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 <[email protected]>
   9  * Parts copyright 2001 Steven Whitehouse <[email protected]>
  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         lockdep_assert_held(&cmd->lock);
 185
 186         /*
 187          * Clear INFLIGHT flag so that this cmd won't be completed in
 188          * normal completion path
 189          *
 190          * INFLIGHT flag will be set when the cmd is queued to nbd next
 191          * time.
 192          */
 193         __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 194
 195         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 196                 blk_mq_requeue_request(req, true);
 197 }
 198
 199 #define NBD_COOKIE_BITS 32
 200
 201 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 202 {
 203         struct request *req = blk_mq_rq_from_pdu(cmd);
 204         u32 tag = blk_mq_unique_tag(req);
 205         u64 cookie = cmd->cmd_cookie;
 206
 207         return (cookie << NBD_COOKIE_BITS) | tag;
 208 }
 209
 210 static u32 nbd_handle_to_tag(u64 handle)
 211 {
 212         return (u32)handle;
 213 }
 214
 215 static u32 nbd_handle_to_cookie(u64 handle)
 216 {
 217         return (u32)(handle >> NBD_COOKIE_BITS);
 218 }
 219
 220 static const char *nbdcmd_to_ascii(int cmd)
 221 {
 222         switch (cmd) {
 223         case  NBD_CMD_READ: return "read";
 224         case NBD_CMD_WRITE: return "write";
 225         case  NBD_CMD_DISC: return "disconnect";
 226         case NBD_CMD_FLUSH: return "flush";
 227         case  NBD_CMD_TRIM: return "trim/discard";
 228         }
 229         return "invalid";
 230 }
 231
 232 static ssize_t pid_show(struct device *dev,
 233                         struct device_attribute *attr, char *buf)
 234 {
 235         struct gendisk *disk = dev_to_disk(dev);
 236         struct nbd_device *nbd = disk->private_data;
 237
 238         return sprintf(buf, "%d\n", nbd->pid);
 239 }
 240
 241 static const struct device_attribute pid_attr = {
 242         .attr = { .name = "pid", .mode = 0444},
 243         .show = pid_show,
 244 };
 245
 246 static ssize_t backend_show(struct device *dev,
 247                 struct device_attribute *attr, char *buf)
 248 {
 249         struct gendisk *disk = dev_to_disk(dev);
 250         struct nbd_device *nbd = disk->private_data;
 251
 252         return sprintf(buf, "%s\n", nbd->backend ?: "");
 253 }
 254
 255 static const struct device_attribute backend_attr = {
 256         .attr = { .name = "backend", .mode = 0444},
 257         .show = backend_show,
 258 };
 259
 260 static void nbd_dev_remove(struct nbd_device *nbd)
 261 {
 262         struct gendisk *disk = nbd->disk;
 263
 264         del_gendisk(disk);
 265         blk_mq_free_tag_set(&nbd->tag_set);
 266
 267         /*
 268          * Remove from idr after del_gendisk() completes, so if the same ID is
 269          * reused, the following add_disk() will succeed.
 270          */
 271         mutex_lock(&nbd_index_mutex);
 272         idr_remove(&nbd_index_idr, nbd->index);
 273         mutex_unlock(&nbd_index_mutex);
 274         destroy_workqueue(nbd->recv_workq);
 275         put_disk(disk);
 276 }
 277
 278 static void nbd_dev_remove_work(struct work_struct *work)
 279 {
 280         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 281 }
 282
 283 static void nbd_put(struct nbd_device *nbd)
 284 {
 285         if (!refcount_dec_and_test(&nbd->refs))
 286                 return;
 287
 288         /* Call del_gendisk() asynchrounously to prevent deadlock */
 289         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 290                 queue_work(nbd_del_wq, &nbd->remove_work);
 291         else
 292                 nbd_dev_remove(nbd);
 293 }
 294
 295 static int nbd_disconnected(struct nbd_config *config)
 296 {
 297         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 298                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 299 }
 300
 301 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 302                                 int notify)
 303 {
 304         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 305                 struct link_dead_args *args;
 306                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 307                 if (args) {
 308                         INIT_WORK(&args->work, nbd_dead_link_work);
 309                         args->index = nbd->index;
 310                         queue_work(system_wq, &args->work);
 311                 }
 312         }
 313         if (!nsock->dead) {
 314                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 315                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 316                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 317                                                &nbd->config->runtime_flags)) {
 318                                 set_bit(NBD_RT_DISCONNECTED,
 319                                         &nbd->config->runtime_flags);
 320                                 dev_info(nbd_to_dev(nbd),
 321                                         "Disconnected due to user request.\n");
 322                         }
 323                 }
 324         }
 325         nsock->dead = true;
 326         nsock->pending = NULL;
 327         nsock->sent = 0;
 328 }
 329
 330 static int __nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 331                 loff_t blksize)
 332 {
 333         struct queue_limits lim;
 334         int error;
 335
 336         if (!blksize)
 337                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 338
 339         if (blk_validate_block_size(blksize))
 340                 return -EINVAL;
 341
 342         if (bytesize < 0)
 343                 return -EINVAL;
 344
 345         nbd->config->bytesize = bytesize;
 346         nbd->config->blksize_bits = __ffs(blksize);
 347
 348         if (!nbd->pid)
 349                 return 0;
 350
 351         lim = queue_limits_start_update(nbd->disk->queue);
 352         if (nbd->config->flags & NBD_FLAG_SEND_TRIM)
 353                 lim.max_hw_discard_sectors = UINT_MAX >> SECTOR_SHIFT;
 354         else
 355                 lim.max_hw_discard_sectors = 0;
 356         if (!(nbd->config->flags & NBD_FLAG_SEND_FLUSH)) {
 357                 lim.features &= ~(BLK_FEAT_WRITE_CACHE | BLK_FEAT_FUA);
 358         } else if (nbd->config->flags & NBD_FLAG_SEND_FUA) {
 359                 lim.features |= BLK_FEAT_WRITE_CACHE | BLK_FEAT_FUA;
 360         } else {
 361                 lim.features |= BLK_FEAT_WRITE_CACHE;
 362                 lim.features &= ~BLK_FEAT_FUA;
 363         }
 364         if (nbd->config->flags & NBD_FLAG_ROTATIONAL)
 365                 lim.features |= BLK_FEAT_ROTATIONAL;
 366         if (nbd->config->flags & NBD_FLAG_SEND_WRITE_ZEROES)
 367                 lim.max_write_zeroes_sectors = UINT_MAX >> SECTOR_SHIFT;
 368
 369         lim.logical_block_size = blksize;
 370         lim.physical_block_size = blksize;
 371         error = queue_limits_commit_update(nbd->disk->queue, &lim);
 372         if (error)
 373                 return error;
 374
 375         if (max_part)
 376                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 377         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 378                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 379         return 0;
 380 }
 381
 382 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 383                 loff_t blksize)
 384 {
 385         int error;
 386
 387         blk_mq_freeze_queue(nbd->disk->queue);
 388         error = __nbd_set_size(nbd, bytesize, blksize);
 389         blk_mq_unfreeze_queue(nbd->disk->queue);
 390
 391         return error;
 392 }
 393
 394 static void nbd_complete_rq(struct request *req)
 395 {
 396         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 397
 398         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 399                 cmd->status ? "failed" : "done");
 400
 401         blk_mq_end_request(req, cmd->status);
 402 }
 403
 404 /*
 405  * Forcibly shutdown the socket causing all listeners to error
 406  */
 407 static void sock_shutdown(struct nbd_device *nbd)
 408 {
 409         struct nbd_config *config = nbd->config;
 410         int i;
 411
 412         if (config->num_connections == 0)
 413                 return;
 414         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 415                 return;
 416
 417         for (i = 0; i < config->num_connections; i++) {
 418                 struct nbd_sock *nsock = config->socks[i];
 419                 mutex_lock(&nsock->tx_lock);
 420                 nbd_mark_nsock_dead(nbd, nsock, 0);
 421                 mutex_unlock(&nsock->tx_lock);
 422         }
 423         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 424 }
 425
 426 static u32 req_to_nbd_cmd_type(struct request *req)
 427 {
 428         switch (req_op(req)) {
 429         case REQ_OP_DISCARD:
 430                 return NBD_CMD_TRIM;
 431         case REQ_OP_FLUSH:
 432                 return NBD_CMD_FLUSH;
 433         case REQ_OP_WRITE:
 434                 return NBD_CMD_WRITE;
 435         case REQ_OP_READ:
 436                 return NBD_CMD_READ;
 437         case REQ_OP_WRITE_ZEROES:
 438                 return NBD_CMD_WRITE_ZEROES;
 439         default:
 440                 return U32_MAX;
 441         }
 442 }
 443
 444 static struct nbd_config *nbd_get_config_unlocked(struct nbd_device *nbd)
 445 {
 446         if (refcount_inc_not_zero(&nbd->config_refs)) {
 447                 /*
 448                  * Add smp_mb__after_atomic to ensure that reading nbd->config_refs
 449                  * and reading nbd->config is ordered. The pair is the barrier in
 450                  * nbd_alloc_and_init_config(), avoid nbd->config_refs is set
 451                  * before nbd->config.
 452                  */
 453                 smp_mb__after_atomic();
 454                 return nbd->config;
 455         }
 456
 457         return NULL;
 458 }
 459
 460 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req)
 461 {
 462         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 463         struct nbd_device *nbd = cmd->nbd;
 464         struct nbd_config *config;
 465
 466         if (!mutex_trylock(&cmd->lock))
 467                 return BLK_EH_RESET_TIMER;
 468
 469         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 470                 mutex_unlock(&cmd->lock);
 471                 return BLK_EH_DONE;
 472         }
 473
 474         config = nbd_get_config_unlocked(nbd);
 475         if (!config) {
 476                 cmd->status = BLK_STS_TIMEOUT;
 477                 __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 478                 mutex_unlock(&cmd->lock);
 479                 goto done;
 480         }
 481
 482         if (config->num_connections > 1 ||
 483             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 484                 dev_err_ratelimited(nbd_to_dev(nbd),
 485                                     "Connection timed out, retrying (%d/%d alive)\n",
 486                                     atomic_read(&config->live_connections),
 487                                     config->num_connections);
 488                 /*
 489                  * Hooray we have more connections, requeue this IO, the submit
 490                  * path will put it on a real connection. Or if only one
 491                  * connection is configured, the submit path will wait util
 492                  * a new connection is reconfigured or util dead timeout.
