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_DISCONNECT_ON_CLOSE      6
  82
  83 #define NBD_DESTROY_ON_DISCONNECT       0
  84 #define NBD_DISCONNECT_REQUESTED        1
  85
  86 struct nbd_config {
  87         u32 flags;
  88         unsigned long runtime_flags;
  89         u64 dead_conn_timeout;
  90
  91         struct nbd_sock **socks;
  92         int num_connections;
  93         atomic_t live_connections;
  94         wait_queue_head_t conn_wait;
  95
  96         atomic_t recv_threads;
  97         wait_queue_head_t recv_wq;
  98         loff_t blksize;
  99         loff_t bytesize;
 100 #if IS_ENABLED(CONFIG_DEBUG_FS)
 101         struct dentry *dbg_dir;
 102 #endif
 103 };
 104
 105 struct nbd_device {
 106         struct blk_mq_tag_set tag_set;
 107
 108         int index;
 109         refcount_t config_refs;
 110         refcount_t refs;
 111         struct nbd_config *config;
 112         struct mutex config_lock;
 113         struct gendisk *disk;
 114         struct workqueue_struct *recv_workq;
 115
 116         struct list_head list;
 117         struct task_struct *task_recv;
 118         struct task_struct *task_setup;
 119
 120         struct completion *destroy_complete;
 121         unsigned long flags;
 122 };
 123
 124 #define NBD_CMD_REQUEUED        1
 125
 126 struct nbd_cmd {
 127         struct nbd_device *nbd;
 128         struct mutex lock;
 129         int index;
 130         int cookie;
 131         int retries;
 132         blk_status_t status;
 133         unsigned long flags;
 134         u32 cmd_cookie;
 135 };
 136
 137 #if IS_ENABLED(CONFIG_DEBUG_FS)
 138 static struct dentry *nbd_dbg_dir;
 139 #endif
 140
 141 #define nbd_name(nbd) ((nbd)->disk->disk_name)
 142
 143 #define NBD_MAGIC 0x68797548
 144
 145 #define NBD_DEF_BLKSIZE 1024
 146
 147 static unsigned int nbds_max = 16;
 148 static int max_part = 16;
 149 static int part_shift;
 150
 151 static int nbd_dev_dbg_init(struct nbd_device *nbd);
 152 static void nbd_dev_dbg_close(struct nbd_device *nbd);
 153 static void nbd_config_put(struct nbd_device *nbd);
 154 static void nbd_connect_reply(struct genl_info *info, int index);
 155 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info);
 156 static void nbd_dead_link_work(struct work_struct *work);
 157 static void nbd_disconnect_and_put(struct nbd_device *nbd);
 158
 159 static inline struct device *nbd_to_dev(struct nbd_device *nbd)
 160 {
 161         return disk_to_dev(nbd->disk);
 162 }
 163
 164 static void nbd_requeue_cmd(struct nbd_cmd *cmd)
 165 {
 166         struct request *req = blk_mq_rq_from_pdu(cmd);
 167
 168         if (!test_and_set_bit(NBD_CMD_REQUEUED, &cmd->flags))
 169                 blk_mq_requeue_request(req, true);
 170 }
 171
 172 #define NBD_COOKIE_BITS 32
 173
 174 static u64 nbd_cmd_handle(struct nbd_cmd *cmd)
 175 {
 176         struct request *req = blk_mq_rq_from_pdu(cmd);
 177         u32 tag = blk_mq_unique_tag(req);
 178         u64 cookie = cmd->cmd_cookie;
 179
 180         return (cookie << NBD_COOKIE_BITS) | tag;
 181 }
 182
 183 static u32 nbd_handle_to_tag(u64 handle)
 184 {
 185         return (u32)handle;
 186 }
 187
 188 static u32 nbd_handle_to_cookie(u64 handle)
 189 {
 190         return (u32)(handle >> NBD_COOKIE_BITS);
 191 }
 192
 193 static const char *nbdcmd_to_ascii(int cmd)
 194 {
 195         switch (cmd) {
 196         case  NBD_CMD_READ: return "read";
 197         case NBD_CMD_WRITE: return "write";
 198         case  NBD_CMD_DISC: return "disconnect";
 199         case NBD_CMD_FLUSH: return "flush";
 200         case  NBD_CMD_TRIM: return "trim/discard";
 201         }
 202         return "invalid";
 203 }
 204
 205 static ssize_t pid_show(struct device *dev,
 206                         struct device_attribute *attr, char *buf)
 207 {
 208         struct gendisk *disk = dev_to_disk(dev);
 209         struct nbd_device *nbd = (struct nbd_device *)disk->private_data;
 210
 211         return sprintf(buf, "%d\n", task_pid_nr(nbd->task_recv));
 212 }
 213
 214 static const struct device_attribute pid_attr = {
 215         .attr = { .name = "pid", .mode = 0444},
 216         .show = pid_show,
 217 };
 218
 219 static void nbd_dev_remove(struct nbd_device *nbd)
 220 {
 221         struct gendisk *disk = nbd->disk;
 222         struct request_queue *q;
 223
 224         if (disk) {
 225                 q = disk->queue;
 226                 del_gendisk(disk);
 227                 blk_cleanup_queue(q);
 228                 blk_mq_free_tag_set(&nbd->tag_set);
 229                 disk->private_data = NULL;
 230                 put_disk(disk);
 231         }
 232
 233         /*
 234          * Place this in the last just before the nbd is freed to
 235          * make sure that the disk and the related kobject are also
 236          * totally removed to avoid duplicate creation of the same
 237          * one.
 238          */
 239         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) && nbd->destroy_complete)
 240                 complete(nbd->destroy_complete);
 241
 242         kfree(nbd);
 243 }
 244
 245 static void nbd_put(struct nbd_device *nbd)
 246 {
 247         if (refcount_dec_and_mutex_lock(&nbd->refs,
 248                                         &nbd_index_mutex)) {
 249                 idr_remove(&nbd_index_idr, nbd->index);
 250                 nbd_dev_remove(nbd);
 251                 mutex_unlock(&nbd_index_mutex);
 252         }
 253 }
 254
 255 static int nbd_disconnected(struct nbd_config *config)
 256 {
 257         return test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags) ||
 258                 test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags);
 259 }
 260
 261 static void nbd_mark_nsock_dead(struct nbd_device *nbd, struct nbd_sock *nsock,
 262                                 int notify)
 263 {
 264         if (!nsock->dead && notify && !nbd_disconnected(nbd->config)) {
 265                 struct link_dead_args *args;
 266                 args = kmalloc(sizeof(struct link_dead_args), GFP_NOIO);
 267                 if (args) {
 268                         INIT_WORK(&args->work, nbd_dead_link_work);
 269                         args->index = nbd->index;
 270                         queue_work(system_wq, &args->work);
 271                 }
 272         }
 273         if (!nsock->dead) {
 274                 kernel_sock_shutdown(nsock->sock, SHUT_RDWR);
 275                 if (atomic_dec_return(&nbd->config->live_connections) == 0) {
 276                         if (test_and_clear_bit(NBD_RT_DISCONNECT_REQUESTED,
 277                                                &nbd->config->runtime_flags)) {
 278                                 set_bit(NBD_RT_DISCONNECTED,
 279                                         &nbd->config->runtime_flags);
 280                                 dev_info(nbd_to_dev(nbd),
 281                                         "Disconnected due to user request.\n");
 282                         }
 283                 }
 284         }
 285         nsock->dead = true;
 286         nsock->pending = NULL;
 287         nsock->sent = 0;
 288 }
 289
 290 static void nbd_size_clear(struct nbd_device *nbd)
 291 {
 292         if (nbd->config->bytesize) {
 293                 set_capacity(nbd->disk, 0);
 294                 kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 295         }
 296 }
 297
 298 static void nbd_size_update(struct nbd_device *nbd, bool start)
 299 {
 300         struct nbd_config *config = nbd->config;
 301         struct block_device *bdev = bdget_disk(nbd->disk, 0);
 302
 303         if (config->flags & NBD_FLAG_SEND_TRIM) {
 304                 nbd->disk->queue->limits.discard_granularity = config->blksize;
 305                 nbd->disk->queue->limits.discard_alignment = config->blksize;
 306                 blk_queue_max_discard_sectors(nbd->disk->queue, UINT_MAX);
 307         }
 308         blk_queue_logical_block_size(nbd->disk->queue, config->blksize);
 309         blk_queue_physical_block_size(nbd->disk->queue, config->blksize);
 310         set_capacity(nbd->disk, config->bytesize >> 9);
 311         if (bdev) {
 312                 if (bdev->bd_disk) {
 313                         bd_set_size(bdev, config->bytesize);
 314                         if (start)
 315                                 set_blocksize(bdev, config->blksize);
 316                 } else
 317                         bdev->bd_invalidated = 1;
 318                 bdput(bdev);
 319         }
 320         kobject_uevent(&nbd_to_dev(nbd)->kobj, KOBJ_CHANGE);
 321 }
 322
 323 static void nbd_size_set(struct nbd_device *nbd, loff_t blocksize,
 324                          loff_t nr_blocks)
 325 {
 326         struct nbd_config *config = nbd->config;
 327         config->blksize = blocksize;
 328         config->bytesize = blocksize * nr_blocks;
 329         if (nbd->task_recv != NULL)
 330                 nbd_size_update(nbd, false);
 331 }
 332
 333 static void nbd_complete_rq(struct request *req)
 334 {
 335         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 336
 337         dev_dbg(nbd_to_dev(cmd->nbd), "request %p: %s\n", req,
 338                 cmd->status ? "failed" : "done");
 339
 340         blk_mq_end_request(req, cmd->status);
 341 }
 342
 343 /*
 344  * Forcibly shutdown the socket causing all listeners to error
 345  */
 346 static void sock_shutdown(struct nbd_device *nbd)
 347 {
 348         struct nbd_config *config = nbd->config;
 349         int i;
 350
 351         if (config->num_connections == 0)
 352                 return;
 353         if (test_and_set_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 354                 return;
 355
 356         for (i = 0; i < config->num_connections; i++) {
 357                 struct nbd_sock *nsock = config->socks[i];
 358                 mutex_lock(&nsock->tx_lock);
 359                 nbd_mark_nsock_dead(nbd, nsock, 0);
 360                 mutex_unlock(&nsock->tx_lock);
 361         }
 362         dev_warn(disk_to_dev(nbd->disk), "shutting down sockets\n");
 363 }
 364
 365 static u32 req_to_nbd_cmd_type(struct request *req)
 366 {
 367         switch (req_op(req)) {
 368         case REQ_OP_DISCARD:
 369                 return NBD_CMD_TRIM;
 370         case REQ_OP_FLUSH:
 371                 return NBD_CMD_FLUSH;
 372         case REQ_OP_WRITE:
 373                 return NBD_CMD_WRITE;
 374         case REQ_OP_READ:
 375                 return NBD_CMD_READ;
 376         default:
 377                 return U32_MAX;
 378         }
 379 }
 380
 381 static enum blk_eh_timer_return nbd_xmit_timeout(struct request *req,
 382                                                  bool reserved)
 383 {
 384         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 385         struct nbd_device *nbd = cmd->nbd;
 386         struct nbd_config *config;
 387
 388         if (!mutex_trylock(&cmd->lock))
 389                 return BLK_EH_RESET_TIMER;
 390
 391         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 392                 cmd->status = BLK_STS_TIMEOUT;
 393                 mutex_unlock(&cmd->lock);
 394                 goto done;
 395         }
 396         config = nbd->config;
 397
 398         if (config->num_connections > 1) {
 399                 dev_err_ratelimited(nbd_to_dev(nbd),
 400                                     "Connection timed out, retrying (%d/%d alive)\n",
 401                                     atomic_read(&config->live_connections),
 402                                     config->num_connections);
 403                 /*
 404                  * Hooray we have more connections, requeue this IO, the submit
 405                  * path will put it on a real connection.
