drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   9  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #include <linux/major.h>
  15
  16 #include <linux/blkdev.h>
  17 #include <linux/module.h>
  18 #include <linux/init.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/mm.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/completion.h>
  30 #include <linux/err.h>
  31 #include <linux/kernel.h>
  32 #include <linux/slab.h>
  33 #include <net/sock.h>
  34 #include <linux/net.h>
  35 #include <linux/kthread.h>
  36 #include <linux/types.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/blk-mq.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/types.h>
  42
  43 #include <linux/nbd.h>
  44 #include <linux/nbd-netlink.h>
  45 #include <net/genetlink.h>
  46
  47 #define CREATE_TRACE_POINTS
  48 #include <trace/events/nbd.h>
  49
  50 static DEFINE_IDR(nbd_index_idr);
  51 static DEFINE_MUTEX(nbd_index_mutex);
  52 static struct workqueue_struct *nbd_del_wq;
  53 static int nbd_total_devices = 0;
  54
  55 struct nbd_sock {
  56         struct socket *sock;
  57         struct mutex tx_lock;
  58         struct request *pending;
  59         int sent;
  60         bool dead;
  61         int fallback_index;
  62         int cookie;
  63 };
  64
  65 struct recv_thread_args {
  66         struct work_struct work;
  67         struct nbd_device *nbd;
  68         int index;
  69 };
  70
  71 struct link_dead_args {
  72         struct work_struct work;
  73         int index;
  74 };
  75
  76 #define NBD_RT_TIMEDOUT                 0
  77 #define NBD_RT_DISCONNECT_REQUESTED     1
  78 #define NBD_RT_DISCONNECTED             2
  79 #define NBD_RT_HAS_PID_FILE             3
  80 #define NBD_RT_HAS_CONFIG_REF           4
  81 #define NBD_RT_BOUND                    5
  82 #define NBD_RT_DISCONNECT_ON_CLOSE      6
  83 #define NBD_RT_HAS_BACKEND_FILE         7
  84
  85 #define NBD_DESTROY_ON_DISCONNECT       0
  86 #define NBD_DISCONNECT_REQUESTED        1
  87
  88 struct nbd_config {
  89         u32 flags;
  90         unsigned long runtime_flags;
  91         u64 dead_conn_timeout;
  92
  93         struct nbd_sock **socks;
  94         int num_connections;
  95         atomic_t live_connections;
  96         wait_queue_head_t conn_wait;
  97
  98         atomic_t recv_threads;
  99         wait_queue_head_t recv_wq;
 100         unsigned int blksize_bits;
 101         loff_t bytesize;
 102 #if IS_ENABLED(CONFIG_DEBUG_FS)
 103         struct dentry *dbg_dir;
 104 #endif
 105 };
 106
 107 static inline unsigned int nbd_blksize(struct nbd_config *config)
 108 {
 109         return 1u << config->blksize_bits;
 110 }
 111
 112 struct nbd_device {
 113         struct blk_mq_tag_set tag_set;
 114
 115         int index;
 116         refcount_t config_refs;
 117         refcount_t refs;
 118         struct nbd_config *config;
 119         struct mutex config_lock;
 120         struct gendisk *disk;
 121         struct workqueue_struct *recv_workq;
 122         struct work_struct remove_work;
 123
 124         struct list_head list;
 125         struct task_struct *task_setup;
 126
 127         unsigned long flags;
 128         pid_t pid; /* pid of nbd-client, if attached */
 129
 130         char *backend;
 131 };
 132
 133 #define NBD_CMD_REQUEUED        1
 134 /*
 135  * This flag will be set if nbd_queue_rq() succeed, and will be checked and
 136  * cleared in completion. Both setting and clearing of the flag are protected
 137  * by cmd->lock.
 138  */
 139 #define NBD_CMD_INFLIGHT        2
 140
 141 struct nbd_cmd {
 142         struct nbd_device *nbd;
 143         struct mutex lock;
 144         int index;
 145         int cookie;
 146         int retries;
 147         blk_status_t status;
 148         unsigned long flags;
 149         u32 cmd_cookie;
 150 };
 151
 152 #if IS_ENABLED(CONFIG_DEBUG_FS)
 153 static struct dentry *nbd_dbg_dir;
 154 #endif
 155
 156 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 157
 158 #define NBD_MAGIC 0x68797548
 159
 160 #define NBD_DEF_BLKSIZE_BITS 10
 161
 162 static unsigned int nbds_max = 16;
 163 static int max_part = 16;
 164 static int part_shift;
 165
 166 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 167 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 168 static void nbd_config_put(struct nbd_device *nbd);
 169 static void nbd_connect_reply(struct genl_info *info, int index);
 170 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 171 static void nbd_dead_link_work(struct work_struct *work);
 172 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 173
 174 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 175 {
 176         return disk_to_dev(nbd->disk);
 177 }
 178
 179 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 180 {
 181         struct request *req = blk_mq_rq_from_pdu(cmd);
 182
 183         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 184                 blk_mq_requeue_request(req, true);
 185 }
 186
 187 #define NBD_COOKIE_BITS 32
 188
 189 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 190 {
 191         struct request *req = blk_mq_rq_from_pdu(cmd);
 192         u32 tag = blk_mq_unique_tag(req);
 193         u64 cookie = cmd->cmd_cookie;
 194
 195         return (cookie << NBD_COOKIE_BITS) | tag;
 196 }
 197
 198 static u32 nbd_handle_to_tag(u64 handle)
 199 {
 200         return (u32)handle;
 201 }
 202
 203 static u32 nbd_handle_to_cookie(u64 handle)
 204 {
 205         return (u32)(handle >> NBD_COOKIE_BITS);
 206 }
 207
 208 static const char *nbdcmd_to_ascii(int cmd)
 209 {
 210         switch (cmd) {
 211         case  NBD_CMD_READ: return "read";
 212         case NBD_CMD_WRITE: return "write";
 213         case  NBD_CMD_DISC: return "disconnect";
 214         case NBD_CMD_FLUSH: return "flush";
 215         case  NBD_CMD_TRIM: return "trim/discard";
 216         }
 217         return "invalid";
 218 }
 219
 220 static ssize_t pid_show(struct device *dev,
 221                         struct device_attribute *attr, char *buf)
 222 {
 223         struct gendisk *disk = dev_to_disk(dev);
 224         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 225
 226         return sprintf(buf, "%d\n", nbd->pid);
 227 }
 228
 229 static const struct device_attribute pid_attr = {
 230         .attr = { .name = "pid", .mode = 0444},
 231         .show = pid_show,
 232 };
 233
 234 static ssize_t backend_show(struct device *dev,
 235                 struct device_attribute *attr, char *buf)
 236 {
 237         struct gendisk *disk = dev_to_disk(dev);
 238         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 239
 240         return sprintf(buf, "%s\n", nbd->backend ?: "");
 241 }
 242
 243 static const struct device_attribute backend_attr = {
 244         .attr = { .name = "backend", .mode = 0444},
 245         .show = backend_show,
 246 };
 247
 248 static void nbd_dev_remove(struct nbd_device *nbd)
 249 {
 250         struct gendisk *disk = nbd->disk;
 251
 252         del_gendisk(disk);
 253         blk_cleanup_disk(disk);
 254         blk_mq_free_tag_set(&nbd->tag_set);
 255
 256         /*
 257          * Remove from idr after del_gendisk() completes, so if the same ID is
 258          * reused, the following add_disk() will succeed.
