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 #define pr_fmt(fmt) "nbd: " fmt
  15
  16 #include <linux/major.h>
  17
  18 #include <linux/blkdev.h>
  19 #include <linux/module.h>
  20 #include <linux/init.h>
  21 #include <linux/sched.h>
  22 #include <linux/sched/mm.h>
  23 #include <linux/fs.h>
  24 #include <linux/bio.h>
  25 #include <linux/stat.h>
  26 #include <linux/errno.h>
  27 #include <linux/file.h>
  28 #include <linux/ioctl.h>
  29 #include <linux/mutex.h>
  30 #include <linux/compiler.h>
  31 #include <linux/completion.h>
  32 #include <linux/err.h>
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <net/sock.h>
  36 #include <linux/net.h>
  37 #include <linux/kthread.h>
  38 #include <linux/types.h>
  39 #include <linux/debugfs.h>
  40 #include <linux/blk-mq.h>
  41
  42 #include <linux/uaccess.h>
  43 #include <asm/types.h>
  44
  45 #include <linux/nbd.h>
  46 #include <linux/nbd-netlink.h>
  47 #include <net/genetlink.h>
  48
  49 #define CREATE_TRACE_POINTS
  50 #include <trace/events/nbd.h>
  51
  52 static DEFINE_IDR(nbd_index_idr);
  53 static DEFINE_MUTEX(nbd_index_mutex);
  54 static struct workqueue_struct *nbd_del_wq;
  55 static int nbd_total_devices = 0;
  56
  57 struct nbd_sock {
  58         struct socket *sock;
  59         struct mutex tx_lock;
  60         struct request *pending;
  61         int sent;
  62         bool dead;
  63         int fallback_index;
  64         int cookie;
  65 };
  66
  67 struct recv_thread_args {
  68         struct work_struct work;
  69         struct nbd_device *nbd;
  70         struct nbd_sock *nsock;
  71         int index;
  72 };
  73
  74 struct link_dead_args {
  75         struct work_struct work;
  76         int index;
  77 };
  78
  79 #define NBD_RT_TIMEDOUT                 0
  80 #define NBD_RT_DISCONNECT_REQUESTED     1
  81 #define NBD_RT_DISCONNECTED             2
  82 #define NBD_RT_HAS_PID_FILE             3
  83 #define NBD_RT_HAS_CONFIG_REF           4
  84 #define NBD_RT_BOUND                    5
  85 #define NBD_RT_DISCONNECT_ON_CLOSE      6
  86 #define NBD_RT_HAS_BACKEND_FILE         7
  87
  88 #define NBD_DESTROY_ON_DISCONNECT       0
  89 #define NBD_DISCONNECT_REQUESTED        1
  90
  91 struct nbd_config {
  92         u32 flags;
  93         unsigned long runtime_flags;
  94         u64 dead_conn_timeout;
  95
  96         struct nbd_sock **socks;
  97         int num_connections;
  98         atomic_t live_connections;
  99         wait_queue_head_t conn_wait;
 100
 101         atomic_t recv_threads;
 102         wait_queue_head_t recv_wq;
 103         unsigned int blksize_bits;
 104         loff_t bytesize;
 105 #if IS_ENABLED(CONFIG_DEBUG_FS)
 106         struct dentry *dbg_dir;
 107 #endif
 108 };
 109
 110 static inline unsigned int nbd_blksize(struct nbd_config *config)
 111 {
 112         return 1u << config->blksize_bits;
 113 }
 114
 115 struct nbd_device {
 116         struct blk_mq_tag_set tag_set;
 117
 118         int index;
 119         refcount_t config_refs;
 120         refcount_t refs;
 121         struct nbd_config *config;
 122         struct mutex config_lock;
 123         struct gendisk *disk;
 124         struct workqueue_struct *recv_workq;
 125         struct work_struct remove_work;
 126
 127         struct list_head list;
 128         struct task_struct *task_setup;
 129
 130         unsigned long flags;
 131         pid_t pid; /* pid of nbd-client, if attached */
 132
 133         char *backend;
 134 };
 135
 136 #define NBD_CMD_REQUEUED        1
 137 /*
 138  * This flag will be set if nbd_queue_rq() succeed, and will be checked and
 139  * cleared in completion. Both setting and clearing of the flag are protected
 140  * by cmd->lock.
 141  */
 142 #define NBD_CMD_INFLIGHT        2
 143
 144 struct nbd_cmd {
 145         struct nbd_device *nbd;
 146         struct mutex lock;
 147         int index;
 148         int cookie;
 149         int retries;
 150         blk_status_t status;
 151         unsigned long flags;
 152         u32 cmd_cookie;
 153 };
 154
 155 #if IS_ENABLED(CONFIG_DEBUG_FS)
 156 static struct dentry *nbd_dbg_dir;
 157 #endif
 158
 159 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 160
 161 #define NBD_DEF_BLKSIZE_BITS 10
 162
 163 static unsigned int nbds_max = 16;
 164 static int max_part = 16;
 165 static int part_shift;
 166
 167 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 168 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 169 static void nbd_config_put(struct nbd_device *nbd);
 170 static void nbd_connect_reply(struct genl_info *info, int index);
 171 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 172 static void nbd_dead_link_work(struct work_struct *work);
 173 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 174
 175 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 176 {
 177         return disk_to_dev(nbd->disk);
 178 }
 179
 180 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 181 {
 182         struct request *req = blk_mq_rq_from_pdu(cmd);
 183
 184         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 185                 blk_mq_requeue_request(req, true);
 186 }
 187
 188 #define NBD_COOKIE_BITS 32
 189
 190 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 191 {
 192         struct request *req = blk_mq_rq_from_pdu(cmd);
 193         u32 tag = blk_mq_unique_tag(req);
 194         u64 cookie = cmd->cmd_cookie;
 195
 196         return (cookie << NBD_COOKIE_BITS) | tag;
 197 }
 198
 199 static u32 nbd_handle_to_tag(u64 handle)
 200 {
 201         return (u32)handle;
 202 }
 203
 204 static u32 nbd_handle_to_cookie(u64 handle)
 205 {
 206         return (u32)(handle >> NBD_COOKIE_BITS);
 207 }
 208
 209 static const char *nbdcmd_to_ascii(int cmd)
 210 {
 211         switch (cmd) {
 212         case  NBD_CMD_READ: return "read";
 213         case NBD_CMD_WRITE: return "write";
 214         case  NBD_CMD_DISC: return "disconnect";
 215         case NBD_CMD_FLUSH: return "flush";
 216         case  NBD_CMD_TRIM: return "trim/discard";
 217         }
 218         return "invalid";
 219 }
 220
 221 static ssize_t pid_show(struct device *dev,
 222                         struct device_attribute *attr, char *buf)
 223 {
 224         struct gendisk *disk = dev_to_disk(dev);
 225         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 226
 227         return sprintf(buf, "%d\n", nbd->pid);
 228 }
 229
 230 static const struct device_attribute pid_attr = {
 231         .attr = { .name = "pid", .mode = 0444},
 232         .show = pid_show,
 233 };
 234
 235 static ssize_t backend_show(struct device *dev,
 236                 struct device_attribute *attr, char *buf)
 237 {
 238         struct gendisk *disk = dev_to_disk(dev);
 239         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 240
 241         return sprintf(buf, "%s\n", nbd->backend ?: "");
 242 }
 243
 244 static const struct device_attribute backend_attr = {
 245         .attr = { .name = "backend", .mode = 0444},
 246         .show = backend_show,
 247 };
 248
 249 static void nbd_dev_remove(struct nbd_device *nbd)
 250 {
 251         struct gendisk *disk = nbd->disk;
 252
 253         del_gendisk(disk);
 254         blk_mq_free_tag_set(&nbd->tag_set);
 255
 256         /*
 257          * Remove from idr after del_gendisk() completes, so if the same ID is
 258          * reused, the following add_disk() will succeed.