 406                  */
 407                 if (config->socks && config->num_connections > 1) {
 408                         if (cmd->index < config->num_connections) {
 409                                 struct nbd_sock *nsock =
 410                                         config->socks[cmd->index];
 411                                 mutex_lock(&nsock->tx_lock);
 412                                 /* We can have multiple outstanding requests, so
 413                                  * we don't want to mark the nsock dead if we've
 414                                  * already reconnected with a new socket, so
 415                                  * only mark it dead if its the same socket we
 416                                  * were sent out on.
 417                                  */
 418                                 if (cmd->cookie == nsock->cookie)
 419                                         nbd_mark_nsock_dead(nbd, nsock, 1);
 420                                 mutex_unlock(&nsock->tx_lock);
 421                         }
 422                         mutex_unlock(&cmd->lock);
 423                         nbd_requeue_cmd(cmd);
 424                         nbd_config_put(nbd);
 425                         return BLK_EH_DONE;
 426                 }
 427         }
 428
 429         if (!nbd->tag_set.timeout) {
 430                 /*
 431                  * Userspace sets timeout=0 to disable socket disconnection,
 432                  * so just warn and reset the timer.
 433                  */
 434                 cmd->retries++;
 435                 dev_info(nbd_to_dev(nbd), "Possible stuck request %p: control (%s@%llu,%uB). Runtime %u seconds\n",
 436                         req, nbdcmd_to_ascii(req_to_nbd_cmd_type(req)),
 437                         (unsigned long long)blk_rq_pos(req) << 9,
 438                         blk_rq_bytes(req), (req->timeout / HZ) * cmd->retries);
 439
 440                 mutex_unlock(&cmd->lock);
 441                 nbd_config_put(nbd);
 442                 return BLK_EH_RESET_TIMER;
 443         }
 444
 445         dev_err_ratelimited(nbd_to_dev(nbd), "Connection timed out\n");
 446         set_bit(NBD_RT_TIMEDOUT, &config->runtime_flags);
 447         cmd->status = BLK_STS_IOERR;
 448         mutex_unlock(&cmd->lock);
 449         sock_shutdown(nbd);
 450         nbd_config_put(nbd);
 451 done:
 452         blk_mq_complete_request(req);
 453         return BLK_EH_DONE;
 454 }
 455
 456 /*
 457  *  Send or receive packet.
 458  */
 459 static int sock_xmit(struct nbd_device *nbd, int index, int send,
 460                      struct iov_iter *iter, int msg_flags, int *sent)
 461 {
 462         struct nbd_config *config = nbd->config;
 463         struct socket *sock = config->socks[index]->sock;
 464         int result;
 465         struct msghdr msg;
 466         unsigned int noreclaim_flag;
 467
 468         if (unlikely(!sock)) {
 469                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 470                         "Attempted %s on closed socket in sock_xmit\n",
 471                         (send ? "send" : "recv"));
 472                 return -EINVAL;
 473         }
 474
 475         msg.msg_iter = *iter;
 476
 477         noreclaim_flag = memalloc_noreclaim_save();
 478         do {
 479                 sock->sk->sk_allocation = GFP_NOIO | __GFP_MEMALLOC;
 480                 msg.msg_name = NULL;
 481                 msg.msg_namelen = 0;
 482                 msg.msg_control = NULL;
 483                 msg.msg_controllen = 0;
 484                 msg.msg_flags = msg_flags | MSG_NOSIGNAL;
 485
 486                 if (send)
 487                         result = sock_sendmsg(sock, &msg);
 488                 else
 489                         result = sock_recvmsg(sock, &msg, msg.msg_flags);
 490
 491                 if (result <= 0) {
 492                         if (result == 0)
 493                                 result = -EPIPE; /* short read */
 494                         break;
 495                 }
 496                 if (sent)
 497                         *sent += result;
 498         } while (msg_data_left(&msg));
 499
 500         memalloc_noreclaim_restore(noreclaim_flag);
 501
 502         return result;
 503 }
 504
 505 /*
 506  * Different settings for sk->sk_sndtimeo can result in different return values
 507  * if there is a signal pending when we enter sendmsg, because reasons?
 508  */
 509 static inline int was_interrupted(int result)
 510 {
 511         return result == -ERESTARTSYS || result == -EINTR;
 512 }
 513
 514 /* always call with the tx_lock held */
 515 static int nbd_send_cmd(struct nbd_device *nbd, struct nbd_cmd *cmd, int index)
 516 {
 517         struct request *req = blk_mq_rq_from_pdu(cmd);
 518         struct nbd_config *config = nbd->config;
 519         struct nbd_sock *nsock = config->socks[index];
 520         int result;
 521         struct nbd_request request = {.magic = htonl(NBD_REQUEST_MAGIC)};
 522         struct kvec iov = {.iov_base = &request, .iov_len = sizeof(request)};
 523         struct iov_iter from;
 524         unsigned long size = blk_rq_bytes(req);
 525         struct bio *bio;
 526         u64 handle;
 527         u32 type;
 528         u32 nbd_cmd_flags = 0;
 529         int sent = nsock->sent, skip = 0;
 530
 531         iov_iter_kvec(&from, WRITE, &iov, 1, sizeof(request));
 532
 533         type = req_to_nbd_cmd_type(req);
 534         if (type == U32_MAX)
 535                 return -EIO;
 536
 537         if (rq_data_dir(req) == WRITE &&
 538             (config->flags & NBD_FLAG_READ_ONLY)) {
 539                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 540                                     "Write on read-only\n");
 541                 return -EIO;
 542         }
 543
 544         if (req->cmd_flags & REQ_FUA)
 545                 nbd_cmd_flags |= NBD_CMD_FLAG_FUA;
 546
 547         /* We did a partial send previously, and we at least sent the whole
 548          * request struct, so just go and send the rest of the pages in the
 549          * request.
 550          */
 551         if (sent) {
 552                 if (sent >= sizeof(request)) {
 553                         skip = sent - sizeof(request);
 554
 555                         /* initialize handle for tracing purposes */
 556                         handle = nbd_cmd_handle(cmd);
 557
 558                         goto send_pages;
 559                 }
 560                 iov_iter_advance(&from, sent);
 561         } else {
 562                 cmd->cmd_cookie++;
 563         }
 564         cmd->index = index;
 565         cmd->cookie = nsock->cookie;
 566         cmd->retries = 0;
 567         request.type = htonl(type | nbd_cmd_flags);
 568         if (type != NBD_CMD_FLUSH) {
 569                 request.from = cpu_to_be64((u64)blk_rq_pos(req) << 9);
 570                 request.len = htonl(size);
 571         }
 572         handle = nbd_cmd_handle(cmd);
 573         memcpy(request.handle, &handle, sizeof(handle));
 574
 575         trace_nbd_send_request(&request, nbd->index, blk_mq_rq_from_pdu(cmd));
 576
 577         dev_dbg(nbd_to_dev(nbd), "request %p: sending control (%s@%llu,%uB)\n",
 578                 req, nbdcmd_to_ascii(type),
 579                 (unsigned long long)blk_rq_pos(req) << 9, blk_rq_bytes(req));
 580         result = sock_xmit(nbd, index, 1, &from,
 581                         (type == NBD_CMD_WRITE) ? MSG_MORE : 0, &sent);
 582         trace_nbd_header_sent(req, handle);
 583         if (result <= 0) {
 584                 if (was_interrupted(result)) {
 585                         /* If we havne't sent anything we can just return BUSY,
 586                          * however if we have sent something we need to make
 587                          * sure we only allow this req to be sent until we are
 588                          * completely done.
 589                          */
 590                         if (sent) {
 591                                 nsock->pending = req;
 592                                 nsock->sent = sent;
 593                         }
 594                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 595                         return BLK_STS_RESOURCE;
 596                 }
 597                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 598                         "Send control failed (result %d)\n", result);
 599                 return -EAGAIN;
 600         }
 601 send_pages:
 602         if (type != NBD_CMD_WRITE)
 603                 goto out;
 604
 605         bio = req->bio;
 606         while (bio) {
 607                 struct bio *next = bio->bi_next;
 608                 struct bvec_iter iter;
 609                 struct bio_vec bvec;
 610
 611                 bio_for_each_segment(bvec, bio, iter) {
 612                         bool is_last = !next && bio_iter_last(bvec, iter);
 613                         int flags = is_last ? 0 : MSG_MORE;
 614
 615                         dev_dbg(nbd_to_dev(nbd), "request %p: sending %d bytes data\n",
 616                                 req, bvec.bv_len);
 617                         iov_iter_bvec(&from, WRITE, &bvec, 1, bvec.bv_len);
 618                         if (skip) {
 619                                 if (skip >= iov_iter_count(&from)) {
 620                                         skip -= iov_iter_count(&from);
 621                                         continue;
 622                                 }
 623                                 iov_iter_advance(&from, skip);
 624                                 skip = 0;
 625                         }
 626                         result = sock_xmit(nbd, index, 1, &from, flags, &sent);
 627                         if (result <= 0) {
 628                                 if (was_interrupted(result)) {
 629                                         /* We've already sent the header, we
 630                                          * have no choice but to set pending and
 631                                          * return BUSY.
