drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   9  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #include <linux/major.h>
  15
  16 #include <linux/blkdev.h>
  17 #include <linux/module.h>
  18 #include <linux/init.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/mm.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/completion.h>
  30 #include <linux/err.h>
  31 #include <linux/kernel.h>
  32 #include <linux/slab.h>
  33 #include <net/sock.h>
  34 #include <linux/net.h>
  35 #include <linux/kthread.h>
  36 #include <linux/types.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/blk-mq.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/types.h>
  42
  43 #include <linux/nbd.h>
  44 #include <linux/nbd-netlink.h>
  45 #include <net/genetlink.h>
  46
  47 #define CREATE_TRACE_POINTS
  48 #include <trace/events/nbd.h>
  49
  50 static DEFINE_IDR(nbd_index_idr);
  51 static DEFINE_MUTEX(nbd_index_mutex);
  52 static int nbd_total_devices = 0;
  53
  54 struct nbd_sock {
  55         struct socket *sock;
  56         struct mutex tx_lock;
  57         struct request *pending;
  58         int sent;
  59         bool dead;
  60         int fallback_index;
  61         int cookie;
  62 };
  63
  64 struct recv_thread_args {
  65         struct work_struct work;
  66         struct nbd_device *nbd;
  67         int index;
  68 };
  69
  70 struct link_dead_args {
  71         struct work_struct work;
  72         int index;
  73 };
  74
  75 #define NBD_RT_TIMEDOUT                 0
  76 #define NBD_RT_DISCONNECT_REQUESTED     1
  77 #define NBD_RT_DISCONNECTED             2
  78 #define NBD_RT_HAS_PID_FILE             3
  79 #define NBD_RT_HAS_CONFIG_REF           4
  80 #define NBD_RT_BOUND                    5
  81 #define NBD_RT_DESTROY_ON_DISCONNECT    6
  82 #define NBD_RT_DISCONNECT_ON_CLOSE      7
  83
  84 #define NBD_DESTROY_ON_DISCONNECT       0
  85 #define NBD_DISCONNECT_REQUESTED        1
  86
  87 struct nbd_config {
  88         u32 flags;
  89         unsigned long runtime_flags;
  90         u64 dead_conn_timeout;
  91
  92         struct nbd_sock **socks;
  93         int num_connections;
  94         atomic_t live_connections;
  95         wait_queue_head_t conn_wait;
  96
  97         atomic_t recv_threads;
  98         wait_queue_head_t recv_wq;
  99         loff_t blksize;
 100         loff_t bytesize;
 101 #if IS_ENABLED(CONFIG_DEBUG_FS)
 102         struct dentry *dbg_dir;
 103 #endif
 104 };
 105
 106 struct nbd_device {
 107         struct blk_mq_tag_set tag_set;
 108
 109         int index;
 110         refcount_t config_refs;
 111         refcount_t refs;
 112         struct nbd_config *config;
 113         struct mutex config_lock;
 114         struct gendisk *disk;
 115         struct workqueue_struct *recv_workq;
 116
 117         struct list_head list;
 118         struct task_struct *task_recv;
 119         struct task_struct *task_setup;
 120
 121         struct completion *destroy_complete;
 122         unsigned long flags;
 123 };
 124
 125 #define NBD_CMD_REQUEUED        1
 126
 127 struct nbd_cmd {
 128         struct nbd_device *nbd;
 129         struct mutex lock;
 130         int index;
 131         int cookie;
 132         int retries;
 133         blk_status_t status;
 134         unsigned long flags;
 135         u32 cmd_cookie;
 136 };
 137
 138 #if IS_ENABLED(CONFIG_DEBUG_FS)
 139 static struct dentry *nbd_dbg_dir;
 140 #endif
 141
 142 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 143
 144 #define NBD_MAGIC 0x68797548
 145
 146 #define NBD_DEF_BLKSIZE 1024
 147
 148 static unsigned int nbds_max = 16;
 149 static int max_part = 16;
 150 static int part_shift;
 151
 152 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 153 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 154 static void nbd_config_put(struct nbd_device *nbd);
 155 static void nbd_connect_reply(struct genl_info *info, int index);
 156 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 157 static void nbd_dead_link_work(struct work_struct *work);
 158 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 159
 160 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 161 {
 162         return disk_to_dev(nbd->disk);
 163 }
 164
 165 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 166 {
 167         struct request *req = blk_mq_rq_from_pdu(cmd);
 168
 169         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 170                 blk_mq_requeue_request(req, true);
 171 }
 172
 173 #define NBD_COOKIE_BITS 32
 174
 175 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 176 {
 177         struct request *req = blk_mq_rq_from_pdu(cmd);
 178         u32 tag = blk_mq_unique_tag(req);
 179         u64 cookie = cmd->cmd_cookie;
 180
 181         return (cookie << NBD_COOKIE_BITS) | tag;
 182 }
 183
 184 static u32 nbd_handle_to_tag(u64 handle)
 185 {
 186         return (u32)handle;
 187 }
 188
 189 static u32 nbd_handle_to_cookie(u64 handle)
 190 {
 191         return (u32)(handle >> NBD_COOKIE_BITS);
 192 }
 193
 194 static const char *nbdcmd_to_ascii(int cmd)
 195 {
 196         switch (cmd) {
 197         case  NBD_CMD_READ: return "read";
 198         case NBD_CMD_WRITE: return "write";
 199         case  NBD_CMD_DISC: return "disconnect";
 200         case NBD_CMD_FLUSH: return "flush";
 201         case  NBD_CMD_TRIM: return "trim/discard";
 202         }
 203         return "invalid";
 204 }
 205
 206 static ssize_t pid_show(struct device *dev,
 207                         struct device_attribute *attr, char *buf)
 208 {
 209         struct gendisk *disk = dev_to_disk(dev);
 210         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 211
 212         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 213 }
 214
 215 static const struct device_attribute pid_attr = {
 216         .attr = { .name = "pid", .mode = 0444},
 217         .show = pid_show,
 218 };
 219
 220 static void nbd_dev_remove(struct nbd_device *nbd)
 221 {
 222         struct gendisk *disk = nbd->disk;
 223         struct request_queue *q;
 224
 225         if (disk) {
 226                 q = disk->queue;
 227                 del_gendisk(disk);
 228                 blk_cleanup_queue(q);
 229                 blk_mq_free_tag_set(&nbd->tag_set);
 230                 disk->private_data = NULL;
 231                 put_disk(disk);
 232         }
 233
 234         /*
 235          * Place this in the last just before the nbd is freed to
 236          * make sure that the disk and the related kobject are also
 237          * totally removed to avoid duplicate creation of the same
 238          * one.
