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                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 250                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 251         }
 252         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 253         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 254         set_capacity(nbd->disk, config->bytesize >> 9);
 255         if (bdev) {
 256                 if (bdev->bd_disk)
 257                         bd_set_size(bdev, config->bytesize);
 258                 else
 259                         bdev->bd_invalidated = 1;
 260                 bdput(bdev);
 261         }
 262         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 263 }
 264
 265 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 266                          loff_t nr_blocks)
 267 {
 268         struct nbd_config *config = nbd->config;
 269         config->blksize = blocksize;
 270         config->bytesize = blocksize * nr_blocks;
 271         if (nbd->task_recv != NULL)
 272                 nbd_size_update(nbd);
 273 }
 274
 275 static void nbd_complete_rq(struct request *req)
 276 {
 277         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 278
 279         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 280                 cmd->status ? "failed" : "done");
 281
 282         blk_mq_end_request(req, cmd->status);
 283 }
 284
 285 /*
 286  * Forcibly shutdown the socket causing all listeners to error
 287  */
 288 static void sock_shutdown(struct nbd_device *nbd)
 289 {
 290         struct nbd_config *config = nbd->config;
 291         int i;
 292
 293         if (config->num_connections == 0)
 294                 return;
 295         if (test_and_set_bit(NBD_DISCONNECTED, &config->runtime_flags))
 296                 return;
 297
 298         for (i = 0; i < config->num_connections; i++) {
 299                 struct nbd_sock *nsock = config->socks[i];
 300                 mutex_lock(&nsock->tx_lock);
 301                 nbd_mark_nsock_dead(nbd, nsock, 0);
 302                 mutex_unlock(&nsock->tx_lock);
 303         }
 304         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 305 }
 306
 307 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 308                                                  bool reserved)
 309 {
 310         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 311         struct nbd_device *nbd = cmd->nbd;
 312         struct nbd_config *config;
 313
 314         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 315                 cmd->status = BLK_STS_TIMEOUT;
 316                 goto done;
 317         }
 318         config = nbd->config;
 319
 320         if (config->num_connections > 1) {
 321                 dev_err_ratelimited(nbd_to_dev(nbd),
 322                                     "Connection timed out, retrying (%d/%d alive)\n",
 323                                     atomic_read(&config->live_connections),
 324                                     config->num_connections);
 325                 /*
 326                  * Hooray we have more connections, requeue this IO, the submit
 327                  * path will put it on a real connection.
 328                  */
 329                 if (config->socks && config->num_connections > 1) {
 330                         if (cmd->index < config->num_connections) {
 331                                 struct nbd_sock *nsock =
 332                                         config->socks[cmd->index];
 333                                 mutex_lock(&nsock->tx_lock);
 334                                 /* We can have multiple outstanding requests, so
 335                                  * we don't want to mark the nsock dead if we've
 336                                  * already reconnected with a new socket, so
 337                                  * only mark it dead if its the same socket we
 338                                  * were sent out on.
 339                                  */
 340                                 if (cmd->cookie == nsock->cookie)
 341                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 342                                 mutex_unlock(&nsock->tx_lock);
 343                         }
 344                         blk_mq_requeue_request(req, true);
 345                         nbd_config_put(nbd);
 346                         return BLK_EH_DONE;
 347                 }
 348         } else {
 349                 dev_err_ratelimited(nbd_to_dev(nbd),
 350                                     "Connection timed out\n");
 351         }
 352         set_bit(NBD_TIMEDOUT, &config->runtime_flags);
 353         cmd->status = BLK_STS_IOERR;
 354         sock_shutdown(nbd);
 355         nbd_config_put(nbd);
 356 done:
 357         blk_mq_complete_request(req);
 358         return BLK_EH_DONE;
 359 }
 360
 361 /*
 362  *  Send or receive packet.
 363  */
 364 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 365                      struct iov_iter *iter, int msg_flags, int *sent)
 366 {
 367         struct nbd_config *config = nbd->config;
 368         struct socket *sock = config->socks[index]->sock;
 369         int result;
 370         struct msghdr msg;
 371         unsigned int noreclaim_flag;
 372
 373         if (unlikely(!sock)) {
 374                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 375                         "Attempted %s on closed socket in sock_xmit\n",
 376                         (send ? "send" : "recv"));
 377                 return -EINVAL;
 378         }
 379
 380         msg.msg_iter = *iter;
 381
 382         noreclaim_flag = memalloc_noreclaim_save();
 383         do {
 384                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 385                 msg.msg_name = NULL;
 386                 msg.msg_namelen = 0;
 387                 msg.msg_control = NULL;
 388                 msg.msg_controllen = 0;
 389                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 390
 391                 if (send)
 392                         result = sock_sendmsg(sock, &msg);
 393                 else
 394                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 395
 396                 if (result <= 0) {
 397                         if (result == 0)
 398                                 result = -EPIPE; /* short read */
 399                         break;
 400                 }
 401                 if (sent)
 402                         *sent += result;
 403         } while (msg_data_left(&msg));
 404
 405         memalloc_noreclaim_restore(noreclaim_flag);
 406
 407         return result;
 408 }
 409
 410 /*
 411  * Different settings for sk->sk_sndtimeo can result in different return values
 412  * if there is a signal pending when we enter sendmsg, because reasons?
 413  */
 414 static inline int was_interrupted(int result)
 415 {
 416         return result == -ERESTARTSYS || result == -EINTR;
 417 }
 418
 419 /* always call with the tx_lock held */
 420 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 421 {
 422         struct request *req = blk_mq_rq_from_pdu(cmd);
 423         struct nbd_config *config = nbd->config;
 424         struct nbd_sock *nsock = config->socks[index];
 425         int result;
 426         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 427         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 428         struct iov_iter from;
 429         unsigned long size = blk_rq_bytes(req);
 430         struct bio *bio;
 431         u32 type;
 432         u32 nbd_cmd_flags = 0;
 433         u32 tag = blk_mq_unique_tag(req);
 434         int sent = nsock->sent, skip = 0;
 435
 436         iov_iter_kvec(&from, WRITE | ITER_KVEC, &iov, 1, sizeof(request));
 437
 438         switch (req_op(req)) {
 439         case REQ_OP_DISCARD:
 440                 type = NBD_CMD_TRIM;
 441                 break;
 442         case REQ_OP_FLUSH:
 443                 type = NBD_CMD_FLUSH;
 444                 break;
 445         case REQ_OP_WRITE:
 446                 type = NBD_CMD_WRITE;
 447                 break;
 448         case REQ_OP_READ:
 449                 type = NBD_CMD_READ;
 450                 break;
 451         default:
 452                 return -EIO;
 453         }
 454
 455         if (rq_data_dir(req) == WRITE &&
 456             (config->flags & NBD_FLAG_READ_ONLY)) {
 457                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 458                                     "Write on read-only\n");
 459                 return -EIO;
 460         }
 461
 462         if (req->cmd_flags & REQ_FUA)
 463                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 464
 465         /* We did a partial send previously, and we at least sent the whole
 466          * request struct, so just go and send the rest of the pages in the
 467          * request.
 468          */
 469         if (sent) {
 470                 if (sent >= sizeof(request)) {
 471                         skip = sent - sizeof(request);
 472                         goto send_pages;
 473                 }
 474                 iov_iter_advance(&from, sent);
 475         }
 476         cmd->index = index;
 477         cmd->cookie = nsock->cookie;
 478         request.type = htonl(type | nbd_cmd_flags);
 479         if (type != NBD_CMD_FLUSH) {
 480                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 481                 request.len = htonl(size);
 482         }
 483         memcpy(request.handle, &tag, sizeof(tag));
 484
 485         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 486                 req, nbdcmd_to_ascii(type),
 487                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 488         result = sock_xmit(nbd, index, 1, &from,
 489                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 490         if (result <= 0) {
 491                 if (was_interrupted(result)) {
 492                         /* If we havne't sent anything we can just return BUSY,
 493                          * however if we have sent something we need to make
 494                          * sure we only allow this req to be sent until we are
 495                          * completely done.
