drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   9  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #define pr_fmt(fmt) "nbd: " fmt
  15
  16 #include <linux/major.h>
  17
  18 #include <linux/blkdev.h>
  19 #include <linux/module.h>
  20 #include <linux/init.h>
  21 #include <linux/sched.h>
  22 #include <linux/sched/mm.h>
  23 #include <linux/fs.h>
  24 #include <linux/bio.h>
  25 #include <linux/stat.h>
  26 #include <linux/errno.h>
  27 #include <linux/file.h>
  28 #include <linux/ioctl.h>
  29 #include <linux/mutex.h>
  30 #include <linux/compiler.h>
  31 #include <linux/completion.h>
  32 #include <linux/err.h>
  33 #include <linux/kernel.h>
  34 #include <linux/slab.h>
  35 #include <net/sock.h>
  36 #include <linux/net.h>
  37 #include <linux/kthread.h>
  38 #include <linux/types.h>
  39 #include <linux/debugfs.h>
  40 #include <linux/blk-mq.h>
  41
  42 #include <linux/uaccess.h>
  43 #include <asm/types.h>
  44
  45 #include <linux/nbd.h>
  46 #include <linux/nbd-netlink.h>
  47 #include <net/genetlink.h>
  48
  49 #define CREATE_TRACE_POINTS
  50 #include <trace/events/nbd.h>
  51
  52 static DEFINE_IDR(nbd_index_idr);
  53 static DEFINE_MUTEX(nbd_index_mutex);
  54 static struct workqueue_struct *nbd_del_wq;
  55 static int nbd_total_devices = 0;
  56
  57 struct nbd_sock {
  58         struct socket *sock;
  59         struct mutex tx_lock;
  60         struct request *pending;
  61         int sent;
  62         bool dead;
  63         int fallback_index;
  64         int cookie;
  65 };
  66
  67 struct recv_thread_args {
  68         struct work_struct work;
  69         struct nbd_device *nbd;
  70         int index;
  71 };
  72
  73 struct link_dead_args {
  74         struct work_struct work;
  75         int index;
  76 };
  77
  78 #define NBD_RT_TIMEDOUT                 0
  79 #define NBD_RT_DISCONNECT_REQUESTED     1
  80 #define NBD_RT_DISCONNECTED             2
  81 #define NBD_RT_HAS_PID_FILE             3
  82 #define NBD_RT_HAS_CONFIG_REF           4
  83 #define NBD_RT_BOUND                    5
  84 #define NBD_RT_DISCONNECT_ON_CLOSE      6
  85 #define NBD_RT_HAS_BACKEND_FILE         7
  86
  87 #define NBD_DESTROY_ON_DISCONNECT       0
  88 #define NBD_DISCONNECT_REQUESTED        1
  89
  90 struct nbd_config {
  91         u32 flags;
  92         unsigned long runtime_flags;
  93         u64 dead_conn_timeout;
  94
  95         struct nbd_sock **socks;
  96         int num_connections;
  97         atomic_t live_connections;
  98         wait_queue_head_t conn_wait;
  99
 100         atomic_t recv_threads;
 101         wait_queue_head_t recv_wq;
 102         unsigned int blksize_bits;
 103         loff_t bytesize;
 104 #if IS_ENABLED(CONFIG_DEBUG_FS)
 105         struct dentry *dbg_dir;
 106 #endif
 107 };
 108
 109 static inline unsigned int nbd_blksize(struct nbd_config *config)
 110 {
 111         return 1u << config->blksize_bits;
 112 }
 113
 114 struct nbd_device {
 115         struct blk_mq_tag_set tag_set;
 116
 117         int index;
 118         refcount_t config_refs;
 119         refcount_t refs;
 120         struct nbd_config *config;
 121         struct mutex config_lock;
 122         struct gendisk *disk;
 123         struct workqueue_struct *recv_workq;
 124         struct work_struct remove_work;
 125
 126         struct list_head list;
 127         struct task_struct *task_setup;
 128
 129         unsigned long flags;
 130         pid_t pid; /* pid of nbd-client, if attached */
 131
 132         char *backend;
 133 };
 134
 135 #define NBD_CMD_REQUEUED        1
 136 /*
 137  * This flag will be set if nbd_queue_rq() succeed, and will be checked and
 138  * cleared in completion. Both setting and clearing of the flag are protected
 139  * by cmd->lock.