 239          */
 240         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) && nbd->destroy_complete)
 241                 complete(nbd->destroy_complete);
 242
 243         kfree(nbd);
 244 }
 245
 246 static void nbd_put(struct nbd_device *nbd)
 247 {
 248         if (refcount_dec_and_mutex_lock(&nbd->refs,
 249                                         &nbd_index_mutex)) {
 250                 idr_remove(&nbd_index_idr, nbd->index);
 251                 nbd_dev_remove(nbd);
 252                 mutex_unlock(&nbd_index_mutex);
 253         }
 254 }
 255
 256 static int nbd_disconnected(struct nbd_config *config)
 257 {
 258         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 259                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 260 }
 261
 262 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 263                                 int notify)
 264 {
 265         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 266                 struct link_dead_args *args;
 267                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 268                 if (args) {
 269                         INIT_WORK(&args->work, nbd_dead_link_work);
 270                         args->index = nbd->index;
 271                         queue_work(system_wq, &args->work);
 272                 }
 273         }
 274         if (!nsock->dead) {
 275                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 276                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 277                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 278                                                &nbd->config->runtime_flags)) {
 279                                 set_bit(NBD_RT_DISCONNECTED,
 280                                         &nbd->config->runtime_flags);
 281                                 dev_info(nbd_to_dev(nbd),
 282                                         "Disconnected due to user request.\n");
 283                         }
 284                 }
 285         }
 286         nsock->dead = true;
 287         nsock->pending = NULL;
 288         nsock->sent = 0;
 289 }
 290
 291 static void nbd_size_clear(struct nbd_device *nbd)
 292 {
 293         if (nbd->config->bytesize) {
 294                 set_capacity(nbd->disk, 0);
 295                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 296         }
 297 }
 298
 299 static int nbd_set_size(struct nbd_device *nbd, loff_t bytesize,
 300                 loff_t blksize)
 301 {
 302         if (!blksize)
 303                 blksize = NBD_DEF_BLKSIZE;
 304         if (blksize < 512 || blksize > PAGE_SIZE || !is_power_of_2(blksize))
 305                 return -EINVAL;
 306
 307         nbd->config->bytesize = bytesize;
 308         nbd->config->blksize = blksize;
 309
 310         if (!nbd->task_recv)
 311                 return 0;
 312
 313         if (nbd->config->flags & NBD_FLAG_SEND_TRIM) {
 314                 nbd->disk->queue->limits.discard_granularity = blksize;
 315                 nbd->disk->queue->limits.discard_alignment = blksize;
 316                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 317         }
 318         blk_queue_logical_block_size(nbd->disk->queue, blksize);
 319         blk_queue_physical_block_size(nbd->disk->queue, blksize);
 320
 321         if (max_part)
 322                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 323         if (!set_capacity_and_notify(nbd->disk, bytesize >> 9))
 324                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 325         return 0;
 326 }
 327
 328 static void nbd_complete_rq(struct request *req)
 329 {
 330         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 331
 332         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 333                 cmd->status ? "failed" : "done");
 334
 335         blk_mq_end_request(req, cmd->status);
 336 }
 337
 338 /*
 339  * Forcibly shutdown the socket causing all listeners to error
 340  */
 341 static void sock_shutdown(struct nbd_device *nbd)
 342 {
 343         struct nbd_config *config = nbd->config;
 344         int i;
 345
 346         if (config->num_connections == 0)
 347                 return;
 348         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 349                 return;
 350
 351         for (i = 0; i < config->num_connections; i++) {
 352                 struct nbd_sock *nsock = config->socks[i];
 353                 mutex_lock(&nsock->tx_lock);
 354                 nbd_mark_nsock_dead(nbd, nsock, 0);
 355                 mutex_unlock(&nsock->tx_lock);
 356         }
 357         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 358 }
 359
 360 static u32 req_to_nbd_cmd_type(struct request *req)
 361 {
 362         switch (req_op(req)) {
 363         case REQ_OP_DISCARD:
 364                 return NBD_CMD_TRIM;
 365         case REQ_OP_FLUSH:
 366                 return NBD_CMD_FLUSH;
 367         case REQ_OP_WRITE:
 368                 return NBD_CMD_WRITE;
 369         case REQ_OP_READ:
 370                 return NBD_CMD_READ;
 371         default:
 372                 return U32_MAX;
 373         }
 374 }
 375
 376 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 377                                                  bool reserved)
 378 {
 379         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 380         struct nbd_device *nbd = cmd->nbd;
 381         struct nbd_config *config;
 382
 383         if (!mutex_trylock(&cmd->lock))
 384                 return BLK_EH_RESET_TIMER;
 385
 386         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 387                 cmd->status = BLK_STS_TIMEOUT;
 388                 mutex_unlock(&cmd->lock);
 389                 goto done;
 390         }
 391         config = nbd->config;
 392
 393         if (config->num_connections > 1 ||
 394             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 395                 dev_err_ratelimited(nbd_to_dev(nbd),
 396                                     "Connection timed out, retrying (%d/%d alive)\n",
 397                                     atomic_read(&config->live_connections),
 398                                     config->num_connections);
 399                 /*
 400                  * Hooray we have more connections, requeue this IO, the submit
 401                  * path will put it on a real connection. Or if only one
 402                  * connection is configured, the submit path will wait util
 403                  * a new connection is reconfigured or util dead timeout.
 404                  */
 405                 if (config->socks) {
 406                         if (cmd->index < config->num_connections) {
 407                                 struct nbd_sock *nsock =
 408                                         config->socks[cmd->index];
 409                                 mutex_lock(&nsock->tx_lock);
 410                                 /* We can have multiple outstanding requests, so
 411                                  * we don't want to mark the nsock dead if we've
 412                                  * already reconnected with a new socket, so
 413                                  * only mark it dead if its the same socket we
 414                                  * were sent out on.
 415                                  */
 416                                 if (cmd->cookie == nsock->cookie)
 417                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 418                                 mutex_unlock(&nsock->tx_lock);
 419                         }
 420                         mutex_unlock(&cmd->lock);
 421                         nbd_requeue_cmd(cmd);
 422                         nbd_config_put(nbd);
 423                         return BLK_EH_DONE;
 424                 }
 425         }
 426
 427         if (!nbd->tag_set.timeout) {
 428                 /*
 429                  * Userspace sets timeout=0 to disable socket disconnection,
 430                  * so just warn and reset the timer.
 431                  */
 432                 struct nbd_sock *nsock = config->socks[cmd->index];
 433                 cmd->retries++;
 434                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 435                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 436                         (unsigned long long)blk_rq_pos(req) << 9,
 437                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 438
 439                 mutex_lock(&nsock->tx_lock);
 440                 if (cmd->cookie != nsock->cookie) {
 441                         nbd_requeue_cmd(cmd);
 442                         mutex_unlock(&nsock->tx_lock);
 443                         mutex_unlock(&cmd->lock);
 444                         nbd_config_put(nbd);
 445                         return BLK_EH_DONE;
 446                 }
 447                 mutex_unlock(&nsock->tx_lock);
 448                 mutex_unlock(&cmd->lock);
 449                 nbd_config_put(nbd);
 450                 return BLK_EH_RESET_TIMER;
 451         }
 452
 453         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 454         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 455         cmd->status = BLK_STS_IOERR;
 456         mutex_unlock(&cmd->lock);
 457         sock_shutdown(nbd);
 458         nbd_config_put(nbd);
 459 done:
 460         blk_mq_complete_request(req);
 461         return BLK_EH_DONE;
 462 }
 463
 464 /*
 465  *  Send or receive packet.
 466  */
 467 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 468                      struct iov_iter *iter, int msg_flags, int *sent)
 469 {
 470         struct nbd_config *config = nbd->config;
 471         struct socket *sock = config->socks[index]->sock;
 472         int result;
 473         struct msghdr msg;
 474         unsigned int noreclaim_flag;
 475
 476         if (unlikely(!sock)) {
 477                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 478                         "Attempted %s on closed socket in sock_xmit\n",
 479                         (send ? "send" : "recv"));
 480                 return -EINVAL;
 481         }
 482
 483         msg.msg_iter = *iter;
 484
 485         noreclaim_flag = memalloc_noreclaim_save();
 486         do {
 487                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 488                 msg.msg_name = NULL;
 489                 msg.msg_namelen = 0;
 490                 msg.msg_control = NULL;
 491                 msg.msg_controllen = 0;
 492                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 493
 494                 if (send)
 495                         result = sock_sendmsg(sock, &msg);
 496                 else
 497                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 498
 499                 if (result <= 0) {
 500                         if (result == 0)
 501                                 result = -EPIPE; /* short read */
 502                         break;
 503                 }
 504                 if (sent)
 505                         *sent += result;
 506         } while (msg_data_left(&msg));
 507
 508         memalloc_noreclaim_restore(noreclaim_flag);
 509
 510         return result;
 511 }
 512
 513 /*
 514  * Different settings for sk->sk_sndtimeo can result in different return values
 515  * if there is a signal pending when we enter sendmsg, because reasons?
 516  */
 517 static inline int was_interrupted(int result)
 518 {
 519         return result == -ERESTARTSYS || result == -EINTR;
 520 }
 521
 522 /* always call with the tx_lock held */
 523 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 524 {
 525         struct request *req = blk_mq_rq_from_pdu(cmd);
 526         struct nbd_config *config = nbd->config;
 527         struct nbd_sock *nsock = config->socks[index];
 528         int result;
 529         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 530         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 531         struct iov_iter from;
 532         unsigned long size = blk_rq_bytes(req);
 533         struct bio *bio;
 534         u64 handle;
 535         u32 type;
 536         u32 nbd_cmd_flags = 0;
 537         int sent = nsock->sent, skip = 0;
 538
 539         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 540
 541         type = req_to_nbd_cmd_type(req);
 542         if (type == U32_MAX)
 543                 return -EIO;
 544
 545         if (rq_data_dir(req) == WRITE &&
 546             (config->flags & NBD_FLAG_READ_ONLY)) {
 547                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 548                                     "Write on read-only\n");
 549                 return -EIO;
 550         }
 551
 552         if (req->cmd_flags & REQ_FUA)
 553                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 554
 555         /* We did a partial send previously, and we at least sent the whole
 556          * request struct, so just go and send the rest of the pages in the
 557          * request.
 558          */
 559         if (sent) {
 560                 if (sent >= sizeof(request)) {
 561                         skip = sent - sizeof(request);
 562
 563                         /* initialize handle for tracing purposes */
 564                         handle = nbd_cmd_handle(cmd);
 565
 566                         goto send_pages;
 567                 }
 568                 iov_iter_advance(&from, sent);
 569         } else {
 570                 cmd->cmd_cookie++;
 571         }
 572         cmd->index = index;
 573         cmd->cookie = nsock->cookie;
 574         cmd->retries = 0;
 575         request.type = htonl(type | nbd_cmd_flags);
 576         if (type != NBD_CMD_FLUSH) {
 577                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 578                 request.len = htonl(size);
 579         }
 580         handle = nbd_cmd_handle(cmd);
 581         memcpy(request.handle, &handle, sizeof(handle));
 582
 583         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 584
 585         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 586                 req, nbdcmd_to_ascii(type),
 587                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 588         result = sock_xmit(nbd, index, 1, &from,
 589                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 590         trace_nbd_header_sent(req, handle);
 591         if (result <= 0) {
 592                 if (was_interrupted(result)) {
 593                         /* If we havne't sent anything we can just return BUSY,
 594                          * however if we have sent something we need to make
 595                          * sure we only allow this req to be sent until we are
 596                          * completely done.