 496                          */
 497                         if (sent) {
 498                                 nsock->pending = req;
 499                                 nsock->sent = sent;
 500                         }
 501                         return BLK_STS_RESOURCE;
 502                 }
 503                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 504                         "Send control failed (result %d)\n", result);
 505                 return -EAGAIN;
 506         }
 507 send_pages:
 508         if (type != NBD_CMD_WRITE)
 509                 goto out;
 510
 511         bio = req->bio;
 512         while (bio) {
 513                 struct bio *next = bio->bi_next;
 514                 struct bvec_iter iter;
 515                 struct bio_vec bvec;
 516
 517                 bio_for_each_segment(bvec, bio, iter) {
 518                         bool is_last = !next && bio_iter_last(bvec, iter);
 519                         int flags = is_last ? 0 : MSG_MORE;
 520
 521                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 522                                 req, bvec.bv_len);
 523                         iov_iter_bvec(&from, ITER_BVEC | WRITE,
 524                                       &bvec, 1, bvec.bv_len);
 525                         if (skip) {
 526                                 if (skip >= iov_iter_count(&from)) {
 527                                         skip -= iov_iter_count(&from);
 528                                         continue;
 529                                 }
 530                                 iov_iter_advance(&from, skip);
 531                                 skip = 0;
 532                         }
 533                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 534                         if (result <= 0) {
 535                                 if (was_interrupted(result)) {
 536                                         /* We've already sent the header, we
 537                                          * have no choice but to set pending and
 538                                          * return BUSY.
 539                                          */
 540                                         nsock->pending = req;
 541                                         nsock->sent = sent;
 542                                         return BLK_STS_RESOURCE;
 543                                 }
 544                                 dev_err(disk_to_dev(nbd->disk),
 545                                         "Send data failed (result %d)\n",
 546                                         result);
 547                                 return -EAGAIN;
 548                         }
 549                         /*
 550                          * The completion might already have come in,
 551                          * so break for the last one instead of letting
 552                          * the iterator do it. This prevents use-after-free
 553                          * of the bio.
 554                          */
 555                         if (is_last)
 556                                 break;
 557                 }
 558                 bio = next;
 559         }
 560 out:
 561         nsock->pending = NULL;
 562         nsock->sent = 0;
 563         return 0;
 564 }
 565
 566 /* NULL returned = something went wrong, inform userspace */
 567 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 568 {
 569         struct nbd_config *config = nbd->config;
 570         int result;
 571         struct nbd_reply reply;
 572         struct nbd_cmd *cmd;
 573         struct request *req = NULL;
 574         u16 hwq;
 575         u32 tag;
 576         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 577         struct iov_iter to;
 578
 579         reply.magic = 0;
 580         iov_iter_kvec(&to, READ | ITER_KVEC, &iov, 1, sizeof(reply));
 581         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 582         if (result <= 0) {
 583                 if (!nbd_disconnected(config))
 584                         dev_err(disk_to_dev(nbd->disk),
 585                                 "Receive control failed (result %d)\n", result);
 586                 return ERR_PTR(result);
 587         }
 588
 589         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 590                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 591                                 (unsigned long)ntohl(reply.magic));
 592                 return ERR_PTR(-EPROTO);
 593         }
 594
 595         memcpy(&tag, reply.handle, sizeof(u32));
 596
 597         hwq = blk_mq_unique_tag_to_hwq(tag);
 598         if (hwq < nbd->tag_set.nr_hw_queues)
 599                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 600                                        blk_mq_unique_tag_to_tag(tag));
 601         if (!req || !blk_mq_request_started(req)) {
 602                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 603                         tag, req);
 604                 return ERR_PTR(-ENOENT);
 605         }
 606         cmd = blk_mq_rq_to_pdu(req);
 607         if (ntohl(reply.error)) {
 608                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 609                         ntohl(reply.error));
 610                 cmd->status = BLK_STS_IOERR;
 611                 return cmd;
 612         }
 613
 614         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 615         if (rq_data_dir(req) != WRITE) {
 616                 struct req_iterator iter;
 617                 struct bio_vec bvec;
 618
 619                 rq_for_each_segment(bvec, req, iter) {
 620                         iov_iter_bvec(&to, ITER_BVEC | READ,
 621                                       &bvec, 1, bvec.bv_len);
 622                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 623                         if (result <= 0) {
 624                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 625                                         result);
 626                                 /*
 627                                  * If we've disconnected or we only have 1
 628                                  * connection then we need to make sure we
 629                                  * complete this request, otherwise error out
 630                                  * and let the timeout stuff handle resubmitting
 631                                  * this request onto another connection.
 632                                  */
 633                                 if (nbd_disconnected(config) ||
 634                                     config->num_connections <= 1) {
 635                                         cmd->status = BLK_STS_IOERR;
 636                                         return cmd;
 637                                 }
 638                                 return ERR_PTR(-EIO);
 639                         }
 640                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 641                                 req, bvec.bv_len);
 642                 }
 643         } else {
 644                 /* See the comment in nbd_queue_rq. */
 645                 wait_for_completion(&cmd->send_complete);
 646         }
 647         return cmd;
 648 }
 649
 650 static void recv_work(struct work_struct *work)
 651 {
 652         struct recv_thread_args *args = container_of(work,
 653                                                      struct recv_thread_args,
 654                                                      work);
 655         struct nbd_device *nbd = args->nbd;
 656         struct nbd_config *config = nbd->config;
 657         struct nbd_cmd *cmd;
 658
 659         while (1) {
 660                 cmd = nbd_read_stat(nbd, args->index);
 661                 if (IS_ERR(cmd)) {
 662                         struct nbd_sock *nsock = config->socks[args->index];
 663
 664                         mutex_lock(&nsock->tx_lock);
 665                         nbd_mark_nsock_dead(nbd, nsock, 1);
 666                         mutex_unlock(&nsock->tx_lock);
 667                         break;
 668                 }
 669
 670                 blk_mq_complete_request(blk_mq_rq_from_pdu(cmd));
 671         }
 672         atomic_dec(&config->recv_threads);
 673         wake_up(&config->recv_wq);
 674         nbd_config_put(nbd);
 675         kfree(args);
 676 }
 677
 678 static void nbd_clear_req(struct request *req, void *data, bool reserved)
 679 {
 680         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 681
 682         cmd->status = BLK_STS_IOERR;
 683         blk_mq_complete_request(req);
 684 }
 685
 686 static void nbd_clear_que(struct nbd_device *nbd)
 687 {
 688         blk_mq_quiesce_queue(nbd->disk->queue);
 689         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 690         blk_mq_unquiesce_queue(nbd->disk->queue);
 691         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 692 }
 693
 694 static int find_fallback(struct nbd_device *nbd, int index)
 695 {
 696         struct nbd_config *config = nbd->config;
 697         int new_index = -1;
 698         struct nbd_sock *nsock = config->socks[index];
 699         int fallback = nsock->fallback_index;
 700
 701         if (test_bit(NBD_DISCONNECTED, &config->runtime_flags))
 702                 return new_index;
 703
 704         if (config->num_connections <= 1) {
 705                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 706                                     "Attempted send on invalid socket\n");
 707                 return new_index;
 708         }
 709
 710         if (fallback >= 0 && fallback < config->num_connections &&
 711             !config->socks[fallback]->dead)
 712                 return fallback;
 713
 714         if (nsock->fallback_index < 0 ||
 715             nsock->fallback_index >= config->num_connections ||
 716             config->socks[nsock->fallback_index]->dead) {
 717                 int i;
