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 void nbd_size_update(struct nbd_device *nbd)
 300 {
 301         struct nbd_config *config = nbd->config;
 302         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 303         sector_t nr_sectors = config->bytesize >> 9;
 304
 305         if (config->flags & NBD_FLAG_SEND_TRIM) {
 306                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 307                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 308                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 309         }
 310         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 311         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 312         set_capacity(nbd->disk, nr_sectors);
 313         if (bdev) {
 314                 if (bdev->bd_disk) {
 315                         bd_set_nr_sectors(bdev, nr_sectors);
 316                         set_blocksize(bdev, config->blksize);
 317                 } else
 318                         set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 319                 bdput(bdev);
 320         }
 321         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 322 }
 323
 324 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 325                          loff_t nr_blocks)
 326 {
 327         struct nbd_config *config = nbd->config;
 328         config->blksize = blocksize;
 329         config->bytesize = blocksize * nr_blocks;
 330         if (nbd->task_recv != NULL)
 331                 nbd_size_update(nbd);
 332 }
 333
 334 static void nbd_complete_rq(struct request *req)
 335 {
 336         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 337
 338         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 339                 cmd->status ? "failed" : "done");
 340
 341         blk_mq_end_request(req, cmd->status);
 342 }
 343
 344 /*
 345  * Forcibly shutdown the socket causing all listeners to error
 346  */
 347 static void sock_shutdown(struct nbd_device *nbd)
 348 {
 349         struct nbd_config *config = nbd->config;
 350         int i;
 351
 352         if (config->num_connections == 0)
 353                 return;
 354         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 355                 return;
 356
 357         for (i = 0; i < config->num_connections; i++) {
 358                 struct nbd_sock *nsock = config->socks[i];
 359                 mutex_lock(&nsock->tx_lock);
 360                 nbd_mark_nsock_dead(nbd, nsock, 0);
 361                 mutex_unlock(&nsock->tx_lock);
 362         }
 363         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 364 }
 365
 366 static u32 req_to_nbd_cmd_type(struct request *req)
 367 {
 368         switch (req_op(req)) {
 369         case REQ_OP_DISCARD:
 370                 return NBD_CMD_TRIM;
 371         case REQ_OP_FLUSH:
 372                 return NBD_CMD_FLUSH;
 373         case REQ_OP_WRITE:
 374                 return NBD_CMD_WRITE;
 375         case REQ_OP_READ:
 376                 return NBD_CMD_READ;
 377         default:
 378                 return U32_MAX;
 379         }
 380 }
 381
 382 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 383                                                  bool reserved)
 384 {
 385         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 386         struct nbd_device *nbd = cmd->nbd;
 387         struct nbd_config *config;
 388
 389         if (!mutex_trylock(&cmd->lock))
 390                 return BLK_EH_RESET_TIMER;
 391
 392         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 393                 cmd->status = BLK_STS_TIMEOUT;
 394                 mutex_unlock(&cmd->lock);
 395                 goto done;
 396         }
 397         config = nbd->config;
 398
 399         if (config->num_connections > 1 ||
 400             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 401                 dev_err_ratelimited(nbd_to_dev(nbd),
 402                                     "Connection timed out, retrying (%d/%d alive)\n",
 403                                     atomic_read(&config->live_connections),
 404                                     config->num_connections);
 405                 /*
 406                  * Hooray we have more connections, requeue this IO, the submit
 407                  * path will put it on a real connection. Or if only one
 408                  * connection is configured, the submit path will wait util
 409                  * a new connection is reconfigured or util dead timeout.
 410                  */
 411                 if (config->socks) {
 412                         if (cmd->index < config->num_connections) {
 413                                 struct nbd_sock *nsock =
 414                                         config->socks[cmd->index];
 415                                 mutex_lock(&nsock->tx_lock);
 416                                 /* We can have multiple outstanding requests, so
 417                                  * we don't want to mark the nsock dead if we've
 418                                  * already reconnected with a new socket, so
 419                                  * only mark it dead if its the same socket we
 420                                  * were sent out on.
 421                                  */
 422                                 if (cmd->cookie == nsock->cookie)
 423                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 424                                 mutex_unlock(&nsock->tx_lock);
 425                         }
 426                         mutex_unlock(&cmd->lock);
 427                         nbd_requeue_cmd(cmd);
 428                         nbd_config_put(nbd);
 429                         return BLK_EH_DONE;
 430                 }
 431         }
 432
 433         if (!nbd->tag_set.timeout) {
 434                 /*
 435                  * Userspace sets timeout=0 to disable socket disconnection,
 436                  * so just warn and reset the timer.
 437                  */
 438                 struct nbd_sock *nsock = config->socks[cmd->index];
 439                 cmd->retries++;
 440                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 441                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 442                         (unsigned long long)blk_rq_pos(req) << 9,
 443                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 444
 445                 mutex_lock(&nsock->tx_lock);
 446                 if (cmd->cookie != nsock->cookie) {
 447                         nbd_requeue_cmd(cmd);
 448                         mutex_unlock(&nsock->tx_lock);
 449                         mutex_unlock(&cmd->lock);
 450                         nbd_config_put(nbd);
 451                         return BLK_EH_DONE;
 452                 }
 453                 mutex_unlock(&nsock->tx_lock);
 454                 mutex_unlock(&cmd->lock);
 455                 nbd_config_put(nbd);
 456                 return BLK_EH_RESET_TIMER;
 457         }
 458
 459         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 460         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 461         cmd->status = BLK_STS_IOERR;
 462         mutex_unlock(&cmd->lock);
 463         sock_shutdown(nbd);
 464         nbd_config_put(nbd);
 465 done:
 466         blk_mq_complete_request(req);
 467         return BLK_EH_DONE;
 468 }
 469
 470 /*
 471  *  Send or receive packet.
 472  */
 473 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 474                      struct iov_iter *iter, int msg_flags, int *sent)
 475 {
 476         struct nbd_config *config = nbd->config;
 477         struct socket *sock = config->socks[index]->sock;
 478         int result;
 479         struct msghdr msg;
 480         unsigned int noreclaim_flag;
 481
 482         if (unlikely(!sock)) {
 483                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 484                         "Attempted %s on closed socket in sock_xmit\n",
 485                         (send ? "send" : "recv"));
 486                 return -EINVAL;
 487         }
 488
 489         msg.msg_iter = *iter;
 490
 491         noreclaim_flag = memalloc_noreclaim_save();
 492         do {
 493                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 494                 msg.msg_name = NULL;
 495                 msg.msg_namelen = 0;
 496                 msg.msg_control = NULL;
 497                 msg.msg_controllen = 0;
 498                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 499
 500                 if (send)
 501                         result = sock_sendmsg(sock, &msg);
 502                 else
 503                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 504
 505                 if (result <= 0) {
 506                         if (result == 0)
 507                                 result = -EPIPE; /* short read */
 508                         break;
 509                 }
 510                 if (sent)
 511                         *sent += result;
 512         } while (msg_data_left(&msg));
 513
 514         memalloc_noreclaim_restore(noreclaim_flag);
 515
 516         return result;
 517 }
 518
 519 /*
 520  * Different settings for sk->sk_sndtimeo can result in different return values
 521  * if there is a signal pending when we enter sendmsg, because reasons?
 522  */
 523 static inline int was_interrupted(int result)
 524 {
 525         return result == -ERESTARTSYS || result == -EINTR;
 526 }
 527
 528 /* always call with the tx_lock held */
 529 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 530 {
 531         struct request *req = blk_mq_rq_from_pdu(cmd);
 532         struct nbd_config *config = nbd->config;
 533         struct nbd_sock *nsock = config->socks[index];
 534         int result;
 535         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 536         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 537         struct iov_iter from;
 538         unsigned long size = blk_rq_bytes(req);
 539         struct bio *bio;
 540         u64 handle;
 541         u32 type;
 542         u32 nbd_cmd_flags = 0;
 543         int sent = nsock->sent, skip = 0;
 544
 545         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 546
 547         type = req_to_nbd_cmd_type(req);
 548         if (type == U32_MAX)
 549                 return -EIO;
 550
 551         if (rq_data_dir(req) == WRITE &&
 552             (config->flags & NBD_FLAG_READ_ONLY)) {
 553                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 554                                     "Write on read-only\n");
 555                 return -EIO;
 556         }
 557
 558         if (req->cmd_flags & REQ_FUA)
 559                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 560
 561         /* We did a partial send previously, and we at least sent the whole
 562          * request struct, so just go and send the rest of the pages in the
 563          * request.
