits.discard_alignment = 0;
1742         blk_queue_max_discard_sectors(disk->queue, 0);
1743         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1744         blk_queue_max_segments(disk->queue, USHRT_MAX);
1745         blk_queue_max_hw_sectors(disk->queue, 65536);
1746         disk->queue->limits.max_sectors = 256;
1747
1748         mutex_init(&nbd->config_lock);
1749         refcount_set(&nbd->config_refs, 0);
1750         /*
1751          * Start out with a zero references to keep other threads from using
1752          * this device until it is fully initialized.
1753          */
1754         refcount_set(&nbd->refs, 0);
1755         INIT_LIST_HEAD(&nbd->list);
1756         disk->major = NBD_MAJOR;
1757
1758         /* Too big first_minor can cause duplicate creation of
1759          * sysfs files/links, since index << part_shift might overflow, or
1760          * MKDEV() expect that the max bits of first_minor is 20.
1761          */
1762         disk->first_minor = index << part_shift;
1763         if (disk->first_minor < index || disk->first_minor > MINORMASK) {
1764                 err = -EINVAL;
1765                 goto out_free_work;
1766         }
1767
1768         disk->minors = 1 << part_shift;
1769         disk->fops = &nbd_fops;
1770         disk->private_data = nbd;
1771         sprintf(disk->disk_name, "nbd%d", index);
1772         err = add_disk(disk);
1773         if (err)
1774                 goto out_free_work;
1775
1776         /*
1777          * Now publish the device.
1778          */
1779         refcount_set(&nbd->refs, refs);
1780         nbd_total_devices++;
1781         return nbd;
1782
1783 out_free_work:
1784         destroy_workqueue(nbd->recv_workq);
1785 out_err_disk:
1786         blk_cleanup_disk(disk);
1787 out_free_idr:
1788         mutex_lock(&nbd_index_mutex);
1789         idr_remove(&nbd_index_idr, index);
1790         mutex_unlock(&nbd_index_mutex);
1791 out_free_tags:
1792         blk_mq_free_tag_set(&nbd->tag_set);
1793 out_free_nbd:
1794         kfree(nbd);
1795 out:
1796         return ERR_PTR(err);
1797 }
1798
1799 static struct nbd_device *nbd_find_get_unused(void)
1800 {
1801         struct nbd_device *nbd;
1802         int id;
1803
1804         lockdep_assert_held(&nbd_index_mutex);
1805
1806         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1807                 if (refcount_read(&nbd->config_refs) ||
1808                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1809                         continue;
1810                 if (refcount_inc_not_zero(&nbd->refs))
1811                         return nbd;
1812         }
1813
1814         return NULL;
1815 }
1816
1817 /* Netlink interface. */
1818 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1819         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1820         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1821         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1822         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1823         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1824         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1825         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1826         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1827         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1828         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1829 };
1830
1831 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1832         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1833 };
1834
1835 /* We don't use this right now since we don't parse the incoming list, but we
1836  * still want it here so userspace knows what to expect.