 718                 for (i = 0; i < config->num_connections; i++) {
 719                         if (i == index)
 720                                 continue;
 721                         if (!config->socks[i]->dead) {
 722                                 new_index = i;
 723                                 break;
 724                         }
 725                 }
 726                 nsock->fallback_index = new_index;
 727                 if (new_index < 0) {
 728                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 729                                             "Dead connection, failed to find a fallback\n");
 730                         return new_index;
 731                 }
 732         }
 733         new_index = nsock->fallback_index;
 734         return new_index;
 735 }
 736
 737 static int wait_for_reconnect(struct nbd_device *nbd)
 738 {
 739         struct nbd_config *config = nbd->config;
 740         if (!config->dead_conn_timeout)
 741                 return 0;
 742         if (test_bit(NBD_DISCONNECTED, &config->runtime_flags))
 743                 return 0;
 744         return wait_event_timeout(config->conn_wait,
 745                                   atomic_read(&config->live_connections) > 0,
 746                                   config->dead_conn_timeout) > 0;
 747 }
 748
 749 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 750 {
 751         struct request *req = blk_mq_rq_from_pdu(cmd);
 752         struct nbd_device *nbd = cmd->nbd;
 753         struct nbd_config *config;
 754         struct nbd_sock *nsock;
 755         int ret;
 756
 757         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 758                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 759                                     "Socks array is empty\n");
 760                 blk_mq_start_request(req);
 761                 return -EINVAL;
 762         }
 763         config = nbd->config;
 764
 765         if (index >= config->num_connections) {
 766                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 767                                     "Attempted send on invalid socket\n");
 768                 nbd_config_put(nbd);
 769                 blk_mq_start_request(req);
 770                 return -EINVAL;
 771         }
 772         cmd->status = BLK_STS_OK;
 773 again:
 774         nsock = config->socks[index];
 775         mutex_lock(&nsock->tx_lock);
 776         if (nsock->dead) {
 777                 int old_index = index;
 778                 index = find_fallback(nbd, index);
 779                 mutex_unlock(&nsock->tx_lock);
 780                 if (index < 0) {
 781                         if (wait_for_reconnect(nbd)) {
 782                                 index = old_index;
 783                                 goto again;
 784                         }
 785                         /* All the sockets should already be down at this point,
 786                          * we just want to make sure that DISCONNECTED is set so
 787                          * any requests that come in that were queue'ed waiting
 788                          * for the reconnect timer don't trigger the timer again
 789                          * and instead just error out.
 790                          */
 791                         sock_shutdown(nbd);
 792                         nbd_config_put(nbd);
 793                         blk_mq_start_request(req);
 794                         return -EIO;
 795                 }
 796                 goto again;
 797         }
 798
 799         /* Handle the case that we have a pending request that was partially
 800          * transmitted that _has_ to be serviced first.  We need to call requeue
 801          * here so that it gets put _after_ the request that is already on the
 802          * dispatch list.
 803          */
 804         blk_mq_start_request(req);
 805         if (unlikely(nsock->pending && nsock->pending != req)) {
 806                 blk_mq_requeue_request(req, true);
 807                 ret = 0;
 808                 goto out;
 809         }
 810         /*
 811          * Some failures are related to the link going down, so anything that
 812          * returns EAGAIN can be retried on a different socket.
 813          */
 814         ret = nbd_send_cmd(nbd, cmd, index);
 815         if (ret == -EAGAIN) {
 816                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 817                                     "Request send failed, requeueing\n");
 818                 nbd_mark_nsock_dead(nbd, nsock, 1);
 819                 blk_mq_requeue_request(req, true);
 820                 ret = 0;
 821         }
 822 out:
 823         mutex_unlock(&nsock->tx_lock);
 824         nbd_config_put(nbd);
 825         return ret;
 826 }
 827
 828 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 829                         const struct blk_mq_queue_data *bd)
 830 {
 831         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 832         int ret;
 833
 834         /*
 835          * Since we look at the bio's to send the request over the network we
 836          * need to make sure the completion work doesn't mark this request done
 837          * before we are done doing our send.  This keeps us from dereferencing
 838          * freed data if we have particularly fast completions (ie we get the
 839          * completion before we exit sock_xmit on the last bvec) or in the case
 840          * that the server is misbehaving (or there was an error) before we're
 841          * done sending everything over the wire.
 842          */
 843         init_completion(&cmd->send_complete);
 844
 845         /* We can be called directly from the user space process, which means we
 846          * could possibly have signals pending so our sendmsg will fail.  In
 847          * this case we need to return that we are busy, otherwise error out as
 848          * appropriate.
 849          */
 850         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 851         if (ret < 0)
 852                 ret = BLK_STS_IOERR;
 853         else if (!ret)
 854                 ret = BLK_STS_OK;
 855         complete(&cmd->send_complete);
 856
 857         return ret;
 858 }
 859
 860 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
 861                           bool netlink)
 862 {
 863         struct nbd_config *config = nbd->config;
 864         struct socket *sock;
 865         struct nbd_sock **socks;
 866         struct nbd_sock *nsock;
 867         int err;
 868
 869         sock = sockfd_lookup(arg, &err);
 870         if (!sock)
 871                 return err;
 872
 873         if (!netlink && !nbd->task_setup &&
 874             !test_bit(NBD_BOUND, &config->runtime_flags))
 875                 nbd->task_setup = current;
 876
 877         if (!netlink &&
 878             (nbd->task_setup != current ||
 879              test_bit(NBD_BOUND, &config->runtime_flags))) {
 880                 dev_err(disk_to_dev(nbd->disk),
 881                         "Device being setup by another task");
 882                 sockfd_put(sock);
 883                 return -EBUSY;
 884         }
 885
 886         socks = krealloc(config->socks, (config->num_connections + 1) *
 887                          sizeof(struct nbd_sock *), GFP_KERNEL);
 888         if (!socks) {
 889                 sockfd_put(sock);
 890                 return -ENOMEM;
 891         }
 892         nsock = kzalloc(sizeof(struct nbd_sock), GFP_KERNEL);
 893         if (!nsock) {
 894                 sockfd_put(sock);
 895                 return -ENOMEM;
 896         }
 897
 898         config->socks = socks;
 899
 900         nsock->fallback_index = -1;
 901         nsock->dead = false;
 902         mutex_init(&nsock->tx_lock);
 903         nsock->sock = sock;
 904         nsock->pending = NULL;
 905         nsock->sent = 0;
 906         nsock->cookie = 0;
 907         socks[config->num_connections++] = nsock;
 908         atomic_inc(&config->live_connections);
 909
 910         return 0;
 911 }
 912
 913 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
 914 {
 915         struct nbd_config *config = nbd->config;
 916         struct socket *sock, *old;
 917         struct recv_thread_args *args;
 918         int i;
 919         int err;
 920
 921         sock = sockfd_lookup(arg, &err);
 922         if (!sock)
 923                 return err;
 924
 925         args = kzalloc(sizeof(*args), GFP_KERNEL);
 926         if (!args) {
 927                 sockfd_put(sock);
 928                 return -ENOMEM;
 929         }
 930
 931         for (i = 0; i < config->num_connections; i++) {
 932                 struct nbd_sock *nsock = config->socks[i];
 933
 934                 if (!nsock->dead)
 935                         continue;
 936
 937                 mutex_lock(&nsock->tx_lock);
 938                 if (!nsock->dead) {
 939                         mutex_unlock(&nsock->tx_lock);
 940                         continue;
 941                 }
 942                 sk_set_memalloc(sock->sk);
 943                 if (nbd->tag_set.timeout)
 944                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
 945                 atomic_inc(&config->recv_threads);
 946                 refcount_inc(&nbd->config_refs);
 947                 old = nsock->sock;
 948                 nsock->fallback_index = -1;
 949                 nsock->sock = sock;
 950                 nsock->dead = false;
 951                 INIT_WORK(&args->work, recv_work);
 952                 args->index = i;
 953                 args->nbd = nbd;
 954                 nsock->cookie++;
 955                 mutex_unlock(&nsock->tx_lock);
 956                 sockfd_put(old);
 957
 958                 clear_bit(NBD_DISCONNECTED, &config->runtime_flags);
 959
 960                 /* We take the tx_mutex in an error path in the recv_work, so we
 961                  * need to queue_work outside of the tx_mutex.
