drivers/block/nbd.c

   1 /*
   2  * Network block device - make block devices work over TCP
   3  *
   4  * Note that you can not swap over this thing, yet. Seems to work but
   5  * deadlocks sometimes - you can not swap over TCP in general.
   6  *
   7  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   8  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
   9  *
  10  * This file is released under GPLv2 or later.
  11  *
  12  * (part of code stolen from loop.c)
  13  */
  14
  15 #include <linux/major.h>
  16
  17 #include <linux/blkdev.h>
  18 #include <linux/module.h>
  19 #include <linux/init.h>
  20 #include <linux/sched.h>
  21 #include <linux/sched/mm.h>
  22 #include <linux/fs.h>
  23 #include <linux/bio.h>
  24 #include <linux/stat.h>
  25 #include <linux/errno.h>
  26 #include <linux/file.h>
  27 #include <linux/ioctl.h>
  28 #include <linux/mutex.h>
  29 #include <linux/compiler.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 static DEFINE_IDR(nbd_index_idr);
  48 static DEFINE_MUTEX(nbd_index_mutex);
  49 static int nbd_total_devices = 0;
  50
  51 struct nbd_sock {
  52         struct socket *sock;
  53         struct mutex tx_lock;
  54         struct request *pending;
  55         int sent;
  56         bool dead;
  57         int fallback_index;
  58         int cookie;
  59 };
  60
  61 struct recv_thread_args {
  62         struct work_struct work;
  63         struct nbd_device *nbd;
  64         int index;
  65 };
  66
  67 struct link_dead_args {
  68         struct work_struct work;
  69         int index;
  70 };
  71
  72 #define NBD_TIMEDOUT                    0
  73 #define NBD_DISCONNECT_REQUESTED        1
  74 #define NBD_DISCONNECTED                2
  75 #define NBD_HAS_PID_FILE                3
  76 #define NBD_HAS_CONFIG_REF              4
  77 #define NBD_BOUND                       5
  78 #define NBD_DESTROY_ON_DISCONNECT       6
  79
  80 struct nbd_config {
  81         u32 flags;
  82         unsigned long runtime_flags;
  83         u64 dead_conn_timeout;
  84
  85         struct nbd_sock **socks;
  86         int num_connections;
  87         atomic_t live_connections;
  88         wait_queue_head_t conn_wait;
  89
  90         atomic_t recv_threads;
  91         wait_queue_head_t recv_wq;
  92         loff_t blksize;
  93         loff_t bytesize;
  94 #if IS_ENABLED(CONFIG_DEBUG_FS)
  95         struct dentry *dbg_dir;
  96 #endif
  97 };
  98
  99 struct nbd_device {
 100         struct blk_mq_tag_set tag_set;
 101
 102         int index;
 103         refcount_t config_refs;
 104         refcount_t refs;
 105         struct nbd_config *config;
 106         struct mutex config_lock;
 107         struct gendisk *disk;
 108
 109         struct list_head list;
 110         struct task_struct *task_recv;
 111         struct task_struct *task_setup;
 112 };
 113
 114 struct nbd_cmd {
 115         struct nbd_device *nbd;
 116         int index;
 117         int cookie;
 118         struct completion send_complete;
 119         blk_status_t status;
 120 };
 121
 122 #if IS_ENABLED(CONFIG_DEBUG_FS)
 123 static struct dentry *nbd_dbg_dir;
 124 #endif
 125
 126 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 127
 128 #define NBD_MAGIC 0x68797548
 129
 130 static unsigned int nbds_max = 16;
 131 static int max_part = 16;
 132 static struct workqueue_struct *recv_workqueue;
 133 static int part_shift;
 134
 135 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 136 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 137 static void nbd_config_put(struct nbd_device *nbd);
 138 static void nbd_connect_reply(struct genl_info *info, int index);
 139 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 140 static void nbd_dead_link_work(struct work_struct *work);
 141
 142 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 143 {
 144         return disk_to_dev(nbd->disk);
 145 }
 146
 147 static const char *nbdcmd_to_ascii(int cmd)
 148 {
 149         switch (cmd) {
 150         case  NBD_CMD_READ: return "read";
 151         case NBD_CMD_WRITE: return "write";
 152         case  NBD_CMD_DISC: return "disconnect";
 153         case NBD_CMD_FLUSH: return "flush";
 154         case  NBD_CMD_TRIM: return "trim/discard";
 155         }
 156         return "invalid";
 157 }
 158
 159 static ssize_t pid_show(struct device *dev,
 160                         struct device_attribute *attr, char *buf)
 161 {
 162         struct gendisk *disk = dev_to_disk(dev);
 163         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 164
 165         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 166 }
 167
 168 static const struct device_attribute pid_attr = {
 169         .attr = { .name = "pid", .mode = 0444},
 170         .show = pid_show,
 171 };
 172
 173 static void nbd_dev_remove(struct nbd_device *nbd)
 174 {
 175         struct gendisk *disk = nbd->disk;
 176         struct request_queue *q;
 177
 178         if (disk) {
 179                 q = disk->queue;
 180                 del_gendisk(disk);
 181                 blk_cleanup_queue(q);
 182                 blk_mq_free_tag_set(&nbd->tag_set);
 183                 disk->private_data = NULL;
 184                 put_disk(disk);
 185         }
 186         kfree(nbd);
 187 }
 188
 189 static void nbd_put(struct nbd_device *nbd)
 190 {
 191         if (refcount_dec_and_mutex_lock(&nbd->refs,
 192                                         &nbd_index_mutex)) {
 193                 idr_remove(&nbd_index_idr, nbd->index);
 194                 mutex_unlock(&nbd_index_mutex);
 195                 nbd_dev_remove(nbd);
 196         }
 197 }
 198
 199 static int nbd_disconnected(struct nbd_config *config)
 200 {
 201         return test_bit(NBD_DISCONNECTED, &config->runtime_flags) ||
 202                 test_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags);
 203 }
 204
 205 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 206                                 int notify)
 207 {
 208         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 209                 struct link_dead_args *args;
 210                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 211                 if (args) {
 212                         INIT_WORK(&args->work, nbd_dead_link_work);
 213                         args->index = nbd->index;
 214                         queue_work(system_wq, &args->work);
 215                 }
 216         }
 217         if (!nsock->dead) {
 218                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 219                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 220                         if (test_and_clear_bit(NBD_DISCONNECT_REQUESTED,
 221                                                &nbd->config->runtime_flags)) {
 222                                 set_bit(NBD_DISCONNECTED,
 223                                         &nbd->config->runtime_flags);
 224                                 dev_info(nbd_to_dev(nbd),
 225                                         "Disconnected due to user request.\n");
 226                         }
 227                 }
 228         }
 229         nsock->dead = true;
 230         nsock->pending = NULL;
 231         nsock->sent = 0;
 232 }
 233
 234 static void nbd_size_clear(struct nbd_device *nbd)
 235 {
 236         if (nbd->config->bytesize) {
 237                 set_capacity(nbd->disk, 0);
 238                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 239         }
 240 }
 241
 242 static void nbd_size_update(struct nbd_device *nbd)
 243 {
 244         struct nbd_config *config = nbd->config;
 245         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 246
 247         if (config->flags & NBD_FLAG_SEND_TRIM) {
 248                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 249                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 250         }
 251         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 252         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 253         set_capacity(nbd->disk, config->bytesize >> 9);
 254         if (bdev) {
 255                 if (bdev->bd_disk)
 256                         bd_set_size(bdev, config->bytesize);
 257                 else
 258                         bdev->bd_invalidated = 1;
 259                 bdput(bdev);
 260         }
 261         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 262 }
 263
 264 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 265                          loff_t nr_blocks)
 266 {
 267         struct nbd_config *config = nbd->config;
 268         config->blksize = blocksize;
 269         config->bytesize = blocksize * nr_blocks;
 270         if (nbd->task_recv != NULL)
 271                 nbd_size_update(nbd);
 272 }
 273
 274 static void nbd_complete_rq(struct request *req)
 275 {
 276         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 277
 278         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", cmd,
 279                 cmd->status ? "failed" : "done");
 280
 281         blk_mq_end_request(req, cmd->status);
 282 }
 283
 284 /*
 285  * Forcibly shutdown the socket causing all listeners to error
 286  */
 287 static void sock_shutdown(struct nbd_device *nbd)
 288 {
 289         struct nbd_config *config = nbd->config;
 290         int i;
 291
 292         if (config->num_connections == 0)
 293                 return;
 294         if (test_and_set_bit(NBD_DISCONNECTED, &config->runtime_flags))
 295                 return;
 296
 297         for (i = 0; i < config->num_connections; i++) {
 298                 struct nbd_sock *nsock = config->socks[i];
 299                 mutex_lock(&nsock->tx_lock);
 300                 nbd_mark_nsock_dead(nbd, nsock, 0);
 301                 mutex_unlock(&nsock->tx_lock);
 302         }
 303         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 304 }
 305
 306 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 307                                                  bool reserved)
 308 {
 309         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 310         struct nbd_device *nbd = cmd->nbd;
 311         struct nbd_config *config;
 312
 313         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 314                 cmd->status = BLK_STS_TIMEOUT;
 315                 goto done;
 316         }
 317         config = nbd->config;
 318
 319         if (config->num_connections > 1) {
 320                 dev_err_ratelimited(nbd_to_dev(nbd),
 321                                     "Connection timed out, retrying (%d/%d alive)\n",
 322                                     atomic_read(&config->live_connections),
 323                                     config->num_connections);
 324                 /*
 325                  * Hooray we have more connections, requeue this IO, the submit
 326                  * path will put it on a real connection.
