drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   9  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #include <linux/major.h>
  15
  16 #include <linux/blkdev.h>
  17 #include <linux/module.h>
  18 #include <linux/init.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/mm.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/completion.h>
  30 #include <linux/err.h>
  31 #include <linux/kernel.h>
  32 #include <linux/slab.h>
  33 #include <net/sock.h>
  34 #include <linux/net.h>
  35 #include <linux/kthread.h>
  36 #include <linux/types.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/blk-mq.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/types.h>
  42
  43 #include <linux/nbd.h>
  44 #include <linux/nbd-netlink.h>
  45 #include <net/genetlink.h>
  46
  47 #define CREATE_TRACE_POINTS
  48 #include <trace/events/nbd.h>
  49
  50 static DEFINE_IDR(nbd_index_idr);
  51 static DEFINE_MUTEX(nbd_index_mutex);
  52 static struct workqueue_struct *nbd_del_wq;
  53 static int nbd_total_devices = 0;
  54
  55 struct nbd_sock {
  56         struct socket *sock;
  57         struct mutex tx_lock;
  58         struct request *pending;
  59         int sent;
  60         bool dead;
  61         int fallback_index;
  62         int cookie;
  63 };
  64
  65 struct recv_thread_args {
  66         struct work_struct work;
  67         struct nbd_device *nbd;
  68         int index;
  69 };
  70
  71 struct link_dead_args {
  72         struct work_struct work;
  73         int index;
  74 };
  75
  76 #define NBD_RT_TIMEDOUT                 0
  77 #define NBD_RT_DISCONNECT_REQUESTED     1
  78 #define NBD_RT_DISCONNECTED             2
  79 #define NBD_RT_HAS_PID_FILE             3
  80 #define NBD_RT_HAS_CONFIG_REF           4
  81 #define NBD_RT_BOUND                    5
  82 #define NBD_RT_DISCONNECT_ON_CLOSE      6
  83 #define NBD_RT_HAS_BACKEND_FILE         7
  84
  85 #define NBD_DESTROY_ON_DISCONNECT       0
  86 #define NBD_DISCONNECT_REQUESTED        1
  87
  88 struct nbd_config {
  89         u32 flags;
  90         unsigned long runtime_flags;
  91         u64 dead_conn_timeout;
  92
  93         struct nbd_sock **socks;
  94         int num_connections;
  95         atomic_t live_connections;
  96         wait_queue_head_t conn_wait;
  97
  98         atomic_t recv_threads;
  99         wait_queue_head_t recv_wq;
 100         unsigned int blksize_bits;
 101         loff_t bytesize;
 102 #if IS_ENABLED(CONFIG_DEBUG_FS)
 103         struct dentry *dbg_dir;
 104 #endif
 105 };
 106
 107 static inline unsigned int nbd_blksize(struct nbd_config *config)
 108 {
 109         return 1u << config->blksize_bits;
 110 }
 111
 112 struct nbd_device {
 113         struct blk_mq_tag_set tag_set;
 114
 115         int index;
 116         refcount_t config_refs;
 117         refcount_t refs;
 118         struct nbd_config *config;
 119         struct mutex config_lock;
 120         struct gendisk *disk;
 121         struct workqueue_struct *recv_workq;
 122         struct work_struct remove_work;
 123
 124         struct list_head list;
 125         struct task_struct *task_setup;
 126
 127         unsigned long flags;
 128         pid_t pid; /* pid of nbd-client, if attached */
 129
 130         char *backend;
 131 };
 132
 133 #define NBD_CMD_REQUEUED        1
 134 /*
 135  * This flag will be set if nbd_queue_rq() succeed, and will be checked and
 136  * cleared in completion. Both setting and clearing of the flag are protected
 137  * by cmd->lock.
 138  */
 139 #define NBD_CMD_INFLIGHT        2
 140
 141 struct nbd_cmd {
 142         struct nbd_device *nbd;
 143         struct mutex lock;
 144         int index;
 145         int cookie;
 146         int retries;
 147         blk_status_t status;
 148         unsigned long flags;
 149         u32 cmd_cookie;
 150 };
 151
 152 #if IS_ENABLED(CONFIG_DEBUG_FS)
 153 static struct dentry *nbd_dbg_dir;
 154 #endif
 155
 156 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 157
 158 #define NBD_MAGIC 0x68797548
 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         blk_cleanup_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         nbd->config->bytesize = bytesize;
 329         nbd->config->blksize_bits = __ffs(blksize);
 330
 331         if (!nbd->pid)
 332                 return 0;
 333
 334         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 335                 nbd->disk->queue->limits.discard_granularity = blksize;
 336                 nbd->disk->queue->limits.discard_alignment = blksize;
 337                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 338         }
 339         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 340         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 341
 342         if (max_part)
 343                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 344         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 345                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 346         return 0;
 347 }
 348
 349 static void nbd_complete_rq(struct request *req)
 350 {
 351         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 352
 353         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 354                 cmd->status ? "failed" : "done");
 355
 356         blk_mq_end_request(req, cmd->status);
 357 }
 358
 359 /*
 360  * Forcibly shutdown the socket causing all listeners to error
 361  */
 362 static void sock_shutdown(struct nbd_device *nbd)
 363 {
 364         struct nbd_config *config = nbd->config;
 365         int i;
 366
 367         if (config->num_connections == 0)
 368                 return;
 369         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 370                 return;
 371
 372         for (i = 0; i < config->num_connections; i++) {
 373                 struct nbd_sock *nsock = config->socks[i];
 374                 mutex_lock(&nsock->tx_lock);
 375                 nbd_mark_nsock_dead(nbd, nsock, 0);
 376                 mutex_unlock(&nsock->tx_lock);
 377         }
 378         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 379 }
 380
 381 static u32 req_to_nbd_cmd_type(struct request *req)
 382 {
 383         switch (req_op(req)) {
 384         case REQ_OP_DISCARD:
 385                 return NBD_CMD_TRIM;
 386         case REQ_OP_FLUSH:
 387                 return NBD_CMD_FLUSH;
 388         case REQ_OP_WRITE:
 389                 return NBD_CMD_WRITE;
 390         case REQ_OP_READ:
 391                 return NBD_CMD_READ;
 392         default:
 393                 return U32_MAX;
 394         }
 395 }
 396
 397 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 398                                                  bool reserved)
 399 {
 400         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 401         struct nbd_device *nbd = cmd->nbd;
 402         struct nbd_config *config;
 403
 404         if (!mutex_trylock(&cmd->lock))
 405                 return BLK_EH_RESET_TIMER;
 406
 407         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 408                 mutex_unlock(&cmd->lock);
 409                 return BLK_EH_DONE;
 410         }
 411
 412         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 413                 cmd->status = BLK_STS_TIMEOUT;
 414                 mutex_unlock(&cmd->lock);
 415                 goto done;
 416         }
 417         config = nbd->config;
 418
 419         if (config->num_connections > 1 ||
 420             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 421                 dev_err_ratelimited(nbd_to_dev(nbd),
 422                                     "Connection timed out, retrying (%d/%d alive)\n",
 423                                     atomic_read(&config->live_connections),
 424                                     config->num_connections);
 425                 /*
 426                  * Hooray we have more connections, requeue this IO, the submit
 427                  * path will put it on a real connection. Or if only one
 428                  * connection is configured, the submit path will wait util
 429                  * a new connection is reconfigured or util dead timeout.
 430                  */
 431                 if (config->socks) {
 432                         if (cmd->index < config->num_connections) {
 433                                 struct nbd_sock *nsock =
 434                                         config->socks[cmd->index];
 435                                 mutex_lock(&nsock->tx_lock);
 436                                 /* We can have multiple outstanding requests, so
 437                                  * we don't want to mark the nsock dead if we've
 438                                  * already reconnected with a new socket, so
 439                                  * only mark it dead if its the same socket we
 440                                  * were sent out on.
 441                                  */
 442                                 if (cmd->cookie == nsock->cookie)
 443                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 444                                 mutex_unlock(&nsock->tx_lock);
 445                         }
 446                         mutex_unlock(&cmd->lock);
 447                         nbd_requeue_cmd(cmd);
 448                         nbd_config_put(nbd);
 449                         return BLK_EH_DONE;
 450                 }
 451         }
 452
 453         if (!nbd->tag_set.timeout) {
 454                 /*
 455                  * Userspace sets timeout=0 to disable socket disconnection,
 456                  * so just warn and reset the timer.