 564          */
 565         if (sent) {
 566                 if (sent >= sizeof(request)) {
 567                         skip = sent - sizeof(request);
 568
 569                         /* initialize handle for tracing purposes */
 570                         handle = nbd_cmd_handle(cmd);
 571
 572                         goto send_pages;
 573                 }
 574                 iov_iter_advance(&from, sent);
 575         } else {
 576                 cmd->cmd_cookie++;
 577         }
 578         cmd->index = index;
 579         cmd->cookie = nsock->cookie;
 580         cmd->retries = 0;
 581         request.type = htonl(type | nbd_cmd_flags);
 582         if (type != NBD_CMD_FLUSH) {
 583                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 584                 request.len = htonl(size);
 585         }
 586         handle = nbd_cmd_handle(cmd);
 587         memcpy(request.handle, &handle, sizeof(handle));
 588
 589         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 590
 591         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 592                 req, nbdcmd_to_ascii(type),
 593                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 594         result = sock_xmit(nbd, index, 1, &from,
 595                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 596         trace_nbd_header_sent(req, handle);
 597         if (result <= 0) {
 598                 if (was_interrupted(result)) {
 599                         /* If we havne't sent anything we can just return BUSY,
 600                          * however if we have sent something we need to make
 601                          * sure we only allow this req to be sent until we are
 602                          * completely done.
 603                          */
 604                         if (sent) {
 605                                 nsock->pending = req;
 606                                 nsock->sent = sent;
 607                         }
 608                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 609                         return BLK_STS_RESOURCE;
 610                 }
 611                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 612                         "Send control failed (result %d)\n", result);
 613                 return -EAGAIN;
 614         }
 615 send_pages:
 616         if (type != NBD_CMD_WRITE)
 617                 goto out;
 618
 619         bio = req->bio;
 620         while (bio) {
 621                 struct bio *next = bio->bi_next;
 622                 struct bvec_iter iter;
 623                 struct bio_vec bvec;
 624
 625                 bio_for_each_segment(bvec, bio, iter) {
 626                         bool is_last = !next && bio_iter_last(bvec, iter);
 627                         int flags = is_last ? 0 : MSG_MORE;
 628
 629                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 630                                 req, bvec.bv_len);
 631                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 632                         if (skip) {
 633                                 if (skip >= iov_iter_count(&from)) {
 634                                         skip -= iov_iter_count(&from);
 635                                         continue;
 636                                 }
 637                                 iov_iter_advance(&from, skip);
 638                                 skip = 0;
 639                         }
 640                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 641                         if (result <= 0) {
 642                                 if (was_interrupted(result)) {
 643                                         /* We've already sent the header, we
 644                                          * have no choice but to set pending and
 645                                          * return BUSY.
 646                                          */
 647                                         nsock->pending = req;
 648                                         nsock->sent = sent;
 649                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 650                                         return BLK_STS_RESOURCE;
 651                                 }
 652                                 dev_err(disk_to_dev(nbd->disk),
 653                                         "Send data failed (result %d)\n",
 654                                         result);
 655                                 return -EAGAIN;
 656                         }
 657                         /*
 658                          * The completion might already have come in,
 659                          * so break for the last one instead of letting
 660                          * the iterator do it. This prevents use-after-free
 661                          * of the bio.
 662                          */
 663                         if (is_last)
 664                                 break;
 665                 }
 666                 bio = next;
 667         }
 668 out:
 669         trace_nbd_payload_sent(req, handle);
 670         nsock->pending = NULL;
 671         nsock->sent = 0;
 672         return 0;
 673 }
 674
 675 /* NULL returned = something went wrong, inform userspace */
 676 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 677 {
 678         struct nbd_config *config = nbd->config;
 679         int result;
 680         struct nbd_reply reply;
 681         struct nbd_cmd *cmd;
 682         struct request *req = NULL;
 683         u64 handle;
 684         u16 hwq;
 685         u32 tag;
 686         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 687         struct iov_iter to;
 688         int ret = 0;
 689
 690         reply.magic = 0;
 691         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 692         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 693         if (result <= 0) {
 694                 if (!nbd_disconnected(config))
 695                         dev_err(disk_to_dev(nbd->disk),
 696                                 "Receive control failed (result %d)\n", result);
 697                 return ERR_PTR(result);
 698         }
 699
 700         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 701                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 702                                 (unsigned long)ntohl(reply.magic));
 703                 return ERR_PTR(-EPROTO);
 704         }
 705
 706         memcpy(&handle, reply.handle, sizeof(handle));
 707         tag = nbd_handle_to_tag(handle);
 708         hwq = blk_mq_unique_tag_to_hwq(tag);
 709         if (hwq < nbd->tag_set.nr_hw_queues)
 710                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 711                                        blk_mq_unique_tag_to_tag(tag));
 712         if (!req || !blk_mq_request_started(req)) {
 713                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 714                         tag, req);
 715                 return ERR_PTR(-ENOENT);
 716         }
 717         trace_nbd_header_received(req, handle);
 718         cmd = blk_mq_rq_to_pdu(req);
 719
 720         mutex_lock(&cmd->lock);
 721         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 722                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 723                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 724                 ret = -ENOENT;
 725                 goto out;
 726         }
 727         if (cmd->status != BLK_STS_OK) {
 728                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 729                         req);
 730                 ret = -ENOENT;
 731                 goto out;
 732         }
 733         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 734                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 735                         req);
 736                 ret = -ENOENT;
 737                 goto out;
 738         }
 739         if (ntohl(reply.error)) {
 740                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 741                         ntohl(reply.error));
 742                 cmd->status = BLK_STS_IOERR;
 743                 goto out;
 744         }
 745
 746         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 747         if (rq_data_dir(req) != WRITE) {
 748                 struct req_iterator iter;
 749                 struct bio_vec bvec;
 750
 751                 rq_for_each_segment(bvec, req, iter) {
 752                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 753                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 754                         if (result <= 0) {
 755                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 756                                         result);
 757                                 /*
 758                                  * If we've disconnected, we need to make sure we
 759                                  * complete this request, otherwise error out
 760                                  * and let the timeout stuff handle resubmitting
 761                                  * this request onto another connection.