1837  */
1838 static const struct nla_policy __attribute__((unused))
1839 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1840         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1841         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1842 };
1843
1844 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1845 {
1846         struct nbd_config *config = nbd->config;
1847         u64 bsize = nbd_blksize(config);
1848         u64 bytes = config->bytesize;
1849
1850         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1851                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1852
1853         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1854                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1855
1856         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1857                 return nbd_set_size(nbd, bytes, bsize);
1858         return 0;
1859 }
1860
1861 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1862 {
1863         struct nbd_device *nbd;
1864         struct nbd_config *config;
1865         int index = -1;
1866         int ret;
1867         bool put_dev = false;
1868
1869         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1870                 return -EPERM;
1871
1872         if (info->attrs[NBD_ATTR_INDEX])
1873                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1874         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1875                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1876                 return -EINVAL;
1877         }
1878         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1879                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1880                 return -EINVAL;
1881         }
1882 again:
1883         mutex_lock(&nbd_index_mutex);
1884         if (index == -1) {
1885                 nbd = nbd_find_get_unused();
1886         } else {
1887                 nbd = idr_find(&nbd_index_idr, index);
1888                 if (nbd) {
1889                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1890                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
1891                             !refcount_inc_not_zero(&nbd->refs)) {
1892                                 mutex_unlock(&nbd_index_mutex);
1893                                 pr_err("nbd: device at index %d is going down\n",
1894                                         index);
1895                                 return -EINVAL;
1896                         }
1897                 }
1898         }
1899         mutex_unlock(&nbd_index_mutex);
1900
1901         if (!nbd) {
1902                 nbd = nbd_dev_add(index, 2);
1903                 if (IS_ERR(nbd)) {
1904                         pr_err("nbd: failed to add new device\n");
1905                         return PTR_ERR(nbd);
1906                 }
1907         }
1908
1909         mutex_lock(&nbd->config_lock);
1910         if (refcount_read(&nbd->config_refs)) {
1911                 mutex_unlock(&nbd->config_lock);
1912                 nbd_put(nbd);
1913                 if (index == -1)
1914                         goto again;
1915                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1916                 return -EBUSY;
1917         }
1918         if (WARN_ON(nbd->config)) {
1919                 mutex_unlock(&nbd->config_lock);
1920                 nbd_put(nbd);
1921                 return -EINVAL;
1922         }
1923         config = nbd_alloc_config();
1924         if (IS_ERR(config)) {
1925                 mutex_unlock(&nbd->config_lock);
1926                 nbd_put(nbd);
1927                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1928                 return PTR_ERR(config);
1929         }
1930         nbd->config = config;
1931         refcount_set(&nbd->config_refs, 1);
1932         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1933
1934         ret = nbd_genl_size_set(info, nbd);
1935         if (ret)
1936                 goto out;
1937
1938         if (info->attrs[NBD_ATTR_TIMEOUT])
1939                 nbd_set_cmd_timeout(nbd,
1940                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1941         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1942                 config->dead_conn_timeout =
1943                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1944                 config->dead_conn_timeout *= HZ;
1945         }
1946         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1947                 config->flags =
1948                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1949         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1950                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1951                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1952                         /*
1953                          * We have 1 ref to keep the device around, and then 1
1954                          * ref for our current operation here, which will be
1955                          * inherited by the config.  If we already have
1956                          * DESTROY_ON_DISCONNECT set then we know we don't have
1957                          * that extra ref already held so we don't need the
1958                          * put_dev.
1959                          */
1960                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1961                                               &nbd->flags))
1962                                 put_dev = true;
1963                 } else {
1964                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1965                                                &nbd->flags))
1966                                 refcount_inc(&nbd->refs);
1967                 }
1968                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1969                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1970                                 &config->runtime_flags);
1971                 }
1972         }
1973
1974         if (info->attrs[NBD_ATTR_SOCKETS]) {
1975                 struct nlattr *attr;
1976                 int rem, fd;
1977
1978                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1979                                     rem) {
1980                         struct nlattr *socks[NBD_SOCK_MAX+1];
1981
1982                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1983                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1984                                 ret = -EINVAL;
1985                                 goto out;
1986                         }
1987                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1988                                                           attr,
1989                                                           nbd_sock_policy,
1990                                                           info->extack);
1991                         if (ret != 0) {
1992                                 printk(KERN_ERR "nbd: error processing sock list\n");
1993                                 ret = -EINVAL;
1994                                 goto out;
1995                         }
1996                         if (!socks[NBD_SOCK_FD])
1997                                 continue;
1998                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1999                         ret = nbd_add_socket(nbd, fd, true);
2000                         if (ret)
2001                                 goto out;
2002                 }
2003         }
2004         ret = nbd_start_device(nbd);
2005         if (ret)
2006                 goto out;
2007         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2008                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2009                                           GFP_KERNEL);
2010                 if (!nbd->backend) {
2011                         ret = -ENOMEM;
2012                         goto out;
2013                 }
2014         }
2015         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2016         if (ret) {
2017                 dev_err(disk_to_dev(nbd->disk),
2018                         "device_create_file failed for backend!\n");
2019                 goto out;
2020         }
2021         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2022 out:
2023         mutex_unlock(&nbd->config_lock);
2024         if (!ret) {
2025                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2026                 refcount_inc(&nbd->config_refs);
2027                 nbd_connect_reply(info, nbd->index);
2028         }
2029         nbd_config_put(nbd);
2030         if (put_dev)
2031                 nbd_put(nbd);
2032         return ret;
2033 }
2034
2035 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2036 {
2037         mutex_lock(&nbd->config_lock);
2038         nbd_disconnect(nbd);
2039         sock_shutdown(nbd);
2040         wake_up(&nbd->config->conn_wait);
2041         /*
2042          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2043          * to cancel the inflight I/Os.
