drivers/block/nbd.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Network block device - make block devices work over TCP
   4  *
   5  * Note that you can not swap over this thing, yet. Seems to work but
   6  * deadlocks sometimes - you can not swap over TCP in general.
   7  *
   8  * Copyright 1997-2000, 2008 Pavel Machek <pavel@ucw.cz>
   9  * Parts copyright 2001 Steven Whitehouse <steve@chygwyn.com>
  10  *
  11  * (part of code stolen from loop.c)
  12  */
  13
  14 #include <linux/major.h>
  15
  16 #include <linux/blkdev.h>
  17 #include <linux/module.h>
  18 #include <linux/init.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/mm.h>
  21 #include <linux/fs.h>
  22 #include <linux/bio.h>
  23 #include <linux/stat.h>
  24 #include <linux/errno.h>
  25 #include <linux/file.h>
  26 #include <linux/ioctl.h>
  27 #include <linux/mutex.h>
  28 #include <linux/compiler.h>
  29 #include <linux/completion.h>
  30 #include <linux/err.h>
  31 #include <linux/kernel.h>
  32 #include <linux/slab.h>
  33 #include <net/sock.h>
  34 #include <linux/net.h>
  35 #include <linux/kthread.h>
  36 #include <linux/types.h>
  37 #include <linux/debugfs.h>
  38 #include <linux/blk-mq.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/types.h>
  42
  43 #include <linux/nbd.h>
  44 #include <linux/nbd-netlink.h>
  45 #include <net/genetlink.h>
  46
  47 #define CREATE_TRACE_POINTS
  48 #include <trace/events/nbd.h>
  49
  50 static DEFINE_IDR(nbd_index_idr);
  51 static DEFINE_MUTEX(nbd_index_mutex);
  52 static int nbd_total_devices = 0;
  53
  54 struct nbd_sock {
  55         struct socket *sock;
  56         struct mutex tx_lock;
  57         struct request *pending;
  58         int sent;
  59         bool dead;
  60         int fallback_index;
  61         int cookie;
  62 };
  63
  64 struct recv_thread_args {
  65         struct work_struct work;
  66         struct nbd_device *nbd;
  67         int index;
  68 };
  69
  70 struct link_dead_args {
  71         struct work_struct work;
  72         int index;
  73 };
  74
  75 #define NBD_RT_TIMEDOUT                 0
  76 #define NBD_RT_DISCONNECT_REQUESTED     1
  77 #define NBD_RT_DISCONNECTED             2
  78 #define NBD_RT_HAS_PID_FILE             3
  79 #define NBD_RT_HAS_CONFIG_REF           4
  80 #define NBD_RT_BOUND                    5
  81 #define NBD_RT_DESTROY_ON_DISCONNECT    6
  82 #define NBD_RT_DISCONNECT_ON_CLOSE      7
  83
  84 #define NBD_DESTROY_ON_DISCONNECT       0
  85 #define NBD_DISCONNECT_REQUESTED        1
  86
  87 struct nbd_config {
  88         u32 flags;
  89         unsigned long runtime_flags;
  90         u64 dead_conn_timeout;
  91
  92         struct nbd_sock **socks;
  93         int num_connections;
  94         atomic_t live_connections;
  95         wait_queue_head_t conn_wait;
  96
  97         atomic_t recv_threads;
  98         wait_queue_head_t recv_wq;
  99         loff_t blksize;
 100         loff_t bytesize;
 101 #if IS_ENABLED(CONFIG_DEBUG_FS)
 102         struct dentry *dbg_dir;
 103 #endif
 104 };
 105
 106 struct nbd_device {
 107         struct blk_mq_tag_set tag_set;
 108
 109         int index;
 110         refcount_t config_refs;
 111         refcount_t refs;
 112         struct nbd_config *config;
 113         struct mutex config_lock;
 114         struct gendisk *disk;
 115         struct workqueue_struct *recv_workq;
 116
 117         struct list_head list;
 118         struct task_struct *task_recv;
 119         struct task_struct *task_setup;
 120
 121         struct completion *destroy_complete;
 122         unsigned long flags;
 123 };
 124
 125 #define NBD_CMD_REQUEUED        1
 126
 127 struct nbd_cmd {
 128         struct nbd_device *nbd;
 129         struct mutex lock;
 130         int index;
 131         int cookie;
 132         int retries;
 133         blk_status_t status;
 134         unsigned long flags;
 135         u32 cmd_cookie;
 136 };
 137
 138 #if IS_ENABLED(CONFIG_DEBUG_FS)
 139 static struct dentry *nbd_dbg_dir;
 140 #endif
 141
 142 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 143
 144 #define NBD_MAGIC 0x68797548
 145
 146 #define NBD_DEF_BLKSIZE 1024
 147
 148 static unsigned int nbds_max = 16;
 149 static int max_part = 16;
 150 static int part_shift;
 151
 152 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 153 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 154 static void nbd_config_put(struct nbd_device *nbd);
 155 static void nbd_connect_reply(struct genl_info *info, int index);
 156 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 157 static void nbd_dead_link_work(struct work_struct *work);
 158 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 159
 160 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 161 {
 162         return disk_to_dev(nbd->disk);
 163 }
 164
 165 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 166 {
 167         struct request *req = blk_mq_rq_from_pdu(cmd);
 168
 169         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 170                 blk_mq_requeue_request(req, true);
 171 }
 172
 173 #define NBD_COOKIE_BITS 32
 174
 175 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 176 {
 177         struct request *req = blk_mq_rq_from_pdu(cmd);
 178         u32 tag = blk_mq_unique_tag(req);
 179         u64 cookie = cmd->cmd_cookie;
 180
 181         return (cookie << NBD_COOKIE_BITS) | tag;
 182 }
 183
 184 static u32 nbd_handle_to_tag(u64 handle)
 185 {
 186         return (u32)handle;
 187 }
 188
 189 static u32 nbd_handle_to_cookie(u64 handle)
 190 {
 191         return (u32)(handle >> NBD_COOKIE_BITS);
 192 }
 193
 194 static const char *nbdcmd_to_ascii(int cmd)
 195 {
 196         switch (cmd) {
 197         case  NBD_CMD_READ: return "read";
 198         case NBD_CMD_WRITE: return "write";
 199         case  NBD_CMD_DISC: return "disconnect";
 200         case NBD_CMD_FLUSH: return "flush";
 201         case  NBD_CMD_TRIM: return "trim/discard";
 202         }
 203         return "invalid";
 204 }
 205
 206 static ssize_t pid_show(struct device *dev,
 207                         struct device_attribute *attr, char *buf)
 208 {
 209         struct gendisk *disk = dev_to_disk(dev);
 210         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 211
 212         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 213 }
 214
 215 static const struct device_attribute pid_attr = {
 216         .attr = { .name = "pid", .mode = 0444},
 217         .show = pid_show,
 218 };
 219
 220 static void nbd_dev_remove(struct nbd_device *nbd)
 221 {
 222         struct gendisk *disk = nbd->disk;
 223         struct request_queue *q;
 224
 225         if (disk) {
 226                 q = disk->queue;
 227                 del_gendisk(disk);
 228                 blk_cleanup_queue(q);
 229                 blk_mq_free_tag_set(&nbd->tag_set);
 230                 disk->private_data = NULL;
 231                 put_disk(disk);
 232         }
 233
 234         /*
 235          * Place this in the last just before the nbd is freed to
 236          * make sure that the disk and the related kobject are also
 237          * totally removed to avoid duplicate creation of the same
 238          * one.
