drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <[email protected]>
   9  * Parts copyright 2001 Steven Whitehouse <[email protected]>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #include <linux/major.h>
  15
  16 #include <linux/blkdev.h>
  17 #include <linux/module.h>
  18 #include <linux/init.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/mm.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/completion.h>
  30 #include <linux/err.h>
  31 #include <linux/kernel.h>
  32 #include <linux/slab.h>
  33 #include <net/sock.h>
  34 #include <linux/net.h>
  35 #include <linux/kthread.h>
  36 #include <linux/types.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/blk-mq.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/types.h>
  42
  43 #include <linux/nbd.h>
  44 #include <linux/nbd-netlink.h>
  45 #include <net/genetlink.h>
  46
  47 #define CREATE_TRACE_POINTS
  48 #include <trace/events/nbd.h>
  49
  50 static DEFINE_IDR(nbd_index_idr);
  51 static DEFINE_MUTEX(nbd_index_mutex);
  52 static int nbd_total_devices = 0;
  53
  54 struct nbd_sock {
  55         struct socket *sock;
  56         struct mutex tx_lock;
  57         struct request *pending;
  58         int sent;
  59         bool dead;
  60         int fallback_index;
  61         int cookie;
  62 };
  63
  64 struct recv_thread_args {
  65         struct work_struct work;
  66         struct nbd_device *nbd;
  67         int index;
  68 };
  69
  70 struct link_dead_args {
  71         struct work_struct work;
  72         int index;
  73 };
  74
  75 #define NBD_RT_TIMEDOUT                 0
  76 #define NBD_RT_DISCONNECT_REQUESTED     1
  77 #define NBD_RT_DISCONNECTED             2
  78 #define NBD_RT_HAS_PID_FILE             3
  79 #define NBD_RT_HAS_CONFIG_REF           4
  80 #define NBD_RT_BOUND                    5
  81 #define NBD_RT_DESTROY_ON_DISCONNECT    6
  82 #define NBD_RT_DISCONNECT_ON_CLOSE      7
  83
  84 #define NBD_DESTROY_ON_DISCONNECT       0
  85 #define NBD_DISCONNECT_REQUESTED        1
  86
  87 struct nbd_config {
  88         u32 flags;
  89         unsigned long runtime_flags;
  90         u64 dead_conn_timeout;
  91
  92         struct nbd_sock **socks;
  93         int num_connections;
  94         atomic_t live_connections;
  95         wait_queue_head_t conn_wait;
  96
  97         atomic_t recv_threads;
  98         wait_queue_head_t recv_wq;
  99         loff_t blksize;
 100         loff_t bytesize;
 101 #if IS_ENABLED(CONFIG_DEBUG_FS)
 102         struct dentry *dbg_dir;
 103 #endif
 104 };
 105
 106 struct nbd_device {
 107         struct blk_mq_tag_set tag_set;
 108
 109         int index;
 110         refcount_t config_refs;
 111         refcount_t refs;
 112         struct nbd_config *config;
 113         struct mutex config_lock;
 114         struct gendisk *disk;
 115         struct workqueue_struct *recv_workq;
 116
 117         struct list_head list;
 118         struct task_struct *task_recv;
 119         struct task_struct *task_setup;
 120
 121         struct completion *destroy_complete;
 122         unsigned long flags;
 123 };
 124
 125 #define NBD_CMD_REQUEUED        1
 126
 127 struct nbd_cmd {
 128         struct nbd_device *nbd;
 129         struct mutex lock;
 130         int index;
 131         int cookie;
 132         int retries;
 133         blk_status_t status;
 134         unsigned long flags;
 135         u32 cmd_cookie;
 136 };
 137
 138 #if IS_ENABLED(CONFIG_DEBUG_FS)
 139 static struct dentry *nbd_dbg_dir;
 140 #endif
 141
 142 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 143
 144 #define NBD_MAGIC 0x68797548
 145
 146 #define NBD_DEF_BLKSIZE 1024
 147
 148 static unsigned int nbds_max = 16;
 149 static int max_part = 16;
 150 static int part_shift;
 151
 152 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 153 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 154 static void nbd_config_put(struct nbd_device *nbd);
 155 static void nbd_connect_reply(struct genl_info *info, int index);
 156 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 157 static void nbd_dead_link_work(struct work_struct *work);
 158 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 159
 160 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 161 {
 162         return disk_to_dev(nbd->disk);
 163 }
 164
 165 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 166 {
 167         struct request *req = blk_mq_rq_from_pdu(cmd);
 168
 169         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 170                 blk_mq_requeue_request(req, true);
 171 }
 172
 173 #define NBD_COOKIE_BITS 32
 174
 175 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 176 {
 177         struct request *req = blk_mq_rq_from_pdu(cmd);
 178         u32 tag = blk_mq_unique_tag(req);
 179         u64 cookie = cmd->cmd_cookie;
 180
 181         return (cookie << NBD_COOKIE_BITS) | tag;
 182 }
 183
 184 static u32 nbd_handle_to_tag(u64 handle)
 185 {
 186         return (u32)handle;
 187 }
 188
 189 static u32 nbd_handle_to_cookie(u64 handle)
 190 {
 191         return (u32)(handle >> NBD_COOKIE_BITS);
 192 }
 193
 194 static const char *nbdcmd_to_ascii(int cmd)
 195 {
 196         switch (cmd) {
 197         case  NBD_CMD_READ: return "read";
 198         case NBD_CMD_WRITE: return "write";
 199         case  NBD_CMD_DISC: return "disconnect";
 200         case NBD_CMD_FLUSH: return "flush";
 201         case  NBD_CMD_TRIM: return "trim/discard";
 202         }
 203         return "invalid";
 204 }
 205
 206 static ssize_t pid_show(struct device *dev,
 207                         struct device_attribute *attr, char *buf)
 208 {
 209         struct gendisk *disk = dev_to_disk(dev);
 210         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 211
 212         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 213 }
 214
 215 static const struct device_attribute pid_attr = {
 216         .attr = { .name = "pid", .mode = 0444},
 217         .show = pid_show,
 218 };
 219
 220 static void nbd_dev_remove(struct nbd_device *nbd)
 221 {
 222         struct gendisk *disk = nbd->disk;
 223         struct request_queue *q;
 224
 225         if (disk) {
 226                 q = disk->queue;
 227                 del_gendisk(disk);
 228                 blk_cleanup_queue(q);
 229                 blk_mq_free_tag_set(&nbd->tag_set);
 230                 disk->private_data = NULL;
 231                 put_disk(disk);
 232         }
 233
 234         /*
 235          * Place this in the last just before the nbd is freed to
 236          * make sure that the disk and the related kobject are also
 237          * totally removed to avoid duplicate creation of the same
 238          * one.