 457                  */
 458                 struct nbd_sock *nsock = config->socks[cmd->index];
 459                 cmd->retries++;
 460                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 461                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 462                         (unsigned long long)blk_rq_pos(req) << 9,
 463                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 464
 465                 mutex_lock(&nsock->tx_lock);
 466                 if (cmd->cookie != nsock->cookie) {
 467                         nbd_requeue_cmd(cmd);
 468                         mutex_unlock(&nsock->tx_lock);
 469                         mutex_unlock(&cmd->lock);
 470                         nbd_config_put(nbd);
 471                         return BLK_EH_DONE;
 472                 }
 473                 mutex_unlock(&nsock->tx_lock);
 474                 mutex_unlock(&cmd->lock);
 475                 nbd_config_put(nbd);
 476                 return BLK_EH_RESET_TIMER;
 477         }
 478
 479         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 480         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 481         cmd->status = BLK_STS_IOERR;
 482         mutex_unlock(&cmd->lock);
 483         sock_shutdown(nbd);
 484         nbd_config_put(nbd);
 485 done:
 486         blk_mq_complete_request(req);
 487         return BLK_EH_DONE;
 488 }
 489
 490 /*
 491  *  Send or receive packet. Return a positive value on success and
 492  *  negtive value on failue, and never return 0.
 493  */
 494 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 495                      struct iov_iter *iter, int msg_flags, int *sent)
 496 {
 497         struct nbd_config *config = nbd->config;
 498         struct socket *sock = config->socks[index]->sock;
 499         int result;
 500         struct msghdr msg;
 501         unsigned int noreclaim_flag;
 502
 503         if (unlikely(!sock)) {
 504                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 505                         "Attempted %s on closed socket in sock_xmit\n",
 506                         (send ? "send" : "recv"));
 507                 return -EINVAL;
 508         }
 509
 510         msg.msg_iter = *iter;
 511
 512         noreclaim_flag = memalloc_noreclaim_save();
 513         do {
 514                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 515                 msg.msg_name = NULL;
 516                 msg.msg_namelen = 0;
 517                 msg.msg_control = NULL;
 518                 msg.msg_controllen = 0;
 519                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 520
 521                 if (send)
 522                         result = sock_sendmsg(sock, &msg);
 523                 else
 524                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 525
 526                 if (result <= 0) {
 527                         if (result == 0)
 528                                 result = -EPIPE; /* short read */
 529                         break;
 530                 }
 531                 if (sent)
 532                         *sent += result;
 533         } while (msg_data_left(&msg));
 534
 535         memalloc_noreclaim_restore(noreclaim_flag);
 536
 537         return result;
 538 }
 539
 540 /*
 541  * Different settings for sk->sk_sndtimeo can result in different return values
 542  * if there is a signal pending when we enter sendmsg, because reasons?
 543  */
 544 static inline int was_interrupted(int result)
 545 {
 546         return result == -ERESTARTSYS || result == -EINTR;
 547 }
 548
 549 /* always call with the tx_lock held */
 550 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 551 {
 552         struct request *req = blk_mq_rq_from_pdu(cmd);
 553         struct nbd_config *config = nbd->config;
 554         struct nbd_sock *nsock = config->socks[index];
 555         int result;
 556         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 557         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 558         struct iov_iter from;
 559         unsigned long size = blk_rq_bytes(req);
 560         struct bio *bio;
 561         u64 handle;
 562         u32 type;
 563         u32 nbd_cmd_flags = 0;
 564         int sent = nsock->sent, skip = 0;
 565
 566         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 567
 568         type = req_to_nbd_cmd_type(req);
 569         if (type == U32_MAX)
 570                 return -EIO;
 571
 572         if (rq_data_dir(req) == WRITE &&
 573             (config->flags & NBD_FLAG_READ_ONLY)) {
 574                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 575                                     "Write on read-only\n");
 576                 return -EIO;
 577         }
 578
 579         if (req->cmd_flags & REQ_FUA)
 580                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 581
 582         /* We did a partial send previously, and we at least sent the whole
 583          * request struct, so just go and send the rest of the pages in the
 584          * request.
 585          */
 586         if (sent) {
 587                 if (sent >= sizeof(request)) {
 588                         skip = sent - sizeof(request);
 589
 590                         /* initialize handle for tracing purposes */
 591                         handle = nbd_cmd_handle(cmd);
 592
 593                         goto send_pages;
 594                 }
 595                 iov_iter_advance(&from, sent);
 596         } else {
 597                 cmd->cmd_cookie++;
 598         }
 599         cmd->index = index;
 600         cmd->cookie = nsock->cookie;
 601         cmd->retries = 0;
 602         request.type = htonl(type | nbd_cmd_flags);
 603         if (type != NBD_CMD_FLUSH) {
 604                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 605                 request.len = htonl(size);
 606         }
 607         handle = nbd_cmd_handle(cmd);
 608         memcpy(request.handle, &handle, sizeof(handle));
 609
 610         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 611
 612         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 613                 req, nbdcmd_to_ascii(type),
 614                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 615         result = sock_xmit(nbd, index, 1, &from,
 616                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 617         trace_nbd_header_sent(req, handle);
 618         if (result < 0) {
 619                 if (was_interrupted(result)) {
 620                         /* If we havne't sent anything we can just return BUSY,
 621                          * however if we have sent something we need to make
 622                          * sure we only allow this req to be sent until we are
 623                          * completely done.
 624                          */
 625                         if (sent) {
 626                                 nsock->pending = req;
 627                                 nsock->sent = sent;
 628                         }
 629                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 630                         return BLK_STS_RESOURCE;
 631                 }
 632                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 633                         "Send control failed (result %d)\n", result);
 634                 return -EAGAIN;
 635         }
 636 send_pages:
 637         if (type != NBD_CMD_WRITE)
 638                 goto out;
 639
 640         bio = req->bio;
 641         while (bio) {
 642                 struct bio *next = bio->bi_next;
 643                 struct bvec_iter iter;
 644                 struct bio_vec bvec;
 645
 646                 bio_for_each_segment(bvec, bio, iter) {
 647                         bool is_last = !next && bio_iter_last(bvec, iter);
 648                         int flags = is_last ? 0 : MSG_MORE;
 649
 650                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 651                                 req, bvec.bv_len);
 652                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 653                         if (skip) {
 654                                 if (skip >= iov_iter_count(&from)) {
 655                                         skip -= iov_iter_count(&from);
 656                                         continue;
 657                                 }
 658                                 iov_iter_advance(&from, skip);
 659                                 skip = 0;
 660                         }
 661                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 662                         if (result < 0) {
 663                                 if (was_interrupted(result)) {
 664                                         /* We've already sent the header, we
 665                                          * have no choice but to set pending and
 666                                          * return BUSY.
 667                                          */
 668                                         nsock->pending = req;
 669                                         nsock->sent = sent;
 670                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 671                                         return BLK_STS_RESOURCE;
 672                                 }
 673                                 dev_err(disk_to_dev(nbd->disk),
 674                                         "Send data failed (result %d)\n",
 675                                         result);
 676                                 return -EAGAIN;
 677                         }
 678                         /*
 679                          * The completion might already have come in,
 680                          * so break for the last one instead of letting
 681                          * the iterator do it. This prevents use-after-free
 682                          * of the bio.
 683                          */
 684                         if (is_last)
 685                                 break;
 686                 }
 687                 bio = next;
 688         }
 689 out:
 690         trace_nbd_payload_sent(req, handle);
 691         nsock->pending = NULL;
 692         nsock->sent = 0;
 693         return 0;
 694 }
 695
 696 static int nbd_read_reply(struct nbd_device *nbd, int index,
 697                           struct nbd_reply *reply)
 698 {
 699         struct kvec iov = {.iov_base = reply, .iov_len = sizeof(*reply)};
 700         struct iov_iter to;
 701         int result;
 702
 703         reply->magic = 0;
 704         iov_iter_kvec(&to, READ, &iov, 1, sizeof(*reply));
 705         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 706         if (result < 0) {
 707                 if (!nbd_disconnected(nbd->config))
 708                         dev_err(disk_to_dev(nbd->disk),
 709                                 "Receive control failed (result %d)\n", result);
 710                 return result;
 711         }
 712
 713         if (ntohl(reply->magic) != NBD_REPLY_MAGIC) {
 714                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 715                                 (unsigned long)ntohl(reply->magic));
 716                 return -EPROTO;
 717         }
 718
 719         return 0;
 720 }
 721
 722 /* NULL returned = something went wrong, inform userspace */
 723 static struct nbd_cmd *nbd_handle_reply(struct nbd_device *nbd, int index,
 724                                         struct nbd_reply *reply)
 725 {
 726         int result;
 727         struct nbd_cmd *cmd;
 728         struct request *req = NULL;
 729         u64 handle;
 730         u16 hwq;
 731         u32 tag;
 732         int ret = 0;
 733
 734         memcpy(&handle, reply->handle, sizeof(handle));
 735         tag = nbd_handle_to_tag(handle);
 736         hwq = blk_mq_unique_tag_to_hwq(tag);
 737         if (hwq < nbd->tag_set.nr_hw_queues)
 738                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 739                                        blk_mq_unique_tag_to_tag(tag));
 740         if (!req || !blk_mq_request_started(req)) {
 741                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 742                         tag, req);
 743                 return ERR_PTR(-ENOENT);
 744         }
 745         trace_nbd_header_received(req, handle);
 746         cmd = blk_mq_rq_to_pdu(req);
 747
 748         mutex_lock(&cmd->lock);
 749         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 750                 dev_err(disk_to_dev(nbd->disk), "Suspicious reply %d (status %u flags %lu)",
 751                         tag, cmd->status, cmd->flags);
 752                 ret = -ENOENT;
 753                 goto out;
 754         }
 755         if (cmd->index != index) {
 756                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply %d from different sock %d (expected %d)",
 757                         tag, index, cmd->index);
 758                 ret = -ENOENT;
 759                 goto out;
 760         }
 761         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 762                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 763                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 764                 ret = -ENOENT;
 765                 goto out;
 766         }
 767         if (cmd->status != BLK_STS_OK) {
 768                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 769                         req);
 770                 ret = -ENOENT;
 771                 goto out;
 772         }
 773         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 774                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 775                         req);
 776                 ret = -ENOENT;
 777                 goto out;
 778         }
 779         if (ntohl(reply->error)) {
 780                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 781                         ntohl(reply->error));
 782                 cmd->status = BLK_STS_IOERR;
 783                 goto out;
 784         }
 785
 786         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 787         if (rq_data_dir(req) != WRITE) {
 788                 struct req_iterator iter;
 789                 struct bio_vec bvec;
 790                 struct iov_iter to;
 791
 792                 rq_for_each_segment(bvec, req, iter) {
 793                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 794                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 795                         if (result < 0) {
 796                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 797                                         result);
 798                                 /*
 799                                  * If we've disconnected, we need to make sure we
 800                                  * complete this request, otherwise error out
 801                                  * and let the timeout stuff handle resubmitting
 802                                  * this request onto another connection.