 597                          */
 598                         if (sent) {
 599                                 nsock->pending = req;
 600                                 nsock->sent = sent;
 601                         }
 602                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 603                         return BLK_STS_RESOURCE;
 604                 }
 605                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 606                         "Send control failed (result %d)\n", result);
 607                 return -EAGAIN;
 608         }
 609 send_pages:
 610         if (type != NBD_CMD_WRITE)
 611                 goto out;
 612
 613         bio = req->bio;
 614         while (bio) {
 615                 struct bio *next = bio->bi_next;
 616                 struct bvec_iter iter;
 617                 struct bio_vec bvec;
 618
 619                 bio_for_each_segment(bvec, bio, iter) {
 620                         bool is_last = !next && bio_iter_last(bvec, iter);
 621                         int flags = is_last ? 0 : MSG_MORE;
 622
 623                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 624                                 req, bvec.bv_len);
 625                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 626                         if (skip) {
 627                                 if (skip >= iov_iter_count(&from)) {
 628                                         skip -= iov_iter_count(&from);
 629                                         continue;
 630                                 }
 631                                 iov_iter_advance(&from, skip);
 632                                 skip = 0;
 633                         }
 634                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 635                         if (result <= 0) {
 636                                 if (was_interrupted(result)) {
 637                                         /* We've already sent the header, we
 638                                          * have no choice but to set pending and
 639                                          * return BUSY.
 640                                          */
 641                                         nsock->pending = req;
 642                                         nsock->sent = sent;
 643                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 644                                         return BLK_STS_RESOURCE;
 645                                 }
 646                                 dev_err(disk_to_dev(nbd->disk),
 647                                         "Send data failed (result %d)\n",
 648                                         result);
 649                                 return -EAGAIN;
 650                         }
 651                         /*
 652                          * The completion might already have come in,
 653                          * so break for the last one instead of letting
 654                          * the iterator do it. This prevents use-after-free
 655                          * of the bio.
 656                          */
 657                         if (is_last)
 658                                 break;
 659                 }
 660                 bio = next;
 661         }
 662 out:
 663         trace_nbd_payload_sent(req, handle);
 664         nsock->pending = NULL;
 665         nsock->sent = 0;
 666         return 0;
 667 }
 668
 669 /* NULL returned = something went wrong, inform userspace */
 670 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 671 {
 672         struct nbd_config *config = nbd->config;
 673         int result;
 674         struct nbd_reply reply;
 675         struct nbd_cmd *cmd;
 676         struct request *req = NULL;
 677         u64 handle;
 678         u16 hwq;
 679         u32 tag;
 680         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 681         struct iov_iter to;
 682         int ret = 0;
 683
 684         reply.magic = 0;
 685         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 686         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 687         if (result <= 0) {
 688                 if (!nbd_disconnected(config))
 689                         dev_err(disk_to_dev(nbd->disk),
 690                                 "Receive control failed (result %d)\n", result);
 691                 return ERR_PTR(result);
 692         }
 693
 694         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 695                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 696                                 (unsigned long)ntohl(reply.magic));
 697                 return ERR_PTR(-EPROTO);
 698         }
 699
 700         memcpy(&handle, reply.handle, sizeof(handle));
 701         tag = nbd_handle_to_tag(handle);
 702         hwq = blk_mq_unique_tag_to_hwq(tag);
 703         if (hwq < nbd->tag_set.nr_hw_queues)
 704                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 705                                        blk_mq_unique_tag_to_tag(tag));
 706         if (!req || !blk_mq_request_started(req)) {
 707                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 708                         tag, req);
 709                 return ERR_PTR(-ENOENT);
 710         }
 711         trace_nbd_header_received(req, handle);
 712         cmd = blk_mq_rq_to_pdu(req);
 713
 714         mutex_lock(&cmd->lock);
 715         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 716                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 717                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 718                 ret = -ENOENT;
 719                 goto out;
 720         }
 721         if (cmd->status != BLK_STS_OK) {
 722                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 723                         req);
 724                 ret = -ENOENT;
 725                 goto out;
 726         }
 727         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 728                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 729                         req);
 730                 ret = -ENOENT;
 731                 goto out;
 732         }
 733         if (ntohl(reply.error)) {
 734                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 735                         ntohl(reply.error));
 736                 cmd->status = BLK_STS_IOERR;
 737                 goto out;
 738         }
 739
 740         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 741         if (rq_data_dir(req) != WRITE) {
 742                 struct req_iterator iter;
 743                 struct bio_vec bvec;
 744
 745                 rq_for_each_segment(bvec, req, iter) {
 746                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 747                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 748                         if (result <= 0) {
 749                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 750                                         result);
 751                                 /*
 752                                  * If we've disconnected, we need to make sure we
 753                                  * complete this request, otherwise error out
 754                                  * and let the timeout stuff handle resubmitting
 755                                  * this request onto another connection.
 756                                  */
 757                                 if (nbd_disconnected(config)) {
 758                                         cmd->status = BLK_STS_IOERR;
 759                                         goto out;
 760                                 }
 761                                 ret = -EIO;
 762                                 goto out;
 763                         }
 764                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 765                                 req, bvec.bv_len);
 766                 }
 767         }
 768 out:
 769         trace_nbd_payload_received(req, handle);
 770         mutex_unlock(&cmd->lock);
 771         return ret ? ERR_PTR(ret) : cmd;
 772 }
 773
 774 static void recv_work(struct work_struct *work)
 775 {
 776         struct recv_thread_args *args = container_of(work,
 777                                                      struct recv_thread_args,
 778                                                      work);
 779         struct nbd_device *nbd = args->nbd;
 780         struct nbd_config *config = nbd->config;
 781         struct nbd_cmd *cmd;
 782         struct request *rq;
 783
 784         while (1) {
 785                 cmd = nbd_read_stat(nbd, args->index);
 786                 if (IS_ERR(cmd)) {
 787                         struct nbd_sock *nsock = config->socks[args->index];
 788
 789                         mutex_lock(&nsock->tx_lock);
 790                         nbd_mark_nsock_dead(nbd, nsock, 1);
 791                         mutex_unlock(&nsock->tx_lock);
 792                         break;
 793                 }
 794
 795                 rq = blk_mq_rq_from_pdu(cmd);
 796                 if (likely(!blk_should_fake_timeout(rq->q)))
 797                         blk_mq_complete_request(rq);
 798         }
 799         nbd_config_put(nbd);
 800         atomic_dec(&config->recv_threads);
 801         wake_up(&config->recv_wq);
 802         kfree(args);
 803 }
 804
 805 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 806 {
 807         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 808
 809         mutex_lock(&cmd->lock);
 810         cmd->status = BLK_STS_IOERR;
 811         mutex_unlock(&cmd->lock);
 812
 813         blk_mq_complete_request(req);
 814         return true;
 815 }
 816
 817 static void nbd_clear_que(struct nbd_device *nbd)
 818 {
 819         blk_mq_quiesce_queue(nbd->disk->queue);
 820         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 821         blk_mq_unquiesce_queue(nbd->disk->queue);
 822         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 823 }
 824
 825 static int find_fallback(struct nbd_device *nbd, int index)
 826 {
 827         struct nbd_config *config = nbd->config;
 828         int new_index = -1;
 829         struct nbd_sock *nsock = config->socks[index];
 830         int fallback = nsock->fallback_index;
 831
 832         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 833                 return new_index;
 834
 835         if (config->num_connections <= 1) {
 836                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 837                                     "Dead connection, failed to find a fallback\n");
 838                 return new_index;
 839         }
 840
 841         if (fallback >= 0 && fallback < config->num_connections &&
 842             !config->socks[fallback]->dead)
 843                 return fallback;
 844
 845         if (nsock->fallback_index < 0 ||
 846             nsock->fallback_index >= config->num_connections ||
 847             config->socks[nsock->fallback_index]->dead) {
 848                 int i;
 849                 for (i = 0; i < config->num_connections; i++) {
 850                         if (i == index)
 851                                 continue;
 852                         if (!config->socks[i]->dead) {
 853                                 new_index = i;
 854                                 break;
 855                         }
 856                 }
 857                 nsock->fallback_index = new_index;
 858                 if (new_index < 0) {
 859                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 860                                             "Dead connection, failed to find a fallback\n");
 861                         return new_index;
 862                 }
 863         }
 864         new_index = nsock->fallback_index;
 865         return new_index;
 866 }
 867
 868 static int wait_for_reconnect(struct nbd_device *nbd)
 869 {
 870         struct nbd_config *config = nbd->config;
 871         if (!config->dead_conn_timeout)
 872                 return 0;
 873         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 874                 return 0;
 875         return wait_event_timeout(config->conn_wait,
 876                                   atomic_read(&config->live_connections) > 0,
 877                                   config->dead_conn_timeout) > 0;
 878 }
 879
 880 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 881 {
 882         struct request *req = blk_mq_rq_from_pdu(cmd);
 883         struct nbd_device *nbd = cmd->nbd;
 884         struct nbd_config *config;
 885         struct nbd_sock *nsock;
 886         int ret;
 887
 888         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 889                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 890                                     "Socks array is empty\n");
 891                 blk_mq_start_request(req);
 892                 return -EINVAL;
 893         }
 894         config = nbd->config;
 895
 896         if (index >= config->num_connections) {
 897                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 898                                     "Attempted send on invalid socket\n");
 899                 nbd_config_put(nbd);
 900                 blk_mq_start_request(req);
 901                 return -EINVAL;
 902         }
 903         cmd->status = BLK_STS_OK;
 904 again:
 905         nsock = config->socks[index];
 906         mutex_lock(&nsock->tx_lock);
 907         if (nsock->dead) {
 908                 int old_index = index;
 909                 index = find_fallback(nbd, index);
 910                 mutex_unlock(&nsock->tx_lock);
 911                 if (index < 0) {
 912                         if (wait_for_reconnect(nbd)) {
 913                                 index = old_index;
 914                                 goto again;
 915                         }
 916                         /* All the sockets should already be down at this point,
 917                          * we just want to make sure that DISCONNECTED is set so
 918                          * any requests that come in that were queue'ed waiting
 919                          * for the reconnect timer don't trigger the timer again
 920                          * and instead just error out.