 239          */
 240         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) && nbd->destroy_complete)
 241                 complete(nbd->destroy_complete);
 242
 243         kfree(nbd);
 244 }
 245
 246 static void nbd_put(struct nbd_device *nbd)
 247 {
 248         if (refcount_dec_and_mutex_lock(&nbd->refs,
 249                                         &nbd_index_mutex)) {
 250                 idr_remove(&nbd_index_idr, nbd->index);
 251                 nbd_dev_remove(nbd);
 252                 mutex_unlock(&nbd_index_mutex);
 253         }
 254 }
 255
 256 static int nbd_disconnected(struct nbd_config *config)
 257 {
 258         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 259                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 260 }
 261
 262 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 263                                 int notify)
 264 {
 265         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 266                 struct link_dead_args *args;
 267                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 268                 if (args) {
 269                         INIT_WORK(&args->work, nbd_dead_link_work);
 270                         args->index = nbd->index;
 271                         queue_work(system_wq, &args->work);
 272                 }
 273         }
 274         if (!nsock->dead) {
 275                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 276                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 277                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 278                                                &nbd->config->runtime_flags)) {
 279                                 set_bit(NBD_RT_DISCONNECTED,
 280                                         &nbd->config->runtime_flags);
 281                                 dev_info(nbd_to_dev(nbd),
 282                                         "Disconnected due to user request.\n");
 283                         }
 284                 }
 285         }
 286         nsock->dead = true;
 287         nsock->pending = NULL;
 288         nsock->sent = 0;
 289 }
 290
 291 static void nbd_size_clear(struct nbd_device *nbd)
 292 {
 293         if (nbd->config->bytesize) {
 294                 set_capacity(nbd->disk, 0);
 295                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 296         }
 297 }
 298
 299 static void nbd_size_update(struct nbd_device *nbd, bool start)
 300 {
 301         struct nbd_config *config = nbd->config;
 302         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 303         sector_t nr_sectors = config->bytesize >> 9;
 304
 305         if (config->flags & NBD_FLAG_SEND_TRIM) {
 306                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 307                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 308                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 309         }
 310         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 311         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 312         set_capacity(nbd->disk, nr_sectors);
 313         if (bdev) {
 314                 if (bdev->bd_disk) {
 315                         bd_set_nr_sectors(bdev, nr_sectors);
 316                         if (start)
 317                                 set_blocksize(bdev, config->blksize);
 318                 } else
 319                         set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
 320                 bdput(bdev);
 321         }
 322         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 323 }
 324
 325 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 326                          loff_t nr_blocks)
 327 {
 328         struct nbd_config *config = nbd->config;
 329         config->blksize = blocksize;
 330         config->bytesize = blocksize * nr_blocks;
 331         if (nbd->task_recv != NULL)
 332                 nbd_size_update(nbd, false);
 333 }
 334
 335 static void nbd_complete_rq(struct request *req)
 336 {
 337         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 338
 339         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 340                 cmd->status ? "failed" : "done");
 341
 342         blk_mq_end_request(req, cmd->status);
 343 }
 344
 345 /*
 346  * Forcibly shutdown the socket causing all listeners to error
 347  */
 348 static void sock_shutdown(struct nbd_device *nbd)
 349 {
 350         struct nbd_config *config = nbd->config;
 351         int i;
 352
 353         if (config->num_connections == 0)
 354                 return;
 355         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 356                 return;
 357
 358         for (i = 0; i < config->num_connections; i++) {
 359                 struct nbd_sock *nsock = config->socks[i];
 360                 mutex_lock(&nsock->tx_lock);
 361                 nbd_mark_nsock_dead(nbd, nsock, 0);
 362                 mutex_unlock(&nsock->tx_lock);
 363         }
 364         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 365 }
 366
 367 static u32 req_to_nbd_cmd_type(struct request *req)
 368 {
 369         switch (req_op(req)) {
 370         case REQ_OP_DISCARD:
 371                 return NBD_CMD_TRIM;
 372         case REQ_OP_FLUSH:
 373                 return NBD_CMD_FLUSH;
 374         case REQ_OP_WRITE:
 375                 return NBD_CMD_WRITE;
 376         case REQ_OP_READ:
 377                 return NBD_CMD_READ;
 378         default:
 379                 return U32_MAX;
 380         }
 381 }
 382
 383 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 384                                                  bool reserved)
 385 {
 386         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 387         struct nbd_device *nbd = cmd->nbd;
 388         struct nbd_config *config;
 389
 390         if (!mutex_trylock(&cmd->lock))
 391                 return BLK_EH_RESET_TIMER;
 392
 393         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 394                 cmd->status = BLK_STS_TIMEOUT;
 395                 mutex_unlock(&cmd->lock);
 396                 goto done;
 397         }
 398         config = nbd->config;
 399
 400         if (config->num_connections > 1 ||
 401             (config->num_connections == 1 && nbd->tag_set.timeout)) {
 402                 dev_err_ratelimited(nbd_to_dev(nbd),
 403                                     "Connection timed out, retrying (%d/%d alive)\n",
 404                                     atomic_read(&config->live_connections),
 405                                     config->num_connections);
 406                 /*
 407                  * Hooray we have more connections, requeue this IO, the submit
 408                  * path will put it on a real connection. Or if only one
 409                  * connection is configured, the submit path will wait util
 410                  * a new connection is reconfigured or util dead timeout.
 411                  */
 412                 if (config->socks) {
 413                         if (cmd->index < config->num_connections) {
 414                                 struct nbd_sock *nsock =
 415                                         config->socks[cmd->index];
 416                                 mutex_lock(&nsock->tx_lock);
 417                                 /* We can have multiple outstanding requests, so
 418                                  * we don't want to mark the nsock dead if we've
 419                                  * already reconnected with a new socket, so
 420                                  * only mark it dead if its the same socket we
 421                                  * were sent out on.
 422                                  */
 423                                 if (cmd->cookie == nsock->cookie)
 424                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 425                                 mutex_unlock(&nsock->tx_lock);
 426                         }
 427                         mutex_unlock(&cmd->lock);
 428                         nbd_requeue_cmd(cmd);
 429                         nbd_config_put(nbd);
 430                         return BLK_EH_DONE;
 431                 }
 432         }
 433
 434         if (!nbd->tag_set.timeout) {
 435                 /*
 436                  * Userspace sets timeout=0 to disable socket disconnection,
 437                  * so just warn and reset the timer.
 438                  */
 439                 struct nbd_sock *nsock = config->socks[cmd->index];
 440                 cmd->retries++;
 441                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 442                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 443                         (unsigned long long)blk_rq_pos(req) << 9,
 444                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 445
 446                 mutex_lock(&nsock->tx_lock);
 447                 if (cmd->cookie != nsock->cookie) {
 448                         nbd_requeue_cmd(cmd);
 449                         mutex_unlock(&nsock->tx_lock);
 450                         mutex_unlock(&cmd->lock);
 451                         nbd_config_put(nbd);
 452                         return BLK_EH_DONE;
 453                 }
 454                 mutex_unlock(&nsock->tx_lock);
 455                 mutex_unlock(&cmd->lock);
 456                 nbd_config_put(nbd);
 457                 return BLK_EH_RESET_TIMER;
 458         }
 459
 460         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 461         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 462         cmd->status = BLK_STS_IOERR;
 463         mutex_unlock(&cmd->lock);
 464         sock_shutdown(nbd);
 465         nbd_config_put(nbd);
 466 done:
 467         blk_mq_complete_request(req);
 468         return BLK_EH_DONE;
 469 }
 470
 471 /*
 472  *  Send or receive packet.
 473  */
 474 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 475                      struct iov_iter *iter, int msg_flags, int *sent)
 476 {
 477         struct nbd_config *config = nbd->config;
 478         struct socket *sock = config->socks[index]->sock;
 479         int result;
 480         struct msghdr msg;
 481         unsigned int noreclaim_flag;
 482
 483         if (unlikely(!sock)) {
 484                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 485                         "Attempted %s on closed socket in sock_xmit\n",
 486                         (send ? "send" : "recv"));
 487                 return -EINVAL;
 488         }
 489
 490         msg.msg_iter = *iter;
 491
 492         noreclaim_flag = memalloc_noreclaim_save();
 493         do {
 494                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 495                 msg.msg_name = NULL;
 496                 msg.msg_namelen = 0;
 497                 msg.msg_control = NULL;
 498                 msg.msg_controllen = 0;
 499                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 500
 501                 if (send)
 502                         result = sock_sendmsg(sock, &msg);
 503                 else
 504                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 505
 506                 if (result <= 0) {
 507                         if (result == 0)
 508                                 result = -EPIPE; /* short read */
 509                         break;
 510                 }
 511                 if (sent)
 512                         *sent += result;
 513         } while (msg_data_left(&msg));
 514
 515         memalloc_noreclaim_restore(noreclaim_flag);
 516
 517         return result;
 518 }
 519
 520 /*
 521  * Different settings for sk->sk_sndtimeo can result in different return values
 522  * if there is a signal pending when we enter sendmsg, because reasons?
 523  */
 524 static inline int was_interrupted(int result)
 525 {
 526         return result == -ERESTARTSYS || result == -EINTR;
 527 }
 528
 529 /* always call with the tx_lock held */
 530 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 531 {
 532         struct request *req = blk_mq_rq_from_pdu(cmd);
 533         struct nbd_config *config = nbd->config;
 534         struct nbd_sock *nsock = config->socks[index];
 535         int result;
 536         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 537         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 538         struct iov_iter from;
 539         unsigned long size = blk_rq_bytes(req);
 540         struct bio *bio;
 541         u64 handle;
 542         u32 type;
 543         u32 nbd_cmd_flags = 0;
 544         int sent = nsock->sent, skip = 0;
 545
 546         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 547
 548         type = req_to_nbd_cmd_type(req);
 549         if (type == U32_MAX)
 550                 return -EIO;
 551
 552         if (rq_data_dir(req) == WRITE &&
 553             (config->flags & NBD_FLAG_READ_ONLY)) {
 554                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 555                                     "Write on read-only\n");
 556                 return -EIO;
 557         }
 558
 559         if (req->cmd_flags & REQ_FUA)
 560                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 561
 562         /* We did a partial send previously, and we at least sent the whole
 563          * request struct, so just go and send the rest of the pages in the
 564          * request.
 565          */
 566         if (sent) {
 567                 if (sent >= sizeof(request)) {
 568                         skip = sent - sizeof(request);
 569
 570                         /* initialize handle for tracing purposes */
 571                         handle = nbd_cmd_handle(cmd);
 572
 573                         goto send_pages;
 574                 }
 575                 iov_iter_advance(&from, sent);
 576         } else {
 577                 cmd->cmd_cookie++;
 578         }
 579         cmd->index = index;
 580         cmd->cookie = nsock->cookie;
 581         cmd->retries = 0;
 582         request.type = htonl(type | nbd_cmd_flags);
 583         if (type != NBD_CMD_FLUSH) {
 584                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 585                 request.len = htonl(size);
 586         }
 587         handle = nbd_cmd_handle(cmd);
 588         memcpy(request.handle, &handle, sizeof(handle));
 589
 590         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 591
 592         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 593                 req, nbdcmd_to_ascii(type),
 594                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 595         result = sock_xmit(nbd, index, 1, &from,
 596                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 597         trace_nbd_header_sent(req, handle);
 598         if (result <= 0) {
 599                 if (was_interrupted(result)) {
 600                         /* If we havne't sent anything we can just return BUSY,
 601                          * however if we have sent something we need to make
 602                          * sure we only allow this req to be sent until we are
 603                          * completely done.