 259          */
 260         mutex_lock(&nbd_index_mutex);
 261         idr_remove(&nbd_index_idr, nbd->index);
 262         mutex_unlock(&nbd_index_mutex);
 263         destroy_workqueue(nbd->recv_workq);
 264         kfree(nbd);
 265 }
 266
 267 static void nbd_dev_remove_work(struct work_struct *work)
 268 {
 269         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 270 }
 271
 272 static void nbd_put(struct nbd_device *nbd)
 273 {
 274         if (!refcount_dec_and_test(&nbd->refs))
 275                 return;
 276
 277         /* Call del_gendisk() asynchrounously to prevent deadlock */
 278         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 279                 queue_work(nbd_del_wq, &nbd->remove_work);
 280         else
 281                 nbd_dev_remove(nbd);
 282 }
 283
 284 static int nbd_disconnected(struct nbd_config *config)
 285 {
 286         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 287                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 288 }
 289
 290 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 291                                 int notify)
 292 {
 293         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 294                 struct link_dead_args *args;
 295                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 296                 if (args) {
 297                         INIT_WORK(&args->work, nbd_dead_link_work);
 298                         args->index = nbd->index;
 299                         queue_work(system_wq, &args->work);
 300                 }
 301         }
 302         if (!nsock->dead) {
 303                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 304                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 305                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 306                                                &nbd->config->runtime_flags)) {
 307                                 set_bit(NBD_RT_DISCONNECTED,
 308                                         &nbd->config->runtime_flags);
 309                                 dev_info(nbd_to_dev(nbd),
 310                                         "Disconnected due to user request.\n");
 311                         }
 312                 }
 313         }
 314         nsock->dead = true;
 315         nsock->pending = NULL;
 316         nsock->sent = 0;
 317 }
 318
 319 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 320                 loff_t blksize)
 321 {
 322         if (!blksize)
 323                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 324
 325         if (blk_validate_block_size(blksize))
 326                 return -EINVAL;
 327
 328         nbd->config->bytesize = bytesize;
 329         nbd->config->blksize_bits = __ffs(blksize);
 330
 331         if (!nbd->pid)
 332                 return 0;
 333
 334         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 335                 nbd->disk->queue->limits.discard_granularity = blksize;
 336                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 337         }
 338         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 339         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 340
 341         if (max_part)
 342                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 343         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 344                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 345         return 0;
 346 }
 347
 348 static void nbd_complete_rq(struct request *req)
 349 {
 350         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 351
 352         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 353                 cmd->status ? "failed" : "done");
 354
 355         blk_mq_end_request(req, cmd->status);
 356 }
 357
 358 /*
 359  * Forcibly shutdown the socket causing all listeners to error
 360  */
 361 static void sock_shutdown(struct nbd_device *nbd)
 362 {
 363         struct nbd_config *config = nbd->config;
 364         int i;
 365
 366         if (config->num_connections == 0)
 367                 return;
 368         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 369                 return;
 370
 371         for (i = 0; i < config->num_connections; i++) {
 372                 struct nbd_sock *nsock = config->socks[i];
 373                 mutex_lock(&nsock->tx_lock);
 374                 nbd_mark_nsock_dead(nbd, nsock, 0);
 375                 mutex_unlock(&nsock->tx_lock);
 376         }
 377         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 378 }
 379
 380 static u32 req_to_nbd_cmd_type(struct request *req)
 381 {
 382         switch (req_op(req)) {
 383         case REQ_OP_DISCARD:
 384                 return NBD_CMD_TRIM;
 385         case REQ_OP_FLUSH:
 386                 return NBD_CMD_FLUSH;
 387         case REQ_OP_WRITE:
 388                 return NBD_CMD_WRITE;
 389         case REQ_OP_READ:
 390                 return NBD_CMD_READ;
 391         default:
 392                 return U32_MAX;
 393         }
 394 }
 395
 396 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 397                                                  bool reserved)
 398 {
 399         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 400         struct nbd_device *nbd = cmd->nbd;
 401         struct nbd_config *config;
 402
 403         if (!mutex_trylock(&cmd->lock))
 404                 return BLK_EH_RESET_TIMER;
 405
 406         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 407                 mutex_unlock(&cmd->lock);
 408                 return BLK_EH_DONE;
 409         }
 410
 411         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 412                 cmd->status = BLK_STS_TIMEOUT;
 413                 mutex_unlock(&cmd->lock);
 414                 goto done;
 415         }
 416         config = nbd->config;
 417
 418         if (config->num_connections > 1 ||
 419             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 420                 dev_err_ratelimited(nbd_to_dev(nbd),
 421                                     "Connection timed out, retrying (%d/%d alive)\n",
 422                                     atomic_read(&config->live_connections),
 423                                     config->num_connections);
 424                 /*
 425                  * Hooray we have more connections, requeue this IO, the submit
 426                  * path will put it on a real connection. Or if only one
 427                  * connection is configured, the submit path will wait util
 428                  * a new connection is reconfigured or util dead timeout.
 429                  */
 430                 if (config->socks) {
 431                         if (cmd->index < config->num_connections) {
 432                                 struct nbd_sock *nsock =
 433                                         config->socks[cmd->index];
 434                                 mutex_lock(&nsock->tx_lock);
 435                                 /* We can have multiple outstanding requests, so
 436                                  * we don't want to mark the nsock dead if we've
 437                                  * already reconnected with a new socket, so
 438                                  * only mark it dead if its the same socket we
 439                                  * were sent out on.
 440                                  */
 441                                 if (cmd->cookie == nsock->cookie)
 442                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 443                                 mutex_unlock(&nsock->tx_lock);
 444                         }
 445                         mutex_unlock(&cmd->lock);
 446                         nbd_requeue_cmd(cmd);
 447                         nbd_config_put(nbd);
 448                         return BLK_EH_DONE;
 449                 }
 450         }
 451
 452         if (!nbd->tag_set.timeout) {
 453                 /*
 454                  * Userspace sets timeout=0 to disable socket disconnection,
 455                  * so just warn and reset the timer.
 456                  */
 457                 struct nbd_sock *nsock = config->socks[cmd->index];
 458                 cmd->retries++;
 459                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 460                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 461                         (unsigned long long)blk_rq_pos(req) << 9,
 462                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 463
 464                 mutex_lock(&nsock->tx_lock);
 465                 if (cmd->cookie != nsock->cookie) {
 466                         nbd_requeue_cmd(cmd);
 467                         mutex_unlock(&nsock->tx_lock);
 468                         mutex_unlock(&cmd->lock);
 469                         nbd_config_put(nbd);
 470                         return BLK_EH_DONE;
 471                 }
 472                 mutex_unlock(&nsock->tx_lock);
 473                 mutex_unlock(&cmd->lock);
 474                 nbd_config_put(nbd);
 475                 return BLK_EH_RESET_TIMER;
 476         }
 477
 478         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 479         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 480         cmd->status = BLK_STS_IOERR;
 481         mutex_unlock(&cmd->lock);
 482         sock_shutdown(nbd);
 483         nbd_config_put(nbd);
 484 done:
 485         blk_mq_complete_request(req);
 486         return BLK_EH_DONE;
 487 }
 488
 489 /*
 490  *  Send or receive packet. Return a positive value on success and
 491  *  negtive value on failue, and never return 0.
 492  */
 493 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 494                      struct iov_iter *iter, int msg_flags, int *sent)
 495 {
 496         struct nbd_config *config = nbd->config;
 497         struct socket *sock = config->socks[index]->sock;
 498         int result;
 499         struct msghdr msg;
 500         unsigned int noreclaim_flag;
 501
 502         if (unlikely(!sock)) {
 503                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 504                         "Attempted %s on closed socket in sock_xmit\n",
 505                         (send ? "send" : "recv"));
 506                 return -EINVAL;
 507         }
 508
 509         msg.msg_iter = *iter;
 510
 511         noreclaim_flag = memalloc_noreclaim_save();
 512         do {
 513                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 514                 msg.msg_name = NULL;
 515                 msg.msg_namelen = 0;
 516                 msg.msg_control = NULL;
 517                 msg.msg_controllen = 0;
 518                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 519
 520                 if (send)
 521                         result = sock_sendmsg(sock, &msg);
 522                 else
 523                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 524
 525                 if (result <= 0) {
 526                         if (result == 0)
 527                                 result = -EPIPE; /* short read */
 528                         break;
 529                 }
 530                 if (sent)
 531                         *sent += result;
 532         } while (msg_data_left(&msg));
 533
 534         memalloc_noreclaim_restore(noreclaim_flag);
 535
 536         return result;
 537 }
 538
 539 /*
 540  * Different settings for sk->sk_sndtimeo can result in different return values
 541  * if there is a signal pending when we enter sendmsg, because reasons?
 542  */
 543 static inline int was_interrupted(int result)
 544 {
 545         return result == -ERESTARTSYS || result == -EINTR;
 546 }
 547
 548 /* always call with the tx_lock held */
 549 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 550 {
 551         struct request *req = blk_mq_rq_from_pdu(cmd);
 552         struct nbd_config *config = nbd->config;
 553         struct nbd_sock *nsock = config->socks[index];
 554         int result;
 555         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 556         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 557         struct iov_iter from;
 558         unsigned long size = blk_rq_bytes(req);
 559         struct bio *bio;
 560         u64 handle;
 561         u32 type;
 562         u32 nbd_cmd_flags = 0;
 563         int sent = nsock->sent, skip = 0;
 564
 565         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 566
 567         type = req_to_nbd_cmd_type(req);
 568         if (type == U32_MAX)
 569                 return -EIO;
 570
 571         if (rq_data_dir(req) == WRITE &&
 572             (config->flags & NBD_FLAG_READ_ONLY)) {
 573                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 574                                     "Write on read-only\n");
 575                 return -EIO;
 576         }
 577
 578         if (req->cmd_flags & REQ_FUA)
 579                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 580
 581         /* We did a partial send previously, and we at least sent the whole
 582          * request struct, so just go and send the rest of the pages in the
 583          * request.
 584          */
 585         if (sent) {
 586                 if (sent >= sizeof(request)) {
 587                         skip = sent - sizeof(request);
 588
 589                         /* initialize handle for tracing purposes */
 590                         handle = nbd_cmd_handle(cmd);
 591
 592                         goto send_pages;
 593                 }
 594                 iov_iter_advance(&from, sent);
 595         } else {
 596                 cmd->cmd_cookie++;
 597         }
 598         cmd->index = index;
 599         cmd->cookie = nsock->cookie;
 600         cmd->retries = 0;
 601         request.type = htonl(type | nbd_cmd_flags);
 602         if (type != NBD_CMD_FLUSH) {
 603                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 604                 request.len = htonl(size);
 605         }
 606         handle = nbd_cmd_handle(cmd);
 607         memcpy(request.handle, &handle, sizeof(handle));
 608
 609         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 610
 611         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 612                 req, nbdcmd_to_ascii(type),
 613                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 614         result = sock_xmit(nbd, index, 1, &from,
 615                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 616         trace_nbd_header_sent(req, handle);
 617         if (result < 0) {
 618                 if (was_interrupted(result)) {
 619                         /* If we havne't sent anything we can just return BUSY,
 620                          * however if we have sent something we need to make
 621                          * sure we only allow this req to be sent until we are
 622                          * completely done.