 962                  */
 963                 queue_work(recv_workqueue, &args->work);
 964
 965                 atomic_inc(&config->live_connections);
 966                 wake_up(&config->conn_wait);
 967                 return 0;
 968         }
 969         sockfd_put(sock);
 970         kfree(args);
 971         return -ENOSPC;
 972 }
 973
 974 static void nbd_bdev_reset(struct block_device *bdev)
 975 {
 976         if (bdev->bd_openers > 1)
 977                 return;
 978         bd_set_size(bdev, 0);
 979 }
 980
 981 static void nbd_parse_flags(struct nbd_device *nbd)
 982 {
 983         struct nbd_config *config = nbd->config;
 984         if (config->flags & NBD_FLAG_READ_ONLY)
 985                 set_disk_ro(nbd->disk, true);
 986         else
 987                 set_disk_ro(nbd->disk, false);
 988         if (config->flags & NBD_FLAG_SEND_TRIM)
 989                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
 990         if (config->flags & NBD_FLAG_SEND_FLUSH) {
 991                 if (config->flags & NBD_FLAG_SEND_FUA)
 992                         blk_queue_write_cache(nbd->disk->queue, true, true);
 993                 else
 994                         blk_queue_write_cache(nbd->disk->queue, true, false);
 995         }
 996         else
 997                 blk_queue_write_cache(nbd->disk->queue, false, false);
 998 }
 999
1000 static void send_disconnects(struct nbd_device *nbd)
1001 {
1002         struct nbd_config *config = nbd->config;
1003         struct nbd_request request = {
1004                 .magic = htonl(NBD_REQUEST_MAGIC),
1005                 .type = htonl(NBD_CMD_DISC),
1006         };
1007         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1008         struct iov_iter from;
1009         int i, ret;
1010
1011         for (i = 0; i < config->num_connections; i++) {
1012                 struct nbd_sock *nsock = config->socks[i];
1013
1014                 iov_iter_kvec(&from, WRITE | ITER_KVEC, &iov, 1, sizeof(request));
1015                 mutex_lock(&nsock->tx_lock);
1016                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1017                 if (ret <= 0)
1018                         dev_err(disk_to_dev(nbd->disk),
1019                                 "Send disconnect failed %d\n", ret);
1020                 mutex_unlock(&nsock->tx_lock);
1021         }
1022 }
1023
1024 static int nbd_disconnect(struct nbd_device *nbd)
1025 {
1026         struct nbd_config *config = nbd->config;
1027
1028         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1029         set_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags);
1030         send_disconnects(nbd);
1031         return 0;
1032 }
1033
1034 static void nbd_clear_sock(struct nbd_device *nbd)
1035 {
1036         sock_shutdown(nbd);
1037         nbd_clear_que(nbd);
1038         nbd->task_setup = NULL;
1039 }
1040
1041 static void nbd_config_put(struct nbd_device *nbd)
1042 {
1043         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1044                                         &nbd->config_lock)) {
1045                 struct nbd_config *config = nbd->config;
1046                 nbd_dev_dbg_close(nbd);
1047                 nbd_size_clear(nbd);
1048                 if (test_and_clear_bit(NBD_HAS_PID_FILE,
1049                                        &config->runtime_flags))
1050                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1051                 nbd->task_recv = NULL;
1052                 nbd_clear_sock(nbd);
1053                 if (config->num_connections) {
1054                         int i;
1055                         for (i = 0; i < config->num_connections; i++) {
1056                                 sockfd_put(config->socks[i]->sock);
1057                                 kfree(config->socks[i]);
1058                         }
1059                         kfree(config->socks);
1060                 }
1061                 kfree(nbd->config);
1062                 nbd->config = NULL;
1063
1064                 nbd->tag_set.timeout = 0;
1065                 nbd->disk->queue->limits.discard_granularity = 0;
1066                 nbd->disk->queue->limits.discard_alignment = 0;
1067                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1068                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1069
1070                 mutex_unlock(&nbd->config_lock);
1071                 nbd_put(nbd);
1072                 module_put(THIS_MODULE);
1073         }
1074 }
1075
1076 static int nbd_start_device(struct nbd_device *nbd)
1077 {
1078         struct nbd_config *config = nbd->config;
1079         int num_connections = config->num_connections;
1080         int error = 0, i;
1081
1082         if (nbd->task_recv)
1083                 return -EBUSY;
1084         if (!config->socks)
1085                 return -EINVAL;
1086         if (num_connections > 1 &&
1087             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1088                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1089                 return -EINVAL;
1090         }
1091
1092         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1093         nbd->task_recv = current;
1094
1095         nbd_parse_flags(nbd);
1096
1097         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1098         if (error) {
1099                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1100                 return error;
1101         }
1102         set_bit(NBD_HAS_PID_FILE, &config->runtime_flags);
1103
1104         nbd_dev_dbg_init(nbd);
1105         for (i = 0; i < num_connections; i++) {
1106                 struct recv_thread_args *args;
1107
1108                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1109                 if (!args) {
1110                         sock_shutdown(nbd);
1111                         return -ENOMEM;
1112                 }
1113                 sk_set_memalloc(config->socks[i]->sock->sk);
1114                 if (nbd->tag_set.timeout)
1115                         config->socks[i]->sock->sk->sk_sndtimeo =
1116                                 nbd->tag_set.timeout;
1117                 atomic_inc(&config->recv_threads);
1118                 refcount_inc(&nbd->config_refs);
1119                 INIT_WORK(&args->work, recv_work);
1120                 args->nbd = nbd;
1121                 args->index = i;
1122                 queue_work(recv_workqueue, &args->work);
1123         }
1124         nbd_size_update(nbd);
1125         return error;
1126 }
1127
1128 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1129 {
1130         struct nbd_config *config = nbd->config;
1131         int ret;
1132
1133         ret = nbd_start_device(nbd);
1134         if (ret)
1135                 return ret;
1136
1137         if (max_part)
1138                 bdev->bd_invalidated = 1;
1139         mutex_unlock(&nbd->config_lock);
1140         ret = wait_event_interruptible(config->recv_wq,
1141                                          atomic_read(&config->recv_threads) == 0);
1142         if (ret)
1143                 sock_shutdown(nbd);
1144         mutex_lock(&nbd->config_lock);
1145         nbd_bdev_reset(bdev);
1146         /* user requested, ignore socket errors */
1147         if (test_bit(NBD_DISCONNECT_REQUESTED, &config->runtime_flags))
1148                 ret = 0;
1149         if (test_bit(NBD_TIMEDOUT, &config->runtime_flags))
1150                 ret = -ETIMEDOUT;
1151         return ret;
1152 }
1153
1154 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1155                                  struct block_device *bdev)
1156 {
1157         sock_shutdown(nbd);
1158         kill_bdev(bdev);
1159         nbd_bdev_reset(bdev);
1160         if (test_and_clear_bit(NBD_HAS_CONFIG_REF,
1161                                &nbd->config->runtime_flags))
1162                 nbd_config_put(nbd);
1163 }
1164
1165 /* Must be called with config_lock held */
1166 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1167                        unsigned int cmd, unsigned long arg)
1168 {
1169         struct nbd_config *config = nbd->config;
1170
1171         switch (cmd) {
1172         case NBD_DISCONNECT:
1173                 return nbd_disconnect(nbd);
1174         case NBD_CLEAR_SOCK:
1175                 nbd_clear_sock_ioctl(nbd, bdev);
1176                 return 0;
1177         case NBD_SET_SOCK:
1178                 return nbd_add_socket(nbd, arg, false);
1179         case NBD_SET_BLKSIZE:
1180                 nbd_size_set(nbd, arg,
1181                              div_s64(config->bytesize, arg));
1182                 return 0;
1183         case NBD_SET_SIZE:
1184                 nbd_size_set(nbd, config->blksize,
1185                              div_s64(arg, config->blksize));
1186                 return 0;
1187         case NBD_SET_SIZE_BLOCKS:
1188                 nbd_size_set(nbd, config->blksize, arg);
1189                 return 0;
1190         case NBD_SET_TIMEOUT:
1191                 if (arg) {
1192                         nbd->tag_set.timeout = arg * HZ;
1193                         blk_queue_rq_timeout(nbd->disk->queue, arg * HZ);
1194                 }
1195                 return 0;
1196
1197         case NBD_SET_FLAGS:
1198                 config->flags = arg;
1199                 return 0;
1200         case NBD_DO_IT:
1201                 return nbd_start_device_ioctl(nbd, bdev);
1202         case NBD_CLEAR_QUE:
1203                 /*
1204                  * This is for compatibility only.  The queue is always cleared
1205                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1206                  */
1207                 return 0;
1208         case NBD_PRINT_DEBUG:
1209                 /*
1210                  * For compatibility only, we no longer keep a list of
1211                  * outstanding requests.