 259          */
 260         mutex_lock(&nbd_index_mutex);
 261         idr_remove(&nbd_index_idr, nbd->index);
 262         mutex_unlock(&nbd_index_mutex);
 263         destroy_workqueue(nbd->recv_workq);
 264         put_disk(disk);
 265 }
 266
 267 static void nbd_dev_remove_work(struct work_struct *work)
 268 {
 269         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 270 }
 271
 272 static void nbd_put(struct nbd_device *nbd)
 273 {
 274         if (!refcount_dec_and_test(&nbd->refs))
 275                 return;
 276
 277         /* Call del_gendisk() asynchrounously to prevent deadlock */
 278         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 279                 queue_work(nbd_del_wq, &nbd->remove_work);
 280         else
 281                 nbd_dev_remove(nbd);
 282 }
 283
 284 static int nbd_disconnected(struct nbd_config *config)
 285 {
 286         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 287                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 288 }
 289
 290 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 291                                 int notify)
 292 {
 293         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 294                 struct link_dead_args *args;
 295                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 296                 if (args) {
 297                         INIT_WORK(&args->work, nbd_dead_link_work);
 298                         args->index = nbd->index;
 299                         queue_work(system_wq, &args->work);
 300                 }
 301         }
 302         if (!nsock->dead) {
 303                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 304                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 305                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 306                                                &nbd->config->runtime_flags)) {
 307                                 set_bit(NBD_RT_DISCONNECTED,
 308                                         &nbd->config->runtime_flags);
 309                                 dev_info(nbd_to_dev(nbd),
 310                                         "Disconnected due to user request.\n");
 311                         }
 312                 }
 313         }
 314         nsock->dead = true;
 315         nsock->pending = NULL;
 316         nsock->sent = 0;
 317 }
 318
 319 static int __nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 320                 loff_t blksize)
 321 {
 322         struct queue_limits lim;
 323         int error;
 324
 325         if (!blksize)
 326                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 327
 328         if (blk_validate_block_size(blksize))
 329                 return -EINVAL;
 330
 331         if (bytesize < 0)
 332                 return -EINVAL;
 333
 334         nbd->config->bytesize = bytesize;
 335         nbd->config->blksize_bits = __ffs(blksize);
 336
 337         if (!nbd->pid)
 338                 return 0;
 339
 340         lim = queue_limits_start_update(nbd->disk->queue);
 341         if (nbd->config->flags & NBD_FLAG_SEND_TRIM)
 342                 lim.max_hw_discard_sectors = UINT_MAX;
 343         else
 344                 lim.max_hw_discard_sectors = 0;
 345         lim.logical_block_size = blksize;
 346         lim.physical_block_size = blksize;
 347         error = queue_limits_commit_update(nbd->disk->queue, &lim);
 348         if (error)
 349                 return error;
 350
 351         if (max_part)
 352                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 353         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 354                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 355         return 0;
 356 }
 357
 358 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 359                 loff_t blksize)
 360 {
 361         int error;
 362
 363         blk_mq_freeze_queue(nbd->disk->queue);
 364         error = __nbd_set_size(nbd, bytesize, blksize);
 365         blk_mq_unfreeze_queue(nbd->disk->queue);
 366
 367         return error;
 368 }
 369
 370 static void nbd_complete_rq(struct request *req)
 371 {
 372         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 373
 374         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 375                 cmd->status ? "failed" : "done");
 376
 377         blk_mq_end_request(req, cmd->status);
 378 }
 379
 380 /*
 381  * Forcibly shutdown the socket causing all listeners to error
 382  */
 383 static void sock_shutdown(struct nbd_device *nbd)
 384 {
 385         struct nbd_config *config = nbd->config;
 386         int i;
 387
 388         if (config->num_connections == 0)
 389                 return;
 390         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 391                 return;
 392
 393         for (i = 0; i < config->num_connections; i++) {
 394                 struct nbd_sock *nsock = config->socks[i];
 395                 mutex_lock(&nsock->tx_lock);
 396                 nbd_mark_nsock_dead(nbd, nsock, 0);
 397                 mutex_unlock(&nsock->tx_lock);
 398         }
 399         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 400 }
 401
 402 static u32 req_to_nbd_cmd_type(struct request *req)
 403 {
 404         switch (req_op(req)) {
 405         case REQ_OP_DISCARD:
 406                 return NBD_CMD_TRIM;
 407         case REQ_OP_FLUSH:
 408                 return NBD_CMD_FLUSH;
 409         case REQ_OP_WRITE:
 410                 return NBD_CMD_WRITE;
 411         case REQ_OP_READ:
 412                 return NBD_CMD_READ;
 413         default:
 414                 return U32_MAX;
 415         }
 416 }
 417
 418 static struct nbd_config *nbd_get_config_unlocked(struct nbd_device *nbd)
 419 {
 420         if (refcount_inc_not_zero(&nbd->config_refs)) {
 421                 /*
 422                  * Add smp_mb__after_atomic to ensure that reading nbd->config_refs
 423                  * and reading nbd->config is ordered. The pair is the barrier in
 424                  * nbd_alloc_and_init_config(), avoid nbd->config_refs is set
 425                  * before nbd->config.
 426                  */
 427                 smp_mb__after_atomic();
 428                 return nbd->config;
 429         }
 430
 431         return NULL;
 432 }
 433
 434 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req)
 435 {
 436         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 437         struct nbd_device *nbd = cmd->nbd;
 438         struct nbd_config *config;
 439
 440         if (!mutex_trylock(&cmd->lock))
 441                 return BLK_EH_RESET_TIMER;
 442
 443         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 444                 mutex_unlock(&cmd->lock);
 445                 return BLK_EH_DONE;
 446         }
 447
 448         config = nbd_get_config_unlocked(nbd);
 449         if (!config) {
 450                 cmd->status = BLK_STS_TIMEOUT;
 451                 __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 452                 mutex_unlock(&cmd->lock);
 453                 goto done;
 454         }
 455
 456         if (config->num_connections > 1 ||
 457             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 458                 dev_err_ratelimited(nbd_to_dev(nbd),
 459                                     "Connection timed out, retrying (%d/%d alive)\n",
 460                                     atomic_read(&config->live_connections),
 461                                     config->num_connections);
 462                 /*
 463                  * Hooray we have more connections, requeue this IO, the submit
 464                  * path will put it on a real connection. Or if only one
 465                  * connection is configured, the submit path will wait util
 466                  * a new connection is reconfigured or util dead timeout.
 467                  */
 468                 if (config->socks) {
 469                         if (cmd->index < config->num_connections) {
 470                                 struct nbd_sock *nsock =
 471                                         config->socks[cmd->index];
 472                                 mutex_lock(&nsock->tx_lock);
 473                                 /* We can have multiple outstanding requests, so
 474                                  * we don't want to mark the nsock dead if we've
 475                                  * already reconnected with a new socket, so
 476                                  * only mark it dead if its the same socket we
 477                                  * were sent out on.
 478                                  */
 479                                 if (cmd->cookie == nsock->cookie)
 480                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 481                                 mutex_unlock(&nsock->tx_lock);
 482                         }
 483                         mutex_unlock(&cmd->lock);
 484                         nbd_requeue_cmd(cmd);
 485                         nbd_config_put(nbd);
 486                         return BLK_EH_DONE;
 487                 }
 488         }
 489
 490         if (!nbd->tag_set.timeout) {
 491                 /*
 492                  * Userspace sets timeout=0 to disable socket disconnection,
 493                  * so just warn and reset the timer.
 494                  */
 495                 struct nbd_sock *nsock = config->socks[cmd->index];
 496                 cmd->retries++;
 497                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 498                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 499                         (unsigned long long)blk_rq_pos(req) << 9,
 500                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 501
 502                 mutex_lock(&nsock->tx_lock);
 503                 if (cmd->cookie != nsock->cookie) {
 504                         nbd_requeue_cmd(cmd);
 505                         mutex_unlock(&nsock->tx_lock);
 506                         mutex_unlock(&cmd->lock);
 507                         nbd_config_put(nbd);
 508                         return BLK_EH_DONE;
 509                 }
 510                 mutex_unlock(&nsock->tx_lock);
 511                 mutex_unlock(&cmd->lock);
 512                 nbd_config_put(nbd);
 513                 return BLK_EH_RESET_TIMER;
 514         }
 515
 516         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 517         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 518         cmd->status = BLK_STS_IOERR;
 519         __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 520         mutex_unlock(&cmd->lock);
 521         sock_shutdown(nbd);
 522         nbd_config_put(nbd);
 523 done:
 524         blk_mq_complete_request(req);
 525         return BLK_EH_DONE;
 526 }
 527
 528 static int __sock_xmit(struct nbd_device *nbd, struct socket *sock, int send,
 529                        struct iov_iter *iter, int msg_flags, int *sent)
 530 {
 531         int result;
 532         struct msghdr msg = {} ;
 533         unsigned int noreclaim_flag;
 534
 535         if (unlikely(!sock)) {
 536                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 537                         "Attempted %s on closed socket in sock_xmit\n",
 538                         (send ? "send" : "recv"));
 539                 return -EINVAL;
 540         }
 541
 542         msg.msg_iter = *iter;
 543
 544         noreclaim_flag = memalloc_noreclaim_save();
 545         do {
 546                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 547                 sock->sk->sk_use_task_frag = false;
 548                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 549
 550                 if (send)
 551                         result = sock_sendmsg(sock, &msg);
 552                 else
 553                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 554
 555                 if (result <= 0) {
 556                         if (result == 0)
 557                                 result = -EPIPE; /* short read */
 558                         break;
 559                 }
 560                 if (sent)
 561                         *sent += result;
 562         } while (msg_data_left(&msg));
 563
 564         memalloc_noreclaim_restore(noreclaim_flag);
 565
 566         return result;
 567 }
 568
 569 /*
 570  *  Send or receive packet. Return a positive value on success and
 571  *  negtive value on failure, and never return 0.
 572  */
 573 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 574                      struct iov_iter *iter, int msg_flags, int *sent)
 575 {
 576         struct nbd_config *config = nbd->config;
 577         struct socket *sock = config->socks[index]->sock;
 578
 579         return __sock_xmit(nbd, sock, send, iter, msg_flags, sent);
 580 }
 581
 582 /*
 583  * Different settings for sk->sk_sndtimeo can result in different return values
 584  * if there is a signal pending when we enter sendmsg, because reasons?
 585  */
 586 static inline int was_interrupted(int result)
 587 {
 588         return result == -ERESTARTSYS || result == -EINTR;
 589 }
 590
 591 /* always call with the tx_lock held */
 592 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 593 {
 594         struct request *req = blk_mq_rq_from_pdu(cmd);
 595         struct nbd_config *config = nbd->config;
 596         struct nbd_sock *nsock = config->socks[index];
 597         int result;
 598         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 599         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 600         struct iov_iter from;
 601         unsigned long size = blk_rq_bytes(req);
 602         struct bio *bio;
 603         u64 handle;
 604         u32 type;
 605         u32 nbd_cmd_flags = 0;
 606         int sent = nsock->sent, skip = 0;
 607
 608         iov_iter_kvec(&from, ITER_SOURCE, &iov, 1, sizeof(request));
 609
 610         type = req_to_nbd_cmd_type(req);
 611         if (type == U32_MAX)
 612                 return -EIO;
 613
 614         if (rq_data_dir(req) == WRITE &&
 615             (config->flags & NBD_FLAG_READ_ONLY)) {
 616                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 617                                     "Write on read-only\n");
 618                 return -EIO;
 619         }
 620
 621         if (req->cmd_flags & REQ_FUA)
 622                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 623
 624         /* We did a partial send previously, and we at least sent the whole
 625          * request struct, so just go and send the rest of the pages in the
 626          * request.
 627          */
 628         if (sent) {
 629                 if (sent >= sizeof(request)) {
 630                         skip = sent - sizeof(request);
 631
 632                         /* initialize handle for tracing purposes */
 633                         handle = nbd_cmd_handle(cmd);
 634
 635                         goto send_pages;
 636                 }
 637                 iov_iter_advance(&from, sent);
 638         } else {
 639                 cmd->cmd_cookie++;
 640         }
 641         cmd->index = index;
 642         cmd->cookie = nsock->cookie;
 643         cmd->retries = 0;
 644         request.type = htonl(type | nbd_cmd_flags);
 645         if (type != NBD_CMD_FLUSH) {
 646                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 647                 request.len = htonl(size);
 648         }
 649         handle = nbd_cmd_handle(cmd);
 650         request.cookie = cpu_to_be64(handle);
 651
 652         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 653
 654         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 655                 req, nbdcmd_to_ascii(type),
 656                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 657         result = sock_xmit(nbd, index, 1, &from,
 658                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 659         trace_nbd_header_sent(req, handle);
 660         if (result < 0) {
 661                 if (was_interrupted(result)) {
 662                         /* If we haven't sent anything we can just return BUSY,
 663                          * however if we have sent something we need to make
 664                          * sure we only allow this req to be sent until we are
 665                          * completely done.