 327                  */
 328                 if (config->socks && config->num_connections > 1) {
 329                         if (cmd->index < config->num_connections) {
 330                                 struct nbd_sock *nsock =
 331                                         config->socks[cmd->index];
 332                                 mutex_lock(&nsock->tx_lock);
 333                                 /* We can have multiple outstanding requests, so
 334                                  * we don't want to mark the nsock dead if we've
 335                                  * already reconnected with a new socket, so
 336                                  * only mark it dead if its the same socket we
 337                                  * were sent out on.
 338                                  */
 339                                 if (cmd->cookie == nsock->cookie)
 340                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 341                                 mutex_unlock(&nsock->tx_lock);
 342                         }
 343                         blk_mq_requeue_request(req, true);
 344                         nbd_config_put(nbd);
 345                         return BLK_EH_DONE;
 346                 }
 347         } else {
 348                 dev_err_ratelimited(nbd_to_dev(nbd),
 349                                     "Connection timed out\n");
 350         }
 351         set_bit(NBD_TIMEDOUT, &config->runtime_flags);
 352         cmd->status = BLK_STS_IOERR;
 353         sock_shutdown(nbd);
 354         nbd_config_put(nbd);
 355 done:
 356         blk_mq_complete_request(req);
 357         return BLK_EH_DONE;
 358 }
 359
 360 /*
 361  *  Send or receive packet.
 362  */
 363 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 364                      struct iov_iter *iter, int msg_flags, int *sent)
 365 {
 366         struct nbd_config *config = nbd->config;
 367         struct socket *sock = config->socks[index]->sock;
 368         int result;
 369         struct msghdr msg;
 370         unsigned int noreclaim_flag;
 371
 372         if (unlikely(!sock)) {
 373                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 374                         "Attempted %s on closed socket in sock_xmit\n",
 375                         (send ? "send" : "recv"));
 376                 return -EINVAL;
 377         }
 378
 379         msg.msg_iter = *iter;
 380
 381         noreclaim_flag = memalloc_noreclaim_save();
 382         do {
 383                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 384                 msg.msg_name = NULL;
 385                 msg.msg_namelen = 0;
 386                 msg.msg_control = NULL;
 387                 msg.msg_controllen = 0;
 388                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 389
 390                 if (send)
 391                         result = sock_sendmsg(sock, &msg);
 392                 else
 393                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 394
 395                 if (result <= 0) {
 396                         if (result == 0)
 397                                 result = -EPIPE; /* short read */
 398                         break;
 399                 }
 400                 if (sent)
 401                         *sent += result;
 402         } while (msg_data_left(&msg));
 403
 404         memalloc_noreclaim_restore(noreclaim_flag);
 405
 406         return result;
 407 }
 408
 409 /*
 410  * Different settings for sk->sk_sndtimeo can result in different return values
 411  * if there is a signal pending when we enter sendmsg, because reasons?
 412  */
 413 static inline int was_interrupted(int result)
 414 {
 415         return result == -ERESTARTSYS || result == -EINTR;
 416 }
 417
 418 /* always call with the tx_lock held */
 419 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 420 {
 421         struct request *req = blk_mq_rq_from_pdu(cmd);
 422         struct nbd_config *config = nbd->config;
 423         struct nbd_sock *nsock = config->socks[index];
 424         int result;
 425         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 426         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 427         struct iov_iter from;
 428         unsigned long size = blk_rq_bytes(req);
 429         struct bio *bio;
 430         u32 type;
 431         u32 nbd_cmd_flags = 0;
 432         u32 tag = blk_mq_unique_tag(req);
 433         int sent = nsock->sent, skip = 0;
 434
 435         iov_iter_kvec(&from, WRITE | ITER_KVEC, &iov, 1, sizeof(request));
 436
 437         switch (req_op(req)) {
 438         case REQ_OP_DISCARD:
 439                 type = NBD_CMD_TRIM;
 440                 break;
 441         case REQ_OP_FLUSH:
 442                 type = NBD_CMD_FLUSH;
 443                 break;
 444         case REQ_OP_WRITE:
 445                 type = NBD_CMD_WRITE;
 446                 break;
 447         case REQ_OP_READ:
 448                 type = NBD_CMD_READ;
 449                 break;
 450         default:
 451                 return -EIO;
 452         }
 453
 454         if (rq_data_dir(req) == WRITE &&
 455             (config->flags & NBD_FLAG_READ_ONLY)) {
 456                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 457                                     "Write on read-only\n");
 458                 return -EIO;
 459         }
 460
 461         if (req->cmd_flags & REQ_FUA)
 462                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 463
 464         /* We did a partial send previously, and we at least sent the whole
 465          * request struct, so just go and send the rest of the pages in the
 466          * request.
 467          */
 468         if (sent) {
 469                 if (sent >= sizeof(request)) {
 470                         skip = sent - sizeof(request);
 471                         goto send_pages;
 472                 }
 473                 iov_iter_advance(&from, sent);
 474         }
 475         cmd->index = index;
 476         cmd->cookie = nsock->cookie;
 477         request.type = htonl(type | nbd_cmd_flags);
 478         if (type != NBD_CMD_FLUSH) {
 479                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 480                 request.len = htonl(size);
 481         }
 482         memcpy(request.handle, &tag, sizeof(tag));
 483
 484         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 485                 cmd, nbdcmd_to_ascii(type),
 486                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 487         result = sock_xmit(nbd, index, 1, &from,
 488                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 489         if (result <= 0) {
 490                 if (was_interrupted(result)) {
 491                         /* If we havne't sent anything we can just return BUSY,
 492                          * however if we have sent something we need to make
 493                          * sure we only allow this req to be sent until we are
 494                          * completely done.
 495                          */
 496                         if (sent) {
 497                                 nsock->pending = req;
 498                                 nsock->sent = sent;
 499                         }
 500                         return BLK_STS_RESOURCE;
 501                 }
 502                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 503                         "Send control failed (result %d)\n", result);
 504                 return -EAGAIN;
 505         }
 506 send_pages:
 507         if (type != NBD_CMD_WRITE)
 508                 goto out;
 509
 510         bio = req->bio;
 511         while (bio) {
 512                 struct bio *next = bio->bi_next;
 513                 struct bvec_iter iter;
 514                 struct bio_vec bvec;
 515
 516                 bio_for_each_segment(bvec, bio, iter) {
 517                         bool is_last = !next && bio_iter_last(bvec, iter);
 518                         int flags = is_last ? 0 : MSG_MORE;
 519
 520                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 521                                 cmd, bvec.bv_len);
 522                         iov_iter_bvec(&from, ITER_BVEC | WRITE,
 523                                       &bvec, 1, bvec.bv_len);
 524                         if (skip) {
 525                                 if (skip >= iov_iter_count(&from)) {
 526                                         skip -= iov_iter_count(&from);
 527                                         continue;
 528                                 }
 529                                 iov_iter_advance(&from, skip);
 530                                 skip = 0;
 531                         }
 532                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 533                         if (result <= 0) {
 534                                 if (was_interrupted(result)) {
 535                                         /* We've already sent the header, we
 536                                          * have no choice but to set pending and
 537                                          * return BUSY.
 538                                          */
 539                                         nsock->pending = req;
 540                                         nsock->sent = sent;
 541                                         return BLK_STS_RESOURCE;
 542                                 }
 543                                 dev_err(disk_to_dev(nbd->disk),
 544                                         "Send data failed (result %d)\n",
 545                                         result);
 546                                 return -EAGAIN;
 547                         }
 548                         /*
 549                          * The completion might already have come in,
 550                          * so break for the last one instead of letting
 551                          * the iterator do it. This prevents use-after-free
 552                          * of the bio.