1212                  */
1213                 return 0;
1214         }
1215         return -ENOTTY;
1216 }
1217
1218 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1219                      unsigned int cmd, unsigned long arg)
1220 {
1221         struct nbd_device *nbd = bdev->bd_disk->private_data;
1222         struct nbd_config *config = nbd->config;
1223         int error = -EINVAL;
1224
1225         if (!capable(CAP_SYS_ADMIN))
1226                 return -EPERM;
1227
1228         /* The block layer will pass back some non-nbd ioctls in case we have
1229          * special handling for them, but we don't so just return an error.
1230          */
1231         if (_IOC_TYPE(cmd) != 0xab)
1232                 return -EINVAL;
1233
1234         mutex_lock(&nbd->config_lock);
1235
1236         /* Don't allow ioctl operations on a nbd device that was created with
1237          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1238          */
1239         if (!test_bit(NBD_BOUND, &config->runtime_flags) ||
1240             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1241                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1242         else
1243                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1244         mutex_unlock(&nbd->config_lock);
1245         return error;
1246 }
1247
1248 static struct nbd_config *nbd_alloc_config(void)
1249 {
1250         struct nbd_config *config;
1251
1252         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1253         if (!config)
1254                 return NULL;
1255         atomic_set(&config->recv_threads, 0);
1256         init_waitqueue_head(&config->recv_wq);
1257         init_waitqueue_head(&config->conn_wait);
1258         config->blksize = 1024;
1259         atomic_set(&config->live_connections, 0);
1260         try_module_get(THIS_MODULE);
1261         return config;
1262 }
1263
1264 static int nbd_open(struct block_device *bdev, fmode_t mode)
1265 {
1266         struct nbd_device *nbd;
1267         int ret = 0;
1268
1269         mutex_lock(&nbd_index_mutex);
1270         nbd = bdev->bd_disk->private_data;
1271         if (!nbd) {
1272                 ret = -ENXIO;
1273                 goto out;
1274         }
1275         if (!refcount_inc_not_zero(&nbd->refs)) {
1276                 ret = -ENXIO;
1277                 goto out;
1278         }
1279         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1280                 struct nbd_config *config;
1281
1282                 mutex_lock(&nbd->config_lock);
1283                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1284                         mutex_unlock(&nbd->config_lock);
1285                         goto out;
1286                 }
1287                 config = nbd->config = nbd_alloc_config();
1288                 if (!config) {
1289                         ret = -ENOMEM;
1290                         mutex_unlock(&nbd->config_lock);
1291                         goto out;
1292                 }
1293                 refcount_set(&nbd->config_refs, 1);
1294                 refcount_inc(&nbd->refs);
1295                 mutex_unlock(&nbd->config_lock);
1296                 bdev->bd_invalidated = 1;
1297         } else if (nbd_disconnected(nbd->config)) {
1298                 bdev->bd_invalidated = 1;
1299         }
1300 out:
1301         mutex_unlock(&nbd_index_mutex);
1302         return ret;
1303 }
1304
1305 static void nbd_release(struct gendisk *disk, fmode_t mode)
1306 {
1307         struct nbd_device *nbd = disk->private_data;
1308         nbd_config_put(nbd);
1309         nbd_put(nbd);
1310 }
1311
1312 static const struct block_device_operations nbd_fops =
1313 {
1314         .owner =        THIS_MODULE,
1315         .open =         nbd_open,
1316         .release =      nbd_release,
1317         .ioctl =        nbd_ioctl,
1318         .compat_ioctl = nbd_ioctl,
1319 };
1320
1321 #if IS_ENABLED(CONFIG_DEBUG_FS)
1322
1323 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1324 {
1325         struct nbd_device *nbd = s->private;
1326
1327         if (nbd->task_recv)
1328                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1329
1330         return 0;
1331 }
1332
1333 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1334 {
1335         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1336 }
1337
1338 static const struct file_operations nbd_dbg_tasks_ops = {
1339         .open = nbd_dbg_tasks_open,
1340         .read = seq_read,
1341         .llseek = seq_lseek,
1342         .release = single_release,
1343 };
1344
1345 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1346 {
1347         struct nbd_device *nbd = s->private;
1348         u32 flags = nbd->config->flags;
1349
1350         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1351
1352         seq_puts(s, "Known flags:\n");
1353
1354         if (flags & NBD_FLAG_HAS_FLAGS)
1355                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1356         if (flags & NBD_FLAG_READ_ONLY)
1357                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1358         if (flags & NBD_FLAG_SEND_FLUSH)
1359                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1360         if (flags & NBD_FLAG_SEND_FUA)
1361                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1362         if (flags & NBD_FLAG_SEND_TRIM)
1363                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1364
1365         return 0;
1366 }
1367
1368 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1369 {
1370         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1371 }
1372
1373 static const struct file_operations nbd_dbg_flags_ops = {
1374         .open = nbd_dbg_flags_open,
1375         .read = seq_read,
1376         .llseek = seq_lseek,
1377         .release = single_release,
1378 };
1379
1380 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1381 {
1382         struct dentry *dir;
1383         struct nbd_config *config = nbd->config;
1384
1385         if (!nbd_dbg_dir)
1386                 return -EIO;
1387
1388         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1389         if (!dir) {
1390                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1391                         nbd_name(nbd));
1392                 return -EIO;
1393         }
1394         config->dbg_dir = dir;
1395
1396         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1397         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1398         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1399         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1400         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1401
1402         return 0;
1403 }
1404
1405 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1406 {
1407         debugfs_remove_recursive(nbd->config->dbg_dir);
1408 }
1409
1410 static int nbd_dbg_init(void)
1411 {
1412         struct dentry *dbg_dir;
1413
1414         dbg_dir = debugfs_create_dir("nbd", NULL);
1415         if (!dbg_dir)
1416                 return -EIO;
1417
1418         nbd_dbg_dir = dbg_dir;
1419
1420         return 0;
1421 }
1422
1423 static void nbd_dbg_close(void)
1424 {
1425         debugfs_remove_recursive(nbd_dbg_dir);
1426 }
1427
1428 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1429
1430 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1431 {
1432         return 0;
1433 }
1434
1435 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1436 {
1437 }
1438
1439 static int nbd_dbg_init(void)
1440 {
1441         return 0;
1442 }
1443
1444 static void nbd_dbg_close(void)
1445 {
1446 }
1447
1448 #endif
1449
1450 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1451                             unsigned int hctx_idx, unsigned int numa_node)
1452 {
1453         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1454         cmd->nbd = set->driver_data;
1455         return 0;
1456 }
1457
1458 static const struct blk_mq_ops nbd_mq_ops = {
1459         .queue_rq       = nbd_queue_rq,
1460         .complete       = nbd_complete_rq,
1461         .init_request   = nbd_init_request,
1462         .timeout        = nbd_xmit_timeout,
1463 };
1464
1465 static int nbd_dev_add(int index)
1466 {
1467         struct nbd_device *nbd;
1468         struct gendisk *disk;
1469         struct request_queue *q;
1470         int err = -ENOMEM;
1471
1472         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1473         if (!nbd)
1474                 goto out;
1475
1476         disk = alloc_disk(1 << part_shift);
1477         if (!disk)
1478                 goto out_free_nbd;
1479
1480         if (index >= 0) {
1481                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1482                                 GFP_KERNEL);
1483                 if (err == -ENOSPC)
1484                         err = -EEXIST;
1485         } else {
1486                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1487                 if (err >= 0)
1488                         index = err;
1489         }
1490         if (err < 0)
1491                 goto out_free_disk;
1492
1493         nbd->index = index;
1494         nbd->disk = disk;
1495         nbd->tag_set.ops = &nbd_mq_ops;
1496         nbd->tag_set.nr_hw_queues = 1;
1497         nbd->tag_set.queue_depth = 128;
1498         nbd->tag_set.numa_node = NUMA_NO_NODE;
1499         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1500         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1501                 BLK_MQ_F_SG_MERGE | BLK_MQ_F_BLOCKING;
1502         nbd->tag_set.driver_data = nbd;