2044          */
2045         flush_workqueue(nbd->recv_workq);
2046         nbd_clear_que(nbd);
2047         nbd->task_setup = NULL;
2048         mutex_unlock(&nbd->config_lock);
2049
2050         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2051                                &nbd->config->runtime_flags))
2052                 nbd_config_put(nbd);
2053 }
2054
2055 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2056 {
2057         struct nbd_device *nbd;
2058         int index;
2059
2060         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2061                 return -EPERM;
2062
2063         if (!info->attrs[NBD_ATTR_INDEX]) {
2064                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2065                 return -EINVAL;
2066         }
2067         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2068         mutex_lock(&nbd_index_mutex);
2069         nbd = idr_find(&nbd_index_idr, index);
2070         if (!nbd) {
2071                 mutex_unlock(&nbd_index_mutex);
2072                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2073                        index);
2074                 return -EINVAL;
2075         }
2076         if (!refcount_inc_not_zero(&nbd->refs)) {
2077                 mutex_unlock(&nbd_index_mutex);
2078                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2079                        index);
2080                 return -EINVAL;
2081         }
2082         mutex_unlock(&nbd_index_mutex);
2083         if (!refcount_inc_not_zero(&nbd->config_refs))
2084                 goto put_nbd;
2085         nbd_disconnect_and_put(nbd);
2086         nbd_config_put(nbd);
2087 put_nbd:
2088         nbd_put(nbd);
2089         return 0;
2090 }
2091
2092 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2093 {
2094         struct nbd_device *nbd = NULL;
2095         struct nbd_config *config;
2096         int index;
2097         int ret = 0;
2098         bool put_dev = false;
2099
2100         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2101                 return -EPERM;
2102
2103         if (!info->attrs[NBD_ATTR_INDEX]) {
2104                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2105                 return -EINVAL;
2106         }
2107         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2108         mutex_lock(&nbd_index_mutex);
2109         nbd = idr_find(&nbd_index_idr, index);
2110         if (!nbd) {
2111                 mutex_unlock(&nbd_index_mutex);
2112                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2113                        index);
2114                 return -EINVAL;
2115         }
2116         if (nbd->backend) {
2117                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2118                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2119                                        nbd->backend)) {
2120                                 mutex_unlock(&nbd_index_mutex);
2121                                 dev_err(nbd_to_dev(nbd),
2122                                         "backend image doesn't match with %s\n",
2123                                         nbd->backend);
2124                                 return -EINVAL;
2125                         }
2126                 } else {
2127                         mutex_unlock(&nbd_index_mutex);
2128                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2129                         return -EINVAL;
2130                 }
2131         }
2132         if (!refcount_inc_not_zero(&nbd->refs)) {
2133                 mutex_unlock(&nbd_index_mutex);
2134                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2135                        index);
2136                 return -EINVAL;
2137         }
2138         mutex_unlock(&nbd_index_mutex);
2139
2140         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2141                 dev_err(nbd_to_dev(nbd),
2142                         "not configured, cannot reconfigure\n");
2143                 nbd_put(nbd);
2144                 return -EINVAL;
2145         }
2146
2147         mutex_lock(&nbd->config_lock);
2148         config = nbd->config;
2149         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2150             !nbd->pid) {
2151                 dev_err(nbd_to_dev(nbd),
2152                         "not configured, cannot reconfigure\n");
2153                 ret = -EINVAL;
2154                 goto out;
2155         }
2156
2157         ret = nbd_genl_size_set(info, nbd);
2158         if (ret)
2159                 goto out;
2160
2161         if (info->attrs[NBD_ATTR_TIMEOUT])
2162                 nbd_set_cmd_timeout(nbd,
2163                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2164         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2165                 config->dead_conn_timeout =
2166                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2167                 config->dead_conn_timeout *= HZ;
2168         }