 553                          */
 554                         if (is_last)
 555                                 break;
 556                 }
 557                 bio = next;
 558         }
 559 out:
 560         nsock->pending = NULL;
 561         nsock->sent = 0;
 562         return 0;
 563 }
 564
 565 /* NULL returned = something went wrong, inform userspace */
 566 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 567 {
 568         struct nbd_config *config = nbd->config;
 569         int result;
 570         struct nbd_reply reply;
 571         struct nbd_cmd *cmd;
 572         struct request *req = NULL;
 573         u16 hwq;
 574         u32 tag;
 575         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 576         struct iov_iter to;
 577
 578         reply.magic = 0;
 579         iov_iter_kvec(&to, READ | ITER_KVEC, &iov, 1, sizeof(reply));
 580         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 581         if (result <= 0) {
 582                 if (!nbd_disconnected(config))
 583                         dev_err(disk_to_dev(nbd->disk),
 584                                 "Receive control failed (result %d)\n", result);
 585                 return ERR_PTR(result);
 586         }
 587
 588         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 589                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 590                                 (unsigned long)ntohl(reply.magic));
 591                 return ERR_PTR(-EPROTO);
 592         }
 593
 594         memcpy(&tag, reply.handle, sizeof(u32));
 595
 596         hwq = blk_mq_unique_tag_to_hwq(tag);
 597         if (hwq < nbd->tag_set.nr_hw_queues)
 598                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 599                                        blk_mq_unique_tag_to_tag(tag));
 600         if (!req || !blk_mq_request_started(req)) {
 601                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 602                         tag, req);
 603                 return ERR_PTR(-ENOENT);
 604         }
 605         cmd = blk_mq_rq_to_pdu(req);
 606         if (ntohl(reply.error)) {
 607                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 608                         ntohl(reply.error));
 609                 cmd->status = BLK_STS_IOERR;
 610                 return cmd;
 611         }
 612
 613         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", cmd);
 614         if (rq_data_dir(req) != WRITE) {
 615                 struct req_iterator iter;
 616                 struct bio_vec bvec;
 617
 618                 rq_for_each_segment(bvec, req, iter) {
 619                         iov_iter_bvec(&to, ITER_BVEC | READ,
 620                                       &bvec, 1, bvec.bv_len);
 621                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 622                         if (result <= 0) {
 623                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 624                                         result);
 625                                 /*
 626                                  * If we've disconnected or we only have 1
 627                                  * connection then we need to make sure we
 628                                  * complete this request, otherwise error out
 629                                  * and let the timeout stuff handle resubmitting
 630                                  * this request onto another connection.
 631                                  */
 632                                 if (nbd_disconnected(config) ||
 633                                     config->num_connections <= 1) {
 634                                         cmd->status = BLK_STS_IOERR;
 635                                         return cmd;
 636                                 }
 637                                 return ERR_PTR(-EIO);
 638                         }
 639                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 640                                 cmd, bvec.bv_len);
 641                 }
 642         } else {
 643                 /* See the comment in nbd_queue_rq. */
 644                 wait_for_completion(&cmd->send_complete);
 645         }
 646         return cmd;
 647 }
 648
 649 static void recv_work(struct work_struct *work)
 650 {
 651         struct recv_thread_args *args = container_of(work,
 652                                                      struct recv_thread_args,
 653                                                      work);
 654         struct nbd_device *nbd = args->nbd;
 655         struct nbd_config *config = nbd->config;
 656         struct nbd_cmd *cmd;
 657
 658         while (1) {
 659                 cmd = nbd_read_stat(nbd, args->index);
 660                 if (IS_ERR(cmd)) {
 661                         struct nbd_sock *nsock = config->socks[args->index];
 662
 663                         mutex_lock(&nsock->tx_lock);
 664                         nbd_mark_nsock_dead(nbd, nsock, 1);
 665                         mutex_unlock(&nsock->tx_lock);
 666                         break;
 667                 }
 668
 669                 blk_mq_complete_request(blk_mq_rq_from_pdu(cmd));
 670         }
 671         atomic_dec(&config->recv_threads);
 672         wake_up(&config->recv_wq);
 673         nbd_config_put(nbd);
 674         kfree(args);
 675 }
 676
 677 static void nbd_clear_req(struct request *req, void *data, bool reserved)
 678 {
 679         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 680
 681         cmd->status = BLK_STS_IOERR;
 682         blk_mq_complete_request(req);
 683 }
 684
 685 static void nbd_clear_que(struct nbd_device *nbd)
 686 {
 687         blk_mq_quiesce_queue(nbd->disk->queue);
 688         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 689         blk_mq_unquiesce_queue(nbd->disk->queue);
 690         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 691 }
 692
 693 static int find_fallback(struct nbd_device *nbd, int index)
 694 {
 695         struct nbd_config *config = nbd->config;
 696         int new_index = -1;
 697         struct nbd_sock *nsock = config->socks[index];
 698         int fallback = nsock->fallback_index;
 699
 700         if (test_bit(NBD_DISCONNECTED, &config->runtime_flags))
 701                 return new_index;
 702
 703         if (config->num_connections <= 1) {
 704                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 705                                     "Attempted send on invalid socket\n");
 706                 return new_index;
 707         }
 708
 709         if (fallback >= 0 && fallback < config->num_connections &&
 710             !config->socks[fallback]->dead)
 711                 return fallback;
 712
 713         if (nsock->fallback_index < 0 ||
 714             nsock->fallback_index >= config->num_connections ||
 715             config->socks[nsock->fallback_index]->dead) {
 716                 int i;
 717                 for (i = 0; i < config->num_connections; i++) {
 718                         if (i == index)
 719                                 continue;
 720                         if (!config->socks[i]->dead) {
 721                                 new_index = i;
 722                                 break;
 723                         }
 724                 }
 725                 nsock->fallback_index = new_index;
 726                 if (new_index < 0) {
 727                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 728                                             "Dead connection, failed to find a fallback\n");
 729                         return new_index;
 730                 }
 731         }
 732         new_index = nsock->fallback_index;
 733         return new_index;
 734 }
 735
 736 static int wait_for_reconnect(struct nbd_device *nbd)
 737 {
 738         struct nbd_config *config = nbd->config;
 739         if (!config->dead_conn_timeout)
 740                 return 0;
 741         if (test_bit(NBD_DISCONNECTED, &config->runtime_flags))
 742                 return 0;
 743         return wait_event_timeout(config->conn_wait,
 744                                   atomic_read(&config->live_connections) > 0,
 745                                   config->dead_conn_timeout) > 0;
 746 }
 747
 748 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 749 {
 750         struct request *req = blk_mq_rq_from_pdu(cmd);
 751         struct nbd_device *nbd = cmd->nbd;
 752         struct nbd_config *config;
 753         struct nbd_sock *nsock;
 754         int ret;
 755
 756         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 757                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 758                                     "Socks array is empty\n");
 759                 blk_mq_start_request(req);
 760                 return -EINVAL;
 761         }
 762         config = nbd->config;
 763
 764         if (index >= config->num_connections) {
 765                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 766                                     "Attempted send on invalid socket\n");
 767                 nbd_config_put(nbd);
 768                 blk_mq_start_request(req);
 769                 return -EINVAL;
 770         }
 771         cmd->status = BLK_STS_OK;
 772 again:
 773         nsock = config->socks[index];
 774         mutex_lock(&nsock->tx_lock);
 775         if (nsock->dead) {
 776                 int old_index = index;
 777                 index = find_fallback(nbd, index);
 778                 mutex_unlock(&nsock->tx_lock);
 779                 if (index < 0) {
 780                         if (wait_for_reconnect(nbd)) {
 781                                 index = old_index;
 782                                 goto again;
 783                         }
 784                         /* All the sockets should already be down at this point,
 785                          * we just want to make sure that DISCONNECTED is set so
 786                          * any requests that come in that were queue'ed waiting
 787                          * for the reconnect timer don't trigger the timer again
 788                          * and instead just error out.
 789                          */
 790                         sock_shutdown(nbd);
 791                         nbd_config_put(nbd);
 792                         blk_mq_start_request(req);
 793                         return -EIO;
 794                 }
 795                 goto again;
 796         }
 797
 798         /* Handle the case that we have a pending request that was partially
 799          * transmitted that _has_ to be serviced first.  We need to call requeue
 800          * here so that it gets put _after_ the request that is already on the
 801          * dispatch list.
 802          */
 803         blk_mq_start_request(req);
 804         if (unlikely(nsock->pending && nsock->pending != req)) {
 805                 blk_mq_requeue_request(req, true);
 806                 ret = 0;
 807                 goto out;
 808         }
 809         /*
 810          * Some failures are related to the link going down, so anything that
 811          * returns EAGAIN can be retried on a different socket.