1503
1504         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1505         if (err)
1506                 goto out_free_idr;
1507
1508         q = blk_mq_init_queue(&nbd->tag_set);
1509         if (IS_ERR(q)) {
1510                 err = PTR_ERR(q);
1511                 goto out_free_tags;
1512         }
1513         disk->queue = q;
1514
1515         /*
1516          * Tell the block layer that we are not a rotational device
1517          */
1518         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1519         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1520         disk->queue->limits.discard_granularity = 0;
1521         disk->queue->limits.discard_alignment = 0;
1522         blk_queue_max_discard_sectors(disk->queue, 0);
1523         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1524         blk_queue_max_segments(disk->queue, USHRT_MAX);
1525         blk_queue_max_hw_sectors(disk->queue, 65536);
1526         disk->queue->limits.max_sectors = 256;
1527
1528         mutex_init(&nbd->config_lock);
1529         refcount_set(&nbd->config_refs, 0);
1530         refcount_set(&nbd->refs, 1);
1531         INIT_LIST_HEAD(&nbd->list);
1532         disk->major = NBD_MAJOR;
1533         disk->first_minor = index << part_shift;
1534         disk->fops = &nbd_fops;
1535         disk->private_data = nbd;
1536         sprintf(disk->disk_name, "nbd%d", index);
1537         add_disk(disk);
1538         nbd_total_devices++;
1539         return index;
1540
1541 out_free_tags:
1542         blk_mq_free_tag_set(&nbd->tag_set);
1543 out_free_idr:
1544         idr_remove(&nbd_index_idr, index);
1545 out_free_disk:
1546         put_disk(disk);
1547 out_free_nbd:
1548         kfree(nbd);
1549 out:
1550         return err;
1551 }
1552
1553 static int find_free_cb(int id, void *ptr, void *data)
1554 {
1555         struct nbd_device *nbd = ptr;
1556         struct nbd_device **found = data;
1557
1558         if (!refcount_read(&nbd->config_refs)) {
1559                 *found = nbd;
1560                 return 1;
1561         }
1562         return 0;
1563 }
1564
1565 /* Netlink interface. */
1566 static struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1567         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1568         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1569         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1570         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1571         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1572         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1573         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1574         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1575         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1576 };
1577
1578 static struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1579         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1580 };
1581
1582 /* We don't use this right now since we don't parse the incoming list, but we
1583  * still want it here so userspace knows what to expect.
1584  */
1585 static struct nla_policy __attribute__((unused))
1586 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1587         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1588         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1589 };
1590
1591 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1592 {
1593         struct nbd_device *nbd = NULL;
1594         struct nbd_config *config;
1595         int index = -1;
1596         int ret;
1597         bool put_dev = false;
1598
1599         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1600                 return -EPERM;
1601
1602         if (info->attrs[NBD_ATTR_INDEX])
1603                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1604         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1605                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1606                 return -EINVAL;
1607         }
1608         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1609                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1610                 return -EINVAL;
1611         }
1612 again:
1613         mutex_lock(&nbd_index_mutex);
1614         if (index == -1) {
1615                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1616                 if (ret == 0) {
1617                         int new_index;
1618                         new_index = nbd_dev_add(-1);
1619                         if (new_index < 0) {
1620                                 mutex_unlock(&nbd_index_mutex);
1621                                 printk(KERN_ERR "nbd: failed to add new device\n");
1622                                 return new_index;
1623                         }
1624                         nbd = idr_find(&nbd_index_idr, new_index);
1625                 }
1626         } else {
1627                 nbd = idr_find(&nbd_index_idr, index);
1628                 if (!nbd) {
1629                         ret = nbd_dev_add(index);
1630                         if (ret < 0) {
1631                                 mutex_unlock(&nbd_index_mutex);
1632                                 printk(KERN_ERR "nbd: failed to add new device\n");
1633                                 return ret;
1634                         }
1635                         nbd = idr_find(&nbd_index_idr, index);
1636                 }
1637         }
1638         if (!nbd) {
1639                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1640                        index);
1641                 mutex_unlock(&nbd_index_mutex);
1642                 return -EINVAL;
1643         }
1644         if (!refcount_inc_not_zero(&nbd->refs)) {
1645                 mutex_unlock(&nbd_index_mutex);
1646                 if (index == -1)
1647                         goto again;
1648                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1649                        index);
1650                 return -EINVAL;
1651         }
1652         mutex_unlock(&nbd_index_mutex);
1653
1654         mutex_lock(&nbd->config_lock);
1655         if (refcount_read(&nbd->config_refs)) {
1656                 mutex_unlock(&nbd->config_lock);
1657                 nbd_put(nbd);
1658                 if (index == -1)
1659                         goto again;
1660                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1661                 return -EBUSY;
1662         }
1663         if (WARN_ON(nbd->config)) {
1664                 mutex_unlock(&nbd->config_lock);
1665                 nbd_put(nbd);
1666                 return -EINVAL;
1667         }
1668         config = nbd->config = nbd_alloc_config();
1669         if (!nbd->config) {
1670                 mutex_unlock(&nbd->config_lock);
1671                 nbd_put(nbd);
1672                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1673                 return -ENOMEM;
1674         }
1675         refcount_set(&nbd->config_refs, 1);
1676         set_bit(NBD_BOUND, &config->runtime_flags);
1677
1678         if (info->attrs[NBD_ATTR_SIZE_BYTES]) {
1679                 u64 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1680                 nbd_size_set(nbd, config->blksize,
1681                              div64_u64(bytes, config->blksize));
1682         }
1683         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1684                 u64 bsize =
1685                         nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1686                 nbd_size_set(nbd, bsize, div64_u64(config->bytesize, bsize));
1687         }
1688         if (info->attrs[NBD_ATTR_TIMEOUT]) {
1689                 u64 timeout = nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]);
1690                 nbd->tag_set.timeout = timeout * HZ;
1691                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1692         }
1693         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1694                 config->dead_conn_timeout =
1695                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1696                 config->dead_conn_timeout *= HZ;
1697         }
1698         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1699                 config->flags =
1700                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1701         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1702                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1703                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1704                         set_bit(NBD_DESTROY_ON_DISCONNECT,
1705                                 &config->runtime_flags);
1706                         put_dev = true;
1707                 }
1708         }
1709
1710         if (info->attrs[NBD_ATTR_SOCKETS]) {
1711                 struct nlattr *attr;
1712                 int rem, fd;
1713
1714                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1715                                     rem) {
1716                         struct nlattr *socks[NBD_SOCK_MAX+1];
1717
1718                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1719                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1720                                 ret = -EINVAL;
1721                                 goto out;
1722                         }
1723                         ret = nla_parse_nested(socks, NBD_SOCK_MAX, attr,
1724                                                nbd_sock_policy, info->extack);
1725                         if (ret != 0) {