 666                          */
 667                         if (sent) {
 668                                 nsock->pending = req;
 669                                 nsock->sent = sent;
 670                         }
 671                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 672                         return BLK_STS_RESOURCE;
 673                 }
 674                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 675                         "Send control failed (result %d)\n", result);
 676                 return -EAGAIN;
 677         }
 678 send_pages:
 679         if (type != NBD_CMD_WRITE)
 680                 goto out;
 681
 682         bio = req->bio;
 683         while (bio) {
 684                 struct bio *next = bio->bi_next;
 685                 struct bvec_iter iter;
 686                 struct bio_vec bvec;
 687
 688                 bio_for_each_segment(bvec, bio, iter) {
 689                         bool is_last = !next && bio_iter_last(bvec, iter);
 690                         int flags = is_last ? 0 : MSG_MORE;
 691
 692                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 693                                 req, bvec.bv_len);
 694                         iov_iter_bvec(&from, ITER_SOURCE, &bvec, 1, bvec.bv_len);
 695                         if (skip) {
 696                                 if (skip >= iov_iter_count(&from)) {
 697                                         skip -= iov_iter_count(&from);
 698                                         continue;
 699                                 }
 700                                 iov_iter_advance(&from, skip);
 701                                 skip = 0;
 702                         }
 703                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 704                         if (result < 0) {
 705                                 if (was_interrupted(result)) {
 706                                         /* We've already sent the header, we
 707                                          * have no choice but to set pending and
 708                                          * return BUSY.
 709                                          */
 710                                         nsock->pending = req;
 711                                         nsock->sent = sent;
 712                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 713                                         return BLK_STS_RESOURCE;
 714                                 }
 715                                 dev_err(disk_to_dev(nbd->disk),
 716                                         "Send data failed (result %d)\n",
 717                                         result);
 718                                 return -EAGAIN;
 719                         }
 720                         /*
 721                          * The completion might already have come in,
 722                          * so break for the last one instead of letting
 723                          * the iterator do it. This prevents use-after-free
 724                          * of the bio.
 725                          */
 726                         if (is_last)
 727                                 break;
 728                 }
 729                 bio = next;
 730         }
 731 out:
 732         trace_nbd_payload_sent(req, handle);
 733         nsock->pending = NULL;
 734         nsock->sent = 0;
 735         return 0;
 736 }
 737
 738 static int nbd_read_reply(struct nbd_device *nbd, struct socket *sock,
 739                           struct nbd_reply *reply)
 740 {
 741         struct kvec iov = {.iov_base = reply, .iov_len = sizeof(*reply)};
 742         struct iov_iter to;
 743         int result;
 744
 745         reply->magic = 0;
 746         iov_iter_kvec(&to, ITER_DEST, &iov, 1, sizeof(*reply));
 747         result = __sock_xmit(nbd, sock, 0, &to, MSG_WAITALL, NULL);
 748         if (result < 0) {
 749                 if (!nbd_disconnected(nbd->config))
 750                         dev_err(disk_to_dev(nbd->disk),
 751                                 "Receive control failed (result %d)\n", result);
 752                 return result;
 753         }
 754
 755         if (ntohl(reply->magic) != NBD_REPLY_MAGIC) {
 756                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 757                                 (unsigned long)ntohl(reply->magic));
 758                 return -EPROTO;
 759         }
 760
 761         return 0;
 762 }
 763
 764 /* NULL returned = something went wrong, inform userspace */
 765 static struct nbd_cmd *nbd_handle_reply(struct nbd_device *nbd, int index,
 766                                         struct nbd_reply *reply)
 767 {
 768         int result;
 769         struct nbd_cmd *cmd;
 770         struct request *req = NULL;
 771         u64 handle;
 772         u16 hwq;
 773         u32 tag;
 774         int ret = 0;
 775
 776         handle = be64_to_cpu(reply->cookie);
 777         tag = nbd_handle_to_tag(handle);
 778         hwq = blk_mq_unique_tag_to_hwq(tag);
 779         if (hwq < nbd->tag_set.nr_hw_queues)
 780                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 781                                        blk_mq_unique_tag_to_tag(tag));
 782         if (!req || !blk_mq_request_started(req)) {
 783                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 784                         tag, req);
 785                 return ERR_PTR(-ENOENT);
 786         }
 787         trace_nbd_header_received(req, handle);
 788         cmd = blk_mq_rq_to_pdu(req);
 789
 790         mutex_lock(&cmd->lock);
 791         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 792                 dev_err(disk_to_dev(nbd->disk), "Suspicious reply %d (status %u flags %lu)",
 793                         tag, cmd->status, cmd->flags);
 794                 ret = -ENOENT;
 795                 goto out;
 796         }
 797         if (cmd->index != index) {
 798                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply %d from different sock %d (expected %d)",
 799                         tag, index, cmd->index);
 800                 ret = -ENOENT;
 801                 goto out;
 802         }
 803         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 804                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 805                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 806                 ret = -ENOENT;
 807                 goto out;
 808         }
 809         if (cmd->status != BLK_STS_OK) {
 810                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 811                         req);
 812                 ret = -ENOENT;
 813                 goto out;
 814         }
 815         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 816                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 817                         req);
 818                 ret = -ENOENT;
 819                 goto out;
 820         }
 821         if (ntohl(reply->error)) {
 822                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 823                         ntohl(reply->error));
 824                 cmd->status = BLK_STS_IOERR;
 825                 goto out;
 826         }
 827
 828         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 829         if (rq_data_dir(req) != WRITE) {
 830                 struct req_iterator iter;
 831                 struct bio_vec bvec;
 832                 struct iov_iter to;
 833
 834                 rq_for_each_segment(bvec, req, iter) {
 835                         iov_iter_bvec(&to, ITER_DEST, &bvec, 1, bvec.bv_len);
 836                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 837                         if (result < 0) {
 838                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 839                                         result);
 840                                 /*
 841                                  * If we've disconnected, we need to make sure we
 842                                  * complete this request, otherwise error out
 843                                  * and let the timeout stuff handle resubmitting
 844                                  * this request onto another connection.
 845                                  */
 846                                 if (nbd_disconnected(nbd->config)) {
 847                                         cmd->status = BLK_STS_IOERR;
 848                                         goto out;
 849                                 }
 850                                 ret = -EIO;
 851                                 goto out;
 852                         }
 853                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 854                                 req, bvec.bv_len);
 855                 }
 856         }
 857 out:
 858         trace_nbd_payload_received(req, handle);
 859         mutex_unlock(&cmd->lock);
 860         return ret ? ERR_PTR(ret) : cmd;
 861 }
 862
 863 static void recv_work(struct work_struct *work)
 864 {
 865         struct recv_thread_args *args = container_of(work,
 866                                                      struct recv_thread_args,
 867                                                      work);
 868         struct nbd_device *nbd = args->nbd;
 869         struct nbd_config *config = nbd->config;
 870         struct request_queue *q = nbd->disk->queue;
 871         struct nbd_sock *nsock = args->nsock;
 872         struct nbd_cmd *cmd;
 873         struct request *rq;
 874
 875         while (1) {
 876                 struct nbd_reply reply;
 877
 878                 if (nbd_read_reply(nbd, nsock->sock, &reply))
 879                         break;
 880
 881                 /*
 882                  * Grab .q_usage_counter so request pool won't go away, then no
 883                  * request use-after-free is possible during nbd_handle_reply().
 884                  * If queue is frozen, there won't be any inflight requests, we
 885                  * needn't to handle the incoming garbage message.