 812          */
 813         ret = nbd_send_cmd(nbd, cmd, index);
 814         if (ret == -EAGAIN) {
 815                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 816                                     "Request send failed, requeueing\n");
 817                 nbd_mark_nsock_dead(nbd, nsock, 1);
 818                 blk_mq_requeue_request(req, true);
 819                 ret = 0;
 820         }
 821 out:
 822         mutex_unlock(&nsock->tx_lock);
 823         nbd_config_put(nbd);
 824         return ret;
 825 }
 826
 827 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 828                         const struct blk_mq_queue_data *bd)
 829 {
 830         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 831         int ret;
 832
 833         /*
 834          * Since we look at the bio's to send the request over the network we
 835          * need to make sure the completion work doesn't mark this request done
 836          * before we are done doing our send.  This keeps us from dereferencing
 837          * freed data if we have particularly fast completions (ie we get the
 838          * completion before we exit sock_xmit on the last bvec) or in the case
 839          * that the server is misbehaving (or there was an error) before we're
 840          * done sending everything over the wire.
 841          */
 842         init_completion(&cmd->send_complete);
 843
 844         /* We can be called directly from the user space process, which means we
 845          * could possibly have signals pending so our sendmsg will fail.  In
 846          * this case we need to return that we are busy, otherwise error out as
 847          * appropriate.
 848          */
 849         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 850         if (ret < 0)
 851                 ret = BLK_STS_IOERR;
 852         else if (!ret)
 853                 ret = BLK_STS_OK;
 854         complete(&cmd->send_complete);
 855
 856         return ret;
 857 }
 858
 859 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
 860                           bool netlink)
 861 {
 862         struct nbd_config *config = nbd->config;
 863         struct socket *sock;
 864         struct nbd_sock **socks;
 865         struct nbd_sock *nsock;
 866         int err;
 867
 868         sock = sockfd_lookup(arg, &err);
 869         if (!sock)
 870                 return err;
 871
 872         if (!netlink && !nbd->task_setup &&
 873             !test_bit(NBD_BOUND, &config->runtime_flags))
 874                 nbd->task_setup = current;
 875
 876         if (!netlink &&
 877             (nbd->task_setup != current ||
 878              test_bit(NBD_BOUND, &config->runtime_flags))) {
 879                 dev_err(disk_to_dev(nbd->disk),
 880                         "Device being setup by another task");
 881                 sockfd_put(sock);
 882                 return -EBUSY;
 883         }
 884
 885         socks = krealloc(config->socks, (config->num_connections + 1) *
 886                          sizeof(struct nbd_sock *), GFP_KERNEL);
 887         if (!socks) {
 888                 sockfd_put(sock);
 889                 return -ENOMEM;
 890         }
 891         nsock = kzalloc(sizeof(struct nbd_sock), GFP_KERNEL);
 892         if (!nsock) {
 893                 sockfd_put(sock);
 894                 return -ENOMEM;
 895         }
 896
 897         config->socks = socks;
 898
 899         nsock->fallback_index = -1;
 900         nsock->dead = false;
 901         mutex_init(&nsock->tx_lock);
 902         nsock->sock = sock;
 903         nsock->pending = NULL;
 904         nsock->sent = 0;
 905         nsock->cookie = 0;
 906         socks[config->num_connections++] = nsock;
 907         atomic_inc(&config->live_connections);
 908
 909         return 0;
 910 }
 911
 912 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
 913 {
 914         struct nbd_config *config = nbd->config;
 915         struct socket *sock, *old;
 916         struct recv_thread_args *args;
 917         int i;
 918         int err;
 919
 920         sock = sockfd_lookup(arg, &err);
 921         if (!sock)
 922                 return err;
 923
 924         args = kzalloc(sizeof(*args), GFP_KERNEL);
 925         if (!args) {
 926                 sockfd_put(sock);
 927                 return -ENOMEM;
 928         }
 929
 930         for (i = 0; i < config->num_connections; i++) {
 931                 struct nbd_sock *nsock = config->socks[i];
 932
 933                 if (!nsock->dead)
 934                         continue;
 935
 936                 mutex_lock(&nsock->tx_lock);
 937                 if (!nsock->dead) {
 938                         mutex_unlock(&nsock->tx_lock);
 939                         continue;
 940                 }
 941                 sk_set_memalloc(sock->sk);
 942                 if (nbd->tag_set.timeout)
 943                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
 944                 atomic_inc(&config->recv_threads);
 945                 refcount_inc(&nbd->config_refs);
 946                 old = nsock->sock;
 947                 nsock->fallback_index = -1;
 948                 nsock->sock = sock;
 949                 nsock->dead = false;
 950                 INIT_WORK(&args->work, recv_work);
 951                 args->index = i;
 952                 args->nbd = nbd;
 953                 nsock->cookie++;
 954                 mutex_unlock(&nsock->tx_lock);
 955                 sockfd_put(old);
 956
 957                 clear_bit(NBD_DISCONNECTED, &config->runtime_flags);
 958
 959                 /* We take the tx_mutex in an error path in the recv_work, so we
 960                  * need to queue_work outside of the tx_mutex.
 961                  */
 962                 queue_work(recv_workqueue, &args->work);
 963
 964                 atomic_inc(&config->live_connections);
 965                 wake_up(&config->conn_wait);
 966                 return 0;
 967         }
 968         sockfd_put(sock);
 969         kfree(args);
 970         return -ENOSPC;
 971 }
 972
 973 static void nbd_bdev_reset(struct block_device *bdev)
 974 {
 975         if (bdev->bd_openers > 1)
 976                 return;
 977         bd_set_size(bdev, 0);
 978 }
 979
 980 static void nbd_parse_flags(struct nbd_device *nbd)
 981 {
 982         struct nbd_config *config = nbd->config;
 983         if (config->flags & NBD_FLAG_READ_ONLY)
 984                 set_disk_ro(nbd->disk, true);
 985         else
 986                 set_disk_ro(nbd->disk, false);
 987         if (config->flags & NBD_FLAG_SEND_TRIM)
 988                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 989         if (config->flags & NBD_FLAG_SEND_FLUSH) {
 990                 if (config->flags & NBD_FLAG_SEND_FUA)
 991                         blk_queue_write_cache(nbd->disk->queue, true, true);
 992                 else
 993                         blk_queue_write_cache(nbd->disk->queue, true, false);
 994         }
 995         else
 996                 blk_queue_write_cache(nbd->disk->queue, false, false);
 997 }
 998
 999 static void send_disconnects(struct nbd_device *nbd)
1000 {
1001         struct nbd_config *config = nbd->config;
1002         struct nbd_request request = {
1003                 .magic = htonl(NBD_REQUEST_MAGIC),
1004                 .type = htonl(NBD_CMD_DISC),
1005         };
1006         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1007         struct iov_iter from;
1008         int i, ret;
1009
1010         for (i = 0; i < config->num_connections; i++) {
1011                 struct nbd_sock *nsock = config->socks[i];
1012
1013                 iov_iter_kvec(&from, WRITE | ITER_KVEC, &iov, 1, sizeof(request));
1014                 mutex_lock(&nsock->tx_lock);
1015                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1016                 if (ret <= 0)
1017                         dev_err(disk_to_dev(nbd->disk),
1018                                 "Send disconnect failed %d\n", ret);
1019                 mutex_unlock(&nsock->tx_lock);
1020         }
1021 }
1022
1023 static int nbd_disconnect(struct nbd_device *nbd)
1024 {
1025         struct nbd_config *config = nbd->config;
1026
1027         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1028         set_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags);
1029         send_disconnects(nbd);
1030         return 0;
1031 }
1032
1033 static void nbd_clear_sock(struct nbd_device *nbd)
1034 {
1035         sock_shutdown(nbd);
1036         nbd_clear_que(nbd);
1037         nbd->task_setup = NULL;
1038 }
1039
1040 static void nbd_config_put(struct nbd_device *nbd)
1041 {
1042         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1043                                         &nbd->config_lock)) {
1044                 struct nbd_config *config = nbd->config;
1045                 nbd_dev_dbg_close(nbd);
1046                 nbd_size_clear(nbd);
1047                 if (test_and_clear_bit(NBD_HAS_PID_FILE,
1048                                        &config->runtime_flags))
1049                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1050                 nbd->task_recv = NULL;
1051                 nbd_clear_sock(nbd);
1052                 if (config->num_connections) {
1053                         int i;
1054                         for (i = 0; i < config->num_connections; i++) {