 239          */
 240         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) && nbd->destroy_complete)
 241                 complete(nbd->destroy_complete);
 242
 243         kfree(nbd);
 244 }
 245
 246 static void nbd_put(struct nbd_device *nbd)
 247 {
 248         if (refcount_dec_and_mutex_lock(&nbd->refs,
 249                                         &nbd_index_mutex)) {
 250                 idr_remove(&nbd_index_idr, nbd->index);
 251                 nbd_dev_remove(nbd);
 252                 mutex_unlock(&nbd_index_mutex);
 253         }
 254 }
 255
 256 static int nbd_disconnected(struct nbd_config *config)
 257 {
 258         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 259                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 260 }
 261
 262 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 263                                 int notify)
 264 {
 265         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 266                 struct link_dead_args *args;
 267                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 268                 if (args) {
 269                         INIT_WORK(&args->work, nbd_dead_link_work);
 270                         args->index = nbd->index;
 271                         queue_work(system_wq, &args->work);
 272                 }
 273         }
 274         if (!nsock->dead) {
 275                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 276                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 277                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 278                                                &nbd->config->runtime_flags)) {
 279                                 set_bit(NBD_RT_DISCONNECTED,
 280                                         &nbd->config->runtime_flags);
 281                                 dev_info(nbd_to_dev(nbd),
 282                                         "Disconnected due to user request.\n");
 283                         }
 284                 }
 285         }
 286         nsock->dead = true;
 287         nsock->pending = NULL;
 288         nsock->sent = 0;
 289 }
 290
 291 static void nbd_size_clear(struct nbd_device *nbd)
 292 {
 293         if (nbd->config->bytesize) {
 294                 set_capacity(nbd->disk, 0);
 295                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 296         }
 297 }
 298
 299 static void nbd_size_update(struct nbd_device *nbd)
 300 {
 301         struct nbd_config *config = nbd->config;
 302         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 303
 304         if (config->flags & NBD_FLAG_SEND_TRIM) {
 305                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 306                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 307                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 308         }
 309         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 310         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 311         set_capacity(nbd->disk, config->bytesize >> 9);
 312         if (bdev) {
 313                 if (bdev->bd_disk) {
 314                         bd_set_size(bdev, config->bytesize);
 315                         set_blocksize(bdev, config->blksize);
 316                 } else
 317                         bdev->bd_invalidated = 1;
 318                 bdput(bdev);
 319         }
 320         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 321 }
 322
 323 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 324                          loff_t nr_blocks)
 325 {
 326         struct nbd_config *config = nbd->config;
 327         config->blksize = blocksize;
 328         config->bytesize = blocksize * nr_blocks;
 329         if (nbd->task_recv != NULL)
 330                 nbd_size_update(nbd);
 331 }
 332
 333 static void nbd_complete_rq(struct request *req)
 334 {
 335         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 336
 337         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 338                 cmd->status ? "failed" : "done");
 339
 340         blk_mq_end_request(req, cmd->status);
 341 }
 342
 343 /*
 344  * Forcibly shutdown the socket causing all listeners to error
 345  */
 346 static void sock_shutdown(struct nbd_device *nbd)
 347 {
 348         struct nbd_config *config = nbd->config;
 349         int i;
 350
 351         if (config->num_connections == 0)
 352                 return;
 353         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 354                 return;
 355
 356         for (i = 0; i < config->num_connections; i++) {
 357                 struct nbd_sock *nsock = config->socks[i];
 358                 mutex_lock(&nsock->tx_lock);
 359                 nbd_mark_nsock_dead(nbd, nsock, 0);
 360                 mutex_unlock(&nsock->tx_lock);
 361         }
 362         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 363 }
 364
 365 static u32 req_to_nbd_cmd_type(struct request *req)
 366 {
 367         switch (req_op(req)) {
 368         case REQ_OP_DISCARD:
 369                 return NBD_CMD_TRIM;
 370         case REQ_OP_FLUSH:
 371                 return NBD_CMD_FLUSH;
 372         case REQ_OP_WRITE:
 373                 return NBD_CMD_WRITE;
 374         case REQ_OP_READ:
 375                 return NBD_CMD_READ;
 376         default:
 377                 return U32_MAX;
 378         }
 379 }
 380
 381 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 382                                                  bool reserved)
 383 {
 384         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 385         struct nbd_device *nbd = cmd->nbd;
 386         struct nbd_config *config;
 387
 388         if (!mutex_trylock(&cmd->lock))
 389                 return BLK_EH_RESET_TIMER;
 390
 391         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 392                 cmd->status = BLK_STS_TIMEOUT;
 393                 mutex_unlock(&cmd->lock);
 394                 goto done;
 395         }
 396         config = nbd->config;
 397
 398         if (config->num_connections > 1 ||
 399             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 400                 dev_err_ratelimited(nbd_to_dev(nbd),
 401                                     "Connection timed out, retrying (%d/%d alive)\n",
 402                                     atomic_read(&config->live_connections),
 403                                     config->num_connections);
 404                 /*
 405                  * Hooray we have more connections, requeue this IO, the submit
 406                  * path will put it on a real connection. Or if only one
 407                  * connection is configured, the submit path will wait util
 408                  * a new connection is reconfigured or util dead timeout.
 409                  */
 410                 if (config->socks) {
 411                         if (cmd->index < config->num_connections) {
 412                                 struct nbd_sock *nsock =
 413                                         config->socks[cmd->index];
 414                                 mutex_lock(&nsock->tx_lock);
 415                                 /* We can have multiple outstanding requests, so
 416                                  * we don't want to mark the nsock dead if we've
 417                                  * already reconnected with a new socket, so
 418                                  * only mark it dead if its the same socket we
 419                                  * were sent out on.
 420                                  */
 421                                 if (cmd->cookie == nsock->cookie)
 422                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 423                                 mutex_unlock(&nsock->tx_lock);
 424                         }
 425                         mutex_unlock(&cmd->lock);
 426                         nbd_requeue_cmd(cmd);
 427                         nbd_config_put(nbd);
 428                         return BLK_EH_DONE;
 429                 }
 430         }
 431
 432         if (!nbd->tag_set.timeout) {
 433                 /*
 434                  * Userspace sets timeout=0 to disable socket disconnection,
 435                  * so just warn and reset the timer.
 436                  */
 437                 struct nbd_sock *nsock = config->socks[cmd->index];
 438                 cmd->retries++;
 439                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 440                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 441                         (unsigned long long)blk_rq_pos(req) << 9,
 442                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 443
 444                 mutex_lock(&nsock->tx_lock);
 445                 if (cmd->cookie != nsock->cookie) {
 446                         nbd_requeue_cmd(cmd);
 447                         mutex_unlock(&nsock->tx_lock);
 448                         mutex_unlock(&cmd->lock);
 449                         nbd_config_put(nbd);
 450                         return BLK_EH_DONE;
 451                 }
 452                 mutex_unlock(&nsock->tx_lock);
 453                 mutex_unlock(&cmd->lock);
 454                 nbd_config_put(nbd);
 455                 return BLK_EH_RESET_TIMER;
 456         }
 457
 458         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 459         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 460         cmd->status = BLK_STS_IOERR;
 461         mutex_unlock(&cmd->lock);
 462         sock_shutdown(nbd);
 463         nbd_config_put(nbd);
 464 done:
 465         blk_mq_complete_request(req);
 466         return BLK_EH_DONE;
 467 }
 468
 469 /*
 470  *  Send or receive packet.
 471  */
 472 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 473                      struct iov_iter *iter, int msg_flags, int *sent)
 474 {
 475         struct nbd_config *config = nbd->config;
 476         struct socket *sock = config->socks[index]->sock;
 477         int result;
 478         struct msghdr msg;
 479         unsigned int noreclaim_flag;
 480
 481         if (unlikely(!sock)) {
 482                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 483                         "Attempted %s on closed socket in sock_xmit\n",
 484                         (send ? "send" : "recv"));
 485                 return -EINVAL;
 486         }
 487
 488         msg.msg_iter = *iter;
 489
 490         noreclaim_flag = memalloc_noreclaim_save();
 491         do {
 492                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 493                 msg.msg_name = NULL;
 494                 msg.msg_namelen = 0;
 495                 msg.msg_control = NULL;
 496                 msg.msg_controllen = 0;
 497                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 498
 499                 if (send)
 500                         result = sock_sendmsg(sock, &msg);
 501                 else
 502                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 503
 504                 if (result <= 0) {
 505                         if (result == 0)
 506                                 result = -EPIPE; /* short read */
 507                         break;
 508                 }
 509                 if (sent)
 510                         *sent += result;
 511         } while (msg_data_left(&msg));
 512
 513         memalloc_noreclaim_restore(noreclaim_flag);
 514
 515         return result;
 516 }
 517
 518 /*
 519  * Different settings for sk->sk_sndtimeo can result in different return values
 520  * if there is a signal pending when we enter sendmsg, because reasons?
 521  */
 522 static inline int was_interrupted(int result)
 523 {
 524         return result == -ERESTARTSYS || result == -EINTR;
 525 }
 526
 527 /* always call with the tx_lock held */
 528 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 529 {
 530         struct request *req = blk_mq_rq_from_pdu(cmd);
 531         struct nbd_config *config = nbd->config;
 532         struct nbd_sock *nsock = config->socks[index];
 533         int result;
 534         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 535         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 536         struct iov_iter from;
 537         unsigned long size = blk_rq_bytes(req);
 538         struct bio *bio;
 539         u64 handle;
 540         u32 type;
 541         u32 nbd_cmd_flags = 0;
 542         int sent = nsock->sent, skip = 0;
 543
 544         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 545
 546         type = req_to_nbd_cmd_type(req);
 547         if (type == U32_MAX)
 548                 return -EIO;
 549
 550         if (rq_data_dir(req) == WRITE &&
 551             (config->flags & NBD_FLAG_READ_ONLY)) {
 552                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 553                                     "Write on read-only\n");
 554                 return -EIO;
 555         }
 556
 557         if (req->cmd_flags & REQ_FUA)
 558                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 559
 560         /* We did a partial send previously, and we at least sent the whole
 561          * request struct, so just go and send the rest of the pages in the
 562          * request.
 563          */
 564         if (sent) {
 565                 if (sent >= sizeof(request)) {
 566                         skip = sent - sizeof(request);
 567
 568                         /* initialize handle for tracing purposes */
 569                         handle = nbd_cmd_handle(cmd);
 570
 571                         goto send_pages;
 572                 }
 573                 iov_iter_advance(&from, sent);
 574         } else {
 575                 cmd->cmd_cookie++;
 576         }
 577         cmd->index = index;
 578         cmd->cookie = nsock->cookie;
 579         cmd->retries = 0;
 580         request.type = htonl(type | nbd_cmd_flags);
 581         if (type != NBD_CMD_FLUSH) {
 582                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 583                 request.len = htonl(size);
 584         }
 585         handle = nbd_cmd_handle(cmd);
 586         memcpy(request.handle, &handle, sizeof(handle));
 587
 588         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 589
 590         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 591                 req, nbdcmd_to_ascii(type),
 592                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 593         result = sock_xmit(nbd, index, 1, &from,
 594                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 595         trace_nbd_header_sent(req, handle);
 596         if (result <= 0) {
 597                 if (was_interrupted(result)) {
 598                         /* If we havne't sent anything we can just return BUSY,
 599                          * however if we have sent something we need to make
 600                          * sure we only allow this req to be sent until we are
 601                          * completely done.