 604                          */
 605                         if (sent) {
 606                                 nsock->pending = req;
 607                                 nsock->sent = sent;
 608                         }
 609                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 610                         return BLK_STS_RESOURCE;
 611                 }
 612                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 613                         "Send control failed (result %d)\n", result);
 614                 return -EAGAIN;
 615         }
 616 send_pages:
 617         if (type != NBD_CMD_WRITE)
 618                 goto out;
 619
 620         bio = req->bio;
 621         while (bio) {
 622                 struct bio *next = bio->bi_next;
 623                 struct bvec_iter iter;
 624                 struct bio_vec bvec;
 625
 626                 bio_for_each_segment(bvec, bio, iter) {
 627                         bool is_last = !next && bio_iter_last(bvec, iter);
 628                         int flags = is_last ? 0 : MSG_MORE;
 629
 630                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 631                                 req, bvec.bv_len);
 632                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 633                         if (skip) {
 634                                 if (skip >= iov_iter_count(&from)) {
 635                                         skip -= iov_iter_count(&from);
 636                                         continue;
 637                                 }
 638                                 iov_iter_advance(&from, skip);
 639                                 skip = 0;
 640                         }
 641                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 642                         if (result <= 0) {
 643                                 if (was_interrupted(result)) {
 644                                         /* We've already sent the header, we
 645                                          * have no choice but to set pending and
 646                                          * return BUSY.
 647                                          */
 648                                         nsock->pending = req;
 649                                         nsock->sent = sent;
 650                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 651                                         return BLK_STS_RESOURCE;
 652                                 }
 653                                 dev_err(disk_to_dev(nbd->disk),
 654                                         "Send data failed (result %d)\n",
 655                                         result);
 656                                 return -EAGAIN;
 657                         }
 658                         /*
 659                          * The completion might already have come in,
 660                          * so break for the last one instead of letting
 661                          * the iterator do it. This prevents use-after-free
 662                          * of the bio.
 663                          */
 664                         if (is_last)
 665                                 break;
 666                 }
 667                 bio = next;
 668         }
 669 out:
 670         trace_nbd_payload_sent(req, handle);
 671         nsock->pending = NULL;
 672         nsock->sent = 0;
 673         return 0;
 674 }
 675
 676 /* NULL returned = something went wrong, inform userspace */
 677 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 678 {
 679         struct nbd_config *config = nbd->config;
 680         int result;
 681         struct nbd_reply reply;
 682         struct nbd_cmd *cmd;
 683         struct request *req = NULL;
 684         u64 handle;
 685         u16 hwq;
 686         u32 tag;
 687         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 688         struct iov_iter to;
 689         int ret = 0;
 690
 691         reply.magic = 0;
 692         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 693         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 694         if (result <= 0) {
 695                 if (!nbd_disconnected(config))
 696                         dev_err(disk_to_dev(nbd->disk),
 697                                 "Receive control failed (result %d)\n", result);
 698                 return ERR_PTR(result);
 699         }
 700
 701         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 702                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 703                                 (unsigned long)ntohl(reply.magic));
 704                 return ERR_PTR(-EPROTO);
 705         }
 706
 707         memcpy(&handle, reply.handle, sizeof(handle));
 708         tag = nbd_handle_to_tag(handle);
 709         hwq = blk_mq_unique_tag_to_hwq(tag);
 710         if (hwq < nbd->tag_set.nr_hw_queues)
 711                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 712                                        blk_mq_unique_tag_to_tag(tag));
 713         if (!req || !blk_mq_request_started(req)) {
 714                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 715                         tag, req);
 716                 return ERR_PTR(-ENOENT);
 717         }
 718         trace_nbd_header_received(req, handle);
 719         cmd = blk_mq_rq_to_pdu(req);
 720
 721         mutex_lock(&cmd->lock);
 722         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 723                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 724                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 725                 ret = -ENOENT;
 726                 goto out;
 727         }
 728         if (cmd->status != BLK_STS_OK) {
 729                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 730                         req);
 731                 ret = -ENOENT;
 732                 goto out;
 733         }
 734         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 735                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 736                         req);
 737                 ret = -ENOENT;
 738                 goto out;
 739         }
 740         if (ntohl(reply.error)) {
 741                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 742                         ntohl(reply.error));
 743                 cmd->status = BLK_STS_IOERR;
 744                 goto out;
 745         }
 746
 747         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 748         if (rq_data_dir(req) != WRITE) {
 749                 struct req_iterator iter;
 750                 struct bio_vec bvec;
 751
 752                 rq_for_each_segment(bvec, req, iter) {
 753                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 754                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 755                         if (result <= 0) {
 756                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 757                                         result);
 758                                 /*
 759                                  * If we've disconnected, we need to make sure we
 760                                  * complete this request, otherwise error out
 761                                  * and let the timeout stuff handle resubmitting
 762                                  * this request onto another connection.
 763                                  */
 764                                 if (nbd_disconnected(config)) {
 765                                         cmd->status = BLK_STS_IOERR;
 766                                         goto out;
 767                                 }
 768                                 ret = -EIO;
 769                                 goto out;
 770                         }
 771                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 772                                 req, bvec.bv_len);
 773                 }
 774         }
 775 out:
 776         trace_nbd_payload_received(req, handle);
 777         mutex_unlock(&cmd->lock);
 778         return ret ? ERR_PTR(ret) : cmd;
 779 }
 780
 781 static void recv_work(struct work_struct *work)
 782 {
 783         struct recv_thread_args *args = container_of(work,
 784                                                      struct recv_thread_args,
 785                                                      work);
 786         struct nbd_device *nbd = args->nbd;
 787         struct nbd_config *config = nbd->config;
 788         struct nbd_cmd *cmd;
 789         struct request *rq;
 790
 791         while (1) {
 792                 cmd = nbd_read_stat(nbd, args->index);
 793                 if (IS_ERR(cmd)) {
 794                         struct nbd_sock *nsock = config->socks[args->index];
 795
 796                         mutex_lock(&nsock->tx_lock);
 797                         nbd_mark_nsock_dead(nbd, nsock, 1);
 798                         mutex_unlock(&nsock->tx_lock);
 799                         break;
 800                 }
 801
 802                 rq = blk_mq_rq_from_pdu(cmd);
 803                 if (likely(!blk_should_fake_timeout(rq->q)))
 804                         blk_mq_complete_request(rq);
 805         }
 806         nbd_config_put(nbd);
 807         atomic_dec(&config->recv_threads);
 808         wake_up(&config->recv_wq);
 809         kfree(args);
 810 }
 811
 812 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 813 {
 814         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 815
 816         mutex_lock(&cmd->lock);
 817         cmd->status = BLK_STS_IOERR;
 818         mutex_unlock(&cmd->lock);
 819
 820         blk_mq_complete_request(req);
 821         return true;
 822 }
 823
 824 static void nbd_clear_que(struct nbd_device *nbd)
 825 {
 826         blk_mq_quiesce_queue(nbd->disk->queue);
 827         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 828         blk_mq_unquiesce_queue(nbd->disk->queue);
 829         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 830 }
 831
 832 static int find_fallback(struct nbd_device *nbd, int index)
 833 {
 834         struct nbd_config *config = nbd->config;
 835         int new_index = -1;
 836         struct nbd_sock *nsock = config->socks[index];
 837         int fallback = nsock->fallback_index;
 838
 839         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 840                 return new_index;
 841
 842         if (config->num_connections <= 1) {
 843                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 844                                     "Dead connection, failed to find a fallback\n");
 845                 return new_index;
 846         }
 847
 848         if (fallback >= 0 && fallback < config->num_connections &&
 849             !config->socks[fallback]->dead)
 850                 return fallback;
 851
 852         if (nsock->fallback_index < 0 ||
 853             nsock->fallback_index >= config->num_connections ||
 854             config->socks[nsock->fallback_index]->dead) {
 855                 int i;
 856                 for (i = 0; i < config->num_connections; i++) {
 857                         if (i == index)
 858                                 continue;
 859                         if (!config->socks[i]->dead) {
 860                                 new_index = i;
 861                                 break;
 862                         }
 863                 }
 864                 nsock->fallback_index = new_index;
 865                 if (new_index < 0) {
 866                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 867                                             "Dead connection, failed to find a fallback\n");
 868                         return new_index;
 869                 }
 870         }
 871         new_index = nsock->fallback_index;
 872         return new_index;
 873 }
 874
 875 static int wait_for_reconnect(struct nbd_device *nbd)
 876 {
 877         struct nbd_config *config = nbd->config;
 878         if (!config->dead_conn_timeout)
 879                 return 0;
 880         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 881                 return 0;
 882         return wait_event_timeout(config->conn_wait,
 883                                   atomic_read(&config->live_connections) > 0,
 884                                   config->dead_conn_timeout) > 0;
 885 }
 886
 887 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 888 {
 889         struct request *req = blk_mq_rq_from_pdu(cmd);
 890         struct nbd_device *nbd = cmd->nbd;
 891         struct nbd_config *config;
 892         struct nbd_sock *nsock;
 893         int ret;
 894
 895         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 896                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 897                                     "Socks array is empty\n");
 898                 blk_mq_start_request(req);
 899                 return -EINVAL;
 900         }
 901         config = nbd->config;
 902
 903         if (index >= config->num_connections) {
 904                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 905                                     "Attempted send on invalid socket\n");
 906                 nbd_config_put(nbd);
 907                 blk_mq_start_request(req);
 908                 return -EINVAL;
 909         }
 910         cmd->status = BLK_STS_OK;
 911 again:
 912         nsock = config->socks[index];
 913         mutex_lock(&nsock->tx_lock);
 914         if (nsock->dead) {
 915                 int old_index = index;
 916                 index = find_fallback(nbd, index);
 917                 mutex_unlock(&nsock->tx_lock);
 918                 if (index < 0) {
 919                         if (wait_for_reconnect(nbd)) {
 920                                 index = old_index;
 921                                 goto again;
 922                         }
 923                         /* All the sockets should already be down at this point,
 924                          * we just want to make sure that DISCONNECTED is set so
 925                          * any requests that come in that were queue'ed waiting
 926                          * for the reconnect timer don't trigger the timer again
 927                          * and instead just error out.