 623                          */
 624                         if (sent) {
 625                                 nsock->pending = req;
 626                                 nsock->sent = sent;
 627                         }
 628                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 629                         return BLK_STS_RESOURCE;
 630                 }
 631                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 632                         "Send control failed (result %d)\n", result);
 633                 return -EAGAIN;
 634         }
 635 send_pages:
 636         if (type != NBD_CMD_WRITE)
 637                 goto out;
 638
 639         bio = req->bio;
 640         while (bio) {
 641                 struct bio *next = bio->bi_next;
 642                 struct bvec_iter iter;
 643                 struct bio_vec bvec;
 644
 645                 bio_for_each_segment(bvec, bio, iter) {
 646                         bool is_last = !next && bio_iter_last(bvec, iter);
 647                         int flags = is_last ? 0 : MSG_MORE;
 648
 649                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 650                                 req, bvec.bv_len);
 651                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 652                         if (skip) {
 653                                 if (skip >= iov_iter_count(&from)) {
 654                                         skip -= iov_iter_count(&from);
 655                                         continue;
 656                                 }
 657                                 iov_iter_advance(&from, skip);
 658                                 skip = 0;
 659                         }
 660                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 661                         if (result < 0) {
 662                                 if (was_interrupted(result)) {
 663                                         /* We've already sent the header, we
 664                                          * have no choice but to set pending and
 665                                          * return BUSY.
 666                                          */
 667                                         nsock->pending = req;
 668                                         nsock->sent = sent;
 669                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 670                                         return BLK_STS_RESOURCE;
 671                                 }
 672                                 dev_err(disk_to_dev(nbd->disk),
 673                                         "Send data failed (result %d)\n",
 674                                         result);
 675                                 return -EAGAIN;
 676                         }
 677                         /*
 678                          * The completion might already have come in,
 679                          * so break for the last one instead of letting
 680                          * the iterator do it. This prevents use-after-free
 681                          * of the bio.
 682                          */
 683                         if (is_last)
 684                                 break;
 685                 }
 686                 bio = next;
 687         }
 688 out:
 689         trace_nbd_payload_sent(req, handle);
 690         nsock->pending = NULL;
 691         nsock->sent = 0;
 692         return 0;
 693 }
 694
 695 static int nbd_read_reply(struct nbd_device *nbd, int index,
 696                           struct nbd_reply *reply)
 697 {
 698         struct kvec iov = {.iov_base = reply, .iov_len = sizeof(*reply)};
 699         struct iov_iter to;
 700         int result;
 701
 702         reply->magic = 0;
 703         iov_iter_kvec(&to, READ, &iov, 1, sizeof(*reply));
 704         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 705         if (result < 0) {
 706                 if (!nbd_disconnected(nbd->config))
 707                         dev_err(disk_to_dev(nbd->disk),
 708                                 "Receive control failed (result %d)\n", result);
 709                 return result;
 710         }
 711
 712         if (ntohl(reply->magic) != NBD_REPLY_MAGIC) {
 713                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 714                                 (unsigned long)ntohl(reply->magic));
 715                 return -EPROTO;
 716         }
 717
 718         return 0;
 719 }
 720
 721 /* NULL returned = something went wrong, inform userspace */
 722 static struct nbd_cmd *nbd_handle_reply(struct nbd_device *nbd, int index,
 723                                         struct nbd_reply *reply)
 724 {
 725         int result;
 726         struct nbd_cmd *cmd;
 727         struct request *req = NULL;
 728         u64 handle;
 729         u16 hwq;
 730         u32 tag;
 731         int ret = 0;
 732
 733         memcpy(&handle, reply->handle, sizeof(handle));
 734         tag = nbd_handle_to_tag(handle);
 735         hwq = blk_mq_unique_tag_to_hwq(tag);
 736         if (hwq < nbd->tag_set.nr_hw_queues)
 737                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 738                                        blk_mq_unique_tag_to_tag(tag));
 739         if (!req || !blk_mq_request_started(req)) {
 740                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 741                         tag, req);
 742                 return ERR_PTR(-ENOENT);
 743         }
 744         trace_nbd_header_received(req, handle);
 745         cmd = blk_mq_rq_to_pdu(req);
 746
 747         mutex_lock(&cmd->lock);
 748         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 749                 dev_err(disk_to_dev(nbd->disk), "Suspicious reply %d (status %u flags %lu)",
 750                         tag, cmd->status, cmd->flags);
 751                 ret = -ENOENT;
 752                 goto out;
 753         }
 754         if (cmd->index != index) {
 755                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply %d from different sock %d (expected %d)",
 756                         tag, index, cmd->index);
 757                 ret = -ENOENT;
 758                 goto out;
 759         }
 760         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 761                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 762                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 763                 ret = -ENOENT;
 764                 goto out;
 765         }
 766         if (cmd->status != BLK_STS_OK) {
 767                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 768                         req);
 769                 ret = -ENOENT;
 770                 goto out;
 771         }
 772         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 773                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 774                         req);
 775                 ret = -ENOENT;
 776                 goto out;
 777         }
 778         if (ntohl(reply->error)) {
 779                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 780                         ntohl(reply->error));
 781                 cmd->status = BLK_STS_IOERR;
 782                 goto out;
 783         }
 784
 785         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 786         if (rq_data_dir(req) != WRITE) {
 787                 struct req_iterator iter;
 788                 struct bio_vec bvec;
 789                 struct iov_iter to;
 790
 791                 rq_for_each_segment(bvec, req, iter) {
 792                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 793                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 794                         if (result < 0) {
 795                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 796                                         result);
 797                                 /*
 798                                  * If we've disconnected, we need to make sure we
 799                                  * complete this request, otherwise error out
 800                                  * and let the timeout stuff handle resubmitting
 801                                  * this request onto another connection.
 802                                  */
 803                                 if (nbd_disconnected(nbd->config)) {
 804                                         cmd->status = BLK_STS_IOERR;
 805                                         goto out;
 806                                 }
 807                                 ret = -EIO;
 808                                 goto out;
 809                         }
 810                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 811                                 req, bvec.bv_len);
 812                 }
 813         }
 814 out:
 815         trace_nbd_payload_received(req, handle);
 816         mutex_unlock(&cmd->lock);
 817         return ret ? ERR_PTR(ret) : cmd;
 818 }
 819
 820 static void recv_work(struct work_struct *work)
 821 {
 822         struct recv_thread_args *args = container_of(work,
 823                                                      struct recv_thread_args,
 824                                                      work);
 825         struct nbd_device *nbd = args->nbd;
 826         struct nbd_config *config = nbd->config;
 827         struct request_queue *q = nbd->disk->queue;
 828         struct nbd_sock *nsock;
 829         struct nbd_cmd *cmd;
 830         struct request *rq;
 831
 832         while (1) {
 833                 struct nbd_reply reply;
 834
 835                 if (nbd_read_reply(nbd, args->index, &reply))
 836                         break;
 837
 838                 /*
 839                  * Grab .q_usage_counter so request pool won't go away, then no
 840                  * request use-after-free is possible during nbd_handle_reply().
 841                  * If queue is frozen, there won't be any inflight requests, we
 842                  * needn't to handle the incoming garbage message.
 843                  */
 844                 if (!percpu_ref_tryget(&q->q_usage_counter)) {
 845                         dev_err(disk_to_dev(nbd->disk), "%s: no io inflight\n",
 846                                 __func__);
 847                         break;
 848                 }
 849
 850                 cmd = nbd_handle_reply(nbd, args->index, &reply);
 851                 if (IS_ERR(cmd)) {
 852                         percpu_ref_put(&q->q_usage_counter);
 853                         break;
 854                 }
 855
 856                 rq = blk_mq_rq_from_pdu(cmd);
 857                 if (likely(!blk_should_fake_timeout(rq->q)))
 858                         blk_mq_complete_request(rq);
 859                 percpu_ref_put(&q->q_usage_counter);
 860         }
 861
 862         nsock = config->socks[args->index];
 863         mutex_lock(&nsock->tx_lock);
 864         nbd_mark_nsock_dead(nbd, nsock, 1);
 865         mutex_unlock(&nsock->tx_lock);
 866
 867         nbd_config_put(nbd);
 868         atomic_dec(&config->recv_threads);
 869         wake_up(&config->recv_wq);
 870         kfree(args);
 871 }
 872
 873 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 874 {
 875         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 876
 877         /* don't abort one completed request */
 878         if (blk_mq_request_completed(req))
 879                 return true;
 880
 881         mutex_lock(&cmd->lock);
 882         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 883                 mutex_unlock(&cmd->lock);
 884                 return true;
 885         }
 886         cmd->status = BLK_STS_IOERR;
 887         mutex_unlock(&cmd->lock);
 888
 889         blk_mq_complete_request(req);
 890         return true;
 891 }
 892
 893 static void nbd_clear_que(struct nbd_device *nbd)
 894 {
 895         blk_mq_quiesce_queue(nbd->disk->queue);
 896         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 897         blk_mq_unquiesce_queue(nbd->disk->queue);
 898         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 899 }
 900
 901 static int find_fallback(struct nbd_device *nbd, int index)
 902 {
 903         struct nbd_config *config = nbd->config;
 904         int new_index = -1;
 905         struct nbd_sock *nsock = config->socks[index];
 906         int fallback = nsock->fallback_index;
 907
 908         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 909                 return new_index;
 910
 911         if (config->num_connections <= 1) {
 912                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 913                                     "Dead connection, failed to find a fallback\n");
 914                 return new_index;
 915         }
 916
 917         if (fallback >= 0 && fallback < config->num_connections &&
 918             !config->socks[fallback]->dead)
 919                 return fallback;
 920
 921         if (nsock->fallback_index < 0 ||
 922             nsock->fallback_index >= config->num_connections ||
 923             config->socks[nsock->fallback_index]->dead) {
 924                 int i;
 925                 for (i = 0; i < config->num_connections; i++) {
 926                         if (i == index)
 927                                 continue;
 928                         if (!config->socks[i]->dead) {
 929                                 new_index = i;
 930                                 break;
 931                         }
 932                 }
 933                 nsock->fallback_index = new_index;
 934                 if (new_index < 0) {
 935                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 936                                             "Dead connection, failed to find a fallback\n");
 937                         return new_index;
 938                 }
 939         }
 940         new_index = nsock->fallback_index;
 941         return new_index;
 942 }
 943
 944 static int wait_for_reconnect(struct nbd_device *nbd)
 945 {
 946         struct nbd_config *config = nbd->config;
 947         if (!config->dead_conn_timeout)
 948                 return 0;
 949
 950         if (!wait_event_timeout(config->conn_wait,
 951                                 test_bit(NBD_RT_DISCONNECTED,
 952                                          &config->runtime_flags) ||
 953                                 atomic_read(&config->live_connections) > 0,
 954                                 config->dead_conn_timeout))
 955                 return 0;
 956
 957         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
 958 }
 959
 960 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 961 {
 962         struct request *req = blk_mq_rq_from_pdu(cmd);
 963         struct nbd_device *nbd = cmd->nbd;
 964         struct nbd_config *config;
 965         struct nbd_sock *nsock;
 966         int ret;
 967
 968         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 969                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 970                                     "Socks array is empty\n");
 971                 return -EINVAL;
 972         }
 973         config = nbd->config;
 974
 975         if (index >= config->num_connections) {
 976                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 977                                     "Attempted send on invalid socket\n");
 978                 nbd_config_put(nbd);
 979                 return -EINVAL;
 980         }
 981         cmd->status = BLK_STS_OK;
 982 again:
 983         nsock = config->socks[index];
 984         mutex_lock(&nsock->tx_lock);
 985         if (nsock->dead) {
 986                 int old_index = index;
 987                 index = find_fallback(nbd, index);
 988                 mutex_unlock(&nsock->tx_lock);
 989                 if (index < 0) {
 990                         if (wait_for_reconnect(nbd)) {
 991                                 index = old_index;
 992                                 goto again;
 993                         }
 994                         /* All the sockets should already be down at this point,
 995                          * we just want to make sure that DISCONNECTED is set so
 996                          * any requests that come in that were queue'ed waiting
 997                          * for the reconnect timer don't trigger the timer again
 998                          * and instead just error out.