 602                          */
 603                         if (sent) {
 604                                 nsock->pending = req;
 605                                 nsock->sent = sent;
 606                         }
 607                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 608                         return BLK_STS_RESOURCE;
 609                 }
 610                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 611                         "Send control failed (result %d)\n", result);
 612                 return -EAGAIN;
 613         }
 614 send_pages:
 615         if (type != NBD_CMD_WRITE)
 616                 goto out;
 617
 618         bio = req->bio;
 619         while (bio) {
 620                 struct bio *next = bio->bi_next;
 621                 struct bvec_iter iter;
 622                 struct bio_vec bvec;
 623
 624                 bio_for_each_segment(bvec, bio, iter) {
 625                         bool is_last = !next && bio_iter_last(bvec, iter);
 626                         int flags = is_last ? 0 : MSG_MORE;
 627
 628                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 629                                 req, bvec.bv_len);
 630                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 631                         if (skip) {
 632                                 if (skip >= iov_iter_count(&from)) {
 633                                         skip -= iov_iter_count(&from);
 634                                         continue;
 635                                 }
 636                                 iov_iter_advance(&from, skip);
 637                                 skip = 0;
 638                         }
 639                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 640                         if (result <= 0) {
 641                                 if (was_interrupted(result)) {
 642                                         /* We've already sent the header, we
 643                                          * have no choice but to set pending and
 644                                          * return BUSY.
 645                                          */
 646                                         nsock->pending = req;
 647                                         nsock->sent = sent;
 648                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 649                                         return BLK_STS_RESOURCE;
 650                                 }
 651                                 dev_err(disk_to_dev(nbd->disk),
 652                                         "Send data failed (result %d)\n",
 653                                         result);
 654                                 return -EAGAIN;
 655                         }
 656                         /*
 657                          * The completion might already have come in,
 658                          * so break for the last one instead of letting
 659                          * the iterator do it. This prevents use-after-free
 660                          * of the bio.
 661                          */
 662                         if (is_last)
 663                                 break;
 664                 }
 665                 bio = next;
 666         }
 667 out:
 668         trace_nbd_payload_sent(req, handle);
 669         nsock->pending = NULL;
 670         nsock->sent = 0;
 671         return 0;
 672 }
 673
 674 /* NULL returned = something went wrong, inform userspace */
 675 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 676 {
 677         struct nbd_config *config = nbd->config;
 678         int result;
 679         struct nbd_reply reply;
 680         struct nbd_cmd *cmd;
 681         struct request *req = NULL;
 682         u64 handle;
 683         u16 hwq;
 684         u32 tag;
 685         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 686         struct iov_iter to;
 687         int ret = 0;
 688
 689         reply.magic = 0;
 690         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 691         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 692         if (result <= 0) {
 693                 if (!nbd_disconnected(config))
 694                         dev_err(disk_to_dev(nbd->disk),
 695                                 "Receive control failed (result %d)\n", result);
 696                 return ERR_PTR(result);
 697         }
 698
 699         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 700                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 701                                 (unsigned long)ntohl(reply.magic));
 702                 return ERR_PTR(-EPROTO);
 703         }
 704
 705         memcpy(&handle, reply.handle, sizeof(handle));
 706         tag = nbd_handle_to_tag(handle);
 707         hwq = blk_mq_unique_tag_to_hwq(tag);
 708         if (hwq < nbd->tag_set.nr_hw_queues)
 709                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 710                                        blk_mq_unique_tag_to_tag(tag));
 711         if (!req || !blk_mq_request_started(req)) {
 712                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 713                         tag, req);
 714                 return ERR_PTR(-ENOENT);
 715         }
 716         trace_nbd_header_received(req, handle);
 717         cmd = blk_mq_rq_to_pdu(req);
 718
 719         mutex_lock(&cmd->lock);
 720         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 721                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 722                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 723                 ret = -ENOENT;
 724                 goto out;
 725         }
 726         if (cmd->status != BLK_STS_OK) {
 727                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 728                         req);
 729                 ret = -ENOENT;
 730                 goto out;
 731         }
 732         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 733                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 734                         req);
 735                 ret = -ENOENT;
 736                 goto out;
 737         }
 738         if (ntohl(reply.error)) {
 739                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 740                         ntohl(reply.error));
 741                 cmd->status = BLK_STS_IOERR;
 742                 goto out;
 743         }
 744
 745         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 746         if (rq_data_dir(req) != WRITE) {
 747                 struct req_iterator iter;
 748                 struct bio_vec bvec;
 749
 750                 rq_for_each_segment(bvec, req, iter) {
 751                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 752                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 753                         if (result <= 0) {
 754                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 755                                         result);
 756                                 /*
 757                                  * If we've disconnected, we need to make sure we
 758                                  * complete this request, otherwise error out
 759                                  * and let the timeout stuff handle resubmitting
 760                                  * this request onto another connection.
 761                                  */
 762                                 if (nbd_disconnected(config)) {
 763                                         cmd->status = BLK_STS_IOERR;
 764                                         goto out;
 765                                 }
 766                                 ret = -EIO;
 767                                 goto out;
 768                         }
 769                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 770                                 req, bvec.bv_len);
 771                 }
 772         }
 773 out:
 774         trace_nbd_payload_received(req, handle);
 775         mutex_unlock(&cmd->lock);
 776         return ret ? ERR_PTR(ret) : cmd;
 777 }
 778
 779 static void recv_work(struct work_struct *work)
 780 {
 781         struct recv_thread_args *args = container_of(work,
 782                                                      struct recv_thread_args,
 783                                                      work);
 784         struct nbd_device *nbd = args->nbd;
 785         struct nbd_config *config = nbd->config;
 786         struct nbd_cmd *cmd;
 787         struct request *rq;
 788
 789         while (1) {
 790                 cmd = nbd_read_stat(nbd, args->index);
 791                 if (IS_ERR(cmd)) {
 792                         struct nbd_sock *nsock = config->socks[args->index];
 793
 794                         mutex_lock(&nsock->tx_lock);
 795                         nbd_mark_nsock_dead(nbd, nsock, 1);
 796                         mutex_unlock(&nsock->tx_lock);
 797                         break;
 798                 }
 799
 800                 rq = blk_mq_rq_from_pdu(cmd);
 801                 if (likely(!blk_should_fake_timeout(rq->q)))
 802                         blk_mq_complete_request(rq);
 803         }
 804         atomic_dec(&config->recv_threads);
 805         wake_up(&config->recv_wq);
 806         nbd_config_put(nbd);
 807         kfree(args);
 808 }
 809
 810 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 811 {
 812         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 813
 814         mutex_lock(&cmd->lock);
 815         cmd->status = BLK_STS_IOERR;
 816         mutex_unlock(&cmd->lock);
 817
 818         blk_mq_complete_request(req);
 819         return true;
 820 }
 821
 822 static void nbd_clear_que(struct nbd_device *nbd)
 823 {
 824         blk_mq_quiesce_queue(nbd->disk->queue);
 825         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 826         blk_mq_unquiesce_queue(nbd->disk->queue);
 827         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 828 }
 829
 830 static int find_fallback(struct nbd_device *nbd, int index)
 831 {
 832         struct nbd_config *config = nbd->config;
 833         int new_index = -1;
 834         struct nbd_sock *nsock = config->socks[index];
 835         int fallback = nsock->fallback_index;
 836
 837         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 838                 return new_index;
 839
 840         if (config->num_connections <= 1) {
 841                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 842                                     "Dead connection, failed to find a fallback\n");
 843                 return new_index;
 844         }
 845
 846         if (fallback >= 0 && fallback < config->num_connections &&
 847             !config->socks[fallback]->dead)
 848                 return fallback;
 849
 850         if (nsock->fallback_index < 0 ||
 851             nsock->fallback_index >= config->num_connections ||
 852             config->socks[nsock->fallback_index]->dead) {
 853                 int i;
 854                 for (i = 0; i < config->num_connections; i++) {
 855                         if (i == index)
 856                                 continue;
 857                         if (!config->socks[i]->dead) {
 858                                 new_index = i;
 859                                 break;
 860                         }
 861                 }
 862                 nsock->fallback_index = new_index;
 863                 if (new_index < 0) {
 864                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 865                                             "Dead connection, failed to find a fallback\n");
 866                         return new_index;
 867                 }
 868         }
 869         new_index = nsock->fallback_index;
 870         return new_index;
 871 }
 872
 873 static int wait_for_reconnect(struct nbd_device *nbd)
 874 {
 875         struct nbd_config *config = nbd->config;
 876         if (!config->dead_conn_timeout)
 877                 return 0;
 878         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 879                 return 0;
 880         return wait_event_timeout(config->conn_wait,
 881                                   atomic_read(&config->live_connections) > 0,
 882                                   config->dead_conn_timeout) > 0;
 883 }
 884
 885 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 886 {
 887         struct request *req = blk_mq_rq_from_pdu(cmd);
 888         struct nbd_device *nbd = cmd->nbd;
 889         struct nbd_config *config;
 890         struct nbd_sock *nsock;
 891         int ret;
 892
 893         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 894                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 895                                     "Socks array is empty\n");
 896                 blk_mq_start_request(req);
 897                 return -EINVAL;
 898         }
 899         config = nbd->config;
 900
 901         if (index >= config->num_connections) {
 902                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 903                                     "Attempted send on invalid socket\n");
 904                 nbd_config_put(nbd);
 905                 blk_mq_start_request(req);
 906                 return -EINVAL;
 907         }
 908         cmd->status = BLK_STS_OK;
 909 again:
 910         nsock = config->socks[index];
 911         mutex_lock(&nsock->tx_lock);
 912         if (nsock->dead) {
 913                 int old_index = index;
 914                 index = find_fallback(nbd, index);
 915                 mutex_unlock(&nsock->tx_lock);
 916                 if (index < 0) {
 917                         if (wait_for_reconnect(nbd)) {
 918                                 index = old_index;
 919                                 goto again;
 920                         }
 921                         /* All the sockets should already be down at this point,
 922                          * we just want to make sure that DISCONNECTED is set so
 923                          * any requests that come in that were queue'ed waiting
 924                          * for the reconnect timer don't trigger the timer again
 925                          * and instead just error out.