1726                                 printk(KERN_ERR "nbd: error processing sock list\n");
1727                                 ret = -EINVAL;
1728                                 goto out;
1729                         }
1730                         if (!socks[NBD_SOCK_FD])
1731                                 continue;
1732                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1733                         ret = nbd_add_socket(nbd, fd, true);
1734                         if (ret)
1735                                 goto out;
1736                 }
1737         }
1738         ret = nbd_start_device(nbd);
1739 out:
1740         mutex_unlock(&nbd->config_lock);
1741         if (!ret) {
1742                 set_bit(NBD_HAS_CONFIG_REF, &config->runtime_flags);
1743                 refcount_inc(&nbd->config_refs);
1744                 nbd_connect_reply(info, nbd->index);
1745         }
1746         nbd_config_put(nbd);
1747         if (put_dev)
1748                 nbd_put(nbd);
1749         return ret;
1750 }
1751
1752 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
1753 {
1754         struct nbd_device *nbd;
1755         int index;
1756
1757         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1758                 return -EPERM;
1759
1760         if (!info->attrs[NBD_ATTR_INDEX]) {
1761                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
1762                 return -EINVAL;
1763         }
1764         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1765         mutex_lock(&nbd_index_mutex);
1766         nbd = idr_find(&nbd_index_idr, index);
1767         if (!nbd) {
1768                 mutex_unlock(&nbd_index_mutex);
1769                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1770                        index);
1771                 return -EINVAL;
1772         }
1773         if (!refcount_inc_not_zero(&nbd->refs)) {
1774                 mutex_unlock(&nbd_index_mutex);
1775                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1776                        index);
1777                 return -EINVAL;
1778         }
1779         mutex_unlock(&nbd_index_mutex);
1780         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1781                 nbd_put(nbd);
1782                 return 0;
1783         }
1784         mutex_lock(&nbd->config_lock);
1785         nbd_disconnect(nbd);
1786         nbd_clear_sock(nbd);
1787         mutex_unlock(&nbd->config_lock);
1788         if (test_and_clear_bit(NBD_HAS_CONFIG_REF,
1789                                &nbd->config->runtime_flags))
1790                 nbd_config_put(nbd);
1791         nbd_config_put(nbd);
1792         nbd_put(nbd);
1793         return 0;
1794 }
1795
1796 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
1797 {
1798         struct nbd_device *nbd = NULL;
1799         struct nbd_config *config;
1800         int index;
1801         int ret = -EINVAL;
1802         bool put_dev = false;
1803
1804         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1805                 return -EPERM;
1806
1807         if (!info->attrs[NBD_ATTR_INDEX]) {
1808                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
1809                 return -EINVAL;
1810         }
1811         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1812         mutex_lock(&nbd_index_mutex);
1813         nbd = idr_find(&nbd_index_idr, index);
1814         if (!nbd) {
1815                 mutex_unlock(&nbd_index_mutex);
1816                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
1817                        index);
1818                 return -EINVAL;
1819         }
1820         if (!refcount_inc_not_zero(&nbd->refs)) {
1821                 mutex_unlock(&nbd_index_mutex);
1822                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1823                        index);
1824                 return -EINVAL;
1825         }
1826         mutex_unlock(&nbd_index_mutex);
1827
1828         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1829                 dev_err(nbd_to_dev(nbd),
1830                         "not configured, cannot reconfigure\n");
1831                 nbd_put(nbd);
1832                 return -EINVAL;
1833         }
1834
1835         mutex_lock(&nbd->config_lock);
1836         config = nbd->config;
1837         if (!test_bit(NBD_BOUND, &config->runtime_flags) ||
1838             !nbd->task_recv) {
1839                 dev_err(nbd_to_dev(nbd),
1840                         "not configured, cannot reconfigure\n");
1841                 goto out;
1842         }
1843
1844         if (info->attrs[NBD_ATTR_TIMEOUT]) {
1845                 u64 timeout = nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]);
1846                 nbd->tag_set.timeout = timeout * HZ;
1847                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1848         }
1849         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1850                 config->dead_conn_timeout =
1851                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1852                 config->dead_conn_timeout *= HZ;
1853         }
1854         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1855                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1856                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1857                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1858                                               &config->runtime_flags))
1859                                 put_dev = true;
1860                 } else {
1861                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1862                                                &config->runtime_flags))
1863                                 refcount_inc(&nbd->refs);
1864                 }
1865         }
1866
1867         if (info->attrs[NBD_ATTR_SOCKETS]) {
1868                 struct nlattr *attr;
1869                 int rem, fd;
1870
1871                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1872                                     rem) {
1873                         struct nlattr *socks[NBD_SOCK_MAX+1];
1874
1875                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1876                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1877                                 ret = -EINVAL;
1878                                 goto out;
1879                         }
1880                         ret = nla_parse_nested(socks, NBD_SOCK_MAX, attr,
1881                                                nbd_sock_policy, info->extack);
1882                         if (ret != 0) {
1883                                 printk(KERN_ERR "nbd: error processing sock list\n");
1884                                 ret = -EINVAL;
1885                                 goto out;
1886                         }
1887                         if (!socks[NBD_SOCK_FD])
1888                                 continue;
1889                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1890                         ret = nbd_reconnect_socket(nbd, fd);
1891                         if (ret) {
1892                                 if (ret == -ENOSPC)
1893                                         ret = 0;
1894                                 goto out;
1895                         }
1896                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
1897                 }
1898         }
1899 out:
1900         mutex_unlock(&nbd->config_lock);
1901         nbd_config_put(nbd);
1902         nbd_put(nbd);
1903         if (put_dev)
1904                 nbd_put(nbd);
1905         return ret;
1906 }
1907
1908 static const struct genl_ops nbd_connect_genl_ops[] = {
1909         {
1910                 .cmd    = NBD_CMD_CONNECT,
1911                 .policy = nbd_attr_policy,
1912                 .doit   = nbd_genl_connect,
1913         },
1914         {
1915                 .cmd    = NBD_CMD_DISCONNECT,
1916                 .policy = nbd_attr_policy,
1917                 .doit   = nbd_genl_disconnect,
1918         },
1919         {
1920                 .cmd    = NBD_CMD_RECONFIGURE,
1921                 .policy = nbd_attr_policy,
1922                 .doit   = nbd_genl_reconfigure,
1923         },
1924         {
1925                 .cmd    = NBD_CMD_STATUS,
1926                 .policy = nbd_attr_policy,
1927                 .doit   = nbd_genl_status,
1928         },
1929 };
1930
1931 static const struct genl_multicast_group nbd_mcast_grps[] = {
1932         { .name = NBD_GENL_MCAST_GROUP_NAME, },
1933 };
1934
1935 static struct genl_family nbd_genl_family __ro_after_init = {
1936         .hdrsize        = 0,
1937         .name           = NBD_GENL_FAMILY_NAME,
1938         .version        = NBD_GENL_VERSION,
1939         .module         = THIS_MODULE,
1940         .ops            = nbd_connect_genl_ops,
1941         .n_ops          = ARRAY_SIZE(nbd_connect_genl_ops),
1942         .maxattr        = NBD_ATTR_MAX,
1943         .mcgrps         = nbd_mcast_grps,
1944         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
1945 };
1946
1947 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
1948 {
1949         struct nlattr *dev_opt;
1950         u8 connected = 0;
1951         int ret;
1952
1953         /* This is a little racey, but for status it's ok.  The
1954          * reason we don't take a ref here is because we can't
1955          * take a ref in the index == -1 case as we would need
1956          * to put under the nbd_index_mutex, which could
1957          * deadlock if we are configured to remove ourselves
1958          * once we're disconnected.
