drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   9  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #define pr_fmt(fmt) "nbd: " fmt
  15
  16 #include <linux/major.h>
  17
  18 #include <linux/blkdev.h>
  19 #include <linux/module.h>
  20 #include <linux/init.h>
  21 #include <linux/sched.h>
  22 #include <linux/sched/mm.h>
  23 #include <linux/fs.h>
  24 #include <linux/bio.h>
  25 #include <linux/stat.h>
  26 #include <linux/errno.h>
  27 #include <linux/file.h>
  28 #include <linux/ioctl.h>
  29 #include <linux/mutex.h>
  30 #include <linux/compiler.h>
  31 #include <linux/completion.h>
  32 #include <linux/err.h>
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <net/sock.h>
  36 #include <linux/net.h>
  37 #include <linux/kthread.h>
  38 #include <linux/types.h>
  39 #include <linux/debugfs.h>
  40 #include <linux/blk-mq.h>
  41
  42 #include <linux/uaccess.h>
  43 #include <asm/types.h>
  44
  45 #include <linux/nbd.h>
  46 #include <linux/nbd-netlink.h>
  47 #include <net/genetlink.h>
  48
  49 #define CREATE_TRACE_POINTS
  50 #include <trace/events/nbd.h>
  51
  52 static DEFINE_IDR(nbd_index_idr);
  53 static DEFINE_MUTEX(nbd_index_mutex);
  54 static struct workqueue_struct *nbd_del_wq;
  55 static int nbd_total_devices = 0;
  56
  57 struct nbd_sock {
  58         struct socket *sock;
  59         struct mutex tx_lock;
  60         struct request *pending;
  61         int sent;
  62         bool dead;
  63         int fallback_index;
  64         int cookie;
  65 };
  66
  67 struct recv_thread_args {
  68         struct work_struct work;
  69         struct nbd_device *nbd;
  70         struct nbd_sock *nsock;
  71         int index;
  72 };
  73
  74 struct link_dead_args {
  75         struct work_struct work;
  76         int index;
  77 };
  78
  79 #define NBD_RT_TIMEDOUT                 0
  80 #define NBD_RT_DISCONNECT_REQUESTED     1
  81 #define NBD_RT_DISCONNECTED             2
  82 #define NBD_RT_HAS_PID_FILE             3
  83 #define NBD_RT_HAS_CONFIG_REF           4
  84 #define NBD_RT_BOUND                    5
  85 #define NBD_RT_DISCONNECT_ON_CLOSE      6
  86 #define NBD_RT_HAS_BACKEND_FILE         7
  87
  88 #define NBD_DESTROY_ON_DISCONNECT       0
  89 #define NBD_DISCONNECT_REQUESTED        1
  90
  91 struct nbd_config {
  92         u32 flags;
  93         unsigned long runtime_flags;
  94         u64 dead_conn_timeout;
  95
  96         struct nbd_sock **socks;
  97         int num_connections;
  98         atomic_t live_connections;
  99         wait_queue_head_t conn_wait;
 100
 101         atomic_t recv_threads;
 102         wait_queue_head_t recv_wq;
 103         unsigned int blksize_bits;
 104         loff_t bytesize;
 105 #if IS_ENABLED(CONFIG_DEBUG_FS)
 106         struct dentry *dbg_dir;
 107 #endif
 108 };
 109
 110 static inline unsigned int nbd_blksize(struct nbd_config *config)
 111 {
 112         return 1u << config->blksize_bits;
 113 }
 114
 115 struct nbd_device {
 116         struct blk_mq_tag_set tag_set;
 117
 118         int index;
 119         refcount_t config_refs;
 120         refcount_t refs;
 121         struct nbd_config *config;
 122         struct mutex config_lock;
 123         struct gendisk *disk;
 124         struct workqueue_struct *recv_workq;
 125         struct work_struct remove_work;
 126
 127         struct list_head list;
 128         struct task_struct *task_setup;
 129
 130         unsigned long flags;
 131         pid_t pid; /* pid of nbd-client, if attached */
 132
 133         char *backend;
 134 };
 135
 136 #define NBD_CMD_REQUEUED        1
 137 /*
 138  * This flag will be set if nbd_queue_rq() succeed, and will be checked and
 139  * cleared in completion. Both setting and clearing of the flag are protected
 140  * by cmd->lock.
 141  */
 142 #define NBD_CMD_INFLIGHT        2
 143
 144 struct nbd_cmd {
 145         struct nbd_device *nbd;
 146         struct mutex lock;
 147         int index;
 148         int cookie;
 149         int retries;
 150         blk_status_t status;
 151         unsigned long flags;
 152         u32 cmd_cookie;
 153 };
 154
 155 #if IS_ENABLED(CONFIG_DEBUG_FS)
 156 static struct dentry *nbd_dbg_dir;
 157 #endif
 158
 159 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 160
 161 #define NBD_DEF_BLKSIZE_BITS 10
 162
 163 static unsigned int nbds_max = 16;
 164 static int max_part = 16;
 165 static int part_shift;
 166
 167 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 168 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 169 static void nbd_config_put(struct nbd_device *nbd);
 170 static void nbd_connect_reply(struct genl_info *info, int index);
 171 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 172 static void nbd_dead_link_work(struct work_struct *work);
 173 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 174
 175 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 176 {
 177         return disk_to_dev(nbd->disk);
 178 }
 179
 180 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 181 {
 182         struct request *req = blk_mq_rq_from_pdu(cmd);
 183
 184         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 185                 blk_mq_requeue_request(req, true);
 186 }
 187
 188 #define NBD_COOKIE_BITS 32
 189
 190 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 191 {
 192         struct request *req = blk_mq_rq_from_pdu(cmd);
 193         u32 tag = blk_mq_unique_tag(req);
 194         u64 cookie = cmd->cmd_cookie;
 195
 196         return (cookie << NBD_COOKIE_BITS) | tag;
 197 }
 198
 199 static u32 nbd_handle_to_tag(u64 handle)
 200 {
 201         return (u32)handle;
 202 }
 203
 204 static u32 nbd_handle_to_cookie(u64 handle)
 205 {
 206         return (u32)(handle >> NBD_COOKIE_BITS);
 207 }
 208
 209 static const char *nbdcmd_to_ascii(int cmd)
 210 {
 211         switch (cmd) {
 212         case  NBD_CMD_READ: return "read";
 213         case NBD_CMD_WRITE: return "write";
 214         case  NBD_CMD_DISC: return "disconnect";
 215         case NBD_CMD_FLUSH: return "flush";
 216         case  NBD_CMD_TRIM: return "trim/discard";
 217         }
 218         return "invalid";
 219 }
 220
 221 static ssize_t pid_show(struct device *dev,
 222                         struct device_attribute *attr, char *buf)
 223 {
 224         struct gendisk *disk = dev_to_disk(dev);
 225         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 226
 227         return sprintf(buf, "%d\n", nbd->pid);
 228 }
 229
 230 static const struct device_attribute pid_attr = {
 231         .attr = { .name = "pid", .mode = 0444},
 232         .show = pid_show,
 233 };
 234
 235 static ssize_t backend_show(struct device *dev,
 236                 struct device_attribute *attr, char *buf)
 237 {
 238         struct gendisk *disk = dev_to_disk(dev);
 239         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 240
 241         return sprintf(buf, "%s\n", nbd->backend ?: "");
 242 }
 243
 244 static const struct device_attribute backend_attr = {
 245         .attr = { .name = "backend", .mode = 0444},
 246         .show = backend_show,
 247 };
 248
 249 static void nbd_dev_remove(struct nbd_device *nbd)
 250 {
 251         struct gendisk *disk = nbd->disk;
 252
 253         del_gendisk(disk);
 254         blk_mq_free_tag_set(&nbd->tag_set);
 255
 256         /*
 257          * Remove from idr after del_gendisk() completes, so if the same ID is
 258          * reused, the following add_disk() will succeed.
 259          */
 260         mutex_lock(&nbd_index_mutex);
 261         idr_remove(&nbd_index_idr, nbd->index);
 262         mutex_unlock(&nbd_index_mutex);
 263         destroy_workqueue(nbd->recv_workq);
 264         put_disk(disk);
 265 }
 266
 267 static void nbd_dev_remove_work(struct work_struct *work)
 268 {
 269         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 270 }
 271
 272 static void nbd_put(struct nbd_device *nbd)
 273 {
 274         if (!refcount_dec_and_test(&nbd->refs))
 275                 return;
 276
 277         /* Call del_gendisk() asynchrounously to prevent deadlock */
 278         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 279                 queue_work(nbd_del_wq, &nbd->remove_work);
 280         else
 281                 nbd_dev_remove(nbd);
 282 }
 283
 284 static int nbd_disconnected(struct nbd_config *config)
 285 {
 286         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 287                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 288 }
 289
 290 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 291                                 int notify)
 292 {
 293         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 294                 struct link_dead_args *args;
 295                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 296                 if (args) {
 297                         INIT_WORK(&args->work, nbd_dead_link_work);
 298                         args->index = nbd->index;
 299                         queue_work(system_wq, &args->work);
 300                 }
 301         }
 302         if (!nsock->dead) {
 303                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 304                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 305                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 306                                                &nbd->config->runtime_flags)) {
 307                                 set_bit(NBD_RT_DISCONNECTED,
 308                                         &nbd->config->runtime_flags);
 309                                 dev_info(nbd_to_dev(nbd),
 310                                         "Disconnected due to user request.\n");
 311                         }
 312                 }
 313         }
 314         nsock->dead = true;
 315         nsock->pending = NULL;
 316         nsock->sent = 0;
 317 }
 318
 319 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 320                 loff_t blksize)
 321 {
 322         if (!blksize)
 323                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 324
 325         if (blk_validate_block_size(blksize))
 326                 return -EINVAL;
 327
 328         if (bytesize < 0)
 329                 return -EINVAL;
 330
 331         nbd->config->bytesize = bytesize;
 332         nbd->config->blksize_bits = __ffs(blksize);
 333
 334         if (!nbd->pid)
 335                 return 0;
 336
 337         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 338                 nbd->disk->queue->limits.discard_granularity = blksize;
 339                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 340         }
 341         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 342         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 343
 344         if (max_part)
 345                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 346         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 347                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 348         return 0;
 349 }
 350
 351 static void nbd_complete_rq(struct request *req)
 352 {
 353         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 354
 355         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 356                 cmd->status ? "failed" : "done");
 357
 358         blk_mq_end_request(req, cmd->status);
 359 }
 360
 361 /*
 362  * Forcibly shutdown the socket causing all listeners to error
 363  */
 364 static void sock_shutdown(struct nbd_device *nbd)
 365 {
 366         struct nbd_config *config = nbd->config;
 367         int i;
 368
 369         if (config->num_connections == 0)
 370                 return;
 371         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 372                 return;
 373
 374         for (i = 0; i < config->num_connections; i++) {
 375                 struct nbd_sock *nsock = config->socks[i];
 376                 mutex_lock(&nsock->tx_lock);
 377                 nbd_mark_nsock_dead(nbd, nsock, 0);
 378                 mutex_unlock(&nsock->tx_lock);
 379         }
 380         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 381 }
 382
 383 static u32 req_to_nbd_cmd_type(struct request *req)
 384 {
 385         switch (req_op(req)) {
 386         case REQ_OP_DISCARD:
 387                 return NBD_CMD_TRIM;
 388         case REQ_OP_FLUSH:
 389                 return NBD_CMD_FLUSH;
 390         case REQ_OP_WRITE:
 391                 return NBD_CMD_WRITE;
 392         case REQ_OP_READ:
 393                 return NBD_CMD_READ;
 394         default:
 395                 return U32_MAX;
 396         }
 397 }
 398
 399 static struct nbd_config *nbd_get_config_unlocked(struct nbd_device *nbd)
 400 {
 401         if (refcount_inc_not_zero(&nbd->config_refs)) {
 402                 /*
 403                  * Add smp_mb__after_atomic to ensure that reading nbd->config_refs
 404                  * and reading nbd->config is ordered. The pair is the barrier in
 405                  * nbd_alloc_and_init_config(), avoid nbd->config_refs is set
 406                  * before nbd->config.
