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         put_disk(disk);
 254         blk_mq_free_tag_set(&nbd->tag_set);
 255
 256         /*
 257          * Remove from idr after del_gendisk() completes, so if the same ID is
 258          * reused, the following add_disk() will succeed.
 259          */
 260         mutex_lock(&nbd_index_mutex);
 261         idr_remove(&nbd_index_idr, nbd->index);
 262         mutex_unlock(&nbd_index_mutex);
 263         destroy_workqueue(nbd->recv_workq);
 264         kfree(nbd);
 265 }
 266
 267 static void nbd_dev_remove_work(struct work_struct *work)
 268 {
 269         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 270 }
 271
 272 static void nbd_put(struct nbd_device *nbd)
 273 {
 274         if (!refcount_dec_and_test(&nbd->refs))
 275                 return;
 276
 277         /* Call del_gendisk() asynchrounously to prevent deadlock */
 278         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 279                 queue_work(nbd_del_wq, &nbd->remove_work);
 280         else
 281                 nbd_dev_remove(nbd);
 282 }
 283
 284 static int nbd_disconnected(struct nbd_config *config)
 285 {
 286         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 287                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 288 }
 289
 290 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 291                                 int notify)
 292 {
 293         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 294                 struct link_dead_args *args;
 295                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 296                 if (args) {
 297                         INIT_WORK(&args->work, nbd_dead_link_work);
 298                         args->index = nbd->index;
 299                         queue_work(system_wq, &args->work);
 300                 }
 301         }
 302         if (!nsock->dead) {
 303                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 304                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 305                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 306                                                &nbd->config->runtime_flags)) {
 307                                 set_bit(NBD_RT_DISCONNECTED,
 308                                         &nbd->config->runtime_flags);
 309                                 dev_info(nbd_to_dev(nbd),
 310                                         "Disconnected due to user request.\n");
 311                         }
 312                 }
 313         }
 314         nsock->dead = true;
 315         nsock->pending = NULL;
 316         nsock->sent = 0;
 317 }
 318
 319 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 320                 loff_t blksize)
 321 {
 322         if (!blksize)
 323                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 324
 325         if (blk_validate_block_size(blksize))
 326                 return -EINVAL;
 327
 328         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                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 337         }
 338         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 339         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 340
 341         if (max_part)
 342                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 343         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 344                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 345         return 0;
 346 }
 347
 348 static void nbd_complete_rq(struct request *req)
 349 {
 350         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 351
 352         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 353                 cmd->status ? "failed" : "done");
 354
 355         blk_mq_end_request(req, cmd->status);
 356 }
 357
 358 /*
 359  * Forcibly shutdown the socket causing all listeners to error
 360  */
 361 static void sock_shutdown(struct nbd_device *nbd)
 362 {
 363         struct nbd_config *config = nbd->config;
 364         int i;
 365
 366         if (config->num_connections == 0)
 367                 return;
 368         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 369                 return;
 370
 371         for (i = 0; i < config->num_connections; i++) {
 372                 struct nbd_sock *nsock = config->socks[i];
 373                 mutex_lock(&nsock->tx_lock);
 374                 nbd_mark_nsock_dead(nbd, nsock, 0);
 375                 mutex_unlock(&nsock->tx_lock);
 376         }
 377         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 378 }
 379
 380 static u32 req_to_nbd_cmd_type(struct request *req)
 381 {
 382         switch (req_op(req)) {
 383         case REQ_OP_DISCARD:
 384                 return NBD_CMD_TRIM;
 385         case REQ_OP_FLUSH:
 386                 return NBD_CMD_FLUSH;
 387         case REQ_OP_WRITE:
 388                 return NBD_CMD_WRITE;
 389         case REQ_OP_READ:
 390                 return NBD_CMD_READ;
 391         default:
 392                 return U32_MAX;
 393         }
 394 }
 395
 396 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req)
 397 {
 398         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 399         struct nbd_device *nbd = cmd->nbd;
 400         struct nbd_config *config;
 401
 402         if (!mutex_trylock(&cmd->lock))
 403                 return BLK_EH_RESET_TIMER;
 404
 405         if (!test_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 406                 mutex_unlock(&cmd->lock);
 407                 return BLK_EH_DONE;
 408         }
 409
 410         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 411                 cmd->status = BLK_STS_TIMEOUT;
 412                 __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 413                 mutex_unlock(&cmd->lock);
 414                 goto done;
 415         }
 416         config = nbd->config;
 417
 418         if (config->num_connections > 1 ||
 419             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 420                 dev_err_ratelimited(nbd_to_dev(nbd),
 421                                     "Connection timed out, retrying (%d/%d alive)\n",
 422                                     atomic_read(&config->live_connections),
 423                                     config->num_connections);
 424                 /*
 425                  * Hooray we have more connections, requeue this IO, the submit
 426                  * path will put it on a real connection. Or if only one
 427                  * connection is configured, the submit path will wait util
 428                  * a new connection is reconfigured or util dead timeout.
 429                  */
 430                 if (config->socks) {
 431                         if (cmd->index < config->num_connections) {
 432                                 struct nbd_sock *nsock =
 433                                         config->socks[cmd->index];
 434                                 mutex_lock(&nsock->tx_lock);
 435                                 /* We can have multiple outstanding requests, so
 436                                  * we don't want to mark the nsock dead if we've
 437                                  * already reconnected with a new socket, so
 438                                  * only mark it dead if its the same socket we
 439                                  * were sent out on.
 440                                  */
 441                                 if (cmd->cookie == nsock->cookie)
 442                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 443                                 mutex_unlock(&nsock->tx_lock);
 444                         }
 445                         mutex_unlock(&cmd->lock);
 446                         nbd_requeue_cmd(cmd);
 447                         nbd_config_put(nbd);
 448                         return BLK_EH_DONE;
 449                 }
 450         }
 451
 452         if (!nbd->tag_set.timeout) {
 453                 /*
 454                  * Userspace sets timeout=0 to disable socket disconnection,
 455                  * so just warn and reset the timer.