 762                                  */
 763                                 if (nbd_disconnected(config)) {
 764                                         cmd->status = BLK_STS_IOERR;
 765                                         goto out;
 766                                 }
 767                                 ret = -EIO;
 768                                 goto out;
 769                         }
 770                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 771                                 req, bvec.bv_len);
 772                 }
 773         }
 774 out:
 775         trace_nbd_payload_received(req, handle);
 776         mutex_unlock(&cmd->lock);
 777         return ret ? ERR_PTR(ret) : cmd;
 778 }
 779
 780 static void recv_work(struct work_struct *work)
 781 {
 782         struct recv_thread_args *args = container_of(work,
 783                                                      struct recv_thread_args,
 784                                                      work);
 785         struct nbd_device *nbd = args->nbd;
 786         struct nbd_config *config = nbd->config;
 787         struct nbd_cmd *cmd;
 788         struct request *rq;
 789
 790         while (1) {
 791                 cmd = nbd_read_stat(nbd, args->index);
 792                 if (IS_ERR(cmd)) {
 793                         struct nbd_sock *nsock = config->socks[args->index];
 794
 795                         mutex_lock(&nsock->tx_lock);
 796                         nbd_mark_nsock_dead(nbd, nsock, 1);
 797                         mutex_unlock(&nsock->tx_lock);
 798                         break;
 799                 }
 800
 801                 rq = blk_mq_rq_from_pdu(cmd);
 802                 if (likely(!blk_should_fake_timeout(rq->q)))
 803                         blk_mq_complete_request(rq);
 804         }
 805         nbd_config_put(nbd);
 806         atomic_dec(&config->recv_threads);
 807         wake_up(&config->recv_wq);
 808         kfree(args);
 809 }
 810
 811 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 812 {
 813         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 814
 815         mutex_lock(&cmd->lock);
 816         cmd->status = BLK_STS_IOERR;
 817         mutex_unlock(&cmd->lock);
 818
 819         blk_mq_complete_request(req);
 820         return true;
 821 }
 822
 823 static void nbd_clear_que(struct nbd_device *nbd)
 824 {
 825         blk_mq_quiesce_queue(nbd->disk->queue);
 826         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 827         blk_mq_unquiesce_queue(nbd->disk->queue);
 828         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 829 }
 830
 831 static int find_fallback(struct nbd_device *nbd, int index)
 832 {
 833         struct nbd_config *config = nbd->config;
 834         int new_index = -1;
 835         struct nbd_sock *nsock = config->socks[index];
 836         int fallback = nsock->fallback_index;
 837
 838         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 839                 return new_index;
 840
 841         if (config->num_connections <= 1) {
 842                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 843                                     "Dead connection, failed to find a fallback\n");
 844                 return new_index;
 845         }
 846
 847         if (fallback >= 0 && fallback < config->num_connections &&
 848             !config->socks[fallback]->dead)
 849                 return fallback;
 850
 851         if (nsock->fallback_index < 0 ||
 852             nsock->fallback_index >= config->num_connections ||
 853             config->socks[nsock->fallback_index]->dead) {
 854                 int i;
 855                 for (i = 0; i < config->num_connections; i++) {
 856                         if (i == index)
 857                                 continue;
 858                         if (!config->socks[i]->dead) {
 859                                 new_index = i;
 860                                 break;
 861                         }
 862                 }
 863                 nsock->fallback_index = new_index;
 864                 if (new_index < 0) {
 865                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 866                                             "Dead connection, failed to find a fallback\n");
 867                         return new_index;
 868                 }
 869         }
 870         new_index = nsock->fallback_index;
 871         return new_index;
 872 }
 873
 874 static int wait_for_reconnect(struct nbd_device *nbd)
 875 {
 876         struct nbd_config *config = nbd->config;
 877         if (!config->dead_conn_timeout)
 878                 return 0;
 879         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 880                 return 0;
 881         return wait_event_timeout(config->conn_wait,
 882                                   atomic_read(&config->live_connections) > 0,
 883                                   config->dead_conn_timeout) > 0;
 884 }
 885
 886 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 887 {
 888         struct request *req = blk_mq_rq_from_pdu(cmd);
 889         struct nbd_device *nbd = cmd->nbd;
 890         struct nbd_config *config;
 891         struct nbd_sock *nsock;
 892         int ret;
 893
 894         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 895                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 896                                     "Socks array is empty\n");
 897                 blk_mq_start_request(req);
 898                 return -EINVAL;
 899         }
 900         config = nbd->config;
 901
 902         if (index >= config->num_connections) {
 903                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 904                                     "Attempted send on invalid socket\n");
 905                 nbd_config_put(nbd);
 906                 blk_mq_start_request(req);
 907                 return -EINVAL;
 908         }
 909         cmd->status = BLK_STS_OK;
 910 again:
 911         nsock = config->socks[index];
 912         mutex_lock(&nsock->tx_lock);
 913         if (nsock->dead) {
 914                 int old_index = index;
 915                 index = find_fallback(nbd, index);
 916                 mutex_unlock(&nsock->tx_lock);
 917                 if (index < 0) {
 918                         if (wait_for_reconnect(nbd)) {
 919                                 index = old_index;
 920                                 goto again;
 921                         }
 922                         /* All the sockets should already be down at this point,
 923                          * we just want to make sure that DISCONNECTED is set so
 924                          * any requests that come in that were queue'ed waiting
 925                          * for the reconnect timer don't trigger the timer again
 926                          * and instead just error out.
 927                          */
 928                         sock_shutdown(nbd);
 929                         nbd_config_put(nbd);
 930                         blk_mq_start_request(req);
 931                         return -EIO;
 932                 }
 933                 goto again;
 934         }
 935
 936         /* Handle the case that we have a pending request that was partially
 937          * transmitted that _has_ to be serviced first.  We need to call requeue
 938          * here so that it gets put _after_ the request that is already on the
 939          * dispatch list.
 940          */
 941         blk_mq_start_request(req);
 942         if (unlikely(nsock->pending && nsock->pending != req)) {
 943                 nbd_requeue_cmd(cmd);
 944                 ret = 0;
 945                 goto out;
 946         }
 947         /*
 948          * Some failures are related to the link going down, so anything that
 949          * returns EAGAIN can be retried on a different socket.
 950          */
 951         ret = nbd_send_cmd(nbd, cmd, index);
 952         if (ret == -EAGAIN) {
 953                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 954                                     "Request send failed, requeueing\n");
 955                 nbd_mark_nsock_dead(nbd, nsock, 1);
 956                 nbd_requeue_cmd(cmd);
 957                 ret = 0;
 958         }
 959 out:
 960         mutex_unlock(&nsock->tx_lock);
 961         nbd_config_put(nbd);
 962         return ret;
 963 }
 964
 965 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 966                         const struct blk_mq_queue_data *bd)
 967 {
 968         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 969         int ret;
 970
 971         /*
 972          * Since we look at the bio's to send the request over the network we
 973          * need to make sure the completion work doesn't mark this request done
 974          * before we are done doing our send.  This keeps us from dereferencing
 975          * freed data if we have particularly fast completions (ie we get the
 976          * completion before we exit sock_xmit on the last bvec) or in the case
 977          * that the server is misbehaving (or there was an error) before we're
 978          * done sending everything over the wire.
 979          */
 980         mutex_lock(&cmd->lock);
 981         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 982
 983         /* We can be called directly from the user space process, which means we
 984          * could possibly have signals pending so our sendmsg will fail.  In
 985          * this case we need to return that we are busy, otherwise error out as
 986          * appropriate.
 987          */
 988         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 989         if (ret < 0)
 990                 ret = BLK_STS_IOERR;
 991         else if (!ret)
 992                 ret = BLK_STS_OK;
 993         mutex_unlock(&cmd->lock);
 994
 995         return ret;
 996 }
 997
 998 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
 999                                      int *err)
1000 {
1001         struct socket *sock;
1002
1003         *err = 0;
1004         sock = sockfd_lookup(fd, err);
1005         if (!sock)
1006                 return NULL;
1007
1008         if (sock->ops->shutdown == sock_no_shutdown) {
1009                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1010                 *err = -EINVAL;
1011                 sockfd_put(sock);
1012                 return NULL;
1013         }
1014
1015         return sock;
1016 }
1017
1018 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1019                           bool netlink)
1020 {
1021         struct nbd_config *config = nbd->config;
1022         struct socket *sock;
1023         struct nbd_sock **socks;
1024         struct nbd_sock *nsock;
1025         int err;
1026
1027         sock = nbd_get_socket(nbd, arg, &err);
1028         if (!sock)
1029                 return err;
1030
1031         if (!netlink && !nbd->task_setup &&
1032             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1033                 nbd->task_setup = current;
1034
1035         if (!netlink &&
1036             (nbd->task_setup != current ||
1037              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1038                 dev_err(disk_to_dev(nbd->disk),
1039                         "Device being setup by another task");
1040                 err = -EBUSY;
1041                 goto put_socket;
1042         }
1043
1044         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1045         if (!nsock) {
1046                 err = -ENOMEM;
1047                 goto put_socket;
1048         }
1049
1050         socks = krealloc(config->socks, (config->num_connections + 1) *
1051                          sizeof(struct nbd_sock *), GFP_KERNEL);
1052         if (!socks) {
1053                 kfree(nsock);
1054                 err = -ENOMEM;
1055                 goto put_socket;
1056         }
1057
1058         config->socks = socks;
1059
1060         nsock->fallback_index = -1;
1061         nsock->dead = false;
1062         mutex_init(&nsock->tx_lock);
1063         nsock->sock = sock;
1064         nsock->pending = NULL;
1065         nsock->sent = 0;
1066         nsock->cookie = 0;
1067         socks[config->num_connections++] = nsock;
1068         atomic_inc(&config->live_connections);
1069
1070         return 0;
1071
1072 put_socket:
1073         sockfd_put(sock);
1074         return err;
1075 }
1076
1077 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1078 {
1079         struct nbd_config *config = nbd->config;
1080         struct socket *sock, *old;
1081         struct recv_thread_args *args;
1082         int i;
1083         int err;
1084
1085         sock = nbd_get_socket(nbd, arg, &err);
1086         if (!sock)
1087                 return err;
1088
1089         args = kzalloc(sizeof(*args), GFP_KERNEL);
1090         if (!args) {
1091                 sockfd_put(sock);
1092                 return -ENOMEM;
1093         }
1094
1095         for (i = 0; i < config->num_connections; i++) {
1096                 struct nbd_sock *nsock = config->socks[i];
1097
1098                 if (!nsock->dead)
1099                         continue;
1100
1101                 mutex_lock(&nsock->tx_lock);
1102                 if (!nsock->dead) {
1103                         mutex_unlock(&nsock->tx_lock);
1104                         continue;
1105                 }
1106                 sk_set_memalloc(sock->sk);
1107                 if (nbd->tag_set.timeout)
1108                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1109                 atomic_inc(&config->recv_threads);
1110                 refcount_inc(&nbd->config_refs);
1111                 old = nsock->sock;
1112                 nsock->fallback_index = -1;
1113                 nsock->sock = sock;
1114                 nsock->dead = false;
1115                 INIT_WORK(&args->work, recv_work);
1116                 args->index = i;
1117                 args->nbd = nbd;
1118                 nsock->cookie++;
1119                 mutex_unlock(&nsock->tx_lock);
1120                 sockfd_put(old);
1121
1122                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1123
1124                 /* We take the tx_mutex in an error path in the recv_work, so we
1125                  * need to queue_work outside of the tx_mutex.