 921                          */
 922                         sock_shutdown(nbd);
 923                         nbd_config_put(nbd);
 924                         blk_mq_start_request(req);
 925                         return -EIO;
 926                 }
 927                 goto again;
 928         }
 929
 930         /* Handle the case that we have a pending request that was partially
 931          * transmitted that _has_ to be serviced first.  We need to call requeue
 932          * here so that it gets put _after_ the request that is already on the
 933          * dispatch list.
 934          */
 935         blk_mq_start_request(req);
 936         if (unlikely(nsock->pending && nsock->pending != req)) {
 937                 nbd_requeue_cmd(cmd);
 938                 ret = 0;
 939                 goto out;
 940         }
 941         /*
 942          * Some failures are related to the link going down, so anything that
 943          * returns EAGAIN can be retried on a different socket.
 944          */
 945         ret = nbd_send_cmd(nbd, cmd, index);
 946         if (ret == -EAGAIN) {
 947                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 948                                     "Request send failed, requeueing\n");
 949                 nbd_mark_nsock_dead(nbd, nsock, 1);
 950                 nbd_requeue_cmd(cmd);
 951                 ret = 0;
 952         }
 953 out:
 954         mutex_unlock(&nsock->tx_lock);
 955         nbd_config_put(nbd);
 956         return ret;
 957 }
 958
 959 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 960                         const struct blk_mq_queue_data *bd)
 961 {
 962         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 963         int ret;
 964
 965         /*
 966          * Since we look at the bio's to send the request over the network we
 967          * need to make sure the completion work doesn't mark this request done
 968          * before we are done doing our send.  This keeps us from dereferencing
 969          * freed data if we have particularly fast completions (ie we get the
 970          * completion before we exit sock_xmit on the last bvec) or in the case
 971          * that the server is misbehaving (or there was an error) before we're
 972          * done sending everything over the wire.
 973          */
 974         mutex_lock(&cmd->lock);
 975         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 976
 977         /* We can be called directly from the user space process, which means we
 978          * could possibly have signals pending so our sendmsg will fail.  In
 979          * this case we need to return that we are busy, otherwise error out as
 980          * appropriate.
 981          */
 982         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 983         if (ret < 0)
 984                 ret = BLK_STS_IOERR;
 985         else if (!ret)
 986                 ret = BLK_STS_OK;
 987         mutex_unlock(&cmd->lock);
 988
 989         return ret;
 990 }
 991
 992 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
 993                                      int *err)
 994 {
 995         struct socket *sock;
 996
 997         *err = 0;
 998         sock = sockfd_lookup(fd, err);
 999         if (!sock)
1000                 return NULL;
1001
1002         if (sock->ops->shutdown == sock_no_shutdown) {
1003                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1004                 *err = -EINVAL;
1005                 sockfd_put(sock);
1006                 return NULL;
1007         }
1008
1009         return sock;
1010 }
1011
1012 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1013                           bool netlink)
1014 {
1015         struct nbd_config *config = nbd->config;
1016         struct socket *sock;
1017         struct nbd_sock **socks;
1018         struct nbd_sock *nsock;
1019         int err;
1020
1021         sock = nbd_get_socket(nbd, arg, &err);
1022         if (!sock)
1023                 return err;
1024
1025         if (!netlink && !nbd->task_setup &&
1026             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1027                 nbd->task_setup = current;
1028
1029         if (!netlink &&
1030             (nbd->task_setup != current ||
1031              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1032                 dev_err(disk_to_dev(nbd->disk),
1033                         "Device being setup by another task");
1034                 err = -EBUSY;
1035                 goto put_socket;
1036         }
1037
1038         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1039         if (!nsock) {
1040                 err = -ENOMEM;
1041                 goto put_socket;
1042         }
1043
1044         socks = krealloc(config->socks, (config->num_connections + 1) *
1045                          sizeof(struct nbd_sock *), GFP_KERNEL);
1046         if (!socks) {
1047                 kfree(nsock);
1048                 err = -ENOMEM;
1049                 goto put_socket;
1050         }
1051
1052         config->socks = socks;
1053
1054         nsock->fallback_index = -1;
1055         nsock->dead = false;
1056         mutex_init(&nsock->tx_lock);
1057         nsock->sock = sock;
1058         nsock->pending = NULL;
1059         nsock->sent = 0;
1060         nsock->cookie = 0;
1061         socks[config->num_connections++] = nsock;
1062         atomic_inc(&config->live_connections);
1063
1064         return 0;
1065
1066 put_socket:
1067         sockfd_put(sock);
1068         return err;
1069 }
1070
1071 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1072 {
1073         struct nbd_config *config = nbd->config;
1074         struct socket *sock, *old;
1075         struct recv_thread_args *args;
1076         int i;
1077         int err;
1078
1079         sock = nbd_get_socket(nbd, arg, &err);
1080         if (!sock)
1081                 return err;
1082
1083         args = kzalloc(sizeof(*args), GFP_KERNEL);
1084         if (!args) {
1085                 sockfd_put(sock);
1086                 return -ENOMEM;
1087         }
1088
1089         for (i = 0; i < config->num_connections; i++) {
1090                 struct nbd_sock *nsock = config->socks[i];
1091
1092                 if (!nsock->dead)
1093                         continue;
1094
1095                 mutex_lock(&nsock->tx_lock);
1096                 if (!nsock->dead) {
1097                         mutex_unlock(&nsock->tx_lock);
1098                         continue;
1099                 }
1100                 sk_set_memalloc(sock->sk);
1101                 if (nbd->tag_set.timeout)
1102                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1103                 atomic_inc(&config->recv_threads);
1104                 refcount_inc(&nbd->config_refs);
1105                 old = nsock->sock;
1106                 nsock->fallback_index = -1;
1107                 nsock->sock = sock;
1108                 nsock->dead = false;
1109                 INIT_WORK(&args->work, recv_work);
1110                 args->index = i;
1111                 args->nbd = nbd;
1112                 nsock->cookie++;
1113                 mutex_unlock(&nsock->tx_lock);
1114                 sockfd_put(old);
1115
1116                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1117
1118                 /* We take the tx_mutex in an error path in the recv_work, so we
1119                  * need to queue_work outside of the tx_mutex.
