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 <[email protected]>
   9  * Parts copyright 2001 Steven Whitehouse <[email protected]>
  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
 264         kfree(nbd);
 265 }
 266
 267 static void nbd_dev_remove_work(struct work_struct *work)
 268 {
 269         nbd_dev_remove(container_of(work, struct nbd_device, remove_work));
 270 }
 271
 272 static void nbd_put(struct nbd_device *nbd)
 273 {
 274         if (!refcount_dec_and_test(&nbd->refs))
 275                 return;
 276
 277         /* Call del_gendisk() asynchrounously to prevent deadlock */
 278         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
 279                 queue_work(nbd_del_wq, &nbd->remove_work);
 280         else
 281                 nbd_dev_remove(nbd);
 282 }
 283
 284 static int nbd_disconnected(struct nbd_config *config)
 285 {
 286         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 287                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 288 }
 289
 290 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 291                                 int notify)
 292 {
 293         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 294                 struct link_dead_args *args;
 295                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 296                 if (args) {
 297                         INIT_WORK(&args->work, nbd_dead_link_work);
 298                         args->index = nbd->index;
 299                         queue_work(system_wq, &args->work);
 300                 }
 301         }
 302         if (!nsock->dead) {
 303                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 304                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 305                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 306                                                &nbd->config->runtime_flags)) {
 307                                 set_bit(NBD_RT_DISCONNECTED,
 308                                         &nbd->config->runtime_flags);
 309                                 dev_info(nbd_to_dev(nbd),
 310                                         "Disconnected due to user request.\n");
 311                         }
 312                 }
 313         }
 314         nsock->dead = true;
 315         nsock->pending = NULL;
 316         nsock->sent = 0;
 317 }
 318
 319 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 320                 loff_t blksize)
 321 {
 322         if (!blksize)
 323                 blksize = 1u << NBD_DEF_BLKSIZE_BITS;
 324
 325         if (blk_validate_block_size(blksize))
 326                 return -EINVAL;
 327
 328         nbd->config->bytesize = bytesize;
 329         nbd->config->blksize_bits = __ffs(blksize);
 330
 331         if (!nbd->pid)
 332                 return 0;
 333
 334         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 335                 nbd->disk->queue->limits.discard_granularity = blksize;
 336                 nbd->disk->queue->limits.discard_alignment = blksize;
 337                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 338         }
 339         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 340         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 341
 342         if (max_part)
 343                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 344         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 345                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 346         return 0;
 347 }
 348
 349 static void nbd_complete_rq(struct request *req)
 350 {
 351         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 352
 353         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 354                 cmd->status ? "failed" : "done");
 355
 356         blk_mq_end_request(req, cmd->status);
 357 }
 358
 359 /*
 360  * Forcibly shutdown the socket causing all listeners to error
 361  */
 362 static void sock_shutdown(struct nbd_device *nbd)
 363 {
 364         struct nbd_config *config = nbd->config;
 365         int i;
 366
 367         if (config->num_connections == 0)
 368                 return;
 369         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 370                 return;
 371
 372         for (i = 0; i < config->num_connections; i++) {
 373                 struct nbd_sock *nsock = config->socks[i];
 374                 mutex_lock(&nsock->tx_lock);
 375                 nbd_mark_nsock_dead(nbd, nsock, 0);
 376                 mutex_unlock(&nsock->tx_lock);
 377         }
 378         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 379 }
 380
 381 static u32 req_to_nbd_cmd_type(struct request *req)
 382 {
 383         switch (req_op(req)) {
 384         case REQ_OP_DISCARD:
 385                 return NBD_CMD_TRIM;
 386         case REQ_OP_FLUSH:
 387                 return NBD_CMD_FLUSH;
 388         case REQ_OP_WRITE:
 389                 return NBD_CMD_WRITE;
 390         case REQ_OP_READ:
 391                 return NBD_CMD_READ;
 392         default:
 393                 return U32_MAX;
 394         }
 395 }
 396
 397 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 398                                                  bool reserved)
 399 {
 400         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 401         struct nbd_device *nbd = cmd->nbd;
 402         struct nbd_config *config;
 403
 404         if (!mutex_trylock(&cmd->lock))
 405                 return BLK_EH_RESET_TIMER;
 406
 407         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 408                 mutex_unlock(&cmd->lock);
 409                 return BLK_EH_DONE;
 410         }
 411
 412         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 413                 cmd->status = BLK_STS_TIMEOUT;
 414                 mutex_unlock(&cmd->lock);
 415                 goto done;
 416         }
 417         config = nbd->config;
 418
 419         if (config->num_connections > 1 ||
 420             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 421                 dev_err_ratelimited(nbd_to_dev(nbd),
 422                                     "Connection timed out, retrying (%d/%d alive)\n",
 423                                     atomic_read(&config->live_connections),
 424                                     config->num_connections);
 425                 /*
 426                  * Hooray we have more connections, requeue this IO, the submit
 427                  * path will put it on a real connection. Or if only one
 428                  * connection is configured, the submit path will wait util
 429                  * a new connection is reconfigured or util dead timeout.
 430                  */
 431                 if (config->socks) {
 432                         if (cmd->index < config->num_connections) {
 433                                 struct nbd_sock *nsock =
 434                                         config->socks[cmd->index];
 435                                 mutex_lock(&nsock->tx_lock);
 436                                 /* We can have multiple outstanding requests, so
 437                                  * we don't want to mark the nsock dead if we've
 438                                  * already reconnected with a new socket, so
 439                                  * only mark it dead if its the same socket we
 440                                  * were sent out on.
 441                                  */
 442                                 if (cmd->cookie == nsock->cookie)
 443                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 444                                 mutex_unlock(&nsock->tx_lock);
 445                         }
 446                         mutex_unlock(&cmd->lock);
 447                         nbd_requeue_cmd(cmd);
 448                         nbd_config_put(nbd);
 449                         return BLK_EH_DONE;
 450                 }
 451         }
 452
 453         if (!nbd->tag_set.timeout) {
 454                 /*
 455                  * Userspace sets timeout=0 to disable socket disconnection,
 456                  * so just warn and reset the timer.
 457                  */
 458                 struct nbd_sock *nsock = config->socks[cmd->index];
 459                 cmd->retries++;
 460                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 461                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 462                         (unsigned long long)blk_rq_pos(req) << 9,
 463                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 464
 465                 mutex_lock(&nsock->tx_lock);
 466                 if (cmd->cookie != nsock->cookie) {
 467                         nbd_requeue_cmd(cmd);
 468                         mutex_unlock(&nsock->tx_lock);
 469                         mutex_unlock(&cmd->lock);
 470                         nbd_config_put(nbd);
 471                         return BLK_EH_DONE;
 472                 }
 473                 mutex_unlock(&nsock->tx_lock);
 474                 mutex_unlock(&cmd->lock);
 475                 nbd_config_put(nbd);
 476                 return BLK_EH_RESET_TIMER;
 477         }
 478
 479         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 480         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 481         cmd->status = BLK_STS_IOERR;
 482         mutex_unlock(&cmd->lock);
 483         sock_shutdown(nbd);
 484         nbd_config_put(nbd);
 485 done:
 486         blk_mq_complete_request(req);
 487         return BLK_EH_DONE;
 488 }
 489
 490 /*
 491  *  Send or receive packet. Return a positive value on success and
 492  *  negtive value on failue, and never return 0.
 493  */
 494 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 495                      struct iov_iter *iter, int msg_flags, int *sent)
 496 {
 497         struct nbd_config *config = nbd->config;
 498         struct socket *sock = config->socks[index]->sock;
 499         int result;
 500         struct msghdr msg;
 501         unsigned int noreclaim_flag;
 502
 503         if (unlikely(!sock)) {
 504                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 505                         "Attempted %s on closed socket in sock_xmit\n",
 506                         (send ? "send" : "recv"));
 507                 return -EINVAL;
 508         }
 509
 510         msg.msg_iter = *iter;
 511
 512         noreclaim_flag = memalloc_noreclaim_save();
 513         do {
 514                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 515                 msg.msg_name = NULL;
 516                 msg.msg_namelen = 0;
 517                 msg.msg_control = NULL;
 518                 msg.msg_controllen = 0;
 519                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 520
 521                 if (send)
 522                         result = sock_sendmsg(sock, &msg);
 523                 else
 524                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 525
 526                 if (result <= 0) {
 527                         if (result == 0)
 528                                 result = -EPIPE; /* short read */
 529                         break;
 530                 }
 531                 if (sent)
 532                         *sent += result;
 533         } while (msg_data_left(&msg));
 534
 535         memalloc_noreclaim_restore(noreclaim_flag);
 536
 537         return result;
 538 }
 539
 540 /*
 541  * Different settings for sk->sk_sndtimeo can result in different return values
 542  * if there is a signal pending when we enter sendmsg, because reasons?