 140  */
 141 #define NBD_CMD_INFLIGHT        2
 142
 143 struct nbd_cmd {
 144         struct nbd_device *nbd;
 145         struct mutex lock;
 146         int index;
 147         int cookie;
 148         int retries;
 149         blk_status_t status;
 150         unsigned long flags;
 151         u32 cmd_cookie;
 152 };
 153
 154 #if IS_ENABLED(CONFIG_DEBUG_FS)
 155 static struct dentry *nbd_dbg_dir;
 156 #endif
 157
 158 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 159
 160 #define NBD_DEF_BLKSIZE_BITS 10
 161
 162 static unsigned int nbds_max = 16;
 163 static int max_part = 16;
 164 static int part_shift;
 165
 166 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 167 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 168 static void nbd_config_put(struct nbd_device *nbd);
 169 static void nbd_connect_reply(struct genl_info *info, int index);
 170 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 171 static void nbd_dead_link_work(struct work_struct *work);
 172 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 173
 174 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 175 {
 176         return disk_to_dev(nbd->disk);
 177 }
 178
 179 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 180 {
 181         struct request *req = blk_mq_rq_from_pdu(cmd);
 182
 183         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 184                 blk_mq_requeue_request(req, true);
 185 }
 186
 187 #define NBD_COOKIE_BITS 32
 188
 189 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 190 {
 191         struct request *req = blk_mq_rq_from_pdu(cmd);
 192         u32 tag = blk_mq_unique_tag(req);
 193         u64 cookie = cmd->cmd_cookie;
 194
 195         return (cookie << NBD_COOKIE_BITS) | tag;
 196 }
 197
 198 static u32 nbd_handle_to_tag(u64 handle)
 199 {
 200         return (u32)handle;
 201 }
 202
 203 static u32 nbd_handle_to_cookie(u64 handle)
 204 {
 205         return (u32)(handle >> NBD_COOKIE_BITS);
 206 }
 207
 208 static const char *nbdcmd_to_ascii(int cmd)
 209 {
 210         switch (cmd) {
 211         case  NBD_CMD_READ: return "read";
 212         case NBD_CMD_WRITE: return "write";
 213         case  NBD_CMD_DISC: return "disconnect";
 214         case NBD_CMD_FLUSH: return "flush";
 215         case  NBD_CMD_TRIM: return "trim/discard";
 216         }
 217         return "invalid";
 218 }
 219
 220 static ssize_t pid_show(struct device *dev,
 221                         struct device_attribute *attr, char *buf)
 222 {
 223         struct gendisk *disk = dev_to_disk(dev);
 224         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 225
 226         return sprintf(buf, "%d\n", nbd->pid);
 227 }
 228
 229 static const struct device_attribute pid_attr = {
 230         .attr = { .name = "pid", .mode = 0444},
 231         .show = pid_show,
 232 };
 233
 234 static ssize_t backend_show(struct device *dev,
 235                 struct device_attribute *attr, char *buf)
 236 {
 237         struct gendisk *disk = dev_to_disk(dev);
 238         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 239
 240         return sprintf(buf, "%s\n", nbd->backend ?: "");
 241 }
 242
 243 static const struct device_attribute backend_attr = {
 244         .attr = { .name = "backend", .mode = 0444},
 245         .show = backend_show,
 246 };
 247
 248 static void nbd_dev_remove(struct nbd_device *nbd)
 249 {
 250         struct gendisk *disk = nbd->disk;
 251
 252         del_gendisk(disk);
 253         put_disk(disk);
 254         blk_mq_free_tag_set(&nbd->tag_set);
 255
 256         /*
 257          * Remove from idr after del_gendisk() completes, so if the same ID is
 258          * reused, the following add_disk() will succeed.
 259          */
 260         mutex_lock(&nbd_index_mutex);
 261         idr_remove(&nbd_index_idr, nbd->index);
 262         mutex_unlock(&nbd_index_mutex);
 263         destroy_workqueue(nbd->recv_workq);
 264         kfree(nbd);
 265 }
 266
 267 static void nbd_dev_remove_work(struct work_struct *work)
 268 {
 269         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 270 }
 271
 272 static void nbd_put(struct nbd_device *nbd)
 273 {
 274         if (!refcount_dec_and_test(&nbd->refs))
 275                 return;
 276
 277         /* Call del_gendisk() asynchrounously to prevent deadlock */
 278         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 279                 queue_work(nbd_del_wq, &nbd->remove_work);
 280         else
 281                 nbd_dev_remove(nbd);
 282 }
 283
 284 static int nbd_disconnected(struct nbd_config *config)
 285 {
 286         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 287                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 288 }
 289
 290 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 291                                 int notify)
 292 {
 293         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 294                 struct link_dead_args *args;
 295                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 296                 if (args) {
 297                         INIT_WORK(&args->work, nbd_dead_link_work);
 298                         args->index = nbd->index;
 299                         queue_work(system_wq, &args->work);
 300                 }
 301         }
 302         if (!nsock->dead) {
 303                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 304                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 305                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 306                                                &nbd->config->runtime_flags)) {
 307                                 set_bit(NBD_RT_DISCONNECTED,
 308                                         &nbd->config->runtime_flags);
 309                                 dev_info(nbd_to_dev(nbd),
 310                                         "Disconnected due to user request.\n");
 311                         }
 312                 }
 313         }
 314         nsock->dead = true;
 315         nsock->pending = NULL;
 316         nsock->sent = 0;
 317 }
 318
 319 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 320                 loff_t blksize)
 321 {
 322         if (!blksize)
 323                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 324
 325         if (blk_validate_block_size(blksize))
 326                 return -EINVAL;
 327
 328         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                 nbd_clear_que(nbd);
1417         }
1418
1419         flush_workqueue(nbd->recv_workq);
1420         mutex_lock(&nbd->config_lock);
1421         nbd_bdev_reset(nbd);
1422         /* user requested, ignore socket errors */
1423         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1424                 ret = 0;
1425         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1426                 ret = -ETIMEDOUT;
1427         return ret;
1428 }
1429
1430 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1431                                  struct block_device *bdev)
1432 {
1433         nbd_clear_sock(nbd);
1434         __invalidate_device(bdev, true);
1435         nbd_bdev_reset(nbd);
1436         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1437                                &nbd->config->runtime_flags))
1438                 nbd_config_put(nbd);
1439 }
1440
1441 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1442 {
1443         nbd->tag_set.timeout = timeout * HZ;
1444         if (timeout)
1445                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1446         else
1447                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1448 }
1449
1450 /* Must be called with config_lock held */
1451 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1452                        unsigned int cmd, unsigned long arg)
1453 {
1454         struct nbd_config *config = nbd->config;
1455         loff_t bytesize;
1456
1457         switch (cmd) {
1458         case NBD_DISCONNECT:
1459                 return nbd_disconnect(nbd);
1460         case NBD_CLEAR_SOCK:
1461                 nbd_clear_sock_ioctl(nbd, bdev);
1462                 return 0;
1463         case NBD_SET_SOCK:
1464                 return nbd_add_socket(nbd, arg, false);
1465         case NBD_SET_BLKSIZE:
1466                 return nbd_set_size(nbd, config->bytesize, arg);
1467         case NBD_SET_SIZE:
1468                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1469         case NBD_SET_SIZE_BLOCKS:
1470                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1471                         return -EINVAL;
1472                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1473         case NBD_SET_TIMEOUT:
1474                 nbd_set_cmd_timeout(nbd, arg);
1475                 return 0;
1476
1477         case NBD_SET_FLAGS:
1478                 config->flags = arg;
1479                 return 0;
1480         case NBD_DO_IT:
1481                 return nbd_start_device_ioctl(nbd);
1482         case NBD_CLEAR_QUE:
1483                 /*
1484                  * This is for compatibility only.  The queue is always cleared
1485                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1486                  */
1487                 return 0;
1488         case NBD_PRINT_DEBUG:
1489                 /*
1490                  * For compatibility only, we no longer keep a list of
1491                  * outstanding requests.