 493                  */
 494                 if (config->socks) {
 495                         if (cmd->index < config->num_connections) {
 496                                 struct nbd_sock *nsock =
 497                                         config->socks[cmd->index];
 498                                 mutex_lock(&nsock->tx_lock);
 499                                 /* We can have multiple outstanding requests, so
 500                                  * we don't want to mark the nsock dead if we've
 501                                  * already reconnected with a new socket, so
 502                                  * only mark it dead if its the same socket we
 503                                  * were sent out on.
 504                                  */
 505                                 if (cmd->cookie == nsock->cookie)
 506                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 507                                 mutex_unlock(&nsock->tx_lock);
 508                         }
 509                         nbd_requeue_cmd(cmd);
 510                         mutex_unlock(&cmd->lock);
 511                         nbd_config_put(nbd);
 512                         return BLK_EH_DONE;
 513                 }
 514         }
 515
 516         if (!nbd->tag_set.timeout) {
 517                 /*
 518                  * Userspace sets timeout=0 to disable socket disconnection,
 519                  * so just warn and reset the timer.
 520                  */
 521                 struct nbd_sock *nsock = config->socks[cmd->index];
 522                 cmd->retries++;
 523                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 524                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 525                         (unsigned long long)blk_rq_pos(req) << 9,
 526                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 527
 528                 mutex_lock(&nsock->tx_lock);
 529                 if (cmd->cookie != nsock->cookie) {
 530                         nbd_requeue_cmd(cmd);
 531                         mutex_unlock(&nsock->tx_lock);
 532                         mutex_unlock(&cmd->lock);
 533                         nbd_config_put(nbd);
 534                         return BLK_EH_DONE;
 535                 }
 536                 mutex_unlock(&nsock->tx_lock);
 537                 mutex_unlock(&cmd->lock);
 538                 nbd_config_put(nbd);
 539                 return BLK_EH_RESET_TIMER;
 540         }
 541
 542         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 543         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 544         cmd->status = BLK_STS_IOERR;
 545         __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 546         mutex_unlock(&cmd->lock);
 547         sock_shutdown(nbd);
 548         nbd_config_put(nbd);
 549 done:
 550         blk_mq_complete_request(req);
 551         return BLK_EH_DONE;
 552 }
 553
 554 static int __sock_xmit(struct nbd_device *nbd, struct socket *sock, int send,
 555                        struct iov_iter *iter, int msg_flags, int *sent)
 556 {
 557         int result;
 558         struct msghdr msg = {} ;
 559         unsigned int noreclaim_flag;
 560
 561         if (unlikely(!sock)) {
 562                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 563                         "Attempted %s on closed socket in sock_xmit\n",
 564                         (send ? "send" : "recv"));
 565                 return -EINVAL;
 566         }
 567
 568         msg.msg_iter = *iter;
 569
 570         noreclaim_flag = memalloc_noreclaim_save();
 571         do {
 572                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 573                 sock->sk->sk_use_task_frag = false;
 574                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 575
 576                 if (send)
 577                         result = sock_sendmsg(sock, &msg);
 578                 else
 579                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 580
 581                 if (result <= 0) {
 582                         if (result == 0)
 583                                 result = -EPIPE; /* short read */
 584                         break;
 585                 }
 586                 if (sent)
 587                         *sent += result;
 588         } while (msg_data_left(&msg));
 589
 590         memalloc_noreclaim_restore(noreclaim_flag);
 591
 592         return result;
 593 }
 594
 595 /*
 596  *  Send or receive packet. Return a positive value on success and
 597  *  negtive value on failure, and never return 0.
 598  */
 599 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 600                      struct iov_iter *iter, int msg_flags, int *sent)
 601 {
 602         struct nbd_config *config = nbd->config;
 603         struct socket *sock = config->socks[index]->sock;
 604
 605         return __sock_xmit(nbd, sock, send, iter, msg_flags, sent);
 606 }
 607
 608 /*
 609  * Different settings for sk->sk_sndtimeo can result in different return values
 610  * if there is a signal pending when we enter sendmsg, because reasons?
 611  */
 612 static inline int was_interrupted(int result)
 613 {
 614         return result == -ERESTARTSYS || result == -EINTR;
 615 }
 616
 617 /*
 618  * Returns BLK_STS_RESOURCE if the caller should retry after a delay.
 619  * Returns BLK_STS_IOERR if sending failed.
 620  */
 621 static blk_status_t nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd,
 622                                  int index)
 623 {
 624         struct request *req = blk_mq_rq_from_pdu(cmd);
 625         struct nbd_config *config = nbd->config;
 626         struct nbd_sock *nsock = config->socks[index];
 627         int result;
 628         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 629         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 630         struct iov_iter from;
 631         struct bio *bio;
 632         u64 handle;
 633         u32 type;
 634         u32 nbd_cmd_flags = 0;
 635         int sent = nsock->sent, skip = 0;
 636
 637         lockdep_assert_held(&cmd->lock);
 638         lockdep_assert_held(&nsock->tx_lock);
 639
 640         iov_iter_kvec(&from, ITER_SOURCE, &iov, 1, sizeof(request));
 641
 642         type = req_to_nbd_cmd_type(req);
 643         if (type == U32_MAX)
 644                 return BLK_STS_IOERR;
 645
 646         if (rq_data_dir(req) == WRITE &&
 647             (config->flags & NBD_FLAG_READ_ONLY)) {
 648                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 649                                     "Write on read-only\n");
 650                 return BLK_STS_IOERR;
 651         }
 652
 653         if (req->cmd_flags & REQ_FUA)
 654                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 655         if ((req->cmd_flags & REQ_NOUNMAP) && (type == NBD_CMD_WRITE_ZEROES))
 656                 nbd_cmd_flags |= NBD_CMD_FLAG_NO_HOLE;
 657
 658         /* We did a partial send previously, and we at least sent the whole
 659          * request struct, so just go and send the rest of the pages in the
 660          * request.
 661          */
 662         if (sent) {
 663                 if (sent >= sizeof(request)) {
 664                         skip = sent - sizeof(request);
 665
 666                         /* initialize handle for tracing purposes */
 667                         handle = nbd_cmd_handle(cmd);
 668
 669                         goto send_pages;
 670                 }
 671                 iov_iter_advance(&from, sent);
 672         } else {
 673                 cmd->cmd_cookie++;
 674         }
 675         cmd->index = index;
 676         cmd->cookie = nsock->cookie;
 677         cmd->retries = 0;
 678         request.type = htonl(type | nbd_cmd_flags);
 679         if (type != NBD_CMD_FLUSH) {
 680                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 681                 request.len = htonl(blk_rq_bytes(req));
 682         }
 683         handle = nbd_cmd_handle(cmd);
 684         request.cookie = cpu_to_be64(handle);
 685
 686         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 687
 688         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 689                 req, nbdcmd_to_ascii(type),
 690                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 691         result = sock_xmit(nbd, index, 1, &from,
 692                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 693         trace_nbd_header_sent(req, handle);
 694         if (result < 0) {
 695                 if (was_interrupted(result)) {
 696                         /* If we haven't sent anything we can just return BUSY,
 697                          * however if we have sent something we need to make
 698                          * sure we only allow this req to be sent until we are
 699                          * completely done.
 700                          */
 701                         if (sent) {
 702                                 nsock->pending = req;
 703                                 nsock->sent = sent;
 704                         }
 705                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 706                         return BLK_STS_RESOURCE;
 707                 }
 708                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 709                         "Send control failed (result %d)\n", result);
 710                 goto requeue;
 711         }
 712 send_pages:
 713         if (type != NBD_CMD_WRITE)
 714                 goto out;
 715
 716         bio = req->bio;
 717         while (bio) {
 718                 struct bio *next = bio->bi_next;
 719                 struct bvec_iter iter;
 720                 struct bio_vec bvec;
 721
 722                 bio_for_each_segment(bvec, bio, iter) {
 723                         bool is_last = !next && bio_iter_last(bvec, iter);
 724                         int flags = is_last ? 0 : MSG_MORE;
 725
 726                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 727                                 req, bvec.bv_len);
 728                         iov_iter_bvec(&from, ITER_SOURCE, &bvec, 1, bvec.bv_len);
 729                         if (skip) {
 730                                 if (skip >= iov_iter_count(&from)) {
 731                                         skip -= iov_iter_count(&from);
 732                                         continue;
 733                                 }
 734                                 iov_iter_advance(&from, skip);
 735                                 skip = 0;
 736                         }
 737                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 738                         if (result < 0) {
 739                                 if (was_interrupted(result)) {
 740                                         /* We've already sent the header, we
 741                                          * have no choice but to set pending and
 742                                          * return BUSY.
 743                                          */
 744                                         nsock->pending = req;
 745                                         nsock->sent = sent;
 746                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 747                                         return BLK_STS_RESOURCE;
 748                                 }
 749                                 dev_err(disk_to_dev(nbd->disk),
 750                                         "Send data failed (result %d)\n",
 751                                         result);
 752                                 goto requeue;
 753                         }
 754                         /*
 755                          * The completion might already have come in,
 756                          * so break for the last one instead of letting
 757                          * the iterator do it. This prevents use-after-free
 758                          * of the bio.