 407                  */
 408                 smp_mb__after_atomic();
 409                 return nbd->config;
 410         }
 411
 412         return NULL;
 413 }
 414
 415 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req)
 416 {
 417         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 418         struct nbd_device *nbd = cmd->nbd;
 419         struct nbd_config *config;
 420
 421         if (!mutex_trylock(&cmd->lock))
 422                 return BLK_EH_RESET_TIMER;
 423
 424         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 425                 mutex_unlock(&cmd->lock);
 426                 return BLK_EH_DONE;
 427         }
 428
 429         config = nbd_get_config_unlocked(nbd);
 430         if (!config) {
 431                 cmd->status = BLK_STS_TIMEOUT;
 432                 __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 433                 mutex_unlock(&cmd->lock);
 434                 goto done;
 435         }
 436
 437         if (config->num_connections > 1 ||
 438             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 439                 dev_err_ratelimited(nbd_to_dev(nbd),
 440                                     "Connection timed out, retrying (%d/%d alive)\n",
 441                                     atomic_read(&config->live_connections),
 442                                     config->num_connections);
 443                 /*
 444                  * Hooray we have more connections, requeue this IO, the submit
 445                  * path will put it on a real connection. Or if only one
 446                  * connection is configured, the submit path will wait util
 447                  * a new connection is reconfigured or util dead timeout.
 448                  */
 449                 if (config->socks) {
 450                         if (cmd->index < config->num_connections) {
 451                                 struct nbd_sock *nsock =
 452                                         config->socks[cmd->index];
 453                                 mutex_lock(&nsock->tx_lock);
 454                                 /* We can have multiple outstanding requests, so
 455                                  * we don't want to mark the nsock dead if we've
 456                                  * already reconnected with a new socket, so
 457                                  * only mark it dead if its the same socket we
 458                                  * were sent out on.
 459                                  */
 460                                 if (cmd->cookie == nsock->cookie)
 461                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 462                                 mutex_unlock(&nsock->tx_lock);
 463                         }
 464                         mutex_unlock(&cmd->lock);
 465                         nbd_requeue_cmd(cmd);
 466                         nbd_config_put(nbd);
 467                         return BLK_EH_DONE;
 468                 }
 469         }
 470
 471         if (!nbd->tag_set.timeout) {
 472                 /*
 473                  * Userspace sets timeout=0 to disable socket disconnection,
 474                  * so just warn and reset the timer.
 475                  */
 476                 struct nbd_sock *nsock = config->socks[cmd->index];
 477                 cmd->retries++;
 478                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 479                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 480                         (unsigned long long)blk_rq_pos(req) << 9,
 481                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 482
 483                 mutex_lock(&nsock->tx_lock);
 484                 if (cmd->cookie != nsock->cookie) {
 485                         nbd_requeue_cmd(cmd);
 486                         mutex_unlock(&nsock->tx_lock);
 487                         mutex_unlock(&cmd->lock);
 488                         nbd_config_put(nbd);
 489                         return BLK_EH_DONE;
 490                 }
 491                 mutex_unlock(&nsock->tx_lock);
 492                 mutex_unlock(&cmd->lock);
 493                 nbd_config_put(nbd);
 494                 return BLK_EH_RESET_TIMER;
 495         }
 496
 497         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 498         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 499         cmd->status = BLK_STS_IOERR;
 500         __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 501         mutex_unlock(&cmd->lock);
 502         sock_shutdown(nbd);
 503         nbd_config_put(nbd);
 504 done:
 505         blk_mq_complete_request(req);
 506         return BLK_EH_DONE;
 507 }
 508
 509 static int __sock_xmit(struct nbd_device *nbd, struct socket *sock, int send,
 510                        struct iov_iter *iter, int msg_flags, int *sent)
 511 {
 512         int result;
 513         struct msghdr msg;
 514         unsigned int noreclaim_flag;
 515
 516         if (unlikely(!sock)) {
 517                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 518                         "Attempted %s on closed socket in sock_xmit\n",
 519                         (send ? "send" : "recv"));
 520                 return -EINVAL;
 521         }
 522
 523         msg.msg_iter = *iter;
 524
 525         noreclaim_flag = memalloc_noreclaim_save();
 526         do {
 527                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 528                 sock->sk->sk_use_task_frag = false;
 529                 msg.msg_name = NULL;
 530                 msg.msg_namelen = 0;
 531                 msg.msg_control = NULL;
 532                 msg.msg_controllen = 0;
 533                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 534
 535                 if (send)
 536                         result = sock_sendmsg(sock, &msg);
 537                 else
 538                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 539
 540                 if (result <= 0) {
 541                         if (result == 0)
 542                                 result = -EPIPE; /* short read */
 543                         break;
 544                 }
 545                 if (sent)
 546                         *sent += result;
 547         } while (msg_data_left(&msg));
 548
 549         memalloc_noreclaim_restore(noreclaim_flag);
 550
 551         return result;
 552 }
 553
 554 /*
 555  *  Send or receive packet. Return a positive value on success and
 556  *  negtive value on failure, and never return 0.
 557  */
 558 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 559                      struct iov_iter *iter, int msg_flags, int *sent)
 560 {
 561         struct nbd_config *config = nbd->config;
 562         struct socket *sock = config->socks[index]->sock;
 563
 564         return __sock_xmit(nbd, sock, send, iter, msg_flags, sent);
 565 }
 566
 567 /*
 568  * Different settings for sk->sk_sndtimeo can result in different return values
 569  * if there is a signal pending when we enter sendmsg, because reasons?
 570  */
 571 static inline int was_interrupted(int result)
 572 {
 573         return result == -ERESTARTSYS || result == -EINTR;
 574 }
 575
 576 /* always call with the tx_lock held */
 577 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 578 {
 579         struct request *req = blk_mq_rq_from_pdu(cmd);
 580         struct nbd_config *config = nbd->config;
 581         struct nbd_sock *nsock = config->socks[index];
 582         int result;
 583         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 584         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 585         struct iov_iter from;
 586         unsigned long size = blk_rq_bytes(req);
 587         struct bio *bio;
 588         u64 handle;
 589         u32 type;
 590         u32 nbd_cmd_flags = 0;
 591         int sent = nsock->sent, skip = 0;
 592
 593         iov_iter_kvec(&from, ITER_SOURCE, &iov, 1, sizeof(request));
 594
 595         type = req_to_nbd_cmd_type(req);
 596         if (type == U32_MAX)
 597                 return -EIO;
 598
 599         if (rq_data_dir(req) == WRITE &&
 600             (config->flags & NBD_FLAG_READ_ONLY)) {
 601                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 602                                     "Write on read-only\n");
 603                 return -EIO;
 604         }
 605
 606         if (req->cmd_flags & REQ_FUA)
 607                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 608
 609         /* We did a partial send previously, and we at least sent the whole
 610          * request struct, so just go and send the rest of the pages in the
 611          * request.
 612          */
 613         if (sent) {
 614                 if (sent >= sizeof(request)) {
 615                         skip = sent - sizeof(request);
 616
 617                         /* initialize handle for tracing purposes */
 618                         handle = nbd_cmd_handle(cmd);
 619
 620                         goto send_pages;
 621                 }
 622                 iov_iter_advance(&from, sent);
 623         } else {
 624                 cmd->cmd_cookie++;
 625         }
 626         cmd->index = index;
 627         cmd->cookie = nsock->cookie;
 628         cmd->retries = 0;
 629         request.type = htonl(type | nbd_cmd_flags);
 630         if (type != NBD_CMD_FLUSH) {
 631                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 632                 request.len = htonl(size);
 633         }
 634         handle = nbd_cmd_handle(cmd);
 635         request.cookie = cpu_to_be64(handle);
 636
 637         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 638
 639         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 640                 req, nbdcmd_to_ascii(type),
 641                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 642         result = sock_xmit(nbd, index, 1, &from,
 643                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 644         trace_nbd_header_sent(req, handle);
 645         if (result < 0) {
 646                 if (was_interrupted(result)) {
 647                         /* If we haven't sent anything we can just return BUSY,
 648                          * however if we have sent something we need to make
 649                          * sure we only allow this req to be sent until we are
 650                          * completely done.