1492                  */
1493                 return 0;
1494         }
1495         return -ENOTTY;
1496 }
1497
1498 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1499                      unsigned int cmd, unsigned long arg)
1500 {
1501         struct nbd_device *nbd = bdev->bd_disk->private_data;
1502         struct nbd_config *config = nbd->config;
1503         int error = -EINVAL;
1504
1505         if (!capable(CAP_SYS_ADMIN))
1506                 return -EPERM;
1507
1508         /* The block layer will pass back some non-nbd ioctls in case we have
1509          * special handling for them, but we don't so just return an error.
1510          */
1511         if (_IOC_TYPE(cmd) != 0xab)
1512                 return -EINVAL;
1513
1514         mutex_lock(&nbd->config_lock);
1515
1516         /* Don't allow ioctl operations on a nbd device that was created with
1517          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1518          */
1519         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1520             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1521                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1522         else
1523                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1524         mutex_unlock(&nbd->config_lock);
1525         return error;
1526 }
1527
1528 static struct nbd_config *nbd_alloc_config(void)
1529 {
1530         struct nbd_config *config;
1531
1532         if (!try_module_get(THIS_MODULE))
1533                 return ERR_PTR(-ENODEV);
1534
1535         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1536         if (!config) {
1537                 module_put(THIS_MODULE);
1538                 return ERR_PTR(-ENOMEM);
1539         }
1540
1541         atomic_set(&config->recv_threads, 0);
1542         init_waitqueue_head(&config->recv_wq);
1543         init_waitqueue_head(&config->conn_wait);
1544         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1545         atomic_set(&config->live_connections, 0);
1546         return config;
1547 }
1548
1549 static int nbd_open(struct block_device *bdev, fmode_t mode)
1550 {
1551         struct nbd_device *nbd;
1552         int ret = 0;
1553
1554         mutex_lock(&nbd_index_mutex);
1555         nbd = bdev->bd_disk->private_data;
1556         if (!nbd) {
1557                 ret = -ENXIO;
1558                 goto out;
1559         }
1560         if (!refcount_inc_not_zero(&nbd->refs)) {
1561                 ret = -ENXIO;
1562                 goto out;
1563         }
1564         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1565                 struct nbd_config *config;
1566
1567                 mutex_lock(&nbd->config_lock);
1568                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1569                         mutex_unlock(&nbd->config_lock);
1570                         goto out;
1571                 }
1572                 config = nbd_alloc_config();
1573                 if (IS_ERR(config)) {
1574                         ret = PTR_ERR(config);
1575                         mutex_unlock(&nbd->config_lock);
1576                         goto out;
1577                 }
1578                 nbd->config = config;
1579                 refcount_set(&nbd->config_refs, 1);
1580                 refcount_inc(&nbd->refs);
1581                 mutex_unlock(&nbd->config_lock);
1582                 if (max_part)
1583                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1584         } else if (nbd_disconnected(nbd->config)) {
1585                 if (max_part)
1586                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1587         }
1588 out:
1589         mutex_unlock(&nbd_index_mutex);
1590         return ret;
1591 }
1592
1593 static void nbd_release(struct gendisk *disk, fmode_t mode)
1594 {
1595         struct nbd_device *nbd = disk->private_data;
1596
1597         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1598                         disk_openers(disk) == 0)
1599                 nbd_disconnect_and_put(nbd);
1600
1601         nbd_config_put(nbd);
1602         nbd_put(nbd);
1603 }
1604
1605 static const struct block_device_operations nbd_fops =
1606 {
1607         .owner =        THIS_MODULE,
1608         .open =         nbd_open,
1609         .release =      nbd_release,
1610         .ioctl =        nbd_ioctl,
1611         .compat_ioctl = nbd_ioctl,
1612 };
1613
1614 #if IS_ENABLED(CONFIG_DEBUG_FS)
1615
1616 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1617 {
1618         struct nbd_device *nbd = s->private;
1619
1620         if (nbd->pid)
1621                 seq_printf(s, "recv: %d\n", nbd->pid);
1622
1623         return 0;
1624 }
1625
1626 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1627
1628 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1629 {
1630         struct nbd_device *nbd = s->private;
1631         u32 flags = nbd->config->flags;
1632
1633         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1634
1635         seq_puts(s, "Known flags:\n");
1636
1637         if (flags & NBD_FLAG_HAS_FLAGS)
1638                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1639         if (flags & NBD_FLAG_READ_ONLY)
1640                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1641         if (flags & NBD_FLAG_SEND_FLUSH)
1642                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1643         if (flags & NBD_FLAG_SEND_FUA)
1644                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1645         if (flags & NBD_FLAG_SEND_TRIM)
1646                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1647
1648         return 0;
1649 }
1650
1651 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1652
1653 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1654 {
1655         struct dentry *dir;