 759                          */
 760                         if (is_last)
 761                                 break;
 762                 }
 763                 bio = next;
 764         }
 765 out:
 766         trace_nbd_payload_sent(req, handle);
 767         nsock->pending = NULL;
 768         nsock->sent = 0;
 769         __set_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 770         return BLK_STS_OK;
 771
 772 requeue:
 773         /* retry on a different socket */
 774         dev_err_ratelimited(disk_to_dev(nbd->disk),
 775                             "Request send failed, requeueing\n");
 776         nbd_mark_nsock_dead(nbd, nsock, 1);
 777         nbd_requeue_cmd(cmd);
 778         return BLK_STS_OK;
 779 }
 780
 781 static int nbd_read_reply(struct nbd_device *nbd, struct socket *sock,
 782                           struct nbd_reply *reply)
 783 {
 784         struct kvec iov = {.iov_base = reply, .iov_len = sizeof(*reply)};
 785         struct iov_iter to;
 786         int result;
 787
 788         reply->magic = 0;
 789         iov_iter_kvec(&to, ITER_DEST, &iov, 1, sizeof(*reply));
 790         result = __sock_xmit(nbd, sock, 0, &to, MSG_WAITALL, NULL);
 791         if (result < 0) {
 792                 if (!nbd_disconnected(nbd->config))
 793                         dev_err(disk_to_dev(nbd->disk),
 794                                 "Receive control failed (result %d)\n", result);
 795                 return result;
 796         }
 797
 798         if (ntohl(reply->magic) != NBD_REPLY_MAGIC) {
 799                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 800                                 (unsigned long)ntohl(reply->magic));
 801                 return -EPROTO;
 802         }
 803
 804         return 0;
 805 }
 806
 807 /* NULL returned = something went wrong, inform userspace */
 808 static struct nbd_cmd *nbd_handle_reply(struct nbd_device *nbd, int index,
 809                                         struct nbd_reply *reply)
 810 {
 811         int result;
 812         struct nbd_cmd *cmd;
 813         struct request *req = NULL;
 814         u64 handle;
 815         u16 hwq;
 816         u32 tag;
 817         int ret = 0;
 818
 819         handle = be64_to_cpu(reply->cookie);
 820         tag = nbd_handle_to_tag(handle);
 821         hwq = blk_mq_unique_tag_to_hwq(tag);
 822         if (hwq < nbd->tag_set.nr_hw_queues)
 823                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 824                                        blk_mq_unique_tag_to_tag(tag));
 825         if (!req || !blk_mq_request_started(req)) {
 826                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 827                         tag, req);
 828                 return ERR_PTR(-ENOENT);
 829         }
 830         trace_nbd_header_received(req, handle);
 831         cmd = blk_mq_rq_to_pdu(req);
 832
 833         mutex_lock(&cmd->lock);
 834         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 835                 dev_err(disk_to_dev(nbd->disk), "Suspicious reply %d (status %u flags %lu)",
 836                         tag, cmd->status, cmd->flags);
 837                 ret = -ENOENT;
 838                 goto out;
 839         }
 840         if (cmd->index != index) {
 841                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply %d from different sock %d (expected %d)",
 842                         tag, index, cmd->index);
 843                 ret = -ENOENT;
 844                 goto out;
 845         }
 846         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 847                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 848                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 849                 ret = -ENOENT;
 850                 goto out;
 851         }
 852         if (cmd->status != BLK_STS_OK) {
 853                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 854                         req);
 855                 ret = -ENOENT;
 856                 goto out;
 857         }
 858         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 859                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 860                         req);
 861                 ret = -ENOENT;
 862                 goto out;
 863         }
 864         if (ntohl(reply->error)) {
 865                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 866                         ntohl(reply->error));
 867                 cmd->status = BLK_STS_IOERR;
 868                 goto out;
 869         }
 870
 871         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 872         if (rq_data_dir(req) != WRITE) {
 873                 struct req_iterator iter;
 874                 struct bio_vec bvec;
 875                 struct iov_iter to;
 876
 877                 rq_for_each_segment(bvec, req, iter) {
 878                         iov_iter_bvec(&to, ITER_DEST, &bvec, 1, bvec.bv_len);
 879                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 880                         if (result < 0) {
 881                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 882                                         result);
 883                                 /*
 884                                  * If we've disconnected, we need to make sure we
 885                                  * complete this request, otherwise error out
 886                                  * and let the timeout stuff handle resubmitting
 887                                  * this request onto another connection.
 888                                  */
 889                                 if (nbd_disconnected(nbd->config)) {
 890                                         cmd->status = BLK_STS_IOERR;
 891                                         goto out;
 892                                 }
 893                                 ret = -EIO;
 894                                 goto out;
 895                         }
 896                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 897                                 req, bvec.bv_len);
 898                 }
 899         }
 900 out:
 901         trace_nbd_payload_received(req, handle);
 902         mutex_unlock(&cmd->lock);
 903         return ret ? ERR_PTR(ret) : cmd;
 904 }
 905
 906 static void recv_work(struct work_struct *work)
 907 {
 908         struct recv_thread_args *args = container_of(work,
 909                                                      struct recv_thread_args,
 910                                                      work);
 911         struct nbd_device *nbd = args->nbd;
 912         struct nbd_config *config = nbd->config;
 913         struct request_queue *q = nbd->disk->queue;
 914         struct nbd_sock *nsock = args->nsock;
 915         struct nbd_cmd *cmd;
 916         struct request *rq;
 917
 918         while (1) {
 919                 struct nbd_reply reply;
 920
 921                 if (nbd_read_reply(nbd, nsock->sock, &reply))
 922                         break;
 923
 924                 /*
 925                  * Grab .q_usage_counter so request pool won't go away, then no
 926                  * request use-after-free is possible during nbd_handle_reply().
 927                  * If queue is frozen, there won't be any inflight requests, we
 928                  * needn't to handle the incoming garbage message.
 929                  */
 930                 if (!percpu_ref_tryget(&q->q_usage_counter)) {
 931                         dev_err(disk_to_dev(nbd->disk), "%s: no io inflight\n",
 932                                 __func__);
 933                         break;
 934                 }
 935
 936                 cmd = nbd_handle_reply(nbd, args->index, &reply);
 937                 if (IS_ERR(cmd)) {
 938                         percpu_ref_put(&q->q_usage_counter);
 939                         break;
 940                 }
 941
 942                 rq = blk_mq_rq_from_pdu(cmd);
 943                 if (likely(!blk_should_fake_timeout(rq->q))) {
 944                         bool complete;
 945
 946                         mutex_lock(&cmd->lock);
 947                         complete = __test_and_clear_bit(NBD_CMD_INFLIGHT,
 948                                                         &cmd->flags);
 949                         mutex_unlock(&cmd->lock);
 950                         if (complete)
 951                                 blk_mq_complete_request(rq);
 952                 }
 953                 percpu_ref_put(&q->q_usage_counter);
 954         }
 955
 956         mutex_lock(&nsock->tx_lock);
 957         nbd_mark_nsock_dead(nbd, nsock, 1);
 958         mutex_unlock(&nsock->tx_lock);
 959
 960         nbd_config_put(nbd);
 961         atomic_dec(&config->recv_threads);
 962         wake_up(&config->recv_wq);
 963         kfree(args);
 964 }
 965
 966 static bool nbd_clear_req(struct request *req, void *data)
 967 {
 968         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 969
 970         /* don't abort one completed request */
 971         if (blk_mq_request_completed(req))
 972                 return true;
 973
 974         mutex_lock(&cmd->lock);
 975         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 976                 mutex_unlock(&cmd->lock);
 977                 return true;
 978         }
 979         cmd->status = BLK_STS_IOERR;
 980         mutex_unlock(&cmd->lock);
 981
 982         blk_mq_complete_request(req);
 983         return true;
 984 }
 985
 986 static void nbd_clear_que(struct nbd_device *nbd)
 987 {
 988         blk_mq_quiesce_queue(nbd->disk->queue);
 989         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 990         blk_mq_unquiesce_queue(nbd->disk->queue);
 991         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 992 }
 993
 994 static int find_fallback(struct nbd_device *nbd, int index)
 995 {
 996         struct nbd_config *config = nbd->config;
 997         int new_index = -1;
 998         struct nbd_sock *nsock = config->socks[index];
 999         int fallback = nsock->fallback_index;
1000
1001         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
1002                 return new_index;
1003
1004         if (config->num_connections <= 1) {
1005                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1006                                     "Dead connection, failed to find a fallback\n");
1007                 return new_index;
1008         }
1009
1010         if (fallback >= 0 && fallback < config->num_connections &&
1011             !config->socks[fallback]->dead)
1012                 return fallback;
1013
1014         if (nsock->fallback_index < 0 ||
1015             nsock->fallback_index >= config->num_connections ||
1016             config->socks[nsock->fallback_index]->dead) {
1017                 int i;
1018                 for (i = 0; i < config->num_connections; i++) {
1019                         if (i == index)
1020                                 continue;
1021                         if (!config->socks[i]->dead) {
1022                                 new_index = i;
1023                                 break;
1024                         }
1025                 }
1026                 nsock->fallback_index = new_index;
1027                 if (new_index < 0) {
1028                         dev_err_ratelimited(disk_to_dev(nbd->disk),
1029                                             "Dead connection, failed to find a fallback\n");
1030                         return new_index;
1031                 }
1032         }
1033         new_index = nsock->fallback_index;
1034         return new_index;
1035 }
1036
1037 static int wait_for_reconnect(struct nbd_device *nbd)
1038 {
1039         struct nbd_config *config = nbd->config;
1040         if (!config->dead_conn_timeout)
1041                 return 0;
1042
1043         if (!wait_event_timeout(config->conn_wait,
1044                                 test_bit(NBD_RT_DISCONNECTED,
1045                                          &config->runtime_flags) ||
1046                                 atomic_read(&config->live_connections) > 0,
1047                                 config->dead_conn_timeout))
1048                 return 0;
1049
1050         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1051 }
1052
1053 static blk_status_t nbd_handle_cmd(struct nbd_cmd *cmd, int index)
1054 {
1055         struct request *req = blk_mq_rq_from_pdu(cmd);
1056         struct nbd_device *nbd = cmd->nbd;
1057         struct nbd_config *config;
1058         struct nbd_sock *nsock;
1059         blk_status_t ret;
1060
1061         lockdep_assert_held(&cmd->lock);
1062
1063         config = nbd_get_config_unlocked(nbd);
1064         if (!config) {
1065                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1066                                     "Socks array is empty\n");
1067                 return BLK_STS_IOERR;
1068         }
1069
1070         if (index >= config->num_connections) {
1071                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1072                                     "Attempted send on invalid socket\n");
1073                 nbd_config_put(nbd);
1074                 return BLK_STS_IOERR;
1075         }
1076         cmd->status = BLK_STS_OK;
1077 again:
1078         nsock = config->socks[index];
1079         mutex_lock(&nsock->tx_lock);
1080         if (nsock->dead) {
1081                 int old_index = index;
1082                 index = find_fallback(nbd, index);
1083                 mutex_unlock(&nsock->tx_lock);
1084                 if (index < 0) {
1085                         if (wait_for_reconnect(nbd)) {
1086                                 index = old_index;
1087                                 goto again;
1088                         }
1089                         /* All the sockets should already be down at this point,
1090                          * we just want to make sure that DISCONNECTED is set so
1091                          * any requests that come in that were queue'ed waiting
1092                          * for the reconnect timer don't trigger the timer again
1093                          * and instead just error out.