 632                                          */
 633                                         nsock->pending = req;
 634                                         nsock->sent = sent;
 635                                         set_bit(NBD_CMD_REQUEUED, &cmd->flags);
 636                                         return BLK_STS_RESOURCE;
 637                                 }
 638                                 dev_err(disk_to_dev(nbd->disk),
 639                                         "Send data failed (result %d)\n",
 640                                         result);
 641                                 return -EAGAIN;
 642                         }
 643                         /*
 644                          * The completion might already have come in,
 645                          * so break for the last one instead of letting
 646                          * the iterator do it. This prevents use-after-free
 647                          * of the bio.
 648                          */
 649                         if (is_last)
 650                                 break;
 651                 }
 652                 bio = next;
 653         }
 654 out:
 655         trace_nbd_payload_sent(req, handle);
 656         nsock->pending = NULL;
 657         nsock->sent = 0;
 658         return 0;
 659 }
 660
 661 /* NULL returned = something went wrong, inform userspace */
 662 static struct nbd_cmd *nbd_read_stat(struct nbd_device *nbd, int index)
 663 {
 664         struct nbd_config *config = nbd->config;
 665         int result;
 666         struct nbd_reply reply;
 667         struct nbd_cmd *cmd;
 668         struct request *req = NULL;
 669         u64 handle;
 670         u16 hwq;
 671         u32 tag;
 672         struct kvec iov = {.iov_base = &reply, .iov_len = sizeof(reply)};
 673         struct iov_iter to;
 674         int ret = 0;
 675
 676         reply.magic = 0;
 677         iov_iter_kvec(&to, READ, &iov, 1, sizeof(reply));
 678         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 679         if (result <= 0) {
 680                 if (!nbd_disconnected(config))
 681                         dev_err(disk_to_dev(nbd->disk),
 682                                 "Receive control failed (result %d)\n", result);
 683                 return ERR_PTR(result);
 684         }
 685
 686         if (ntohl(reply.magic) != NBD_REPLY_MAGIC) {
 687                 dev_err(disk_to_dev(nbd->disk), "Wrong magic (0x%lx)\n",
 688                                 (unsigned long)ntohl(reply.magic));
 689                 return ERR_PTR(-EPROTO);
 690         }
 691
 692         memcpy(&handle, reply.handle, sizeof(handle));
 693         tag = nbd_handle_to_tag(handle);
 694         hwq = blk_mq_unique_tag_to_hwq(tag);
 695         if (hwq < nbd->tag_set.nr_hw_queues)
 696                 req = blk_mq_tag_to_rq(nbd->tag_set.tags[hwq],
 697                                        blk_mq_unique_tag_to_tag(tag));
 698         if (!req || !blk_mq_request_started(req)) {
 699                 dev_err(disk_to_dev(nbd->disk), "Unexpected reply (%d) %p\n",
 700                         tag, req);
 701                 return ERR_PTR(-ENOENT);
 702         }
 703         trace_nbd_header_received(req, handle);
 704         cmd = blk_mq_rq_to_pdu(req);
 705
 706         mutex_lock(&cmd->lock);
 707         if (cmd->cmd_cookie != nbd_handle_to_cookie(handle)) {
 708                 dev_err(disk_to_dev(nbd->disk), "Double reply on req %p, cmd_cookie %u, handle cookie %u\n",
 709                         req, cmd->cmd_cookie, nbd_handle_to_cookie(handle));
 710                 ret = -ENOENT;
 711                 goto out;
 712         }
 713         if (cmd->status != BLK_STS_OK) {
 714                 dev_err(disk_to_dev(nbd->disk), "Command already handled %p\n",
 715                         req);
 716                 ret = -ENOENT;
 717                 goto out;
 718         }
 719         if (test_bit(NBD_CMD_REQUEUED, &cmd->flags)) {
 720                 dev_err(disk_to_dev(nbd->disk), "Raced with timeout on req %p\n",
 721                         req);
 722                 ret = -ENOENT;
 723                 goto out;
 724         }
 725         if (ntohl(reply.error)) {
 726                 dev_err(disk_to_dev(nbd->disk), "Other side returned error (%d)\n",
 727                         ntohl(reply.error));
 728                 cmd->status = BLK_STS_IOERR;
 729                 goto out;
 730         }
 731
 732         dev_dbg(nbd_to_dev(nbd), "request %p: got reply\n", req);
 733         if (rq_data_dir(req) != WRITE) {
 734                 struct req_iterator iter;
 735                 struct bio_vec bvec;
 736
 737                 rq_for_each_segment(bvec, req, iter) {
 738                         iov_iter_bvec(&to, READ, &bvec, 1, bvec.bv_len);
 739                         result = sock_xmit(nbd, index, 0, &to, MSG_WAITALL, NULL);
 740                         if (result <= 0) {
 741                                 dev_err(disk_to_dev(nbd->disk), "Receive data failed (result %d)\n",
 742                                         result);
 743                                 /*
 744                                  * If we've disconnected or we only have 1
 745                                  * connection then we need to make sure we
 746                                  * complete this request, otherwise error out
 747                                  * and let the timeout stuff handle resubmitting
 748                                  * this request onto another connection.
 749                                  */
 750                                 if (nbd_disconnected(config) ||
 751                                     config->num_connections <= 1) {
 752                                         cmd->status = BLK_STS_IOERR;
 753                                         goto out;
 754                                 }
 755                                 ret = -EIO;
 756                                 goto out;
 757                         }
 758                         dev_dbg(nbd_to_dev(nbd), "request %p: got %d bytes data\n",
 759                                 req, bvec.bv_len);
 760                 }
 761         }
 762 out:
 763         trace_nbd_payload_received(req, handle);
 764         mutex_unlock(&cmd->lock);
 765         return ret ? ERR_PTR(ret) : cmd;
 766 }
 767
 768 static void recv_work(struct work_struct *work)
 769 {
 770         struct recv_thread_args *args = container_of(work,
 771                                                      struct recv_thread_args,
 772                                                      work);
 773         struct nbd_device *nbd = args->nbd;
 774         struct nbd_config *config = nbd->config;
 775         struct nbd_cmd *cmd;
 776
 777         while (1) {
 778                 cmd = nbd_read_stat(nbd, args->index);
 779                 if (IS_ERR(cmd)) {
 780                         struct nbd_sock *nsock = config->socks[args->index];
 781
 782                         mutex_lock(&nsock->tx_lock);
 783                         nbd_mark_nsock_dead(nbd, nsock, 1);
 784                         mutex_unlock(&nsock->tx_lock);
 785                         break;
 786                 }
 787
 788                 blk_mq_complete_request(blk_mq_rq_from_pdu(cmd));
 789         }
 790         nbd_config_put(nbd);
 791         atomic_dec(&config->recv_threads);
 792         wake_up(&config->recv_wq);
 793         kfree(args);
 794 }
 795
 796 static bool nbd_clear_req(struct request *req, void *data, bool reserved)
 797 {
 798         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(req);
 799
 800         /* don't abort one completed request */
 801         if (blk_mq_request_completed(req))
 802                 return true;
 803
 804         mutex_lock(&cmd->lock);
 805         cmd->status = BLK_STS_IOERR;
 806         mutex_unlock(&cmd->lock);
 807
 808         blk_mq_complete_request(req);
 809         return true;
 810 }
 811
 812 static void nbd_clear_que(struct nbd_device *nbd)
 813 {
 814         blk_mq_quiesce_queue(nbd->disk->queue);
 815         blk_mq_tagset_busy_iter(&nbd->tag_set, nbd_clear_req, NULL);
 816         blk_mq_unquiesce_queue(nbd->disk->queue);
 817         dev_dbg(disk_to_dev(nbd->disk), "queue cleared\n");
 818 }
 819
 820 static int find_fallback(struct nbd_device *nbd, int index)
 821 {
 822         struct nbd_config *config = nbd->config;
 823         int new_index = -1;
 824         struct nbd_sock *nsock = config->socks[index];
 825         int fallback = nsock->fallback_index;
 826
 827         if (test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags))
 828                 return new_index;
 829
 830         if (config->num_connections <= 1) {
 831                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 832                                     "Attempted send on invalid socket\n");
 833                 return new_index;
 834         }
 835
 836         if (fallback >= 0 && fallback < config->num_connections &&
 837             !config->socks[fallback]->dead)
 838                 return fallback;
 839
 840         if (nsock->fallback_index < 0 ||
 841             nsock->fallback_index >= config->num_connections ||
 842             config->socks[nsock->fallback_index]->dead) {
 843                 int i;
 844                 for (i = 0; i < config->num_connections; i++) {
 845                         if (i == index)
 846                                 continue;
 847                         if (!config->socks[i]->dead) {
 848                                 new_index = i;
 849                                 break;
 850                         }
 851                 }
 852                 nsock->fallback_index = new_index;
 853                 if (new_index < 0) {
 854                         dev_err_ratelimited(disk_to_dev(nbd->disk),
 855                                             "Dead connection, failed to find a fallback\n");
 856                         return new_index;
 857                 }
 858         }
 859         new_index = nsock->fallback_index;
 860         return new_index;
 861 }
 862
 863 static int wait_for_reconnect(struct nbd_device *nbd)
 864 {
 865         struct nbd_config *config = nbd->config;
 866         if (!config->dead_conn_timeout)
 867                 return 0;
 868
 869         if (!wait_event_timeout(config->conn_wait,
 870                                 test_bit(NBD_RT_DISCONNECTED,
 871                                          &config->runtime_flags) ||
 872                                 atomic_read(&config->live_connections) > 0,
 873                                 config->dead_conn_timeout))
 874                 return 0;
 875
 876         return !test_bit(NBD_RT_DISCONNECTED, &config->runtime_flags);
 877 }
 878
 879 static int nbd_handle_cmd(struct nbd_cmd *cmd, int index)
 880 {
 881         struct request *req = blk_mq_rq_from_pdu(cmd);
 882         struct nbd_device *nbd = cmd->nbd;
 883         struct nbd_config *config;
 884         struct nbd_sock *nsock;
 885         int ret;
 886
 887         if (!refcount_inc_not_zero(&nbd->config_refs)) {
 888                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 889                                     "Socks array is empty\n");
 890                 blk_mq_start_request(req);
 891                 return -EINVAL;
 892         }
 893         config = nbd->config;
 894
 895         if (index >= config->num_connections) {
 896                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 897                                     "Attempted send on invalid socket\n");
 898                 nbd_config_put(nbd);
 899                 blk_mq_start_request(req);
 900                 return -EINVAL;
 901         }
 902         cmd->status = BLK_STS_OK;
 903 again:
 904         nsock = config->socks[index];
 905         mutex_lock(&nsock->tx_lock);
 906         if (nsock->dead) {
 907                 int old_index = index;
 908                 index = find_fallback(nbd, index);
 909                 mutex_unlock(&nsock->tx_lock);
 910                 if (index < 0) {
 911                         if (wait_for_reconnect(nbd)) {
 912                                 index = old_index;
 913                                 goto again;
 914                         }
 915                         /* All the sockets should already be down at this point,
 916                          * we just want to make sure that DISCONNECTED is set so
 917                          * any requests that come in that were queue'ed waiting
 918                          * for the reconnect timer don't trigger the timer again
 919                          * and instead just error out.