 543  */
 544 static inline int was_interrupted(int result)
 545 {
 546         return result == -ERESTARTSYS || result == -EINTR;
 547 }
 548
 549 /* always call with the tx_lock held */
 550 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 551 {
 552         struct request *req = blk_mq_rq_from_pdu(cmd);
 553         struct nbd_config *config = nbd->config;
 554         struct nbd_sock *nsock = config->socks[index];
 555         int result;
 556         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 557         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 558         struct iov_iter from;
 559         unsigned long size = blk_rq_bytes(req);
 560         struct bio *bio;
 561         u64 handle;
 562         u32 type;
 563         u32 nbd_cmd_flags = 0;
 564         int sent = nsock->sent, skip = 0;
 565
 566         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 567
 568         type = req_to_nbd_cmd_type(req);
 569         if (type == U32_MAX)
 570                 return -EIO;
 571
 572         if (rq_data_dir(req) == WRITE &&
 573             (config->flags & NBD_FLAG_READ_ONLY)) {
 574                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 575                                     "Write on read-only\n");
 576                 return -EIO;
 577         }
 578
 579         if (req->cmd_flags & REQ_FUA)
 580                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 581
 582         /* We did a partial send previously, and we at least sent the whole
 583          * request struct, so just go and send the rest of the pages in the
 584          * request.
 585          */
 586         if (sent) {
 587                 if (sent >= sizeof(request)) {
 588                         skip = sent - sizeof(request);
 589
 590                         /* initialize handle for tracing purposes */
 591                         handle = nbd_cmd_handle(cmd);
 592
 593                         goto send_pages;
 594                 }
 595                 iov_iter_advance(&from, sent);
 596         } else {
 597                 cmd->cmd_cookie++;
 598         }
 599         cmd->index = index;
 600         cmd->cookie = nsock->cookie;
 601         cmd->retries = 0;
 602         request.type = htonl(type | nbd_cmd_flags);
 603         if (type != NBD_CMD_FLUSH) {
 604                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 605                 request.len = htonl(size);
 606         }
 607         handle = nbd_cmd_handle(cmd);
 608         memcpy(request.handle, &handle, sizeof(handle));
 609
 610         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 611
 612         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 613                 req, nbdcmd_to_ascii(type),
 614                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 615         result = sock_xmit(nbd, index, 1, &from,
 616                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 617         trace_nbd_header_sent(req, handle);
 618         if (result < 0) {
 619                 if (was_interrupted(result)) {
 620                         /* If we havne't sent anything we can just return BUSY,
 621                          * however if we have sent something we need to make
 622                          * sure we only allow this req to be sent until we are
 623                          * completely done.
 624                          */
 625                         if (sent) {
 626                                 nsock->pending = req;
 627                                 nsock->sent = sent;
 628                         }
 629                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 630                         return BLK_STS_RESOURCE;
 631                 }
 632                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 633                         "Send control failed (result %d)\n", result);
 634                 return -EAGAIN;
 635         }
 636 send_pages:
 637         if (type != NBD_CMD_WRITE)
 638                 goto out;
 639
 640         bio = req->bio;
 641         while (bio) {
 642                 struct bio *next = bio->bi_next;
 643                 struct bvec_iter iter;
 644                 struct bio_vec bvec;
 645
 646                 bio_for_each_segment(bvec, bio, iter) {
 647                         bool is_last = !next && bio_iter_last(bvec, iter);
 648                         int flags = is_last ? 0 : MSG_MORE;
 649
 650                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 651                                 req, bvec.bv_len);
 652                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 653                         if (skip) {
 654                                 if (skip >= iov_iter_count(&from)) {
 655                                         skip -= iov_iter_count(&from);
 656                                         continue;
 657                                 }
 658                                 iov_iter_advance(&from, skip);
 659                                 skip = 0;
 660                         }
 661                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 662                         if (result < 0) {
 663                                 if (was_interrupted(result)) {
 664                                         /* We've already sent the header, we
 665                                          * have no choice but to set pending and
 666                                          * return BUSY.
 667                                          */
 668                                         nsock->pending = req;
 669                                         nsock->sent = sent;
 670                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 671                                         return BLK_STS_RESOURCE;
 672                                 }
 673                                 dev_err(disk_to_dev(nbd->disk),
 674                                         "Send data failed (result %d)\n",
 675                                         result);
 676                                 return -EAGAIN;
 677                         }
 678                         /*
 679                          * The completion might already have come in,
 680                          * so break for the last one instead of letting
 681                          * the iterator do it. This prevents use-after-free
 682                          * of the bio.
 683                          */
 684                         if (is_last)
 685                                 break;
 686                 }
 687                 bio = next;
 688         }
 689 out:
 690         trace_nbd_payload_sent(req, handle);
 691         nsock->pending = NULL;
 692         nsock->sent = 0;
 693         return 0;
 694 }
 695
 696 static int nbd_read_reply(struct nbd_device *nbd, int index,
 697                           struct nbd_reply *reply)
 698 {
 699         struct kvec iov = {.iov_base = reply, .iov_len = sizeof(*reply)};
 700         struct iov_iter to;
 701         int result;
 702
 703         reply->magic = 0;
 704         iov_iter_kvec(&to, READ, &iov, 1, sizeof(*reply));
 705         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 706         if (result < 0) {
 707                 if (!nbd_disconnected(nbd->config))
 708                         dev_err(disk_to_dev(nbd->disk),
 709                                 "Receive control failed (result %d)\n", result);
 710                 return result;
 711         }
 712
 713         if (ntohl(reply->magic) != NBD_REPLY_MAGIC) {
 714                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 715                                 (unsigned long)ntohl(reply->magic));
 716                 return -EPROTO;
 717         }
 718
 719         return 0;
 720 }
 721
 722 /* NULL returned = something went wrong, inform userspace */
 723 static struct nbd_cmd *nbd_handle_reply(struct nbd_device *nbd, int index,
 724                                         struct nbd_reply *reply)
 725 {
 726         int result;
 727         struct nbd_cmd *cmd;
 728         struct request *req = NULL;
 729         u64 handle;
 730         u16 hwq;
 731         u32 tag;
 732         int ret = 0;
 733
 734         memcpy(&handle, reply->handle, sizeof(handle));
 735         tag = nbd_handle_to_tag(handle);
 736         hwq = blk_mq_unique_tag_to_hwq(tag);
 737         if (hwq < nbd->tag_set.nr_hw_queues)
 738                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 739                                        blk_mq_unique_tag_to_tag(tag));
 740         if (!req || !blk_mq_request_started(req)) {
 741                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 742                         tag, req);
 743                 return ERR_PTR(-ENOENT);
 744         }
 745         trace_nbd_header_received(req, handle);
 746         cmd = blk_mq_rq_to_pdu(req);
 747
 748         mutex_lock(&cmd->lock);
 749         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 750                 dev_err(disk_to_dev(nbd->disk), "Suspicious reply %d (status %u flags %lu)",
 751                         tag, cmd->status, cmd->flags);
 752                 ret = -ENOENT;
 753                 goto out;
 754         }
 755         if (cmd->index != index) {
 756                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply %d from different sock %d (expected %d)",
 757                         tag, index, cmd->index);
 758         }
 759         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 760                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 761                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 762                 ret = -ENOENT;
 763                 goto out;
 764         }
 765         if (cmd->status != BLK_STS_OK) {
 766                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 767                         req);
 768                 ret = -ENOENT;
 769                 goto out;
 770         }
 771         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 772                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 773                         req);
 774                 ret = -ENOENT;
 775                 goto out;
 776         }
 777         if (ntohl(reply->error)) {
 778                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 779                         ntohl(reply->error));
 780                 cmd->status = BLK_STS_IOERR;
 781                 goto out;
 782         }
 783
 784         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 785         if (rq_data_dir(req) != WRITE) {
 786                 struct req_iterator iter;
 787                 struct bio_vec bvec;
 788                 struct iov_iter to;
 789
 790                 rq_for_each_segment(bvec, req, iter) {
 791                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 792                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 793                         if (result < 0) {
 794                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 795                                         result);
 796                                 /*
 797                                  * If we've disconnected, we need to make sure we
 798                                  * complete this request, otherwise error out
 799                                  * and let the timeout stuff handle resubmitting
 800                                  * this request onto another connection.
 801                                  */
 802                                 if (nbd_disconnected(nbd->config)) {
 803                                         cmd->status = BLK_STS_IOERR;
 804                                         goto out;
 805                                 }
 806                                 ret = -EIO;
 807                                 goto out;
 808                         }
 809                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 810                                 req, bvec.bv_len);
 811                 }
 812         }
 813 out:
 814         trace_nbd_payload_received(req, handle);
 815         mutex_unlock(&cmd->lock);
 816         return ret ? ERR_PTR(ret) : cmd;
 817 }
 818
 819 static void recv_work(struct work_struct *work)
 820 {
 821         struct recv_thread_args *args = container_of(work,
 822                                                      struct recv_thread_args,
 823                                                      work);
 824         struct nbd_device *nbd = args->nbd;
 825         struct nbd_config *config = nbd->config;
 826         struct request_queue *q = nbd->disk->queue;
 827         struct nbd_sock *nsock;
 828         struct nbd_cmd *cmd;
 829         struct request *rq;
 830
 831         while (1) {
 832                 struct nbd_reply reply;
 833
 834                 if (nbd_read_reply(nbd, args->index, &reply))
 835                         break;
 836
 837                 /*
 838                  * Grab .q_usage_counter so request pool won't go away, then no
 839                  * request use-after-free is possible during nbd_handle_reply().