 456                  */
 457                 struct nbd_sock *nsock = config->socks[cmd->index];
 458                 cmd->retries++;
 459                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 460                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 461                         (unsigned long long)blk_rq_pos(req) << 9,
 462                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 463
 464                 mutex_lock(&nsock->tx_lock);
 465                 if (cmd->cookie != nsock->cookie) {
 466                         nbd_requeue_cmd(cmd);
 467                         mutex_unlock(&nsock->tx_lock);
 468                         mutex_unlock(&cmd->lock);
 469                         nbd_config_put(nbd);
 470                         return BLK_EH_DONE;
 471                 }
 472                 mutex_unlock(&nsock->tx_lock);
 473                 mutex_unlock(&cmd->lock);
 474                 nbd_config_put(nbd);
 475                 return BLK_EH_RESET_TIMER;
 476         }
 477
 478         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 479         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 480         cmd->status = BLK_STS_IOERR;
 481         __clear_bit(NBD_CMD_INFLIGHT, &cmd->flags);
 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_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                         bool complete;
 860
 861                         mutex_lock(&cmd->lock);
 862                         complete = __test_and_clear_bit(NBD_CMD_INFLIGHT,
 863                                                         &cmd->flags);
 864                         mutex_unlock(&cmd->lock);
 865                         if (complete)
 866                                 blk_mq_complete_request(rq);
 867                 }
 868                 percpu_ref_put(&q->q_usage_counter);
 869         }
 870
 871         nsock = config->socks[args->index];
 872         mutex_lock(&nsock->tx_lock);
 873         nbd_mark_nsock_dead(nbd, nsock, 1);
 874         mutex_unlock(&nsock->tx_lock);
 875
 876         nbd_config_put(nbd);
 877         atomic_dec(&config->recv_threads);
 878         wake_up(&config->recv_wq);
 879         kfree(args);
 880 }
 881
 882 static bool nbd_clear_req(struct request *req, void *data)
 883 {
 884         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 885
 886         /* don't abort one completed request */
 887         if (blk_mq_request_completed(req))
 888                 return true;
 889
 890         mutex_lock(&cmd->lock);
 891         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 892                 mutex_unlock(&cmd->lock);
 893                 return true;
 894         }
 895         cmd->status = BLK_STS_IOERR;
 896         mutex_unlock(&cmd->lock);
 897
 898         blk_mq_complete_request(req);
 899         return true;
 900 }
 901
 902 static void nbd_clear_que(struct nbd_device *nbd)
 903 {
 904         blk_mq_quiesce_queue(nbd->disk->queue);
 905         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 906         blk_mq_unquiesce_queue(nbd->disk->queue);
 907         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 908 }
 909
 910 static int find_fallback(struct nbd_device *nbd, int index)
 911 {
 912         struct nbd_config *config = nbd->config;
 913         int new_index = -1;
 914         struct nbd_sock *nsock = config->socks[index];
 915         int fallback = nsock->fallback_index;
 916
 917         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 918                 return new_index;
 919
 920         if (config->num_connections <= 1) {
 921                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 922                                     "Dead connection, failed to find a fallback\n");
 923                 return new_index;
 924         }
 925
 926         if (fallback >= 0 && fallback < config->num_connections &&
 927             !config->socks[fallback]->dead)
 928                 return fallback;
 929
 930         if (nsock->fallback_index < 0 ||
 931             nsock->fallback_index >= config->num_connections ||
 932             config->socks[nsock->fallback_index]->dead) {
 933                 int i;
 934                 for (i = 0; i < config->num_connections; i++) {
 935                         if (i == index)
 936                                 continue;
 937                         if (!config->socks[i]->dead) {
 938                                 new_index = i;
 939                                 break;
 940                         }
 941                 }
 942                 nsock->fallback_index = new_index;
 943                 if (new_index < 0) {
 944                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 945                                             "Dead connection, failed to find a fallback\n");
 946                         return new_index;
 947                 }
 948         }
 949         new_index = nsock->fallback_index;
 950         return new_index;
 951 }
 952
 953 static int wait_for_reconnect(struct nbd_device *nbd)
 954 {
 955         struct nbd_config *config = nbd->config;
 956         if (!config->dead_conn_timeout)
 957                 return 0;
 958
 959         if (!wait_event_timeout(config->conn_wait,
 960                                 test_bit(NBD_RT_DISCONNECTED,
 961                                          &config->runtime_flags) ||
 962                                 atomic_read(&config->live_connections) > 0,
 963                                 config->dead_conn_timeout))
 964                 return 0;
 965
 966         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
 967 }
 968
 969 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 970 {
 971         struct request *req = blk_mq_rq_from_pdu(cmd);
 972         struct nbd_device *nbd = cmd->nbd;
 973         struct nbd_config *config;
 974         struct nbd_sock *nsock;
 975         int ret;
 976
 977         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 978                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 979                                     "Socks array is empty\n");
 980                 return -EINVAL;
 981         }
 982         config = nbd->config;
 983
 984         if (index >= config->num_connections) {
 985                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 986                                     "Attempted send on invalid socket\n");
 987                 nbd_config_put(nbd);
 988                 return -EINVAL;
 989         }
 990         cmd->status = BLK_STS_OK;
 991 again:
 992         nsock = config->socks[index];
 993         mutex_lock(&nsock->tx_lock);
 994         if (nsock->dead) {
 995                 int old_index = index;
 996                 index = find_fallback(nbd, index);
 997                 mutex_unlock(&nsock->tx_lock);
 998                 if (index < 0) {
 999                         if (wait_for_reconnect(nbd)) {
1000                                 index = old_index;
1001                                 goto again;
1002                         }
1003                         /* All the sockets should already be down at this point,
1004                          * we just want to make sure that DISCONNECTED is set so
1005                          * any requests that come in that were queue'ed waiting
1006                          * for the reconnect timer don't trigger the timer again
1007                          * and instead just error out.
1008                          */
1009                         sock_shutdown(nbd);
1010                         nbd_config_put(nbd);
1011                         return -EIO;
1012                 }
1013                 goto again;
1014         }
1015
1016         /* Handle the case that we have a pending request that was partially
1017          * transmitted that _has_ to be serviced first.  We need to call requeue
1018          * here so that it gets put _after_ the request that is already on the
1019          * dispatch list.
1020          */
1021         blk_mq_start_request(req);
1022         if (unlikely(nsock->pending && nsock->pending != req)) {
1023                 nbd_requeue_cmd(cmd);
1024                 ret = 0;
1025                 goto out;
1026         }
1027         /*
1028          * Some failures are related to the link going down, so anything that
1029          * returns EAGAIN can be retried on a different socket.
1030          */
1031         ret = nbd_send_cmd(nbd, cmd, index);
1032         /*
1033          * Access to this flag is protected by cmd->lock, thus it's safe to set
1034          * the flag after nbd_send_cmd() succeed to send request to server.
1035          */
1036         if (!ret)
1037                 __set_bit(NBD_CMD_INFLIGHT, &cmd->flags);
1038         else if (ret == -EAGAIN) {
1039                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1040                                     "Request send failed, requeueing\n");
1041                 nbd_mark_nsock_dead(nbd, nsock, 1);
1042                 nbd_requeue_cmd(cmd);
1043                 ret = 0;
1044         }
1045 out:
1046         mutex_unlock(&nsock->tx_lock);
1047         nbd_config_put(nbd);
1048         return ret;
1049 }
1050
1051 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1052                         const struct blk_mq_queue_data *bd)
1053 {
1054         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1055         int ret;
1056
1057         /*
1058          * Since we look at the bio's to send the request over the network we
1059          * need to make sure the completion work doesn't mark this request done
1060          * before we are done doing our send.  This keeps us from dereferencing
1061          * freed data if we have particularly fast completions (ie we get the
1062          * completion before we exit sock_xmit on the last bvec) or in the case
1063          * that the server is misbehaving (or there was an error) before we're
1064          * done sending everything over the wire.
1065          */
1066         mutex_lock(&cmd->lock);
1067         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1068
1069         /* We can be called directly from the user space process, which means we
1070          * could possibly have signals pending so our sendmsg will fail.  In
1071          * this case we need to return that we are busy, otherwise error out as
1072          * appropriate.