 803                                  */
 804                                 if (nbd_disconnected(nbd->config)) {
 805                                         cmd->status = BLK_STS_IOERR;
 806                                         goto out;
 807                                 }
 808                                 ret = -EIO;
 809                                 goto out;
 810                         }
 811                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 812                                 req, bvec.bv_len);
 813                 }
 814         }
 815 out:
 816         trace_nbd_payload_received(req, handle);
 817         mutex_unlock(&cmd->lock);
 818         return ret ? ERR_PTR(ret) : cmd;
 819 }
 820
 821 static void recv_work(struct work_struct *work)
 822 {
 823         struct recv_thread_args *args = container_of(work,
 824                                                      struct recv_thread_args,
 825                                                      work);
 826         struct nbd_device *nbd = args->nbd;
 827         struct nbd_config *config = nbd->config;
 828         struct request_queue *q = nbd->disk->queue;
 829         struct nbd_sock *nsock;
 830         struct nbd_cmd *cmd;
 831         struct request *rq;
 832
 833         while (1) {
 834                 struct nbd_reply reply;
 835
 836                 if (nbd_read_reply(nbd, args->index, &reply))
 837                         break;
 838
 839                 /*
 840                  * Grab .q_usage_counter so request pool won't go away, then no
 841                  * request use-after-free is possible during nbd_handle_reply().
 842                  * If queue is frozen, there won't be any inflight requests, we
 843                  * needn't to handle the incoming garbage message.
 844                  */
 845                 if (!percpu_ref_tryget(&q->q_usage_counter)) {
 846                         dev_err(disk_to_dev(nbd->disk), "%s: no io inflight\n",
 847                                 __func__);
 848                         break;
 849                 }
 850
 851                 cmd = nbd_handle_reply(nbd, args->index, &reply);
 852                 if (IS_ERR(cmd)) {
 853                         percpu_ref_put(&q->q_usage_counter);
 854                         break;
 855                 }
 856
 857                 rq = blk_mq_rq_from_pdu(cmd);
 858                 if (likely(!blk_should_fake_timeout(rq->q)))
 859                         blk_mq_complete_request(rq);
 860                 percpu_ref_put(&q->q_usage_counter);
 861         }
 862
 863         nsock = config->socks[args->index];
 864         mutex_lock(&nsock->tx_lock);
 865         nbd_mark_nsock_dead(nbd, nsock, 1);
 866         mutex_unlock(&nsock->tx_lock);
 867
 868         nbd_config_put(nbd);
 869         atomic_dec(&config->recv_threads);
 870         wake_up(&config->recv_wq);
 871         kfree(args);
 872 }
 873
 874 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 875 {
 876         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 877
 878         /* don't abort one completed request */
 879         if (blk_mq_request_completed(req))
 880                 return true;
 881
 882         mutex_lock(&cmd->lock);
 883         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 884                 mutex_unlock(&cmd->lock);
 885                 return true;
 886         }
 887         cmd->status = BLK_STS_IOERR;
 888         mutex_unlock(&cmd->lock);
 889
 890         blk_mq_complete_request(req);
 891         return true;
 892 }
 893
 894 static void nbd_clear_que(struct nbd_device *nbd)
 895 {
 896         blk_mq_quiesce_queue(nbd->disk->queue);
 897         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 898         blk_mq_unquiesce_queue(nbd->disk->queue);
 899         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 900 }
 901
 902 static int find_fallback(struct nbd_device *nbd, int index)
 903 {
 904         struct nbd_config *config = nbd->config;
 905         int new_index = -1;
 906         struct nbd_sock *nsock = config->socks[index];
 907         int fallback = nsock->fallback_index;
 908
 909         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 910                 return new_index;
 911
 912         if (config->num_connections <= 1) {
 913                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 914                                     "Dead connection, failed to find a fallback\n");
 915                 return new_index;
 916         }
 917
 918         if (fallback >= 0 && fallback < config->num_connections &&
 919             !config->socks[fallback]->dead)
 920                 return fallback;
 921
 922         if (nsock->fallback_index < 0 ||
 923             nsock->fallback_index >= config->num_connections ||
 924             config->socks[nsock->fallback_index]->dead) {
 925                 int i;
 926                 for (i = 0; i < config->num_connections; i++) {
 927                         if (i == index)
 928                                 continue;
 929                         if (!config->socks[i]->dead) {
 930                                 new_index = i;
 931                                 break;
 932                         }
 933                 }
 934                 nsock->fallback_index = new_index;
 935                 if (new_index < 0) {
 936                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 937                                             "Dead connection, failed to find a fallback\n");
 938                         return new_index;
 939                 }
 940         }
 941         new_index = nsock->fallback_index;
 942         return new_index;
 943 }
 944
 945 static int wait_for_reconnect(struct nbd_device *nbd)
 946 {
 947         struct nbd_config *config = nbd->config;
 948         if (!config->dead_conn_timeout)
 949                 return 0;
 950         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 951                 return 0;
 952         return wait_event_timeout(config->conn_wait,
 953                                   atomic_read(&config->live_connections) > 0,
 954                                   config->dead_conn_timeout) > 0;
 955 }
 956
 957 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 958 {
 959         struct request *req = blk_mq_rq_from_pdu(cmd);
 960         struct nbd_device *nbd = cmd->nbd;
 961         struct nbd_config *config;
 962         struct nbd_sock *nsock;
 963         int ret;
 964
 965         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 966                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 967                                     "Socks array is empty\n");
 968                 return -EINVAL;
 969         }
 970         config = nbd->config;
 971
 972         if (index >= config->num_connections) {
 973                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 974                                     "Attempted send on invalid socket\n");
 975                 nbd_config_put(nbd);
 976                 return -EINVAL;
 977         }
 978         cmd->status = BLK_STS_OK;
 979 again:
 980         nsock = config->socks[index];
 981         mutex_lock(&nsock->tx_lock);
 982         if (nsock->dead) {
 983                 int old_index = index;
 984                 index = find_fallback(nbd, index);
 985                 mutex_unlock(&nsock->tx_lock);
 986                 if (index < 0) {
 987                         if (wait_for_reconnect(nbd)) {
 988                                 index = old_index;
 989                                 goto again;
 990                         }
 991                         /* All the sockets should already be down at this point,
 992                          * we just want to make sure that DISCONNECTED is set so
 993                          * any requests that come in that were queue'ed waiting
 994                          * for the reconnect timer don't trigger the timer again
 995                          * and instead just error out.
 996                          */
 997                         sock_shutdown(nbd);
 998                         nbd_config_put(nbd);
 999                         return -EIO;
1000                 }
1001                 goto again;
1002         }
1003
1004         /* Handle the case that we have a pending request that was partially
1005          * transmitted that _has_ to be serviced first.  We need to call requeue
1006          * here so that it gets put _after_ the request that is already on the
1007          * dispatch list.
1008          */
1009         blk_mq_start_request(req);
1010         if (unlikely(nsock->pending && nsock->pending != req)) {
1011                 nbd_requeue_cmd(cmd);
1012                 ret = 0;
1013                 goto out;
1014         }
1015         /*
1016          * Some failures are related to the link going down, so anything that
1017          * returns EAGAIN can be retried on a different socket.
1018          */
1019         ret = nbd_send_cmd(nbd, cmd, index);
1020         /*
1021          * Access to this flag is protected by cmd->lock, thus it's safe to set
1022          * the flag after nbd_send_cmd() succeed to send request to server.