1094                          */
1095                         sock_shutdown(nbd);
1096                         nbd_config_put(nbd);
1097                         return BLK_STS_IOERR;
1098                 }
1099                 goto again;
1100         }
1101
1102         /* Handle the case that we have a pending request that was partially
1103          * transmitted that _has_ to be serviced first.  We need to call requeue
1104          * here so that it gets put _after_ the request that is already on the
1105          * dispatch list.
1106          */
1107         blk_mq_start_request(req);
1108         if (unlikely(nsock->pending && nsock->pending != req)) {
1109                 nbd_requeue_cmd(cmd);
1110                 ret = BLK_STS_OK;
1111                 goto out;
1112         }
1113         ret = nbd_send_cmd(nbd, cmd, index);
1114 out:
1115         mutex_unlock(&nsock->tx_lock);
1116         nbd_config_put(nbd);
1117         return ret;
1118 }
1119
1120 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1121                         const struct blk_mq_queue_data *bd)
1122 {
1123         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1124         blk_status_t ret;
1125
1126         /*
1127          * Since we look at the bio's to send the request over the network we
1128          * need to make sure the completion work doesn't mark this request done
1129          * before we are done doing our send.  This keeps us from dereferencing
1130          * freed data if we have particularly fast completions (ie we get the
1131          * completion before we exit sock_xmit on the last bvec) or in the case
1132          * that the server is misbehaving (or there was an error) before we're
1133          * done sending everything over the wire.
1134          */
1135         mutex_lock(&cmd->lock);
1136         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1137
1138         /* We can be called directly from the user space process, which means we
1139          * could possibly have signals pending so our sendmsg will fail.  In
1140          * this case we need to return that we are busy, otherwise error out as
1141          * appropriate.
1142          */
1143         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1144         mutex_unlock(&cmd->lock);
1145
1146         return ret;
1147 }
1148
1149 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1150                                      int *err)
1151 {
1152         struct socket *sock;
1153
1154         *err = 0;
1155         sock = sockfd_lookup(fd, err);
1156         if (!sock)
1157                 return NULL;
1158
1159         if (sock->ops->shutdown == sock_no_shutdown) {
1160                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1161                 *err = -EINVAL;
1162                 sockfd_put(sock);
1163                 return NULL;
1164         }
1165
1166         return sock;
1167 }
1168
1169 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1170                           bool netlink)
1171 {
1172         struct nbd_config *config = nbd->config;
1173         struct socket *sock;
1174         struct nbd_sock **socks;
1175         struct nbd_sock *nsock;
1176         int err;
1177
1178         /* Arg will be cast to int, check it to avoid overflow */
1179         if (arg > INT_MAX)
1180                 return -EINVAL;
1181         sock = nbd_get_socket(nbd, arg, &err);
1182         if (!sock)
1183                 return err;
1184
1185         /*
1186          * We need to make sure we don't get any errant requests while we're
1187          * reallocating the ->socks array.
1188          */
1189         blk_mq_freeze_queue(nbd->disk->queue);
1190
1191         if (!netlink && !nbd->task_setup &&
1192             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1193                 nbd->task_setup = current;
1194
1195         if (!netlink &&
1196             (nbd->task_setup != current ||
1197              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1198                 dev_err(disk_to_dev(nbd->disk),
1199                         "Device being setup by another task");
1200                 err = -EBUSY;
1201                 goto put_socket;
1202         }
1203
1204         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1205         if (!nsock) {
1206                 err = -ENOMEM;
1207                 goto put_socket;
1208         }
1209
1210         socks = krealloc(config->socks, (config->num_connections + 1) *
1211                          sizeof(struct nbd_sock *), GFP_KERNEL);
1212         if (!socks) {
1213                 kfree(nsock);
1214                 err = -ENOMEM;
1215                 goto put_socket;
1216         }
1217
1218         config->socks = socks;
1219
1220         nsock->fallback_index = -1;
1221         nsock->dead = false;
1222         mutex_init(&nsock->tx_lock);
1223         nsock->sock = sock;
1224         nsock->pending = NULL;
1225         nsock->sent = 0;
1226         nsock->cookie = 0;
1227         socks[config->num_connections++] = nsock;
1228         atomic_inc(&config->live_connections);
1229         blk_mq_unfreeze_queue(nbd->disk->queue);
1230
1231         return 0;
1232
1233 put_socket:
1234         blk_mq_unfreeze_queue(nbd->disk->queue);
1235         sockfd_put(sock);
1236         return err;
1237 }
1238
1239 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1240 {
1241         struct nbd_config *config = nbd->config;
1242         struct socket *sock, *old;
1243         struct recv_thread_args *args;
1244         int i;
1245         int err;
1246
1247         sock = nbd_get_socket(nbd, arg, &err);
1248         if (!sock)
1249                 return err;
1250
1251         args = kzalloc(sizeof(*args), GFP_KERNEL);
1252         if (!args) {
1253                 sockfd_put(sock);
1254                 return -ENOMEM;
1255         }
1256
1257         for (i = 0; i < config->num_connections; i++) {
1258                 struct nbd_sock *nsock = config->socks[i];
1259
1260                 if (!nsock->dead)
1261                         continue;
1262
1263                 mutex_lock(&nsock->tx_lock);
1264                 if (!nsock->dead) {
1265                         mutex_unlock(&nsock->tx_lock);
1266                         continue;
1267                 }
1268                 sk_set_memalloc(sock->sk);
1269                 if (nbd->tag_set.timeout)
1270                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1271                 atomic_inc(&config->recv_threads);
1272                 refcount_inc(&nbd->config_refs);
1273                 old = nsock->sock;
1274                 nsock->fallback_index = -1;
1275                 nsock->sock = sock;
1276                 nsock->dead = false;
1277                 INIT_WORK(&args->work, recv_work);
1278                 args->index = i;
1279                 args->nbd = nbd;
1280                 args->nsock = nsock;
1281                 nsock->cookie++;
1282                 mutex_unlock(&nsock->tx_lock);
1283                 sockfd_put(old);
1284
1285                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1286
1287                 /* We take the tx_mutex in an error path in the recv_work, so we
1288                  * need to queue_work outside of the tx_mutex.