 651                          */
 652                         if (sent) {
 653                                 nsock->pending = req;
 654                                 nsock->sent = sent;
 655                         }
 656                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 657                         return BLK_STS_RESOURCE;
 658                 }
 659                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 660                         "Send control failed (result %d)\n", result);
 661                 return -EAGAIN;
 662         }
 663 send_pages:
 664         if (type != NBD_CMD_WRITE)
 665                 goto out;
 666
 667         bio = req->bio;
 668         while (bio) {
 669                 struct bio *next = bio->bi_next;
 670                 struct bvec_iter iter;
 671                 struct bio_vec bvec;
 672
 673                 bio_for_each_segment(bvec, bio, iter) {
 674                         bool is_last = !next && bio_iter_last(bvec, iter);
 675                         int flags = is_last ? 0 : MSG_MORE;
 676
 677                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 678                                 req, bvec.bv_len);
 679                         iov_iter_bvec(&from, ITER_SOURCE, &bvec, 1, bvec.bv_len);
 680                         if (skip) {
 681                                 if (skip >= iov_iter_count(&from)) {
 682                                         skip -= iov_iter_count(&from);
 683                                         continue;
 684                                 }
 685                                 iov_iter_advance(&from, skip);
 686                                 skip = 0;
 687                         }
 688                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 689                         if (result < 0) {
 690                                 if (was_interrupted(result)) {
 691                                         /* We've already sent the header, we
 692                                          * have no choice but to set pending and
 693                                          * return BUSY.
 694                                          */
 695                                         nsock->pending = req;
 696                                         nsock->sent = sent;
 697                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 698                                         return BLK_STS_RESOURCE;
 699                                 }
 700                                 dev_err(disk_to_dev(nbd->disk),
 701                                         "Send data failed (result %d)\n",
 702                                         result);
 703                                 return -EAGAIN;
 704                         }
 705                         /*
 706                          * The completion might already have come in,
 707                          * so break for the last one instead of letting
 708                          * the iterator do it. This prevents use-after-free
 709                          * of the bio.
 710                          */
 711                         if (is_last)
 712                                 break;
 713                 }
 714                 bio = next;
 715         }
 716 out:
 717         trace_nbd_payload_sent(req, handle);
 718         nsock->pending = NULL;
 719         nsock->sent = 0;
 720         return 0;
 721 }
 722
 723 static int nbd_read_reply(struct nbd_device *nbd, struct socket *sock,
 724                           struct nbd_reply *reply)
 725 {
 726         struct kvec iov = {.iov_base = reply, .iov_len = sizeof(*reply)};
 727         struct iov_iter to;
 728         int result;
 729
 730         reply->magic = 0;
 731         iov_iter_kvec(&to, ITER_DEST, &iov, 1, sizeof(*reply));
 732         result = __sock_xmit(nbd, sock, 0, &to, MSG_WAITALL, NULL);
 733         if (result < 0) {
 734                 if (!nbd_disconnected(nbd->config))
 735                         dev_err(disk_to_dev(nbd->disk),
 736                                 "Receive control failed (result %d)\n", result);
 737                 return result;
 738         }
 739
 740         if (ntohl(reply->magic) != NBD_REPLY_MAGIC) {
 741                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 742                                 (unsigned long)ntohl(reply->magic));
 743                 return -EPROTO;
 744         }
 745
 746         return 0;
 747 }
 748
 749 /* NULL returned = something went wrong, inform userspace */
 750 static struct nbd_cmd *nbd_handle_reply(struct nbd_device *nbd, int index,
 751                                         struct nbd_reply *reply)
 752 {
 753         int result;
 754         struct nbd_cmd *cmd;
 755         struct request *req = NULL;
 756         u64 handle;
 757         u16 hwq;
 758         u32 tag;
 759         int ret = 0;
 760
 761         handle = be64_to_cpu(reply->cookie);
 762         tag = nbd_handle_to_tag(handle);
 763         hwq = blk_mq_unique_tag_to_hwq(tag);
 764         if (hwq < nbd->tag_set.nr_hw_queues)
 765                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 766                                        blk_mq_unique_tag_to_tag(tag));
 767         if (!req || !blk_mq_request_started(req)) {
 768                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 769                         tag, req);
 770                 return ERR_PTR(-ENOENT);
 771         }
 772         trace_nbd_header_received(req, handle);
 773         cmd = blk_mq_rq_to_pdu(req);
 774
 775         mutex_lock(&cmd->lock);
 776         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 777                 dev_err(disk_to_dev(nbd->disk), "Suspicious reply %d (status %u flags %lu)",
 778                         tag, cmd->status, cmd->flags);
 779                 ret = -ENOENT;
 780                 goto out;
 781         }
 782         if (cmd->index != index) {
 783                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply %d from different sock %d (expected %d)",
 784                         tag, index, cmd->index);
 785                 ret = -ENOENT;
 786                 goto out;
 787         }
 788         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 789                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 790                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 791                 ret = -ENOENT;
 792                 goto out;
 793         }
 794         if (cmd->status != BLK_STS_OK) {
 795                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 796                         req);
 797                 ret = -ENOENT;
 798                 goto out;
 799         }
 800         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 801                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 802                         req);
 803                 ret = -ENOENT;
 804                 goto out;
 805         }
 806         if (ntohl(reply->error)) {
 807                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 808                         ntohl(reply->error));
 809                 cmd->status = BLK_STS_IOERR;
 810                 goto out;
 811         }
 812
 813         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 814         if (rq_data_dir(req) != WRITE) {
 815                 struct req_iterator iter;
 816                 struct bio_vec bvec;
 817                 struct iov_iter to;
 818
 819                 rq_for_each_segment(bvec, req, iter) {
 820                         iov_iter_bvec(&to, ITER_DEST, &bvec, 1, bvec.bv_len);
 821                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 822                         if (result < 0) {
 823                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 824                                         result);
 825                                 /*
 826                                  * If we've disconnected, we need to make sure we
 827                                  * complete this request, otherwise error out
 828                                  * and let the timeout stuff handle resubmitting
 829                                  * this request onto another connection.
 830                                  */
 831                                 if (nbd_disconnected(nbd->config)) {
 832                                         cmd->status = BLK_STS_IOERR;
 833                                         goto out;
 834                                 }
 835                                 ret = -EIO;
 836                                 goto out;
 837                         }
 838                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 839                                 req, bvec.bv_len);
 840                 }
 841         }
 842 out:
 843         trace_nbd_payload_received(req, handle);
 844         mutex_unlock(&cmd->lock);
 845         return ret ? ERR_PTR(ret) : cmd;
 846 }
 847
 848 static void recv_work(struct work_struct *work)
 849 {
 850         struct recv_thread_args *args = container_of(work,
 851                                                      struct recv_thread_args,
 852                                                      work);
 853         struct nbd_device *nbd = args->nbd;
 854         struct nbd_config *config = nbd->config;
 855         struct request_queue *q = nbd->disk->queue;
 856         struct nbd_sock *nsock = args->nsock;
 857         struct nbd_cmd *cmd;
 858         struct request *rq;
 859
 860         while (1) {
 861                 struct nbd_reply reply;
 862
 863                 if (nbd_read_reply(nbd, nsock->sock, &reply))
 864                         break;
 865
 866                 /*
 867                  * Grab .q_usage_counter so request pool won't go away, then no
 868                  * request use-after-free is possible during nbd_handle_reply().
 869                  * If queue is frozen, there won't be any inflight requests, we
 870                  * needn't to handle the incoming garbage message.
 871                  */
 872                 if (!percpu_ref_tryget(&q->q_usage_counter)) {
 873                         dev_err(disk_to_dev(nbd->disk), "%s: no io inflight\n",
 874                                 __func__);
 875                         break;
 876                 }
 877
 878                 cmd = nbd_handle_reply(nbd, args->index, &reply);
 879                 if (IS_ERR(cmd)) {
 880                         percpu_ref_put(&q->q_usage_counter);
 881                         break;
 882                 }
 883
 884                 rq = blk_mq_rq_from_pdu(cmd);
 885                 if (likely(!blk_should_fake_timeout(rq->q))) {
 886                         bool complete;
 887
 888                         mutex_lock(&cmd->lock);
 889                         complete = __test_and_clear_bit(NBD_CMD_INFLIGHT,
 890                                                         &cmd->flags);
 891                         mutex_unlock(&cmd->lock);
 892                         if (complete)
 893                                 blk_mq_complete_request(rq);
 894                 }
 895                 percpu_ref_put(&q->q_usage_counter);
 896         }
 897
 898         mutex_lock(&nsock->tx_lock);
 899         nbd_mark_nsock_dead(nbd, nsock, 1);
 900         mutex_unlock(&nsock->tx_lock);
 901
 902         nbd_config_put(nbd);
 903         atomic_dec(&config->recv_threads);
 904         wake_up(&config->recv_wq);
 905         kfree(args);
 906 }
 907
 908 static bool nbd_clear_req(struct request *req, void *data)
 909 {
 910         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 911
 912         /* don't abort one completed request */
 913         if (blk_mq_request_completed(req))
 914                 return true;
 915
 916         mutex_lock(&cmd->lock);
 917         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 918                 mutex_unlock(&cmd->lock);
 919                 return true;
 920         }
 921         cmd->status = BLK_STS_IOERR;
 922         mutex_unlock(&cmd->lock);
 923
 924         blk_mq_complete_request(req);
 925         return true;
 926 }
 927
 928 static void nbd_clear_que(struct nbd_device *nbd)
 929 {
 930         blk_mq_quiesce_queue(nbd->disk->queue);
 931         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 932         blk_mq_unquiesce_queue(nbd->disk->queue);
 933         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 934 }
 935
 936 static int find_fallback(struct nbd_device *nbd, int index)
 937 {
 938         struct nbd_config *config = nbd->config;
 939         int new_index = -1;
 940         struct nbd_sock *nsock = config->socks[index];
 941         int fallback = nsock->fallback_index;
 942
 943         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 944                 return new_index;
 945
 946         if (config->num_connections <= 1) {
 947                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 948                                     "Dead connection, failed to find a fallback\n");
 949                 return new_index;
 950         }
 951
 952         if (fallback >= 0 && fallback < config->num_connections &&
 953             !config->socks[fallback]->dead)
 954                 return fallback;
 955
 956         if (nsock->fallback_index < 0 ||
 957             nsock->fallback_index >= config->num_connections ||
 958             config->socks[nsock->fallback_index]->dead) {
 959                 int i;
 960                 for (i = 0; i < config->num_connections; i++) {
 961                         if (i == index)
 962                                 continue;
 963                         if (!config->socks[i]->dead) {
 964                                 new_index = i;
 965                                 break;
 966                         }
 967                 }
 968                 nsock->fallback_index = new_index;
 969                 if (new_index < 0) {
 970                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 971                                             "Dead connection, failed to find a fallback\n");
 972                         return new_index;
 973                 }
 974         }
 975         new_index = nsock->fallback_index;
 976         return new_index;
 977 }
 978
 979 static int wait_for_reconnect(struct nbd_device *nbd)
 980 {
 981         struct nbd_config *config = nbd->config;
 982         if (!config->dead_conn_timeout)
 983                 return 0;
 984
 985         if (!wait_event_timeout(config->conn_wait,
 986                                 test_bit(NBD_RT_DISCONNECTED,
 987                                          &config->runtime_flags) ||
 988                                 atomic_read(&config->live_connections) > 0,
 989                                 config->dead_conn_timeout))
 990                 return 0;
 991
 992         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
 993 }
 994
 995 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 996 {
 997         struct request *req = blk_mq_rq_from_pdu(cmd);
 998         struct nbd_device *nbd = cmd->nbd;
 999         struct nbd_config *config;
1000         struct nbd_sock *nsock;
1001         int ret;
1002
1003         config = nbd_get_config_unlocked(nbd);
1004         if (!config) {
1005                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1006                                     "Socks array is empty\n");
1007                 return -EINVAL;
1008         }
1009
1010         if (index >= config->num_connections) {
1011                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1012                                     "Attempted send on invalid socket\n");
1013                 nbd_config_put(nbd);
1014                 return -EINVAL;
1015         }
1016         cmd->status = BLK_STS_OK;
1017 again:
1018         nsock = config->socks[index];
1019         mutex_lock(&nsock->tx_lock);
1020         if (nsock->dead) {
1021                 int old_index = index;
1022                 index = find_fallback(nbd, index);
1023                 mutex_unlock(&nsock->tx_lock);
1024                 if (index < 0) {
1025                         if (wait_for_reconnect(nbd)) {
1026                                 index = old_index;
1027                                 goto again;
1028                         }
1029                         /* All the sockets should already be down at this point,
1030                          * we just want to make sure that DISCONNECTED is set so
1031                          * any requests that come in that were queue'ed waiting
1032                          * for the reconnect timer don't trigger the timer again
1033                          * and instead just error out.