1656         struct nbd_config *config = nbd->config;
1657
1658         if (!nbd_dbg_dir)
1659                 return -EIO;
1660
1661         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1662         if (!dir) {
1663                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1664                         nbd_name(nbd));
1665                 return -EIO;
1666         }
1667         config->dbg_dir = dir;
1668
1669         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1670         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1671         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1672         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1673         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1674
1675         return 0;
1676 }
1677
1678 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1679 {
1680         debugfs_remove_recursive(nbd->config->dbg_dir);
1681 }
1682
1683 static int nbd_dbg_init(void)
1684 {
1685         struct dentry *dbg_dir;
1686
1687         dbg_dir = debugfs_create_dir("nbd", NULL);
1688         if (!dbg_dir)
1689                 return -EIO;
1690
1691         nbd_dbg_dir = dbg_dir;
1692
1693         return 0;
1694 }
1695
1696 static void nbd_dbg_close(void)
1697 {
1698         debugfs_remove_recursive(nbd_dbg_dir);
1699 }
1700
1701 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1702
1703 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1704 {
1705         return 0;
1706 }
1707
1708 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1709 {
1710 }
1711
1712 static int nbd_dbg_init(void)
1713 {
1714         return 0;
1715 }
1716
1717 static void nbd_dbg_close(void)
1718 {
1719 }
1720
1721 #endif
1722
1723 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1724                             unsigned int hctx_idx, unsigned int numa_node)
1725 {
1726         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1727         cmd->nbd = set->driver_data;
1728         cmd->flags = 0;
1729         mutex_init(&cmd->lock);
1730         return 0;
1731 }
1732
1733 static const struct blk_mq_ops nbd_mq_ops = {
1734         .queue_rq       = nbd_queue_rq,
1735         .complete       = nbd_complete_rq,
1736         .init_request   = nbd_init_request,
1737         .timeout        = nbd_xmit_timeout,
1738 };
1739
1740 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1741 {
1742         struct nbd_device *nbd;
1743         struct gendisk *disk;
1744         int err = -ENOMEM;
1745
1746         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1747         if (!nbd)
1748                 goto out;
1749
1750         nbd->tag_set.ops = &nbd_mq_ops;
1751         nbd->tag_set.nr_hw_queues = 1;
1752         nbd->tag_set.queue_depth = 128;
1753         nbd->tag_set.numa_node = NUMA_NO_NODE;
1754         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1755         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1756                 BLK_MQ_F_BLOCKING;
1757         nbd->tag_set.driver_data = nbd;
1758         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1759         nbd->backend = NULL;
1760
1761         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1762         if (err)
1763                 goto out_free_nbd;
1764
1765         mutex_lock(&nbd_index_mutex);
1766         if (index >= 0) {
1767                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1768                                 GFP_KERNEL);
1769                 if (err == -ENOSPC)
1770                         err = -EEXIST;
1771         } else {
1772                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1773                 if (err >= 0)
1774                         index = err;
1775         }
1776         nbd->index = index;
1777         mutex_unlock(&nbd_index_mutex);
1778         if (err < 0)
1779                 goto out_free_tags;
1780
1781         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1782         if (IS_ERR(disk)) {
1783                 err = PTR_ERR(disk);
1784                 goto out_free_idr;
1785         }
1786         nbd->disk = disk;
1787
1788         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1789                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1790                                           WQ_UNBOUND, 0, nbd->index);
1791         if (!nbd->recv_workq) {
1792                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1793                 err = -ENOMEM;
1794                 goto out_err_disk;
1795         }
1796
1797         /*
1798          * Tell the block layer that we are not a rotational device
1799          */
1800         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1801         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1802         disk->queue->limits.discard_granularity = 0;
1803         blk_queue_max_discard_sectors(disk->queue, 0);
1804         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1805         blk_queue_max_segments(disk->queue, USHRT_MAX);
1806         blk_queue_max_hw_sectors(disk->queue, 65536);
1807         disk->queue->limits.max_sectors = 256;
1808
1809         mutex_init(&nbd->config_lock);
1810         refcount_set(&nbd->config_refs, 0);
1811         /*
1812          * Start out with a zero references to keep other threads from using
1813          * this device until it is fully initialized.
1814          */
1815         refcount_set(&nbd->refs, 0);
1816         INIT_LIST_HEAD(&nbd->list);
1817         disk->major = NBD_MAJOR;
1818         disk->first_minor = index << part_shift;
1819         disk->minors = 1 << part_shift;
1820         disk->fops = &nbd_fops;
1821         disk->private_data = nbd;
1822         sprintf(disk->disk_name, "nbd%d", index);
1823         err = add_disk(disk);
1824         if (err)
1825                 goto out_free_work;
1826
1827         /*
1828          * Now publish the device.