1289                  */
1290                 queue_work(nbd->recv_workq, &args->work);
1291
1292                 atomic_inc(&config->live_connections);
1293                 wake_up(&config->conn_wait);
1294                 return 0;
1295         }
1296         sockfd_put(sock);
1297         kfree(args);
1298         return -ENOSPC;
1299 }
1300
1301 static void nbd_bdev_reset(struct nbd_device *nbd)
1302 {
1303         if (disk_openers(nbd->disk) > 1)
1304                 return;
1305         set_capacity(nbd->disk, 0);
1306 }
1307
1308 static void nbd_parse_flags(struct nbd_device *nbd)
1309 {
1310         if (nbd->config->flags & NBD_FLAG_READ_ONLY)
1311                 set_disk_ro(nbd->disk, true);
1312         else
1313                 set_disk_ro(nbd->disk, false);
1314 }
1315
1316 static void send_disconnects(struct nbd_device *nbd)
1317 {
1318         struct nbd_config *config = nbd->config;
1319         struct nbd_request request = {
1320                 .magic = htonl(NBD_REQUEST_MAGIC),
1321                 .type = htonl(NBD_CMD_DISC),
1322         };
1323         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1324         struct iov_iter from;
1325         int i, ret;
1326
1327         for (i = 0; i < config->num_connections; i++) {
1328                 struct nbd_sock *nsock = config->socks[i];
1329
1330                 iov_iter_kvec(&from, ITER_SOURCE, &iov, 1, sizeof(request));
1331                 mutex_lock(&nsock->tx_lock);
1332                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1333                 if (ret < 0)
1334                         dev_err(disk_to_dev(nbd->disk),
1335                                 "Send disconnect failed %d\n", ret);
1336                 mutex_unlock(&nsock->tx_lock);
1337         }
1338 }
1339
1340 static int nbd_disconnect(struct nbd_device *nbd)
1341 {
1342         struct nbd_config *config = nbd->config;
1343
1344         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1345         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1346         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1347         send_disconnects(nbd);
1348         return 0;
1349 }
1350
1351 static void nbd_clear_sock(struct nbd_device *nbd)
1352 {
1353         sock_shutdown(nbd);
1354         nbd_clear_que(nbd);
1355         nbd->task_setup = NULL;
1356 }
1357
1358 static void nbd_config_put(struct nbd_device *nbd)
1359 {
1360         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1361                                         &nbd->config_lock)) {
1362                 struct nbd_config *config = nbd->config;
1363                 nbd_dev_dbg_close(nbd);
1364                 invalidate_disk(nbd->disk);
1365                 if (nbd->config->bytesize)
1366                         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
1367                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1368                                        &config->runtime_flags))
1369                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1370                 nbd->pid = 0;
1371                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1372                                        &config->runtime_flags)) {
1373                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1374                         kfree(nbd->backend);
1375                         nbd->backend = NULL;
1376                 }
1377                 nbd_clear_sock(nbd);
1378                 if (config->num_connections) {
1379                         int i;
1380                         for (i = 0; i < config->num_connections; i++) {
1381                                 sockfd_put(config->socks[i]->sock);
1382                                 kfree(config->socks[i]);
1383                         }
1384                         kfree(config->socks);
1385                 }
1386                 kfree(nbd->config);
1387                 nbd->config = NULL;
1388
1389                 nbd->tag_set.timeout = 0;
1390
1391                 mutex_unlock(&nbd->config_lock);
1392                 nbd_put(nbd);
1393                 module_put(THIS_MODULE);
1394         }
1395 }
1396
1397 static int nbd_start_device(struct nbd_device *nbd)
1398 {
1399         struct nbd_config *config = nbd->config;
1400         int num_connections = config->num_connections;
1401         int error = 0, i;
1402
1403         if (nbd->pid)
1404                 return -EBUSY;
1405         if (!config->socks)
1406                 return -EINVAL;
1407         if (num_connections > 1 &&
1408             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1409                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1410                 return -EINVAL;
1411         }
1412
1413         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1414         nbd->pid = task_pid_nr(current);
1415
1416         nbd_parse_flags(nbd);
1417
1418         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1419         if (error) {
1420                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1421                 return error;
1422         }
1423         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1424
1425         nbd_dev_dbg_init(nbd);
1426         for (i = 0; i < num_connections; i++) {
1427                 struct recv_thread_args *args;
1428
1429                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1430                 if (!args) {
1431                         sock_shutdown(nbd);
1432                         /*
1433                          * If num_connections is m (2 < m),
1434                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1435                          * But NO.(n + 1) failed. We still have n recv threads.
1436                          * So, add flush_workqueue here to prevent recv threads
1437                          * dropping the last config_refs and trying to destroy
1438                          * the workqueue from inside the workqueue.
1439                          */
1440                         if (i)
1441                                 flush_workqueue(nbd->recv_workq);
1442                         return -ENOMEM;
1443                 }
1444                 sk_set_memalloc(config->socks[i]->sock->sk);
1445                 if (nbd->tag_set.timeout)
1446                         config->socks[i]->sock->sk->sk_sndtimeo =
1447                                 nbd->tag_set.timeout;
1448                 atomic_inc(&config->recv_threads);
1449                 refcount_inc(&nbd->config_refs);
1450                 INIT_WORK(&args->work, recv_work);
1451                 args->nbd = nbd;
1452                 args->nsock = config->socks[i];
1453                 args->index = i;
1454                 queue_work(nbd->recv_workq, &args->work);
1455         }
1456         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1457 }
1458
1459 static int nbd_start_device_ioctl(struct nbd_device *nbd)
1460 {
1461         struct nbd_config *config = nbd->config;
1462         int ret;
1463
1464         ret = nbd_start_device(nbd);
1465         if (ret)
1466                 return ret;
1467
1468         if (max_part)
1469                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1470         mutex_unlock(&nbd->config_lock);
1471         ret = wait_event_interruptible(config->recv_wq,
1472                                          atomic_read(&config->recv_threads) == 0);
1473         if (ret) {
1474                 sock_shutdown(nbd);
1475                 nbd_clear_que(nbd);
1476         }
1477
1478         flush_workqueue(nbd->recv_workq);
1479         mutex_lock(&nbd->config_lock);
1480         nbd_bdev_reset(nbd);
1481         /* user requested, ignore socket errors */
1482         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1483                 ret = 0;
1484         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1485                 ret = -ETIMEDOUT;
1486         return ret;
1487 }
1488
1489 static void nbd_clear_sock_ioctl(struct nbd_device *nbd)
1490 {
1491         nbd_clear_sock(nbd);
1492         disk_force_media_change(nbd->disk);
1493         nbd_bdev_reset(nbd);
1494         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1495                                &nbd->config->runtime_flags))
1496                 nbd_config_put(nbd);
1497 }
1498
1499 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1500 {
1501         nbd->tag_set.timeout = timeout * HZ;
1502         if (timeout)
1503                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1504         else
1505                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1506 }
1507
1508 /* Must be called with config_lock held */
1509 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1510                        unsigned int cmd, unsigned long arg)
1511 {
1512         struct nbd_config *config = nbd->config;
1513         loff_t bytesize;
1514
1515         switch (cmd) {
1516         case NBD_DISCONNECT:
1517                 return nbd_disconnect(nbd);
1518         case NBD_CLEAR_SOCK:
1519                 nbd_clear_sock_ioctl(nbd);
1520                 return 0;
1521         case NBD_SET_SOCK:
1522                 return nbd_add_socket(nbd, arg, false);
1523         case NBD_SET_BLKSIZE:
1524                 return nbd_set_size(nbd, config->bytesize, arg);
1525         case NBD_SET_SIZE:
1526                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1527         case NBD_SET_SIZE_BLOCKS:
1528                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1529                         return -EINVAL;
1530                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1531         case NBD_SET_TIMEOUT:
1532                 nbd_set_cmd_timeout(nbd, arg);
1533                 return 0;
1534
1535         case NBD_SET_FLAGS:
1536                 config->flags = arg;
1537                 return 0;
1538         case NBD_DO_IT:
1539                 return nbd_start_device_ioctl(nbd);
1540         case NBD_CLEAR_QUE:
1541                 /*
1542                  * This is for compatibility only.  The queue is always cleared
1543                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1544                  */
1545                 return 0;
1546         case NBD_PRINT_DEBUG:
1547                 /*
1548                  * For compatibility only, we no longer keep a list of
1549                  * outstanding requests.
1550                  */
1551                 return 0;
1552         }
1553         return -ENOTTY;
1554 }
1555
1556 static int nbd_ioctl(struct block_device *bdev, blk_mode_t mode,
1557                      unsigned int cmd, unsigned long arg)
1558 {
1559         struct nbd_device *nbd = bdev->bd_disk->private_data;
1560         struct nbd_config *config = nbd->config;
1561         int error = -EINVAL;
1562
1563         if (!capable(CAP_SYS_ADMIN))
1564                 return -EPERM;
1565
1566         /* The block layer will pass back some non-nbd ioctls in case we have
1567          * special handling for them, but we don't so just return an error.
1568          */
1569         if (_IOC_TYPE(cmd) != 0xab)
1570                 return -EINVAL;
1571
1572         mutex_lock(&nbd->config_lock);
1573
1574         /* Don't allow ioctl operations on a nbd device that was created with
1575          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1576          */
1577         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1578             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1579                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1580         else
1581                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1582         mutex_unlock(&nbd->config_lock);
1583         return error;
1584 }
1585
1586 static int nbd_alloc_and_init_config(struct nbd_device *nbd)
1587 {
1588         struct nbd_config *config;
1589
1590         if (WARN_ON(nbd->config))
1591                 return -EINVAL;
1592
1593         if (!try_module_get(THIS_MODULE))
1594                 return -ENODEV;
1595
1596         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1597         if (!config) {
1598                 module_put(THIS_MODULE);
1599                 return -ENOMEM;
1600         }
1601
1602         atomic_set(&config->recv_threads, 0);
1603         init_waitqueue_head(&config->recv_wq);
1604         init_waitqueue_head(&config->conn_wait);
1605         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1606         atomic_set(&config->live_connections, 0);
1607
1608         nbd->config = config;
1609         /*
1610          * Order refcount_set(&nbd->config_refs, 1) and nbd->config assignment,
1611          * its pair is the barrier in nbd_get_config_unlocked().
1612          * So nbd_get_config_unlocked() won't see nbd->config as null after
1613          * refcount_inc_not_zero() succeed.