 928                          */
 929                         sock_shutdown(nbd);
 930                         nbd_config_put(nbd);
 931                         blk_mq_start_request(req);
 932                         return -EIO;
 933                 }
 934                 goto again;
 935         }
 936
 937         /* Handle the case that we have a pending request that was partially
 938          * transmitted that _has_ to be serviced first.  We need to call requeue
 939          * here so that it gets put _after_ the request that is already on the
 940          * dispatch list.
 941          */
 942         blk_mq_start_request(req);
 943         if (unlikely(nsock->pending && nsock->pending != req)) {
 944                 nbd_requeue_cmd(cmd);
 945                 ret = 0;
 946                 goto out;
 947         }
 948         /*
 949          * Some failures are related to the link going down, so anything that
 950          * returns EAGAIN can be retried on a different socket.
 951          */
 952         ret = nbd_send_cmd(nbd, cmd, index);
 953         if (ret == -EAGAIN) {
 954                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 955                                     "Request send failed, requeueing\n");
 956                 nbd_mark_nsock_dead(nbd, nsock, 1);
 957                 nbd_requeue_cmd(cmd);
 958                 ret = 0;
 959         }
 960 out:
 961         mutex_unlock(&nsock->tx_lock);
 962         nbd_config_put(nbd);
 963         return ret;
 964 }
 965
 966 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 967                         const struct blk_mq_queue_data *bd)
 968 {
 969         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 970         int ret;
 971
 972         /*
 973          * Since we look at the bio's to send the request over the network we
 974          * need to make sure the completion work doesn't mark this request done
 975          * before we are done doing our send.  This keeps us from dereferencing
 976          * freed data if we have particularly fast completions (ie we get the
 977          * completion before we exit sock_xmit on the last bvec) or in the case
 978          * that the server is misbehaving (or there was an error) before we're
 979          * done sending everything over the wire.
 980          */
 981         mutex_lock(&cmd->lock);
 982         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 983
 984         /* We can be called directly from the user space process, which means we
 985          * could possibly have signals pending so our sendmsg will fail.  In
 986          * this case we need to return that we are busy, otherwise error out as
 987          * appropriate.
 988          */
 989         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 990         if (ret < 0)
 991                 ret = BLK_STS_IOERR;
 992         else if (!ret)
 993                 ret = BLK_STS_OK;
 994         mutex_unlock(&cmd->lock);
 995
 996         return ret;
 997 }
 998
 999 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
1000                                      int *err)
1001 {
1002         struct socket *sock;
1003
1004         *err = 0;
1005         sock = sockfd_lookup(fd, err);
1006         if (!sock)
1007                 return NULL;
1008
1009         if (sock->ops->shutdown == sock_no_shutdown) {
1010                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1011                 *err = -EINVAL;
1012                 sockfd_put(sock);
1013                 return NULL;
1014         }
1015
1016         return sock;
1017 }
1018
1019 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1020                           bool netlink)
1021 {
1022         struct nbd_config *config = nbd->config;
1023         struct socket *sock;
1024         struct nbd_sock **socks;
1025         struct nbd_sock *nsock;
1026         int err;
1027
1028         sock = nbd_get_socket(nbd, arg, &err);
1029         if (!sock)
1030                 return err;
1031
1032         if (!netlink && !nbd->task_setup &&
1033             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1034                 nbd->task_setup = current;
1035
1036         if (!netlink &&
1037             (nbd->task_setup != current ||
1038              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1039                 dev_err(disk_to_dev(nbd->disk),
1040                         "Device being setup by another task");
1041                 err = -EBUSY;
1042                 goto put_socket;
1043         }
1044
1045         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1046         if (!nsock) {
1047                 err = -ENOMEM;
1048                 goto put_socket;
1049         }
1050
1051         socks = krealloc(config->socks, (config->num_connections + 1) *
1052                          sizeof(struct nbd_sock *), GFP_KERNEL);
1053         if (!socks) {
1054                 kfree(nsock);
1055                 err = -ENOMEM;
1056                 goto put_socket;
1057         }
1058
1059         config->socks = socks;
1060
1061         nsock->fallback_index = -1;
1062         nsock->dead = false;
1063         mutex_init(&nsock->tx_lock);
1064         nsock->sock = sock;
1065         nsock->pending = NULL;
1066         nsock->sent = 0;
1067         nsock->cookie = 0;
1068         socks[config->num_connections++] = nsock;
1069         atomic_inc(&config->live_connections);
1070
1071         return 0;
1072
1073 put_socket:
1074         sockfd_put(sock);
1075         return err;
1076 }
1077
1078 static int nbd_reconnect_socket(struct nbd_device *nbd, unsigned long arg)
1079 {
1080         struct nbd_config *config = nbd->config;
1081         struct socket *sock, *old;
1082         struct recv_thread_args *args;
1083         int i;
1084         int err;
1085
1086         sock = nbd_get_socket(nbd, arg, &err);
1087         if (!sock)
1088                 return err;
1089
1090         args = kzalloc(sizeof(*args), GFP_KERNEL);
1091         if (!args) {
1092                 sockfd_put(sock);
1093                 return -ENOMEM;
1094         }
1095
1096         for (i = 0; i < config->num_connections; i++) {
1097                 struct nbd_sock *nsock = config->socks[i];
1098
1099                 if (!nsock->dead)
1100                         continue;
1101
1102                 mutex_lock(&nsock->tx_lock);
1103                 if (!nsock->dead) {
1104                         mutex_unlock(&nsock->tx_lock);
1105                         continue;
1106                 }
1107                 sk_set_memalloc(sock->sk);
1108                 if (nbd->tag_set.timeout)
1109                         sock->sk->sk_sndtimeo = nbd->tag_set.timeout;
1110                 atomic_inc(&config->recv_threads);
1111                 refcount_inc(&nbd->config_refs);
1112                 old = nsock->sock;
1113                 nsock->fallback_index = -1;
1114                 nsock->sock = sock;
1115                 nsock->dead = false;
1116                 INIT_WORK(&args->work, recv_work);
1117                 args->index = i;
1118                 args->nbd = nbd;
1119                 nsock->cookie++;
1120                 mutex_unlock(&nsock->tx_lock);
1121                 sockfd_put(old);
1122
1123                 clear_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
1124
1125                 /* We take the tx_mutex in an error path in the recv_work, so we
1126                  * need to queue_work outside of the tx_mutex.