 886                  */
 887                 if (!percpu_ref_tryget(&q->q_usage_counter)) {
 888                         dev_err(disk_to_dev(nbd->disk), "%s: no io inflight\n",
 889                                 __func__);
 890                         break;
 891                 }
 892
 893                 cmd = nbd_handle_reply(nbd, args->index, &reply);
 894                 if (IS_ERR(cmd)) {
 895                         percpu_ref_put(&q->q_usage_counter);
 896                         break;
 897                 }
 898
 899                 rq = blk_mq_rq_from_pdu(cmd);
 900                 if (likely(!blk_should_fake_timeout(rq->q))) {
 901                         bool complete;
 902
 903                         mutex_lock(&cmd->lock);
 904                         complete = __test_and_clear_bit(NBD_CMD_INFLIGHT,
 905                                                         &cmd->flags);
 906                         mutex_unlock(&cmd->lock);
 907                         if (complete)
 908                                 blk_mq_complete_request(rq);
 909                 }
 910                 percpu_ref_put(&q->q_usage_counter);
 911         }
 912
 913         mutex_lock(&nsock->tx_lock);
 914         nbd_mark_nsock_dead(nbd, nsock, 1);
 915         mutex_unlock(&nsock->tx_lock);
 916
 917         nbd_config_put(nbd);
 918         atomic_dec(&config->recv_threads);
 919         wake_up(&config->recv_wq);
 920         kfree(args);
 921 }
 922
 923 static bool nbd_clear_req(struct request *req, void *data)
 924 {
 925         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 926
 927         /* don't abort one completed request */
 928         if (blk_mq_request_completed(req))
 929                 return true;
 930
 931         mutex_lock(&cmd->lock);
 932         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 933                 mutex_unlock(&cmd->lock);
 934                 return true;
 935         }
 936         cmd->status = BLK_STS_IOERR;
 937         mutex_unlock(&cmd->lock);
 938
 939         blk_mq_complete_request(req);
 940         return true;
 941 }
 942
 943 static void nbd_clear_que(struct nbd_device *nbd)
 944 {
 945         blk_mq_quiesce_queue(nbd->disk->queue);
 946         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 947         blk_mq_unquiesce_queue(nbd->disk->queue);
 948         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 949 }
 950
 951 static int find_fallback(struct nbd_device *nbd, int index)
 952 {
 953         struct nbd_config *config = nbd->config;
 954         int new_index = -1;
 955         struct nbd_sock *nsock = config->socks[index];
 956         int fallback = nsock->fallback_index;
 957
 958         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 959                 return new_index;
 960
 961         if (config->num_connections <= 1) {
 962                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 963                                     "Dead connection, failed to find a fallback\n");
 964                 return new_index;
 965         }
 966
 967         if (fallback >= 0 && fallback < config->num_connections &&
 968             !config->socks[fallback]->dead)
 969                 return fallback;
 970
 971         if (nsock->fallback_index < 0 ||
 972             nsock->fallback_index >= config->num_connections ||
 973             config->socks[nsock->fallback_index]->dead) {
 974                 int i;
 975                 for (i = 0; i < config->num_connections; i++) {
 976                         if (i == index)
 977                                 continue;
 978                         if (!config->socks[i]->dead) {
 979                                 new_index = i;
 980                                 break;
 981                         }
 982                 }
 983                 nsock->fallback_index = new_index;
 984                 if (new_index < 0) {
 985                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 986                                             "Dead connection, failed to find a fallback\n");
 987                         return new_index;
 988                 }
 989         }
 990         new_index = nsock->fallback_index;
 991         return new_index;
 992 }
 993
 994 static int wait_for_reconnect(struct nbd_device *nbd)
 995 {
 996         struct nbd_config *config = nbd->config;
 997         if (!config->dead_conn_timeout)
 998                 return 0;
 999
1000         if (!wait_event_timeout(config->conn_wait,
1001                                 test_bit(NBD_RT_DISCONNECTED,
1002                                          &config->runtime_flags) ||
1003                                 atomic_read(&config->live_connections) > 0,
1004                                 config->dead_conn_timeout))
1005                 return 0;
1006
1007         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1008 }
1009
1010 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
1011 {
1012         struct request *req = blk_mq_rq_from_pdu(cmd);
1013         struct nbd_device *nbd = cmd->nbd;
1014         struct nbd_config *config;
1015         struct nbd_sock *nsock;
1016         int ret;
1017
1018         config = nbd_get_config_unlocked(nbd);
1019         if (!config) {
1020                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1021                                     "Socks array is empty\n");
1022                 return -EINVAL;
1023         }
1024
1025         if (index >= config->num_connections) {
1026                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1027                                     "Attempted send on invalid socket\n");
1028                 nbd_config_put(nbd);
1029                 return -EINVAL;
1030         }
1031         cmd->status = BLK_STS_OK;
1032 again:
1033         nsock = config->socks[index];
1034         mutex_lock(&nsock->tx_lock);
1035         if (nsock->dead) {
1036                 int old_index = index;
1037                 index = find_fallback(nbd, index);
1038                 mutex_unlock(&nsock->tx_lock);
1039                 if (index < 0) {
1040                         if (wait_for_reconnect(nbd)) {
1041                                 index = old_index;
1042                                 goto again;
1043                         }
1044                         /* All the sockets should already be down at this point,
1045                          * we just want to make sure that DISCONNECTED is set so
1046                          * any requests that come in that were queue'ed waiting
1047                          * for the reconnect timer don't trigger the timer again
1048                          * and instead just error out.
1049                          */
1050                         sock_shutdown(nbd);
1051                         nbd_config_put(nbd);
1052                         return -EIO;
1053                 }
1054                 goto again;
1055         }
1056
1057         /* Handle the case that we have a pending request that was partially
1058          * transmitted that _has_ to be serviced first.  We need to call requeue
1059          * here so that it gets put _after_ the request that is already on the
1060          * dispatch list.
1061          */
1062         blk_mq_start_request(req);
1063         if (unlikely(nsock->pending && nsock->pending != req)) {
1064                 nbd_requeue_cmd(cmd);
1065                 ret = 0;
1066                 goto out;
1067         }
1068         /*
1069          * Some failures are related to the link going down, so anything that
1070          * returns EAGAIN can be retried on a different socket.
1071          */
1072         ret = nbd_send_cmd(nbd, cmd, index);
1073         /*
1074          * Access to this flag is protected by cmd->lock, thus it's safe to set
1075          * the flag after nbd_send_cmd() succeed to send request to server.
1076          */
1077         if (!ret)
1078                 __set_bit(NBD_CMD_INFLIGHT, &cmd->flags);
1079         else if (ret == -EAGAIN) {
1080                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1081                                     "Request send failed, requeueing\n");
1082                 nbd_mark_nsock_dead(nbd, nsock, 1);
1083                 nbd_requeue_cmd(cmd);
1084                 ret = 0;
1085         }
1086 out:
1087         mutex_unlock(&nsock->tx_lock);
1088         nbd_config_put(nbd);
1089         return ret;
1090 }
1091
1092 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1093                         const struct blk_mq_queue_data *bd)
1094 {
1095         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1096         int ret;
1097
1098         /*
1099          * Since we look at the bio's to send the request over the network we
1100          * need to make sure the completion work doesn't mark this request done
1101          * before we are done doing our send.  This keeps us from dereferencing
1102          * freed data if we have particularly fast completions (ie we get the
1103          * completion before we exit sock_xmit on the last bvec) or in the case
1104          * that the server is misbehaving (or there was an error) before we're
1105          * done sending everything over the wire.
1106          */
1107         mutex_lock(&cmd->lock);
1108         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1109
1110         /* We can be called directly from the user space process, which means we
1111          * could possibly have signals pending so our sendmsg will fail.  In
1112          * this case we need to return that we are busy, otherwise error out as
1113          * appropriate.
1114          */
1115         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1116         if (ret < 0)
1117                 ret = BLK_STS_IOERR;
1118         else if (!ret)
1119                 ret = BLK_STS_OK;
1120         mutex_unlock(&cmd->lock);
1121
1122         return ret;
1123 }
1124
1125 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1126                                      int *err)
1127 {
1128         struct socket *sock;
1129
1130         *err = 0;
1131         sock = sockfd_lookup(fd, err);
1132         if (!sock)
1133                 return NULL;
1134
1135         if (sock->ops->shutdown == sock_no_shutdown) {
1136                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1137                 *err = -EINVAL;
1138                 sockfd_put(sock);
1139                 return NULL;
1140         }
1141
1142         return sock;
1143 }
1144
1145 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1146                           bool netlink)
1147 {
1148         struct nbd_config *config = nbd->config;
1149         struct socket *sock;
1150         struct nbd_sock **socks;
1151         struct nbd_sock *nsock;
1152         int err;
1153
1154         /* Arg will be cast to int, check it to avoid overflow */
1155         if (arg > INT_MAX)
1156                 return -EINVAL;
1157         sock = nbd_get_socket(nbd, arg, &err);
1158         if (!sock)
1159                 return err;
1160
1161         /*
1162          * We need to make sure we don't get any errant requests while we're
1163          * reallocating the ->socks array.
1164          */
1165         blk_mq_freeze_queue(nbd->disk->queue);
1166
1167         if (!netlink && !nbd->task_setup &&
1168             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1169                 nbd->task_setup = current;
1170
1171         if (!netlink &&
1172             (nbd->task_setup != current ||
1173              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1174                 dev_err(disk_to_dev(nbd->disk),
1175                         "Device being setup by another task");
1176                 err = -EBUSY;
1177                 goto put_socket;
1178         }
1179
1180         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1181         if (!nsock) {
1182                 err = -ENOMEM;
1183                 goto put_socket;
1184         }
1185
1186         socks = krealloc(config->socks, (config->num_connections + 1) *
1187                          sizeof(struct nbd_sock *), GFP_KERNEL);
1188         if (!socks) {
1189                 kfree(nsock);
1190                 err = -ENOMEM;
1191                 goto put_socket;
1192         }
1193
1194         config->socks = socks;
1195
1196         nsock->fallback_index = -1;
1197         nsock->dead = false;
1198         mutex_init(&nsock->tx_lock);
1199         nsock->sock = sock;
1200         nsock->pending = NULL;
1201         nsock->sent = 0;
1202         nsock->cookie = 0;
1203         socks[config->num_connections++] = nsock;
1204         atomic_inc(&config->live_connections);
1205         blk_mq_unfreeze_queue(nbd->disk->queue);
1206
1207         return 0;
1208
1209 put_socket:
1210         blk_mq_unfreeze_queue(nbd->disk->queue);
1211         sockfd_put(sock);
1212         return err;
1213 }
1214
1215 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1216 {
1217         struct nbd_config *config = nbd->config;
1218         struct socket *sock, *old;
1219         struct recv_thread_args *args;
1220         int i;
1221         int err;
1222
1223         sock = nbd_get_socket(nbd, arg, &err);
1224         if (!sock)
1225                 return err;
1226
1227         args = kzalloc(sizeof(*args), GFP_KERNEL);
1228         if (!args) {
1229                 sockfd_put(sock);
1230                 return -ENOMEM;
1231         }
1232
1233         for (i = 0; i < config->num_connections; i++) {
1234                 struct nbd_sock *nsock = config->socks[i];
1235
1236                 if (!nsock->dead)
1237                         continue;
1238
1239                 mutex_lock(&nsock->tx_lock);
1240                 if (!nsock->dead) {
1241                         mutex_unlock(&nsock->tx_lock);
1242                         continue;
1243                 }
1244                 sk_set_memalloc(sock->sk);
1245                 if (nbd->tag_set.timeout)
1246                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1247                 atomic_inc(&config->recv_threads);
1248                 refcount_inc(&nbd->config_refs);
1249                 old = nsock->sock;
1250                 nsock->fallback_index = -1;
1251                 nsock->sock = sock;
1252                 nsock->dead = false;
1253                 INIT_WORK(&args->work, recv_work);
1254                 args->index = i;
1255                 args->nbd = nbd;
1256                 args->nsock = nsock;
1257                 nsock->cookie++;
1258                 mutex_unlock(&nsock->tx_lock);
1259                 sockfd_put(old);
1260
1261                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1262
1263                 /* We take the tx_mutex in an error path in the recv_work, so we
1264                  * need to queue_work outside of the tx_mutex.