1034                          */
1035                         sock_shutdown(nbd);
1036                         nbd_config_put(nbd);
1037                         return -EIO;
1038                 }
1039                 goto again;
1040         }
1041
1042         /* Handle the case that we have a pending request that was partially
1043          * transmitted that _has_ to be serviced first.  We need to call requeue
1044          * here so that it gets put _after_ the request that is already on the
1045          * dispatch list.
1046          */
1047         blk_mq_start_request(req);
1048         if (unlikely(nsock->pending && nsock->pending != req)) {
1049                 nbd_requeue_cmd(cmd);
1050                 ret = 0;
1051                 goto out;
1052         }
1053         /*
1054          * Some failures are related to the link going down, so anything that
1055          * returns EAGAIN can be retried on a different socket.
1056          */
1057         ret = nbd_send_cmd(nbd, cmd, index);
1058         /*
1059          * Access to this flag is protected by cmd->lock, thus it's safe to set
1060          * the flag after nbd_send_cmd() succeed to send request to server.
1061          */
1062         if (!ret)
1063                 __set_bit(NBD_CMD_INFLIGHT, &cmd->flags);
1064         else if (ret == -EAGAIN) {
1065                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1066                                     "Request send failed, requeueing\n");
1067                 nbd_mark_nsock_dead(nbd, nsock, 1);
1068                 nbd_requeue_cmd(cmd);
1069                 ret = 0;
1070         }
1071 out:
1072         mutex_unlock(&nsock->tx_lock);
1073         nbd_config_put(nbd);
1074         return ret;
1075 }
1076
1077 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1078                         const struct blk_mq_queue_data *bd)
1079 {
1080         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1081         int ret;
1082
1083         /*
1084          * Since we look at the bio's to send the request over the network we
1085          * need to make sure the completion work doesn't mark this request done
1086          * before we are done doing our send.  This keeps us from dereferencing
1087          * freed data if we have particularly fast completions (ie we get the
1088          * completion before we exit sock_xmit on the last bvec) or in the case
1089          * that the server is misbehaving (or there was an error) before we're
1090          * done sending everything over the wire.
1091          */
1092         mutex_lock(&cmd->lock);
1093         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1094
1095         /* We can be called directly from the user space process, which means we
1096          * could possibly have signals pending so our sendmsg will fail.  In
1097          * this case we need to return that we are busy, otherwise error out as
1098          * appropriate.
1099          */
1100         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1101         if (ret < 0)
1102                 ret = BLK_STS_IOERR;
1103         else if (!ret)
1104                 ret = BLK_STS_OK;
1105         mutex_unlock(&cmd->lock);
1106
1107         return ret;
1108 }
1109
1110 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1111                                      int *err)
1112 {
1113         struct socket *sock;
1114
1115         *err = 0;
1116         sock = sockfd_lookup(fd, err);
1117         if (!sock)
1118                 return NULL;
1119
1120         if (sock->ops->shutdown == sock_no_shutdown) {
1121                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1122                 *err = -EINVAL;
1123                 sockfd_put(sock);
1124                 return NULL;
1125         }
1126
1127         return sock;
1128 }
1129
1130 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1131                           bool netlink)
1132 {
1133         struct nbd_config *config = nbd->config;
1134         struct socket *sock;
1135         struct nbd_sock **socks;
1136         struct nbd_sock *nsock;
1137         int err;
1138
1139         /* Arg will be cast to int, check it to avoid overflow */
1140         if (arg > INT_MAX)
1141                 return -EINVAL;
1142         sock = nbd_get_socket(nbd, arg, &err);
1143         if (!sock)
1144                 return err;
1145
1146         /*
1147          * We need to make sure we don't get any errant requests while we're
1148          * reallocating the ->socks array.
1149          */
1150         blk_mq_freeze_queue(nbd->disk->queue);
1151
1152         if (!netlink && !nbd->task_setup &&
1153             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1154                 nbd->task_setup = current;
1155
1156         if (!netlink &&
1157             (nbd->task_setup != current ||
1158              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1159                 dev_err(disk_to_dev(nbd->disk),
1160                         "Device being setup by another task");
1161                 err = -EBUSY;
1162                 goto put_socket;
1163         }
1164
1165         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1166         if (!nsock) {
1167                 err = -ENOMEM;
1168                 goto put_socket;
1169         }
1170
1171         socks = krealloc(config->socks, (config->num_connections + 1) *
1172                          sizeof(struct nbd_sock *), GFP_KERNEL);
1173         if (!socks) {
1174                 kfree(nsock);
1175                 err = -ENOMEM;
1176                 goto put_socket;
1177         }
1178
1179         config->socks = socks;
1180
1181         nsock->fallback_index = -1;
1182         nsock->dead = false;
1183         mutex_init(&nsock->tx_lock);
1184         nsock->sock = sock;
1185         nsock->pending = NULL;
1186         nsock->sent = 0;
1187         nsock->cookie = 0;
1188         socks[config->num_connections++] = nsock;
1189         atomic_inc(&config->live_connections);
1190         blk_mq_unfreeze_queue(nbd->disk->queue);
1191
1192         return 0;
1193
1194 put_socket:
1195         blk_mq_unfreeze_queue(nbd->disk->queue);
1196         sockfd_put(sock);
1197         return err;
1198 }
1199
1200 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1201 {
1202         struct nbd_config *config = nbd->config;
1203         struct socket *sock, *old;
1204         struct recv_thread_args *args;
1205         int i;
1206         int err;
1207
1208         sock = nbd_get_socket(nbd, arg, &err);
1209         if (!sock)
1210                 return err;
1211
1212         args = kzalloc(sizeof(*args), GFP_KERNEL);
1213         if (!args) {
1214                 sockfd_put(sock);
1215                 return -ENOMEM;
1216         }
1217
1218         for (i = 0; i < config->num_connections; i++) {
1219                 struct nbd_sock *nsock = config->socks[i];
1220
1221                 if (!nsock->dead)
1222                         continue;
1223
1224                 mutex_lock(&nsock->tx_lock);
1225                 if (!nsock->dead) {
1226                         mutex_unlock(&nsock->tx_lock);
1227                         continue;
1228                 }
1229                 sk_set_memalloc(sock->sk);
1230                 if (nbd->tag_set.timeout)
1231                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1232                 atomic_inc(&config->recv_threads);
1233                 refcount_inc(&nbd->config_refs);
1234                 old = nsock->sock;
1235                 nsock->fallback_index = -1;
1236                 nsock->sock = sock;
1237                 nsock->dead = false;
1238                 INIT_WORK(&args->work, recv_work);
1239                 args->index = i;
1240                 args->nbd = nbd;
1241                 args->nsock = nsock;
1242                 nsock->cookie++;
1243                 mutex_unlock(&nsock->tx_lock);
1244                 sockfd_put(old);
1245
1246                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1247
1248                 /* We take the tx_mutex in an error path in the recv_work, so we
1249                  * need to queue_work outside of the tx_mutex.