1023          */
1024         if (!ret)
1025                 __set_bit(NBD_CMD_INFLIGHT, &cmd->flags);
1026         else if (ret == -EAGAIN) {
1027                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1028                                     "Request send failed, requeueing\n");
1029                 nbd_mark_nsock_dead(nbd, nsock, 1);
1030                 nbd_requeue_cmd(cmd);
1031                 ret = 0;
1032         }
1033 out:
1034         mutex_unlock(&nsock->tx_lock);
1035         nbd_config_put(nbd);
1036         return ret;
1037 }
1038
1039 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1040                         const struct blk_mq_queue_data *bd)
1041 {
1042         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1043         int ret;
1044
1045         /*
1046          * Since we look at the bio's to send the request over the network we
1047          * need to make sure the completion work doesn't mark this request done
1048          * before we are done doing our send.  This keeps us from dereferencing
1049          * freed data if we have particularly fast completions (ie we get the
1050          * completion before we exit sock_xmit on the last bvec) or in the case
1051          * that the server is misbehaving (or there was an error) before we're
1052          * done sending everything over the wire.
1053          */
1054         mutex_lock(&cmd->lock);
1055         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1056
1057         /* We can be called directly from the user space process, which means we
1058          * could possibly have signals pending so our sendmsg will fail.  In
1059          * this case we need to return that we are busy, otherwise error out as
1060          * appropriate.
1061          */
1062         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1063         if (ret < 0)
1064                 ret = BLK_STS_IOERR;
1065         else if (!ret)
1066                 ret = BLK_STS_OK;
1067         mutex_unlock(&cmd->lock);
1068
1069         return ret;
1070 }
1071
1072 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1073                                      int *err)
1074 {
1075         struct socket *sock;
1076
1077         *err = 0;
1078         sock = sockfd_lookup(fd, err);
1079         if (!sock)
1080                 return NULL;
1081
1082         if (sock->ops->shutdown == sock_no_shutdown) {
1083                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1084                 *err = -EINVAL;
1085                 sockfd_put(sock);
1086                 return NULL;
1087         }
1088
1089         return sock;
1090 }
1091
1092 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1093                           bool netlink)
1094 {
1095         struct nbd_config *config = nbd->config;
1096         struct socket *sock;
1097         struct nbd_sock **socks;
1098         struct nbd_sock *nsock;
1099         int err;
1100
1101         sock = nbd_get_socket(nbd, arg, &err);
1102         if (!sock)
1103                 return err;
1104
1105         /*
1106          * We need to make sure we don't get any errant requests while we're
1107          * reallocating the ->socks array.
1108          */
1109         blk_mq_freeze_queue(nbd->disk->queue);
1110
1111         if (!netlink && !nbd->task_setup &&
1112             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1113                 nbd->task_setup = current;
1114
1115         if (!netlink &&
1116             (nbd->task_setup != current ||
1117              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1118                 dev_err(disk_to_dev(nbd->disk),
1119                         "Device being setup by another task");
1120                 err = -EBUSY;
1121                 goto put_socket;
1122         }
1123
1124         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1125         if (!nsock) {
1126                 err = -ENOMEM;
1127                 goto put_socket;
1128         }
1129
1130         socks = krealloc(config->socks, (config->num_connections + 1) *
1131                          sizeof(struct nbd_sock *), GFP_KERNEL);
1132         if (!socks) {
1133                 kfree(nsock);
1134                 err = -ENOMEM;
1135                 goto put_socket;
1136         }
1137
1138         config->socks = socks;
1139
1140         nsock->fallback_index = -1;
1141         nsock->dead = false;
1142         mutex_init(&nsock->tx_lock);
1143         nsock->sock = sock;
1144         nsock->pending = NULL;
1145         nsock->sent = 0;
1146         nsock->cookie = 0;
1147         socks[config->num_connections++] = nsock;
1148         atomic_inc(&config->live_connections);
1149         blk_mq_unfreeze_queue(nbd->disk->queue);
1150
1151         return 0;
1152
1153 put_socket:
1154         blk_mq_unfreeze_queue(nbd->disk->queue);
1155         sockfd_put(sock);
1156         return err;
1157 }
1158
1159 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1160 {
1161         struct nbd_config *config = nbd->config;
1162         struct socket *sock, *old;
1163         struct recv_thread_args *args;
1164         int i;
1165         int err;
1166
1167         sock = nbd_get_socket(nbd, arg, &err);
1168         if (!sock)
1169                 return err;
1170
1171         args = kzalloc(sizeof(*args), GFP_KERNEL);
1172         if (!args) {
1173                 sockfd_put(sock);
1174                 return -ENOMEM;
1175         }
1176
1177         for (i = 0; i < config->num_connections; i++) {
1178                 struct nbd_sock *nsock = config->socks[i];
1179
1180                 if (!nsock->dead)
1181                         continue;
1182
1183                 mutex_lock(&nsock->tx_lock);
1184                 if (!nsock->dead) {
1185                         mutex_unlock(&nsock->tx_lock);
1186                         continue;
1187                 }
1188                 sk_set_memalloc(sock->sk);
1189                 if (nbd->tag_set.timeout)
1190                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1191                 atomic_inc(&config->recv_threads);
1192                 refcount_inc(&nbd->config_refs);
1193                 old = nsock->sock;
1194                 nsock->fallback_index = -1;
1195                 nsock->sock = sock;
1196                 nsock->dead = false;
1197                 INIT_WORK(&args->work, recv_work);
1198                 args->index = i;
1199                 args->nbd = nbd;
1200                 nsock->cookie++;
1201                 mutex_unlock(&nsock->tx_lock);
1202                 sockfd_put(old);
1203
1204                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1205
1206                 /* We take the tx_mutex in an error path in the recv_work, so we
1207                  * need to queue_work outside of the tx_mutex.
1208                  */
1209                 queue_work(nbd->recv_workq, &args->work);
1210
1211                 atomic_inc(&config->live_connections);
1212                 wake_up(&config->conn_wait);
1213                 return 0;
1214         }
1215         sockfd_put(sock);
1216         kfree(args);
1217         return -ENOSPC;
1218 }
1219
1220 static void nbd_bdev_reset(struct block_device *bdev)
1221 {
1222         if (bdev->bd_openers > 1)
1223                 return;
1224         set_capacity(bdev->bd_disk, 0);
1225 }
1226
1227 static void nbd_parse_flags(struct nbd_device *nbd)
1228 {
1229         struct nbd_config *config = nbd->config;
1230         if (config->flags & NBD_FLAG_READ_ONLY)
1231                 set_disk_ro(nbd->disk, true);
1232         else
1233                 set_disk_ro(nbd->disk, false);
1234         if (config->flags & NBD_FLAG_SEND_TRIM)
1235                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1236         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1237                 if (config->flags & NBD_FLAG_SEND_FUA)
1238                         blk_queue_write_cache(nbd->disk->queue, true, true);
1239                 else
1240                         blk_queue_write_cache(nbd->disk->queue, true, false);
1241         }
1242         else
1243                 blk_queue_write_cache(nbd->disk->queue, false, false);
1244 }
1245
1246 static void send_disconnects(struct nbd_device *nbd)
1247 {
1248         struct nbd_config *config = nbd->config;
1249         struct nbd_request request = {
1250                 .magic = htonl(NBD_REQUEST_MAGIC),
1251                 .type = htonl(NBD_CMD_DISC),
1252         };
1253         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1254         struct iov_iter from;
1255         int i, ret;
1256
1257         for (i = 0; i < config->num_connections; i++) {
1258                 struct nbd_sock *nsock = config->socks[i];
1259
1260                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1261                 mutex_lock(&nsock->tx_lock);
1262                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1263                 if (ret < 0)
1264                         dev_err(disk_to_dev(nbd->disk),
1265                                 "Send disconnect failed %d\n", ret);
1266                 mutex_unlock(&nsock->tx_lock);
1267         }
1268 }
1269
1270 static int nbd_disconnect(struct nbd_device *nbd)
1271 {
1272         struct nbd_config *config = nbd->config;
1273
1274         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1275         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1276         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1277         send_disconnects(nbd);
1278         return 0;
1279 }
1280
1281 static void nbd_clear_sock(struct nbd_device *nbd)
1282 {
1283         sock_shutdown(nbd);
1284         nbd_clear_que(nbd);
1285         nbd->task_setup = NULL;
1286 }
1287
1288 static void nbd_config_put(struct nbd_device *nbd)
1289 {
1290         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1291                                         &nbd->config_lock)) {
1292                 struct nbd_config *config = nbd->config;
1293                 nbd_dev_dbg_close(nbd);
1294                 invalidate_disk(nbd->disk);
1295                 if (nbd->config->bytesize)
1296                         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
1297                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1298                                        &config->runtime_flags))
1299                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1300                 nbd->pid = 0;
1301                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1302                                        &config->runtime_flags)) {
1303                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1304                         kfree(nbd->backend);
1305                         nbd->backend = NULL;
1306                 }
1307                 nbd_clear_sock(nbd);
1308                 if (config->num_connections) {
1309                         int i;
1310                         for (i = 0; i < config->num_connections; i++) {
1311                                 sockfd_put(config->socks[i]->sock);
1312                                 kfree(config->socks[i]);
1313                         }
1314                         kfree(config->socks);
1315                 }
1316                 kfree(nbd->config);
1317                 nbd->config = NULL;
1318
1319                 nbd->tag_set.timeout = 0;
1320                 nbd->disk->queue->limits.discard_granularity = 0;
1321                 nbd->disk->queue->limits.discard_alignment = 0;
1322                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1323                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1324
1325                 mutex_unlock(&nbd->config_lock);
1326                 nbd_put(nbd);
1327                 module_put(THIS_MODULE);
1328         }
1329 }
1330
1331 static int nbd_start_device(struct nbd_device *nbd)
1332 {
1333         struct nbd_config *config = nbd->config;
1334         int num_connections = config->num_connections;
1335         int error = 0, i;
1336
1337         if (nbd->pid)
1338                 return -EBUSY;
1339         if (!config->socks)
1340                 return -EINVAL;
1341         if (num_connections > 1 &&
1342             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1343                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1344                 return -EINVAL;
1345         }
1346
1347         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1348         nbd->pid = task_pid_nr(current);
1349
1350         nbd_parse_flags(nbd);
1351
1352         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1353         if (error) {
1354                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1355                 return error;
1356         }
1357         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1358
1359         nbd_dev_dbg_init(nbd);
1360         for (i = 0; i < num_connections; i++) {
1361                 struct recv_thread_args *args;
1362
1363                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1364                 if (!args) {
1365                         sock_shutdown(nbd);
1366                         /*
1367                          * If num_connections is m (2 < m),
1368                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1369                          * But NO.(n + 1) failed. We still have n recv threads.