1265                  */
1266                 queue_work(nbd->recv_workq, &args->work);
1267
1268                 atomic_inc(&config->live_connections);
1269                 wake_up(&config->conn_wait);
1270                 return 0;
1271         }
1272         sockfd_put(sock);
1273         kfree(args);
1274         return -ENOSPC;
1275 }
1276
1277 static void nbd_bdev_reset(struct nbd_device *nbd)
1278 {
1279         if (disk_openers(nbd->disk) > 1)
1280                 return;
1281         set_capacity(nbd->disk, 0);
1282 }
1283
1284 static void nbd_parse_flags(struct nbd_device *nbd)
1285 {
1286         struct nbd_config *config = nbd->config;
1287         if (config->flags & NBD_FLAG_READ_ONLY)
1288                 set_disk_ro(nbd->disk, true);
1289         else
1290                 set_disk_ro(nbd->disk, false);
1291         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1292                 if (config->flags & NBD_FLAG_SEND_FUA)
1293                         blk_queue_write_cache(nbd->disk->queue, true, true);
1294                 else
1295                         blk_queue_write_cache(nbd->disk->queue, true, false);
1296         }
1297         else
1298                 blk_queue_write_cache(nbd->disk->queue, false, false);
1299 }
1300
1301 static void send_disconnects(struct nbd_device *nbd)
1302 {
1303         struct nbd_config *config = nbd->config;
1304         struct nbd_request request = {
1305                 .magic = htonl(NBD_REQUEST_MAGIC),
1306                 .type = htonl(NBD_CMD_DISC),
1307         };
1308         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1309         struct iov_iter from;
1310         int i, ret;
1311
1312         for (i = 0; i < config->num_connections; i++) {
1313                 struct nbd_sock *nsock = config->socks[i];
1314
1315                 iov_iter_kvec(&from, ITER_SOURCE, &iov, 1, sizeof(request));
1316                 mutex_lock(&nsock->tx_lock);
1317                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1318                 if (ret < 0)
1319                         dev_err(disk_to_dev(nbd->disk),
1320                                 "Send disconnect failed %d\n", ret);
1321                 mutex_unlock(&nsock->tx_lock);
1322         }
1323 }
1324
1325 static int nbd_disconnect(struct nbd_device *nbd)
1326 {
1327         struct nbd_config *config = nbd->config;
1328
1329         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1330         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1331         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1332         send_disconnects(nbd);
1333         return 0;
1334 }
1335
1336 static void nbd_clear_sock(struct nbd_device *nbd)
1337 {
1338         sock_shutdown(nbd);
1339         nbd_clear_que(nbd);
1340         nbd->task_setup = NULL;
1341 }
1342
1343 static void nbd_config_put(struct nbd_device *nbd)
1344 {
1345         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1346                                         &nbd->config_lock)) {
1347                 struct nbd_config *config = nbd->config;
1348                 nbd_dev_dbg_close(nbd);
1349                 invalidate_disk(nbd->disk);
1350                 if (nbd->config->bytesize)
1351                         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
1352                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1353                                        &config->runtime_flags))
1354                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1355                 nbd->pid = 0;
1356                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1357                                        &config->runtime_flags)) {
1358                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1359                         kfree(nbd->backend);
1360                         nbd->backend = NULL;
1361                 }
1362                 nbd_clear_sock(nbd);
1363                 if (config->num_connections) {
1364                         int i;
1365                         for (i = 0; i < config->num_connections; i++) {
1366                                 sockfd_put(config->socks[i]->sock);
1367                                 kfree(config->socks[i]);
1368                         }
1369                         kfree(config->socks);
1370                 }
1371                 kfree(nbd->config);
1372                 nbd->config = NULL;
1373
1374                 nbd->tag_set.timeout = 0;
1375
1376                 mutex_unlock(&nbd->config_lock);
1377                 nbd_put(nbd);
1378                 module_put(THIS_MODULE);
1379         }
1380 }
1381
1382 static int nbd_start_device(struct nbd_device *nbd)
1383 {
1384         struct nbd_config *config = nbd->config;
1385         int num_connections = config->num_connections;
1386         int error = 0, i;
1387
1388         if (nbd->pid)
1389                 return -EBUSY;
1390         if (!config->socks)
1391                 return -EINVAL;
1392         if (num_connections > 1 &&
1393             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1394                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1395                 return -EINVAL;
1396         }
1397
1398         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1399         nbd->pid = task_pid_nr(current);
1400
1401         nbd_parse_flags(nbd);
1402
1403         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1404         if (error) {
1405                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1406                 return error;
1407         }
1408         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1409
1410         nbd_dev_dbg_init(nbd);
1411         for (i = 0; i < num_connections; i++) {
1412                 struct recv_thread_args *args;
1413
1414                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1415                 if (!args) {
1416                         sock_shutdown(nbd);
1417                         /*
1418                          * If num_connections is m (2 < m),
1419                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1420                          * But NO.(n + 1) failed. We still have n recv threads.
1421                          * So, add flush_workqueue here to prevent recv threads
1422                          * dropping the last config_refs and trying to destroy
1423                          * the workqueue from inside the workqueue.
1424                          */
1425                         if (i)
1426                                 flush_workqueue(nbd->recv_workq);
1427                         return -ENOMEM;
1428                 }
1429                 sk_set_memalloc(config->socks[i]->sock->sk);
1430                 if (nbd->tag_set.timeout)
1431                         config->socks[i]->sock->sk->sk_sndtimeo =
1432                                 nbd->tag_set.timeout;
1433                 atomic_inc(&config->recv_threads);
1434                 refcount_inc(&nbd->config_refs);
1435                 INIT_WORK(&args->work, recv_work);
1436                 args->nbd = nbd;
1437                 args->nsock = config->socks[i];
1438                 args->index = i;
1439                 queue_work(nbd->recv_workq, &args->work);
1440         }
1441         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1442 }
1443
1444 static int nbd_start_device_ioctl(struct nbd_device *nbd)
1445 {
1446         struct nbd_config *config = nbd->config;
1447         int ret;
1448
1449         ret = nbd_start_device(nbd);
1450         if (ret)
1451                 return ret;
1452
1453         if (max_part)
1454                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1455         mutex_unlock(&nbd->config_lock);
1456         ret = wait_event_interruptible(config->recv_wq,
1457                                          atomic_read(&config->recv_threads) == 0);
1458         if (ret) {
1459                 sock_shutdown(nbd);
1460                 nbd_clear_que(nbd);
1461         }
1462
1463         flush_workqueue(nbd->recv_workq);
1464         mutex_lock(&nbd->config_lock);
1465         nbd_bdev_reset(nbd);
1466         /* user requested, ignore socket errors */
1467         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1468                 ret = 0;
1469         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1470                 ret = -ETIMEDOUT;
1471         return ret;
1472 }
1473
1474 static void nbd_clear_sock_ioctl(struct nbd_device *nbd)
1475 {
1476         nbd_clear_sock(nbd);
1477         disk_force_media_change(nbd->disk);
1478         nbd_bdev_reset(nbd);
1479         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1480                                &nbd->config->runtime_flags))
1481                 nbd_config_put(nbd);
1482 }
1483
1484 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1485 {
1486         nbd->tag_set.timeout = timeout * HZ;
1487         if (timeout)
1488                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1489         else
1490                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1491 }
1492
1493 /* Must be called with config_lock held */
1494 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1495                        unsigned int cmd, unsigned long arg)
1496 {
1497         struct nbd_config *config = nbd->config;
1498         loff_t bytesize;
1499
1500         switch (cmd) {
1501         case NBD_DISCONNECT:
1502                 return nbd_disconnect(nbd);
1503         case NBD_CLEAR_SOCK:
1504                 nbd_clear_sock_ioctl(nbd);
1505                 return 0;
1506         case NBD_SET_SOCK:
1507                 return nbd_add_socket(nbd, arg, false);
1508         case NBD_SET_BLKSIZE:
1509                 return nbd_set_size(nbd, config->bytesize, arg);
1510         case NBD_SET_SIZE:
1511                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1512         case NBD_SET_SIZE_BLOCKS:
1513                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1514                         return -EINVAL;
1515                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1516         case NBD_SET_TIMEOUT:
1517                 nbd_set_cmd_timeout(nbd, arg);
1518                 return 0;
1519
1520         case NBD_SET_FLAGS:
1521                 config->flags = arg;
1522                 return 0;
1523         case NBD_DO_IT:
1524                 return nbd_start_device_ioctl(nbd);
1525         case NBD_CLEAR_QUE:
1526                 /*
1527                  * This is for compatibility only.  The queue is always cleared
1528                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1529                  */
1530                 return 0;
1531         case NBD_PRINT_DEBUG:
1532                 /*
1533                  * For compatibility only, we no longer keep a list of
1534                  * outstanding requests.
1535                  */
1536                 return 0;
1537         }
1538         return -ENOTTY;
1539 }
1540
1541 static int nbd_ioctl(struct block_device *bdev, blk_mode_t mode,
1542                      unsigned int cmd, unsigned long arg)
1543 {
1544         struct nbd_device *nbd = bdev->bd_disk->private_data;
1545         struct nbd_config *config = nbd->config;
1546         int error = -EINVAL;
1547
1548         if (!capable(CAP_SYS_ADMIN))
1549                 return -EPERM;
1550
1551         /* The block layer will pass back some non-nbd ioctls in case we have
1552          * special handling for them, but we don't so just return an error.
1553          */
1554         if (_IOC_TYPE(cmd) != 0xab)
1555                 return -EINVAL;
1556
1557         mutex_lock(&nbd->config_lock);
1558
1559         /* Don't allow ioctl operations on a nbd device that was created with
1560          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1561          */
1562         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1563             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1564                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1565         else
1566                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1567         mutex_unlock(&nbd->config_lock);
1568         return error;
1569 }
1570
1571 static int nbd_alloc_and_init_config(struct nbd_device *nbd)
1572 {
1573         struct nbd_config *config;
1574
1575         if (WARN_ON(nbd->config))
1576                 return -EINVAL;
1577
1578         if (!try_module_get(THIS_MODULE))
1579                 return -ENODEV;
1580
1581         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1582         if (!config) {
1583                 module_put(THIS_MODULE);
1584                 return -ENOMEM;
1585         }
1586
1587         atomic_set(&config->recv_threads, 0);
1588         init_waitqueue_head(&config->recv_wq);
1589         init_waitqueue_head(&config->conn_wait);
1590         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1591         atomic_set(&config->live_connections, 0);
1592
1593         nbd->config = config;
1594         /*
1595          * Order refcount_set(&nbd->config_refs, 1) and nbd->config assignment,
1596          * its pair is the barrier in nbd_get_config_unlocked().