1126                  */
1127                 queue_work(nbd->recv_workq, &args->work);
1128
1129                 atomic_inc(&config->live_connections);
1130                 wake_up(&config->conn_wait);
1131                 return 0;
1132         }
1133         sockfd_put(sock);
1134         kfree(args);
1135         return -ENOSPC;
1136 }
1137
1138 static void nbd_bdev_reset(struct block_device *bdev)
1139 {
1140         if (bdev->bd_openers > 1)
1141                 return;
1142         bd_set_nr_sectors(bdev, 0);
1143 }
1144
1145 static void nbd_parse_flags(struct nbd_device *nbd)
1146 {
1147         struct nbd_config *config = nbd->config;
1148         if (config->flags & NBD_FLAG_READ_ONLY)
1149                 set_disk_ro(nbd->disk, true);
1150         else
1151                 set_disk_ro(nbd->disk, false);
1152         if (config->flags & NBD_FLAG_SEND_TRIM)
1153                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1154         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1155                 if (config->flags & NBD_FLAG_SEND_FUA)
1156                         blk_queue_write_cache(nbd->disk->queue, true, true);
1157                 else
1158                         blk_queue_write_cache(nbd->disk->queue, true, false);
1159         }
1160         else
1161                 blk_queue_write_cache(nbd->disk->queue, false, false);
1162 }
1163
1164 static void send_disconnects(struct nbd_device *nbd)
1165 {
1166         struct nbd_config *config = nbd->config;
1167         struct nbd_request request = {
1168                 .magic = htonl(NBD_REQUEST_MAGIC),
1169                 .type = htonl(NBD_CMD_DISC),
1170         };
1171         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1172         struct iov_iter from;
1173         int i, ret;
1174
1175         for (i = 0; i < config->num_connections; i++) {
1176                 struct nbd_sock *nsock = config->socks[i];
1177
1178                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1179                 mutex_lock(&nsock->tx_lock);
1180                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1181                 if (ret <= 0)
1182                         dev_err(disk_to_dev(nbd->disk),
1183                                 "Send disconnect failed %d\n", ret);
1184                 mutex_unlock(&nsock->tx_lock);
1185         }
1186 }
1187
1188 static int nbd_disconnect(struct nbd_device *nbd)
1189 {
1190         struct nbd_config *config = nbd->config;
1191
1192         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1193         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1194         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1195         send_disconnects(nbd);
1196         return 0;
1197 }
1198
1199 static void nbd_clear_sock(struct nbd_device *nbd)
1200 {
1201         sock_shutdown(nbd);
1202         nbd_clear_que(nbd);
1203         nbd->task_setup = NULL;
1204 }
1205
1206 static void nbd_config_put(struct nbd_device *nbd)
1207 {
1208         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1209                                         &nbd->config_lock)) {
1210                 struct nbd_config *config = nbd->config;
1211                 nbd_dev_dbg_close(nbd);
1212                 nbd_size_clear(nbd);
1213                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1214                                        &config->runtime_flags))
1215                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1216                 nbd->task_recv = NULL;
1217                 nbd_clear_sock(nbd);
1218                 if (config->num_connections) {
1219                         int i;
1220                         for (i = 0; i < config->num_connections; i++) {
1221                                 sockfd_put(config->socks[i]->sock);
1222                                 kfree(config->socks[i]);
1223                         }
1224                         kfree(config->socks);
1225                 }
1226                 kfree(nbd->config);
1227                 nbd->config = NULL;
1228
1229                 if (nbd->recv_workq)
1230                         destroy_workqueue(nbd->recv_workq);
1231                 nbd->recv_workq = NULL;
1232
1233                 nbd->tag_set.timeout = 0;
1234                 nbd->disk->queue->limits.discard_granularity = 0;
1235                 nbd->disk->queue->limits.discard_alignment = 0;
1236                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1237                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1238
1239                 mutex_unlock(&nbd->config_lock);
1240                 nbd_put(nbd);
1241                 module_put(THIS_MODULE);
1242         }
1243 }
1244
1245 static int nbd_start_device(struct nbd_device *nbd)
1246 {
1247         struct nbd_config *config = nbd->config;
1248         int num_connections = config->num_connections;
1249         int error = 0, i;
1250
1251         if (nbd->task_recv)
1252                 return -EBUSY;
1253         if (!config->socks)
1254                 return -EINVAL;
1255         if (num_connections > 1 &&
1256             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1257                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1258                 return -EINVAL;
1259         }
1260
1261         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1262                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1263                                           WQ_UNBOUND, 0, nbd->index);
1264         if (!nbd->recv_workq) {
1265                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1266                 return -ENOMEM;
1267         }
1268
1269         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1270         nbd->task_recv = current;
1271
1272         nbd_parse_flags(nbd);
1273
1274         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1275         if (error) {
1276                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1277                 return error;
1278         }
1279         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1280
1281         nbd_dev_dbg_init(nbd);
1282         for (i = 0; i < num_connections; i++) {
1283                 struct recv_thread_args *args;
1284
1285                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1286                 if (!args) {
1287                         sock_shutdown(nbd);
1288                         /*
1289                          * If num_connections is m (2 < m),
1290                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1291                          * But NO.(n + 1) failed. We still have n recv threads.
1292                          * So, add flush_workqueue here to prevent recv threads
1293                          * dropping the last config_refs and trying to destroy
1294                          * the workqueue from inside the workqueue.