1055                                 sockfd_put(config->socks[i]->sock);
1056                                 kfree(config->socks[i]);
1057                         }
1058                         kfree(config->socks);
1059                 }
1060                 kfree(nbd->config);
1061                 nbd->config = NULL;
1062
1063                 nbd->tag_set.timeout = 0;
1064                 nbd->disk->queue->limits.discard_granularity = 0;
1065                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1066                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1067
1068                 mutex_unlock(&nbd->config_lock);
1069                 nbd_put(nbd);
1070                 module_put(THIS_MODULE);
1071         }
1072 }
1073
1074 static int nbd_start_device(struct nbd_device *nbd)
1075 {
1076         struct nbd_config *config = nbd->config;
1077         int num_connections = config->num_connections;
1078         int error = 0, i;
1079
1080         if (nbd->task_recv)
1081                 return -EBUSY;
1082         if (!config->socks)
1083                 return -EINVAL;
1084         if (num_connections > 1 &&
1085             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1086                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1087                 return -EINVAL;
1088         }
1089
1090         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1091         nbd->task_recv = current;
1092
1093         nbd_parse_flags(nbd);
1094
1095         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1096         if (error) {
1097                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1098                 return error;
1099         }
1100         set_bit(NBD_HAS_PID_FILE, &config->runtime_flags);
1101
1102         nbd_dev_dbg_init(nbd);
1103         for (i = 0; i < num_connections; i++) {
1104                 struct recv_thread_args *args;
1105
1106                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1107                 if (!args) {
1108                         sock_shutdown(nbd);
1109                         return -ENOMEM;
1110                 }
1111                 sk_set_memalloc(config->socks[i]->sock->sk);
1112                 if (nbd->tag_set.timeout)
1113                         config->socks[i]->sock->sk->sk_sndtimeo =
1114                                 nbd->tag_set.timeout;
1115                 atomic_inc(&config->recv_threads);
1116                 refcount_inc(&nbd->config_refs);
1117                 INIT_WORK(&args->work, recv_work);
1118                 args->nbd = nbd;
1119                 args->index = i;
1120                 queue_work(recv_workqueue, &args->work);
1121         }
1122         nbd_size_update(nbd);
1123         return error;
1124 }
1125
1126 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1127 {
1128         struct nbd_config *config = nbd->config;
1129         int ret;
1130
1131         ret = nbd_start_device(nbd);
1132         if (ret)
1133                 return ret;
1134
1135         if (max_part)
1136                 bdev->bd_invalidated = 1;
1137         mutex_unlock(&nbd->config_lock);
1138         ret = wait_event_interruptible(config->recv_wq,
1139                                          atomic_read(&config->recv_threads) == 0);
1140         if (ret)
1141                 sock_shutdown(nbd);
1142         mutex_lock(&nbd->config_lock);
1143         nbd_bdev_reset(bdev);
1144         /* user requested, ignore socket errors */
1145         if (test_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags))
1146                 ret = 0;
1147         if (test_bit(NBD_TIMEDOUT, &config->runtime_flags))
1148                 ret = -ETIMEDOUT;
1149         return ret;
1150 }
1151
1152 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1153                                  struct block_device *bdev)
1154 {
1155         sock_shutdown(nbd);
1156         kill_bdev(bdev);
1157         nbd_bdev_reset(bdev);
1158         if (test_and_clear_bit(NBD_HAS_CONFIG_REF,
1159                                &nbd->config->runtime_flags))
1160                 nbd_config_put(nbd);
1161 }
1162
1163 /* Must be called with config_lock held */
1164 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1165                        unsigned int cmd, unsigned long arg)
1166 {
1167         struct nbd_config *config = nbd->config;
1168
1169         switch (cmd) {
1170         case NBD_DISCONNECT:
1171                 return nbd_disconnect(nbd);
1172         case NBD_CLEAR_SOCK:
1173                 nbd_clear_sock_ioctl(nbd, bdev);
1174                 return 0;
1175         case NBD_SET_SOCK:
1176                 return nbd_add_socket(nbd, arg, false);
1177         case NBD_SET_BLKSIZE:
1178                 nbd_size_set(nbd, arg,
1179                              div_s64(config->bytesize, arg));
1180                 return 0;
1181         case NBD_SET_SIZE:
1182                 nbd_size_set(nbd, config->blksize,
1183                              div_s64(arg, config->blksize));
1184                 return 0;
1185         case NBD_SET_SIZE_BLOCKS:
1186                 nbd_size_set(nbd, config->blksize, arg);
1187                 return 0;
1188         case NBD_SET_TIMEOUT:
1189                 if (arg) {
1190                         nbd->tag_set.timeout = arg * HZ;
1191                         blk_queue_rq_timeout(nbd->disk->queue, arg * HZ);
1192                 }
1193                 return 0;
1194
1195         case NBD_SET_FLAGS:
1196                 config->flags = arg;
1197                 return 0;
1198         case NBD_DO_IT:
1199                 return nbd_start_device_ioctl(nbd, bdev);
1200         case NBD_CLEAR_QUE:
1201                 /*
1202                  * This is for compatibility only.  The queue is always cleared
1203                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1204                  */
1205                 return 0;
1206         case NBD_PRINT_DEBUG:
1207                 /*
1208                  * For compatibility only, we no longer keep a list of
1209                  * outstanding requests.
1210                  */
1211                 return 0;
1212         }
1213         return -ENOTTY;
1214 }
1215
1216 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1217                      unsigned int cmd, unsigned long arg)
1218 {
1219         struct nbd_device *nbd = bdev->bd_disk->private_data;
1220         struct nbd_config *config = nbd->config;
1221         int error = -EINVAL;
1222
1223         if (!capable(CAP_SYS_ADMIN))
1224                 return -EPERM;
1225
1226         /* The block layer will pass back some non-nbd ioctls in case we have
1227          * special handling for them, but we don't so just return an error.
1228          */
1229         if (_IOC_TYPE(cmd) != 0xab)
1230                 return -EINVAL;
1231
1232         mutex_lock(&nbd->config_lock);
1233
1234         /* Don't allow ioctl operations on a nbd device that was created with
1235          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1236          */
1237         if (!test_bit(NBD_BOUND, &config->runtime_flags) ||
1238             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1239                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1240         else
1241                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1242         mutex_unlock(&nbd->config_lock);
1243         return error;
1244 }
1245
1246 static struct nbd_config *nbd_alloc_config(void)
1247 {
1248         struct nbd_config *config;
1249
1250         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1251         if (!config)
1252                 return NULL;
1253         atomic_set(&config->recv_threads, 0);
1254         init_waitqueue_head(&config->recv_wq);
1255         init_waitqueue_head(&config->conn_wait);
1256         config->blksize = 1024;
1257         atomic_set(&config->live_connections, 0);
1258         try_module_get(THIS_MODULE);
1259         return config;
1260 }
1261
1262 static int nbd_open(struct block_device *bdev, fmode_t mode)
1263 {
1264         struct nbd_device *nbd;
1265         int ret = 0;
1266
1267         mutex_lock(&nbd_index_mutex);
1268         nbd = bdev->bd_disk->private_data;
1269         if (!nbd) {
1270                 ret = -ENXIO;
1271                 goto out;
1272         }
1273         if (!refcount_inc_not_zero(&nbd->refs)) {
1274                 ret = -ENXIO;
1275                 goto out;
1276         }
1277         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1278                 struct nbd_config *config;
1279
1280                 mutex_lock(&nbd->config_lock);
1281                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1282                         mutex_unlock(&nbd->config_lock);
1283                         goto out;
1284                 }
1285                 config = nbd->config = nbd_alloc_config();
1286                 if (!config) {
1287                         ret = -ENOMEM;
1288                         mutex_unlock(&nbd->config_lock);
1289                         goto out;
1290                 }
1291                 refcount_set(&nbd->config_refs, 1);
1292                 refcount_inc(&nbd->refs);
1293                 mutex_unlock(&nbd->config_lock);
1294                 bdev->bd_invalidated = 1;
1295         } else if (nbd_disconnected(nbd->config)) {