1120                  */
1121                 queue_work(nbd->recv_workq, &args->work);
1122
1123                 atomic_inc(&config->live_connections);
1124                 wake_up(&config->conn_wait);
1125                 return 0;
1126         }
1127         sockfd_put(sock);
1128         kfree(args);
1129         return -ENOSPC;
1130 }
1131
1132 static void nbd_bdev_reset(struct block_device *bdev)
1133 {
1134         if (bdev->bd_openers > 1)
1135                 return;
1136         set_capacity(bdev->bd_disk, 0);
1137 }
1138
1139 static void nbd_parse_flags(struct nbd_device *nbd)
1140 {
1141         struct nbd_config *config = nbd->config;
1142         if (config->flags & NBD_FLAG_READ_ONLY)
1143                 set_disk_ro(nbd->disk, true);
1144         else
1145                 set_disk_ro(nbd->disk, false);
1146         if (config->flags & NBD_FLAG_SEND_TRIM)
1147                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1148         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1149                 if (config->flags & NBD_FLAG_SEND_FUA)
1150                         blk_queue_write_cache(nbd->disk->queue, true, true);
1151                 else
1152                         blk_queue_write_cache(nbd->disk->queue, true, false);
1153         }
1154         else
1155                 blk_queue_write_cache(nbd->disk->queue, false, false);
1156 }
1157
1158 static void send_disconnects(struct nbd_device *nbd)
1159 {
1160         struct nbd_config *config = nbd->config;
1161         struct nbd_request request = {
1162                 .magic = htonl(NBD_REQUEST_MAGIC),
1163                 .type = htonl(NBD_CMD_DISC),
1164         };
1165         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1166         struct iov_iter from;
1167         int i, ret;
1168
1169         for (i = 0; i < config->num_connections; i++) {
1170                 struct nbd_sock *nsock = config->socks[i];
1171
1172                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1173                 mutex_lock(&nsock->tx_lock);
1174                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1175                 if (ret <= 0)
1176                         dev_err(disk_to_dev(nbd->disk),
1177                                 "Send disconnect failed %d\n", ret);
1178                 mutex_unlock(&nsock->tx_lock);
1179         }
1180 }
1181
1182 static int nbd_disconnect(struct nbd_device *nbd)
1183 {
1184         struct nbd_config *config = nbd->config;
1185
1186         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1187         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1188         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1189         send_disconnects(nbd);
1190         return 0;
1191 }
1192
1193 static void nbd_clear_sock(struct nbd_device *nbd)
1194 {
1195         sock_shutdown(nbd);
1196         nbd_clear_que(nbd);
1197         nbd->task_setup = NULL;
1198 }
1199
1200 static void nbd_config_put(struct nbd_device *nbd)
1201 {
1202         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1203                                         &nbd->config_lock)) {
1204                 struct nbd_config *config = nbd->config;
1205                 nbd_dev_dbg_close(nbd);
1206                 nbd_size_clear(nbd);
1207                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1208                                        &config->runtime_flags))
1209                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1210                 nbd->task_recv = NULL;
1211                 nbd_clear_sock(nbd);
1212                 if (config->num_connections) {
1213                         int i;
1214                         for (i = 0; i < config->num_connections; i++) {
1215                                 sockfd_put(config->socks[i]->sock);
1216                                 kfree(config->socks[i]);
1217                         }
1218                         kfree(config->socks);
1219                 }
1220                 kfree(nbd->config);
1221                 nbd->config = NULL;
1222
1223                 if (nbd->recv_workq)
1224                         destroy_workqueue(nbd->recv_workq);
1225                 nbd->recv_workq = NULL;
1226
1227                 nbd->tag_set.timeout = 0;
1228                 nbd->disk->queue->limits.discard_granularity = 0;
1229                 nbd->disk->queue->limits.discard_alignment = 0;
1230                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1231                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1232
1233                 mutex_unlock(&nbd->config_lock);
1234                 nbd_put(nbd);
1235                 module_put(THIS_MODULE);
1236         }
1237 }
1238
1239 static int nbd_start_device(struct nbd_device *nbd)
1240 {
1241         struct nbd_config *config = nbd->config;
1242         int num_connections = config->num_connections;
1243         int error = 0, i;
1244
1245         if (nbd->task_recv)
1246                 return -EBUSY;
1247         if (!config->socks)
1248                 return -EINVAL;
1249         if (num_connections > 1 &&
1250             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1251                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1252                 return -EINVAL;
1253         }
1254
1255         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1256                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1257                                           WQ_UNBOUND, 0, nbd->index);
1258         if (!nbd->recv_workq) {
1259                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1260                 return -ENOMEM;
1261         }
1262
1263         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1264         nbd->task_recv = current;
1265
1266         nbd_parse_flags(nbd);
1267
1268         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1269         if (error) {
1270                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1271                 return error;
1272         }
1273         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1274
1275         nbd_dev_dbg_init(nbd);
1276         for (i = 0; i < num_connections; i++) {
1277                 struct recv_thread_args *args;
1278
1279                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1280                 if (!args) {
1281                         sock_shutdown(nbd);
1282                         /*
1283                          * If num_connections is m (2 < m),
1284                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1285                          * But NO.(n + 1) failed. We still have n recv threads.
1286                          * So, add flush_workqueue here to prevent recv threads
1287                          * dropping the last config_refs and trying to destroy
1288                          * the workqueue from inside the workqueue.
1289                          */
1290                         if (i)
1291                                 flush_workqueue(nbd->recv_workq);
1292                         return -ENOMEM;
1293                 }
1294                 sk_set_memalloc(config->socks[i]->sock->sk);
1295                 if (nbd->tag_set.timeout)
1296                         config->socks[i]->sock->sk->sk_sndtimeo =
1297                                 nbd->tag_set.timeout;
1298                 atomic_inc(&config->recv_threads);
1299                 refcount_inc(&nbd->config_refs);
1300                 INIT_WORK(&args->work, recv_work);
1301                 args->nbd = nbd;
1302                 args->index = i;
1303                 queue_work(nbd->recv_workq, &args->work);
1304         }
1305         return nbd_set_size(nbd, config->bytesize, config->blksize);
1306 }
1307
1308 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1309 {
1310         struct nbd_config *config = nbd->config;
1311         int ret;
1312
1313         ret = nbd_start_device(nbd);
1314         if (ret)
1315                 return ret;
1316
1317         if (max_part)
1318                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1319         mutex_unlock(&nbd->config_lock);
1320         ret = wait_event_interruptible(config->recv_wq,
1321                                          atomic_read(&config->recv_threads) == 0);
1322         if (ret)
1323                 sock_shutdown(nbd);
1324         flush_workqueue(nbd->recv_workq);
1325
1326         mutex_lock(&nbd->config_lock);
1327         nbd_bdev_reset(bdev);
1328         /* user requested, ignore socket errors */
1329         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1330                 ret = 0;
1331         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1332                 ret = -ETIMEDOUT;
1333         return ret;
1334 }
1335
1336 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1337                                  struct block_device *bdev)
1338 {
1339         sock_shutdown(nbd);
1340         __invalidate_device(bdev, true);
1341         nbd_bdev_reset(bdev);
1342         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1343                                &nbd->config->runtime_flags))
1344                 nbd_config_put(nbd);
1345 }
1346
1347 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1348 {
1349         nbd->tag_set.timeout = timeout * HZ;
1350         if (timeout)
1351                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1352         else
1353                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1354 }
1355
1356 /* Must be called with config_lock held */
1357 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1358                        unsigned int cmd, unsigned long arg)
1359 {
1360         struct nbd_config *config = nbd->config;
1361
1362         switch (cmd) {
1363         case NBD_DISCONNECT:
1364                 return nbd_disconnect(nbd);
1365         case NBD_CLEAR_SOCK:
1366                 nbd_clear_sock_ioctl(nbd, bdev);
1367                 return 0;
1368         case NBD_SET_SOCK:
1369                 return nbd_add_socket(nbd, arg, false);
1370         case NBD_SET_BLKSIZE:
1371                 return nbd_set_size(nbd, config->bytesize, arg);
1372         case NBD_SET_SIZE:
1373                 return nbd_set_size(nbd, arg, config->blksize);
1374         case NBD_SET_SIZE_BLOCKS:
1375                 return nbd_set_size(nbd, arg * config->blksize,
1376                                     config->blksize);
1377         case NBD_SET_TIMEOUT:
1378                 nbd_set_cmd_timeout(nbd, arg);
1379                 return 0;
1380
1381         case NBD_SET_FLAGS:
1382                 config->flags = arg;
1383                 return 0;
1384         case NBD_DO_IT:
1385                 return nbd_start_device_ioctl(nbd, bdev);
1386         case NBD_CLEAR_QUE:
1387                 /*
1388                  * This is for compatibility only.  The queue is always cleared
1389                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1390                  */
1391                 return 0;
1392         case NBD_PRINT_DEBUG:
1393                 /*
1394                  * For compatibility only, we no longer keep a list of
1395                  * outstanding requests.
1396                  */
1397                 return 0;
1398         }
1399         return -ENOTTY;
1400 }
1401
1402 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1403                      unsigned int cmd, unsigned long arg)
1404 {
1405         struct nbd_device *nbd = bdev->bd_disk->private_data;
1406         struct nbd_config *config = nbd->config;
1407         int error = -EINVAL;
1408
1409         if (!capable(CAP_SYS_ADMIN))
1410                 return -EPERM;
1411
1412         /* The block layer will pass back some non-nbd ioctls in case we have
1413          * special handling for them, but we don't so just return an error.