 999                          */
1000                         sock_shutdown(nbd);
1001                         nbd_config_put(nbd);
1002                         return -EIO;
1003                 }
1004                 goto again;
1005         }
1006
1007         /* Handle the case that we have a pending request that was partially
1008          * transmitted that _has_ to be serviced first.  We need to call requeue
1009          * here so that it gets put _after_ the request that is already on the
1010          * dispatch list.
1011          */
1012         blk_mq_start_request(req);
1013         if (unlikely(nsock->pending && nsock->pending != req)) {
1014                 nbd_requeue_cmd(cmd);
1015                 ret = 0;
1016                 goto out;
1017         }
1018         /*
1019          * Some failures are related to the link going down, so anything that
1020          * returns EAGAIN can be retried on a different socket.
1021          */
1022         ret = nbd_send_cmd(nbd, cmd, index);
1023         /*
1024          * Access to this flag is protected by cmd->lock, thus it's safe to set
1025          * the flag after nbd_send_cmd() succeed to send request to server.
1026          */
1027         if (!ret)
1028                 __set_bit(NBD_CMD_INFLIGHT, &cmd->flags);
1029         else if (ret == -EAGAIN) {
1030                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1031                                     "Request send failed, requeueing\n");
1032                 nbd_mark_nsock_dead(nbd, nsock, 1);
1033                 nbd_requeue_cmd(cmd);
1034                 ret = 0;
1035         }
1036 out:
1037         mutex_unlock(&nsock->tx_lock);
1038         nbd_config_put(nbd);
1039         return ret;
1040 }
1041
1042 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1043                         const struct blk_mq_queue_data *bd)
1044 {
1045         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1046         int ret;
1047
1048         /*
1049          * Since we look at the bio's to send the request over the network we
1050          * need to make sure the completion work doesn't mark this request done
1051          * before we are done doing our send.  This keeps us from dereferencing
1052          * freed data if we have particularly fast completions (ie we get the
1053          * completion before we exit sock_xmit on the last bvec) or in the case
1054          * that the server is misbehaving (or there was an error) before we're
1055          * done sending everything over the wire.
1056          */
1057         mutex_lock(&cmd->lock);
1058         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1059
1060         /* We can be called directly from the user space process, which means we
1061          * could possibly have signals pending so our sendmsg will fail.  In
1062          * this case we need to return that we are busy, otherwise error out as
1063          * appropriate.
1064          */
1065         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1066         if (ret < 0)
1067                 ret = BLK_STS_IOERR;
1068         else if (!ret)
1069                 ret = BLK_STS_OK;
1070         mutex_unlock(&cmd->lock);
1071
1072         return ret;
1073 }
1074
1075 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1076                                      int *err)
1077 {
1078         struct socket *sock;
1079
1080         *err = 0;
1081         sock = sockfd_lookup(fd, err);
1082         if (!sock)
1083                 return NULL;
1084
1085         if (sock->ops->shutdown == sock_no_shutdown) {
1086                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1087                 *err = -EINVAL;
1088                 sockfd_put(sock);
1089                 return NULL;
1090         }
1091
1092         return sock;
1093 }
1094
1095 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1096                           bool netlink)
1097 {
1098         struct nbd_config *config = nbd->config;
1099         struct socket *sock;
1100         struct nbd_sock **socks;
1101         struct nbd_sock *nsock;
1102         int err;
1103
1104         sock = nbd_get_socket(nbd, arg, &err);
1105         if (!sock)
1106                 return err;
1107
1108         /*
1109          * We need to make sure we don't get any errant requests while we're
1110          * reallocating the ->socks array.
1111          */
1112         blk_mq_freeze_queue(nbd->disk->queue);
1113
1114         if (!netlink && !nbd->task_setup &&
1115             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1116                 nbd->task_setup = current;
1117
1118         if (!netlink &&
1119             (nbd->task_setup != current ||
1120              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1121                 dev_err(disk_to_dev(nbd->disk),
1122                         "Device being setup by another task");
1123                 err = -EBUSY;
1124                 goto put_socket;
1125         }
1126
1127         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1128         if (!nsock) {
1129                 err = -ENOMEM;
1130                 goto put_socket;
1131         }
1132
1133         socks = krealloc(config->socks, (config->num_connections + 1) *
1134                          sizeof(struct nbd_sock *), GFP_KERNEL);
1135         if (!socks) {
1136                 kfree(nsock);
1137                 err = -ENOMEM;
1138                 goto put_socket;
1139         }
1140
1141         config->socks = socks;
1142
1143         nsock->fallback_index = -1;
1144         nsock->dead = false;
1145         mutex_init(&nsock->tx_lock);
1146         nsock->sock = sock;
1147         nsock->pending = NULL;
1148         nsock->sent = 0;
1149         nsock->cookie = 0;
1150         socks[config->num_connections++] = nsock;
1151         atomic_inc(&config->live_connections);
1152         blk_mq_unfreeze_queue(nbd->disk->queue);
1153
1154         return 0;
1155
1156 put_socket:
1157         blk_mq_unfreeze_queue(nbd->disk->queue);
1158         sockfd_put(sock);
1159         return err;
1160 }
1161
1162 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1163 {
1164         struct nbd_config *config = nbd->config;
1165         struct socket *sock, *old;
1166         struct recv_thread_args *args;
1167         int i;
1168         int err;
1169
1170         sock = nbd_get_socket(nbd, arg, &err);
1171         if (!sock)
1172                 return err;
1173
1174         args = kzalloc(sizeof(*args), GFP_KERNEL);
1175         if (!args) {
1176                 sockfd_put(sock);
1177                 return -ENOMEM;
1178         }
1179
1180         for (i = 0; i < config->num_connections; i++) {
1181                 struct nbd_sock *nsock = config->socks[i];
1182
1183                 if (!nsock->dead)
1184                         continue;
1185
1186                 mutex_lock(&nsock->tx_lock);
1187                 if (!nsock->dead) {
1188                         mutex_unlock(&nsock->tx_lock);
1189                         continue;
1190                 }
1191                 sk_set_memalloc(sock->sk);
1192                 if (nbd->tag_set.timeout)
1193                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1194                 atomic_inc(&config->recv_threads);
1195                 refcount_inc(&nbd->config_refs);
1196                 old = nsock->sock;
1197                 nsock->fallback_index = -1;
1198                 nsock->sock = sock;
1199                 nsock->dead = false;
1200                 INIT_WORK(&args->work, recv_work);
1201                 args->index = i;
1202                 args->nbd = nbd;
1203                 nsock->cookie++;
1204                 mutex_unlock(&nsock->tx_lock);
1205                 sockfd_put(old);
1206
1207                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1208
1209                 /* We take the tx_mutex in an error path in the recv_work, so we
1210                  * need to queue_work outside of the tx_mutex.