 840                  * If queue is frozen, there won't be any inflight requests, we
 841                  * needn't to handle the incoming garbage message.
 842                  */
 843                 if (!percpu_ref_tryget(&q->q_usage_counter)) {
 844                         dev_err(disk_to_dev(nbd->disk), "%s: no io inflight\n",
 845                                 __func__);
 846                         break;
 847                 }
 848
 849                 cmd = nbd_handle_reply(nbd, args->index, &reply);
 850                 if (IS_ERR(cmd)) {
 851                         percpu_ref_put(&q->q_usage_counter);
 852                         break;
 853                 }
 854
 855                 rq = blk_mq_rq_from_pdu(cmd);
 856                 if (likely(!blk_should_fake_timeout(rq->q)))
 857                         blk_mq_complete_request(rq);
 858                 percpu_ref_put(&q->q_usage_counter);
 859         }
 860
 861         nsock = config->socks[args->index];
 862         mutex_lock(&nsock->tx_lock);
 863         nbd_mark_nsock_dead(nbd, nsock, 1);
 864         mutex_unlock(&nsock->tx_lock);
 865
 866         nbd_config_put(nbd);
 867         atomic_dec(&config->recv_threads);
 868         wake_up(&config->recv_wq);
 869         kfree(args);
 870 }
 871
 872 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 873 {
 874         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 875
 876         /* don't abort one completed request */
 877         if (blk_mq_request_completed(req))
 878                 return true;
 879
 880         mutex_lock(&cmd->lock);
 881         if (!__test_and_clear_bit(NBD_CMD_INFLIGHT, &cmd->flags)) {
 882                 mutex_unlock(&cmd->lock);
 883                 return true;
 884         }
 885         cmd->status = BLK_STS_IOERR;
 886         mutex_unlock(&cmd->lock);
 887
 888         blk_mq_complete_request(req);
 889         return true;
 890 }
 891
 892 static void nbd_clear_que(struct nbd_device *nbd)
 893 {
 894         blk_mq_quiesce_queue(nbd->disk->queue);
 895         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 896         blk_mq_unquiesce_queue(nbd->disk->queue);
 897         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 898 }
 899
 900 static int find_fallback(struct nbd_device *nbd, int index)
 901 {
 902         struct nbd_config *config = nbd->config;
 903         int new_index = -1;
 904         struct nbd_sock *nsock = config->socks[index];
 905         int fallback = nsock->fallback_index;
 906
 907         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 908                 return new_index;
 909
 910         if (config->num_connections <= 1) {
 911                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 912                                     "Dead connection, failed to find a fallback\n");
 913                 return new_index;
 914         }
 915
 916         if (fallback >= 0 && fallback < config->num_connections &&
 917             !config->socks[fallback]->dead)
 918                 return fallback;
 919
 920         if (nsock->fallback_index < 0 ||
 921             nsock->fallback_index >= config->num_connections ||
 922             config->socks[nsock->fallback_index]->dead) {
 923                 int i;
 924                 for (i = 0; i < config->num_connections; i++) {
 925                         if (i == index)
 926                                 continue;
 927                         if (!config->socks[i]->dead) {
 928                                 new_index = i;
 929                                 break;
 930                         }
 931                 }
 932                 nsock->fallback_index = new_index;
 933                 if (new_index < 0) {
 934                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 935                                             "Dead connection, failed to find a fallback\n");
 936                         return new_index;
 937                 }
 938         }
 939         new_index = nsock->fallback_index;
 940         return new_index;
 941 }
 942
 943 static int wait_for_reconnect(struct nbd_device *nbd)
 944 {
 945         struct nbd_config *config = nbd->config;
 946         if (!config->dead_conn_timeout)
 947                 return 0;
 948         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 949                 return 0;
 950         return wait_event_timeout(config->conn_wait,
 951                                   atomic_read(&config->live_connections) > 0,
 952                                   config->dead_conn_timeout) > 0;
 953 }
 954
 955 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 956 {
 957         struct request *req = blk_mq_rq_from_pdu(cmd);
 958         struct nbd_device *nbd = cmd->nbd;
 959         struct nbd_config *config;
 960         struct nbd_sock *nsock;
 961         int ret;
 962
 963         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 964                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 965                                     "Socks array is empty\n");
 966                 return -EINVAL;
 967         }
 968         config = nbd->config;
 969
 970         if (index >= config->num_connections) {
 971                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 972                                     "Attempted send on invalid socket\n");
 973                 nbd_config_put(nbd);
 974                 return -EINVAL;
 975         }
 976         cmd->status = BLK_STS_OK;
 977 again:
 978         nsock = config->socks[index];
 979         mutex_lock(&nsock->tx_lock);
 980         if (nsock->dead) {
 981                 int old_index = index;
 982                 index = find_fallback(nbd, index);
 983                 mutex_unlock(&nsock->tx_lock);
 984                 if (index < 0) {
 985                         if (wait_for_reconnect(nbd)) {
 986                                 index = old_index;
 987                                 goto again;
 988                         }
 989                         /* All the sockets should already be down at this point,
 990                          * we just want to make sure that DISCONNECTED is set so
 991                          * any requests that come in that were queue'ed waiting
 992                          * for the reconnect timer don't trigger the timer again
 993                          * and instead just error out.
 994                          */
 995                         sock_shutdown(nbd);
 996                         nbd_config_put(nbd);
 997                         return -EIO;
 998                 }
 999                 goto again;
1000         }
1001
1002         /* Handle the case that we have a pending request that was partially
1003          * transmitted that _has_ to be serviced first.  We need to call requeue
1004          * here so that it gets put _after_ the request that is already on the
1005          * dispatch list.
1006          */
1007         blk_mq_start_request(req);
1008         if (unlikely(nsock->pending && nsock->pending != req)) {
1009                 nbd_requeue_cmd(cmd);
1010                 ret = 0;
1011                 goto out;
1012         }
1013         /*
1014          * Some failures are related to the link going down, so anything that
1015          * returns EAGAIN can be retried on a different socket.
1016          */
1017         ret = nbd_send_cmd(nbd, cmd, index);
1018         /*
1019          * Access to this flag is protected by cmd->lock, thus it's safe to set
1020          * the flag after nbd_send_cmd() succeed to send request to server.
1021          */
1022         if (!ret)
1023                 __set_bit(NBD_CMD_INFLIGHT, &cmd->flags);
1024         else if (ret == -EAGAIN) {
1025                 dev_err_ratelimited(disk_to_dev(nbd->disk),
1026                                     "Request send failed, requeueing\n");
1027                 nbd_mark_nsock_dead(nbd, nsock, 1);
1028                 nbd_requeue_cmd(cmd);
1029                 ret = 0;
1030         }
1031 out:
1032         mutex_unlock(&nsock->tx_lock);
1033         nbd_config_put(nbd);
1034         return ret;
1035 }
1036
1037 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1038                         const struct blk_mq_queue_data *bd)
1039 {
1040         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
1041         int ret;
1042
1043         /*
1044          * Since we look at the bio's to send the request over the network we
1045          * need to make sure the completion work doesn't mark this request done
1046          * before we are done doing our send.  This keeps us from dereferencing
1047          * freed data if we have particularly fast completions (ie we get the
1048          * completion before we exit sock_xmit on the last bvec) or in the case
1049          * that the server is misbehaving (or there was an error) before we're
1050          * done sending everything over the wire.
1051          */
1052         mutex_lock(&cmd->lock);
1053         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
1054
1055         /* We can be called directly from the user space process, which means we
1056          * could possibly have signals pending so our sendmsg will fail.  In
1057          * this case we need to return that we are busy, otherwise error out as
1058          * appropriate.
1059          */
1060         ret = nbd_handle_cmd(cmd, hctx->queue_num);
1061         if (ret < 0)
1062                 ret = BLK_STS_IOERR;
1063         else if (!ret)
1064                 ret = BLK_STS_OK;
1065         mutex_unlock(&cmd->lock);
1066
1067         return ret;
1068 }
1069
1070 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1071                                      int *err)
1072 {
1073         struct socket *sock;
1074
1075         *err = 0;
1076         sock = sockfd_lookup(fd, err);
1077         if (!sock)
1078                 return NULL;
1079
1080         if (sock->ops->shutdown == sock_no_shutdown) {
1081                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1082                 *err = -EINVAL;
1083                 sockfd_put(sock);
1084                 return NULL;
1085         }
1086
1087         return sock;
1088 }
1089
1090 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1091                           bool netlink)
1092 {
1093         struct nbd_config *config = nbd->config;
1094         struct socket *sock;
1095         struct nbd_sock **socks;
1096         struct nbd_sock *nsock;
1097         int err;
1098
1099         sock = nbd_get_socket(nbd, arg, &err);
1100         if (!sock)
1101                 return err;
1102
1103         /*
1104          * We need to make sure we don't get any errant requests while we're
1105          * reallocating the ->socks array.