1614          */
1615         smp_mb__before_atomic();
1616         refcount_set(&nbd->config_refs, 1);
1617
1618         return 0;
1619 }
1620
1621 static int nbd_open(struct gendisk *disk, blk_mode_t mode)
1622 {
1623         struct nbd_device *nbd;
1624         struct nbd_config *config;
1625         int ret = 0;
1626
1627         mutex_lock(&nbd_index_mutex);
1628         nbd = disk->private_data;
1629         if (!nbd) {
1630                 ret = -ENXIO;
1631                 goto out;
1632         }
1633         if (!refcount_inc_not_zero(&nbd->refs)) {
1634                 ret = -ENXIO;
1635                 goto out;
1636         }
1637
1638         config = nbd_get_config_unlocked(nbd);
1639         if (!config) {
1640                 mutex_lock(&nbd->config_lock);
1641                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1642                         mutex_unlock(&nbd->config_lock);
1643                         goto out;
1644                 }
1645                 ret = nbd_alloc_and_init_config(nbd);
1646                 if (ret) {
1647                         mutex_unlock(&nbd->config_lock);
1648                         goto out;
1649                 }
1650
1651                 refcount_inc(&nbd->refs);
1652                 mutex_unlock(&nbd->config_lock);
1653                 if (max_part)
1654                         set_bit(GD_NEED_PART_SCAN, &disk->state);
1655         } else if (nbd_disconnected(config)) {
1656                 if (max_part)
1657                         set_bit(GD_NEED_PART_SCAN, &disk->state);
1658         }
1659 out:
1660         mutex_unlock(&nbd_index_mutex);
1661         return ret;
1662 }
1663
1664 static void nbd_release(struct gendisk *disk)
1665 {
1666         struct nbd_device *nbd = disk->private_data;
1667
1668         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1669                         disk_openers(disk) == 0)
1670                 nbd_disconnect_and_put(nbd);
1671
1672         nbd_config_put(nbd);
1673         nbd_put(nbd);
1674 }
1675
1676 static void nbd_free_disk(struct gendisk *disk)
1677 {
1678         struct nbd_device *nbd = disk->private_data;
1679
1680         kfree(nbd);
1681 }
1682
1683 static const struct block_device_operations nbd_fops =
1684 {
1685         .owner =        THIS_MODULE,
1686         .open =         nbd_open,
1687         .release =      nbd_release,
1688         .ioctl =        nbd_ioctl,
1689         .compat_ioctl = nbd_ioctl,
1690         .free_disk =    nbd_free_disk,
1691 };
1692
1693 #if IS_ENABLED(CONFIG_DEBUG_FS)
1694
1695 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1696 {
1697         struct nbd_device *nbd = s->private;
1698
1699         if (nbd->pid)
1700                 seq_printf(s, "recv: %d\n", nbd->pid);
1701
1702         return 0;
1703 }
1704
1705 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1706
1707 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1708 {
1709         struct nbd_device *nbd = s->private;
1710         u32 flags = nbd->config->flags;
1711
1712         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1713
1714         seq_puts(s, "Known flags:\n");
1715
1716         if (flags & NBD_FLAG_HAS_FLAGS)
1717                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1718         if (flags & NBD_FLAG_READ_ONLY)
1719                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1720         if (flags & NBD_FLAG_SEND_FLUSH)
1721                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1722         if (flags & NBD_FLAG_SEND_FUA)
1723                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1724         if (flags & NBD_FLAG_SEND_TRIM)
1725                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1726         if (flags & NBD_FLAG_SEND_WRITE_ZEROES)
1727                 seq_puts(s, "NBD_FLAG_SEND_WRITE_ZEROES\n");
1728         if (flags & NBD_FLAG_ROTATIONAL)
1729                 seq_puts(s, "NBD_FLAG_ROTATIONAL\n");
1730
1731         return 0;
1732 }
1733
1734 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1735
1736 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1737 {
1738         struct dentry *dir;
1739         struct nbd_config *config = nbd->config;
1740
1741         if (!nbd_dbg_dir)
1742                 return -EIO;
1743
1744         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1745         if (IS_ERR(dir)) {
1746                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1747                         nbd_name(nbd));
1748                 return -EIO;
1749         }
1750         config->dbg_dir = dir;
1751
1752         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1753         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1754         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1755         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1756         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1757
1758         return 0;
1759 }
1760
1761 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1762 {
1763         debugfs_remove_recursive(nbd->config->dbg_dir);
1764 }
1765
1766 static int nbd_dbg_init(void)
1767 {
1768         struct dentry *dbg_dir;
1769
1770         dbg_dir = debugfs_create_dir("nbd", NULL);
1771         if (IS_ERR(dbg_dir))
1772                 return -EIO;
1773
1774         nbd_dbg_dir = dbg_dir;
1775
1776         return 0;
1777 }
1778
1779 static void nbd_dbg_close(void)
1780 {
1781         debugfs_remove_recursive(nbd_dbg_dir);
1782 }
1783
1784 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1785
1786 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1787 {
1788         return 0;
1789 }
1790
1791 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1792 {
1793 }
1794
1795 static int nbd_dbg_init(void)
1796 {
1797         return 0;
1798 }
1799
1800 static void nbd_dbg_close(void)
1801 {
1802 }
1803
1804 #endif
1805
1806 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1807                             unsigned int hctx_idx, unsigned int numa_node)
1808 {
1809         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1810         cmd->nbd = set->driver_data;
1811         cmd->flags = 0;
1812         mutex_init(&cmd->lock);
1813         return 0;
1814 }
1815
1816 static const struct blk_mq_ops nbd_mq_ops = {
1817         .queue_rq       = nbd_queue_rq,
1818         .complete       = nbd_complete_rq,
1819         .init_request   = nbd_init_request,
1820         .timeout        = nbd_xmit_timeout,
1821 };
1822
1823 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1824 {
1825         struct queue_limits lim = {
1826                 .max_hw_sectors         = 65536,
1827                 .io_opt                 = 256 << SECTOR_SHIFT,
1828                 .max_segments           = USHRT_MAX,
1829                 .max_segment_size       = UINT_MAX,
1830         };
1831         struct nbd_device *nbd;
1832         struct gendisk *disk;
1833         int err = -ENOMEM;
1834
1835         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1836         if (!nbd)
1837                 goto out;
1838
1839         nbd->tag_set.ops = &nbd_mq_ops;
1840         nbd->tag_set.nr_hw_queues = 1;
1841         nbd->tag_set.queue_depth = 128;
1842         nbd->tag_set.numa_node = NUMA_NO_NODE;
1843         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1844         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1845                 BLK_MQ_F_BLOCKING;
1846         nbd->tag_set.driver_data = nbd;
1847         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1848         nbd->backend = NULL;
1849
1850         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1851         if (err)
1852                 goto out_free_nbd;
1853
1854         mutex_lock(&nbd_index_mutex);
1855         if (index >= 0) {
1856                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1857                                 GFP_KERNEL);
1858                 if (err == -ENOSPC)
1859                         err = -EEXIST;
1860         } else {
1861                 err = idr_alloc(&nbd_index_idr, nbd, 0,
1862                                 (MINORMASK >> part_shift) + 1, GFP_KERNEL);
1863                 if (err >= 0)
1864                         index = err;
1865         }
1866         nbd->index = index;
1867         mutex_unlock(&nbd_index_mutex);
1868         if (err < 0)
1869                 goto out_free_tags;
1870
1871         disk = blk_mq_alloc_disk(&nbd->tag_set, &lim, NULL);
1872         if (IS_ERR(disk)) {
1873                 err = PTR_ERR(disk);
1874                 goto out_free_idr;
1875         }
1876         nbd->disk = disk;
1877
1878         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1879                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1880                                           WQ_UNBOUND, 0, nbd->index);
1881         if (!nbd->recv_workq) {
1882                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1883                 err = -ENOMEM;
1884                 goto out_err_disk;
1885         }
1886
1887         mutex_init(&nbd->config_lock);
1888         refcount_set(&nbd->config_refs, 0);
1889         /*
1890          * Start out with a zero references to keep other threads from using
1891          * this device until it is fully initialized.
1892          */
1893         refcount_set(&nbd->refs, 0);
1894         INIT_LIST_HEAD(&nbd->list);
1895         disk->major = NBD_MAJOR;
1896         disk->first_minor = index << part_shift;
1897         disk->minors = 1 << part_shift;
1898         disk->fops = &nbd_fops;
1899         disk->private_data = nbd;
1900         sprintf(disk->disk_name, "nbd%d", index);
1901         err = add_disk(disk);
1902         if (err)
1903                 goto out_free_work;
1904
1905         /*
1906          * Now publish the device.
1907          */
1908         refcount_set(&nbd->refs, refs);
1909         nbd_total_devices++;
1910         return nbd;
1911
1912 out_free_work:
1913         destroy_workqueue(nbd->recv_workq);
1914 out_err_disk:
1915         put_disk(disk);
1916 out_free_idr:
1917         mutex_lock(&nbd_index_mutex);
1918         idr_remove(&nbd_index_idr, index);
1919         mutex_unlock(&nbd_index_mutex);
1920 out_free_tags:
1921         blk_mq_free_tag_set(&nbd->tag_set);
1922 out_free_nbd:
1923         kfree(nbd);
1924 out:
1925         return ERR_PTR(err);
1926 }
1927
1928 static struct nbd_device *nbd_find_get_unused(void)
1929 {
1930         struct nbd_device *nbd;
1931         int id;
1932
1933         lockdep_assert_held(&nbd_index_mutex);
1934
1935         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1936                 if (refcount_read(&nbd->config_refs) ||
1937                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1938                         continue;
1939                 if (refcount_inc_not_zero(&nbd->refs))
1940                         return nbd;
1941         }
1942
1943         return NULL;
1944 }
1945
1946 /* Netlink interface. */
1947 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1948         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1949         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1950         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1951         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1952         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1953         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1954         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1955         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1956         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1957         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1958 };
1959
1960 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1961         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1962 };
1963
1964 /* We don't use this right now since we don't parse the incoming list, but we
1965  * still want it here so userspace knows what to expect.
1966  */
1967 static const struct nla_policy __attribute__((unused))
1968 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1969         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1970         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1971 };
1972
1973 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1974 {
1975         struct nbd_config *config = nbd->config;
1976         u64 bsize = nbd_blksize(config);
1977         u64 bytes = config->bytesize;
1978
1979         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1980                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1981
1982         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1983                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1984
1985         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1986                 return nbd_set_size(nbd, bytes, bsize);
1987         return 0;
1988 }
1989
1990 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1991 {
1992         struct nbd_device *nbd;
1993         struct nbd_config *config;
1994         int index = -1;
1995         int ret;
1996         bool put_dev = false;
1997
1998         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1999                 return -EPERM;
2000
2001         if (info->attrs[NBD_ATTR_INDEX]) {
2002                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2003
2004                 /*
2005                  * Too big first_minor can cause duplicate creation of
2006                  * sysfs files/links, since index << part_shift might overflow, or
2007                  * MKDEV() expect that the max bits of first_minor is 20.
