drivers/block/nbd.c

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