1127                  */
1128                 queue_work(nbd->recv_workq, &args->work);
1129
1130                 atomic_inc(&config->live_connections);
1131                 wake_up(&config->conn_wait);
1132                 return 0;
1133         }
1134         sockfd_put(sock);
1135         kfree(args);
1136         return -ENOSPC;
1137 }
1138
1139 static void nbd_bdev_reset(struct block_device *bdev)
1140 {
1141         if (bdev->bd_openers > 1)
1142                 return;
1143         bd_set_nr_sectors(bdev, 0);
1144 }
1145
1146 static void nbd_parse_flags(struct nbd_device *nbd)
1147 {
1148         struct nbd_config *config = nbd->config;
1149         if (config->flags & NBD_FLAG_READ_ONLY)
1150                 set_disk_ro(nbd->disk, true);
1151         else
1152                 set_disk_ro(nbd->disk, false);
1153         if (config->flags & NBD_FLAG_SEND_TRIM)
1154                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1155         if (config->flags & NBD_FLAG_SEND_FLUSH) {
1156                 if (config->flags & NBD_FLAG_SEND_FUA)
1157                         blk_queue_write_cache(nbd->disk->queue, true, true);
1158                 else
1159                         blk_queue_write_cache(nbd->disk->queue, true, false);
1160         }
1161         else
1162                 blk_queue_write_cache(nbd->disk->queue, false, false);
1163 }
1164
1165 static void send_disconnects(struct nbd_device *nbd)
1166 {
1167         struct nbd_config *config = nbd->config;
1168         struct nbd_request request = {
1169                 .magic = htonl(NBD_REQUEST_MAGIC),
1170                 .type = htonl(NBD_CMD_DISC),
1171         };
1172         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
1173         struct iov_iter from;
1174         int i, ret;
1175
1176         for (i = 0; i < config->num_connections; i++) {
1177                 struct nbd_sock *nsock = config->socks[i];
1178
1179                 iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
1180                 mutex_lock(&nsock->tx_lock);
1181                 ret = sock_xmit(nbd, i, 1, &from, 0, NULL);
1182                 if (ret <= 0)
1183                         dev_err(disk_to_dev(nbd->disk),
1184                                 "Send disconnect failed %d\n", ret);
1185                 mutex_unlock(&nsock->tx_lock);
1186         }
1187 }
1188
1189 static int nbd_disconnect(struct nbd_device *nbd)
1190 {
1191         struct nbd_config *config = nbd->config;
1192
1193         dev_info(disk_to_dev(nbd->disk), "NBD_DISCONNECT\n");
1194         set_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
1195         set_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags);
1196         send_disconnects(nbd);
1197         return 0;
1198 }
1199
1200 static void nbd_clear_sock(struct nbd_device *nbd)
1201 {
1202         sock_shutdown(nbd);
1203         nbd_clear_que(nbd);
1204         nbd->task_setup = NULL;
1205 }
1206
1207 static void nbd_config_put(struct nbd_device *nbd)
1208 {
1209         if (refcount_dec_and_mutex_lock(&nbd->config_refs,
1210                                         &nbd->config_lock)) {
1211                 struct nbd_config *config = nbd->config;
1212                 nbd_dev_dbg_close(nbd);
1213                 nbd_size_clear(nbd);
1214                 if (test_and_clear_bit(NBD_RT_HAS_PID_FILE,
1215                                        &config->runtime_flags))
1216                         device_remove_file(disk_to_dev(nbd->disk), &pid_attr);
1217                 nbd->task_recv = NULL;
1218                 nbd_clear_sock(nbd);
1219                 if (config->num_connections) {
1220                         int i;
1221                         for (i = 0; i < config->num_connections; i++) {
1222                                 sockfd_put(config->socks[i]->sock);
1223                                 kfree(config->socks[i]);
1224                         }
1225                         kfree(config->socks);
1226                 }
1227                 kfree(nbd->config);
1228                 nbd->config = NULL;
1229
1230                 if (nbd->recv_workq)
1231                         destroy_workqueue(nbd->recv_workq);
1232                 nbd->recv_workq = NULL;
1233
1234                 nbd->tag_set.timeout = 0;
1235                 nbd->disk->queue->limits.discard_granularity = 0;
1236                 nbd->disk->queue->limits.discard_alignment = 0;
1237                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
1238                 blk_queue_flag_clear(QUEUE_FLAG_DISCARD, nbd->disk->queue);
1239
1240                 mutex_unlock(&nbd->config_lock);
1241                 nbd_put(nbd);
1242                 module_put(THIS_MODULE);
1243         }
1244 }
1245
1246 static int nbd_start_device(struct nbd_device *nbd)
1247 {
1248         struct nbd_config *config = nbd->config;
1249         int num_connections = config->num_connections;
1250         int error = 0, i;
1251
1252         if (nbd->task_recv)
1253                 return -EBUSY;
1254         if (!config->socks)
1255                 return -EINVAL;
1256         if (num_connections > 1 &&
1257             !(config->flags & NBD_FLAG_CAN_MULTI_CONN)) {
1258                 dev_err(disk_to_dev(nbd->disk), "server does not support multiple connections per device.\n");
1259                 return -EINVAL;
1260         }
1261
1262         nbd->recv_workq = alloc_workqueue("knbd%d-recv",
1263                                           WQ_MEM_RECLAIM | WQ_HIGHPRI |
1264                                           WQ_UNBOUND, 0, nbd->index);
1265         if (!nbd->recv_workq) {
1266                 dev_err(disk_to_dev(nbd->disk), "Could not allocate knbd recv work queue.\n");
1267                 return -ENOMEM;
1268         }
1269
1270         blk_mq_update_nr_hw_queues(&nbd->tag_set, config->num_connections);
1271         nbd->task_recv = current;
1272
1273         nbd_parse_flags(nbd);
1274
1275         error = device_create_file(disk_to_dev(nbd->disk), &pid_attr);
1276         if (error) {
1277                 dev_err(disk_to_dev(nbd->disk), "device_create_file failed!\n");
1278                 return error;
1279         }
1280         set_bit(NBD_RT_HAS_PID_FILE, &config->runtime_flags);
1281
1282         nbd_dev_dbg_init(nbd);
1283         for (i = 0; i < num_connections; i++) {
1284                 struct recv_thread_args *args;
1285
1286                 args = kzalloc(sizeof(*args), GFP_KERNEL);
1287                 if (!args) {
1288                         sock_shutdown(nbd);
1289                         /*
1290                          * If num_connections is m (2 < m),
1291                          * and NO.1 ~ NO.n(1 < n < m) kzallocs are successful.
1292                          * But NO.(n + 1) failed. We still have n recv threads.
1293                          * So, add flush_workqueue here to prevent recv threads
1294                          * dropping the last config_refs and trying to destroy
1295                          * the workqueue from inside the workqueue.
1296                          */
1297                         if (i)
1298                                 flush_workqueue(nbd->recv_workq);
1299                         return -ENOMEM;
1300                 }
1301                 sk_set_memalloc(config->socks[i]->sock->sk);
1302                 if (nbd->tag_set.timeout)
1303                         config->socks[i]->sock->sk->sk_sndtimeo =
1304                                 nbd->tag_set.timeout;
1305                 atomic_inc(&config->recv_threads);
1306                 refcount_inc(&nbd->config_refs);
1307                 INIT_WORK(&args->work, recv_work);
1308                 args->nbd = nbd;
1309                 args->index = i;
1310                 queue_work(nbd->recv_workq, &args->work);
1311         }
1312         nbd_size_update(nbd, true);
1313         return error;
1314 }
1315
1316 static int nbd_start_device_ioctl(struct nbd_device *nbd, struct block_device *bdev)
1317 {
1318         struct nbd_config *config = nbd->config;
1319         int ret;
1320
1321         ret = nbd_start_device(nbd);
1322         if (ret)
1323                 return ret;
1324
1325         if (max_part)
1326                 set_bit(GD_NEED_PART_SCAN, &nbd->disk->state);
1327         mutex_unlock(&nbd->config_lock);
1328         ret = wait_event_interruptible(config->recv_wq,
1329                                          atomic_read(&config->recv_threads) == 0);
1330         if (ret)
1331                 sock_shutdown(nbd);
1332         flush_workqueue(nbd->recv_workq);
1333
1334         mutex_lock(&nbd->config_lock);
1335         nbd_bdev_reset(bdev);
1336         /* user requested, ignore socket errors */
1337         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1338                 ret = 0;
1339         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1340                 ret = -ETIMEDOUT;
1341         return ret;
1342 }
1343
1344 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1345                                  struct block_device *bdev)
1346 {
1347         sock_shutdown(nbd);
1348         __invalidate_device(bdev, true);
1349         nbd_bdev_reset(bdev);
1350         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1351                                &nbd->config->runtime_flags))
1352                 nbd_config_put(nbd);
1353 }
1354
1355 static bool nbd_is_valid_blksize(unsigned long blksize)
1356 {
1357         if (!blksize || !is_power_of_2(blksize) || blksize < 512 ||
1358             blksize > PAGE_SIZE)
1359                 return false;
1360         return true;
1361 }
1362
1363 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1364 {
1365         nbd->tag_set.timeout = timeout * HZ;
1366         if (timeout)
1367                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1368         else
1369                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1370 }
1371
1372 /* Must be called with config_lock held */
1373 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1374                        unsigned int cmd, unsigned long arg)
1375 {
1376         struct nbd_config *config = nbd->config;
1377
1378         switch (cmd) {
1379         case NBD_DISCONNECT:
1380                 return nbd_disconnect(nbd);
1381         case NBD_CLEAR_SOCK:
1382                 nbd_clear_sock_ioctl(nbd, bdev);
1383                 return 0;
1384         case NBD_SET_SOCK:
1385                 return nbd_add_socket(nbd, arg, false);
1386         case NBD_SET_BLKSIZE:
1387                 if (!arg)
1388                         arg = NBD_DEF_BLKSIZE;
1389                 if (!nbd_is_valid_blksize(arg))
1390                         return -EINVAL;
1391                 nbd_size_set(nbd, arg,
1392                              div_s64(config->bytesize, arg));
1393                 return 0;
1394         case NBD_SET_SIZE:
1395                 nbd_size_set(nbd, config->blksize,
1396                              div_s64(arg, config->blksize));
1397                 return 0;
1398         case NBD_SET_SIZE_BLOCKS:
1399                 nbd_size_set(nbd, config->blksize, arg);
1400                 return 0;
1401         case NBD_SET_TIMEOUT:
1402                 nbd_set_cmd_timeout(nbd, arg);
1403                 return 0;
1404
1405         case NBD_SET_FLAGS:
1406                 config->flags = arg;
1407                 return 0;
1408         case NBD_DO_IT:
1409                 return nbd_start_device_ioctl(nbd, bdev);
1410         case NBD_CLEAR_QUE:
1411                 /*
1412                  * This is for compatibility only.  The queue is always cleared
1413                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1414                  */
1415                 return 0;
1416         case NBD_PRINT_DEBUG:
1417                 /*
1418                  * For compatibility only, we no longer keep a list of
1419                  * outstanding requests.
1420                  */
1421                 return 0;
1422         }
1423         return -ENOTTY;
1424 }
1425
1426 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1427                      unsigned int cmd, unsigned long arg)
1428 {
1429         struct nbd_device *nbd = bdev->bd_disk->private_data;
1430         struct nbd_config *config = nbd->config;
1431         int error = -EINVAL;
1432
1433         if (!capable(CAP_SYS_ADMIN))
1434                 return -EPERM;
1435
1436         /* The block layer will pass back some non-nbd ioctls in case we have
1437          * special handling for them, but we don't so just return an error.