1106          */
1107         blk_mq_freeze_queue(nbd->disk->queue);
1108
1109         if (!netlink && !nbd->task_setup &&
1110             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1111                 nbd->task_setup = current;
1112
1113         if (!netlink &&
1114             (nbd->task_setup != current ||
1115              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1116                 dev_err(disk_to_dev(nbd->disk),
1117                         "Device being setup by another task");
1118                 err = -EBUSY;
1119                 goto put_socket;
1120         }
1121
1122         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1123         if (!nsock) {
1124                 err = -ENOMEM;
1125                 goto put_socket;
1126         }
1127
1128         socks = krealloc(config->socks, (config->num_connections + 1) *
1129                          sizeof(struct nbd_sock *), GFP_KERNEL);
1130         if (!socks) {
1131                 kfree(nsock);
1132                 err = -ENOMEM;
1133                 goto put_socket;
1134         }
1135
1136         config->socks = socks;
1137
1138         nsock->fallback_index = -1;
1139         nsock->dead = false;
1140         mutex_init(&nsock->tx_lock);
1141         nsock->sock = sock;
1142         nsock->pending = NULL;
1143         nsock->sent = 0;
1144         nsock->cookie = 0;
1145         socks[config->num_connections++] = nsock;
1146         atomic_inc(&config->live_connections);
1147         blk_mq_unfreeze_queue(nbd->disk->queue);
1148
1149         return 0;
1150
1151 put_socket:
1152         blk_mq_unfreeze_queue(nbd->disk->queue);
1153         sockfd_put(sock);
1154         return err;
1155 }
1156
1157 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1158 {
1159         struct nbd_config *config = nbd->config;
1160         struct socket *sock, *old;
1161         struct recv_thread_args *args;
1162         int i;
1163         int err;
1164
1165         sock = nbd_get_socket(nbd, arg, &err);
1166         if (!sock)
1167                 return err;
1168
1169         args = kzalloc(sizeof(*args), GFP_KERNEL);
1170         if (!args) {
1171                 sockfd_put(sock);
1172                 return -ENOMEM;
1173         }
1174
1175         for (i = 0; i < config->num_connections; i++) {
1176                 struct nbd_sock *nsock = config->socks[i];
1177
1178                 if (!nsock->dead)
1179                         continue;
1180
1181                 mutex_lock(&nsock->tx_lock);
1182                 if (!nsock->dead) {
1183                         mutex_unlock(&nsock->tx_lock);
1184                         continue;
1185                 }
1186                 sk_set_memalloc(sock->sk);
1187                 if (nbd->tag_set.timeout)
1188                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1189                 atomic_inc(&config->recv_threads);
1190                 refcount_inc(&nbd->config_refs);
1191                 old = nsock->sock;
1192                 nsock->fallback_index = -1;
1193                 nsock->sock = sock;
1194                 nsock->dead = false;
1195                 INIT_WORK(&args->work, recv_work);
1196                 args->index = i;
1197                 args->nbd = nbd;
1198                 nsock->cookie++;
1199                 mutex_unlock(&nsock->tx_lock);
1200                 sockfd_put(old);
1201
1202                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1203
1204                 /* We take the tx_mutex in an error path in the recv_work, so we
1205                  * need to queue_work outside of the tx_mutex.
1206                  */
1207                 queue_work(nbd->recv_workq, &args->work);
1208
1209                 atomic_inc(&config->live_connections);
1210                 wake_up(&config->conn_wait);
1211                 return 0;
1212         }
1213         sockfd_put(sock);
1214         kfree(args);
1215         return -ENOSPC;
1216 }
1217
1218 static void nbd_bdev_reset(struct block_device *bdev)
1219 {
1220         if (bdev->bd_openers > 1)
1221                 return;
1222         set_capacity(bdev->bd_disk, 0);
1223 }
1224
1225 static void nbd_parse_flags(struct nbd_device *nbd)
1226 {
1227         struct nbd_config *config = nbd->config;
1228         if (config->flags & NBD_FLAG_READ_ONLY)
1229                 set_disk_ro(nbd->disk, true);
1230         else
1231                 set_disk_ro(nbd->disk, false);
1232         if (config->flags & NBD_FLAG_SEND_TRIM)
1233                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1234         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1235                 if (config->flags & NBD_FLAG_SEND_FUA)
1236                         blk_queue_write_cache(nbd->disk->queue, true, true);
1237                 else
1238                         blk_queue_write_cache(nbd->disk->queue, true, false);
1239         }
1240         else
1241                 blk_queue_write_cache(nbd->disk->queue, false, false);
1242 }
1243
1244 static void send_disconnects(struct nbd_device *nbd)
1245 {
1246         struct nbd_config *config = nbd->config;
1247         struct nbd_request request = {
1248                 .magic = htonl(NBD_REQUEST_MAGIC),
1249                 .type = htonl(NBD_CMD_DISC),
1250         };
1251         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1252         struct iov_iter from;
1253         int i, ret;
1254
1255         for (i = 0; i < config->num_connections; i++) {
1256                 struct nbd_sock *nsock = config->socks[i];
1257
1258                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1259                 mutex_lock(&nsock->tx_lock);
1260                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1261                 if (ret < 0)
1262                         dev_err(disk_to_dev(nbd->disk),
1263                                 "Send disconnect failed %d\n", ret);
1264                 mutex_unlock(&nsock->tx_lock);
1265         }
1266 }
1267
1268 static int nbd_disconnect(struct nbd_device *nbd)
1269 {
1270         struct nbd_config *config = nbd->config;
1271
1272         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1273         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1274         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1275         send_disconnects(nbd);
1276         return 0;
1277 }
1278
1279 static void nbd_clear_sock(struct nbd_device *nbd)
1280 {
1281         sock_shutdown(nbd);
1282         nbd_clear_que(nbd);
1283         nbd->task_setup = NULL;
1284 }
1285
1286 static void nbd_config_put(struct nbd_device *nbd)
1287 {
1288         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1289                                         &nbd->config_lock)) {
1290                 struct nbd_config *config = nbd->config;
1291                 nbd_dev_dbg_close(nbd);
1292                 invalidate_disk(nbd->disk);
1293                 if (nbd->config->bytesize)
1294                         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
1295                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1296                                        &config->runtime_flags))
1297                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1298                 nbd->pid = 0;
1299                 if (test_and_clear_bit(NBD_RT_HAS_BACKEND_FILE,
1300                                        &config->runtime_flags)) {
1301                         device_remove_file(disk_to_dev(nbd->disk), &backend_attr);
1302                         kfree(nbd->backend);
1303                         nbd->backend = NULL;
1304                 }
1305                 nbd_clear_sock(nbd);
1306                 if (config->num_connections) {
1307                         int i;
1308                         for (i = 0; i < config->num_connections; i++) {
1309                                 sockfd_put(config->socks[i]->sock);
1310                                 kfree(config->socks[i]);
1311                         }
1312                         kfree(config->socks);
1313                 }
1314                 kfree(nbd->config);
1315                 nbd->config = NULL;
1316
1317                 if (nbd->recv_workq)
1318                         destroy_workqueue(nbd->recv_workq);
1319                 nbd->recv_workq = NULL;
1320
1321                 nbd->tag_set.timeout = 0;
1322                 nbd->disk->queue->limits.discard_granularity = 0;
1323                 nbd->disk->queue->limits.discard_alignment = 0;
1324                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1325                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1326
1327                 mutex_unlock(&nbd->config_lock);
1328                 nbd_put(nbd);
1329                 module_put(THIS_MODULE);
1330         }
1331 }
1332
1333 static int nbd_start_device(struct nbd_device *nbd)
1334 {
1335         struct nbd_config *config = nbd->config;
1336         int num_connections = config->num_connections;
1337         int error = 0, i;
1338
1339         if (nbd->pid)
1340                 return -EBUSY;
1341         if (!config->socks)
1342                 return -EINVAL;
1343         if (num_connections > 1 &&
1344             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1345                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1346                 return -EINVAL;
1347         }
1348
1349         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1350                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1351                                           WQ_UNBOUND, 0, nbd->index);
1352         if (!nbd->recv_workq) {
1353                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1354                 return -ENOMEM;
1355         }
1356
1357         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1358         nbd->pid = task_pid_nr(current);
1359
1360         nbd_parse_flags(nbd);
1361
1362         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1363         if (error) {
1364                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed for pid!\n");
1365                 return error;
1366         }
1367         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1368
1369         nbd_dev_dbg_init(nbd);
1370         for (i = 0; i < num_connections; i++) {
1371                 struct recv_thread_args *args;
1372
1373                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1374                 if (!args) {
1375                         sock_shutdown(nbd);
1376                         /*
1377                          * If num_connections is m (2 < m),
1378                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1379                          * But NO.(n + 1) failed. We still have n recv threads.