2169         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2170                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2171                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2172                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2173                                               &nbd->flags))
2174                                 put_dev = true;
2175                 } else {
2176                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2177                                                &nbd->flags))
2178                                 refcount_inc(&nbd->refs);
2179                 }
2180
2181                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2182                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2183                                         &config->runtime_flags);
2184                 } else {
2185                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2186                                         &config->runtime_flags);
2187                 }
2188         }
2189
2190         if (info->attrs[NBD_ATTR_SOCKETS]) {
2191                 struct nlattr *attr;
2192                 int rem, fd;
2193
2194                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2195                                     rem) {
2196                         struct nlattr *socks[NBD_SOCK_MAX+1];
2197
2198                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2199                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2200                                 ret = -EINVAL;
2201                                 goto out;
2202                         }
2203                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2204                                                           attr,
2205                                                           nbd_sock_policy,
2206                                                           info->extack);
2207                         if (ret != 0) {
2208                                 printk(KERN_ERR "nbd: error processing sock list\n");
2209                                 ret = -EINVAL;
2210                                 goto out;
2211                         }
2212                         if (!socks[NBD_SOCK_FD])
2213                                 continue;
2214                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2215                         ret = nbd_reconnect_socket(nbd, fd);
2216                         if (ret) {
2217                                 if (ret == -ENOSPC)
2218                                         ret = 0;
2219                                 goto out;
2220                         }
2221                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2222                 }
2223         }
2224 out:
2225         mutex_unlock(&nbd->config_lock);
2226         nbd_config_put(nbd);
2227         nbd_put(nbd);
2228         if (put_dev)
2229                 nbd_put(nbd);
2230         return ret;
2231 }
2232
2233 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2234         {
2235                 .cmd    = NBD_CMD_CONNECT,
2236                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2237                 .doit   = nbd_genl_connect,
2238         },
2239         {
2240                 .cmd    = NBD_CMD_DISCONNECT,
2241                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2242                 .doit   = nbd_genl_disconnect,
2243         },
2244         {
2245                 .cmd    = NBD_CMD_RECONFIGURE,
2246                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2247                 .doit   = nbd_genl_reconfigure,
2248         },
2249         {
2250                 .cmd    = NBD_CMD_STATUS,
2251                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2252                 .doit   = nbd_genl_status,
2253         },
2254 };
2255
2256 static const struct genl_multicast_group nbd_mcast_grps[] = {
2257         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2258 };
2259
2260 static struct genl_family nbd_genl_family __ro_after_init = {
2261         .hdrsize        = 0,
2262         .name           = NBD_GENL_FAMILY_NAME,
2263         .version        = NBD_GENL_VERSION,
2264         .module         = THIS_MODULE,
2265         .small_ops      = nbd_connect_genl_ops,
2266         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2267         .maxattr        = NBD_ATTR_MAX,
2268         .policy = nbd_attr_policy,
2269         .mcgrps         = nbd_mcast_grps,
2270         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2271 };
2272
2273 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2274 {
2275         struct nlattr *dev_opt;
2276         u8 connected = 0;
2277         int ret;
2278
2279         /* This is a little racey, but for status it's ok.  The
2280          * reason we don't take a ref here is because we can't
2281          * take a ref in the index == -1 case as we would need
2282          * to put under the nbd_index_mutex, which could
2283          * deadlock if we are configured to remove ourselves
2284          * once we're disconnected.
