drivers/block/nbd.c

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