1829          */
1830         refcount_set(&nbd->refs, refs);
1831         nbd_total_devices++;
1832         return nbd;
1833
1834 out_free_work:
1835         destroy_workqueue(nbd->recv_workq);
1836 out_err_disk:
1837         put_disk(disk);
1838 out_free_idr:
1839         mutex_lock(&nbd_index_mutex);
1840         idr_remove(&nbd_index_idr, index);
1841         mutex_unlock(&nbd_index_mutex);
1842 out_free_tags:
1843         blk_mq_free_tag_set(&nbd->tag_set);
1844 out_free_nbd:
1845         kfree(nbd);
1846 out:
1847         return ERR_PTR(err);
1848 }
1849
1850 static struct nbd_device *nbd_find_get_unused(void)
1851 {
1852         struct nbd_device *nbd;
1853         int id;
1854
1855         lockdep_assert_held(&nbd_index_mutex);
1856
1857         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1858                 if (refcount_read(&nbd->config_refs) ||
1859                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1860                         continue;
1861                 if (refcount_inc_not_zero(&nbd->refs))
1862                         return nbd;
1863         }
1864
1865         return NULL;
1866 }
1867
1868 /* Netlink interface. */
1869 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1870         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1871         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1872         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1873         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1874         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1875         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1876         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1877         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1878         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1879         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1880 };
1881
1882 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1883         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1884 };
1885
1886 /* We don't use this right now since we don't parse the incoming list, but we
1887  * still want it here so userspace knows what to expect.
1888  */
1889 static const struct nla_policy __attribute__((unused))
1890 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1891         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1892         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1893 };
1894
1895 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1896 {
1897         struct nbd_config *config = nbd->config;
1898         u64 bsize = nbd_blksize(config);
1899         u64 bytes = config->bytesize;
1900
1901         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1902                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1903
1904         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1905                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1906
1907         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1908                 return nbd_set_size(nbd, bytes, bsize);
1909         return 0;
1910 }
1911
1912 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1913 {
1914         struct nbd_device *nbd;
1915         struct nbd_config *config;
1916         int index = -1;
1917         int ret;
1918         bool put_dev = false;
1919
1920         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1921                 return -EPERM;
1922
1923         if (info->attrs[NBD_ATTR_INDEX]) {
1924                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1925
1926                 /*
1927                  * Too big first_minor can cause duplicate creation of
1928                  * sysfs files/links, since index << part_shift might overflow, or
1929                  * MKDEV() expect that the max bits of first_minor is 20.
1930                  */
1931                 if (index < 0 || index > MINORMASK >> part_shift) {
1932                         pr_err("illegal input index %d\n", index);
1933                         return -EINVAL;
1934                 }
1935         }
1936         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1937                 pr_err("must specify at least one socket\n");
1938                 return -EINVAL;
1939         }
1940         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1941                 pr_err("must specify a size in bytes for the device\n");
1942                 return -EINVAL;
1943         }
1944 again:
1945         mutex_lock(&nbd_index_mutex);
1946         if (index == -1) {
1947                 nbd = nbd_find_get_unused();
1948         } else {
1949                 nbd = idr_find(&nbd_index_idr, index);
1950                 if (nbd) {
1951                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1952                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
1953                             !refcount_inc_not_zero(&nbd->refs)) {
1954                                 mutex_unlock(&nbd_index_mutex);
1955                                 pr_err("device at index %d is going down\n",
1956                                         index);
1957                                 return -EINVAL;
1958                         }
1959                 }
1960         }
1961         mutex_unlock(&nbd_index_mutex);
1962
1963         if (!nbd) {
1964                 nbd = nbd_dev_add(index, 2);
1965                 if (IS_ERR(nbd)) {
1966                         pr_err("failed to add new device\n");
1967                         return PTR_ERR(nbd);
1968                 }
1969         }
1970
1971         mutex_lock(&nbd->config_lock);
1972         if (refcount_read(&nbd->config_refs)) {
1973                 mutex_unlock(&nbd->config_lock);
1974                 nbd_put(nbd);
1975                 if (index == -1)
1976                         goto again;
1977                 pr_err("nbd%d already in use\n", index);
1978                 return -EBUSY;
1979         }
1980         if (WARN_ON(nbd->config)) {
1981                 mutex_unlock(&nbd->config_lock);
1982                 nbd_put(nbd);
1983                 return -EINVAL;
1984         }
1985         config = nbd_alloc_config();
1986         if (IS_ERR(config)) {
1987                 mutex_unlock(&nbd->config_lock);
1988                 nbd_put(nbd);
1989                 pr_err("couldn't allocate config\n");
1990                 return PTR_ERR(config);
1991         }
1992         nbd->config = config;
1993         refcount_set(&nbd->config_refs, 1);
1994         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1995
1996         ret = nbd_genl_size_set(info, nbd);
1997         if (ret)
1998                 goto out;
1999
2000         if (info->attrs[NBD_ATTR_TIMEOUT])
2001                 nbd_set_cmd_timeout(nbd,
2002                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2003         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2004                 config->dead_conn_timeout =
2005                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2006                 config->dead_conn_timeout *= HZ;
2007         }
2008         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
2009                 config->flags =
2010                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
2011         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2012                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2013                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2014                         /*
2015                          * We have 1 ref to keep the device around, and then 1
2016                          * ref for our current operation here, which will be
2017                          * inherited by the config.  If we already have
2018                          * DESTROY_ON_DISCONNECT set then we know we don't have
2019                          * that extra ref already held so we don't need the
2020                          * put_dev.