 920                          */
 921                         sock_shutdown(nbd);
 922                         nbd_config_put(nbd);
 923                         blk_mq_start_request(req);
 924                         return -EIO;
 925                 }
 926                 goto again;
 927         }
 928
 929         /* Handle the case that we have a pending request that was partially
 930          * transmitted that _has_ to be serviced first.  We need to call requeue
 931          * here so that it gets put _after_ the request that is already on the
 932          * dispatch list.
 933          */
 934         blk_mq_start_request(req);
 935         if (unlikely(nsock->pending && nsock->pending != req)) {
 936                 nbd_requeue_cmd(cmd);
 937                 ret = 0;
 938                 goto out;
 939         }
 940         /*
 941          * Some failures are related to the link going down, so anything that
 942          * returns EAGAIN can be retried on a different socket.
 943          */
 944         ret = nbd_send_cmd(nbd, cmd, index);
 945         if (ret == -EAGAIN) {
 946                 dev_err_ratelimited(disk_to_dev(nbd->disk),
 947                                     "Request send failed, requeueing\n");
 948                 nbd_mark_nsock_dead(nbd, nsock, 1);
 949                 nbd_requeue_cmd(cmd);
 950                 ret = 0;
 951         }
 952 out:
 953         mutex_unlock(&nsock->tx_lock);
 954         nbd_config_put(nbd);
 955         return ret;
 956 }
 957
 958 static blk_status_t nbd_queue_rq(struct blk_mq_hw_ctx *hctx,
 959                         const struct blk_mq_queue_data *bd)
 960 {
 961         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(bd->rq);
 962         int ret;
 963
 964         /*
 965          * Since we look at the bio's to send the request over the network we
 966          * need to make sure the completion work doesn't mark this request done
 967          * before we are done doing our send.  This keeps us from dereferencing
 968          * freed data if we have particularly fast completions (ie we get the
 969          * completion before we exit sock_xmit on the last bvec) or in the case
 970          * that the server is misbehaving (or there was an error) before we're
 971          * done sending everything over the wire.
 972          */
 973         mutex_lock(&cmd->lock);
 974         clear_bit(NBD_CMD_REQUEUED, &cmd->flags);
 975
 976         /* We can be called directly from the user space process, which means we
 977          * could possibly have signals pending so our sendmsg will fail.  In
 978          * this case we need to return that we are busy, otherwise error out as
 979          * appropriate.
 980          */
 981         ret = nbd_handle_cmd(cmd, hctx->queue_num);
 982         if (ret < 0)
 983                 ret = BLK_STS_IOERR;
 984         else if (!ret)
 985                 ret = BLK_STS_OK;
 986         mutex_unlock(&cmd->lock);
 987
 988         return ret;
 989 }
 990
 991 static struct socket *nbd_get_socket(struct nbd_device *nbd, unsigned long fd,
 992                                      int *err)
 993 {
 994         struct socket *sock;
 995
 996         *err = 0;
 997         sock = sockfd_lookup(fd, err);
 998         if (!sock)
 999                 return NULL;
1000
1001         if (sock->ops->shutdown == sock_no_shutdown) {
1002                 dev_err(disk_to_dev(nbd->disk), "Unsupported socket: shutdown callout must be supported.\n");
1003                 *err = -EINVAL;
1004                 sockfd_put(sock);
1005                 return NULL;
1006         }
1007
1008         return sock;
1009 }
1010
1011 static int nbd_add_socket(struct nbd_device *nbd, unsigned long arg,
1012                           bool netlink)
1013 {
1014         struct nbd_config *config = nbd->config;
1015         struct socket *sock;
1016         struct nbd_sock **socks;
1017         struct nbd_sock *nsock;
1018         int err;
1019
1020         sock = nbd_get_socket(nbd, arg, &err);
1021         if (!sock)
1022                 return err;
1023
1024         /*
1025          * We need to make sure we don't get any errant requests while we're
1026          * reallocating the ->socks array.
1027          */
1028         blk_mq_freeze_queue(nbd->disk->queue);
1029
1030         if (!netlink && !nbd->task_setup &&
1031             !test_bit(NBD_RT_BOUND, &config->runtime_flags))
1032                 nbd->task_setup = current;
1033
1034         if (!netlink &&
1035             (nbd->task_setup != current ||
1036              test_bit(NBD_RT_BOUND, &config->runtime_flags))) {
1037                 dev_err(disk_to_dev(nbd->disk),
1038                         "Device being setup by another task");
1039                 err = -EBUSY;
1040                 goto put_socket;
1041         }
1042
1043         nsock = kzalloc(sizeof(*nsock), GFP_KERNEL);
1044         if (!nsock) {
1045                 err = -ENOMEM;
1046                 goto put_socket;
1047         }
1048
1049         socks = krealloc(config->socks, (config->num_connections + 1) *
1050                          sizeof(struct nbd_sock *), GFP_KERNEL);
1051         if (!socks) {
1052                 kfree(nsock);
1053                 err = -ENOMEM;
1054                 goto put_socket;
1055         }
1056
1057         config->socks = socks;
1058
1059         nsock->fallback_index = -1;
1060         nsock->dead = false;
1061         mutex_init(&nsock->tx_lock);
1062         nsock->sock = sock;
1063         nsock->pending = NULL;
1064         nsock->sent = 0;
1065         nsock->cookie = 0;
1066         socks[config->num_connections++] = nsock;
1067         atomic_inc(&config->live_connections);
1068         blk_mq_unfreeze_queue(nbd->disk->queue);
1069
1070         return 0;
1071
1072 put_socket:
1073         blk_mq_unfreeze_queue(nbd->disk->queue);
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_size(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                 bdev->bd_invalidated = 1;
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                 nbd_clear_que(nbd);
1333         }
1334
1335         flush_workqueue(nbd->recv_workq);
1336         mutex_lock(&nbd->config_lock);
1337         nbd_bdev_reset(bdev);
1338         /* user requested, ignore socket errors */
1339         if (test_bit(NBD_RT_DISCONNECT_REQUESTED, &config->runtime_flags))
1340                 ret = 0;
1341         if (test_bit(NBD_RT_TIMEDOUT, &config->runtime_flags))
1342                 ret = -ETIMEDOUT;
1343         return ret;
1344 }
1345
1346 static void nbd_clear_sock_ioctl(struct nbd_device *nbd,
1347                                  struct block_device *bdev)
1348 {
1349         nbd_clear_sock(nbd);
1350         __invalidate_device(bdev, true);
1351         nbd_bdev_reset(bdev);
1352         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
1353                                &nbd->config->runtime_flags))
1354                 nbd_config_put(nbd);
1355 }
1356
1357 static bool nbd_is_valid_blksize(unsigned long blksize)
1358 {
1359         if (!blksize || !is_power_of_2(blksize) || blksize < 512 ||
1360             blksize > PAGE_SIZE)
1361                 return false;
1362         return true;
1363 }
1364
1365 static void nbd_set_cmd_timeout(struct nbd_device *nbd, u64 timeout)
1366 {
1367         nbd->tag_set.timeout = timeout * HZ;
1368         if (timeout)
1369                 blk_queue_rq_timeout(nbd->disk->queue, timeout * HZ);
1370         else
1371                 blk_queue_rq_timeout(nbd->disk->queue, 30 * HZ);
1372 }
1373
1374 /* Must be called with config_lock held */
1375 static int __nbd_ioctl(struct block_device *bdev, struct nbd_device *nbd,
1376                        unsigned int cmd, unsigned long arg)
1377 {
1378         struct nbd_config *config = nbd->config;
1379
1380         switch (cmd) {
1381         case NBD_DISCONNECT:
1382                 return nbd_disconnect(nbd);
1383         case NBD_CLEAR_SOCK:
1384                 nbd_clear_sock_ioctl(nbd, bdev);
1385                 return 0;
1386         case NBD_SET_SOCK:
1387                 return nbd_add_socket(nbd, arg, false);
1388         case NBD_SET_BLKSIZE:
1389                 if (!arg)
1390                         arg = NBD_DEF_BLKSIZE;
1391                 if (!nbd_is_valid_blksize(arg))
1392                         return -EINVAL;
1393                 nbd_size_set(nbd, arg,
1394                              div_s64(config->bytesize, arg));
1395                 return 0;
1396         case NBD_SET_SIZE:
1397                 nbd_size_set(nbd, config->blksize,
1398                              div_s64(arg, config->blksize));
1399                 return 0;
1400         case NBD_SET_SIZE_BLOCKS:
1401                 nbd_size_set(nbd, config->blksize, arg);
1402                 return 0;
1403         case NBD_SET_TIMEOUT:
1404                 nbd_set_cmd_timeout(nbd, arg);
1405                 return 0;
1406
1407         case NBD_SET_FLAGS:
1408                 config->flags = arg;
1409                 return 0;
1410         case NBD_DO_IT:
1411                 return nbd_start_device_ioctl(nbd, bdev);
1412         case NBD_CLEAR_QUE:
1413                 /*
1414                  * This is for compatibility only.  The queue is always cleared
1415                  * by NBD_DO_IT or NBD_CLEAR_SOCK.
1416                  */
1417                 return 0;
1418         case NBD_PRINT_DEBUG:
1419                 /*
1420                  * For compatibility only, we no longer keep a list of
1421                  * outstanding requests.