1073          */
1074         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1075         if (ret < 0)
1076                 ret = BLK_STS_IOERR;
1077         else if (!ret)
1078                 ret = BLK_STS_OK;
1079         mutex_unlock(&cmd->lock);
1080
1081         return ret;
1082 }
1083
1084 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1085                                      int *err)
1086 {
1087         struct socket *sock;
1088
1089         *err = 0;
1090         sock = sockfd_lookup(fd, err);
1091         if (!sock)
1092                 return NULL;
1093
1094         if (sock->ops->shutdown == sock_no_shutdown) {
1095                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1096                 *err = -EINVAL;
1097                 sockfd_put(sock);
1098                 return NULL;
1099         }
1100
1101         return sock;
1102 }
1103
1104 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1105                           bool netlink)
1106 {
1107         struct nbd_config *config = nbd->config;
1108         struct socket *sock;
1109         struct nbd_sock **socks;
1110         struct nbd_sock *nsock;
1111         int err;
1112
1113         sock = nbd_get_socket(nbd, arg, &err);
1114         if (!sock)
1115                 return err;
1116
1117         /*
1118          * We need to make sure we don't get any errant requests while we're
1119          * reallocating the ->socks array.
1120          */
1121         blk_mq_freeze_queue(nbd->disk->queue);
1122
1123         if (!netlink && !nbd->task_setup &&
1124             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1125                 nbd->task_setup = current;
1126
1127         if (!netlink &&
1128             (nbd->task_setup != current ||
1129              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1130                 dev_err(disk_to_dev(nbd->disk),
1131                         "Device being setup by another task");
1132                 err = -EBUSY;
1133                 goto put_socket;
1134         }
1135
1136         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1137         if (!nsock) {
1138                 err = -ENOMEM;
1139                 goto put_socket;
1140         }
1141
1142         socks = krealloc(config->socks, (config->num_connections + 1) *
1143                          sizeof(struct nbd_sock *), GFP_KERNEL);
1144         if (!socks) {
1145                 kfree(nsock);
1146                 err = -ENOMEM;
1147                 goto put_socket;
1148         }
1149
1150         config->socks = socks;
1151
1152         nsock->fallback_index = -1;
1153         nsock->dead = false;
1154         mutex_init(&nsock->tx_lock);
1155         nsock->sock = sock;
1156         nsock->pending = NULL;
1157         nsock->sent = 0;
1158         nsock->cookie = 0;
1159         socks[config->num_connections++] = nsock;
1160         atomic_inc(&config->live_connections);
1161         blk_mq_unfreeze_queue(nbd->disk->queue);
1162
1163         return 0;
1164
1165 put_socket:
1166         blk_mq_unfreeze_queue(nbd->disk->queue);
1167         sockfd_put(sock);
1168         return err;
1169 }
1170
1171 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1172 {
1173         struct nbd_config *config = nbd->config;
1174         struct socket *sock, *old;
1175         struct recv_thread_args *args;
1176         int i;
1177         int err;
1178
1179         sock = nbd_get_socket(nbd, arg, &err);
1180         if (!sock)
1181                 return err;
1182
1183         args = kzalloc(sizeof(*args), GFP_KERNEL);
1184         if (!args) {
1185                 sockfd_put(sock);
1186                 return -ENOMEM;
1187         }
1188
1189         for (i = 0; i < config->num_connections; i++) {
1190                 struct nbd_sock *nsock = config->socks[i];
1191
1192                 if (!nsock->dead)
1193                         continue;
1194
1195                 mutex_lock(&nsock->tx_lock);
1196                 if (!nsock->dead) {
1197                         mutex_unlock(&nsock->tx_lock);
1198                         continue;
1199                 }
1200                 sk_set_memalloc(sock->sk);
1201                 if (nbd->tag_set.timeout)
1202                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1203                 atomic_inc(&config->recv_threads);
1204                 refcount_inc(&nbd->config_refs);
1205                 old = nsock->sock;
1206                 nsock->fallback_index = -1;
1207                 nsock->sock = sock;
1208                 nsock->dead = false;
1209                 INIT_WORK(&args->work, recv_work);
1210                 args->index = i;
1211                 args->nbd = nbd;
1212                 nsock->cookie++;
1213                 mutex_unlock(&nsock->tx_lock);
1214                 sockfd_put(old);
1215
1216                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1217
1218                 /* We take the tx_mutex in an error path in the recv_work, so we
1219                  * need to queue_work outside of the tx_mutex.
1220                  */
1221                 queue_work(nbd->recv_workq, &args->work);
1222
1223                 atomic_inc(&config->live_connections);
1224                 wake_up(&config->conn_wait);
1225                 return 0;
1226         }
1227         sockfd_put(sock);
1228         kfree(args);
1229         return -ENOSPC;
1230 }
1231
1232 static void nbd_bdev_reset(struct nbd_device *nbd)
1233 {
1234         if (disk_openers(nbd->disk) > 1)
1235                 return;
1236         set_capacity(nbd->disk, 0);
1237 }
1238
1239 static void nbd_parse_flags(struct nbd_device *nbd)
1240 {
1241         struct nbd_config *config = nbd->config;
1242         if (config->flags & NBD_FLAG_READ_ONLY)
1243                 set_disk_ro(nbd->disk, true);
1244         else
1245                 set_disk_ro(nbd->disk, false);
1246         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1247                 if (config->flags & NBD_FLAG_SEND_FUA)
1248                         blk_queue_write_cache(nbd->disk->queue, true, true);
1249                 else
1250                         blk_queue_write_cache(nbd->disk->queue, true, false);
1251         }
1252         else
1253                 blk_queue_write_cache(nbd->disk->queue, false, false);
1254 }
1255
1256 static void send_disconnects(struct nbd_device *nbd)
1257 {
1258         struct nbd_config *config = nbd->config;
1259         struct nbd_request request = {
1260                 .magic = htonl(NBD_REQUEST_MAGIC),
1261                 .type = htonl(NBD_CMD_DISC),
1262         };
1263         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1264         struct iov_iter from;
1265         int i, ret;
1266
1267         for (i = 0; i < config->num_connections; i++) {
1268                 struct nbd_sock *nsock = config->socks[i];
1269
1270                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1271                 mutex_lock(&nsock->tx_lock);
1272                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1273                 if (ret < 0)
1274                         dev_err(disk_to_dev(nbd->disk),
1275                                 "Send disconnect failed %d\n", ret);
1276                 mutex_unlock(&nsock->tx_lock);
1277         }
1278 }
1279
1280 static int nbd_disconnect(struct nbd_device *nbd)
1281 {
1282         struct nbd_config *config = nbd->config;
1283
1284         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1285         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1286         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1287         send_disconnects(nbd);
1288         return 0;
1289 }
1290
1291 static void nbd_clear_sock(struct nbd_device *nbd)
1292 {
1293         sock_shutdown(nbd);
1294         nbd_clear_que(nbd);
1295         nbd->task_setup = NULL;
1296 }
1297
1298 static void nbd_config_put(struct nbd_device *nbd)
1299 {
1300         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1301                                         &nbd->config_lock)) {
1302                 struct nbd_config *config = nbd->config;
1303                 nbd_dev_dbg_close(nbd);
1304                 invalidate_disk(nbd->disk);
1305                 if (nbd->config->bytesize)
1306                         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
1307                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1308                                        &config->runtime_flags))
1309                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1310                 nbd->pid = 0;
1311                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1312                                        &config->runtime_flags)) {
1313                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1314                         kfree(nbd->backend);
1315                         nbd->backend = NULL;
1316                 }
1317                 nbd_clear_sock(nbd);
1318                 if (config->num_connections) {
1319                         int i;
1320                         for (i = 0; i < config->num_connections; i++) {
1321                                 sockfd_put(config->socks[i]->sock);
1322                                 kfree(config->socks[i]);
1323                         }
1324                         kfree(config->socks);
1325                 }
1326                 kfree(nbd->config);
1327                 nbd->config = NULL;
1328
1329                 nbd->tag_set.timeout = 0;
1330                 nbd->disk->queue->limits.discard_granularity = 0;
1331                 blk_queue_max_discard_sectors(nbd->disk->queue, 0);
1332
1333                 mutex_unlock(&nbd->config_lock);
1334                 nbd_put(nbd);
1335                 module_put(THIS_MODULE);
1336         }
1337 }
1338
1339 static int nbd_start_device(struct nbd_device *nbd)
1340 {
1341         struct nbd_config *config = nbd->config;
1342         int num_connections = config->num_connections;
1343         int error = 0, i;
1344
1345         if (nbd->pid)
1346                 return -EBUSY;
1347         if (!config->socks)
1348                 return -EINVAL;
1349         if (num_connections > 1 &&
1350             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1351                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1352                 return -EINVAL;
1353         }
1354
1355         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1356         nbd->pid = task_pid_nr(current);
1357
1358         nbd_parse_flags(nbd);
1359
1360         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1361         if (error) {
1362                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1363                 return error;
1364         }
1365         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1366
1367         nbd_dev_dbg_init(nbd);
1368         for (i = 0; i < num_connections; i++) {
1369                 struct recv_thread_args *args;
1370
1371                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1372                 if (!args) {
1373                         sock_shutdown(nbd);
1374                         /*
1375                          * If num_connections is m (2 < m),
1376                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1377                          * But NO.(n + 1) failed. We still have n recv threads.