2285          */
2286         if (refcount_read(&nbd->config_refs))
2287                 connected = 1;
2288         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2289         if (!dev_opt)
2290                 return -EMSGSIZE;
2291         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2292         if (ret)
2293                 return -EMSGSIZE;
2294         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2295                          connected);
2296         if (ret)
2297                 return -EMSGSIZE;
2298         nla_nest_end(reply, dev_opt);
2299         return 0;
2300 }
2301
2302 static int status_cb(int id, void *ptr, void *data)
2303 {
2304         struct nbd_device *nbd = ptr;
2305         return populate_nbd_status(nbd, (struct sk_buff *)data);
2306 }
2307
2308 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2309 {
2310         struct nlattr *dev_list;
2311         struct sk_buff *reply;
2312         void *reply_head;
2313         size_t msg_size;
2314         int index = -1;
2315         int ret = -ENOMEM;
2316
2317         if (info->attrs[NBD_ATTR_INDEX])
2318                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2319
2320         mutex_lock(&nbd_index_mutex);
2321
2322         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2323                                   nla_attr_size(sizeof(u8)));
2324         msg_size *= (index == -1) ? nbd_total_devices : 1;
2325
2326         reply = genlmsg_new(msg_size, GFP_KERNEL);
2327         if (!reply)
2328                 goto out;
2329         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2330                                        NBD_CMD_STATUS);
2331         if (!reply_head) {
2332                 nlmsg_free(reply);
2333                 goto out;
2334         }
2335
2336         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2337         if (index == -1) {
2338                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2339                 if (ret) {
2340                         nlmsg_free(reply);
2341                         goto out;
2342                 }
2343         } else {
2344                 struct nbd_device *nbd;
2345                 nbd = idr_find(&nbd_index_idr, index);
2346                 if (nbd) {
2347                         ret = populate_nbd_status(nbd, reply);
2348                         if (ret) {
2349                                 nlmsg_free(reply);
2350                                 goto out;
2351                         }
2352                 }
2353         }
2354         nla_nest_end(reply, dev_list);
2355         genlmsg_end(reply, reply_head);
2356         ret = genlmsg_reply(reply, info);
2357 out:
2358         mutex_unlock(&nbd_index_mutex);
2359         return ret;
2360 }
2361
2362 static void nbd_connect_reply(struct genl_info *info, int index)
2363 {
2364         struct sk_buff *skb;
2365         void *msg_head;
2366         int ret;
2367
2368         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2369         if (!skb)
2370                 return;
2371         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2372                                      NBD_CMD_CONNECT);
2373         if (!msg_head) {
2374                 nlmsg_free(skb);
2375                 return;
2376         }
2377         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2378         if (ret) {
2379                 nlmsg_free(skb);
2380                 return;
2381         }
2382         genlmsg_end(skb, msg_head);
2383         genlmsg_reply(skb, info);
2384 }
2385
2386 static void nbd_mcast_index(int index)
2387 {
2388         struct sk_buff *skb;
2389         void *msg_head;
2390         int ret;
2391
2392         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2393         if (!skb)
2394                 return;
2395         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2396                                      NBD_CMD_LINK_DEAD);
2397         if (!msg_head) {
2398                 nlmsg_free(skb);
2399                 return;
2400         }
2401         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2402         if (ret) {
2403                 nlmsg_free(skb);
2404                 return;
2405         }
2406         genlmsg_end(skb, msg_head);
2407         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2408 }
2409
2410 static void nbd_dead_link_work(struct work_struct *work)
2411 {
2412         struct link_dead_args *args = container_of(work, struct link_dead_args,
2413                                                    work);
2414         nbd_mcast_index(args->index);
2415         kfree(args);
2416 }
2417
2418 static int __init nbd_init(void)
2419 {
2420         int i;
2421
2422         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2423
2424         if (max_part < 0) {
2425                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2426                 return -EINVAL;
2427         }
2428
2429         part_shift = 0;
2430         if (max_part > 0) {
2431                 part_shift = fls(max_part);
2432
2433                 /*
2434                  * Adjust max_part according to part_shift as it is exported
2435                  * to user space so that user can know the max number of
2436                  * partition kernel should be able to manage.