1250                  */
1251                 queue_work(nbd->recv_workq, &args->work);
1252
1253                 atomic_inc(&config->live_connections);
1254                 wake_up(&config->conn_wait);
1255                 return 0;
1256         }
1257         sockfd_put(sock);
1258         kfree(args);
1259         return -ENOSPC;
1260 }
1261
1262 static void nbd_bdev_reset(struct nbd_device *nbd)
1263 {
1264         if (disk_openers(nbd->disk) > 1)
1265                 return;
1266         set_capacity(nbd->disk, 0);
1267 }
1268
1269 static void nbd_parse_flags(struct nbd_device *nbd)
1270 {
1271         struct nbd_config *config = nbd->config;
1272         if (config->flags & NBD_FLAG_READ_ONLY)
1273                 set_disk_ro(nbd->disk, true);
1274         else
1275                 set_disk_ro(nbd->disk, false);
1276         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1277                 if (config->flags & NBD_FLAG_SEND_FUA)
1278                         blk_queue_write_cache(nbd->disk->queue, true, true);
1279                 else
1280                         blk_queue_write_cache(nbd->disk->queue, true, false);
1281         }
1282         else
1283                 blk_queue_write_cache(nbd->disk->queue, false, false);
1284 }
1285
1286 static void send_disconnects(struct nbd_device *nbd)
1287 {
1288         struct nbd_config *config = nbd->config;
1289         struct nbd_request request = {
1290                 .magic = htonl(NBD_REQUEST_MAGIC),
1291                 .type = htonl(NBD_CMD_DISC),
1292         };
1293         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1294         struct iov_iter from;
1295         int i, ret;
1296
1297         for (i = 0; i < config->num_connections; i++) {
1298                 struct nbd_sock *nsock = config->socks[i];
1299
1300                 iov_iter_kvec(&from, ITER_SOURCE, &iov, 1, sizeof(request));
1301                 mutex_lock(&nsock->tx_lock);
1302                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1303                 if (ret < 0)
1304                         dev_err(disk_to_dev(nbd->disk),
1305                                 "Send disconnect failed %d\n", ret);
1306                 mutex_unlock(&nsock->tx_lock);
1307         }
1308 }
1309
1310 static int nbd_disconnect(struct nbd_device *nbd)
1311 {
1312         struct nbd_config *config = nbd->config;
1313
1314         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1315         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1316         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1317         send_disconnects(nbd);
1318         return 0;
1319 }
1320
1321 static void nbd_clear_sock(struct nbd_device *nbd)
1322 {
1323         sock_shutdown(nbd);
1324         nbd_clear_que(nbd);
1325         nbd->task_setup = NULL;
1326 }
1327
1328 static void nbd_config_put(struct nbd_device *nbd)
1329 {
1330         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1331                                         &nbd->config_lock)) {
1332                 struct nbd_config *config = nbd->config;
1333                 nbd_dev_dbg_close(nbd);
1334                 invalidate_disk(nbd->disk);
1335                 if (nbd->config->bytesize)
1336                         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
1337                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1338                                        &config->runtime_flags))
1339                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1340                 nbd->pid = 0;
1341                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1342                                        &config->runtime_flags)) {
1343                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1344                         kfree(nbd->backend);
1345                         nbd->backend = NULL;
1346                 }
1347                 nbd_clear_sock(nbd);
1348                 if (config->num_connections) {
1349                         int i;
1350                         for (i = 0; i < config->num_connections; i++) {
1351                                 sockfd_put(config->socks[i]->sock);
1352                                 kfree(config->socks[i]);
1353                         }
1354                         kfree(config->socks);
1355                 }
1356                 kfree(nbd->config);
1357                 nbd->config = NULL;
1358
1359                 nbd->tag_set.timeout = 0;
1360                 nbd->disk->queue->limits.discard_granularity = 0;
1361                 blk_queue_max_discard_sectors(nbd->disk->queue, 0);
1362
1363                 mutex_unlock(&nbd->config_lock);
1364                 nbd_put(nbd);
1365                 module_put(THIS_MODULE);
1366         }
1367 }
1368
1369 static int nbd_start_device(struct nbd_device *nbd)
1370 {
1371         struct nbd_config *config = nbd->config;
1372         int num_connections = config->num_connections;
1373         int error = 0, i;
1374
1375         if (nbd->pid)
1376                 return -EBUSY;
1377         if (!config->socks)
1378                 return -EINVAL;
1379         if (num_connections > 1 &&
1380             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1381                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1382                 return -EINVAL;
1383         }
1384
1385         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1386         nbd->pid = task_pid_nr(current);
1387
1388         nbd_parse_flags(nbd);
1389
1390         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1391         if (error) {
1392                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1393                 return error;
1394         }
1395         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1396
1397         nbd_dev_dbg_init(nbd);
1398         for (i = 0; i < num_connections; i++) {
1399                 struct recv_thread_args *args;
1400
1401                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1402                 if (!args) {
1403                         sock_shutdown(nbd);
1404                         /*
1405                          * If num_connections is m (2 < m),
1406                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1407                          * But NO.(n + 1) failed. We still have n recv threads.
1408                          * So, add flush_workqueue here to prevent recv threads
1409                          * dropping the last config_refs and trying to destroy
1410                          * the workqueue from inside the workqueue.
1411                          */
1412                         if (i)
1413                                 flush_workqueue(nbd->recv_workq);
1414                         return -ENOMEM;
1415                 }
1416                 sk_set_memalloc(config->socks[i]->sock->sk);
1417                 if (nbd->tag_set.timeout)
1418                         config->socks[i]->sock->sk->sk_sndtimeo =
1419                                 nbd->tag_set.timeout;
1420                 atomic_inc(&config->recv_threads);
1421                 refcount_inc(&nbd->config_refs);
1422                 INIT_WORK(&args->work, recv_work);
1423                 args->nbd = nbd;
1424                 args->nsock = config->socks[i];
1425                 args->index = i;
1426                 queue_work(nbd->recv_workq, &args->work);
1427         }
1428         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1429 }
1430
1431 static int nbd_start_device_ioctl(struct nbd_device *nbd)
1432 {
1433         struct nbd_config *config = nbd->config;
1434         int ret;
1435
1436         ret = nbd_start_device(nbd);
1437         if (ret)
1438                 return ret;
1439
1440         if (max_part)
1441                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1442         mutex_unlock(&nbd->config_lock);
1443         ret = wait_event_interruptible(config->recv_wq,
1444                                          atomic_read(&config->recv_threads) == 0);
1445         if (ret) {
1446                 sock_shutdown(nbd);
1447                 nbd_clear_que(nbd);
1448         }
1449
1450         flush_workqueue(nbd->recv_workq);
1451         mutex_lock(&nbd->config_lock);
1452         nbd_bdev_reset(nbd);
1453         /* user requested, ignore socket errors */
1454         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1455                 ret = 0;
1456         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1457                 ret = -ETIMEDOUT;
1458         return ret;
1459 }
1460
1461 static void nbd_clear_sock_ioctl(struct nbd_device *nbd)
1462 {
1463         nbd_clear_sock(nbd);
1464         disk_force_media_change(nbd->disk);
1465         nbd_bdev_reset(nbd);
1466         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1467                                &nbd->config->runtime_flags))
1468                 nbd_config_put(nbd);
1469 }
1470
1471 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1472 {
1473         nbd->tag_set.timeout = timeout * HZ;
1474         if (timeout)
1475                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1476         else
1477                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1478 }
1479
1480 /* Must be called with config_lock held */
1481 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1482                        unsigned int cmd, unsigned long arg)
1483 {
1484         struct nbd_config *config = nbd->config;
1485         loff_t bytesize;
1486
1487         switch (cmd) {
1488         case NBD_DISCONNECT:
1489                 return nbd_disconnect(nbd);
1490         case NBD_CLEAR_SOCK:
1491                 nbd_clear_sock_ioctl(nbd);
1492                 return 0;
1493         case NBD_SET_SOCK:
1494                 return nbd_add_socket(nbd, arg, false);
1495         case NBD_SET_BLKSIZE:
1496                 return nbd_set_size(nbd, config->bytesize, arg);
1497         case NBD_SET_SIZE:
1498                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1499         case NBD_SET_SIZE_BLOCKS:
1500                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1501                         return -EINVAL;
1502                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1503         case NBD_SET_TIMEOUT:
1504                 nbd_set_cmd_timeout(nbd, arg);
1505                 return 0;
1506
1507         case NBD_SET_FLAGS:
1508                 config->flags = arg;
1509                 return 0;
1510         case NBD_DO_IT:
1511                 return nbd_start_device_ioctl(nbd);
1512         case NBD_CLEAR_QUE:
1513                 /*
1514                  * This is for compatibility only.  The queue is always cleared
1515                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1516                  */
1517                 return 0;
1518         case NBD_PRINT_DEBUG:
1519                 /*
1520                  * For compatibility only, we no longer keep a list of
1521                  * outstanding requests.
1522                  */
1523                 return 0;
1524         }
1525         return -ENOTTY;
1526 }
1527
1528 static int nbd_ioctl(struct block_device *bdev, blk_mode_t mode,
1529                      unsigned int cmd, unsigned long arg)
1530 {
1531         struct nbd_device *nbd = bdev->bd_disk->private_data;
1532         struct nbd_config *config = nbd->config;
1533         int error = -EINVAL;
1534
1535         if (!capable(CAP_SYS_ADMIN))
1536                 return -EPERM;
1537
1538         /* The block layer will pass back some non-nbd ioctls in case we have
1539          * special handling for them, but we don't so just return an error.
1540          */
1541         if (_IOC_TYPE(cmd) != 0xab)
1542                 return -EINVAL;
1543
1544         mutex_lock(&nbd->config_lock);
1545
1546         /* Don't allow ioctl operations on a nbd device that was created with
1547          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1548          */
1549         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1550             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1551                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1552         else
1553                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1554         mutex_unlock(&nbd->config_lock);
1555         return error;
1556 }
1557
1558 static int nbd_alloc_and_init_config(struct nbd_device *nbd)
1559 {
1560         struct nbd_config *config;
1561
1562         if (WARN_ON(nbd->config))
1563                 return -EINVAL;
1564
1565         if (!try_module_get(THIS_MODULE))
1566                 return -ENODEV;
1567
1568         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1569         if (!config) {
1570                 module_put(THIS_MODULE);
1571                 return -ENOMEM;
1572         }
1573
1574         atomic_set(&config->recv_threads, 0);
1575         init_waitqueue_head(&config->recv_wq);
1576         init_waitqueue_head(&config->conn_wait);
1577         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1578         atomic_set(&config->live_connections, 0);
1579
1580         nbd->config = config;
1581         /*
1582          * Order refcount_set(&nbd->config_refs, 1) and nbd->config assignment,
1583          * its pair is the barrier in nbd_get_config_unlocked().
1584          * So nbd_get_config_unlocked() won't see nbd->config as null after
1585          * refcount_inc_not_zero() succeed.