1438          */
1439         if (_IOC_TYPE(cmd) != 0xab)
1440                 return -EINVAL;
1441
1442         mutex_lock(&nbd->config_lock);
1443
1444         /* Don't allow ioctl operations on a nbd device that was created with
1445          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1446          */
1447         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1448             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1449                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1450         else
1451                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1452         mutex_unlock(&nbd->config_lock);
1453         return error;
1454 }
1455
1456 static struct nbd_config *nbd_alloc_config(void)
1457 {
1458         struct nbd_config *config;
1459
1460         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1461         if (!config)
1462                 return NULL;
1463         atomic_set(&config->recv_threads, 0);
1464         init_waitqueue_head(&config->recv_wq);
1465         init_waitqueue_head(&config->conn_wait);
1466         config->blksize = NBD_DEF_BLKSIZE;
1467         atomic_set(&config->live_connections, 0);
1468         try_module_get(THIS_MODULE);
1469         return config;
1470 }
1471
1472 static int nbd_open(struct block_device *bdev, fmode_t mode)
1473 {
1474         struct nbd_device *nbd;
1475         int ret = 0;
1476
1477         mutex_lock(&nbd_index_mutex);
1478         nbd = bdev->bd_disk->private_data;
1479         if (!nbd) {
1480                 ret = -ENXIO;
1481                 goto out;
1482         }
1483         if (!refcount_inc_not_zero(&nbd->refs)) {
1484                 ret = -ENXIO;
1485                 goto out;
1486         }
1487         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1488                 struct nbd_config *config;
1489
1490                 mutex_lock(&nbd->config_lock);
1491                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1492                         mutex_unlock(&nbd->config_lock);
1493                         goto out;
1494                 }
1495                 config = nbd->config = nbd_alloc_config();
1496                 if (!config) {
1497                         ret = -ENOMEM;
1498                         mutex_unlock(&nbd->config_lock);
1499                         goto out;
1500                 }
1501                 refcount_set(&nbd->config_refs, 1);
1502                 refcount_inc(&nbd->refs);
1503                 mutex_unlock(&nbd->config_lock);
1504                 set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1505         } else if (nbd_disconnected(nbd->config)) {
1506                 set_bit(GD_NEED_PART_SCAN, &bdev->bd_disk->state);
1507         }
1508 out:
1509         mutex_unlock(&nbd_index_mutex);
1510         return ret;
1511 }
1512
1513 static void nbd_release(struct gendisk *disk, fmode_t mode)
1514 {
1515         struct nbd_device *nbd = disk->private_data;
1516         struct block_device *bdev = bdget_disk(disk, 0);
1517
1518         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1519                         bdev->bd_openers == 0)
1520                 nbd_disconnect_and_put(nbd);
1521
1522         nbd_config_put(nbd);
1523         nbd_put(nbd);
1524 }
1525
1526 static const struct block_device_operations nbd_fops =
1527 {
1528         .owner =        THIS_MODULE,
1529         .open =         nbd_open,
1530         .release =      nbd_release,
1531         .ioctl =        nbd_ioctl,
1532         .compat_ioctl = nbd_ioctl,
1533 };
1534
1535 #if IS_ENABLED(CONFIG_DEBUG_FS)
1536
1537 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1538 {
1539         struct nbd_device *nbd = s->private;
1540
1541         if (nbd->task_recv)
1542                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1543
1544         return 0;
1545 }
1546
1547 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1548 {
1549         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1550 }
1551
1552 static const struct file_operations nbd_dbg_tasks_ops = {
1553         .open = nbd_dbg_tasks_open,
1554         .read = seq_read,
1555         .llseek = seq_lseek,
1556         .release = single_release,
1557 };
1558
1559 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1560 {
1561         struct nbd_device *nbd = s->private;
1562         u32 flags = nbd->config->flags;
1563
1564         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1565
1566         seq_puts(s, "Known flags:\n");
1567
1568         if (flags & NBD_FLAG_HAS_FLAGS)
1569                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1570         if (flags & NBD_FLAG_READ_ONLY)
1571                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1572         if (flags & NBD_FLAG_SEND_FLUSH)
1573                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1574         if (flags & NBD_FLAG_SEND_FUA)
1575                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1576         if (flags & NBD_FLAG_SEND_TRIM)
1577                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1578
1579         return 0;
1580 }
1581
1582 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1583 {
1584         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1585 }
1586
1587 static const struct file_operations nbd_dbg_flags_ops = {
1588         .open = nbd_dbg_flags_open,
1589         .read = seq_read,
1590         .llseek = seq_lseek,
1591         .release = single_release,
1592 };
1593
1594 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1595 {
1596         struct dentry *dir;
1597         struct nbd_config *config = nbd->config;
1598
1599         if (!nbd_dbg_dir)
1600                 return -EIO;
1601
1602         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1603         if (!dir) {
1604                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1605                         nbd_name(nbd));
1606                 return -EIO;
1607         }
1608         config->dbg_dir = dir;
1609
1610         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1611         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1612         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1613         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1614         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1615
1616         return 0;
1617 }
1618
1619 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1620 {
1621         debugfs_remove_recursive(nbd->config->dbg_dir);
1622 }
1623
1624 static int nbd_dbg_init(void)
1625 {
1626         struct dentry *dbg_dir;
1627
1628         dbg_dir = debugfs_create_dir("nbd", NULL);
1629         if (!dbg_dir)
1630                 return -EIO;
1631
1632         nbd_dbg_dir = dbg_dir;
1633
1634         return 0;
1635 }
1636
1637 static void nbd_dbg_close(void)
1638 {
1639         debugfs_remove_recursive(nbd_dbg_dir);
1640 }
1641
1642 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1643
1644 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1645 {
1646         return 0;
1647 }
1648
1649 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1650 {
1651 }
1652
1653 static int nbd_dbg_init(void)
1654 {
1655         return 0;
1656 }
1657
1658 static void nbd_dbg_close(void)
1659 {
1660 }
1661
1662 #endif
1663
1664 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1665                             unsigned int hctx_idx, unsigned int numa_node)
1666 {
1667         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1668         cmd->nbd = set->driver_data;
1669         cmd->flags = 0;
1670         mutex_init(&cmd->lock);
1671         return 0;
1672 }
1673
1674 static const struct blk_mq_ops nbd_mq_ops = {
1675         .queue_rq       = nbd_queue_rq,
1676         .complete       = nbd_complete_rq,
1677         .init_request   = nbd_init_request,
1678         .timeout        = nbd_xmit_timeout,
1679 };
1680
1681 static int nbd_dev_add(int index)
1682 {
1683         struct nbd_device *nbd;
1684         struct gendisk *disk;
1685         struct request_queue *q;
1686         int err = -ENOMEM;
1687
1688         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1689         if (!nbd)
1690                 goto out;
1691
1692         disk = alloc_disk(1 << part_shift);
1693         if (!disk)
1694                 goto out_free_nbd;
1695
1696         if (index >= 0) {
1697                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1698                                 GFP_KERNEL);
1699                 if (err == -ENOSPC)
1700                         err = -EEXIST;
1701         } else {
1702                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1703                 if (err >= 0)
1704                         index = err;
1705         }
1706         if (err < 0)
1707                 goto out_free_disk;
1708
1709         nbd->index = index;
1710         nbd->disk = disk;
1711         nbd->tag_set.ops = &nbd_mq_ops;
1712         nbd->tag_set.nr_hw_queues = 1;
1713         nbd->tag_set.queue_depth = 128;
1714         nbd->tag_set.numa_node = NUMA_NO_NODE;
1715         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1716         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1717                 BLK_MQ_F_BLOCKING;
1718         nbd->tag_set.driver_data = nbd;
1719         nbd->destroy_complete = NULL;
1720
1721         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1722         if (err)
1723                 goto out_free_idr;
1724
1725         q = blk_mq_init_queue(&nbd->tag_set);
1726         if (IS_ERR(q)) {
1727                 err = PTR_ERR(q);
1728                 goto out_free_tags;
1729         }
1730         disk->queue = q;
1731
1732         /*
1733          * Tell the block layer that we are not a rotational device
1734          */
1735         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1736         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1737         disk->queue->limits.discard_granularity = 0;
1738         disk->queue->limits.discard_alignment = 0;
1739         blk_queue_max_discard_sectors(disk->queue, 0);
1740         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1741         blk_queue_max_segments(disk->queue, USHRT_MAX);
1742         blk_queue_max_hw_sectors(disk->queue, 65536);
1743         disk->queue->limits.max_sectors = 256;
1744
1745         mutex_init(&nbd->config_lock);
1746         refcount_set(&nbd->config_refs, 0);
1747         refcount_set(&nbd->refs, 1);
1748         INIT_LIST_HEAD(&nbd->list);
1749         disk->major = NBD_MAJOR;
1750         disk->first_minor = index << part_shift;
1751         disk->fops = &nbd_fops;
1752         disk->private_data = nbd;
1753         sprintf(disk->disk_name, "nbd%d", index);
1754         add_disk(disk);
1755         nbd_total_devices++;
1756         return index;
1757
1758 out_free_tags:
1759         blk_mq_free_tag_set(&nbd->tag_set);
1760 out_free_idr:
1761         idr_remove(&nbd_index_idr, index);
1762 out_free_disk:
1763         put_disk(disk);
1764 out_free_nbd:
1765         kfree(nbd);
1766 out:
1767         return err;
1768 }
1769
1770 static int find_free_cb(int id, void *ptr, void *data)
1771 {
1772         struct nbd_device *nbd = ptr;
1773         struct nbd_device **found = data;
1774
1775         if (!refcount_read(&nbd->config_refs)) {
1776                 *found = nbd;
1777                 return 1;
1778         }
1779         return 0;
1780 }
1781
1782 /* Netlink interface. */
1783 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1784         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1785         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1786         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1787         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1788         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1789         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1790         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1791         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1792         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1793 };
1794
1795 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1796         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1797 };
1798
1799 /* We don't use this right now since we don't parse the incoming list, but we
1800  * still want it here so userspace knows what to expect.