2021                          */
2022                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2023                                               &nbd->flags))
2024                                 put_dev = true;
2025                 } else {
2026                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2027                                                &nbd->flags))
2028                                 refcount_inc(&nbd->refs);
2029                 }
2030                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2031                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2032                                 &config->runtime_flags);
2033                 }
2034         }
2035
2036         if (info->attrs[NBD_ATTR_SOCKETS]) {
2037                 struct nlattr *attr;
2038                 int rem, fd;
2039
2040                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2041                                     rem) {
2042                         struct nlattr *socks[NBD_SOCK_MAX+1];
2043
2044                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2045                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2046                                 ret = -EINVAL;
2047                                 goto out;
2048                         }
2049                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2050                                                           attr,
2051                                                           nbd_sock_policy,
2052                                                           info->extack);
2053                         if (ret != 0) {
2054                                 pr_err("error processing sock list\n");
2055                                 ret = -EINVAL;
2056                                 goto out;
2057                         }
2058                         if (!socks[NBD_SOCK_FD])
2059                                 continue;
2060                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2061                         ret = nbd_add_socket(nbd, fd, true);
2062                         if (ret)
2063                                 goto out;
2064                 }
2065         }
2066         ret = nbd_start_device(nbd);
2067         if (ret)
2068                 goto out;
2069         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2070                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2071                                           GFP_KERNEL);
2072                 if (!nbd->backend) {
2073                         ret = -ENOMEM;
2074                         goto out;
2075                 }
2076         }
2077         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2078         if (ret) {
2079                 dev_err(disk_to_dev(nbd->disk),
2080                         "device_create_file failed for backend!\n");
2081                 goto out;
2082         }
2083         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2084 out:
2085         mutex_unlock(&nbd->config_lock);
2086         if (!ret) {
2087                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2088                 refcount_inc(&nbd->config_refs);
2089                 nbd_connect_reply(info, nbd->index);
2090         }
2091         nbd_config_put(nbd);
2092         if (put_dev)
2093                 nbd_put(nbd);
2094         return ret;
2095 }
2096
2097 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2098 {
2099         mutex_lock(&nbd->config_lock);
2100         nbd_disconnect(nbd);
2101         sock_shutdown(nbd);
2102         wake_up(&nbd->config->conn_wait);
2103         /*
2104          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2105          * to cancel the inflight I/Os.
2106          */
2107         flush_workqueue(nbd->recv_workq);
2108         nbd_clear_que(nbd);
2109         nbd->task_setup = NULL;
2110         mutex_unlock(&nbd->config_lock);
2111
2112         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2113                                &nbd->config->runtime_flags))
2114                 nbd_config_put(nbd);
2115 }
2116
2117 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2118 {
2119         struct nbd_device *nbd;
2120         int index;
2121
2122         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2123                 return -EPERM;
2124
2125         if (!info->attrs[NBD_ATTR_INDEX]) {
2126                 pr_err("must specify an index to disconnect\n");
2127                 return -EINVAL;
2128         }
2129         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2130         mutex_lock(&nbd_index_mutex);
2131         nbd = idr_find(&nbd_index_idr, index);
2132         if (!nbd) {
2133                 mutex_unlock(&nbd_index_mutex);
2134                 pr_err("couldn't find device at index %d\n", index);
2135                 return -EINVAL;
2136         }
2137         if (!refcount_inc_not_zero(&nbd->refs)) {
2138                 mutex_unlock(&nbd_index_mutex);
2139                 pr_err("device at index %d is going down\n", index);
2140                 return -EINVAL;
2141         }
2142         mutex_unlock(&nbd_index_mutex);
2143         if (!refcount_inc_not_zero(&nbd->config_refs))
2144                 goto put_nbd;
2145         nbd_disconnect_and_put(nbd);
2146         nbd_config_put(nbd);
2147 put_nbd:
2148         nbd_put(nbd);
2149         return 0;
2150 }
2151
2152 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2153 {
2154         struct nbd_device *nbd = NULL;
2155         struct nbd_config *config;
2156         int index;
2157         int ret = 0;
2158         bool put_dev = false;
2159
2160         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2161                 return -EPERM;
2162
2163         if (!info->attrs[NBD_ATTR_INDEX]) {
2164                 pr_err("must specify a device to reconfigure\n");
2165                 return -EINVAL;
2166         }
2167         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2168         mutex_lock(&nbd_index_mutex);
2169         nbd = idr_find(&nbd_index_idr, index);
2170         if (!nbd) {
2171                 mutex_unlock(&nbd_index_mutex);
2172                 pr_err("couldn't find a device at index %d\n", index);
2173                 return -EINVAL;
2174         }
2175         if (nbd->backend) {
2176                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2177                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2178                                        nbd->backend)) {
2179                                 mutex_unlock(&nbd_index_mutex);
2180                                 dev_err(nbd_to_dev(nbd),
2181                                         "backend image doesn't match with %s\n",
2182                                         nbd->backend);
2183                                 return -EINVAL;
2184                         }
2185                 } else {
2186                         mutex_unlock(&nbd_index_mutex);
2187                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2188                         return -EINVAL;
2189                 }
2190         }
2191         if (!refcount_inc_not_zero(&nbd->refs)) {
2192                 mutex_unlock(&nbd_index_mutex);
2193                 pr_err("device at index %d is going down\n", index);
2194                 return -EINVAL;
2195         }
2196         mutex_unlock(&nbd_index_mutex);