2008                  */
2009                 if (index < 0 || index > MINORMASK >> part_shift) {
2010                         pr_err("illegal input index %d\n", index);
2011                         return -EINVAL;
2012                 }
2013         }
2014         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_SOCKETS)) {
2015                 pr_err("must specify at least one socket\n");
2016                 return -EINVAL;
2017         }
2018         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_SIZE_BYTES)) {
2019                 pr_err("must specify a size in bytes for the device\n");
2020                 return -EINVAL;
2021         }
2022 again:
2023         mutex_lock(&nbd_index_mutex);
2024         if (index == -1) {
2025                 nbd = nbd_find_get_unused();
2026         } else {
2027                 nbd = idr_find(&nbd_index_idr, index);
2028                 if (nbd) {
2029                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
2030                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
2031                             !refcount_inc_not_zero(&nbd->refs)) {
2032                                 mutex_unlock(&nbd_index_mutex);
2033                                 pr_err("device at index %d is going down\n",
2034                                         index);
2035                                 return -EINVAL;
2036                         }
2037                 }
2038         }
2039         mutex_unlock(&nbd_index_mutex);
2040
2041         if (!nbd) {
2042                 nbd = nbd_dev_add(index, 2);
2043                 if (IS_ERR(nbd)) {
2044                         pr_err("failed to add new device\n");
2045                         return PTR_ERR(nbd);
2046                 }
2047         }
2048
2049         mutex_lock(&nbd->config_lock);
2050         if (refcount_read(&nbd->config_refs)) {
2051                 mutex_unlock(&nbd->config_lock);
2052                 nbd_put(nbd);
2053                 if (index == -1)
2054                         goto again;
2055                 pr_err("nbd%d already in use\n", index);
2056                 return -EBUSY;
2057         }
2058
2059         ret = nbd_alloc_and_init_config(nbd);
2060         if (ret) {
2061                 mutex_unlock(&nbd->config_lock);
2062                 nbd_put(nbd);
2063                 pr_err("couldn't allocate config\n");
2064                 return ret;
2065         }
2066
2067         config = nbd->config;
2068         set_bit(NBD_RT_BOUND, &config->runtime_flags);
2069         ret = nbd_genl_size_set(info, nbd);
2070         if (ret)
2071                 goto out;
2072
2073         if (info->attrs[NBD_ATTR_TIMEOUT])
2074                 nbd_set_cmd_timeout(nbd,
2075                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2076         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2077                 config->dead_conn_timeout =
2078                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2079                 config->dead_conn_timeout *= HZ;
2080         }
2081         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
2082                 config->flags =
2083                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
2084         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2085                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2086                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2087                         /*
2088                          * We have 1 ref to keep the device around, and then 1
2089                          * ref for our current operation here, which will be
2090                          * inherited by the config.  If we already have
2091                          * DESTROY_ON_DISCONNECT set then we know we don't have
2092                          * that extra ref already held so we don't need the
2093                          * put_dev.
2094                          */
2095                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2096                                               &nbd->flags))
2097                                 put_dev = true;
2098                 } else {
2099                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2100                                                &nbd->flags))
2101                                 refcount_inc(&nbd->refs);
2102                 }
2103                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2104                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2105                                 &config->runtime_flags);
2106                 }
2107         }
2108
2109         if (info->attrs[NBD_ATTR_SOCKETS]) {
2110                 struct nlattr *attr;
2111                 int rem, fd;
2112
2113                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2114                                     rem) {
2115                         struct nlattr *socks[NBD_SOCK_MAX+1];
2116
2117                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2118                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2119                                 ret = -EINVAL;
2120                                 goto out;
2121                         }
2122                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2123                                                           attr,
2124                                                           nbd_sock_policy,
2125                                                           info->extack);
2126                         if (ret != 0) {
2127                                 pr_err("error processing sock list\n");
2128                                 ret = -EINVAL;
2129                                 goto out;
2130                         }
2131                         if (!socks[NBD_SOCK_FD])
2132                                 continue;
2133                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2134                         ret = nbd_add_socket(nbd, fd, true);
2135                         if (ret)
2136                                 goto out;
2137                 }
2138         }
2139         ret = nbd_start_device(nbd);
2140         if (ret)
2141                 goto out;
2142         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2143                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2144                                           GFP_KERNEL);
2145                 if (!nbd->backend) {
2146                         ret = -ENOMEM;
2147                         goto out;
2148                 }
2149         }
2150         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2151         if (ret) {
2152                 dev_err(disk_to_dev(nbd->disk),
2153                         "device_create_file failed for backend!\n");
2154                 goto out;
2155         }
2156         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2157 out:
2158         mutex_unlock(&nbd->config_lock);
2159         if (!ret) {
2160                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2161                 refcount_inc(&nbd->config_refs);
2162                 nbd_connect_reply(info, nbd->index);
2163         }
2164         nbd_config_put(nbd);
2165         if (put_dev)
2166                 nbd_put(nbd);
2167         return ret;
2168 }
2169
2170 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2171 {
2172         mutex_lock(&nbd->config_lock);
2173         nbd_disconnect(nbd);
2174         sock_shutdown(nbd);
2175         wake_up(&nbd->config->conn_wait);
2176         /*
2177          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2178          * to cancel the inflight I/Os.
2179          */
2180         flush_workqueue(nbd->recv_workq);
2181         nbd_clear_que(nbd);
2182         nbd->task_setup = NULL;
2183         mutex_unlock(&nbd->config_lock);
2184
2185         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2186                                &nbd->config->runtime_flags))
2187                 nbd_config_put(nbd);
2188 }
2189
2190 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2191 {
2192         struct nbd_device *nbd;
2193         int index;
2194
2195         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2196                 return -EPERM;
2197
2198         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_INDEX)) {
2199                 pr_err("must specify an index to disconnect\n");
2200                 return -EINVAL;
2201         }
2202         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2203         mutex_lock(&nbd_index_mutex);
2204         nbd = idr_find(&nbd_index_idr, index);
2205         if (!nbd) {
2206                 mutex_unlock(&nbd_index_mutex);
2207                 pr_err("couldn't find device at index %d\n", index);
2208                 return -EINVAL;
2209         }
2210         if (!refcount_inc_not_zero(&nbd->refs)) {
2211                 mutex_unlock(&nbd_index_mutex);
2212                 pr_err("device at index %d is going down\n", index);
2213                 return -EINVAL;
2214         }
2215         mutex_unlock(&nbd_index_mutex);
2216         if (!refcount_inc_not_zero(&nbd->config_refs))
2217                 goto put_nbd;
2218         nbd_disconnect_and_put(nbd);
2219         nbd_config_put(nbd);
2220 put_nbd:
2221         nbd_put(nbd);
2222         return 0;
2223 }
2224
2225 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2226 {
2227         struct nbd_device *nbd = NULL;
2228         struct nbd_config *config;
2229         int index;
2230         int ret = 0;
2231         bool put_dev = false;
2232
2233         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2234                 return -EPERM;
2235
2236         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_INDEX)) {
2237                 pr_err("must specify a device to reconfigure\n");
2238                 return -EINVAL;
2239         }
2240         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2241         mutex_lock(&nbd_index_mutex);
2242         nbd = idr_find(&nbd_index_idr, index);
2243         if (!nbd) {
2244                 mutex_unlock(&nbd_index_mutex);
2245                 pr_err("couldn't find a device at index %d\n", index);
2246                 return -EINVAL;
2247         }
2248         if (nbd->backend) {
2249                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2250                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2251                                        nbd->backend)) {
2252                                 mutex_unlock(&nbd_index_mutex);
2253                                 dev_err(nbd_to_dev(nbd),
2254                                         "backend image doesn't match with %s\n",
2255                                         nbd->backend);
2256                                 return -EINVAL;
2257                         }
2258                 } else {
2259                         mutex_unlock(&nbd_index_mutex);
2260                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2261                         return -EINVAL;
2262                 }
2263         }
2264         if (!refcount_inc_not_zero(&nbd->refs)) {
2265                 mutex_unlock(&nbd_index_mutex);
2266                 pr_err("device at index %d is going down\n", index);
2267                 return -EINVAL;
2268         }
2269         mutex_unlock(&nbd_index_mutex);
2270
2271         config = nbd_get_config_unlocked(nbd);
2272         if (!config) {
2273                 dev_err(nbd_to_dev(nbd),
2274                         "not configured, cannot reconfigure\n");
2275                 nbd_put(nbd);
2276                 return -EINVAL;
2277         }
2278
2279         mutex_lock(&nbd->config_lock);
2280         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2281             !nbd->pid) {
2282                 dev_err(nbd_to_dev(nbd),
2283                         "not configured, cannot reconfigure\n");
2284                 ret = -EINVAL;
2285                 goto out;
2286         }
2287
2288         ret = nbd_genl_size_set(info, nbd);
2289         if (ret)
2290                 goto out;
2291
2292         if (info->attrs[NBD_ATTR_TIMEOUT])
2293                 nbd_set_cmd_timeout(nbd,
2294                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2295         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2296                 config->dead_conn_timeout =
2297                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2298                 config->dead_conn_timeout *= HZ;
2299         }
2300         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2301                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2302                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2303                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2304                                               &nbd->flags))
2305                                 put_dev = true;
2306                 } else {
2307                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2308                                                &nbd->flags))
2309                                 refcount_inc(&nbd->refs);
2310                 }
2311
2312                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2313                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2314                                         &config->runtime_flags);
2315                 } else {
2316                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2317                                         &config->runtime_flags);
2318                 }
2319         }
2320
2321         if (info->attrs[NBD_ATTR_SOCKETS]) {
2322                 struct nlattr *attr;
2323                 int rem, fd;
2324
2325                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2326                                     rem) {
2327                         struct nlattr *socks[NBD_SOCK_MAX+1];
2328
2329                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2330                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2331                                 ret = -EINVAL;
2332                                 goto out;
2333                         }
2334                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2335                                                           attr,
2336                                                           nbd_sock_policy,
2337                                                           info->extack);
2338                         if (ret != 0) {
2339                                 pr_err("error processing sock list\n");
2340                                 ret = -EINVAL;
2341                                 goto out;
2342                         }
2343                         if (!socks[NBD_SOCK_FD])
2344                                 continue;
2345                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2346                         ret = nbd_reconnect_socket(nbd, fd);
2347                         if (ret) {
2348                                 if (ret == -ENOSPC)
2349                                         ret = 0;
2350                                 goto out;
2351                         }
2352                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2353                 }
2354         }
2355 out:
2356         mutex_unlock(&nbd->config_lock);
2357         nbd_config_put(nbd);
2358         nbd_put(nbd);
2359         if (put_dev)
2360                 nbd_put(nbd);
2361         return ret;
2362 }
2363
2364 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2365         {
2366                 .cmd    = NBD_CMD_CONNECT,
2367                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2368                 .doit   = nbd_genl_connect,
2369         },
2370         {
2371                 .cmd    = NBD_CMD_DISCONNECT,
2372                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2373                 .doit   = nbd_genl_disconnect,
2374         },
2375         {
2376                 .cmd    = NBD_CMD_RECONFIGURE,
2377                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2378                 .doit   = nbd_genl_reconfigure,
2379         },
2380         {
2381                 .cmd    = NBD_CMD_STATUS,
2382                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2383                 .doit   = nbd_genl_status,
2384         },
2385 };
2386
2387 static const struct genl_multicast_group nbd_mcast_grps[] = {
2388         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2389 };
2390
2391 static struct genl_family nbd_genl_family __ro_after_init = {
2392         .hdrsize        = 0,
2393         .name           = NBD_GENL_FAMILY_NAME,
2394         .version        = NBD_GENL_VERSION,
2395         .module         = THIS_MODULE,
2396         .small_ops      = nbd_connect_genl_ops,
2397         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2398         .resv_start_op  = NBD_CMD_STATUS + 1,
2399         .maxattr        = NBD_ATTR_MAX,
2400         .netnsok        = 1,
2401         .policy = nbd_attr_policy,
2402         .mcgrps         = nbd_mcast_grps,
2403         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2404 };
2405 MODULE_ALIAS_GENL_FAMILY(NBD_GENL_FAMILY_NAME);
2406
2407 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2408 {
2409         struct nlattr *dev_opt;
2410         u8 connected = 0;
2411         int ret;
2412
2413         /* This is a little racey, but for status it's ok.  The
2414          * reason we don't take a ref here is because we can't
2415          * take a ref in the index == -1 case as we would need
2416          * to put under the nbd_index_mutex, which could
2417          * deadlock if we are configured to remove ourselves
2418          * once we're disconnected.