1586          */
1587         smp_mb__before_atomic();
1588         refcount_set(&nbd->config_refs, 1);
1589
1590         return 0;
1591 }
1592
1593 static int nbd_open(struct gendisk *disk, blk_mode_t mode)
1594 {
1595         struct nbd_device *nbd;
1596         struct nbd_config *config;
1597         int ret = 0;
1598
1599         mutex_lock(&nbd_index_mutex);
1600         nbd = disk->private_data;
1601         if (!nbd) {
1602                 ret = -ENXIO;
1603                 goto out;
1604         }
1605         if (!refcount_inc_not_zero(&nbd->refs)) {
1606                 ret = -ENXIO;
1607                 goto out;
1608         }
1609
1610         config = nbd_get_config_unlocked(nbd);
1611         if (!config) {
1612                 mutex_lock(&nbd->config_lock);
1613                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1614                         mutex_unlock(&nbd->config_lock);
1615                         goto out;
1616                 }
1617                 ret = nbd_alloc_and_init_config(nbd);
1618                 if (ret) {
1619                         mutex_unlock(&nbd->config_lock);
1620                         goto out;
1621                 }
1622
1623                 refcount_inc(&nbd->refs);
1624                 mutex_unlock(&nbd->config_lock);
1625                 if (max_part)
1626                         set_bit(GD_NEED_PART_SCAN, &disk->state);
1627         } else if (nbd_disconnected(config)) {
1628                 if (max_part)
1629                         set_bit(GD_NEED_PART_SCAN, &disk->state);
1630         }
1631 out:
1632         mutex_unlock(&nbd_index_mutex);
1633         return ret;
1634 }
1635
1636 static void nbd_release(struct gendisk *disk)
1637 {
1638         struct nbd_device *nbd = disk->private_data;
1639
1640         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1641                         disk_openers(disk) == 0)
1642                 nbd_disconnect_and_put(nbd);
1643
1644         nbd_config_put(nbd);
1645         nbd_put(nbd);
1646 }
1647
1648 static void nbd_free_disk(struct gendisk *disk)
1649 {
1650         struct nbd_device *nbd = disk->private_data;
1651
1652         kfree(nbd);
1653 }
1654
1655 static const struct block_device_operations nbd_fops =
1656 {
1657         .owner =        THIS_MODULE,
1658         .open =         nbd_open,
1659         .release =      nbd_release,
1660         .ioctl =        nbd_ioctl,
1661         .compat_ioctl = nbd_ioctl,
1662         .free_disk =    nbd_free_disk,
1663 };
1664
1665 #if IS_ENABLED(CONFIG_DEBUG_FS)
1666
1667 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1668 {
1669         struct nbd_device *nbd = s->private;
1670
1671         if (nbd->pid)
1672                 seq_printf(s, "recv: %d\n", nbd->pid);
1673
1674         return 0;
1675 }
1676
1677 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1678
1679 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1680 {
1681         struct nbd_device *nbd = s->private;
1682         u32 flags = nbd->config->flags;
1683
1684         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1685
1686         seq_puts(s, "Known flags:\n");
1687
1688         if (flags & NBD_FLAG_HAS_FLAGS)
1689                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1690         if (flags & NBD_FLAG_READ_ONLY)
1691                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1692         if (flags & NBD_FLAG_SEND_FLUSH)
1693                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1694         if (flags & NBD_FLAG_SEND_FUA)
1695                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1696         if (flags & NBD_FLAG_SEND_TRIM)
1697                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1698
1699         return 0;
1700 }
1701
1702 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1703
1704 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1705 {
1706         struct dentry *dir;
1707         struct nbd_config *config = nbd->config;
1708
1709         if (!nbd_dbg_dir)
1710                 return -EIO;
1711
1712         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1713         if (IS_ERR(dir)) {
1714                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1715                         nbd_name(nbd));
1716                 return -EIO;
1717         }
1718         config->dbg_dir = dir;
1719
1720         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1721         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1722         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1723         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1724         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1725
1726         return 0;
1727 }
1728
1729 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1730 {
1731         debugfs_remove_recursive(nbd->config->dbg_dir);
1732 }
1733
1734 static int nbd_dbg_init(void)
1735 {
1736         struct dentry *dbg_dir;
1737
1738         dbg_dir = debugfs_create_dir("nbd", NULL);
1739         if (IS_ERR(dbg_dir))
1740                 return -EIO;
1741
1742         nbd_dbg_dir = dbg_dir;
1743
1744         return 0;
1745 }
1746
1747 static void nbd_dbg_close(void)
1748 {
1749         debugfs_remove_recursive(nbd_dbg_dir);
1750 }
1751
1752 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1753
1754 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1755 {
1756         return 0;
1757 }
1758
1759 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1760 {
1761 }
1762
1763 static int nbd_dbg_init(void)
1764 {
1765         return 0;
1766 }
1767
1768 static void nbd_dbg_close(void)
1769 {
1770 }
1771
1772 #endif
1773
1774 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1775                             unsigned int hctx_idx, unsigned int numa_node)
1776 {
1777         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1778         cmd->nbd = set->driver_data;
1779         cmd->flags = 0;
1780         mutex_init(&cmd->lock);
1781         return 0;
1782 }
1783
1784 static const struct blk_mq_ops nbd_mq_ops = {
1785         .queue_rq       = nbd_queue_rq,
1786         .complete       = nbd_complete_rq,
1787         .init_request   = nbd_init_request,
1788         .timeout        = nbd_xmit_timeout,
1789 };
1790
1791 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1792 {
1793         struct nbd_device *nbd;
1794         struct gendisk *disk;
1795         int err = -ENOMEM;
1796
1797         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1798         if (!nbd)
1799                 goto out;
1800
1801         nbd->tag_set.ops = &nbd_mq_ops;
1802         nbd->tag_set.nr_hw_queues = 1;
1803         nbd->tag_set.queue_depth = 128;
1804         nbd->tag_set.numa_node = NUMA_NO_NODE;
1805         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1806         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1807                 BLK_MQ_F_BLOCKING;
1808         nbd->tag_set.driver_data = nbd;
1809         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1810         nbd->backend = NULL;
1811
1812         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1813         if (err)
1814                 goto out_free_nbd;
1815
1816         mutex_lock(&nbd_index_mutex);
1817         if (index >= 0) {
1818                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1819                                 GFP_KERNEL);
1820                 if (err == -ENOSPC)
1821                         err = -EEXIST;
1822         } else {
1823                 err = idr_alloc(&nbd_index_idr, nbd, 0,
1824                                 (MINORMASK >> part_shift) + 1, GFP_KERNEL);
1825                 if (err >= 0)
1826                         index = err;
1827         }
1828         nbd->index = index;
1829         mutex_unlock(&nbd_index_mutex);
1830         if (err < 0)
1831                 goto out_free_tags;
1832
1833         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1834         if (IS_ERR(disk)) {
1835                 err = PTR_ERR(disk);
1836                 goto out_free_idr;
1837         }
1838         nbd->disk = disk;
1839
1840         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1841                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1842                                           WQ_UNBOUND, 0, nbd->index);
1843         if (!nbd->recv_workq) {
1844                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1845                 err = -ENOMEM;
1846                 goto out_err_disk;
1847         }
1848
1849         /*
1850          * Tell the block layer that we are not a rotational device
1851          */
1852         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1853         disk->queue->limits.discard_granularity = 0;
1854         blk_queue_max_discard_sectors(disk->queue, 0);
1855         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1856         blk_queue_max_segments(disk->queue, USHRT_MAX);
1857         blk_queue_max_hw_sectors(disk->queue, 65536);
1858         disk->queue->limits.max_sectors = 256;
1859
1860         mutex_init(&nbd->config_lock);
1861         refcount_set(&nbd->config_refs, 0);
1862         /*
1863          * Start out with a zero references to keep other threads from using
1864          * this device until it is fully initialized.
1865          */
1866         refcount_set(&nbd->refs, 0);
1867         INIT_LIST_HEAD(&nbd->list);
1868         disk->major = NBD_MAJOR;
1869         disk->first_minor = index << part_shift;
1870         disk->minors = 1 << part_shift;
1871         disk->fops = &nbd_fops;
1872         disk->private_data = nbd;
1873         sprintf(disk->disk_name, "nbd%d", index);
1874         err = add_disk(disk);
1875         if (err)
1876                 goto out_free_work;
1877
1878         /*
1879          * Now publish the device.
1880          */
1881         refcount_set(&nbd->refs, refs);
1882         nbd_total_devices++;
1883         return nbd;
1884
1885 out_free_work:
1886         destroy_workqueue(nbd->recv_workq);
1887 out_err_disk:
1888         put_disk(disk);
1889 out_free_idr:
1890         mutex_lock(&nbd_index_mutex);
1891         idr_remove(&nbd_index_idr, index);
1892         mutex_unlock(&nbd_index_mutex);
1893 out_free_tags:
1894         blk_mq_free_tag_set(&nbd->tag_set);
1895 out_free_nbd:
1896         kfree(nbd);
1897 out:
1898         return ERR_PTR(err);
1899 }
1900
1901 static struct nbd_device *nbd_find_get_unused(void)
1902 {
1903         struct nbd_device *nbd;
1904         int id;
1905
1906         lockdep_assert_held(&nbd_index_mutex);
1907
1908         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1909                 if (refcount_read(&nbd->config_refs) ||
1910                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1911                         continue;
1912                 if (refcount_inc_not_zero(&nbd->refs))
1913                         return nbd;
1914         }
1915
1916         return NULL;
1917 }
1918
1919 /* Netlink interface. */
1920 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1921         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1922         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1923         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1924         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1925         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1926         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1927         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1928         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1929         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1930         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1931 };
1932
1933 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1934         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1935 };
1936
1937 /* We don't use this right now since we don't parse the incoming list, but we
1938  * still want it here so userspace knows what to expect.