2197
2198         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2199                 dev_err(nbd_to_dev(nbd),
2200                         "not configured, cannot reconfigure\n");
2201                 nbd_put(nbd);
2202                 return -EINVAL;
2203         }
2204
2205         mutex_lock(&nbd->config_lock);
2206         config = nbd->config;
2207         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2208             !nbd->pid) {
2209                 dev_err(nbd_to_dev(nbd),
2210                         "not configured, cannot reconfigure\n");
2211                 ret = -EINVAL;
2212                 goto out;
2213         }
2214
2215         ret = nbd_genl_size_set(info, nbd);
2216         if (ret)
2217                 goto out;
2218
2219         if (info->attrs[NBD_ATTR_TIMEOUT])
2220                 nbd_set_cmd_timeout(nbd,
2221                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2222         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2223                 config->dead_conn_timeout =
2224                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2225                 config->dead_conn_timeout *= HZ;
2226         }
2227         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2228                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2229                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2230                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2231                                               &nbd->flags))
2232                                 put_dev = true;
2233                 } else {
2234                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2235                                                &nbd->flags))
2236                                 refcount_inc(&nbd->refs);
2237                 }
2238
2239                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2240                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2241                                         &config->runtime_flags);
2242                 } else {
2243                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2244                                         &config->runtime_flags);
2245                 }
2246         }
2247
2248         if (info->attrs[NBD_ATTR_SOCKETS]) {
2249                 struct nlattr *attr;
2250                 int rem, fd;
2251
2252                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2253                                     rem) {
2254                         struct nlattr *socks[NBD_SOCK_MAX+1];
2255
2256                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2257                                 pr_err("socks must be embedded in a SOCK_ITEM attr\n");
2258                                 ret = -EINVAL;
2259                                 goto out;
2260                         }
2261                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2262                                                           attr,
2263                                                           nbd_sock_policy,
2264                                                           info->extack);
2265                         if (ret != 0) {
2266                                 pr_err("error processing sock list\n");
2267                                 ret = -EINVAL;
2268                                 goto out;
2269                         }
2270                         if (!socks[NBD_SOCK_FD])
2271                                 continue;
2272                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2273                         ret = nbd_reconnect_socket(nbd, fd);
2274                         if (ret) {
2275                                 if (ret == -ENOSPC)
2276                                         ret = 0;
2277                                 goto out;
2278                         }
2279                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2280                 }
2281         }
2282 out:
2283         mutex_unlock(&nbd->config_lock);
2284         nbd_config_put(nbd);
2285         nbd_put(nbd);
2286         if (put_dev)
2287                 nbd_put(nbd);
2288         return ret;
2289 }
2290
2291 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2292         {
2293                 .cmd    = NBD_CMD_CONNECT,
2294                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2295                 .doit   = nbd_genl_connect,
2296         },
2297         {
2298                 .cmd    = NBD_CMD_DISCONNECT,
2299                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2300                 .doit   = nbd_genl_disconnect,
2301         },
2302         {
2303                 .cmd    = NBD_CMD_RECONFIGURE,
2304                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2305                 .doit   = nbd_genl_reconfigure,
2306         },
2307         {
2308                 .cmd    = NBD_CMD_STATUS,
2309                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2310                 .doit   = nbd_genl_status,
2311         },
2312 };
2313
2314 static const struct genl_multicast_group nbd_mcast_grps[] = {
2315         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2316 };
2317
2318 static struct genl_family nbd_genl_family __ro_after_init = {
2319         .hdrsize        = 0,
2320         .name           = NBD_GENL_FAMILY_NAME,
2321         .version        = NBD_GENL_VERSION,
2322         .module         = THIS_MODULE,
2323         .small_ops      = nbd_connect_genl_ops,
2324         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2325         .resv_start_op  = NBD_CMD_STATUS + 1,
2326         .maxattr        = NBD_ATTR_MAX,
2327         .policy = nbd_attr_policy,
2328         .mcgrps         = nbd_mcast_grps,
2329         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2330 };
2331
2332 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2333 {
2334         struct nlattr *dev_opt;
2335         u8 connected = 0;
2336         int ret;
2337
2338         /* This is a little racey, but for status it's ok.  The
2339          * reason we don't take a ref here is because we can't
2340          * take a ref in the index == -1 case as we would need
2341          * to put under the nbd_index_mutex, which could
2342          * deadlock if we are configured to remove ourselves
2343          * once we're disconnected.
2344          */
2345         if (refcount_read(&nbd->config_refs))
2346                 connected = 1;
2347         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2348         if (!dev_opt)
2349                 return -EMSGSIZE;
2350         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2351         if (ret)
2352                 return -EMSGSIZE;
2353         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2354                          connected);
2355         if (ret)
2356                 return -EMSGSIZE;
2357         nla_nest_end(reply, dev_opt);
2358         return 0;
2359 }
2360
2361 static int status_cb(int id, void *ptr, void *data)
2362 {
2363         struct nbd_device *nbd = ptr;
2364         return populate_nbd_status(nbd, (struct sk_buff *)data);
2365 }
2366
2367 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2368 {
2369         struct nlattr *dev_list;
2370         struct sk_buff *reply;
2371         void *reply_head;
2372         size_t msg_size;