1370                          * So, add flush_workqueue here to prevent recv threads
1371                          * dropping the last config_refs and trying to destroy
1372                          * the workqueue from inside the workqueue.
1373                          */
1374                         if (i)
1375                                 flush_workqueue(nbd->recv_workq);
1376                         return -ENOMEM;
1377                 }
1378                 sk_set_memalloc(config->socks[i]->sock->sk);
1379                 if (nbd->tag_set.timeout)
1380                         config->socks[i]->sock->sk->sk_sndtimeo =
1381                                 nbd->tag_set.timeout;
1382                 atomic_inc(&config->recv_threads);
1383                 refcount_inc(&nbd->config_refs);
1384                 INIT_WORK(&args->work, recv_work);
1385                 args->nbd = nbd;
1386                 args->index = i;
1387                 queue_work(nbd->recv_workq, &args->work);
1388         }
1389         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1390 }
1391
1392 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1393 {
1394         struct nbd_config *config = nbd->config;
1395         int ret;
1396
1397         ret = nbd_start_device(nbd);
1398         if (ret)
1399                 return ret;
1400
1401         if (max_part)
1402                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1403         mutex_unlock(&nbd->config_lock);
1404         ret = wait_event_interruptible(config->recv_wq,
1405                                          atomic_read(&config->recv_threads) == 0);
1406         if (ret)
1407                 sock_shutdown(nbd);
1408         flush_workqueue(nbd->recv_workq);
1409
1410         mutex_lock(&nbd->config_lock);
1411         nbd_bdev_reset(bdev);
1412         /* user requested, ignore socket errors */
1413         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1414                 ret = 0;
1415         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1416                 ret = -ETIMEDOUT;
1417         return ret;
1418 }
1419
1420 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1421                                  struct block_device *bdev)
1422 {
1423         sock_shutdown(nbd);
1424         __invalidate_device(bdev, true);
1425         nbd_bdev_reset(bdev);
1426         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1427                                &nbd->config->runtime_flags))
1428                 nbd_config_put(nbd);
1429 }
1430
1431 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1432 {
1433         nbd->tag_set.timeout = timeout * HZ;
1434         if (timeout)
1435                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1436         else
1437                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1438 }
1439
1440 /* Must be called with config_lock held */
1441 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1442                        unsigned int cmd, unsigned long arg)
1443 {
1444         struct nbd_config *config = nbd->config;
1445         loff_t bytesize;
1446
1447         switch (cmd) {
1448         case NBD_DISCONNECT:
1449                 return nbd_disconnect(nbd);
1450         case NBD_CLEAR_SOCK:
1451                 nbd_clear_sock_ioctl(nbd, bdev);
1452                 return 0;
1453         case NBD_SET_SOCK:
1454                 return nbd_add_socket(nbd, arg, false);
1455         case NBD_SET_BLKSIZE:
1456                 return nbd_set_size(nbd, config->bytesize, arg);
1457         case NBD_SET_SIZE:
1458                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1459         case NBD_SET_SIZE_BLOCKS:
1460                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1461                         return -EINVAL;
1462                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1463         case NBD_SET_TIMEOUT:
1464                 nbd_set_cmd_timeout(nbd, arg);
1465                 return 0;
1466
1467         case NBD_SET_FLAGS:
1468                 config->flags = arg;
1469                 return 0;
1470         case NBD_DO_IT:
1471                 return nbd_start_device_ioctl(nbd, bdev);
1472         case NBD_CLEAR_QUE:
1473                 /*
1474                  * This is for compatibility only.  The queue is always cleared
1475                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1476                  */
1477                 return 0;
1478         case NBD_PRINT_DEBUG:
1479                 /*
1480                  * For compatibility only, we no longer keep a list of
1481                  * outstanding requests.
1482                  */
1483                 return 0;
1484         }
1485         return -ENOTTY;
1486 }
1487
1488 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1489                      unsigned int cmd, unsigned long arg)
1490 {
1491         struct nbd_device *nbd = bdev->bd_disk->private_data;
1492         struct nbd_config *config = nbd->config;
1493         int error = -EINVAL;
1494
1495         if (!capable(CAP_SYS_ADMIN))
1496                 return -EPERM;
1497
1498         /* The block layer will pass back some non-nbd ioctls in case we have
1499          * special handling for them, but we don't so just return an error.
1500          */
1501         if (_IOC_TYPE(cmd) != 0xab)
1502                 return -EINVAL;
1503
1504         mutex_lock(&nbd->config_lock);
1505
1506         /* Don't allow ioctl operations on a nbd device that was created with
1507          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1508          */
1509         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1510             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1511                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1512         else
1513                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1514         mutex_unlock(&nbd->config_lock);
1515         return error;
1516 }
1517
1518 static struct nbd_config *nbd_alloc_config(void)
1519 {
1520         struct nbd_config *config;
1521
1522         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1523         if (!config)
1524                 return NULL;
1525         atomic_set(&config->recv_threads, 0);
1526         init_waitqueue_head(&config->recv_wq);
1527         init_waitqueue_head(&config->conn_wait);
1528         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1529         atomic_set(&config->live_connections, 0);
1530         try_module_get(THIS_MODULE);
1531         return config;
1532 }
1533
1534 static int nbd_open(struct block_device *bdev, fmode_t mode)
1535 {
1536         struct nbd_device *nbd;
1537         int ret = 0;
1538
1539         mutex_lock(&nbd_index_mutex);
1540         nbd = bdev->bd_disk->private_data;
1541         if (!nbd) {
1542                 ret = -ENXIO;
1543                 goto out;
1544         }
1545         if (!refcount_inc_not_zero(&nbd->refs)) {
1546                 ret = -ENXIO;
1547                 goto out;
1548         }
1549         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1550                 struct nbd_config *config;
1551
1552                 mutex_lock(&nbd->config_lock);
1553                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1554                         mutex_unlock(&nbd->config_lock);
1555                         goto out;
1556                 }
1557                 config = nbd->config = nbd_alloc_config();
1558                 if (!config) {
1559                         ret = -ENOMEM;
1560                         mutex_unlock(&nbd->config_lock);
1561                         goto out;
1562                 }
1563                 refcount_set(&nbd->config_refs, 1);
1564                 refcount_inc(&nbd->refs);
1565                 mutex_unlock(&nbd->config_lock);
1566                 if (max_part)
1567                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1568         } else if (nbd_disconnected(nbd->config)) {
1569                 if (max_part)
1570                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1571         }
1572 out:
1573         mutex_unlock(&nbd_index_mutex);
1574         return ret;
1575 }
1576
1577 static void nbd_release(struct gendisk *disk, fmode_t mode)
1578 {
1579         struct nbd_device *nbd = disk->private_data;
1580
1581         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1582                         disk->part0->bd_openers == 0)
1583                 nbd_disconnect_and_put(nbd);
1584
1585         nbd_config_put(nbd);
1586         nbd_put(nbd);
1587 }
1588
1589 static const struct block_device_operations nbd_fops =
1590 {
1591         .owner =        THIS_MODULE,
1592         .open =         nbd_open,
1593         .release =      nbd_release,
1594         .ioctl =        nbd_ioctl,
1595         .compat_ioctl = nbd_ioctl,
1596 };
1597
1598 #if IS_ENABLED(CONFIG_DEBUG_FS)
1599
1600 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1601 {
1602         struct nbd_device *nbd = s->private;
1603
1604         if (nbd->pid)
1605                 seq_printf(s, "recv: %d\n", nbd->pid);
1606
1607         return 0;
1608 }
1609
1610 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1611
1612 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1613 {
1614         struct nbd_device *nbd = s->private;
1615         u32 flags = nbd->config->flags;
1616