1939  */
1940 static const struct nla_policy __attribute__((unused))
1941 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1942         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1943         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1944 };
1945
1946 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1947 {
1948         struct nbd_config *config = nbd->config;
1949         u64 bsize = nbd_blksize(config);
1950         u64 bytes = config->bytesize;
1951
1952         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1953                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1954
1955         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1956                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1957
1958         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1959                 return nbd_set_size(nbd, bytes, bsize);
1960         return 0;
1961 }
1962
1963 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1964 {
1965         struct nbd_device *nbd;
1966         struct nbd_config *config;
1967         int index = -1;
1968         int ret;
1969         bool put_dev = false;
1970
1971         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1972                 return -EPERM;
1973
1974         if (info->attrs[NBD_ATTR_INDEX]) {
1975                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1976
1977                 /*
1978                  * Too big first_minor can cause duplicate creation of
1979                  * sysfs files/links, since index << part_shift might overflow, or
1980                  * MKDEV() expect that the max bits of first_minor is 20.
1981                  */
1982                 if (index < 0 || index > MINORMASK >> part_shift) {
1983                         pr_err("illegal input index %d\n", index);
1984                         return -EINVAL;
1985                 }
1986         }
1987         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_SOCKETS)) {
1988                 pr_err("must specify at least one socket\n");
1989                 return -EINVAL;
1990         }
1991         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_SIZE_BYTES)) {
1992                 pr_err("must specify a size in bytes for the device\n");
1993                 return -EINVAL;
1994         }
1995 again:
1996         mutex_lock(&nbd_index_mutex);
1997         if (index == -1) {
1998                 nbd = nbd_find_get_unused();
1999         } else {
2000                 nbd = idr_find(&nbd_index_idr, index);
2001                 if (nbd) {
2002                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
2003                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
2004                             !refcount_inc_not_zero(&nbd->refs)) {
2005                                 mutex_unlock(&nbd_index_mutex);
2006                                 pr_err("device at index %d is going down\n",
2007                                         index);
2008                                 return -EINVAL;
2009                         }
2010                 }
2011         }
2012         mutex_unlock(&nbd_index_mutex);
2013
2014         if (!nbd) {
2015                 nbd = nbd_dev_add(index, 2);
2016                 if (IS_ERR(nbd)) {
2017                         pr_err("failed to add new device\n");
2018                         return PTR_ERR(nbd);
2019                 }
2020         }
2021
2022         mutex_lock(&nbd->config_lock);
2023         if (refcount_read(&nbd->config_refs)) {
2024                 mutex_unlock(&nbd->config_lock);
2025                 nbd_put(nbd);
2026                 if (index == -1)
2027                         goto again;
2028                 pr_err("nbd%d already in use\n", index);
2029                 return -EBUSY;
2030         }
2031
2032         ret = nbd_alloc_and_init_config(nbd);
2033         if (ret) {
2034                 mutex_unlock(&nbd->config_lock);
2035                 nbd_put(nbd);
2036                 pr_err("couldn't allocate config\n");
2037                 return ret;
2038         }
2039
2040         config = nbd->config;
2041         set_bit(NBD_RT_BOUND, &config->runtime_flags);
2042         ret = nbd_genl_size_set(info, nbd);
2043         if (ret)
2044                 goto out;
2045
2046         if (info->attrs[NBD_ATTR_TIMEOUT])
2047                 nbd_set_cmd_timeout(nbd,
2048                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2049         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2050                 config->dead_conn_timeout =
2051                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2052                 config->dead_conn_timeout *= HZ;
2053         }
2054         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
2055                 config->flags =
2056                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
2057         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2058                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2059                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2060                         /*
2061                          * We have 1 ref to keep the device around, and then 1
2062                          * ref for our current operation here, which will be
2063                          * inherited by the config.  If we already have
2064                          * DESTROY_ON_DISCONNECT set then we know we don't have
2065                          * that extra ref already held so we don't need the
2066                          * put_dev.
2067                          */
2068                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2069                                               &nbd->flags))
2070                                 put_dev = true;
2071                 } else {
2072                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2073                                                &nbd->flags))
2074                                 refcount_inc(&nbd->refs);
2075                 }
2076                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2077                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2078                                 &config->runtime_flags);
2079                 }
2080         }
2081
2082         if (info->attrs[NBD_ATTR_SOCKETS]) {
2083                 struct nlattr *attr;
2084                 int rem, fd;
2085
2086                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2087                                     rem) {
2088                         struct nlattr *socks[NBD_SOCK_MAX+1];
2089
2090                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2091                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2092                                 ret = -EINVAL;
2093                                 goto out;
2094                         }
2095                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2096                                                           attr,
2097                                                           nbd_sock_policy,
2098                                                           info->extack);
2099                         if (ret != 0) {
2100                                 pr_err("error processing sock list\n");
2101                                 ret = -EINVAL;
2102                                 goto out;
2103                         }
2104                         if (!socks[NBD_SOCK_FD])
2105                                 continue;
2106                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2107                         ret = nbd_add_socket(nbd, fd, true);
2108                         if (ret)
2109                                 goto out;
2110                 }
2111         }
2112         ret = nbd_start_device(nbd);
2113         if (ret)
2114                 goto out;
2115         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2116                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2117                                           GFP_KERNEL);
2118                 if (!nbd->backend) {
2119                         ret = -ENOMEM;
2120                         goto out;
2121                 }
2122         }
2123         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2124         if (ret) {
2125                 dev_err(disk_to_dev(nbd->disk),
2126                         "device_create_file failed for backend!\n");
2127                 goto out;
2128         }
2129         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2130 out:
2131         mutex_unlock(&nbd->config_lock);
2132         if (!ret) {
2133                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2134                 refcount_inc(&nbd->config_refs);
2135                 nbd_connect_reply(info, nbd->index);
2136         }
2137         nbd_config_put(nbd);
2138         if (put_dev)
2139                 nbd_put(nbd);
2140         return ret;
2141 }
2142
2143 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2144 {
2145         mutex_lock(&nbd->config_lock);
2146         nbd_disconnect(nbd);
2147         sock_shutdown(nbd);
2148         wake_up(&nbd->config->conn_wait);
2149         /*
2150          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2151          * to cancel the inflight I/Os.
2152          */
2153         flush_workqueue(nbd->recv_workq);
2154         nbd_clear_que(nbd);
2155         nbd->task_setup = NULL;
2156         mutex_unlock(&nbd->config_lock);
2157
2158         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2159                                &nbd->config->runtime_flags))
2160                 nbd_config_put(nbd);
2161 }
2162
2163 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2164 {
2165         struct nbd_device *nbd;
2166         int index;
2167
2168         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2169                 return -EPERM;
2170
2171         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_INDEX)) {
2172                 pr_err("must specify an index to disconnect\n");
2173                 return -EINVAL;
2174         }
2175         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2176         mutex_lock(&nbd_index_mutex);
2177         nbd = idr_find(&nbd_index_idr, index);
2178         if (!nbd) {
2179                 mutex_unlock(&nbd_index_mutex);
2180                 pr_err("couldn't find device at index %d\n", index);
2181                 return -EINVAL;
2182         }
2183         if (!refcount_inc_not_zero(&nbd->refs)) {
2184                 mutex_unlock(&nbd_index_mutex);
2185                 pr_err("device at index %d is going down\n", index);
2186                 return -EINVAL;
2187         }
2188         mutex_unlock(&nbd_index_mutex);
2189         if (!refcount_inc_not_zero(&nbd->config_refs))
2190                 goto put_nbd;
2191         nbd_disconnect_and_put(nbd);
2192         nbd_config_put(nbd);
2193 put_nbd:
2194         nbd_put(nbd);
2195         return 0;
2196 }
2197
2198 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2199 {
2200         struct nbd_device *nbd = NULL;
2201         struct nbd_config *config;
2202         int index;
2203         int ret = 0;
2204         bool put_dev = false;
2205
2206         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2207                 return -EPERM;
2208
2209         if (GENL_REQ_ATTR_CHECK(info, NBD_ATTR_INDEX)) {
2210                 pr_err("must specify a device to reconfigure\n");
2211                 return -EINVAL;
2212         }
2213         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2214         mutex_lock(&nbd_index_mutex);
2215         nbd = idr_find(&nbd_index_idr, index);
2216         if (!nbd) {
2217                 mutex_unlock(&nbd_index_mutex);
2218                 pr_err("couldn't find a device at index %d\n", index);
2219                 return -EINVAL;
2220         }
2221         if (nbd->backend) {
2222                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2223                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2224                                        nbd->backend)) {
2225                                 mutex_unlock(&nbd_index_mutex);
2226                                 dev_err(nbd_to_dev(nbd),
2227                                         "backend image doesn't match with %s\n",
2228                                         nbd->backend);
2229                                 return -EINVAL;
2230                         }
2231                 } else {
2232                         mutex_unlock(&nbd_index_mutex);
2233                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2234                         return -EINVAL;
2235                 }
2236         }
2237         if (!refcount_inc_not_zero(&nbd->refs)) {
2238                 mutex_unlock(&nbd_index_mutex);
2239                 pr_err("device at index %d is going down\n", index);
2240                 return -EINVAL;
2241         }
2242         mutex_unlock(&nbd_index_mutex);
2243
2244         config = nbd_get_config_unlocked(nbd);
2245         if (!config) {
2246                 dev_err(nbd_to_dev(nbd),
2247                         "not configured, cannot reconfigure\n");
2248                 nbd_put(nbd);
2249                 return -EINVAL;
2250         }
2251
2252         mutex_lock(&nbd->config_lock);
2253         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2254             !nbd->pid) {
2255                 dev_err(nbd_to_dev(nbd),
2256                         "not configured, cannot reconfigure\n");
2257                 ret = -EINVAL;
2258                 goto out;
2259         }
2260
2261         ret = nbd_genl_size_set(info, nbd);
2262         if (ret)
2263                 goto out;
2264
2265         if (info->attrs[NBD_ATTR_TIMEOUT])
2266                 nbd_set_cmd_timeout(nbd,
2267                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2268         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2269                 config->dead_conn_timeout =
2270                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2271                 config->dead_conn_timeout *= HZ;
2272         }
2273         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2274                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2275                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2276                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2277                                               &nbd->flags))
2278                                 put_dev = true;
2279                 } else {
2280                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2281                                                &nbd->flags))
2282                                 refcount_inc(&nbd->refs);
2283                 }
2284
2285                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2286                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2287                                         &config->runtime_flags);
2288                 } else {
2289                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2290                                         &config->runtime_flags);
2291                 }
2292         }
2293
2294         if (info->attrs[NBD_ATTR_SOCKETS]) {
2295                 struct nlattr *attr;
2296                 int rem, fd;
2297
2298                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2299                                     rem) {
2300                         struct nlattr *socks[NBD_SOCK_MAX+1];
2301
2302                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2303                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2304                                 ret = -EINVAL;
2305                                 goto out;
2306                         }
2307                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2308                                                           attr,
2309                                                           nbd_sock_policy,
2310                                                           info->extack);
2311                         if (ret != 0) {
2312                                 pr_err("error processing sock list\n");
2313                                 ret = -EINVAL;
2314                                 goto out;
2315                         }
2316                         if (!socks[NBD_SOCK_FD])
2317                                 continue;
2318                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2319                         ret = nbd_reconnect_socket(nbd, fd);
2320                         if (ret) {
2321                                 if (ret == -ENOSPC)
2322                                         ret = 0;
2323                                 goto out;
2324                         }
2325                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2326                 }
2327         }
2328 out:
2329         mutex_unlock(&nbd->config_lock);
2330         nbd_config_put(nbd);
2331         nbd_put(nbd);
2332         if (put_dev)
2333                 nbd_put(nbd);
2334         return ret;
2335 }
2336
2337 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2338         {
2339                 .cmd    = NBD_CMD_CONNECT,
2340                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2341                 .doit   = nbd_genl_connect,
2342         },
2343         {
2344                 .cmd    = NBD_CMD_DISCONNECT,
2345                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2346                 .doit   = nbd_genl_disconnect,
2347         },
2348         {
2349                 .cmd    = NBD_CMD_RECONFIGURE,
2350                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2351                 .doit   = nbd_genl_reconfigure,
2352         },
2353         {
2354                 .cmd    = NBD_CMD_STATUS,
2355                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2356                 .doit   = nbd_genl_status,
2357         },
2358 };
2359
2360 static const struct genl_multicast_group nbd_mcast_grps[] = {
2361         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2362 };
2363
2364 static struct genl_family nbd_genl_family __ro_after_init = {
2365         .hdrsize        = 0,
2366         .name           = NBD_GENL_FAMILY_NAME,
2367         .version        = NBD_GENL_VERSION,
2368         .module         = THIS_MODULE,
2369         .small_ops      = nbd_connect_genl_ops,
2370         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2371         .resv_start_op  = NBD_CMD_STATUS + 1,
2372         .maxattr        = NBD_ATTR_MAX,
2373         .netnsok        = 1,
2374         .policy = nbd_attr_policy,
2375         .mcgrps         = nbd_mcast_grps,
2376         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2377 };
2378 MODULE_ALIAS_GENL_FAMILY(NBD_GENL_FAMILY_NAME);
2379
2380 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2381 {
2382         struct nlattr *dev_opt;
2383         u8 connected = 0;
2384         int ret;
2385
2386         /* This is a little racey, but for status it's ok.  The
2387          * reason we don't take a ref here is because we can't
2388          * take a ref in the index == -1 case as we would need
2389          * to put under the nbd_index_mutex, which could
2390          * deadlock if we are configured to remove ourselves
2391          * once we're disconnected.