1380                          * So, add flush_workqueue here to prevent recv threads
1381                          * dropping the last config_refs and trying to destroy
1382                          * the workqueue from inside the workqueue.
1383                          */
1384                         if (i)
1385                                 flush_workqueue(nbd->recv_workq);
1386                         return -ENOMEM;
1387                 }
1388                 sk_set_memalloc(config->socks[i]->sock->sk);
1389                 if (nbd->tag_set.timeout)
1390                         config->socks[i]->sock->sk->sk_sndtimeo =
1391                                 nbd->tag_set.timeout;
1392                 atomic_inc(&config->recv_threads);
1393                 refcount_inc(&nbd->config_refs);
1394                 INIT_WORK(&args->work, recv_work);
1395                 args->nbd = nbd;
1396                 args->index = i;
1397                 queue_work(nbd->recv_workq, &args->work);
1398         }
1399         return nbd_set_size(nbd, config->bytesize, nbd_blksize(config));
1400 }
1401
1402 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1403 {
1404         struct nbd_config *config = nbd->config;
1405         int ret;
1406
1407         ret = nbd_start_device(nbd);
1408         if (ret)
1409                 return ret;
1410
1411         if (max_part)
1412                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1413         mutex_unlock(&nbd->config_lock);
1414         ret = wait_event_interruptible(config->recv_wq,
1415                                          atomic_read(&config->recv_threads) == 0);
1416         if (ret)
1417                 sock_shutdown(nbd);
1418         flush_workqueue(nbd->recv_workq);
1419
1420         mutex_lock(&nbd->config_lock);
1421         nbd_bdev_reset(bdev);
1422         /* user requested, ignore socket errors */
1423         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1424                 ret = 0;
1425         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1426                 ret = -ETIMEDOUT;
1427         return ret;
1428 }
1429
1430 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1431                                  struct block_device *bdev)
1432 {
1433         sock_shutdown(nbd);
1434         __invalidate_device(bdev, true);
1435         nbd_bdev_reset(bdev);
1436         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1437                                &nbd->config->runtime_flags))
1438                 nbd_config_put(nbd);
1439 }
1440
1441 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1442 {
1443         nbd->tag_set.timeout = timeout * HZ;
1444         if (timeout)
1445                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1446         else
1447                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1448 }
1449
1450 /* Must be called with config_lock held */
1451 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1452                        unsigned int cmd, unsigned long arg)
1453 {
1454         struct nbd_config *config = nbd->config;
1455         loff_t bytesize;
1456
1457         switch (cmd) {
1458         case NBD_DISCONNECT:
1459                 return nbd_disconnect(nbd);
1460         case NBD_CLEAR_SOCK:
1461                 nbd_clear_sock_ioctl(nbd, bdev);
1462                 return 0;
1463         case NBD_SET_SOCK:
1464                 return nbd_add_socket(nbd, arg, false);
1465         case NBD_SET_BLKSIZE:
1466                 return nbd_set_size(nbd, config->bytesize, arg);
1467         case NBD_SET_SIZE:
1468                 return nbd_set_size(nbd, arg, nbd_blksize(config));
1469         case NBD_SET_SIZE_BLOCKS:
1470                 if (check_shl_overflow(arg, config->blksize_bits, &bytesize))
1471                         return -EINVAL;
1472                 return nbd_set_size(nbd, bytesize, nbd_blksize(config));
1473         case NBD_SET_TIMEOUT:
1474                 nbd_set_cmd_timeout(nbd, arg);
1475                 return 0;
1476
1477         case NBD_SET_FLAGS:
1478                 config->flags = arg;
1479                 return 0;
1480         case NBD_DO_IT:
1481                 return nbd_start_device_ioctl(nbd, bdev);
1482         case NBD_CLEAR_QUE:
1483                 /*
1484                  * This is for compatibility only.  The queue is always cleared
1485                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1486                  */
1487                 return 0;
1488         case NBD_PRINT_DEBUG:
1489                 /*
1490                  * For compatibility only, we no longer keep a list of
1491                  * outstanding requests.
1492                  */
1493                 return 0;
1494         }
1495         return -ENOTTY;
1496 }
1497
1498 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1499                      unsigned int cmd, unsigned long arg)
1500 {
1501         struct nbd_device *nbd = bdev->bd_disk->private_data;
1502         struct nbd_config *config = nbd->config;
1503         int error = -EINVAL;
1504
1505         if (!capable(CAP_SYS_ADMIN))
1506                 return -EPERM;
1507
1508         /* The block layer will pass back some non-nbd ioctls in case we have
1509          * special handling for them, but we don't so just return an error.
1510          */
1511         if (_IOC_TYPE(cmd) != 0xab)
1512                 return -EINVAL;
1513
1514         mutex_lock(&nbd->config_lock);
1515
1516         /* Don't allow ioctl operations on a nbd device that was created with
1517          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1518          */
1519         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1520             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1521                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1522         else
1523                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1524         mutex_unlock(&nbd->config_lock);
1525         return error;
1526 }
1527
1528 static struct nbd_config *nbd_alloc_config(void)
1529 {
1530         struct nbd_config *config;
1531
1532         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1533         if (!config)
1534                 return NULL;
1535         atomic_set(&config->recv_threads, 0);
1536         init_waitqueue_head(&config->recv_wq);
1537         init_waitqueue_head(&config->conn_wait);
1538         config->blksize_bits = NBD_DEF_BLKSIZE_BITS;
1539         atomic_set(&config->live_connections, 0);
1540         try_module_get(THIS_MODULE);
1541         return config;
1542 }
1543
1544 static int nbd_open(struct block_device *bdev, fmode_t mode)
1545 {
1546         struct nbd_device *nbd;
1547         int ret = 0;
1548
1549         mutex_lock(&nbd_index_mutex);
1550         nbd = bdev->bd_disk->private_data;
1551         if (!nbd) {
1552                 ret = -ENXIO;
1553                 goto out;
1554         }
1555         if (!refcount_inc_not_zero(&nbd->refs)) {
1556                 ret = -ENXIO;
1557                 goto out;
1558         }
1559         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1560                 struct nbd_config *config;
1561
1562                 mutex_lock(&nbd->config_lock);
1563                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1564                         mutex_unlock(&nbd->config_lock);
1565                         goto out;
1566                 }
1567                 config = nbd->config = nbd_alloc_config();
1568                 if (!config) {
1569                         ret = -ENOMEM;
1570                         mutex_unlock(&nbd->config_lock);
1571                         goto out;
1572                 }
1573                 refcount_set(&nbd->config_refs, 1);
1574                 refcount_inc(&nbd->refs);
1575                 mutex_unlock(&nbd->config_lock);
1576                 if (max_part)
1577                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1578         } else if (nbd_disconnected(nbd->config)) {
1579                 if (max_part)
1580                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1581         }
1582 out:
1583         mutex_unlock(&nbd_index_mutex);
1584         return ret;
1585 }
1586
1587 static void nbd_release(struct gendisk *disk, fmode_t mode)
1588 {
1589         struct nbd_device *nbd = disk->private_data;
1590
1591         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1592                         disk->part0->bd_openers == 0)
1593                 nbd_disconnect_and_put(nbd);
1594
1595         nbd_config_put(nbd);
1596         nbd_put(nbd);
1597 }
1598
1599 static const struct block_device_operations nbd_fops =
1600 {
1601         .owner =        THIS_MODULE,
1602         .open =         nbd_open,
1603         .release =      nbd_release,
1604         .ioctl =        nbd_ioctl,
1605         .compat_ioctl = nbd_ioctl,
1606 };
1607
1608 #if IS_ENABLED(CONFIG_DEBUG_FS)
1609
1610 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1611 {
1612         struct nbd_device *nbd = s->private;
1613
1614         if (nbd->pid)
1615                 seq_printf(s, "recv: %d\n", nbd->pid);
1616
1617         return 0;
1618 }
1619
1620 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1621
1622 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1623 {
1624         struct nbd_device *nbd = s->private;
1625         u32 flags = nbd->config->flags;
1626
1627         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1628
1629         seq_puts(s, "Known flags:\n");
1630
1631         if (flags & NBD_FLAG_HAS_FLAGS)
1632                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1633         if (flags & NBD_FLAG_READ_ONLY)
1634                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1635         if (flags & NBD_FLAG_SEND_FLUSH)
1636                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1637         if (flags & NBD_FLAG_SEND_FUA)
1638                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1639         if (flags & NBD_FLAG_SEND_TRIM)
1640                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1641
1642         return 0;
1643 }
1644
1645 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1646
1647 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1648 {
1649         struct dentry *dir;
1650         struct nbd_config *config = nbd->config;
1651
1652         if (!nbd_dbg_dir)
1653                 return -EIO;
1654
1655         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1656         if (!dir) {
1657                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1658                         nbd_name(nbd));
1659                 return -EIO;
1660         }
1661         config->dbg_dir = dir;
1662
1663         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1664         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1665         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1666         debugfs_create_u32("blocksize_bits", 0444, dir, &config->blksize_bits);
1667         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1668
1669         return 0;
1670 }
1671
1672 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1673 {
1674         debugfs_remove_recursive(nbd->config->dbg_dir);
1675 }
1676
1677 static int nbd_dbg_init(void)
1678 {
1679         struct dentry *dbg_dir;
1680
1681         dbg_dir = debugfs_create_dir("nbd", NULL);
1682         if (!dbg_dir)
1683                 return -EIO;
1684
1685         nbd_dbg_dir = dbg_dir;
1686
1687         return 0;
1688 }
1689
1690 static void nbd_dbg_close(void)
1691 {
1692         debugfs_remove_recursive(nbd_dbg_dir);
1693 }
1694
1695 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1696
1697 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1698 {
1699         return 0;
1700 }
1701