1422                  */
1423                 return 0;
1424         }
1425         return -ENOTTY;
1426 }
1427
1428 static int nbd_ioctl(struct block_device *bdev, fmode_t mode,
1429                      unsigned int cmd, unsigned long arg)
1430 {
1431         struct nbd_device *nbd = bdev->bd_disk->private_data;
1432         struct nbd_config *config = nbd->config;
1433         int error = -EINVAL;
1434
1435         if (!capable(CAP_SYS_ADMIN))
1436                 return -EPERM;
1437
1438         /* The block layer will pass back some non-nbd ioctls in case we have
1439          * special handling for them, but we don't so just return an error.
1440          */
1441         if (_IOC_TYPE(cmd) != 0xab)
1442                 return -EINVAL;
1443
1444         mutex_lock(&nbd->config_lock);
1445
1446         /* Don't allow ioctl operations on a nbd device that was created with
1447          * netlink, unless it's DISCONNECT or CLEAR_SOCK, which are fine.
1448          */
1449         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
1450             (cmd == NBD_DISCONNECT || cmd == NBD_CLEAR_SOCK))
1451                 error = __nbd_ioctl(bdev, nbd, cmd, arg);
1452         else
1453                 dev_err(nbd_to_dev(nbd), "Cannot use ioctl interface on a netlink controlled device.\n");
1454         mutex_unlock(&nbd->config_lock);
1455         return error;
1456 }
1457
1458 static struct nbd_config *nbd_alloc_config(void)
1459 {
1460         struct nbd_config *config;
1461
1462         if (!try_module_get(THIS_MODULE))
1463                 return ERR_PTR(-ENODEV);
1464
1465         config = kzalloc(sizeof(struct nbd_config), GFP_NOFS);
1466         if (!config) {
1467                 module_put(THIS_MODULE);
1468                 return ERR_PTR(-ENOMEM);
1469         }
1470
1471         atomic_set(&config->recv_threads, 0);
1472         init_waitqueue_head(&config->recv_wq);
1473         init_waitqueue_head(&config->conn_wait);
1474         config->blksize = NBD_DEF_BLKSIZE;
1475         atomic_set(&config->live_connections, 0);
1476         return config;
1477 }
1478
1479 static int nbd_open(struct block_device *bdev, fmode_t mode)
1480 {
1481         struct nbd_device *nbd;
1482         int ret = 0;
1483
1484         mutex_lock(&nbd_index_mutex);
1485         nbd = bdev->bd_disk->private_data;
1486         if (!nbd) {
1487                 ret = -ENXIO;
1488                 goto out;
1489         }
1490         if (!refcount_inc_not_zero(&nbd->refs)) {
1491                 ret = -ENXIO;
1492                 goto out;
1493         }
1494         if (!refcount_inc_not_zero(&nbd->config_refs)) {
1495                 struct nbd_config *config;
1496
1497                 mutex_lock(&nbd->config_lock);
1498                 if (refcount_inc_not_zero(&nbd->config_refs)) {
1499                         mutex_unlock(&nbd->config_lock);
1500                         goto out;
1501                 }
1502                 config = nbd_alloc_config();
1503                 if (IS_ERR(config)) {
1504                         ret = PTR_ERR(config);
1505                         mutex_unlock(&nbd->config_lock);
1506                         goto out;
1507                 }
1508                 nbd->config = config;
1509                 refcount_set(&nbd->config_refs, 1);
1510                 refcount_inc(&nbd->refs);
1511                 mutex_unlock(&nbd->config_lock);
1512                 bdev->bd_invalidated = 1;
1513         } else if (nbd_disconnected(nbd->config)) {
1514                 bdev->bd_invalidated = 1;
1515         }
1516 out:
1517         mutex_unlock(&nbd_index_mutex);
1518         return ret;
1519 }
1520
1521 static void nbd_release(struct gendisk *disk, fmode_t mode)
1522 {
1523         struct nbd_device *nbd = disk->private_data;
1524         struct block_device *bdev = bdget_disk(disk, 0);
1525
1526         if (test_bit(NBD_RT_DISCONNECT_ON_CLOSE, &nbd->config->runtime_flags) &&
1527                         bdev->bd_openers == 0)
1528                 nbd_disconnect_and_put(nbd);
1529         bdput(bdev);
1530
1531         nbd_config_put(nbd);
1532         nbd_put(nbd);
1533 }
1534
1535 static const struct block_device_operations nbd_fops =
1536 {
1537         .owner =        THIS_MODULE,
1538         .open =         nbd_open,
1539         .release =      nbd_release,
1540         .ioctl =        nbd_ioctl,
1541         .compat_ioctl = nbd_ioctl,
1542 };
1543
1544 #if IS_ENABLED(CONFIG_DEBUG_FS)
1545
1546 static int nbd_dbg_tasks_show(struct seq_file *s, void *unused)
1547 {
1548         struct nbd_device *nbd = s->private;
1549
1550         if (nbd->task_recv)
1551                 seq_printf(s, "recv: %d\n", task_pid_nr(nbd->task_recv));
1552
1553         return 0;
1554 }
1555
1556 static int nbd_dbg_tasks_open(struct inode *inode, struct file *file)
1557 {
1558         return single_open(file, nbd_dbg_tasks_show, inode->i_private);
1559 }
1560
1561 static const struct file_operations nbd_dbg_tasks_ops = {
1562         .open = nbd_dbg_tasks_open,
1563         .read = seq_read,
1564         .llseek = seq_lseek,
1565         .release = single_release,
1566 };
1567
1568 static int nbd_dbg_flags_show(struct seq_file *s, void *unused)
1569 {
1570         struct nbd_device *nbd = s->private;
1571         u32 flags = nbd->config->flags;
1572
1573         seq_printf(s, "Hex: 0x%08x\n\n", flags);
1574
1575         seq_puts(s, "Known flags:\n");
1576
1577         if (flags & NBD_FLAG_HAS_FLAGS)
1578                 seq_puts(s, "NBD_FLAG_HAS_FLAGS\n");
1579         if (flags & NBD_FLAG_READ_ONLY)
1580                 seq_puts(s, "NBD_FLAG_READ_ONLY\n");
1581         if (flags & NBD_FLAG_SEND_FLUSH)
1582                 seq_puts(s, "NBD_FLAG_SEND_FLUSH\n");
1583         if (flags & NBD_FLAG_SEND_FUA)
1584                 seq_puts(s, "NBD_FLAG_SEND_FUA\n");
1585         if (flags & NBD_FLAG_SEND_TRIM)
1586                 seq_puts(s, "NBD_FLAG_SEND_TRIM\n");
1587
1588         return 0;
1589 }
1590
1591 static int nbd_dbg_flags_open(struct inode *inode, struct file *file)
1592 {
1593         return single_open(file, nbd_dbg_flags_show, inode->i_private);
1594 }
1595
1596 static const struct file_operations nbd_dbg_flags_ops = {
1597         .open = nbd_dbg_flags_open,
1598         .read = seq_read,
1599         .llseek = seq_lseek,
1600         .release = single_release,
1601 };
1602
1603 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1604 {
1605         struct dentry *dir;
1606         struct nbd_config *config = nbd->config;
1607
1608         if (!nbd_dbg_dir)
1609                 return -EIO;
1610
1611         dir = debugfs_create_dir(nbd_name(nbd), nbd_dbg_dir);
1612         if (!dir) {
1613                 dev_err(nbd_to_dev(nbd), "Failed to create debugfs dir for '%s'\n",
1614                         nbd_name(nbd));
1615                 return -EIO;
1616         }
1617         config->dbg_dir = dir;
1618
1619         debugfs_create_file("tasks", 0444, dir, nbd, &nbd_dbg_tasks_ops);
1620         debugfs_create_u64("size_bytes", 0444, dir, &config->bytesize);
1621         debugfs_create_u32("timeout", 0444, dir, &nbd->tag_set.timeout);
1622         debugfs_create_u64("blocksize", 0444, dir, &config->blksize);
1623         debugfs_create_file("flags", 0444, dir, nbd, &nbd_dbg_flags_ops);
1624
1625         return 0;
1626 }
1627
1628 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1629 {
1630         debugfs_remove_recursive(nbd->config->dbg_dir);
1631 }
1632
1633 static int nbd_dbg_init(void)
1634 {
1635         struct dentry *dbg_dir;
1636
1637         dbg_dir = debugfs_create_dir("nbd", NULL);
1638         if (!dbg_dir)
1639                 return -EIO;
1640
1641         nbd_dbg_dir = dbg_dir;
1642
1643         return 0;
1644 }
1645
1646 static void nbd_dbg_close(void)
1647 {
1648         debugfs_remove_recursive(nbd_dbg_dir);
1649 }
1650
1651 #else  /* IS_ENABLED(CONFIG_DEBUG_FS) */
1652
1653 static int nbd_dev_dbg_init(struct nbd_device *nbd)
1654 {
1655         return 0;
1656 }
1657
1658 static void nbd_dev_dbg_close(struct nbd_device *nbd)
1659 {
1660 }
1661
1662 static int nbd_dbg_init(void)
1663 {
1664         return 0;
1665 }
1666
1667 static void nbd_dbg_close(void)
1668 {
1669 }
1670
1671 #endif
1672
1673 static int nbd_init_request(struct blk_mq_tag_set *set, struct request *rq,
1674                             unsigned int hctx_idx, unsigned int numa_node)
1675 {
1676         struct nbd_cmd *cmd = blk_mq_rq_to_pdu(rq);
1677         cmd->nbd = set->driver_data;
1678         cmd->flags = 0;
1679         mutex_init(&cmd->lock);
1680         return 0;
1681 }
1682
1683 static const struct blk_mq_ops nbd_mq_ops = {
1684         .queue_rq       = nbd_queue_rq,
1685         .complete       = nbd_complete_rq,
1686         .init_request   = nbd_init_request,
1687         .timeout        = nbd_xmit_timeout,
1688 };
1689
1690 static int nbd_dev_add(int index)
1691 {
1692         struct nbd_device *nbd;
1693         struct gendisk *disk;
1694         struct request_queue *q;
1695         int err = -ENOMEM;
1696
1697         nbd = kzalloc(sizeof(struct nbd_device), GFP_KERNEL);
1698         if (!nbd)
1699                 goto out;
1700
1701         disk = alloc_disk(1 << part_shift);
1702         if (!disk)
1703                 goto out_free_nbd;
1704
1705         if (index >= 0) {
1706                 err = idr_alloc(&nbd_index_idr, nbd, index, index + 1,
1707                                 GFP_KERNEL);
1708                 if (err == -ENOSPC)
1709                         err = -EEXIST;
1710         } else {
1711                 err = idr_alloc(&nbd_index_idr, nbd, 0, 0, GFP_KERNEL);
1712                 if (err >= 0)
1713                         index = err;
1714         }
1715         if (err < 0)
1716                 goto out_free_disk;
1717
1718         nbd->index = index;
1719         nbd->disk = disk;
1720         nbd->tag_set.ops = &nbd_mq_ops;
1721         nbd->tag_set.nr_hw_queues = 1;
1722         nbd->tag_set.queue_depth = 128;
1723         nbd->tag_set.numa_node = NUMA_NO_NODE;
1724         nbd->tag_set.cmd_size = sizeof(struct nbd_cmd);
1725         nbd->tag_set.flags = BLK_MQ_F_SHOULD_MERGE |
1726                 BLK_MQ_F_BLOCKING;