1211                  */
1212                 queue_work(nbd->recv_workq, &args->work);
1213
1214                 atomic_inc(&config->live_connections);
1215                 wake_up(&config->conn_wait);
1216                 return 0;
1217         }
1218         sockfd_put(sock);
1219         kfree(args);
1220         return -ENOSPC;
1221 }
1222
1223 static void nbd_bdev_reset(struct nbd_device *nbd)
1224 {
1225         if (disk_openers(nbd->disk) > 1)
1226                 return;
1227         set_capacity(nbd->disk, 0);
1228 }
1229
1230 static void nbd_parse_flags(struct nbd_device *nbd)
1231 {
1232         struct nbd_config *config = nbd->config;
1233         if (config->flags & NBD_FLAG_READ_ONLY)
1234                 set_disk_ro(nbd->disk, true);
1235         else
1236                 set_disk_ro(nbd->disk, false);
1237         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1238                 if (config->flags & NBD_FLAG_SEND_FUA)
1239                         blk_queue_write_cache(nbd->disk->queue, true, true);
1240                 else
1241                         blk_queue_write_cache(nbd->disk->queue, true, false);
1242         }
1243         else
1244                 blk_queue_write_cache(nbd->disk->queue, false, false);
1245 }
1246
1247 static void send_disconnects(struct nbd_device *nbd)
1248 {
1249         struct nbd_config *config = nbd->config;
1250         struct nbd_request request = {
1251                 .magic = htonl(NBD_REQUEST_MAGIC),
1252                 .type = htonl(NBD_CMD_DISC),
1253         };
1254         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1255         struct iov_iter from;
1256         int i, ret;
1257
1258         for (i = 0; i < config->num_connections; i++) {
1259                 struct nbd_sock *nsock = config->socks[i];
1260
1261                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1262                 mutex_lock(&nsock->tx_lock);
1263                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1264                 if (ret < 0)
1265                         dev_err(disk_to_dev(nbd->disk),
1266                                 "Send disconnect failed %d\n", ret);
1267                 mutex_unlock(&nsock->tx_lock);
1268         }
1269 }
1270
1271 static int nbd_disconnect(struct nbd_device *nbd)
1272 {
1273         struct nbd_config *config = nbd->config;
1274
1275         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1276         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1277         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1278         send_disconnects(nbd);
1279         return 0;
1280 }
1281
1282 static void nbd_clear_sock(struct nbd_device *nbd)
1283 {
1284         sock_shutdown(nbd);
1285         nbd_clear_que(nbd);
1286         nbd->task_setup = NULL;
1287 }
1288
1289 static void nbd_config_put(struct nbd_device *nbd)
1290 {
1291         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1292                                         &nbd->config_lock)) {
1293                 struct nbd_config *config = nbd->config;
1294                 nbd_dev_dbg_close(nbd);
1295                 invalidate_disk(nbd->disk);
1296                 if (nbd->config->bytesize)
1297                         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
1298                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1299                                        &config->runtime_flags))
1300                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1301                 nbd->pid = 0;
1302                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1303                                        &config->runtime_flags)) {
1304                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1305                         kfree(nbd->backend);
1306                         nbd->backend = NULL;
1307                 }
1308                 nbd_clear_sock(nbd);
1309                 if (config->num_connections) {
1310                         int i;
1311                         for (i = 0; i < config->num_connections; i++) {
1312                                 sockfd_put(config->socks[i]->sock);
1313                                 kfree(config->socks[i]);
1314                         }
1315                         kfree(config->socks);
1316                 }
1317                 kfree(nbd->config);
1318                 nbd->config = NULL;
1319
1320                 nbd->tag_set.timeout = 0;
1321                 nbd->disk->queue->limits.discard_granularity = 0;
1322                 blk_queue_max_discard_sectors(nbd->disk->queue, 0);
1323
1324                 mutex_unlock(&nbd->config_lock);
1325                 nbd_put(nbd);
1326                 module_put(THIS_MODULE);
1327         }
1328 }
1329
1330 static int nbd_start_device(struct nbd_device *nbd)
1331 {
1332         struct nbd_config *config = nbd->config;
1333         int num_connections = config->num_connections;
1334         int error = 0, i;
1335
1336         if (nbd->pid)
1337                 return -EBUSY;
1338         if (!config->socks)
1339                 return -EINVAL;
1340         if (num_connections > 1 &&
1341             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1342                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1343                 return -EINVAL;
1344         }
1345
1346         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1347         nbd->pid = task_pid_nr(current);
1348
1349         nbd_parse_flags(nbd);
1350
1351         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1352         if (error) {
1353                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1354                 return error;
1355         }
1356         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1357
1358         nbd_dev_dbg_init(nbd);
1359         for (i = 0; i < num_connections; i++) {
1360                 struct recv_thread_args *args;
1361
1362                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1363                 if (!args) {
1364                         sock_shutdown(nbd);
1365                         /*
1366                          * If num_connections is m (2 < m),
1367                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1368                          * But NO.(n + 1) failed. We still have n recv threads.
1369                          * So, add flush_workqueue here to prevent recv threads
1370                          * dropping the last config_refs and trying to destroy
1371                          * the workqueue from inside the workqueue.
1372                          */
1373                         if (i)
1374                                 flush_workqueue(nbd->recv_workq);
1375                         return -ENOMEM;
1376                 }
1377                 sk_set_memalloc(config->socks[i]->sock->sk);
1378                 if (nbd->tag_set.timeout)
1379                         config->socks[i]->sock->sk->sk_sndtimeo =
1380                                 nbd->tag_set.timeout;
1381                 atomic_inc(&config->recv_threads);
1382                 refcount_inc(&nbd->config_refs);
1383                 INIT_WORK(&args->work, recv_work);
1384                 args->nbd = nbd;
1385                 args->index = i;
1386                 queue_work(nbd->recv_workq, &args->work);
1387         }
1388         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1389 }
1390
1391 static int nbd_start_device_ioctl(struct nbd_device *nbd)
1392 {
1393         struct nbd_config *config = nbd->config;
1394         int ret;
1395
1396         ret = nbd_start_device(nbd);
1397         if (ret)
1398                 return ret;
1399
1400         if (max_part)
1401                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1402         mutex_unlock(&nbd->config_lock);
1403         ret = wait_event_interruptible(config->recv_wq,
1404                                          atomic_read(&config->recv_threads) == 0);
1405         if (ret)
1406                 sock_shutdown(nbd);
1407         flush_workqueue(nbd->recv_workq);
1408
1409         mutex_lock(&nbd->config_lock);
1410         nbd_bdev_reset(nbd);
1411         /* user requested, ignore socket errors */
1412         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1413                 ret = 0;
1414         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1415                 ret = -ETIMEDOUT;
1416         return ret;
1417 }
1418
1419 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1420                                  struct block_device *bdev)
1421 {
1422         sock_shutdown(nbd);
1423         __invalidate_device(bdev, true);
1424         nbd_bdev_reset(nbd);
1425         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1426                                &nbd->config->runtime_flags))
1427                 nbd_config_put(nbd);
1428 }
1429
1430 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1431 {
1432         nbd->tag_set.timeout = timeout * HZ;
1433         if (timeout)
1434                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1435         else
1436                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1437 }
1438
1439 /* Must be called with config_lock held */
1440 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1441                        unsigned int cmd, unsigned long arg)
1442 {
1443         struct nbd_config *config = nbd->config;
1444         loff_t bytesize;
1445
1446         switch (cmd) {
1447         case NBD_DISCONNECT:
1448                 return nbd_disconnect(nbd);
1449         case NBD_CLEAR_SOCK:
1450                 nbd_clear_sock_ioctl(nbd, bdev);
1451                 return 0;
1452         case NBD_SET_SOCK:
1453                 return nbd_add_socket(nbd, arg, false);
1454         case NBD_SET_BLKSIZE:
1455                 return nbd_set_size(nbd, config->bytesize, arg);
1456         case NBD_SET_SIZE:
1457                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1458         case NBD_SET_SIZE_BLOCKS:
1459                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1460                         return -EINVAL;
1461                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1462         case NBD_SET_TIMEOUT:
1463                 nbd_set_cmd_timeout(nbd, arg);
1464                 return 0;
1465
1466         case NBD_SET_FLAGS:
1467                 config->flags = arg;
1468                 return 0;
1469         case NBD_DO_IT:
1470                 return nbd_start_device_ioctl(nbd);
1471         case NBD_CLEAR_QUE:
1472                 /*
1473                  * This is for compatibility only.  The queue is always cleared
1474                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1475                  */
1476                 return 0;
1477         case NBD_PRINT_DEBUG:
1478                 /*
1479                  * For compatibility only, we no longer keep a list of
1480                  * outstanding requests.
1481                  */
1482                 return 0;
1483         }
1484         return -ENOTTY;
1485 }
1486
1487 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1488                      unsigned int cmd, unsigned long arg)
1489 {
1490         struct nbd_device *nbd = bdev->bd_disk->private_data;
1491         struct nbd_config *config = nbd->config;
1492         int error = -EINVAL;
1493
1494         if (!capable(CAP_SYS_ADMIN))
1495                 return -EPERM;
1496
1497         /* The block layer will pass back some non-nbd ioctls in case we have
1498          * special handling for them, but we don't so just return an error.