2419          */
2420         if (refcount_read(&nbd->config_refs))
2421                 connected = 1;
2422         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2423         if (!dev_opt)
2424                 return -EMSGSIZE;
2425         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2426         if (ret)
2427                 return -EMSGSIZE;
2428         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2429                          connected);
2430         if (ret)
2431                 return -EMSGSIZE;
2432         nla_nest_end(reply, dev_opt);
2433         return 0;
2434 }
2435
2436 static int status_cb(int id, void *ptr, void *data)
2437 {
2438         struct nbd_device *nbd = ptr;
2439         return populate_nbd_status(nbd, (struct sk_buff *)data);
2440 }
2441
2442 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2443 {
2444         struct nlattr *dev_list;
2445         struct sk_buff *reply;
2446         void *reply_head;
2447         size_t msg_size;
2448         int index = -1;
2449         int ret = -ENOMEM;
2450
2451         if (info->attrs[NBD_ATTR_INDEX])
2452                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2453
2454         mutex_lock(&nbd_index_mutex);
2455
2456         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2457                                   nla_attr_size(sizeof(u8)));
2458         msg_size *= (index == -1) ? nbd_total_devices : 1;
2459
2460         reply = genlmsg_new(msg_size, GFP_KERNEL);
2461         if (!reply)
2462                 goto out;
2463         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2464                                        NBD_CMD_STATUS);
2465         if (!reply_head) {
2466                 nlmsg_free(reply);
2467                 goto out;
2468         }
2469
2470         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2471         if (!dev_list) {
2472                 nlmsg_free(reply);
2473                 ret = -EMSGSIZE;
2474                 goto out;
2475         }
2476
2477         if (index == -1) {
2478                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2479                 if (ret) {
2480                         nlmsg_free(reply);
2481                         goto out;
2482                 }
2483         } else {
2484                 struct nbd_device *nbd;
2485                 nbd = idr_find(&nbd_index_idr, index);
2486                 if (nbd) {
2487                         ret = populate_nbd_status(nbd, reply);
2488                         if (ret) {
2489                                 nlmsg_free(reply);
2490                                 goto out;
2491                         }
2492                 }
2493         }
2494         nla_nest_end(reply, dev_list);
2495         genlmsg_end(reply, reply_head);
2496         ret = genlmsg_reply(reply, info);
2497 out:
2498         mutex_unlock(&nbd_index_mutex);
2499         return ret;
2500 }
2501
2502 static void nbd_connect_reply(struct genl_info *info, int index)
2503 {
2504         struct sk_buff *skb;
2505         void *msg_head;
2506         int ret;
2507
2508         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2509         if (!skb)
2510                 return;
2511         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2512                                      NBD_CMD_CONNECT);
2513         if (!msg_head) {
2514                 nlmsg_free(skb);
2515                 return;
2516         }
2517         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2518         if (ret) {
2519                 nlmsg_free(skb);
2520                 return;
2521         }
2522         genlmsg_end(skb, msg_head);
2523         genlmsg_reply(skb, info);
2524 }
2525
2526 static void nbd_mcast_index(int index)
2527 {
2528         struct sk_buff *skb;
2529         void *msg_head;
2530         int ret;
2531
2532         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2533         if (!skb)
2534                 return;
2535         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2536                                      NBD_CMD_LINK_DEAD);
2537         if (!msg_head) {
2538                 nlmsg_free(skb);
2539                 return;
2540         }
2541         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2542         if (ret) {
2543                 nlmsg_free(skb);
2544                 return;
2545         }
2546         genlmsg_end(skb, msg_head);
2547         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2548 }
2549
2550 static void nbd_dead_link_work(struct work_struct *work)
2551 {
2552         struct link_dead_args *args = container_of(work, struct link_dead_args,
2553                                                    work);
2554         nbd_mcast_index(args->index);
2555         kfree(args);
2556 }
2557
2558 static int __init nbd_init(void)
2559 {
2560         int i;
2561
2562         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2563
2564         if (max_part < 0) {
2565                 pr_err("max_part must be >= 0\n");
2566                 return -EINVAL;
2567         }
2568
2569         part_shift = 0;
2570         if (max_part > 0) {
2571                 part_shift = fls(max_part);
2572
2573                 /*
2574                  * Adjust max_part according to part_shift as it is exported
2575                  * to user space so that user can know the max number of
2576                  * partition kernel should be able to manage.
2577                  *
2578                  * Note that -1 is required because partition 0 is reserved
2579                  * for the whole disk.
2580                  */
2581                 max_part = (1UL << part_shift) - 1;
2582         }
2583
2584         if ((1UL << part_shift) > DISK_MAX_PARTS)
2585                 return -EINVAL;
2586
2587         if (nbds_max > 1UL << (MINORBITS - part_shift))
2588                 return -EINVAL;
2589
2590         if (register_blkdev(NBD_MAJOR, "nbd"))
2591                 return -EIO;
2592
2593         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2594         if (!nbd_del_wq) {
2595                 unregister_blkdev(NBD_MAJOR, "nbd");
2596                 return -ENOMEM;
2597         }
2598
2599         if (genl_register_family(&nbd_genl_family)) {
2600                 destroy_workqueue(nbd_del_wq);
2601                 unregister_blkdev(NBD_MAJOR, "nbd");
2602                 return -EINVAL;
2603         }
2604         nbd_dbg_init();
2605
2606         for (i = 0; i < nbds_max; i++)
2607                 nbd_dev_add(i, 1);
2608         return 0;
2609 }
2610
2611 static int nbd_exit_cb(int id, void *ptr, void *data)
2612 {
2613         struct list_head *list = (struct list_head *)data;
2614         struct nbd_device *nbd = ptr;
2615
2616         /* Skip nbd that is being removed asynchronously */
2617         if (refcount_read(&nbd->refs))
2618                 list_add_tail(&nbd->list, list);
2619
2620         return 0;
2621 }
2622
2623 static void __exit nbd_cleanup(void)
2624 {
2625         struct nbd_device *nbd;
2626         LIST_HEAD(del_list);
2627
2628         /*
2629          * Unregister netlink interface prior to waiting
2630          * for the completion of netlink commands.
2631          */
2632         genl_unregister_family(&nbd_genl_family);
2633
2634         nbd_dbg_close();
2635
2636         mutex_lock(&nbd_index_mutex);
2637         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2638         mutex_unlock(&nbd_index_mutex);
2639
2640         while (!list_empty(&del_list)) {
2641                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2642                 list_del_init(&nbd->list);
2643                 if (refcount_read(&nbd->config_refs))
2644                         pr_err("possibly leaking nbd_config (ref %d)\n",
2645                                         refcount_read(&nbd->config_refs));
2646                 if (refcount_read(&nbd->refs) != 1)
2647                         pr_err("possibly leaking a device\n");
2648                 nbd_put(nbd);
2649         }
2650
2651         /* Also wait for nbd_dev_remove_work() completes */
2652         destroy_workqueue(nbd_del_wq);
2653
2654         idr_destroy(&nbd_index_idr);
2655         unregister_blkdev(NBD_MAJOR, "nbd");
2656 }
2657
2658 module_init(nbd_init);
2659 module_exit(nbd_cleanup);
2660
2661 MODULE_DESCRIPTION("Network Block Device");
2662 MODULE_LICENSE("GPL");
2663
2664 module_param(nbds_max, int, 0444);
2665 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2666 module_param(max_part, int, 0444);
2667 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");