1378                          * So, add flush_workqueue here to prevent recv threads
1379                          * dropping the last config_refs and trying to destroy
1380                          * the workqueue from inside the workqueue.
1381                          */
1382                         if (i)
1383                                 flush_workqueue(nbd->recv_workq);
1384                         return -ENOMEM;
1385                 }
1386                 sk_set_memalloc(config->socks[i]->sock->sk);
1387                 if (nbd->tag_set.timeout)
1388                         config->socks[i]->sock->sk->sk_sndtimeo =
1389                                 nbd->tag_set.timeout;
1390                 atomic_inc(&config->recv_threads);
1391                 refcount_inc(&nbd->config_refs);
1392                 INIT_WORK(&args->work, recv_work);
1393                 args->nbd = nbd;
1394                 args->index = i;
1395                 queue_work(nbd->recv_workq, &args->work);
1396         }
1397         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1398 }
1399
1400 static int nbd_start_device_ioctl(struct nbd_device *nbd)
1401 {
1402         struct nbd_config *config = nbd->config;
1403         int ret;
1404
1405         ret = nbd_start_device(nbd);
1406         if (ret)
1407                 return ret;
1408
1409         if (max_part)
1410                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1411         mutex_unlock(&nbd->config_lock);
1412         ret = wait_event_interruptible(config->recv_wq,
1413                                          atomic_read(&config->recv_threads) == 0);
1414         if (ret)
1415                 sock_shutdown(nbd);
1416         flush_workqueue(nbd->recv_workq);
1417
1418         mutex_lock(&nbd->config_lock);
1419         nbd_bdev_reset(nbd);
1420         /* user requested, ignore socket errors */
1421         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1422                 ret = 0;
1423         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1424                 ret = -ETIMEDOUT;
1425         return ret;
1426 }
1427
1428 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1429                                  struct block_device *bdev)
1430 {
1431         nbd_clear_sock(nbd);
1432         __invalidate_device(bdev, true);
1433         nbd_bdev_reset(nbd);
1434         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1435                                &nbd->config->runtime_flags))
1436                 nbd_config_put(nbd);
1437 }
1438
1439 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1440 {
1441         nbd->tag_set.timeout = timeout * HZ;
1442         if (timeout)
1443                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1444         else
1445                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1446 }
1447
1448 /* Must be called with config_lock held */
1449 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1450                        unsigned int cmd, unsigned long arg)
1451 {
1452         struct nbd_config *config = nbd->config;
1453         loff_t bytesize;
1454
1455         switch (cmd) {
1456         case NBD_DISCONNECT:
1457                 return nbd_disconnect(nbd);
1458         case NBD_CLEAR_SOCK:
1459                 nbd_clear_sock_ioctl(nbd, bdev);
1460                 return 0;
1461         case NBD_SET_SOCK:
1462                 return nbd_add_socket(nbd, arg, false);
1463         case NBD_SET_BLKSIZE:
1464                 return nbd_set_size(nbd, config->bytesize, arg);
1465         case NBD_SET_SIZE:
1466                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1467         case NBD_SET_SIZE_BLOCKS:
1468                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1469                         return -EINVAL;
1470                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1471         case NBD_SET_TIMEOUT:
1472                 nbd_set_cmd_timeout(nbd, arg);
1473                 return 0;
1474
1475         case NBD_SET_FLAGS:
1476                 config->flags = arg;
1477                 return 0;
1478         case NBD_DO_IT:
1479                 return nbd_start_device_ioctl(nbd);
1480         case NBD_CLEAR_QUE:
1481                 /*
1482                  * This is for compatibility only.  The queue is always cleared
1483                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1484                  */
1485                 return 0;
1486         case NBD_PRINT_DEBUG:
1487                 /*
1488                  * For compatibility only, we no longer keep a list of
1489                  * outstanding requests.
1490                  */
1491                 return 0;
1492         }
1493         return -ENOTTY;
1494 }
1495
1496 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1497                      unsigned int cmd, unsigned long arg)
1498 {
1499         struct nbd_device *nbd = bdev->bd_disk->private_data;
1500         struct nbd_config *config = nbd->config;
1501         int error = -EINVAL;
1502
1503         if (!capable(CAP_SYS_ADMIN))
1504                 return -EPERM;
1505
1506         /* The block layer will pass back some non-nbd ioctls in case we have
1507          * special handling for them, but we don't so just return an error.