 926                          */
 927                         sock_shutdown(nbd);
 928                         nbd_config_put(nbd);
 929                         blk_mq_start_request(req);
 930                         return -EIO;
 931                 }
 932                 goto again;
 933         }
 934
 935         /* Handle the case that we have a pending request that was partially
 936          * transmitted that _has_ to be serviced first.  We need to call requeue
 937          * here so that it gets put _after_ the request that is already on the
 938          * dispatch list.
 939          */
 940         blk_mq_start_request(req);
 941         if (unlikely(nsock->pending && nsock->pending != req)) {
 942                 nbd_requeue_cmd(cmd);
 943                 ret = 0;
 944                 goto out;
 945         }
 946         /*
 947          * Some failures are related to the link going down, so anything that
 948          * returns EAGAIN can be retried on a different socket.
 949          */
 950         ret = nbd_send_cmd(nbd, cmd, index);
 951         if (ret == -EAGAIN) {
 952                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 953                                     "Request send failed, requeueing\n");
 954                 nbd_mark_nsock_dead(nbd, nsock, 1);
 955                 nbd_requeue_cmd(cmd);
 956                 ret = 0;
 957         }
 958 out:
 959         mutex_unlock(&nsock->tx_lock);
 960         nbd_config_put(nbd);
 961         return ret;
 962 }
 963
 964 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 965                         const struct blk_mq_queue_data *bd)
 966 {
 967         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 968         int ret;
 969
 970         /*
 971          * Since we look at the bio's to send the request over the network we
 972          * need to make sure the completion work doesn't mark this request done
 973          * before we are done doing our send.  This keeps us from dereferencing
 974          * freed data if we have particularly fast completions (ie we get the
 975          * completion before we exit sock_xmit on the last bvec) or in the case
 976          * that the server is misbehaving (or there was an error) before we're
 977          * done sending everything over the wire.
 978          */
 979         mutex_lock(&cmd->lock);
 980         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 981
 982         /* We can be called directly from the user space process, which means we
 983          * could possibly have signals pending so our sendmsg will fail.  In
 984          * this case we need to return that we are busy, otherwise error out as
 985          * appropriate.
 986          */
 987         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 988         if (ret < 0)
 989                 ret = BLK_STS_IOERR;
 990         else if (!ret)
 991                 ret = BLK_STS_OK;
 992         mutex_unlock(&cmd->lock);
 993
 994         return ret;
 995 }
 996
 997 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
 998                                      int *err)
 999 {
1000         struct socket *sock;
1001
1002         *err = 0;
1003         sock = sockfd_lookup(fd, err);
1004         if (!sock)
1005                 return NULL;
1006
1007         if (sock->ops->shutdown == sock_no_shutdown) {
1008                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1009                 *err = -EINVAL;
1010                 sockfd_put(sock);
1011                 return NULL;
1012         }
1013
1014         return sock;
1015 }
1016
1017 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1018                           bool netlink)
1019 {
1020         struct nbd_config *config = nbd->config;
1021         struct socket *sock;
1022         struct nbd_sock **socks;
1023         struct nbd_sock *nsock;
1024         int err;
1025
1026         sock = nbd_get_socket(nbd, arg, &err);
1027         if (!sock)
1028                 return err;
1029
1030         if (!netlink && !nbd->task_setup &&
1031             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1032                 nbd->task_setup = current;
1033
1034         if (!netlink &&
1035             (nbd->task_setup != current ||
1036              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1037                 dev_err(disk_to_dev(nbd->disk),
1038                         "Device being setup by another task");
1039                 err = -EBUSY;
1040                 goto put_socket;
1041         }
1042
1043         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1044         if (!nsock) {
1045                 err = -ENOMEM;
1046                 goto put_socket;
1047         }
1048
1049         socks = krealloc(config->socks, (config->num_connections + 1) *
1050                          sizeof(struct nbd_sock *), GFP_KERNEL);
1051         if (!socks) {
1052                 kfree(nsock);
1053                 err = -ENOMEM;
1054                 goto put_socket;
1055         }
1056
1057         config->socks = socks;
1058
1059         nsock->fallback_index = -1;
1060         nsock->dead = false;
1061         mutex_init(&nsock->tx_lock);
1062         nsock->sock = sock;
1063         nsock->pending = NULL;
1064         nsock->sent = 0;
1065         nsock->cookie = 0;
1066         socks[config->num_connections++] = nsock;
1067         atomic_inc(&config->live_connections);
1068
1069         return 0;
1070
1071 put_socket:
1072         sockfd_put(sock);
1073         return err;
1074 }
1075
1076 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1077 {
1078         struct nbd_config *config = nbd->config;
1079         struct socket *sock, *old;
1080         struct recv_thread_args *args;
1081         int i;
1082         int err;
1083
1084         sock = nbd_get_socket(nbd, arg, &err);
1085         if (!sock)
1086                 return err;
1087
1088         args = kzalloc(sizeof(*args), GFP_KERNEL);
1089         if (!args) {
1090                 sockfd_put(sock);
1091                 return -ENOMEM;
1092         }
1093
1094         for (i = 0; i < config->num_connections; i++) {
1095                 struct nbd_sock *nsock = config->socks[i];
1096
1097                 if (!nsock->dead)
1098                         continue;
1099
1100                 mutex_lock(&nsock->tx_lock);
1101                 if (!nsock->dead) {
1102                         mutex_unlock(&nsock->tx_lock);
1103                         continue;
1104                 }
1105                 sk_set_memalloc(sock->sk);
1106                 if (nbd->tag_set.timeout)
1107                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1108                 atomic_inc(&config->recv_threads);
1109                 refcount_inc(&nbd->config_refs);
1110                 old = nsock->sock;
1111                 nsock->fallback_index = -1;
1112                 nsock->sock = sock;
1113                 nsock->dead = false;
1114                 INIT_WORK(&args->work, recv_work);
1115                 args->index = i;
1116                 args->nbd = nbd;
1117                 nsock->cookie++;
1118                 mutex_unlock(&nsock->tx_lock);
1119                 sockfd_put(old);
1120
1121                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1122
1123                 /* We take the tx_mutex in an error path in the recv_work, so we
1124                  * need to queue_work outside of the tx_mutex.