1597          * So nbd_get_config_unlocked() won't see nbd->config as null after
1598          * refcount_inc_not_zero() succeed.
1599          */
1600         smp_mb__before_atomic();
1601         refcount_set(&nbd->config_refs, 1);
1602
1603         return 0;
1604 }
1605
1606 static int nbd_open(struct gendisk *disk, blk_mode_t mode)
1607 {
1608         struct nbd_device *nbd;
1609         struct nbd_config *config;
1610         int ret = 0;
1611
1612         mutex_lock(&nbd_index_mutex);
1613         nbd = disk->private_data;
1614         if (!nbd) {
1615                 ret = -ENXIO;
1616                 goto out;
1617         }
1618         if (!refcount_inc_not_zero(&nbd->refs)) {
1619                 ret = -ENXIO;
1620                 goto out;
1621         }
1622
1623         config = nbd_get_config_unlocked(nbd);
1624         if (!config) {
1625                 mutex_lock(&nbd->config_lock);
1626                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1627                         mutex_unlock(&nbd->config_lock);
1628                         goto out;
1629                 }
1630                 ret = nbd_alloc_and_init_config(nbd);
1631                 if (ret) {
1632                         mutex_unlock(&nbd->config_lock);
1633                         goto out;
1634                 }
1635
1636                 refcount_inc(&nbd->refs);
1637                 mutex_unlock(&nbd->config_lock);
1638                 if (max_part)
1639                         set_bit(GD_NEED_PART_SCAN, &disk->state);
1640         } else if (nbd_disconnected(config)) {
1641                 if (max_part)
1642                         set_bit(GD_NEED_PART_SCAN, &disk->state);
1643         }
1644 out:
1645         mutex_unlock(&nbd_index_mutex);
1646         return ret;
1647 }
1648
1649 static void nbd_release(struct gendisk *disk)
1650 {
1651         struct nbd_device *nbd = disk->private_data;
1652
1653         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1654                         disk_openers(disk) == 0)
1655                 nbd_disconnect_and_put(nbd);
1656
1657         nbd_config_put(nbd);
1658         nbd_put(nbd);
1659 }
1660
1661 static void nbd_free_disk(struct gendisk *disk)
1662 {
1663         struct nbd_device *nbd = disk->private_data;
1664
1665         kfree(nbd);
1666 }
1667
1668 static const struct block_device_operations nbd_fops =
1669 {
1670         .owner =        THIS_MODULE,
1671         .open =         nbd_open,
1672         .release =      nbd_release,
1673         .ioctl =        nbd_ioctl,
1674         .compat_ioctl = nbd_ioctl,
1675         .free_disk =    nbd_free_disk,
1676 };
1677
1678 #if IS_ENABLED(CONFIG_DEBUG_FS)
1679
1680 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1681 {
1682         struct nbd_device *nbd = s->private;
1683
1684         if (nbd->pid)
1685                 seq_printf(s, "recv: %d\n", nbd->pid);
1686
1687         return 0;
1688 }
1689
1690 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1691
1692 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1693 {
1694         struct nbd_device *nbd = s->private;
1695         u32 flags = nbd->config->flags;
1696
1697         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1698
1699         seq_puts(s, "Known flags:\n");
1700
1701         if (flags & NBD_FLAG_HAS_FLAGS)
1702                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1703         if (flags & NBD_FLAG_READ_ONLY)
1704                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1705         if (flags & NBD_FLAG_SEND_FLUSH)
1706                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1707         if (flags & NBD_FLAG_SEND_FUA)
1708                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1709         if (flags & NBD_FLAG_SEND_TRIM)
1710                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1711
1712         return 0;
1713 }
1714
1715 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1716
1717 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1718 {
1719         struct dentry *dir;
1720         struct nbd_config *config = nbd->config;
1721
1722         if (!nbd_dbg_dir)
1723                 return -EIO;
1724
1725         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1726         if (IS_ERR(dir)) {
1727                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1728                         nbd_name(nbd));
1729                 return -EIO;
1730         }
1731         config->dbg_dir = dir;
1732
1733         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1734         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1735         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1736         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1737         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1738
1739         return 0;
1740 }
1741
1742 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1743 {
1744         debugfs_remove_recursive(nbd->config->dbg_dir);
1745 }
1746
1747 static int nbd_dbg_init(void)
1748 {
1749         struct dentry *dbg_dir;
1750
1751         dbg_dir = debugfs_create_dir("nbd", NULL);
1752         if (IS_ERR(dbg_dir))
1753                 return -EIO;
1754
1755         nbd_dbg_dir = dbg_dir;
1756
1757         return 0;
1758 }
1759
1760 static void nbd_dbg_close(void)
1761 {
1762         debugfs_remove_recursive(nbd_dbg_dir);
1763 }
1764
1765 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1766
1767 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1768 {
1769         return 0;
1770 }
1771
1772 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1773 {
1774 }
1775
1776 static int nbd_dbg_init(void)
1777 {
1778         return 0;
1779 }
1780
1781 static void nbd_dbg_close(void)
1782 {
1783 }
1784
1785 #endif
1786
1787 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1788                             unsigned int hctx_idx, unsigned int numa_node)
1789 {
1790         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1791         cmd->nbd = set->driver_data;
1792         cmd->flags = 0;
1793         mutex_init(&cmd->lock);
1794         return 0;
1795 }
1796
1797 static const struct blk_mq_ops nbd_mq_ops = {
1798         .queue_rq       = nbd_queue_rq,
1799         .complete       = nbd_complete_rq,
1800         .init_request   = nbd_init_request,
1801         .timeout        = nbd_xmit_timeout,
1802 };
1803
1804 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1805 {
1806         struct queue_limits lim = {
1807                 .max_hw_sectors         = 65536,
1808                 .max_user_sectors       = 256,
1809                 .max_segments           = USHRT_MAX,
1810                 .max_segment_size       = UINT_MAX,
1811         };
1812         struct nbd_device *nbd;
1813         struct gendisk *disk;
1814         int err = -ENOMEM;
1815
1816         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1817         if (!nbd)
1818                 goto out;
1819
1820         nbd->tag_set.ops = &nbd_mq_ops;
1821         nbd->tag_set.nr_hw_queues = 1;
1822         nbd->tag_set.queue_depth = 128;
1823         nbd->tag_set.numa_node = NUMA_NO_NODE;
1824         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1825         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1826                 BLK_MQ_F_BLOCKING;
1827         nbd->tag_set.driver_data = nbd;
1828         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1829         nbd->backend = NULL;
1830
1831         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1832         if (err)
1833                 goto out_free_nbd;
1834
1835         mutex_lock(&nbd_index_mutex);
1836         if (index >= 0) {
1837                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1838                                 GFP_KERNEL);
1839                 if (err == -ENOSPC)
1840                         err = -EEXIST;
1841         } else {
1842                 err = idr_alloc(&nbd_index_idr, nbd, 0,
1843                                 (MINORMASK >> part_shift) + 1, GFP_KERNEL);
1844                 if (err >= 0)
1845                         index = err;
1846         }
1847         nbd->index = index;
1848         mutex_unlock(&nbd_index_mutex);
1849         if (err < 0)
1850                 goto out_free_tags;
1851
1852         disk = blk_mq_alloc_disk(&nbd->tag_set, &lim, NULL);
1853         if (IS_ERR(disk)) {
1854                 err = PTR_ERR(disk);
1855                 goto out_free_idr;
1856         }
1857         nbd->disk = disk;
1858
1859         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1860                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1861                                           WQ_UNBOUND, 0, nbd->index);
1862         if (!nbd->recv_workq) {
1863                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1864                 err = -ENOMEM;
1865                 goto out_err_disk;
1866         }
1867
1868         /*
1869          * Tell the block layer that we are not a rotational device
1870          */
1871         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1872
1873         mutex_init(&nbd->config_lock);
1874         refcount_set(&nbd->config_refs, 0);
1875         /*
1876          * Start out with a zero references to keep other threads from using
1877          * this device until it is fully initialized.
1878          */
1879         refcount_set(&nbd->refs, 0);
1880         INIT_LIST_HEAD(&nbd->list);
1881         disk->major = NBD_MAJOR;
1882         disk->first_minor = index << part_shift;
1883         disk->minors = 1 << part_shift;
1884         disk->fops = &nbd_fops;
1885         disk->private_data = nbd;
1886         sprintf(disk->disk_name, "nbd%d", index);
1887         err = add_disk(disk);
1888         if (err)
1889                 goto out_free_work;
1890
1891         /*
1892          * Now publish the device.
1893          */
1894         refcount_set(&nbd->refs, refs);
1895         nbd_total_devices++;
1896         return nbd;
1897
1898 out_free_work:
1899         destroy_workqueue(nbd->recv_workq);
1900 out_err_disk:
1901         put_disk(disk);
1902 out_free_idr:
1903         mutex_lock(&nbd_index_mutex);
1904         idr_remove(&nbd_index_idr, index);
1905         mutex_unlock(&nbd_index_mutex);
1906 out_free_tags:
1907         blk_mq_free_tag_set(&nbd->tag_set);
1908 out_free_nbd:
1909         kfree(nbd);
1910 out:
1911         return ERR_PTR(err);
1912 }
1913
1914 static struct nbd_device *nbd_find_get_unused(void)
1915 {
1916         struct nbd_device *nbd;
1917         int id;
1918
1919         lockdep_assert_held(&nbd_index_mutex);
1920
1921         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1922                 if (refcount_read(&nbd->config_refs) ||
1923                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1924                         continue;
1925                 if (refcount_inc_not_zero(&nbd->refs))
1926                         return nbd;
1927         }
1928
1929         return NULL;
1930 }
1931
1932 /* Netlink interface. */
1933 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1934         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1935         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1936         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1937         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1938         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1939         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1940         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1941         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1942         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1943         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1944 };
1945
1946 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1947         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1948 };
1949
1950 /* We don't use this right now since we don't parse the incoming list, but we
1951  * still want it here so userspace knows what to expect.