1499          */
1500         if (_IOC_TYPE(cmd) != 0xab)
1501                 return -EINVAL;
1502
1503         mutex_lock(&nbd->config_lock);
1504
1505         /* Don't allow ioctl operations on a nbd device that was created with
1506          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1507          */
1508         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1509             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1510                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1511         else
1512                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1513         mutex_unlock(&nbd->config_lock);
1514         return error;
1515 }
1516
1517 static struct nbd_config *nbd_alloc_config(void)
1518 {
1519         struct nbd_config *config;
1520
1521         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1522         if (!config)
1523                 return NULL;
1524         atomic_set(&config->recv_threads, 0);
1525         init_waitqueue_head(&config->recv_wq);
1526         init_waitqueue_head(&config->conn_wait);
1527         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1528         atomic_set(&config->live_connections, 0);
1529         try_module_get(THIS_MODULE);
1530         return config;
1531 }
1532
1533 static int nbd_open(struct block_device *bdev, fmode_t mode)
1534 {
1535         struct nbd_device *nbd;
1536         int ret = 0;
1537
1538         mutex_lock(&nbd_index_mutex);
1539         nbd = bdev->bd_disk->private_data;
1540         if (!nbd) {
1541                 ret = -ENXIO;
1542                 goto out;
1543         }
1544         if (!refcount_inc_not_zero(&nbd->refs)) {
1545                 ret = -ENXIO;
1546                 goto out;
1547         }
1548         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1549                 struct nbd_config *config;
1550
1551                 mutex_lock(&nbd->config_lock);
1552                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1553                         mutex_unlock(&nbd->config_lock);
1554                         goto out;
1555                 }
1556                 config = nbd->config = nbd_alloc_config();
1557                 if (!config) {
1558                         ret = -ENOMEM;
1559                         mutex_unlock(&nbd->config_lock);
1560                         goto out;
1561                 }
1562                 refcount_set(&nbd->config_refs, 1);
1563                 refcount_inc(&nbd->refs);
1564                 mutex_unlock(&nbd->config_lock);
1565                 if (max_part)
1566                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1567         } else if (nbd_disconnected(nbd->config)) {
1568                 if (max_part)
1569                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1570         }
1571 out:
1572         mutex_unlock(&nbd_index_mutex);
1573         return ret;
1574 }
1575
1576 static void nbd_release(struct gendisk *disk, fmode_t mode)
1577 {
1578         struct nbd_device *nbd = disk->private_data;
1579
1580         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1581                         disk_openers(disk) == 0)
1582                 nbd_disconnect_and_put(nbd);
1583
1584         nbd_config_put(nbd);
1585         nbd_put(nbd);
1586 }
1587
1588 static const struct block_device_operations nbd_fops =
1589 {
1590         .owner =        THIS_MODULE,
1591         .open =         nbd_open,
1592         .release =      nbd_release,
1593         .ioctl =        nbd_ioctl,
1594         .compat_ioctl = nbd_ioctl,
1595 };
1596
1597 #if IS_ENABLED(CONFIG_DEBUG_FS)
1598
1599 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1600 {
1601         struct nbd_device *nbd = s->private;
1602
1603         if (nbd->pid)
1604                 seq_printf(s, "recv: %d\n", nbd->pid);
1605
1606         return 0;
1607 }
1608
1609 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1610
1611 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1612 {
1613         struct nbd_device *nbd = s->private;
1614         u32 flags = nbd->config->flags;
1615
1616         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1617
1618         seq_puts(s, "Known flags:\n");
1619
1620         if (flags & NBD_FLAG_HAS_FLAGS)
1621                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1622         if (flags & NBD_FLAG_READ_ONLY)
1623                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1624         if (flags & NBD_FLAG_SEND_FLUSH)
1625                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1626         if (flags & NBD_FLAG_SEND_FUA)
1627                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1628         if (flags & NBD_FLAG_SEND_TRIM)
1629                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1630
1631         return 0;
1632 }
1633
1634 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1635
1636 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1637 {
1638         struct dentry *dir;
1639         struct nbd_config *config = nbd->config;
1640
1641         if (!nbd_dbg_dir)
1642                 return -EIO;
1643
1644         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1645         if (!dir) {
1646                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1647                         nbd_name(nbd));
1648                 return -EIO;
1649         }
1650         config->dbg_dir = dir;
1651
1652         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1653         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1654         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1655         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1656         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1657
1658         return 0;
1659 }
1660
1661 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1662 {
1663         debugfs_remove_recursive(nbd->config->dbg_dir);
1664 }
1665
1666 static int nbd_dbg_init(void)
1667 {
1668         struct dentry *dbg_dir;
1669
1670         dbg_dir = debugfs_create_dir("nbd", NULL);
1671         if (!dbg_dir)
1672                 return -EIO;
1673
1674         nbd_dbg_dir = dbg_dir;
1675
1676         return 0;
1677 }
1678
1679 static void nbd_dbg_close(void)
1680 {
1681         debugfs_remove_recursive(nbd_dbg_dir);
1682 }
1683
1684 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1685
1686 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1687 {
1688         return 0;
1689 }
1690
1691 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1692 {
1693 }
1694
1695 static int nbd_dbg_init(void)
1696 {
1697         return 0;
1698 }
1699
1700 static void nbd_dbg_close(void)
1701 {
1702 }
1703
1704 #endif
1705
1706 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1707                             unsigned int hctx_idx, unsigned int numa_node)
1708 {
1709         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1710         cmd->nbd = set->driver_data;
1711         cmd->flags = 0;
1712         mutex_init(&cmd->lock);
1713         return 0;
1714 }
1715
1716 static const struct blk_mq_ops nbd_mq_ops = {
1717         .queue_rq       = nbd_queue_rq,
1718         .complete       = nbd_complete_rq,
1719         .init_request   = nbd_init_request,
1720         .timeout        = nbd_xmit_timeout,
1721 };
1722
1723 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1724 {
1725         struct nbd_device *nbd;
1726         struct gendisk *disk;
1727         int err = -ENOMEM;
1728
1729         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1730         if (!nbd)
1731                 goto out;
1732
1733         nbd->tag_set.ops = &nbd_mq_ops;
1734         nbd->tag_set.nr_hw_queues = 1;
1735         nbd->tag_set.queue_depth = 128;
1736         nbd->tag_set.numa_node = NUMA_NO_NODE;
1737         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1738         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1739                 BLK_MQ_F_BLOCKING;
1740         nbd->tag_set.driver_data = nbd;
1741         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1742         nbd->backend = NULL;
1743
1744         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1745         if (err)
1746                 goto out_free_nbd;
1747
1748         mutex_lock(&nbd_index_mutex);
1749         if (index >= 0) {
1750                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1751                                 GFP_KERNEL);
1752                 if (err == -ENOSPC)
1753                         err = -EEXIST;
1754         } else {
1755                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1756                 if (err >= 0)
1757                         index = err;
1758         }
1759         nbd->index = index;
1760         mutex_unlock(&nbd_index_mutex);
1761         if (err < 0)
1762                 goto out_free_tags;
1763
1764         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1765         if (IS_ERR(disk)) {
1766                 err = PTR_ERR(disk);
1767                 goto out_free_idr;
1768         }
1769         nbd->disk = disk;
1770
1771         nbd->recv_workq = alloc_workqueue("nbd%d-recv",
1772                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1773                                           WQ_UNBOUND, 0, nbd->index);
1774         if (!nbd->recv_workq) {
1775                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1776                 err = -ENOMEM;
1777                 goto out_err_disk;
1778         }
1779
1780         /*
1781          * Tell the block layer that we are not a rotational device
1782          */
1783         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1784         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1785         disk->queue->limits.discard_granularity = 0;
1786         blk_queue_max_discard_sectors(disk->queue, 0);
1787         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1788         blk_queue_max_segments(disk->queue, USHRT_MAX);
1789         blk_queue_max_hw_sectors(disk->queue, 65536);
1790         disk->queue->limits.max_sectors = 256;
1791
1792         mutex_init(&nbd->config_lock);
1793         refcount_set(&nbd->config_refs, 0);
1794         /*
1795          * Start out with a zero references to keep other threads from using
1796          * this device until it is fully initialized.
1797          */
1798         refcount_set(&nbd->refs, 0);
1799         INIT_LIST_HEAD(&nbd->list);
1800         disk->major = NBD_MAJOR;
1801
1802         /* Too big first_minor can cause duplicate creation of
1803          * sysfs files/links, since index << part_shift might overflow, or
1804          * MKDEV() expect that the max bits of first_minor is 20.
1805          */
1806         disk->first_minor = index << part_shift;
1807         if (disk->first_minor < index || disk->first_minor > MINORMASK) {
1808                 err = -EINVAL;
1809                 goto out_free_work;
1810         }
1811
1812         disk->minors = 1 << part_shift;
1813         disk->fops = &nbd_fops;
1814         disk->private_data = nbd;
1815         sprintf(disk->disk_name, "nbd%d", index);
1816         err = add_disk(disk);
1817         if (err)
1818                 goto out_free_work;
1819
1820         /*
1821          * Now publish the device.
1822          */
1823         refcount_set(&nbd->refs, refs);
1824         nbd_total_devices++;
1825         return nbd;
1826
1827 out_free_work:
1828         destroy_workqueue(nbd->recv_workq);
1829 out_err_disk:
1830         blk_cleanup_disk(disk);
1831 out_free_idr:
1832         mutex_lock(&nbd_index_mutex);
1833         idr_remove(&nbd_index_idr, index);
1834         mutex_unlock(&nbd_index_mutex);
1835 out_free_tags:
1836         blk_mq_free_tag_set(&nbd->tag_set);
1837 out_free_nbd:
1838         kfree(nbd);
1839 out:
1840         return ERR_PTR(err);
1841 }
1842
1843 static struct nbd_device *nbd_find_get_unused(void)
1844 {
1845         struct nbd_device *nbd;
1846         int id;
1847
1848         lockdep_assert_held(&nbd_index_mutex);
1849
1850         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1851                 if (refcount_read(&nbd->config_refs) ||
1852                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1853                         continue;
1854                 if (refcount_inc_not_zero(&nbd->refs))
1855                         return nbd;
1856         }
1857
1858         return NULL;
1859 }
1860
1861 /* Netlink interface. */
1862 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1863         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1864         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1865         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1866         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1867         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1868         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1869         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1870         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1871         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1872         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1873 };
1874
1875 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1876         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1877 };
1878
1879 /* We don't use this right now since we don't parse the incoming list, but we
1880  * still want it here so userspace knows what to expect.