1508          */
1509         if (_IOC_TYPE(cmd) != 0xab)
1510                 return -EINVAL;
1511
1512         mutex_lock(&nbd->config_lock);
1513
1514         /* Don't allow ioctl operations on a nbd device that was created with
1515          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1516          */
1517         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1518             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1519                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1520         else
1521                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1522         mutex_unlock(&nbd->config_lock);
1523         return error;
1524 }
1525
1526 static struct nbd_config *nbd_alloc_config(void)
1527 {
1528         struct nbd_config *config;
1529
1530         if (!try_module_get(THIS_MODULE))
1531                 return ERR_PTR(-ENODEV);
1532
1533         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1534         if (!config) {
1535                 module_put(THIS_MODULE);
1536                 return ERR_PTR(-ENOMEM);
1537         }
1538
1539         atomic_set(&config->recv_threads, 0);
1540         init_waitqueue_head(&config->recv_wq);
1541         init_waitqueue_head(&config->conn_wait);
1542         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1543         atomic_set(&config->live_connections, 0);
1544         return config;
1545 }
1546
1547 static int nbd_open(struct block_device *bdev, fmode_t mode)
1548 {
1549         struct nbd_device *nbd;
1550         int ret = 0;
1551
1552         mutex_lock(&nbd_index_mutex);
1553         nbd = bdev->bd_disk->private_data;
1554         if (!nbd) {
1555                 ret = -ENXIO;
1556                 goto out;
1557         }
1558         if (!refcount_inc_not_zero(&nbd->refs)) {
1559                 ret = -ENXIO;
1560                 goto out;
1561         }
1562         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1563                 struct nbd_config *config;
1564
1565                 mutex_lock(&nbd->config_lock);
1566                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1567                         mutex_unlock(&nbd->config_lock);
1568                         goto out;
1569                 }
1570                 config = nbd_alloc_config();
1571                 if (IS_ERR(config)) {
1572                         ret = PTR_ERR(config);
1573                         mutex_unlock(&nbd->config_lock);
1574                         goto out;
1575                 }
1576                 nbd->config = config;
1577                 refcount_set(&nbd->config_refs, 1);
1578                 refcount_inc(&nbd->refs);
1579                 mutex_unlock(&nbd->config_lock);
1580                 if (max_part)
1581                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1582         } else if (nbd_disconnected(nbd->config)) {
1583                 if (max_part)
1584                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1585         }
1586 out:
1587         mutex_unlock(&nbd_index_mutex);
1588         return ret;
1589 }
1590
1591 static void nbd_release(struct gendisk *disk, fmode_t mode)
1592 {
1593         struct nbd_device *nbd = disk->private_data;
1594
1595         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1596                         disk_openers(disk) == 0)
1597                 nbd_disconnect_and_put(nbd);
1598
1599         nbd_config_put(nbd);
1600         nbd_put(nbd);
1601 }
1602
1603 static const struct block_device_operations nbd_fops =
1604 {
1605         .owner =        THIS_MODULE,
1606         .open =         nbd_open,
1607         .release =      nbd_release,
1608         .ioctl =        nbd_ioctl,
1609         .compat_ioctl = nbd_ioctl,
1610 };
1611
1612 #if IS_ENABLED(CONFIG_DEBUG_FS)
1613
1614 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1615 {
1616         struct nbd_device *nbd = s->private;
1617
1618         if (nbd->pid)
1619                 seq_printf(s, "recv: %d\n", nbd->pid);
1620
1621         return 0;
1622 }
1623
1624 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1625
1626 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1627 {
1628         struct nbd_device *nbd = s->private;
1629         u32 flags = nbd->config->flags;
1630
1631         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1632
1633         seq_puts(s, "Known flags:\n");
1634
1635         if (flags & NBD_FLAG_HAS_FLAGS)
1636                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1637         if (flags & NBD_FLAG_READ_ONLY)
1638                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1639         if (flags & NBD_FLAG_SEND_FLUSH)
1640                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1641         if (flags & NBD_FLAG_SEND_FUA)
1642                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1643         if (flags & NBD_FLAG_SEND_TRIM)
1644                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1645
1646         return 0;
1647 }
1648
1649 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1650
1651 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1652 {
1653         struct dentry *dir;
1654         struct nbd_config *config = nbd->config;
1655
1656         if (!nbd_dbg_dir)
1657                 return -EIO;
1658
1659         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1660         if (!dir) {
1661                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1662                         nbd_name(nbd));
1663                 return -EIO;
1664         }
1665         config->dbg_dir = dir;
1666
1667         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1668         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1669         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1670         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1671         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1672
1673         return 0;
1674 }
1675
1676 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1677 {
1678         debugfs_remove_recursive(nbd->config->dbg_dir);
1679 }
1680
1681 static int nbd_dbg_init(void)
1682 {
1683         struct dentry *dbg_dir;
1684
1685         dbg_dir = debugfs_create_dir("nbd", NULL);
1686         if (!dbg_dir)
1687                 return -EIO;
1688
1689         nbd_dbg_dir = dbg_dir;
1690
1691         return 0;
1692 }
1693
1694 static void nbd_dbg_close(void)
1695 {
1696         debugfs_remove_recursive(nbd_dbg_dir);
1697 }
1698
1699 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1700
1701 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1702 {
1703         return 0;
1704 }
1705
1706 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1707 {
1708 }
1709
1710 static int nbd_dbg_init(void)
1711 {
1712         return 0;
1713 }
1714
1715 static void nbd_dbg_close(void)
1716 {
1717 }
1718
1719 #endif
1720
1721 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1722                             unsigned int hctx_idx, unsigned int numa_node)
1723 {
1724         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1725         cmd->nbd = set->driver_data;
1726         cmd->flags = 0;
1727         mutex_init(&cmd->lock);
1728         return 0;
1729 }
1730
1731 static const struct blk_mq_ops nbd_mq_ops = {
1732         .queue_rq       = nbd_queue_rq,
1733         .complete       = nbd_complete_rq,
1734         .init_request   = nbd_init_request,
1735         .timeout        = nbd_xmit_timeout,
1736 };
1737
1738 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1739 {
1740         struct nbd_device *nbd;
1741         struct gendisk *disk;
1742         int err = -ENOMEM;
1743
1744         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1745         if (!nbd)
1746                 goto out;
1747
1748         nbd->tag_set.ops = &nbd_mq_ops;
1749         nbd->tag_set.nr_hw_queues = 1;
1750         nbd->tag_set.queue_depth = 128;
1751         nbd->tag_set.numa_node = NUMA_NO_NODE;
1752         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1753         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1754                 BLK_MQ_F_BLOCKING;
1755         nbd->tag_set.driver_data = nbd;
1756         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1757         nbd->backend = NULL;
1758
1759         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1760         if (err)
1761                 goto out_free_nbd;
1762
1763         mutex_lock(&nbd_index_mutex);
1764         if (index >= 0) {
1765                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1766                                 GFP_KERNEL);
1767                 if (err == -ENOSPC)
1768                         err = -EEXIST;
1769         } else {
1770                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1771                 if (err >= 0)
1772                         index = err;
1773         }
1774         nbd->index = index;
1775         mutex_unlock(&nbd_index_mutex);
1776         if (err < 0)
1777                 goto out_free_tags;
1778
1779         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1780         if (IS_ERR(disk)) {
1781                 err = PTR_ERR(disk);
1782                 goto out_free_idr;
1783         }
1784         nbd->disk = disk;
1785
1786         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1787                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1788                                           WQ_UNBOUND, 0, nbd->index);
1789         if (!nbd->recv_workq) {
1790                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1791                 err = -ENOMEM;
1792                 goto out_err_disk;
1793         }
1794
1795         /*
1796          * Tell the block layer that we are not a rotational device
1797          */
1798         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1799         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1800         disk->queue->limits.discard_granularity = 0;
1801         blk_queue_max_discard_sectors(disk->queue, 0);
1802         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1803         blk_queue_max_segments(disk->queue, USHRT_MAX);
1804         blk_queue_max_hw_sectors(disk->queue, 65536);
1805         disk->queue->limits.max_sectors = 256;
1806
1807         mutex_init(&nbd->config_lock);
1808         refcount_set(&nbd->config_refs, 0);
1809         /*
1810          * Start out with a zero references to keep other threads from using
1811          * this device until it is fully initialized.
1812          */
1813         refcount_set(&nbd->refs, 0);
1814         INIT_LIST_HEAD(&nbd->list);
1815         disk->major = NBD_MAJOR;
1816         disk->first_minor = index << part_shift;
1817         disk->minors = 1 << part_shift;
1818         disk->fops = &nbd_fops;
1819         disk->private_data = nbd;
1820         sprintf(disk->disk_name, "nbd%d", index);
1821         err = add_disk(disk);
1822         if (err)
1823                 goto out_free_work;
1824
1825         /*
1826          * Now publish the device.