1617         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1618
1619         seq_puts(s, "Known flags:\n");
1620
1621         if (flags & NBD_FLAG_HAS_FLAGS)
1622                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1623         if (flags & NBD_FLAG_READ_ONLY)
1624                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1625         if (flags & NBD_FLAG_SEND_FLUSH)
1626                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1627         if (flags & NBD_FLAG_SEND_FUA)
1628                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1629         if (flags & NBD_FLAG_SEND_TRIM)
1630                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1631
1632         return 0;
1633 }
1634
1635 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1636
1637 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1638 {
1639         struct dentry *dir;
1640         struct nbd_config *config = nbd->config;
1641
1642         if (!nbd_dbg_dir)
1643                 return -EIO;
1644
1645         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1646         if (!dir) {
1647                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1648                         nbd_name(nbd));
1649                 return -EIO;
1650         }
1651         config->dbg_dir = dir;
1652
1653         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1654         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1655         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1656         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1657         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1658
1659         return 0;
1660 }
1661
1662 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1663 {
1664         debugfs_remove_recursive(nbd->config->dbg_dir);
1665 }
1666
1667 static int nbd_dbg_init(void)
1668 {
1669         struct dentry *dbg_dir;
1670
1671         dbg_dir = debugfs_create_dir("nbd", NULL);
1672         if (!dbg_dir)
1673                 return -EIO;
1674
1675         nbd_dbg_dir = dbg_dir;
1676
1677         return 0;
1678 }
1679
1680 static void nbd_dbg_close(void)
1681 {
1682         debugfs_remove_recursive(nbd_dbg_dir);
1683 }
1684
1685 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1686
1687 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1688 {
1689         return 0;
1690 }
1691
1692 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1693 {
1694 }
1695
1696 static int nbd_dbg_init(void)
1697 {
1698         return 0;
1699 }
1700
1701 static void nbd_dbg_close(void)
1702 {
1703 }
1704
1705 #endif
1706
1707 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1708                             unsigned int hctx_idx, unsigned int numa_node)
1709 {
1710         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1711         cmd->nbd = set->driver_data;
1712         cmd->flags = 0;
1713         mutex_init(&cmd->lock);
1714         return 0;
1715 }
1716
1717 static const struct blk_mq_ops nbd_mq_ops = {
1718         .queue_rq       = nbd_queue_rq,
1719         .complete       = nbd_complete_rq,
1720         .init_request   = nbd_init_request,
1721         .timeout        = nbd_xmit_timeout,
1722 };
1723
1724 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1725 {
1726         struct nbd_device *nbd;
1727         struct gendisk *disk;
1728         int err = -ENOMEM;
1729
1730         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1731         if (!nbd)
1732                 goto out;
1733
1734         nbd->tag_set.ops = &nbd_mq_ops;
1735         nbd->tag_set.nr_hw_queues = 1;
1736         nbd->tag_set.queue_depth = 128;
1737         nbd->tag_set.numa_node = NUMA_NO_NODE;
1738         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1739         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1740                 BLK_MQ_F_BLOCKING;
1741         nbd->tag_set.driver_data = nbd;
1742         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1743         nbd->backend = NULL;
1744
1745         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1746         if (err)
1747                 goto out_free_nbd;
1748
1749         mutex_lock(&nbd_index_mutex);
1750         if (index >= 0) {
1751                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1752                                 GFP_KERNEL);
1753                 if (err == -ENOSPC)
1754                         err = -EEXIST;
1755         } else {
1756                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1757                 if (err >= 0)
1758                         index = err;
1759         }
1760         nbd->index = index;
1761         mutex_unlock(&nbd_index_mutex);
1762         if (err < 0)
1763                 goto out_free_tags;
1764
1765         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1766         if (IS_ERR(disk)) {
1767                 err = PTR_ERR(disk);
1768                 goto out_free_idr;
1769         }
1770         nbd->disk = disk;
1771
1772         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1773                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1774                                           WQ_UNBOUND, 0, nbd->index);
1775         if (!nbd->recv_workq) {
1776                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1777                 err = -ENOMEM;
1778                 goto out_err_disk;
1779         }
1780
1781         /*
1782          * Tell the block layer that we are not a rotational device
1783          */
1784         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1785         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1786         disk->queue->limits.discard_granularity = 0;
1787         disk->queue->limits.discard_alignment = 0;
1788         blk_queue_max_discard_sectors(disk->queue, 0);
1789         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1790         blk_queue_max_segments(disk->queue, USHRT_MAX);
1791         blk_queue_max_hw_sectors(disk->queue, 65536);
1792         disk->queue->limits.max_sectors = 256;
1793
1794         mutex_init(&nbd->config_lock);
1795         refcount_set(&nbd->config_refs, 0);
1796         /*
1797          * Start out with a zero references to keep other threads from using
1798          * this device until it is fully initialized.
1799          */
1800         refcount_set(&nbd->refs, 0);
1801         INIT_LIST_HEAD(&nbd->list);
1802         disk->major = NBD_MAJOR;
1803
1804         /* Too big first_minor can cause duplicate creation of
1805          * sysfs files/links, since index << part_shift might overflow, or
1806          * MKDEV() expect that the max bits of first_minor is 20.
1807          */
1808         disk->first_minor = index << part_shift;
1809         if (disk->first_minor < index || disk->first_minor > MINORMASK) {
1810                 err = -EINVAL;
1811                 goto out_free_work;
1812         }
1813
1814         disk->minors = 1 << part_shift;
1815         disk->fops = &nbd_fops;
1816         disk->private_data = nbd;
1817         sprintf(disk->disk_name, "nbd%d", index);
1818         err = add_disk(disk);
1819         if (err)
1820                 goto out_free_work;
1821
1822         /*
1823          * Now publish the device.
1824          */
1825         refcount_set(&nbd->refs, refs);
1826         nbd_total_devices++;
1827         return nbd;
1828
1829 out_free_work:
1830         destroy_workqueue(nbd->recv_workq);
1831 out_err_disk:
1832         blk_cleanup_disk(disk);
1833 out_free_idr:
1834         mutex_lock(&nbd_index_mutex);
1835         idr_remove(&nbd_index_idr, index);
1836         mutex_unlock(&nbd_index_mutex);
1837 out_free_tags:
1838         blk_mq_free_tag_set(&nbd->tag_set);
1839 out_free_nbd:
1840         kfree(nbd);
1841 out:
1842         return ERR_PTR(err);
1843 }
1844
1845 static struct nbd_device *nbd_find_get_unused(void)
1846 {
1847         struct nbd_device *nbd;
1848         int id;
1849
1850         lockdep_assert_held(&nbd_index_mutex);
1851
1852         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1853                 if (refcount_read(&nbd->config_refs) ||
1854                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1855                         continue;
1856                 if (refcount_inc_not_zero(&nbd->refs))
1857                         return nbd;
1858         }
1859
1860         return NULL;
1861 }
1862
1863 /* Netlink interface. */
1864 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1865         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1866         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1867         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1868         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1869         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1870         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1871         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1872         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1873         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1874         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1875 };
1876
1877 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1878         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1879 };
1880
1881 /* We don't use this right now since we don't parse the incoming list, but we
1882  * still want it here so userspace knows what to expect.