1959          */
1960         if (refcount_read(&nbd->config_refs))
1961                 connected = 1;
1962         dev_opt = nla_nest_start(reply, NBD_DEVICE_ITEM);
1963         if (!dev_opt)
1964                 return -EMSGSIZE;
1965         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
1966         if (ret)
1967                 return -EMSGSIZE;
1968         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
1969                          connected);
1970         if (ret)
1971                 return -EMSGSIZE;
1972         nla_nest_end(reply, dev_opt);
1973         return 0;
1974 }
1975
1976 static int status_cb(int id, void *ptr, void *data)
1977 {
1978         struct nbd_device *nbd = ptr;
1979         return populate_nbd_status(nbd, (struct sk_buff *)data);
1980 }
1981
1982 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
1983 {
1984         struct nlattr *dev_list;
1985         struct sk_buff *reply;
1986         void *reply_head;
1987         size_t msg_size;
1988         int index = -1;
1989         int ret = -ENOMEM;
1990
1991         if (info->attrs[NBD_ATTR_INDEX])
1992                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1993
1994         mutex_lock(&nbd_index_mutex);
1995
1996         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
1997                                   nla_attr_size(sizeof(u8)));
1998         msg_size *= (index == -1) ? nbd_total_devices : 1;
1999
2000         reply = genlmsg_new(msg_size, GFP_KERNEL);
2001         if (!reply)
2002                 goto out;
2003         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2004                                        NBD_CMD_STATUS);
2005         if (!reply_head) {
2006                 nlmsg_free(reply);
2007                 goto out;
2008         }
2009
2010         dev_list = nla_nest_start(reply, NBD_ATTR_DEVICE_LIST);
2011         if (index == -1) {
2012                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2013                 if (ret) {
2014                         nlmsg_free(reply);
2015                         goto out;
2016                 }
2017         } else {
2018                 struct nbd_device *nbd;
2019                 nbd = idr_find(&nbd_index_idr, index);
2020                 if (nbd) {
2021                         ret = populate_nbd_status(nbd, reply);
2022                         if (ret) {
2023                                 nlmsg_free(reply);
2024                                 goto out;
2025                         }
2026                 }
2027         }
2028         nla_nest_end(reply, dev_list);
2029         genlmsg_end(reply, reply_head);
2030         genlmsg_reply(reply, info);
2031         ret = 0;
2032 out:
2033         mutex_unlock(&nbd_index_mutex);
2034         return ret;
2035 }
2036
2037 static void nbd_connect_reply(struct genl_info *info, int index)
2038 {
2039         struct sk_buff *skb;
2040         void *msg_head;
2041         int ret;
2042
2043         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2044         if (!skb)
2045                 return;
2046         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2047                                      NBD_CMD_CONNECT);
2048         if (!msg_head) {
2049                 nlmsg_free(skb);
2050                 return;
2051         }
2052         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2053         if (ret) {
2054                 nlmsg_free(skb);
2055                 return;
2056         }
2057         genlmsg_end(skb, msg_head);
2058         genlmsg_reply(skb, info);
2059 }
2060
2061 static void nbd_mcast_index(int index)
2062 {
2063         struct sk_buff *skb;
2064         void *msg_head;
2065         int ret;
2066
2067         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2068         if (!skb)
2069                 return;
2070         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2071                                      NBD_CMD_LINK_DEAD);
2072         if (!msg_head) {
2073                 nlmsg_free(skb);
2074                 return;
2075         }
2076         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2077         if (ret) {
2078                 nlmsg_free(skb);
2079                 return;
2080         }
2081         genlmsg_end(skb, msg_head);
2082         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2083 }
2084
2085 static void nbd_dead_link_work(struct work_struct *work)
2086 {
2087         struct link_dead_args *args = container_of(work, struct link_dead_args,
2088                                                    work);
2089         nbd_mcast_index(args->index);
2090         kfree(args);
2091 }
2092
2093 static int __init nbd_init(void)
2094 {
2095         int i;
2096
2097         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2098
2099         if (max_part < 0) {
2100                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2101                 return -EINVAL;
2102         }
2103
2104         part_shift = 0;
2105         if (max_part > 0) {
2106                 part_shift = fls(max_part);
2107
2108                 /*
2109                  * Adjust max_part according to part_shift as it is exported
2110                  * to user space so that user can know the max number of
2111                  * partition kernel should be able to manage.
2112                  *
2113                  * Note that -1 is required because partition 0 is reserved
2114                  * for the whole disk.
2115                  */
2116                 max_part = (1UL << part_shift) - 1;
2117         }
2118
2119         if ((1UL << part_shift) > DISK_MAX_PARTS)
2120                 return -EINVAL;
2121
2122         if (nbds_max > 1UL << (MINORBITS - part_shift))
2123                 return -EINVAL;
2124         recv_workqueue = alloc_workqueue("knbd-recv",
2125                                          WQ_MEM_RECLAIM | WQ_HIGHPRI |
2126                                          WQ_UNBOUND, 0);
2127         if (!recv_workqueue)
2128                 return -ENOMEM;
2129
2130         if (register_blkdev(NBD_MAJOR, "nbd")) {
2131                 destroy_workqueue(recv_workqueue);
2132                 return -EIO;
2133         }
2134
2135         if (genl_register_family(&nbd_genl_family)) {
2136                 unregister_blkdev(NBD_MAJOR, "nbd");
2137                 destroy_workqueue(recv_workqueue);
2138                 return -EINVAL;
2139         }
2140         nbd_dbg_init();
2141
2142         mutex_lock(&nbd_index_mutex);
2143         for (i = 0; i < nbds_max; i++)
2144                 nbd_dev_add(i);
2145         mutex_unlock(&nbd_index_mutex);
2146         return 0;
2147 }
2148
2149 static int nbd_exit_cb(int id, void *ptr, void *data)
2150 {
2151         struct list_head *list = (struct list_head *)data;
2152         struct nbd_device *nbd = ptr;
2153
2154         list_add_tail(&nbd->list, list);
2155         return 0;
2156 }
2157
2158 static void __exit nbd_cleanup(void)
2159 {
2160         struct nbd_device *nbd;
2161         LIST_HEAD(del_list);
2162
2163         nbd_dbg_close();
2164
2165         mutex_lock(&nbd_index_mutex);
2166         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2167         mutex_unlock(&nbd_index_mutex);
2168
2169         while (!list_empty(&del_list)) {
2170                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2171                 list_del_init(&nbd->list);
2172                 if (refcount_read(&nbd->refs) != 1)
2173                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2174                 nbd_put(nbd);
2175         }
2176
2177         idr_destroy(&nbd_index_idr);
2178         genl_unregister_family(&nbd_genl_family);
2179         destroy_workqueue(recv_workqueue);
2180         unregister_blkdev(NBD_MAJOR, "nbd");
2181 }
2182
2183 module_init(nbd_init);
2184 module_exit(nbd_cleanup);
2185
2186 MODULE_DESCRIPTION("Network Block Device");
2187 MODULE_LICENSE("GPL");
2188
2189 module_param(nbds_max, int, 0444);
2190 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2191 module_param(max_part, int, 0444);
2192 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");