1827          */
1828         refcount_set(&nbd->refs, refs);
1829         nbd_total_devices++;
1830         return nbd;
1831
1832 out_free_work:
1833         destroy_workqueue(nbd->recv_workq);
1834 out_err_disk:
1835         put_disk(disk);
1836 out_free_idr:
1837         mutex_lock(&nbd_index_mutex);
1838         idr_remove(&nbd_index_idr, index);
1839         mutex_unlock(&nbd_index_mutex);
1840 out_free_tags:
1841         blk_mq_free_tag_set(&nbd->tag_set);
1842 out_free_nbd:
1843         kfree(nbd);
1844 out:
1845         return ERR_PTR(err);
1846 }
1847
1848 static struct nbd_device *nbd_find_get_unused(void)
1849 {
1850         struct nbd_device *nbd;
1851         int id;
1852
1853         lockdep_assert_held(&nbd_index_mutex);
1854
1855         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1856                 if (refcount_read(&nbd->config_refs) ||
1857                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1858                         continue;
1859                 if (refcount_inc_not_zero(&nbd->refs))
1860                         return nbd;
1861         }
1862
1863         return NULL;
1864 }
1865
1866 /* Netlink interface. */
1867 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1868         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1869         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1870         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1871         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1872         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1873         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1874         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1875         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1876         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1877         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1878 };
1879
1880 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1881         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1882 };
1883
1884 /* We don't use this right now since we don't parse the incoming list, but we
1885  * still want it here so userspace knows what to expect.
1886  */
1887 static const struct nla_policy __attribute__((unused))
1888 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1889         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1890         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1891 };
1892
1893 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1894 {
1895         struct nbd_config *config = nbd->config;
1896         u64 bsize = nbd_blksize(config);
1897         u64 bytes = config->bytesize;
1898
1899         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1900                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1901
1902         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1903                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1904
1905         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1906                 return nbd_set_size(nbd, bytes, bsize);
1907         return 0;
1908 }
1909
1910 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1911 {
1912         struct nbd_device *nbd;
1913         struct nbd_config *config;
1914         int index = -1;
1915         int ret;
1916         bool put_dev = false;
1917
1918         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1919                 return -EPERM;
1920
1921         if (info->attrs[NBD_ATTR_INDEX]) {
1922                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1923
1924                 /*
1925                  * Too big first_minor can cause duplicate creation of
1926                  * sysfs files/links, since index << part_shift might overflow, or
1927                  * MKDEV() expect that the max bits of first_minor is 20.
1928                  */
1929                 if (index < 0 || index > MINORMASK >> part_shift) {
1930                         pr_err("illegal input index %d\n", index);
1931                         return -EINVAL;
1932                 }
1933         }
1934         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1935                 pr_err("must specify at least one socket\n");
1936                 return -EINVAL;
1937         }
1938         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1939                 pr_err("must specify a size in bytes for the device\n");
1940                 return -EINVAL;
1941         }
1942 again:
1943         mutex_lock(&nbd_index_mutex);
1944         if (index == -1) {
1945                 nbd = nbd_find_get_unused();
1946         } else {
1947                 nbd = idr_find(&nbd_index_idr, index);
1948                 if (nbd) {
1949                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1950                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
1951                             !refcount_inc_not_zero(&nbd->refs)) {
1952                                 mutex_unlock(&nbd_index_mutex);
1953                                 pr_err("nbd: device at index %d is going down\n",
1954                                         index);
1955                                 return -EINVAL;
1956                         }
1957                 }
1958         }
1959         mutex_unlock(&nbd_index_mutex);
1960
1961         if (!nbd) {
1962                 nbd = nbd_dev_add(index, 2);
1963                 if (IS_ERR(nbd)) {
1964                         pr_err("nbd: failed to add new device\n");
1965                         return PTR_ERR(nbd);
1966                 }
1967         }
1968
1969         mutex_lock(&nbd->config_lock);
1970         if (refcount_read(&nbd->config_refs)) {
1971                 mutex_unlock(&nbd->config_lock);
1972                 nbd_put(nbd);
1973                 if (index == -1)
1974                         goto again;
1975                 pr_err("nbd%d already in use\n", index);
1976                 return -EBUSY;
1977         }
1978         if (WARN_ON(nbd->config)) {
1979                 mutex_unlock(&nbd->config_lock);
1980                 nbd_put(nbd);
1981                 return -EINVAL;
1982         }
1983         config = nbd_alloc_config();
1984         if (IS_ERR(config)) {
1985                 mutex_unlock(&nbd->config_lock);
1986                 nbd_put(nbd);
1987                 pr_err("couldn't allocate config\n");
1988                 return PTR_ERR(config);
1989         }
1990         nbd->config = config;
1991         refcount_set(&nbd->config_refs, 1);
1992         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1993
1994         ret = nbd_genl_size_set(info, nbd);
1995         if (ret)
1996                 goto out;
1997
1998         if (info->attrs[NBD_ATTR_TIMEOUT])
1999                 nbd_set_cmd_timeout(nbd,
2000                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2001         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2002                 config->dead_conn_timeout =
2003                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2004                 config->dead_conn_timeout *= HZ;
2005         }
2006         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
2007                 config->flags =
2008                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
2009         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2010                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2011                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2012                         /*
2013                          * We have 1 ref to keep the device around, and then 1
2014                          * ref for our current operation here, which will be
2015                          * inherited by the config.  If we already have
2016                          * DESTROY_ON_DISCONNECT set then we know we don't have
2017                          * that extra ref already held so we don't need the
2018                          * put_dev.
2019                          */
2020                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2021                                               &nbd->flags))
2022                                 put_dev = true;
2023                 } else {
2024                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2025                                                &nbd->flags))
2026                                 refcount_inc(&nbd->refs);
2027                 }
2028                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2029                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2030                                 &config->runtime_flags);
2031                 }
2032         }
2033
2034         if (info->attrs[NBD_ATTR_SOCKETS]) {
2035                 struct nlattr *attr;
2036                 int rem, fd;
2037
2038                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2039                                     rem) {
2040                         struct nlattr *socks[NBD_SOCK_MAX+1];
2041
2042                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2043                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2044                                 ret = -EINVAL;
2045                                 goto out;
2046                         }
2047                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2048                                                           attr,
2049                                                           nbd_sock_policy,
2050                                                           info->extack);
2051                         if (ret != 0) {
2052                                 pr_err("error processing sock list\n");
2053                                 ret = -EINVAL;
2054                                 goto out;
2055                         }
2056                         if (!socks[NBD_SOCK_FD])
2057                                 continue;
2058                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2059                         ret = nbd_add_socket(nbd, fd, true);
2060                         if (ret)
2061                                 goto out;
2062                 }
2063         }
2064         ret = nbd_start_device(nbd);
2065         if (ret)
2066                 goto out;
2067         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2068                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2069                                           GFP_KERNEL);
2070                 if (!nbd->backend) {
2071                         ret = -ENOMEM;
2072                         goto out;
2073                 }
2074         }
2075         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2076         if (ret) {
2077                 dev_err(disk_to_dev(nbd->disk),
2078                         "device_create_file failed for backend!\n");
2079                 goto out;
2080         }
2081         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2082 out:
2083         mutex_unlock(&nbd->config_lock);
2084         if (!ret) {
2085                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2086                 refcount_inc(&nbd->config_refs);
2087                 nbd_connect_reply(info, nbd->index);
2088         }
2089         nbd_config_put(nbd);
2090         if (put_dev)
2091                 nbd_put(nbd);
2092         return ret;
2093 }
2094
2095 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2096 {
2097         mutex_lock(&nbd->config_lock);
2098         nbd_disconnect(nbd);
2099         sock_shutdown(nbd);
2100         wake_up(&nbd->config->conn_wait);
2101         /*
2102          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2103          * to cancel the inflight I/Os.