1883  */
1884 static const struct nla_policy __attribute__((unused))
1885 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1886         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1887         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1888 };
1889
1890 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1891 {
1892         struct nbd_config *config = nbd->config;
1893         u64 bsize = nbd_blksize(config);
1894         u64 bytes = config->bytesize;
1895
1896         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1897                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1898
1899         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1900                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1901
1902         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1903                 return nbd_set_size(nbd, bytes, bsize);
1904         return 0;
1905 }
1906
1907 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1908 {
1909         struct nbd_device *nbd;
1910         struct nbd_config *config;
1911         int index = -1;
1912         int ret;
1913         bool put_dev = false;
1914
1915         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1916                 return -EPERM;
1917
1918         if (info->attrs[NBD_ATTR_INDEX])
1919                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1920         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1921                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1922                 return -EINVAL;
1923         }
1924         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1925                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1926                 return -EINVAL;
1927         }
1928 again:
1929         mutex_lock(&nbd_index_mutex);
1930         if (index == -1) {
1931                 nbd = nbd_find_get_unused();
1932         } else {
1933                 nbd = idr_find(&nbd_index_idr, index);
1934                 if (nbd) {
1935                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1936                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
1937                             !refcount_inc_not_zero(&nbd->refs)) {
1938                                 mutex_unlock(&nbd_index_mutex);
1939                                 pr_err("nbd: device at index %d is going down\n",
1940                                         index);
1941                                 return -EINVAL;
1942                         }
1943                 }
1944         }
1945         mutex_unlock(&nbd_index_mutex);
1946
1947         if (!nbd) {
1948                 nbd = nbd_dev_add(index, 2);
1949                 if (IS_ERR(nbd)) {
1950                         pr_err("nbd: failed to add new device\n");
1951                         return PTR_ERR(nbd);
1952                 }
1953         }
1954
1955         mutex_lock(&nbd->config_lock);
1956         if (refcount_read(&nbd->config_refs)) {
1957                 mutex_unlock(&nbd->config_lock);
1958                 nbd_put(nbd);
1959                 if (index == -1)
1960                         goto again;
1961                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1962                 return -EBUSY;
1963         }
1964         if (WARN_ON(nbd->config)) {
1965                 mutex_unlock(&nbd->config_lock);
1966                 nbd_put(nbd);
1967                 return -EINVAL;
1968         }
1969         config = nbd->config = nbd_alloc_config();
1970         if (!nbd->config) {
1971                 mutex_unlock(&nbd->config_lock);
1972                 nbd_put(nbd);
1973                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1974                 return -ENOMEM;
1975         }
1976         refcount_set(&nbd->config_refs, 1);
1977         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1978
1979         ret = nbd_genl_size_set(info, nbd);
1980         if (ret)
1981                 goto out;
1982
1983         if (info->attrs[NBD_ATTR_TIMEOUT])
1984                 nbd_set_cmd_timeout(nbd,
1985                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1986         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1987                 config->dead_conn_timeout =
1988                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1989                 config->dead_conn_timeout *= HZ;
1990         }
1991         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1992                 config->flags =
1993                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1994         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1995                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1996                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1997                         /*
1998                          * We have 1 ref to keep the device around, and then 1
1999                          * ref for our current operation here, which will be
2000                          * inherited by the config.  If we already have
2001                          * DESTROY_ON_DISCONNECT set then we know we don't have
2002                          * that extra ref already held so we don't need the
2003                          * put_dev.
2004                          */
2005                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2006                                               &nbd->flags))
2007                                 put_dev = true;
2008                 } else {
2009                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2010                                                &nbd->flags))
2011                                 refcount_inc(&nbd->refs);
2012                 }
2013                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2014                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2015                                 &config->runtime_flags);
2016                 }
2017         }
2018
2019         if (info->attrs[NBD_ATTR_SOCKETS]) {
2020                 struct nlattr *attr;
2021                 int rem, fd;
2022
2023                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2024                                     rem) {
2025                         struct nlattr *socks[NBD_SOCK_MAX+1];
2026
2027                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2028                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2029                                 ret = -EINVAL;
2030                                 goto out;
2031                         }
2032                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2033                                                           attr,
2034                                                           nbd_sock_policy,
2035                                                           info->extack);
2036                         if (ret != 0) {
2037                                 printk(KERN_ERR "nbd: error processing sock list\n");
2038                                 ret = -EINVAL;
2039                                 goto out;
2040                         }
2041                         if (!socks[NBD_SOCK_FD])
2042                                 continue;
2043                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2044                         ret = nbd_add_socket(nbd, fd, true);
2045                         if (ret)
2046                                 goto out;
2047                 }
2048         }
2049         ret = nbd_start_device(nbd);
2050         if (ret)
2051                 goto out;
2052         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2053                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2054                                           GFP_KERNEL);
2055                 if (!nbd->backend) {
2056                         ret = -ENOMEM;
2057                         goto out;
2058                 }
2059         }
2060         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2061         if (ret) {
2062                 dev_err(disk_to_dev(nbd->disk),
2063                         "device_create_file failed for backend!\n");
2064                 goto out;
2065         }
2066         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2067 out:
2068         mutex_unlock(&nbd->config_lock);
2069         if (!ret) {
2070                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2071                 refcount_inc(&nbd->config_refs);
2072                 nbd_connect_reply(info, nbd->index);
2073         }
2074         nbd_config_put(nbd);
2075         if (put_dev)
2076                 nbd_put(nbd);
2077         return ret;
2078 }
2079
2080 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2081 {
2082         mutex_lock(&nbd->config_lock);
2083         nbd_disconnect(nbd);
2084         sock_shutdown(nbd);
2085         /*
2086          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2087          * to cancel the inflight I/Os.
2088          */
2089         flush_workqueue(nbd->recv_workq);
2090         nbd_clear_que(nbd);
2091         nbd->task_setup = NULL;
2092         mutex_unlock(&nbd->config_lock);
2093
2094         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2095                                &nbd->config->runtime_flags))
2096                 nbd_config_put(nbd);
2097 }
2098
2099 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2100 {
2101         struct nbd_device *nbd;
2102         int index;
2103
2104         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2105                 return -EPERM;
2106
2107         if (!info->attrs[NBD_ATTR_INDEX]) {
2108                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2109                 return -EINVAL;
2110         }
2111         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2112         mutex_lock(&nbd_index_mutex);
2113         nbd = idr_find(&nbd_index_idr, index);
2114         if (!nbd) {
2115                 mutex_unlock(&nbd_index_mutex);
2116                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2117                        index);
2118                 return -EINVAL;
2119         }
2120         if (!refcount_inc_not_zero(&nbd->refs)) {
2121                 mutex_unlock(&nbd_index_mutex);
2122                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2123                        index);
2124                 return -EINVAL;
2125         }
2126         mutex_unlock(&nbd_index_mutex);
2127         if (!refcount_inc_not_zero(&nbd->config_refs))
2128                 goto put_nbd;
2129         nbd_disconnect_and_put(nbd);
2130         nbd_config_put(nbd);
2131 put_nbd:
2132         nbd_put(nbd);
2133         return 0;
2134 }
2135
2136 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2137 {
2138         struct nbd_device *nbd = NULL;
2139         struct nbd_config *config;
2140         int index;
2141         int ret = 0;
2142         bool put_dev = false;
2143
2144         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2145                 return -EPERM;
2146
2147         if (!info->attrs[NBD_ATTR_INDEX]) {
2148                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2149                 return -EINVAL;
2150         }
2151         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2152         mutex_lock(&nbd_index_mutex);
2153         nbd = idr_find(&nbd_index_idr, index);
2154         if (!nbd) {
2155                 mutex_unlock(&nbd_index_mutex);
2156                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2157                        index);
2158                 return -EINVAL;
2159         }
2160         if (nbd->backend) {
2161                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2162                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2163                                        nbd->backend)) {
2164                                 mutex_unlock(&nbd_index_mutex);
2165                                 dev_err(nbd_to_dev(nbd),
2166                                         "backend image doesn't match with %s\n",
2167                                         nbd->backend);
2168                                 return -EINVAL;
2169                         }
2170                 } else {
2171                         mutex_unlock(&nbd_index_mutex);
2172                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2173                         return -EINVAL;
2174                 }
2175         }
2176         if (!refcount_inc_not_zero(&nbd->refs)) {
2177                 mutex_unlock(&nbd_index_mutex);
2178                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2179                        index);
2180                 return -EINVAL;
2181         }
2182         mutex_unlock(&nbd_index_mutex);
2183
2184         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2185                 dev_err(nbd_to_dev(nbd),
2186                         "not configured, cannot reconfigure\n");
2187                 nbd_put(nbd);
2188                 return -EINVAL;
2189         }
2190
2191         mutex_lock(&nbd->config_lock);
2192         config = nbd->config;
2193         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2194             !nbd->pid) {
2195                 dev_err(nbd_to_dev(nbd),
2196                         "not configured, cannot reconfigure\n");
2197                 ret = -EINVAL;
2198                 goto out;
2199         }
2200
2201         ret = nbd_genl_size_set(info, nbd);
2202         if (ret)
2203                 goto out;
2204
2205         if (info->attrs[NBD_ATTR_TIMEOUT])
2206                 nbd_set_cmd_timeout(nbd,
2207                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2208         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2209                 config->dead_conn_timeout =
2210                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2211                 config->dead_conn_timeout *= HZ;
2212         }
2213         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2214                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2215                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2216                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2217                                               &nbd->flags))
2218                                 put_dev = true;
2219                 } else {
2220                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2221                                                &nbd->flags))
2222                                 refcount_inc(&nbd->refs);
2223                 }
2224
2225                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2226                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2227                                         &config->runtime_flags);
2228                 } else {
2229                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2230                                         &config->runtime_flags);
2231                 }
2232         }
2233
2234         if (info->attrs[NBD_ATTR_SOCKETS]) {
2235                 struct nlattr *attr;
2236                 int rem, fd;
2237
2238                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2239                                     rem) {
2240                         struct nlattr *socks[NBD_SOCK_MAX+1];
2241
2242                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2243                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2244                                 ret = -EINVAL;
2245                                 goto out;
2246                         }
2247                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2248                                                           attr,
2249                                                           nbd_sock_policy,
2250                                                           info->extack);
2251                         if (ret != 0) {
2252                                 printk(KERN_ERR "nbd: error processing sock list\n");
2253                                 ret = -EINVAL;
2254                                 goto out;
2255                         }
2256                         if (!socks[NBD_SOCK_FD])
2257                                 continue;
2258                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2259                         ret = nbd_reconnect_socket(nbd, fd);
2260                         if (ret) {
2261                                 if (ret == -ENOSPC)
2262                                         ret = 0;
2263                                 goto out;
2264                         }
2265                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2266                 }
2267         }
2268 out:
2269         mutex_unlock(&nbd->config_lock);
2270         nbd_config_put(nbd);
2271         nbd_put(nbd);
2272         if (put_dev)
2273                 nbd_put(nbd);
2274         return ret;
2275 }
2276
2277 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2278         {
2279                 .cmd    = NBD_CMD_CONNECT,
2280                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2281                 .doit   = nbd_genl_connect,
2282         },
2283         {
2284                 .cmd    = NBD_CMD_DISCONNECT,
2285                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2286                 .doit   = nbd_genl_disconnect,
2287         },
2288         {
2289                 .cmd    = NBD_CMD_RECONFIGURE,
2290                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2291                 .doit   = nbd_genl_reconfigure,
2292         },
2293         {
2294                 .cmd    = NBD_CMD_STATUS,
2295                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2296                 .doit   = nbd_genl_status,
2297         },
2298 };
2299
2300 static const struct genl_multicast_group nbd_mcast_grps[] = {
2301         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2302 };
2303
2304 static struct genl_family nbd_genl_family __ro_after_init = {
2305         .hdrsize        = 0,
2306         .name           = NBD_GENL_FAMILY_NAME,
2307         .version        = NBD_GENL_VERSION,
2308         .module         = THIS_MODULE,
2309         .small_ops      = nbd_connect_genl_ops,
2310         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2311         .maxattr        = NBD_ATTR_MAX,
2312         .policy = nbd_attr_policy,
2313         .mcgrps         = nbd_mcast_grps,
2314         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2315 };
2316
2317 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2318 {
2319         struct nlattr *dev_opt;
2320         u8 connected = 0;
2321         int ret;
2322
2323         /* This is a little racey, but for status it's ok.  The
2324          * reason we don't take a ref here is because we can't
2325          * take a ref in the index == -1 case as we would need
2326          * to put under the nbd_index_mutex, which could
2327          * deadlock if we are configured to remove ourselves
2328          * once we're disconnected.