1881  */
1882 static const struct nla_policy __attribute__((unused))
1883 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1884         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1885         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1886 };
1887
1888 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1889 {
1890         struct nbd_config *config = nbd->config;
1891         u64 bsize = nbd_blksize(config);
1892         u64 bytes = config->bytesize;
1893
1894         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1895                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1896
1897         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1898                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1899
1900         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1901                 return nbd_set_size(nbd, bytes, bsize);
1902         return 0;
1903 }
1904
1905 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1906 {
1907         struct nbd_device *nbd;
1908         struct nbd_config *config;
1909         int index = -1;
1910         int ret;
1911         bool put_dev = false;
1912
1913         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1914                 return -EPERM;
1915
1916         if (info->attrs[NBD_ATTR_INDEX])
1917                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1918         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1919                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1920                 return -EINVAL;
1921         }
1922         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1923                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1924                 return -EINVAL;
1925         }
1926 again:
1927         mutex_lock(&nbd_index_mutex);
1928         if (index == -1) {
1929                 nbd = nbd_find_get_unused();
1930         } else {
1931                 nbd = idr_find(&nbd_index_idr, index);
1932                 if (nbd) {
1933                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1934                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
1935                             !refcount_inc_not_zero(&nbd->refs)) {
1936                                 mutex_unlock(&nbd_index_mutex);
1937                                 pr_err("nbd: device at index %d is going down\n",
1938                                         index);
1939                                 return -EINVAL;
1940                         }
1941                 }
1942         }
1943         mutex_unlock(&nbd_index_mutex);
1944
1945         if (!nbd) {
1946                 nbd = nbd_dev_add(index, 2);
1947                 if (IS_ERR(nbd)) {
1948                         pr_err("nbd: failed to add new device\n");
1949                         return PTR_ERR(nbd);
1950                 }
1951         }
1952
1953         mutex_lock(&nbd->config_lock);
1954         if (refcount_read(&nbd->config_refs)) {
1955                 mutex_unlock(&nbd->config_lock);
1956                 nbd_put(nbd);
1957                 if (index == -1)
1958                         goto again;
1959                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1960                 return -EBUSY;
1961         }
1962         if (WARN_ON(nbd->config)) {
1963                 mutex_unlock(&nbd->config_lock);
1964                 nbd_put(nbd);
1965                 return -EINVAL;
1966         }
1967         config = nbd->config = nbd_alloc_config();
1968         if (!nbd->config) {
1969                 mutex_unlock(&nbd->config_lock);
1970                 nbd_put(nbd);
1971                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1972                 return -ENOMEM;
1973         }
1974         refcount_set(&nbd->config_refs, 1);
1975         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1976
1977         ret = nbd_genl_size_set(info, nbd);
1978         if (ret)
1979                 goto out;
1980
1981         if (info->attrs[NBD_ATTR_TIMEOUT])
1982                 nbd_set_cmd_timeout(nbd,
1983                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1984         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1985                 config->dead_conn_timeout =
1986                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1987                 config->dead_conn_timeout *= HZ;
1988         }
1989         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1990                 config->flags =
1991                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1992         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1993                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1994                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1995                         /*
1996                          * We have 1 ref to keep the device around, and then 1
1997                          * ref for our current operation here, which will be
1998                          * inherited by the config.  If we already have
1999                          * DESTROY_ON_DISCONNECT set then we know we don't have
2000                          * that extra ref already held so we don't need the
2001                          * put_dev.
2002                          */
2003                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2004                                               &nbd->flags))
2005                                 put_dev = true;
2006                 } else {
2007                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2008                                                &nbd->flags))
2009                                 refcount_inc(&nbd->refs);
2010                 }
2011                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2012                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2013                                 &config->runtime_flags);
2014                 }
2015         }
2016
2017         if (info->attrs[NBD_ATTR_SOCKETS]) {
2018                 struct nlattr *attr;
2019                 int rem, fd;
2020
2021                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2022                                     rem) {
2023                         struct nlattr *socks[NBD_SOCK_MAX+1];
2024
2025                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2026                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2027                                 ret = -EINVAL;
2028                                 goto out;
2029                         }
2030                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2031                                                           attr,
2032                                                           nbd_sock_policy,
2033                                                           info->extack);
2034                         if (ret != 0) {
2035                                 printk(KERN_ERR "nbd: error processing sock list\n");
2036                                 ret = -EINVAL;
2037                                 goto out;
2038                         }
2039                         if (!socks[NBD_SOCK_FD])
2040                                 continue;
2041                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2042                         ret = nbd_add_socket(nbd, fd, true);
2043                         if (ret)
2044                                 goto out;
2045                 }
2046         }
2047         ret = nbd_start_device(nbd);
2048         if (ret)
2049                 goto out;
2050         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2051                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2052                                           GFP_KERNEL);
2053                 if (!nbd->backend) {
2054                         ret = -ENOMEM;
2055                         goto out;
2056                 }
2057         }
2058         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2059         if (ret) {
2060                 dev_err(disk_to_dev(nbd->disk),
2061                         "device_create_file failed for backend!\n");
2062                 goto out;
2063         }
2064         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2065 out:
2066         mutex_unlock(&nbd->config_lock);
2067         if (!ret) {
2068                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2069                 refcount_inc(&nbd->config_refs);
2070                 nbd_connect_reply(info, nbd->index);
2071         }
2072         nbd_config_put(nbd);
2073         if (put_dev)
2074                 nbd_put(nbd);
2075         return ret;
2076 }
2077
2078 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2079 {
2080         mutex_lock(&nbd->config_lock);
2081         nbd_disconnect(nbd);
2082         sock_shutdown(nbd);
2083         wake_up(&nbd->config->conn_wait);
2084         /*
2085          * Make sure recv thread has finished, we can safely call nbd_clear_que()
2086          * to cancel the inflight I/Os.
2087          */
2088         flush_workqueue(nbd->recv_workq);
2089         nbd_clear_que(nbd);
2090         nbd->task_setup = NULL;
2091         mutex_unlock(&nbd->config_lock);
2092
2093         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2094                                &nbd->config->runtime_flags))
2095                 nbd_config_put(nbd);
2096 }
2097
2098 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2099 {
2100         struct nbd_device *nbd;
2101         int index;
2102
2103         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2104                 return -EPERM;
2105
2106         if (!info->attrs[NBD_ATTR_INDEX]) {
2107                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2108                 return -EINVAL;
2109         }
2110         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2111         mutex_lock(&nbd_index_mutex);
2112         nbd = idr_find(&nbd_index_idr, index);
2113         if (!nbd) {
2114                 mutex_unlock(&nbd_index_mutex);
2115                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2116                        index);
2117                 return -EINVAL;
2118         }
2119         if (!refcount_inc_not_zero(&nbd->refs)) {
2120                 mutex_unlock(&nbd_index_mutex);
2121                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2122                        index);
2123                 return -EINVAL;
2124         }
2125         mutex_unlock(&nbd_index_mutex);
2126         if (!refcount_inc_not_zero(&nbd->config_refs))
2127                 goto put_nbd;
2128         nbd_disconnect_and_put(nbd);
2129         nbd_config_put(nbd);
2130 put_nbd:
2131         nbd_put(nbd);
2132         return 0;
2133 }
2134
2135 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2136 {
2137         struct nbd_device *nbd = NULL;
2138         struct nbd_config *config;
2139         int index;
2140         int ret = 0;
2141         bool put_dev = false;
2142
2143         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2144                 return -EPERM;
2145
2146         if (!info->attrs[NBD_ATTR_INDEX]) {
2147                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2148                 return -EINVAL;
2149         }
2150         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2151         mutex_lock(&nbd_index_mutex);
2152         nbd = idr_find(&nbd_index_idr, index);
2153         if (!nbd) {
2154                 mutex_unlock(&nbd_index_mutex);
2155                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2156                        index);
2157                 return -EINVAL;
2158         }
2159         if (nbd->backend) {
2160                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2161                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2162                                        nbd->backend)) {
2163                                 mutex_unlock(&nbd_index_mutex);
2164                                 dev_err(nbd_to_dev(nbd),
2165                                         "backend image doesn't match with %s\n",
2166                                         nbd->backend);
2167                                 return -EINVAL;
2168                         }
2169                 } else {
2170                         mutex_unlock(&nbd_index_mutex);
2171                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2172                         return -EINVAL;
2173                 }
2174         }
2175         if (!refcount_inc_not_zero(&nbd->refs)) {
2176                 mutex_unlock(&nbd_index_mutex);
2177                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2178                        index);
2179                 return -EINVAL;
2180         }
2181         mutex_unlock(&nbd_index_mutex);
2182
2183         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2184                 dev_err(nbd_to_dev(nbd),
2185                         "not configured, cannot reconfigure\n");
2186                 nbd_put(nbd);
2187                 return -EINVAL;
2188         }
2189
2190         mutex_lock(&nbd->config_lock);
2191         config = nbd->config;
2192         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2193             !nbd->pid) {
2194                 dev_err(nbd_to_dev(nbd),
2195                         "not configured, cannot reconfigure\n");
2196                 ret = -EINVAL;
2197                 goto out;
2198         }
2199
2200         ret = nbd_genl_size_set(info, nbd);
2201         if (ret)
2202                 goto out;
2203
2204         if (info->attrs[NBD_ATTR_TIMEOUT])
2205                 nbd_set_cmd_timeout(nbd,
2206                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2207         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2208                 config->dead_conn_timeout =
2209                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2210                 config->dead_conn_timeout *= HZ;
2211         }
2212         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2213                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2214                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2215                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2216                                               &nbd->flags))
2217                                 put_dev = true;
2218                 } else {
2219                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2220                                                &nbd->flags))
2221                                 refcount_inc(&nbd->refs);
2222                 }
2223
2224                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2225                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2226                                         &config->runtime_flags);
2227                 } else {
2228                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2229                                         &config->runtime_flags);
2230                 }
2231         }
2232
2233         if (info->attrs[NBD_ATTR_SOCKETS]) {
2234                 struct nlattr *attr;
2235                 int rem, fd;
2236
2237                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2238                                     rem) {
2239                         struct nlattr *socks[NBD_SOCK_MAX+1];
2240
2241                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2242                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2243                                 ret = -EINVAL;
2244                                 goto out;
2245                         }
2246                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2247                                                           attr,
2248                                                           nbd_sock_policy,
2249                                                           info->extack);
2250                         if (ret != 0) {
2251                                 printk(KERN_ERR "nbd: error processing sock list\n");
2252                                 ret = -EINVAL;
2253                                 goto out;
2254                         }
2255                         if (!socks[NBD_SOCK_FD])
2256                                 continue;
2257                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2258                         ret = nbd_reconnect_socket(nbd, fd);
2259                         if (ret) {
2260                                 if (ret == -ENOSPC)
2261                                         ret = 0;
2262                                 goto out;
2263                         }
2264                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2265                 }
2266         }
2267 out:
2268         mutex_unlock(&nbd->config_lock);
2269         nbd_config_put(nbd);
2270         nbd_put(nbd);
2271         if (put_dev)
2272                 nbd_put(nbd);
2273         return ret;
2274 }
2275
2276 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2277         {
2278                 .cmd    = NBD_CMD_CONNECT,
2279                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2280                 .doit   = nbd_genl_connect,
2281         },
2282         {
2283                 .cmd    = NBD_CMD_DISCONNECT,
2284                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2285                 .doit   = nbd_genl_disconnect,
2286         },
2287         {
2288                 .cmd    = NBD_CMD_RECONFIGURE,
2289                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2290                 .doit   = nbd_genl_reconfigure,
2291         },
2292         {
2293                 .cmd    = NBD_CMD_STATUS,
2294                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2295                 .doit   = nbd_genl_status,
2296         },
2297 };
2298
2299 static const struct genl_multicast_group nbd_mcast_grps[] = {
2300         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2301 };
2302
2303 static struct genl_family nbd_genl_family __ro_after_init = {
2304         .hdrsize        = 0,
2305         .name           = NBD_GENL_FAMILY_NAME,
2306         .version        = NBD_GENL_VERSION,
2307         .module         = THIS_MODULE,
2308         .small_ops      = nbd_connect_genl_ops,
2309         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2310         .maxattr        = NBD_ATTR_MAX,
2311         .policy = nbd_attr_policy,
2312         .mcgrps         = nbd_mcast_grps,
2313         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2314 };
2315
2316 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2317 {
2318         struct nlattr *dev_opt;
2319         u8 connected = 0;
2320         int ret;
2321
2322         /* This is a little racey, but for status it's ok.  The
2323          * reason we don't take a ref here is because we can't
2324          * take a ref in the index == -1 case as we would need
2325          * to put under the nbd_index_mutex, which could
2326          * deadlock if we are configured to remove ourselves
2327          * once we're disconnected.