1295                          */
1296                         if (i)
1297                                 flush_workqueue(nbd->recv_workq);
1298                         return -ENOMEM;
1299                 }
1300                 sk_set_memalloc(config->socks[i]->sock->sk);
1301                 if (nbd->tag_set.timeout)
1302                         config->socks[i]->sock->sk->sk_sndtimeo =
1303                                 nbd->tag_set.timeout;
1304                 atomic_inc(&config->recv_threads);
1305                 refcount_inc(&nbd->config_refs);
1306                 INIT_WORK(&args->work, recv_work);
1307                 args->nbd = nbd;
1308                 args->index = i;
1309                 queue_work(nbd->recv_workq, &args->work);
1310         }
1311         nbd_size_update(nbd);
1312         return error;
1313 }
1314
1315 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1316 {
1317         struct nbd_config *config = nbd->config;
1318         int ret;
1319
1320         ret = nbd_start_device(nbd);
1321         if (ret)
1322                 return ret;
1323
1324         if (max_part)
1325                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1326         mutex_unlock(&nbd->config_lock);
1327         ret = wait_event_interruptible(config->recv_wq,
1328                                          atomic_read(&config->recv_threads) == 0);
1329         if (ret)
1330                 sock_shutdown(nbd);
1331         flush_workqueue(nbd->recv_workq);
1332
1333         mutex_lock(&nbd->config_lock);
1334         nbd_bdev_reset(bdev);
1335         /* user requested, ignore socket errors */
1336         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1337                 ret = 0;
1338         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1339                 ret = -ETIMEDOUT;
1340         return ret;
1341 }
1342
1343 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1344                                  struct block_device *bdev)
1345 {
1346         sock_shutdown(nbd);
1347         __invalidate_device(bdev, true);
1348         nbd_bdev_reset(bdev);
1349         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1350                                &nbd->config->runtime_flags))
1351                 nbd_config_put(nbd);
1352 }
1353
1354 static bool nbd_is_valid_blksize(unsigned long blksize)
1355 {
1356         if (!blksize || !is_power_of_2(blksize) || blksize < 512 ||
1357             blksize > PAGE_SIZE)
1358                 return false;
1359         return true;
1360 }
1361
1362 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1363 {
1364         nbd->tag_set.timeout = timeout * HZ;
1365         if (timeout)
1366                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1367         else
1368                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1369 }
1370
1371 /* Must be called with config_lock held */
1372 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1373                        unsigned int cmd, unsigned long arg)
1374 {
1375         struct nbd_config *config = nbd->config;
1376
1377         switch (cmd) {
1378         case NBD_DISCONNECT:
1379                 return nbd_disconnect(nbd);
1380         case NBD_CLEAR_SOCK:
1381                 nbd_clear_sock_ioctl(nbd, bdev);
1382                 return 0;
1383         case NBD_SET_SOCK:
1384                 return nbd_add_socket(nbd, arg, false);
1385         case NBD_SET_BLKSIZE:
1386                 if (!arg)
1387                         arg = NBD_DEF_BLKSIZE;
1388                 if (!nbd_is_valid_blksize(arg))
1389                         return -EINVAL;
1390                 nbd_size_set(nbd, arg,
1391                              div_s64(config->bytesize, arg));
1392                 return 0;
1393         case NBD_SET_SIZE:
1394                 nbd_size_set(nbd, config->blksize,
1395                              div_s64(arg, config->blksize));
1396                 return 0;
1397         case NBD_SET_SIZE_BLOCKS:
1398                 nbd_size_set(nbd, config->blksize, arg);
1399                 return 0;
1400         case NBD_SET_TIMEOUT:
1401                 nbd_set_cmd_timeout(nbd, arg);
1402                 return 0;
1403
1404         case NBD_SET_FLAGS:
1405                 config->flags = arg;
1406                 return 0;
1407         case NBD_DO_IT:
1408                 return nbd_start_device_ioctl(nbd, bdev);
1409         case NBD_CLEAR_QUE:
1410                 /*
1411                  * This is for compatibility only.  The queue is always cleared
1412                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1413                  */
1414                 return 0;
1415         case NBD_PRINT_DEBUG:
1416                 /*
1417                  * For compatibility only, we no longer keep a list of
1418                  * outstanding requests.
1419                  */
1420                 return 0;
1421         }
1422         return -ENOTTY;
1423 }
1424
1425 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1426                      unsigned int cmd, unsigned long arg)
1427 {
1428         struct nbd_device *nbd = bdev->bd_disk->private_data;
1429         struct nbd_config *config = nbd->config;
1430         int error = -EINVAL;
1431
1432         if (!capable(CAP_SYS_ADMIN))
1433                 return -EPERM;
1434
1435         /* The block layer will pass back some non-nbd ioctls in case we have
1436          * special handling for them, but we don't so just return an error.
1437          */
1438         if (_IOC_TYPE(cmd) != 0xab)
1439                 return -EINVAL;
1440
1441         mutex_lock(&nbd->config_lock);
1442
1443         /* Don't allow ioctl operations on a nbd device that was created with
1444          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1445          */
1446         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1447             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1448                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1449         else
1450                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1451         mutex_unlock(&nbd->config_lock);
1452         return error;
1453 }
1454
1455 static struct nbd_config *nbd_alloc_config(void)
1456 {
1457         struct nbd_config *config;
1458
1459         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1460         if (!config)
1461                 return NULL;
1462         atomic_set(&config->recv_threads, 0);
1463         init_waitqueue_head(&config->recv_wq);
1464         init_waitqueue_head(&config->conn_wait);
1465         config->blksize = NBD_DEF_BLKSIZE;
1466         atomic_set(&config->live_connections, 0);
1467         try_module_get(THIS_MODULE);
1468         return config;
1469 }
1470
1471 static int nbd_open(struct block_device *bdev, fmode_t mode)
1472 {
1473         struct nbd_device *nbd;
1474         int ret = 0;
1475
1476         mutex_lock(&nbd_index_mutex);
1477         nbd = bdev->bd_disk->private_data;
1478         if (!nbd) {
1479                 ret = -ENXIO;
1480                 goto out;
1481         }
1482         if (!refcount_inc_not_zero(&nbd->refs)) {
1483                 ret = -ENXIO;
1484                 goto out;
1485         }
1486         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1487                 struct nbd_config *config;
1488
1489                 mutex_lock(&nbd->config_lock);
1490                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1491                         mutex_unlock(&nbd->config_lock);
1492                         goto out;
1493                 }
1494                 config = nbd->config = nbd_alloc_config();
1495                 if (!config) {
1496                         ret = -ENOMEM;
1497                         mutex_unlock(&nbd->config_lock);
1498                         goto out;
1499                 }
1500                 refcount_set(&nbd->config_refs, 1);
1501                 refcount_inc(&nbd->refs);
1502                 mutex_unlock(&nbd->config_lock);
1503                 set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1504         } else if (nbd_disconnected(nbd->config)) {
1505                 set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1506         }
1507 out:
1508         mutex_unlock(&nbd_index_mutex);
1509         return ret;
1510 }
1511
1512 static void nbd_release(struct gendisk *disk, fmode_t mode)
1513 {
1514         struct nbd_device *nbd = disk->private_data;
1515         struct block_device *bdev = bdget_disk(disk, 0);
1516
1517         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1518                         bdev->bd_openers == 0)
1519                 nbd_disconnect_and_put(nbd);
1520
1521         nbd_config_put(nbd);
1522         nbd_put(nbd);
1523 }
1524
1525 static const struct block_device_operations nbd_fops =
1526 {
1527         .owner =        THIS_MODULE,
1528         .open =         nbd_open,
1529         .release =      nbd_release,
1530         .ioctl =        nbd_ioctl,
1531         .compat_ioctl = nbd_ioctl,
1532 };
1533
1534 #if IS_ENABLED(CONFIG_DEBUG_FS)
1535
1536 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1537 {
1538         struct nbd_device *nbd = s->private;
1539
1540         if (nbd->task_recv)
1541                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1542
1543         return 0;
1544 }
1545
1546 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1547 {
1548         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1549 }
1550
1551 static const struct file_operations nbd_dbg_tasks_ops = {
1552         .open = nbd_dbg_tasks_open,
1553         .read = seq_read,
1554         .llseek = seq_lseek,
1555         .release = single_release,
1556 };
1557
1558 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1559 {
1560         struct nbd_device *nbd = s->private;
1561         u32 flags = nbd->config->flags;
1562
1563         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1564
1565         seq_puts(s, "Known flags:\n");
1566
1567         if (flags & NBD_FLAG_HAS_FLAGS)
1568                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1569         if (flags & NBD_FLAG_READ_ONLY)
1570                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1571         if (flags & NBD_FLAG_SEND_FLUSH)
1572                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1573         if (flags & NBD_FLAG_SEND_FUA)
1574                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1575         if (flags & NBD_FLAG_SEND_TRIM)
1576                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1577
1578         return 0;
1579 }
1580
1581 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1582 {
1583         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1584 }
1585
1586 static const struct file_operations nbd_dbg_flags_ops = {
1587         .open = nbd_dbg_flags_open,
1588         .read = seq_read,
1589         .llseek = seq_lseek,
1590         .release = single_release,
1591 };
1592
1593 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1594 {
1595         struct dentry *dir;
1596         struct nbd_config *config = nbd->config;
1597
1598         if (!nbd_dbg_dir)
1599                 return -EIO;
1600
1601         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1602         if (!dir) {
1603                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1604                         nbd_name(nbd));
1605                 return -EIO;
1606         }
1607         config->dbg_dir = dir;
1608
1609         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1610         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1611         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1612         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1613         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1614
1615         return 0;
1616 }
1617
1618 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1619 {
1620         debugfs_remove_recursive(nbd->config->dbg_dir);
1621 }
1622
1623 static int nbd_dbg_init(void)
1624 {
1625         struct dentry *dbg_dir;
1626
1627         dbg_dir = debugfs_create_dir("nbd", NULL);
1628         if (!dbg_dir)
1629                 return -EIO;
1630
1631         nbd_dbg_dir = dbg_dir;