1414          */
1415         if (_IOC_TYPE(cmd) != 0xab)
1416                 return -EINVAL;
1417
1418         mutex_lock(&nbd->config_lock);
1419
1420         /* Don't allow ioctl operations on a nbd device that was created with
1421          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1422          */
1423         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1424             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1425                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1426         else
1427                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1428         mutex_unlock(&nbd->config_lock);
1429         return error;
1430 }
1431
1432 static struct nbd_config *nbd_alloc_config(void)
1433 {
1434         struct nbd_config *config;
1435
1436         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1437         if (!config)
1438                 return NULL;
1439         atomic_set(&config->recv_threads, 0);
1440         init_waitqueue_head(&config->recv_wq);
1441         init_waitqueue_head(&config->conn_wait);
1442         config->blksize = NBD_DEF_BLKSIZE;
1443         atomic_set(&config->live_connections, 0);
1444         try_module_get(THIS_MODULE);
1445         return config;
1446 }
1447
1448 static int nbd_open(struct block_device *bdev, fmode_t mode)
1449 {
1450         struct nbd_device *nbd;
1451         int ret = 0;
1452
1453         mutex_lock(&nbd_index_mutex);
1454         nbd = bdev->bd_disk->private_data;
1455         if (!nbd) {
1456                 ret = -ENXIO;
1457                 goto out;
1458         }
1459         if (!refcount_inc_not_zero(&nbd->refs)) {
1460                 ret = -ENXIO;
1461                 goto out;
1462         }
1463         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1464                 struct nbd_config *config;
1465
1466                 mutex_lock(&nbd->config_lock);
1467                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1468                         mutex_unlock(&nbd->config_lock);
1469                         goto out;
1470                 }
1471                 config = nbd->config = nbd_alloc_config();
1472                 if (!config) {
1473                         ret = -ENOMEM;
1474                         mutex_unlock(&nbd->config_lock);
1475                         goto out;
1476                 }
1477                 refcount_set(&nbd->config_refs, 1);
1478                 refcount_inc(&nbd->refs);
1479                 mutex_unlock(&nbd->config_lock);
1480                 if (max_part)
1481                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1482         } else if (nbd_disconnected(nbd->config)) {
1483                 if (max_part)
1484                         set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1485         }
1486 out:
1487         mutex_unlock(&nbd_index_mutex);
1488         return ret;
1489 }
1490
1491 static void nbd_release(struct gendisk *disk, fmode_t mode)
1492 {
1493         struct nbd_device *nbd = disk->private_data;
1494
1495         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1496                         disk->part0->bd_openers == 0)
1497                 nbd_disconnect_and_put(nbd);
1498
1499         nbd_config_put(nbd);
1500         nbd_put(nbd);
1501 }
1502
1503 static const struct block_device_operations nbd_fops =
1504 {
1505         .owner =        THIS_MODULE,
1506         .open =         nbd_open,
1507         .release =      nbd_release,
1508         .ioctl =        nbd_ioctl,
1509         .compat_ioctl = nbd_ioctl,
1510 };
1511
1512 #if IS_ENABLED(CONFIG_DEBUG_FS)
1513
1514 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1515 {
1516         struct nbd_device *nbd = s->private;
1517
1518         if (nbd->task_recv)
1519                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1520
1521         return 0;
1522 }
1523
1524 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_tasks);
1525
1526 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1527 {
1528         struct nbd_device *nbd = s->private;
1529         u32 flags = nbd->config->flags;
1530
1531         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1532
1533         seq_puts(s, "Known flags:\n");
1534
1535         if (flags & NBD_FLAG_HAS_FLAGS)
1536                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1537         if (flags & NBD_FLAG_READ_ONLY)
1538                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1539         if (flags & NBD_FLAG_SEND_FLUSH)
1540                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1541         if (flags & NBD_FLAG_SEND_FUA)
1542                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1543         if (flags & NBD_FLAG_SEND_TRIM)
1544                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1545
1546         return 0;
1547 }
1548
1549 DEFINE_SHOW_ATTRIBUTE(nbd_dbg_flags);
1550
1551 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1552 {
1553         struct dentry *dir;
1554         struct nbd_config *config = nbd->config;
1555
1556         if (!nbd_dbg_dir)
1557                 return -EIO;
1558
1559         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1560         if (!dir) {
1561                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1562                         nbd_name(nbd));
1563                 return -EIO;
1564         }
1565         config->dbg_dir = dir;
1566
1567         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_fops);
1568         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1569         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1570         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1571         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_fops);
1572
1573         return 0;
1574 }
1575
1576 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1577 {
1578         debugfs_remove_recursive(nbd->config->dbg_dir);
1579 }
1580
1581 static int nbd_dbg_init(void)
1582 {
1583         struct dentry *dbg_dir;
1584
1585         dbg_dir = debugfs_create_dir("nbd", NULL);
1586         if (!dbg_dir)
1587                 return -EIO;
1588
1589         nbd_dbg_dir = dbg_dir;
1590
1591         return 0;
1592 }
1593
1594 static void nbd_dbg_close(void)
1595 {
1596         debugfs_remove_recursive(nbd_dbg_dir);
1597 }
1598
1599 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1600
1601 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1602 {
1603         return 0;
1604 }
1605
1606 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1607 {
1608 }
1609
1610 static int nbd_dbg_init(void)
1611 {
1612         return 0;
1613 }
1614
1615 static void nbd_dbg_close(void)
1616 {
1617 }
1618
1619 #endif
1620
1621 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1622                             unsigned int hctx_idx, unsigned int numa_node)
1623 {
1624         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1625         cmd->nbd = set->driver_data;
1626         cmd->flags = 0;
1627         mutex_init(&cmd->lock);
1628         return 0;
1629 }
1630
1631 static const struct blk_mq_ops nbd_mq_ops = {
1632         .queue_rq       = nbd_queue_rq,
1633         .complete       = nbd_complete_rq,
1634         .init_request   = nbd_init_request,
1635         .timeout        = nbd_xmit_timeout,
1636 };
1637
1638 static int nbd_dev_add(int index)
1639 {
1640         struct nbd_device *nbd;
1641         struct gendisk *disk;
1642         struct request_queue *q;
1643         int err = -ENOMEM;
1644
1645         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1646         if (!nbd)
1647                 goto out;
1648
1649         disk = alloc_disk(1 << part_shift);
1650         if (!disk)
1651                 goto out_free_nbd;
1652
1653         if (index >= 0) {
1654                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1655                                 GFP_KERNEL);
1656                 if (err == -ENOSPC)
1657                         err = -EEXIST;
1658         } else {
1659                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1660                 if (err >= 0)
1661                         index = err;
1662         }
1663         if (err < 0)
1664                 goto out_free_disk;
1665
1666         nbd->index = index;
1667         nbd->disk = disk;
1668         nbd->tag_set.ops = &nbd_mq_ops;
1669         nbd->tag_set.nr_hw_queues = 1;
1670         nbd->tag_set.queue_depth = 128;
1671         nbd->tag_set.numa_node = NUMA_NO_NODE;
1672         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1673         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1674                 BLK_MQ_F_BLOCKING;
1675         nbd->tag_set.driver_data = nbd;
1676         nbd->destroy_complete = NULL;
1677
1678         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1679         if (err)
1680                 goto out_free_idr;
1681
1682         q = blk_mq_init_queue(&nbd->tag_set);
1683         if (IS_ERR(q)) {
1684                 err = PTR_ERR(q);
1685                 goto out_free_tags;
1686         }
1687         disk->queue = q;
1688
1689         /*
1690          * Tell the block layer that we are not a rotational device
1691          */
1692         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1693         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1694         disk->queue->limits.discard_granularity = 0;
1695         disk->queue->limits.discard_alignment = 0;
1696         blk_queue_max_discard_sectors(disk->queue, 0);
1697         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1698         blk_queue_max_segments(disk->queue, USHRT_MAX);
1699         blk_queue_max_hw_sectors(disk->queue, 65536);
1700         disk->queue->limits.max_sectors = 256;
1701
1702         mutex_init(&nbd->config_lock);
1703         refcount_set(&nbd->config_refs, 0);
1704         refcount_set(&nbd->refs, 1);
1705         INIT_LIST_HEAD(&nbd->list);
1706         disk->major = NBD_MAJOR;
1707         disk->first_minor = index << part_shift;
1708         disk->fops = &nbd_fops;
1709         disk->private_data = nbd;
1710         sprintf(disk->disk_name, "nbd%d", index);
1711         add_disk(disk);
1712         nbd_total_devices++;
1713         return index;
1714
1715 out_free_tags:
1716         blk_mq_free_tag_set(&nbd->tag_set);
1717 out_free_idr:
1718         idr_remove(&nbd_index_idr, index);
1719 out_free_disk:
1720         put_disk(disk);
1721 out_free_nbd:
1722         kfree(nbd);
1723 out:
1724         return err;
1725 }
1726
1727 static int find_free_cb(int id, void *ptr, void *data)
1728 {
1729         struct nbd_device *nbd = ptr;
1730         struct nbd_device **found = data;
1731
1732         if (!refcount_read(&nbd->config_refs)) {
1733                 *found = nbd;
1734                 return 1;
1735         }
1736         return 0;
1737 }
1738
1739 /* Netlink interface. */
1740 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1741         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1742         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1743         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1744         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1745         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1746         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1747         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1748         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1749         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1750 };
1751
1752 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1753         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1754 };
1755
1756 /* We don't use this right now since we don't parse the incoming list, but we
1757  * still want it here so userspace knows what to expect.