1801  */
1802 static const struct nla_policy __attribute__((unused))
1803 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1804         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1805         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1806 };
1807
1808 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1809 {
1810         struct nbd_config *config = nbd->config;
1811         u64 bsize = config->blksize;
1812         u64 bytes = config->bytesize;
1813
1814         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1815                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1816
1817         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1818                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1819                 if (!bsize)
1820                         bsize = NBD_DEF_BLKSIZE;
1821                 if (!nbd_is_valid_blksize(bsize)) {
1822                         printk(KERN_ERR "Invalid block size %llu\n", bsize);
1823                         return -EINVAL;
1824                 }
1825         }
1826
1827         if (bytes != config->bytesize || bsize != config->blksize)
1828                 nbd_size_set(nbd, bsize, div64_u64(bytes, bsize));
1829         return 0;
1830 }
1831
1832 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1833 {
1834         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1835         struct nbd_device *nbd = NULL;
1836         struct nbd_config *config;
1837         int index = -1;
1838         int ret;
1839         bool put_dev = false;
1840
1841         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1842                 return -EPERM;
1843
1844         if (info->attrs[NBD_ATTR_INDEX])
1845                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1846         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1847                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1848                 return -EINVAL;
1849         }
1850         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1851                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1852                 return -EINVAL;
1853         }
1854 again:
1855         mutex_lock(&nbd_index_mutex);
1856         if (index == -1) {
1857                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1858                 if (ret == 0) {
1859                         int new_index;
1860                         new_index = nbd_dev_add(-1);
1861                         if (new_index < 0) {
1862                                 mutex_unlock(&nbd_index_mutex);
1863                                 printk(KERN_ERR "nbd: failed to add new device\n");
1864                                 return new_index;
1865                         }
1866                         nbd = idr_find(&nbd_index_idr, new_index);
1867                 }
1868         } else {
1869                 nbd = idr_find(&nbd_index_idr, index);
1870                 if (!nbd) {
1871                         ret = nbd_dev_add(index);
1872                         if (ret < 0) {
1873                                 mutex_unlock(&nbd_index_mutex);
1874                                 printk(KERN_ERR "nbd: failed to add new device\n");
1875                                 return ret;
1876                         }
1877                         nbd = idr_find(&nbd_index_idr, index);
1878                 }
1879         }
1880         if (!nbd) {
1881                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1882                        index);
1883                 mutex_unlock(&nbd_index_mutex);
1884                 return -EINVAL;
1885         }
1886
1887         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1888             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1889                 nbd->destroy_complete = &destroy_complete;
1890                 mutex_unlock(&nbd_index_mutex);
1891
1892                 /* Wait untill the the nbd stuff is totally destroyed */
1893                 wait_for_completion(&destroy_complete);
1894                 goto again;
1895         }
1896
1897         if (!refcount_inc_not_zero(&nbd->refs)) {
1898                 mutex_unlock(&nbd_index_mutex);
1899                 if (index == -1)
1900                         goto again;
1901                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1902                        index);
1903                 return -EINVAL;
1904         }
1905         mutex_unlock(&nbd_index_mutex);
1906
1907         mutex_lock(&nbd->config_lock);
1908         if (refcount_read(&nbd->config_refs)) {
1909                 mutex_unlock(&nbd->config_lock);
1910                 nbd_put(nbd);
1911                 if (index == -1)
1912                         goto again;
1913                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1914                 return -EBUSY;
1915         }
1916         if (WARN_ON(nbd->config)) {
1917                 mutex_unlock(&nbd->config_lock);
1918                 nbd_put(nbd);
1919                 return -EINVAL;
1920         }
1921         config = nbd->config = nbd_alloc_config();
1922         if (!nbd->config) {
1923                 mutex_unlock(&nbd->config_lock);
1924                 nbd_put(nbd);
1925                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1926                 return -ENOMEM;
1927         }
1928         refcount_set(&nbd->config_refs, 1);
1929         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1930
1931         ret = nbd_genl_size_set(info, nbd);
1932         if (ret)
1933                 goto out;
1934
1935         if (info->attrs[NBD_ATTR_TIMEOUT])
1936                 nbd_set_cmd_timeout(nbd,
1937                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1938         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1939                 config->dead_conn_timeout =
1940                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1941                 config->dead_conn_timeout *= HZ;
1942         }
1943         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1944                 config->flags =
1945                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1946         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1947                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1948                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1949                         set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
1950                                 &config->runtime_flags);
1951                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1952                         put_dev = true;
1953                 } else {
1954                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
1955                 }
1956                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1957                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1958                                 &config->runtime_flags);
1959                 }
1960         }
1961
1962         if (info->attrs[NBD_ATTR_SOCKETS]) {
1963                 struct nlattr *attr;
1964                 int rem, fd;
1965
1966                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1967                                     rem) {
1968                         struct nlattr *socks[NBD_SOCK_MAX+1];
1969
1970                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1971                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1972                                 ret = -EINVAL;
1973                                 goto out;
1974                         }
1975                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1976                                                           attr,
1977                                                           nbd_sock_policy,
1978                                                           info->extack);
1979                         if (ret != 0) {
1980                                 printk(KERN_ERR "nbd: error processing sock list\n");
1981                                 ret = -EINVAL;
1982                                 goto out;
1983                         }
1984                         if (!socks[NBD_SOCK_FD])
1985                                 continue;
1986                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
1987                         ret = nbd_add_socket(nbd, fd, true);
1988                         if (ret)
1989                                 goto out;
1990                 }
1991         }
1992         ret = nbd_start_device(nbd);
1993 out:
1994         mutex_unlock(&nbd->config_lock);
1995         if (!ret) {
1996                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
1997                 refcount_inc(&nbd->config_refs);
1998                 nbd_connect_reply(info, nbd->index);
1999         }
2000         nbd_config_put(nbd);
2001         if (put_dev)
2002                 nbd_put(nbd);
2003         return ret;
2004 }
2005
2006 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2007 {
2008         mutex_lock(&nbd->config_lock);
2009         nbd_disconnect(nbd);
2010         nbd_clear_sock(nbd);
2011         mutex_unlock(&nbd->config_lock);
2012         /*
2013          * Make sure recv thread has finished, so it does not drop the last
2014          * config ref and try to destroy the workqueue from inside the work
2015          * queue.
2016          */
2017         flush_workqueue(nbd->recv_workq);
2018         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2019                                &nbd->config->runtime_flags))
2020                 nbd_config_put(nbd);
2021 }
2022
2023 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2024 {
2025         struct nbd_device *nbd;
2026         int index;
2027
2028         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2029                 return -EPERM;
2030
2031         if (!info->attrs[NBD_ATTR_INDEX]) {
2032                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2033                 return -EINVAL;
2034         }
2035         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2036         mutex_lock(&nbd_index_mutex);
2037         nbd = idr_find(&nbd_index_idr, index);
2038         if (!nbd) {
2039                 mutex_unlock(&nbd_index_mutex);
2040                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2041                        index);
2042                 return -EINVAL;
2043         }
2044         if (!refcount_inc_not_zero(&nbd->refs)) {
2045                 mutex_unlock(&nbd_index_mutex);
2046                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2047                        index);
2048                 return -EINVAL;
2049         }
2050         mutex_unlock(&nbd_index_mutex);
2051         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2052                 nbd_put(nbd);
2053                 return 0;
2054         }
2055         nbd_disconnect_and_put(nbd);
2056         nbd_config_put(nbd);
2057         nbd_put(nbd);
2058         return 0;
2059 }
2060
2061 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2062 {
2063         struct nbd_device *nbd = NULL;
2064         struct nbd_config *config;
2065         int index;
2066         int ret = 0;
2067         bool put_dev = false;
2068
2069         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2070                 return -EPERM;
2071
2072         if (!info->attrs[NBD_ATTR_INDEX]) {
2073                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2074                 return -EINVAL;
2075         }
2076         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2077         mutex_lock(&nbd_index_mutex);
2078         nbd = idr_find(&nbd_index_idr, index);
2079         if (!nbd) {
2080                 mutex_unlock(&nbd_index_mutex);
2081                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2082                        index);
2083                 return -EINVAL;
2084         }
2085         if (!refcount_inc_not_zero(&nbd->refs)) {
2086                 mutex_unlock(&nbd_index_mutex);
2087                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2088                        index);
2089                 return -EINVAL;
2090         }
2091         mutex_unlock(&nbd_index_mutex);
2092
2093         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2094                 dev_err(nbd_to_dev(nbd),
2095                         "not configured, cannot reconfigure\n");
2096                 nbd_put(nbd);
2097                 return -EINVAL;
2098         }
2099
2100         mutex_lock(&nbd->config_lock);
2101         config = nbd->config;
2102         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2103             !nbd->task_recv) {
2104                 dev_err(nbd_to_dev(nbd),
2105                         "not configured, cannot reconfigure\n");
2106                 ret = -EINVAL;
2107                 goto out;
2108         }
2109
2110         ret = nbd_genl_size_set(info, nbd);
2111         if (ret)
2112                 goto out;
2113
2114         if (info->attrs[NBD_ATTR_TIMEOUT])
2115                 nbd_set_cmd_timeout(nbd,
2116                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2117         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2118                 config->dead_conn_timeout =