1632
1633         return 0;
1634 }
1635
1636 static void nbd_dbg_close(void)
1637 {
1638         debugfs_remove_recursive(nbd_dbg_dir);
1639 }
1640
1641 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1642
1643 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1644 {
1645         return 0;
1646 }
1647
1648 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1649 {
1650 }
1651
1652 static int nbd_dbg_init(void)
1653 {
1654         return 0;
1655 }
1656
1657 static void nbd_dbg_close(void)
1658 {
1659 }
1660
1661 #endif
1662
1663 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1664                             unsigned int hctx_idx, unsigned int numa_node)
1665 {
1666         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1667         cmd->nbd = set->driver_data;
1668         cmd->flags = 0;
1669         mutex_init(&cmd->lock);
1670         return 0;
1671 }
1672
1673 static const struct blk_mq_ops nbd_mq_ops = {
1674         .queue_rq       = nbd_queue_rq,
1675         .complete       = nbd_complete_rq,
1676         .init_request   = nbd_init_request,
1677         .timeout        = nbd_xmit_timeout,
1678 };
1679
1680 static int nbd_dev_add(int index)
1681 {
1682         struct nbd_device *nbd;
1683         struct gendisk *disk;
1684         struct request_queue *q;
1685         int err = -ENOMEM;
1686
1687         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1688         if (!nbd)
1689                 goto out;
1690
1691         disk = alloc_disk(1 << part_shift);
1692         if (!disk)
1693                 goto out_free_nbd;
1694
1695         if (index >= 0) {
1696                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1697                                 GFP_KERNEL);
1698                 if (err == -ENOSPC)
1699                         err = -EEXIST;
1700         } else {
1701                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1702                 if (err >= 0)
1703                         index = err;
1704         }
1705         if (err < 0)
1706                 goto out_free_disk;
1707
1708         nbd->index = index;
1709         nbd->disk = disk;
1710         nbd->tag_set.ops = &nbd_mq_ops;
1711         nbd->tag_set.nr_hw_queues = 1;
1712         nbd->tag_set.queue_depth = 128;
1713         nbd->tag_set.numa_node = NUMA_NO_NODE;
1714         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1715         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1716                 BLK_MQ_F_BLOCKING;
1717         nbd->tag_set.driver_data = nbd;
1718         nbd->destroy_complete = NULL;
1719
1720         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1721         if (err)
1722                 goto out_free_idr;
1723
1724         q = blk_mq_init_queue(&nbd->tag_set);
1725         if (IS_ERR(q)) {
1726                 err = PTR_ERR(q);
1727                 goto out_free_tags;
1728         }
1729         disk->queue = q;
1730
1731         /*
1732          * Tell the block layer that we are not a rotational device
1733          */
1734         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1735         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1736         disk->queue->limits.discard_granularity = 0;
1737         disk->queue->limits.discard_alignment = 0;
1738         blk_queue_max_discard_sectors(disk->queue, 0);
1739         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1740         blk_queue_max_segments(disk->queue, USHRT_MAX);
1741         blk_queue_max_hw_sectors(disk->queue, 65536);
1742         disk->queue->limits.max_sectors = 256;
1743
1744         mutex_init(&nbd->config_lock);
1745         refcount_set(&nbd->config_refs, 0);
1746         refcount_set(&nbd->refs, 1);
1747         INIT_LIST_HEAD(&nbd->list);
1748         disk->major = NBD_MAJOR;
1749         disk->first_minor = index << part_shift;
1750         disk->fops = &nbd_fops;
1751         disk->private_data = nbd;
1752         sprintf(disk->disk_name, "nbd%d", index);
1753         add_disk(disk);
1754         nbd_total_devices++;
1755         return index;
1756
1757 out_free_tags:
1758         blk_mq_free_tag_set(&nbd->tag_set);
1759 out_free_idr:
1760         idr_remove(&nbd_index_idr, index);
1761 out_free_disk:
1762         put_disk(disk);
1763 out_free_nbd:
1764         kfree(nbd);
1765 out:
1766         return err;
1767 }
1768
1769 static int find_free_cb(int id, void *ptr, void *data)
1770 {
1771         struct nbd_device *nbd = ptr;
1772         struct nbd_device **found = data;
1773
1774         if (!refcount_read(&nbd->config_refs)) {
1775                 *found = nbd;
1776                 return 1;
1777         }
1778         return 0;
1779 }
1780
1781 /* Netlink interface. */
1782 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1783         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1784         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1785         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1786         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1787         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1788         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1789         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1790         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1791         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1792 };
1793
1794 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1795         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1796 };
1797
1798 /* We don't use this right now since we don't parse the incoming list, but we
1799  * still want it here so userspace knows what to expect.
1800  */
1801 static const struct nla_policy __attribute__((unused))
1802 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1803         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1804         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1805 };
1806
1807 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1808 {
1809         struct nbd_config *config = nbd->config;
1810         u64 bsize = config->blksize;
1811         u64 bytes = config->bytesize;
1812
1813         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1814                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1815
1816         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1817                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1818                 if (!bsize)
1819                         bsize = NBD_DEF_BLKSIZE;
1820                 if (!nbd_is_valid_blksize(bsize)) {
1821                         printk(KERN_ERR "Invalid block size %llu\n", bsize);
1822                         return -EINVAL;
1823                 }
1824         }
1825
1826         if (bytes != config->bytesize || bsize != config->blksize)
1827                 nbd_size_set(nbd, bsize, div64_u64(bytes, bsize));
1828         return 0;
1829 }
1830
1831 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1832 {
1833         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1834         struct nbd_device *nbd = NULL;
1835         struct nbd_config *config;
1836         int index = -1;
1837         int ret;
1838         bool put_dev = false;
1839
1840         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1841                 return -EPERM;
1842
1843         if (info->attrs[NBD_ATTR_INDEX])
1844                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1845         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1846                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1847                 return -EINVAL;
1848         }
1849         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1850                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1851                 return -EINVAL;
1852         }
1853 again:
1854         mutex_lock(&nbd_index_mutex);
1855         if (index == -1) {
1856                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1857                 if (ret == 0) {
1858                         int new_index;
1859                         new_index = nbd_dev_add(-1);
1860                         if (new_index < 0) {
1861                                 mutex_unlock(&nbd_index_mutex);
1862                                 printk(KERN_ERR "nbd: failed to add new device\n");
1863                                 return new_index;
1864                         }
1865                         nbd = idr_find(&nbd_index_idr, new_index);
1866                 }
1867         } else {
1868                 nbd = idr_find(&nbd_index_idr, index);
1869                 if (!nbd) {
1870                         ret = nbd_dev_add(index);
1871                         if (ret < 0) {
1872                                 mutex_unlock(&nbd_index_mutex);
1873                                 printk(KERN_ERR "nbd: failed to add new device\n");
1874                                 return ret;
1875                         }
1876                         nbd = idr_find(&nbd_index_idr, index);
1877                 }
1878         }
1879         if (!nbd) {
1880                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1881                        index);
1882                 mutex_unlock(&nbd_index_mutex);
1883                 return -EINVAL;
1884         }
1885
1886         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1887             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1888                 nbd->destroy_complete = &destroy_complete;
1889                 mutex_unlock(&nbd_index_mutex);
1890
1891                 /* Wait untill the the nbd stuff is totally destroyed */
1892                 wait_for_completion(&destroy_complete);
1893                 goto again;
1894         }
1895
1896         if (!refcount_inc_not_zero(&nbd->refs)) {
1897                 mutex_unlock(&nbd_index_mutex);
1898                 if (index == -1)
1899                         goto again;
1900                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1901                        index);
1902                 return -EINVAL;
1903         }
1904         mutex_unlock(&nbd_index_mutex);
1905
1906         mutex_lock(&nbd->config_lock);
1907         if (refcount_read(&nbd->config_refs)) {
1908                 mutex_unlock(&nbd->config_lock);
1909                 nbd_put(nbd);
1910                 if (index == -1)
1911                         goto again;
1912                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1913                 return -EBUSY;
1914         }
1915         if (WARN_ON(nbd->config)) {
1916                 mutex_unlock(&nbd->config_lock);
1917                 nbd_put(nbd);
1918                 return -EINVAL;
1919         }
1920         config = nbd->config = nbd_alloc_config();
1921         if (!nbd->config) {
1922                 mutex_unlock(&nbd->config_lock);
1923                 nbd_put(nbd);
1924                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1925                 return -ENOMEM;
1926         }
1927         refcount_set(&nbd->config_refs, 1);
1928         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1929
1930         ret = nbd_genl_size_set(info, nbd);
1931         if (ret)
1932                 goto out;
1933
1934         if (info->attrs[NBD_ATTR_TIMEOUT])
1935                 nbd_set_cmd_timeout(nbd,
1936                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1937         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1938                 config->dead_conn_timeout =
1939                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1940                 config->dead_conn_timeout *= HZ;
1941         }
1942         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1943                 config->flags =
1944                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1945         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1946                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1947                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1948                         set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
1949                                 &config->runtime_flags);
1950                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1951                         put_dev = true;