1758  */
1759 static const struct nla_policy __attribute__((unused))
1760 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1761         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1762         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1763 };
1764
1765 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1766 {
1767         struct nbd_config *config = nbd->config;
1768         u64 bsize = config->blksize;
1769         u64 bytes = config->bytesize;
1770
1771         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1772                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1773
1774         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES])
1775                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1776
1777         if (bytes != config->bytesize || bsize != config->blksize)
1778                 return nbd_set_size(nbd, bytes, bsize);
1779         return 0;
1780 }
1781
1782 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1783 {
1784         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1785         struct nbd_device *nbd = NULL;
1786         struct nbd_config *config;
1787         int index = -1;
1788         int ret;
1789         bool put_dev = false;
1790
1791         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1792                 return -EPERM;
1793
1794         if (info->attrs[NBD_ATTR_INDEX])
1795                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1796         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1797                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1798                 return -EINVAL;
1799         }
1800         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1801                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1802                 return -EINVAL;
1803         }
1804 again:
1805         mutex_lock(&nbd_index_mutex);
1806         if (index == -1) {
1807                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1808                 if (ret == 0) {
1809                         int new_index;
1810                         new_index = nbd_dev_add(-1);
1811                         if (new_index < 0) {
1812                                 mutex_unlock(&nbd_index_mutex);
1813                                 printk(KERN_ERR "nbd: failed to add new device\n");
1814                                 return new_index;
1815                         }
1816                         nbd = idr_find(&nbd_index_idr, new_index);
1817                 }
1818         } else {
1819                 nbd = idr_find(&nbd_index_idr, index);
1820                 if (!nbd) {
1821                         ret = nbd_dev_add(index);
1822                         if (ret < 0) {
1823                                 mutex_unlock(&nbd_index_mutex);
1824                                 printk(KERN_ERR "nbd: failed to add new device\n");
1825                                 return ret;
1826                         }
1827                         nbd = idr_find(&nbd_index_idr, index);
1828                 }
1829         }
1830         if (!nbd) {
1831                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1832                        index);
1833                 mutex_unlock(&nbd_index_mutex);
1834                 return -EINVAL;
1835         }
1836
1837         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1838             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1839                 nbd->destroy_complete = &destroy_complete;
1840                 mutex_unlock(&nbd_index_mutex);
1841
1842                 /* Wait untill the the nbd stuff is totally destroyed */
1843                 wait_for_completion(&destroy_complete);
1844                 goto again;
1845         }
1846
1847         if (!refcount_inc_not_zero(&nbd->refs)) {
1848                 mutex_unlock(&nbd_index_mutex);
1849                 if (index == -1)
1850                         goto again;
1851                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1852                        index);
1853                 return -EINVAL;
1854         }
1855         mutex_unlock(&nbd_index_mutex);
1856
1857         mutex_lock(&nbd->config_lock);
1858         if (refcount_read(&nbd->config_refs)) {
1859                 mutex_unlock(&nbd->config_lock);
1860                 nbd_put(nbd);
1861                 if (index == -1)
1862                         goto again;
1863                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1864                 return -EBUSY;
1865         }
1866         if (WARN_ON(nbd->config)) {
1867                 mutex_unlock(&nbd->config_lock);
1868                 nbd_put(nbd);
1869                 return -EINVAL;
1870         }
1871         config = nbd->config = nbd_alloc_config();
1872         if (!nbd->config) {
1873                 mutex_unlock(&nbd->config_lock);
1874                 nbd_put(nbd);
1875                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1876                 return -ENOMEM;
1877         }
1878         refcount_set(&nbd->config_refs, 1);
1879         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1880
1881         ret = nbd_genl_size_set(info, nbd);
1882         if (ret)
1883                 goto out;
1884
1885         if (info->attrs[NBD_ATTR_TIMEOUT])
1886                 nbd_set_cmd_timeout(nbd,
1887                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1888         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1889                 config->dead_conn_timeout =
1890                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1891                 config->dead_conn_timeout *= HZ;
1892         }
1893         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1894                 config->flags =
1895                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1896         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1897                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1898                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1899                         set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
1900                                 &config->runtime_flags);
1901                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1902                         put_dev = true;
1903                 } else {
1904                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1905                 }
1906                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1907                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1908                                 &config->runtime_flags);
1909                 }
1910         }
1911
1912         if (info->attrs[NBD_ATTR_SOCKETS]) {
1913                 struct nlattr *attr;
1914                 int rem, fd;
1915
1916                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1917                                     rem) {
1918                         struct nlattr *socks[NBD_SOCK_MAX+1];
1919
1920                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1921                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1922                                 ret = -EINVAL;
1923                                 goto out;
1924                         }
1925                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1926                                                           attr,
1927                                                           nbd_sock_policy,
1928                                                           info->extack);
1929                         if (ret != 0) {
1930                                 printk(KERN_ERR "nbd: error processing sock list\n");
1931                                 ret = -EINVAL;
1932                                 goto out;
1933                         }
1934                         if (!socks[NBD_SOCK_FD])
1935                                 continue;
1936                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1937                         ret = nbd_add_socket(nbd, fd, true);
1938                         if (ret)
1939                                 goto out;
1940                 }
1941         }
1942         ret = nbd_start_device(nbd);
1943 out:
1944         mutex_unlock(&nbd->config_lock);
1945         if (!ret) {
1946                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
1947                 refcount_inc(&nbd->config_refs);
1948                 nbd_connect_reply(info, nbd->index);
1949         }
1950         nbd_config_put(nbd);
1951         if (put_dev)
1952                 nbd_put(nbd);
1953         return ret;
1954 }
1955
1956 static void nbd_disconnect_and_put(struct nbd_device *nbd)
1957 {
1958         mutex_lock(&nbd->config_lock);
1959         nbd_disconnect(nbd);
1960         nbd_clear_sock(nbd);
1961         mutex_unlock(&nbd->config_lock);
1962         /*
1963          * Make sure recv thread has finished, so it does not drop the last
1964          * config ref and try to destroy the workqueue from inside the work
1965          * queue.
1966          */
1967         flush_workqueue(nbd->recv_workq);
1968         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1969                                &nbd->config->runtime_flags))
1970                 nbd_config_put(nbd);
1971 }
1972
1973 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
1974 {
1975         struct nbd_device *nbd;
1976         int index;
1977
1978         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1979                 return -EPERM;
1980
1981         if (!info->attrs[NBD_ATTR_INDEX]) {
1982                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
1983                 return -EINVAL;
1984         }
1985         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1986         mutex_lock(&nbd_index_mutex);
1987         nbd = idr_find(&nbd_index_idr, index);
1988         if (!nbd) {
1989                 mutex_unlock(&nbd_index_mutex);
1990                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1991                        index);
1992                 return -EINVAL;
1993         }
1994         if (!refcount_inc_not_zero(&nbd->refs)) {
1995                 mutex_unlock(&nbd_index_mutex);
1996                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1997                        index);
1998                 return -EINVAL;
1999         }
2000         mutex_unlock(&nbd_index_mutex);
2001         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2002                 nbd_put(nbd);
2003                 return 0;
2004         }
2005         nbd_disconnect_and_put(nbd);
2006         nbd_config_put(nbd);
2007         nbd_put(nbd);
2008         return 0;
2009 }
2010
2011 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2012 {
2013         struct nbd_device *nbd = NULL;
2014         struct nbd_config *config;
2015         int index;
2016         int ret = 0;
2017         bool put_dev = false;
2018
2019         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2020                 return -EPERM;
2021
2022         if (!info->attrs[NBD_ATTR_INDEX]) {
2023                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2024                 return -EINVAL;
2025         }
2026         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2027         mutex_lock(&nbd_index_mutex);
2028         nbd = idr_find(&nbd_index_idr, index);
2029         if (!nbd) {
2030                 mutex_unlock(&nbd_index_mutex);
2031                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2032                        index);
2033                 return -EINVAL;
2034         }
2035         if (!refcount_inc_not_zero(&nbd->refs)) {
2036                 mutex_unlock(&nbd_index_mutex);
2037                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2038                        index);
2039                 return -EINVAL;
2040         }
2041         mutex_unlock(&nbd_index_mutex);
2042
2043         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2044                 dev_err(nbd_to_dev(nbd),
2045                         "not configured, cannot reconfigure\n");
2046                 nbd_put(nbd);
2047                 return -EINVAL;
2048         }
2049
2050         mutex_lock(&nbd->config_lock);
2051         config = nbd->config;
2052         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2053             !nbd->task_recv) {
2054                 dev_err(nbd_to_dev(nbd),
2055                         "not configured, cannot reconfigure\n");
2056                 ret = -EINVAL;
2057                 goto out;
2058         }
2059
2060         ret = nbd_genl_size_set(info, nbd);
2061         if (ret)
2062                 goto out;
2063
2064         if (info->attrs[NBD_ATTR_TIMEOUT])
2065                 nbd_set_cmd_timeout(nbd,
2066                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2067         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2068                 config->dead_conn_timeout =
2069                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2070                 config->dead_conn_timeout *= HZ;
2071         }
2072         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2073                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2074                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2075                         if (!test_and_set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2076                                               &config->runtime_flags))
2077                                 put_dev = true;
2078                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2079                 } else {
2080                         if (test_and_clear_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2081                                                &config->runtime_flags))
2082                                 refcount_inc(&nbd->refs);
2083                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2084                 }
2085
2086                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2087                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2088                                         &config->runtime_flags);
2089                 } else {
2090                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2091                                         &config->runtime_flags);
2092                 }
2093         }
2094
2095         if (info->attrs[NBD_ATTR_SOCKETS]) {
2096                 struct nlattr *attr;
2097                 int rem, fd;
2098
2099                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2100                                     rem) {
2101                         struct nlattr *socks[NBD_SOCK_MAX+1];
2102
2103                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2104                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2105                                 ret = -EINVAL;
2106                                 goto out;
2107                         }
2108                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2109                                                           attr,
2110                                                           nbd_sock_policy,
2111                                                           info->extack);
2112                         if (ret != 0) {
2113                                 printk(KERN_ERR "nbd: error processing sock list\n");
2114                                 ret = -EINVAL;
2115                                 goto out;
2116                         }
2117                         if (!socks[NBD_SOCK_FD])
2118                                 continue;
2119                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2120                         ret = nbd_reconnect_socket(nbd, fd);
2121                         if (ret) {
2122                                 if (ret == -ENOSPC)
2123                                         ret = 0;
2124                                 goto out;
2125                         }
2126                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2127                 }
2128         }
2129 out:
2130         mutex_unlock(&nbd->config_lock);
2131         nbd_config_put(nbd);
2132         nbd_put(nbd);
2133         if (put_dev)
2134                 nbd_put(nbd);
2135         return ret;
2136 }
2137
2138 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2139         {
2140                 .cmd    = NBD_CMD_CONNECT,
2141                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2142                 .doit   = nbd_genl_connect,
2143         },
2144         {
2145                 .cmd    = NBD_CMD_DISCONNECT,
2146                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2147                 .doit   = nbd_genl_disconnect,
2148         },
2149         {
2150                 .cmd    = NBD_CMD_RECONFIGURE,
2151                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2152                 .doit   = nbd_genl_reconfigure,
2153         },
2154         {
2155                 .cmd    = NBD_CMD_STATUS,
2156                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2157                 .doit   = nbd_genl_status,
2158         },
2159 };
2160
2161 static const struct genl_multicast_group nbd_mcast_grps[] = {
2162         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2163 };
2164
2165 static struct genl_family nbd_genl_family __ro_after_init = {
2166         .hdrsize        = 0,
2167         .name           = NBD_GENL_FAMILY_NAME,
2168         .version        = NBD_GENL_VERSION,
2169         .module         = THIS_MODULE,
2170         .small_ops      = nbd_connect_genl_ops,
2171         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2172         .maxattr        = NBD_ATTR_MAX,
2173         .policy = nbd_attr_policy,
2174         .mcgrps         = nbd_mcast_grps,
2175         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2176 };
2177
2178 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2179 {
2180         struct nlattr *dev_opt;
2181         u8 connected = 0;
2182         int ret;
2183
2184         /* This is a little racey, but for status it's ok.  The
2185          * reason we don't take a ref here is because we can't
2186          * take a ref in the index == -1 case as we would need
2187          * to put under the nbd_index_mutex, which could
2188          * deadlock if we are configured to remove ourselves
2189          * once we're disconnected.