1727         nbd->tag_set.driver_data = nbd;
1728         nbd->destroy_complete = NULL;
1729
1730         err = blk_mq_alloc_tag_set(&nbd->tag_set);
1731         if (err)
1732                 goto out_free_idr;
1733
1734         q = blk_mq_init_queue(&nbd->tag_set);
1735         if (IS_ERR(q)) {
1736                 err = PTR_ERR(q);
1737                 goto out_free_tags;
1738         }
1739         disk->queue = q;
1740
1741         /*
1742          * Tell the block layer that we are not a rotational device
1743          */
1744         blk_queue_flag_set(QUEUE_FLAG_NONROT, disk->queue);
1745         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, disk->queue);
1746         disk->queue->limits.discard_granularity = 0;
1747         disk->queue->limits.discard_alignment = 0;
1748         blk_queue_max_discard_sectors(disk->queue, 0);
1749         blk_queue_max_segment_size(disk->queue, UINT_MAX);
1750         blk_queue_max_segments(disk->queue, USHRT_MAX);
1751         blk_queue_max_hw_sectors(disk->queue, 65536);
1752         disk->queue->limits.max_sectors = 256;
1753
1754         mutex_init(&nbd->config_lock);
1755         refcount_set(&nbd->config_refs, 0);
1756         refcount_set(&nbd->refs, 1);
1757         INIT_LIST_HEAD(&nbd->list);
1758         disk->major = NBD_MAJOR;
1759         disk->first_minor = index << part_shift;
1760         disk->fops = &nbd_fops;
1761         disk->private_data = nbd;
1762         sprintf(disk->disk_name, "nbd%d", index);
1763         add_disk(disk);
1764         nbd_total_devices++;
1765         return index;
1766
1767 out_free_tags:
1768         blk_mq_free_tag_set(&nbd->tag_set);
1769 out_free_idr:
1770         idr_remove(&nbd_index_idr, index);
1771 out_free_disk:
1772         put_disk(disk);
1773 out_free_nbd:
1774         kfree(nbd);
1775 out:
1776         return err;
1777 }
1778
1779 static int find_free_cb(int id, void *ptr, void *data)
1780 {
1781         struct nbd_device *nbd = ptr;
1782         struct nbd_device **found = data;
1783
1784         if (!refcount_read(&nbd->config_refs)) {
1785                 *found = nbd;
1786                 return 1;
1787         }
1788         return 0;
1789 }
1790
1791 /* Netlink interface. */
1792 static const struct nla_policy nbd_attr_policy[NBD_ATTR_MAX + 1] = {
1793         [NBD_ATTR_INDEX]                =       { .type = NLA_U32 },
1794         [NBD_ATTR_SIZE_BYTES]           =       { .type = NLA_U64 },
1795         [NBD_ATTR_BLOCK_SIZE_BYTES]     =       { .type = NLA_U64 },
1796         [NBD_ATTR_TIMEOUT]              =       { .type = NLA_U64 },
1797         [NBD_ATTR_SERVER_FLAGS]         =       { .type = NLA_U64 },
1798         [NBD_ATTR_CLIENT_FLAGS]         =       { .type = NLA_U64 },
1799         [NBD_ATTR_SOCKETS]              =       { .type = NLA_NESTED},
1800         [NBD_ATTR_DEAD_CONN_TIMEOUT]    =       { .type = NLA_U64 },
1801         [NBD_ATTR_DEVICE_LIST]          =       { .type = NLA_NESTED},
1802 };
1803
1804 static const struct nla_policy nbd_sock_policy[NBD_SOCK_MAX + 1] = {
1805         [NBD_SOCK_FD]                   =       { .type = NLA_U32 },
1806 };
1807
1808 /* We don't use this right now since we don't parse the incoming list, but we
1809  * still want it here so userspace knows what to expect.
1810  */
1811 static const struct nla_policy __attribute__((unused))
1812 nbd_device_policy[NBD_DEVICE_ATTR_MAX + 1] = {
1813         [NBD_DEVICE_INDEX]              =       { .type = NLA_U32 },
1814         [NBD_DEVICE_CONNECTED]          =       { .type = NLA_U8 },
1815 };
1816
1817 static int nbd_genl_size_set(struct genl_info *info, struct nbd_device *nbd)
1818 {
1819         struct nbd_config *config = nbd->config;
1820         u64 bsize = config->blksize;
1821         u64 bytes = config->bytesize;
1822
1823         if (info->attrs[NBD_ATTR_SIZE_BYTES])
1824                 bytes = nla_get_u64(info->attrs[NBD_ATTR_SIZE_BYTES]);
1825
1826         if (info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]) {
1827                 bsize = nla_get_u64(info->attrs[NBD_ATTR_BLOCK_SIZE_BYTES]);
1828                 if (!bsize)
1829                         bsize = NBD_DEF_BLKSIZE;
1830                 if (!nbd_is_valid_blksize(bsize)) {
1831                         printk(KERN_ERR "Invalid block size %llu\n", bsize);
1832                         return -EINVAL;
1833                 }
1834         }
1835
1836         if (bytes != config->bytesize || bsize != config->blksize)
1837                 nbd_size_set(nbd, bsize, div64_u64(bytes, bsize));
1838         return 0;
1839 }
1840
1841 static int nbd_genl_connect(struct sk_buff *skb, struct genl_info *info)
1842 {
1843         DECLARE_COMPLETION_ONSTACK(destroy_complete);
1844         struct nbd_device *nbd = NULL;
1845         struct nbd_config *config;
1846         int index = -1;
1847         int ret;
1848         bool put_dev = false;
1849
1850         if (!netlink_capable(skb, CAP_SYS_ADMIN))
1851                 return -EPERM;
1852
1853         if (info->attrs[NBD_ATTR_INDEX])
1854                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
1855         if (!info->attrs[NBD_ATTR_SOCKETS]) {
1856                 printk(KERN_ERR "nbd: must specify at least one socket\n");
1857                 return -EINVAL;
1858         }
1859         if (!info->attrs[NBD_ATTR_SIZE_BYTES]) {
1860                 printk(KERN_ERR "nbd: must specify a size in bytes for the device\n");
1861                 return -EINVAL;
1862         }
1863 again:
1864         mutex_lock(&nbd_index_mutex);
1865         if (index == -1) {
1866                 ret = idr_for_each(&nbd_index_idr, &find_free_cb, &nbd);
1867                 if (ret == 0) {
1868                         int new_index;
1869                         new_index = nbd_dev_add(-1);
1870                         if (new_index < 0) {
1871                                 mutex_unlock(&nbd_index_mutex);
1872                                 printk(KERN_ERR "nbd: failed to add new device\n");
1873                                 return new_index;
1874                         }
1875                         nbd = idr_find(&nbd_index_idr, new_index);
1876                 }
1877         } else {
1878                 nbd = idr_find(&nbd_index_idr, index);
1879                 if (!nbd) {
1880                         ret = nbd_dev_add(index);
1881                         if (ret < 0) {
1882                                 mutex_unlock(&nbd_index_mutex);
1883                                 printk(KERN_ERR "nbd: failed to add new device\n");
1884                                 return ret;
1885                         }
1886                         nbd = idr_find(&nbd_index_idr, index);
1887                 }
1888         }
1889         if (!nbd) {
1890                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
1891                        index);
1892                 mutex_unlock(&nbd_index_mutex);
1893                 return -EINVAL;
1894         }
1895
1896         if (test_bit(NBD_DESTROY_ON_DISCONNECT, &nbd->flags) &&
1897             test_bit(NBD_DISCONNECT_REQUESTED, &nbd->flags)) {
1898                 nbd->destroy_complete = &destroy_complete;
1899                 mutex_unlock(&nbd_index_mutex);
1900
1901                 /* Wait untill the the nbd stuff is totally destroyed */
1902                 wait_for_completion(&destroy_complete);
1903                 goto again;
1904         }
1905
1906         if (!refcount_inc_not_zero(&nbd->refs)) {
1907                 mutex_unlock(&nbd_index_mutex);
1908                 if (index == -1)
1909                         goto again;
1910                 printk(KERN_ERR "nbd: device at index %d is going down\n",
1911                        index);
1912                 return -EINVAL;
1913         }
1914         mutex_unlock(&nbd_index_mutex);
1915
1916         mutex_lock(&nbd->config_lock);
1917         if (refcount_read(&nbd->config_refs)) {
1918                 mutex_unlock(&nbd->config_lock);
1919                 nbd_put(nbd);
1920                 if (index == -1)
1921                         goto again;
1922                 printk(KERN_ERR "nbd: nbd%d already in use\n", index);
1923                 return -EBUSY;
1924         }
1925         if (WARN_ON(nbd->config)) {
1926                 mutex_unlock(&nbd->config_lock);
1927                 nbd_put(nbd);
1928                 return -EINVAL;
1929         }
1930         config = nbd_alloc_config();
1931         if (IS_ERR(config)) {
1932                 mutex_unlock(&nbd->config_lock);
1933                 nbd_put(nbd);
1934                 printk(KERN_ERR "nbd: couldn't allocate config\n");
1935                 return PTR_ERR(config);
1936         }
1937         nbd->config = config;
1938         refcount_set(&nbd->config_refs, 1);
1939         set_bit(NBD_RT_BOUND, &config->runtime_flags);
1940
1941         ret = nbd_genl_size_set(info, nbd);
1942         if (ret)
1943                 goto out;
1944
1945         if (info->attrs[NBD_ATTR_TIMEOUT])
1946                 nbd_set_cmd_timeout(nbd,
1947                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
1948         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
1949                 config->dead_conn_timeout =
1950                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
1951                 config->dead_conn_timeout *= HZ;
1952         }
1953         if (info->attrs[NBD_ATTR_SERVER_FLAGS])
1954                 config->flags =
1955                         nla_get_u64(info->attrs[NBD_ATTR_SERVER_FLAGS]);
1956         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
1957                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
1958                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
1959                         /*
1960                          * We have 1 ref to keep the device around, and then 1
1961                          * ref for our current operation here, which will be
1962                          * inherited by the config.  If we already have
1963                          * DESTROY_ON_DISCONNECT set then we know we don't have
1964                          * that extra ref already held so we don't need the
1965                          * put_dev.