1296                 bdev->bd_invalidated = 1;
1297         }
1298 out:
1299         mutex_unlock(&nbd_index_mutex);
1300         return ret;
1301 }
1302
1303 static void nbd_release(struct gendisk *disk, fmode_t mode)
1304 {
1305         struct nbd_device *nbd = disk->private_data;
1306         nbd_config_put(nbd);
1307         nbd_put(nbd);
1308 }
1309
1310 static const struct block_device_operations nbd_fops =
1311 {
1312         .owner =        THIS_MODULE,
1313         .open =         nbd_open,
1314         .release =      nbd_release,
1315         .ioctl =        nbd_ioctl,
1316         .compat_ioctl = nbd_ioctl,
1317 };
1318
1319 #if IS_ENABLED(CONFIG_DEBUG_FS)
1320
1321 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1322 {
1323         struct nbd_device *nbd = s->private;
1324
1325         if (nbd->task_recv)
1326                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1327
1328         return 0;
1329 }
1330
1331 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1332 {
1333         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1334 }
1335
1336 static const struct file_operations nbd_dbg_tasks_ops = {
1337         .open = nbd_dbg_tasks_open,
1338         .read = seq_read,
1339         .llseek = seq_lseek,
1340         .release = single_release,
1341 };
1342
1343 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1344 {
1345         struct nbd_device *nbd = s->private;
1346         u32 flags = nbd->config->flags;
1347
1348         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1349
1350         seq_puts(s, "Known flags:\n");
1351
1352         if (flags & NBD_FLAG_HAS_FLAGS)
1353                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1354         if (flags & NBD_FLAG_READ_ONLY)
1355                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1356         if (flags & NBD_FLAG_SEND_FLUSH)
1357                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1358         if (flags & NBD_FLAG_SEND_FUA)
1359                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1360         if (flags & NBD_FLAG_SEND_TRIM)
1361                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1362
1363         return 0;
1364 }
1365
1366 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1367 {
1368         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1369 }
1370
1371 static const struct file_operations nbd_dbg_flags_ops = {
1372         .open = nbd_dbg_flags_open,
1373         .read = seq_read,
1374         .llseek = seq_lseek,
1375         .release = single_release,
1376 };
1377
1378 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1379 {
1380         struct dentry *dir;
1381         struct nbd_config *config = nbd->config;
1382
1383         if (!nbd_dbg_dir)
1384                 return -EIO;
1385
1386         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1387         if (!dir) {
1388                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1389                         nbd_name(nbd));
1390                 return -EIO;
1391         }
1392         config->dbg_dir = dir;
1393
1394         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1395         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1396         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1397         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1398         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1399
1400         return 0;
1401 }
1402
1403 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1404 {
1405         debugfs_remove_recursive(nbd->config->dbg_dir);
1406 }
1407
1408 static int nbd_dbg_init(void)
1409 {
1410         struct dentry *dbg_dir;
1411
1412         dbg_dir = debugfs_create_dir("nbd", NULL);
1413         if (!dbg_dir)
1414                 return -EIO;
1415
1416         nbd_dbg_dir = dbg_dir;
1417
1418         return 0;
1419 }
1420
1421 static void nbd_dbg_close(void)
1422 {
1423         debugfs_remove_recursive(nbd_dbg_dir);
1424 }
1425
1426 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1427
1428 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1429 {
1430         return 0;
1431 }
1432
1433 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1434 {
1435 }
1436
1437 static int nbd_dbg_init(void)
1438 {
1439         return 0;
1440 }
1441
1442 static void nbd_dbg_close(void)
1443 {
1444 }
1445
1446 #endif
1447
1448 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1449                             unsigned int hctx_idx, unsigned int numa_node)
1450 {
1451         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1452         cmd->nbd = set->driver_data;
1453         return 0;
1454 }
1455
1456 static const struct blk_mq_ops nbd_mq_ops = {
1457         .queue_rq       = nbd_queue_rq,
1458         .complete       = nbd_complete_rq,
1459         .init_request   = nbd_init_request,
1460         .timeout        = nbd_xmit_timeout,
1461 };
1462
1463 static int nbd_dev_add(int index)
1464 {
1465         struct nbd_device *nbd;
1466         struct gendisk *disk;
1467         struct request_queue *q;
1468         int err = -ENOMEM;
1469
1470         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1471         if (!nbd)
1472                 goto out;
1473
1474         disk = alloc_disk(1 << part_shift);
1475         if (!disk)
1476                 goto out_free_nbd;
1477
1478         if (index >= 0) {
1479                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1480                                 GFP_KERNEL);
1481                 if (err == -ENOSPC)
1482                         err = -EEXIST;
1483         } else {
1484                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1485                 if (err >= 0)
1486                         index = err;
1487         }
1488         if (err < 0)
1489                 goto out_free_disk;
1490
1491         nbd->index = index;
1492         nbd->disk = disk;
1493         nbd->tag_set.ops = &nbd_mq_ops;
1494         nbd->tag_set.nr_hw_queues = 1;
1495         nbd->tag_set.queue_depth = 128;
1496         nbd->tag_set.numa_node = NUMA_NO_NODE;
1497         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1498         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1499                 BLK_MQ_F_SG_MERGE | BLK_MQ_F_BLOCKING;
1500         nbd->tag_set.driver_data = nbd;
1501
1502         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1503         if (err)
1504                 goto out_free_idr;
1505
1506         q = blk_mq_init_queue(&nbd->tag_set);
1507         if (IS_ERR(q)) {
1508                 err = PTR_ERR(q);
1509                 goto out_free_tags;
1510         }
1511         disk->queue = q;
1512
1513         /*
1514          * Tell the block layer that we are not a rotational device
1515          */
1516         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1517         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1518         disk->queue->limits.discard_granularity = 0;
1519         blk_queue_max_discard_sectors(disk->queue, 0);
1520         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1521         blk_queue_max_segments(disk->queue, USHRT_MAX);
1522         blk_queue_max_hw_sectors(disk->queue, 65536);
1523         disk->queue->limits.max_sectors = 256;
1524
1525         mutex_init(&nbd->config_lock);
1526         refcount_set(&nbd->config_refs, 0);
1527         refcount_set(&nbd->refs, 1);
1528         INIT_LIST_HEAD(&nbd->list);
1529         disk->major = NBD_MAJOR;
1530         disk->first_minor = index << part_shift;
1531         disk->fops = &nbd_fops;
1532         disk->private_data = nbd;
1533         sprintf(disk->disk_name, "nbd%d", index);
1534         add_disk(disk);
1535         nbd_total_devices++;
1536         return index;
1537
1538 out_free_tags:
1539         blk_mq_free_tag_set(&nbd->tag_set);
1540 out_free_idr:
1541         idr_remove(&nbd_index_idr, index);
1542 out_free_disk:
1543         put_disk(disk);
1544 out_free_nbd:
1545         kfree(nbd);
1546 out:
1547         return err;
1548 }
1549
1550 static int find_free_cb(int id, void *ptr, void *data)
1551 {
1552         struct nbd_device *nbd = ptr;
1553         struct nbd_device **found = data;
1554
1555         if (!refcount_read(&nbd->config_refs)) {
1556                 *found = nbd;
1557                 return 1;
1558         }
1559         return 0;
1560 }
1561
1562 /* Netlink interface. */
1563 static struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1564         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1565         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1566         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1567         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1568         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1569         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1570         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1571         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1572         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1573 };
1574
1575 static struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1576         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1577 };
1578
1579 /* We don't use this right now since we don't parse the incoming list, but we
1580  * still want it here so userspace knows what to expect.