1702 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1703 {
1704 }
1705
1706 static int nbd_dbg_init(void)
1707 {
1708         return 0;
1709 }
1710
1711 static void nbd_dbg_close(void)
1712 {
1713 }
1714
1715 #endif
1716
1717 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1718                             unsigned int hctx_idx, unsigned int numa_node)
1719 {
1720         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1721         cmd->nbd = set->driver_data;
1722         cmd->flags = 0;
1723         mutex_init(&cmd->lock);
1724         return 0;
1725 }
1726
1727 static const struct blk_mq_ops nbd_mq_ops = {
1728         .queue_rq       = nbd_queue_rq,
1729         .complete       = nbd_complete_rq,
1730         .init_request   = nbd_init_request,
1731         .timeout        = nbd_xmit_timeout,
1732 };
1733
1734 static struct nbd_device *nbd_dev_add(int index, unsigned int refs)
1735 {
1736         struct nbd_device *nbd;
1737         struct gendisk *disk;
1738         int err = -ENOMEM;
1739
1740         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1741         if (!nbd)
1742                 goto out;
1743
1744         nbd->tag_set.ops = &nbd_mq_ops;
1745         nbd->tag_set.nr_hw_queues = 1;
1746         nbd->tag_set.queue_depth = 128;
1747         nbd->tag_set.numa_node = NUMA_NO_NODE;
1748         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1749         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1750                 BLK_MQ_F_BLOCKING;
1751         nbd->tag_set.driver_data = nbd;
1752         INIT_WORK(&nbd->remove_work, nbd_dev_remove_work);
1753         nbd->backend = NULL;
1754
1755         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1756         if (err)
1757                 goto out_free_nbd;
1758
1759         mutex_lock(&nbd_index_mutex);
1760         if (index >= 0) {
1761                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1762                                 GFP_KERNEL);
1763                 if (err == -ENOSPC)
1764                         err = -EEXIST;
1765         } else {
1766                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1767                 if (err >= 0)
1768                         index = err;
1769         }
1770         nbd->index = index;
1771         mutex_unlock(&nbd_index_mutex);
1772         if (err < 0)
1773                 goto out_free_tags;
1774
1775         disk = blk_mq_alloc_disk(&nbd->tag_set, NULL);
1776         if (IS_ERR(disk)) {
1777                 err = PTR_ERR(disk);
1778                 goto out_free_idr;
1779         }
1780         nbd->disk = disk;
1781
1782         /*
1783          * Tell the block layer that we are not a rotational device
1784          */
1785         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1786         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1787         disk->queue->limits.discard_granularity = 0;
1788         disk->queue->limits.discard_alignment = 0;
1789         blk_queue_max_discard_sectors(disk->queue, 0);
1790         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1791         blk_queue_max_segments(disk->queue, USHRT_MAX);
1792         blk_queue_max_hw_sectors(disk->queue, 65536);
1793         disk->queue->limits.max_sectors = 256;
1794
1795         mutex_init(&nbd->config_lock);
1796         refcount_set(&nbd->config_refs, 0);
1797         /*
1798          * Start out with a zero references to keep other threads from using
1799          * this device until it is fully initialized.
1800          */
1801         refcount_set(&nbd->refs, 0);
1802         INIT_LIST_HEAD(&nbd->list);
1803         disk->major = NBD_MAJOR;
1804
1805         /* Too big first_minor can cause duplicate creation of
1806          * sysfs files/links, since first_minor will be truncated to
1807          * byte in __device_add_disk().
1808          */
1809         disk->first_minor = index << part_shift;
1810         if (disk->first_minor > 0xff) {
1811                 err = -EINVAL;
1812                 goto out_free_idr;
1813         }
1814
1815         disk->minors = 1 << part_shift;
1816         disk->fops = &nbd_fops;
1817         disk->private_data = nbd;
1818         sprintf(disk->disk_name, "nbd%d", index);
1819         err = add_disk(disk);
1820         if (err)
1821                 goto out_err_disk;
1822
1823         /*
1824          * Now publish the device.
1825          */
1826         refcount_set(&nbd->refs, refs);
1827         nbd_total_devices++;
1828         return nbd;
1829
1830 out_err_disk:
1831         blk_cleanup_disk(disk);
1832 out_free_idr:
1833         mutex_lock(&nbd_index_mutex);
1834         idr_remove(&nbd_index_idr, index);
1835         mutex_unlock(&nbd_index_mutex);
1836 out_free_tags:
1837         blk_mq_free_tag_set(&nbd->tag_set);
1838 out_free_nbd:
1839         kfree(nbd);
1840 out:
1841         return ERR_PTR(err);
1842 }
1843
1844 static struct nbd_device *nbd_find_get_unused(void)
1845 {
1846         struct nbd_device *nbd;
1847         int id;
1848
1849         lockdep_assert_held(&nbd_index_mutex);
1850
1851         idr_for_each_entry(&nbd_index_idr, nbd, id) {
1852                 if (refcount_read(&nbd->config_refs) ||
1853                     test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags))
1854                         continue;
1855                 if (refcount_inc_not_zero(&nbd->refs))
1856                         return nbd;
1857         }
1858
1859         return NULL;
1860 }
1861
1862 /* Netlink interface. */
1863 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1864         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1865         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1866         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1867         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1868         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1869         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1870         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1871         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1872         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1873         [NBD_ATTR_BACKEND_IDENTIFIER]   =       { .type = NLA_STRING},
1874 };
1875
1876 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1877         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1878 };
1879
1880 /* We don't use this right now since we don't parse the incoming list, but we
1881  * still want it here so userspace knows what to expect.
1882  */
1883 static const struct nla_policy __attribute__((unused))
1884 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1885         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1886         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1887 };
1888
1889 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1890 {
1891         struct nbd_config *config = nbd->config;
1892         u64 bsize = nbd_blksize(config);
1893         u64 bytes = config->bytesize;
1894
1895         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1896                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1897
1898         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1899                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1900
1901         if (bytes != config->bytesize || bsize != nbd_blksize(config))
1902                 return nbd_set_size(nbd, bytes, bsize);
1903         return 0;
1904 }
1905
1906 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1907 {
1908         struct nbd_device *nbd;
1909         struct nbd_config *config;
1910         int index = -1;
1911         int ret;
1912         bool put_dev = false;
1913
1914         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1915                 return -EPERM;
1916
1917         if (info->attrs[NBD_ATTR_INDEX])
1918                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1919         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1920                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1921                 return -EINVAL;
1922         }
1923         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1924                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1925                 return -EINVAL;
1926         }
1927 again:
1928         mutex_lock(&nbd_index_mutex);
1929         if (index == -1) {
1930                 nbd = nbd_find_get_unused();
1931         } else {
1932                 nbd = idr_find(&nbd_index_idr, index);
1933                 if (nbd) {
1934                         if ((test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1935                              test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) ||
1936                             !refcount_inc_not_zero(&nbd->refs)) {
1937                                 mutex_unlock(&nbd_index_mutex);
1938                                 pr_err("nbd: device at index %d is going down\n",
1939                                         index);
1940                                 return -EINVAL;
1941                         }
1942                 }
1943         }
1944         mutex_unlock(&nbd_index_mutex);
1945
1946         if (!nbd) {
1947                 nbd = nbd_dev_add(index, 2);
1948                 if (IS_ERR(nbd)) {
1949                         pr_err("nbd: failed to add new device\n");
1950                         return PTR_ERR(nbd);
1951                 }
1952         }
1953
1954         mutex_lock(&nbd->config_lock);
1955         if (refcount_read(&nbd->config_refs)) {
1956                 mutex_unlock(&nbd->config_lock);
1957                 nbd_put(nbd);
1958                 if (index == -1)
1959                         goto again;
1960                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1961                 return -EBUSY;
1962         }
1963         if (WARN_ON(nbd->config)) {
1964                 mutex_unlock(&nbd->config_lock);
1965                 nbd_put(nbd);
1966                 return -EINVAL;
1967         }
1968         config = nbd->config = nbd_alloc_config();
1969         if (!nbd->config) {
1970                 mutex_unlock(&nbd->config_lock);
1971                 nbd_put(nbd);
1972                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1973                 return -ENOMEM;
1974         }
1975         refcount_set(&nbd->config_refs, 1);
1976         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1977
1978         ret = nbd_genl_size_set(info, nbd);
1979         if (ret)
1980                 goto out;
1981
1982         if (info->attrs[NBD_ATTR_TIMEOUT])
1983                 nbd_set_cmd_timeout(nbd,
1984                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1985         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1986                 config->dead_conn_timeout =
1987                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1988                 config->dead_conn_timeout *= HZ;
1989         }
1990         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1991                 config->flags =
1992                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1993         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1994                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1995                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1996                         /*
1997                          * We have 1 ref to keep the device around, and then 1
1998                          * ref for our current operation here, which will be
1999                          * inherited by the config.  If we already have
2000                          * DESTROY_ON_DISCONNECT set then we know we don't have
2001                          * that extra ref already held so we don't need the
2002                          * put_dev.