1125                  */
1126                 queue_work(nbd->recv_workq, &args->work);
1127
1128                 atomic_inc(&config->live_connections);
1129                 wake_up(&config->conn_wait);
1130                 return 0;
1131         }
1132         sockfd_put(sock);
1133         kfree(args);
1134         return -ENOSPC;
1135 }
1136
1137 static void nbd_bdev_reset(struct block_device *bdev)
1138 {
1139         if (bdev->bd_openers > 1)
1140                 return;
1141         bd_set_size(bdev, 0);
1142 }
1143
1144 static void nbd_parse_flags(struct nbd_device *nbd)
1145 {
1146         struct nbd_config *config = nbd->config;
1147         if (config->flags & NBD_FLAG_READ_ONLY)
1148                 set_disk_ro(nbd->disk, true);
1149         else
1150                 set_disk_ro(nbd->disk, false);
1151         if (config->flags & NBD_FLAG_SEND_TRIM)
1152                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1153         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1154                 if (config->flags & NBD_FLAG_SEND_FUA)
1155                         blk_queue_write_cache(nbd->disk->queue, true, true);
1156                 else
1157                         blk_queue_write_cache(nbd->disk->queue, true, false);
1158         }
1159         else
1160                 blk_queue_write_cache(nbd->disk->queue, false, false);
1161 }
1162
1163 static void send_disconnects(struct nbd_device *nbd)
1164 {
1165         struct nbd_config *config = nbd->config;
1166         struct nbd_request request = {
1167                 .magic = htonl(NBD_REQUEST_MAGIC),
1168                 .type = htonl(NBD_CMD_DISC),
1169         };
1170         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1171         struct iov_iter from;
1172         int i, ret;
1173
1174         for (i = 0; i < config->num_connections; i++) {
1175                 struct nbd_sock *nsock = config->socks[i];
1176
1177                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1178                 mutex_lock(&nsock->tx_lock);
1179                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1180                 if (ret <= 0)
1181                         dev_err(disk_to_dev(nbd->disk),
1182                                 "Send disconnect failed %d\n", ret);
1183                 mutex_unlock(&nsock->tx_lock);
1184         }
1185 }
1186
1187 static int nbd_disconnect(struct nbd_device *nbd)
1188 {
1189         struct nbd_config *config = nbd->config;
1190
1191         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1192         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1193         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1194         send_disconnects(nbd);
1195         return 0;
1196 }
1197
1198 static void nbd_clear_sock(struct nbd_device *nbd)
1199 {
1200         sock_shutdown(nbd);
1201         nbd_clear_que(nbd);
1202         nbd->task_setup = NULL;
1203 }
1204
1205 static void nbd_config_put(struct nbd_device *nbd)
1206 {
1207         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1208                                         &nbd->config_lock)) {
1209                 struct nbd_config *config = nbd->config;
1210                 nbd_dev_dbg_close(nbd);
1211                 nbd_size_clear(nbd);
1212                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1213                                        &config->runtime_flags))
1214                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1215                 nbd->task_recv = NULL;
1216                 nbd_clear_sock(nbd);
1217                 if (config->num_connections) {
1218                         int i;
1219                         for (i = 0; i < config->num_connections; i++) {
1220                                 sockfd_put(config->socks[i]->sock);
1221                                 kfree(config->socks[i]);
1222                         }
1223                         kfree(config->socks);
1224                 }
1225                 kfree(nbd->config);
1226                 nbd->config = NULL;
1227
1228                 if (nbd->recv_workq)
1229                         destroy_workqueue(nbd->recv_workq);
1230                 nbd->recv_workq = NULL;
1231
1232                 nbd->tag_set.timeout = 0;
1233                 nbd->disk->queue->limits.discard_granularity = 0;
1234                 nbd->disk->queue->limits.discard_alignment = 0;
1235                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1236                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1237
1238                 mutex_unlock(&nbd->config_lock);
1239                 nbd_put(nbd);
1240                 module_put(THIS_MODULE);
1241         }
1242 }
1243
1244 static int nbd_start_device(struct nbd_device *nbd)
1245 {
1246         struct nbd_config *config = nbd->config;
1247         int num_connections = config->num_connections;
1248         int error = 0, i;
1249
1250         if (nbd->task_recv)
1251                 return -EBUSY;
1252         if (!config->socks)
1253                 return -EINVAL;
1254         if (num_connections > 1 &&
1255             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1256                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1257                 return -EINVAL;
1258         }
1259
1260         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1261                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1262                                           WQ_UNBOUND, 0, nbd->index);
1263         if (!nbd->recv_workq) {
1264                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1265                 return -ENOMEM;
1266         }
1267
1268         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1269         nbd->task_recv = current;
1270
1271         nbd_parse_flags(nbd);
1272
1273         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1274         if (error) {
1275                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1276                 return error;
1277         }
1278         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1279
1280         nbd_dev_dbg_init(nbd);
1281         for (i = 0; i < num_connections; i++) {
1282                 struct recv_thread_args *args;
1283
1284                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1285                 if (!args) {
1286                         sock_shutdown(nbd);
1287                         /*
1288                          * If num_connections is m (2 < m),
1289                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1290                          * But NO.(n + 1) failed. We still have n recv threads.
1291                          * So, add flush_workqueue here to prevent recv threads
1292                          * dropping the last config_refs and trying to destroy
1293                          * the workqueue from inside the workqueue.
1294                          */
1295                         if (i)
1296                                 flush_workqueue(nbd->recv_workq);
1297                         return -ENOMEM;
1298                 }
1299                 sk_set_memalloc(config->socks[i]->sock->sk);
1300                 if (nbd->tag_set.timeout)
1301                         config->socks[i]->sock->sk->sk_sndtimeo =
1302                                 nbd->tag_set.timeout;
1303                 atomic_inc(&config->recv_threads);
1304                 refcount_inc(&nbd->config_refs);
1305                 INIT_WORK(&args->work, recv_work);
1306                 args->nbd = nbd;
1307                 args->index = i;
1308                 queue_work(nbd->recv_workq, &args->work);
1309         }
1310         nbd_size_update(nbd);
1311         return error;
1312 }
1313
1314 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1315 {
1316         struct nbd_config *config = nbd->config;
1317         int ret;
1318
1319         ret = nbd_start_device(nbd);
1320         if (ret)
1321                 return ret;
1322
1323         if (max_part)
1324                 bdev->bd_invalidated = 1;
1325         mutex_unlock(&nbd->config_lock);
1326         ret = wait_event_interruptible(config->recv_wq,
1327                                          atomic_read(&config->recv_threads) == 0);
1328         if (ret)
1329                 sock_shutdown(nbd);
1330         flush_workqueue(nbd->recv_workq);
1331
1332         mutex_lock(&nbd->config_lock);
1333         nbd_bdev_reset(bdev);
1334         /* user requested, ignore socket errors */
1335         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1336                 ret = 0;
1337         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1338                 ret = -ETIMEDOUT;
1339         return ret;
1340 }
1341
1342 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1343                                  struct block_device *bdev)
1344 {
1345         sock_shutdown(nbd);
1346         __invalidate_device(bdev, true);
1347         nbd_bdev_reset(bdev);
1348         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1349                                &nbd->config->runtime_flags))
1350                 nbd_config_put(nbd);
1351 }
1352
1353 static bool nbd_is_valid_blksize(unsigned long blksize)
1354 {
1355         if (!blksize || !is_power_of_2(blksize) || blksize < 512 ||
1356             blksize > PAGE_SIZE)
1357                 return false;
1358         return true;
1359 }
1360
1361 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1362 {
1363         nbd->tag_set.timeout = timeout * HZ;
1364         if (timeout)
1365                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1366 }
1367
1368 /* Must be called with config_lock held */
1369 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1370                        unsigned int cmd, unsigned long arg)
1371 {
1372         struct nbd_config *config = nbd->config;
1373
1374         switch (cmd) {
1375         case NBD_DISCONNECT:
1376                 return nbd_disconnect(nbd);
1377         case NBD_CLEAR_SOCK:
1378                 nbd_clear_sock_ioctl(nbd, bdev);
1379                 return 0;
1380         case NBD_SET_SOCK:
1381                 return nbd_add_socket(nbd, arg, false);
1382         case NBD_SET_BLKSIZE:
1383                 if (!arg)
1384                         arg = NBD_DEF_BLKSIZE;
1385                 if (!nbd_is_valid_blksize(arg))
1386                         return -EINVAL;
1387                 nbd_size_set(nbd, arg,
1388                              div_s64(config->bytesize, arg));
1389                 return 0;
1390         case NBD_SET_SIZE:
1391                 nbd_size_set(nbd, config->blksize,
1392                              div_s64(arg, config->blksize));
1393                 return 0;
1394         case NBD_SET_SIZE_BLOCKS:
1395                 nbd_size_set(nbd, config->blksize, arg);
1396                 return 0;
1397         case NBD_SET_TIMEOUT:
1398                 nbd_set_cmd_timeout(nbd, arg);
1399                 return 0;
1400
1401         case NBD_SET_FLAGS:
1402                 config->flags = arg;
1403                 return 0;
1404         case NBD_DO_IT:
1405                 return nbd_start_device_ioctl(nbd, bdev);
1406         case NBD_CLEAR_QUE:
1407                 /*
1408                  * This is for compatibility only.  The queue is always cleared
1409                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1410                  */
1411                 return 0;
1412         case NBD_PRINT_DEBUG:
1413                 /*
1414                  * For compatibility only, we no longer keep a list of
1415                  * outstanding requests.
1416                  */
1417                 return 0;
1418         }
1419         return -ENOTTY;
1420 }
1421
1422 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1423                      unsigned int cmd, unsigned long arg)
1424 {
1425         struct nbd_device *nbd = bdev->bd_disk->private_data;
1426         struct nbd_config *config = nbd->config;
1427         int error = -EINVAL;
1428
1429         if (!capable(CAP_SYS_ADMIN))
1430                 return -EPERM;
1431
1432         /* The block layer will pass back some non-nbd ioctls in case we have
1433          * special handling for them, but we don't so just return an error.