2392          */
2393         if (refcount_read(&nbd->config_refs))
2394                 connected = 1;
2395         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2396         if (!dev_opt)
2397                 return -EMSGSIZE;
2398         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2399         if (ret)
2400                 return -EMSGSIZE;
2401         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2402                          connected);
2403         if (ret)
2404                 return -EMSGSIZE;
2405         nla_nest_end(reply, dev_opt);
2406         return 0;
2407 }
2408
2409 static int status_cb(int id, void *ptr, void *data)
2410 {
2411         struct nbd_device *nbd = ptr;
2412         return populate_nbd_status(nbd, (struct sk_buff *)data);
2413 }
2414
2415 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2416 {
2417         struct nlattr *dev_list;
2418         struct sk_buff *reply;
2419         void *reply_head;
2420         size_t msg_size;
2421         int index = -1;
2422         int ret = -ENOMEM;
2423
2424         if (info->attrs[NBD_ATTR_INDEX])
2425                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2426
2427         mutex_lock(&nbd_index_mutex);
2428
2429         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2430                                   nla_attr_size(sizeof(u8)));
2431         msg_size *= (index == -1) ? nbd_total_devices : 1;
2432
2433         reply = genlmsg_new(msg_size, GFP_KERNEL);
2434         if (!reply)
2435                 goto out;
2436         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2437                                        NBD_CMD_STATUS);
2438         if (!reply_head) {
2439                 nlmsg_free(reply);
2440                 goto out;
2441         }
2442
2443         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2444         if (index == -1) {
2445                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2446                 if (ret) {
2447                         nlmsg_free(reply);
2448                         goto out;
2449                 }
2450         } else {
2451                 struct nbd_device *nbd;
2452                 nbd = idr_find(&nbd_index_idr, index);
2453                 if (nbd) {
2454                         ret = populate_nbd_status(nbd, reply);
2455                         if (ret) {
2456                                 nlmsg_free(reply);
2457                                 goto out;
2458                         }
2459                 }
2460         }
2461         nla_nest_end(reply, dev_list);
2462         genlmsg_end(reply, reply_head);
2463         ret = genlmsg_reply(reply, info);
2464 out:
2465         mutex_unlock(&nbd_index_mutex);
2466         return ret;
2467 }
2468
2469 static void nbd_connect_reply(struct genl_info *info, int index)
2470 {
2471         struct sk_buff *skb;
2472         void *msg_head;
2473         int ret;
2474
2475         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2476         if (!skb)
2477                 return;
2478         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2479                                      NBD_CMD_CONNECT);
2480         if (!msg_head) {
2481                 nlmsg_free(skb);
2482                 return;
2483         }
2484         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2485         if (ret) {
2486                 nlmsg_free(skb);
2487                 return;
2488         }
2489         genlmsg_end(skb, msg_head);
2490         genlmsg_reply(skb, info);
2491 }
2492
2493 static void nbd_mcast_index(int index)
2494 {
2495         struct sk_buff *skb;
2496         void *msg_head;
2497         int ret;
2498
2499         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2500         if (!skb)
2501                 return;
2502         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2503                                      NBD_CMD_LINK_DEAD);
2504         if (!msg_head) {
2505                 nlmsg_free(skb);
2506                 return;
2507         }
2508         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2509         if (ret) {
2510                 nlmsg_free(skb);
2511                 return;
2512         }
2513         genlmsg_end(skb, msg_head);
2514         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2515 }
2516
2517 static void nbd_dead_link_work(struct work_struct *work)
2518 {
2519         struct link_dead_args *args = container_of(work, struct link_dead_args,
2520                                                    work);
2521         nbd_mcast_index(args->index);
2522         kfree(args);
2523 }
2524
2525 static int __init nbd_init(void)
2526 {
2527         int i;
2528
2529         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2530
2531         if (max_part < 0) {
2532                 pr_err("max_part must be >= 0\n");
2533                 return -EINVAL;
2534         }
2535
2536         part_shift = 0;
2537         if (max_part > 0) {
2538                 part_shift = fls(max_part);
2539
2540                 /*
2541                  * Adjust max_part according to part_shift as it is exported
2542                  * to user space so that user can know the max number of
2543                  * partition kernel should be able to manage.
2544                  *
2545                  * Note that -1 is required because partition 0 is reserved
2546                  * for the whole disk.
2547                  */
2548                 max_part = (1UL << part_shift) - 1;
2549         }
2550
2551         if ((1UL << part_shift) > DISK_MAX_PARTS)
2552                 return -EINVAL;
2553
2554         if (nbds_max > 1UL << (MINORBITS - part_shift))
2555                 return -EINVAL;
2556
2557         if (register_blkdev(NBD_MAJOR, "nbd"))
2558                 return -EIO;
2559
2560         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2561         if (!nbd_del_wq) {
2562                 unregister_blkdev(NBD_MAJOR, "nbd");
2563                 return -ENOMEM;
2564         }
2565
2566         if (genl_register_family(&nbd_genl_family)) {
2567                 destroy_workqueue(nbd_del_wq);
2568                 unregister_blkdev(NBD_MAJOR, "nbd");
2569                 return -EINVAL;
2570         }
2571         nbd_dbg_init();
2572
2573         for (i = 0; i < nbds_max; i++)
2574                 nbd_dev_add(i, 1);
2575         return 0;
2576 }
2577
2578 static int nbd_exit_cb(int id, void *ptr, void *data)
2579 {
2580         struct list_head *list = (struct list_head *)data;
2581         struct nbd_device *nbd = ptr;
2582
2583         /* Skip nbd that is being removed asynchronously */
2584         if (refcount_read(&nbd->refs))
2585                 list_add_tail(&nbd->list, list);
2586
2587         return 0;
2588 }
2589
2590 static void __exit nbd_cleanup(void)
2591 {
2592         struct nbd_device *nbd;
2593         LIST_HEAD(del_list);
2594
2595         /*
2596          * Unregister netlink interface prior to waiting
2597          * for the completion of netlink commands.
2598          */
2599         genl_unregister_family(&nbd_genl_family);
2600
2601         nbd_dbg_close();
2602
2603         mutex_lock(&nbd_index_mutex);
2604         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2605         mutex_unlock(&nbd_index_mutex);
2606
2607         while (!list_empty(&del_list)) {
2608                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2609                 list_del_init(&nbd->list);
2610                 if (refcount_read(&nbd->config_refs))
2611                         pr_err("possibly leaking nbd_config (ref %d)\n",
2612                                         refcount_read(&nbd->config_refs));
2613                 if (refcount_read(&nbd->refs) != 1)
2614                         pr_err("possibly leaking a device\n");
2615                 nbd_put(nbd);
2616         }
2617
2618         /* Also wait for nbd_dev_remove_work() completes */
2619         destroy_workqueue(nbd_del_wq);
2620
2621         idr_destroy(&nbd_index_idr);
2622         unregister_blkdev(NBD_MAJOR, "nbd");
2623 }
2624
2625 module_init(nbd_init);
2626 module_exit(nbd_cleanup);
2627
2628 MODULE_DESCRIPTION("Network Block Device");
2629 MODULE_LICENSE("GPL");
2630
2631 module_param(nbds_max, int, 0444);
2632 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2633 module_param(max_part, int, 0444);
2634 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");