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