1966                          */
1967                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
1968                                               &nbd->flags))
1969                                 put_dev = true;
1970                 } else {
1971                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
1972                                                &nbd->flags))
1973                                 refcount_inc(&nbd->refs);
1974                 }
1975                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
1976                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
1977                                 &config->runtime_flags);
1978                 }
1979         }
1980
1981         if (info->attrs[NBD_ATTR_SOCKETS]) {
1982                 struct nlattr *attr;
1983                 int rem, fd;
1984
1985                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
1986                                     rem) {
1987                         struct nlattr *socks[NBD_SOCK_MAX+1];
1988
1989                         if (nla_type(attr) != NBD_SOCK_ITEM) {
1990                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
1991                                 ret = -EINVAL;
1992                                 goto out;
1993                         }
1994                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
1995                                                           attr,
1996                                                           nbd_sock_policy,
1997                                                           info->extack);
1998                         if (ret != 0) {
1999                                 printk(KERN_ERR "nbd: error processing sock list\n");
2000                                 ret = -EINVAL;
2001                                 goto out;
2002                         }
2003                         if (!socks[NBD_SOCK_FD])
2004                                 continue;
2005                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2006                         ret = nbd_add_socket(nbd, fd, true);
2007                         if (ret)
2008                                 goto out;
2009                 }
2010         }
2011         ret = nbd_start_device(nbd);
2012 out:
2013         mutex_unlock(&nbd->config_lock);
2014         if (!ret) {
2015                 set_bit(NBD_RT_HAS_CONFIG_REF, &config->runtime_flags);
2016                 refcount_inc(&nbd->config_refs);
2017                 nbd_connect_reply(info, nbd->index);
2018         }
2019         nbd_config_put(nbd);
2020         if (put_dev)
2021                 nbd_put(nbd);
2022         return ret;
2023 }
2024
2025 static void nbd_disconnect_and_put(struct nbd_device *nbd)
2026 {
2027         mutex_lock(&nbd->config_lock);
2028         nbd_disconnect(nbd);
2029         sock_shutdown(nbd);
2030         wake_up(&nbd->config->conn_wait);
2031         /*
2032          * Make sure recv thread has finished, so it does not drop the last
2033          * config ref and try to destroy the workqueue from inside the work
2034          * queue. And this also ensure that we can safely call nbd_clear_que()
2035          * to cancel the inflight I/Os.
2036          */
2037         if (nbd->recv_workq)
2038                 flush_workqueue(nbd->recv_workq);
2039         nbd_clear_que(nbd);
2040         nbd->task_setup = NULL;
2041         mutex_unlock(&nbd->config_lock);
2042
2043         if (test_and_clear_bit(NBD_RT_HAS_CONFIG_REF,
2044                                &nbd->config->runtime_flags))
2045                 nbd_config_put(nbd);
2046 }
2047
2048 static int nbd_genl_disconnect(struct sk_buff *skb, struct genl_info *info)
2049 {
2050         struct nbd_device *nbd;
2051         int index;
2052
2053         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2054                 return -EPERM;
2055
2056         if (!info->attrs[NBD_ATTR_INDEX]) {
2057                 printk(KERN_ERR "nbd: must specify an index to disconnect\n");
2058                 return -EINVAL;
2059         }
2060         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2061         mutex_lock(&nbd_index_mutex);
2062         nbd = idr_find(&nbd_index_idr, index);
2063         if (!nbd) {
2064                 mutex_unlock(&nbd_index_mutex);
2065                 printk(KERN_ERR "nbd: couldn't find device at index %d\n",
2066                        index);
2067                 return -EINVAL;
2068         }
2069         if (!refcount_inc_not_zero(&nbd->refs)) {
2070                 mutex_unlock(&nbd_index_mutex);
2071                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2072                        index);
2073                 return -EINVAL;
2074         }
2075         mutex_unlock(&nbd_index_mutex);
2076         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2077                 nbd_put(nbd);
2078                 return 0;
2079         }
2080         nbd_disconnect_and_put(nbd);
2081         nbd_config_put(nbd);
2082         nbd_put(nbd);
2083         return 0;
2084 }
2085
2086 static int nbd_genl_reconfigure(struct sk_buff *skb, struct genl_info *info)
2087 {
2088         struct nbd_device *nbd = NULL;
2089         struct nbd_config *config;
2090         int index;
2091         int ret = 0;
2092         bool put_dev = false;
2093
2094         if (!netlink_capable(skb, CAP_SYS_ADMIN))
2095                 return -EPERM;
2096
2097         if (!info->attrs[NBD_ATTR_INDEX]) {
2098                 printk(KERN_ERR "nbd: must specify a device to reconfigure\n");
2099                 return -EINVAL;
2100         }
2101         index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2102         mutex_lock(&nbd_index_mutex);
2103         nbd = idr_find(&nbd_index_idr, index);
2104         if (!nbd) {
2105                 mutex_unlock(&nbd_index_mutex);
2106                 printk(KERN_ERR "nbd: couldn't find a device at index %d\n",
2107                        index);
2108                 return -EINVAL;
2109         }
2110         if (!refcount_inc_not_zero(&nbd->refs)) {
2111                 mutex_unlock(&nbd_index_mutex);
2112                 printk(KERN_ERR "nbd: device at index %d is going down\n",
2113                        index);
2114                 return -EINVAL;
2115         }
2116         mutex_unlock(&nbd_index_mutex);
2117
2118         if (!refcount_inc_not_zero(&nbd->config_refs)) {
2119                 dev_err(nbd_to_dev(nbd),
2120                         "not configured, cannot reconfigure\n");
2121                 nbd_put(nbd);
2122                 return -EINVAL;
2123         }
2124
2125         mutex_lock(&nbd->config_lock);
2126         config = nbd->config;
2127         if (!test_bit(NBD_RT_BOUND, &config->runtime_flags) ||
2128             !nbd->task_recv) {
2129                 dev_err(nbd_to_dev(nbd),
2130                         "not configured, cannot reconfigure\n");
2131                 ret = -EINVAL;
2132                 goto out;
2133         }
2134
2135         ret = nbd_genl_size_set(info, nbd);
2136         if (ret)
2137                 goto out;
2138
2139         if (info->attrs[NBD_ATTR_TIMEOUT])
2140                 nbd_set_cmd_timeout(nbd,
2141                                     nla_get_u64(info->attrs[NBD_ATTR_TIMEOUT]));
2142         if (info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]) {
2143                 config->dead_conn_timeout =
2144                         nla_get_u64(info->attrs[NBD_ATTR_DEAD_CONN_TIMEOUT]);
2145                 config->dead_conn_timeout *= HZ;
2146         }
2147         if (info->attrs[NBD_ATTR_CLIENT_FLAGS]) {
2148                 u64 flags = nla_get_u64(info->attrs[NBD_ATTR_CLIENT_FLAGS]);
2149                 if (flags & NBD_CFLAG_DESTROY_ON_DISCONNECT) {
2150                         if (!test_and_set_bit(NBD_DESTROY_ON_DISCONNECT,
2151                                               &nbd->flags))
2152                                 put_dev = true;
2153                 } else {
2154                         if (test_and_clear_bit(NBD_DESTROY_ON_DISCONNECT,
2155                                                &nbd->flags))
2156                                 refcount_inc(&nbd->refs);
2157                 }
2158
2159                 if (flags & NBD_CFLAG_DISCONNECT_ON_CLOSE) {
2160                         set_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2161                                         &config->runtime_flags);
2162                 } else {
2163                         clear_bit(NBD_RT_DISCONNECT_ON_CLOSE,
2164                                         &config->runtime_flags);
2165                 }
2166         }
2167
2168         if (info->attrs[NBD_ATTR_SOCKETS]) {
2169                 struct nlattr *attr;
2170                 int rem, fd;
2171
2172                 nla_for_each_nested(attr, info->attrs[NBD_ATTR_SOCKETS],
2173                                     rem) {
2174                         struct nlattr *socks[NBD_SOCK_MAX+1];
2175
2176                         if (nla_type(attr) != NBD_SOCK_ITEM) {
2177                                 printk(KERN_ERR "nbd: socks must be embedded in a SOCK_ITEM attr\n");
2178                                 ret = -EINVAL;
2179                                 goto out;
2180                         }
2181                         ret = nla_parse_nested_deprecated(socks, NBD_SOCK_MAX,
2182                                                           attr,
2183                                                           nbd_sock_policy,
2184                                                           info->extack);
2185                         if (ret != 0) {
2186                                 printk(KERN_ERR "nbd: error processing sock list\n");
2187                                 ret = -EINVAL;
2188                                 goto out;
2189                         }
2190                         if (!socks[NBD_SOCK_FD])
2191                                 continue;
2192                         fd = (int)nla_get_u32(socks[NBD_SOCK_FD]);
2193                         ret = nbd_reconnect_socket(nbd, fd);
2194                         if (ret) {
2195                                 if (ret == -ENOSPC)
2196                                         ret = 0;
2197                                 goto out;
2198                         }
2199                         dev_info(nbd_to_dev(nbd), "reconnected socket\n");
2200                 }
2201         }
2202 out:
2203         mutex_unlock(&nbd->config_lock);
2204         nbd_config_put(nbd);
2205         nbd_put(nbd);
2206         if (put_dev)
2207                 nbd_put(nbd);
2208         return ret;
2209 }
2210
2211 static const struct genl_ops nbd_connect_genl_ops[] = {
2212         {
2213                 .cmd    = NBD_CMD_CONNECT,
2214                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2215                 .doit   = nbd_genl_connect,
2216         },
2217         {
2218                 .cmd    = NBD_CMD_DISCONNECT,
2219                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2220                 .doit   = nbd_genl_disconnect,
2221         },
2222         {
2223                 .cmd    = NBD_CMD_RECONFIGURE,
2224                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2225                 .doit   = nbd_genl_reconfigure,
2226         },
2227         {
2228                 .cmd    = NBD_CMD_STATUS,
2229                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
2230                 .doit   = nbd_genl_status,
2231         },
2232 };
2233
2234 static const struct genl_multicast_group nbd_mcast_grps[] = {
2235         { .name = NBD_GENL_MCAST_GROUP_NAME, },
2236 };
2237
2238 static struct genl_family nbd_genl_family __ro_after_init = {
2239         .hdrsize        = 0,
2240         .name           = NBD_GENL_FAMILY_NAME,
2241         .version        = NBD_GENL_VERSION,
2242         .module         = THIS_MODULE,
2243         .ops            = nbd_connect_genl_ops,
2244         .n_ops          = ARRAY_SIZE(nbd_connect_genl_ops),
2245         .maxattr        = NBD_ATTR_MAX,
2246         .policy = nbd_attr_policy,
2247         .mcgrps         = nbd_mcast_grps,
2248         .n_mcgrps       = ARRAY_SIZE(nbd_mcast_grps),
2249 };
2250
2251 static int populate_nbd_status(struct nbd_device *nbd, struct sk_buff *reply)
2252 {
2253         struct nlattr *dev_opt;
2254         u8 connected = 0;
2255         int ret;
2256
2257         /* This is a little racey, but for status it's ok.  The
2258          * reason we don't take a ref here is because we can't
2259          * take a ref in the index == -1 case as we would need
2260          * to put under the nbd_index_mutex, which could
2261          * deadlock if we are configured to remove ourselves
2262          * once we're disconnected.