1952  */
1953 static const struct nla_policy __attribute__((unused))
1954 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1955         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1956         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1957 };
1958
1959 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1960 {
1961         struct nbd_config *config = nbd->config;
1962         u64 bsize = nbd_blksize(config);
1963         u64 bytes = config->bytesize;
1964
1965         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1966                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1967
1968         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1969                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1970
1971         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1972                 return nbd_set_size(nbd, bytes, bsize);
1973         return 0;
1974 }
1975
1976 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1977 {
1978         struct nbd_device *nbd;
1979         struct nbd_config *config;
1980         int index = -1;
1981         int ret;
1982         bool put_dev = false;
1983
1984         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1985                 return -EPERM;
1986
1987         if (info->attrs[NBD_ATTR_INDEX]) {
1988                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1989
1990                 /*
1991                  * Too big first_minor can cause duplicate creation of
1992                  * sysfs files/links, since index << part_shift might overflow, or
1993                  * MKDEV() expect that the max bits of first_minor is 20.
1994                  */
1995                 if (index < 0 || index > MINORMASK >> part_shift) {
1996                         pr_err("illegal input index %d\n", index);
1997                         return -EINVAL;
1998                 }
1999         }
2000         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_SOCKETS)) {
2001                 pr_err("must specify at least one socket\n");
2002                 return -EINVAL;
2003         }
2004         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_SIZE_BYTES)) {
2005                 pr_err("must specify a size in bytes for the device\n");
2006                 return -EINVAL;
2007         }
2008 again:
2009         mutex_lock(&nbd_index_mutex);
2010         if (index == -1) {
2011                 nbd = nbd_find_get_unused();
2012         } else {
2013                 nbd = idr_find(&nbd_index_idr, index);
2014                 if (nbd) {
2015                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
2016                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
2017                             !refcount_inc_not_zero(&nbd->refs)) {
2018                                 mutex_unlock(&nbd_index_mutex);
2019                                 pr_err("device at index %d is going down\n",
2020                                         index);
2021                                 return -EINVAL;
2022                         }
2023                 }
2024         }
2025         mutex_unlock(&nbd_index_mutex);
2026
2027         if (!nbd) {
2028                 nbd = nbd_dev_add(index, 2);
2029                 if (IS_ERR(nbd)) {
2030                         pr_err("failed to add new device\n");
2031                         return PTR_ERR(nbd);
2032                 }
2033         }
2034
2035         mutex_lock(&nbd->config_lock);
2036         if (refcount_read(&nbd->config_refs)) {
2037                 mutex_unlock(&nbd->config_lock);
2038                 nbd_put(nbd);
2039                 if (index == -1)
2040                         goto again;
2041                 pr_err("nbd%d already in use\n", index);
2042                 return -EBUSY;
2043         }
2044
2045         ret = nbd_alloc_and_init_config(nbd);
2046         if (ret) {
2047                 mutex_unlock(&nbd->config_lock);
2048                 nbd_put(nbd);
2049                 pr_err("couldn't allocate config\n");
2050                 return ret;
2051         }
2052
2053         config = nbd->config;
2054         set_bit(NBD_RT_BOUND, &config->runtime_flags);
2055         ret = nbd_genl_size_set(info, nbd);
2056         if (ret)
2057                 goto out;
2058
2059         if (info->attrs[NBD_ATTR_TIMEOUT])
2060                 nbd_set_cmd_timeout(nbd,
2061                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2062         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2063                 config->dead_conn_timeout =
2064                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2065                 config->dead_conn_timeout *= HZ;
2066         }
2067         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
2068                 config->flags =
2069                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
2070         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2071                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2072                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2073                         /*
2074                          * We have 1 ref to keep the device around, and then 1
2075                          * ref for our current operation here, which will be
2076                          * inherited by the config.  If we already have
2077                          * DESTROY_ON_DISCONNECT set then we know we don't have
2078                          * that extra ref already held so we don't need the
2079                          * put_dev.
2080                          */
2081                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2082                                               &nbd->flags))
2083                                 put_dev = true;
2084                 } else {
2085                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2086                                                &nbd->flags))
2087                                 refcount_inc(&nbd->refs);
2088                 }
2089                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2090                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2091                                 &config->runtime_flags);
2092                 }
2093         }
2094
2095         if (info->attrs[NBD_ATTR_SOCKETS]) {
2096                 struct nlattr *attr;
2097                 int rem, fd;
2098
2099                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2100                                     rem) {
2101                         struct nlattr *socks[NBD_SOCK_MAX+1];
2102
2103                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2104                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2105                                 ret = -EINVAL;
2106                                 goto out;
2107                         }
2108                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2109                                                           attr,
2110                                                           nbd_sock_policy,
2111                                                           info->extack);
2112                         if (ret != 0) {
2113                                 pr_err("error processing sock list\n");
2114                                 ret = -EINVAL;
2115                                 goto out;
2116                         }
2117                         if (!socks[NBD_SOCK_FD])
2118                                 continue;
2119                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2120                         ret = nbd_add_socket(nbd, fd, true);
2121                         if (ret)
2122                                 goto out;
2123                 }
2124         }
2125         ret = nbd_start_device(nbd);
2126         if (ret)
2127                 goto out;
2128         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2129                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2130                                           GFP_KERNEL);
2131                 if (!nbd->backend) {
2132                         ret = -ENOMEM;
2133                         goto out;
2134                 }
2135         }
2136         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2137         if (ret) {
2138                 dev_err(disk_to_dev(nbd->disk),
2139                         "device_create_file failed for backend!\n");
2140                 goto out;
2141         }
2142         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2143 out:
2144         mutex_unlock(&nbd->config_lock);
2145         if (!ret) {
2146                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2147                 refcount_inc(&nbd->config_refs);
2148                 nbd_connect_reply(info, nbd->index);
2149         }
2150         nbd_config_put(nbd);
2151         if (put_dev)
2152                 nbd_put(nbd);
2153         return ret;
2154 }
2155
2156 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2157 {
2158         mutex_lock(&nbd->config_lock);
2159         nbd_disconnect(nbd);
2160         sock_shutdown(nbd);
2161         wake_up(&nbd->config->conn_wait);
2162         /*
2163          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2164          * to cancel the inflight I/Os.
2165          */
2166         flush_workqueue(nbd->recv_workq);
2167         nbd_clear_que(nbd);
2168         nbd->task_setup = NULL;
2169         mutex_unlock(&nbd->config_lock);
2170
2171         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2172                                &nbd->config->runtime_flags))
2173                 nbd_config_put(nbd);
2174 }
2175
2176 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2177 {
2178         struct nbd_device *nbd;
2179         int index;
2180
2181         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2182                 return -EPERM;
2183
2184         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_INDEX)) {
2185                 pr_err("must specify an index to disconnect\n");
2186                 return -EINVAL;
2187         }
2188         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2189         mutex_lock(&nbd_index_mutex);
2190         nbd = idr_find(&nbd_index_idr, index);
2191         if (!nbd) {
2192                 mutex_unlock(&nbd_index_mutex);
2193                 pr_err("couldn't find device at index %d\n", index);
2194                 return -EINVAL;
2195         }
2196         if (!refcount_inc_not_zero(&nbd->refs)) {
2197                 mutex_unlock(&nbd_index_mutex);
2198                 pr_err("device at index %d is going down\n", index);
2199                 return -EINVAL;
2200         }
2201         mutex_unlock(&nbd_index_mutex);
2202         if (!refcount_inc_not_zero(&nbd->config_refs))
2203                 goto put_nbd;
2204         nbd_disconnect_and_put(nbd);
2205         nbd_config_put(nbd);
2206 put_nbd:
2207         nbd_put(nbd);
2208         return 0;
2209 }
2210
2211 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2212 {
2213         struct nbd_device *nbd = NULL;
2214         struct nbd_config *config;
2215         int index;
2216         int ret = 0;
2217         bool put_dev = false;
2218
2219         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2220                 return -EPERM;
2221
2222         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_INDEX)) {
2223                 pr_err("must specify a device to reconfigure\n");
2224                 return -EINVAL;
2225         }
2226         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2227         mutex_lock(&nbd_index_mutex);
2228         nbd = idr_find(&nbd_index_idr, index);
2229         if (!nbd) {
2230                 mutex_unlock(&nbd_index_mutex);
2231                 pr_err("couldn't find a device at index %d\n", index);
2232                 return -EINVAL;
2233         }
2234         if (nbd->backend) {
2235                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2236                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2237                                        nbd->backend)) {
2238                                 mutex_unlock(&nbd_index_mutex);
2239                                 dev_err(nbd_to_dev(nbd),
2240                                         "backend image doesn't match with %s\n",
2241                                         nbd->backend);
2242                                 return -EINVAL;
2243                         }
2244                 } else {
2245                         mutex_unlock(&nbd_index_mutex);
2246                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2247                         return -EINVAL;
2248                 }
2249         }
2250         if (!refcount_inc_not_zero(&nbd->refs)) {
2251                 mutex_unlock(&nbd_index_mutex);
2252                 pr_err("device at index %d is going down\n", index);
2253                 return -EINVAL;
2254         }
2255         mutex_unlock(&nbd_index_mutex);
2256
2257         config = nbd_get_config_unlocked(nbd);
2258         if (!config) {
2259                 dev_err(nbd_to_dev(nbd),
2260                         "not configured, cannot reconfigure\n");
2261                 nbd_put(nbd);
2262                 return -EINVAL;
2263         }
2264
2265         mutex_lock(&nbd->config_lock);
2266         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2267             !nbd->pid) {
2268                 dev_err(nbd_to_dev(nbd),
2269                         "not configured, cannot reconfigure\n");
2270                 ret = -EINVAL;
2271                 goto out;
2272         }
2273
2274         ret = nbd_genl_size_set(info, nbd);
2275         if (ret)
2276                 goto out;
2277
2278         if (info->attrs[NBD_ATTR_TIMEOUT])
2279                 nbd_set_cmd_timeout(nbd,
2280                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2281         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2282                 config->dead_conn_timeout =
2283                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2284                 config->dead_conn_timeout *= HZ;
2285         }
2286         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2287                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2288                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2289                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2290                                               &nbd->flags))
2291                                 put_dev = true;
2292                 } else {
2293                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2294                                                &nbd->flags))
2295                                 refcount_inc(&nbd->refs);
2296                 }
2297
2298                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2299                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2300                                         &config->runtime_flags);
2301                 } else {
2302                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2303                                         &config->runtime_flags);
2304                 }
2305         }
2306
2307         if (info->attrs[NBD_ATTR_SOCKETS]) {
2308                 struct nlattr *attr;
2309                 int rem, fd;
2310
2311                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2312                                     rem) {
2313                         struct nlattr *socks[NBD_SOCK_MAX+1];
2314
2315                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2316                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2317                                 ret = -EINVAL;
2318                                 goto out;
2319                         }
2320                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2321                                                           attr,
2322                                                           nbd_sock_policy,
2323                                                           info->extack);
2324                         if (ret != 0) {
2325                                 pr_err("error processing sock list\n");
2326                                 ret = -EINVAL;
2327                                 goto out;
2328                         }
2329                         if (!socks[NBD_SOCK_FD])
2330                                 continue;
2331                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2332                         ret = nbd_reconnect_socket(nbd, fd);
2333                         if (ret) {
2334                                 if (ret == -ENOSPC)
2335                                         ret = 0;
2336                                 goto out;
2337                         }
2338                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2339                 }
2340         }
2341 out:
2342         mutex_unlock(&nbd->config_lock);
2343         nbd_config_put(nbd);
2344         nbd_put(nbd);
2345         if (put_dev)
2346                 nbd_put(nbd);
2347         return ret;
2348 }
2349
2350 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2351         {
2352                 .cmd    = NBD_CMD_CONNECT,
2353                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2354                 .doit   = nbd_genl_connect,
2355         },
2356         {
2357                 .cmd    = NBD_CMD_DISCONNECT,
2358                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2359                 .doit   = nbd_genl_disconnect,
2360         },
2361         {
2362                 .cmd    = NBD_CMD_RECONFIGURE,
2363                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2364                 .doit   = nbd_genl_reconfigure,
2365         },
2366         {
2367                 .cmd    = NBD_CMD_STATUS,
2368                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2369                 .doit   = nbd_genl_status,
2370         },
2371 };
2372
2373 static const struct genl_multicast_group nbd_mcast_grps[] = {
2374         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2375 };
2376
2377 static struct genl_family nbd_genl_family __ro_after_init = {
2378         .hdrsize        = 0,
2379         .name           = NBD_GENL_FAMILY_NAME,
2380         .version        = NBD_GENL_VERSION,
2381         .module         = THIS_MODULE,
2382         .small_ops      = nbd_connect_genl_ops,
2383         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2384         .resv_start_op  = NBD_CMD_STATUS + 1,
2385         .maxattr        = NBD_ATTR_MAX,
2386         .netnsok        = 1,
2387         .policy = nbd_attr_policy,
2388         .mcgrps         = nbd_mcast_grps,
2389         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2390 };
2391 MODULE_ALIAS_GENL_FAMILY(NBD_GENL_FAMILY_NAME);
2392
2393 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2394 {
2395         struct nlattr *dev_opt;
2396         u8 connected = 0;
2397         int ret;
2398
2399         /* This is a little racey, but for status it's ok.  The
2400          * reason we don't take a ref here is because we can't
2401          * take a ref in the index == -1 case as we would need
2402          * to put under the nbd_index_mutex, which could
2403          * deadlock if we are configured to remove ourselves
2404          * once we're disconnected.