2328          */
2329         if (refcount_read(&nbd->config_refs))
2330                 connected = 1;
2331         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2332         if (!dev_opt)
2333                 return -EMSGSIZE;
2334         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2335         if (ret)
2336                 return -EMSGSIZE;
2337         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2338                          connected);
2339         if (ret)
2340                 return -EMSGSIZE;
2341         nla_nest_end(reply, dev_opt);
2342         return 0;
2343 }
2344
2345 static int status_cb(int id, void *ptr, void *data)
2346 {
2347         struct nbd_device *nbd = ptr;
2348         return populate_nbd_status(nbd, (struct sk_buff *)data);
2349 }
2350
2351 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2352 {
2353         struct nlattr *dev_list;
2354         struct sk_buff *reply;
2355         void *reply_head;
2356         size_t msg_size;
2357         int index = -1;
2358         int ret = -ENOMEM;
2359
2360         if (info->attrs[NBD_ATTR_INDEX])
2361                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2362
2363         mutex_lock(&nbd_index_mutex);
2364
2365         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2366                                   nla_attr_size(sizeof(u8)));
2367         msg_size *= (index == -1) ? nbd_total_devices : 1;
2368
2369         reply = genlmsg_new(msg_size, GFP_KERNEL);
2370         if (!reply)
2371                 goto out;
2372         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2373                                        NBD_CMD_STATUS);
2374         if (!reply_head) {
2375                 nlmsg_free(reply);
2376                 goto out;
2377         }
2378
2379         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2380         if (index == -1) {
2381                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2382                 if (ret) {
2383                         nlmsg_free(reply);
2384                         goto out;
2385                 }
2386         } else {
2387                 struct nbd_device *nbd;
2388                 nbd = idr_find(&nbd_index_idr, index);
2389                 if (nbd) {
2390                         ret = populate_nbd_status(nbd, reply);
2391                         if (ret) {
2392                                 nlmsg_free(reply);
2393                                 goto out;
2394                         }
2395                 }
2396         }
2397         nla_nest_end(reply, dev_list);
2398         genlmsg_end(reply, reply_head);
2399         ret = genlmsg_reply(reply, info);
2400 out:
2401         mutex_unlock(&nbd_index_mutex);
2402         return ret;
2403 }
2404
2405 static void nbd_connect_reply(struct genl_info *info, int index)
2406 {
2407         struct sk_buff *skb;
2408         void *msg_head;
2409         int ret;
2410
2411         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2412         if (!skb)
2413                 return;
2414         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2415                                      NBD_CMD_CONNECT);
2416         if (!msg_head) {
2417                 nlmsg_free(skb);
2418                 return;
2419         }
2420         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2421         if (ret) {
2422                 nlmsg_free(skb);
2423                 return;
2424         }
2425         genlmsg_end(skb, msg_head);
2426         genlmsg_reply(skb, info);
2427 }
2428
2429 static void nbd_mcast_index(int index)
2430 {
2431         struct sk_buff *skb;
2432         void *msg_head;
2433         int ret;
2434
2435         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2436         if (!skb)
2437                 return;
2438         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2439                                      NBD_CMD_LINK_DEAD);
2440         if (!msg_head) {
2441                 nlmsg_free(skb);
2442                 return;
2443         }
2444         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2445         if (ret) {
2446                 nlmsg_free(skb);
2447                 return;
2448         }
2449         genlmsg_end(skb, msg_head);
2450         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2451 }
2452
2453 static void nbd_dead_link_work(struct work_struct *work)
2454 {
2455         struct link_dead_args *args = container_of(work, struct link_dead_args,
2456                                                    work);
2457         nbd_mcast_index(args->index);
2458         kfree(args);
2459 }
2460
2461 static int __init nbd_init(void)
2462 {
2463         int i;
2464
2465         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2466
2467         if (max_part < 0) {
2468                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2469                 return -EINVAL;
2470         }
2471
2472         part_shift = 0;
2473         if (max_part > 0) {
2474                 part_shift = fls(max_part);
2475
2476                 /*
2477                  * Adjust max_part according to part_shift as it is exported
2478                  * to user space so that user can know the max number of
2479                  * partition kernel should be able to manage.
2480                  *
2481                  * Note that -1 is required because partition 0 is reserved
2482                  * for the whole disk.
2483                  */
2484                 max_part = (1UL << part_shift) - 1;
2485         }
2486
2487         if ((1UL << part_shift) > DISK_MAX_PARTS)
2488                 return -EINVAL;
2489
2490         if (nbds_max > 1UL << (MINORBITS - part_shift))
2491                 return -EINVAL;
2492
2493         if (register_blkdev(NBD_MAJOR, "nbd"))
2494                 return -EIO;
2495
2496         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2497         if (!nbd_del_wq) {
2498                 unregister_blkdev(NBD_MAJOR, "nbd");
2499                 return -ENOMEM;
2500         }
2501
2502         if (genl_register_family(&nbd_genl_family)) {
2503                 destroy_workqueue(nbd_del_wq);
2504                 unregister_blkdev(NBD_MAJOR, "nbd");
2505                 return -EINVAL;
2506         }
2507         nbd_dbg_init();
2508
2509         for (i = 0; i < nbds_max; i++)
2510                 nbd_dev_add(i, 1);
2511         return 0;
2512 }
2513
2514 static int nbd_exit_cb(int id, void *ptr, void *data)
2515 {
2516         struct list_head *list = (struct list_head *)data;
2517         struct nbd_device *nbd = ptr;
2518
2519         /* Skip nbd that is being removed asynchronously */
2520         if (refcount_read(&nbd->refs))
2521                 list_add_tail(&nbd->list, list);
2522
2523         return 0;
2524 }
2525
2526 static void __exit nbd_cleanup(void)
2527 {
2528         struct nbd_device *nbd;
2529         LIST_HEAD(del_list);
2530
2531         nbd_dbg_close();
2532
2533         mutex_lock(&nbd_index_mutex);
2534         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2535         mutex_unlock(&nbd_index_mutex);
2536
2537         while (!list_empty(&del_list)) {
2538                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2539                 list_del_init(&nbd->list);
2540                 if (refcount_read(&nbd->refs) != 1)
2541                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2542                 nbd_put(nbd);
2543         }
2544
2545         /* Also wait for nbd_dev_remove_work() completes */
2546         destroy_workqueue(nbd_del_wq);
2547
2548         idr_destroy(&nbd_index_idr);
2549         genl_unregister_family(&nbd_genl_family);
2550         unregister_blkdev(NBD_MAJOR, "nbd");
2551 }
2552
2553 module_init(nbd_init);
2554 module_exit(nbd_cleanup);
2555
2556 MODULE_DESCRIPTION("Network Block Device");
2557 MODULE_LICENSE("GPL");
2558
2559 module_param(nbds_max, int, 0444);
2560 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2561 module_param(max_part, int, 0444);
2562 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");