1434          */
1435         if (_IOC_TYPE(cmd) != 0xab)
1436                 return -EINVAL;
1437
1438         mutex_lock(&nbd->config_lock);
1439
1440         /* Don't allow ioctl operations on a nbd device that was created with
1441          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1442          */
1443         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1444             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1445                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1446         else
1447                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1448         mutex_unlock(&nbd->config_lock);
1449         return error;
1450 }
1451
1452 static struct nbd_config *nbd_alloc_config(void)
1453 {
1454         struct nbd_config *config;
1455
1456         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1457         if (!config)
1458                 return NULL;
1459         atomic_set(&config->recv_threads, 0);
1460         init_waitqueue_head(&config->recv_wq);
1461         init_waitqueue_head(&config->conn_wait);
1462         config->blksize = NBD_DEF_BLKSIZE;
1463         atomic_set(&config->live_connections, 0);
1464         try_module_get(THIS_MODULE);
1465         return config;
1466 }
1467
1468 static int nbd_open(struct block_device *bdev, fmode_t mode)
1469 {
1470         struct nbd_device *nbd;
1471         int ret = 0;
1472
1473         mutex_lock(&nbd_index_mutex);
1474         nbd = bdev->bd_disk->private_data;
1475         if (!nbd) {
1476                 ret = -ENXIO;
1477                 goto out;
1478         }
1479         if (!refcount_inc_not_zero(&nbd->refs)) {
1480                 ret = -ENXIO;
1481                 goto out;
1482         }
1483         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1484                 struct nbd_config *config;
1485
1486                 mutex_lock(&nbd->config_lock);
1487                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1488                         mutex_unlock(&nbd->config_lock);
1489                         goto out;
1490                 }
1491                 config = nbd->config = nbd_alloc_config();
1492                 if (!config) {
1493                         ret = -ENOMEM;
1494                         mutex_unlock(&nbd->config_lock);
1495                         goto out;
1496                 }
1497                 refcount_set(&nbd->config_refs, 1);
1498                 refcount_inc(&nbd->refs);
1499                 mutex_unlock(&nbd->config_lock);
1500                 bdev->bd_invalidated = 1;
1501         } else if (nbd_disconnected(nbd->config)) {
1502                 bdev->bd_invalidated = 1;
1503         }
1504 out:
1505         mutex_unlock(&nbd_index_mutex);
1506         return ret;
1507 }
1508
1509 static void nbd_release(struct gendisk *disk, fmode_t mode)
1510 {
1511         struct nbd_device *nbd = disk->private_data;
1512         struct block_device *bdev = bdget_disk(disk, 0);
1513
1514         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1515                         bdev->bd_openers == 0)
1516                 nbd_disconnect_and_put(nbd);
1517
1518         nbd_config_put(nbd);
1519         nbd_put(nbd);
1520 }
1521
1522 static const struct block_device_operations nbd_fops =
1523 {
1524         .owner =        THIS_MODULE,
1525         .open =         nbd_open,
1526         .release =      nbd_release,
1527         .ioctl =        nbd_ioctl,
1528         .compat_ioctl = nbd_ioctl,
1529 };
1530
1531 #if IS_ENABLED(CONFIG_DEBUG_FS)
1532
1533 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1534 {
1535         struct nbd_device *nbd = s->private;
1536
1537         if (nbd->task_recv)
1538                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1539
1540         return 0;
1541 }
1542
1543 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1544 {
1545         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1546 }
1547
1548 static const struct file_operations nbd_dbg_tasks_ops = {
1549         .open = nbd_dbg_tasks_open,
1550         .read = seq_read,
1551         .llseek = seq_lseek,
1552         .release = single_release,
1553 };
1554
1555 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1556 {
1557         struct nbd_device *nbd = s->private;
1558         u32 flags = nbd->config->flags;
1559
1560         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1561
1562         seq_puts(s, "Known flags:\n");
1563
1564         if (flags & NBD_FLAG_HAS_FLAGS)
1565                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1566         if (flags & NBD_FLAG_READ_ONLY)
1567                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1568         if (flags & NBD_FLAG_SEND_FLUSH)
1569                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1570         if (flags & NBD_FLAG_SEND_FUA)
1571                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1572         if (flags & NBD_FLAG_SEND_TRIM)
1573                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1574
1575         return 0;
1576 }
1577
1578 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1579 {
1580         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1581 }
1582
1583 static const struct file_operations nbd_dbg_flags_ops = {
1584         .open = nbd_dbg_flags_open,
1585         .read = seq_read,
1586         .llseek = seq_lseek,
1587         .release = single_release,
1588 };
1589
1590 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1591 {
1592         struct dentry *dir;
1593         struct nbd_config *config = nbd->config;
1594
1595         if (!nbd_dbg_dir)
1596                 return -EIO;
1597
1598         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1599         if (!dir) {
1600                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1601                         nbd_name(nbd));
1602                 return -EIO;
1603         }
1604         config->dbg_dir = dir;
1605
1606         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1607         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1608         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1609         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1610         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1611
1612         return 0;
1613 }
1614
1615 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1616 {
1617         debugfs_remove_recursive(nbd->config->dbg_dir);
1618 }
1619
1620 static int nbd_dbg_init(void)
1621 {
1622         struct dentry *dbg_dir;
1623
1624         dbg_dir = debugfs_create_dir("nbd", NULL);
1625         if (!dbg_dir)
1626                 return -EIO;
1627
1628         nbd_dbg_dir = dbg_dir;
1629
1630         return 0;
1631 }
1632
1633 static void nbd_dbg_close(void)
1634 {
1635         debugfs_remove_recursive(nbd_dbg_dir);
1636 }
1637
1638 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1639
1640 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1641 {
1642         return 0;
1643 }
1644
1645 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1646 {
1647 }
1648
1649 static int nbd_dbg_init(void)
1650 {
1651         return 0;
1652 }
1653
1654 static void nbd_dbg_close(void)
1655 {
1656 }
1657
1658 #endif
1659
1660 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1661                             unsigned int hctx_idx, unsigned int numa_node)
1662 {
1663         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1664         cmd->nbd = set->driver_data;
1665         cmd->flags = 0;
1666         mutex_init(&cmd->lock);
1667         return 0;
1668 }
1669
1670 static const struct blk_mq_ops nbd_mq_ops = {
1671         .queue_rq       = nbd_queue_rq,
1672         .complete       = nbd_complete_rq,
1673         .init_request   = nbd_init_request,
1674         .timeout        = nbd_xmit_timeout,
1675 };
1676
1677 static int nbd_dev_add(int index)
1678 {
1679         struct nbd_device *nbd;
1680         struct gendisk *disk;
1681         struct request_queue *q;
1682         int err = -ENOMEM;
1683
1684         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1685         if (!nbd)
1686                 goto out;
1687
1688         disk = alloc_disk(1 << part_shift);
1689         if (!disk)
1690                 goto out_free_nbd;
1691
1692         if (index >= 0) {
1693                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1694                                 GFP_KERNEL);
1695                 if (err == -ENOSPC)
1696                         err = -EEXIST;
1697         } else {
1698                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1699                 if (err >= 0)
1700                         index = err;
1701         }
1702         if (err < 0)
1703                 goto out_free_disk;
1704
1705         nbd->index = index;
1706         nbd->disk = disk;
1707         nbd->tag_set.ops = &nbd_mq_ops;
1708         nbd->tag_set.nr_hw_queues = 1;
1709         nbd->tag_set.queue_depth = 128;
1710         nbd->tag_set.numa_node = NUMA_NO_NODE;
1711         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1712         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1713                 BLK_MQ_F_BLOCKING;
1714         nbd->tag_set.driver_data = nbd;
1715         nbd->destroy_complete = NULL;
1716
1717         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1718         if (err)
1719                 goto out_free_idr;
1720
1721         q = blk_mq_init_queue(&nbd->tag_set);
1722         if (IS_ERR(q)) {
1723                 err = PTR_ERR(q);
1724                 goto out_free_tags;
1725         }
1726         disk->queue = q;
1727
1728         /*
1729          * Tell the block layer that we are not a rotational device
1730          */
1731         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1732         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1733         disk->queue->limits.discard_granularity = 0;
1734         disk->queue->limits.discard_alignment = 0;
1735         blk_queue_max_discard_sectors(disk->queue, 0);
1736         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1737         blk_queue_max_segments(disk->queue, USHRT_MAX);
1738         blk_queue_max_hw_sectors(disk->queue, 65536);
1739         disk->queue->limits.max_sectors = 256;
1740
1741         mutex_init(&nbd->config_lock);
1742         refcount_set(&nbd->config_refs, 0);
1743         refcount_set(&nbd->refs, 1);
1744         INIT_LIST_HEAD(&nbd->list);
1745         disk->major = NBD_MAJOR;
1746         disk->first_minor = index << part_shift;
1747         disk->fops = &nbd_fops;
1748         disk->private_data = nbd;
1749         sprintf(disk->disk_name, "nbd%d", index);
1750         add_disk(disk);
1751         nbd_total_devices++;
1752         return index;
1753
1754 out_free_tags:
1755         blk_mq_free_tag_set(&nbd->tag_set);
1756 out_free_idr:
1757         idr_remove(&nbd_index_idr, index);
1758 out_free_disk:
1759         put_disk(disk);
1760 out_free_nbd:
1761         kfree(nbd);
1762 out:
1763         return err;
1764 }
1765
1766 static int find_free_cb(int id, void *ptr, void *data)
1767 {
1768         struct nbd_device *nbd = ptr;
1769         struct nbd_device **found = data;
1770
1771         if (!refcount_read(&nbd->config_refs)) {
1772                 *found = nbd;
1773                 return 1;
1774         }
1775         return 0;
1776 }
1777
1778 /* Netlink interface. */
1779 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1780         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1781         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1782         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1783         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1784         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1785         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1786         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1787         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1788         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1789 };
1790
1791 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1792         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1793 };
1794
1795 /* We don't use this right now since we don't parse the incoming list, but we
1796  * still want it here so userspace knows what to expect.