2190          */
2191         if (refcount_read(&nbd->config_refs))
2192                 connected = 1;
2193         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2194         if (!dev_opt)
2195                 return -EMSGSIZE;
2196         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2197         if (ret)
2198                 return -EMSGSIZE;
2199         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2200                          connected);
2201         if (ret)
2202                 return -EMSGSIZE;
2203         nla_nest_end(reply, dev_opt);
2204         return 0;
2205 }
2206
2207 static int status_cb(int id, void *ptr, void *data)
2208 {
2209         struct nbd_device *nbd = ptr;
2210         return populate_nbd_status(nbd, (struct sk_buff *)data);
2211 }
2212
2213 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2214 {
2215         struct nlattr *dev_list;
2216         struct sk_buff *reply;
2217         void *reply_head;
2218         size_t msg_size;
2219         int index = -1;
2220         int ret = -ENOMEM;
2221
2222         if (info->attrs[NBD_ATTR_INDEX])
2223                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2224
2225         mutex_lock(&nbd_index_mutex);
2226
2227         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2228                                   nla_attr_size(sizeof(u8)));
2229         msg_size *= (index == -1) ? nbd_total_devices : 1;
2230
2231         reply = genlmsg_new(msg_size, GFP_KERNEL);
2232         if (!reply)
2233                 goto out;
2234         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2235                                        NBD_CMD_STATUS);
2236         if (!reply_head) {
2237                 nlmsg_free(reply);
2238                 goto out;
2239         }
2240
2241         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2242         if (index == -1) {
2243                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2244                 if (ret) {
2245                         nlmsg_free(reply);
2246                         goto out;
2247                 }
2248         } else {
2249                 struct nbd_device *nbd;
2250                 nbd = idr_find(&nbd_index_idr, index);
2251                 if (nbd) {
2252                         ret = populate_nbd_status(nbd, reply);
2253                         if (ret) {
2254                                 nlmsg_free(reply);
2255                                 goto out;
2256                         }
2257                 }
2258         }
2259         nla_nest_end(reply, dev_list);
2260         genlmsg_end(reply, reply_head);
2261         ret = genlmsg_reply(reply, info);
2262 out:
2263         mutex_unlock(&nbd_index_mutex);
2264         return ret;
2265 }
2266
2267 static void nbd_connect_reply(struct genl_info *info, int index)
2268 {
2269         struct sk_buff *skb;
2270         void *msg_head;
2271         int ret;
2272
2273         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2274         if (!skb)
2275                 return;
2276         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2277                                      NBD_CMD_CONNECT);
2278         if (!msg_head) {
2279                 nlmsg_free(skb);
2280                 return;
2281         }
2282         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2283         if (ret) {
2284                 nlmsg_free(skb);
2285                 return;
2286         }
2287         genlmsg_end(skb, msg_head);
2288         genlmsg_reply(skb, info);
2289 }
2290
2291 static void nbd_mcast_index(int index)
2292 {
2293         struct sk_buff *skb;
2294         void *msg_head;
2295         int ret;
2296
2297         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2298         if (!skb)
2299                 return;
2300         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2301                                      NBD_CMD_LINK_DEAD);
2302         if (!msg_head) {
2303                 nlmsg_free(skb);
2304                 return;
2305         }
2306         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2307         if (ret) {
2308                 nlmsg_free(skb);
2309                 return;
2310         }
2311         genlmsg_end(skb, msg_head);
2312         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2313 }
2314
2315 static void nbd_dead_link_work(struct work_struct *work)
2316 {
2317         struct link_dead_args *args = container_of(work, struct link_dead_args,
2318                                                    work);
2319         nbd_mcast_index(args->index);
2320         kfree(args);
2321 }
2322
2323 static int __init nbd_init(void)
2324 {
2325         int i;
2326
2327         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2328
2329         if (max_part < 0) {
2330                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2331                 return -EINVAL;
2332         }
2333
2334         part_shift = 0;
2335         if (max_part > 0) {
2336                 part_shift = fls(max_part);
2337
2338                 /*
2339                  * Adjust max_part according to part_shift as it is exported
2340                  * to user space so that user can know the max number of
2341                  * partition kernel should be able to manage.
2342                  *
2343                  * Note that -1 is required because partition 0 is reserved
2344                  * for the whole disk.
2345                  */
2346                 max_part = (1UL << part_shift) - 1;
2347         }
2348
2349         if ((1UL << part_shift) > DISK_MAX_PARTS)
2350                 return -EINVAL;
2351
2352         if (nbds_max > 1UL << (MINORBITS - part_shift))
2353                 return -EINVAL;
2354
2355         if (register_blkdev(NBD_MAJOR, "nbd"))
2356                 return -EIO;
2357
2358         if (genl_register_family(&nbd_genl_family)) {
2359                 unregister_blkdev(NBD_MAJOR, "nbd");
2360                 return -EINVAL;
2361         }
2362         nbd_dbg_init();
2363
2364         mutex_lock(&nbd_index_mutex);
2365         for (i = 0; i < nbds_max; i++)
2366                 nbd_dev_add(i);
2367         mutex_unlock(&nbd_index_mutex);
2368         return 0;
2369 }
2370
2371 static int nbd_exit_cb(int id, void *ptr, void *data)
2372 {
2373         struct list_head *list = (struct list_head *)data;
2374         struct nbd_device *nbd = ptr;
2375
2376         list_add_tail(&nbd->list, list);
2377         return 0;
2378 }
2379
2380 static void __exit nbd_cleanup(void)
2381 {
2382         struct nbd_device *nbd;
2383         LIST_HEAD(del_list);
2384
2385         nbd_dbg_close();
2386
2387         mutex_lock(&nbd_index_mutex);
2388         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2389         mutex_unlock(&nbd_index_mutex);
2390
2391         while (!list_empty(&del_list)) {
2392                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2393                 list_del_init(&nbd->list);
2394                 if (refcount_read(&nbd->refs) != 1)
2395                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2396                 nbd_put(nbd);
2397         }
2398
2399         idr_destroy(&nbd_index_idr);
2400         genl_unregister_family(&nbd_genl_family);
2401         unregister_blkdev(NBD_MAJOR, "nbd");
2402 }
2403
2404 module_init(nbd_init);
2405 module_exit(nbd_cleanup);
2406
2407 MODULE_DESCRIPTION("Network Block Device");
2408 MODULE_LICENSE("GPL");
2409
2410 module_param(nbds_max, int, 0444);
2411 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2412 module_param(max_part, int, 0444);
2413 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");