2003                          */
2004                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2005                                               &nbd->flags))
2006                                 put_dev = true;
2007                 } else {
2008                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2009                                                &nbd->flags))
2010                                 refcount_inc(&nbd->refs);
2011                 }
2012                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2013                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2014                                 &config->runtime_flags);
2015                 }
2016         }
2017
2018         if (info->attrs[NBD_ATTR_SOCKETS]) {
2019                 struct nlattr *attr;
2020                 int rem, fd;
2021
2022                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2023                                     rem) {
2024                         struct nlattr *socks[NBD_SOCK_MAX+1];
2025
2026                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2027                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2028                                 ret = -EINVAL;
2029                                 goto out;
2030                         }
2031                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2032                                                           attr,
2033                                                           nbd_sock_policy,
2034                                                           info->extack);
2035                         if (ret != 0) {
2036                                 printk(KERN_ERR "nbd: error processing sock list\n");
2037                                 ret = -EINVAL;
2038                                 goto out;
2039                         }
2040                         if (!socks[NBD_SOCK_FD])
2041                                 continue;
2042                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2043                         ret = nbd_add_socket(nbd, fd, true);
2044                         if (ret)
2045                                 goto out;
2046                 }
2047         }
2048         ret = nbd_start_device(nbd);
2049         if (ret)
2050                 goto out;
2051         if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2052                 nbd->backend = nla_strdup(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2053                                           GFP_KERNEL);
2054                 if (!nbd->backend) {
2055                         ret = -ENOMEM;
2056                         goto out;
2057                 }
2058         }
2059         ret = device_create_file(disk_to_dev(nbd->disk), &backend_attr);
2060         if (ret) {
2061                 dev_err(disk_to_dev(nbd->disk),
2062                         "device_create_file failed for backend!\n");
2063                 goto out;
2064         }
2065         set_bit(NBD_RT_HAS_BACKEND_FILE, &config->runtime_flags);
2066 out:
2067         mutex_unlock(&nbd->config_lock);
2068         if (!ret) {
2069                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2070                 refcount_inc(&nbd->config_refs);
2071                 nbd_connect_reply(info, nbd->index);
2072         }
2073         nbd_config_put(nbd);
2074         if (put_dev)
2075                 nbd_put(nbd);
2076         return ret;
2077 }
2078
2079 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2080 {
2081         mutex_lock(&nbd->config_lock);
2082         nbd_disconnect(nbd);
2083         sock_shutdown(nbd);
2084         /*
2085          * Make sure recv thread has finished, so it does not drop the last
2086          * config ref and try to destroy the workqueue from inside the work
2087          * queue. And this also ensure that we can safely call nbd_clear_que()
2088          * to cancel the inflight I/Os.
2089          */
2090         if (nbd->recv_workq)
2091                 flush_workqueue(nbd->recv_workq);
2092         nbd_clear_que(nbd);
2093         nbd->task_setup = NULL;
2094         mutex_unlock(&nbd->config_lock);
2095
2096         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2097                                &nbd->config->runtime_flags))
2098                 nbd_config_put(nbd);
2099 }
2100
2101 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2102 {
2103         struct nbd_device *nbd;
2104         int index;
2105
2106         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2107                 return -EPERM;
2108
2109         if (!info->attrs[NBD_ATTR_INDEX]) {
2110                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2111                 return -EINVAL;
2112         }
2113         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2114         mutex_lock(&nbd_index_mutex);
2115         nbd = idr_find(&nbd_index_idr, index);
2116         if (!nbd) {
2117                 mutex_unlock(&nbd_index_mutex);
2118                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2119                        index);
2120                 return -EINVAL;
2121         }
2122         if (!refcount_inc_not_zero(&nbd->refs)) {
2123                 mutex_unlock(&nbd_index_mutex);
2124                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2125                        index);
2126                 return -EINVAL;
2127         }
2128         mutex_unlock(&nbd_index_mutex);
2129         if (!refcount_inc_not_zero(&nbd->config_refs))
2130                 goto put_nbd;
2131         nbd_disconnect_and_put(nbd);
2132         nbd_config_put(nbd);
2133 put_nbd:
2134         nbd_put(nbd);
2135         return 0;
2136 }
2137
2138 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2139 {
2140         struct nbd_device *nbd = NULL;
2141         struct nbd_config *config;
2142         int index;
2143         int ret = 0;
2144         bool put_dev = false;
2145
2146         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2147                 return -EPERM;
2148
2149         if (!info->attrs[NBD_ATTR_INDEX]) {
2150                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2151                 return -EINVAL;
2152         }
2153         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2154         mutex_lock(&nbd_index_mutex);
2155         nbd = idr_find(&nbd_index_idr, index);
2156         if (!nbd) {
2157                 mutex_unlock(&nbd_index_mutex);
2158                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2159                        index);
2160                 return -EINVAL;
2161         }
2162         if (nbd->backend) {
2163                 if (info->attrs[NBD_ATTR_BACKEND_IDENTIFIER]) {
2164                         if (nla_strcmp(info->attrs[NBD_ATTR_BACKEND_IDENTIFIER],
2165                                        nbd->backend)) {
2166                                 mutex_unlock(&nbd_index_mutex);
2167                                 dev_err(nbd_to_dev(nbd),
2168                                         "backend image doesn't match with %s\n",
2169                                         nbd->backend);
2170                                 return -EINVAL;
2171                         }
2172                 } else {
2173                         mutex_unlock(&nbd_index_mutex);
2174                         dev_err(nbd_to_dev(nbd), "must specify backend\n");
2175                         return -EINVAL;
2176                 }
2177         }
2178         if (!refcount_inc_not_zero(&nbd->refs)) {
2179                 mutex_unlock(&nbd_index_mutex);
2180                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2181                        index);
2182                 return -EINVAL;
2183         }
2184         mutex_unlock(&nbd_index_mutex);
2185
2186         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2187                 dev_err(nbd_to_dev(nbd),
2188                         "not configured, cannot reconfigure\n");
2189                 nbd_put(nbd);
2190                 return -EINVAL;
2191         }
2192
2193         mutex_lock(&nbd->config_lock);
2194         config = nbd->config;
2195         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2196             !nbd->pid) {
2197                 dev_err(nbd_to_dev(nbd),
2198                         "not configured, cannot reconfigure\n");
2199                 ret = -EINVAL;
2200                 goto out;
2201         }
2202
2203         ret = nbd_genl_size_set(info, nbd);
2204         if (ret)
2205                 goto out;
2206
2207         if (info->attrs[NBD_ATTR_TIMEOUT])
2208                 nbd_set_cmd_timeout(nbd,
2209                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2210         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2211                 config->dead_conn_timeout =
2212                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2213                 config->dead_conn_timeout *= HZ;
2214         }
2215         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2216                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2217                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2218                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2219                                               &nbd->flags))
2220                                 put_dev = true;
2221                 } else {
2222                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2223                                                &nbd->flags))
2224                                 refcount_inc(&nbd->refs);
2225                 }
2226
2227                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2228                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2229                                         &config->runtime_flags);
2230                 } else {
2231                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2232                                         &config->runtime_flags);
2233                 }
2234         }
2235
2236         if (info->attrs[NBD_ATTR_SOCKETS]) {
2237                 struct nlattr *attr;
2238                 int rem, fd;
2239
2240                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2241                                     rem) {
2242                         struct nlattr *socks[NBD_SOCK_MAX+1];
2243
2244                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2245                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2246                                 ret = -EINVAL;
2247                                 goto out;
2248                         }
2249                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2250                                                           attr,
2251                                                           nbd_sock_policy,
2252                                                           info->extack);
2253                         if (ret != 0) {
2254                                 printk(KERN_ERR "nbd: error processing sock list\n");
2255                                 ret = -EINVAL;
2256                                 goto out;
2257                         }
2258                         if (!socks[NBD_SOCK_FD])
2259                                 continue;
2260                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2261                         ret = nbd_reconnect_socket(nbd, fd);
2262                         if (ret) {
2263                                 if (ret == -ENOSPC)
2264                                         ret = 0;
2265                                 goto out;
2266                         }
2267                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2268                 }
2269         }
2270 out:
2271         mutex_unlock(&nbd->config_lock);
2272         nbd_config_put(nbd);
2273         nbd_put(nbd);
2274         if (put_dev)
2275                 nbd_put(nbd);
2276         return ret;
2277 }
2278
2279 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2280         {
2281                 .cmd    = NBD_CMD_CONNECT,
2282                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2283                 .doit   = nbd_genl_connect,
2284         },
2285         {
2286                 .cmd    = NBD_CMD_DISCONNECT,
2287                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2288                 .doit   = nbd_genl_disconnect,
2289         },
2290         {
2291                 .cmd    = NBD_CMD_RECONFIGURE,
2292                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2293                 .doit   = nbd_genl_reconfigure,
2294         },
2295         {
2296                 .cmd    = NBD_CMD_STATUS,
2297                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2298                 .doit   = nbd_genl_status,
2299         },
2300 };
2301
2302 static const struct genl_multicast_group nbd_mcast_grps[] = {
2303         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2304 };
2305
2306 static struct genl_family nbd_genl_family __ro_after_init = {
2307         .hdrsize        = 0,
2308         .name           = NBD_GENL_FAMILY_NAME,
2309         .version        = NBD_GENL_VERSION,
2310         .module         = THIS_MODULE,
2311         .small_ops      = nbd_connect_genl_ops,
2312         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2313         .maxattr        = NBD_ATTR_MAX,
2314         .policy = nbd_attr_policy,
2315         .mcgrps         = nbd_mcast_grps,
2316         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2317 };
2318
2319 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2320 {
2321         struct nlattr *dev_opt;
2322         u8 connected = 0;
2323         int ret;
2324
2325         /* This is a little racey, but for status it's ok.  The
2326          * reason we don't take a ref here is because we can't
2327          * take a ref in the index == -1 case as we would need
2328          * to put under the nbd_index_mutex, which could
2329          * deadlock if we are configured to remove ourselves
2330          * once we're disconnected.