1797  */
1798 static const struct nla_policy __attribute__((unused))
1799 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1800         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1801         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1802 };
1803
1804 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1805 {
1806         struct nbd_config *config = nbd->config;
1807         u64 bsize = config->blksize;
1808         u64 bytes = config->bytesize;
1809
1810         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1811                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1812
1813         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1814                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1815                 if (!bsize)
1816                         bsize = NBD_DEF_BLKSIZE;
1817                 if (!nbd_is_valid_blksize(bsize)) {
1818                         printk(KERN_ERR "Invalid block size %llu\n", bsize);
1819                         return -EINVAL;
1820                 }
1821         }
1822
1823         if (bytes != config->bytesize || bsize != config->blksize)
1824                 nbd_size_set(nbd, bsize, div64_u64(bytes, bsize));
1825         return 0;
1826 }
1827
1828 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1829 {
1830         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1831         struct nbd_device *nbd = NULL;
1832         struct nbd_config *config;
1833         int index = -1;
1834         int ret;
1835         bool put_dev = false;
1836
1837         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1838                 return -EPERM;
1839
1840         if (info->attrs[NBD_ATTR_INDEX])
1841                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1842         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1843                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1844                 return -EINVAL;
1845         }
1846         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1847                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1848                 return -EINVAL;
1849         }
1850 again:
1851         mutex_lock(&nbd_index_mutex);
1852         if (index == -1) {
1853                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1854                 if (ret == 0) {
1855                         int new_index;
1856                         new_index = nbd_dev_add(-1);
1857                         if (new_index < 0) {
1858                                 mutex_unlock(&nbd_index_mutex);
1859                                 printk(KERN_ERR "nbd: failed to add new device\n");
1860                                 return new_index;
1861                         }
1862                         nbd = idr_find(&nbd_index_idr, new_index);
1863                 }
1864         } else {
1865                 nbd = idr_find(&nbd_index_idr, index);
1866                 if (!nbd) {
1867                         ret = nbd_dev_add(index);
1868                         if (ret < 0) {
1869                                 mutex_unlock(&nbd_index_mutex);
1870                                 printk(KERN_ERR "nbd: failed to add new device\n");
1871                                 return ret;
1872                         }
1873                         nbd = idr_find(&nbd_index_idr, index);
1874                 }
1875         }
1876         if (!nbd) {
1877                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1878                        index);
1879                 mutex_unlock(&nbd_index_mutex);
1880                 return -EINVAL;
1881         }
1882
1883         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1884             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1885                 nbd->destroy_complete = &destroy_complete;
1886                 mutex_unlock(&nbd_index_mutex);
1887
1888                 /* Wait untill the the nbd stuff is totally destroyed */
1889                 wait_for_completion(&destroy_complete);
1890                 goto again;
1891         }
1892
1893         if (!refcount_inc_not_zero(&nbd->refs)) {
1894                 mutex_unlock(&nbd_index_mutex);
1895                 if (index == -1)
1896                         goto again;
1897                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1898                        index);
1899                 return -EINVAL;
1900         }
1901         mutex_unlock(&nbd_index_mutex);
1902
1903         mutex_lock(&nbd->config_lock);
1904         if (refcount_read(&nbd->config_refs)) {
1905                 mutex_unlock(&nbd->config_lock);
1906                 nbd_put(nbd);
1907                 if (index == -1)
1908                         goto again;
1909                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1910                 return -EBUSY;
1911         }
1912         if (WARN_ON(nbd->config)) {
1913                 mutex_unlock(&nbd->config_lock);
1914                 nbd_put(nbd);
1915                 return -EINVAL;
1916         }
1917         config = nbd->config = nbd_alloc_config();
1918         if (!nbd->config) {
1919                 mutex_unlock(&nbd->config_lock);
1920                 nbd_put(nbd);
1921                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1922                 return -ENOMEM;
1923         }
1924         refcount_set(&nbd->config_refs, 1);
1925         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1926
1927         ret = nbd_genl_size_set(info, nbd);
1928         if (ret)
1929                 goto out;
1930
1931         if (info->attrs[NBD_ATTR_TIMEOUT])
1932                 nbd_set_cmd_timeout(nbd,
1933                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1934         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1935                 config->dead_conn_timeout =
1936                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1937                 config->dead_conn_timeout *= HZ;
1938         }
1939         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1940                 config->flags =
1941                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1942         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1943                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1944                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1945                         set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
1946                                 &config->runtime_flags);
1947                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1948                         put_dev = true;
1949                 } else {
1950                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1951                 }
1952                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1953                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1954                                 &config->runtime_flags);
1955                 }
1956         }
1957
1958         if (info->attrs[NBD_ATTR_SOCKETS]) {
1959                 struct nlattr *attr;
1960                 int rem, fd;
1961
1962                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1963                                     rem) {
1964                         struct nlattr *socks[NBD_SOCK_MAX+1];
1965
1966                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1967                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1968                                 ret = -EINVAL;
1969                                 goto out;
1970                         }
1971                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1972                                                           attr,
1973                                                           nbd_sock_policy,
1974                                                           info->extack);
1975                         if (ret != 0) {
1976                                 printk(KERN_ERR "nbd: error processing sock list\n");
1977                                 ret = -EINVAL;
1978                                 goto out;
1979                         }
1980                         if (!socks[NBD_SOCK_FD])
1981                                 continue;
1982                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1983                         ret = nbd_add_socket(nbd, fd, true);
1984                         if (ret)
1985                                 goto out;
1986                 }
1987         }
1988         ret = nbd_start_device(nbd);
1989 out:
1990         mutex_unlock(&nbd->config_lock);
1991         if (!ret) {
1992                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
1993                 refcount_inc(&nbd->config_refs);
1994                 nbd_connect_reply(info, nbd->index);
1995         }
1996         nbd_config_put(nbd);
1997         if (put_dev)
1998                 nbd_put(nbd);
1999         return ret;
2000 }
2001
2002 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2003 {
2004         mutex_lock(&nbd->config_lock);
2005         nbd_disconnect(nbd);
2006         nbd_clear_sock(nbd);
2007         mutex_unlock(&nbd->config_lock);
2008         /*
2009          * Make sure recv thread has finished, so it does not drop the last
2010          * config ref and try to destroy the workqueue from inside the work
2011          * queue.
2012          */
2013         flush_workqueue(nbd->recv_workq);
2014         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2015                                &nbd->config->runtime_flags))
2016                 nbd_config_put(nbd);
2017 }
2018
2019 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2020 {
2021         struct nbd_device *nbd;
2022         int index;
2023
2024         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2025                 return -EPERM;
2026
2027         if (!info->attrs[NBD_ATTR_INDEX]) {
2028                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2029                 return -EINVAL;
2030         }
2031         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2032         mutex_lock(&nbd_index_mutex);
2033         nbd = idr_find(&nbd_index_idr, index);
2034         if (!nbd) {
2035                 mutex_unlock(&nbd_index_mutex);
2036                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2037                        index);
2038                 return -EINVAL;
2039         }
2040         if (!refcount_inc_not_zero(&nbd->refs)) {
2041                 mutex_unlock(&nbd_index_mutex);
2042                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2043                        index);
2044                 return -EINVAL;
2045         }
2046         mutex_unlock(&nbd_index_mutex);
2047         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2048                 nbd_put(nbd);
2049                 return 0;
2050         }
2051         nbd_disconnect_and_put(nbd);
2052         nbd_config_put(nbd);
2053         nbd_put(nbd);
2054         return 0;
2055 }
2056
2057 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2058 {
2059         struct nbd_device *nbd = NULL;
2060         struct nbd_config *config;
2061         int index;
2062         int ret = 0;
2063         bool put_dev = false;
2064
2065         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2066                 return -EPERM;
2067
2068         if (!info->attrs[NBD_ATTR_INDEX]) {
2069                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2070                 return -EINVAL;
2071         }
2072         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2073         mutex_lock(&nbd_index_mutex);
2074         nbd = idr_find(&nbd_index_idr, index);
2075         if (!nbd) {
2076                 mutex_unlock(&nbd_index_mutex);
2077                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2078                        index);
2079                 return -EINVAL;
2080         }
2081         if (!refcount_inc_not_zero(&nbd->refs)) {
2082                 mutex_unlock(&nbd_index_mutex);
2083                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2084                        index);
2085                 return -EINVAL;
2086         }
2087         mutex_unlock(&nbd_index_mutex);
2088
2089         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2090                 dev_err(nbd_to_dev(nbd),
2091                         "not configured, cannot reconfigure\n");
2092                 nbd_put(nbd);
2093                 return -EINVAL;
2094         }
2095
2096         mutex_lock(&nbd->config_lock);
2097         config = nbd->config;
2098         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2099             !nbd->task_recv) {
2100                 dev_err(nbd_to_dev(nbd),
2101                         "not configured, cannot reconfigure\n");
2102                 ret = -EINVAL;
2103                 goto out;
2104         }
2105
2106         ret = nbd_genl_size_set(info, nbd);
2107         if (ret)
2108                 goto out;
2109
2110         if (info->attrs[NBD_ATTR_TIMEOUT])
2111                 nbd_set_cmd_timeout(nbd,
2112                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2113         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2114                 config->dead_conn_timeout =