2329          */
2330         if (refcount_read(&nbd->config_refs))
2331                 connected = 1;
2332         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2333         if (!dev_opt)
2334                 return -EMSGSIZE;
2335         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2336         if (ret)
2337                 return -EMSGSIZE;
2338         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2339                          connected);
2340         if (ret)
2341                 return -EMSGSIZE;
2342         nla_nest_end(reply, dev_opt);
2343         return 0;
2344 }
2345
2346 static int status_cb(int id, void *ptr, void *data)
2347 {
2348         struct nbd_device *nbd = ptr;
2349         return populate_nbd_status(nbd, (struct sk_buff *)data);
2350 }
2351
2352 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2353 {
2354         struct nlattr *dev_list;
2355         struct sk_buff *reply;
2356         void *reply_head;
2357         size_t msg_size;
2358         int index = -1;
2359         int ret = -ENOMEM;
2360
2361         if (info->attrs[NBD_ATTR_INDEX])
2362                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2363
2364         mutex_lock(&nbd_index_mutex);
2365
2366         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2367                                   nla_attr_size(sizeof(u8)));
2368         msg_size *= (index == -1) ? nbd_total_devices : 1;
2369
2370         reply = genlmsg_new(msg_size, GFP_KERNEL);
2371         if (!reply)
2372                 goto out;
2373         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2374                                        NBD_CMD_STATUS);
2375         if (!reply_head) {
2376                 nlmsg_free(reply);
2377                 goto out;
2378         }
2379
2380         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2381         if (index == -1) {
2382                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2383                 if (ret) {
2384                         nlmsg_free(reply);
2385                         goto out;
2386                 }
2387         } else {
2388                 struct nbd_device *nbd;
2389                 nbd = idr_find(&nbd_index_idr, index);
2390                 if (nbd) {
2391                         ret = populate_nbd_status(nbd, reply);
2392                         if (ret) {
2393                                 nlmsg_free(reply);
2394                                 goto out;
2395                         }
2396                 }
2397         }
2398         nla_nest_end(reply, dev_list);
2399         genlmsg_end(reply, reply_head);
2400         ret = genlmsg_reply(reply, info);
2401 out:
2402         mutex_unlock(&nbd_index_mutex);
2403         return ret;
2404 }
2405
2406 static void nbd_connect_reply(struct genl_info *info, int index)
2407 {
2408         struct sk_buff *skb;
2409         void *msg_head;
2410         int ret;
2411
2412         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2413         if (!skb)
2414                 return;
2415         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2416                                      NBD_CMD_CONNECT);
2417         if (!msg_head) {
2418                 nlmsg_free(skb);
2419                 return;
2420         }
2421         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2422         if (ret) {
2423                 nlmsg_free(skb);
2424                 return;
2425         }
2426         genlmsg_end(skb, msg_head);
2427         genlmsg_reply(skb, info);
2428 }
2429
2430 static void nbd_mcast_index(int index)
2431 {
2432         struct sk_buff *skb;
2433         void *msg_head;
2434         int ret;
2435
2436         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2437         if (!skb)
2438                 return;
2439         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2440                                      NBD_CMD_LINK_DEAD);
2441         if (!msg_head) {
2442                 nlmsg_free(skb);
2443                 return;
2444         }
2445         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2446         if (ret) {
2447                 nlmsg_free(skb);
2448                 return;
2449         }
2450         genlmsg_end(skb, msg_head);
2451         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2452 }
2453
2454 static void nbd_dead_link_work(struct work_struct *work)
2455 {
2456         struct link_dead_args *args = container_of(work, struct link_dead_args,
2457                                                    work);
2458         nbd_mcast_index(args->index);
2459         kfree(args);
2460 }
2461
2462 static int __init nbd_init(void)
2463 {
2464         int i;
2465
2466         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2467
2468         if (max_part < 0) {
2469                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2470                 return -EINVAL;
2471         }
2472
2473         part_shift = 0;
2474         if (max_part > 0) {
2475                 part_shift = fls(max_part);
2476
2477                 /*
2478                  * Adjust max_part according to part_shift as it is exported
2479                  * to user space so that user can know the max number of
2480                  * partition kernel should be able to manage.
2481                  *
2482                  * Note that -1 is required because partition 0 is reserved
2483                  * for the whole disk.
2484                  */
2485                 max_part = (1UL << part_shift) - 1;
2486         }
2487
2488         if ((1UL << part_shift) > DISK_MAX_PARTS)
2489                 return -EINVAL;
2490
2491         if (nbds_max > 1UL << (MINORBITS - part_shift))
2492                 return -EINVAL;
2493
2494         if (register_blkdev(NBD_MAJOR, "nbd"))
2495                 return -EIO;
2496
2497         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2498         if (!nbd_del_wq) {
2499                 unregister_blkdev(NBD_MAJOR, "nbd");
2500                 return -ENOMEM;
2501         }
2502
2503         if (genl_register_family(&nbd_genl_family)) {
2504                 destroy_workqueue(nbd_del_wq);
2505                 unregister_blkdev(NBD_MAJOR, "nbd");
2506                 return -EINVAL;
2507         }
2508         nbd_dbg_init();
2509
2510         for (i = 0; i < nbds_max; i++)
2511                 nbd_dev_add(i, 1);
2512         return 0;
2513 }
2514
2515 static int nbd_exit_cb(int id, void *ptr, void *data)
2516 {
2517         struct list_head *list = (struct list_head *)data;
2518         struct nbd_device *nbd = ptr;
2519
2520         /* Skip nbd that is being removed asynchronously */
2521         if (refcount_read(&nbd->refs))
2522                 list_add_tail(&nbd->list, list);
2523
2524         return 0;
2525 }
2526
2527 static void __exit nbd_cleanup(void)
2528 {
2529         struct nbd_device *nbd;
2530         LIST_HEAD(del_list);
2531
2532         nbd_dbg_close();
2533
2534         mutex_lock(&nbd_index_mutex);
2535         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2536         mutex_unlock(&nbd_index_mutex);
2537
2538         while (!list_empty(&del_list)) {
2539                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2540                 list_del_init(&nbd->list);
2541                 if (refcount_read(&nbd->refs) != 1)
2542                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2543                 nbd_put(nbd);
2544         }
2545
2546         /* Also wait for nbd_dev_remove_work() completes */
2547         destroy_workqueue(nbd_del_wq);
2548
2549         idr_destroy(&nbd_index_idr);
2550         genl_unregister_family(&nbd_genl_family);
2551         unregister_blkdev(NBD_MAJOR, "nbd");
2552 }
2553
2554 module_init(nbd_init);
2555 module_exit(nbd_cleanup);
2556
2557 MODULE_DESCRIPTION("Network Block Device");
2558 MODULE_LICENSE("GPL");
2559
2560 module_param(nbds_max, int, 0444);
2561 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2562 module_param(max_part, int, 0444);
2563 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");