2331          */
2332         if (refcount_read(&nbd->config_refs))
2333                 connected = 1;
2334         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2335         if (!dev_opt)
2336                 return -EMSGSIZE;
2337         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2338         if (ret)
2339                 return -EMSGSIZE;
2340         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2341                          connected);
2342         if (ret)
2343                 return -EMSGSIZE;
2344         nla_nest_end(reply, dev_opt);
2345         return 0;
2346 }
2347
2348 static int status_cb(int id, void *ptr, void *data)
2349 {
2350         struct nbd_device *nbd = ptr;
2351         return populate_nbd_status(nbd, (struct sk_buff *)data);
2352 }
2353
2354 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2355 {
2356         struct nlattr *dev_list;
2357         struct sk_buff *reply;
2358         void *reply_head;
2359         size_t msg_size;
2360         int index = -1;
2361         int ret = -ENOMEM;
2362
2363         if (info->attrs[NBD_ATTR_INDEX])
2364                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2365
2366         mutex_lock(&nbd_index_mutex);
2367
2368         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2369                                   nla_attr_size(sizeof(u8)));
2370         msg_size *= (index == -1) ? nbd_total_devices : 1;
2371
2372         reply = genlmsg_new(msg_size, GFP_KERNEL);
2373         if (!reply)
2374                 goto out;
2375         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2376                                        NBD_CMD_STATUS);
2377         if (!reply_head) {
2378                 nlmsg_free(reply);
2379                 goto out;
2380         }
2381
2382         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2383         if (index == -1) {
2384                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2385                 if (ret) {
2386                         nlmsg_free(reply);
2387                         goto out;
2388                 }
2389         } else {
2390                 struct nbd_device *nbd;
2391                 nbd = idr_find(&nbd_index_idr, index);
2392                 if (nbd) {
2393                         ret = populate_nbd_status(nbd, reply);
2394                         if (ret) {
2395                                 nlmsg_free(reply);
2396                                 goto out;
2397                         }
2398                 }
2399         }
2400         nla_nest_end(reply, dev_list);
2401         genlmsg_end(reply, reply_head);
2402         ret = genlmsg_reply(reply, info);
2403 out:
2404         mutex_unlock(&nbd_index_mutex);
2405         return ret;
2406 }
2407
2408 static void nbd_connect_reply(struct genl_info *info, int index)
2409 {
2410         struct sk_buff *skb;
2411         void *msg_head;
2412         int ret;
2413
2414         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2415         if (!skb)
2416                 return;
2417         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2418                                      NBD_CMD_CONNECT);
2419         if (!msg_head) {
2420                 nlmsg_free(skb);
2421                 return;
2422         }
2423         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2424         if (ret) {
2425                 nlmsg_free(skb);
2426                 return;
2427         }
2428         genlmsg_end(skb, msg_head);
2429         genlmsg_reply(skb, info);
2430 }
2431
2432 static void nbd_mcast_index(int index)
2433 {
2434         struct sk_buff *skb;
2435         void *msg_head;
2436         int ret;
2437
2438         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2439         if (!skb)
2440                 return;
2441         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2442                                      NBD_CMD_LINK_DEAD);
2443         if (!msg_head) {
2444                 nlmsg_free(skb);
2445                 return;
2446         }
2447         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2448         if (ret) {
2449                 nlmsg_free(skb);
2450                 return;
2451         }
2452         genlmsg_end(skb, msg_head);
2453         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2454 }
2455
2456 static void nbd_dead_link_work(struct work_struct *work)
2457 {
2458         struct link_dead_args *args = container_of(work, struct link_dead_args,
2459                                                    work);
2460         nbd_mcast_index(args->index);
2461         kfree(args);
2462 }
2463
2464 static int __init nbd_init(void)
2465 {
2466         int i;
2467
2468         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2469
2470         if (max_part < 0) {
2471                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2472                 return -EINVAL;
2473         }
2474
2475         part_shift = 0;
2476         if (max_part > 0) {
2477                 part_shift = fls(max_part);
2478
2479                 /*
2480                  * Adjust max_part according to part_shift as it is exported
2481                  * to user space so that user can know the max number of
2482                  * partition kernel should be able to manage.
2483                  *
2484                  * Note that -1 is required because partition 0 is reserved
2485                  * for the whole disk.
2486                  */
2487                 max_part = (1UL << part_shift) - 1;
2488         }
2489
2490         if ((1UL << part_shift) > DISK_MAX_PARTS)
2491                 return -EINVAL;
2492
2493         if (nbds_max > 1UL << (MINORBITS - part_shift))
2494                 return -EINVAL;
2495
2496         if (register_blkdev(NBD_MAJOR, "nbd"))
2497                 return -EIO;
2498
2499         nbd_del_wq = alloc_workqueue("nbd-del", WQ_UNBOUND, 0);
2500         if (!nbd_del_wq) {
2501                 unregister_blkdev(NBD_MAJOR, "nbd");
2502                 return -ENOMEM;
2503         }
2504
2505         if (genl_register_family(&nbd_genl_family)) {
2506                 destroy_workqueue(nbd_del_wq);
2507                 unregister_blkdev(NBD_MAJOR, "nbd");
2508                 return -EINVAL;
2509         }
2510         nbd_dbg_init();
2511
2512         for (i = 0; i < nbds_max; i++)
2513                 nbd_dev_add(i, 1);
2514         return 0;
2515 }
2516
2517 static int nbd_exit_cb(int id, void *ptr, void *data)
2518 {
2519         struct list_head *list = (struct list_head *)data;
2520         struct nbd_device *nbd = ptr;
2521
2522         /* Skip nbd that is being removed asynchronously */
2523         if (refcount_read(&nbd->refs))
2524                 list_add_tail(&nbd->list, list);
2525
2526         return 0;
2527 }
2528
2529 static void __exit nbd_cleanup(void)
2530 {
2531         struct nbd_device *nbd;
2532         LIST_HEAD(del_list);
2533
2534         nbd_dbg_close();
2535
2536         mutex_lock(&nbd_index_mutex);
2537         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2538         mutex_unlock(&nbd_index_mutex);
2539
2540         while (!list_empty(&del_list)) {
2541                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2542                 list_del_init(&nbd->list);
2543                 if (refcount_read(&nbd->refs) != 1)
2544                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2545                 nbd_put(nbd);
2546         }
2547
2548         /* Also wait for nbd_dev_remove_work() completes */
2549         destroy_workqueue(nbd_del_wq);
2550
2551         idr_destroy(&nbd_index_idr);
2552         genl_unregister_family(&nbd_genl_family);
2553         unregister_blkdev(NBD_MAJOR, "nbd");
2554 }
2555
2556 module_init(nbd_init);
2557 module_exit(nbd_cleanup);
2558
2559 MODULE_DESCRIPTION("Network Block Device");
2560 MODULE_LICENSE("GPL");
2561
2562 module_param(nbds_max, int, 0444);
2563 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2564 module_param(max_part, int, 0444);
2565 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");