1581  */
1582 static struct nla_policy __attribute__((unused))
1583 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1584         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1585         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1586 };
1587
1588 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1589 {
1590         struct nbd_device *nbd = NULL;
1591         struct nbd_config *config;
1592         int index = -1;
1593         int ret;
1594         bool put_dev = false;
1595
1596         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1597                 return -EPERM;
1598
1599         if (info->attrs[NBD_ATTR_INDEX])
1600                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1601         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1602                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1603                 return -EINVAL;
1604         }
1605         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1606                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1607                 return -EINVAL;
1608         }
1609 again:
1610         mutex_lock(&nbd_index_mutex);
1611         if (index == -1) {
1612                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1613                 if (ret == 0) {
1614                         int new_index;
1615                         new_index = nbd_dev_add(-1);
1616                         if (new_index < 0) {
1617                                 mutex_unlock(&nbd_index_mutex);
1618                                 printk(KERN_ERR "nbd: failed to add new device\n");
1619                                 return new_index;
1620                         }
1621                         nbd = idr_find(&nbd_index_idr, new_index);
1622                 }
1623         } else {
1624                 nbd = idr_find(&nbd_index_idr, index);
1625                 if (!nbd) {
1626                         ret = nbd_dev_add(index);
1627                         if (ret < 0) {
1628                                 mutex_unlock(&nbd_index_mutex);
1629                                 printk(KERN_ERR "nbd: failed to add new device\n");
1630                                 return ret;
1631                         }
1632                         nbd = idr_find(&nbd_index_idr, index);
1633                 }
1634         }
1635         if (!nbd) {
1636                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1637                        index);
1638                 mutex_unlock(&nbd_index_mutex);
1639                 return -EINVAL;
1640         }
1641         if (!refcount_inc_not_zero(&nbd->refs)) {
1642                 mutex_unlock(&nbd_index_mutex);
1643                 if (index == -1)
1644                         goto again;
1645                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1646                        index);
1647                 return -EINVAL;
1648         }
1649         mutex_unlock(&nbd_index_mutex);
1650
1651         mutex_lock(&nbd->config_lock);
1652         if (refcount_read(&nbd->config_refs)) {
1653                 mutex_unlock(&nbd->config_lock);
1654                 nbd_put(nbd);
1655                 if (index == -1)
1656                         goto again;
1657                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1658                 return -EBUSY;
1659         }
1660         if (WARN_ON(nbd->config)) {
1661                 mutex_unlock(&nbd->config_lock);
1662                 nbd_put(nbd);
1663                 return -EINVAL;
1664         }
1665         config = nbd->config = nbd_alloc_config();
1666         if (!nbd->config) {
1667                 mutex_unlock(&nbd->config_lock);
1668                 nbd_put(nbd);
1669                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1670                 return -ENOMEM;
1671         }
1672         refcount_set(&nbd->config_refs, 1);
1673         set_bit(NBD_BOUND, &config->runtime_flags);
1674
1675         if (info->attrs[NBD_ATTR_SIZE_BYTES]) {
1676                 u64 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1677                 nbd_size_set(nbd, config->blksize,
1678                              div64_u64(bytes, config->blksize));
1679         }
1680         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1681                 u64 bsize =
1682                         nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1683                 nbd_size_set(nbd, bsize, div64_u64(config->bytesize, bsize));
1684         }
1685         if (info->attrs[NBD_ATTR_TIMEOUT]) {
1686                 u64 timeout = nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]);
1687                 nbd->tag_set.timeout = timeout * HZ;
1688                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1689         }
1690         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1691                 config->dead_conn_timeout =
1692                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1693                 config->dead_conn_timeout *= HZ;
1694         }
1695         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1696                 config->flags =
1697                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1698         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1699                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1700                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1701                         set_bit(NBD_DESTROY_ON_DISCONNECT,
1702                                 &config->runtime_flags);
1703                         put_dev = true;
1704                 }
1705         }
1706
1707         if (info->attrs[NBD_ATTR_SOCKETS]) {
1708                 struct nlattr *attr;
1709                 int rem, fd;
1710
1711                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1712                                     rem) {
1713                         struct nlattr *socks[NBD_SOCK_MAX+1];
1714
1715                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1716                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1717                                 ret = -EINVAL;
1718                                 goto out;
1719                         }
1720                         ret = nla_parse_nested(socks, NBD_SOCK_MAX, attr,
1721                                                nbd_sock_policy, info->extack);
1722                         if (ret != 0) {
1723                                 printk(KERN_ERR "nbd: error processing sock list\n");
1724                                 ret = -EINVAL;
1725                                 goto out;
1726                         }
1727                         if (!socks[NBD_SOCK_FD])
1728                                 continue;
1729                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1730                         ret = nbd_add_socket(nbd, fd, true);
1731                         if (ret)
1732                                 goto out;
1733                 }
1734         }
1735         ret = nbd_start_device(nbd);
1736 out:
1737         mutex_unlock(&nbd->config_lock);
1738         if (!ret) {
1739                 set_bit(NBD_HAS_CONFIG_REF, &config->runtime_flags);
1740                 refcount_inc(&nbd->config_refs);
1741                 nbd_connect_reply(info, nbd->index);
1742         }
1743         nbd_config_put(nbd);
1744         if (put_dev)
1745                 nbd_put(nbd);
1746         return ret;
1747 }
1748
1749 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
1750 {
1751         struct nbd_device *nbd;
1752         int index;
1753
1754         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1755                 return -EPERM;
1756
1757         if (!info->attrs[NBD_ATTR_INDEX]) {
1758                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
1759                 return -EINVAL;
1760         }
1761         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1762         mutex_lock(&nbd_index_mutex);
1763         nbd = idr_find(&nbd_index_idr, index);
1764         if (!nbd) {
1765                 mutex_unlock(&nbd_index_mutex);
1766                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1767                        index);
1768                 return -EINVAL;
1769         }
1770         if (!refcount_inc_not_zero(&nbd->refs)) {
1771                 mutex_unlock(&nbd_index_mutex);
1772                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1773                        index);
1774                 return -EINVAL;
1775         }
1776         mutex_unlock(&nbd_index_mutex);
1777         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1778                 nbd_put(nbd);
1779                 return 0;
1780         }
1781         mutex_lock(&nbd->config_lock);
1782         nbd_disconnect(nbd);
1783         nbd_clear_sock(nbd);
1784         mutex_unlock(&nbd->config_lock);
1785         if (test_and_clear_bit(NBD_HAS_CONFIG_REF,
1786                                &nbd->config->runtime_flags))
1787                 nbd_config_put(nbd);
1788         nbd_config_put(nbd);
1789         nbd_put(nbd);
1790         return 0;
1791 }
1792
1793 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
1794 {
1795         struct nbd_device *nbd = NULL;
1796         struct nbd_config *config;
1797         int index;
1798         int ret = -EINVAL;
1799         bool put_dev = false;
1800
1801         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1802                 return -EPERM;
1803
1804         if (!info->attrs[NBD_ATTR_INDEX]) {
1805                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
1806                 return -EINVAL;
1807         }
1808         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1809         mutex_lock(&nbd_index_mutex);
1810         nbd = idr_find(&nbd_index_idr, index);
1811         if (!nbd) {
1812                 mutex_unlock(&nbd_index_mutex);
1813                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
1814                        index);
1815                 return -EINVAL;
1816         }
1817         if (!refcount_inc_not_zero(&nbd->refs)) {
1818                 mutex_unlock(&nbd_index_mutex);
1819                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1820                        index);
1821                 return -EINVAL;
1822         }
1823         mutex_unlock(&nbd_index_mutex);
1824
1825         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1826                 dev_err(nbd_to_dev(nbd),
1827                         "not configured, cannot reconfigure\n");
1828                 nbd_put(nbd);
1829                 return -EINVAL;
1830         }
1831
1832         mutex_lock(&nbd->config_lock);
1833         config = nbd->config;
1834         if (!test_bit(NBD_BOUND, &config->runtime_flags) ||
1835             !nbd->task_recv) {
1836                 dev_err(nbd_to_dev(nbd),
1837                         "not configured, cannot reconfigure\n");
1838                 goto out;
1839         }
1840
1841         if (info->attrs[NBD_ATTR_TIMEOUT]) {
1842                 u64 timeout = nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]);
1843                 nbd->tag_set.timeout = timeout * HZ;
1844                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1845         }
1846         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1847                 config->dead_conn_timeout =
1848                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1849                 config->dead_conn_timeout *= HZ;
1850         }
1851         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1852                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1853                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1854                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1855                                               &config->runtime_flags))
1856                                 put_dev = true;
1857                 } else {
1858                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1859                                                &config->runtime_flags))
1860                                 refcount_inc(&nbd->refs);
1861                 }
1862         }
1863
1864         if (info->attrs[NBD_ATTR_SOCKETS]) {
1865                 struct nlattr *attr;
1866                 int rem, fd;
1867
1868                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1869                                     rem) {
1870                         struct nlattr *socks[NBD_SOCK_MAX+1];
1871
1872                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1873                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1874                                 ret = -EINVAL;
1875                                 goto out;
1876                         }
1877                         ret = nla_parse_nested(socks, NBD_SOCK_MAX, attr,
1878                                                nbd_sock_policy, info->extack);
1879                         if (ret != 0) {
1880                                 printk(KERN_ERR "nbd: error processing sock list\n");
1881                                 ret = -EINVAL;
1882                                 goto out;
1883                         }
1884                         if (!socks[NBD_SOCK_FD])
1885                                 continue;
1886                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1887                         ret = nbd_reconnect_socket(nbd, fd);
1888                         if (ret) {
1889                                 if (ret == -ENOSPC)
1890                                         ret = 0;
1891                                 goto out;
1892                         }
1893                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
1894                 }
1895         }
1896 out:
1897         mutex_unlock(&nbd->config_lock);
1898         nbd_config_put(nbd);
1899         nbd_put(nbd);
1900         if (put_dev)
1901                 nbd_put(nbd);
1902         return ret;
1903 }
1904
1905 static const struct genl_ops nbd_connect_genl_ops[] = {
1906         {
1907                 .cmd    = NBD_CMD_CONNECT,
1908                 .policy = nbd_attr_policy,
1909                 .doit   = nbd_genl_connect,
1910         },
1911         {
1912                 .cmd    = NBD_CMD_DISCONNECT,
1913                 .policy = nbd_attr_policy,
1914                 .doit   = nbd_genl_disconnect,
1915         },
1916         {
1917                 .cmd    = NBD_CMD_RECONFIGURE,
1918                 .policy = nbd_attr_policy,
1919                 .doit   = nbd_genl_reconfigure,
1920         },
1921         {
1922                 .cmd    = NBD_CMD_STATUS,
1923                 .policy = nbd_attr_policy,
1924                 .doit   = nbd_genl_status,
1925         },
1926 };
1927
1928 static const struct genl_multicast_group nbd_mcast_grps[] = {
1929         { .name = NBD_GENL_MCAST_GROUP_NAME, },
1930 };
1931
1932 static struct genl_family nbd_genl_family __ro_after_init = {
1933         .hdrsize        = 0,
1934         .name           = NBD_GENL_FAMILY_NAME,
1935         .version        = NBD_GENL_VERSION,
1936         .module         = THIS_MODULE,
1937         .ops            = nbd_connect_genl_ops,
1938         .n_ops          = ARRAY_SIZE(nbd_connect_genl_ops),
1939         .maxattr        = NBD_ATTR_MAX,
1940         .mcgrps         = nbd_mcast_grps,
1941         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
1942 };
1943
1944 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
1945 {
1946         struct nlattr *dev_opt;
1947         u8 connected = 0;
1948         int ret;
1949
1950         /* This is a little racey, but for status it's ok.  The
1951          * reason we don't take a ref here is because we can't
1952          * take a ref in the index == -1 case as we would need
1953          * to put under the nbd_index_mutex, which could
1954          * deadlock if we are configured to remove ourselves
1955          * once we're disconnected.