1952                 } else {
1953                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1954                 }
1955                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1956                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1957                                 &config->runtime_flags);
1958                 }
1959         }
1960
1961         if (info->attrs[NBD_ATTR_SOCKETS]) {
1962                 struct nlattr *attr;
1963                 int rem, fd;
1964
1965                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1966                                     rem) {
1967                         struct nlattr *socks[NBD_SOCK_MAX+1];
1968
1969                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1970                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1971                                 ret = -EINVAL;
1972                                 goto out;
1973                         }
1974                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1975                                                           attr,
1976                                                           nbd_sock_policy,
1977                                                           info->extack);
1978                         if (ret != 0) {
1979                                 printk(KERN_ERR "nbd: error processing sock list\n");
1980                                 ret = -EINVAL;
1981                                 goto out;
1982                         }
1983                         if (!socks[NBD_SOCK_FD])
1984                                 continue;
1985                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1986                         ret = nbd_add_socket(nbd, fd, true);
1987                         if (ret)
1988                                 goto out;
1989                 }
1990         }
1991         ret = nbd_start_device(nbd);
1992 out:
1993         mutex_unlock(&nbd->config_lock);
1994         if (!ret) {
1995                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
1996                 refcount_inc(&nbd->config_refs);
1997                 nbd_connect_reply(info, nbd->index);
1998         }
1999         nbd_config_put(nbd);
2000         if (put_dev)
2001                 nbd_put(nbd);
2002         return ret;
2003 }
2004
2005 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2006 {
2007         mutex_lock(&nbd->config_lock);
2008         nbd_disconnect(nbd);
2009         nbd_clear_sock(nbd);
2010         mutex_unlock(&nbd->config_lock);
2011         /*
2012          * Make sure recv thread has finished, so it does not drop the last
2013          * config ref and try to destroy the workqueue from inside the work
2014          * queue.
2015          */
2016         flush_workqueue(nbd->recv_workq);
2017         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2018                                &nbd->config->runtime_flags))
2019                 nbd_config_put(nbd);
2020 }
2021
2022 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2023 {
2024         struct nbd_device *nbd;
2025         int index;
2026
2027         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2028                 return -EPERM;
2029
2030         if (!info->attrs[NBD_ATTR_INDEX]) {
2031                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2032                 return -EINVAL;
2033         }
2034         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2035         mutex_lock(&nbd_index_mutex);
2036         nbd = idr_find(&nbd_index_idr, index);
2037         if (!nbd) {
2038                 mutex_unlock(&nbd_index_mutex);
2039                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2040                        index);
2041                 return -EINVAL;
2042         }
2043         if (!refcount_inc_not_zero(&nbd->refs)) {
2044                 mutex_unlock(&nbd_index_mutex);
2045                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2046                        index);
2047                 return -EINVAL;
2048         }
2049         mutex_unlock(&nbd_index_mutex);
2050         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2051                 nbd_put(nbd);
2052                 return 0;
2053         }
2054         nbd_disconnect_and_put(nbd);
2055         nbd_config_put(nbd);
2056         nbd_put(nbd);
2057         return 0;
2058 }
2059
2060 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2061 {
2062         struct nbd_device *nbd = NULL;
2063         struct nbd_config *config;
2064         int index;
2065         int ret = 0;
2066         bool put_dev = false;
2067
2068         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2069                 return -EPERM;
2070
2071         if (!info->attrs[NBD_ATTR_INDEX]) {
2072                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2073                 return -EINVAL;
2074         }
2075         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2076         mutex_lock(&nbd_index_mutex);
2077         nbd = idr_find(&nbd_index_idr, index);
2078         if (!nbd) {
2079                 mutex_unlock(&nbd_index_mutex);
2080                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2081                        index);
2082                 return -EINVAL;
2083         }
2084         if (!refcount_inc_not_zero(&nbd->refs)) {
2085                 mutex_unlock(&nbd_index_mutex);
2086                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2087                        index);
2088                 return -EINVAL;
2089         }
2090         mutex_unlock(&nbd_index_mutex);
2091
2092         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2093                 dev_err(nbd_to_dev(nbd),
2094                         "not configured, cannot reconfigure\n");
2095                 nbd_put(nbd);
2096                 return -EINVAL;
2097         }
2098
2099         mutex_lock(&nbd->config_lock);
2100         config = nbd->config;
2101         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2102             !nbd->task_recv) {
2103                 dev_err(nbd_to_dev(nbd),
2104                         "not configured, cannot reconfigure\n");
2105                 ret = -EINVAL;
2106                 goto out;
2107         }
2108
2109         ret = nbd_genl_size_set(info, nbd);
2110         if (ret)
2111                 goto out;
2112
2113         if (info->attrs[NBD_ATTR_TIMEOUT])
2114                 nbd_set_cmd_timeout(nbd,
2115                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2116         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2117                 config->dead_conn_timeout =
2118                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2119                 config->dead_conn_timeout *= HZ;
2120         }
2121         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2122                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2123                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2124                         if (!test_and_set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2125                                               &config->runtime_flags))
2126                                 put_dev = true;
2127                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2128                 } else {
2129                         if (test_and_clear_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2130                                                &config->runtime_flags))
2131                                 refcount_inc(&nbd->refs);
2132                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2133                 }
2134
2135                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2136                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2137                                         &config->runtime_flags);
2138                 } else {
2139                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2140                                         &config->runtime_flags);
2141                 }
2142         }
2143
2144         if (info->attrs[NBD_ATTR_SOCKETS]) {
2145                 struct nlattr *attr;
2146                 int rem, fd;
2147
2148                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2149                                     rem) {
2150                         struct nlattr *socks[NBD_SOCK_MAX+1];
2151
2152                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2153                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2154                                 ret = -EINVAL;
2155                                 goto out;
2156                         }
2157                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2158                                                           attr,
2159                                                           nbd_sock_policy,
2160                                                           info->extack);
2161                         if (ret != 0) {
2162                                 printk(KERN_ERR "nbd: error processing sock list\n");
2163                                 ret = -EINVAL;
2164                                 goto out;
2165                         }
2166                         if (!socks[NBD_SOCK_FD])
2167                                 continue;
2168                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2169                         ret = nbd_reconnect_socket(nbd, fd);
2170                         if (ret) {
2171                                 if (ret == -ENOSPC)
2172                                         ret = 0;
2173                                 goto out;
2174                         }
2175                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2176                 }
2177         }
2178 out:
2179         mutex_unlock(&nbd->config_lock);
2180         nbd_config_put(nbd);
2181         nbd_put(nbd);
2182         if (put_dev)
2183                 nbd_put(nbd);
2184         return ret;
2185 }
2186
2187 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2188         {
2189                 .cmd    = NBD_CMD_CONNECT,
2190                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2191                 .doit   = nbd_genl_connect,
2192         },
2193         {
2194                 .cmd    = NBD_CMD_DISCONNECT,
2195                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2196                 .doit   = nbd_genl_disconnect,
2197         },
2198         {
2199                 .cmd    = NBD_CMD_RECONFIGURE,
2200                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2201                 .doit   = nbd_genl_reconfigure,
2202         },
2203         {
2204                 .cmd    = NBD_CMD_STATUS,
2205                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2206                 .doit   = nbd_genl_status,
2207         },
2208 };
2209
2210 static const struct genl_multicast_group nbd_mcast_grps[] = {
2211         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2212 };
2213
2214 static struct genl_family nbd_genl_family __ro_after_init = {
2215         .hdrsize        = 0,
2216         .name           = NBD_GENL_FAMILY_NAME,
2217         .version        = NBD_GENL_VERSION,
2218         .module         = THIS_MODULE,
2219         .small_ops      = nbd_connect_genl_ops,
2220         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2221         .maxattr        = NBD_ATTR_MAX,
2222         .policy = nbd_attr_policy,
2223         .mcgrps         = nbd_mcast_grps,
2224         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2225 };
2226
2227 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2228 {
2229         struct nlattr *dev_opt;
2230         u8 connected = 0;
2231         int ret;
2232
2233         /* This is a little racey, but for status it's ok.  The
2234          * reason we don't take a ref here is because we can't
2235          * take a ref in the index == -1 case as we would need
2236          * to put under the nbd_index_mutex, which could
2237          * deadlock if we are configured to remove ourselves
2238          * once we're disconnected.