2115                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2116                 config->dead_conn_timeout *= HZ;
2117         }
2118         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2119                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2120                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2121                         if (!test_and_set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2122                                               &config->runtime_flags))
2123                                 put_dev = true;
2124                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2125                 } else {
2126                         if (test_and_clear_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2127                                                &config->runtime_flags))
2128                                 refcount_inc(&nbd->refs);
2129                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2130                 }
2131
2132                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2133                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2134                                         &config->runtime_flags);
2135                 } else {
2136                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2137                                         &config->runtime_flags);
2138                 }
2139         }
2140
2141         if (info->attrs[NBD_ATTR_SOCKETS]) {
2142                 struct nlattr *attr;
2143                 int rem, fd;
2144
2145                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2146                                     rem) {
2147                         struct nlattr *socks[NBD_SOCK_MAX+1];
2148
2149                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2150                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2151                                 ret = -EINVAL;
2152                                 goto out;
2153                         }
2154                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2155                                                           attr,
2156                                                           nbd_sock_policy,
2157                                                           info->extack);
2158                         if (ret != 0) {
2159                                 printk(KERN_ERR "nbd: error processing sock list\n");
2160                                 ret = -EINVAL;
2161                                 goto out;
2162                         }
2163                         if (!socks[NBD_SOCK_FD])
2164                                 continue;
2165                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2166                         ret = nbd_reconnect_socket(nbd, fd);
2167                         if (ret) {
2168                                 if (ret == -ENOSPC)
2169                                         ret = 0;
2170                                 goto out;
2171                         }
2172                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2173                 }
2174         }
2175 out:
2176         mutex_unlock(&nbd->config_lock);
2177         nbd_config_put(nbd);
2178         nbd_put(nbd);
2179         if (put_dev)
2180                 nbd_put(nbd);
2181         return ret;
2182 }
2183
2184 static const struct genl_ops nbd_connect_genl_ops[] = {
2185         {
2186                 .cmd    = NBD_CMD_CONNECT,
2187                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2188                 .doit   = nbd_genl_connect,
2189         },
2190         {
2191                 .cmd    = NBD_CMD_DISCONNECT,
2192                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2193                 .doit   = nbd_genl_disconnect,
2194         },
2195         {
2196                 .cmd    = NBD_CMD_RECONFIGURE,
2197                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2198                 .doit   = nbd_genl_reconfigure,
2199         },
2200         {
2201                 .cmd    = NBD_CMD_STATUS,
2202                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2203                 .doit   = nbd_genl_status,
2204         },
2205 };
2206
2207 static const struct genl_multicast_group nbd_mcast_grps[] = {
2208         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2209 };
2210
2211 static struct genl_family nbd_genl_family __ro_after_init = {
2212         .hdrsize        = 0,
2213         .name           = NBD_GENL_FAMILY_NAME,
2214         .version        = NBD_GENL_VERSION,
2215         .module         = THIS_MODULE,
2216         .ops            = nbd_connect_genl_ops,
2217         .n_ops          = ARRAY_SIZE(nbd_connect_genl_ops),
2218         .maxattr        = NBD_ATTR_MAX,
2219         .policy = nbd_attr_policy,
2220         .mcgrps         = nbd_mcast_grps,
2221         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2222 };
2223
2224 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2225 {
2226         struct nlattr *dev_opt;
2227         u8 connected = 0;
2228         int ret;
2229
2230         /* This is a little racey, but for status it's ok.  The
2231          * reason we don't take a ref here is because we can't
2232          * take a ref in the index == -1 case as we would need
2233          * to put under the nbd_index_mutex, which could
2234          * deadlock if we are configured to remove ourselves
2235          * once we're disconnected.
2236          */
2237         if (refcount_read(&nbd->config_refs))
2238                 connected = 1;
2239         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2240         if (!dev_opt)
2241                 return -EMSGSIZE;
2242         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2243         if (ret)
2244                 return -EMSGSIZE;
2245         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2246                          connected);
2247         if (ret)
2248                 return -EMSGSIZE;
2249         nla_nest_end(reply, dev_opt);
2250         return 0;
2251 }
2252
2253 static int status_cb(int id, void *ptr, void *data)
2254 {
2255         struct nbd_device *nbd = ptr;
2256         return populate_nbd_status(nbd, (struct sk_buff *)data);
2257 }
2258
2259 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2260 {
2261         struct nlattr *dev_list;
2262         struct sk_buff *reply;
2263         void *reply_head;
2264         size_t msg_size;
2265         int index = -1;
2266         int ret = -ENOMEM;
2267
2268         if (info->attrs[NBD_ATTR_INDEX])
2269                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2270
2271         mutex_lock(&nbd_index_mutex);
2272
2273         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2274                                   nla_attr_size(sizeof(u8)));
2275         msg_size *= (index == -1) ? nbd_total_devices : 1;
2276
2277         reply = genlmsg_new(msg_size, GFP_KERNEL);
2278         if (!reply)
2279                 goto out;
2280         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2281                                        NBD_CMD_STATUS);
2282         if (!reply_head) {
2283                 nlmsg_free(reply);
2284                 goto out;
2285         }
2286
2287         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2288         if (index == -1) {
2289                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2290                 if (ret) {
2291                         nlmsg_free(reply);
2292                         goto out;
2293                 }
2294         } else {
2295                 struct nbd_device *nbd;
2296                 nbd = idr_find(&nbd_index_idr, index);
2297                 if (nbd) {
2298                         ret = populate_nbd_status(nbd, reply);
2299                         if (ret) {
2300                                 nlmsg_free(reply);
2301                                 goto out;
2302                         }
2303                 }
2304         }
2305         nla_nest_end(reply, dev_list);
2306         genlmsg_end(reply, reply_head);
2307         ret = genlmsg_reply(reply, info);
2308 out:
2309         mutex_unlock(&nbd_index_mutex);
2310         return ret;
2311 }
2312
2313 static void nbd_connect_reply(struct genl_info *info, int index)
2314 {
2315         struct sk_buff *skb;
2316         void *msg_head;
2317         int ret;
2318
2319         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2320         if (!skb)
2321                 return;
2322         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2323                                      NBD_CMD_CONNECT);
2324         if (!msg_head) {
2325                 nlmsg_free(skb);
2326                 return;
2327         }
2328         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2329         if (ret) {
2330                 nlmsg_free(skb);
2331                 return;
2332         }
2333         genlmsg_end(skb, msg_head);
2334         genlmsg_reply(skb, info);
2335 }
2336
2337 static void nbd_mcast_index(int index)
2338 {
2339         struct sk_buff *skb;
2340         void *msg_head;
2341         int ret;
2342
2343         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2344         if (!skb)
2345                 return;
2346         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2347                                      NBD_CMD_LINK_DEAD);
2348         if (!msg_head) {
2349                 nlmsg_free(skb);
2350                 return;
2351         }
2352         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2353         if (ret) {
2354                 nlmsg_free(skb);
2355                 return;
2356         }
2357         genlmsg_end(skb, msg_head);
2358         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2359 }
2360
2361 static void nbd_dead_link_work(struct work_struct *work)
2362 {
2363         struct link_dead_args *args = container_of(work, struct link_dead_args,
2364                                                    work);
2365         nbd_mcast_index(args->index);
2366         kfree(args);
2367 }
2368
2369 static int __init nbd_init(void)
2370 {
2371         int i;
2372
2373         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2374
2375         if (max_part < 0) {
2376                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2377                 return -EINVAL;
2378         }
2379
2380         part_shift = 0;
2381         if (max_part > 0) {
2382                 part_shift = fls(max_part);
2383
2384                 /*
2385                  * Adjust max_part according to part_shift as it is exported
2386                  * to user space so that user can know the max number of
2387                  * partition kernel should be able to manage.
2388                  *
2389                  * Note that -1 is required because partition 0 is reserved
2390                  * for the whole disk.
2391                  */
2392                 max_part = (1UL << part_shift) - 1;
2393         }
2394
2395         if ((1UL << part_shift) > DISK_MAX_PARTS)
2396                 return -EINVAL;
2397
2398         if (nbds_max > 1UL << (MINORBITS - part_shift))
2399                 return -EINVAL;
2400
2401         if (register_blkdev(NBD_MAJOR, "nbd"))
2402                 return -EIO;
2403
2404         if (genl_register_family(&nbd_genl_family)) {
2405                 unregister_blkdev(NBD_MAJOR, "nbd");
2406                 return -EINVAL;
2407         }
2408         nbd_dbg_init();
2409
2410         mutex_lock(&nbd_index_mutex);
2411         for (i = 0; i < nbds_max; i++)
2412                 nbd_dev_add(i);
2413         mutex_unlock(&nbd_index_mutex);
2414         return 0;
2415 }
2416
2417 static int nbd_exit_cb(int id, void *ptr, void *data)
2418 {
2419         struct list_head *list = (struct list_head *)data;
2420         struct nbd_device *nbd = ptr;
2421
2422         list_add_tail(&nbd->list, list);
2423         return 0;
2424 }
2425
2426 static void __exit nbd_cleanup(void)
2427 {
2428         struct nbd_device *nbd;
2429         LIST_HEAD(del_list);
2430
2431         nbd_dbg_close();
2432
2433         mutex_lock(&nbd_index_mutex);
2434         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2435         mutex_unlock(&nbd_index_mutex);
2436
2437         while (!list_empty(&del_list)) {
2438                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2439                 list_del_init(&nbd->list);
2440                 if (refcount_read(&nbd->refs) != 1)
2441                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2442                 nbd_put(nbd);
2443         }
2444
2445         idr_destroy(&nbd_index_idr);
2446         genl_unregister_family(&nbd_genl_family);
2447         unregister_blkdev(NBD_MAJOR, "nbd");
2448 }
2449
2450 module_init(nbd_init);
2451 module_exit(nbd_cleanup);
2452
2453 MODULE_DESCRIPTION("Network Block Device");
2454 MODULE_LICENSE("GPL");
2455
2456 module_param(nbds_max, int, 0444);
2457 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2458 module_param(max_part, int, 0444);
2459 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");