2263          */
2264         if (refcount_read(&nbd->config_refs))
2265                 connected = 1;
2266         dev_opt = nla_nest_start_noflag(reply, NBD_DEVICE_ITEM);
2267         if (!dev_opt)
2268                 return -EMSGSIZE;
2269         ret = nla_put_u32(reply, NBD_DEVICE_INDEX, nbd->index);
2270         if (ret)
2271                 return -EMSGSIZE;
2272         ret = nla_put_u8(reply, NBD_DEVICE_CONNECTED,
2273                          connected);
2274         if (ret)
2275                 return -EMSGSIZE;
2276         nla_nest_end(reply, dev_opt);
2277         return 0;
2278 }
2279
2280 static int status_cb(int id, void *ptr, void *data)
2281 {
2282         struct nbd_device *nbd = ptr;
2283         return populate_nbd_status(nbd, (struct sk_buff *)data);
2284 }
2285
2286 static int nbd_genl_status(struct sk_buff *skb, struct genl_info *info)
2287 {
2288         struct nlattr *dev_list;
2289         struct sk_buff *reply;
2290         void *reply_head;
2291         size_t msg_size;
2292         int index = -1;
2293         int ret = -ENOMEM;
2294
2295         if (info->attrs[NBD_ATTR_INDEX])
2296                 index = nla_get_u32(info->attrs[NBD_ATTR_INDEX]);
2297
2298         mutex_lock(&nbd_index_mutex);
2299
2300         msg_size = nla_total_size(nla_attr_size(sizeof(u32)) +
2301                                   nla_attr_size(sizeof(u8)));
2302         msg_size *= (index == -1) ? nbd_total_devices : 1;
2303
2304         reply = genlmsg_new(msg_size, GFP_KERNEL);
2305         if (!reply)
2306                 goto out;
2307         reply_head = genlmsg_put_reply(reply, info, &nbd_genl_family, 0,
2308                                        NBD_CMD_STATUS);
2309         if (!reply_head) {
2310                 nlmsg_free(reply);
2311                 goto out;
2312         }
2313
2314         dev_list = nla_nest_start_noflag(reply, NBD_ATTR_DEVICE_LIST);
2315         if (index == -1) {
2316                 ret = idr_for_each(&nbd_index_idr, &status_cb, reply);
2317                 if (ret) {
2318                         nlmsg_free(reply);
2319                         goto out;
2320                 }
2321         } else {
2322                 struct nbd_device *nbd;
2323                 nbd = idr_find(&nbd_index_idr, index);
2324                 if (nbd) {
2325                         ret = populate_nbd_status(nbd, reply);
2326                         if (ret) {
2327                                 nlmsg_free(reply);
2328                                 goto out;
2329                         }
2330                 }
2331         }
2332         nla_nest_end(reply, dev_list);
2333         genlmsg_end(reply, reply_head);
2334         ret = genlmsg_reply(reply, info);
2335 out:
2336         mutex_unlock(&nbd_index_mutex);
2337         return ret;
2338 }
2339
2340 static void nbd_connect_reply(struct genl_info *info, int index)
2341 {
2342         struct sk_buff *skb;
2343         void *msg_head;
2344         int ret;
2345
2346         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2347         if (!skb)
2348                 return;
2349         msg_head = genlmsg_put_reply(skb, info, &nbd_genl_family, 0,
2350                                      NBD_CMD_CONNECT);
2351         if (!msg_head) {
2352                 nlmsg_free(skb);
2353                 return;
2354         }
2355         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2356         if (ret) {
2357                 nlmsg_free(skb);
2358                 return;
2359         }
2360         genlmsg_end(skb, msg_head);
2361         genlmsg_reply(skb, info);
2362 }
2363
2364 static void nbd_mcast_index(int index)
2365 {
2366         struct sk_buff *skb;
2367         void *msg_head;
2368         int ret;
2369
2370         skb = genlmsg_new(nla_total_size(sizeof(u32)), GFP_KERNEL);
2371         if (!skb)
2372                 return;
2373         msg_head = genlmsg_put(skb, 0, 0, &nbd_genl_family, 0,
2374                                      NBD_CMD_LINK_DEAD);
2375         if (!msg_head) {
2376                 nlmsg_free(skb);
2377                 return;
2378         }
2379         ret = nla_put_u32(skb, NBD_ATTR_INDEX, index);
2380         if (ret) {
2381                 nlmsg_free(skb);
2382                 return;
2383         }
2384         genlmsg_end(skb, msg_head);
2385         genlmsg_multicast(&nbd_genl_family, skb, 0, 0, GFP_KERNEL);
2386 }
2387
2388 static void nbd_dead_link_work(struct work_struct *work)
2389 {
2390         struct link_dead_args *args = container_of(work, struct link_dead_args,
2391                                                    work);
2392         nbd_mcast_index(args->index);
2393         kfree(args);
2394 }
2395
2396 static int __init nbd_init(void)
2397 {
2398         int i;
2399
2400         BUILD_BUG_ON(sizeof(struct nbd_request) != 28);
2401
2402         if (max_part < 0) {
2403                 printk(KERN_ERR "nbd: max_part must be >= 0\n");
2404                 return -EINVAL;
2405         }
2406
2407         part_shift = 0;
2408         if (max_part > 0) {
2409                 part_shift = fls(max_part);
2410
2411                 /*
2412                  * Adjust max_part according to part_shift as it is exported
2413                  * to user space so that user can know the max number of
2414                  * partition kernel should be able to manage.
2415                  *
2416                  * Note that -1 is required because partition 0 is reserved
2417                  * for the whole disk.
2418                  */
2419                 max_part = (1UL << part_shift) - 1;
2420         }
2421
2422         if ((1UL << part_shift) > DISK_MAX_PARTS)
2423                 return -EINVAL;
2424
2425         if (nbds_max > 1UL << (MINORBITS - part_shift))
2426                 return -EINVAL;
2427
2428         if (register_blkdev(NBD_MAJOR, "nbd"))
2429                 return -EIO;
2430
2431         if (genl_register_family(&nbd_genl_family)) {
2432                 unregister_blkdev(NBD_MAJOR, "nbd");
2433                 return -EINVAL;
2434         }
2435         nbd_dbg_init();
2436
2437         mutex_lock(&nbd_index_mutex);
2438         for (i = 0; i < nbds_max; i++)
2439                 nbd_dev_add(i);
2440         mutex_unlock(&nbd_index_mutex);
2441         return 0;
2442 }
2443
2444 static int nbd_exit_cb(int id, void *ptr, void *data)
2445 {
2446         struct list_head *list = (struct list_head *)data;
2447         struct nbd_device *nbd = ptr;
2448
2449         list_add_tail(&nbd->list, list);
2450         return 0;
2451 }
2452
2453 static void __exit nbd_cleanup(void)
2454 {
2455         struct nbd_device *nbd;
2456         LIST_HEAD(del_list);
2457
2458         /*
2459          * Unregister netlink interface prior to waiting
2460          * for the completion of netlink commands.
2461          */
2462         genl_unregister_family(&nbd_genl_family);
2463
2464         nbd_dbg_close();
2465
2466         mutex_lock(&nbd_index_mutex);
2467         idr_for_each(&nbd_index_idr, &nbd_exit_cb, &del_list);
2468         mutex_unlock(&nbd_index_mutex);
2469
2470         while (!list_empty(&del_list)) {
2471                 nbd = list_first_entry(&del_list, struct nbd_device, list);
2472                 list_del_init(&nbd->list);
2473                 if (refcount_read(&nbd->config_refs))
2474                         printk(KERN_ERR "nbd: possibly leaking nbd_config (ref %d)\n",
2475                                         refcount_read(&nbd->config_refs));
2476                 if (refcount_read(&nbd->refs) != 1)
2477                         printk(KERN_ERR "nbd: possibly leaking a device\n");
2478                 nbd_put(nbd);
2479         }
2480
2481         idr_destroy(&nbd_index_idr);
2482         unregister_blkdev(NBD_MAJOR, "nbd");
2483 }
2484
2485 module_init(nbd_init);
2486 module_exit(nbd_cleanup);
2487
2488 MODULE_DESCRIPTION("Network Block Device");
2489 MODULE_LICENSE("GPL");
2490
2491 module_param(nbds_max, int, 0444);
2492 MODULE_PARM_DESC(nbds_max, "number of network block devices to initialize (default: 16)");
2493 module_param(max_part, int, 0444);
2494 MODULE_PARM_DESC(max_part, "number of partitions per device (default: 16)");