2373         int index = -1;
2374         int ret = -ENOMEM;
2375
2376         if (info->attrs[NBD_ATTR_INDEX])
2377                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2378
2379         mutex_lock(&nbd_index_mutex);
2380
2381         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2382                                   nla_attr_size(sizeof(u8)));
2383         msg_size *= (index == -1) ? nbd_total_devices : 1;
2384
2385         reply = genlmsg_new(msg_size, GFP_KERNEL);
2386         if (!reply)
2387                 goto out;
2388         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2389                                        NBD_CMD_STATUS);
2390         if (!reply_head) {
2391                 nlmsg_free(reply);
2392                 goto out;
2393         }
2394
2395         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2396         if (index == -1) {
2397                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2398                 if (ret) {
2399                         nlmsg_free(reply);
2400                         goto out;
2401                 }
2402         } else {
2403                 struct nbd_device *nbd;
2404                 nbd = idr_find(&nbd_index_idr, index);
2405                 if (nbd) {
2406                         ret = populate_nbd_status(nbd, reply);
2407                         if (ret) {
2408                                 nlmsg_free(reply);
2409                                 goto out;
2410                         }
2411                 }
2412         }
2413         nla_nest_end(reply, dev_list);
2414         genlmsg_end(reply, reply_head);
2415         ret = genlmsg_reply(reply, info);
2416 out:
2417         mutex_unlock(&nbd_index_mutex);
2418         return ret;
2419 }
2420
2421 static void nbd_connect_reply(struct genl_info *info, int index)
2422 {
2423         struct sk_buff *skb;
2424         void *msg_head;
2425         int ret;
2426
2427         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2428         if (!skb)
2429                 return;
2430         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2431                                      NBD_CMD_CONNECT);
2432         if (!msg_head) {
2433                 nlmsg_free(skb);
2434                 return;
2435         }
2436         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2437         if (ret) {
2438                 nlmsg_free(skb);
2439                 return;
2440         }
2441         genlmsg_end(skb, msg_head);
2442         genlmsg_reply(skb, info);
2443 }
2444
2445 static void nbd_mcast_index(int index)
2446 {
2447         struct sk_buff *skb;
2448         void *msg_head;
2449         int ret;
2450
2451         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2452         if (!skb)
2453                 return;
2454         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2455                                      NBD_CMD_LINK_DEAD);
2456         if (!msg_head) {
2457                 nlmsg_free(skb);
2458                 return;
2459         }
2460         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2461         if (ret) {
2462                 nlmsg_free(skb);
2463                 return;
2464         }
2465         genlmsg_end(skb, msg_head);
2466         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2467 }
2468
2469 static void nbd_dead_link_work(struct work_struct *work)
2470 {
2471         struct link_dead_args *args = container_of(work, struct link_dead_args,
2472                                                    work);
2473         nbd_mcast_index(args->index);
2474         kfree(args);
2475 }
2476
2477 static int __init nbd_init(void)
2478 {
2479         int i;
2480
2481         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2482
2483         if (max_part < 0) {
2484                 pr_err("max_part must be >= 0\n");
2485                 return -EINVAL;
2486         }
2487
2488         part_shift = 0;
2489         if (max_part > 0) {
2490                 part_shift = fls(max_part);
2491
2492                 /*
2493                  * Adjust max_part according to part_shift as it is exported
2494                  * to user space so that user can know the max number of
2495                  * partition kernel should be able to manage.
2496                  *
2497                  * Note that -1 is required because partition 0 is reserved
2498                  * for the whole disk.
2499                  */
2500                 max_part = (1UL << part_shift) - 1;
2501         }
2502
2503         if ((1UL << part_shift) > DISK_MAX_PARTS)
2504                 return -EINVAL;
2505
2506         if (nbds_max > 1UL << (MINORBITS - part_shift))
2507                 return -EINVAL;
2508
2509         if (register_blkdev(NBD_MAJOR, "nbd"))
2510                 return -EIO;
2511
2512         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2513         if (!nbd_del_wq) {
2514                 unregister_blkdev(NBD_MAJOR, "nbd");
2515                 return -ENOMEM;
2516         }
2517
2518         if (genl_register_family(&nbd_genl_family)) {
2519                 destroy_workqueue(nbd_del_wq);
2520                 unregister_blkdev(NBD_MAJOR, "nbd");
2521                 return -EINVAL;
2522         }
2523         nbd_dbg_init();
2524
2525         for (i = 0; i < nbds_max; i++)
2526                 nbd_dev_add(i, 1);
2527         return 0;
2528 }
2529
2530 static int nbd_exit_cb(int id, void *ptr, void *data)
2531 {
2532         struct list_head *list = (struct list_head *)data;
2533         struct nbd_device *nbd = ptr;
2534
2535         /* Skip nbd that is being removed asynchronously */
2536         if (refcount_read(&nbd->refs))
2537                 list_add_tail(&nbd->list, list);
2538
2539         return 0;
2540 }
2541
2542 static void __exit nbd_cleanup(void)
2543 {
2544         struct nbd_device *nbd;
2545         LIST_HEAD(del_list);
2546
2547         /*
2548          * Unregister netlink interface prior to waiting
2549          * for the completion of netlink commands.
2550          */
2551         genl_unregister_family(&nbd_genl_family);
2552
2553         nbd_dbg_close();
2554
2555         mutex_lock(&nbd_index_mutex);
2556         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2557         mutex_unlock(&nbd_index_mutex);
2558
2559         while (!list_empty(&del_list)) {
2560                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2561                 list_del_init(&nbd->list);
2562                 if (refcount_read(&nbd->config_refs))
2563                         pr_err("possibly leaking nbd_config (ref %d)\n",
2564                                         refcount_read(&nbd->config_refs));
2565                 if (refcount_read(&nbd->refs) != 1)
2566                         pr_err("possibly leaking a device\n");
2567                 nbd_put(nbd);
2568         }
2569
2570         /* Also wait for nbd_dev_remove_work() completes */
2571         destroy_workqueue(nbd_del_wq);
2572
2573         idr_destroy(&nbd_index_idr);
2574         unregister_blkdev(NBD_MAJOR, "nbd");
2575 }
2576
2577 module_init(nbd_init);
2578 module_exit(nbd_cleanup);
2579
2580 MODULE_DESCRIPTION("Network Block Device");
2581 MODULE_LICENSE("GPL");
2582
2583 module_param(nbds_max, int, 0444);
2584 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2585 module_param(max_part, int, 0444);
2586 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");