2405          */
2406         if (refcount_read(&nbd->config_refs))
2407                 connected = 1;
2408         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2409         if (!dev_opt)
2410                 return -EMSGSIZE;
2411         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2412         if (ret)
2413                 return -EMSGSIZE;
2414         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2415                          connected);
2416         if (ret)
2417                 return -EMSGSIZE;
2418         nla_nest_end(reply, dev_opt);
2419         return 0;
2420 }
2421
2422 static int status_cb(int id, void *ptr, void *data)
2423 {
2424         struct nbd_device *nbd = ptr;
2425         return populate_nbd_status(nbd, (struct sk_buff *)data);
2426 }
2427
2428 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2429 {
2430         struct nlattr *dev_list;
2431         struct sk_buff *reply;
2432         void *reply_head;
2433         size_t msg_size;
2434         int index = -1;
2435         int ret = -ENOMEM;
2436
2437         if (info->attrs[NBD_ATTR_INDEX])
2438                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2439
2440         mutex_lock(&nbd_index_mutex);
2441
2442         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2443                                   nla_attr_size(sizeof(u8)));
2444         msg_size *= (index == -1) ? nbd_total_devices : 1;
2445
2446         reply = genlmsg_new(msg_size, GFP_KERNEL);
2447         if (!reply)
2448                 goto out;
2449         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2450                                        NBD_CMD_STATUS);
2451         if (!reply_head) {
2452                 nlmsg_free(reply);
2453                 goto out;
2454         }
2455
2456         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2457         if (!dev_list) {
2458                 nlmsg_free(reply);
2459                 ret = -EMSGSIZE;
2460                 goto out;
2461         }
2462
2463         if (index == -1) {
2464                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2465                 if (ret) {
2466                         nlmsg_free(reply);
2467                         goto out;
2468                 }
2469         } else {
2470                 struct nbd_device *nbd;
2471                 nbd = idr_find(&nbd_index_idr, index);
2472                 if (nbd) {
2473                         ret = populate_nbd_status(nbd, reply);
2474                         if (ret) {
2475                                 nlmsg_free(reply);
2476                                 goto out;
2477                         }
2478                 }
2479         }
2480         nla_nest_end(reply, dev_list);
2481         genlmsg_end(reply, reply_head);
2482         ret = genlmsg_reply(reply, info);
2483 out:
2484         mutex_unlock(&nbd_index_mutex);
2485         return ret;
2486 }
2487
2488 static void nbd_connect_reply(struct genl_info *info, int index)
2489 {
2490         struct sk_buff *skb;
2491         void *msg_head;
2492         int ret;
2493
2494         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2495         if (!skb)
2496                 return;
2497         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2498                                      NBD_CMD_CONNECT);
2499         if (!msg_head) {
2500                 nlmsg_free(skb);
2501                 return;
2502         }
2503         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2504         if (ret) {
2505                 nlmsg_free(skb);
2506                 return;
2507         }
2508         genlmsg_end(skb, msg_head);
2509         genlmsg_reply(skb, info);
2510 }
2511
2512 static void nbd_mcast_index(int index)
2513 {
2514         struct sk_buff *skb;
2515         void *msg_head;
2516         int ret;
2517
2518         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2519         if (!skb)
2520                 return;
2521         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2522                                      NBD_CMD_LINK_DEAD);
2523         if (!msg_head) {
2524                 nlmsg_free(skb);
2525                 return;
2526         }
2527         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2528         if (ret) {
2529                 nlmsg_free(skb);
2530                 return;
2531         }
2532         genlmsg_end(skb, msg_head);
2533         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2534 }
2535
2536 static void nbd_dead_link_work(struct work_struct *work)
2537 {
2538         struct link_dead_args *args = container_of(work, struct link_dead_args,
2539                                                    work);
2540         nbd_mcast_index(args->index);
2541         kfree(args);
2542 }
2543
2544 static int __init nbd_init(void)
2545 {
2546         int i;
2547
2548         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2549
2550         if (max_part < 0) {
2551                 pr_err("max_part must be >= 0\n");
2552                 return -EINVAL;
2553         }
2554
2555         part_shift = 0;
2556         if (max_part > 0) {
2557                 part_shift = fls(max_part);
2558
2559                 /*
2560                  * Adjust max_part according to part_shift as it is exported
2561                  * to user space so that user can know the max number of
2562                  * partition kernel should be able to manage.
2563                  *
2564                  * Note that -1 is required because partition 0 is reserved
2565                  * for the whole disk.
2566                  */
2567                 max_part = (1UL << part_shift) - 1;
2568         }
2569
2570         if ((1UL << part_shift) > DISK_MAX_PARTS)
2571                 return -EINVAL;
2572
2573         if (nbds_max > 1UL << (MINORBITS - part_shift))
2574                 return -EINVAL;
2575
2576         if (register_blkdev(NBD_MAJOR, "nbd"))
2577                 return -EIO;
2578
2579         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2580         if (!nbd_del_wq) {
2581                 unregister_blkdev(NBD_MAJOR, "nbd");
2582                 return -ENOMEM;
2583         }
2584
2585         if (genl_register_family(&nbd_genl_family)) {
2586                 destroy_workqueue(nbd_del_wq);
2587                 unregister_blkdev(NBD_MAJOR, "nbd");
2588                 return -EINVAL;
2589         }
2590         nbd_dbg_init();
2591
2592         for (i = 0; i < nbds_max; i++)
2593                 nbd_dev_add(i, 1);
2594         return 0;
2595 }
2596
2597 static int nbd_exit_cb(int id, void *ptr, void *data)
2598 {
2599         struct list_head *list = (struct list_head *)data;
2600         struct nbd_device *nbd = ptr;
2601
2602         /* Skip nbd that is being removed asynchronously */
2603         if (refcount_read(&nbd->refs))
2604                 list_add_tail(&nbd->list, list);
2605
2606         return 0;
2607 }
2608
2609 static void __exit nbd_cleanup(void)
2610 {
2611         struct nbd_device *nbd;
2612         LIST_HEAD(del_list);
2613
2614         /*
2615          * Unregister netlink interface prior to waiting
2616          * for the completion of netlink commands.
2617          */
2618         genl_unregister_family(&nbd_genl_family);
2619
2620         nbd_dbg_close();
2621
2622         mutex_lock(&nbd_index_mutex);
2623         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2624         mutex_unlock(&nbd_index_mutex);
2625
2626         while (!list_empty(&del_list)) {
2627                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2628                 list_del_init(&nbd->list);
2629                 if (refcount_read(&nbd->config_refs))
2630                         pr_err("possibly leaking nbd_config (ref %d)\n",
2631                                         refcount_read(&nbd->config_refs));
2632                 if (refcount_read(&nbd->refs) != 1)
2633                         pr_err("possibly leaking a device\n");
2634                 nbd_put(nbd);
2635         }
2636
2637         /* Also wait for nbd_dev_remove_work() completes */
2638         destroy_workqueue(nbd_del_wq);
2639
2640         idr_destroy(&nbd_index_idr);
2641         unregister_blkdev(NBD_MAJOR, "nbd");
2642 }
2643
2644 module_init(nbd_init);
2645 module_exit(nbd_cleanup);
2646
2647 MODULE_DESCRIPTION("Network Block Device");
2648 MODULE_LICENSE("GPL");
2649
2650 module_param(nbds_max, int, 0444);
2651 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2652 module_param(max_part, int, 0444);
2653 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");