2119                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2120                 config->dead_conn_timeout *= HZ;
2121         }
2122         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2123                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2124                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2125                         if (!test_and_set_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2126                                               &config->runtime_flags))
2127                                 put_dev = true;
2128                         set_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2129                 } else {
2130                         if (test_and_clear_bit(NBD_RT_DESTROY_ON_DISCONNECT,
2131                                                &config->runtime_flags))
2132                                 refcount_inc(&nbd->refs);
2133                         clear_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags);
2134                 }
2135
2136                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2137                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2138                                         &config->runtime_flags);
2139                 } else {
2140                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2141                                         &config->runtime_flags);
2142                 }
2143         }
2144
2145         if (info->attrs[NBD_ATTR_SOCKETS]) {
2146                 struct nlattr *attr;
2147                 int rem, fd;
2148
2149                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2150                                     rem) {
2151                         struct nlattr *socks[NBD_SOCK_MAX+1];
2152
2153                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2154                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2155                                 ret = -EINVAL;
2156                                 goto out;
2157                         }
2158                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2159                                                           attr,
2160                                                           nbd_sock_policy,
2161                                                           info->extack);
2162                         if (ret != 0) {
2163                                 printk(KERN_ERR "nbd: error processing sock list\n");
2164                                 ret = -EINVAL;
2165                                 goto out;
2166                         }
2167                         if (!socks[NBD_SOCK_FD])
2168                                 continue;
2169                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2170                         ret = nbd_reconnect_socket(nbd, fd);
2171                         if (ret) {
2172                                 if (ret == -ENOSPC)
2173                                         ret = 0;
2174                                 goto out;
2175                         }
2176                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2177                 }
2178         }
2179 out:
2180         mutex_unlock(&nbd->config_lock);
2181         nbd_config_put(nbd);
2182         nbd_put(nbd);
2183         if (put_dev)
2184                 nbd_put(nbd);
2185         return ret;
2186 }
2187
2188 static const struct genl_small_ops nbd_connect_genl_ops[] = {
2189         {
2190                 .cmd    = NBD_CMD_CONNECT,
2191                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2192                 .doit   = nbd_genl_connect,
2193         },
2194         {
2195                 .cmd    = NBD_CMD_DISCONNECT,
2196                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2197                 .doit   = nbd_genl_disconnect,
2198         },
2199         {
2200                 .cmd    = NBD_CMD_RECONFIGURE,
2201                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2202                 .doit   = nbd_genl_reconfigure,
2203         },
2204         {
2205                 .cmd    = NBD_CMD_STATUS,
2206                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2207                 .doit   = nbd_genl_status,
2208         },
2209 };
2210
2211 static const struct genl_multicast_group nbd_mcast_grps[] = {
2212         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2213 };
2214
2215 static struct genl_family nbd_genl_family __ro_after_init = {
2216         .hdrsize        = 0,
2217         .name           = NBD_GENL_FAMILY_NAME,
2218         .version        = NBD_GENL_VERSION,
2219         .module         = THIS_MODULE,
2220         .small_ops      = nbd_connect_genl_ops,
2221         .n_small_ops    = ARRAY_SIZE(nbd_connect_genl_ops),
2222         .maxattr        = NBD_ATTR_MAX,
2223         .policy = nbd_attr_policy,
2224         .mcgrps         = nbd_mcast_grps,
2225         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2226 };
2227
2228 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2229 {
2230         struct nlattr *dev_opt;
2231         u8 connected = 0;
2232         int ret;
2233
2234         /* This is a little racey, but for status it's ok.  The
2235          * reason we don't take a ref here is because we can't
2236          * take a ref in the index == -1 case as we would need
2237          * to put under the nbd_index_mutex, which could
2238          * deadlock if we are configured to remove ourselves
2239          * once we're disconnected.
2240          */
2241         if (refcount_read(&nbd->config_refs))
2242                 connected = 1;
2243         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2244         if (!dev_opt)
2245                 return -EMSGSIZE;
2246         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2247         if (ret)
2248                 return -EMSGSIZE;
2249         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2250                          connected);
2251         if (ret)
2252                 return -EMSGSIZE;
2253         nla_nest_end(reply, dev_opt);
2254         return 0;
2255 }
2256
2257 static int status_cb(int id, void *ptr, void *data)
2258 {
2259         struct nbd_device *nbd = ptr;
2260         return populate_nbd_status(nbd, (struct sk_buff *)data);
2261 }
2262
2263 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2264 {
2265         struct nlattr *dev_list;
2266         struct sk_buff *reply;
2267         void *reply_head;
2268         size_t msg_size;
2269         int index = -1;
2270         int ret = -ENOMEM;
2271
2272         if (info->attrs[NBD_ATTR_INDEX])
2273                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2274
2275         mutex_lock(&nbd_index_mutex);
2276
2277         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2278                                   nla_attr_size(sizeof(u8)));
2279         msg_size *= (index == -1) ? nbd_total_devices : 1;
2280
2281         reply = genlmsg_new(msg_size, GFP_KERNEL);
2282         if (!reply)
2283                 goto out;
2284         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2285                                        NBD_CMD_STATUS);
2286         if (!reply_head) {
2287                 nlmsg_free(reply);
2288                 goto out;
2289         }
2290
2291         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2292         if (index == -1) {
2293                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2294                 if (ret) {
2295                         nlmsg_free(reply);
2296                         goto out;
2297                 }
2298         } else {
2299                 struct nbd_device *nbd;
2300                 nbd = idr_find(&nbd_index_idr, index);
2301                 if (nbd) {
2302                         ret = populate_nbd_status(nbd, reply);
2303                         if (ret) {
2304                                 nlmsg_free(reply);
2305                                 goto out;
2306                         }
2307                 }
2308         }
2309         nla_nest_end(reply, dev_list);
2310         genlmsg_end(reply, reply_head);
2311         ret = genlmsg_reply(reply, info);
2312 out:
2313         mutex_unlock(&nbd_index_mutex);
2314         return ret;
2315 }
2316
2317 static void nbd_connect_reply(struct genl_info *info, int index)
2318 {
2319         struct sk_buff *skb;
2320         void *msg_head;
2321         int ret;
2322
2323         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2324         if (!skb)
2325                 return;
2326         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2327                                      NBD_CMD_CONNECT);
2328         if (!msg_head) {
2329                 nlmsg_free(skb);
2330                 return;
2331         }
2332         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2333         if (ret) {
2334                 nlmsg_free(skb);
2335                 return;
2336         }
2337         genlmsg_end(skb, msg_head);
2338         genlmsg_reply(skb, info);
2339 }
2340
2341 static void nbd_mcast_index(int index)
2342 {
2343         struct sk_buff *skb;
2344         void *msg_head;
2345         int ret;
2346
2347         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2348         if (!skb)
2349                 return;
2350         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2351                                      NBD_CMD_LINK_DEAD);
2352         if (!msg_head) {
2353                 nlmsg_free(skb);
2354                 return;
2355         }
2356         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2357         if (ret) {
2358                 nlmsg_free(skb);
2359                 return;
2360         }
2361         genlmsg_end(skb, msg_head);
2362         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2363 }
2364
2365 static void nbd_dead_link_work(struct work_struct *work)
2366 {
2367         struct link_dead_args *args = container_of(work, struct link_dead_args,
2368                                                    work);
2369         nbd_mcast_index(args->index);
2370         kfree(args);
2371 }
2372
2373 static int __init nbd_init(void)
2374 {
2375         int i;
2376
2377         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2378
2379         if (max_part < 0) {
2380                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2381                 return -EINVAL;
2382         }
2383
2384         part_shift = 0;
2385         if (max_part > 0) {
2386                 part_shift = fls(max_part);
2387
2388                 /*
2389                  * Adjust max_part according to part_shift as it is exported
2390                  * to user space so that user can know the max number of
2391                  * partition kernel should be able to manage.
2392                  *
2393                  * Note that -1 is required because partition 0 is reserved
2394                  * for the whole disk.
2395                  */
2396                 max_part = (1UL << part_shift) - 1;
2397         }
2398
2399         if ((1UL << part_shift) > DISK_MAX_PARTS)
2400                 return -EINVAL;
2401
2402         if (nbds_max > 1UL << (MINORBITS - part_shift))
2403                 return -EINVAL;
2404
2405         if (register_blkdev(NBD_MAJOR, "nbd"))
2406                 return -EIO;
2407
2408         if (genl_register_family(&nbd_genl_family)) {
2409                 unregister_blkdev(NBD_MAJOR, "nbd");
2410                 return -EINVAL;
2411         }
2412         nbd_dbg_init();
2413
2414         mutex_lock(&nbd_index_mutex);
2415         for (i = 0; i < nbds_max; i++)
2416                 nbd_dev_add(i);
2417         mutex_unlock(&nbd_index_mutex);
2418         return 0;
2419 }
2420
2421 static int nbd_exit_cb(int id, void *ptr, void *data)
2422 {
2423         struct list_head *list = (struct list_head *)data;
2424         struct nbd_device *nbd = ptr;
2425
2426         list_add_tail(&nbd->list, list);
2427         return 0;
2428 }
2429
2430 static void __exit nbd_cleanup(void)
2431 {
2432         struct nbd_device *nbd;
2433         LIST_HEAD(del_list);
2434
2435         nbd_dbg_close();
2436
2437         mutex_lock(&nbd_index_mutex);
2438         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2439         mutex_unlock(&nbd_index_mutex);
2440
2441         while (!list_empty(&del_list)) {
2442                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2443                 list_del_init(&nbd->list);
2444                 if (refcount_read(&nbd->refs) != 1)
2445                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2446                 nbd_put(nbd);
2447         }
2448
2449         idr_destroy(&nbd_index_idr);
2450         genl_unregister_family(&nbd_genl_family);
2451         unregister_blkdev(NBD_MAJOR, "nbd");
2452 }
2453
2454 module_init(nbd_init);
2455 module_exit(nbd_cleanup);
2456
2457 MODULE_DESCRIPTION("Network Block Device");
2458 MODULE_LICENSE("GPL");
2459
2460 module_param(nbds_max, int, 0444);
2461 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2462 module_param(max_part, int, 0444);
2463 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");