2437                  *
2438                  * Note that -1 is required because partition 0 is reserved
2439                  * for the whole disk.
2440                  */
2441                 max_part = (1UL << part_shift) - 1;
2442         }
2443
2444         if ((1UL << part_shift) > DISK_MAX_PARTS)
2445                 return -EINVAL;
2446
2447         if (nbds_max > 1UL << (MINORBITS - part_shift))
2448                 return -EINVAL;
2449
2450         if (register_blkdev(NBD_MAJOR, "nbd"))
2451                 return -EIO;
2452
2453         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2454         if (!nbd_del_wq) {
2455                 unregister_blkdev(NBD_MAJOR, "nbd");
2456                 return -ENOMEM;
2457         }
2458
2459         if (genl_register_family(&nbd_genl_family)) {
2460                 destroy_workqueue(nbd_del_wq);
2461                 unregister_blkdev(NBD_MAJOR, "nbd");
2462                 return -EINVAL;
2463         }
2464         nbd_dbg_init();
2465
2466         for (i = 0; i < nbds_max; i++)
2467                 nbd_dev_add(i, 1);
2468         return 0;
2469 }
2470
2471 static int nbd_exit_cb(int id, void *ptr, void *data)
2472 {
2473         struct list_head *list = (struct list_head *)data;
2474         struct nbd_device *nbd = ptr;
2475
2476         /* Skip nbd that is being removed asynchronously */
2477         if (refcount_read(&nbd->refs))
2478                 list_add_tail(&nbd->list, list);
2479
2480         return 0;
2481 }
2482
2483 static void __exit nbd_cleanup(void)
2484 {
2485         struct nbd_device *nbd;
2486         LIST_HEAD(del_list);
2487
2488         /*
2489          * Unregister netlink interface prior to waiting
2490          * for the completion of netlink commands.
2491          */
2492         genl_unregister_family(&nbd_genl_family);
2493
2494         nbd_dbg_close();
2495
2496         mutex_lock(&nbd_index_mutex);
2497         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2498         mutex_unlock(&nbd_index_mutex);
2499
2500         while (!list_empty(&del_list)) {
2501                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2502                 list_del_init(&nbd->list);
2503                 if (refcount_read(&nbd->config_refs))
2504                         printk(KERN_ERR "nbd: possibly leaking nbd_config (ref %d)\n",
2505                                         refcount_read(&nbd->config_refs));
2506                 if (refcount_read(&nbd->refs) != 1)
2507                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2508                 nbd_put(nbd);
2509         }
2510
2511         /* Also wait for nbd_dev_remove_work() completes */
2512         destroy_workqueue(nbd_del_wq);
2513
2514         idr_destroy(&nbd_index_idr);
2515         unregister_blkdev(NBD_MAJOR, "nbd");
2516 }
2517
2518 module_init(nbd_init);
2519 module_exit(nbd_cleanup);
2520
2521 MODULE_DESCRIPTION("Network Block Device");
2522 MODULE_LICENSE("GPL");
2523
2524 module_param(nbds_max, int, 0444);
2525 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2526 module_param(max_part, int, 0444);
2527 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");