1956          */
1957         if (refcount_read(&nbd->config_refs))
1958                 connected = 1;
1959         dev_opt = nla_nest_start(reply, NBD_DEVICE_ITEM);
1960         if (!dev_opt)
1961                 return -EMSGSIZE;
1962         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
1963         if (ret)
1964                 return -EMSGSIZE;
1965         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
1966                          connected);
1967         if (ret)
1968                 return -EMSGSIZE;
1969         nla_nest_end(reply, dev_opt);
1970         return 0;
1971 }
1972
1973 static int status_cb(int id, void *ptr, void *data)
1974 {
1975         struct nbd_device *nbd = ptr;
1976         return populate_nbd_status(nbd, (struct sk_buff *)data);
1977 }
1978
1979 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
1980 {
1981         struct nlattr *dev_list;
1982         struct sk_buff *reply;
1983         void *reply_head;
1984         size_t msg_size;
1985         int index = -1;
1986         int ret = -ENOMEM;
1987
1988         if (info->attrs[NBD_ATTR_INDEX])
1989                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1990
1991         mutex_lock(&nbd_index_mutex);
1992
1993         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
1994                                   nla_attr_size(sizeof(u8)));
1995         msg_size *= (index == -1) ? nbd_total_devices : 1;
1996
1997         reply = genlmsg_new(msg_size, GFP_KERNEL);
1998         if (!reply)
1999                 goto out;
2000         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2001                                        NBD_CMD_STATUS);
2002         if (!reply_head) {
2003                 nlmsg_free(reply);
2004                 goto out;
2005         }
2006
2007         dev_list = nla_nest_start(reply, NBD_ATTR_DEVICE_LIST);
2008         if (index == -1) {
2009                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2010                 if (ret) {
2011                         nlmsg_free(reply);
2012                         goto out;
2013                 }
2014         } else {
2015                 struct nbd_device *nbd;
2016                 nbd = idr_find(&nbd_index_idr, index);
2017                 if (nbd) {
2018                         ret = populate_nbd_status(nbd, reply);
2019                         if (ret) {
2020                                 nlmsg_free(reply);
2021                                 goto out;
2022                         }
2023                 }
2024         }
2025         nla_nest_end(reply, dev_list);
2026         genlmsg_end(reply, reply_head);
2027         genlmsg_reply(reply, info);
2028         ret = 0;
2029 out:
2030         mutex_unlock(&nbd_index_mutex);
2031         return ret;
2032 }
2033
2034 static void nbd_connect_reply(struct genl_info *info, int index)
2035 {
2036         struct sk_buff *skb;
2037         void *msg_head;
2038         int ret;
2039
2040         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2041         if (!skb)
2042                 return;
2043         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2044                                      NBD_CMD_CONNECT);
2045         if (!msg_head) {
2046                 nlmsg_free(skb);
2047                 return;
2048         }
2049         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2050         if (ret) {
2051                 nlmsg_free(skb);
2052                 return;
2053         }
2054         genlmsg_end(skb, msg_head);
2055         genlmsg_reply(skb, info);
2056 }
2057
2058 static void nbd_mcast_index(int index)
2059 {
2060         struct sk_buff *skb;
2061         void *msg_head;
2062         int ret;
2063
2064         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2065         if (!skb)
2066                 return;
2067         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2068                                      NBD_CMD_LINK_DEAD);
2069         if (!msg_head) {
2070                 nlmsg_free(skb);
2071                 return;
2072         }
2073         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2074         if (ret) {
2075                 nlmsg_free(skb);
2076                 return;
2077         }
2078         genlmsg_end(skb, msg_head);
2079         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2080 }
2081
2082 static void nbd_dead_link_work(struct work_struct *work)
2083 {
2084         struct link_dead_args *args = container_of(work, struct link_dead_args,
2085                                                    work);
2086         nbd_mcast_index(args->index);
2087         kfree(args);
2088 }
2089
2090 static int __init nbd_init(void)
2091 {
2092         int i;
2093
2094         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2095
2096         if (max_part < 0) {
2097                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2098                 return -EINVAL;
2099         }
2100
2101         part_shift = 0;
2102         if (max_part > 0) {
2103                 part_shift = fls(max_part);
2104
2105                 /*
2106                  * Adjust max_part according to part_shift as it is exported
2107                  * to user space so that user can know the max number of
2108                  * partition kernel should be able to manage.
2109                  *
2110                  * Note that -1 is required because partition 0 is reserved
2111                  * for the whole disk.
2112                  */
2113                 max_part = (1UL << part_shift) - 1;
2114         }
2115
2116         if ((1UL << part_shift) > DISK_MAX_PARTS)
2117                 return -EINVAL;
2118
2119         if (nbds_max > 1UL << (MINORBITS - part_shift))
2120                 return -EINVAL;
2121         recv_workqueue = alloc_workqueue("knbd-recv",
2122                                          WQ_MEM_RECLAIM | WQ_HIGHPRI |
2123                                          WQ_UNBOUND, 0);
2124         if (!recv_workqueue)
2125                 return -ENOMEM;
2126
2127         if (register_blkdev(NBD_MAJOR, "nbd")) {
2128                 destroy_workqueue(recv_workqueue);
2129                 return -EIO;
2130         }
2131
2132         if (genl_register_family(&nbd_genl_family)) {
2133                 unregister_blkdev(NBD_MAJOR, "nbd");
2134                 destroy_workqueue(recv_workqueue);
2135                 return -EINVAL;
2136         }
2137         nbd_dbg_init();
2138
2139         mutex_lock(&nbd_index_mutex);
2140         for (i = 0; i < nbds_max; i++)
2141                 nbd_dev_add(i);
2142         mutex_unlock(&nbd_index_mutex);
2143         return 0;
2144 }
2145
2146 static int nbd_exit_cb(int id, void *ptr, void *data)
2147 {
2148         struct list_head *list = (struct list_head *)data;
2149         struct nbd_device *nbd = ptr;
2150
2151         list_add_tail(&nbd->list, list);
2152         return 0;
2153 }
2154
2155 static void __exit nbd_cleanup(void)
2156 {
2157         struct nbd_device *nbd;
2158         LIST_HEAD(del_list);
2159
2160         nbd_dbg_close();
2161
2162         mutex_lock(&nbd_index_mutex);
2163         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2164         mutex_unlock(&nbd_index_mutex);
2165
2166         while (!list_empty(&del_list)) {
2167                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2168                 list_del_init(&nbd->list);
2169                 if (refcount_read(&nbd->refs) != 1)
2170                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2171                 nbd_put(nbd);
2172         }
2173
2174         idr_destroy(&nbd_index_idr);
2175         genl_unregister_family(&nbd_genl_family);
2176         destroy_workqueue(recv_workqueue);
2177         unregister_blkdev(NBD_MAJOR, "nbd");
2178 }
2179
2180 module_init(nbd_init);
2181 module_exit(nbd_cleanup);
2182
2183 MODULE_DESCRIPTION("Network Block Device");
2184 MODULE_LICENSE("GPL");
2185
2186 module_param(nbds_max, int, 0444);
2187 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2188 module_param(max_part, int, 0444);
2189 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");