2239          */
2240         if (refcount_read(&nbd->config_refs))
2241                 connected = 1;
2242         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2243         if (!dev_opt)
2244                 return -EMSGSIZE;
2245         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2246         if (ret)
2247                 return -EMSGSIZE;
2248         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2249                          connected);
2250         if (ret)
2251                 return -EMSGSIZE;
2252         nla_nest_end(reply, dev_opt);
2253         return 0;
2254 }
2255
2256 static int status_cb(int id, void *ptr, void *data)
2257 {
2258         struct nbd_device *nbd = ptr;
2259         return populate_nbd_status(nbd, (struct sk_buff *)data);
2260 }
2261
2262 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2263 {
2264         struct nlattr *dev_list;
2265         struct sk_buff *reply;
2266         void *reply_head;
2267         size_t msg_size;
2268         int index = -1;
2269         int ret = -ENOMEM;
2270
2271         if (info->attrs[NBD_ATTR_INDEX])
2272                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2273
2274         mutex_lock(&nbd_index_mutex);
2275
2276         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2277                                   nla_attr_size(sizeof(u8)));
2278         msg_size *= (index == -1) ? nbd_total_devices : 1;
2279
2280         reply = genlmsg_new(msg_size, GFP_KERNEL);
2281         if (!reply)
2282                 goto out;
2283         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2284                                        NBD_CMD_STATUS);
2285         if (!reply_head) {
2286                 nlmsg_free(reply);
2287                 goto out;
2288         }
2289
2290         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2291         if (index == -1) {
2292                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2293                 if (ret) {
2294                         nlmsg_free(reply);
2295                         goto out;
2296                 }
2297         } else {
2298                 struct nbd_device *nbd;
2299                 nbd = idr_find(&nbd_index_idr, index);
2300                 if (nbd) {
2301                         ret = populate_nbd_status(nbd, reply);
2302                         if (ret) {
2303                                 nlmsg_free(reply);
2304                                 goto out;
2305                         }
2306                 }
2307         }
2308         nla_nest_end(reply, dev_list);
2309         genlmsg_end(reply, reply_head);
2310         ret = genlmsg_reply(reply, info);
2311 out:
2312         mutex_unlock(&nbd_index_mutex);
2313         return ret;
2314 }
2315
2316 static void nbd_connect_reply(struct genl_info *info, int index)
2317 {
2318         struct sk_buff *skb;
2319         void *msg_head;
2320         int ret;
2321
2322         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2323         if (!skb)
2324                 return;
2325         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2326                                      NBD_CMD_CONNECT);
2327         if (!msg_head) {
2328                 nlmsg_free(skb);
2329                 return;
2330         }
2331         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2332         if (ret) {
2333                 nlmsg_free(skb);
2334                 return;
2335         }
2336         genlmsg_end(skb, msg_head);
2337         genlmsg_reply(skb, info);
2338 }
2339
2340 static void nbd_mcast_index(int index)
2341 {
2342         struct sk_buff *skb;
2343         void *msg_head;
2344         int ret;
2345
2346         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2347         if (!skb)
2348                 return;
2349         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2350                                      NBD_CMD_LINK_DEAD);
2351         if (!msg_head) {
2352                 nlmsg_free(skb);
2353                 return;
2354         }
2355         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2356         if (ret) {
2357                 nlmsg_free(skb);
2358                 return;
2359         }
2360         genlmsg_end(skb, msg_head);
2361         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2362 }
2363
2364 static void nbd_dead_link_work(struct work_struct *work)
2365 {
2366         struct link_dead_args *args = container_of(work, struct link_dead_args,
2367                                                    work);
2368         nbd_mcast_index(args->index);
2369         kfree(args);
2370 }
2371
2372 static int __init nbd_init(void)
2373 {
2374         int i;
2375
2376         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2377
2378         if (max_part < 0) {
2379                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2380                 return -EINVAL;
2381         }
2382
2383         part_shift = 0;
2384         if (max_part > 0) {
2385                 part_shift = fls(max_part);
2386
2387                 /*
2388                  * Adjust max_part according to part_shift as it is exported
2389                  * to user space so that user can know the max number of
2390                  * partition kernel should be able to manage.
2391                  *
2392                  * Note that -1 is required because partition 0 is reserved
2393                  * for the whole disk.
2394                  */
2395                 max_part = (1UL << part_shift) - 1;
2396         }
2397
2398         if ((1UL << part_shift) > DISK_MAX_PARTS)
2399                 return -EINVAL;
2400
2401         if (nbds_max > 1UL << (MINORBITS - part_shift))
2402                 return -EINVAL;
2403
2404         if (register_blkdev(NBD_MAJOR, "nbd"))
2405                 return -EIO;
2406
2407         if (genl_register_family(&nbd_genl_family)) {
2408                 unregister_blkdev(NBD_MAJOR, "nbd");
2409                 return -EINVAL;
2410         }
2411         nbd_dbg_init();
2412
2413         mutex_lock(&nbd_index_mutex);
2414         for (i = 0; i < nbds_max; i++)
2415                 nbd_dev_add(i);
2416         mutex_unlock(&nbd_index_mutex);
2417         return 0;
2418 }
2419
2420 static int nbd_exit_cb(int id, void *ptr, void *data)
2421 {
2422         struct list_head *list = (struct list_head *)data;
2423         struct nbd_device *nbd = ptr;
2424
2425         list_add_tail(&nbd->list, list);
2426         return 0;
2427 }
2428
2429 static void __exit nbd_cleanup(void)
2430 {
2431         struct nbd_device *nbd;
2432         LIST_HEAD(del_list);
2433
2434         nbd_dbg_close();
2435
2436         mutex_lock(&nbd_index_mutex);
2437         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2438         mutex_unlock(&nbd_index_mutex);
2439
2440         while (!list_empty(&del_list)) {
2441                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2442                 list_del_init(&nbd->list);
2443                 if (refcount_read(&nbd->refs) != 1)
2444                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2445                 nbd_put(nbd);
2446         }
2447
2448         idr_destroy(&nbd_index_idr);
2449         genl_unregister_family(&nbd_genl_family);
2450         unregister_blkdev(NBD_MAJOR, "nbd");
2451 }
2452
2453 module_init(nbd_init);
2454 module_exit(nbd_cleanup);
2455
2456 MODULE_DESCRIPTION("Network Block Device");
2457 MODULE_LICENSE("GPL");
2458
2459 module_param(nbds_max, int, 0444);
2460 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2461 module_param(max_part, int, 0444);
2462 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");