2104          */
2105         flush_workqueue(nbd->recv_workq);
2106         nbd_clear_que(nbd);
2107         nbd->task_setup = NULL;
2108         mutex_unlock(&nbd->config_lock);
2109
2110         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2111                                &nbd->config->runtime_flags))
2112                 nbd_config_put(nbd);
2113 }
2114
2115 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2116 {
2117         struct nbd_device *nbd;
2118         int index;
2119
2120         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2121                 return -EPERM;
2122
2123         if (!info->attrs[NBD_ATTR_INDEX]) {
2124                 pr_err("must specify an index to disconnect\n");
2125                 return -EINVAL;
2126         }
2127         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2128         mutex_lock(&nbd_index_mutex);
2129         nbd = idr_find(&nbd_index_idr, index);
2130         if (!nbd) {
2131                 mutex_unlock(&nbd_index_mutex);
2132                 pr_err("couldn't find device at index %d\n", index);
2133                 return -EINVAL;
2134         }
2135         if (!refcount_inc_not_zero(&nbd->refs)) {
2136                 mutex_unlock(&nbd_index_mutex);
2137                 pr_err("device at index %d is going down\n", index);
2138                 return -EINVAL;
2139         }
2140         mutex_unlock(&nbd_index_mutex);
2141         if (!refcount_inc_not_zero(&nbd->config_refs))
2142                 goto put_nbd;
2143         nbd_disconnect_and_put(nbd);
2144         nbd_config_put(nbd);
2145 put_nbd:
2146         nbd_put(nbd);
2147         return 0;
2148 }
2149
2150 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2151 {
2152         struct nbd_device *nbd = NULL;
2153         struct nbd_config *config;
2154         int index;
2155         int ret = 0;
2156         bool put_dev = false;
2157
2158         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2159                 return -EPERM;
2160
2161         if (!info->attrs[NBD_ATTR_INDEX]) {
2162                 pr_err("must specify a device to reconfigure\n");
2163                 return -EINVAL;
2164         }
2165         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2166         mutex_lock(&nbd_index_mutex);
2167         nbd = idr_find(&nbd_index_idr, index);
2168         if (!nbd) {
2169                 mutex_unlock(&nbd_index_mutex);
2170                 pr_err("couldn't find a device at index %d\n", index);
2171                 return -EINVAL;
2172         }
2173         if (nbd->backend) {
2174                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2175                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2176                                        nbd->backend)) {
2177                                 mutex_unlock(&nbd_index_mutex);
2178                                 dev_err(nbd_to_dev(nbd),
2179                                         "backend image doesn't match with %s\n",
2180                                         nbd->backend);
2181                                 return -EINVAL;
2182                         }
2183                 } else {
2184                         mutex_unlock(&nbd_index_mutex);
2185                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2186                         return -EINVAL;
2187                 }
2188         }
2189         if (!refcount_inc_not_zero(&nbd->refs)) {
2190                 mutex_unlock(&nbd_index_mutex);
2191                 pr_err("device at index %d is going down\n", index);
2192                 return -EINVAL;
2193         }
2194         mutex_unlock(&nbd_index_mutex);
2195
2196         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2197                 dev_err(nbd_to_dev(nbd),
2198                         "not configured, cannot reconfigure\n");
2199                 nbd_put(nbd);
2200                 return -EINVAL;
2201         }
2202
2203         mutex_lock(&nbd->config_lock);
2204         config = nbd->config;
2205         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2206             !nbd->pid) {
2207                 dev_err(nbd_to_dev(nbd),
2208                         "not configured, cannot reconfigure\n");
2209                 ret = -EINVAL;
2210                 goto out;
2211         }
2212
2213         ret = nbd_genl_size_set(info, nbd);
2214         if (ret)
2215                 goto out;
2216
2217         if (info->attrs[NBD_ATTR_TIMEOUT])
2218                 nbd_set_cmd_timeout(nbd,
2219                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2220         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2221                 config->dead_conn_timeout =
2222                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2223                 config->dead_conn_timeout *= HZ;
2224         }
2225         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2226                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2227                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2228                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2229                                               &nbd->flags))
2230                                 put_dev = true;
2231                 } else {
2232                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2233                                                &nbd->flags))
2234                                 refcount_inc(&nbd->refs);
2235                 }
2236
2237                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2238                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2239                                         &config->runtime_flags);
2240                 } else {
2241                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2242                                         &config->runtime_flags);
2243                 }
2244         }
2245
2246         if (info->attrs[NBD_ATTR_SOCKETS]) {
2247                 struct nlattr *attr;
2248                 int rem, fd;
2249
2250                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2251                                     rem) {
2252                         struct nlattr *socks[NBD_SOCK_MAX+1];
2253
2254                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2255                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2256                                 ret = -EINVAL;
2257                                 goto out;
2258                         }
2259                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2260                                                           attr,
2261                                                           nbd_sock_policy,
2262                                                           info->extack);
2263                         if (ret != 0) {
2264                                 pr_err("error processing sock list\n");
2265                                 ret = -EINVAL;
2266                                 goto out;
2267                         }
2268                         if (!socks[NBD_SOCK_FD])
2269                                 continue;
2270                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2271                         ret = nbd_reconnect_socket(nbd, fd);
2272                         if (ret) {
2273                                 if (ret == -ENOSPC)
2274                                         ret = 0;
2275                                 goto out;
2276                         }
2277                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2278                 }
2279         }
2280 out:
2281         mutex_unlock(&nbd->config_lock);
2282         nbd_config_put(nbd);
2283         nbd_put(nbd);
2284         if (put_dev)
2285                 nbd_put(nbd);
2286         return ret;
2287 }
2288
2289 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2290         {
2291                 .cmd    = NBD_CMD_CONNECT,
2292                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2293                 .doit   = nbd_genl_connect,
2294         },
2295         {
2296                 .cmd    = NBD_CMD_DISCONNECT,
2297                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2298                 .doit   = nbd_genl_disconnect,
2299         },
2300         {
2301                 .cmd    = NBD_CMD_RECONFIGURE,
2302                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2303                 .doit   = nbd_genl_reconfigure,
2304         },
2305         {
2306                 .cmd    = NBD_CMD_STATUS,
2307                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2308                 .doit   = nbd_genl_status,
2309         },
2310 };
2311
2312 static const struct genl_multicast_group nbd_mcast_grps[] = {
2313         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2314 };
2315
2316 static struct genl_family nbd_genl_family __ro_after_init = {
2317         .hdrsize        = 0,
2318         .name           = NBD_GENL_FAMILY_NAME,
2319         .version        = NBD_GENL_VERSION,
2320         .module         = THIS_MODULE,
2321         .small_ops      = nbd_connect_genl_ops,
2322         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2323         .maxattr        = NBD_ATTR_MAX,
2324         .policy = nbd_attr_policy,
2325         .mcgrps         = nbd_mcast_grps,
2326         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2327 };
2328
2329 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2330 {
2331         struct nlattr *dev_opt;
2332         u8 connected = 0;
2333         int ret;
2334
2335         /* This is a little racey, but for status it's ok.  The
2336          * reason we don't take a ref here is because we can't
2337          * take a ref in the index == -1 case as we would need
2338          * to put under the nbd_index_mutex, which could
2339          * deadlock if we are configured to remove ourselves
2340          * once we're disconnected.
