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