2341          */
2342         if (refcount_read(&nbd->config_refs))
2343                 connected = 1;
2344         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2345         if (!dev_opt)
2346                 return -EMSGSIZE;
2347         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2348         if (ret)
2349                 return -EMSGSIZE;
2350         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2351                          connected);
2352         if (ret)
2353                 return -EMSGSIZE;
2354         nla_nest_end(reply, dev_opt);
2355         return 0;
2356 }
2357
2358 static int status_cb(int id, void *ptr, void *data)
2359 {
2360         struct nbd_device *nbd = ptr;
2361         return populate_nbd_status(nbd, (struct sk_buff *)data);
2362 }
2363
2364 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2365 {
2366         struct nlattr *dev_list;
2367         struct sk_buff *reply;
2368         void *reply_head;
2369         size_t msg_size;
2370         int index = -1;
2371         int ret = -ENOMEM;
2372
2373         if (info->attrs[NBD_ATTR_INDEX])
2374                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2375
2376         mutex_lock(&nbd_index_mutex);
2377
2378         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2379                                   nla_attr_size(sizeof(u8)));
2380         msg_size *= (index == -1) ? nbd_total_devices : 1;
2381
2382         reply = genlmsg_new(msg_size, GFP_KERNEL);
2383         if (!reply)
2384                 goto out;
2385         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2386                                        NBD_CMD_STATUS);
2387         if (!reply_head) {
2388                 nlmsg_free(reply);
2389                 goto out;
2390         }
2391
2392         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2393         if (index == -1) {
2394                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2395                 if (ret) {
2396                         nlmsg_free(reply);
2397                         goto out;
2398                 }
2399         } else {
2400                 struct nbd_device *nbd;
2401                 nbd = idr_find(&nbd_index_idr, index);
2402                 if (nbd) {
2403                         ret = populate_nbd_status(nbd, reply);
2404                         if (ret) {
2405                                 nlmsg_free(reply);
2406                                 goto out;
2407                         }
2408                 }
2409         }
2410         nla_nest_end(reply, dev_list);
2411         genlmsg_end(reply, reply_head);
2412         ret = genlmsg_reply(reply, info);
2413 out:
2414         mutex_unlock(&nbd_index_mutex);
2415         return ret;
2416 }
2417
2418 static void nbd_connect_reply(struct genl_info *info, int index)
2419 {
2420         struct sk_buff *skb;
2421         void *msg_head;
2422         int ret;
2423
2424         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2425         if (!skb)
2426                 return;
2427         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2428                                      NBD_CMD_CONNECT);
2429         if (!msg_head) {
2430                 nlmsg_free(skb);
2431                 return;
2432         }
2433         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2434         if (ret) {
2435                 nlmsg_free(skb);
2436                 return;
2437         }
2438         genlmsg_end(skb, msg_head);
2439         genlmsg_reply(skb, info);
2440 }
2441
2442 static void nbd_mcast_index(int index)
2443 {
2444         struct sk_buff *skb;
2445         void *msg_head;
2446         int ret;
2447
2448         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2449         if (!skb)
2450                 return;
2451         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2452                                      NBD_CMD_LINK_DEAD);
2453         if (!msg_head) {
2454                 nlmsg_free(skb);
2455                 return;
2456         }
2457         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2458         if (ret) {
2459                 nlmsg_free(skb);
2460                 return;
2461         }
2462         genlmsg_end(skb, msg_head);
2463         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2464 }
2465
2466 static void nbd_dead_link_work(struct work_struct *work)
2467 {
2468         struct link_dead_args *args = container_of(work, struct link_dead_args,
2469                                                    work);
2470         nbd_mcast_index(args->index);
2471         kfree(args);
2472 }
2473
2474 static int __init nbd_init(void)
2475 {
2476         int i;
2477
2478         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2479
2480         if (max_part < 0) {
2481                 pr_err("max_part must be >= 0\n");
2482                 return -EINVAL;
2483         }
2484
2485         part_shift = 0;
2486         if (max_part > 0) {
2487                 part_shift = fls(max_part);
2488
2489                 /*
2490                  * Adjust max_part according to part_shift as it is exported
2491                  * to user space so that user can know the max number of
2492                  * partition kernel should be able to manage.
2493                  *
2494                  * Note that -1 is required because partition 0 is reserved
2495                  * for the whole disk.
2496                  */
2497                 max_part = (1UL << part_shift) - 1;
2498         }
2499
2500         if ((1UL << part_shift) > DISK_MAX_PARTS)
2501                 return -EINVAL;
2502
2503         if (nbds_max > 1UL << (MINORBITS - part_shift))
2504                 return -EINVAL;
2505
2506         if (register_blkdev(NBD_MAJOR, "nbd"))
2507                 return -EIO;
2508
2509         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2510         if (!nbd_del_wq) {
2511                 unregister_blkdev(NBD_MAJOR, "nbd");
2512                 return -ENOMEM;
2513         }
2514
2515         if (genl_register_family(&nbd_genl_family)) {
2516                 destroy_workqueue(nbd_del_wq);
2517                 unregister_blkdev(NBD_MAJOR, "nbd");
2518                 return -EINVAL;
2519         }
2520         nbd_dbg_init();
2521
2522         for (i = 0; i < nbds_max; i++)
2523                 nbd_dev_add(i, 1);
2524         return 0;
2525 }
2526
2527 static int nbd_exit_cb(int id, void *ptr, void *data)
2528 {
2529         struct list_head *list = (struct list_head *)data;
2530         struct nbd_device *nbd = ptr;
2531
2532         /* Skip nbd that is being removed asynchronously */
2533         if (refcount_read(&nbd->refs))
2534                 list_add_tail(&nbd->list, list);
2535
2536         return 0;
2537 }
2538
2539 static void __exit nbd_cleanup(void)
2540 {
2541         struct nbd_device *nbd;
2542         LIST_HEAD(del_list);
2543
2544         /*
2545          * Unregister netlink interface prior to waiting
2546          * for the completion of netlink commands.
2547          */
2548         genl_unregister_family(&nbd_genl_family);
2549
2550         nbd_dbg_close();
2551
2552         mutex_lock(&nbd_index_mutex);
2553         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2554         mutex_unlock(&nbd_index_mutex);
2555
2556         while (!list_empty(&del_list)) {
2557                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2558                 list_del_init(&nbd->list);
2559                 if (refcount_read(&nbd->config_refs))
2560                         pr_err("possibly leaking nbd_config (ref %d)\n",
2561                                         refcount_read(&nbd->config_refs));
2562                 if (refcount_read(&nbd->refs) != 1)
2563                         pr_err("possibly leaking a device\n");
2564                 nbd_put(nbd);
2565         }
2566
2567         /* Also wait for nbd_dev_remove_work() completes */
2568         destroy_workqueue(nbd_del_wq);
2569
2570         idr_destroy(&nbd_index_idr);
2571         unregister_blkdev(NBD_MAJOR, "nbd");
2572 }
2573
2574 module_init(nbd_init);
2575 module_exit(nbd_cleanup);
2576
2577 MODULE_DESCRIPTION("Network Block Device");
2578 MODULE_LICENSE("GPL");
2579
2580 module_param(nbds_max, int, 0444);
2581 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2582 module_param(max_part, int, 0444);
2583 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");